Home | History | Annotate | Line # | Download | only in aarch64
      1 ;; Machine description for AArch64 SVE.
      2 ;; Copyright (C) 2009-2022 Free Software Foundation, Inc.
      3 ;; Contributed by ARM Ltd.
      4 ;;
      5 ;; This file is part of GCC.
      6 ;;
      7 ;; GCC is free software; you can redistribute it and/or modify it
      8 ;; under the terms of the GNU General Public License as published by
      9 ;; the Free Software Foundation; either version 3, or (at your option)
     10 ;; any later version.
     11 ;;
     12 ;; GCC is distributed in the hope that it will be useful, but
     13 ;; WITHOUT ANY WARRANTY; without even the implied warranty of
     14 ;; MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
     15 ;; General Public License for more details.
     16 ;;
     17 ;; You should have received a copy of the GNU General Public License
     18 ;; along with GCC; see the file COPYING3.  If not see
     19 ;; <http://www.gnu.org/licenses/>.
     20 
     21 ;; The file is organised into the following sections (search for the full
     22 ;; line):
     23 ;;
     24 ;; == General notes
     25 ;; ---- Note on the handling of big-endian SVE
     26 ;; ---- Description of UNSPEC_PTEST
     27 ;; ---- Description of UNSPEC_PRED_Z
     28 ;; ---- Note on predicated integer arithemtic and UNSPEC_PRED_X
     29 ;; ---- Note on predicated FP arithmetic patterns and GP "strictness"
     30 ;; ---- Note on FFR handling
     31 ;;
     32 ;; == Moves
     33 ;; ---- Moves of single vectors
     34 ;; ---- Moves of multiple vectors
     35 ;; ---- Moves of predicates
     36 ;; ---- Moves relating to the FFR
     37 ;;
     38 ;; == Loads
     39 ;; ---- Normal contiguous loads
     40 ;; ---- Extending contiguous loads
     41 ;; ---- First-faulting contiguous loads
     42 ;; ---- First-faulting extending contiguous loads
     43 ;; ---- Non-temporal contiguous loads
     44 ;; ---- Normal gather loads
     45 ;; ---- Extending gather loads
     46 ;; ---- First-faulting gather loads
     47 ;; ---- First-faulting extending gather loads
     48 ;;
     49 ;; == Prefetches
     50 ;; ---- Contiguous prefetches
     51 ;; ---- Gather prefetches
     52 ;;
     53 ;; == Stores
     54 ;; ---- Normal contiguous stores
     55 ;; ---- Truncating contiguous stores
     56 ;; ---- Non-temporal contiguous stores
     57 ;; ---- Normal scatter stores
     58 ;; ---- Truncating scatter stores
     59 ;;
     60 ;; == Vector creation
     61 ;; ---- [INT,FP] Duplicate element
     62 ;; ---- [INT,FP] Initialize from individual elements
     63 ;; ---- [INT] Linear series
     64 ;; ---- [PRED] Duplicate element
     65 ;;
     66 ;; == Vector decomposition
     67 ;; ---- [INT,FP] Extract index
     68 ;; ---- [INT,FP] Extract active element
     69 ;; ---- [PRED] Extract index
     70 ;;
     71 ;; == Unary arithmetic
     72 ;; ---- [INT] General unary arithmetic corresponding to rtx codes
     73 ;; ---- [INT] General unary arithmetic corresponding to unspecs
     74 ;; ---- [INT] Sign and zero extension
     75 ;; ---- [INT] Truncation
     76 ;; ---- [INT] Logical inverse
     77 ;; ---- [FP<-INT] General unary arithmetic that maps to unspecs
     78 ;; ---- [FP] General unary arithmetic corresponding to unspecs
     79 ;; ---- [FP] Square root
     80 ;; ---- [FP] Reciprocal square root
     81 ;; ---- [PRED] Inverse
     82 
     83 ;; == Binary arithmetic
     84 ;; ---- [INT] General binary arithmetic corresponding to rtx codes
     85 ;; ---- [INT] Addition
     86 ;; ---- [INT] Subtraction
     87 ;; ---- [INT] Take address
     88 ;; ---- [INT] Absolute difference
     89 ;; ---- [INT] Saturating addition and subtraction
     90 ;; ---- [INT] Highpart multiplication
     91 ;; ---- [INT] Division
     92 ;; ---- [INT] Binary logical operations
     93 ;; ---- [INT] Binary logical operations (inverted second input)
     94 ;; ---- [INT] Shifts (rounding towards -Inf)
     95 ;; ---- [INT] Shifts (rounding towards 0)
     96 ;; ---- [FP<-INT] General binary arithmetic corresponding to unspecs
     97 ;; ---- [FP] General binary arithmetic corresponding to rtx codes
     98 ;; ---- [FP] General binary arithmetic corresponding to unspecs
     99 ;; ---- [FP] Addition
    100 ;; ---- [FP] Complex addition
    101 ;; ---- [FP] Subtraction
    102 ;; ---- [FP] Absolute difference
    103 ;; ---- [FP] Multiplication
    104 ;; ---- [FP] Division
    105 ;; ---- [FP] Binary logical operations
    106 ;; ---- [FP] Sign copying
    107 ;; ---- [FP] Maximum and minimum
    108 ;; ---- [PRED] Binary logical operations
    109 ;; ---- [PRED] Binary logical operations (inverted second input)
    110 ;; ---- [PRED] Binary logical operations (inverted result)
    111 ;;
    112 ;; == Ternary arithmetic
    113 ;; ---- [INT] MLA and MAD
    114 ;; ---- [INT] MLS and MSB
    115 ;; ---- [INT] Dot product
    116 ;; ---- [INT] Sum of absolute differences
    117 ;; ---- [INT] Matrix multiply-accumulate
    118 ;; ---- [FP] General ternary arithmetic corresponding to unspecs
    119 ;; ---- [FP] Complex multiply-add
    120 ;; ---- [FP] Trigonometric multiply-add
    121 ;; ---- [FP] Bfloat16 long ternary arithmetic (SF,BF,BF)
    122 ;; ---- [FP] Matrix multiply-accumulate
    123 ;;
    124 ;; == Comparisons and selects
    125 ;; ---- [INT,FP] Select based on predicates
    126 ;; ---- [INT,FP] Compare and select
    127 ;; ---- [INT] Comparisons
    128 ;; ---- [INT] While tests
    129 ;; ---- [FP] Direct comparisons
    130 ;; ---- [FP] Absolute comparisons
    131 ;; ---- [PRED] Select
    132 ;; ---- [PRED] Test bits
    133 ;;
    134 ;; == Reductions
    135 ;; ---- [INT,FP] Conditional reductions
    136 ;; ---- [INT] Tree reductions
    137 ;; ---- [FP] Tree reductions
    138 ;; ---- [FP] Left-to-right reductions
    139 ;;
    140 ;; == Permutes
    141 ;; ---- [INT,FP] General permutes
    142 ;; ---- [INT,FP] Special-purpose unary permutes
    143 ;; ---- [INT,FP] Special-purpose binary permutes
    144 ;; ---- [PRED] Special-purpose unary permutes
    145 ;; ---- [PRED] Special-purpose binary permutes
    146 ;;
    147 ;; == Conversions
    148 ;; ---- [INT<-INT] Packs
    149 ;; ---- [INT<-INT] Unpacks
    150 ;; ---- [INT<-FP] Conversions
    151 ;; ---- [INT<-FP] Packs
    152 ;; ---- [INT<-FP] Unpacks
    153 ;; ---- [FP<-INT] Conversions
    154 ;; ---- [FP<-INT] Packs
    155 ;; ---- [FP<-INT] Unpacks
    156 ;; ---- [FP<-FP] Packs
    157 ;; ---- [FP<-FP] Packs (bfloat16)
    158 ;; ---- [FP<-FP] Unpacks
    159 ;; ---- [PRED<-PRED] Packs
    160 ;; ---- [PRED<-PRED] Unpacks
    161 ;;
    162 ;; == Vector partitioning
    163 ;; ---- [PRED] Unary partitioning
    164 ;; ---- [PRED] Binary partitioning
    165 ;; ---- [PRED] Scalarization
    166 ;;
    167 ;; == Counting elements
    168 ;; ---- [INT] Count elements in a pattern (scalar)
    169 ;; ---- [INT] Increment by the number of elements in a pattern (scalar)
    170 ;; ---- [INT] Increment by the number of elements in a pattern (vector)
    171 ;; ---- [INT] Decrement by the number of elements in a pattern (scalar)
    172 ;; ---- [INT] Decrement by the number of elements in a pattern (vector)
    173 ;; ---- [INT] Count elements in a predicate (scalar)
    174 ;; ---- [INT] Increment by the number of elements in a predicate (scalar)
    175 ;; ---- [INT] Increment by the number of elements in a predicate (vector)
    176 ;; ---- [INT] Decrement by the number of elements in a predicate (scalar)
    177 ;; ---- [INT] Decrement by the number of elements in a predicate (vector)
    178 
    179 ;; =========================================================================
    180 ;; == General notes
    181 ;; =========================================================================
    182 ;;
    183 ;; -------------------------------------------------------------------------
    184 ;; ---- Note on the handling of big-endian SVE
    185 ;; -------------------------------------------------------------------------
    186 ;;
    187 ;; On big-endian systems, Advanced SIMD mov<mode> patterns act in the
    188 ;; same way as movdi or movti would: the first byte of memory goes
    189 ;; into the most significant byte of the register and the last byte
    190 ;; of memory goes into the least significant byte of the register.
    191 ;; This is the most natural ordering for Advanced SIMD and matches
    192 ;; the ABI layout for 64-bit and 128-bit vector types.
    193 ;;
    194 ;; As a result, the order of bytes within the register is what GCC
    195 ;; expects for a big-endian target, and subreg offsets therefore work
    196 ;; as expected, with the first element in memory having subreg offset 0
    197 ;; and the last element in memory having the subreg offset associated
    198 ;; with a big-endian lowpart.  However, this ordering also means that
    199 ;; GCC's lane numbering does not match the architecture's numbering:
    200 ;; GCC always treats the element at the lowest address in memory
    201 ;; (subreg offset 0) as element 0, while the architecture treats
    202 ;; the least significant end of the register as element 0.
    203 ;;
    204 ;; The situation for SVE is different.  We want the layout of the
    205 ;; SVE register to be same for mov<mode> as it is for maskload<mode>:
    206 ;; logically, a mov<mode> load must be indistinguishable from a
    207 ;; maskload<mode> whose mask is all true.  We therefore need the
    208 ;; register layout to match LD1 rather than LDR.  The ABI layout of
    209 ;; SVE types also matches LD1 byte ordering rather than LDR byte ordering.
    210 ;;
    211 ;; As a result, the architecture lane numbering matches GCC's lane
    212 ;; numbering, with element 0 always being the first in memory.
    213 ;; However:
    214 ;;
    215 ;; - Applying a subreg offset to a register does not give the element
    216 ;;   that GCC expects: the first element in memory has the subreg offset
    217 ;;   associated with a big-endian lowpart while the last element in memory
    218 ;;   has subreg offset 0.  We handle this via TARGET_CAN_CHANGE_MODE_CLASS.
    219 ;;
    220 ;; - We cannot use LDR and STR for spill slots that might be accessed
    221 ;;   via subregs, since although the elements have the order GCC expects,
    222 ;;   the order of the bytes within the elements is different.  We instead
    223 ;;   access spill slots via LD1 and ST1, using secondary reloads to
    224 ;;   reserve a predicate register.
    225 ;;
    226 ;; -------------------------------------------------------------------------
    227 ;; ---- Description of UNSPEC_PTEST
    228 ;; -------------------------------------------------------------------------
    229 ;;
    230 ;; SVE provides a PTEST instruction for testing the active lanes of a
    231 ;; predicate and setting the flags based on the result.  The associated
    232 ;; condition code tests are:
    233 ;;
    234 ;; - any   (= ne): at least one active bit is set
    235 ;; - none  (= eq): all active bits are clear (*)
    236 ;; - first (= mi): the first active bit is set
    237 ;; - nfrst (= pl): the first active bit is clear (*)
    238 ;; - last  (= cc): the last active bit is set
    239 ;; - nlast (= cs): the last active bit is clear (*)
    240 ;;
    241 ;; where the conditions marked (*) are also true when there are no active
    242 ;; lanes (i.e. when the governing predicate is a PFALSE).  The flags results
    243 ;; of a PTEST use the condition code mode CC_NZC.
    244 ;;
    245 ;; PTEST is always a .B operation (i.e. it always operates on VNx16BI).
    246 ;; This means that for other predicate modes, we need a governing predicate
    247 ;; in which all bits are defined.
    248 ;;
    249 ;; For example, most predicated .H operations ignore the odd bits of the
    250 ;; governing predicate, so that an active lane is represented by the
    251 ;; bits "1x" and an inactive lane by the bits "0x", where "x" can be
    252 ;; any value.  To test a .H predicate, we instead need "10" and "00"
    253 ;; respectively, so that the condition only tests the even bits of the
    254 ;; predicate.
    255 ;;
    256 ;; Several instructions set the flags as a side-effect, in the same way
    257 ;; that a separate PTEST would.  It's important for code quality that we
    258 ;; use these flags results as often as possible, particularly in the case
    259 ;; of WHILE* and RDFFR.
    260 ;;
    261 ;; Also, some of the instructions that set the flags are unpredicated
    262 ;; and instead implicitly test all .B, .H, .S or .D elements, as though
    263 ;; they were predicated on a PTRUE of that size.  For example, a .S
    264 ;; WHILELO sets the flags in the same way as a PTEST with a .S PTRUE
    265 ;; would.
    266 ;;
    267 ;; We therefore need to represent PTEST operations in a way that
    268 ;; makes it easy to combine them with both predicated and unpredicated
    269 ;; operations, while using a VNx16BI governing predicate for all
    270 ;; predicate modes.  We do this using:
    271 ;;
    272 ;;   (unspec:CC_NZC [gp cast_gp ptrue_flag op] UNSPEC_PTEST)
    273 ;;
    274 ;; where:
    275 ;;
    276 ;; - GP is the real VNx16BI governing predicate
    277 ;;
    278 ;; - CAST_GP is GP cast to the mode of OP.  All bits dropped by casting
    279 ;;   GP to CAST_GP are guaranteed to be clear in GP.
    280 ;;
    281 ;; - PTRUE_FLAG is a CONST_INT (conceptually of mode SI) that has the value
    282 ;;   SVE_KNOWN_PTRUE if we know that CAST_GP (rather than GP) is all-true and
    283 ;;   SVE_MAYBE_NOT_PTRUE otherwise.
    284 ;;
    285 ;; - OP is the predicate we want to test, of the same mode as CAST_GP.
    286 ;;
    287 ;; -------------------------------------------------------------------------
    288 ;; ---- Description of UNSPEC_PRED_Z
    289 ;; -------------------------------------------------------------------------
    290 ;;
    291 ;; SVE integer comparisons are predicated and return zero for inactive
    292 ;; lanes.  Sometimes we use them with predicates that are all-true and
    293 ;; sometimes we use them with general predicates.
    294 ;;
    295 ;; The integer comparisons also set the flags and so build-in the effect
    296 ;; of a PTEST.  We therefore want to be able to combine integer comparison
    297 ;; patterns with PTESTs of the result.  One difficulty with doing this is
    298 ;; that (as noted above) the PTEST is always a .B operation and so can place
    299 ;; stronger requirements on the governing predicate than the comparison does.
    300 ;;
    301 ;; For example, when applying a separate PTEST to the result of a full-vector
    302 ;; .H comparison, the PTEST must be predicated on a .H PTRUE instead of a
    303 ;; .B PTRUE.  In constrast, the comparison might be predicated on either
    304 ;; a .H PTRUE or a .B PTRUE, since the values of odd-indexed predicate
    305 ;; bits don't matter for .H operations.
    306 ;;
    307 ;; We therefore can't rely on a full-vector comparison using the same
    308 ;; predicate register as a following PTEST.  We instead need to remember
    309 ;; whether a comparison is known to be a full-vector comparison and use
    310 ;; this information in addition to a check for equal predicate registers.
    311 ;; At the same time, it's useful to have a common representation for all
    312 ;; integer comparisons, so that they can be handled by a single set of
    313 ;; patterns.
    314 ;;
    315 ;; We therefore take a similar approach to UNSPEC_PTEST above and use:
    316 ;;
    317 ;;   (unspec:<M:VPRED> [gp ptrue_flag (code:M op0 op1)] UNSPEC_PRED_Z)
    318 ;;
    319 ;; where:
    320 ;;
    321 ;; - GP is the governing predicate, of mode <M:VPRED>
    322 ;;
    323 ;; - PTRUE_FLAG is a CONST_INT (conceptually of mode SI) that has the value
    324 ;;   SVE_KNOWN_PTRUE if we know that GP is all-true and SVE_MAYBE_NOT_PTRUE
    325 ;;   otherwise
    326 ;;
    327 ;; - CODE is the comparison code
    328 ;;
    329 ;; - OP0 and OP1 are the values being compared, of mode M
    330 ;;
    331 ;; The "Z" in UNSPEC_PRED_Z indicates that inactive lanes are zero.
    332 ;;
    333 ;; -------------------------------------------------------------------------
    334 ;; ---- Note on predicated integer arithemtic and UNSPEC_PRED_X
    335 ;; -------------------------------------------------------------------------
    336 ;;
    337 ;; Many SVE integer operations are predicated.  We can generate them
    338 ;; from four sources:
    339 ;;
    340 ;; (1) Using normal unpredicated optabs.  In this case we need to create
    341 ;;     an all-true predicate register to act as the governing predicate
    342 ;;     for the SVE instruction.  There are no inactive lanes, and thus
    343 ;;     the values of inactive lanes don't matter.
    344 ;;
    345 ;; (2) Using _x ACLE functions.  In this case the function provides a
    346 ;;     specific predicate and some lanes might be inactive.  However,
    347 ;;     as for (1), the values of the inactive lanes don't matter.
    348 ;;     We can make extra lanes active without changing the behavior
    349 ;;     (although for code-quality reasons we should avoid doing so
    350 ;;     needlessly).
    351 ;;
    352 ;; (3) Using cond_* optabs that correspond to IFN_COND_* internal functions.
    353 ;;     These optabs have a predicate operand that specifies which lanes are
    354 ;;     active and another operand that provides the values of inactive lanes.
    355 ;;
    356 ;; (4) Using _m and _z ACLE functions.  These functions map to the same
    357 ;;     patterns as (3), with the _z functions setting inactive lanes to zero
    358 ;;     and the _m functions setting the inactive lanes to one of the function
    359 ;;     arguments.
    360 ;;
    361 ;; For (1) and (2) we need a way of attaching the predicate to a normal
    362 ;; unpredicated integer operation.  We do this using:
    363 ;;
    364 ;;   (unspec:M [pred (code:M (op0 op1 ...))] UNSPEC_PRED_X)
    365 ;;
    366 ;; where (code:M (op0 op1 ...)) is the normal integer operation and PRED
    367 ;; is a predicate of mode <M:VPRED>.  PRED might or might not be a PTRUE;
    368 ;; it always is for (1), but might not be for (2).
    369 ;;
    370 ;; The unspec as a whole has the same value as (code:M ...) when PRED is
    371 ;; all-true.  It is always semantically valid to replace PRED with a PTRUE,
    372 ;; but as noted above, we should only do so if there's a specific benefit.
    373 ;;
    374 ;; (The "_X" in the unspec is named after the ACLE functions in (2).)
    375 ;;
    376 ;; For (3) and (4) we can simply use the SVE port's normal representation
    377 ;; of a predicate-based select:
    378 ;;
    379 ;;   (unspec:M [pred (code:M (op0 op1 ...)) inactive] UNSPEC_SEL)
    380 ;;
    381 ;; where INACTIVE specifies the values of inactive lanes.
    382 ;;
    383 ;; We can also use the UNSPEC_PRED_X wrapper in the UNSPEC_SEL rather
    384 ;; than inserting the integer operation directly.  This is mostly useful
    385 ;; if we want the combine pass to merge an integer operation with an explicit
    386 ;; vcond_mask (in other words, with a following SEL instruction).  However,
    387 ;; it's generally better to merge such operations at the gimple level
    388 ;; using (3).
    389 ;;
    390 ;; -------------------------------------------------------------------------
    391 ;; ---- Note on predicated FP arithmetic patterns and GP "strictness"
    392 ;; -------------------------------------------------------------------------
    393 ;;
    394 ;; Most SVE floating-point operations are predicated.  We can generate
    395 ;; them from four sources:
    396 ;;
    397 ;; (1) Using normal unpredicated optabs.  In this case we need to create
    398 ;;     an all-true predicate register to act as the governing predicate
    399 ;;     for the SVE instruction.  There are no inactive lanes, and thus
    400 ;;     the values of inactive lanes don't matter.
    401 ;;
    402 ;; (2) Using _x ACLE functions.  In this case the function provides a
    403 ;;     specific predicate and some lanes might be inactive.  However,
    404 ;;     as for (1), the values of the inactive lanes don't matter.
    405 ;;
    406 ;;     The instruction must have the same exception behavior as the
    407 ;;     function call unless things like command-line flags specifically
    408 ;;     allow otherwise.  For example, with -ffast-math, it is OK to
    409 ;;     raise exceptions for inactive lanes, but normally it isn't.
    410 ;;
    411 ;; (3) Using cond_* optabs that correspond to IFN_COND_* internal functions.
    412 ;;     These optabs have a predicate operand that specifies which lanes are
    413 ;;     active and another operand that provides the values of inactive lanes.
    414 ;;
    415 ;; (4) Using _m and _z ACLE functions.  These functions map to the same
    416 ;;     patterns as (3), with the _z functions setting inactive lanes to zero
    417 ;;     and the _m functions setting the inactive lanes to one of the function
    418 ;;     arguments.
    419 ;;
    420 ;; So:
    421 ;;
    422 ;; - In (1), the predicate is known to be all true and the pattern can use
    423 ;;   unpredicated operations where available.
    424 ;;
    425 ;; - In (2), the predicate might or might not be all true.  The pattern can
    426 ;;   use unpredicated instructions if the predicate is all-true or if things
    427 ;;   like command-line flags allow exceptions for inactive lanes.
    428 ;;
    429 ;; - (3) and (4) represent a native SVE predicated operation.  Some lanes
    430 ;;   might be inactive and inactive lanes of the result must have specific
    431 ;;   values.  There is no scope for using unpredicated instructions (and no
    432 ;;   reason to want to), so the question about command-line flags doesn't
    433 ;;   arise.
    434 ;;
    435 ;; It would be inaccurate to model (2) as an rtx code like (sqrt ...)
    436 ;; in combination with a separate predicate operand, e.g.
    437 ;;
    438 ;;   (unspec [(match_operand:<VPRED> 1 "register_operand" "Upl")
    439 ;;	      (sqrt:SVE_FULL_F 2 "register_operand" "w")]
    440 ;;	     ....)
    441 ;;
    442 ;; because (sqrt ...) can raise an exception for any lane, including
    443 ;; inactive ones.  We therefore need to use an unspec instead.
    444 ;;
    445 ;; Also, (2) requires some way of distinguishing the case in which the
    446 ;; predicate might have inactive lanes and cannot be changed from the
    447 ;; case in which the predicate has no inactive lanes or can be changed.
    448 ;; This information is also useful when matching combined FP patterns
    449 ;; in which the predicates might not be equal.
    450 ;;
    451 ;; We therefore model FP operations as an unspec of the form:
    452 ;;
    453 ;;   (unspec [pred strictness op0 op1 ...] UNSPEC_COND_<MNEMONIC>)
    454 ;;
    455 ;; where:
    456 ;;
    457 ;; - PRED is the governing predicate.
    458 ;;
    459 ;; - STRICTNESS is a CONST_INT that conceptually has mode SI.  It has the
    460 ;;   value SVE_STRICT_GP if PRED might have inactive lanes and if those
    461 ;;   lanes must remain inactive.  It has the value SVE_RELAXED_GP otherwise.
    462 ;;
    463 ;; - OP0 OP1 ... are the normal input operands to the operation.
    464 ;;
    465 ;; - MNEMONIC is the mnemonic of the associated SVE instruction.
    466 ;;
    467 ;; For (3) and (4), we combine these operations with an UNSPEC_SEL
    468 ;; that selects between the result of the FP operation and the "else"
    469 ;; value.  (This else value is a merge input for _m ACLE functions
    470 ;; and zero for _z ACLE functions.)  The outer pattern then has the form:
    471 ;;
    472 ;;   (unspec [pred fp_operation else_value] UNSPEC_SEL)
    473 ;;
    474 ;; This means that the patterns for (3) and (4) have two predicates:
    475 ;; one for the FP operation itself and one for the UNSPEC_SEL.
    476 ;; This pattern is equivalent to the result of combining an instance
    477 ;; of (1) or (2) with a separate vcond instruction, so these patterns
    478 ;; are useful as combine targets too.
    479 ;;
    480 ;; However, in the combine case, the instructions that we want to
    481 ;; combine might use different predicates.  Then:
    482 ;;
    483 ;; - Some of the active lanes of the FP operation might be discarded
    484 ;;   by the UNSPEC_SEL.  It's OK to drop the FP operation on those lanes,
    485 ;;   even for SVE_STRICT_GP, since the operations on those lanes are
    486 ;;   effectively dead code.
    487 ;;
    488 ;; - Some of the inactive lanes of the FP operation might be selected
    489 ;;   by the UNSPEC_SEL, giving unspecified values for those lanes.
    490 ;;   SVE_RELAXED_GP lets us extend the FP operation to cover these
    491 ;;   extra lanes, but SVE_STRICT_GP does not.
    492 ;;
    493 ;; Thus SVE_RELAXED_GP allows us to ignore the predicate on the FP operation
    494 ;; and operate on exactly the lanes selected by the UNSPEC_SEL predicate.
    495 ;; This typically leads to patterns like:
    496 ;;
    497 ;;    (unspec [(match_operand 1 "register_operand" "Upl")
    498 ;;             (unspec [(match_operand N)
    499 ;;                      (const_int SVE_RELAXED_GP)
    500 ;;                      ...]
    501 ;;                     UNSPEC_COND_<MNEMONIC>)
    502 ;;             ...])
    503 ;;
    504 ;; where operand N is allowed to be anything.  These instructions then
    505 ;; have rewrite rules to replace operand N with operand 1, which gives the
    506 ;; instructions a canonical form and means that the original operand N is
    507 ;; not kept live unnecessarily.
    508 ;;
    509 ;; In contrast, SVE_STRICT_GP only allows the UNSPEC_SEL predicate to be
    510 ;; a subset of the FP operation predicate.  This case isn't interesting
    511 ;; for FP operations that have an all-true predicate, since such operations
    512 ;; use SVE_RELAXED_GP instead.  And it is not possible for instruction
    513 ;; conditions to track the subset relationship for arbitrary registers.
    514 ;; So in practice, the only useful case for SVE_STRICT_GP is the one
    515 ;; in which the predicates match:
    516 ;;
    517 ;;    (unspec [(match_operand 1 "register_operand" "Upl")
    518 ;;             (unspec [(match_dup 1)
    519 ;;                      (const_int SVE_STRICT_GP)
    520 ;;                      ...]
    521 ;;                     UNSPEC_COND_<MNEMONIC>)
    522 ;;             ...])
    523 ;;
    524 ;; This pattern would also be correct for SVE_RELAXED_GP, but it would
    525 ;; be redundant with the one above.  However, if the combine pattern
    526 ;; has multiple FP operations, using a match_operand allows combinations
    527 ;; of SVE_STRICT_GP and SVE_RELAXED_GP in the same operation, provided
    528 ;; that the predicates are the same:
    529 ;;
    530 ;;    (unspec [(match_operand 1 "register_operand" "Upl")
    531 ;;             (...
    532 ;;                (unspec [(match_dup 1)
    533 ;;                         (match_operand:SI N "aarch64_sve_gp_strictness")
    534 ;;                         ...]
    535 ;;                        UNSPEC_COND_<MNEMONIC1>)
    536 ;;                (unspec [(match_dup 1)
    537 ;;                         (match_operand:SI M "aarch64_sve_gp_strictness")
    538 ;;                         ...]
    539 ;;                        UNSPEC_COND_<MNEMONIC2>) ...)
    540 ;;             ...])
    541 ;;
    542 ;; The fully-relaxed version of this pattern is:
    543 ;;
    544 ;;    (unspec [(match_operand 1 "register_operand" "Upl")
    545 ;;             (...
    546 ;;                (unspec [(match_operand:SI N)
    547 ;;                         (const_int SVE_RELAXED_GP)
    548 ;;                         ...]
    549 ;;                        UNSPEC_COND_<MNEMONIC1>)
    550 ;;                (unspec [(match_operand:SI M)
    551 ;;                         (const_int SVE_RELAXED_GP)
    552 ;;                         ...]
    553 ;;                        UNSPEC_COND_<MNEMONIC2>) ...)
    554 ;;             ...])
    555 ;;
    556 ;; -------------------------------------------------------------------------
    557 ;; ---- Note on FFR handling
    558 ;; -------------------------------------------------------------------------
    559 ;;
    560 ;; Logically we want to divide FFR-related instructions into regions
    561 ;; that contain exactly one of:
    562 ;;
    563 ;; - a single write to the FFR
    564 ;; - any number of reads from the FFR (but only one read is likely)
    565 ;; - any number of LDFF1 and LDNF1 instructions
    566 ;;
    567 ;; However, LDFF1 and LDNF1 instructions should otherwise behave like
    568 ;; normal loads as far as possible.  This means that they should be
    569 ;; schedulable within a region in the same way that LD1 would be,
    570 ;; and they should be deleted as dead if the result is unused.  The loads
    571 ;; should therefore not write to the FFR, since that would both serialize
    572 ;; the loads with respect to each other and keep the loads live for any
    573 ;; later RDFFR.
    574 ;;
    575 ;; We get around this by using a fake "FFR token" (FFRT) to help describe
    576 ;; the dependencies.  Writing to the FFRT starts a new "FFRT region",
    577 ;; while using the FFRT keeps the instruction within its region.
    578 ;; Specifically:
    579 ;;
    580 ;; - Writes start a new FFRT region as well as setting the FFR:
    581 ;;
    582 ;;       W1: parallel (FFRT = <new value>, FFR = <actual FFR value>)
    583 ;;
    584 ;; - Loads use an LD1-like instruction that also uses the FFRT, so that the
    585 ;;   loads stay within the same FFRT region:
    586 ;;
    587 ;;       L1: load data while using the FFRT
    588 ;;
    589 ;;   In addition, any FFRT region that includes a load also has at least one
    590 ;;   instance of:
    591 ;;
    592 ;;       L2: FFR = update(FFR, FFRT)  [type == no_insn]
    593 ;;
    594 ;;   to make it clear that the region both reads from and writes to the FFR.
    595 ;;
    596 ;; - Reads do the following:
    597 ;;
    598 ;;       R1: FFRT = FFR               [type == no_insn]
    599 ;;       R2: read from the FFRT
    600 ;;       R3: FFRT = update(FFRT)      [type == no_insn]
    601 ;;
    602 ;;   R1 and R3 both create new FFRT regions, so that previous LDFF1s and
    603 ;;   LDNF1s cannot move forwards across R1 and later LDFF1s and LDNF1s
    604 ;;   cannot move backwards across R3.
    605 ;;
    606 ;; This way, writes are only kept alive by later loads or reads,
    607 ;; and write/read pairs fold normally.  For two consecutive reads,
    608 ;; the first R3 is made dead by the second R1, which in turn becomes
    609 ;; redundant with the first R1.  We then have:
    610 ;;
    611 ;;     first R1: FFRT = FFR
    612 ;;     first read from the FFRT
    613 ;;     second read from the FFRT
    614 ;;     second R3: FFRT = update(FFRT)
    615 ;;
    616 ;; i.e. the two FFRT regions collapse into a single one with two
    617 ;; independent reads.
    618 ;;
    619 ;; The model still prevents some valid optimizations though.  For example,
    620 ;; if all loads in an FFRT region are deleted as dead, nothing would remove
    621 ;; the L2 instructions.
    622 
    623 ;; =========================================================================
    624 ;; == Moves
    625 ;; =========================================================================
    626 
    627 ;; -------------------------------------------------------------------------
    628 ;; ---- Moves of single vectors
    629 ;; -------------------------------------------------------------------------
    630 ;; Includes:
    631 ;; - MOV  (including aliases)
    632 ;; - LD1B (contiguous form)
    633 ;; - LD1D (    "    "     )
    634 ;; - LD1H (    "    "     )
    635 ;; - LD1W (    "    "     )
    636 ;; - LDR
    637 ;; - ST1B (contiguous form)
    638 ;; - ST1D (    "    "     )
    639 ;; - ST1H (    "    "     )
    640 ;; - ST1W (    "    "     )
    641 ;; - STR
    642 ;; -------------------------------------------------------------------------
    643 
    644 (define_expand "mov<mode>"
    645   [(set (match_operand:SVE_ALL 0 "nonimmediate_operand")
    646 	(match_operand:SVE_ALL 1 "general_operand"))]
    647   "TARGET_SVE"
    648   {
    649     /* Use the predicated load and store patterns where possible.
    650        This is required for big-endian targets (see the comment at the
    651        head of the file) and increases the addressing choices for
    652        little-endian.  */
    653     if ((MEM_P (operands[0]) || MEM_P (operands[1]))
    654 	&& can_create_pseudo_p ())
    655       {
    656 	aarch64_expand_sve_mem_move (operands[0], operands[1], <VPRED>mode);
    657 	DONE;
    658       }
    659 
    660     if (CONSTANT_P (operands[1]))
    661       {
    662 	aarch64_expand_mov_immediate (operands[0], operands[1]);
    663 	DONE;
    664       }
    665 
    666     /* Optimize subregs on big-endian targets: we can use REV[BHW]
    667        instead of going through memory.  */
    668     if (BYTES_BIG_ENDIAN
    669 	&& aarch64_maybe_expand_sve_subreg_move (operands[0], operands[1]))
    670       DONE;
    671   }
    672 )
    673 
    674 (define_expand "movmisalign<mode>"
    675   [(set (match_operand:SVE_ALL 0 "nonimmediate_operand")
    676 	(match_operand:SVE_ALL 1 "general_operand"))]
    677   "TARGET_SVE"
    678   {
    679     /* Equivalent to a normal move for our purpooses.  */
    680     emit_move_insn (operands[0], operands[1]);
    681     DONE;
    682   }
    683 )
    684 
    685 ;; Unpredicated moves that can use LDR and STR, i.e. full vectors for which
    686 ;; little-endian ordering is acceptable.  Only allow memory operations during
    687 ;; and after RA; before RA we want the predicated load and store patterns to
    688 ;; be used instead.
    689 (define_insn "*aarch64_sve_mov<mode>_ldr_str"
    690   [(set (match_operand:SVE_FULL 0 "aarch64_sve_nonimmediate_operand" "=w, Utr, w, w")
    691 	(match_operand:SVE_FULL 1 "aarch64_sve_general_operand" "Utr, w, w, Dn"))]
    692   "TARGET_SVE
    693    && (<MODE>mode == VNx16QImode || !BYTES_BIG_ENDIAN)
    694    && ((lra_in_progress || reload_completed)
    695        || (register_operand (operands[0], <MODE>mode)
    696 	   && nonmemory_operand (operands[1], <MODE>mode)))"
    697   "@
    698    ldr\t%0, %1
    699    str\t%1, %0
    700    mov\t%0.d, %1.d
    701    * return aarch64_output_sve_mov_immediate (operands[1]);"
    702 )
    703 
    704 ;; Unpredicated moves that cannot use LDR and STR, i.e. partial vectors
    705 ;; or vectors for which little-endian ordering isn't acceptable.  Memory
    706 ;; accesses require secondary reloads.
    707 (define_insn "*aarch64_sve_mov<mode>_no_ldr_str"
    708   [(set (match_operand:SVE_ALL 0 "register_operand" "=w, w")
    709 	(match_operand:SVE_ALL 1 "aarch64_nonmemory_operand" "w, Dn"))]
    710   "TARGET_SVE
    711    && <MODE>mode != VNx16QImode
    712    && (BYTES_BIG_ENDIAN
    713        || maybe_ne (BYTES_PER_SVE_VECTOR, GET_MODE_SIZE (<MODE>mode)))"
    714   "@
    715    mov\t%0.d, %1.d
    716    * return aarch64_output_sve_mov_immediate (operands[1]);"
    717 )
    718 
    719 ;; Handle memory reloads for modes that can't use LDR and STR.  We use
    720 ;; byte PTRUE for all modes to try to encourage reuse.  This pattern
    721 ;; needs constraints because it is returned by TARGET_SECONDARY_RELOAD.
    722 (define_expand "aarch64_sve_reload_mem"
    723   [(parallel
    724      [(set (match_operand 0)
    725 	   (match_operand 1))
    726       (clobber (match_operand:VNx16BI 2 "register_operand" "=Upl"))])]
    727   "TARGET_SVE"
    728   {
    729     /* Create a PTRUE.  */
    730     emit_move_insn (operands[2], CONSTM1_RTX (VNx16BImode));
    731 
    732     /* Refer to the PTRUE in the appropriate mode for this move.  */
    733     machine_mode mode = GET_MODE (operands[0]);
    734     rtx pred = gen_lowpart (aarch64_sve_pred_mode (mode), operands[2]);
    735 
    736     /* Emit a predicated load or store.  */
    737     aarch64_emit_sve_pred_move (operands[0], pred, operands[1]);
    738     DONE;
    739   }
    740 )
    741 
    742 ;; A predicated move in which the predicate is known to be all-true.
    743 ;; Note that this pattern is generated directly by aarch64_emit_sve_pred_move,
    744 ;; so changes to this pattern will need changes there as well.
    745 (define_insn_and_split "@aarch64_pred_mov<mode>"
    746   [(set (match_operand:SVE_ALL 0 "nonimmediate_operand" "=w, w, m")
    747 	(unspec:SVE_ALL
    748 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
    749 	   (match_operand:SVE_ALL 2 "nonimmediate_operand" "w, m, w")]
    750 	  UNSPEC_PRED_X))]
    751   "TARGET_SVE
    752    && (register_operand (operands[0], <MODE>mode)
    753        || register_operand (operands[2], <MODE>mode))"
    754   "@
    755    #
    756    ld1<Vesize>\t%0.<Vctype>, %1/z, %2
    757    st1<Vesize>\t%2.<Vctype>, %1, %0"
    758   "&& register_operand (operands[0], <MODE>mode)
    759    && register_operand (operands[2], <MODE>mode)"
    760   [(set (match_dup 0) (match_dup 2))]
    761 )
    762 
    763 ;; A pattern for optimizing SUBREGs that have a reinterpreting effect
    764 ;; on big-endian targets; see aarch64_maybe_expand_sve_subreg_move
    765 ;; for details.  We use a special predicate for operand 2 to reduce
    766 ;; the number of patterns.
    767 (define_insn_and_split "*aarch64_sve_mov<mode>_subreg_be"
    768   [(set (match_operand:SVE_ALL 0 "aarch64_sve_nonimmediate_operand" "=w")
    769 	(unspec:SVE_ALL
    770 	  [(match_operand:VNx16BI 1 "register_operand" "Upl")
    771 	   (match_operand 2 "aarch64_any_register_operand" "w")]
    772 	  UNSPEC_REV_SUBREG))]
    773   "TARGET_SVE && BYTES_BIG_ENDIAN"
    774   "#"
    775   "&& reload_completed"
    776   [(const_int 0)]
    777   {
    778     aarch64_split_sve_subreg_move (operands[0], operands[1], operands[2]);
    779     DONE;
    780   }
    781 )
    782 
    783 ;; Reinterpret operand 1 in operand 0's mode, without changing its contents.
    784 ;; This is equivalent to a subreg on little-endian targets but not for
    785 ;; big-endian; see the comment at the head of the file for details.
    786 (define_expand "@aarch64_sve_reinterpret<mode>"
    787   [(set (match_operand:SVE_ALL 0 "register_operand")
    788 	(unspec:SVE_ALL
    789 	  [(match_operand 1 "aarch64_any_register_operand")]
    790 	  UNSPEC_REINTERPRET))]
    791   "TARGET_SVE"
    792   {
    793     machine_mode src_mode = GET_MODE (operands[1]);
    794     if (targetm.can_change_mode_class (<MODE>mode, src_mode, FP_REGS))
    795       {
    796 	emit_move_insn (operands[0], gen_lowpart (<MODE>mode, operands[1]));
    797 	DONE;
    798       }
    799   }
    800 )
    801 
    802 ;; A pattern for handling type punning on big-endian targets.  We use a
    803 ;; special predicate for operand 1 to reduce the number of patterns.
    804 (define_insn_and_split "*aarch64_sve_reinterpret<mode>"
    805   [(set (match_operand:SVE_ALL 0 "register_operand" "=w")
    806 	(unspec:SVE_ALL
    807 	  [(match_operand 1 "aarch64_any_register_operand" "w")]
    808 	  UNSPEC_REINTERPRET))]
    809   "TARGET_SVE"
    810   "#"
    811   "&& reload_completed"
    812   [(set (match_dup 0) (match_dup 1))]
    813   {
    814     operands[1] = aarch64_replace_reg_mode (operands[1], <MODE>mode);
    815   }
    816 )
    817 
    818 ;; -------------------------------------------------------------------------
    819 ;; ---- Moves of multiple vectors
    820 ;; -------------------------------------------------------------------------
    821 ;; All patterns in this section are synthetic and split to real
    822 ;; instructions after reload.
    823 ;; -------------------------------------------------------------------------
    824 
    825 (define_expand "mov<mode>"
    826   [(set (match_operand:SVE_STRUCT 0 "nonimmediate_operand")
    827 	(match_operand:SVE_STRUCT 1 "general_operand"))]
    828   "TARGET_SVE"
    829   {
    830     /* Big-endian loads and stores need to be done via LD1 and ST1;
    831        see the comment at the head of the file for details.  */
    832     if ((MEM_P (operands[0]) || MEM_P (operands[1]))
    833 	&& BYTES_BIG_ENDIAN)
    834       {
    835 	gcc_assert (can_create_pseudo_p ());
    836 	aarch64_expand_sve_mem_move (operands[0], operands[1], <VPRED>mode);
    837 	DONE;
    838       }
    839 
    840     if (CONSTANT_P (operands[1]))
    841       {
    842 	aarch64_expand_mov_immediate (operands[0], operands[1]);
    843 	DONE;
    844       }
    845   }
    846 )
    847 
    848 ;; Unpredicated structure moves (little-endian).
    849 (define_insn "*aarch64_sve_mov<mode>_le"
    850   [(set (match_operand:SVE_STRUCT 0 "aarch64_sve_nonimmediate_operand" "=w, Utr, w, w")
    851 	(match_operand:SVE_STRUCT 1 "aarch64_sve_general_operand" "Utr, w, w, Dn"))]
    852   "TARGET_SVE && !BYTES_BIG_ENDIAN"
    853   "#"
    854   [(set_attr "length" "<insn_length>")]
    855 )
    856 
    857 ;; Unpredicated structure moves (big-endian).  Memory accesses require
    858 ;; secondary reloads.
    859 (define_insn "*aarch64_sve_mov<mode>_be"
    860   [(set (match_operand:SVE_STRUCT 0 "register_operand" "=w, w")
    861 	(match_operand:SVE_STRUCT 1 "aarch64_nonmemory_operand" "w, Dn"))]
    862   "TARGET_SVE && BYTES_BIG_ENDIAN"
    863   "#"
    864   [(set_attr "length" "<insn_length>")]
    865 )
    866 
    867 ;; Split unpredicated structure moves into pieces.  This is the same
    868 ;; for both big-endian and little-endian code, although it only needs
    869 ;; to handle memory operands for little-endian code.
    870 (define_split
    871   [(set (match_operand:SVE_STRUCT 0 "aarch64_sve_nonimmediate_operand")
    872 	(match_operand:SVE_STRUCT 1 "aarch64_sve_general_operand"))]
    873   "TARGET_SVE && reload_completed"
    874   [(const_int 0)]
    875   {
    876     rtx dest = operands[0];
    877     rtx src = operands[1];
    878     if (REG_P (dest) && REG_P (src))
    879       aarch64_simd_emit_reg_reg_move (operands, <VSINGLE>mode, <vector_count>);
    880     else
    881       for (unsigned int i = 0; i < <vector_count>; ++i)
    882 	{
    883 	  rtx subdest = simplify_gen_subreg (<VSINGLE>mode, dest, <MODE>mode,
    884 					     i * BYTES_PER_SVE_VECTOR);
    885 	  rtx subsrc = simplify_gen_subreg (<VSINGLE>mode, src, <MODE>mode,
    886 					    i * BYTES_PER_SVE_VECTOR);
    887 	  emit_insn (gen_rtx_SET (subdest, subsrc));
    888 	}
    889     DONE;
    890   }
    891 )
    892 
    893 ;; Predicated structure moves.  This works for both endiannesses but in
    894 ;; practice is only useful for big-endian.
    895 (define_insn_and_split "@aarch64_pred_mov<mode>"
    896   [(set (match_operand:SVE_STRUCT 0 "aarch64_sve_struct_nonimmediate_operand" "=w, w, Utx")
    897 	(unspec:SVE_STRUCT
    898 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
    899 	   (match_operand:SVE_STRUCT 2 "aarch64_sve_struct_nonimmediate_operand" "w, Utx, w")]
    900 	  UNSPEC_PRED_X))]
    901   "TARGET_SVE
    902    && (register_operand (operands[0], <MODE>mode)
    903        || register_operand (operands[2], <MODE>mode))"
    904   "#"
    905   "&& reload_completed"
    906   [(const_int 0)]
    907   {
    908     for (unsigned int i = 0; i < <vector_count>; ++i)
    909       {
    910 	rtx subdest = simplify_gen_subreg (<VSINGLE>mode, operands[0],
    911 					   <MODE>mode,
    912 					   i * BYTES_PER_SVE_VECTOR);
    913 	rtx subsrc = simplify_gen_subreg (<VSINGLE>mode, operands[2],
    914 					  <MODE>mode,
    915 					  i * BYTES_PER_SVE_VECTOR);
    916 	aarch64_emit_sve_pred_move (subdest, operands[1], subsrc);
    917       }
    918     DONE;
    919   }
    920   [(set_attr "length" "<insn_length>")]
    921 )
    922 
    923 ;; -------------------------------------------------------------------------
    924 ;; ---- Moves of predicates
    925 ;; -------------------------------------------------------------------------
    926 ;; Includes:
    927 ;; - MOV
    928 ;; - LDR
    929 ;; - PFALSE
    930 ;; - PTRUE
    931 ;; - PTRUES
    932 ;; - STR
    933 ;; -------------------------------------------------------------------------
    934 
    935 (define_expand "mov<mode>"
    936   [(set (match_operand:PRED_ALL 0 "nonimmediate_operand")
    937 	(match_operand:PRED_ALL 1 "general_operand"))]
    938   "TARGET_SVE"
    939   {
    940     if (GET_CODE (operands[0]) == MEM)
    941       operands[1] = force_reg (<MODE>mode, operands[1]);
    942 
    943     if (CONSTANT_P (operands[1]))
    944       {
    945 	aarch64_expand_mov_immediate (operands[0], operands[1]);
    946 	DONE;
    947       }
    948   }
    949 )
    950 
    951 (define_insn "*aarch64_sve_mov<mode>"
    952   [(set (match_operand:PRED_ALL 0 "nonimmediate_operand" "=Upa, m, Upa, Upa")
    953 	(match_operand:PRED_ALL 1 "aarch64_mov_operand" "Upa, Upa, m, Dn"))]
    954   "TARGET_SVE
    955    && (register_operand (operands[0], <MODE>mode)
    956        || register_operand (operands[1], <MODE>mode))"
    957   "@
    958    mov\t%0.b, %1.b
    959    str\t%1, %0
    960    ldr\t%0, %1
    961    * return aarch64_output_sve_mov_immediate (operands[1]);"
    962 )
    963 
    964 ;; Match PTRUES Pn.B when both the predicate and flags are useful.
    965 (define_insn_and_rewrite "*aarch64_sve_ptruevnx16bi_cc"
    966   [(set (reg:CC_NZC CC_REGNUM)
    967 	(unspec:CC_NZC
    968 	  [(match_operand 2)
    969 	   (match_operand 3)
    970 	   (const_int SVE_KNOWN_PTRUE)
    971 	   (match_operator:VNx16BI 1 "aarch64_sve_ptrue_svpattern_immediate"
    972 	     [(unspec:VNx16BI
    973 		[(match_operand:SI 4 "const_int_operand")
    974 		 (match_operand:VNx16BI 5 "aarch64_simd_imm_zero")]
    975 		UNSPEC_PTRUE)])]
    976 	  UNSPEC_PTEST))
    977    (set (match_operand:VNx16BI 0 "register_operand" "=Upa")
    978 	(match_dup 1))]
    979   "TARGET_SVE"
    980   {
    981     return aarch64_output_sve_ptrues (operands[1]);
    982   }
    983   "&& (!CONSTANT_P (operands[2]) || !CONSTANT_P (operands[3]))"
    984   {
    985     operands[2] = operands[3] = CONSTM1_RTX (VNx16BImode);
    986   }
    987 )
    988 
    989 ;; Match PTRUES Pn.[HSD] when both the predicate and flags are useful.
    990 (define_insn_and_rewrite "*aarch64_sve_ptrue<mode>_cc"
    991   [(set (reg:CC_NZC CC_REGNUM)
    992 	(unspec:CC_NZC
    993 	  [(match_operand 2)
    994 	   (match_operand 3)
    995 	   (const_int SVE_KNOWN_PTRUE)
    996 	   (subreg:PRED_HSD
    997 	     (match_operator:VNx16BI 1 "aarch64_sve_ptrue_svpattern_immediate"
    998 	       [(unspec:VNx16BI
    999 		  [(match_operand:SI 4 "const_int_operand")
   1000 		   (match_operand:PRED_HSD 5 "aarch64_simd_imm_zero")]
   1001 		  UNSPEC_PTRUE)]) 0)]
   1002 	  UNSPEC_PTEST))
   1003    (set (match_operand:VNx16BI 0 "register_operand" "=Upa")
   1004 	(match_dup 1))]
   1005   "TARGET_SVE"
   1006   {
   1007     return aarch64_output_sve_ptrues (operands[1]);
   1008   }
   1009   "&& (!CONSTANT_P (operands[2]) || !CONSTANT_P (operands[3]))"
   1010   {
   1011     operands[2] = CONSTM1_RTX (VNx16BImode);
   1012     operands[3] = CONSTM1_RTX (<MODE>mode);
   1013   }
   1014 )
   1015 
   1016 ;; Match PTRUES Pn.B when only the flags result is useful (which is
   1017 ;; a way of testing VL).
   1018 (define_insn_and_rewrite "*aarch64_sve_ptruevnx16bi_ptest"
   1019   [(set (reg:CC_NZC CC_REGNUM)
   1020 	(unspec:CC_NZC
   1021 	  [(match_operand 2)
   1022 	   (match_operand 3)
   1023 	   (const_int SVE_KNOWN_PTRUE)
   1024 	   (match_operator:VNx16BI 1 "aarch64_sve_ptrue_svpattern_immediate"
   1025 	     [(unspec:VNx16BI
   1026 		[(match_operand:SI 4 "const_int_operand")
   1027 		 (match_operand:VNx16BI 5 "aarch64_simd_imm_zero")]
   1028 		UNSPEC_PTRUE)])]
   1029 	  UNSPEC_PTEST))
   1030    (clobber (match_scratch:VNx16BI 0 "=Upa"))]
   1031   "TARGET_SVE"
   1032   {
   1033     return aarch64_output_sve_ptrues (operands[1]);
   1034   }
   1035   "&& (!CONSTANT_P (operands[2]) || !CONSTANT_P (operands[3]))"
   1036   {
   1037     operands[2] = operands[3] = CONSTM1_RTX (VNx16BImode);
   1038   }
   1039 )
   1040 
   1041 ;; Match PTRUES Pn.[HWD] when only the flags result is useful (which is
   1042 ;; a way of testing VL).
   1043 (define_insn_and_rewrite "*aarch64_sve_ptrue<mode>_ptest"
   1044   [(set (reg:CC_NZC CC_REGNUM)
   1045 	(unspec:CC_NZC
   1046 	  [(match_operand 2)
   1047 	   (match_operand 3)
   1048 	   (const_int SVE_KNOWN_PTRUE)
   1049 	   (subreg:PRED_HSD
   1050 	     (match_operator:VNx16BI 1 "aarch64_sve_ptrue_svpattern_immediate"
   1051 	       [(unspec:VNx16BI
   1052 		  [(match_operand:SI 4 "const_int_operand")
   1053 		   (match_operand:PRED_HSD 5 "aarch64_simd_imm_zero")]
   1054 		  UNSPEC_PTRUE)]) 0)]
   1055 	  UNSPEC_PTEST))
   1056    (clobber (match_scratch:VNx16BI 0 "=Upa"))]
   1057   "TARGET_SVE"
   1058   {
   1059     return aarch64_output_sve_ptrues (operands[1]);
   1060   }
   1061   "&& (!CONSTANT_P (operands[2]) || !CONSTANT_P (operands[3]))"
   1062   {
   1063     operands[2] = CONSTM1_RTX (VNx16BImode);
   1064     operands[3] = CONSTM1_RTX (<MODE>mode);
   1065   }
   1066 )
   1067 
   1068 ;; -------------------------------------------------------------------------
   1069 ;; ---- Moves relating to the FFR
   1070 ;; -------------------------------------------------------------------------
   1071 ;; RDFFR
   1072 ;; RDFFRS
   1073 ;; SETFFR
   1074 ;; WRFFR
   1075 ;; -------------------------------------------------------------------------
   1076 
   1077 ;; [W1 in the block comment above about FFR handling]
   1078 ;;
   1079 ;; Write to the FFR and start a new FFRT scheduling region.
   1080 (define_insn "aarch64_wrffr"
   1081   [(set (reg:VNx16BI FFR_REGNUM)
   1082 	(match_operand:VNx16BI 0 "aarch64_simd_reg_or_minus_one" "Dm, Upa"))
   1083    (set (reg:VNx16BI FFRT_REGNUM)
   1084 	(unspec:VNx16BI [(match_dup 0)] UNSPEC_WRFFR))]
   1085   "TARGET_SVE"
   1086   "@
   1087    setffr
   1088    wrffr\t%0.b"
   1089 )
   1090 
   1091 ;; [L2 in the block comment above about FFR handling]
   1092 ;;
   1093 ;; Introduce a read from and write to the FFR in the current FFRT region,
   1094 ;; so that the FFR value is live on entry to the region and so that the FFR
   1095 ;; value visibly changes within the region.  This is used (possibly multiple
   1096 ;; times) in an FFRT region that includes LDFF1 or LDNF1 instructions.
   1097 (define_insn "aarch64_update_ffr_for_load"
   1098   [(set (reg:VNx16BI FFR_REGNUM)
   1099 	(unspec:VNx16BI [(reg:VNx16BI FFRT_REGNUM)
   1100 			 (reg:VNx16BI FFR_REGNUM)] UNSPEC_UPDATE_FFR))]
   1101   "TARGET_SVE"
   1102   ""
   1103   [(set_attr "type" "no_insn")]
   1104 )
   1105 
   1106 ;; [R1 in the block comment above about FFR handling]
   1107 ;;
   1108 ;; Notionally copy the FFR to the FFRT, so that the current FFR value
   1109 ;; can be read from there by the RDFFR instructions below.  This acts
   1110 ;; as a scheduling barrier for earlier LDFF1 and LDNF1 instructions and
   1111 ;; creates a natural dependency with earlier writes.
   1112 (define_insn "aarch64_copy_ffr_to_ffrt"
   1113   [(set (reg:VNx16BI FFRT_REGNUM)
   1114 	(reg:VNx16BI FFR_REGNUM))]
   1115   "TARGET_SVE"
   1116   ""
   1117   [(set_attr "type" "no_insn")]
   1118 )
   1119 
   1120 ;; [R2 in the block comment above about FFR handling]
   1121 ;;
   1122 ;; Read the FFR via the FFRT.
   1123 (define_insn "aarch64_rdffr"
   1124   [(set (match_operand:VNx16BI 0 "register_operand" "=Upa")
   1125 	(reg:VNx16BI FFRT_REGNUM))]
   1126   "TARGET_SVE"
   1127   "rdffr\t%0.b"
   1128 )
   1129 
   1130 ;; Likewise with zero predication.
   1131 (define_insn "aarch64_rdffr_z"
   1132   [(set (match_operand:VNx16BI 0 "register_operand" "=Upa")
   1133 	(and:VNx16BI
   1134 	  (reg:VNx16BI FFRT_REGNUM)
   1135 	  (match_operand:VNx16BI 1 "register_operand" "Upa")))]
   1136   "TARGET_SVE"
   1137   "rdffr\t%0.b, %1/z"
   1138 )
   1139 
   1140 ;; Read the FFR to test for a fault, without using the predicate result.
   1141 (define_insn "*aarch64_rdffr_z_ptest"
   1142   [(set (reg:CC_NZC CC_REGNUM)
   1143 	(unspec:CC_NZC
   1144 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   1145 	   (match_dup 1)
   1146 	   (match_operand:SI 2 "aarch64_sve_ptrue_flag")
   1147 	   (and:VNx16BI
   1148 	     (reg:VNx16BI FFRT_REGNUM)
   1149 	     (match_dup 1))]
   1150 	  UNSPEC_PTEST))
   1151    (clobber (match_scratch:VNx16BI 0 "=Upa"))]
   1152   "TARGET_SVE"
   1153   "rdffrs\t%0.b, %1/z"
   1154 )
   1155 
   1156 ;; Same for unpredicated RDFFR when tested with a known PTRUE.
   1157 (define_insn "*aarch64_rdffr_ptest"
   1158   [(set (reg:CC_NZC CC_REGNUM)
   1159 	(unspec:CC_NZC
   1160 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   1161 	   (match_dup 1)
   1162 	   (const_int SVE_KNOWN_PTRUE)
   1163 	   (reg:VNx16BI FFRT_REGNUM)]
   1164 	  UNSPEC_PTEST))
   1165    (clobber (match_scratch:VNx16BI 0 "=Upa"))]
   1166   "TARGET_SVE"
   1167   "rdffrs\t%0.b, %1/z"
   1168 )
   1169 
   1170 ;; Read the FFR with zero predication and test the result.
   1171 (define_insn "*aarch64_rdffr_z_cc"
   1172   [(set (reg:CC_NZC CC_REGNUM)
   1173 	(unspec:CC_NZC
   1174 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   1175 	   (match_dup 1)
   1176 	   (match_operand:SI 2 "aarch64_sve_ptrue_flag")
   1177 	   (and:VNx16BI
   1178 	     (reg:VNx16BI FFRT_REGNUM)
   1179 	     (match_dup 1))]
   1180 	  UNSPEC_PTEST))
   1181    (set (match_operand:VNx16BI 0 "register_operand" "=Upa")
   1182 	(and:VNx16BI
   1183 	  (reg:VNx16BI FFRT_REGNUM)
   1184 	  (match_dup 1)))]
   1185   "TARGET_SVE"
   1186   "rdffrs\t%0.b, %1/z"
   1187 )
   1188 
   1189 ;; Same for unpredicated RDFFR when tested with a known PTRUE.
   1190 (define_insn "*aarch64_rdffr_cc"
   1191   [(set (reg:CC_NZC CC_REGNUM)
   1192 	(unspec:CC_NZC
   1193 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   1194 	   (match_dup 1)
   1195 	   (const_int SVE_KNOWN_PTRUE)
   1196 	   (reg:VNx16BI FFRT_REGNUM)]
   1197 	  UNSPEC_PTEST))
   1198    (set (match_operand:VNx16BI 0 "register_operand" "=Upa")
   1199 	(reg:VNx16BI FFRT_REGNUM))]
   1200   "TARGET_SVE"
   1201   "rdffrs\t%0.b, %1/z"
   1202 )
   1203 
   1204 ;; [R3 in the block comment above about FFR handling]
   1205 ;;
   1206 ;; Arbitrarily update the FFRT after a read from the FFR.  This acts as
   1207 ;; a scheduling barrier for later LDFF1 and LDNF1 instructions.
   1208 (define_insn "aarch64_update_ffrt"
   1209   [(set (reg:VNx16BI FFRT_REGNUM)
   1210 	(unspec:VNx16BI [(reg:VNx16BI FFRT_REGNUM)] UNSPEC_UPDATE_FFRT))]
   1211   "TARGET_SVE"
   1212   ""
   1213   [(set_attr "type" "no_insn")]
   1214 )
   1215 
   1216 ;; =========================================================================
   1217 ;; == Loads
   1218 ;; =========================================================================
   1219 
   1220 ;; -------------------------------------------------------------------------
   1221 ;; ---- Normal contiguous loads
   1222 ;; -------------------------------------------------------------------------
   1223 ;; Includes contiguous forms of:
   1224 ;; - LD1B
   1225 ;; - LD1D
   1226 ;; - LD1H
   1227 ;; - LD1W
   1228 ;; - LD2B
   1229 ;; - LD2D
   1230 ;; - LD2H
   1231 ;; - LD2W
   1232 ;; - LD3B
   1233 ;; - LD3D
   1234 ;; - LD3H
   1235 ;; - LD3W
   1236 ;; - LD4B
   1237 ;; - LD4D
   1238 ;; - LD4H
   1239 ;; - LD4W
   1240 ;; -------------------------------------------------------------------------
   1241 
   1242 ;; Predicated LD1.
   1243 (define_insn "maskload<mode><vpred>"
   1244   [(set (match_operand:SVE_ALL 0 "register_operand" "=w")
   1245 	(unspec:SVE_ALL
   1246 	  [(match_operand:<VPRED> 2 "register_operand" "Upl")
   1247 	   (match_operand:SVE_ALL 1 "memory_operand" "m")]
   1248 	  UNSPEC_LD1_SVE))]
   1249   "TARGET_SVE"
   1250   "ld1<Vesize>\t%0.<Vctype>, %2/z, %1"
   1251 )
   1252 
   1253 ;; Unpredicated LD[234].
   1254 (define_expand "vec_load_lanes<mode><vsingle>"
   1255   [(set (match_operand:SVE_STRUCT 0 "register_operand")
   1256 	(unspec:SVE_STRUCT
   1257 	  [(match_dup 2)
   1258 	   (match_operand:SVE_STRUCT 1 "memory_operand")]
   1259 	  UNSPEC_LDN))]
   1260   "TARGET_SVE"
   1261   {
   1262     operands[2] = aarch64_ptrue_reg (<VPRED>mode);
   1263   }
   1264 )
   1265 
   1266 ;; Predicated LD[234].
   1267 (define_insn "vec_mask_load_lanes<mode><vsingle>"
   1268   [(set (match_operand:SVE_STRUCT 0 "register_operand" "=w")
   1269 	(unspec:SVE_STRUCT
   1270 	  [(match_operand:<VPRED> 2 "register_operand" "Upl")
   1271 	   (match_operand:SVE_STRUCT 1 "memory_operand" "m")]
   1272 	  UNSPEC_LDN))]
   1273   "TARGET_SVE"
   1274   "ld<vector_count><Vesize>\t%0, %2/z, %1"
   1275 )
   1276 
   1277 ;; -------------------------------------------------------------------------
   1278 ;; ---- Extending contiguous loads
   1279 ;; -------------------------------------------------------------------------
   1280 ;; Includes contiguous forms of:
   1281 ;; LD1B
   1282 ;; LD1H
   1283 ;; LD1SB
   1284 ;; LD1SH
   1285 ;; LD1SW
   1286 ;; LD1W
   1287 ;; -------------------------------------------------------------------------
   1288 
   1289 ;; Predicated load and extend, with 8 elements per 128-bit block.
   1290 (define_insn_and_rewrite "@aarch64_load<SVE_PRED_LOAD:pred_load>_<ANY_EXTEND:optab><SVE_HSDI:mode><SVE_PARTIAL_I:mode>"
   1291   [(set (match_operand:SVE_HSDI 0 "register_operand" "=w")
   1292 	(unspec:SVE_HSDI
   1293 	  [(match_operand:<SVE_HSDI:VPRED> 3 "general_operand" "UplDnm")
   1294 	   (ANY_EXTEND:SVE_HSDI
   1295 	     (unspec:SVE_PARTIAL_I
   1296 	       [(match_operand:<SVE_PARTIAL_I:VPRED> 2 "register_operand" "Upl")
   1297 		(match_operand:SVE_PARTIAL_I 1 "memory_operand" "m")]
   1298 	       SVE_PRED_LOAD))]
   1299 	  UNSPEC_PRED_X))]
   1300   "TARGET_SVE && (~<SVE_HSDI:narrower_mask> & <SVE_PARTIAL_I:self_mask>) == 0"
   1301   "ld1<ANY_EXTEND:s><SVE_PARTIAL_I:Vesize>\t%0.<SVE_HSDI:Vctype>, %2/z, %1"
   1302   "&& !CONSTANT_P (operands[3])"
   1303   {
   1304     operands[3] = CONSTM1_RTX (<SVE_HSDI:VPRED>mode);
   1305   }
   1306 )
   1307 
   1308 ;; -------------------------------------------------------------------------
   1309 ;; ---- First-faulting contiguous loads
   1310 ;; -------------------------------------------------------------------------
   1311 ;; Includes contiguous forms of:
   1312 ;; - LDFF1B
   1313 ;; - LDFF1D
   1314 ;; - LDFF1H
   1315 ;; - LDFF1W
   1316 ;; - LDNF1B
   1317 ;; - LDNF1D
   1318 ;; - LDNF1H
   1319 ;; - LDNF1W
   1320 ;; -------------------------------------------------------------------------
   1321 
   1322 ;; Contiguous non-extending first-faulting or non-faulting loads.
   1323 (define_insn "@aarch64_ld<fn>f1<mode>"
   1324   [(set (match_operand:SVE_FULL 0 "register_operand" "=w")
   1325 	(unspec:SVE_FULL
   1326 	  [(match_operand:<VPRED> 2 "register_operand" "Upl")
   1327 	   (match_operand:SVE_FULL 1 "aarch64_sve_ld<fn>f1_operand" "Ut<fn>")
   1328 	   (reg:VNx16BI FFRT_REGNUM)]
   1329 	  SVE_LDFF1_LDNF1))]
   1330   "TARGET_SVE"
   1331   "ld<fn>f1<Vesize>\t%0.<Vetype>, %2/z, %1"
   1332 )
   1333 
   1334 ;; -------------------------------------------------------------------------
   1335 ;; ---- First-faulting extending contiguous loads
   1336 ;; -------------------------------------------------------------------------
   1337 ;; Includes contiguous forms of:
   1338 ;; - LDFF1B
   1339 ;; - LDFF1H
   1340 ;; - LDFF1SB
   1341 ;; - LDFF1SH
   1342 ;; - LDFF1SW
   1343 ;; - LDFF1W
   1344 ;; - LDNF1B
   1345 ;; - LDNF1H
   1346 ;; - LDNF1SB
   1347 ;; - LDNF1SH
   1348 ;; - LDNF1SW
   1349 ;; - LDNF1W
   1350 ;; -------------------------------------------------------------------------
   1351 
   1352 ;; Predicated first-faulting or non-faulting load and extend.
   1353 (define_insn_and_rewrite "@aarch64_ld<fn>f1_<ANY_EXTEND:optab><SVE_HSDI:mode><SVE_PARTIAL_I:mode>"
   1354   [(set (match_operand:SVE_HSDI 0 "register_operand" "=w")
   1355 	(unspec:SVE_HSDI
   1356 	  [(match_operand:<SVE_HSDI:VPRED> 3 "general_operand" "UplDnm")
   1357 	   (ANY_EXTEND:SVE_HSDI
   1358 	     (unspec:SVE_PARTIAL_I
   1359 	       [(match_operand:<SVE_PARTIAL_I:VPRED> 2 "register_operand" "Upl")
   1360 		(match_operand:SVE_PARTIAL_I 1 "aarch64_sve_ld<fn>f1_operand" "Ut<fn>")
   1361 		(reg:VNx16BI FFRT_REGNUM)]
   1362 	       SVE_LDFF1_LDNF1))]
   1363 	  UNSPEC_PRED_X))]
   1364   "TARGET_SVE && (~<SVE_HSDI:narrower_mask> & <SVE_PARTIAL_I:self_mask>) == 0"
   1365   "ld<fn>f1<ANY_EXTEND:s><SVE_PARTIAL_I:Vesize>\t%0.<SVE_HSDI:Vctype>, %2/z, %1"
   1366   "&& !CONSTANT_P (operands[3])"
   1367   {
   1368     operands[3] = CONSTM1_RTX (<SVE_HSDI:VPRED>mode);
   1369   }
   1370 )
   1371 
   1372 ;; -------------------------------------------------------------------------
   1373 ;; ---- Non-temporal contiguous loads
   1374 ;; -------------------------------------------------------------------------
   1375 ;; Includes:
   1376 ;; - LDNT1B
   1377 ;; - LDNT1D
   1378 ;; - LDNT1H
   1379 ;; - LDNT1W
   1380 ;; -------------------------------------------------------------------------
   1381 
   1382 ;; Predicated contiguous non-temporal load.
   1383 (define_insn "@aarch64_ldnt1<mode>"
   1384   [(set (match_operand:SVE_FULL 0 "register_operand" "=w")
   1385 	(unspec:SVE_FULL
   1386 	  [(match_operand:<VPRED> 2 "register_operand" "Upl")
   1387 	   (match_operand:SVE_FULL 1 "memory_operand" "m")]
   1388 	  UNSPEC_LDNT1_SVE))]
   1389   "TARGET_SVE"
   1390   "ldnt1<Vesize>\t%0.<Vetype>, %2/z, %1"
   1391 )
   1392 
   1393 ;; -------------------------------------------------------------------------
   1394 ;; ---- Normal gather loads
   1395 ;; -------------------------------------------------------------------------
   1396 ;; Includes gather forms of:
   1397 ;; - LD1D
   1398 ;; - LD1W
   1399 ;; -------------------------------------------------------------------------
   1400 
   1401 ;; Unpredicated gather loads.
   1402 (define_expand "gather_load<mode><v_int_container>"
   1403   [(set (match_operand:SVE_24 0 "register_operand")
   1404 	(unspec:SVE_24
   1405 	  [(match_dup 5)
   1406 	   (match_operand:DI 1 "aarch64_sve_gather_offset_<Vesize>")
   1407 	   (match_operand:<V_INT_CONTAINER> 2 "register_operand")
   1408 	   (match_operand:DI 3 "const_int_operand")
   1409 	   (match_operand:DI 4 "aarch64_gather_scale_operand_<Vesize>")
   1410 	   (mem:BLK (scratch))]
   1411 	  UNSPEC_LD1_GATHER))]
   1412   "TARGET_SVE"
   1413   {
   1414     operands[5] = aarch64_ptrue_reg (<VPRED>mode);
   1415   }
   1416 )
   1417 
   1418 ;; Predicated gather loads for 32-bit elements.  Operand 3 is true for
   1419 ;; unsigned extension and false for signed extension.
   1420 (define_insn "mask_gather_load<mode><v_int_container>"
   1421   [(set (match_operand:SVE_4 0 "register_operand" "=w, w, w, w, w, w")
   1422 	(unspec:SVE_4
   1423 	  [(match_operand:VNx4BI 5 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   1424 	   (match_operand:DI 1 "aarch64_sve_gather_offset_<Vesize>" "Z, vgw, rk, rk, rk, rk")
   1425 	   (match_operand:VNx4SI 2 "register_operand" "w, w, w, w, w, w")
   1426 	   (match_operand:DI 3 "const_int_operand" "Ui1, Ui1, Z, Ui1, Z, Ui1")
   1427 	   (match_operand:DI 4 "aarch64_gather_scale_operand_<Vesize>" "Ui1, Ui1, Ui1, Ui1, i, i")
   1428 	   (mem:BLK (scratch))]
   1429 	  UNSPEC_LD1_GATHER))]
   1430   "TARGET_SVE"
   1431   "@
   1432    ld1<Vesize>\t%0.s, %5/z, [%2.s]
   1433    ld1<Vesize>\t%0.s, %5/z, [%2.s, #%1]
   1434    ld1<Vesize>\t%0.s, %5/z, [%1, %2.s, sxtw]
   1435    ld1<Vesize>\t%0.s, %5/z, [%1, %2.s, uxtw]
   1436    ld1<Vesize>\t%0.s, %5/z, [%1, %2.s, sxtw %p4]
   1437    ld1<Vesize>\t%0.s, %5/z, [%1, %2.s, uxtw %p4]"
   1438 )
   1439 
   1440 ;; Predicated gather loads for 64-bit elements.  The value of operand 3
   1441 ;; doesn't matter in this case.
   1442 (define_insn "mask_gather_load<mode><v_int_container>"
   1443   [(set (match_operand:SVE_2 0 "register_operand" "=w, w, w, w")
   1444 	(unspec:SVE_2
   1445 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl, Upl, Upl")
   1446 	   (match_operand:DI 1 "aarch64_sve_gather_offset_<Vesize>" "Z, vgd, rk, rk")
   1447 	   (match_operand:VNx2DI 2 "register_operand" "w, w, w, w")
   1448 	   (match_operand:DI 3 "const_int_operand")
   1449 	   (match_operand:DI 4 "aarch64_gather_scale_operand_<Vesize>" "Ui1, Ui1, Ui1, i")
   1450 	   (mem:BLK (scratch))]
   1451 	  UNSPEC_LD1_GATHER))]
   1452   "TARGET_SVE"
   1453   "@
   1454    ld1<Vesize>\t%0.d, %5/z, [%2.d]
   1455    ld1<Vesize>\t%0.d, %5/z, [%2.d, #%1]
   1456    ld1<Vesize>\t%0.d, %5/z, [%1, %2.d]
   1457    ld1<Vesize>\t%0.d, %5/z, [%1, %2.d, lsl %p4]"
   1458 )
   1459 
   1460 ;; Likewise, but with the offset being extended from 32 bits.
   1461 (define_insn_and_rewrite "*mask_gather_load<mode><v_int_container>_<su>xtw_unpacked"
   1462   [(set (match_operand:SVE_2 0 "register_operand" "=w, w")
   1463 	(unspec:SVE_2
   1464 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   1465 	   (match_operand:DI 1 "register_operand" "rk, rk")
   1466 	   (unspec:VNx2DI
   1467 	     [(match_operand 6)
   1468 	      (ANY_EXTEND:VNx2DI
   1469 		(match_operand:VNx2SI 2 "register_operand" "w, w"))]
   1470 	     UNSPEC_PRED_X)
   1471 	   (match_operand:DI 3 "const_int_operand")
   1472 	   (match_operand:DI 4 "aarch64_gather_scale_operand_<Vesize>" "Ui1, i")
   1473 	   (mem:BLK (scratch))]
   1474 	  UNSPEC_LD1_GATHER))]
   1475   "TARGET_SVE"
   1476   "@
   1477    ld1<Vesize>\t%0.d, %5/z, [%1, %2.d, <su>xtw]
   1478    ld1<Vesize>\t%0.d, %5/z, [%1, %2.d, <su>xtw %p4]"
   1479   "&& !CONSTANT_P (operands[6])"
   1480   {
   1481     operands[6] = CONSTM1_RTX (VNx2BImode);
   1482   }
   1483 )
   1484 
   1485 ;; Likewise, but with the offset being truncated to 32 bits and then
   1486 ;; sign-extended.
   1487 (define_insn_and_rewrite "*mask_gather_load<mode><v_int_container>_sxtw"
   1488   [(set (match_operand:SVE_2 0 "register_operand" "=w, w")
   1489 	(unspec:SVE_2
   1490 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   1491 	   (match_operand:DI 1 "register_operand" "rk, rk")
   1492 	   (unspec:VNx2DI
   1493 	     [(match_operand 6)
   1494 	      (sign_extend:VNx2DI
   1495 		(truncate:VNx2SI
   1496 		  (match_operand:VNx2DI 2 "register_operand" "w, w")))]
   1497 	     UNSPEC_PRED_X)
   1498 	   (match_operand:DI 3 "const_int_operand")
   1499 	   (match_operand:DI 4 "aarch64_gather_scale_operand_<Vesize>" "Ui1, i")
   1500 	   (mem:BLK (scratch))]
   1501 	  UNSPEC_LD1_GATHER))]
   1502   "TARGET_SVE"
   1503   "@
   1504    ld1<Vesize>\t%0.d, %5/z, [%1, %2.d, sxtw]
   1505    ld1<Vesize>\t%0.d, %5/z, [%1, %2.d, sxtw %p4]"
   1506   "&& !CONSTANT_P (operands[6])"
   1507   {
   1508     operands[6] = CONSTM1_RTX (VNx2BImode);
   1509   }
   1510 )
   1511 
   1512 ;; Likewise, but with the offset being truncated to 32 bits and then
   1513 ;; zero-extended.
   1514 (define_insn "*mask_gather_load<mode><v_int_container>_uxtw"
   1515   [(set (match_operand:SVE_2 0 "register_operand" "=w, w")
   1516 	(unspec:SVE_2
   1517 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   1518 	   (match_operand:DI 1 "register_operand" "rk, rk")
   1519 	   (and:VNx2DI
   1520 	     (match_operand:VNx2DI 2 "register_operand" "w, w")
   1521 	     (match_operand:VNx2DI 6 "aarch64_sve_uxtw_immediate"))
   1522 	   (match_operand:DI 3 "const_int_operand")
   1523 	   (match_operand:DI 4 "aarch64_gather_scale_operand_<Vesize>" "Ui1, i")
   1524 	   (mem:BLK (scratch))]
   1525 	  UNSPEC_LD1_GATHER))]
   1526   "TARGET_SVE"
   1527   "@
   1528    ld1<Vesize>\t%0.d, %5/z, [%1, %2.d, uxtw]
   1529    ld1<Vesize>\t%0.d, %5/z, [%1, %2.d, uxtw %p4]"
   1530 )
   1531 
   1532 ;; -------------------------------------------------------------------------
   1533 ;; ---- Extending gather loads
   1534 ;; -------------------------------------------------------------------------
   1535 ;; Includes gather forms of:
   1536 ;; - LD1B
   1537 ;; - LD1H
   1538 ;; - LD1SB
   1539 ;; - LD1SH
   1540 ;; - LD1SW
   1541 ;; - LD1W
   1542 ;; -------------------------------------------------------------------------
   1543 
   1544 ;; Predicated extending gather loads for 32-bit elements.  Operand 3 is
   1545 ;; true for unsigned extension and false for signed extension.
   1546 (define_insn_and_rewrite "@aarch64_gather_load_<ANY_EXTEND:optab><SVE_4HSI:mode><SVE_4BHI:mode>"
   1547   [(set (match_operand:SVE_4HSI 0 "register_operand" "=w, w, w, w, w, w")
   1548 	(unspec:SVE_4HSI
   1549 	  [(match_operand:VNx4BI 6 "general_operand" "UplDnm, UplDnm, UplDnm, UplDnm, UplDnm, UplDnm")
   1550 	   (ANY_EXTEND:SVE_4HSI
   1551 	     (unspec:SVE_4BHI
   1552 	       [(match_operand:VNx4BI 5 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   1553 		(match_operand:DI 1 "aarch64_sve_gather_offset_<SVE_4BHI:Vesize>" "Z, vg<SVE_4BHI:Vesize>, rk, rk, rk, rk")
   1554 		(match_operand:VNx4SI 2 "register_operand" "w, w, w, w, w, w")
   1555 		(match_operand:DI 3 "const_int_operand" "Ui1, Ui1, Z, Ui1, Z, Ui1")
   1556 		(match_operand:DI 4 "aarch64_gather_scale_operand_<SVE_4BHI:Vesize>" "Ui1, Ui1, Ui1, Ui1, i, i")
   1557 		(mem:BLK (scratch))]
   1558 	       UNSPEC_LD1_GATHER))]
   1559 	  UNSPEC_PRED_X))]
   1560   "TARGET_SVE && (~<SVE_4HSI:narrower_mask> & <SVE_4BHI:self_mask>) == 0"
   1561   "@
   1562    ld1<ANY_EXTEND:s><SVE_4BHI:Vesize>\t%0.s, %5/z, [%2.s]
   1563    ld1<ANY_EXTEND:s><SVE_4BHI:Vesize>\t%0.s, %5/z, [%2.s, #%1]
   1564    ld1<ANY_EXTEND:s><SVE_4BHI:Vesize>\t%0.s, %5/z, [%1, %2.s, sxtw]
   1565    ld1<ANY_EXTEND:s><SVE_4BHI:Vesize>\t%0.s, %5/z, [%1, %2.s, uxtw]
   1566    ld1<ANY_EXTEND:s><SVE_4BHI:Vesize>\t%0.s, %5/z, [%1, %2.s, sxtw %p4]
   1567    ld1<ANY_EXTEND:s><SVE_4BHI:Vesize>\t%0.s, %5/z, [%1, %2.s, uxtw %p4]"
   1568   "&& !CONSTANT_P (operands[6])"
   1569   {
   1570     operands[6] = CONSTM1_RTX (VNx4BImode);
   1571   }
   1572 )
   1573 
   1574 ;; Predicated extending gather loads for 64-bit elements.  The value of
   1575 ;; operand 3 doesn't matter in this case.
   1576 (define_insn_and_rewrite "@aarch64_gather_load_<ANY_EXTEND:optab><SVE_2HSDI:mode><SVE_2BHSI:mode>"
   1577   [(set (match_operand:SVE_2HSDI 0 "register_operand" "=w, w, w, w")
   1578 	(unspec:SVE_2HSDI
   1579 	  [(match_operand:VNx2BI 6 "general_operand" "UplDnm, UplDnm, UplDnm, UplDnm")
   1580 	   (ANY_EXTEND:SVE_2HSDI
   1581 	     (unspec:SVE_2BHSI
   1582 	       [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl, Upl, Upl")
   1583 		(match_operand:DI 1 "aarch64_sve_gather_offset_<SVE_2BHSI:Vesize>" "Z, vg<SVE_2BHSI:Vesize>, rk, rk")
   1584 		(match_operand:VNx2DI 2 "register_operand" "w, w, w, w")
   1585 		(match_operand:DI 3 "const_int_operand")
   1586 		(match_operand:DI 4 "aarch64_gather_scale_operand_<SVE_2BHSI:Vesize>" "Ui1, Ui1, Ui1, i")
   1587 		(mem:BLK (scratch))]
   1588 	       UNSPEC_LD1_GATHER))]
   1589 	  UNSPEC_PRED_X))]
   1590   "TARGET_SVE && (~<SVE_2HSDI:narrower_mask> & <SVE_2BHSI:self_mask>) == 0"
   1591   "@
   1592    ld1<ANY_EXTEND:s><SVE_2BHSI:Vesize>\t%0.d, %5/z, [%2.d]
   1593    ld1<ANY_EXTEND:s><SVE_2BHSI:Vesize>\t%0.d, %5/z, [%2.d, #%1]
   1594    ld1<ANY_EXTEND:s><SVE_2BHSI:Vesize>\t%0.d, %5/z, [%1, %2.d]
   1595    ld1<ANY_EXTEND:s><SVE_2BHSI:Vesize>\t%0.d, %5/z, [%1, %2.d, lsl %p4]"
   1596   "&& !CONSTANT_P (operands[6])"
   1597   {
   1598     operands[6] = CONSTM1_RTX (VNx2BImode);
   1599   }
   1600 )
   1601 
   1602 ;; Likewise, but with the offset being extended from 32 bits.
   1603 (define_insn_and_rewrite "*aarch64_gather_load_<ANY_EXTEND:optab><SVE_2HSDI:mode><SVE_2BHSI:mode>_<ANY_EXTEND2:su>xtw_unpacked"
   1604   [(set (match_operand:SVE_2HSDI 0 "register_operand" "=w, w")
   1605 	(unspec:SVE_2HSDI
   1606 	  [(match_operand 6)
   1607 	   (ANY_EXTEND:SVE_2HSDI
   1608 	     (unspec:SVE_2BHSI
   1609 	       [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   1610 		(match_operand:DI 1 "aarch64_reg_or_zero" "rk, rk")
   1611 		(unspec:VNx2DI
   1612 		  [(match_operand 7)
   1613 		   (ANY_EXTEND2:VNx2DI
   1614 		     (match_operand:VNx2SI 2 "register_operand" "w, w"))]
   1615 		  UNSPEC_PRED_X)
   1616 		(match_operand:DI 3 "const_int_operand")
   1617 		(match_operand:DI 4 "aarch64_gather_scale_operand_<SVE_2BHSI:Vesize>" "Ui1, i")
   1618 		(mem:BLK (scratch))]
   1619 	       UNSPEC_LD1_GATHER))]
   1620 	  UNSPEC_PRED_X))]
   1621   "TARGET_SVE && (~<SVE_2HSDI:narrower_mask> & <SVE_2BHSI:self_mask>) == 0"
   1622   "@
   1623    ld1<ANY_EXTEND:s><SVE_2BHSI:Vesize>\t%0.d, %5/z, [%1, %2.d, <ANY_EXTEND2:su>xtw]
   1624    ld1<ANY_EXTEND:s><SVE_2BHSI:Vesize>\t%0.d, %5/z, [%1, %2.d, <ANY_EXTEND2:su>xtw %p4]"
   1625   "&& (!CONSTANT_P (operands[6]) || !CONSTANT_P (operands[7]))"
   1626   {
   1627     operands[6] = CONSTM1_RTX (VNx2BImode);
   1628     operands[7] = CONSTM1_RTX (VNx2BImode);
   1629   }
   1630 )
   1631 
   1632 ;; Likewise, but with the offset being truncated to 32 bits and then
   1633 ;; sign-extended.
   1634 (define_insn_and_rewrite "*aarch64_gather_load_<ANY_EXTEND:optab><SVE_2HSDI:mode><SVE_2BHSI:mode>_sxtw"
   1635   [(set (match_operand:SVE_2HSDI 0 "register_operand" "=w, w")
   1636 	(unspec:SVE_2HSDI
   1637 	  [(match_operand 6)
   1638 	   (ANY_EXTEND:SVE_2HSDI
   1639 	     (unspec:SVE_2BHSI
   1640 	       [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   1641 		(match_operand:DI 1 "aarch64_reg_or_zero" "rk, rk")
   1642 		(unspec:VNx2DI
   1643 		  [(match_operand 7)
   1644 		   (sign_extend:VNx2DI
   1645 		     (truncate:VNx2SI
   1646 		       (match_operand:VNx2DI 2 "register_operand" "w, w")))]
   1647 		  UNSPEC_PRED_X)
   1648 		(match_operand:DI 3 "const_int_operand")
   1649 		(match_operand:DI 4 "aarch64_gather_scale_operand_<SVE_2BHSI:Vesize>" "Ui1, i")
   1650 		(mem:BLK (scratch))]
   1651 	       UNSPEC_LD1_GATHER))]
   1652 	  UNSPEC_PRED_X))]
   1653   "TARGET_SVE && (~<SVE_2HSDI:narrower_mask> & <SVE_2BHSI:self_mask>) == 0"
   1654   "@
   1655    ld1<ANY_EXTEND:s><SVE_2BHSI:Vesize>\t%0.d, %5/z, [%1, %2.d, sxtw]
   1656    ld1<ANY_EXTEND:s><SVE_2BHSI:Vesize>\t%0.d, %5/z, [%1, %2.d, sxtw %p4]"
   1657   "&& (!CONSTANT_P (operands[6]) || !CONSTANT_P (operands[7]))"
   1658   {
   1659     operands[6] = CONSTM1_RTX (VNx2BImode);
   1660     operands[7] = CONSTM1_RTX (VNx2BImode);
   1661   }
   1662 )
   1663 
   1664 ;; Likewise, but with the offset being truncated to 32 bits and then
   1665 ;; zero-extended.
   1666 (define_insn_and_rewrite "*aarch64_gather_load_<ANY_EXTEND:optab><SVE_2HSDI:mode><SVE_2BHSI:mode>_uxtw"
   1667   [(set (match_operand:SVE_2HSDI 0 "register_operand" "=w, w")
   1668 	(unspec:SVE_2HSDI
   1669 	  [(match_operand 7)
   1670 	   (ANY_EXTEND:SVE_2HSDI
   1671 	     (unspec:SVE_2BHSI
   1672 	       [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   1673 		(match_operand:DI 1 "aarch64_reg_or_zero" "rk, rk")
   1674 		(and:VNx2DI
   1675 		  (match_operand:VNx2DI 2 "register_operand" "w, w")
   1676 		  (match_operand:VNx2DI 6 "aarch64_sve_uxtw_immediate"))
   1677 		(match_operand:DI 3 "const_int_operand")
   1678 		(match_operand:DI 4 "aarch64_gather_scale_operand_<SVE_2BHSI:Vesize>" "Ui1, i")
   1679 		(mem:BLK (scratch))]
   1680 	       UNSPEC_LD1_GATHER))]
   1681 	  UNSPEC_PRED_X))]
   1682   "TARGET_SVE && (~<SVE_2HSDI:narrower_mask> & <SVE_2BHSI:self_mask>) == 0"
   1683   "@
   1684    ld1<ANY_EXTEND:s><SVE_2BHSI:Vesize>\t%0.d, %5/z, [%1, %2.d, uxtw]
   1685    ld1<ANY_EXTEND:s><SVE_2BHSI:Vesize>\t%0.d, %5/z, [%1, %2.d, uxtw %p4]"
   1686   "&& !CONSTANT_P (operands[7])"
   1687   {
   1688     operands[7] = CONSTM1_RTX (VNx2BImode);
   1689   }
   1690 )
   1691 
   1692 ;; -------------------------------------------------------------------------
   1693 ;; ---- First-faulting gather loads
   1694 ;; -------------------------------------------------------------------------
   1695 ;; Includes gather forms of:
   1696 ;; - LDFF1D
   1697 ;; - LDFF1W
   1698 ;; -------------------------------------------------------------------------
   1699 
   1700 ;; Predicated first-faulting gather loads for 32-bit elements.  Operand
   1701 ;; 3 is true for unsigned extension and false for signed extension.
   1702 (define_insn "@aarch64_ldff1_gather<mode>"
   1703   [(set (match_operand:SVE_FULL_S 0 "register_operand" "=w, w, w, w, w, w")
   1704 	(unspec:SVE_FULL_S
   1705 	  [(match_operand:VNx4BI 5 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   1706 	   (match_operand:DI 1 "aarch64_sve_gather_offset_w" "Z, vgw, rk, rk, rk, rk")
   1707 	   (match_operand:VNx4SI 2 "register_operand" "w, w, w, w, w, w")
   1708 	   (match_operand:DI 3 "const_int_operand" "i, i, Z, Ui1, Z, Ui1")
   1709 	   (match_operand:DI 4 "aarch64_gather_scale_operand_w" "Ui1, Ui1, Ui1, Ui1, i, i")
   1710 	   (mem:BLK (scratch))
   1711 	   (reg:VNx16BI FFRT_REGNUM)]
   1712 	  UNSPEC_LDFF1_GATHER))]
   1713   "TARGET_SVE"
   1714   "@
   1715    ldff1w\t%0.s, %5/z, [%2.s]
   1716    ldff1w\t%0.s, %5/z, [%2.s, #%1]
   1717    ldff1w\t%0.s, %5/z, [%1, %2.s, sxtw]
   1718    ldff1w\t%0.s, %5/z, [%1, %2.s, uxtw]
   1719    ldff1w\t%0.s, %5/z, [%1, %2.s, sxtw %p4]
   1720    ldff1w\t%0.s, %5/z, [%1, %2.s, uxtw %p4]"
   1721 )
   1722 
   1723 ;; Predicated first-faulting gather loads for 64-bit elements.  The value
   1724 ;; of operand 3 doesn't matter in this case.
   1725 (define_insn "@aarch64_ldff1_gather<mode>"
   1726   [(set (match_operand:SVE_FULL_D 0 "register_operand" "=w, w, w, w")
   1727 	(unspec:SVE_FULL_D
   1728 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl, Upl, Upl")
   1729 	   (match_operand:DI 1 "aarch64_sve_gather_offset_d" "Z, vgd, rk, rk")
   1730 	   (match_operand:VNx2DI 2 "register_operand" "w, w, w, w")
   1731 	   (match_operand:DI 3 "const_int_operand")
   1732 	   (match_operand:DI 4 "aarch64_gather_scale_operand_d" "Ui1, Ui1, Ui1, i")
   1733 	   (mem:BLK (scratch))
   1734 	   (reg:VNx16BI FFRT_REGNUM)]
   1735 	  UNSPEC_LDFF1_GATHER))]
   1736   "TARGET_SVE"
   1737   "@
   1738    ldff1d\t%0.d, %5/z, [%2.d]
   1739    ldff1d\t%0.d, %5/z, [%2.d, #%1]
   1740    ldff1d\t%0.d, %5/z, [%1, %2.d]
   1741    ldff1d\t%0.d, %5/z, [%1, %2.d, lsl %p4]"
   1742 )
   1743 
   1744 ;; Likewise, but with the offset being sign-extended from 32 bits.
   1745 (define_insn_and_rewrite "*aarch64_ldff1_gather<mode>_sxtw"
   1746   [(set (match_operand:SVE_FULL_D 0 "register_operand" "=w, w")
   1747 	(unspec:SVE_FULL_D
   1748 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   1749 	   (match_operand:DI 1 "register_operand" "rk, rk")
   1750 	   (unspec:VNx2DI
   1751 	     [(match_operand 6)
   1752 	      (sign_extend:VNx2DI
   1753 		(truncate:VNx2SI
   1754 		  (match_operand:VNx2DI 2 "register_operand" "w, w")))]
   1755 	     UNSPEC_PRED_X)
   1756 	   (match_operand:DI 3 "const_int_operand")
   1757 	   (match_operand:DI 4 "aarch64_gather_scale_operand_d" "Ui1, i")
   1758 	   (mem:BLK (scratch))
   1759 	   (reg:VNx16BI FFRT_REGNUM)]
   1760 	  UNSPEC_LDFF1_GATHER))]
   1761   "TARGET_SVE"
   1762   "@
   1763    ldff1d\t%0.d, %5/z, [%1, %2.d, sxtw]
   1764    ldff1d\t%0.d, %5/z, [%1, %2.d, sxtw %p4]"
   1765   "&& !CONSTANT_P (operands[6])"
   1766   {
   1767     operands[6] = CONSTM1_RTX (VNx2BImode);
   1768   }
   1769 )
   1770 
   1771 ;; Likewise, but with the offset being zero-extended from 32 bits.
   1772 (define_insn "*aarch64_ldff1_gather<mode>_uxtw"
   1773   [(set (match_operand:SVE_FULL_D 0 "register_operand" "=w, w")
   1774 	(unspec:SVE_FULL_D
   1775 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   1776 	   (match_operand:DI 1 "register_operand" "rk, rk")
   1777 	   (and:VNx2DI
   1778 	     (match_operand:VNx2DI 2 "register_operand" "w, w")
   1779 	     (match_operand:VNx2DI 6 "aarch64_sve_uxtw_immediate"))
   1780 	   (match_operand:DI 3 "const_int_operand")
   1781 	   (match_operand:DI 4 "aarch64_gather_scale_operand_d" "Ui1, i")
   1782 	   (mem:BLK (scratch))
   1783 	   (reg:VNx16BI FFRT_REGNUM)]
   1784 	  UNSPEC_LDFF1_GATHER))]
   1785   "TARGET_SVE"
   1786   "@
   1787    ldff1d\t%0.d, %5/z, [%1, %2.d, uxtw]
   1788    ldff1d\t%0.d, %5/z, [%1, %2.d, uxtw %p4]"
   1789 )
   1790 
   1791 ;; -------------------------------------------------------------------------
   1792 ;; ---- First-faulting extending gather loads
   1793 ;; -------------------------------------------------------------------------
   1794 ;; Includes gather forms of:
   1795 ;; - LDFF1B
   1796 ;; - LDFF1H
   1797 ;; - LDFF1SB
   1798 ;; - LDFF1SH
   1799 ;; - LDFF1SW
   1800 ;; - LDFF1W
   1801 ;; -------------------------------------------------------------------------
   1802 
   1803 ;; Predicated extending first-faulting gather loads for 32-bit elements.
   1804 ;; Operand 3 is true for unsigned extension and false for signed extension.
   1805 (define_insn_and_rewrite "@aarch64_ldff1_gather_<ANY_EXTEND:optab><VNx4_WIDE:mode><VNx4_NARROW:mode>"
   1806   [(set (match_operand:VNx4_WIDE 0 "register_operand" "=w, w, w, w, w, w")
   1807 	(unspec:VNx4_WIDE
   1808 	  [(match_operand:VNx4BI 6 "general_operand" "UplDnm, UplDnm, UplDnm, UplDnm, UplDnm, UplDnm")
   1809 	   (ANY_EXTEND:VNx4_WIDE
   1810 	     (unspec:VNx4_NARROW
   1811 	       [(match_operand:VNx4BI 5 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   1812 		(match_operand:DI 1 "aarch64_sve_gather_offset_<VNx4_NARROW:Vesize>" "Z, vg<VNx4_NARROW:Vesize>, rk, rk, rk, rk")
   1813 		(match_operand:VNx4_WIDE 2 "register_operand" "w, w, w, w, w, w")
   1814 		(match_operand:DI 3 "const_int_operand" "i, i, Z, Ui1, Z, Ui1")
   1815 		(match_operand:DI 4 "aarch64_gather_scale_operand_<VNx4_NARROW:Vesize>" "Ui1, Ui1, Ui1, Ui1, i, i")
   1816 		(mem:BLK (scratch))
   1817 		(reg:VNx16BI FFRT_REGNUM)]
   1818 	       UNSPEC_LDFF1_GATHER))]
   1819 	  UNSPEC_PRED_X))]
   1820   "TARGET_SVE"
   1821   "@
   1822    ldff1<ANY_EXTEND:s><VNx4_NARROW:Vesize>\t%0.s, %5/z, [%2.s]
   1823    ldff1<ANY_EXTEND:s><VNx4_NARROW:Vesize>\t%0.s, %5/z, [%2.s, #%1]
   1824    ldff1<ANY_EXTEND:s><VNx4_NARROW:Vesize>\t%0.s, %5/z, [%1, %2.s, sxtw]
   1825    ldff1<ANY_EXTEND:s><VNx4_NARROW:Vesize>\t%0.s, %5/z, [%1, %2.s, uxtw]
   1826    ldff1<ANY_EXTEND:s><VNx4_NARROW:Vesize>\t%0.s, %5/z, [%1, %2.s, sxtw %p4]
   1827    ldff1<ANY_EXTEND:s><VNx4_NARROW:Vesize>\t%0.s, %5/z, [%1, %2.s, uxtw %p4]"
   1828   "&& !CONSTANT_P (operands[6])"
   1829   {
   1830     operands[6] = CONSTM1_RTX (VNx4BImode);
   1831   }
   1832 )
   1833 
   1834 ;; Predicated extending first-faulting gather loads for 64-bit elements.
   1835 ;; The value of operand 3 doesn't matter in this case.
   1836 (define_insn_and_rewrite "@aarch64_ldff1_gather_<ANY_EXTEND:optab><VNx2_WIDE:mode><VNx2_NARROW:mode>"
   1837   [(set (match_operand:VNx2_WIDE 0 "register_operand" "=w, w, w, w")
   1838 	(unspec:VNx2_WIDE
   1839 	  [(match_operand:VNx2BI 6 "general_operand" "UplDnm, UplDnm, UplDnm, UplDnm")
   1840 	   (ANY_EXTEND:VNx2_WIDE
   1841 	     (unspec:VNx2_NARROW
   1842 	       [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl, Upl, Upl")
   1843 		(match_operand:DI 1 "aarch64_sve_gather_offset_<VNx2_NARROW:Vesize>" "Z, vg<VNx2_NARROW:Vesize>, rk, rk")
   1844 		(match_operand:VNx2_WIDE 2 "register_operand" "w, w, w, w")
   1845 		(match_operand:DI 3 "const_int_operand")
   1846 		(match_operand:DI 4 "aarch64_gather_scale_operand_<VNx2_NARROW:Vesize>" "Ui1, Ui1, Ui1, i")
   1847 		(mem:BLK (scratch))
   1848 		(reg:VNx16BI FFRT_REGNUM)]
   1849 	       UNSPEC_LDFF1_GATHER))]
   1850 	  UNSPEC_PRED_X))]
   1851   "TARGET_SVE"
   1852   "@
   1853    ldff1<ANY_EXTEND:s><VNx2_NARROW:Vesize>\t%0.d, %5/z, [%2.d]
   1854    ldff1<ANY_EXTEND:s><VNx2_NARROW:Vesize>\t%0.d, %5/z, [%2.d, #%1]
   1855    ldff1<ANY_EXTEND:s><VNx2_NARROW:Vesize>\t%0.d, %5/z, [%1, %2.d]
   1856    ldff1<ANY_EXTEND:s><VNx2_NARROW:Vesize>\t%0.d, %5/z, [%1, %2.d, lsl %p4]"
   1857   "&& !CONSTANT_P (operands[6])"
   1858   {
   1859     operands[6] = CONSTM1_RTX (VNx2BImode);
   1860   }
   1861 )
   1862 
   1863 ;; Likewise, but with the offset being sign-extended from 32 bits.
   1864 (define_insn_and_rewrite "*aarch64_ldff1_gather_<ANY_EXTEND:optab><VNx2_WIDE:mode><VNx2_NARROW:mode>_sxtw"
   1865   [(set (match_operand:VNx2_WIDE 0 "register_operand" "=w, w")
   1866 	(unspec:VNx2_WIDE
   1867 	  [(match_operand 6)
   1868 	   (ANY_EXTEND:VNx2_WIDE
   1869 	     (unspec:VNx2_NARROW
   1870 	       [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   1871 		(match_operand:DI 1 "aarch64_reg_or_zero" "rk, rk")
   1872 		(unspec:VNx2DI
   1873 		  [(match_operand 7)
   1874 		   (sign_extend:VNx2DI
   1875 		     (truncate:VNx2SI
   1876 		       (match_operand:VNx2DI 2 "register_operand" "w, w")))]
   1877 		  UNSPEC_PRED_X)
   1878 		(match_operand:DI 3 "const_int_operand")
   1879 		(match_operand:DI 4 "aarch64_gather_scale_operand_<VNx2_NARROW:Vesize>" "Ui1, i")
   1880 		(mem:BLK (scratch))
   1881 		(reg:VNx16BI FFRT_REGNUM)]
   1882 	       UNSPEC_LDFF1_GATHER))]
   1883 	  UNSPEC_PRED_X))]
   1884   "TARGET_SVE"
   1885   "@
   1886    ldff1<ANY_EXTEND:s><VNx2_NARROW:Vesize>\t%0.d, %5/z, [%1, %2.d, sxtw]
   1887    ldff1<ANY_EXTEND:s><VNx2_NARROW:Vesize>\t%0.d, %5/z, [%1, %2.d, sxtw %p4]"
   1888   "&& (!CONSTANT_P (operands[6]) || !CONSTANT_P (operands[7]))"
   1889   {
   1890     operands[6] = CONSTM1_RTX (VNx2BImode);
   1891     operands[7] = CONSTM1_RTX (VNx2BImode);
   1892   }
   1893 )
   1894 
   1895 ;; Likewise, but with the offset being zero-extended from 32 bits.
   1896 (define_insn_and_rewrite "*aarch64_ldff1_gather_<ANY_EXTEND:optab><VNx2_WIDE:mode><VNx2_NARROW:mode>_uxtw"
   1897   [(set (match_operand:VNx2_WIDE 0 "register_operand" "=w, w")
   1898 	(unspec:VNx2_WIDE
   1899 	  [(match_operand 7)
   1900 	   (ANY_EXTEND:VNx2_WIDE
   1901 	     (unspec:VNx2_NARROW
   1902 	       [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   1903 		(match_operand:DI 1 "aarch64_reg_or_zero" "rk, rk")
   1904 		(and:VNx2DI
   1905 		  (match_operand:VNx2DI 2 "register_operand" "w, w")
   1906 		  (match_operand:VNx2DI 6 "aarch64_sve_uxtw_immediate"))
   1907 		(match_operand:DI 3 "const_int_operand")
   1908 		(match_operand:DI 4 "aarch64_gather_scale_operand_<VNx2_NARROW:Vesize>" "Ui1, i")
   1909 		(mem:BLK (scratch))
   1910 		(reg:VNx16BI FFRT_REGNUM)]
   1911 	       UNSPEC_LDFF1_GATHER))]
   1912 	  UNSPEC_PRED_X))]
   1913   "TARGET_SVE"
   1914   "@
   1915    ldff1<ANY_EXTEND:s><VNx2_NARROW:Vesize>\t%0.d, %5/z, [%1, %2.d, uxtw]
   1916    ldff1<ANY_EXTEND:s><VNx2_NARROW:Vesize>\t%0.d, %5/z, [%1, %2.d, uxtw %p4]"
   1917   "&& !CONSTANT_P (operands[7])"
   1918   {
   1919     operands[7] = CONSTM1_RTX (VNx2BImode);
   1920   }
   1921 )
   1922 
   1923 ;; =========================================================================
   1924 ;; == Prefetches
   1925 ;; =========================================================================
   1926 
   1927 ;; -------------------------------------------------------------------------
   1928 ;; ---- Contiguous prefetches
   1929 ;; -------------------------------------------------------------------------
   1930 ;; Includes contiguous forms of:
   1931 ;; - PRFB
   1932 ;; - PRFD
   1933 ;; - PRFH
   1934 ;; - PRFW
   1935 ;; -------------------------------------------------------------------------
   1936 
   1937 ;; Contiguous predicated prefetches.  Operand 2 gives the real prefetch
   1938 ;; operation (as an svprfop), with operands 3 and 4 providing distilled
   1939 ;; information.
   1940 (define_insn "@aarch64_sve_prefetch<mode>"
   1941   [(prefetch (unspec:DI
   1942 	       [(match_operand:<VPRED> 0 "register_operand" "Upl")
   1943 		(match_operand:SVE_FULL_I 1 "aarch64_sve_prefetch_operand" "UP<Vesize>")
   1944 		(match_operand:DI 2 "const_int_operand")]
   1945 	       UNSPEC_SVE_PREFETCH)
   1946 	     (match_operand:DI 3 "const_int_operand")
   1947 	     (match_operand:DI 4 "const_int_operand"))]
   1948   "TARGET_SVE"
   1949   {
   1950     operands[1] = gen_rtx_MEM (<MODE>mode, operands[1]);
   1951     return aarch64_output_sve_prefetch ("prf<Vesize>", operands[2], "%0, %1");
   1952   }
   1953 )
   1954 
   1955 ;; -------------------------------------------------------------------------
   1956 ;; ---- Gather prefetches
   1957 ;; -------------------------------------------------------------------------
   1958 ;; Includes gather forms of:
   1959 ;; - PRFB
   1960 ;; - PRFD
   1961 ;; - PRFH
   1962 ;; - PRFW
   1963 ;; -------------------------------------------------------------------------
   1964 
   1965 ;; Predicated gather prefetches for 32-bit bases and offsets.  The operands
   1966 ;; are:
   1967 ;; 0: the governing predicate
   1968 ;; 1: the scalar component of the address
   1969 ;; 2: the vector component of the address
   1970 ;; 3: 1 for zero extension, 0 for sign extension
   1971 ;; 4: the scale multiplier
   1972 ;; 5: a vector zero that identifies the mode of data being accessed
   1973 ;; 6: the prefetch operator (an svprfop)
   1974 ;; 7: the normal RTL prefetch rw flag
   1975 ;; 8: the normal RTL prefetch locality value
   1976 (define_insn "@aarch64_sve_gather_prefetch<SVE_FULL_I:mode><VNx4SI_ONLY:mode>"
   1977   [(prefetch (unspec:DI
   1978 	       [(match_operand:VNx4BI 0 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   1979 		(match_operand:DI 1 "aarch64_sve_gather_offset_<SVE_FULL_I:Vesize>" "Z, vg<SVE_FULL_I:Vesize>, rk, rk, rk, rk")
   1980 		(match_operand:VNx4SI_ONLY 2 "register_operand" "w, w, w, w, w, w")
   1981 		(match_operand:DI 3 "const_int_operand" "i, i, Z, Ui1, Z, Ui1")
   1982 		(match_operand:DI 4 "aarch64_gather_scale_operand_<SVE_FULL_I:Vesize>" "Ui1, Ui1, Ui1, Ui1, i, i")
   1983 		(match_operand:SVE_FULL_I 5 "aarch64_simd_imm_zero")
   1984 		(match_operand:DI 6 "const_int_operand")]
   1985 	       UNSPEC_SVE_PREFETCH_GATHER)
   1986 	     (match_operand:DI 7 "const_int_operand")
   1987 	     (match_operand:DI 8 "const_int_operand"))]
   1988   "TARGET_SVE"
   1989   {
   1990     static const char *const insns[][2] = {
   1991       "prf<SVE_FULL_I:Vesize>", "%0, [%2.s]",
   1992       "prf<SVE_FULL_I:Vesize>", "%0, [%2.s, #%1]",
   1993       "prfb", "%0, [%1, %2.s, sxtw]",
   1994       "prfb", "%0, [%1, %2.s, uxtw]",
   1995       "prf<SVE_FULL_I:Vesize>", "%0, [%1, %2.s, sxtw %p4]",
   1996       "prf<SVE_FULL_I:Vesize>", "%0, [%1, %2.s, uxtw %p4]"
   1997     };
   1998     const char *const *parts = insns[which_alternative];
   1999     return aarch64_output_sve_prefetch (parts[0], operands[6], parts[1]);
   2000   }
   2001 )
   2002 
   2003 ;; Predicated gather prefetches for 64-bit elements.  The value of operand 3
   2004 ;; doesn't matter in this case.
   2005 (define_insn "@aarch64_sve_gather_prefetch<SVE_FULL_I:mode><VNx2DI_ONLY:mode>"
   2006   [(prefetch (unspec:DI
   2007 	       [(match_operand:VNx2BI 0 "register_operand" "Upl, Upl, Upl, Upl")
   2008 		(match_operand:DI 1 "aarch64_sve_gather_offset_<SVE_FULL_I:Vesize>" "Z, vg<SVE_FULL_I:Vesize>, rk, rk")
   2009 		(match_operand:VNx2DI_ONLY 2 "register_operand" "w, w, w, w")
   2010 		(match_operand:DI 3 "const_int_operand")
   2011 		(match_operand:DI 4 "aarch64_gather_scale_operand_<SVE_FULL_I:Vesize>" "Ui1, Ui1, Ui1, i")
   2012 		(match_operand:SVE_FULL_I 5 "aarch64_simd_imm_zero")
   2013 		(match_operand:DI 6 "const_int_operand")]
   2014 	       UNSPEC_SVE_PREFETCH_GATHER)
   2015 	     (match_operand:DI 7 "const_int_operand")
   2016 	     (match_operand:DI 8 "const_int_operand"))]
   2017   "TARGET_SVE"
   2018   {
   2019     static const char *const insns[][2] = {
   2020       "prf<SVE_FULL_I:Vesize>", "%0, [%2.d]",
   2021       "prf<SVE_FULL_I:Vesize>", "%0, [%2.d, #%1]",
   2022       "prfb", "%0, [%1, %2.d]",
   2023       "prf<SVE_FULL_I:Vesize>", "%0, [%1, %2.d, lsl %p4]"
   2024     };
   2025     const char *const *parts = insns[which_alternative];
   2026     return aarch64_output_sve_prefetch (parts[0], operands[6], parts[1]);
   2027   }
   2028 )
   2029 
   2030 ;; Likewise, but with the offset being sign-extended from 32 bits.
   2031 (define_insn_and_rewrite "*aarch64_sve_gather_prefetch<SVE_FULL_I:mode><VNx2DI_ONLY:mode>_sxtw"
   2032   [(prefetch (unspec:DI
   2033 	       [(match_operand:VNx2BI 0 "register_operand" "Upl, Upl")
   2034 		(match_operand:DI 1 "register_operand" "rk, rk")
   2035 		(unspec:VNx2DI_ONLY
   2036 		  [(match_operand 9)
   2037 		   (sign_extend:VNx2DI
   2038 		     (truncate:VNx2SI
   2039 		       (match_operand:VNx2DI 2 "register_operand" "w, w")))]
   2040 		  UNSPEC_PRED_X)
   2041 		(match_operand:DI 3 "const_int_operand")
   2042 		(match_operand:DI 4 "aarch64_gather_scale_operand_<SVE_FULL_I:Vesize>" "Ui1, i")
   2043 		(match_operand:SVE_FULL_I 5 "aarch64_simd_imm_zero")
   2044 		(match_operand:DI 6 "const_int_operand")]
   2045 	       UNSPEC_SVE_PREFETCH_GATHER)
   2046 	     (match_operand:DI 7 "const_int_operand")
   2047 	     (match_operand:DI 8 "const_int_operand"))]
   2048   "TARGET_SVE"
   2049   {
   2050     static const char *const insns[][2] = {
   2051       "prfb", "%0, [%1, %2.d, sxtw]",
   2052       "prf<SVE_FULL_I:Vesize>", "%0, [%1, %2.d, sxtw %p4]"
   2053     };
   2054     const char *const *parts = insns[which_alternative];
   2055     return aarch64_output_sve_prefetch (parts[0], operands[6], parts[1]);
   2056   }
   2057   "&& !rtx_equal_p (operands[0], operands[9])"
   2058   {
   2059     operands[9] = copy_rtx (operands[0]);
   2060   }
   2061 )
   2062 
   2063 ;; Likewise, but with the offset being zero-extended from 32 bits.
   2064 (define_insn "*aarch64_sve_gather_prefetch<SVE_FULL_I:mode><VNx2DI_ONLY:mode>_uxtw"
   2065   [(prefetch (unspec:DI
   2066 	       [(match_operand:VNx2BI 0 "register_operand" "Upl, Upl")
   2067 		(match_operand:DI 1 "register_operand" "rk, rk")
   2068 		(and:VNx2DI_ONLY
   2069 		  (match_operand:VNx2DI 2 "register_operand" "w, w")
   2070 		  (match_operand:VNx2DI 9 "aarch64_sve_uxtw_immediate"))
   2071 		(match_operand:DI 3 "const_int_operand")
   2072 		(match_operand:DI 4 "aarch64_gather_scale_operand_<SVE_FULL_I:Vesize>" "Ui1, i")
   2073 		(match_operand:SVE_FULL_I 5 "aarch64_simd_imm_zero")
   2074 		(match_operand:DI 6 "const_int_operand")]
   2075 	       UNSPEC_SVE_PREFETCH_GATHER)
   2076 	     (match_operand:DI 7 "const_int_operand")
   2077 	     (match_operand:DI 8 "const_int_operand"))]
   2078   "TARGET_SVE"
   2079   {
   2080     static const char *const insns[][2] = {
   2081       "prfb", "%0, [%1, %2.d, uxtw]",
   2082       "prf<SVE_FULL_I:Vesize>", "%0, [%1, %2.d, uxtw %p4]"
   2083     };
   2084     const char *const *parts = insns[which_alternative];
   2085     return aarch64_output_sve_prefetch (parts[0], operands[6], parts[1]);
   2086   }
   2087 )
   2088 
   2089 ;; =========================================================================
   2090 ;; == Stores
   2091 ;; =========================================================================
   2092 
   2093 ;; -------------------------------------------------------------------------
   2094 ;; ---- Normal contiguous stores
   2095 ;; -------------------------------------------------------------------------
   2096 ;; Includes contiguous forms of:
   2097 ;; - ST1B
   2098 ;; - ST1D
   2099 ;; - ST1H
   2100 ;; - ST1W
   2101 ;; - ST2B
   2102 ;; - ST2D
   2103 ;; - ST2H
   2104 ;; - ST2W
   2105 ;; - ST3B
   2106 ;; - ST3D
   2107 ;; - ST3H
   2108 ;; - ST3W
   2109 ;; - ST4B
   2110 ;; - ST4D
   2111 ;; - ST4H
   2112 ;; - ST4W
   2113 ;; -------------------------------------------------------------------------
   2114 
   2115 ;; Predicated ST1.
   2116 (define_insn "maskstore<mode><vpred>"
   2117   [(set (match_operand:SVE_ALL 0 "memory_operand" "+m")
   2118 	(unspec:SVE_ALL
   2119 	  [(match_operand:<VPRED> 2 "register_operand" "Upl")
   2120 	   (match_operand:SVE_ALL 1 "register_operand" "w")
   2121 	   (match_dup 0)]
   2122 	  UNSPEC_ST1_SVE))]
   2123   "TARGET_SVE"
   2124   "st1<Vesize>\t%1.<Vctype>, %2, %0"
   2125 )
   2126 
   2127 ;; Unpredicated ST[234].  This is always a full update, so the dependence
   2128 ;; on the old value of the memory location (via (match_dup 0)) is redundant.
   2129 ;; There doesn't seem to be any obvious benefit to treating the all-true
   2130 ;; case differently though.  In particular, it's very unlikely that we'll
   2131 ;; only find out during RTL that a store_lanes is dead.
   2132 (define_expand "vec_store_lanes<mode><vsingle>"
   2133   [(set (match_operand:SVE_STRUCT 0 "memory_operand")
   2134 	(unspec:SVE_STRUCT
   2135 	  [(match_dup 2)
   2136 	   (match_operand:SVE_STRUCT 1 "register_operand")
   2137 	   (match_dup 0)]
   2138 	  UNSPEC_STN))]
   2139   "TARGET_SVE"
   2140   {
   2141     operands[2] = aarch64_ptrue_reg (<VPRED>mode);
   2142   }
   2143 )
   2144 
   2145 ;; Predicated ST[234].
   2146 (define_insn "vec_mask_store_lanes<mode><vsingle>"
   2147   [(set (match_operand:SVE_STRUCT 0 "memory_operand" "+m")
   2148 	(unspec:SVE_STRUCT
   2149 	  [(match_operand:<VPRED> 2 "register_operand" "Upl")
   2150 	   (match_operand:SVE_STRUCT 1 "register_operand" "w")
   2151 	   (match_dup 0)]
   2152 	  UNSPEC_STN))]
   2153   "TARGET_SVE"
   2154   "st<vector_count><Vesize>\t%1, %2, %0"
   2155 )
   2156 
   2157 ;; -------------------------------------------------------------------------
   2158 ;; ---- Truncating contiguous stores
   2159 ;; -------------------------------------------------------------------------
   2160 ;; Includes:
   2161 ;; - ST1B
   2162 ;; - ST1H
   2163 ;; - ST1W
   2164 ;; -------------------------------------------------------------------------
   2165 
   2166 ;; Predicated truncate and store, with 8 elements per 128-bit block.
   2167 (define_insn "@aarch64_store_trunc<VNx8_NARROW:mode><VNx8_WIDE:mode>"
   2168   [(set (match_operand:VNx8_NARROW 0 "memory_operand" "+m")
   2169 	(unspec:VNx8_NARROW
   2170 	  [(match_operand:VNx8BI 2 "register_operand" "Upl")
   2171 	   (truncate:VNx8_NARROW
   2172 	     (match_operand:VNx8_WIDE 1 "register_operand" "w"))
   2173 	   (match_dup 0)]
   2174 	  UNSPEC_ST1_SVE))]
   2175   "TARGET_SVE"
   2176   "st1<VNx8_NARROW:Vesize>\t%1.<VNx8_WIDE:Vetype>, %2, %0"
   2177 )
   2178 
   2179 ;; Predicated truncate and store, with 4 elements per 128-bit block.
   2180 (define_insn "@aarch64_store_trunc<VNx4_NARROW:mode><VNx4_WIDE:mode>"
   2181   [(set (match_operand:VNx4_NARROW 0 "memory_operand" "+m")
   2182 	(unspec:VNx4_NARROW
   2183 	  [(match_operand:VNx4BI 2 "register_operand" "Upl")
   2184 	   (truncate:VNx4_NARROW
   2185 	     (match_operand:VNx4_WIDE 1 "register_operand" "w"))
   2186 	   (match_dup 0)]
   2187 	  UNSPEC_ST1_SVE))]
   2188   "TARGET_SVE"
   2189   "st1<VNx4_NARROW:Vesize>\t%1.<VNx4_WIDE:Vetype>, %2, %0"
   2190 )
   2191 
   2192 ;; Predicated truncate and store, with 2 elements per 128-bit block.
   2193 (define_insn "@aarch64_store_trunc<VNx2_NARROW:mode><VNx2_WIDE:mode>"
   2194   [(set (match_operand:VNx2_NARROW 0 "memory_operand" "+m")
   2195 	(unspec:VNx2_NARROW
   2196 	  [(match_operand:VNx2BI 2 "register_operand" "Upl")
   2197 	   (truncate:VNx2_NARROW
   2198 	     (match_operand:VNx2_WIDE 1 "register_operand" "w"))
   2199 	   (match_dup 0)]
   2200 	  UNSPEC_ST1_SVE))]
   2201   "TARGET_SVE"
   2202   "st1<VNx2_NARROW:Vesize>\t%1.<VNx2_WIDE:Vetype>, %2, %0"
   2203 )
   2204 
   2205 ;; -------------------------------------------------------------------------
   2206 ;; ---- Non-temporal contiguous stores
   2207 ;; -------------------------------------------------------------------------
   2208 ;; Includes:
   2209 ;; - STNT1B
   2210 ;; - STNT1D
   2211 ;; - STNT1H
   2212 ;; - STNT1W
   2213 ;; -------------------------------------------------------------------------
   2214 
   2215 (define_insn "@aarch64_stnt1<mode>"
   2216   [(set (match_operand:SVE_FULL 0 "memory_operand" "+m")
   2217 	(unspec:SVE_FULL
   2218 	  [(match_operand:<VPRED> 2 "register_operand" "Upl")
   2219 	   (match_operand:SVE_FULL 1 "register_operand" "w")
   2220 	   (match_dup 0)]
   2221 	  UNSPEC_STNT1_SVE))]
   2222   "TARGET_SVE"
   2223   "stnt1<Vesize>\t%1.<Vetype>, %2, %0"
   2224 )
   2225 
   2226 ;; -------------------------------------------------------------------------
   2227 ;; ---- Normal scatter stores
   2228 ;; -------------------------------------------------------------------------
   2229 ;; Includes scatter forms of:
   2230 ;; - ST1D
   2231 ;; - ST1W
   2232 ;; -------------------------------------------------------------------------
   2233 
   2234 ;; Unpredicated scatter stores.
   2235 (define_expand "scatter_store<mode><v_int_container>"
   2236   [(set (mem:BLK (scratch))
   2237 	(unspec:BLK
   2238 	  [(match_dup 5)
   2239 	   (match_operand:DI 0 "aarch64_sve_gather_offset_<Vesize>")
   2240 	   (match_operand:<V_INT_CONTAINER> 1 "register_operand")
   2241 	   (match_operand:DI 2 "const_int_operand")
   2242 	   (match_operand:DI 3 "aarch64_gather_scale_operand_<Vesize>")
   2243 	   (match_operand:SVE_24 4 "register_operand")]
   2244 	  UNSPEC_ST1_SCATTER))]
   2245   "TARGET_SVE"
   2246   {
   2247     operands[5] = aarch64_ptrue_reg (<VPRED>mode);
   2248   }
   2249 )
   2250 
   2251 ;; Predicated scatter stores for 32-bit elements.  Operand 2 is true for
   2252 ;; unsigned extension and false for signed extension.
   2253 (define_insn "mask_scatter_store<mode><v_int_container>"
   2254   [(set (mem:BLK (scratch))
   2255 	(unspec:BLK
   2256 	  [(match_operand:VNx4BI 5 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   2257 	   (match_operand:DI 0 "aarch64_sve_gather_offset_<Vesize>" "Z, vgw, rk, rk, rk, rk")
   2258 	   (match_operand:VNx4SI 1 "register_operand" "w, w, w, w, w, w")
   2259 	   (match_operand:DI 2 "const_int_operand" "Ui1, Ui1, Z, Ui1, Z, Ui1")
   2260 	   (match_operand:DI 3 "aarch64_gather_scale_operand_<Vesize>" "Ui1, Ui1, Ui1, Ui1, i, i")
   2261 	   (match_operand:SVE_4 4 "register_operand" "w, w, w, w, w, w")]
   2262 	  UNSPEC_ST1_SCATTER))]
   2263   "TARGET_SVE"
   2264   "@
   2265    st1<Vesize>\t%4.s, %5, [%1.s]
   2266    st1<Vesize>\t%4.s, %5, [%1.s, #%0]
   2267    st1<Vesize>\t%4.s, %5, [%0, %1.s, sxtw]
   2268    st1<Vesize>\t%4.s, %5, [%0, %1.s, uxtw]
   2269    st1<Vesize>\t%4.s, %5, [%0, %1.s, sxtw %p3]
   2270    st1<Vesize>\t%4.s, %5, [%0, %1.s, uxtw %p3]"
   2271 )
   2272 
   2273 ;; Predicated scatter stores for 64-bit elements.  The value of operand 2
   2274 ;; doesn't matter in this case.
   2275 (define_insn "mask_scatter_store<mode><v_int_container>"
   2276   [(set (mem:BLK (scratch))
   2277 	(unspec:BLK
   2278 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl, Upl, Upl")
   2279 	   (match_operand:DI 0 "aarch64_sve_gather_offset_<Vesize>" "Z, vgd, rk, rk")
   2280 	   (match_operand:VNx2DI 1 "register_operand" "w, w, w, w")
   2281 	   (match_operand:DI 2 "const_int_operand")
   2282 	   (match_operand:DI 3 "aarch64_gather_scale_operand_<Vesize>" "Ui1, Ui1, Ui1, i")
   2283 	   (match_operand:SVE_2 4 "register_operand" "w, w, w, w")]
   2284 	  UNSPEC_ST1_SCATTER))]
   2285   "TARGET_SVE"
   2286   "@
   2287    st1<Vesize>\t%4.d, %5, [%1.d]
   2288    st1<Vesize>\t%4.d, %5, [%1.d, #%0]
   2289    st1<Vesize>\t%4.d, %5, [%0, %1.d]
   2290    st1<Vesize>\t%4.d, %5, [%0, %1.d, lsl %p3]"
   2291 )
   2292 
   2293 ;; Likewise, but with the offset being extended from 32 bits.
   2294 (define_insn_and_rewrite "*mask_scatter_store<mode><v_int_container>_<su>xtw_unpacked"
   2295   [(set (mem:BLK (scratch))
   2296 	(unspec:BLK
   2297 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   2298 	   (match_operand:DI 0 "register_operand" "rk, rk")
   2299 	   (unspec:VNx2DI
   2300 	     [(match_operand 6)
   2301 	      (ANY_EXTEND:VNx2DI
   2302 		(match_operand:VNx2SI 1 "register_operand" "w, w"))]
   2303 	     UNSPEC_PRED_X)
   2304 	   (match_operand:DI 2 "const_int_operand")
   2305 	   (match_operand:DI 3 "aarch64_gather_scale_operand_<Vesize>" "Ui1, i")
   2306 	   (match_operand:SVE_2 4 "register_operand" "w, w")]
   2307 	  UNSPEC_ST1_SCATTER))]
   2308   "TARGET_SVE"
   2309   "@
   2310    st1<Vesize>\t%4.d, %5, [%0, %1.d, <su>xtw]
   2311    st1<Vesize>\t%4.d, %5, [%0, %1.d, <su>xtw %p3]"
   2312   "&& !CONSTANT_P (operands[6])"
   2313   {
   2314     operands[6] = CONSTM1_RTX (<VPRED>mode);
   2315   }
   2316 )
   2317 
   2318 ;; Likewise, but with the offset being truncated to 32 bits and then
   2319 ;; sign-extended.
   2320 (define_insn_and_rewrite "*mask_scatter_store<mode><v_int_container>_sxtw"
   2321   [(set (mem:BLK (scratch))
   2322 	(unspec:BLK
   2323 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   2324 	   (match_operand:DI 0 "register_operand" "rk, rk")
   2325 	   (unspec:VNx2DI
   2326 	     [(match_operand 6)
   2327 	      (sign_extend:VNx2DI
   2328 		(truncate:VNx2SI
   2329 		  (match_operand:VNx2DI 1 "register_operand" "w, w")))]
   2330 	     UNSPEC_PRED_X)
   2331 	   (match_operand:DI 2 "const_int_operand")
   2332 	   (match_operand:DI 3 "aarch64_gather_scale_operand_<Vesize>" "Ui1, i")
   2333 	   (match_operand:SVE_2 4 "register_operand" "w, w")]
   2334 	  UNSPEC_ST1_SCATTER))]
   2335   "TARGET_SVE"
   2336   "@
   2337    st1<Vesize>\t%4.d, %5, [%0, %1.d, sxtw]
   2338    st1<Vesize>\t%4.d, %5, [%0, %1.d, sxtw %p3]"
   2339   "&& !CONSTANT_P (operands[6])"
   2340   {
   2341     operands[6] = CONSTM1_RTX (<VPRED>mode);
   2342   }
   2343 )
   2344 
   2345 ;; Likewise, but with the offset being truncated to 32 bits and then
   2346 ;; zero-extended.
   2347 (define_insn "*mask_scatter_store<mode><v_int_container>_uxtw"
   2348   [(set (mem:BLK (scratch))
   2349 	(unspec:BLK
   2350 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   2351 	   (match_operand:DI 0 "aarch64_reg_or_zero" "rk, rk")
   2352 	   (and:VNx2DI
   2353 	     (match_operand:VNx2DI 1 "register_operand" "w, w")
   2354 	     (match_operand:VNx2DI 6 "aarch64_sve_uxtw_immediate"))
   2355 	   (match_operand:DI 2 "const_int_operand")
   2356 	   (match_operand:DI 3 "aarch64_gather_scale_operand_<Vesize>" "Ui1, i")
   2357 	   (match_operand:SVE_2 4 "register_operand" "w, w")]
   2358 	  UNSPEC_ST1_SCATTER))]
   2359   "TARGET_SVE"
   2360   "@
   2361    st1<Vesize>\t%4.d, %5, [%0, %1.d, uxtw]
   2362    st1<Vesize>\t%4.d, %5, [%0, %1.d, uxtw %p3]"
   2363 )
   2364 
   2365 ;; -------------------------------------------------------------------------
   2366 ;; ---- Truncating scatter stores
   2367 ;; -------------------------------------------------------------------------
   2368 ;; Includes scatter forms of:
   2369 ;; - ST1B
   2370 ;; - ST1H
   2371 ;; - ST1W
   2372 ;; -------------------------------------------------------------------------
   2373 
   2374 ;; Predicated truncating scatter stores for 32-bit elements.  Operand 2 is
   2375 ;; true for unsigned extension and false for signed extension.
   2376 (define_insn "@aarch64_scatter_store_trunc<VNx4_NARROW:mode><VNx4_WIDE:mode>"
   2377   [(set (mem:BLK (scratch))
   2378 	(unspec:BLK
   2379 	  [(match_operand:VNx4BI 5 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   2380 	   (match_operand:DI 0 "aarch64_sve_gather_offset_<VNx4_NARROW:Vesize>" "Z, vg<VNx4_NARROW:Vesize>, rk, rk, rk, rk")
   2381 	   (match_operand:VNx4SI 1 "register_operand" "w, w, w, w, w, w")
   2382 	   (match_operand:DI 2 "const_int_operand" "Ui1, Ui1, Z, Ui1, Z, Ui1")
   2383 	   (match_operand:DI 3 "aarch64_gather_scale_operand_<VNx4_NARROW:Vesize>" "Ui1, Ui1, Ui1, Ui1, i, i")
   2384 	   (truncate:VNx4_NARROW
   2385 	     (match_operand:VNx4_WIDE 4 "register_operand" "w, w, w, w, w, w"))]
   2386 	  UNSPEC_ST1_SCATTER))]
   2387   "TARGET_SVE"
   2388   "@
   2389    st1<VNx4_NARROW:Vesize>\t%4.s, %5, [%1.s]
   2390    st1<VNx4_NARROW:Vesize>\t%4.s, %5, [%1.s, #%0]
   2391    st1<VNx4_NARROW:Vesize>\t%4.s, %5, [%0, %1.s, sxtw]
   2392    st1<VNx4_NARROW:Vesize>\t%4.s, %5, [%0, %1.s, uxtw]
   2393    st1<VNx4_NARROW:Vesize>\t%4.s, %5, [%0, %1.s, sxtw %p3]
   2394    st1<VNx4_NARROW:Vesize>\t%4.s, %5, [%0, %1.s, uxtw %p3]"
   2395 )
   2396 
   2397 ;; Predicated truncating scatter stores for 64-bit elements.  The value of
   2398 ;; operand 2 doesn't matter in this case.
   2399 (define_insn "@aarch64_scatter_store_trunc<VNx2_NARROW:mode><VNx2_WIDE:mode>"
   2400   [(set (mem:BLK (scratch))
   2401 	(unspec:BLK
   2402 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl, Upl, Upl")
   2403 	   (match_operand:DI 0 "aarch64_sve_gather_offset_<VNx2_NARROW:Vesize>" "Z, vg<VNx2_NARROW:Vesize>, rk, rk")
   2404 	   (match_operand:VNx2DI 1 "register_operand" "w, w, w, w")
   2405 	   (match_operand:DI 2 "const_int_operand")
   2406 	   (match_operand:DI 3 "aarch64_gather_scale_operand_<VNx2_NARROW:Vesize>" "Ui1, Ui1, Ui1, i")
   2407 	   (truncate:VNx2_NARROW
   2408 	     (match_operand:VNx2_WIDE 4 "register_operand" "w, w, w, w"))]
   2409 	  UNSPEC_ST1_SCATTER))]
   2410   "TARGET_SVE"
   2411   "@
   2412    st1<VNx2_NARROW:Vesize>\t%4.d, %5, [%1.d]
   2413    st1<VNx2_NARROW:Vesize>\t%4.d, %5, [%1.d, #%0]
   2414    st1<VNx2_NARROW:Vesize>\t%4.d, %5, [%0, %1.d]
   2415    st1<VNx2_NARROW:Vesize>\t%4.d, %5, [%0, %1.d, lsl %p3]"
   2416 )
   2417 
   2418 ;; Likewise, but with the offset being sign-extended from 32 bits.
   2419 (define_insn_and_rewrite "*aarch64_scatter_store_trunc<VNx2_NARROW:mode><VNx2_WIDE:mode>_sxtw"
   2420   [(set (mem:BLK (scratch))
   2421 	(unspec:BLK
   2422 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   2423 	   (match_operand:DI 0 "register_operand" "rk, rk")
   2424 	   (unspec:VNx2DI
   2425 	     [(match_operand 6)
   2426 	      (sign_extend:VNx2DI
   2427 		(truncate:VNx2SI
   2428 		  (match_operand:VNx2DI 1 "register_operand" "w, w")))]
   2429 	     UNSPEC_PRED_X)
   2430 	   (match_operand:DI 2 "const_int_operand")
   2431 	   (match_operand:DI 3 "aarch64_gather_scale_operand_<VNx2_NARROW:Vesize>" "Ui1, i")
   2432 	   (truncate:VNx2_NARROW
   2433 	     (match_operand:VNx2_WIDE 4 "register_operand" "w, w"))]
   2434 	  UNSPEC_ST1_SCATTER))]
   2435   "TARGET_SVE"
   2436   "@
   2437    st1<VNx2_NARROW:Vesize>\t%4.d, %5, [%0, %1.d, sxtw]
   2438    st1<VNx2_NARROW:Vesize>\t%4.d, %5, [%0, %1.d, sxtw %p3]"
   2439   "&& !rtx_equal_p (operands[5], operands[6])"
   2440   {
   2441     operands[6] = copy_rtx (operands[5]);
   2442   }
   2443 )
   2444 
   2445 ;; Likewise, but with the offset being zero-extended from 32 bits.
   2446 (define_insn "*aarch64_scatter_store_trunc<VNx2_NARROW:mode><VNx2_WIDE:mode>_uxtw"
   2447   [(set (mem:BLK (scratch))
   2448 	(unspec:BLK
   2449 	  [(match_operand:VNx2BI 5 "register_operand" "Upl, Upl")
   2450 	   (match_operand:DI 0 "aarch64_reg_or_zero" "rk, rk")
   2451 	   (and:VNx2DI
   2452 	     (match_operand:VNx2DI 1 "register_operand" "w, w")
   2453 	     (match_operand:VNx2DI 6 "aarch64_sve_uxtw_immediate"))
   2454 	   (match_operand:DI 2 "const_int_operand")
   2455 	   (match_operand:DI 3 "aarch64_gather_scale_operand_<VNx2_NARROW:Vesize>" "Ui1, i")
   2456 	   (truncate:VNx2_NARROW
   2457 	     (match_operand:VNx2_WIDE 4 "register_operand" "w, w"))]
   2458 	  UNSPEC_ST1_SCATTER))]
   2459   "TARGET_SVE"
   2460   "@
   2461    st1<VNx2_NARROW:Vesize>\t%4.d, %5, [%0, %1.d, uxtw]
   2462    st1<VNx2_NARROW:Vesize>\t%4.d, %5, [%0, %1.d, uxtw %p3]"
   2463 )
   2464 
   2465 ;; =========================================================================
   2466 ;; == Vector creation
   2467 ;; =========================================================================
   2468 
   2469 ;; -------------------------------------------------------------------------
   2470 ;; ---- [INT,FP] Duplicate element
   2471 ;; -------------------------------------------------------------------------
   2472 ;; Includes:
   2473 ;; - DUP
   2474 ;; - MOV
   2475 ;; - LD1RB
   2476 ;; - LD1RD
   2477 ;; - LD1RH
   2478 ;; - LD1RW
   2479 ;; - LD1ROB (F64MM)
   2480 ;; - LD1ROD (F64MM)
   2481 ;; - LD1ROH (F64MM)
   2482 ;; - LD1ROW (F64MM)
   2483 ;; - LD1RQB
   2484 ;; - LD1RQD
   2485 ;; - LD1RQH
   2486 ;; - LD1RQW
   2487 ;; -------------------------------------------------------------------------
   2488 
   2489 (define_expand "vec_duplicate<mode>"
   2490   [(parallel
   2491     [(set (match_operand:SVE_ALL 0 "register_operand")
   2492 	  (vec_duplicate:SVE_ALL
   2493 	    (match_operand:<VEL> 1 "aarch64_sve_dup_operand")))
   2494      (clobber (scratch:VNx16BI))])]
   2495   "TARGET_SVE"
   2496   {
   2497     if (MEM_P (operands[1]))
   2498       {
   2499 	rtx ptrue = aarch64_ptrue_reg (<VPRED>mode);
   2500 	emit_insn (gen_sve_ld1r<mode> (operands[0], ptrue, operands[1],
   2501 				       CONST0_RTX (<MODE>mode)));
   2502 	DONE;
   2503       }
   2504   }
   2505 )
   2506 
   2507 ;; Accept memory operands for the benefit of combine, and also in case
   2508 ;; the scalar input gets spilled to memory during RA.  We want to split
   2509 ;; the load at the first opportunity in order to allow the PTRUE to be
   2510 ;; optimized with surrounding code.
   2511 (define_insn_and_split "*vec_duplicate<mode>_reg"
   2512   [(set (match_operand:SVE_ALL 0 "register_operand" "=w, w, w")
   2513 	(vec_duplicate:SVE_ALL
   2514 	  (match_operand:<VEL> 1 "aarch64_sve_dup_operand" "r, w, Uty")))
   2515    (clobber (match_scratch:VNx16BI 2 "=X, X, Upl"))]
   2516   "TARGET_SVE"
   2517   "@
   2518    mov\t%0.<Vetype>, %<vwcore>1
   2519    mov\t%0.<Vetype>, %<Vetype>1
   2520    #"
   2521   "&& MEM_P (operands[1])"
   2522   [(const_int 0)]
   2523   {
   2524     if (GET_CODE (operands[2]) == SCRATCH)
   2525       operands[2] = gen_reg_rtx (VNx16BImode);
   2526     emit_move_insn (operands[2], CONSTM1_RTX (VNx16BImode));
   2527     rtx gp = gen_lowpart (<VPRED>mode, operands[2]);
   2528     emit_insn (gen_sve_ld1r<mode> (operands[0], gp, operands[1],
   2529 				   CONST0_RTX (<MODE>mode)));
   2530     DONE;
   2531   }
   2532   [(set_attr "length" "4,4,8")]
   2533 )
   2534 
   2535 ;; Duplicate an Advanced SIMD vector to fill an SVE vector (LE version).
   2536 (define_insn "@aarch64_vec_duplicate_vq<mode>_le"
   2537   [(set (match_operand:SVE_FULL 0 "register_operand" "=w")
   2538 	(vec_duplicate:SVE_FULL
   2539 	  (match_operand:<V128> 1 "register_operand" "w")))]
   2540   "TARGET_SVE && !BYTES_BIG_ENDIAN"
   2541   {
   2542     operands[1] = gen_rtx_REG (<MODE>mode, REGNO (operands[1]));
   2543     return "dup\t%0.q, %1.q[0]";
   2544   }
   2545 )
   2546 
   2547 ;; Duplicate an Advanced SIMD vector to fill an SVE vector (BE version).
   2548 ;; The SVE register layout puts memory lane N into (architectural)
   2549 ;; register lane N, whereas the Advanced SIMD layout puts the memory
   2550 ;; lsb into the register lsb.  We therefore have to describe this in rtl
   2551 ;; terms as a reverse of the V128 vector followed by a duplicate.
   2552 (define_insn "@aarch64_vec_duplicate_vq<mode>_be"
   2553   [(set (match_operand:SVE_FULL 0 "register_operand" "=w")
   2554 	(vec_duplicate:SVE_FULL
   2555 	  (vec_select:<V128>
   2556 	    (match_operand:<V128> 1 "register_operand" "w")
   2557 	    (match_operand 2 "descending_int_parallel"))))]
   2558   "TARGET_SVE
   2559    && BYTES_BIG_ENDIAN
   2560    && known_eq (INTVAL (XVECEXP (operands[2], 0, 0)),
   2561 		GET_MODE_NUNITS (<V128>mode) - 1)"
   2562   {
   2563     operands[1] = gen_rtx_REG (<MODE>mode, REGNO (operands[1]));
   2564     return "dup\t%0.q, %1.q[0]";
   2565   }
   2566 )
   2567 
   2568 ;; This is used for vec_duplicate<mode>s from memory, but can also
   2569 ;; be used by combine to optimize selects of a vec_duplicate<mode>
   2570 ;; with zero.
   2571 (define_insn "sve_ld1r<mode>"
   2572   [(set (match_operand:SVE_ALL 0 "register_operand" "=w")
   2573 	(unspec:SVE_ALL
   2574 	  [(match_operand:<VPRED> 1 "register_operand" "Upl")
   2575 	   (vec_duplicate:SVE_ALL
   2576 	     (match_operand:<VEL> 2 "aarch64_sve_ld1r_operand" "Uty"))
   2577 	   (match_operand:SVE_ALL 3 "aarch64_simd_imm_zero")]
   2578 	  UNSPEC_SEL))]
   2579   "TARGET_SVE"
   2580   "ld1r<Vesize>\t%0.<Vetype>, %1/z, %2"
   2581 )
   2582 
   2583 ;; Load 128 bits from memory under predicate control and duplicate to
   2584 ;; fill a vector.
   2585 (define_insn "@aarch64_sve_ld1rq<mode>"
   2586   [(set (match_operand:SVE_FULL 0 "register_operand" "=w")
   2587 	(unspec:SVE_FULL
   2588 	  [(match_operand:<VPRED> 2 "register_operand" "Upl")
   2589 	   (match_operand:<V128> 1 "aarch64_sve_ld1rq_operand" "UtQ")]
   2590 	  UNSPEC_LD1RQ))]
   2591   "TARGET_SVE"
   2592   {
   2593     operands[1] = gen_rtx_MEM (<VEL>mode, XEXP (operands[1], 0));
   2594     return "ld1rq<Vesize>\t%0.<Vetype>, %2/z, %1";
   2595   }
   2596 )
   2597 
   2598 (define_insn "@aarch64_sve_ld1ro<mode>"
   2599   [(set (match_operand:SVE_FULL 0 "register_operand" "=w")
   2600 	(unspec:SVE_FULL
   2601 	  [(match_operand:<VPRED> 2 "register_operand" "Upl")
   2602 	   (match_operand:OI 1 "aarch64_sve_ld1ro_operand_<Vesize>"
   2603 			       "UO<Vesize>")]
   2604 	  UNSPEC_LD1RO))]
   2605   "TARGET_SVE_F64MM"
   2606   {
   2607     operands[1] = gen_rtx_MEM (<VEL>mode, XEXP (operands[1], 0));
   2608     return "ld1ro<Vesize>\t%0.<Vetype>, %2/z, %1";
   2609   }
   2610 )
   2611 
   2612 ;; -------------------------------------------------------------------------
   2613 ;; ---- [INT,FP] Initialize from individual elements
   2614 ;; -------------------------------------------------------------------------
   2615 ;; Includes:
   2616 ;; - INSR
   2617 ;; -------------------------------------------------------------------------
   2618 
   2619 (define_expand "vec_init<mode><Vel>"
   2620   [(match_operand:SVE_FULL 0 "register_operand")
   2621     (match_operand 1 "")]
   2622   "TARGET_SVE"
   2623   {
   2624     aarch64_sve_expand_vector_init (operands[0], operands[1]);
   2625     DONE;
   2626   }
   2627 )
   2628 
   2629 ;; Shift an SVE vector left and insert a scalar into element 0.
   2630 (define_insn "vec_shl_insert_<mode>"
   2631   [(set (match_operand:SVE_FULL 0 "register_operand" "=?w, w, ??&w, ?&w")
   2632 	(unspec:SVE_FULL
   2633 	  [(match_operand:SVE_FULL 1 "register_operand" "0, 0, w, w")
   2634 	   (match_operand:<VEL> 2 "aarch64_reg_or_zero" "rZ, w, rZ, w")]
   2635 	  UNSPEC_INSR))]
   2636   "TARGET_SVE"
   2637   "@
   2638    insr\t%0.<Vetype>, %<vwcore>2
   2639    insr\t%0.<Vetype>, %<Vetype>2
   2640    movprfx\t%0, %1\;insr\t%0.<Vetype>, %<vwcore>2
   2641    movprfx\t%0, %1\;insr\t%0.<Vetype>, %<Vetype>2"
   2642   [(set_attr "movprfx" "*,*,yes,yes")]
   2643 )
   2644 
   2645 ;; -------------------------------------------------------------------------
   2646 ;; ---- [INT] Linear series
   2647 ;; -------------------------------------------------------------------------
   2648 ;; Includes:
   2649 ;; - INDEX
   2650 ;; -------------------------------------------------------------------------
   2651 
   2652 (define_insn "vec_series<mode>"
   2653   [(set (match_operand:SVE_I 0 "register_operand" "=w, w, w")
   2654 	(vec_series:SVE_I
   2655 	  (match_operand:<VEL> 1 "aarch64_sve_index_operand" "Usi, r, r")
   2656 	  (match_operand:<VEL> 2 "aarch64_sve_index_operand" "r, Usi, r")))]
   2657   "TARGET_SVE"
   2658   "@
   2659    index\t%0.<Vctype>, #%1, %<vccore>2
   2660    index\t%0.<Vctype>, %<vccore>1, #%2
   2661    index\t%0.<Vctype>, %<vccore>1, %<vccore>2"
   2662 )
   2663 
   2664 ;; Optimize {x, x, x, x, ...} + {0, n, 2*n, 3*n, ...} if n is in range
   2665 ;; of an INDEX instruction.
   2666 (define_insn "*vec_series<mode>_plus"
   2667   [(set (match_operand:SVE_I 0 "register_operand" "=w")
   2668 	(plus:SVE_I
   2669 	  (vec_duplicate:SVE_I
   2670 	    (match_operand:<VEL> 1 "register_operand" "r"))
   2671 	  (match_operand:SVE_I 2 "immediate_operand")))]
   2672   "TARGET_SVE && aarch64_check_zero_based_sve_index_immediate (operands[2])"
   2673   {
   2674     operands[2] = aarch64_check_zero_based_sve_index_immediate (operands[2]);
   2675     return "index\t%0.<Vctype>, %<vccore>1, #%2";
   2676   }
   2677 )
   2678 
   2679 ;; -------------------------------------------------------------------------
   2680 ;; ---- [PRED] Duplicate element
   2681 ;; -------------------------------------------------------------------------
   2682 ;; The patterns in this section are synthetic.
   2683 ;; -------------------------------------------------------------------------
   2684 
   2685 ;; Implement a predicate broadcast by shifting the low bit of the scalar
   2686 ;; input into the top bit and using a WHILELO.  An alternative would be to
   2687 ;; duplicate the input and do a compare with zero.
   2688 (define_expand "vec_duplicate<mode>"
   2689   [(set (match_operand:PRED_ALL 0 "register_operand")
   2690 	(vec_duplicate:PRED_ALL (match_operand:QI 1 "register_operand")))]
   2691   "TARGET_SVE"
   2692   {
   2693     rtx tmp = gen_reg_rtx (DImode);
   2694     rtx op1 = gen_lowpart (DImode, operands[1]);
   2695     emit_insn (gen_ashldi3 (tmp, op1, gen_int_mode (63, DImode)));
   2696     emit_insn (gen_while_ultdi<mode> (operands[0], const0_rtx, tmp));
   2697     DONE;
   2698   }
   2699 )
   2700 
   2701 ;; =========================================================================
   2702 ;; == Vector decomposition
   2703 ;; =========================================================================
   2704 
   2705 ;; -------------------------------------------------------------------------
   2706 ;; ---- [INT,FP] Extract index
   2707 ;; -------------------------------------------------------------------------
   2708 ;; Includes:
   2709 ;; - DUP    (Advanced SIMD)
   2710 ;; - DUP    (SVE)
   2711 ;; - EXT    (SVE)
   2712 ;; - ST1    (Advanced SIMD)
   2713 ;; - UMOV   (Advanced SIMD)
   2714 ;; -------------------------------------------------------------------------
   2715 
   2716 (define_expand "vec_extract<mode><Vel>"
   2717   [(set (match_operand:<VEL> 0 "register_operand")
   2718 	(vec_select:<VEL>
   2719 	  (match_operand:SVE_FULL 1 "register_operand")
   2720 	  (parallel [(match_operand:SI 2 "nonmemory_operand")])))]
   2721   "TARGET_SVE"
   2722   {
   2723     poly_int64 val;
   2724     if (poly_int_rtx_p (operands[2], &val)
   2725 	&& known_eq (val, GET_MODE_NUNITS (<MODE>mode) - 1))
   2726       {
   2727 	/* The last element can be extracted with a LASTB and a false
   2728 	   predicate.  */
   2729 	rtx sel = aarch64_pfalse_reg (<VPRED>mode);
   2730 	emit_insn (gen_extract_last_<mode> (operands[0], sel, operands[1]));
   2731 	DONE;
   2732       }
   2733     if (!CONST_INT_P (operands[2]))
   2734       {
   2735 	/* Create an index with operand[2] as the base and -1 as the step.
   2736 	   It will then be zero for the element we care about.  */
   2737 	rtx index = gen_lowpart (<VEL_INT>mode, operands[2]);
   2738 	index = force_reg (<VEL_INT>mode, index);
   2739 	rtx series = gen_reg_rtx (<V_INT_EQUIV>mode);
   2740 	emit_insn (gen_vec_series<v_int_equiv> (series, index, constm1_rtx));
   2741 
   2742 	/* Get a predicate that is true for only that element.  */
   2743 	rtx zero = CONST0_RTX (<V_INT_EQUIV>mode);
   2744 	rtx cmp = gen_rtx_EQ (<V_INT_EQUIV>mode, series, zero);
   2745 	rtx sel = gen_reg_rtx (<VPRED>mode);
   2746 	emit_insn (gen_vec_cmp<v_int_equiv><vpred> (sel, cmp, series, zero));
   2747 
   2748 	/* Select the element using LASTB.  */
   2749 	emit_insn (gen_extract_last_<mode> (operands[0], sel, operands[1]));
   2750 	DONE;
   2751       }
   2752   }
   2753 )
   2754 
   2755 ;; Extract element zero.  This is a special case because we want to force
   2756 ;; the registers to be the same for the second alternative, and then
   2757 ;; split the instruction into nothing after RA.
   2758 (define_insn_and_split "*vec_extract<mode><Vel>_0"
   2759   [(set (match_operand:<VEL> 0 "aarch64_simd_nonimmediate_operand" "=r, w, Utv")
   2760 	(vec_select:<VEL>
   2761 	  (match_operand:SVE_FULL 1 "register_operand" "w, 0, w")
   2762 	  (parallel [(const_int 0)])))]
   2763   "TARGET_SVE"
   2764   {
   2765     operands[1] = gen_rtx_REG (<V128>mode, REGNO (operands[1]));
   2766     switch (which_alternative)
   2767       {
   2768 	case 0:
   2769 	  return "umov\\t%<vwcore>0, %1.<Vetype>[0]";
   2770 	case 1:
   2771 	  return "#";
   2772 	case 2:
   2773 	  return "st1\\t{%1.<Vetype>}[0], %0";
   2774 	default:
   2775 	  gcc_unreachable ();
   2776       }
   2777   }
   2778   "&& reload_completed
   2779    && REG_P (operands[0])
   2780    && REGNO (operands[0]) == REGNO (operands[1])"
   2781   [(const_int 0)]
   2782   {
   2783     emit_note (NOTE_INSN_DELETED);
   2784     DONE;
   2785   }
   2786   [(set_attr "type" "neon_to_gp_q, untyped, neon_store1_one_lane_q")]
   2787 )
   2788 
   2789 ;; Extract an element from the Advanced SIMD portion of the register.
   2790 ;; We don't just reuse the aarch64-simd.md pattern because we don't
   2791 ;; want any change in lane number on big-endian targets.
   2792 (define_insn "*vec_extract<mode><Vel>_v128"
   2793   [(set (match_operand:<VEL> 0 "aarch64_simd_nonimmediate_operand" "=r, w, Utv")
   2794 	(vec_select:<VEL>
   2795 	  (match_operand:SVE_FULL 1 "register_operand" "w, w, w")
   2796 	  (parallel [(match_operand:SI 2 "const_int_operand")])))]
   2797   "TARGET_SVE
   2798    && IN_RANGE (INTVAL (operands[2]) * GET_MODE_SIZE (<VEL>mode), 1, 15)"
   2799   {
   2800     operands[1] = gen_rtx_REG (<V128>mode, REGNO (operands[1]));
   2801     switch (which_alternative)
   2802       {
   2803 	case 0:
   2804 	  return "umov\\t%<vwcore>0, %1.<Vetype>[%2]";
   2805 	case 1:
   2806 	  return "dup\\t%<Vetype>0, %1.<Vetype>[%2]";
   2807 	case 2:
   2808 	  return "st1\\t{%1.<Vetype>}[%2], %0";
   2809 	default:
   2810 	  gcc_unreachable ();
   2811       }
   2812   }
   2813   [(set_attr "type" "neon_to_gp_q, neon_dup_q, neon_store1_one_lane_q")]
   2814 )
   2815 
   2816 ;; Extract an element in the range of DUP.  This pattern allows the
   2817 ;; source and destination to be different.
   2818 (define_insn "*vec_extract<mode><Vel>_dup"
   2819   [(set (match_operand:<VEL> 0 "register_operand" "=w")
   2820 	(vec_select:<VEL>
   2821 	  (match_operand:SVE_FULL 1 "register_operand" "w")
   2822 	  (parallel [(match_operand:SI 2 "const_int_operand")])))]
   2823   "TARGET_SVE
   2824    && IN_RANGE (INTVAL (operands[2]) * GET_MODE_SIZE (<VEL>mode), 16, 63)"
   2825   {
   2826     operands[0] = gen_rtx_REG (<MODE>mode, REGNO (operands[0]));
   2827     return "dup\t%0.<Vetype>, %1.<Vetype>[%2]";
   2828   }
   2829 )
   2830 
   2831 ;; Extract an element outside the range of DUP.  This pattern requires the
   2832 ;; source and destination to be the same.
   2833 (define_insn "*vec_extract<mode><Vel>_ext"
   2834   [(set (match_operand:<VEL> 0 "register_operand" "=w, ?&w")
   2835 	(vec_select:<VEL>
   2836 	  (match_operand:SVE_FULL 1 "register_operand" "0, w")
   2837 	  (parallel [(match_operand:SI 2 "const_int_operand")])))]
   2838   "TARGET_SVE && INTVAL (operands[2]) * GET_MODE_SIZE (<VEL>mode) >= 64"
   2839   {
   2840     operands[0] = gen_rtx_REG (<MODE>mode, REGNO (operands[0]));
   2841     operands[2] = GEN_INT (INTVAL (operands[2]) * GET_MODE_SIZE (<VEL>mode));
   2842     return (which_alternative == 0
   2843 	    ? "ext\t%0.b, %0.b, %0.b, #%2"
   2844 	    : "movprfx\t%0, %1\;ext\t%0.b, %0.b, %1.b, #%2");
   2845   }
   2846   [(set_attr "movprfx" "*,yes")]
   2847 )
   2848 
   2849 ;; -------------------------------------------------------------------------
   2850 ;; ---- [INT,FP] Extract active element
   2851 ;; -------------------------------------------------------------------------
   2852 ;; Includes:
   2853 ;; - LASTA
   2854 ;; - LASTB
   2855 ;; -------------------------------------------------------------------------
   2856 
   2857 ;; Extract the last active element of operand 1 into operand 0.
   2858 ;; If no elements are active, extract the last inactive element instead.
   2859 (define_insn "@extract_<last_op>_<mode>"
   2860   [(set (match_operand:<VEL> 0 "register_operand" "=?r, w")
   2861 	(unspec:<VEL>
   2862 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   2863 	   (match_operand:SVE_FULL 2 "register_operand" "w, w")]
   2864 	  LAST))]
   2865   "TARGET_SVE"
   2866   "@
   2867    last<ab>\t%<vwcore>0, %1, %2.<Vetype>
   2868    last<ab>\t%<Vetype>0, %1, %2.<Vetype>"
   2869 )
   2870 
   2871 ;; -------------------------------------------------------------------------
   2872 ;; ---- [PRED] Extract index
   2873 ;; -------------------------------------------------------------------------
   2874 ;; The patterns in this section are synthetic.
   2875 ;; -------------------------------------------------------------------------
   2876 
   2877 ;; Handle extractions from a predicate by converting to an integer vector
   2878 ;; and extracting from there.
   2879 (define_expand "vec_extract<vpred><Vel>"
   2880   [(match_operand:<VEL> 0 "register_operand")
   2881    (match_operand:<VPRED> 1 "register_operand")
   2882    (match_operand:SI 2 "nonmemory_operand")
   2883    ;; Dummy operand to which we can attach the iterator.
   2884    (reg:SVE_FULL_I V0_REGNUM)]
   2885   "TARGET_SVE"
   2886   {
   2887     rtx tmp = gen_reg_rtx (<MODE>mode);
   2888     emit_insn (gen_vcond_mask_<mode><vpred> (tmp, operands[1],
   2889 					     CONST1_RTX (<MODE>mode),
   2890 					     CONST0_RTX (<MODE>mode)));
   2891     emit_insn (gen_vec_extract<mode><Vel> (operands[0], tmp, operands[2]));
   2892     DONE;
   2893   }
   2894 )
   2895 
   2896 ;; =========================================================================
   2897 ;; == Unary arithmetic
   2898 ;; =========================================================================
   2899 
   2900 ;; -------------------------------------------------------------------------
   2901 ;; ---- [INT] General unary arithmetic corresponding to rtx codes
   2902 ;; -------------------------------------------------------------------------
   2903 ;; Includes:
   2904 ;; - ABS
   2905 ;; - CLS (= clrsb)
   2906 ;; - CLZ
   2907 ;; - CNT (= popcount)
   2908 ;; - NEG
   2909 ;; - NOT
   2910 ;; -------------------------------------------------------------------------
   2911 
   2912 ;; Unpredicated integer unary arithmetic.
   2913 (define_expand "<optab><mode>2"
   2914   [(set (match_operand:SVE_I 0 "register_operand")
   2915 	(unspec:SVE_I
   2916 	  [(match_dup 2)
   2917 	   (SVE_INT_UNARY:SVE_I
   2918 	     (match_operand:SVE_I 1 "register_operand"))]
   2919 	  UNSPEC_PRED_X))]
   2920   "TARGET_SVE"
   2921   {
   2922     operands[2] = aarch64_ptrue_reg (<VPRED>mode);
   2923   }
   2924 )
   2925 
   2926 ;; Integer unary arithmetic predicated with a PTRUE.
   2927 (define_insn "@aarch64_pred_<optab><mode>"
   2928   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   2929 	(unspec:SVE_I
   2930 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   2931 	   (SVE_INT_UNARY:SVE_I
   2932 	     (match_operand:SVE_I 2 "register_operand" "0, w"))]
   2933 	  UNSPEC_PRED_X))]
   2934   "TARGET_SVE"
   2935   "@
   2936    <sve_int_op>\t%0.<Vetype>, %1/m, %2.<Vetype>
   2937    movprfx\t%0, %2\;<sve_int_op>\t%0.<Vetype>, %1/m, %2.<Vetype>"
   2938   [(set_attr "movprfx" "*,yes")]
   2939 )
   2940 
   2941 ;; Predicated integer unary arithmetic with merging.
   2942 (define_expand "@cond_<optab><mode>"
   2943   [(set (match_operand:SVE_I 0 "register_operand")
   2944 	(unspec:SVE_I
   2945 	  [(match_operand:<VPRED> 1 "register_operand")
   2946 	   (SVE_INT_UNARY:SVE_I
   2947 	     (match_operand:SVE_I 2 "register_operand"))
   2948 	   (match_operand:SVE_I 3 "aarch64_simd_reg_or_zero")]
   2949 	  UNSPEC_SEL))]
   2950   "TARGET_SVE"
   2951 )
   2952 
   2953 ;; Predicated integer unary arithmetic, merging with the first input.
   2954 (define_insn "*cond_<optab><mode>_2"
   2955   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   2956 	(unspec:SVE_I
   2957 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   2958 	   (SVE_INT_UNARY:SVE_I
   2959 	     (match_operand:SVE_I 2 "register_operand" "0, w"))
   2960 	   (match_dup 2)]
   2961 	  UNSPEC_SEL))]
   2962   "TARGET_SVE"
   2963   "@
   2964    <sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>
   2965    movprfx\t%0, %2\;<sve_int_op>\t%0.<Vetype>, %1/m, %2.<Vetype>"
   2966   [(set_attr "movprfx" "*,yes")]
   2967 )
   2968 
   2969 ;; Predicated integer unary arithmetic, merging with an independent value.
   2970 ;;
   2971 ;; The earlyclobber isn't needed for the first alternative, but omitting
   2972 ;; it would only help the case in which operands 2 and 3 are the same,
   2973 ;; which is handled above rather than here.  Marking all the alternatives
   2974 ;; as earlyclobber helps to make the instruction more regular to the
   2975 ;; register allocator.
   2976 (define_insn "*cond_<optab><mode>_any"
   2977   [(set (match_operand:SVE_I 0 "register_operand" "=&w, ?&w, ?&w")
   2978 	(unspec:SVE_I
   2979 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   2980 	   (SVE_INT_UNARY:SVE_I
   2981 	     (match_operand:SVE_I 2 "register_operand" "w, w, w"))
   2982 	   (match_operand:SVE_I 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   2983 	  UNSPEC_SEL))]
   2984   "TARGET_SVE && !rtx_equal_p (operands[2], operands[3])"
   2985   "@
   2986    <sve_int_op>\t%0.<Vetype>, %1/m, %2.<Vetype>
   2987    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %2.<Vetype>
   2988    movprfx\t%0, %3\;<sve_int_op>\t%0.<Vetype>, %1/m, %2.<Vetype>"
   2989   [(set_attr "movprfx" "*,yes,yes")]
   2990 )
   2991 
   2992 ;; -------------------------------------------------------------------------
   2993 ;; ---- [INT] General unary arithmetic corresponding to unspecs
   2994 ;; -------------------------------------------------------------------------
   2995 ;; Includes
   2996 ;; - RBIT
   2997 ;; - REVB
   2998 ;; - REVH
   2999 ;; - REVW
   3000 ;; -------------------------------------------------------------------------
   3001 
   3002 ;; Predicated integer unary operations.
   3003 (define_insn "@aarch64_pred_<optab><mode>"
   3004   [(set (match_operand:SVE_FULL_I 0 "register_operand" "=w, ?&w")
   3005 	(unspec:SVE_FULL_I
   3006 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   3007 	   (unspec:SVE_FULL_I
   3008 	     [(match_operand:SVE_FULL_I 2 "register_operand" "0, w")]
   3009 	     SVE_INT_UNARY)]
   3010 	  UNSPEC_PRED_X))]
   3011   "TARGET_SVE && <elem_bits> >= <min_elem_bits>"
   3012   "@
   3013    <sve_int_op>\t%0.<Vetype>, %1/m, %2.<Vetype>
   3014    movprfx\t%0, %2\;<sve_int_op>\t%0.<Vetype>, %1/m, %2.<Vetype>"
   3015   [(set_attr "movprfx" "*,yes")]
   3016 )
   3017 
   3018 ;; Another way of expressing the REVB, REVH and REVW patterns, with this
   3019 ;; form being easier for permutes.  The predicate mode determines the number
   3020 ;; of lanes and the data mode decides the granularity of the reversal within
   3021 ;; each lane.
   3022 (define_insn "@aarch64_sve_revbhw_<SVE_ALL:mode><PRED_HSD:mode>"
   3023   [(set (match_operand:SVE_ALL 0 "register_operand" "=w, ?&w")
   3024 	(unspec:SVE_ALL
   3025 	  [(match_operand:PRED_HSD 1 "register_operand" "Upl, Upl")
   3026 	   (unspec:SVE_ALL
   3027 	     [(match_operand:SVE_ALL 2 "register_operand" "0, w")]
   3028 	     UNSPEC_REVBHW)]
   3029 	  UNSPEC_PRED_X))]
   3030   "TARGET_SVE && <PRED_HSD:elem_bits> > <SVE_ALL:container_bits>"
   3031   "@
   3032    rev<SVE_ALL:Vcwtype>\t%0.<PRED_HSD:Vetype>, %1/m, %2.<PRED_HSD:Vetype>
   3033    movprfx\t%0, %2\;rev<SVE_ALL:Vcwtype>\t%0.<PRED_HSD:Vetype>, %1/m, %2.<PRED_HSD:Vetype>"
   3034   [(set_attr "movprfx" "*,yes")]
   3035 )
   3036 
   3037 ;; Predicated integer unary operations with merging.
   3038 (define_insn "@cond_<optab><mode>"
   3039   [(set (match_operand:SVE_FULL_I 0 "register_operand" "=w, ?&w, ?&w")
   3040 	(unspec:SVE_FULL_I
   3041 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   3042 	   (unspec:SVE_FULL_I
   3043 	     [(match_operand:SVE_FULL_I 2 "register_operand" "w, w, w")]
   3044 	     SVE_INT_UNARY)
   3045 	   (match_operand:SVE_FULL_I 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   3046 	  UNSPEC_SEL))]
   3047   "TARGET_SVE && <elem_bits> >= <min_elem_bits>"
   3048   "@
   3049    <sve_int_op>\t%0.<Vetype>, %1/m, %2.<Vetype>
   3050    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %2.<Vetype>
   3051    movprfx\t%0, %3\;<sve_int_op>\t%0.<Vetype>, %1/m, %2.<Vetype>"
   3052   [(set_attr "movprfx" "*,yes,yes")]
   3053 )
   3054 
   3055 ;; -------------------------------------------------------------------------
   3056 ;; ---- [INT] Sign and zero extension
   3057 ;; -------------------------------------------------------------------------
   3058 ;; Includes:
   3059 ;; - SXTB
   3060 ;; - SXTH
   3061 ;; - SXTW
   3062 ;; - UXTB
   3063 ;; - UXTH
   3064 ;; - UXTW
   3065 ;; -------------------------------------------------------------------------
   3066 
   3067 ;; Unpredicated sign and zero extension from a narrower mode.
   3068 (define_expand "<optab><SVE_PARTIAL_I:mode><SVE_HSDI:mode>2"
   3069   [(set (match_operand:SVE_HSDI 0 "register_operand")
   3070 	(unspec:SVE_HSDI
   3071 	  [(match_dup 2)
   3072 	   (ANY_EXTEND:SVE_HSDI
   3073 	     (match_operand:SVE_PARTIAL_I 1 "register_operand"))]
   3074 	  UNSPEC_PRED_X))]
   3075   "TARGET_SVE && (~<SVE_HSDI:narrower_mask> & <SVE_PARTIAL_I:self_mask>) == 0"
   3076   {
   3077     operands[2] = aarch64_ptrue_reg (<SVE_HSDI:VPRED>mode);
   3078   }
   3079 )
   3080 
   3081 ;; Predicated sign and zero extension from a narrower mode.
   3082 (define_insn "*<optab><SVE_PARTIAL_I:mode><SVE_HSDI:mode>2"
   3083   [(set (match_operand:SVE_HSDI 0 "register_operand" "=w, ?&w")
   3084 	(unspec:SVE_HSDI
   3085 	  [(match_operand:<SVE_HSDI:VPRED> 1 "register_operand" "Upl, Upl")
   3086 	   (ANY_EXTEND:SVE_HSDI
   3087 	     (match_operand:SVE_PARTIAL_I 2 "register_operand" "0, w"))]
   3088 	  UNSPEC_PRED_X))]
   3089   "TARGET_SVE && (~<SVE_HSDI:narrower_mask> & <SVE_PARTIAL_I:self_mask>) == 0"
   3090   "@
   3091    <su>xt<SVE_PARTIAL_I:Vesize>\t%0.<SVE_HSDI:Vetype>, %1/m, %2.<SVE_HSDI:Vetype>
   3092    movprfx\t%0, %2\;<su>xt<SVE_PARTIAL_I:Vesize>\t%0.<SVE_HSDI:Vetype>, %1/m, %2.<SVE_HSDI:Vetype>"
   3093   [(set_attr "movprfx" "*,yes")]
   3094 )
   3095 
   3096 ;; Predicated truncate-and-sign-extend operations.
   3097 (define_insn "@aarch64_pred_sxt<SVE_FULL_HSDI:mode><SVE_PARTIAL_I:mode>"
   3098   [(set (match_operand:SVE_FULL_HSDI 0 "register_operand" "=w, ?&w")
   3099 	(unspec:SVE_FULL_HSDI
   3100 	  [(match_operand:<SVE_FULL_HSDI:VPRED> 1 "register_operand" "Upl, Upl")
   3101 	   (sign_extend:SVE_FULL_HSDI
   3102 	     (truncate:SVE_PARTIAL_I
   3103 	       (match_operand:SVE_FULL_HSDI 2 "register_operand" "0, w")))]
   3104 	  UNSPEC_PRED_X))]
   3105   "TARGET_SVE
   3106    && (~<SVE_FULL_HSDI:narrower_mask> & <SVE_PARTIAL_I:self_mask>) == 0"
   3107   "@
   3108    sxt<SVE_PARTIAL_I:Vesize>\t%0.<SVE_FULL_HSDI:Vetype>, %1/m, %2.<SVE_FULL_HSDI:Vetype>
   3109    movprfx\t%0, %2\;sxt<SVE_PARTIAL_I:Vesize>\t%0.<SVE_FULL_HSDI:Vetype>, %1/m, %2.<SVE_FULL_HSDI:Vetype>"
   3110   [(set_attr "movprfx" "*,yes")]
   3111 )
   3112 
   3113 ;; Predicated truncate-and-sign-extend operations with merging.
   3114 (define_insn "@aarch64_cond_sxt<SVE_FULL_HSDI:mode><SVE_PARTIAL_I:mode>"
   3115   [(set (match_operand:SVE_FULL_HSDI 0 "register_operand" "=w, ?&w, ?&w")
   3116 	(unspec:SVE_FULL_HSDI
   3117 	  [(match_operand:<SVE_FULL_HSDI:VPRED> 1 "register_operand" "Upl, Upl, Upl")
   3118 	   (sign_extend:SVE_FULL_HSDI
   3119 	     (truncate:SVE_PARTIAL_I
   3120 	       (match_operand:SVE_FULL_HSDI 2 "register_operand" "w, w, w")))
   3121 	   (match_operand:SVE_FULL_HSDI 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   3122 	  UNSPEC_SEL))]
   3123   "TARGET_SVE
   3124    && (~<SVE_FULL_HSDI:narrower_mask> & <SVE_PARTIAL_I:self_mask>) == 0"
   3125   "@
   3126    sxt<SVE_PARTIAL_I:Vesize>\t%0.<SVE_FULL_HSDI:Vetype>, %1/m, %2.<SVE_FULL_HSDI:Vetype>
   3127    movprfx\t%0.<SVE_FULL_HSDI:Vetype>, %1/z, %2.<SVE_FULL_HSDI:Vetype>\;sxt<SVE_PARTIAL_I:Vesize>\t%0.<SVE_FULL_HSDI:Vetype>, %1/m, %2.<SVE_FULL_HSDI:Vetype>
   3128    movprfx\t%0, %3\;sxt<SVE_PARTIAL_I:Vesize>\t%0.<SVE_FULL_HSDI:Vetype>, %1/m, %2.<SVE_FULL_HSDI:Vetype>"
   3129   [(set_attr "movprfx" "*,yes,yes")]
   3130 )
   3131 
   3132 ;; Predicated truncate-and-zero-extend operations, merging with the
   3133 ;; first input.
   3134 ;;
   3135 ;; The canonical form of this operation is an AND of a constant rather
   3136 ;; than (zero_extend (truncate ...)).
   3137 (define_insn "*cond_uxt<mode>_2"
   3138   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   3139 	(unspec:SVE_I
   3140 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   3141 	   (and:SVE_I
   3142 	     (match_operand:SVE_I 2 "register_operand" "0, w")
   3143 	     (match_operand:SVE_I 3 "aarch64_sve_uxt_immediate"))
   3144 	   (match_dup 2)]
   3145 	  UNSPEC_SEL))]
   3146   "TARGET_SVE"
   3147   "@
   3148    uxt%e3\t%0.<Vetype>, %1/m, %0.<Vetype>
   3149    movprfx\t%0, %2\;uxt%e3\t%0.<Vetype>, %1/m, %2.<Vetype>"
   3150   [(set_attr "movprfx" "*,yes")]
   3151 )
   3152 
   3153 ;; Predicated truncate-and-zero-extend operations, merging with an
   3154 ;; independent value.
   3155 ;;
   3156 ;; The earlyclobber isn't needed for the first alternative, but omitting
   3157 ;; it would only help the case in which operands 2 and 4 are the same,
   3158 ;; which is handled above rather than here.  Marking all the alternatives
   3159 ;; as early-clobber helps to make the instruction more regular to the
   3160 ;; register allocator.
   3161 (define_insn "*cond_uxt<mode>_any"
   3162   [(set (match_operand:SVE_I 0 "register_operand" "=&w, ?&w, ?&w")
   3163 	(unspec:SVE_I
   3164 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   3165 	   (and:SVE_I
   3166 	     (match_operand:SVE_I 2 "register_operand" "w, w, w")
   3167 	     (match_operand:SVE_I 3 "aarch64_sve_uxt_immediate"))
   3168 	   (match_operand:SVE_I 4 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   3169 	  UNSPEC_SEL))]
   3170   "TARGET_SVE && !rtx_equal_p (operands[2], operands[4])"
   3171   "@
   3172    uxt%e3\t%0.<Vetype>, %1/m, %2.<Vetype>
   3173    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;uxt%e3\t%0.<Vetype>, %1/m, %2.<Vetype>
   3174    movprfx\t%0, %4\;uxt%e3\t%0.<Vetype>, %1/m, %2.<Vetype>"
   3175   [(set_attr "movprfx" "*,yes,yes")]
   3176 )
   3177 
   3178 ;; -------------------------------------------------------------------------
   3179 ;; ---- [INT] Truncation
   3180 ;; -------------------------------------------------------------------------
   3181 ;; The patterns in this section are synthetic.
   3182 ;; -------------------------------------------------------------------------
   3183 
   3184 ;; Truncate to a partial SVE vector from either a full vector or a
   3185 ;; wider partial vector.  This is a no-op, because we can just ignore
   3186 ;; the unused upper bits of the source.
   3187 (define_insn_and_split "trunc<SVE_HSDI:mode><SVE_PARTIAL_I:mode>2"
   3188   [(set (match_operand:SVE_PARTIAL_I 0 "register_operand" "=w")
   3189 	(truncate:SVE_PARTIAL_I
   3190 	  (match_operand:SVE_HSDI 1 "register_operand" "w")))]
   3191   "TARGET_SVE && (~<SVE_HSDI:narrower_mask> & <SVE_PARTIAL_I:self_mask>) == 0"
   3192   "#"
   3193   "&& reload_completed"
   3194   [(set (match_dup 0) (match_dup 1))]
   3195   {
   3196     operands[1] = aarch64_replace_reg_mode (operands[1],
   3197 					    <SVE_PARTIAL_I:MODE>mode);
   3198   }
   3199 )
   3200 
   3201 ;; -------------------------------------------------------------------------
   3202 ;; ---- [INT] Logical inverse
   3203 ;; -------------------------------------------------------------------------
   3204 ;; Includes:
   3205 ;; - CNOT
   3206 ;; -------------------------------------------------------------------------
   3207 
   3208 ;; Logical inverse, predicated with a ptrue.
   3209 (define_expand "@aarch64_ptrue_cnot<mode>"
   3210   [(set (match_operand:SVE_FULL_I 0 "register_operand")
   3211 	(unspec:SVE_FULL_I
   3212 	  [(unspec:<VPRED>
   3213 	     [(match_operand:<VPRED> 1 "register_operand")
   3214 	      (const_int SVE_KNOWN_PTRUE)
   3215 	      (eq:<VPRED>
   3216 		(match_operand:SVE_FULL_I 2 "register_operand")
   3217 		(match_dup 3))]
   3218 	     UNSPEC_PRED_Z)
   3219 	   (match_dup 4)
   3220 	   (match_dup 3)]
   3221 	  UNSPEC_SEL))]
   3222   "TARGET_SVE"
   3223   {
   3224     operands[3] = CONST0_RTX (<MODE>mode);
   3225     operands[4] = CONST1_RTX (<MODE>mode);
   3226   }
   3227 )
   3228 
   3229 (define_insn "*cnot<mode>"
   3230   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   3231 	(unspec:SVE_I
   3232 	  [(unspec:<VPRED>
   3233 	     [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   3234 	      (const_int SVE_KNOWN_PTRUE)
   3235 	      (eq:<VPRED>
   3236 		(match_operand:SVE_I 2 "register_operand" "0, w")
   3237 		(match_operand:SVE_I 3 "aarch64_simd_imm_zero"))]
   3238 	     UNSPEC_PRED_Z)
   3239 	   (match_operand:SVE_I 4 "aarch64_simd_imm_one")
   3240 	   (match_dup 3)]
   3241 	  UNSPEC_SEL))]
   3242   "TARGET_SVE"
   3243   "@
   3244    cnot\t%0.<Vetype>, %1/m, %2.<Vetype>
   3245    movprfx\t%0, %2\;cnot\t%0.<Vetype>, %1/m, %2.<Vetype>"
   3246   [(set_attr "movprfx" "*,yes")]
   3247 )
   3248 
   3249 ;; Predicated logical inverse with merging.
   3250 (define_expand "@cond_cnot<mode>"
   3251   [(set (match_operand:SVE_FULL_I 0 "register_operand")
   3252 	(unspec:SVE_FULL_I
   3253 	  [(match_operand:<VPRED> 1 "register_operand")
   3254 	   (unspec:SVE_FULL_I
   3255 	     [(unspec:<VPRED>
   3256 		[(match_dup 4)
   3257 		 (const_int SVE_KNOWN_PTRUE)
   3258 		 (eq:<VPRED>
   3259 		   (match_operand:SVE_FULL_I 2 "register_operand")
   3260 		   (match_dup 5))]
   3261 		UNSPEC_PRED_Z)
   3262 	      (match_dup 6)
   3263 	      (match_dup 5)]
   3264 	     UNSPEC_SEL)
   3265 	   (match_operand:SVE_FULL_I 3 "aarch64_simd_reg_or_zero")]
   3266 	  UNSPEC_SEL))]
   3267   "TARGET_SVE"
   3268   {
   3269     operands[4] = CONSTM1_RTX (<VPRED>mode);
   3270     operands[5] = CONST0_RTX (<MODE>mode);
   3271     operands[6] = CONST1_RTX (<MODE>mode);
   3272   }
   3273 )
   3274 
   3275 ;; Predicated logical inverse, merging with the first input.
   3276 (define_insn_and_rewrite "*cond_cnot<mode>_2"
   3277   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   3278 	(unspec:SVE_I
   3279 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   3280 	   ;; Logical inverse of operand 2 (as above).
   3281 	   (unspec:SVE_I
   3282 	     [(unspec:<VPRED>
   3283 		[(match_operand 5)
   3284 		 (const_int SVE_KNOWN_PTRUE)
   3285 		 (eq:<VPRED>
   3286 		   (match_operand:SVE_I 2 "register_operand" "0, w")
   3287 		   (match_operand:SVE_I 3 "aarch64_simd_imm_zero"))]
   3288 		UNSPEC_PRED_Z)
   3289 	      (match_operand:SVE_I 4 "aarch64_simd_imm_one")
   3290 	      (match_dup 3)]
   3291 	     UNSPEC_SEL)
   3292 	   (match_dup 2)]
   3293 	  UNSPEC_SEL))]
   3294   "TARGET_SVE"
   3295   "@
   3296    cnot\t%0.<Vetype>, %1/m, %0.<Vetype>
   3297    movprfx\t%0, %2\;cnot\t%0.<Vetype>, %1/m, %2.<Vetype>"
   3298   "&& !CONSTANT_P (operands[5])"
   3299   {
   3300     operands[5] = CONSTM1_RTX (<VPRED>mode);
   3301   }
   3302   [(set_attr "movprfx" "*,yes")]
   3303 )
   3304 
   3305 ;; Predicated logical inverse, merging with an independent value.
   3306 ;;
   3307 ;; The earlyclobber isn't needed for the first alternative, but omitting
   3308 ;; it would only help the case in which operands 2 and 6 are the same,
   3309 ;; which is handled above rather than here.  Marking all the alternatives
   3310 ;; as earlyclobber helps to make the instruction more regular to the
   3311 ;; register allocator.
   3312 (define_insn_and_rewrite "*cond_cnot<mode>_any"
   3313   [(set (match_operand:SVE_I 0 "register_operand" "=&w, ?&w, ?&w")
   3314 	(unspec:SVE_I
   3315 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   3316 	   ;; Logical inverse of operand 2 (as above).
   3317 	   (unspec:SVE_I
   3318 	     [(unspec:<VPRED>
   3319 		[(match_operand 5)
   3320 		 (const_int SVE_KNOWN_PTRUE)
   3321 		 (eq:<VPRED>
   3322 		   (match_operand:SVE_I 2 "register_operand" "w, w, w")
   3323 		   (match_operand:SVE_I 3 "aarch64_simd_imm_zero"))]
   3324 		UNSPEC_PRED_Z)
   3325 	      (match_operand:SVE_I 4 "aarch64_simd_imm_one")
   3326 	      (match_dup 3)]
   3327 	     UNSPEC_SEL)
   3328 	   (match_operand:SVE_I 6 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   3329 	  UNSPEC_SEL))]
   3330   "TARGET_SVE && !rtx_equal_p (operands[2], operands[6])"
   3331   "@
   3332    cnot\t%0.<Vetype>, %1/m, %2.<Vetype>
   3333    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;cnot\t%0.<Vetype>, %1/m, %2.<Vetype>
   3334    movprfx\t%0, %6\;cnot\t%0.<Vetype>, %1/m, %2.<Vetype>"
   3335   "&& !CONSTANT_P (operands[5])"
   3336   {
   3337     operands[5] = CONSTM1_RTX (<VPRED>mode);
   3338   }
   3339   [(set_attr "movprfx" "*,yes,yes")]
   3340 )
   3341 
   3342 ;; -------------------------------------------------------------------------
   3343 ;; ---- [FP<-INT] General unary arithmetic that maps to unspecs
   3344 ;; -------------------------------------------------------------------------
   3345 ;; Includes:
   3346 ;; - FEXPA
   3347 ;; -------------------------------------------------------------------------
   3348 
   3349 ;; Unpredicated unary operations that take an integer and return a float.
   3350 (define_insn "@aarch64_sve_<optab><mode>"
   3351   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w")
   3352 	(unspec:SVE_FULL_F
   3353 	  [(match_operand:<V_INT_EQUIV> 1 "register_operand" "w")]
   3354 	  SVE_FP_UNARY_INT))]
   3355   "TARGET_SVE"
   3356   "<sve_fp_op>\t%0.<Vetype>, %1.<Vetype>"
   3357 )
   3358 
   3359 ;; -------------------------------------------------------------------------
   3360 ;; ---- [FP] General unary arithmetic corresponding to unspecs
   3361 ;; -------------------------------------------------------------------------
   3362 ;; Includes:
   3363 ;; - FABS
   3364 ;; - FNEG
   3365 ;; - FRECPE
   3366 ;; - FRECPX
   3367 ;; - FRINTA
   3368 ;; - FRINTI
   3369 ;; - FRINTM
   3370 ;; - FRINTN
   3371 ;; - FRINTP
   3372 ;; - FRINTX
   3373 ;; - FRINTZ
   3374 ;; - FRSQRTE
   3375 ;; - FSQRT
   3376 ;; -------------------------------------------------------------------------
   3377 
   3378 ;; Unpredicated floating-point unary operations.
   3379 (define_insn "@aarch64_sve_<optab><mode>"
   3380   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w")
   3381 	(unspec:SVE_FULL_F
   3382 	  [(match_operand:SVE_FULL_F 1 "register_operand" "w")]
   3383 	  SVE_FP_UNARY))]
   3384   "TARGET_SVE"
   3385   "<sve_fp_op>\t%0.<Vetype>, %1.<Vetype>"
   3386 )
   3387 
   3388 ;; Unpredicated floating-point unary operations.
   3389 (define_expand "<optab><mode>2"
   3390   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   3391 	(unspec:SVE_FULL_F
   3392 	  [(match_dup 2)
   3393 	   (const_int SVE_RELAXED_GP)
   3394 	   (match_operand:SVE_FULL_F 1 "register_operand")]
   3395 	  SVE_COND_FP_UNARY_OPTAB))]
   3396   "TARGET_SVE"
   3397   {
   3398     operands[2] = aarch64_ptrue_reg (<VPRED>mode);
   3399   }
   3400 )
   3401 
   3402 ;; Predicated floating-point unary operations.
   3403 (define_insn "@aarch64_pred_<optab><mode>"
   3404   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   3405 	(unspec:SVE_FULL_F
   3406 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   3407 	   (match_operand:SI 3 "aarch64_sve_gp_strictness")
   3408 	   (match_operand:SVE_FULL_F 2 "register_operand" "0, w")]
   3409 	  SVE_COND_FP_UNARY))]
   3410   "TARGET_SVE"
   3411   "@
   3412    <sve_fp_op>\t%0.<Vetype>, %1/m, %2.<Vetype>
   3413    movprfx\t%0, %2\;<sve_fp_op>\t%0.<Vetype>, %1/m, %2.<Vetype>"
   3414   [(set_attr "movprfx" "*,yes")]
   3415 )
   3416 
   3417 ;; Predicated floating-point unary arithmetic with merging.
   3418 (define_expand "@cond_<optab><mode>"
   3419   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   3420 	(unspec:SVE_FULL_F
   3421 	  [(match_operand:<VPRED> 1 "register_operand")
   3422 	   (unspec:SVE_FULL_F
   3423 	     [(match_dup 1)
   3424 	      (const_int SVE_STRICT_GP)
   3425 	      (match_operand:SVE_FULL_F 2 "register_operand")]
   3426 	     SVE_COND_FP_UNARY)
   3427 	   (match_operand:SVE_FULL_F 3 "aarch64_simd_reg_or_zero")]
   3428 	  UNSPEC_SEL))]
   3429   "TARGET_SVE"
   3430 )
   3431 
   3432 ;; Predicated floating-point unary arithmetic, merging with the first input.
   3433 (define_insn_and_rewrite "*cond_<optab><mode>_2_relaxed"
   3434   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   3435 	(unspec:SVE_FULL_F
   3436 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   3437 	   (unspec:SVE_FULL_F
   3438 	     [(match_operand 3)
   3439 	      (const_int SVE_RELAXED_GP)
   3440 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w")]
   3441 	     SVE_COND_FP_UNARY)
   3442 	   (match_dup 2)]
   3443 	  UNSPEC_SEL))]
   3444   "TARGET_SVE"
   3445   "@
   3446    <sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>
   3447    movprfx\t%0, %2\;<sve_fp_op>\t%0.<Vetype>, %1/m, %2.<Vetype>"
   3448   "&& !rtx_equal_p (operands[1], operands[3])"
   3449   {
   3450     operands[3] = copy_rtx (operands[1]);
   3451   }
   3452   [(set_attr "movprfx" "*,yes")]
   3453 )
   3454 
   3455 (define_insn "*cond_<optab><mode>_2_strict"
   3456   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   3457 	(unspec:SVE_FULL_F
   3458 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   3459 	   (unspec:SVE_FULL_F
   3460 	     [(match_dup 1)
   3461 	      (const_int SVE_STRICT_GP)
   3462 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w")]
   3463 	     SVE_COND_FP_UNARY)
   3464 	   (match_dup 2)]
   3465 	  UNSPEC_SEL))]
   3466   "TARGET_SVE"
   3467   "@
   3468    <sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>
   3469    movprfx\t%0, %2\;<sve_fp_op>\t%0.<Vetype>, %1/m, %2.<Vetype>"
   3470   [(set_attr "movprfx" "*,yes")]
   3471 )
   3472 
   3473 ;; Predicated floating-point unary arithmetic, merging with an independent
   3474 ;; value.
   3475 ;;
   3476 ;; The earlyclobber isn't needed for the first alternative, but omitting
   3477 ;; it would only help the case in which operands 2 and 3 are the same,
   3478 ;; which is handled above rather than here.  Marking all the alternatives
   3479 ;; as earlyclobber helps to make the instruction more regular to the
   3480 ;; register allocator.
   3481 (define_insn_and_rewrite "*cond_<optab><mode>_any_relaxed"
   3482   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, ?&w, ?&w")
   3483 	(unspec:SVE_FULL_F
   3484 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   3485 	   (unspec:SVE_FULL_F
   3486 	     [(match_operand 4)
   3487 	      (const_int SVE_RELAXED_GP)
   3488 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w, w")]
   3489 	     SVE_COND_FP_UNARY)
   3490 	   (match_operand:SVE_FULL_F 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   3491 	  UNSPEC_SEL))]
   3492   "TARGET_SVE && !rtx_equal_p (operands[2], operands[3])"
   3493   "@
   3494    <sve_fp_op>\t%0.<Vetype>, %1/m, %2.<Vetype>
   3495    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %2.<Vetype>
   3496    movprfx\t%0, %3\;<sve_fp_op>\t%0.<Vetype>, %1/m, %2.<Vetype>"
   3497   "&& !rtx_equal_p (operands[1], operands[4])"
   3498   {
   3499     operands[4] = copy_rtx (operands[1]);
   3500   }
   3501   [(set_attr "movprfx" "*,yes,yes")]
   3502 )
   3503 
   3504 (define_insn "*cond_<optab><mode>_any_strict"
   3505   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, ?&w, ?&w")
   3506 	(unspec:SVE_FULL_F
   3507 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   3508 	   (unspec:SVE_FULL_F
   3509 	     [(match_dup 1)
   3510 	      (const_int SVE_STRICT_GP)
   3511 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w, w")]
   3512 	     SVE_COND_FP_UNARY)
   3513 	   (match_operand:SVE_FULL_F 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   3514 	  UNSPEC_SEL))]
   3515   "TARGET_SVE && !rtx_equal_p (operands[2], operands[3])"
   3516   "@
   3517    <sve_fp_op>\t%0.<Vetype>, %1/m, %2.<Vetype>
   3518    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %2.<Vetype>
   3519    movprfx\t%0, %3\;<sve_fp_op>\t%0.<Vetype>, %1/m, %2.<Vetype>"
   3520   [(set_attr "movprfx" "*,yes,yes")]
   3521 )
   3522 
   3523 ;; -------------------------------------------------------------------------
   3524 ;; ---- [FP] Square root
   3525 ;; -------------------------------------------------------------------------
   3526 
   3527 (define_expand "sqrt<mode>2"
   3528   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   3529 	(unspec:SVE_FULL_F
   3530 	  [(match_dup 2)
   3531 	   (const_int SVE_RELAXED_GP)
   3532 	   (match_operand:SVE_FULL_F 1 "register_operand")]
   3533 	  UNSPEC_COND_FSQRT))]
   3534   "TARGET_SVE"
   3535 {
   3536   if (aarch64_emit_approx_sqrt (operands[0], operands[1], false))
   3537     DONE;
   3538   operands[2] = aarch64_ptrue_reg (<VPRED>mode);
   3539 })
   3540 
   3541 ;; -------------------------------------------------------------------------
   3542 ;; ---- [FP] Reciprocal square root
   3543 ;; -------------------------------------------------------------------------
   3544 
   3545 (define_expand "rsqrt<mode>2"
   3546   [(set (match_operand:SVE_FULL_SDF 0 "register_operand")
   3547 	(unspec:SVE_FULL_SDF
   3548 	  [(match_operand:SVE_FULL_SDF 1 "register_operand")]
   3549 	  UNSPEC_RSQRT))]
   3550   "TARGET_SVE"
   3551 {
   3552   aarch64_emit_approx_sqrt (operands[0], operands[1], true);
   3553   DONE;
   3554 })
   3555 
   3556 (define_expand "@aarch64_rsqrte<mode>"
   3557   [(set (match_operand:SVE_FULL_SDF 0 "register_operand")
   3558 	(unspec:SVE_FULL_SDF
   3559 	  [(match_operand:SVE_FULL_SDF 1 "register_operand")]
   3560 	  UNSPEC_RSQRTE))]
   3561   "TARGET_SVE"
   3562 )
   3563 
   3564 (define_expand "@aarch64_rsqrts<mode>"
   3565   [(set (match_operand:SVE_FULL_SDF 0 "register_operand")
   3566 	(unspec:SVE_FULL_SDF
   3567 	  [(match_operand:SVE_FULL_SDF 1 "register_operand")
   3568 	   (match_operand:SVE_FULL_SDF 2 "register_operand")]
   3569 	  UNSPEC_RSQRTS))]
   3570   "TARGET_SVE"
   3571 )
   3572 
   3573 ;; -------------------------------------------------------------------------
   3574 ;; ---- [PRED] Inverse
   3575 ;; -------------------------------------------------------------------------
   3576 ;; Includes:
   3577 ;; - NOT
   3578 ;; -------------------------------------------------------------------------
   3579 
   3580 ;; Unpredicated predicate inverse.
   3581 (define_expand "one_cmpl<mode>2"
   3582   [(set (match_operand:PRED_ALL 0 "register_operand")
   3583 	(and:PRED_ALL
   3584 	  (not:PRED_ALL (match_operand:PRED_ALL 1 "register_operand"))
   3585 	  (match_dup 2)))]
   3586   "TARGET_SVE"
   3587   {
   3588     operands[2] = aarch64_ptrue_reg (<MODE>mode);
   3589   }
   3590 )
   3591 
   3592 ;; Predicated predicate inverse.
   3593 (define_insn "*one_cmpl<mode>3"
   3594   [(set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   3595 	(and:PRED_ALL
   3596 	  (not:PRED_ALL (match_operand:PRED_ALL 2 "register_operand" "Upa"))
   3597 	  (match_operand:PRED_ALL 1 "register_operand" "Upa")))]
   3598   "TARGET_SVE"
   3599   "not\t%0.b, %1/z, %2.b"
   3600 )
   3601 
   3602 ;; =========================================================================
   3603 ;; == Binary arithmetic
   3604 ;; =========================================================================
   3605 
   3606 ;; -------------------------------------------------------------------------
   3607 ;; ---- [INT] General binary arithmetic corresponding to rtx codes
   3608 ;; -------------------------------------------------------------------------
   3609 ;; Includes:
   3610 ;; - ADD    (merging form only)
   3611 ;; - AND    (merging form only)
   3612 ;; - ASR    (merging form only)
   3613 ;; - EOR    (merging form only)
   3614 ;; - LSL    (merging form only)
   3615 ;; - LSR    (merging form only)
   3616 ;; - MUL
   3617 ;; - ORR    (merging form only)
   3618 ;; - SMAX
   3619 ;; - SMIN
   3620 ;; - SQADD  (SVE2 merging form only)
   3621 ;; - SQSUB  (SVE2 merging form only)
   3622 ;; - SUB    (merging form only)
   3623 ;; - UMAX
   3624 ;; - UMIN
   3625 ;; - UQADD  (SVE2 merging form only)
   3626 ;; - UQSUB  (SVE2 merging form only)
   3627 ;; -------------------------------------------------------------------------
   3628 
   3629 ;; Unpredicated integer binary operations that have an immediate form.
   3630 (define_expand "<optab><mode>3"
   3631   [(set (match_operand:SVE_I 0 "register_operand")
   3632 	(unspec:SVE_I
   3633 	  [(match_dup 3)
   3634 	   (SVE_INT_BINARY_IMM:SVE_I
   3635 	     (match_operand:SVE_I 1 "register_operand")
   3636 	     (match_operand:SVE_I 2 "aarch64_sve_<sve_imm_con>_operand"))]
   3637 	  UNSPEC_PRED_X))]
   3638   "TARGET_SVE"
   3639   {
   3640     operands[3] = aarch64_ptrue_reg (<VPRED>mode);
   3641   }
   3642 )
   3643 
   3644 ;; Integer binary operations that have an immediate form, predicated
   3645 ;; with a PTRUE.  We don't actually need the predicate for the first
   3646 ;; and third alternatives, but using Upa or X isn't likely to gain much
   3647 ;; and would make the instruction seem less uniform to the register
   3648 ;; allocator.
   3649 (define_insn_and_split "@aarch64_pred_<optab><mode>"
   3650   [(set (match_operand:SVE_I 0 "register_operand" "=w, w, ?&w, ?&w")
   3651 	(unspec:SVE_I
   3652 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl")
   3653 	   (SVE_INT_BINARY_IMM:SVE_I
   3654 	     (match_operand:SVE_I 2 "register_operand" "%0, 0, w, w")
   3655 	     (match_operand:SVE_I 3 "aarch64_sve_<sve_imm_con>_operand" "<sve_imm_con>, w, <sve_imm_con>, w"))]
   3656 	  UNSPEC_PRED_X))]
   3657   "TARGET_SVE"
   3658   "@
   3659    #
   3660    <sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   3661    #
   3662    movprfx\t%0, %2\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   3663   ; Split the unpredicated form after reload, so that we don't have
   3664   ; the unnecessary PTRUE.
   3665   "&& reload_completed
   3666    && !register_operand (operands[3], <MODE>mode)"
   3667   [(set (match_dup 0)
   3668 	(SVE_INT_BINARY_IMM:SVE_I (match_dup 2) (match_dup 3)))]
   3669   ""
   3670   [(set_attr "movprfx" "*,*,yes,yes")]
   3671 )
   3672 
   3673 ;; Unpredicated binary operations with a constant (post-RA only).
   3674 ;; These are generated by splitting a predicated instruction whose
   3675 ;; predicate is unused.
   3676 (define_insn "*post_ra_<optab><mode>3"
   3677   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   3678 	(SVE_INT_BINARY_IMM:SVE_I
   3679 	  (match_operand:SVE_I 1 "register_operand" "0, w")
   3680 	  (match_operand:SVE_I 2 "aarch64_sve_<sve_imm_con>_immediate")))]
   3681   "TARGET_SVE && reload_completed"
   3682   "@
   3683    <sve_int_op>\t%0.<Vetype>, %0.<Vetype>, #%<sve_imm_prefix>2
   3684    movprfx\t%0, %1\;<sve_int_op>\t%0.<Vetype>, %0.<Vetype>, #%<sve_imm_prefix>2"
   3685   [(set_attr "movprfx" "*,yes")]
   3686 )
   3687 
   3688 ;; Predicated integer operations with merging.
   3689 (define_expand "@cond_<optab><mode>"
   3690   [(set (match_operand:SVE_I 0 "register_operand")
   3691 	(unspec:SVE_I
   3692 	  [(match_operand:<VPRED> 1 "register_operand")
   3693 	   (SVE_INT_BINARY:SVE_I
   3694 	     (match_operand:SVE_I 2 "register_operand")
   3695 	     (match_operand:SVE_I 3 "<sve_pred_int_rhs2_operand>"))
   3696 	   (match_operand:SVE_I 4 "aarch64_simd_reg_or_zero")]
   3697 	  UNSPEC_SEL))]
   3698   "TARGET_SVE"
   3699 )
   3700 
   3701 ;; Predicated integer operations, merging with the first input.
   3702 (define_insn "*cond_<optab><mode>_2"
   3703   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   3704 	(unspec:SVE_I
   3705 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   3706 	   (SVE_INT_BINARY:SVE_I
   3707 	     (match_operand:SVE_I 2 "register_operand" "0, w")
   3708 	     (match_operand:SVE_I 3 "register_operand" "w, w"))
   3709 	   (match_dup 2)]
   3710 	  UNSPEC_SEL))]
   3711   "TARGET_SVE"
   3712   "@
   3713    <sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   3714    movprfx\t%0, %2\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   3715   [(set_attr "movprfx" "*,yes")]
   3716 )
   3717 
   3718 ;; Predicated integer operations, merging with the second input.
   3719 (define_insn "*cond_<optab><mode>_3"
   3720   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   3721 	(unspec:SVE_I
   3722 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   3723 	   (SVE_INT_BINARY:SVE_I
   3724 	     (match_operand:SVE_I 2 "register_operand" "w, w")
   3725 	     (match_operand:SVE_I 3 "register_operand" "0, w"))
   3726 	   (match_dup 3)]
   3727 	  UNSPEC_SEL))]
   3728   "TARGET_SVE"
   3729   "@
   3730    <sve_int_op_rev>\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   3731    movprfx\t%0, %3\;<sve_int_op_rev>\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>"
   3732   [(set_attr "movprfx" "*,yes")]
   3733 )
   3734 
   3735 ;; Predicated integer operations, merging with an independent value.
   3736 (define_insn_and_rewrite "*cond_<optab><mode>_any"
   3737   [(set (match_operand:SVE_I 0 "register_operand" "=&w, &w, &w, &w, ?&w")
   3738 	(unspec:SVE_I
   3739 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl")
   3740 	   (SVE_INT_BINARY:SVE_I
   3741 	     (match_operand:SVE_I 2 "register_operand" "0, w, w, w, w")
   3742 	     (match_operand:SVE_I 3 "register_operand" "w, 0, w, w, w"))
   3743 	   (match_operand:SVE_I 4 "aarch64_simd_reg_or_zero" "Dz, Dz, Dz, 0, w")]
   3744 	  UNSPEC_SEL))]
   3745   "TARGET_SVE
   3746    && !rtx_equal_p (operands[2], operands[4])
   3747    && !rtx_equal_p (operands[3], operands[4])"
   3748   "@
   3749    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   3750    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_int_op_rev>\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   3751    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   3752    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   3753    #"
   3754   "&& reload_completed
   3755    && register_operand (operands[4], <MODE>mode)
   3756    && !rtx_equal_p (operands[0], operands[4])"
   3757   {
   3758     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   3759 					     operands[4], operands[1]));
   3760     operands[4] = operands[2] = operands[0];
   3761   }
   3762   [(set_attr "movprfx" "yes")]
   3763 )
   3764 
   3765 ;; -------------------------------------------------------------------------
   3766 ;; ---- [INT] Addition
   3767 ;; -------------------------------------------------------------------------
   3768 ;; Includes:
   3769 ;; - ADD
   3770 ;; - DECB
   3771 ;; - DECD
   3772 ;; - DECH
   3773 ;; - DECW
   3774 ;; - INCB
   3775 ;; - INCD
   3776 ;; - INCH
   3777 ;; - INCW
   3778 ;; - SUB
   3779 ;; -------------------------------------------------------------------------
   3780 
   3781 (define_insn "add<mode>3"
   3782   [(set (match_operand:SVE_I 0 "register_operand" "=w, w, w, ?w, ?w, w")
   3783 	(plus:SVE_I
   3784 	  (match_operand:SVE_I 1 "register_operand" "%0, 0, 0, w, w, w")
   3785 	  (match_operand:SVE_I 2 "aarch64_sve_add_operand" "vsa, vsn, vsi, vsa, vsn, w")))]
   3786   "TARGET_SVE"
   3787   "@
   3788    add\t%0.<Vetype>, %0.<Vetype>, #%D2
   3789    sub\t%0.<Vetype>, %0.<Vetype>, #%N2
   3790    * return aarch64_output_sve_vector_inc_dec (\"%0.<Vetype>\", operands[2]);
   3791    movprfx\t%0, %1\;add\t%0.<Vetype>, %0.<Vetype>, #%D2
   3792    movprfx\t%0, %1\;sub\t%0.<Vetype>, %0.<Vetype>, #%N2
   3793    add\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>"
   3794   [(set_attr "movprfx" "*,*,*,yes,yes,*")]
   3795 )
   3796 
   3797 ;; Merging forms are handled through SVE_INT_BINARY.
   3798 
   3799 ;; -------------------------------------------------------------------------
   3800 ;; ---- [INT] Subtraction
   3801 ;; -------------------------------------------------------------------------
   3802 ;; Includes:
   3803 ;; - SUB
   3804 ;; - SUBR
   3805 ;; -------------------------------------------------------------------------
   3806 
   3807 (define_insn "sub<mode>3"
   3808   [(set (match_operand:SVE_I 0 "register_operand" "=w, w, ?&w")
   3809 	(minus:SVE_I
   3810 	  (match_operand:SVE_I 1 "aarch64_sve_arith_operand" "w, vsa, vsa")
   3811 	  (match_operand:SVE_I 2 "register_operand" "w, 0, w")))]
   3812   "TARGET_SVE"
   3813   "@
   3814    sub\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>
   3815    subr\t%0.<Vetype>, %0.<Vetype>, #%D1
   3816    movprfx\t%0, %2\;subr\t%0.<Vetype>, %0.<Vetype>, #%D1"
   3817   [(set_attr "movprfx" "*,*,yes")]
   3818 )
   3819 
   3820 ;; Merging forms are handled through SVE_INT_BINARY.
   3821 
   3822 ;; -------------------------------------------------------------------------
   3823 ;; ---- [INT] Take address
   3824 ;; -------------------------------------------------------------------------
   3825 ;; Includes:
   3826 ;; - ADR
   3827 ;; -------------------------------------------------------------------------
   3828 
   3829 ;; An unshifted and unscaled ADR.  This is functionally equivalent to an ADD,
   3830 ;; but the svadrb intrinsics should preserve the user's choice.
   3831 (define_insn "@aarch64_adr<mode>"
   3832   [(set (match_operand:SVE_FULL_SDI 0 "register_operand" "=w")
   3833 	(unspec:SVE_FULL_SDI
   3834 	  [(match_operand:SVE_FULL_SDI 1 "register_operand" "w")
   3835 	   (match_operand:SVE_FULL_SDI 2 "register_operand" "w")]
   3836 	  UNSPEC_ADR))]
   3837   "TARGET_SVE"
   3838   "adr\t%0.<Vetype>, [%1.<Vetype>, %2.<Vetype>]"
   3839 )
   3840 
   3841 ;; Same, but with the offset being sign-extended from the low 32 bits.
   3842 (define_insn_and_rewrite "*aarch64_adr_sxtw"
   3843   [(set (match_operand:VNx2DI 0 "register_operand" "=w")
   3844 	(unspec:VNx2DI
   3845 	  [(match_operand:VNx2DI 1 "register_operand" "w")
   3846 	   (unspec:VNx2DI
   3847 	     [(match_operand 3)
   3848 	      (sign_extend:VNx2DI
   3849 		(truncate:VNx2SI
   3850 		  (match_operand:VNx2DI 2 "register_operand" "w")))]
   3851 	     UNSPEC_PRED_X)]
   3852 	  UNSPEC_ADR))]
   3853   "TARGET_SVE"
   3854   "adr\t%0.d, [%1.d, %2.d, sxtw]"
   3855   "&& !CONSTANT_P (operands[3])"
   3856   {
   3857     operands[3] = CONSTM1_RTX (VNx2BImode);
   3858   }
   3859 )
   3860 
   3861 ;; Same, but with the offset being zero-extended from the low 32 bits.
   3862 (define_insn "*aarch64_adr_uxtw_unspec"
   3863   [(set (match_operand:VNx2DI 0 "register_operand" "=w")
   3864 	(unspec:VNx2DI
   3865 	  [(match_operand:VNx2DI 1 "register_operand" "w")
   3866 	   (and:VNx2DI
   3867 	     (match_operand:VNx2DI 2 "register_operand" "w")
   3868 	     (match_operand:VNx2DI 3 "aarch64_sve_uxtw_immediate"))]
   3869 	  UNSPEC_ADR))]
   3870   "TARGET_SVE"
   3871   "adr\t%0.d, [%1.d, %2.d, uxtw]"
   3872 )
   3873 
   3874 ;; Same, matching as a PLUS rather than unspec.
   3875 (define_insn "*aarch64_adr_uxtw_and"
   3876   [(set (match_operand:VNx2DI 0 "register_operand" "=w")
   3877 	(plus:VNx2DI
   3878 	  (and:VNx2DI
   3879 	    (match_operand:VNx2DI 2 "register_operand" "w")
   3880 	    (match_operand:VNx2DI 3 "aarch64_sve_uxtw_immediate"))
   3881 	  (match_operand:VNx2DI 1 "register_operand" "w")))]
   3882   "TARGET_SVE"
   3883   "adr\t%0.d, [%1.d, %2.d, uxtw]"
   3884 )
   3885 
   3886 ;; ADR with a nonzero shift.
   3887 (define_expand "@aarch64_adr<mode>_shift"
   3888   [(set (match_operand:SVE_FULL_SDI 0 "register_operand")
   3889 	(plus:SVE_FULL_SDI
   3890 	  (unspec:SVE_FULL_SDI
   3891 	    [(match_dup 4)
   3892 	     (ashift:SVE_FULL_SDI
   3893 	       (match_operand:SVE_FULL_SDI 2 "register_operand")
   3894 	       (match_operand:SVE_FULL_SDI 3 "const_1_to_3_operand"))]
   3895 	    UNSPEC_PRED_X)
   3896 	  (match_operand:SVE_FULL_SDI 1 "register_operand")))]
   3897   "TARGET_SVE"
   3898   {
   3899     operands[4] = CONSTM1_RTX (<VPRED>mode);
   3900   }
   3901 )
   3902 
   3903 (define_insn_and_rewrite "*aarch64_adr<mode>_shift"
   3904   [(set (match_operand:SVE_24I 0 "register_operand" "=w")
   3905 	(plus:SVE_24I
   3906 	  (unspec:SVE_24I
   3907 	    [(match_operand 4)
   3908 	     (ashift:SVE_24I
   3909 	       (match_operand:SVE_24I 2 "register_operand" "w")
   3910 	       (match_operand:SVE_24I 3 "const_1_to_3_operand"))]
   3911 	    UNSPEC_PRED_X)
   3912 	  (match_operand:SVE_24I 1 "register_operand" "w")))]
   3913   "TARGET_SVE"
   3914   "adr\t%0.<Vctype>, [%1.<Vctype>, %2.<Vctype>, lsl %3]"
   3915   "&& !CONSTANT_P (operands[4])"
   3916   {
   3917     operands[4] = CONSTM1_RTX (<VPRED>mode);
   3918   }
   3919 )
   3920 
   3921 ;; Same, but with the index being sign-extended from the low 32 bits.
   3922 (define_insn_and_rewrite "*aarch64_adr_shift_sxtw"
   3923   [(set (match_operand:VNx2DI 0 "register_operand" "=w")
   3924 	(plus:VNx2DI
   3925 	  (unspec:VNx2DI
   3926 	    [(match_operand 4)
   3927 	     (ashift:VNx2DI
   3928 	       (unspec:VNx2DI
   3929 		 [(match_operand 5)
   3930 		  (sign_extend:VNx2DI
   3931 		    (truncate:VNx2SI
   3932 		      (match_operand:VNx2DI 2 "register_operand" "w")))]
   3933 		 UNSPEC_PRED_X)
   3934 	       (match_operand:VNx2DI 3 "const_1_to_3_operand"))]
   3935 	    UNSPEC_PRED_X)
   3936 	  (match_operand:VNx2DI 1 "register_operand" "w")))]
   3937   "TARGET_SVE"
   3938   "adr\t%0.d, [%1.d, %2.d, sxtw %3]"
   3939   "&& (!CONSTANT_P (operands[4]) || !CONSTANT_P (operands[5]))"
   3940   {
   3941     operands[5] = operands[4] = CONSTM1_RTX (VNx2BImode);
   3942   }
   3943 )
   3944 
   3945 ;; Same, but with the index being zero-extended from the low 32 bits.
   3946 (define_insn_and_rewrite "*aarch64_adr_shift_uxtw"
   3947   [(set (match_operand:VNx2DI 0 "register_operand" "=w")
   3948 	(plus:VNx2DI
   3949 	  (unspec:VNx2DI
   3950 	    [(match_operand 5)
   3951 	     (ashift:VNx2DI
   3952 	       (and:VNx2DI
   3953 		 (match_operand:VNx2DI 2 "register_operand" "w")
   3954 		 (match_operand:VNx2DI 4 "aarch64_sve_uxtw_immediate"))
   3955 	       (match_operand:VNx2DI 3 "const_1_to_3_operand"))]
   3956 	    UNSPEC_PRED_X)
   3957 	  (match_operand:VNx2DI 1 "register_operand" "w")))]
   3958   "TARGET_SVE"
   3959   "adr\t%0.d, [%1.d, %2.d, uxtw %3]"
   3960   "&& !CONSTANT_P (operands[5])"
   3961   {
   3962     operands[5] = CONSTM1_RTX (VNx2BImode);
   3963   }
   3964 )
   3965 
   3966 ;; -------------------------------------------------------------------------
   3967 ;; ---- [INT] Absolute difference
   3968 ;; -------------------------------------------------------------------------
   3969 ;; Includes:
   3970 ;; - SABD
   3971 ;; - UABD
   3972 ;; -------------------------------------------------------------------------
   3973 
   3974 ;; Unpredicated integer absolute difference.
   3975 (define_expand "<su>abd<mode>_3"
   3976   [(use (match_operand:SVE_I 0 "register_operand"))
   3977    (USMAX:SVE_I
   3978      (match_operand:SVE_I 1 "register_operand")
   3979      (match_operand:SVE_I 2 "register_operand"))]
   3980   "TARGET_SVE"
   3981   {
   3982     rtx pred = aarch64_ptrue_reg (<VPRED>mode);
   3983     emit_insn (gen_aarch64_pred_<su>abd<mode> (operands[0], pred, operands[1],
   3984 					       operands[2]));
   3985     DONE;
   3986   }
   3987 )
   3988 
   3989 ;; Predicated integer absolute difference.
   3990 (define_insn "@aarch64_pred_<su>abd<mode>"
   3991   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   3992 	(minus:SVE_I
   3993 	  (unspec:SVE_I
   3994 	    [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   3995 	     (USMAX:SVE_I
   3996 	       (match_operand:SVE_I 2 "register_operand" "%0, w")
   3997 	       (match_operand:SVE_I 3 "register_operand" "w, w"))]
   3998 	    UNSPEC_PRED_X)
   3999 	  (unspec:SVE_I
   4000 	    [(match_dup 1)
   4001 	     (<max_opp>:SVE_I
   4002 	       (match_dup 2)
   4003 	       (match_dup 3))]
   4004 	    UNSPEC_PRED_X)))]
   4005   "TARGET_SVE"
   4006   "@
   4007    <su>abd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4008    movprfx\t%0, %2\;<su>abd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   4009   [(set_attr "movprfx" "*,yes")]
   4010 )
   4011 
   4012 (define_expand "@aarch64_cond_<su>abd<mode>"
   4013   [(set (match_operand:SVE_FULL_I 0 "register_operand")
   4014 	(unspec:SVE_FULL_I
   4015 	  [(match_operand:<VPRED> 1 "register_operand")
   4016 	   (minus:SVE_FULL_I
   4017 	     (unspec:SVE_FULL_I
   4018 	       [(match_dup 1)
   4019 		(USMAX:SVE_FULL_I
   4020 		  (match_operand:SVE_FULL_I 2 "register_operand")
   4021 		  (match_operand:SVE_FULL_I 3 "register_operand"))]
   4022 	       UNSPEC_PRED_X)
   4023 	     (unspec:SVE_FULL_I
   4024 	       [(match_dup 1)
   4025 		(<max_opp>:SVE_FULL_I
   4026 		  (match_dup 2)
   4027 		  (match_dup 3))]
   4028 	       UNSPEC_PRED_X))
   4029 	   (match_operand:SVE_FULL_I 4 "aarch64_simd_reg_or_zero")]
   4030 	  UNSPEC_SEL))]
   4031   "TARGET_SVE"
   4032 {
   4033   if (rtx_equal_p (operands[3], operands[4]))
   4034     std::swap (operands[2], operands[3]);
   4035 })
   4036 
   4037 ;; Predicated integer absolute difference, merging with the first input.
   4038 (define_insn_and_rewrite "*aarch64_cond_<su>abd<mode>_2"
   4039   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   4040 	(unspec:SVE_I
   4041 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4042 	   (minus:SVE_I
   4043 	     (unspec:SVE_I
   4044 	       [(match_operand 4)
   4045 		(USMAX:SVE_I
   4046 		  (match_operand:SVE_I 2 "register_operand" "0, w")
   4047 		  (match_operand:SVE_I 3 "register_operand" "w, w"))]
   4048 	       UNSPEC_PRED_X)
   4049 	     (unspec:SVE_I
   4050 	       [(match_operand 5)
   4051 		(<max_opp>:SVE_I
   4052 		  (match_dup 2)
   4053 		  (match_dup 3))]
   4054 	       UNSPEC_PRED_X))
   4055 	   (match_dup 2)]
   4056 	  UNSPEC_SEL))]
   4057   "TARGET_SVE"
   4058   "@
   4059    <su>abd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4060    movprfx\t%0, %2\;<su>abd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   4061   "&& (!CONSTANT_P (operands[4]) || !CONSTANT_P (operands[5]))"
   4062   {
   4063     operands[4] = operands[5] = CONSTM1_RTX (<VPRED>mode);
   4064   }
   4065   [(set_attr "movprfx" "*,yes")]
   4066 )
   4067 
   4068 ;; Predicated integer absolute difference, merging with the second input.
   4069 (define_insn_and_rewrite "*aarch64_cond_<su>abd<mode>_3"
   4070   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   4071 	(unspec:SVE_I
   4072 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4073 	   (minus:SVE_I
   4074 	     (unspec:SVE_I
   4075 	       [(match_operand 4)
   4076 		(USMAX:SVE_I
   4077 		  (match_operand:SVE_I 2 "register_operand" "w, w")
   4078 		  (match_operand:SVE_I 3 "register_operand" "0, w"))]
   4079 	       UNSPEC_PRED_X)
   4080 	     (unspec:SVE_I
   4081 	       [(match_operand 5)
   4082 		(<max_opp>:SVE_I
   4083 		  (match_dup 2)
   4084 		  (match_dup 3))]
   4085 	       UNSPEC_PRED_X))
   4086 	   (match_dup 3)]
   4087 	  UNSPEC_SEL))]
   4088   "TARGET_SVE"
   4089   "@
   4090    <su>abd\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   4091    movprfx\t%0, %3\;<su>abd\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>"
   4092   "&& (!CONSTANT_P (operands[4]) || !CONSTANT_P (operands[5]))"
   4093   {
   4094     operands[4] = operands[5] = CONSTM1_RTX (<VPRED>mode);
   4095   }
   4096   [(set_attr "movprfx" "*,yes")]
   4097 )
   4098 
   4099 ;; Predicated integer absolute difference, merging with an independent value.
   4100 (define_insn_and_rewrite "*aarch64_cond_<su>abd<mode>_any"
   4101   [(set (match_operand:SVE_I 0 "register_operand" "=&w, &w, &w, &w, ?&w")
   4102 	(unspec:SVE_I
   4103 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl")
   4104 	   (minus:SVE_I
   4105 	     (unspec:SVE_I
   4106 	       [(match_operand 5)
   4107 		(USMAX:SVE_I
   4108 		  (match_operand:SVE_I 2 "register_operand" "0, w, w, w, w")
   4109 		  (match_operand:SVE_I 3 "register_operand" "w, 0, w, w, w"))]
   4110 	       UNSPEC_PRED_X)
   4111 	     (unspec:SVE_I
   4112 	       [(match_operand 6)
   4113 		(<max_opp>:SVE_I
   4114 		  (match_dup 2)
   4115 		  (match_dup 3))]
   4116 	       UNSPEC_PRED_X))
   4117 	   (match_operand:SVE_I 4 "aarch64_simd_reg_or_zero" "Dz, Dz, Dz, 0, w")]
   4118 	  UNSPEC_SEL))]
   4119   "TARGET_SVE
   4120    && !rtx_equal_p (operands[2], operands[4])
   4121    && !rtx_equal_p (operands[3], operands[4])"
   4122   "@
   4123    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<su>abd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4124    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<su>abd\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   4125    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<su>abd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4126    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;<su>abd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4127    #"
   4128   "&& 1"
   4129   {
   4130     if (!CONSTANT_P (operands[5]) || !CONSTANT_P (operands[6]))
   4131       operands[5] = operands[6] = CONSTM1_RTX (<VPRED>mode);
   4132     else if (reload_completed
   4133 	     && register_operand (operands[4], <MODE>mode)
   4134 	     && !rtx_equal_p (operands[0], operands[4]))
   4135       {
   4136 	emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   4137 						 operands[4], operands[1]));
   4138 	operands[4] = operands[2] = operands[0];
   4139       }
   4140     else
   4141       FAIL;
   4142   }
   4143   [(set_attr "movprfx" "yes")]
   4144 )
   4145 
   4146 ;; -------------------------------------------------------------------------
   4147 ;; ---- [INT] Saturating addition and subtraction
   4148 ;; -------------------------------------------------------------------------
   4149 ;; - SQADD
   4150 ;; - SQSUB
   4151 ;; - UQADD
   4152 ;; - UQSUB
   4153 ;; -------------------------------------------------------------------------
   4154 
   4155 ;; Unpredicated saturating signed addition and subtraction.
   4156 (define_insn "@aarch64_sve_<optab><mode>"
   4157   [(set (match_operand:SVE_FULL_I 0 "register_operand" "=w, w, ?&w, ?&w, w")
   4158 	(SBINQOPS:SVE_FULL_I
   4159 	  (match_operand:SVE_FULL_I 1 "register_operand" "0, 0, w, w, w")
   4160 	  (match_operand:SVE_FULL_I 2 "aarch64_sve_sqadd_operand" "vsQ, vsS, vsQ, vsS, w")))]
   4161   "TARGET_SVE"
   4162   "@
   4163    <binqops_op>\t%0.<Vetype>, %0.<Vetype>, #%D2
   4164    <binqops_op_rev>\t%0.<Vetype>, %0.<Vetype>, #%N2
   4165    movprfx\t%0, %1\;<binqops_op>\t%0.<Vetype>, %0.<Vetype>, #%D2
   4166    movprfx\t%0, %1\;<binqops_op_rev>\t%0.<Vetype>, %0.<Vetype>, #%N2
   4167    <binqops_op>\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>"
   4168   [(set_attr "movprfx" "*,*,yes,yes,*")]
   4169 )
   4170 
   4171 ;; Unpredicated saturating unsigned addition and subtraction.
   4172 (define_insn "@aarch64_sve_<optab><mode>"
   4173   [(set (match_operand:SVE_FULL_I 0 "register_operand" "=w, ?&w, w")
   4174 	(UBINQOPS:SVE_FULL_I
   4175 	  (match_operand:SVE_FULL_I 1 "register_operand" "0, w, w")
   4176 	  (match_operand:SVE_FULL_I 2 "aarch64_sve_arith_operand" "vsa, vsa, w")))]
   4177   "TARGET_SVE"
   4178   "@
   4179    <binqops_op>\t%0.<Vetype>, %0.<Vetype>, #%D2
   4180    movprfx\t%0, %1\;<binqops_op>\t%0.<Vetype>, %0.<Vetype>, #%D2
   4181    <binqops_op>\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>"
   4182   [(set_attr "movprfx" "*,yes,*")]
   4183 )
   4184 
   4185 ;; -------------------------------------------------------------------------
   4186 ;; ---- [INT] Highpart multiplication
   4187 ;; -------------------------------------------------------------------------
   4188 ;; Includes:
   4189 ;; - SMULH
   4190 ;; - UMULH
   4191 ;; -------------------------------------------------------------------------
   4192 
   4193 ;; Unpredicated highpart multiplication.
   4194 (define_expand "<su>mul<mode>3_highpart"
   4195   [(set (match_operand:SVE_I 0 "register_operand")
   4196 	(unspec:SVE_I
   4197 	  [(match_dup 3)
   4198 	   (unspec:SVE_I
   4199 	     [(match_operand:SVE_I 1 "register_operand")
   4200 	      (match_operand:SVE_I 2 "register_operand")]
   4201 	     MUL_HIGHPART)]
   4202 	  UNSPEC_PRED_X))]
   4203   "TARGET_SVE"
   4204   {
   4205     operands[3] = aarch64_ptrue_reg (<VPRED>mode);
   4206   }
   4207 )
   4208 
   4209 ;; Predicated highpart multiplication.
   4210 (define_insn "@aarch64_pred_<optab><mode>"
   4211   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   4212 	(unspec:SVE_I
   4213 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4214 	   (unspec:SVE_I
   4215 	     [(match_operand:SVE_I 2 "register_operand" "%0, w")
   4216 	      (match_operand:SVE_I 3 "register_operand" "w, w")]
   4217 	     MUL_HIGHPART)]
   4218 	  UNSPEC_PRED_X))]
   4219   "TARGET_SVE"
   4220   "@
   4221    <su>mulh\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4222    movprfx\t%0, %2\;<su>mulh\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   4223   [(set_attr "movprfx" "*,yes")]
   4224 )
   4225 
   4226 ;; Predicated highpart multiplications with merging.
   4227 (define_expand "@cond_<optab><mode>"
   4228   [(set (match_operand:SVE_FULL_I 0 "register_operand")
   4229 	(unspec:SVE_FULL_I
   4230 	  [(match_operand:<VPRED> 1 "register_operand")
   4231 	   (unspec:SVE_FULL_I
   4232 	     [(match_operand:SVE_FULL_I 2 "register_operand")
   4233 	      (match_operand:SVE_FULL_I 3 "register_operand")]
   4234 	     MUL_HIGHPART)
   4235 	   (match_operand:SVE_FULL_I 4 "aarch64_simd_reg_or_zero")]
   4236 	  UNSPEC_SEL))]
   4237   "TARGET_SVE"
   4238 {
   4239   /* Only target code is aware of these operations, so we don't need
   4240      to handle the fully-general case.  */
   4241   gcc_assert (rtx_equal_p (operands[2], operands[4])
   4242 	      || CONSTANT_P (operands[4]));
   4243 })
   4244 
   4245 ;; Predicated highpart multiplications, merging with the first input.
   4246 (define_insn "*cond_<optab><mode>_2"
   4247   [(set (match_operand:SVE_FULL_I 0 "register_operand" "=w, ?&w")
   4248 	(unspec:SVE_FULL_I
   4249 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4250 	   (unspec:SVE_FULL_I
   4251 	     [(match_operand:SVE_FULL_I 2 "register_operand" "0, w")
   4252 	      (match_operand:SVE_FULL_I 3 "register_operand" "w, w")]
   4253 	     MUL_HIGHPART)
   4254 	   (match_dup 2)]
   4255 	  UNSPEC_SEL))]
   4256   "TARGET_SVE"
   4257   "@
   4258    <sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4259    movprfx\t%0, %2\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   4260   [(set_attr "movprfx" "*,yes")])
   4261 
   4262 ;; Predicated highpart multiplications, merging with zero.
   4263 (define_insn "*cond_<optab><mode>_z"
   4264   [(set (match_operand:SVE_FULL_I 0 "register_operand" "=&w, &w")
   4265 	(unspec:SVE_FULL_I
   4266 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4267 	   (unspec:SVE_FULL_I
   4268 	     [(match_operand:SVE_FULL_I 2 "register_operand" "%0, w")
   4269 	      (match_operand:SVE_FULL_I 3 "register_operand" "w, w")]
   4270 	     MUL_HIGHPART)
   4271 	   (match_operand:SVE_FULL_I 4 "aarch64_simd_imm_zero")]
   4272 	  UNSPEC_SEL))]
   4273   "TARGET_SVE"
   4274   "@
   4275    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4276    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   4277   [(set_attr "movprfx" "yes")])
   4278 
   4279 ;; -------------------------------------------------------------------------
   4280 ;; ---- [INT] Division
   4281 ;; -------------------------------------------------------------------------
   4282 ;; Includes:
   4283 ;; - SDIV
   4284 ;; - SDIVR
   4285 ;; - UDIV
   4286 ;; - UDIVR
   4287 ;; -------------------------------------------------------------------------
   4288 
   4289 ;; Unpredicated integer division.
   4290 (define_expand "<optab><mode>3"
   4291   [(set (match_operand:SVE_FULL_SDI 0 "register_operand")
   4292 	(unspec:SVE_FULL_SDI
   4293 	  [(match_dup 3)
   4294 	   (SVE_INT_BINARY_SD:SVE_FULL_SDI
   4295 	     (match_operand:SVE_FULL_SDI 1 "register_operand")
   4296 	     (match_operand:SVE_FULL_SDI 2 "register_operand"))]
   4297 	  UNSPEC_PRED_X))]
   4298   "TARGET_SVE"
   4299   {
   4300     operands[3] = aarch64_ptrue_reg (<VPRED>mode);
   4301   }
   4302 )
   4303 
   4304 ;; Integer division predicated with a PTRUE.
   4305 (define_insn "@aarch64_pred_<optab><mode>"
   4306   [(set (match_operand:SVE_FULL_SDI 0 "register_operand" "=w, w, ?&w")
   4307 	(unspec:SVE_FULL_SDI
   4308 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   4309 	   (SVE_INT_BINARY_SD:SVE_FULL_SDI
   4310 	     (match_operand:SVE_FULL_SDI 2 "register_operand" "0, w, w")
   4311 	     (match_operand:SVE_FULL_SDI 3 "register_operand" "w, 0, w"))]
   4312 	  UNSPEC_PRED_X))]
   4313   "TARGET_SVE"
   4314   "@
   4315    <sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4316    <sve_int_op>r\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   4317    movprfx\t%0, %2\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   4318   [(set_attr "movprfx" "*,*,yes")]
   4319 )
   4320 
   4321 ;; Predicated integer division with merging.
   4322 (define_expand "@cond_<optab><mode>"
   4323   [(set (match_operand:SVE_FULL_SDI 0 "register_operand")
   4324 	(unspec:SVE_FULL_SDI
   4325 	  [(match_operand:<VPRED> 1 "register_operand")
   4326 	   (SVE_INT_BINARY_SD:SVE_FULL_SDI
   4327 	     (match_operand:SVE_FULL_SDI 2 "register_operand")
   4328 	     (match_operand:SVE_FULL_SDI 3 "register_operand"))
   4329 	   (match_operand:SVE_FULL_SDI 4 "aarch64_simd_reg_or_zero")]
   4330 	  UNSPEC_SEL))]
   4331   "TARGET_SVE"
   4332 )
   4333 
   4334 ;; Predicated integer division, merging with the first input.
   4335 (define_insn "*cond_<optab><mode>_2"
   4336   [(set (match_operand:SVE_FULL_SDI 0 "register_operand" "=w, ?&w")
   4337 	(unspec:SVE_FULL_SDI
   4338 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4339 	   (SVE_INT_BINARY_SD:SVE_FULL_SDI
   4340 	     (match_operand:SVE_FULL_SDI 2 "register_operand" "0, w")
   4341 	     (match_operand:SVE_FULL_SDI 3 "register_operand" "w, w"))
   4342 	   (match_dup 2)]
   4343 	  UNSPEC_SEL))]
   4344   "TARGET_SVE"
   4345   "@
   4346    <sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4347    movprfx\t%0, %2\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   4348   [(set_attr "movprfx" "*,yes")]
   4349 )
   4350 
   4351 ;; Predicated integer division, merging with the second input.
   4352 (define_insn "*cond_<optab><mode>_3"
   4353   [(set (match_operand:SVE_FULL_SDI 0 "register_operand" "=w, ?&w")
   4354 	(unspec:SVE_FULL_SDI
   4355 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4356 	   (SVE_INT_BINARY_SD:SVE_FULL_SDI
   4357 	     (match_operand:SVE_FULL_SDI 2 "register_operand" "w, w")
   4358 	     (match_operand:SVE_FULL_SDI 3 "register_operand" "0, w"))
   4359 	   (match_dup 3)]
   4360 	  UNSPEC_SEL))]
   4361   "TARGET_SVE"
   4362   "@
   4363    <sve_int_op_rev>\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   4364    movprfx\t%0, %3\;<sve_int_op_rev>\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>"
   4365   [(set_attr "movprfx" "*,yes")]
   4366 )
   4367 
   4368 ;; Predicated integer division, merging with an independent value.
   4369 (define_insn_and_rewrite "*cond_<optab><mode>_any"
   4370   [(set (match_operand:SVE_FULL_SDI 0 "register_operand" "=&w, &w, &w, &w, ?&w")
   4371 	(unspec:SVE_FULL_SDI
   4372 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl")
   4373 	   (SVE_INT_BINARY_SD:SVE_FULL_SDI
   4374 	     (match_operand:SVE_FULL_SDI 2 "register_operand" "0, w, w, w, w")
   4375 	     (match_operand:SVE_FULL_SDI 3 "register_operand" "w, 0, w, w, w"))
   4376 	   (match_operand:SVE_FULL_SDI 4 "aarch64_simd_reg_or_zero" "Dz, Dz, Dz, 0, w")]
   4377 	  UNSPEC_SEL))]
   4378   "TARGET_SVE
   4379    && !rtx_equal_p (operands[2], operands[4])
   4380    && !rtx_equal_p (operands[3], operands[4])"
   4381   "@
   4382    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4383    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_int_op_rev>\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   4384    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4385    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4386    #"
   4387   "&& reload_completed
   4388    && register_operand (operands[4], <MODE>mode)
   4389    && !rtx_equal_p (operands[0], operands[4])"
   4390   {
   4391     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   4392 					     operands[4], operands[1]));
   4393     operands[4] = operands[2] = operands[0];
   4394   }
   4395   [(set_attr "movprfx" "yes")]
   4396 )
   4397 
   4398 ;; -------------------------------------------------------------------------
   4399 ;; ---- [INT] Binary logical operations
   4400 ;; -------------------------------------------------------------------------
   4401 ;; Includes:
   4402 ;; - AND
   4403 ;; - EOR
   4404 ;; - ORR
   4405 ;; -------------------------------------------------------------------------
   4406 
   4407 ;; Unpredicated integer binary logical operations.
   4408 (define_insn "<optab><mode>3"
   4409   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?w, w")
   4410 	(LOGICAL:SVE_I
   4411 	  (match_operand:SVE_I 1 "register_operand" "%0, w, w")
   4412 	  (match_operand:SVE_I 2 "aarch64_sve_logical_operand" "vsl, vsl, w")))]
   4413   "TARGET_SVE"
   4414   "@
   4415    <logical>\t%0.<Vetype>, %0.<Vetype>, #%C2
   4416    movprfx\t%0, %1\;<logical>\t%0.<Vetype>, %0.<Vetype>, #%C2
   4417    <logical>\t%0.d, %1.d, %2.d"
   4418   [(set_attr "movprfx" "*,yes,*")]
   4419 )
   4420 
   4421 ;; Merging forms are handled through SVE_INT_BINARY.
   4422 
   4423 ;; -------------------------------------------------------------------------
   4424 ;; ---- [INT] Binary logical operations (inverted second input)
   4425 ;; -------------------------------------------------------------------------
   4426 ;; Includes:
   4427 ;; - BIC
   4428 ;; -------------------------------------------------------------------------
   4429 
   4430 ;; Unpredicated BIC.
   4431 (define_expand "@aarch64_bic<mode>"
   4432   [(set (match_operand:SVE_I 0 "register_operand")
   4433 	(and:SVE_I
   4434 	  (unspec:SVE_I
   4435 	    [(match_dup 3)
   4436 	     (not:SVE_I (match_operand:SVE_I 2 "register_operand"))]
   4437 	    UNSPEC_PRED_X)
   4438 	  (match_operand:SVE_I 1 "register_operand")))]
   4439   "TARGET_SVE"
   4440   {
   4441     operands[3] = CONSTM1_RTX (<VPRED>mode);
   4442   }
   4443 )
   4444 
   4445 ;; Predicated BIC.
   4446 (define_insn_and_rewrite "*bic<mode>3"
   4447   [(set (match_operand:SVE_I 0 "register_operand" "=w")
   4448 	(and:SVE_I
   4449 	  (unspec:SVE_I
   4450 	    [(match_operand 3)
   4451 	     (not:SVE_I
   4452 	       (match_operand:SVE_I 2 "register_operand" "w"))]
   4453 	    UNSPEC_PRED_X)
   4454 	  (match_operand:SVE_I 1 "register_operand" "w")))]
   4455   "TARGET_SVE"
   4456   "bic\t%0.d, %1.d, %2.d"
   4457   "&& !CONSTANT_P (operands[3])"
   4458   {
   4459     operands[3] = CONSTM1_RTX (<VPRED>mode);
   4460   }
   4461 )
   4462 
   4463 ;; Predicated BIC with merging.
   4464 (define_expand "@cond_bic<mode>"
   4465   [(set (match_operand:SVE_FULL_I 0 "register_operand")
   4466 	(unspec:SVE_FULL_I
   4467 	  [(match_operand:<VPRED> 1 "register_operand")
   4468 	   (and:SVE_FULL_I
   4469 	     (not:SVE_FULL_I (match_operand:SVE_FULL_I 3 "register_operand"))
   4470 	     (match_operand:SVE_FULL_I 2 "register_operand"))
   4471 	   (match_operand:SVE_FULL_I 4 "aarch64_simd_reg_or_zero")]
   4472 	  UNSPEC_SEL))]
   4473   "TARGET_SVE"
   4474 )
   4475 
   4476 ;; Predicated integer BIC, merging with the first input.
   4477 (define_insn "*cond_bic<mode>_2"
   4478   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   4479 	(unspec:SVE_I
   4480 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4481 	   (and:SVE_I
   4482 	     (not:SVE_I
   4483 	       (match_operand:SVE_I 3 "register_operand" "w, w"))
   4484 	     (match_operand:SVE_I 2 "register_operand" "0, w"))
   4485 	   (match_dup 2)]
   4486 	  UNSPEC_SEL))]
   4487   "TARGET_SVE"
   4488   "@
   4489    bic\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4490    movprfx\t%0, %2\;bic\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   4491   [(set_attr "movprfx" "*,yes")]
   4492 )
   4493 
   4494 ;; Predicated integer BIC, merging with an independent value.
   4495 (define_insn_and_rewrite "*cond_bic<mode>_any"
   4496   [(set (match_operand:SVE_I 0 "register_operand" "=&w, &w, &w, ?&w")
   4497 	(unspec:SVE_I
   4498 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl")
   4499 	   (and:SVE_I
   4500 	     (not:SVE_I
   4501 	       (match_operand:SVE_I 3 "register_operand" "w, w, w, w"))
   4502 	     (match_operand:SVE_I 2 "register_operand" "0, w, w, w"))
   4503 	   (match_operand:SVE_I 4 "aarch64_simd_reg_or_zero" "Dz, Dz, 0, w")]
   4504 	  UNSPEC_SEL))]
   4505   "TARGET_SVE && !rtx_equal_p (operands[2], operands[4])"
   4506   "@
   4507    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;bic\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4508    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;bic\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4509    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;bic\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4510    #"
   4511   "&& reload_completed
   4512    && register_operand (operands[4], <MODE>mode)
   4513    && !rtx_equal_p (operands[0], operands[4])"
   4514   {
   4515     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   4516 					     operands[4], operands[1]));
   4517     operands[4] = operands[2] = operands[0];
   4518   }
   4519   [(set_attr "movprfx" "yes")]
   4520 )
   4521 
   4522 ;; -------------------------------------------------------------------------
   4523 ;; ---- [INT] Shifts (rounding towards -Inf)
   4524 ;; -------------------------------------------------------------------------
   4525 ;; Includes:
   4526 ;; - ASR
   4527 ;; - ASRR
   4528 ;; - LSL
   4529 ;; - LSLR
   4530 ;; - LSR
   4531 ;; - LSRR
   4532 ;; -------------------------------------------------------------------------
   4533 
   4534 ;; Unpredicated shift by a scalar, which expands into one of the vector
   4535 ;; shifts below.
   4536 (define_expand "<ASHIFT:optab><mode>3"
   4537   [(set (match_operand:SVE_I 0 "register_operand")
   4538 	(ASHIFT:SVE_I
   4539 	  (match_operand:SVE_I 1 "register_operand")
   4540 	  (match_operand:<VEL> 2 "general_operand")))]
   4541   "TARGET_SVE"
   4542   {
   4543     rtx amount;
   4544     if (CONST_INT_P (operands[2]))
   4545       {
   4546 	amount = gen_const_vec_duplicate (<MODE>mode, operands[2]);
   4547 	if (!aarch64_sve_<lr>shift_operand (operands[2], <MODE>mode))
   4548 	  amount = force_reg (<MODE>mode, amount);
   4549       }
   4550     else
   4551       {
   4552 	amount = convert_to_mode (<VEL>mode, operands[2], 0);
   4553 	amount = expand_vector_broadcast (<MODE>mode, amount);
   4554       }
   4555     emit_insn (gen_v<optab><mode>3 (operands[0], operands[1], amount));
   4556     DONE;
   4557   }
   4558 )
   4559 
   4560 ;; Unpredicated shift by a vector.
   4561 (define_expand "v<optab><mode>3"
   4562   [(set (match_operand:SVE_I 0 "register_operand")
   4563 	(unspec:SVE_I
   4564 	  [(match_dup 3)
   4565 	   (ASHIFT:SVE_I
   4566 	     (match_operand:SVE_I 1 "register_operand")
   4567 	     (match_operand:SVE_I 2 "aarch64_sve_<lr>shift_operand"))]
   4568 	  UNSPEC_PRED_X))]
   4569   "TARGET_SVE"
   4570   {
   4571     operands[3] = aarch64_ptrue_reg (<VPRED>mode);
   4572   }
   4573 )
   4574 
   4575 ;; Shift by a vector, predicated with a PTRUE.  We don't actually need
   4576 ;; the predicate for the first alternative, but using Upa or X isn't
   4577 ;; likely to gain much and would make the instruction seem less uniform
   4578 ;; to the register allocator.
   4579 (define_insn_and_split "@aarch64_pred_<optab><mode>"
   4580   [(set (match_operand:SVE_I 0 "register_operand" "=w, w, w, ?&w")
   4581 	(unspec:SVE_I
   4582 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl")
   4583 	   (ASHIFT:SVE_I
   4584 	     (match_operand:SVE_I 2 "register_operand" "w, 0, w, w")
   4585 	     (match_operand:SVE_I 3 "aarch64_sve_<lr>shift_operand" "D<lr>, w, 0, w"))]
   4586 	  UNSPEC_PRED_X))]
   4587   "TARGET_SVE"
   4588   "@
   4589    #
   4590    <shift>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4591    <shift>r\t%0.<Vetype>, %1/m, %3.<Vetype>, %2.<Vetype>
   4592    movprfx\t%0, %2\;<shift>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   4593   "&& reload_completed
   4594    && !register_operand (operands[3], <MODE>mode)"
   4595   [(set (match_dup 0) (ASHIFT:SVE_I (match_dup 2) (match_dup 3)))]
   4596   ""
   4597   [(set_attr "movprfx" "*,*,*,yes")]
   4598 )
   4599 
   4600 ;; Unpredicated shift operations by a constant (post-RA only).
   4601 ;; These are generated by splitting a predicated instruction whose
   4602 ;; predicate is unused.
   4603 (define_insn "*post_ra_v<optab><mode>3"
   4604   [(set (match_operand:SVE_I 0 "register_operand" "=w")
   4605 	(ASHIFT:SVE_I
   4606 	  (match_operand:SVE_I 1 "register_operand" "w")
   4607 	  (match_operand:SVE_I 2 "aarch64_simd_<lr>shift_imm")))]
   4608   "TARGET_SVE && reload_completed"
   4609   "<shift>\t%0.<Vetype>, %1.<Vetype>, #%2"
   4610 )
   4611 
   4612 ;; Predicated integer shift, merging with the first input.
   4613 (define_insn "*cond_<optab><mode>_2_const"
   4614   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   4615 	(unspec:SVE_I
   4616 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4617 	   (ASHIFT:SVE_I
   4618 	     (match_operand:SVE_I 2 "register_operand" "0, w")
   4619 	     (match_operand:SVE_I 3 "aarch64_simd_<lr>shift_imm"))
   4620 	   (match_dup 2)]
   4621 	 UNSPEC_SEL))]
   4622   "TARGET_SVE"
   4623   "@
   4624    <shift>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   4625    movprfx\t%0, %2\;<shift>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3"
   4626   [(set_attr "movprfx" "*,yes")]
   4627 )
   4628 
   4629 ;; Predicated integer shift, merging with an independent value.
   4630 (define_insn_and_rewrite "*cond_<optab><mode>_any_const"
   4631   [(set (match_operand:SVE_I 0 "register_operand" "=w, &w, ?&w")
   4632 	(unspec:SVE_I
   4633 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   4634 	   (ASHIFT:SVE_I
   4635 	     (match_operand:SVE_I 2 "register_operand" "w, w, w")
   4636 	     (match_operand:SVE_I 3 "aarch64_simd_<lr>shift_imm"))
   4637 	   (match_operand:SVE_I 4 "aarch64_simd_reg_or_zero" "Dz, 0, w")]
   4638 	 UNSPEC_SEL))]
   4639   "TARGET_SVE && !rtx_equal_p (operands[2], operands[4])"
   4640   "@
   4641    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<shift>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   4642    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;<shift>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   4643    #"
   4644   "&& reload_completed
   4645    && register_operand (operands[4], <MODE>mode)
   4646    && !rtx_equal_p (operands[0], operands[4])"
   4647   {
   4648     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   4649 					     operands[4], operands[1]));
   4650     operands[4] = operands[2] = operands[0];
   4651   }
   4652   [(set_attr "movprfx" "yes")]
   4653 )
   4654 
   4655 ;; Unpredicated shifts of narrow elements by 64-bit amounts.
   4656 (define_insn "@aarch64_sve_<sve_int_op><mode>"
   4657   [(set (match_operand:SVE_FULL_BHSI 0 "register_operand" "=w")
   4658 	(unspec:SVE_FULL_BHSI
   4659 	  [(match_operand:SVE_FULL_BHSI 1 "register_operand" "w")
   4660 	   (match_operand:VNx2DI 2 "register_operand" "w")]
   4661 	  SVE_SHIFT_WIDE))]
   4662   "TARGET_SVE"
   4663   "<sve_int_op>\t%0.<Vetype>, %1.<Vetype>, %2.d"
   4664 )
   4665 
   4666 ;; Merging predicated shifts of narrow elements by 64-bit amounts.
   4667 (define_expand "@cond_<sve_int_op><mode>"
   4668   [(set (match_operand:SVE_FULL_BHSI 0 "register_operand")
   4669 	(unspec:SVE_FULL_BHSI
   4670 	  [(match_operand:<VPRED> 1 "register_operand")
   4671 	   (unspec:SVE_FULL_BHSI
   4672 	     [(match_operand:SVE_FULL_BHSI 2 "register_operand")
   4673 	      (match_operand:VNx2DI 3 "register_operand")]
   4674 	     SVE_SHIFT_WIDE)
   4675 	   (match_operand:SVE_FULL_BHSI 4 "aarch64_simd_reg_or_zero")]
   4676 	  UNSPEC_SEL))]
   4677   "TARGET_SVE"
   4678 )
   4679 
   4680 ;; Predicated shifts of narrow elements by 64-bit amounts, merging with
   4681 ;; the first input.
   4682 (define_insn "*cond_<sve_int_op><mode>_m"
   4683   [(set (match_operand:SVE_FULL_BHSI 0 "register_operand" "=w, ?&w")
   4684 	(unspec:SVE_FULL_BHSI
   4685 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4686 	   (unspec:SVE_FULL_BHSI
   4687 	     [(match_operand:SVE_FULL_BHSI 2 "register_operand" "0, w")
   4688 	      (match_operand:VNx2DI 3 "register_operand" "w, w")]
   4689 	     SVE_SHIFT_WIDE)
   4690 	   (match_dup 2)]
   4691 	 UNSPEC_SEL))]
   4692   "TARGET_SVE"
   4693   "@
   4694    <sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.d
   4695    movprfx\t%0, %2\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.d"
   4696   [(set_attr "movprfx" "*, yes")])
   4697 
   4698 ;; Predicated shifts of narrow elements by 64-bit amounts, merging with zero.
   4699 (define_insn "*cond_<sve_int_op><mode>_z"
   4700   [(set (match_operand:SVE_FULL_BHSI 0 "register_operand" "=&w, &w")
   4701 	(unspec:SVE_FULL_BHSI
   4702 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4703 	   (unspec:SVE_FULL_BHSI
   4704 	     [(match_operand:SVE_FULL_BHSI 2 "register_operand" "0, w")
   4705 	      (match_operand:VNx2DI 3 "register_operand" "w, w")]
   4706 	     SVE_SHIFT_WIDE)
   4707 	   (match_operand:SVE_FULL_BHSI 4 "aarch64_simd_imm_zero")]
   4708 	 UNSPEC_SEL))]
   4709   "TARGET_SVE"
   4710   "@
   4711    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.d
   4712    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.d"
   4713   [(set_attr "movprfx" "yes")])
   4714 
   4715 ;; -------------------------------------------------------------------------
   4716 ;; ---- [INT] Shifts (rounding towards 0)
   4717 ;; -------------------------------------------------------------------------
   4718 ;; Includes:
   4719 ;; - ASRD
   4720 ;; - SQSHLU (SVE2)
   4721 ;; - SRSHR (SVE2)
   4722 ;; - URSHR (SVE2)
   4723 ;; -------------------------------------------------------------------------
   4724 
   4725 ;; Unpredicated ASRD.
   4726 (define_expand "sdiv_pow2<mode>3"
   4727   [(set (match_operand:SVE_I 0 "register_operand")
   4728 	(unspec:SVE_I
   4729 	  [(match_dup 3)
   4730 	   (unspec:SVE_I
   4731 	     [(match_operand:SVE_I 1 "register_operand")
   4732 	      (match_operand 2 "aarch64_simd_rshift_imm")]
   4733 	     UNSPEC_ASRD)]
   4734 	 UNSPEC_PRED_X))]
   4735   "TARGET_SVE"
   4736   {
   4737     operands[3] = aarch64_ptrue_reg (<VPRED>mode);
   4738   }
   4739 )
   4740 
   4741 ;; Predicated ASRD.
   4742 (define_insn "*sdiv_pow2<mode>3"
   4743   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   4744 	(unspec:SVE_I
   4745 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4746 	   (unspec:SVE_I
   4747 	     [(match_operand:SVE_I 2 "register_operand" "0, w")
   4748 	      (match_operand:SVE_I 3 "aarch64_simd_rshift_imm")]
   4749 	     UNSPEC_ASRD)]
   4750 	  UNSPEC_PRED_X))]
   4751   "TARGET_SVE"
   4752   "@
   4753    asrd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   4754    movprfx\t%0, %2\;asrd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3"
   4755   [(set_attr "movprfx" "*,yes")])
   4756 
   4757 ;; Predicated shift with merging.
   4758 (define_expand "@cond_<sve_int_op><mode>"
   4759   [(set (match_operand:SVE_I 0 "register_operand")
   4760 	(unspec:SVE_I
   4761 	  [(match_operand:<VPRED> 1 "register_operand")
   4762 	   (unspec:SVE_I
   4763 	     [(match_dup 5)
   4764 	      (unspec:SVE_I
   4765 		[(match_operand:SVE_I 2 "register_operand")
   4766 		 (match_operand:SVE_I 3 "aarch64_simd_<lr>shift_imm")]
   4767 		SVE_INT_SHIFT_IMM)]
   4768 	     UNSPEC_PRED_X)
   4769 	   (match_operand:SVE_I 4 "aarch64_simd_reg_or_zero")]
   4770 	  UNSPEC_SEL))]
   4771   "TARGET_SVE"
   4772   {
   4773     operands[5] = aarch64_ptrue_reg (<VPRED>mode);
   4774   }
   4775 )
   4776 
   4777 ;; Predicated shift, merging with the first input.
   4778 (define_insn_and_rewrite "*cond_<sve_int_op><mode>_2"
   4779   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   4780 	(unspec:SVE_I
   4781 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4782 	   (unspec:SVE_I
   4783 	     [(match_operand 4)
   4784 	      (unspec:SVE_I
   4785 		[(match_operand:SVE_I 2 "register_operand" "0, w")
   4786 		 (match_operand:SVE_I 3 "aarch64_simd_<lr>shift_imm")]
   4787 		SVE_INT_SHIFT_IMM)]
   4788 	     UNSPEC_PRED_X)
   4789 	   (match_dup 2)]
   4790 	  UNSPEC_SEL))]
   4791   "TARGET_SVE"
   4792   "@
   4793    <sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   4794    movprfx\t%0, %2\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3"
   4795   "&& !CONSTANT_P (operands[4])"
   4796   {
   4797     operands[4] = CONSTM1_RTX (<VPRED>mode);
   4798   }
   4799   [(set_attr "movprfx" "*,yes")])
   4800 
   4801 ;; Predicated shift, merging with an independent value.
   4802 (define_insn_and_rewrite "*cond_<sve_int_op><mode>_any"
   4803   [(set (match_operand:SVE_I 0 "register_operand" "=w, &w, ?&w")
   4804 	(unspec:SVE_I
   4805 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   4806 	   (unspec:SVE_I
   4807 	     [(match_operand 5)
   4808 	      (unspec:SVE_I
   4809 		[(match_operand:SVE_I 2 "register_operand" "w, w, w")
   4810 		 (match_operand:SVE_I 3 "aarch64_simd_<lr>shift_imm")]
   4811 		SVE_INT_SHIFT_IMM)]
   4812 	     UNSPEC_PRED_X)
   4813 	   (match_operand:SVE_I 4 "aarch64_simd_reg_or_zero" "Dz, 0, w")]
   4814 	 UNSPEC_SEL))]
   4815   "TARGET_SVE && !rtx_equal_p (operands[2], operands[4])"
   4816   "@
   4817    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   4818    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;<sve_int_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   4819    #"
   4820   "&& reload_completed
   4821    && register_operand (operands[4], <MODE>mode)
   4822    && !rtx_equal_p (operands[0], operands[4])"
   4823   {
   4824     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   4825 					     operands[4], operands[1]));
   4826     operands[4] = operands[2] = operands[0];
   4827   }
   4828   [(set_attr "movprfx" "yes")]
   4829 )
   4830 
   4831 ;; -------------------------------------------------------------------------
   4832 ;; ---- [FP<-INT] General binary arithmetic corresponding to unspecs
   4833 ;; -------------------------------------------------------------------------
   4834 ;; Includes:
   4835 ;; - FSCALE
   4836 ;; - FTSMUL
   4837 ;; - FTSSEL
   4838 ;; -------------------------------------------------------------------------
   4839 
   4840 ;; Unpredicated floating-point binary operations that take an integer as
   4841 ;; their second operand.
   4842 (define_insn "@aarch64_sve_<optab><mode>"
   4843   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w")
   4844 	(unspec:SVE_FULL_F
   4845 	  [(match_operand:SVE_FULL_F 1 "register_operand" "w")
   4846 	   (match_operand:<V_INT_EQUIV> 2 "register_operand" "w")]
   4847 	  SVE_FP_BINARY_INT))]
   4848   "TARGET_SVE"
   4849   "<sve_fp_op>\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>"
   4850 )
   4851 
   4852 ;; Predicated floating-point binary operations that take an integer
   4853 ;; as their second operand.
   4854 (define_insn "@aarch64_pred_<optab><mode>"
   4855   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   4856 	(unspec:SVE_FULL_F
   4857 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4858 	   (match_operand:SI 4 "aarch64_sve_gp_strictness")
   4859 	   (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   4860 	   (match_operand:<V_INT_EQUIV> 3 "register_operand" "w, w")]
   4861 	  SVE_COND_FP_BINARY_INT))]
   4862   "TARGET_SVE"
   4863   "@
   4864    <sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4865    movprfx\t%0, %2\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   4866   [(set_attr "movprfx" "*,yes")]
   4867 )
   4868 
   4869 ;; Predicated floating-point binary operations with merging, taking an
   4870 ;; integer as their second operand.
   4871 (define_expand "@cond_<optab><mode>"
   4872   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   4873 	(unspec:SVE_FULL_F
   4874 	  [(match_operand:<VPRED> 1 "register_operand")
   4875 	   (unspec:SVE_FULL_F
   4876 	     [(match_dup 1)
   4877 	      (const_int SVE_STRICT_GP)
   4878 	      (match_operand:SVE_FULL_F 2 "register_operand")
   4879 	      (match_operand:<V_INT_EQUIV> 3 "register_operand")]
   4880 	     SVE_COND_FP_BINARY_INT)
   4881 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero")]
   4882 	  UNSPEC_SEL))]
   4883   "TARGET_SVE"
   4884 )
   4885 
   4886 ;; Predicated floating-point binary operations that take an integer as their
   4887 ;; second operand, with inactive lanes coming from the first operand.
   4888 (define_insn_and_rewrite "*cond_<optab><mode>_2_relaxed"
   4889   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   4890 	(unspec:SVE_FULL_F
   4891 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4892 	   (unspec:SVE_FULL_F
   4893 	     [(match_operand 4)
   4894 	      (const_int SVE_RELAXED_GP)
   4895 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   4896 	      (match_operand:<V_INT_EQUIV> 3 "register_operand" "w, w")]
   4897 	     SVE_COND_FP_BINARY_INT)
   4898 	   (match_dup 2)]
   4899 	  UNSPEC_SEL))]
   4900   "TARGET_SVE"
   4901   "@
   4902    <sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4903    movprfx\t%0, %2\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   4904   "&& !rtx_equal_p (operands[1], operands[4])"
   4905   {
   4906     operands[4] = copy_rtx (operands[1]);
   4907   }
   4908   [(set_attr "movprfx" "*,yes")]
   4909 )
   4910 
   4911 (define_insn "*cond_<optab><mode>_2_strict"
   4912   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   4913 	(unspec:SVE_FULL_F
   4914 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   4915 	   (unspec:SVE_FULL_F
   4916 	     [(match_dup 1)
   4917 	      (const_int SVE_STRICT_GP)
   4918 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   4919 	      (match_operand:<V_INT_EQUIV> 3 "register_operand" "w, w")]
   4920 	     SVE_COND_FP_BINARY_INT)
   4921 	   (match_dup 2)]
   4922 	  UNSPEC_SEL))]
   4923   "TARGET_SVE"
   4924   "@
   4925    <sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4926    movprfx\t%0, %2\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   4927   [(set_attr "movprfx" "*,yes")]
   4928 )
   4929 
   4930 ;; Predicated floating-point binary operations that take an integer as
   4931 ;; their second operand, with the values of inactive lanes being distinct
   4932 ;; from the other inputs.
   4933 (define_insn_and_rewrite "*cond_<optab><mode>_any_relaxed"
   4934   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, &w, ?&w")
   4935 	(unspec:SVE_FULL_F
   4936 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl")
   4937 	   (unspec:SVE_FULL_F
   4938 	     [(match_operand 5)
   4939 	      (const_int SVE_RELAXED_GP)
   4940 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w, w, w")
   4941 	      (match_operand:<V_INT_EQUIV> 3 "register_operand" "w, w, w, w")]
   4942 	     SVE_COND_FP_BINARY_INT)
   4943 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, Dz, 0, w")]
   4944 	  UNSPEC_SEL))]
   4945   "TARGET_SVE && !rtx_equal_p (operands[2], operands[4])"
   4946   "@
   4947    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4948    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4949    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4950    #"
   4951   "&& 1"
   4952   {
   4953     if (reload_completed
   4954         && register_operand (operands[4], <MODE>mode)
   4955         && !rtx_equal_p (operands[0], operands[4]))
   4956       {
   4957 	emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   4958 						 operands[4], operands[1]));
   4959 	operands[4] = operands[2] = operands[0];
   4960       }
   4961     else if (!rtx_equal_p (operands[1], operands[5]))
   4962       operands[5] = copy_rtx (operands[1]);
   4963     else
   4964       FAIL;
   4965   }
   4966   [(set_attr "movprfx" "yes")]
   4967 )
   4968 
   4969 (define_insn_and_rewrite "*cond_<optab><mode>_any_strict"
   4970   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, &w, ?&w")
   4971 	(unspec:SVE_FULL_F
   4972 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl")
   4973 	   (unspec:SVE_FULL_F
   4974 	     [(match_dup 1)
   4975 	      (const_int SVE_STRICT_GP)
   4976 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w, w, w")
   4977 	      (match_operand:<V_INT_EQUIV> 3 "register_operand" "w, w, w, w")]
   4978 	     SVE_COND_FP_BINARY_INT)
   4979 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, Dz, 0, w")]
   4980 	  UNSPEC_SEL))]
   4981   "TARGET_SVE && !rtx_equal_p (operands[2], operands[4])"
   4982   "@
   4983    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4984    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4985    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   4986    #"
   4987   "&& reload_completed
   4988    && register_operand (operands[4], <MODE>mode)
   4989    && !rtx_equal_p (operands[0], operands[4])"
   4990   {
   4991     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   4992 					     operands[4], operands[1]));
   4993     operands[4] = operands[2] = operands[0];
   4994   }
   4995   [(set_attr "movprfx" "yes")]
   4996 )
   4997 
   4998 ;; -------------------------------------------------------------------------
   4999 ;; ---- [FP] General binary arithmetic corresponding to rtx codes
   5000 ;; -------------------------------------------------------------------------
   5001 ;; Includes post-RA forms of:
   5002 ;; - FADD
   5003 ;; - FMUL
   5004 ;; - FSUB
   5005 ;; -------------------------------------------------------------------------
   5006 
   5007 ;; Unpredicated floating-point binary operations (post-RA only).
   5008 ;; These are generated by splitting a predicated instruction whose
   5009 ;; predicate is unused.
   5010 (define_insn "*post_ra_<sve_fp_op><mode>3"
   5011   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w")
   5012 	(SVE_UNPRED_FP_BINARY:SVE_FULL_F
   5013 	  (match_operand:SVE_FULL_F 1 "register_operand" "w")
   5014 	  (match_operand:SVE_FULL_F 2 "register_operand" "w")))]
   5015   "TARGET_SVE && reload_completed"
   5016   "<sve_fp_op>\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>")
   5017 
   5018 ;; -------------------------------------------------------------------------
   5019 ;; ---- [FP] General binary arithmetic corresponding to unspecs
   5020 ;; -------------------------------------------------------------------------
   5021 ;; Includes merging forms of:
   5022 ;; - FADD    (constant forms handled in the "Addition" section)
   5023 ;; - FDIV
   5024 ;; - FDIVR
   5025 ;; - FMAX
   5026 ;; - FMAXNM  (including #0.0 and #1.0)
   5027 ;; - FMIN
   5028 ;; - FMINNM  (including #0.0 and #1.0)
   5029 ;; - FMUL    (including #0.5 and #2.0)
   5030 ;; - FMULX
   5031 ;; - FRECPS
   5032 ;; - FRSQRTS
   5033 ;; - FSUB    (constant forms handled in the "Addition" section)
   5034 ;; - FSUBR   (constant forms handled in the "Subtraction" section)
   5035 ;; -------------------------------------------------------------------------
   5036 
   5037 ;; Unpredicated floating-point binary operations.
   5038 (define_insn "@aarch64_sve_<optab><mode>"
   5039   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w")
   5040 	(unspec:SVE_FULL_F
   5041 	  [(match_operand:SVE_FULL_F 1 "register_operand" "w")
   5042 	   (match_operand:SVE_FULL_F 2 "register_operand" "w")]
   5043 	  SVE_FP_BINARY))]
   5044   "TARGET_SVE"
   5045   "<sve_fp_op>\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>"
   5046 )
   5047 
   5048 ;; Unpredicated floating-point binary operations that need to be predicated
   5049 ;; for SVE.
   5050 (define_expand "<optab><mode>3"
   5051   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   5052 	(unspec:SVE_FULL_F
   5053 	  [(match_dup 3)
   5054 	   (const_int SVE_RELAXED_GP)
   5055 	   (match_operand:SVE_FULL_F 1 "<sve_pred_fp_rhs1_operand>")
   5056 	   (match_operand:SVE_FULL_F 2 "<sve_pred_fp_rhs2_operand>")]
   5057 	  SVE_COND_FP_BINARY_OPTAB))]
   5058   "TARGET_SVE"
   5059   {
   5060     operands[3] = aarch64_ptrue_reg (<VPRED>mode);
   5061   }
   5062 )
   5063 
   5064 ;; Predicated floating-point binary operations that have no immediate forms.
   5065 (define_insn "@aarch64_pred_<optab><mode>"
   5066   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, ?&w")
   5067 	(unspec:SVE_FULL_F
   5068 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   5069 	   (match_operand:SI 4 "aarch64_sve_gp_strictness")
   5070 	   (match_operand:SVE_FULL_F 2 "register_operand" "0, w, w")
   5071 	   (match_operand:SVE_FULL_F 3 "register_operand" "w, 0, w")]
   5072 	  SVE_COND_FP_BINARY_REG))]
   5073   "TARGET_SVE"
   5074   "@
   5075    <sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5076    <sve_fp_op_rev>\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   5077    movprfx\t%0, %2\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   5078   [(set_attr "movprfx" "*,*,yes")]
   5079 )
   5080 
   5081 ;; Predicated floating-point operations with merging.
   5082 (define_expand "@cond_<optab><mode>"
   5083   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   5084 	(unspec:SVE_FULL_F
   5085 	  [(match_operand:<VPRED> 1 "register_operand")
   5086 	   (unspec:SVE_FULL_F
   5087 	     [(match_dup 1)
   5088 	      (const_int SVE_STRICT_GP)
   5089 	      (match_operand:SVE_FULL_F 2 "<sve_pred_fp_rhs1_operand>")
   5090 	      (match_operand:SVE_FULL_F 3 "<sve_pred_fp_rhs2_operand>")]
   5091 	     SVE_COND_FP_BINARY)
   5092 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero")]
   5093 	  UNSPEC_SEL))]
   5094   "TARGET_SVE"
   5095 )
   5096 
   5097 ;; Predicated floating-point operations, merging with the first input.
   5098 (define_insn_and_rewrite "*cond_<optab><mode>_2_relaxed"
   5099   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   5100 	(unspec:SVE_FULL_F
   5101 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5102 	   (unspec:SVE_FULL_F
   5103 	     [(match_operand 4)
   5104 	      (const_int SVE_RELAXED_GP)
   5105 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   5106 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w")]
   5107 	     SVE_COND_FP_BINARY)
   5108 	   (match_dup 2)]
   5109 	  UNSPEC_SEL))]
   5110   "TARGET_SVE"
   5111   "@
   5112    <sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5113    movprfx\t%0, %2\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   5114   "&& !rtx_equal_p (operands[1], operands[4])"
   5115   {
   5116     operands[4] = copy_rtx (operands[1]);
   5117   }
   5118   [(set_attr "movprfx" "*,yes")]
   5119 )
   5120 
   5121 (define_insn "*cond_<optab><mode>_2_strict"
   5122   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   5123 	(unspec:SVE_FULL_F
   5124 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5125 	   (unspec:SVE_FULL_F
   5126 	     [(match_dup 1)
   5127 	      (const_int SVE_STRICT_GP)
   5128 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   5129 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w")]
   5130 	     SVE_COND_FP_BINARY)
   5131 	   (match_dup 2)]
   5132 	  UNSPEC_SEL))]
   5133   "TARGET_SVE"
   5134   "@
   5135    <sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5136    movprfx\t%0, %2\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   5137   [(set_attr "movprfx" "*,yes")]
   5138 )
   5139 
   5140 ;; Same for operations that take a 1-bit constant.
   5141 (define_insn_and_rewrite "*cond_<optab><mode>_2_const_relaxed"
   5142   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?w")
   5143 	(unspec:SVE_FULL_F
   5144 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5145 	   (unspec:SVE_FULL_F
   5146 	     [(match_operand 4)
   5147 	      (const_int SVE_RELAXED_GP)
   5148 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   5149 	      (match_operand:SVE_FULL_F 3 "<sve_pred_fp_rhs2_immediate>")]
   5150 	     SVE_COND_FP_BINARY_I1)
   5151 	   (match_dup 2)]
   5152 	  UNSPEC_SEL))]
   5153   "TARGET_SVE"
   5154   "@
   5155    <sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5156    movprfx\t%0, %2\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3"
   5157   "&& !rtx_equal_p (operands[1], operands[4])"
   5158   {
   5159     operands[4] = copy_rtx (operands[1]);
   5160   }
   5161   [(set_attr "movprfx" "*,yes")]
   5162 )
   5163 
   5164 (define_insn "*cond_<optab><mode>_2_const_strict"
   5165   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?w")
   5166 	(unspec:SVE_FULL_F
   5167 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5168 	   (unspec:SVE_FULL_F
   5169 	     [(match_dup 1)
   5170 	      (const_int SVE_STRICT_GP)
   5171 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   5172 	      (match_operand:SVE_FULL_F 3 "<sve_pred_fp_rhs2_immediate>")]
   5173 	     SVE_COND_FP_BINARY_I1)
   5174 	   (match_dup 2)]
   5175 	  UNSPEC_SEL))]
   5176   "TARGET_SVE"
   5177   "@
   5178    <sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5179    movprfx\t%0, %2\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3"
   5180   [(set_attr "movprfx" "*,yes")]
   5181 )
   5182 
   5183 ;; Predicated floating-point operations, merging with the second input.
   5184 (define_insn_and_rewrite "*cond_<optab><mode>_3_relaxed"
   5185   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   5186 	(unspec:SVE_FULL_F
   5187 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5188 	   (unspec:SVE_FULL_F
   5189 	     [(match_operand 4)
   5190 	      (const_int SVE_RELAXED_GP)
   5191 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w")
   5192 	      (match_operand:SVE_FULL_F 3 "register_operand" "0, w")]
   5193 	     SVE_COND_FP_BINARY)
   5194 	   (match_dup 3)]
   5195 	  UNSPEC_SEL))]
   5196   "TARGET_SVE"
   5197   "@
   5198    <sve_fp_op_rev>\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   5199    movprfx\t%0, %3\;<sve_fp_op_rev>\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>"
   5200   "&& !rtx_equal_p (operands[1], operands[4])"
   5201   {
   5202     operands[4] = copy_rtx (operands[1]);
   5203   }
   5204   [(set_attr "movprfx" "*,yes")]
   5205 )
   5206 
   5207 (define_insn "*cond_<optab><mode>_3_strict"
   5208   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   5209 	(unspec:SVE_FULL_F
   5210 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5211 	   (unspec:SVE_FULL_F
   5212 	     [(match_dup 1)
   5213 	      (const_int SVE_STRICT_GP)
   5214 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w")
   5215 	      (match_operand:SVE_FULL_F 3 "register_operand" "0, w")]
   5216 	     SVE_COND_FP_BINARY)
   5217 	   (match_dup 3)]
   5218 	  UNSPEC_SEL))]
   5219   "TARGET_SVE"
   5220   "@
   5221    <sve_fp_op_rev>\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   5222    movprfx\t%0, %3\;<sve_fp_op_rev>\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>"
   5223   [(set_attr "movprfx" "*,yes")]
   5224 )
   5225 
   5226 ;; Predicated floating-point operations, merging with an independent value.
   5227 (define_insn_and_rewrite "*cond_<optab><mode>_any_relaxed"
   5228   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, &w, &w, ?&w")
   5229 	(unspec:SVE_FULL_F
   5230 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl")
   5231 	   (unspec:SVE_FULL_F
   5232 	     [(match_operand 5)
   5233 	      (const_int SVE_RELAXED_GP)
   5234 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w, w, w, w")
   5235 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, 0, w, w, w")]
   5236 	     SVE_COND_FP_BINARY)
   5237 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, Dz, Dz, 0, w")]
   5238 	  UNSPEC_SEL))]
   5239   "TARGET_SVE
   5240    && !rtx_equal_p (operands[2], operands[4])
   5241    && !rtx_equal_p (operands[3], operands[4])"
   5242   "@
   5243    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5244    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_fp_op_rev>\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   5245    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5246    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5247    #"
   5248   "&& 1"
   5249   {
   5250     if (reload_completed
   5251         && register_operand (operands[4], <MODE>mode)
   5252         && !rtx_equal_p (operands[0], operands[4]))
   5253       {
   5254 	emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   5255 						 operands[4], operands[1]));
   5256 	operands[4] = operands[2] = operands[0];
   5257       }
   5258     else if (!rtx_equal_p (operands[1], operands[5]))
   5259       operands[5] = copy_rtx (operands[1]);
   5260     else
   5261       FAIL;
   5262   }
   5263   [(set_attr "movprfx" "yes")]
   5264 )
   5265 
   5266 (define_insn_and_rewrite "*cond_<optab><mode>_any_strict"
   5267   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, &w, &w, ?&w")
   5268 	(unspec:SVE_FULL_F
   5269 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl")
   5270 	   (unspec:SVE_FULL_F
   5271 	     [(match_dup 1)
   5272 	      (const_int SVE_STRICT_GP)
   5273 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w, w, w, w")
   5274 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, 0, w, w, w")]
   5275 	     SVE_COND_FP_BINARY)
   5276 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, Dz, Dz, 0, w")]
   5277 	  UNSPEC_SEL))]
   5278   "TARGET_SVE
   5279    && !rtx_equal_p (operands[2], operands[4])
   5280    && !rtx_equal_p (operands[3], operands[4])"
   5281   "@
   5282    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5283    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_fp_op_rev>\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   5284    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5285    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5286    #"
   5287   "&& reload_completed
   5288    && register_operand (operands[4], <MODE>mode)
   5289    && !rtx_equal_p (operands[0], operands[4])"
   5290   {
   5291     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   5292 					     operands[4], operands[1]));
   5293     operands[4] = operands[2] = operands[0];
   5294   }
   5295   [(set_attr "movprfx" "yes")]
   5296 )
   5297 
   5298 ;; Same for operations that take a 1-bit constant.
   5299 (define_insn_and_rewrite "*cond_<optab><mode>_any_const_relaxed"
   5300   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, ?w")
   5301 	(unspec:SVE_FULL_F
   5302 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   5303 	   (unspec:SVE_FULL_F
   5304 	     [(match_operand 5)
   5305 	      (const_int SVE_RELAXED_GP)
   5306 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w, w")
   5307 	      (match_operand:SVE_FULL_F 3 "<sve_pred_fp_rhs2_immediate>")]
   5308 	     SVE_COND_FP_BINARY_I1)
   5309 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, 0, w")]
   5310 	  UNSPEC_SEL))]
   5311   "TARGET_SVE && !rtx_equal_p (operands[2], operands[4])"
   5312   "@
   5313    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5314    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5315    #"
   5316   "&& 1"
   5317   {
   5318     if (reload_completed
   5319         && register_operand (operands[4], <MODE>mode)
   5320         && !rtx_equal_p (operands[0], operands[4]))
   5321       {
   5322 	emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   5323 						 operands[4], operands[1]));
   5324 	operands[4] = operands[2] = operands[0];
   5325       }
   5326     else if (!rtx_equal_p (operands[1], operands[5]))
   5327       operands[5] = copy_rtx (operands[1]);
   5328     else
   5329       FAIL;
   5330   }
   5331   [(set_attr "movprfx" "yes")]
   5332 )
   5333 
   5334 (define_insn_and_rewrite "*cond_<optab><mode>_any_const_strict"
   5335   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, ?w")
   5336 	(unspec:SVE_FULL_F
   5337 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   5338 	   (unspec:SVE_FULL_F
   5339 	     [(match_dup 1)
   5340 	      (const_int SVE_STRICT_GP)
   5341 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w, w")
   5342 	      (match_operand:SVE_FULL_F 3 "<sve_pred_fp_rhs2_immediate>")]
   5343 	     SVE_COND_FP_BINARY_I1)
   5344 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, 0, w")]
   5345 	  UNSPEC_SEL))]
   5346   "TARGET_SVE && !rtx_equal_p (operands[2], operands[4])"
   5347   "@
   5348    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5349    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5350    #"
   5351   "&& reload_completed
   5352    && register_operand (operands[4], <MODE>mode)
   5353    && !rtx_equal_p (operands[0], operands[4])"
   5354   {
   5355     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   5356 					     operands[4], operands[1]));
   5357     operands[4] = operands[2] = operands[0];
   5358   }
   5359   [(set_attr "movprfx" "yes")]
   5360 )
   5361 
   5362 ;; -------------------------------------------------------------------------
   5363 ;; ---- [FP] Addition
   5364 ;; -------------------------------------------------------------------------
   5365 ;; Includes:
   5366 ;; - FADD
   5367 ;; - FSUB
   5368 ;; -------------------------------------------------------------------------
   5369 
   5370 ;; Predicated floating-point addition.
   5371 (define_insn_and_split "@aarch64_pred_<optab><mode>"
   5372   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, w, w, ?&w, ?&w, ?&w")
   5373 	(unspec:SVE_FULL_F
   5374 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl, Upl")
   5375 	   (match_operand:SI 4 "aarch64_sve_gp_strictness" "i, i, Z, Ui1, i, i, Ui1")
   5376 	   (match_operand:SVE_FULL_F 2 "register_operand" "%0, 0, w, 0, w, w, w")
   5377 	   (match_operand:SVE_FULL_F 3 "aarch64_sve_float_arith_with_sub_operand" "vsA, vsN, w, w, vsA, vsN, w")]
   5378 	  SVE_COND_FP_ADD))]
   5379   "TARGET_SVE"
   5380   "@
   5381    fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5382    fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, #%N3
   5383    #
   5384    fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5385    movprfx\t%0, %2\;fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5386    movprfx\t%0, %2\;fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, #%N3
   5387    movprfx\t%0, %2\;fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   5388   ; Split the unpredicated form after reload, so that we don't have
   5389   ; the unnecessary PTRUE.
   5390   "&& reload_completed
   5391    && register_operand (operands[3], <MODE>mode)
   5392    && INTVAL (operands[4]) == SVE_RELAXED_GP"
   5393   [(set (match_dup 0) (plus:SVE_FULL_F (match_dup 2) (match_dup 3)))]
   5394   ""
   5395   [(set_attr "movprfx" "*,*,*,*,yes,yes,yes")]
   5396 )
   5397 
   5398 ;; Predicated floating-point addition of a constant, merging with the
   5399 ;; first input.
   5400 (define_insn_and_rewrite "*cond_add<mode>_2_const_relaxed"
   5401   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, ?w, ?w")
   5402 	(unspec:SVE_FULL_F
   5403 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl")
   5404 	   (unspec:SVE_FULL_F
   5405 	     [(match_operand 4)
   5406 	      (const_int SVE_RELAXED_GP)
   5407 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, 0, w, w")
   5408 	      (match_operand:SVE_FULL_F 3 "aarch64_sve_float_arith_with_sub_immediate" "vsA, vsN, vsA, vsN")]
   5409 	     UNSPEC_COND_FADD)
   5410 	   (match_dup 2)]
   5411 	  UNSPEC_SEL))]
   5412   "TARGET_SVE"
   5413   "@
   5414    fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5415    fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, #%N3
   5416    movprfx\t%0, %2\;fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5417    movprfx\t%0, %2\;fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, #%N3"
   5418   "&& !rtx_equal_p (operands[1], operands[4])"
   5419   {
   5420     operands[4] = copy_rtx (operands[1]);
   5421   }
   5422   [(set_attr "movprfx" "*,*,yes,yes")]
   5423 )
   5424 
   5425 (define_insn "*cond_add<mode>_2_const_strict"
   5426   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, ?w, ?w")
   5427 	(unspec:SVE_FULL_F
   5428 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl")
   5429 	   (unspec:SVE_FULL_F
   5430 	     [(match_dup 1)
   5431 	      (const_int SVE_STRICT_GP)
   5432 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, 0, w, w")
   5433 	      (match_operand:SVE_FULL_F 3 "aarch64_sve_float_arith_with_sub_immediate" "vsA, vsN, vsA, vsN")]
   5434 	     UNSPEC_COND_FADD)
   5435 	   (match_dup 2)]
   5436 	  UNSPEC_SEL))]
   5437   "TARGET_SVE"
   5438   "@
   5439    fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5440    fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, #%N3
   5441    movprfx\t%0, %2\;fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5442    movprfx\t%0, %2\;fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, #%N3"
   5443   [(set_attr "movprfx" "*,*,yes,yes")]
   5444 )
   5445 
   5446 ;; Predicated floating-point addition of a constant, merging with an
   5447 ;; independent value.
   5448 (define_insn_and_rewrite "*cond_add<mode>_any_const_relaxed"
   5449   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, w, w, ?w, ?w")
   5450 	(unspec:SVE_FULL_F
   5451 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   5452 	   (unspec:SVE_FULL_F
   5453 	     [(match_operand 5)
   5454 	      (const_int SVE_RELAXED_GP)
   5455 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w, w, w, w, w")
   5456 	      (match_operand:SVE_FULL_F 3 "aarch64_sve_float_arith_with_sub_immediate" "vsA, vsN, vsA, vsN, vsA, vsN")]
   5457 	     UNSPEC_COND_FADD)
   5458 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, Dz, 0, 0, w, w")]
   5459 	  UNSPEC_SEL))]
   5460   "TARGET_SVE && !rtx_equal_p (operands[2], operands[4])"
   5461   "@
   5462    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5463    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, #%N3
   5464    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5465    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, #%N3
   5466    #
   5467    #"
   5468   "&& 1"
   5469   {
   5470     if (reload_completed
   5471         && register_operand (operands[4], <MODE>mode)
   5472         && !rtx_equal_p (operands[0], operands[4]))
   5473       {
   5474 	emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   5475 						 operands[4], operands[1]));
   5476 	operands[4] = operands[2] = operands[0];
   5477       }
   5478     else if (!rtx_equal_p (operands[1], operands[5]))
   5479       operands[5] = copy_rtx (operands[1]);
   5480     else
   5481       FAIL;
   5482   }
   5483   [(set_attr "movprfx" "yes")]
   5484 )
   5485 
   5486 (define_insn_and_rewrite "*cond_add<mode>_any_const_strict"
   5487   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, w, w, ?w, ?w")
   5488 	(unspec:SVE_FULL_F
   5489 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   5490 	   (unspec:SVE_FULL_F
   5491 	     [(match_dup 1)
   5492 	      (const_int SVE_STRICT_GP)
   5493 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w, w, w, w, w")
   5494 	      (match_operand:SVE_FULL_F 3 "aarch64_sve_float_arith_with_sub_immediate" "vsA, vsN, vsA, vsN, vsA, vsN")]
   5495 	     UNSPEC_COND_FADD)
   5496 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, Dz, 0, 0, w, w")]
   5497 	  UNSPEC_SEL))]
   5498   "TARGET_SVE && !rtx_equal_p (operands[2], operands[4])"
   5499   "@
   5500    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5501    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, #%N3
   5502    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;fadd\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   5503    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, #%N3
   5504    #
   5505    #"
   5506   "&& reload_completed
   5507    && register_operand (operands[4], <MODE>mode)
   5508    && !rtx_equal_p (operands[0], operands[4])"
   5509   {
   5510     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   5511 					     operands[4], operands[1]));
   5512     operands[4] = operands[2] = operands[0];
   5513   }
   5514   [(set_attr "movprfx" "yes")]
   5515 )
   5516 
   5517 ;; Register merging forms are handled through SVE_COND_FP_BINARY.
   5518 
   5519 ;; -------------------------------------------------------------------------
   5520 ;; ---- [FP] Complex addition
   5521 ;; -------------------------------------------------------------------------
   5522 ;; Includes:
   5523 ;; - FCADD
   5524 ;; -------------------------------------------------------------------------
   5525 
   5526 ;; Predicated FCADD.
   5527 (define_insn "@aarch64_pred_<optab><mode>"
   5528   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   5529 	(unspec:SVE_FULL_F
   5530 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5531 	   (match_operand:SI 4 "aarch64_sve_gp_strictness")
   5532 	   (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   5533 	   (match_operand:SVE_FULL_F 3 "register_operand" "w, w")]
   5534 	  SVE_COND_FCADD))]
   5535   "TARGET_SVE"
   5536   "@
   5537    fcadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>, #<rot>
   5538    movprfx\t%0, %2\;fcadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>, #<rot>"
   5539   [(set_attr "movprfx" "*,yes")]
   5540 )
   5541 
   5542 ;; Predicated FCADD with merging.
   5543 (define_expand "@cond_<optab><mode>"
   5544   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   5545 	(unspec:SVE_FULL_F
   5546 	  [(match_operand:<VPRED> 1 "register_operand")
   5547 	   (unspec:SVE_FULL_F
   5548 	     [(match_dup 1)
   5549 	      (const_int SVE_STRICT_GP)
   5550 	      (match_operand:SVE_FULL_F 2 "register_operand")
   5551 	      (match_operand:SVE_FULL_F 3 "register_operand")]
   5552 	     SVE_COND_FCADD)
   5553 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero")]
   5554 	  UNSPEC_SEL))]
   5555   "TARGET_SVE"
   5556 )
   5557 
   5558 ;; Predicated FCADD using ptrue for unpredicated optab for auto-vectorizer
   5559 (define_expand "@cadd<rot><mode>3"
   5560   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   5561 	(unspec:SVE_FULL_F
   5562 	  [(match_dup 3)
   5563 	   (const_int SVE_RELAXED_GP)
   5564 	   (match_operand:SVE_FULL_F 1 "register_operand")
   5565 	   (match_operand:SVE_FULL_F 2 "register_operand")]
   5566 	  SVE_COND_FCADD))]
   5567   "TARGET_SVE"
   5568 {
   5569   operands[3] = aarch64_ptrue_reg (<VPRED>mode);
   5570 })
   5571 
   5572 ;; Predicated FCADD, merging with the first input.
   5573 (define_insn_and_rewrite "*cond_<optab><mode>_2_relaxed"
   5574   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   5575 	(unspec:SVE_FULL_F
   5576 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5577 	   (unspec:SVE_FULL_F
   5578 	     [(match_operand 4)
   5579 	      (const_int SVE_RELAXED_GP)
   5580 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   5581 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w")]
   5582 	     SVE_COND_FCADD)
   5583 	   (match_dup 2)]
   5584 	  UNSPEC_SEL))]
   5585   "TARGET_SVE"
   5586   "@
   5587    fcadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>, #<rot>
   5588    movprfx\t%0, %2\;fcadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>, #<rot>"
   5589   "&& !rtx_equal_p (operands[1], operands[4])"
   5590   {
   5591     operands[4] = copy_rtx (operands[1]);
   5592   }
   5593   [(set_attr "movprfx" "*,yes")]
   5594 )
   5595 
   5596 (define_insn "*cond_<optab><mode>_2_strict"
   5597   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   5598 	(unspec:SVE_FULL_F
   5599 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5600 	   (unspec:SVE_FULL_F
   5601 	     [(match_dup 1)
   5602 	      (const_int SVE_STRICT_GP)
   5603 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   5604 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w")]
   5605 	     SVE_COND_FCADD)
   5606 	   (match_dup 2)]
   5607 	  UNSPEC_SEL))]
   5608   "TARGET_SVE"
   5609   "@
   5610    fcadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>, #<rot>
   5611    movprfx\t%0, %2\;fcadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>, #<rot>"
   5612   [(set_attr "movprfx" "*,yes")]
   5613 )
   5614 
   5615 ;; Predicated FCADD, merging with an independent value.
   5616 (define_insn_and_rewrite "*cond_<optab><mode>_any_relaxed"
   5617   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, &w, ?&w")
   5618 	(unspec:SVE_FULL_F
   5619 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl")
   5620 	   (unspec:SVE_FULL_F
   5621 	     [(match_operand 5)
   5622 	      (const_int SVE_RELAXED_GP)
   5623 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, 0, w, w")
   5624 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w, w, w")]
   5625 	     SVE_COND_FCADD)
   5626 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, Dz, 0, w")]
   5627 	  UNSPEC_SEL))]
   5628   "TARGET_SVE && !rtx_equal_p (operands[2], operands[4])"
   5629   "@
   5630    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;fcadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>, #<rot>
   5631    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;fcadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>, #<rot>
   5632    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;fcadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>, #<rot>
   5633    #"
   5634   "&& 1"
   5635   {
   5636     if (reload_completed
   5637         && register_operand (operands[4], <MODE>mode)
   5638         && !rtx_equal_p (operands[0], operands[4]))
   5639       {
   5640 	emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   5641 						 operands[4], operands[1]));
   5642 	operands[4] = operands[2] = operands[0];
   5643       }
   5644     else if (!rtx_equal_p (operands[1], operands[5]))
   5645       operands[5] = copy_rtx (operands[1]);
   5646     else
   5647       FAIL;
   5648   }
   5649   [(set_attr "movprfx" "yes")]
   5650 )
   5651 
   5652 (define_insn_and_rewrite "*cond_<optab><mode>_any_strict"
   5653   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, &w, ?&w")
   5654 	(unspec:SVE_FULL_F
   5655 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl")
   5656 	   (unspec:SVE_FULL_F
   5657 	     [(match_dup 1)
   5658 	      (const_int SVE_STRICT_GP)
   5659 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, 0, w, w")
   5660 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w, w, w")]
   5661 	     SVE_COND_FCADD)
   5662 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, Dz, 0, w")]
   5663 	  UNSPEC_SEL))]
   5664   "TARGET_SVE && !rtx_equal_p (operands[2], operands[4])"
   5665   "@
   5666    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;fcadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>, #<rot>
   5667    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;fcadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>, #<rot>
   5668    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;fcadd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>, #<rot>
   5669    #"
   5670   "&& reload_completed
   5671    && register_operand (operands[4], <MODE>mode)
   5672    && !rtx_equal_p (operands[0], operands[4])"
   5673   {
   5674     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[2],
   5675 					     operands[4], operands[1]));
   5676     operands[4] = operands[2] = operands[0];
   5677   }
   5678   [(set_attr "movprfx" "yes")]
   5679 )
   5680 
   5681 ;; -------------------------------------------------------------------------
   5682 ;; ---- [FP] Subtraction
   5683 ;; -------------------------------------------------------------------------
   5684 ;; Includes:
   5685 ;; - FSUB
   5686 ;; - FSUBR
   5687 ;; -------------------------------------------------------------------------
   5688 
   5689 ;; Predicated floating-point subtraction.
   5690 (define_insn_and_split "@aarch64_pred_<optab><mode>"
   5691   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, w, w, ?&w, ?&w")
   5692 	(unspec:SVE_FULL_F
   5693 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   5694 	   (match_operand:SI 4 "aarch64_sve_gp_strictness" "i, Z, Ui1, Ui1, i, Ui1")
   5695 	   (match_operand:SVE_FULL_F 2 "aarch64_sve_float_arith_operand" "vsA, w, 0, w, vsA, w")
   5696 	   (match_operand:SVE_FULL_F 3 "register_operand" "0, w, w, 0, w, w")]
   5697 	  SVE_COND_FP_SUB))]
   5698   "TARGET_SVE"
   5699   "@
   5700    fsubr\t%0.<Vetype>, %1/m, %0.<Vetype>, #%2
   5701    #
   5702    fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5703    fsubr\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   5704    movprfx\t%0, %3\;fsubr\t%0.<Vetype>, %1/m, %0.<Vetype>, #%2
   5705    movprfx\t%0, %2\;fsub\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   5706   ; Split the unpredicated form after reload, so that we don't have
   5707   ; the unnecessary PTRUE.
   5708   "&& reload_completed
   5709    && register_operand (operands[2], <MODE>mode)
   5710    && INTVAL (operands[4]) == SVE_RELAXED_GP"
   5711   [(set (match_dup 0) (minus:SVE_FULL_F (match_dup 2) (match_dup 3)))]
   5712   ""
   5713   [(set_attr "movprfx" "*,*,*,*,yes,yes")]
   5714 )
   5715 
   5716 ;; Predicated floating-point subtraction from a constant, merging with the
   5717 ;; second input.
   5718 (define_insn_and_rewrite "*cond_sub<mode>_3_const_relaxed"
   5719   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?w")
   5720 	(unspec:SVE_FULL_F
   5721 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5722 	   (unspec:SVE_FULL_F
   5723 	     [(match_operand 4)
   5724 	      (const_int SVE_RELAXED_GP)
   5725 	      (match_operand:SVE_FULL_F 2 "aarch64_sve_float_arith_immediate")
   5726 	      (match_operand:SVE_FULL_F 3 "register_operand" "0, w")]
   5727 	     UNSPEC_COND_FSUB)
   5728 	   (match_dup 3)]
   5729 	  UNSPEC_SEL))]
   5730   "TARGET_SVE"
   5731   "@
   5732    fsubr\t%0.<Vetype>, %1/m, %0.<Vetype>, #%2
   5733    movprfx\t%0, %3\;fsubr\t%0.<Vetype>, %1/m, %0.<Vetype>, #%2"
   5734   "&& !rtx_equal_p (operands[1], operands[4])"
   5735   {
   5736     operands[4] = copy_rtx (operands[1]);
   5737   }
   5738   [(set_attr "movprfx" "*,yes")]
   5739 )
   5740 
   5741 (define_insn "*cond_sub<mode>_3_const_strict"
   5742   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?w")
   5743 	(unspec:SVE_FULL_F
   5744 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5745 	   (unspec:SVE_FULL_F
   5746 	     [(match_dup 1)
   5747 	      (const_int SVE_STRICT_GP)
   5748 	      (match_operand:SVE_FULL_F 2 "aarch64_sve_float_arith_immediate")
   5749 	      (match_operand:SVE_FULL_F 3 "register_operand" "0, w")]
   5750 	     UNSPEC_COND_FSUB)
   5751 	   (match_dup 3)]
   5752 	  UNSPEC_SEL))]
   5753   "TARGET_SVE"
   5754   "@
   5755    fsubr\t%0.<Vetype>, %1/m, %0.<Vetype>, #%2
   5756    movprfx\t%0, %3\;fsubr\t%0.<Vetype>, %1/m, %0.<Vetype>, #%2"
   5757   [(set_attr "movprfx" "*,yes")]
   5758 )
   5759 
   5760 ;; Predicated floating-point subtraction from a constant, merging with an
   5761 ;; independent value.
   5762 (define_insn_and_rewrite "*cond_sub<mode>_const_relaxed"
   5763   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, ?w")
   5764 	(unspec:SVE_FULL_F
   5765 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   5766 	   (unspec:SVE_FULL_F
   5767 	     [(match_operand 5)
   5768 	      (const_int SVE_RELAXED_GP)
   5769 	      (match_operand:SVE_FULL_F 2 "aarch64_sve_float_arith_immediate")
   5770 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w, w")]
   5771 	     UNSPEC_COND_FSUB)
   5772 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, 0, w")]
   5773 	  UNSPEC_SEL))]
   5774   "TARGET_SVE && !rtx_equal_p (operands[3], operands[4])"
   5775   "@
   5776    movprfx\t%0.<Vetype>, %1/z, %3.<Vetype>\;fsubr\t%0.<Vetype>, %1/m, %0.<Vetype>, #%2
   5777    movprfx\t%0.<Vetype>, %1/m, %3.<Vetype>\;fsubr\t%0.<Vetype>, %1/m, %0.<Vetype>, #%2
   5778    #"
   5779   "&& 1"
   5780   {
   5781     if (reload_completed
   5782         && register_operand (operands[4], <MODE>mode)
   5783         && !rtx_equal_p (operands[0], operands[4]))
   5784       {
   5785 	emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[3],
   5786 						 operands[4], operands[1]));
   5787 	operands[4] = operands[3] = operands[0];
   5788       }
   5789     else if (!rtx_equal_p (operands[1], operands[5]))
   5790       operands[5] = copy_rtx (operands[1]);
   5791     else
   5792       FAIL;
   5793   }
   5794   [(set_attr "movprfx" "yes")]
   5795 )
   5796 
   5797 (define_insn_and_rewrite "*cond_sub<mode>_const_strict"
   5798   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, ?w")
   5799 	(unspec:SVE_FULL_F
   5800 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   5801 	   (unspec:SVE_FULL_F
   5802 	     [(match_dup 1)
   5803 	      (const_int SVE_STRICT_GP)
   5804 	      (match_operand:SVE_FULL_F 2 "aarch64_sve_float_arith_immediate")
   5805 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w, w")]
   5806 	     UNSPEC_COND_FSUB)
   5807 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, 0, w")]
   5808 	  UNSPEC_SEL))]
   5809   "TARGET_SVE && !rtx_equal_p (operands[3], operands[4])"
   5810   "@
   5811    movprfx\t%0.<Vetype>, %1/z, %3.<Vetype>\;fsubr\t%0.<Vetype>, %1/m, %0.<Vetype>, #%2
   5812    movprfx\t%0.<Vetype>, %1/m, %3.<Vetype>\;fsubr\t%0.<Vetype>, %1/m, %0.<Vetype>, #%2
   5813    #"
   5814   "&& reload_completed
   5815    && register_operand (operands[4], <MODE>mode)
   5816    && !rtx_equal_p (operands[0], operands[4])"
   5817   {
   5818     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[3],
   5819                                              operands[4], operands[1]));
   5820     operands[4] = operands[3] = operands[0];
   5821   }
   5822   [(set_attr "movprfx" "yes")]
   5823 )
   5824 ;; Register merging forms are handled through SVE_COND_FP_BINARY.
   5825 
   5826 ;; -------------------------------------------------------------------------
   5827 ;; ---- [FP] Absolute difference
   5828 ;; -------------------------------------------------------------------------
   5829 ;; Includes:
   5830 ;; - FABD
   5831 ;; -------------------------------------------------------------------------
   5832 
   5833 ;; Predicated floating-point absolute difference.
   5834 (define_expand "@aarch64_pred_abd<mode>"
   5835   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   5836 	(unspec:SVE_FULL_F
   5837 	  [(match_operand:<VPRED> 1 "register_operand")
   5838 	   (match_operand:SI 4 "aarch64_sve_gp_strictness")
   5839 	   (unspec:SVE_FULL_F
   5840 	     [(match_dup 1)
   5841 	      (match_dup 4)
   5842 	      (match_operand:SVE_FULL_F 2 "register_operand")
   5843 	      (match_operand:SVE_FULL_F 3 "register_operand")]
   5844 	     UNSPEC_COND_FSUB)]
   5845 	  UNSPEC_COND_FABS))]
   5846   "TARGET_SVE"
   5847 )
   5848 
   5849 ;; Predicated floating-point absolute difference.
   5850 (define_insn_and_rewrite "*aarch64_pred_abd<mode>_relaxed"
   5851   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   5852 	(unspec:SVE_FULL_F
   5853 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5854 	   (match_operand:SI 4 "aarch64_sve_gp_strictness")
   5855 	   (unspec:SVE_FULL_F
   5856 	     [(match_operand 5)
   5857 	      (const_int SVE_RELAXED_GP)
   5858 	      (match_operand:SVE_FULL_F 2 "register_operand" "%0, w")
   5859 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w")]
   5860 	     UNSPEC_COND_FSUB)]
   5861 	  UNSPEC_COND_FABS))]
   5862   "TARGET_SVE"
   5863   "@
   5864    fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5865    movprfx\t%0, %2\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   5866   "&& !rtx_equal_p (operands[1], operands[5])"
   5867   {
   5868     operands[5] = copy_rtx (operands[1]);
   5869   }
   5870   [(set_attr "movprfx" "*,yes")]
   5871 )
   5872 
   5873 (define_insn "*aarch64_pred_abd<mode>_strict"
   5874   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   5875 	(unspec:SVE_FULL_F
   5876 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5877 	   (match_operand:SI 4 "aarch64_sve_gp_strictness")
   5878 	   (unspec:SVE_FULL_F
   5879 	     [(match_dup 1)
   5880 	      (const_int SVE_STRICT_GP)
   5881 	      (match_operand:SVE_FULL_F 2 "register_operand" "%0, w")
   5882 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w")]
   5883 	     UNSPEC_COND_FSUB)]
   5884 	  UNSPEC_COND_FABS))]
   5885   "TARGET_SVE"
   5886   "@
   5887    fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5888    movprfx\t%0, %2\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   5889   [(set_attr "movprfx" "*,yes")]
   5890 )
   5891 
   5892 (define_expand "@aarch64_cond_abd<mode>"
   5893   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   5894 	(unspec:SVE_FULL_F
   5895 	  [(match_operand:<VPRED> 1 "register_operand")
   5896 	   (unspec:SVE_FULL_F
   5897 	     [(match_dup 1)
   5898 	      (const_int SVE_STRICT_GP)
   5899 	      (unspec:SVE_FULL_F
   5900 		[(match_dup 1)
   5901 		 (const_int SVE_STRICT_GP)
   5902 		 (match_operand:SVE_FULL_F 2 "register_operand")
   5903 		 (match_operand:SVE_FULL_F 3 "register_operand")]
   5904 		UNSPEC_COND_FSUB)]
   5905 	     UNSPEC_COND_FABS)
   5906 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero")]
   5907 	  UNSPEC_SEL))]
   5908   "TARGET_SVE"
   5909 {
   5910   if (rtx_equal_p (operands[3], operands[4]))
   5911     std::swap (operands[2], operands[3]);
   5912 })
   5913 
   5914 ;; Predicated floating-point absolute difference, merging with the first
   5915 ;; input.
   5916 (define_insn_and_rewrite "*aarch64_cond_abd<mode>_2_relaxed"
   5917   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   5918 	(unspec:SVE_FULL_F
   5919 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5920 	   (unspec:SVE_FULL_F
   5921 	     [(match_operand 4)
   5922 	      (const_int SVE_RELAXED_GP)
   5923 	      (unspec:SVE_FULL_F
   5924 		[(match_operand 5)
   5925 		 (const_int SVE_RELAXED_GP)
   5926 		 (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   5927 		 (match_operand:SVE_FULL_F 3 "register_operand" "w, w")]
   5928 		UNSPEC_COND_FSUB)]
   5929 	     UNSPEC_COND_FABS)
   5930 	   (match_dup 2)]
   5931 	  UNSPEC_SEL))]
   5932   "TARGET_SVE"
   5933   "@
   5934    fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5935    movprfx\t%0, %2\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   5936   "&& (!rtx_equal_p (operands[1], operands[4])
   5937        || !rtx_equal_p (operands[1], operands[5]))"
   5938   {
   5939     operands[4] = copy_rtx (operands[1]);
   5940     operands[5] = copy_rtx (operands[1]);
   5941   }
   5942   [(set_attr "movprfx" "*,yes")]
   5943 )
   5944 
   5945 (define_insn "*aarch64_cond_abd<mode>_2_strict"
   5946   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   5947 	(unspec:SVE_FULL_F
   5948 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5949 	   (unspec:SVE_FULL_F
   5950 	     [(match_dup 1)
   5951 	      (match_operand:SI 4 "aarch64_sve_gp_strictness")
   5952 	      (unspec:SVE_FULL_F
   5953 		[(match_dup 1)
   5954 		 (match_operand:SI 5 "aarch64_sve_gp_strictness")
   5955 		 (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   5956 		 (match_operand:SVE_FULL_F 3 "register_operand" "w, w")]
   5957 		UNSPEC_COND_FSUB)]
   5958 	     UNSPEC_COND_FABS)
   5959 	   (match_dup 2)]
   5960 	  UNSPEC_SEL))]
   5961   "TARGET_SVE"
   5962   "@
   5963    fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   5964    movprfx\t%0, %2\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   5965   [(set_attr "movprfx" "*,yes")]
   5966 )
   5967 
   5968 ;; Predicated floating-point absolute difference, merging with the second
   5969 ;; input.
   5970 (define_insn_and_rewrite "*aarch64_cond_abd<mode>_3_relaxed"
   5971   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   5972 	(unspec:SVE_FULL_F
   5973 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   5974 	   (unspec:SVE_FULL_F
   5975 	     [(match_operand 4)
   5976 	      (const_int SVE_RELAXED_GP)
   5977 	      (unspec:SVE_FULL_F
   5978 		[(match_operand 5)
   5979 		 (const_int SVE_RELAXED_GP)
   5980 		 (match_operand:SVE_FULL_F 2 "register_operand" "w, w")
   5981 		 (match_operand:SVE_FULL_F 3 "register_operand" "0, w")]
   5982 		UNSPEC_COND_FSUB)]
   5983 	     UNSPEC_COND_FABS)
   5984 	   (match_dup 3)]
   5985 	  UNSPEC_SEL))]
   5986   "TARGET_SVE"
   5987   "@
   5988    fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   5989    movprfx\t%0, %3\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>"
   5990   "&& (!rtx_equal_p (operands[1], operands[4])
   5991        || !rtx_equal_p (operands[1], operands[5]))"
   5992   {
   5993     operands[4] = copy_rtx (operands[1]);
   5994     operands[5] = copy_rtx (operands[1]);
   5995   }
   5996   [(set_attr "movprfx" "*,yes")]
   5997 )
   5998 
   5999 (define_insn "*aarch64_cond_abd<mode>_3_strict"
   6000   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   6001 	(unspec:SVE_FULL_F
   6002 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   6003 	   (unspec:SVE_FULL_F
   6004 	     [(match_dup 1)
   6005 	      (match_operand:SI 4 "aarch64_sve_gp_strictness")
   6006 	      (unspec:SVE_FULL_F
   6007 		[(match_dup 1)
   6008 		 (match_operand:SI 5 "aarch64_sve_gp_strictness")
   6009 		 (match_operand:SVE_FULL_F 2 "register_operand" "w, w")
   6010 		 (match_operand:SVE_FULL_F 3 "register_operand" "0, w")]
   6011 		UNSPEC_COND_FSUB)]
   6012 	     UNSPEC_COND_FABS)
   6013 	   (match_dup 3)]
   6014 	  UNSPEC_SEL))]
   6015   "TARGET_SVE"
   6016   "@
   6017    fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   6018    movprfx\t%0, %3\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>"
   6019   [(set_attr "movprfx" "*,yes")]
   6020 )
   6021 
   6022 ;; Predicated floating-point absolute difference, merging with an
   6023 ;; independent value.
   6024 (define_insn_and_rewrite "*aarch64_cond_abd<mode>_any_relaxed"
   6025   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, &w, &w, ?&w")
   6026 	(unspec:SVE_FULL_F
   6027 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl")
   6028 	   (unspec:SVE_FULL_F
   6029 	     [(match_operand 5)
   6030 	      (const_int SVE_RELAXED_GP)
   6031 	      (unspec:SVE_FULL_F
   6032 		[(match_operand 6)
   6033 		 (const_int SVE_RELAXED_GP)
   6034 		 (match_operand:SVE_FULL_F 2 "register_operand" "0, w, w, w, w")
   6035 		 (match_operand:SVE_FULL_F 3 "register_operand" "w, 0, w, w, w")]
   6036 		UNSPEC_COND_FSUB)]
   6037 	     UNSPEC_COND_FABS)
   6038 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, Dz, Dz, 0, w")]
   6039 	  UNSPEC_SEL))]
   6040   "TARGET_SVE
   6041    && !rtx_equal_p (operands[2], operands[4])
   6042    && !rtx_equal_p (operands[3], operands[4])"
   6043   "@
   6044    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   6045    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   6046    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   6047    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   6048    #"
   6049   "&& 1"
   6050   {
   6051     if (reload_completed
   6052 	&& register_operand (operands[4], <MODE>mode)
   6053 	&& !rtx_equal_p (operands[0], operands[4]))
   6054       {
   6055 	emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[3],
   6056 						 operands[4], operands[1]));
   6057 	operands[4] = operands[3] = operands[0];
   6058       }
   6059     else if (!rtx_equal_p (operands[1], operands[5])
   6060 	     || !rtx_equal_p (operands[1], operands[6]))
   6061       {
   6062 	operands[5] = copy_rtx (operands[1]);
   6063 	operands[6] = copy_rtx (operands[1]);
   6064       }
   6065     else
   6066       FAIL;
   6067   }
   6068   [(set_attr "movprfx" "yes")]
   6069 )
   6070 
   6071 (define_insn_and_rewrite "*aarch64_cond_abd<mode>_any_strict"
   6072   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, &w, &w, ?&w")
   6073 	(unspec:SVE_FULL_F
   6074 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl")
   6075 	   (unspec:SVE_FULL_F
   6076 	     [(match_dup 1)
   6077 	      (match_operand:SI 5 "aarch64_sve_gp_strictness")
   6078 	      (unspec:SVE_FULL_F
   6079 		[(match_dup 1)
   6080 		 (match_operand:SI 6 "aarch64_sve_gp_strictness")
   6081 		 (match_operand:SVE_FULL_F 2 "register_operand" "0, w, w, w, w")
   6082 		 (match_operand:SVE_FULL_F 3 "register_operand" "w, 0, w, w, w")]
   6083 		UNSPEC_COND_FSUB)]
   6084 	     UNSPEC_COND_FABS)
   6085 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, Dz, Dz, 0, w")]
   6086 	  UNSPEC_SEL))]
   6087   "TARGET_SVE
   6088    && !rtx_equal_p (operands[2], operands[4])
   6089    && !rtx_equal_p (operands[3], operands[4])"
   6090   "@
   6091    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   6092    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %2.<Vetype>
   6093    movprfx\t%0.<Vetype>, %1/z, %2.<Vetype>\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   6094    movprfx\t%0.<Vetype>, %1/m, %2.<Vetype>\;fabd\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   6095    #"
   6096   "&& reload_completed
   6097    && register_operand (operands[4], <MODE>mode)
   6098    && !rtx_equal_p (operands[0], operands[4])"
   6099   {
   6100     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[3],
   6101 					     operands[4], operands[1]));
   6102     operands[4] = operands[3] = operands[0];
   6103   }
   6104   [(set_attr "movprfx" "yes")]
   6105 )
   6106 
   6107 ;; -------------------------------------------------------------------------
   6108 ;; ---- [FP] Multiplication
   6109 ;; -------------------------------------------------------------------------
   6110 ;; Includes:
   6111 ;; - FMUL
   6112 ;; -------------------------------------------------------------------------
   6113 
   6114 ;; Predicated floating-point multiplication.
   6115 (define_insn_and_split "@aarch64_pred_<optab><mode>"
   6116   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, w, ?&w, ?&w")
   6117 	(unspec:SVE_FULL_F
   6118 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl")
   6119 	   (match_operand:SI 4 "aarch64_sve_gp_strictness" "i, Z, Ui1, i, Ui1")
   6120 	   (match_operand:SVE_FULL_F 2 "register_operand" "%0, w, 0, w, w")
   6121 	   (match_operand:SVE_FULL_F 3 "aarch64_sve_float_mul_operand" "vsM, w, w, vsM, w")]
   6122 	  SVE_COND_FP_MUL))]
   6123   "TARGET_SVE"
   6124   "@
   6125    fmul\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   6126    #
   6127    fmul\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   6128    movprfx\t%0, %2\;fmul\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   6129    movprfx\t%0, %2\;fmul\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   6130   ; Split the unpredicated form after reload, so that we don't have
   6131   ; the unnecessary PTRUE.
   6132   "&& reload_completed
   6133    && register_operand (operands[3], <MODE>mode)
   6134    && INTVAL (operands[4]) == SVE_RELAXED_GP"
   6135   [(set (match_dup 0) (mult:SVE_FULL_F (match_dup 2) (match_dup 3)))]
   6136   ""
   6137   [(set_attr "movprfx" "*,*,*,yes,yes")]
   6138 )
   6139 
   6140 ;; Merging forms are handled through SVE_COND_FP_BINARY and
   6141 ;; SVE_COND_FP_BINARY_I1.
   6142 
   6143 ;; Unpredicated multiplication by selected lanes.
   6144 (define_insn "@aarch64_mul_lane_<mode>"
   6145   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w")
   6146 	(mult:SVE_FULL_F
   6147 	  (unspec:SVE_FULL_F
   6148 	    [(match_operand:SVE_FULL_F 2 "register_operand" "<sve_lane_con>")
   6149 	     (match_operand:SI 3 "const_int_operand")]
   6150 	    UNSPEC_SVE_LANE_SELECT)
   6151 	  (match_operand:SVE_FULL_F 1 "register_operand" "w")))]
   6152   "TARGET_SVE"
   6153   "fmul\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>[%3]"
   6154 )
   6155 
   6156 ;; -------------------------------------------------------------------------
   6157 ;; ---- [FP] Division
   6158 ;; -------------------------------------------------------------------------
   6159 ;; The patterns in this section are synthetic.
   6160 ;; -------------------------------------------------------------------------
   6161 
   6162 (define_expand "div<mode>3"
   6163   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   6164 	(unspec:SVE_FULL_F
   6165 	  [(match_dup 3)
   6166 	   (const_int SVE_RELAXED_GP)
   6167 	   (match_operand:SVE_FULL_F 1 "nonmemory_operand")
   6168 	   (match_operand:SVE_FULL_F 2 "register_operand")]
   6169 	  UNSPEC_COND_FDIV))]
   6170   "TARGET_SVE"
   6171   {
   6172     if (aarch64_emit_approx_div (operands[0], operands[1], operands[2]))
   6173       DONE;
   6174 
   6175     operands[1] = force_reg (<MODE>mode, operands[1]);
   6176     operands[3] = aarch64_ptrue_reg (<VPRED>mode);
   6177   }
   6178 )
   6179 
   6180 (define_expand "@aarch64_frecpe<mode>"
   6181   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   6182 	(unspec:SVE_FULL_F
   6183 	  [(match_operand:SVE_FULL_F 1 "register_operand")]
   6184 	  UNSPEC_FRECPE))]
   6185   "TARGET_SVE"
   6186 )
   6187 
   6188 (define_expand "@aarch64_frecps<mode>"
   6189   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   6190 	(unspec:SVE_FULL_F
   6191 	  [(match_operand:SVE_FULL_F 1 "register_operand")
   6192 	   (match_operand:SVE_FULL_F 2 "register_operand")]
   6193 	  UNSPEC_FRECPS))]
   6194   "TARGET_SVE"
   6195 )
   6196 
   6197 ;; -------------------------------------------------------------------------
   6198 ;; ---- [FP] Binary logical operations
   6199 ;; -------------------------------------------------------------------------
   6200 ;; Includes
   6201 ;; - AND
   6202 ;; - EOR
   6203 ;; - ORR
   6204 ;; -------------------------------------------------------------------------
   6205 
   6206 ;; Binary logical operations on floating-point modes.  We avoid subregs
   6207 ;; by providing this, but we need to use UNSPECs since rtx logical ops
   6208 ;; aren't defined for floating-point modes.
   6209 (define_insn "*<optab><mode>3"
   6210   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w")
   6211 	(unspec:SVE_FULL_F
   6212 	  [(match_operand:SVE_FULL_F 1 "register_operand" "w")
   6213 	   (match_operand:SVE_FULL_F 2 "register_operand" "w")]
   6214 	  LOGICALF))]
   6215   "TARGET_SVE"
   6216   "<logicalf_op>\t%0.d, %1.d, %2.d"
   6217 )
   6218 
   6219 ;; -------------------------------------------------------------------------
   6220 ;; ---- [FP] Sign copying
   6221 ;; -------------------------------------------------------------------------
   6222 ;; The patterns in this section are synthetic.
   6223 ;; -------------------------------------------------------------------------
   6224 
   6225 (define_expand "copysign<mode>3"
   6226   [(match_operand:SVE_FULL_F 0 "register_operand")
   6227    (match_operand:SVE_FULL_F 1 "register_operand")
   6228    (match_operand:SVE_FULL_F 2 "register_operand")]
   6229   "TARGET_SVE"
   6230   {
   6231     rtx sign = gen_reg_rtx (<V_INT_EQUIV>mode);
   6232     rtx mant = gen_reg_rtx (<V_INT_EQUIV>mode);
   6233     rtx int_res = gen_reg_rtx (<V_INT_EQUIV>mode);
   6234     int bits = GET_MODE_UNIT_BITSIZE (<MODE>mode) - 1;
   6235 
   6236     rtx arg1 = lowpart_subreg (<V_INT_EQUIV>mode, operands[1], <MODE>mode);
   6237     rtx arg2 = lowpart_subreg (<V_INT_EQUIV>mode, operands[2], <MODE>mode);
   6238 
   6239     emit_insn (gen_and<v_int_equiv>3
   6240 	       (sign, arg2,
   6241 		aarch64_simd_gen_const_vector_dup (<V_INT_EQUIV>mode,
   6242 						   HOST_WIDE_INT_M1U
   6243 						   << bits)));
   6244     emit_insn (gen_and<v_int_equiv>3
   6245 	       (mant, arg1,
   6246 		aarch64_simd_gen_const_vector_dup (<V_INT_EQUIV>mode,
   6247 						   ~(HOST_WIDE_INT_M1U
   6248 						     << bits))));
   6249     emit_insn (gen_ior<v_int_equiv>3 (int_res, sign, mant));
   6250     emit_move_insn (operands[0], gen_lowpart (<MODE>mode, int_res));
   6251     DONE;
   6252   }
   6253 )
   6254 
   6255 (define_expand "xorsign<mode>3"
   6256   [(match_operand:SVE_FULL_F 0 "register_operand")
   6257    (match_operand:SVE_FULL_F 1 "register_operand")
   6258    (match_operand:SVE_FULL_F 2 "register_operand")]
   6259   "TARGET_SVE"
   6260   {
   6261     rtx sign = gen_reg_rtx (<V_INT_EQUIV>mode);
   6262     rtx int_res = gen_reg_rtx (<V_INT_EQUIV>mode);
   6263     int bits = GET_MODE_UNIT_BITSIZE (<MODE>mode) - 1;
   6264 
   6265     rtx arg1 = lowpart_subreg (<V_INT_EQUIV>mode, operands[1], <MODE>mode);
   6266     rtx arg2 = lowpart_subreg (<V_INT_EQUIV>mode, operands[2], <MODE>mode);
   6267 
   6268     emit_insn (gen_and<v_int_equiv>3
   6269 	       (sign, arg2,
   6270 		aarch64_simd_gen_const_vector_dup (<V_INT_EQUIV>mode,
   6271 						   HOST_WIDE_INT_M1U
   6272 						   << bits)));
   6273     emit_insn (gen_xor<v_int_equiv>3 (int_res, arg1, sign));
   6274     emit_move_insn (operands[0], gen_lowpart (<MODE>mode, int_res));
   6275     DONE;
   6276   }
   6277 )
   6278 
   6279 ;; -------------------------------------------------------------------------
   6280 ;; ---- [FP] Maximum and minimum
   6281 ;; -------------------------------------------------------------------------
   6282 ;; Includes:
   6283 ;; - FMAX
   6284 ;; - FMAXNM
   6285 ;; - FMIN
   6286 ;; - FMINNM
   6287 ;; -------------------------------------------------------------------------
   6288 
   6289 ;; Unpredicated fmax/fmin (the libm functions).  The optabs for the
   6290 ;; smax/smin rtx codes are handled in the generic section above.
   6291 (define_expand "<fmaxmin><mode>3"
   6292   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   6293 	(unspec:SVE_FULL_F
   6294 	  [(match_dup 3)
   6295 	   (const_int SVE_RELAXED_GP)
   6296 	   (match_operand:SVE_FULL_F 1 "register_operand")
   6297 	   (match_operand:SVE_FULL_F 2 "aarch64_sve_float_maxmin_operand")]
   6298 	  SVE_COND_FP_MAXMIN_PUBLIC))]
   6299   "TARGET_SVE"
   6300   {
   6301     operands[3] = aarch64_ptrue_reg (<VPRED>mode);
   6302   }
   6303 )
   6304 
   6305 ;; Predicated fmax/fmin (the libm functions).  The optabs for the
   6306 ;; smax/smin rtx codes are handled in the generic section above.
   6307 (define_expand "cond_<fmaxmin><mode>"
   6308   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   6309 	(unspec:SVE_FULL_F
   6310 	  [(match_operand:<VPRED> 1 "register_operand")
   6311 	   (unspec:SVE_FULL_F
   6312 	     [(match_dup 1)
   6313 	      (const_int SVE_RELAXED_GP)
   6314 	      (match_operand:SVE_FULL_F 2 "register_operand")
   6315 	      (match_operand:SVE_FULL_F 3 "aarch64_sve_float_maxmin_operand")]
   6316 	     SVE_COND_FP_MAXMIN_PUBLIC)
   6317 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero")]
   6318 	  UNSPEC_SEL))]
   6319   "TARGET_SVE"
   6320 )
   6321 
   6322 ;; Predicated floating-point maximum/minimum.
   6323 (define_insn "@aarch64_pred_<optab><mode>"
   6324   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, ?&w, ?&w")
   6325 	(unspec:SVE_FULL_F
   6326 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl")
   6327 	   (match_operand:SI 4 "aarch64_sve_gp_strictness")
   6328 	   (match_operand:SVE_FULL_F 2 "register_operand" "%0, 0, w, w")
   6329 	   (match_operand:SVE_FULL_F 3 "aarch64_sve_float_maxmin_operand" "vsB, w, vsB, w")]
   6330 	  SVE_COND_FP_MAXMIN))]
   6331   "TARGET_SVE"
   6332   "@
   6333    <sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   6334    <sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>
   6335    movprfx\t%0, %2\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, #%3
   6336    movprfx\t%0, %2\;<sve_fp_op>\t%0.<Vetype>, %1/m, %0.<Vetype>, %3.<Vetype>"
   6337   [(set_attr "movprfx" "*,*,yes,yes")]
   6338 )
   6339 
   6340 ;; Merging forms are handled through SVE_COND_FP_BINARY and
   6341 ;; SVE_COND_FP_BINARY_I1.
   6342 
   6343 ;; -------------------------------------------------------------------------
   6344 ;; ---- [PRED] Binary logical operations
   6345 ;; -------------------------------------------------------------------------
   6346 ;; Includes:
   6347 ;; - AND
   6348 ;; - ANDS
   6349 ;; - EOR
   6350 ;; - EORS
   6351 ;; - ORR
   6352 ;; - ORRS
   6353 ;; -------------------------------------------------------------------------
   6354 
   6355 ;; Predicate AND.  We can reuse one of the inputs as the GP.
   6356 ;; Doubling the second operand is the preferred implementation
   6357 ;; of the MOV alias, so we use that instead of %1/z, %1, %2.
   6358 (define_insn "and<mode>3"
   6359   [(set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   6360 	(and:PRED_ALL (match_operand:PRED_ALL 1 "register_operand" "Upa")
   6361 		      (match_operand:PRED_ALL 2 "register_operand" "Upa")))]
   6362   "TARGET_SVE"
   6363   "and\t%0.b, %1/z, %2.b, %2.b"
   6364 )
   6365 
   6366 ;; Unpredicated predicate EOR and ORR.
   6367 (define_expand "<optab><mode>3"
   6368   [(set (match_operand:PRED_ALL 0 "register_operand")
   6369 	(and:PRED_ALL
   6370 	  (LOGICAL_OR:PRED_ALL
   6371 	    (match_operand:PRED_ALL 1 "register_operand")
   6372 	    (match_operand:PRED_ALL 2 "register_operand"))
   6373 	  (match_dup 3)))]
   6374   "TARGET_SVE"
   6375   {
   6376     operands[3] = aarch64_ptrue_reg (<MODE>mode);
   6377   }
   6378 )
   6379 
   6380 ;; Predicated predicate AND, EOR and ORR.
   6381 (define_insn "@aarch64_pred_<optab><mode>_z"
   6382   [(set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   6383 	(and:PRED_ALL
   6384 	  (LOGICAL:PRED_ALL
   6385 	    (match_operand:PRED_ALL 2 "register_operand" "Upa")
   6386 	    (match_operand:PRED_ALL 3 "register_operand" "Upa"))
   6387 	  (match_operand:PRED_ALL 1 "register_operand" "Upa")))]
   6388   "TARGET_SVE"
   6389   "<logical>\t%0.b, %1/z, %2.b, %3.b"
   6390 )
   6391 
   6392 ;; Perform a logical operation on operands 2 and 3, using operand 1 as
   6393 ;; the GP.  Store the result in operand 0 and set the flags in the same
   6394 ;; way as for PTEST.
   6395 (define_insn "*<optab><mode>3_cc"
   6396   [(set (reg:CC_NZC CC_REGNUM)
   6397 	(unspec:CC_NZC
   6398 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   6399 	   (match_operand 4)
   6400 	   (match_operand:SI 5 "aarch64_sve_ptrue_flag")
   6401 	   (and:PRED_ALL
   6402 	     (LOGICAL:PRED_ALL
   6403 	       (match_operand:PRED_ALL 2 "register_operand" "Upa")
   6404 	       (match_operand:PRED_ALL 3 "register_operand" "Upa"))
   6405 	     (match_dup 4))]
   6406 	  UNSPEC_PTEST))
   6407    (set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   6408 	(and:PRED_ALL (LOGICAL:PRED_ALL (match_dup 2) (match_dup 3))
   6409 		      (match_dup 4)))]
   6410   "TARGET_SVE"
   6411   "<logical>s\t%0.b, %1/z, %2.b, %3.b"
   6412 )
   6413 
   6414 ;; Same with just the flags result.
   6415 (define_insn "*<optab><mode>3_ptest"
   6416   [(set (reg:CC_NZC CC_REGNUM)
   6417 	(unspec:CC_NZC
   6418 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   6419 	   (match_operand 4)
   6420 	   (match_operand:SI 5 "aarch64_sve_ptrue_flag")
   6421 	   (and:PRED_ALL
   6422 	     (LOGICAL:PRED_ALL
   6423 	       (match_operand:PRED_ALL 2 "register_operand" "Upa")
   6424 	       (match_operand:PRED_ALL 3 "register_operand" "Upa"))
   6425 	     (match_dup 4))]
   6426 	  UNSPEC_PTEST))
   6427    (clobber (match_scratch:VNx16BI 0 "=Upa"))]
   6428   "TARGET_SVE"
   6429   "<logical>s\t%0.b, %1/z, %2.b, %3.b"
   6430 )
   6431 
   6432 ;; -------------------------------------------------------------------------
   6433 ;; ---- [PRED] Binary logical operations (inverted second input)
   6434 ;; -------------------------------------------------------------------------
   6435 ;; Includes:
   6436 ;; - BIC
   6437 ;; - ORN
   6438 ;; -------------------------------------------------------------------------
   6439 
   6440 ;; Predicated predicate BIC and ORN.
   6441 (define_insn "aarch64_pred_<nlogical><mode>_z"
   6442   [(set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   6443 	(and:PRED_ALL
   6444 	  (NLOGICAL:PRED_ALL
   6445 	    (not:PRED_ALL (match_operand:PRED_ALL 3 "register_operand" "Upa"))
   6446 	    (match_operand:PRED_ALL 2 "register_operand" "Upa"))
   6447 	  (match_operand:PRED_ALL 1 "register_operand" "Upa")))]
   6448   "TARGET_SVE"
   6449   "<nlogical>\t%0.b, %1/z, %2.b, %3.b"
   6450 )
   6451 
   6452 ;; Same, but set the flags as a side-effect.
   6453 (define_insn "*<nlogical><mode>3_cc"
   6454   [(set (reg:CC_NZC CC_REGNUM)
   6455 	(unspec:CC_NZC
   6456 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   6457 	   (match_operand 4)
   6458 	   (match_operand:SI 5 "aarch64_sve_ptrue_flag")
   6459 	   (and:PRED_ALL
   6460 	     (NLOGICAL:PRED_ALL
   6461 	       (not:PRED_ALL
   6462 		 (match_operand:PRED_ALL 3 "register_operand" "Upa"))
   6463 	       (match_operand:PRED_ALL 2 "register_operand" "Upa"))
   6464 	     (match_dup 4))]
   6465 	  UNSPEC_PTEST))
   6466    (set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   6467 	(and:PRED_ALL (NLOGICAL:PRED_ALL
   6468 			(not:PRED_ALL (match_dup 3))
   6469 			(match_dup 2))
   6470 		      (match_dup 4)))]
   6471   "TARGET_SVE"
   6472   "<nlogical>s\t%0.b, %1/z, %2.b, %3.b"
   6473 )
   6474 
   6475 ;; Same with just the flags result.
   6476 (define_insn "*<nlogical><mode>3_ptest"
   6477   [(set (reg:CC_NZC CC_REGNUM)
   6478 	(unspec:CC_NZC
   6479 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   6480 	   (match_operand 4)
   6481 	   (match_operand:SI 5 "aarch64_sve_ptrue_flag")
   6482 	   (and:PRED_ALL
   6483 	     (NLOGICAL:PRED_ALL
   6484 	       (not:PRED_ALL
   6485 		 (match_operand:PRED_ALL 3 "register_operand" "Upa"))
   6486 	       (match_operand:PRED_ALL 2 "register_operand" "Upa"))
   6487 	     (match_dup 4))]
   6488 	  UNSPEC_PTEST))
   6489    (clobber (match_scratch:VNx16BI 0 "=Upa"))]
   6490   "TARGET_SVE"
   6491   "<nlogical>s\t%0.b, %1/z, %2.b, %3.b"
   6492 )
   6493 
   6494 ;; -------------------------------------------------------------------------
   6495 ;; ---- [PRED] Binary logical operations (inverted result)
   6496 ;; -------------------------------------------------------------------------
   6497 ;; Includes:
   6498 ;; - NAND
   6499 ;; - NOR
   6500 ;; -------------------------------------------------------------------------
   6501 
   6502 ;; Predicated predicate NAND and NOR.
   6503 (define_insn "aarch64_pred_<logical_nn><mode>_z"
   6504   [(set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   6505 	(and:PRED_ALL
   6506 	  (NLOGICAL:PRED_ALL
   6507 	    (not:PRED_ALL (match_operand:PRED_ALL 2 "register_operand" "Upa"))
   6508 	    (not:PRED_ALL (match_operand:PRED_ALL 3 "register_operand" "Upa")))
   6509 	  (match_operand:PRED_ALL 1 "register_operand" "Upa")))]
   6510   "TARGET_SVE"
   6511   "<logical_nn>\t%0.b, %1/z, %2.b, %3.b"
   6512 )
   6513 
   6514 ;; Same, but set the flags as a side-effect.
   6515 (define_insn "*<logical_nn><mode>3_cc"
   6516   [(set (reg:CC_NZC CC_REGNUM)
   6517 	(unspec:CC_NZC
   6518 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   6519 	   (match_operand 4)
   6520 	   (match_operand:SI 5 "aarch64_sve_ptrue_flag")
   6521 	   (and:PRED_ALL
   6522 	     (NLOGICAL:PRED_ALL
   6523 	       (not:PRED_ALL
   6524 		 (match_operand:PRED_ALL 2 "register_operand" "Upa"))
   6525 	       (not:PRED_ALL
   6526 		 (match_operand:PRED_ALL 3 "register_operand" "Upa")))
   6527 	     (match_dup 4))]
   6528 	  UNSPEC_PTEST))
   6529    (set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   6530 	(and:PRED_ALL (NLOGICAL:PRED_ALL
   6531 			(not:PRED_ALL (match_dup 2))
   6532 			(not:PRED_ALL (match_dup 3)))
   6533 		      (match_dup 4)))]
   6534   "TARGET_SVE"
   6535   "<logical_nn>s\t%0.b, %1/z, %2.b, %3.b"
   6536 )
   6537 
   6538 ;; Same with just the flags result.
   6539 (define_insn "*<logical_nn><mode>3_ptest"
   6540   [(set (reg:CC_NZC CC_REGNUM)
   6541 	(unspec:CC_NZC
   6542 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   6543 	   (match_operand 4)
   6544 	   (match_operand:SI 5 "aarch64_sve_ptrue_flag")
   6545 	   (and:PRED_ALL
   6546 	     (NLOGICAL:PRED_ALL
   6547 	       (not:PRED_ALL
   6548 		 (match_operand:PRED_ALL 2 "register_operand" "Upa"))
   6549 	       (not:PRED_ALL
   6550 		 (match_operand:PRED_ALL 3 "register_operand" "Upa")))
   6551 	     (match_dup 4))]
   6552 	  UNSPEC_PTEST))
   6553    (clobber (match_scratch:VNx16BI 0 "=Upa"))]
   6554   "TARGET_SVE"
   6555   "<logical_nn>s\t%0.b, %1/z, %2.b, %3.b"
   6556 )
   6557 
   6558 ;; =========================================================================
   6559 ;; == Ternary arithmetic
   6560 ;; =========================================================================
   6561 
   6562 ;; -------------------------------------------------------------------------
   6563 ;; ---- [INT] MLA and MAD
   6564 ;; -------------------------------------------------------------------------
   6565 ;; Includes:
   6566 ;; - MAD
   6567 ;; - MLA
   6568 ;; -------------------------------------------------------------------------
   6569 
   6570 ;; Unpredicated integer addition of product.
   6571 (define_expand "fma<mode>4"
   6572   [(set (match_operand:SVE_I 0 "register_operand")
   6573 	(plus:SVE_I
   6574 	  (unspec:SVE_I
   6575 	    [(match_dup 4)
   6576 	     (mult:SVE_I
   6577 	       (match_operand:SVE_I 1 "register_operand")
   6578 	       (match_operand:SVE_I 2 "nonmemory_operand"))]
   6579 	    UNSPEC_PRED_X)
   6580 	  (match_operand:SVE_I 3 "register_operand")))]
   6581   "TARGET_SVE"
   6582   {
   6583     if (aarch64_prepare_sve_int_fma (operands, PLUS))
   6584       DONE;
   6585     operands[4] = aarch64_ptrue_reg (<VPRED>mode);
   6586   }
   6587 )
   6588 
   6589 ;; Predicated integer addition of product.
   6590 (define_insn "@aarch64_pred_fma<mode>"
   6591   [(set (match_operand:SVE_I 0 "register_operand" "=w, w, ?&w")
   6592 	(plus:SVE_I
   6593 	  (unspec:SVE_I
   6594 	    [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   6595 	     (mult:SVE_I
   6596 	       (match_operand:SVE_I 2 "register_operand" "%0, w, w")
   6597 	       (match_operand:SVE_I 3 "register_operand" "w, w, w"))]
   6598 	    UNSPEC_PRED_X)
   6599 	  (match_operand:SVE_I 4 "register_operand" "w, 0, w")))]
   6600   "TARGET_SVE"
   6601   "@
   6602    mad\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>
   6603    mla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   6604    movprfx\t%0, %4\;mla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>"
   6605   [(set_attr "movprfx" "*,*,yes")]
   6606 )
   6607 
   6608 ;; Predicated integer addition of product with merging.
   6609 (define_expand "cond_fma<mode>"
   6610   [(set (match_operand:SVE_I 0 "register_operand")
   6611 	(unspec:SVE_I
   6612 	  [(match_operand:<VPRED> 1 "register_operand")
   6613 	   (plus:SVE_I
   6614 	     (mult:SVE_I
   6615 	       (match_operand:SVE_I 2 "register_operand")
   6616 	       (match_operand:SVE_I 3 "general_operand"))
   6617 	     (match_operand:SVE_I 4 "register_operand"))
   6618 	   (match_operand:SVE_I 5 "aarch64_simd_reg_or_zero")]
   6619 	  UNSPEC_SEL))]
   6620   "TARGET_SVE"
   6621   {
   6622     if (aarch64_prepare_sve_cond_int_fma (operands, PLUS))
   6623       DONE;
   6624     /* Swap the multiplication operands if the fallback value is the
   6625        second of the two.  */
   6626     if (rtx_equal_p (operands[3], operands[5]))
   6627       std::swap (operands[2], operands[3]);
   6628   }
   6629 )
   6630 
   6631 ;; Predicated integer addition of product, merging with the first input.
   6632 (define_insn "*cond_fma<mode>_2"
   6633   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   6634 	(unspec:SVE_I
   6635 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   6636 	   (plus:SVE_I
   6637 	     (mult:SVE_I
   6638 	       (match_operand:SVE_I 2 "register_operand" "0, w")
   6639 	       (match_operand:SVE_I 3 "register_operand" "w, w"))
   6640 	     (match_operand:SVE_I 4 "register_operand" "w, w"))
   6641 	   (match_dup 2)]
   6642 	  UNSPEC_SEL))]
   6643   "TARGET_SVE"
   6644   "@
   6645    mad\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>
   6646    movprfx\t%0, %2\;mad\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>"
   6647   [(set_attr "movprfx" "*,yes")]
   6648 )
   6649 
   6650 ;; Predicated integer addition of product, merging with the third input.
   6651 (define_insn "*cond_fma<mode>_4"
   6652   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   6653 	(unspec:SVE_I
   6654 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   6655 	   (plus:SVE_I
   6656 	     (mult:SVE_I
   6657 	       (match_operand:SVE_I 2 "register_operand" "w, w")
   6658 	       (match_operand:SVE_I 3 "register_operand" "w, w"))
   6659 	     (match_operand:SVE_I 4 "register_operand" "0, w"))
   6660 	   (match_dup 4)]
   6661 	  UNSPEC_SEL))]
   6662   "TARGET_SVE"
   6663   "@
   6664    mla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   6665    movprfx\t%0, %4\;mla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>"
   6666   [(set_attr "movprfx" "*,yes")]
   6667 )
   6668 
   6669 ;; Predicated integer addition of product, merging with an independent value.
   6670 (define_insn_and_rewrite "*cond_fma<mode>_any"
   6671   [(set (match_operand:SVE_I 0 "register_operand" "=&w, &w, &w, &w, &w, ?&w")
   6672 	(unspec:SVE_I
   6673 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   6674 	   (plus:SVE_I
   6675 	     (mult:SVE_I
   6676 	       (match_operand:SVE_I 2 "register_operand" "w, w, 0, w, w, w")
   6677 	       (match_operand:SVE_I 3 "register_operand" "w, w, w, 0, w, w"))
   6678 	     (match_operand:SVE_I 4 "register_operand" "w, 0, w, w, w, w"))
   6679 	   (match_operand:SVE_I 5 "aarch64_simd_reg_or_zero" "Dz, Dz, Dz, Dz, 0, w")]
   6680 	  UNSPEC_SEL))]
   6681   "TARGET_SVE
   6682    && !rtx_equal_p (operands[2], operands[5])
   6683    && !rtx_equal_p (operands[3], operands[5])
   6684    && !rtx_equal_p (operands[4], operands[5])"
   6685   "@
   6686    movprfx\t%0.<Vetype>, %1/z, %4.<Vetype>\;mla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   6687    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;mla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   6688    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;mad\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>
   6689    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;mad\t%0.<Vetype>, %1/m, %2.<Vetype>, %4.<Vetype>
   6690    movprfx\t%0.<Vetype>, %1/m, %4.<Vetype>\;mla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   6691    #"
   6692   "&& reload_completed
   6693    && register_operand (operands[5], <MODE>mode)
   6694    && !rtx_equal_p (operands[0], operands[5])"
   6695   {
   6696     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[4],
   6697 					     operands[5], operands[1]));
   6698     operands[5] = operands[4] = operands[0];
   6699   }
   6700   [(set_attr "movprfx" "yes")]
   6701 )
   6702 
   6703 ;; -------------------------------------------------------------------------
   6704 ;; ---- [INT] MLS and MSB
   6705 ;; -------------------------------------------------------------------------
   6706 ;; Includes:
   6707 ;; - MLS
   6708 ;; - MSB
   6709 ;; -------------------------------------------------------------------------
   6710 
   6711 ;; Unpredicated integer subtraction of product.
   6712 (define_expand "fnma<mode>4"
   6713   [(set (match_operand:SVE_I 0 "register_operand")
   6714 	(minus:SVE_I
   6715 	  (match_operand:SVE_I 3 "register_operand")
   6716 	  (unspec:SVE_I
   6717 	    [(match_dup 4)
   6718 	     (mult:SVE_I
   6719 	       (match_operand:SVE_I 1 "register_operand")
   6720 	       (match_operand:SVE_I 2 "general_operand"))]
   6721 	    UNSPEC_PRED_X)))]
   6722   "TARGET_SVE"
   6723   {
   6724     if (aarch64_prepare_sve_int_fma (operands, MINUS))
   6725       DONE;
   6726     operands[4] = aarch64_ptrue_reg (<VPRED>mode);
   6727   }
   6728 )
   6729 
   6730 ;; Predicated integer subtraction of product.
   6731 (define_insn "@aarch64_pred_fnma<mode>"
   6732   [(set (match_operand:SVE_I 0 "register_operand" "=w, w, ?&w")
   6733 	(minus:SVE_I
   6734 	  (match_operand:SVE_I 4 "register_operand" "w, 0, w")
   6735 	  (unspec:SVE_I
   6736 	    [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   6737 	     (mult:SVE_I
   6738 	       (match_operand:SVE_I 2 "register_operand" "%0, w, w")
   6739 	       (match_operand:SVE_I 3 "register_operand" "w, w, w"))]
   6740 	    UNSPEC_PRED_X)))]
   6741   "TARGET_SVE"
   6742   "@
   6743    msb\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>
   6744    mls\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   6745    movprfx\t%0, %4\;mls\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>"
   6746   [(set_attr "movprfx" "*,*,yes")]
   6747 )
   6748 
   6749 ;; Predicated integer subtraction of product with merging.
   6750 (define_expand "cond_fnma<mode>"
   6751   [(set (match_operand:SVE_I 0 "register_operand")
   6752    (unspec:SVE_I
   6753 	[(match_operand:<VPRED> 1 "register_operand")
   6754 	 (minus:SVE_I
   6755 	   (match_operand:SVE_I 4 "register_operand")
   6756 	   (mult:SVE_I
   6757 	     (match_operand:SVE_I 2 "register_operand")
   6758 	     (match_operand:SVE_I 3 "general_operand")))
   6759 	 (match_operand:SVE_I 5 "aarch64_simd_reg_or_zero")]
   6760 	UNSPEC_SEL))]
   6761   "TARGET_SVE"
   6762   {
   6763     if (aarch64_prepare_sve_cond_int_fma (operands, MINUS))
   6764       DONE;
   6765     /* Swap the multiplication operands if the fallback value is the
   6766        second of the two.  */
   6767     if (rtx_equal_p (operands[3], operands[5]))
   6768       std::swap (operands[2], operands[3]);
   6769   }
   6770 )
   6771 
   6772 ;; Predicated integer subtraction of product, merging with the first input.
   6773 (define_insn "*cond_fnma<mode>_2"
   6774   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   6775 	(unspec:SVE_I
   6776 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   6777 	   (minus:SVE_I
   6778 	     (match_operand:SVE_I 4 "register_operand" "w, w")
   6779 	     (mult:SVE_I
   6780 	       (match_operand:SVE_I 2 "register_operand" "0, w")
   6781 	       (match_operand:SVE_I 3 "register_operand" "w, w")))
   6782 	   (match_dup 2)]
   6783 	  UNSPEC_SEL))]
   6784   "TARGET_SVE"
   6785   "@
   6786    msb\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>
   6787    movprfx\t%0, %2\;msb\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>"
   6788   [(set_attr "movprfx" "*,yes")]
   6789 )
   6790 
   6791 ;; Predicated integer subtraction of product, merging with the third input.
   6792 (define_insn "*cond_fnma<mode>_4"
   6793   [(set (match_operand:SVE_I 0 "register_operand" "=w, ?&w")
   6794 	(unspec:SVE_I
   6795 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   6796 	   (minus:SVE_I
   6797 	     (match_operand:SVE_I 4 "register_operand" "0, w")
   6798 	     (mult:SVE_I
   6799 	       (match_operand:SVE_I 2 "register_operand" "w, w")
   6800 	       (match_operand:SVE_I 3 "register_operand" "w, w")))
   6801 	   (match_dup 4)]
   6802 	  UNSPEC_SEL))]
   6803   "TARGET_SVE"
   6804   "@
   6805    mls\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   6806    movprfx\t%0, %4\;mls\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>"
   6807   [(set_attr "movprfx" "*,yes")]
   6808 )
   6809 
   6810 ;; Predicated integer subtraction of product, merging with an
   6811 ;; independent value.
   6812 (define_insn_and_rewrite "*cond_fnma<mode>_any"
   6813   [(set (match_operand:SVE_I 0 "register_operand" "=&w, &w, &w, &w, &w, ?&w")
   6814 	(unspec:SVE_I
   6815 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   6816 	   (minus:SVE_I
   6817 	     (match_operand:SVE_I 4 "register_operand" "w, 0, w, w, w, w")
   6818 	     (mult:SVE_I
   6819 	       (match_operand:SVE_I 2 "register_operand" "w, w, 0, w, w, w")
   6820 	       (match_operand:SVE_I 3 "register_operand" "w, w, w, 0, w, w")))
   6821 	   (match_operand:SVE_I 5 "aarch64_simd_reg_or_zero" "Dz, Dz, Dz, Dz, 0, w")]
   6822 	  UNSPEC_SEL))]
   6823   "TARGET_SVE
   6824    && !rtx_equal_p (operands[2], operands[5])
   6825    && !rtx_equal_p (operands[3], operands[5])
   6826    && !rtx_equal_p (operands[4], operands[5])"
   6827   "@
   6828    movprfx\t%0.<Vetype>, %1/z, %4.<Vetype>\;mls\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   6829    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;mls\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   6830    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;msb\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>
   6831    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;msb\t%0.<Vetype>, %1/m, %2.<Vetype>, %4.<Vetype>
   6832    movprfx\t%0.<Vetype>, %1/m, %4.<Vetype>\;mls\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   6833    #"
   6834   "&& reload_completed
   6835    && register_operand (operands[5], <MODE>mode)
   6836    && !rtx_equal_p (operands[0], operands[5])"
   6837   {
   6838     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[4],
   6839 					     operands[5], operands[1]));
   6840     operands[5] = operands[4] = operands[0];
   6841   }
   6842   [(set_attr "movprfx" "yes")]
   6843 )
   6844 
   6845 ;; -------------------------------------------------------------------------
   6846 ;; ---- [INT] Dot product
   6847 ;; -------------------------------------------------------------------------
   6848 ;; Includes:
   6849 ;; - SDOT
   6850 ;; - SUDOT   (I8MM)
   6851 ;; - UDOT
   6852 ;; - USDOT   (I8MM)
   6853 ;; -------------------------------------------------------------------------
   6854 
   6855 ;; Four-element integer dot-product with accumulation.
   6856 (define_insn "<sur>dot_prod<vsi2qi>"
   6857   [(set (match_operand:SVE_FULL_SDI 0 "register_operand" "=w, ?&w")
   6858 	(plus:SVE_FULL_SDI
   6859 	  (unspec:SVE_FULL_SDI
   6860 	    [(match_operand:<VSI2QI> 1 "register_operand" "w, w")
   6861 	     (match_operand:<VSI2QI> 2 "register_operand" "w, w")]
   6862 	    DOTPROD)
   6863 	  (match_operand:SVE_FULL_SDI 3 "register_operand" "0, w")))]
   6864   "TARGET_SVE"
   6865   "@
   6866    <sur>dot\\t%0.<Vetype>, %1.<Vetype_fourth>, %2.<Vetype_fourth>
   6867    movprfx\t%0, %3\;<sur>dot\\t%0.<Vetype>, %1.<Vetype_fourth>, %2.<Vetype_fourth>"
   6868   [(set_attr "movprfx" "*,yes")]
   6869 )
   6870 
   6871 ;; Four-element integer dot-product by selected lanes with accumulation.
   6872 (define_insn "@aarch64_<sur>dot_prod_lane<vsi2qi>"
   6873   [(set (match_operand:SVE_FULL_SDI 0 "register_operand" "=w, ?&w")
   6874 	(plus:SVE_FULL_SDI
   6875 	  (unspec:SVE_FULL_SDI
   6876 	    [(match_operand:<VSI2QI> 1 "register_operand" "w, w")
   6877 	     (unspec:<VSI2QI>
   6878 	       [(match_operand:<VSI2QI> 2 "register_operand" "<sve_lane_con>, <sve_lane_con>")
   6879 		(match_operand:SI 3 "const_int_operand")]
   6880 	       UNSPEC_SVE_LANE_SELECT)]
   6881 	    DOTPROD)
   6882 	  (match_operand:SVE_FULL_SDI 4 "register_operand" "0, w")))]
   6883   "TARGET_SVE"
   6884   "@
   6885    <sur>dot\\t%0.<Vetype>, %1.<Vetype_fourth>, %2.<Vetype_fourth>[%3]
   6886    movprfx\t%0, %4\;<sur>dot\\t%0.<Vetype>, %1.<Vetype_fourth>, %2.<Vetype_fourth>[%3]"
   6887   [(set_attr "movprfx" "*,yes")]
   6888 )
   6889 
   6890 (define_insn "@<sur>dot_prod<vsi2qi>"
   6891   [(set (match_operand:VNx4SI_ONLY 0 "register_operand" "=w, ?&w")
   6892         (plus:VNx4SI_ONLY
   6893 	  (unspec:VNx4SI_ONLY
   6894 	    [(match_operand:<VSI2QI> 1 "register_operand" "w, w")
   6895 	     (match_operand:<VSI2QI> 2 "register_operand" "w, w")]
   6896 	    DOTPROD_US_ONLY)
   6897 	  (match_operand:VNx4SI_ONLY 3 "register_operand" "0, w")))]
   6898   "TARGET_SVE_I8MM"
   6899   "@
   6900    <sur>dot\\t%0.s, %1.b, %2.b
   6901    movprfx\t%0, %3\;<sur>dot\\t%0.s, %1.b, %2.b"
   6902    [(set_attr "movprfx" "*,yes")]
   6903 )
   6904 
   6905 (define_insn "@aarch64_<sur>dot_prod_lane<vsi2qi>"
   6906   [(set (match_operand:VNx4SI_ONLY 0 "register_operand" "=w, ?&w")
   6907 	(plus:VNx4SI_ONLY
   6908 	  (unspec:VNx4SI_ONLY
   6909 	    [(match_operand:<VSI2QI> 1 "register_operand" "w, w")
   6910 	     (unspec:<VSI2QI>
   6911 	       [(match_operand:<VSI2QI> 2 "register_operand" "y, y")
   6912 		(match_operand:SI 3 "const_int_operand")]
   6913 	       UNSPEC_SVE_LANE_SELECT)]
   6914 	    DOTPROD_I8MM)
   6915 	  (match_operand:VNx4SI_ONLY 4 "register_operand" "0, w")))]
   6916   "TARGET_SVE_I8MM"
   6917   "@
   6918    <sur>dot\\t%0.s, %1.b, %2.b[%3]
   6919    movprfx\t%0, %4\;<sur>dot\\t%0.s, %1.b, %2.b[%3]"
   6920   [(set_attr "movprfx" "*,yes")]
   6921 )
   6922 
   6923 ;; -------------------------------------------------------------------------
   6924 ;; ---- [INT] Sum of absolute differences
   6925 ;; -------------------------------------------------------------------------
   6926 ;; The patterns in this section are synthetic.
   6927 ;; -------------------------------------------------------------------------
   6928 
   6929 ;; Emit a sequence to produce a sum-of-absolute-differences of the inputs in
   6930 ;; operands 1 and 2.  The sequence also has to perform a widening reduction of
   6931 ;; the difference into a vector and accumulate that into operand 3 before
   6932 ;; copying that into the result operand 0.
   6933 ;; Perform that with a sequence of:
   6934 ;; MOV		ones.b, #1
   6935 ;; [SU]ABD	diff.b, p0/m, op1.b, op2.b
   6936 ;; MOVPRFX	op0, op3	// If necessary
   6937 ;; UDOT		op0.s, diff.b, ones.b
   6938 (define_expand "<sur>sad<vsi2qi>"
   6939   [(use (match_operand:SVE_FULL_SDI 0 "register_operand"))
   6940    (unspec:<VSI2QI> [(use (match_operand:<VSI2QI> 1 "register_operand"))
   6941 		    (use (match_operand:<VSI2QI> 2 "register_operand"))] ABAL)
   6942    (use (match_operand:SVE_FULL_SDI 3 "register_operand"))]
   6943   "TARGET_SVE"
   6944   {
   6945     rtx ones = force_reg (<VSI2QI>mode, CONST1_RTX (<VSI2QI>mode));
   6946     rtx diff = gen_reg_rtx (<VSI2QI>mode);
   6947     emit_insn (gen_<sur>abd<vsi2qi>_3 (diff, operands[1], operands[2]));
   6948     emit_insn (gen_udot_prod<vsi2qi> (operands[0], diff, ones, operands[3]));
   6949     DONE;
   6950   }
   6951 )
   6952 
   6953 ;; -------------------------------------------------------------------------
   6954 ;; ---- [INT] Matrix multiply-accumulate
   6955 ;; -------------------------------------------------------------------------
   6956 ;; Includes:
   6957 ;; - SMMLA (I8MM)
   6958 ;; - UMMLA (I8MM)
   6959 ;; - USMMLA (I8MM)
   6960 ;; -------------------------------------------------------------------------
   6961 
   6962 (define_insn "@aarch64_sve_add_<optab><vsi2qi>"
   6963   [(set (match_operand:VNx4SI_ONLY 0 "register_operand" "=w, ?&w")
   6964 	(plus:VNx4SI_ONLY
   6965 	  (unspec:VNx4SI_ONLY
   6966 	    [(match_operand:<VSI2QI> 2 "register_operand" "w, w")
   6967 	     (match_operand:<VSI2QI> 3 "register_operand" "w, w")]
   6968 	    MATMUL)
   6969 	  (match_operand:VNx4SI_ONLY 1 "register_operand" "0, w")))]
   6970   "TARGET_SVE_I8MM"
   6971   "@
   6972    <sur>mmla\\t%0.s, %2.b, %3.b
   6973    movprfx\t%0, %1\;<sur>mmla\\t%0.s, %2.b, %3.b"
   6974   [(set_attr "movprfx" "*,yes")]
   6975 )
   6976 
   6977 ;; -------------------------------------------------------------------------
   6978 ;; ---- [FP] General ternary arithmetic corresponding to unspecs
   6979 ;; -------------------------------------------------------------------------
   6980 ;; Includes merging patterns for:
   6981 ;; - FMAD
   6982 ;; - FMLA
   6983 ;; - FMLS
   6984 ;; - FMSB
   6985 ;; - FNMAD
   6986 ;; - FNMLA
   6987 ;; - FNMLS
   6988 ;; - FNMSB
   6989 ;; -------------------------------------------------------------------------
   6990 
   6991 ;; Unpredicated floating-point ternary operations.
   6992 (define_expand "<optab><mode>4"
   6993   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   6994 	(unspec:SVE_FULL_F
   6995 	  [(match_dup 4)
   6996 	   (const_int SVE_RELAXED_GP)
   6997 	   (match_operand:SVE_FULL_F 1 "register_operand")
   6998 	   (match_operand:SVE_FULL_F 2 "register_operand")
   6999 	   (match_operand:SVE_FULL_F 3 "register_operand")]
   7000 	  SVE_COND_FP_TERNARY))]
   7001   "TARGET_SVE"
   7002   {
   7003     operands[4] = aarch64_ptrue_reg (<VPRED>mode);
   7004   }
   7005 )
   7006 
   7007 ;; Predicated floating-point ternary operations.
   7008 (define_insn "@aarch64_pred_<optab><mode>"
   7009   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, w, ?&w")
   7010 	(unspec:SVE_FULL_F
   7011 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl")
   7012 	   (match_operand:SI 5 "aarch64_sve_gp_strictness")
   7013 	   (match_operand:SVE_FULL_F 2 "register_operand" "%w, 0, w")
   7014 	   (match_operand:SVE_FULL_F 3 "register_operand" "w, w, w")
   7015 	   (match_operand:SVE_FULL_F 4 "register_operand" "0, w, w")]
   7016 	  SVE_COND_FP_TERNARY))]
   7017   "TARGET_SVE"
   7018   "@
   7019    <sve_fmla_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   7020    <sve_fmad_op>\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>
   7021    movprfx\t%0, %4\;<sve_fmla_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>"
   7022   [(set_attr "movprfx" "*,*,yes")]
   7023 )
   7024 
   7025 ;; Predicated floating-point ternary operations with merging.
   7026 (define_expand "@cond_<optab><mode>"
   7027   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   7028 	(unspec:SVE_FULL_F
   7029 	  [(match_operand:<VPRED> 1 "register_operand")
   7030 	   (unspec:SVE_FULL_F
   7031 	     [(match_dup 1)
   7032 	      (const_int SVE_STRICT_GP)
   7033 	      (match_operand:SVE_FULL_F 2 "register_operand")
   7034 	      (match_operand:SVE_FULL_F 3 "register_operand")
   7035 	      (match_operand:SVE_FULL_F 4 "register_operand")]
   7036 	     SVE_COND_FP_TERNARY)
   7037 	   (match_operand:SVE_FULL_F 5 "aarch64_simd_reg_or_zero")]
   7038 	  UNSPEC_SEL))]
   7039   "TARGET_SVE"
   7040 {
   7041   /* Swap the multiplication operands if the fallback value is the
   7042      second of the two.  */
   7043   if (rtx_equal_p (operands[3], operands[5]))
   7044     std::swap (operands[2], operands[3]);
   7045 })
   7046 
   7047 ;; Predicated floating-point ternary operations, merging with the
   7048 ;; first input.
   7049 (define_insn_and_rewrite "*cond_<optab><mode>_2_relaxed"
   7050   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   7051 	(unspec:SVE_FULL_F
   7052 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   7053 	   (unspec:SVE_FULL_F
   7054 	     [(match_operand 5)
   7055 	      (const_int SVE_RELAXED_GP)
   7056 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   7057 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w")
   7058 	      (match_operand:SVE_FULL_F 4 "register_operand" "w, w")]
   7059 	     SVE_COND_FP_TERNARY)
   7060 	   (match_dup 2)]
   7061 	  UNSPEC_SEL))]
   7062   "TARGET_SVE"
   7063   "@
   7064    <sve_fmad_op>\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>
   7065    movprfx\t%0, %2\;<sve_fmad_op>\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>"
   7066   "&& !rtx_equal_p (operands[1], operands[5])"
   7067   {
   7068     operands[5] = copy_rtx (operands[1]);
   7069   }
   7070   [(set_attr "movprfx" "*,yes")]
   7071 )
   7072 
   7073 (define_insn "*cond_<optab><mode>_2_strict"
   7074   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   7075 	(unspec:SVE_FULL_F
   7076 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   7077 	   (unspec:SVE_FULL_F
   7078 	     [(match_dup 1)
   7079 	      (const_int SVE_STRICT_GP)
   7080 	      (match_operand:SVE_FULL_F 2 "register_operand" "0, w")
   7081 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w")
   7082 	      (match_operand:SVE_FULL_F 4 "register_operand" "w, w")]
   7083 	     SVE_COND_FP_TERNARY)
   7084 	   (match_dup 2)]
   7085 	  UNSPEC_SEL))]
   7086   "TARGET_SVE"
   7087   "@
   7088    <sve_fmad_op>\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>
   7089    movprfx\t%0, %2\;<sve_fmad_op>\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>"
   7090   [(set_attr "movprfx" "*,yes")]
   7091 )
   7092 
   7093 ;; Predicated floating-point ternary operations, merging with the
   7094 ;; third input.
   7095 (define_insn_and_rewrite "*cond_<optab><mode>_4_relaxed"
   7096   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   7097 	(unspec:SVE_FULL_F
   7098 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   7099 	   (unspec:SVE_FULL_F
   7100 	     [(match_operand 5)
   7101 	      (const_int SVE_RELAXED_GP)
   7102 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w")
   7103 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w")
   7104 	      (match_operand:SVE_FULL_F 4 "register_operand" "0, w")]
   7105 	     SVE_COND_FP_TERNARY)
   7106 	   (match_dup 4)]
   7107 	  UNSPEC_SEL))]
   7108   "TARGET_SVE"
   7109   "@
   7110    <sve_fmla_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   7111    movprfx\t%0, %4\;<sve_fmla_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>"
   7112   "&& !rtx_equal_p (operands[1], operands[5])"
   7113   {
   7114     operands[5] = copy_rtx (operands[1]);
   7115   }
   7116   [(set_attr "movprfx" "*,yes")]
   7117 )
   7118 
   7119 (define_insn "*cond_<optab><mode>_4_strict"
   7120   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   7121 	(unspec:SVE_FULL_F
   7122 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   7123 	   (unspec:SVE_FULL_F
   7124 	     [(match_dup 1)
   7125 	      (const_int SVE_STRICT_GP)
   7126 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w")
   7127 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w")
   7128 	      (match_operand:SVE_FULL_F 4 "register_operand" "0, w")]
   7129 	     SVE_COND_FP_TERNARY)
   7130 	   (match_dup 4)]
   7131 	  UNSPEC_SEL))]
   7132   "TARGET_SVE"
   7133   "@
   7134    <sve_fmla_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   7135    movprfx\t%0, %4\;<sve_fmla_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>"
   7136   [(set_attr "movprfx" "*,yes")]
   7137 )
   7138 
   7139 ;; Predicated floating-point ternary operations, merging with an
   7140 ;; independent value.
   7141 (define_insn_and_rewrite "*cond_<optab><mode>_any_relaxed"
   7142   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, &w, &w, &w, ?&w")
   7143 	(unspec:SVE_FULL_F
   7144 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   7145 	   (unspec:SVE_FULL_F
   7146 	     [(match_operand 6)
   7147 	      (const_int SVE_RELAXED_GP)
   7148 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w, 0, w, w, w")
   7149 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w, w, 0, w, w")
   7150 	      (match_operand:SVE_FULL_F 4 "register_operand" "w, 0, w, w, w, w")]
   7151 	     SVE_COND_FP_TERNARY)
   7152 	   (match_operand:SVE_FULL_F 5 "aarch64_simd_reg_or_zero" "Dz, Dz, Dz, Dz, 0, w")]
   7153 	  UNSPEC_SEL))]
   7154   "TARGET_SVE
   7155    && !rtx_equal_p (operands[2], operands[5])
   7156    && !rtx_equal_p (operands[3], operands[5])
   7157    && !rtx_equal_p (operands[4], operands[5])"
   7158   "@
   7159    movprfx\t%0.<Vetype>, %1/z, %4.<Vetype>\;<sve_fmla_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   7160    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_fmla_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   7161    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_fmad_op>\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>
   7162    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_fmad_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %4.<Vetype>
   7163    movprfx\t%0.<Vetype>, %1/m, %4.<Vetype>\;<sve_fmla_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   7164    #"
   7165   "&& 1"
   7166   {
   7167     if (reload_completed
   7168         && register_operand (operands[5], <MODE>mode)
   7169         && !rtx_equal_p (operands[0], operands[5]))
   7170       {
   7171 	emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[4],
   7172 						 operands[5], operands[1]));
   7173 	operands[5] = operands[4] = operands[0];
   7174       }
   7175     else if (!rtx_equal_p (operands[1], operands[6]))
   7176       operands[6] = copy_rtx (operands[1]);
   7177     else
   7178       FAIL;
   7179   }
   7180   [(set_attr "movprfx" "yes")]
   7181 )
   7182 
   7183 (define_insn_and_rewrite "*cond_<optab><mode>_any_strict"
   7184   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, &w, &w, &w, ?&w")
   7185 	(unspec:SVE_FULL_F
   7186 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   7187 	   (unspec:SVE_FULL_F
   7188 	     [(match_dup 1)
   7189 	      (const_int SVE_STRICT_GP)
   7190 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w, 0, w, w, w")
   7191 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w, w, 0, w, w")
   7192 	      (match_operand:SVE_FULL_F 4 "register_operand" "w, 0, w, w, w, w")]
   7193 	     SVE_COND_FP_TERNARY)
   7194 	   (match_operand:SVE_FULL_F 5 "aarch64_simd_reg_or_zero" "Dz, Dz, Dz, Dz, 0, w")]
   7195 	  UNSPEC_SEL))]
   7196   "TARGET_SVE
   7197    && !rtx_equal_p (operands[2], operands[5])
   7198    && !rtx_equal_p (operands[3], operands[5])
   7199    && !rtx_equal_p (operands[4], operands[5])"
   7200   "@
   7201    movprfx\t%0.<Vetype>, %1/z, %4.<Vetype>\;<sve_fmla_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   7202    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_fmla_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   7203    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_fmad_op>\t%0.<Vetype>, %1/m, %3.<Vetype>, %4.<Vetype>
   7204    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;<sve_fmad_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %4.<Vetype>
   7205    movprfx\t%0.<Vetype>, %1/m, %4.<Vetype>\;<sve_fmla_op>\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>
   7206    #"
   7207   "&& reload_completed
   7208    && register_operand (operands[5], <MODE>mode)
   7209    && !rtx_equal_p (operands[0], operands[5])"
   7210   {
   7211     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[4],
   7212 					     operands[5], operands[1]));
   7213     operands[5] = operands[4] = operands[0];
   7214   }
   7215   [(set_attr "movprfx" "yes")]
   7216 )
   7217 
   7218 ;; Unpredicated FMLA and FMLS by selected lanes.  It doesn't seem worth using
   7219 ;; (fma ...) since target-independent code won't understand the indexing.
   7220 (define_insn "@aarch64_<optab>_lane_<mode>"
   7221   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   7222 	(unspec:SVE_FULL_F
   7223 	  [(match_operand:SVE_FULL_F 1 "register_operand" "w, w")
   7224 	   (unspec:SVE_FULL_F
   7225 	     [(match_operand:SVE_FULL_F 2 "register_operand" "<sve_lane_con>, <sve_lane_con>")
   7226 	      (match_operand:SI 3 "const_int_operand")]
   7227 	     UNSPEC_SVE_LANE_SELECT)
   7228 	   (match_operand:SVE_FULL_F 4 "register_operand" "0, w")]
   7229 	  SVE_FP_TERNARY_LANE))]
   7230   "TARGET_SVE"
   7231   "@
   7232    <sve_fp_op>\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>[%3]
   7233    movprfx\t%0, %4\;<sve_fp_op>\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>[%3]"
   7234   [(set_attr "movprfx" "*,yes")]
   7235 )
   7236 
   7237 ;; -------------------------------------------------------------------------
   7238 ;; ---- [FP] Complex multiply-add
   7239 ;; -------------------------------------------------------------------------
   7240 ;; Includes merging patterns for:
   7241 ;; - FCMLA
   7242 ;; -------------------------------------------------------------------------
   7243 
   7244 ;; Predicated FCMLA.
   7245 (define_insn "@aarch64_pred_<optab><mode>"
   7246   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   7247 	(unspec:SVE_FULL_F
   7248 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   7249 	   (match_operand:SI 5 "aarch64_sve_gp_strictness")
   7250 	   (match_operand:SVE_FULL_F 2 "register_operand" "w, w")
   7251 	   (match_operand:SVE_FULL_F 3 "register_operand" "w, w")
   7252 	   (match_operand:SVE_FULL_F 4 "register_operand" "0, w")]
   7253 	  SVE_COND_FCMLA))]
   7254   "TARGET_SVE"
   7255   "@
   7256    fcmla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>, #<rot>
   7257    movprfx\t%0, %4\;fcmla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>, #<rot>"
   7258   [(set_attr "movprfx" "*,yes")]
   7259 )
   7260 
   7261 ;; unpredicated optab pattern for auto-vectorizer
   7262 ;; The complex mla/mls operations always need to expand to two instructions.
   7263 ;; The first operation does half the computation and the second does the
   7264 ;; remainder.  Because of this, expand early.
   7265 (define_expand "cml<fcmac1><conj_op><mode>4"
   7266   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   7267 	(unspec:SVE_FULL_F
   7268 	  [(match_dup 4)
   7269 	   (match_dup 5)
   7270 	   (match_operand:SVE_FULL_F 1 "register_operand")
   7271 	   (match_operand:SVE_FULL_F 2 "register_operand")
   7272 	   (match_operand:SVE_FULL_F 3 "register_operand")]
   7273 	  FCMLA_OP))]
   7274   "TARGET_SVE"
   7275 {
   7276   operands[4] = aarch64_ptrue_reg (<VPRED>mode);
   7277   operands[5] = gen_int_mode (SVE_RELAXED_GP, SImode);
   7278   rtx tmp = gen_reg_rtx (<MODE>mode);
   7279   emit_insn
   7280     (gen_aarch64_pred_fcmla<sve_rot1><mode> (tmp, operands[4],
   7281 					     operands[2], operands[1],
   7282 					     operands[3], operands[5]));
   7283   emit_insn
   7284     (gen_aarch64_pred_fcmla<sve_rot2><mode> (operands[0], operands[4],
   7285 					     operands[2], operands[1],
   7286 					     tmp, operands[5]));
   7287   DONE;
   7288 })
   7289 
   7290 ;; unpredicated optab pattern for auto-vectorizer
   7291 ;; The complex mul operations always need to expand to two instructions.
   7292 ;; The first operation does half the computation and the second does the
   7293 ;; remainder.  Because of this, expand early.
   7294 (define_expand "cmul<conj_op><mode>3"
   7295   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   7296 	(unspec:SVE_FULL_F
   7297 	   [(match_operand:SVE_FULL_F 1 "register_operand")
   7298 	    (match_operand:SVE_FULL_F 2 "register_operand")]
   7299 	  FCMUL_OP))]
   7300   "TARGET_SVE"
   7301 {
   7302   rtx pred_reg = aarch64_ptrue_reg (<VPRED>mode);
   7303   rtx gp_mode = gen_int_mode (SVE_RELAXED_GP, SImode);
   7304   rtx accum = force_reg (<MODE>mode, CONST0_RTX (<MODE>mode));
   7305   rtx tmp = gen_reg_rtx (<MODE>mode);
   7306   emit_insn
   7307     (gen_aarch64_pred_fcmla<sve_rot1><mode> (tmp, pred_reg,
   7308 					     operands[2], operands[1],
   7309 					     accum, gp_mode));
   7310   emit_insn
   7311     (gen_aarch64_pred_fcmla<sve_rot2><mode> (operands[0], pred_reg,
   7312 					     operands[2], operands[1],
   7313 					     tmp, gp_mode));
   7314   DONE;
   7315 })
   7316 
   7317 ;; Predicated FCMLA with merging.
   7318 (define_expand "@cond_<optab><mode>"
   7319   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   7320 	(unspec:SVE_FULL_F
   7321 	  [(match_operand:<VPRED> 1 "register_operand")
   7322 	   (unspec:SVE_FULL_F
   7323 	     [(match_dup 1)
   7324 	      (const_int SVE_STRICT_GP)
   7325 	      (match_operand:SVE_FULL_F 2 "register_operand")
   7326 	      (match_operand:SVE_FULL_F 3 "register_operand")
   7327 	      (match_operand:SVE_FULL_F 4 "register_operand")]
   7328 	     SVE_COND_FCMLA)
   7329 	   (match_operand:SVE_FULL_F 5 "aarch64_simd_reg_or_zero")]
   7330 	  UNSPEC_SEL))]
   7331   "TARGET_SVE"
   7332 )
   7333 
   7334 ;; Predicated FCMLA, merging with the third input.
   7335 (define_insn_and_rewrite "*cond_<optab><mode>_4_relaxed"
   7336   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   7337 	(unspec:SVE_FULL_F
   7338 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   7339 	   (unspec:SVE_FULL_F
   7340 	     [(match_operand 5)
   7341 	      (const_int SVE_RELAXED_GP)
   7342 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w")
   7343 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w")
   7344 	      (match_operand:SVE_FULL_F 4 "register_operand" "0, w")]
   7345 	     SVE_COND_FCMLA)
   7346 	   (match_dup 4)]
   7347 	  UNSPEC_SEL))]
   7348   "TARGET_SVE"
   7349   "@
   7350    fcmla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>, #<rot>
   7351    movprfx\t%0, %4\;fcmla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>, #<rot>"
   7352   "&& !rtx_equal_p (operands[1], operands[5])"
   7353   {
   7354     operands[5] = copy_rtx (operands[1]);
   7355   }
   7356   [(set_attr "movprfx" "*,yes")]
   7357 )
   7358 
   7359 (define_insn "*cond_<optab><mode>_4_strict"
   7360   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   7361 	(unspec:SVE_FULL_F
   7362 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   7363 	   (unspec:SVE_FULL_F
   7364 	     [(match_dup 1)
   7365 	      (const_int SVE_STRICT_GP)
   7366 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w")
   7367 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w")
   7368 	      (match_operand:SVE_FULL_F 4 "register_operand" "0, w")]
   7369 	     SVE_COND_FCMLA)
   7370 	   (match_dup 4)]
   7371 	  UNSPEC_SEL))]
   7372   "TARGET_SVE"
   7373   "@
   7374    fcmla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>, #<rot>
   7375    movprfx\t%0, %4\;fcmla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>, #<rot>"
   7376   [(set_attr "movprfx" "*,yes")]
   7377 )
   7378 
   7379 ;; Predicated FCMLA, merging with an independent value.
   7380 (define_insn_and_rewrite "*cond_<optab><mode>_any_relaxed"
   7381   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, &w, ?&w")
   7382 	(unspec:SVE_FULL_F
   7383 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl")
   7384 	   (unspec:SVE_FULL_F
   7385 	     [(match_operand 6)
   7386 	      (const_int SVE_RELAXED_GP)
   7387 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w, w, w")
   7388 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w, w, w")
   7389 	      (match_operand:SVE_FULL_F 4 "register_operand" "w, 0, w, w")]
   7390 	     SVE_COND_FCMLA)
   7391 	   (match_operand:SVE_FULL_F 5 "aarch64_simd_reg_or_zero" "Dz, Dz, 0, w")]
   7392 	  UNSPEC_SEL))]
   7393   "TARGET_SVE && !rtx_equal_p (operands[4], operands[5])"
   7394   "@
   7395    movprfx\t%0.<Vetype>, %1/z, %4.<Vetype>\;fcmla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>, #<rot>
   7396    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;fcmla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>, #<rot>
   7397    movprfx\t%0.<Vetype>, %1/m, %4.<Vetype>\;fcmla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>, #<rot>
   7398    #"
   7399   "&& 1"
   7400   {
   7401     if (reload_completed
   7402         && register_operand (operands[5], <MODE>mode)
   7403         && !rtx_equal_p (operands[0], operands[5]))
   7404       {
   7405 	emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[4],
   7406 						 operands[5], operands[1]));
   7407 	operands[5] = operands[4] = operands[0];
   7408       }
   7409     else if (!rtx_equal_p (operands[1], operands[6]))
   7410       operands[6] = copy_rtx (operands[1]);
   7411     else
   7412       FAIL;
   7413   }
   7414   [(set_attr "movprfx" "yes")]
   7415 )
   7416 
   7417 (define_insn_and_rewrite "*cond_<optab><mode>_any_strict"
   7418   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, &w, ?&w")
   7419 	(unspec:SVE_FULL_F
   7420 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl, Upl, Upl")
   7421 	   (unspec:SVE_FULL_F
   7422 	     [(match_dup 1)
   7423 	      (const_int SVE_STRICT_GP)
   7424 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w, w, w")
   7425 	      (match_operand:SVE_FULL_F 3 "register_operand" "w, w, w, w")
   7426 	      (match_operand:SVE_FULL_F 4 "register_operand" "w, 0, w, w")]
   7427 	     SVE_COND_FCMLA)
   7428 	   (match_operand:SVE_FULL_F 5 "aarch64_simd_reg_or_zero" "Dz, Dz, 0, w")]
   7429 	  UNSPEC_SEL))]
   7430   "TARGET_SVE && !rtx_equal_p (operands[4], operands[5])"
   7431   "@
   7432    movprfx\t%0.<Vetype>, %1/z, %4.<Vetype>\;fcmla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>, #<rot>
   7433    movprfx\t%0.<Vetype>, %1/z, %0.<Vetype>\;fcmla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>, #<rot>
   7434    movprfx\t%0.<Vetype>, %1/m, %4.<Vetype>\;fcmla\t%0.<Vetype>, %1/m, %2.<Vetype>, %3.<Vetype>, #<rot>
   7435    #"
   7436   "&& reload_completed
   7437    && register_operand (operands[5], <MODE>mode)
   7438    && !rtx_equal_p (operands[0], operands[5])"
   7439   {
   7440     emit_insn (gen_vcond_mask_<mode><vpred> (operands[0], operands[4],
   7441 					     operands[5], operands[1]));
   7442     operands[5] = operands[4] = operands[0];
   7443   }
   7444   [(set_attr "movprfx" "yes")]
   7445 )
   7446 
   7447 ;; Unpredicated FCMLA with indexing.
   7448 (define_insn "@aarch64_<optab>_lane_<mode>"
   7449   [(set (match_operand:SVE_FULL_HSF 0 "register_operand" "=w, ?&w")
   7450 	(unspec:SVE_FULL_HSF
   7451 	  [(match_operand:SVE_FULL_HSF 1 "register_operand" "w, w")
   7452 	   (unspec:SVE_FULL_HSF
   7453 	     [(match_operand:SVE_FULL_HSF 2 "register_operand" "<sve_lane_pair_con>, <sve_lane_pair_con>")
   7454 	      (match_operand:SI 3 "const_int_operand")]
   7455 	     UNSPEC_SVE_LANE_SELECT)
   7456 	   (match_operand:SVE_FULL_HSF 4 "register_operand" "0, w")]
   7457 	  FCMLA))]
   7458   "TARGET_SVE"
   7459   "@
   7460    fcmla\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>[%3], #<rot>
   7461    movprfx\t%0, %4\;fcmla\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>[%3], #<rot>"
   7462   [(set_attr "movprfx" "*,yes")]
   7463 )
   7464 
   7465 ;; -------------------------------------------------------------------------
   7466 ;; ---- [FP] Trigonometric multiply-add
   7467 ;; -------------------------------------------------------------------------
   7468 ;; Includes:
   7469 ;; - FTMAD
   7470 ;; -------------------------------------------------------------------------
   7471 
   7472 (define_insn "@aarch64_sve_tmad<mode>"
   7473   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   7474 	(unspec:SVE_FULL_F
   7475 	  [(match_operand:SVE_FULL_F 1 "register_operand" "0, w")
   7476 	   (match_operand:SVE_FULL_F 2 "register_operand" "w, w")
   7477 	   (match_operand:DI 3 "const_int_operand")]
   7478 	  UNSPEC_FTMAD))]
   7479   "TARGET_SVE"
   7480   "@
   7481    ftmad\t%0.<Vetype>, %0.<Vetype>, %2.<Vetype>, #%3
   7482    movprfx\t%0, %1\;ftmad\t%0.<Vetype>, %0.<Vetype>, %2.<Vetype>, #%3"
   7483   [(set_attr "movprfx" "*,yes")]
   7484 )
   7485 
   7486 ;; -------------------------------------------------------------------------
   7487 ;; ---- [FP] Bfloat16 long ternary arithmetic (SF,BF,BF)
   7488 ;; -------------------------------------------------------------------------
   7489 ;; Includes:
   7490 ;; - BFDOT (BF16)
   7491 ;; - BFMLALB (BF16)
   7492 ;; - BFMLALT (BF16)
   7493 ;; - BFMMLA (BF16)
   7494 ;; -------------------------------------------------------------------------
   7495 
   7496 (define_insn "@aarch64_sve_<sve_fp_op>vnx4sf"
   7497   [(set (match_operand:VNx4SF 0 "register_operand" "=w, ?&w")
   7498 	(unspec:VNx4SF
   7499 	  [(match_operand:VNx4SF 1 "register_operand" "0, w")
   7500 	   (match_operand:VNx8BF 2 "register_operand" "w, w")
   7501 	   (match_operand:VNx8BF 3 "register_operand" "w, w")]
   7502 	  SVE_BFLOAT_TERNARY_LONG))]
   7503   "TARGET_SVE_BF16"
   7504   "@
   7505    <sve_fp_op>\t%0.s, %2.h, %3.h
   7506    movprfx\t%0, %1\;<sve_fp_op>\t%0.s, %2.h, %3.h"
   7507   [(set_attr "movprfx" "*,yes")]
   7508 )
   7509 
   7510 ;; The immediate range is enforced before generating the instruction.
   7511 (define_insn "@aarch64_sve_<sve_fp_op>_lanevnx4sf"
   7512   [(set (match_operand:VNx4SF 0 "register_operand" "=w, ?&w")
   7513 	(unspec:VNx4SF
   7514 	  [(match_operand:VNx4SF 1 "register_operand" "0, w")
   7515 	   (match_operand:VNx8BF 2 "register_operand" "w, w")
   7516 	   (match_operand:VNx8BF 3 "register_operand" "y, y")
   7517 	   (match_operand:SI 4 "const_int_operand")]
   7518 	  SVE_BFLOAT_TERNARY_LONG_LANE))]
   7519   "TARGET_SVE_BF16"
   7520   "@
   7521    <sve_fp_op>\t%0.s, %2.h, %3.h[%4]
   7522    movprfx\t%0, %1\;<sve_fp_op>\t%0.s, %2.h, %3.h[%4]"
   7523   [(set_attr "movprfx" "*,yes")]
   7524 )
   7525 
   7526 ;; -------------------------------------------------------------------------
   7527 ;; ---- [FP] Matrix multiply-accumulate
   7528 ;; -------------------------------------------------------------------------
   7529 ;; Includes:
   7530 ;; - FMMLA (F32MM,F64MM)
   7531 ;; -------------------------------------------------------------------------
   7532 
   7533 ;; The mode iterator enforces the target requirements.
   7534 (define_insn "@aarch64_sve_<sve_fp_op><mode>"
   7535   [(set (match_operand:SVE_MATMULF 0 "register_operand" "=w, ?&w")
   7536 	(unspec:SVE_MATMULF
   7537 	  [(match_operand:SVE_MATMULF 2 "register_operand" "w, w")
   7538 	   (match_operand:SVE_MATMULF 3 "register_operand" "w, w")
   7539 	   (match_operand:SVE_MATMULF 1 "register_operand" "0, w")]
   7540 	  FMMLA))]
   7541   "TARGET_SVE"
   7542   "@
   7543    <sve_fp_op>\\t%0.<Vetype>, %2.<Vetype>, %3.<Vetype>
   7544    movprfx\t%0, %1\;<sve_fp_op>\\t%0.<Vetype>, %2.<Vetype>, %3.<Vetype>"
   7545   [(set_attr "movprfx" "*,yes")]
   7546 )
   7547 
   7548 ;; =========================================================================
   7549 ;; == Comparisons and selects
   7550 ;; =========================================================================
   7551 
   7552 ;; -------------------------------------------------------------------------
   7553 ;; ---- [INT,FP] Select based on predicates
   7554 ;; -------------------------------------------------------------------------
   7555 ;; Includes merging patterns for:
   7556 ;; - FMOV
   7557 ;; - MOV
   7558 ;; - SEL
   7559 ;; -------------------------------------------------------------------------
   7560 
   7561 ;; vcond_mask operand order: true, false, mask
   7562 ;; UNSPEC_SEL operand order: mask, true, false (as for VEC_COND_EXPR)
   7563 ;; SEL operand order:        mask, true, false
   7564 (define_expand "@vcond_mask_<mode><vpred>"
   7565   [(set (match_operand:SVE_ALL 0 "register_operand")
   7566 	(unspec:SVE_ALL
   7567 	  [(match_operand:<VPRED> 3 "register_operand")
   7568 	   (match_operand:SVE_ALL 1 "aarch64_sve_reg_or_dup_imm")
   7569 	   (match_operand:SVE_ALL 2 "aarch64_simd_reg_or_zero")]
   7570 	  UNSPEC_SEL))]
   7571   "TARGET_SVE"
   7572   {
   7573     if (register_operand (operands[1], <MODE>mode))
   7574       operands[2] = force_reg (<MODE>mode, operands[2]);
   7575   }
   7576 )
   7577 
   7578 ;; Selects between:
   7579 ;; - two registers
   7580 ;; - a duplicated immediate and a register
   7581 ;; - a duplicated immediate and zero
   7582 ;;
   7583 ;; For unpacked vectors, it doesn't really matter whether SEL uses the
   7584 ;; the container size or the element size.  If SEL used the container size,
   7585 ;; it would ignore undefined bits of the predicate but would copy the
   7586 ;; upper (undefined) bits of each container along with the defined bits.
   7587 ;; If SEL used the element size, it would use undefined bits of the predicate
   7588 ;; to select between undefined elements in each input vector.  Thus the only
   7589 ;; difference is whether the undefined bits in a container always come from
   7590 ;; the same input as the defined bits, or whether the choice can vary
   7591 ;; independently of the defined bits.
   7592 ;;
   7593 ;; For the other instructions, using the element size is more natural,
   7594 ;; so we do that for SEL as well.
   7595 (define_insn "*vcond_mask_<mode><vpred>"
   7596   [(set (match_operand:SVE_ALL 0 "register_operand" "=w, w, w, w, ?w, ?&w, ?&w")
   7597 	(unspec:SVE_ALL
   7598 	  [(match_operand:<VPRED> 3 "register_operand" "Upa, Upa, Upa, Upa, Upl, Upl, Upl")
   7599 	   (match_operand:SVE_ALL 1 "aarch64_sve_reg_or_dup_imm" "w, vss, vss, Ufc, Ufc, vss, Ufc")
   7600 	   (match_operand:SVE_ALL 2 "aarch64_simd_reg_or_zero" "w, 0, Dz, 0, Dz, w, w")]
   7601 	  UNSPEC_SEL))]
   7602   "TARGET_SVE
   7603    && (!register_operand (operands[1], <MODE>mode)
   7604        || register_operand (operands[2], <MODE>mode))"
   7605   "@
   7606    sel\t%0.<Vetype>, %3, %1.<Vetype>, %2.<Vetype>
   7607    mov\t%0.<Vetype>, %3/m, #%I1
   7608    mov\t%0.<Vetype>, %3/z, #%I1
   7609    fmov\t%0.<Vetype>, %3/m, #%1
   7610    movprfx\t%0.<Vetype>, %3/z, %0.<Vetype>\;fmov\t%0.<Vetype>, %3/m, #%1
   7611    movprfx\t%0, %2\;mov\t%0.<Vetype>, %3/m, #%I1
   7612    movprfx\t%0, %2\;fmov\t%0.<Vetype>, %3/m, #%1"
   7613   [(set_attr "movprfx" "*,*,*,*,yes,yes,yes")]
   7614 )
   7615 
   7616 ;; Optimize selects between a duplicated scalar variable and another vector,
   7617 ;; the latter of which can be a zero constant or a variable.  Treat duplicates
   7618 ;; of GPRs as being more expensive than duplicates of FPRs, since they
   7619 ;; involve a cross-file move.
   7620 (define_insn "@aarch64_sel_dup<mode>"
   7621   [(set (match_operand:SVE_ALL 0 "register_operand" "=?w, w, ??w, ?&w, ??&w, ?&w")
   7622 	(unspec:SVE_ALL
   7623 	  [(match_operand:<VPRED> 3 "register_operand" "Upl, Upl, Upl, Upl, Upl, Upl")
   7624 	   (vec_duplicate:SVE_ALL
   7625 	     (match_operand:<VEL> 1 "register_operand" "r, w, r, w, r, w"))
   7626 	   (match_operand:SVE_ALL 2 "aarch64_simd_reg_or_zero" "0, 0, Dz, Dz, w, w")]
   7627 	  UNSPEC_SEL))]
   7628   "TARGET_SVE"
   7629   "@
   7630    mov\t%0.<Vetype>, %3/m, %<vwcore>1
   7631    mov\t%0.<Vetype>, %3/m, %<Vetype>1
   7632    movprfx\t%0.<Vetype>, %3/z, %0.<Vetype>\;mov\t%0.<Vetype>, %3/m, %<vwcore>1
   7633    movprfx\t%0.<Vetype>, %3/z, %0.<Vetype>\;mov\t%0.<Vetype>, %3/m, %<Vetype>1
   7634    movprfx\t%0, %2\;mov\t%0.<Vetype>, %3/m, %<vwcore>1
   7635    movprfx\t%0, %2\;mov\t%0.<Vetype>, %3/m, %<Vetype>1"
   7636   [(set_attr "movprfx" "*,*,yes,yes,yes,yes")]
   7637 )
   7638 
   7639 ;; -------------------------------------------------------------------------
   7640 ;; ---- [INT,FP] Compare and select
   7641 ;; -------------------------------------------------------------------------
   7642 ;; The patterns in this section are synthetic.
   7643 ;; -------------------------------------------------------------------------
   7644 
   7645 ;; Integer (signed) vcond.  Don't enforce an immediate range here, since it
   7646 ;; depends on the comparison; leave it to aarch64_expand_sve_vcond instead.
   7647 (define_expand "vcond<SVE_ALL:mode><SVE_I:mode>"
   7648   [(set (match_operand:SVE_ALL 0 "register_operand")
   7649 	(if_then_else:SVE_ALL
   7650 	  (match_operator 3 "comparison_operator"
   7651 	    [(match_operand:SVE_I 4 "register_operand")
   7652 	     (match_operand:SVE_I 5 "nonmemory_operand")])
   7653 	  (match_operand:SVE_ALL 1 "nonmemory_operand")
   7654 	  (match_operand:SVE_ALL 2 "nonmemory_operand")))]
   7655   "TARGET_SVE && <SVE_ALL:container_bits> == <SVE_I:container_bits>"
   7656   {
   7657     aarch64_expand_sve_vcond (<SVE_ALL:MODE>mode, <SVE_I:MODE>mode, operands);
   7658     DONE;
   7659   }
   7660 )
   7661 
   7662 ;; Integer vcondu.  Don't enforce an immediate range here, since it
   7663 ;; depends on the comparison; leave it to aarch64_expand_sve_vcond instead.
   7664 (define_expand "vcondu<SVE_ALL:mode><SVE_I:mode>"
   7665   [(set (match_operand:SVE_ALL 0 "register_operand")
   7666 	(if_then_else:SVE_ALL
   7667 	  (match_operator 3 "comparison_operator"
   7668 	    [(match_operand:SVE_I 4 "register_operand")
   7669 	     (match_operand:SVE_I 5 "nonmemory_operand")])
   7670 	  (match_operand:SVE_ALL 1 "nonmemory_operand")
   7671 	  (match_operand:SVE_ALL 2 "nonmemory_operand")))]
   7672   "TARGET_SVE && <SVE_ALL:container_bits> == <SVE_I:container_bits>"
   7673   {
   7674     aarch64_expand_sve_vcond (<SVE_ALL:MODE>mode, <SVE_I:MODE>mode, operands);
   7675     DONE;
   7676   }
   7677 )
   7678 
   7679 ;; Floating-point vcond.  All comparisons except FCMUO allow a zero operand;
   7680 ;; aarch64_expand_sve_vcond handles the case of an FCMUO with zero.
   7681 (define_expand "vcond<mode><v_fp_equiv>"
   7682   [(set (match_operand:SVE_FULL_HSD 0 "register_operand")
   7683 	(if_then_else:SVE_FULL_HSD
   7684 	  (match_operator 3 "comparison_operator"
   7685 	    [(match_operand:<V_FP_EQUIV> 4 "register_operand")
   7686 	     (match_operand:<V_FP_EQUIV> 5 "aarch64_simd_reg_or_zero")])
   7687 	  (match_operand:SVE_FULL_HSD 1 "nonmemory_operand")
   7688 	  (match_operand:SVE_FULL_HSD 2 "nonmemory_operand")))]
   7689   "TARGET_SVE"
   7690   {
   7691     aarch64_expand_sve_vcond (<MODE>mode, <V_FP_EQUIV>mode, operands);
   7692     DONE;
   7693   }
   7694 )
   7695 
   7696 ;; -------------------------------------------------------------------------
   7697 ;; ---- [INT] Comparisons
   7698 ;; -------------------------------------------------------------------------
   7699 ;; Includes:
   7700 ;; - CMPEQ
   7701 ;; - CMPGE
   7702 ;; - CMPGT
   7703 ;; - CMPHI
   7704 ;; - CMPHS
   7705 ;; - CMPLE
   7706 ;; - CMPLO
   7707 ;; - CMPLS
   7708 ;; - CMPLT
   7709 ;; - CMPNE
   7710 ;; -------------------------------------------------------------------------
   7711 
   7712 ;; Signed integer comparisons.  Don't enforce an immediate range here, since
   7713 ;; it depends on the comparison; leave it to aarch64_expand_sve_vec_cmp_int
   7714 ;; instead.
   7715 (define_expand "vec_cmp<mode><vpred>"
   7716   [(parallel
   7717     [(set (match_operand:<VPRED> 0 "register_operand")
   7718 	  (match_operator:<VPRED> 1 "comparison_operator"
   7719 	    [(match_operand:SVE_I 2 "register_operand")
   7720 	     (match_operand:SVE_I 3 "nonmemory_operand")]))
   7721      (clobber (reg:CC_NZC CC_REGNUM))])]
   7722   "TARGET_SVE"
   7723   {
   7724     aarch64_expand_sve_vec_cmp_int (operands[0], GET_CODE (operands[1]),
   7725 				    operands[2], operands[3]);
   7726     DONE;
   7727   }
   7728 )
   7729 
   7730 ;; Unsigned integer comparisons.  Don't enforce an immediate range here, since
   7731 ;; it depends on the comparison; leave it to aarch64_expand_sve_vec_cmp_int
   7732 ;; instead.
   7733 (define_expand "vec_cmpu<mode><vpred>"
   7734   [(parallel
   7735     [(set (match_operand:<VPRED> 0 "register_operand")
   7736 	  (match_operator:<VPRED> 1 "comparison_operator"
   7737 	    [(match_operand:SVE_I 2 "register_operand")
   7738 	     (match_operand:SVE_I 3 "nonmemory_operand")]))
   7739      (clobber (reg:CC_NZC CC_REGNUM))])]
   7740   "TARGET_SVE"
   7741   {
   7742     aarch64_expand_sve_vec_cmp_int (operands[0], GET_CODE (operands[1]),
   7743 				    operands[2], operands[3]);
   7744     DONE;
   7745   }
   7746 )
   7747 
   7748 ;; Predicated integer comparisons.
   7749 ;;
   7750 ;; For unpacked vectors, only the lowpart element in each input container
   7751 ;; has a defined value, and only the predicate bits associated with
   7752 ;; those elements are defined.  For example, when comparing two VNx2SIs:
   7753 ;;
   7754 ;; - The VNx2SIs can be seem as VNx2DIs in which the low halves of each
   7755 ;;   DI container store an SI element.  The upper bits of each DI container
   7756 ;;   are undefined.
   7757 ;;
   7758 ;; - Alternatively, the VNx2SIs can be seen as VNx4SIs in which the
   7759 ;;   even elements are defined and the odd elements are undefined.
   7760 ;;
   7761 ;; - The associated predicate mode is VNx2BI.  This means that only the
   7762 ;;   low bit in each predicate byte is defined (on input and on output).
   7763 ;;
   7764 ;; - We use a .s comparison to compare VNx2SIs, under the control of a
   7765 ;;   VNx2BI governing predicate, to produce a VNx2BI result.  If we view
   7766 ;;   the .s operation as operating on VNx4SIs then for odd lanes:
   7767 ;;
   7768 ;;   - the input governing predicate bit is undefined
   7769 ;;   - the SI elements being compared are undefined
   7770 ;;   - the predicate result bit is therefore undefined, but
   7771 ;;   - the predicate result bit is in the undefined part of a VNx2BI,
   7772 ;;     so its value doesn't matter anyway.
   7773 (define_insn "@aarch64_pred_cmp<cmp_op><mode>"
   7774   [(set (match_operand:<VPRED> 0 "register_operand" "=Upa, Upa")
   7775 	(unspec:<VPRED>
   7776 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   7777 	   (match_operand:SI 2 "aarch64_sve_ptrue_flag")
   7778 	   (SVE_INT_CMP:<VPRED>
   7779 	     (match_operand:SVE_I 3 "register_operand" "w, w")
   7780 	     (match_operand:SVE_I 4 "aarch64_sve_cmp_<sve_imm_con>_operand" "<sve_imm_con>, w"))]
   7781 	  UNSPEC_PRED_Z))
   7782    (clobber (reg:CC_NZC CC_REGNUM))]
   7783   "TARGET_SVE"
   7784   "@
   7785    cmp<cmp_op>\t%0.<Vetype>, %1/z, %3.<Vetype>, #%4
   7786    cmp<cmp_op>\t%0.<Vetype>, %1/z, %3.<Vetype>, %4.<Vetype>"
   7787 )
   7788 
   7789 ;; Predicated integer comparisons in which both the flag and predicate
   7790 ;; results are interesting.
   7791 (define_insn_and_rewrite "*cmp<cmp_op><mode>_cc"
   7792   [(set (reg:CC_NZC CC_REGNUM)
   7793 	(unspec:CC_NZC
   7794 	  [(match_operand:VNx16BI 1 "register_operand" "Upl, Upl")
   7795 	   (match_operand 4)
   7796 	   (match_operand:SI 5 "aarch64_sve_ptrue_flag")
   7797 	   (unspec:<VPRED>
   7798 	     [(match_operand 6)
   7799 	      (match_operand:SI 7 "aarch64_sve_ptrue_flag")
   7800 	      (SVE_INT_CMP:<VPRED>
   7801 		(match_operand:SVE_I 2 "register_operand" "w, w")
   7802 		(match_operand:SVE_I 3 "aarch64_sve_cmp_<sve_imm_con>_operand" "<sve_imm_con>, w"))]
   7803 	     UNSPEC_PRED_Z)]
   7804 	  UNSPEC_PTEST))
   7805    (set (match_operand:<VPRED> 0 "register_operand" "=Upa, Upa")
   7806 	(unspec:<VPRED>
   7807 	  [(match_dup 6)
   7808 	   (match_dup 7)
   7809 	   (SVE_INT_CMP:<VPRED>
   7810 	     (match_dup 2)
   7811 	     (match_dup 3))]
   7812 	  UNSPEC_PRED_Z))]
   7813   "TARGET_SVE
   7814    && aarch64_sve_same_pred_for_ptest_p (&operands[4], &operands[6])"
   7815   "@
   7816    cmp<cmp_op>\t%0.<Vetype>, %1/z, %2.<Vetype>, #%3
   7817    cmp<cmp_op>\t%0.<Vetype>, %1/z, %2.<Vetype>, %3.<Vetype>"
   7818   "&& !rtx_equal_p (operands[4], operands[6])"
   7819   {
   7820     operands[6] = copy_rtx (operands[4]);
   7821     operands[7] = operands[5];
   7822   }
   7823 )
   7824 
   7825 ;; Predicated integer comparisons in which only the flags result is
   7826 ;; interesting.
   7827 (define_insn_and_rewrite "*cmp<cmp_op><mode>_ptest"
   7828   [(set (reg:CC_NZC CC_REGNUM)
   7829 	(unspec:CC_NZC
   7830 	  [(match_operand:VNx16BI 1 "register_operand" "Upl, Upl")
   7831 	   (match_operand 4)
   7832 	   (match_operand:SI 5 "aarch64_sve_ptrue_flag")
   7833 	   (unspec:<VPRED>
   7834 	     [(match_operand 6)
   7835 	      (match_operand:SI 7 "aarch64_sve_ptrue_flag")
   7836 	      (SVE_INT_CMP:<VPRED>
   7837 		(match_operand:SVE_I 2 "register_operand" "w, w")
   7838 		(match_operand:SVE_I 3 "aarch64_sve_cmp_<sve_imm_con>_operand" "<sve_imm_con>, w"))]
   7839 	     UNSPEC_PRED_Z)]
   7840 	  UNSPEC_PTEST))
   7841    (clobber (match_scratch:<VPRED> 0 "=Upa, Upa"))]
   7842   "TARGET_SVE
   7843    && aarch64_sve_same_pred_for_ptest_p (&operands[4], &operands[6])"
   7844   "@
   7845    cmp<cmp_op>\t%0.<Vetype>, %1/z, %2.<Vetype>, #%3
   7846    cmp<cmp_op>\t%0.<Vetype>, %1/z, %2.<Vetype>, %3.<Vetype>"
   7847   "&& !rtx_equal_p (operands[4], operands[6])"
   7848   {
   7849     operands[6] = copy_rtx (operands[4]);
   7850     operands[7] = operands[5];
   7851   }
   7852 )
   7853 
   7854 ;; Predicated integer comparisons, formed by combining a PTRUE-predicated
   7855 ;; comparison with an AND.  Split the instruction into its preferred form
   7856 ;; at the earliest opportunity, in order to get rid of the redundant
   7857 ;; operand 4.
   7858 (define_insn_and_split "*cmp<cmp_op><mode>_and"
   7859   [(set (match_operand:<VPRED> 0 "register_operand" "=Upa, Upa")
   7860 	(and:<VPRED>
   7861 	  (unspec:<VPRED>
   7862 	    [(match_operand 4)
   7863 	     (const_int SVE_KNOWN_PTRUE)
   7864 	     (SVE_INT_CMP:<VPRED>
   7865 	       (match_operand:SVE_I 2 "register_operand" "w, w")
   7866 	       (match_operand:SVE_I 3 "aarch64_sve_cmp_<sve_imm_con>_operand" "<sve_imm_con>, w"))]
   7867 	    UNSPEC_PRED_Z)
   7868 	  (match_operand:<VPRED> 1 "register_operand" "Upl, Upl")))
   7869    (clobber (reg:CC_NZC CC_REGNUM))]
   7870   "TARGET_SVE"
   7871   "#"
   7872   "&& 1"
   7873   [(parallel
   7874      [(set (match_dup 0)
   7875 	   (unspec:<VPRED>
   7876 	     [(match_dup 1)
   7877 	      (const_int SVE_MAYBE_NOT_PTRUE)
   7878 	      (SVE_INT_CMP:<VPRED>
   7879 		(match_dup 2)
   7880 		(match_dup 3))]
   7881 	     UNSPEC_PRED_Z))
   7882       (clobber (reg:CC_NZC CC_REGNUM))])]
   7883 )
   7884 
   7885 ;; Predicated integer wide comparisons.
   7886 (define_insn "@aarch64_pred_cmp<cmp_op><mode>_wide"
   7887   [(set (match_operand:<VPRED> 0 "register_operand" "=Upa")
   7888 	(unspec:<VPRED>
   7889 	  [(match_operand:VNx16BI 1 "register_operand" "Upl")
   7890 	   (match_operand:SI 2 "aarch64_sve_ptrue_flag")
   7891 	   (unspec:<VPRED>
   7892 	     [(match_operand:SVE_FULL_BHSI 3 "register_operand" "w")
   7893 	      (match_operand:VNx2DI 4 "register_operand" "w")]
   7894 	     SVE_COND_INT_CMP_WIDE)]
   7895 	  UNSPEC_PRED_Z))
   7896    (clobber (reg:CC_NZC CC_REGNUM))]
   7897   "TARGET_SVE"
   7898   "cmp<cmp_op>\t%0.<Vetype>, %1/z, %3.<Vetype>, %4.d"
   7899 )
   7900 
   7901 ;; Predicated integer wide comparisons in which both the flag and
   7902 ;; predicate results are interesting.
   7903 (define_insn "*aarch64_pred_cmp<cmp_op><mode>_wide_cc"
   7904   [(set (reg:CC_NZC CC_REGNUM)
   7905 	(unspec:CC_NZC
   7906 	  [(match_operand:VNx16BI 1 "register_operand" "Upl")
   7907 	   (match_operand 4)
   7908 	   (match_operand:SI 5 "aarch64_sve_ptrue_flag")
   7909 	   (unspec:<VPRED>
   7910 	     [(match_operand:VNx16BI 6 "register_operand" "Upl")
   7911 	      (match_operand:SI 7 "aarch64_sve_ptrue_flag")
   7912 	      (unspec:<VPRED>
   7913 		[(match_operand:SVE_FULL_BHSI 2 "register_operand" "w")
   7914 		 (match_operand:VNx2DI 3 "register_operand" "w")]
   7915 		SVE_COND_INT_CMP_WIDE)]
   7916 	     UNSPEC_PRED_Z)]
   7917 	  UNSPEC_PTEST))
   7918    (set (match_operand:<VPRED> 0 "register_operand" "=Upa")
   7919 	(unspec:<VPRED>
   7920 	  [(match_dup 6)
   7921 	   (match_dup 7)
   7922 	   (unspec:<VPRED>
   7923 	     [(match_dup 2)
   7924 	      (match_dup 3)]
   7925 	     SVE_COND_INT_CMP_WIDE)]
   7926 	  UNSPEC_PRED_Z))]
   7927   "TARGET_SVE
   7928    && aarch64_sve_same_pred_for_ptest_p (&operands[4], &operands[6])"
   7929   "cmp<cmp_op>\t%0.<Vetype>, %1/z, %2.<Vetype>, %3.d"
   7930 )
   7931 
   7932 ;; Predicated integer wide comparisons in which only the flags result
   7933 ;; is interesting.
   7934 (define_insn "*aarch64_pred_cmp<cmp_op><mode>_wide_ptest"
   7935   [(set (reg:CC_NZC CC_REGNUM)
   7936 	(unspec:CC_NZC
   7937 	  [(match_operand:VNx16BI 1 "register_operand" "Upl")
   7938 	   (match_operand 4)
   7939 	   (match_operand:SI 5 "aarch64_sve_ptrue_flag")
   7940 	   (unspec:<VPRED>
   7941 	     [(match_operand:VNx16BI 6 "register_operand" "Upl")
   7942 	      (match_operand:SI 7 "aarch64_sve_ptrue_flag")
   7943 	      (unspec:<VPRED>
   7944 		[(match_operand:SVE_FULL_BHSI 2 "register_operand" "w")
   7945 		 (match_operand:VNx2DI 3 "register_operand" "w")]
   7946 		SVE_COND_INT_CMP_WIDE)]
   7947 	     UNSPEC_PRED_Z)]
   7948 	  UNSPEC_PTEST))
   7949    (clobber (match_scratch:<VPRED> 0 "=Upa"))]
   7950   "TARGET_SVE
   7951    && aarch64_sve_same_pred_for_ptest_p (&operands[4], &operands[6])"
   7952   "cmp<cmp_op>\t%0.<Vetype>, %1/z, %2.<Vetype>, %3.d"
   7953 )
   7954 
   7955 ;; -------------------------------------------------------------------------
   7956 ;; ---- [INT] While tests
   7957 ;; -------------------------------------------------------------------------
   7958 ;; Includes:
   7959 ;; - WHILEGE (SVE2)
   7960 ;; - WHILEGT (SVE2)
   7961 ;; - WHILEHI (SVE2)
   7962 ;; - WHILEHS (SVE2)
   7963 ;; - WHILELE
   7964 ;; - WHILELO
   7965 ;; - WHILELS
   7966 ;; - WHILELT
   7967 ;; - WHILERW (SVE2)
   7968 ;; - WHILEWR (SVE2)
   7969 ;; -------------------------------------------------------------------------
   7970 
   7971 ;; Set element I of the result if (cmp (plus operand1 J) operand2) is
   7972 ;; true for all J in [0, I].
   7973 (define_insn "@while_<while_optab_cmp><GPI:mode><PRED_ALL:mode>"
   7974   [(set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   7975 	(unspec:PRED_ALL [(match_operand:GPI 1 "aarch64_reg_or_zero" "rZ")
   7976 			  (match_operand:GPI 2 "aarch64_reg_or_zero" "rZ")]
   7977 			 SVE_WHILE))
   7978    (clobber (reg:CC_NZC CC_REGNUM))]
   7979   "TARGET_SVE"
   7980   "while<cmp_op>\t%0.<PRED_ALL:Vetype>, %<w>1, %<w>2"
   7981 )
   7982 
   7983 ;; The WHILE instructions set the flags in the same way as a PTEST with
   7984 ;; a PTRUE GP.  Handle the case in which both results are useful.  The GP
   7985 ;; operands to the PTEST aren't needed, so we allow them to be anything.
   7986 (define_insn_and_rewrite "*while_<while_optab_cmp><GPI:mode><PRED_ALL:mode>_cc"
   7987   [(set (reg:CC_NZC CC_REGNUM)
   7988 	(unspec:CC_NZC
   7989 	  [(match_operand 3)
   7990 	   (match_operand 4)
   7991 	   (const_int SVE_KNOWN_PTRUE)
   7992 	   (unspec:PRED_ALL
   7993 	     [(match_operand:GPI 1 "aarch64_reg_or_zero" "rZ")
   7994 	      (match_operand:GPI 2 "aarch64_reg_or_zero" "rZ")]
   7995 	     SVE_WHILE)]
   7996 	  UNSPEC_PTEST))
   7997    (set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   7998 	(unspec:PRED_ALL [(match_dup 1)
   7999 			  (match_dup 2)]
   8000 			 SVE_WHILE))]
   8001   "TARGET_SVE"
   8002   "while<cmp_op>\t%0.<PRED_ALL:Vetype>, %<w>1, %<w>2"
   8003   ;; Force the compiler to drop the unused predicate operand, so that we
   8004   ;; don't have an unnecessary PTRUE.
   8005   "&& (!CONSTANT_P (operands[3]) || !CONSTANT_P (operands[4]))"
   8006   {
   8007     operands[3] = CONSTM1_RTX (VNx16BImode);
   8008     operands[4] = CONSTM1_RTX (<PRED_ALL:MODE>mode);
   8009   }
   8010 )
   8011 
   8012 ;; Same, but handle the case in which only the flags result is useful.
   8013 (define_insn_and_rewrite "@while_<while_optab_cmp><GPI:mode><PRED_ALL:mode>_ptest"
   8014   [(set (reg:CC_NZC CC_REGNUM)
   8015 	(unspec:CC_NZC
   8016 	  [(match_operand 3)
   8017 	   (match_operand 4)
   8018 	   (const_int SVE_KNOWN_PTRUE)
   8019 	   (unspec:PRED_ALL
   8020 	     [(match_operand:GPI 1 "aarch64_reg_or_zero" "rZ")
   8021 	      (match_operand:GPI 2 "aarch64_reg_or_zero" "rZ")]
   8022 	     SVE_WHILE)]
   8023 	  UNSPEC_PTEST))
   8024    (clobber (match_scratch:PRED_ALL 0 "=Upa"))]
   8025   "TARGET_SVE"
   8026   "while<cmp_op>\t%0.<PRED_ALL:Vetype>, %<w>1, %<w>2"
   8027   ;; Force the compiler to drop the unused predicate operand, so that we
   8028   ;; don't have an unnecessary PTRUE.
   8029   "&& (!CONSTANT_P (operands[3]) || !CONSTANT_P (operands[4]))"
   8030   {
   8031     operands[3] = CONSTM1_RTX (VNx16BImode);
   8032     operands[4] = CONSTM1_RTX (<PRED_ALL:MODE>mode);
   8033   }
   8034 )
   8035 
   8036 ;; -------------------------------------------------------------------------
   8037 ;; ---- [FP] Direct comparisons
   8038 ;; -------------------------------------------------------------------------
   8039 ;; Includes:
   8040 ;; - FCMEQ
   8041 ;; - FCMGE
   8042 ;; - FCMGT
   8043 ;; - FCMLE
   8044 ;; - FCMLT
   8045 ;; - FCMNE
   8046 ;; - FCMUO
   8047 ;; -------------------------------------------------------------------------
   8048 
   8049 ;; Floating-point comparisons.  All comparisons except FCMUO allow a zero
   8050 ;; operand; aarch64_expand_sve_vec_cmp_float handles the case of an FCMUO
   8051 ;; with zero.
   8052 (define_expand "vec_cmp<mode><vpred>"
   8053   [(set (match_operand:<VPRED> 0 "register_operand")
   8054 	(match_operator:<VPRED> 1 "comparison_operator"
   8055 	  [(match_operand:SVE_FULL_F 2 "register_operand")
   8056 	   (match_operand:SVE_FULL_F 3 "aarch64_simd_reg_or_zero")]))]
   8057   "TARGET_SVE"
   8058   {
   8059     aarch64_expand_sve_vec_cmp_float (operands[0], GET_CODE (operands[1]),
   8060 				      operands[2], operands[3], false);
   8061     DONE;
   8062   }
   8063 )
   8064 
   8065 ;; Predicated floating-point comparisons.
   8066 (define_insn "@aarch64_pred_fcm<cmp_op><mode>"
   8067   [(set (match_operand:<VPRED> 0 "register_operand" "=Upa, Upa")
   8068 	(unspec:<VPRED>
   8069 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   8070 	   (match_operand:SI 2 "aarch64_sve_ptrue_flag")
   8071 	   (match_operand:SVE_FULL_F 3 "register_operand" "w, w")
   8072 	   (match_operand:SVE_FULL_F 4 "aarch64_simd_reg_or_zero" "Dz, w")]
   8073 	  SVE_COND_FP_CMP_I0))]
   8074   "TARGET_SVE"
   8075   "@
   8076    fcm<cmp_op>\t%0.<Vetype>, %1/z, %3.<Vetype>, #0.0
   8077    fcm<cmp_op>\t%0.<Vetype>, %1/z, %3.<Vetype>, %4.<Vetype>"
   8078 )
   8079 
   8080 ;; Same for unordered comparisons.
   8081 (define_insn "@aarch64_pred_fcmuo<mode>"
   8082   [(set (match_operand:<VPRED> 0 "register_operand" "=Upa")
   8083 	(unspec:<VPRED>
   8084 	  [(match_operand:<VPRED> 1 "register_operand" "Upl")
   8085 	   (match_operand:SI 2 "aarch64_sve_ptrue_flag")
   8086 	   (match_operand:SVE_FULL_F 3 "register_operand" "w")
   8087 	   (match_operand:SVE_FULL_F 4 "register_operand" "w")]
   8088 	  UNSPEC_COND_FCMUO))]
   8089   "TARGET_SVE"
   8090   "fcmuo\t%0.<Vetype>, %1/z, %3.<Vetype>, %4.<Vetype>"
   8091 )
   8092 
   8093 ;; Floating-point comparisons predicated on a PTRUE, with the results ANDed
   8094 ;; with another predicate P.  This does not have the same trapping behavior
   8095 ;; as predicating the comparison itself on P, but it's a legitimate fold,
   8096 ;; since we can drop any potentially-trapping operations whose results
   8097 ;; are not needed.
   8098 ;;
   8099 ;; Split the instruction into its preferred form (below) at the earliest
   8100 ;; opportunity, in order to get rid of the redundant operand 1.
   8101 (define_insn_and_split "*fcm<cmp_op><mode>_and_combine"
   8102   [(set (match_operand:<VPRED> 0 "register_operand" "=Upa, Upa")
   8103 	(and:<VPRED>
   8104 	  (unspec:<VPRED>
   8105 	    [(match_operand:<VPRED> 1)
   8106 	     (const_int SVE_KNOWN_PTRUE)
   8107 	     (match_operand:SVE_FULL_F 2 "register_operand" "w, w")
   8108 	     (match_operand:SVE_FULL_F 3 "aarch64_simd_reg_or_zero" "Dz, w")]
   8109 	    SVE_COND_FP_CMP_I0)
   8110 	  (match_operand:<VPRED> 4 "register_operand" "Upl, Upl")))]
   8111   "TARGET_SVE"
   8112   "#"
   8113   "&& 1"
   8114   [(set (match_dup 0)
   8115 	(unspec:<VPRED>
   8116 	  [(match_dup 4)
   8117 	   (const_int SVE_MAYBE_NOT_PTRUE)
   8118 	   (match_dup 2)
   8119 	   (match_dup 3)]
   8120 	  SVE_COND_FP_CMP_I0))]
   8121 )
   8122 
   8123 ;; Same for unordered comparisons.
   8124 (define_insn_and_split "*fcmuo<mode>_and_combine"
   8125   [(set (match_operand:<VPRED> 0 "register_operand" "=Upa")
   8126 	(and:<VPRED>
   8127 	  (unspec:<VPRED>
   8128 	    [(match_operand:<VPRED> 1)
   8129 	     (const_int SVE_KNOWN_PTRUE)
   8130 	     (match_operand:SVE_FULL_F 2 "register_operand" "w")
   8131 	     (match_operand:SVE_FULL_F 3 "register_operand" "w")]
   8132 	    UNSPEC_COND_FCMUO)
   8133 	  (match_operand:<VPRED> 4 "register_operand" "Upl")))]
   8134   "TARGET_SVE"
   8135   "#"
   8136   "&& 1"
   8137   [(set (match_dup 0)
   8138 	(unspec:<VPRED>
   8139 	  [(match_dup 4)
   8140 	   (const_int SVE_MAYBE_NOT_PTRUE)
   8141 	   (match_dup 2)
   8142 	   (match_dup 3)]
   8143 	  UNSPEC_COND_FCMUO))]
   8144 )
   8145 
   8146 ;; Similar to *fcm<cmp_op><mode>_and_combine, but for BIC rather than AND.
   8147 ;; In this case, we still need a separate NOT/BIC operation, but predicating
   8148 ;; the comparison on the BIC operand removes the need for a PTRUE.
   8149 (define_insn_and_split "*fcm<cmp_op><mode>_bic_combine"
   8150   [(set (match_operand:<VPRED> 0 "register_operand" "=Upa")
   8151 	(and:<VPRED>
   8152 	  (and:<VPRED>
   8153 	    (not:<VPRED>
   8154 	      (unspec:<VPRED>
   8155 	        [(match_operand:<VPRED> 1)
   8156 	         (const_int SVE_KNOWN_PTRUE)
   8157 	         (match_operand:SVE_FULL_F 2 "register_operand" "w")
   8158 	         (match_operand:SVE_FULL_F 3 "aarch64_simd_reg_or_zero" "wDz")]
   8159 	        SVE_COND_FP_CMP_I0))
   8160 	    (match_operand:<VPRED> 4 "register_operand" "Upa"))
   8161 	  (match_dup:<VPRED> 1)))
   8162    (clobber (match_scratch:<VPRED> 5 "=&Upl"))]
   8163   "TARGET_SVE"
   8164   "#"
   8165   "&& 1"
   8166   [(set (match_dup 5)
   8167 	(unspec:<VPRED>
   8168 	  [(match_dup 4)
   8169 	   (const_int SVE_MAYBE_NOT_PTRUE)
   8170 	   (match_dup 2)
   8171 	   (match_dup 3)]
   8172 	  SVE_COND_FP_CMP_I0))
   8173    (set (match_dup 0)
   8174 	(and:<VPRED>
   8175 	  (not:<VPRED>
   8176 	    (match_dup 5))
   8177 	  (match_dup 4)))]
   8178 {
   8179   if (can_create_pseudo_p ())
   8180     operands[5] = gen_reg_rtx (<VPRED>mode);
   8181 }
   8182 )
   8183 
   8184 ;; Make sure that we expand to a nor when the operand 4 of
   8185 ;; *fcm<cmp_op><mode>_bic_combine is a not.
   8186 (define_insn_and_split "*fcm<cmp_op><mode>_nor_combine"
   8187   [(set (match_operand:<VPRED> 0 "register_operand" "=Upa")
   8188 	(and:<VPRED>
   8189 	  (and:<VPRED>
   8190 	    (not:<VPRED>
   8191 	      (unspec:<VPRED>
   8192 	        [(match_operand:<VPRED> 1)
   8193 	         (const_int SVE_KNOWN_PTRUE)
   8194 	         (match_operand:SVE_FULL_F 2 "register_operand" "w")
   8195 	         (match_operand:SVE_FULL_F 3 "aarch64_simd_reg_or_zero" "wDz")]
   8196 	        SVE_COND_FP_CMP_I0))
   8197 	    (not:<VPRED>
   8198 	      (match_operand:<VPRED> 4 "register_operand" "Upa")))
   8199 	  (match_dup:<VPRED> 1)))
   8200    (clobber (match_scratch:<VPRED> 5 "=&Upl"))]
   8201   "TARGET_SVE"
   8202   "#"
   8203   "&& 1"
   8204   [(set (match_dup 5)
   8205 	(unspec:<VPRED>
   8206 	  [(match_dup 1)
   8207 	   (const_int SVE_KNOWN_PTRUE)
   8208 	   (match_dup 2)
   8209 	   (match_dup 3)]
   8210 	  SVE_COND_FP_CMP_I0))
   8211    (set (match_dup 0)
   8212 	(and:<VPRED>
   8213 	  (and:<VPRED>
   8214 	    (not:<VPRED>
   8215 	      (match_dup 5))
   8216 	    (not:<VPRED>
   8217 	      (match_dup 4)))
   8218 	  (match_dup 1)))]
   8219 {
   8220   if (can_create_pseudo_p ())
   8221     operands[5] = gen_reg_rtx (<VPRED>mode);
   8222 }
   8223 )
   8224 
   8225 (define_insn_and_split "*fcmuo<mode>_bic_combine"
   8226   [(set (match_operand:<VPRED> 0 "register_operand" "=Upa")
   8227 	(and:<VPRED>
   8228 	  (and:<VPRED>
   8229 	    (not:<VPRED>
   8230 	      (unspec:<VPRED>
   8231 	        [(match_operand:<VPRED> 1)
   8232 	         (const_int SVE_KNOWN_PTRUE)
   8233 	         (match_operand:SVE_FULL_F 2 "register_operand" "w")
   8234 	         (match_operand:SVE_FULL_F 3 "register_operand" "w")]
   8235 	        UNSPEC_COND_FCMUO))
   8236 	    (match_operand:<VPRED> 4 "register_operand" "Upa"))
   8237 	  (match_dup:<VPRED> 1)))
   8238    (clobber (match_scratch:<VPRED> 5 "=&Upl"))]
   8239   "TARGET_SVE"
   8240   "#"
   8241   "&& 1"
   8242   [(set (match_dup 5)
   8243 	(unspec:<VPRED>
   8244 	  [(match_dup 4)
   8245 	   (const_int SVE_MAYBE_NOT_PTRUE)
   8246 	   (match_dup 2)
   8247 	   (match_dup 3)]
   8248 	  UNSPEC_COND_FCMUO))
   8249    (set (match_dup 0)
   8250 	(and:<VPRED>
   8251 	  (not:<VPRED>
   8252 	    (match_dup 5))
   8253 	  (match_dup 4)))]
   8254 {
   8255   if (can_create_pseudo_p ())
   8256     operands[5] = gen_reg_rtx (<VPRED>mode);
   8257 }
   8258 )
   8259 
   8260 ;; Same for unordered comparisons.
   8261 (define_insn_and_split "*fcmuo<mode>_nor_combine"
   8262   [(set (match_operand:<VPRED> 0 "register_operand" "=Upa")
   8263 	(and:<VPRED>
   8264 	  (and:<VPRED>
   8265 	    (not:<VPRED>
   8266 	      (unspec:<VPRED>
   8267 	        [(match_operand:<VPRED> 1)
   8268 	         (const_int SVE_KNOWN_PTRUE)
   8269 	         (match_operand:SVE_FULL_F 2 "register_operand" "w")
   8270 	         (match_operand:SVE_FULL_F 3 "register_operand" "w")]
   8271 	        UNSPEC_COND_FCMUO))
   8272 	    (not:<VPRED>
   8273 	      (match_operand:<VPRED> 4 "register_operand" "Upa")))
   8274 	  (match_dup:<VPRED> 1)))
   8275    (clobber (match_scratch:<VPRED> 5 "=&Upl"))]
   8276   "TARGET_SVE"
   8277   "#"
   8278   "&& 1"
   8279   [(set (match_dup 5)
   8280 	(unspec:<VPRED>
   8281 	  [(match_dup 1)
   8282 	   (const_int SVE_KNOWN_PTRUE)
   8283 	   (match_dup 2)
   8284 	   (match_dup 3)]
   8285 	  UNSPEC_COND_FCMUO))
   8286    (set (match_dup 0)
   8287 	(and:<VPRED>
   8288 	  (and:<VPRED>
   8289 	    (not:<VPRED>
   8290 	      (match_dup 5))
   8291 	    (not:<VPRED>
   8292 	      (match_dup 4)))
   8293 	  (match_dup 1)))]
   8294 {
   8295   if (can_create_pseudo_p ())
   8296     operands[5] = gen_reg_rtx (<VPRED>mode);
   8297 }
   8298 )
   8299 
   8300 ;; -------------------------------------------------------------------------
   8301 ;; ---- [FP] Absolute comparisons
   8302 ;; -------------------------------------------------------------------------
   8303 ;; Includes:
   8304 ;; - FACGE
   8305 ;; - FACGT
   8306 ;; - FACLE
   8307 ;; - FACLT
   8308 ;; -------------------------------------------------------------------------
   8309 
   8310 ;; Predicated floating-point absolute comparisons.
   8311 (define_expand "@aarch64_pred_fac<cmp_op><mode>"
   8312   [(set (match_operand:<VPRED> 0 "register_operand")
   8313 	(unspec:<VPRED>
   8314 	  [(match_operand:<VPRED> 1 "register_operand")
   8315 	   (match_operand:SI 2 "aarch64_sve_ptrue_flag")
   8316 	   (unspec:SVE_FULL_F
   8317 	     [(match_dup 1)
   8318 	      (match_dup 2)
   8319 	      (match_operand:SVE_FULL_F 3 "register_operand")]
   8320 	     UNSPEC_COND_FABS)
   8321 	   (unspec:SVE_FULL_F
   8322 	     [(match_dup 1)
   8323 	      (match_dup 2)
   8324 	      (match_operand:SVE_FULL_F 4 "register_operand")]
   8325 	     UNSPEC_COND_FABS)]
   8326 	  SVE_COND_FP_ABS_CMP))]
   8327   "TARGET_SVE"
   8328 )
   8329 
   8330 (define_insn_and_rewrite "*aarch64_pred_fac<cmp_op><mode>_relaxed"
   8331   [(set (match_operand:<VPRED> 0 "register_operand" "=Upa")
   8332 	(unspec:<VPRED>
   8333 	  [(match_operand:<VPRED> 1 "register_operand" "Upl")
   8334 	   (match_operand:SI 4 "aarch64_sve_ptrue_flag")
   8335 	   (unspec:SVE_FULL_F
   8336 	     [(match_operand 5)
   8337 	      (const_int SVE_RELAXED_GP)
   8338 	      (match_operand:SVE_FULL_F 2 "register_operand" "w")]
   8339 	     UNSPEC_COND_FABS)
   8340 	   (unspec:SVE_FULL_F
   8341 	     [(match_operand 6)
   8342 	      (const_int SVE_RELAXED_GP)
   8343 	      (match_operand:SVE_FULL_F 3 "register_operand" "w")]
   8344 	     UNSPEC_COND_FABS)]
   8345 	  SVE_COND_FP_ABS_CMP))]
   8346   "TARGET_SVE"
   8347   "fac<cmp_op>\t%0.<Vetype>, %1/z, %2.<Vetype>, %3.<Vetype>"
   8348   "&& (!rtx_equal_p (operands[1], operands[5])
   8349        || !rtx_equal_p (operands[1], operands[6]))"
   8350   {
   8351     operands[5] = copy_rtx (operands[1]);
   8352     operands[6] = copy_rtx (operands[1]);
   8353   }
   8354 )
   8355 
   8356 (define_insn "*aarch64_pred_fac<cmp_op><mode>_strict"
   8357   [(set (match_operand:<VPRED> 0 "register_operand" "=Upa")
   8358 	(unspec:<VPRED>
   8359 	  [(match_operand:<VPRED> 1 "register_operand" "Upl")
   8360 	   (match_operand:SI 4 "aarch64_sve_ptrue_flag")
   8361 	   (unspec:SVE_FULL_F
   8362 	     [(match_dup 1)
   8363 	      (match_operand:SI 5 "aarch64_sve_gp_strictness")
   8364 	      (match_operand:SVE_FULL_F 2 "register_operand" "w")]
   8365 	     UNSPEC_COND_FABS)
   8366 	   (unspec:SVE_FULL_F
   8367 	     [(match_dup 1)
   8368 	      (match_operand:SI 6 "aarch64_sve_gp_strictness")
   8369 	      (match_operand:SVE_FULL_F 3 "register_operand" "w")]
   8370 	     UNSPEC_COND_FABS)]
   8371 	  SVE_COND_FP_ABS_CMP))]
   8372   "TARGET_SVE"
   8373   "fac<cmp_op>\t%0.<Vetype>, %1/z, %2.<Vetype>, %3.<Vetype>"
   8374 )
   8375 
   8376 ;; -------------------------------------------------------------------------
   8377 ;; ---- [PRED] Select
   8378 ;; -------------------------------------------------------------------------
   8379 ;; Includes:
   8380 ;; - SEL
   8381 ;; -------------------------------------------------------------------------
   8382 
   8383 (define_insn "@vcond_mask_<mode><mode>"
   8384   [(set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   8385 	(ior:PRED_ALL
   8386 	  (and:PRED_ALL
   8387 	    (match_operand:PRED_ALL 3 "register_operand" "Upa")
   8388 	    (match_operand:PRED_ALL 1 "register_operand" "Upa"))
   8389 	  (and:PRED_ALL
   8390 	    (not (match_dup 3))
   8391 	    (match_operand:PRED_ALL 2 "register_operand" "Upa"))))]
   8392   "TARGET_SVE"
   8393   "sel\t%0.b, %3, %1.b, %2.b"
   8394 )
   8395 
   8396 ;; -------------------------------------------------------------------------
   8397 ;; ---- [PRED] Test bits
   8398 ;; -------------------------------------------------------------------------
   8399 ;; Includes:
   8400 ;; - PTEST
   8401 ;; -------------------------------------------------------------------------
   8402 
   8403 ;; Branch based on predicate equality or inequality.
   8404 (define_expand "cbranch<mode>4"
   8405   [(set (pc)
   8406 	(if_then_else
   8407 	  (match_operator 0 "aarch64_equality_operator"
   8408 	    [(match_operand:PRED_ALL 1 "register_operand")
   8409 	     (match_operand:PRED_ALL 2 "aarch64_simd_reg_or_zero")])
   8410 	  (label_ref (match_operand 3 ""))
   8411 	  (pc)))]
   8412   ""
   8413   {
   8414     rtx ptrue = force_reg (VNx16BImode, aarch64_ptrue_all (<data_bytes>));
   8415     rtx cast_ptrue = gen_lowpart (<MODE>mode, ptrue);
   8416     rtx ptrue_flag = gen_int_mode (SVE_KNOWN_PTRUE, SImode);
   8417     rtx pred;
   8418     if (operands[2] == CONST0_RTX (<MODE>mode))
   8419       pred = operands[1];
   8420     else
   8421       {
   8422 	pred = gen_reg_rtx (<MODE>mode);
   8423 	emit_insn (gen_aarch64_pred_xor<mode>_z (pred, cast_ptrue, operands[1],
   8424 						 operands[2]));
   8425       }
   8426     emit_insn (gen_aarch64_ptest<mode> (ptrue, cast_ptrue, ptrue_flag, pred));
   8427     operands[1] = gen_rtx_REG (CC_NZCmode, CC_REGNUM);
   8428     operands[2] = const0_rtx;
   8429   }
   8430 )
   8431 
   8432 ;; See "Description of UNSPEC_PTEST" above for details.
   8433 (define_insn "aarch64_ptest<mode>"
   8434   [(set (reg:CC_NZC CC_REGNUM)
   8435 	(unspec:CC_NZC [(match_operand:VNx16BI 0 "register_operand" "Upa")
   8436 			(match_operand 1)
   8437 			(match_operand:SI 2 "aarch64_sve_ptrue_flag")
   8438 			(match_operand:PRED_ALL 3 "register_operand" "Upa")]
   8439 		       UNSPEC_PTEST))]
   8440   "TARGET_SVE"
   8441   "ptest\t%0, %3.b"
   8442 )
   8443 
   8444 ;; =========================================================================
   8445 ;; == Reductions
   8446 ;; =========================================================================
   8447 
   8448 ;; -------------------------------------------------------------------------
   8449 ;; ---- [INT,FP] Conditional reductions
   8450 ;; -------------------------------------------------------------------------
   8451 ;; Includes:
   8452 ;; - CLASTA
   8453 ;; - CLASTB
   8454 ;; -------------------------------------------------------------------------
   8455 
   8456 ;; Set operand 0 to the last active element in operand 3, or to tied
   8457 ;; operand 1 if no elements are active.
   8458 (define_insn "@fold_extract_<last_op>_<mode>"
   8459   [(set (match_operand:<VEL> 0 "register_operand" "=?r, w")
   8460 	(unspec:<VEL>
   8461 	  [(match_operand:<VEL> 1 "register_operand" "0, 0")
   8462 	   (match_operand:<VPRED> 2 "register_operand" "Upl, Upl")
   8463 	   (match_operand:SVE_FULL 3 "register_operand" "w, w")]
   8464 	  CLAST))]
   8465   "TARGET_SVE"
   8466   "@
   8467    clast<ab>\t%<vwcore>0, %2, %<vwcore>0, %3.<Vetype>
   8468    clast<ab>\t%<Vetype>0, %2, %<Vetype>0, %3.<Vetype>"
   8469 )
   8470 
   8471 (define_insn "@aarch64_fold_extract_vector_<last_op>_<mode>"
   8472   [(set (match_operand:SVE_FULL 0 "register_operand" "=w, ?&w")
   8473 	(unspec:SVE_FULL
   8474 	  [(match_operand:SVE_FULL 1 "register_operand" "0, w")
   8475 	   (match_operand:<VPRED> 2 "register_operand" "Upl, Upl")
   8476 	   (match_operand:SVE_FULL 3 "register_operand" "w, w")]
   8477 	  CLAST))]
   8478   "TARGET_SVE"
   8479   "@
   8480    clast<ab>\t%0.<Vetype>, %2, %0.<Vetype>, %3.<Vetype>
   8481    movprfx\t%0, %1\;clast<ab>\t%0.<Vetype>, %2, %0.<Vetype>, %3.<Vetype>"
   8482 )
   8483 
   8484 ;; -------------------------------------------------------------------------
   8485 ;; ---- [INT] Tree reductions
   8486 ;; -------------------------------------------------------------------------
   8487 ;; Includes:
   8488 ;; - ANDV
   8489 ;; - EORV
   8490 ;; - ORV
   8491 ;; - SADDV
   8492 ;; - SMAXV
   8493 ;; - SMINV
   8494 ;; - UADDV
   8495 ;; - UMAXV
   8496 ;; - UMINV
   8497 ;; -------------------------------------------------------------------------
   8498 
   8499 ;; Unpredicated integer add reduction.
   8500 (define_expand "reduc_plus_scal_<mode>"
   8501   [(match_operand:<VEL> 0 "register_operand")
   8502    (match_operand:SVE_FULL_I 1 "register_operand")]
   8503   "TARGET_SVE"
   8504   {
   8505     rtx pred = aarch64_ptrue_reg (<VPRED>mode);
   8506     rtx tmp = <VEL>mode == DImode ? operands[0] : gen_reg_rtx (DImode);
   8507     emit_insn (gen_aarch64_pred_reduc_uadd_<mode> (tmp, pred, operands[1]));
   8508     if (tmp != operands[0])
   8509       emit_move_insn (operands[0], gen_lowpart (<VEL>mode, tmp));
   8510     DONE;
   8511   }
   8512 )
   8513 
   8514 ;; Predicated integer add reduction.  The result is always 64-bits.
   8515 (define_insn "@aarch64_pred_reduc_<optab>_<mode>"
   8516   [(set (match_operand:DI 0 "register_operand" "=w")
   8517 	(unspec:DI [(match_operand:<VPRED> 1 "register_operand" "Upl")
   8518 		    (match_operand:SVE_FULL_I 2 "register_operand" "w")]
   8519 		   SVE_INT_ADDV))]
   8520   "TARGET_SVE && <max_elem_bits> >= <elem_bits>"
   8521   "<su>addv\t%d0, %1, %2.<Vetype>"
   8522 )
   8523 
   8524 ;; Unpredicated integer reductions.
   8525 (define_expand "reduc_<optab>_scal_<mode>"
   8526   [(set (match_operand:<VEL> 0 "register_operand")
   8527 	(unspec:<VEL> [(match_dup 2)
   8528 		       (match_operand:SVE_FULL_I 1 "register_operand")]
   8529 		      SVE_INT_REDUCTION))]
   8530   "TARGET_SVE"
   8531   {
   8532     operands[2] = aarch64_ptrue_reg (<VPRED>mode);
   8533   }
   8534 )
   8535 
   8536 ;; Predicated integer reductions.
   8537 (define_insn "@aarch64_pred_reduc_<optab>_<mode>"
   8538   [(set (match_operand:<VEL> 0 "register_operand" "=w")
   8539 	(unspec:<VEL> [(match_operand:<VPRED> 1 "register_operand" "Upl")
   8540 		       (match_operand:SVE_FULL_I 2 "register_operand" "w")]
   8541 		      SVE_INT_REDUCTION))]
   8542   "TARGET_SVE"
   8543   "<sve_int_op>\t%<Vetype>0, %1, %2.<Vetype>"
   8544 )
   8545 
   8546 ;; -------------------------------------------------------------------------
   8547 ;; ---- [FP] Tree reductions
   8548 ;; -------------------------------------------------------------------------
   8549 ;; Includes:
   8550 ;; - FADDV
   8551 ;; - FMAXNMV
   8552 ;; - FMAXV
   8553 ;; - FMINNMV
   8554 ;; - FMINV
   8555 ;; -------------------------------------------------------------------------
   8556 
   8557 ;; Unpredicated floating-point tree reductions.
   8558 (define_expand "reduc_<optab>_scal_<mode>"
   8559   [(set (match_operand:<VEL> 0 "register_operand")
   8560 	(unspec:<VEL> [(match_dup 2)
   8561 		       (match_operand:SVE_FULL_F 1 "register_operand")]
   8562 		      SVE_FP_REDUCTION))]
   8563   "TARGET_SVE"
   8564   {
   8565     operands[2] = aarch64_ptrue_reg (<VPRED>mode);
   8566   }
   8567 )
   8568 
   8569 (define_expand "reduc_<fmaxmin>_scal_<mode>"
   8570   [(match_operand:<VEL> 0 "register_operand")
   8571    (unspec:<VEL> [(match_operand:SVE_FULL_F 1 "register_operand")]
   8572 		 FMAXMINNMV)]
   8573   "TARGET_SVE"
   8574   {
   8575     emit_insn (gen_reduc_<optab>_scal_<mode> (operands[0], operands[1]));
   8576     DONE;
   8577   }
   8578 )
   8579 
   8580 ;; Predicated floating-point tree reductions.
   8581 (define_insn "@aarch64_pred_reduc_<optab>_<mode>"
   8582   [(set (match_operand:<VEL> 0 "register_operand" "=w")
   8583 	(unspec:<VEL> [(match_operand:<VPRED> 1 "register_operand" "Upl")
   8584 		       (match_operand:SVE_FULL_F 2 "register_operand" "w")]
   8585 		      SVE_FP_REDUCTION))]
   8586   "TARGET_SVE"
   8587   "<sve_fp_op>\t%<Vetype>0, %1, %2.<Vetype>"
   8588 )
   8589 
   8590 ;; -------------------------------------------------------------------------
   8591 ;; ---- [FP] Left-to-right reductions
   8592 ;; -------------------------------------------------------------------------
   8593 ;; Includes:
   8594 ;; - FADDA
   8595 ;; -------------------------------------------------------------------------
   8596 
   8597 ;; Unpredicated in-order FP reductions.
   8598 (define_expand "fold_left_plus_<mode>"
   8599   [(set (match_operand:<VEL> 0 "register_operand")
   8600 	(unspec:<VEL> [(match_dup 3)
   8601 		       (match_operand:<VEL> 1 "register_operand")
   8602 		       (match_operand:SVE_FULL_F 2 "register_operand")]
   8603 		      UNSPEC_FADDA))]
   8604   "TARGET_SVE"
   8605   {
   8606     operands[3] = aarch64_ptrue_reg (<VPRED>mode);
   8607   }
   8608 )
   8609 
   8610 ;; Predicated in-order FP reductions.
   8611 (define_insn "mask_fold_left_plus_<mode>"
   8612   [(set (match_operand:<VEL> 0 "register_operand" "=w")
   8613 	(unspec:<VEL> [(match_operand:<VPRED> 3 "register_operand" "Upl")
   8614 		       (match_operand:<VEL> 1 "register_operand" "0")
   8615 		       (match_operand:SVE_FULL_F 2 "register_operand" "w")]
   8616 		      UNSPEC_FADDA))]
   8617   "TARGET_SVE"
   8618   "fadda\t%<Vetype>0, %3, %<Vetype>0, %2.<Vetype>"
   8619 )
   8620 
   8621 ;; =========================================================================
   8622 ;; == Permutes
   8623 ;; =========================================================================
   8624 
   8625 ;; -------------------------------------------------------------------------
   8626 ;; ---- [INT,FP] General permutes
   8627 ;; -------------------------------------------------------------------------
   8628 ;; Includes:
   8629 ;; - TBL
   8630 ;; -------------------------------------------------------------------------
   8631 
   8632 (define_expand "vec_perm<mode>"
   8633   [(match_operand:SVE_FULL 0 "register_operand")
   8634    (match_operand:SVE_FULL 1 "register_operand")
   8635    (match_operand:SVE_FULL 2 "register_operand")
   8636    (match_operand:<V_INT_EQUIV> 3 "aarch64_sve_vec_perm_operand")]
   8637   "TARGET_SVE && GET_MODE_NUNITS (<MODE>mode).is_constant ()"
   8638   {
   8639     aarch64_expand_sve_vec_perm (operands[0], operands[1],
   8640 				 operands[2], operands[3]);
   8641     DONE;
   8642   }
   8643 )
   8644 
   8645 (define_insn "@aarch64_sve_tbl<mode>"
   8646   [(set (match_operand:SVE_FULL 0 "register_operand" "=w")
   8647 	(unspec:SVE_FULL
   8648 	  [(match_operand:SVE_FULL 1 "register_operand" "w")
   8649 	   (match_operand:<V_INT_EQUIV> 2 "register_operand" "w")]
   8650 	  UNSPEC_TBL))]
   8651   "TARGET_SVE"
   8652   "tbl\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>"
   8653 )
   8654 
   8655 ;; -------------------------------------------------------------------------
   8656 ;; ---- [INT,FP] Special-purpose unary permutes
   8657 ;; -------------------------------------------------------------------------
   8658 ;; Includes:
   8659 ;; - COMPACT
   8660 ;; - DUP
   8661 ;; - REV
   8662 ;; -------------------------------------------------------------------------
   8663 
   8664 ;; Compact active elements and pad with zeros.
   8665 (define_insn "@aarch64_sve_compact<mode>"
   8666   [(set (match_operand:SVE_FULL_SD 0 "register_operand" "=w")
   8667 	(unspec:SVE_FULL_SD
   8668 	  [(match_operand:<VPRED> 1 "register_operand" "Upl")
   8669 	   (match_operand:SVE_FULL_SD 2 "register_operand" "w")]
   8670 	  UNSPEC_SVE_COMPACT))]
   8671   "TARGET_SVE"
   8672   "compact\t%0.<Vetype>, %1, %2.<Vetype>"
   8673 )
   8674 
   8675 ;; Duplicate one element of a vector.
   8676 (define_insn "@aarch64_sve_dup_lane<mode>"
   8677   [(set (match_operand:SVE_ALL 0 "register_operand" "=w")
   8678 	(vec_duplicate:SVE_ALL
   8679 	  (vec_select:<VEL>
   8680 	    (match_operand:SVE_ALL 1 "register_operand" "w")
   8681 	    (parallel [(match_operand:SI 2 "const_int_operand")]))))]
   8682   "TARGET_SVE
   8683    && IN_RANGE (INTVAL (operands[2]) * <container_bits> / 8, 0, 63)"
   8684   "dup\t%0.<Vctype>, %1.<Vctype>[%2]"
   8685 )
   8686 
   8687 ;; Use DUP.Q to duplicate a 128-bit segment of a register.
   8688 ;;
   8689 ;; The vec_select:<V128> sets memory lane number N of the V128 to lane
   8690 ;; number op2 + N of op1.  (We don't need to distinguish between memory
   8691 ;; and architectural register lane numbering for op1 or op0, since the
   8692 ;; two numbering schemes are the same for SVE.)
   8693 ;;
   8694 ;; The vec_duplicate:SVE_FULL then copies memory lane number N of the
   8695 ;; V128 (and thus lane number op2 + N of op1) to lane numbers N + I * STEP
   8696 ;; of op0.  We therefore get the correct result for both endiannesses.
   8697 ;;
   8698 ;; The wrinkle is that for big-endian V128 registers, memory lane numbering
   8699 ;; is in the opposite order to architectural register lane numbering.
   8700 ;; Thus if we were to do this operation via a V128 temporary register,
   8701 ;; the vec_select and vec_duplicate would both involve a reverse operation
   8702 ;; for big-endian targets.  In this fused pattern the two reverses cancel
   8703 ;; each other out.
   8704 (define_insn "@aarch64_sve_dupq_lane<mode>"
   8705   [(set (match_operand:SVE_FULL 0 "register_operand" "=w")
   8706 	(vec_duplicate:SVE_FULL
   8707 	  (vec_select:<V128>
   8708 	    (match_operand:SVE_FULL 1 "register_operand" "w")
   8709 	    (match_operand 2 "ascending_int_parallel"))))]
   8710   "TARGET_SVE
   8711    && (INTVAL (XVECEXP (operands[2], 0, 0))
   8712        * GET_MODE_SIZE (<VEL>mode)) % 16 == 0
   8713    && IN_RANGE (INTVAL (XVECEXP (operands[2], 0, 0))
   8714 		* GET_MODE_SIZE (<VEL>mode), 0, 63)"
   8715   {
   8716     unsigned int byte = (INTVAL (XVECEXP (operands[2], 0, 0))
   8717 			 * GET_MODE_SIZE (<VEL>mode));
   8718     operands[2] = gen_int_mode (byte / 16, DImode);
   8719     return "dup\t%0.q, %1.q[%2]";
   8720   }
   8721 )
   8722 
   8723 ;; Reverse the order of elements within a full vector.
   8724 (define_insn "@aarch64_sve_rev<mode>"
   8725   [(set (match_operand:SVE_ALL 0 "register_operand" "=w")
   8726 	(unspec:SVE_ALL
   8727 	  [(match_operand:SVE_ALL 1 "register_operand" "w")]
   8728 	  UNSPEC_REV))]
   8729   "TARGET_SVE"
   8730   "rev\t%0.<Vctype>, %1.<Vctype>")
   8731 
   8732 ;; -------------------------------------------------------------------------
   8733 ;; ---- [INT,FP] Special-purpose binary permutes
   8734 ;; -------------------------------------------------------------------------
   8735 ;; Includes:
   8736 ;; - EXT
   8737 ;; - SPLICE
   8738 ;; - TRN1
   8739 ;; - TRN2
   8740 ;; - UZP1
   8741 ;; - UZP2
   8742 ;; - ZIP1
   8743 ;; - ZIP2
   8744 ;; -------------------------------------------------------------------------
   8745 
   8746 ;; Like EXT, but start at the first active element.
   8747 (define_insn "@aarch64_sve_splice<mode>"
   8748   [(set (match_operand:SVE_FULL 0 "register_operand" "=w, ?&w")
   8749 	(unspec:SVE_FULL
   8750 	  [(match_operand:<VPRED> 1 "register_operand" "Upl, Upl")
   8751 	   (match_operand:SVE_FULL 2 "register_operand" "0, w")
   8752 	   (match_operand:SVE_FULL 3 "register_operand" "w, w")]
   8753 	  UNSPEC_SVE_SPLICE))]
   8754   "TARGET_SVE"
   8755   "@
   8756    splice\t%0.<Vetype>, %1, %0.<Vetype>, %3.<Vetype>
   8757    movprfx\t%0, %2\;splice\t%0.<Vetype>, %1, %0.<Vetype>, %3.<Vetype>"
   8758   [(set_attr "movprfx" "*, yes")]
   8759 )
   8760 
   8761 ;; Permutes that take half the elements from one vector and half the
   8762 ;; elements from the other.
   8763 (define_insn "@aarch64_sve_<perm_insn><mode>"
   8764   [(set (match_operand:SVE_ALL 0 "register_operand" "=w")
   8765 	(unspec:SVE_ALL
   8766 	  [(match_operand:SVE_ALL 1 "register_operand" "w")
   8767 	   (match_operand:SVE_ALL 2 "register_operand" "w")]
   8768 	  PERMUTE))]
   8769   "TARGET_SVE"
   8770   "<perm_insn>\t%0.<Vctype>, %1.<Vctype>, %2.<Vctype>"
   8771 )
   8772 
   8773 ;; Apply PERMUTE to 128-bit sequences.  The behavior of these patterns
   8774 ;; doesn't depend on the mode.
   8775 (define_insn "@aarch64_sve_<optab><mode>"
   8776   [(set (match_operand:SVE_FULL 0 "register_operand" "=w")
   8777 	(unspec:SVE_FULL
   8778 	  [(match_operand:SVE_FULL 1 "register_operand" "w")
   8779 	   (match_operand:SVE_FULL 2 "register_operand" "w")]
   8780 	  PERMUTEQ))]
   8781   "TARGET_SVE_F64MM"
   8782   "<perm_insn>\t%0.q, %1.q, %2.q"
   8783 )
   8784 
   8785 ;; Concatenate two vectors and extract a subvector.  Note that the
   8786 ;; immediate (third) operand is the lane index not the byte index.
   8787 (define_insn "@aarch64_sve_ext<mode>"
   8788   [(set (match_operand:SVE_ALL 0 "register_operand" "=w, ?&w")
   8789 	(unspec:SVE_ALL
   8790 	  [(match_operand:SVE_ALL 1 "register_operand" "0, w")
   8791 	   (match_operand:SVE_ALL 2 "register_operand" "w, w")
   8792 	   (match_operand:SI 3 "const_int_operand")]
   8793 	  UNSPEC_EXT))]
   8794   "TARGET_SVE
   8795    && IN_RANGE (INTVAL (operands[3]) * <container_bits> / 8, 0, 255)"
   8796   {
   8797     operands[3] = GEN_INT (INTVAL (operands[3]) * <container_bits> / 8);
   8798     return (which_alternative == 0
   8799 	    ? "ext\\t%0.b, %0.b, %2.b, #%3"
   8800 	    : "movprfx\t%0, %1\;ext\\t%0.b, %0.b, %2.b, #%3");
   8801   }
   8802   [(set_attr "movprfx" "*,yes")]
   8803 )
   8804 
   8805 ;; -------------------------------------------------------------------------
   8806 ;; ---- [PRED] Special-purpose unary permutes
   8807 ;; -------------------------------------------------------------------------
   8808 ;; Includes:
   8809 ;; - REV
   8810 ;; -------------------------------------------------------------------------
   8811 
   8812 (define_insn "@aarch64_sve_rev<mode>"
   8813   [(set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   8814 	(unspec:PRED_ALL [(match_operand:PRED_ALL 1 "register_operand" "Upa")]
   8815 			 UNSPEC_REV))]
   8816   "TARGET_SVE"
   8817   "rev\t%0.<Vetype>, %1.<Vetype>")
   8818 
   8819 ;; -------------------------------------------------------------------------
   8820 ;; ---- [PRED] Special-purpose binary permutes
   8821 ;; -------------------------------------------------------------------------
   8822 ;; Includes:
   8823 ;; - TRN1
   8824 ;; - TRN2
   8825 ;; - UZP1
   8826 ;; - UZP2
   8827 ;; - ZIP1
   8828 ;; - ZIP2
   8829 ;; -------------------------------------------------------------------------
   8830 
   8831 ;; Permutes that take half the elements from one vector and half the
   8832 ;; elements from the other.
   8833 (define_insn "@aarch64_sve_<perm_insn><mode>"
   8834   [(set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   8835 	(unspec:PRED_ALL [(match_operand:PRED_ALL 1 "register_operand" "Upa")
   8836 			  (match_operand:PRED_ALL 2 "register_operand" "Upa")]
   8837 			 PERMUTE))]
   8838   "TARGET_SVE"
   8839   "<perm_insn>\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>"
   8840 )
   8841 
   8842 ;; Special purpose permute used by the predicate generation instructions.
   8843 ;; Unlike the normal permute patterns, these instructions operate on VNx16BI
   8844 ;; regardless of the element size, so that all input and output bits are
   8845 ;; well-defined.  Operand 3 then indicates the size of the permute.
   8846 (define_insn "@aarch64_sve_trn1_conv<mode>"
   8847   [(set (match_operand:VNx16BI 0 "register_operand" "=Upa")
   8848 	(unspec:VNx16BI [(match_operand:VNx16BI 1 "register_operand" "Upa")
   8849 			 (match_operand:VNx16BI 2 "register_operand" "Upa")
   8850 			 (match_operand:PRED_ALL 3 "aarch64_simd_imm_zero")]
   8851 			UNSPEC_TRN1_CONV))]
   8852   "TARGET_SVE"
   8853   "trn1\t%0.<PRED_ALL:Vetype>, %1.<PRED_ALL:Vetype>, %2.<PRED_ALL:Vetype>"
   8854 )
   8855 
   8856 ;; =========================================================================
   8857 ;; == Conversions
   8858 ;; =========================================================================
   8859 
   8860 ;; -------------------------------------------------------------------------
   8861 ;; ---- [INT<-INT] Packs
   8862 ;; -------------------------------------------------------------------------
   8863 ;; Includes:
   8864 ;; - UZP1
   8865 ;; -------------------------------------------------------------------------
   8866 
   8867 ;; Integer pack.  Use UZP1 on the narrower type, which discards
   8868 ;; the high part of each wide element.
   8869 (define_insn "vec_pack_trunc_<Vwide>"
   8870   [(set (match_operand:SVE_FULL_BHSI 0 "register_operand" "=w")
   8871 	(unspec:SVE_FULL_BHSI
   8872 	  [(match_operand:<VWIDE> 1 "register_operand" "w")
   8873 	   (match_operand:<VWIDE> 2 "register_operand" "w")]
   8874 	  UNSPEC_PACK))]
   8875   "TARGET_SVE"
   8876   "uzp1\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>"
   8877 )
   8878 
   8879 ;; -------------------------------------------------------------------------
   8880 ;; ---- [INT<-INT] Unpacks
   8881 ;; -------------------------------------------------------------------------
   8882 ;; Includes:
   8883 ;; - SUNPKHI
   8884 ;; - SUNPKLO
   8885 ;; - UUNPKHI
   8886 ;; - UUNPKLO
   8887 ;; -------------------------------------------------------------------------
   8888 
   8889 ;; Unpack the low or high half of a vector, where "high" refers to
   8890 ;; the low-numbered lanes for big-endian and the high-numbered lanes
   8891 ;; for little-endian.
   8892 (define_expand "vec_unpack<su>_<perm_hilo>_<SVE_FULL_BHSI:mode>"
   8893   [(match_operand:<VWIDE> 0 "register_operand")
   8894    (unspec:<VWIDE>
   8895      [(match_operand:SVE_FULL_BHSI 1 "register_operand")] UNPACK)]
   8896   "TARGET_SVE"
   8897   {
   8898     emit_insn ((<hi_lanes_optab>
   8899 		? gen_aarch64_sve_<su>unpkhi_<SVE_FULL_BHSI:mode>
   8900 		: gen_aarch64_sve_<su>unpklo_<SVE_FULL_BHSI:mode>)
   8901 	       (operands[0], operands[1]));
   8902     DONE;
   8903   }
   8904 )
   8905 
   8906 (define_insn "@aarch64_sve_<su>unpk<perm_hilo>_<SVE_FULL_BHSI:mode>"
   8907   [(set (match_operand:<VWIDE> 0 "register_operand" "=w")
   8908 	(unspec:<VWIDE>
   8909 	  [(match_operand:SVE_FULL_BHSI 1 "register_operand" "w")]
   8910 	  UNPACK))]
   8911   "TARGET_SVE"
   8912   "<su>unpk<perm_hilo>\t%0.<Vewtype>, %1.<Vetype>"
   8913 )
   8914 
   8915 ;; -------------------------------------------------------------------------
   8916 ;; ---- [INT<-FP] Conversions
   8917 ;; -------------------------------------------------------------------------
   8918 ;; Includes:
   8919 ;; - FCVTZS
   8920 ;; - FCVTZU
   8921 ;; -------------------------------------------------------------------------
   8922 
   8923 ;; Unpredicated conversion of floats to integers of the same size (HF to HI,
   8924 ;; SF to SI or DF to DI).
   8925 (define_expand "<optab><mode><v_int_equiv>2"
   8926   [(set (match_operand:<V_INT_EQUIV> 0 "register_operand")
   8927 	(unspec:<V_INT_EQUIV>
   8928 	  [(match_dup 2)
   8929 	   (const_int SVE_RELAXED_GP)
   8930 	   (match_operand:SVE_FULL_F 1 "register_operand")]
   8931 	  SVE_COND_FCVTI))]
   8932   "TARGET_SVE"
   8933   {
   8934     operands[2] = aarch64_ptrue_reg (<VPRED>mode);
   8935   }
   8936 )
   8937 
   8938 ;; Predicated float-to-integer conversion, either to the same width or wider.
   8939 (define_insn "@aarch64_sve_<optab>_nontrunc<SVE_FULL_F:mode><SVE_FULL_HSDI:mode>"
   8940   [(set (match_operand:SVE_FULL_HSDI 0 "register_operand" "=w, ?&w")
   8941 	(unspec:SVE_FULL_HSDI
   8942 	  [(match_operand:<SVE_FULL_HSDI:VPRED> 1 "register_operand" "Upl, Upl")
   8943 	   (match_operand:SI 3 "aarch64_sve_gp_strictness")
   8944 	   (match_operand:SVE_FULL_F 2 "register_operand" "0, w")]
   8945 	  SVE_COND_FCVTI))]
   8946   "TARGET_SVE && <SVE_FULL_HSDI:elem_bits> >= <SVE_FULL_F:elem_bits>"
   8947   "@
   8948    fcvtz<su>\t%0.<SVE_FULL_HSDI:Vetype>, %1/m, %2.<SVE_FULL_F:Vetype>
   8949    movprfx\t%0, %2\;fcvtz<su>\t%0.<SVE_FULL_HSDI:Vetype>, %1/m, %2.<SVE_FULL_F:Vetype>"
   8950   [(set_attr "movprfx" "*,yes")]
   8951 )
   8952 
   8953 ;; Predicated narrowing float-to-integer conversion.
   8954 (define_insn "@aarch64_sve_<optab>_trunc<VNx2DF_ONLY:mode><VNx4SI_ONLY:mode>"
   8955   [(set (match_operand:VNx4SI_ONLY 0 "register_operand" "=w, ?&w")
   8956 	(unspec:VNx4SI_ONLY
   8957 	  [(match_operand:VNx2BI 1 "register_operand" "Upl, Upl")
   8958 	   (match_operand:SI 3 "aarch64_sve_gp_strictness")
   8959 	   (match_operand:VNx2DF_ONLY 2 "register_operand" "0, w")]
   8960 	  SVE_COND_FCVTI))]
   8961   "TARGET_SVE"
   8962   "@
   8963    fcvtz<su>\t%0.<VNx4SI_ONLY:Vetype>, %1/m, %2.<VNx2DF_ONLY:Vetype>
   8964    movprfx\t%0, %2\;fcvtz<su>\t%0.<VNx4SI_ONLY:Vetype>, %1/m, %2.<VNx2DF_ONLY:Vetype>"
   8965   [(set_attr "movprfx" "*,yes")]
   8966 )
   8967 
   8968 ;; Predicated float-to-integer conversion with merging, either to the same
   8969 ;; width or wider.
   8970 (define_expand "@cond_<optab>_nontrunc<SVE_FULL_F:mode><SVE_FULL_HSDI:mode>"
   8971   [(set (match_operand:SVE_FULL_HSDI 0 "register_operand")
   8972 	(unspec:SVE_FULL_HSDI
   8973 	  [(match_operand:<SVE_FULL_HSDI:VPRED> 1 "register_operand")
   8974 	   (unspec:SVE_FULL_HSDI
   8975 	     [(match_dup 1)
   8976 	      (const_int SVE_STRICT_GP)
   8977 	      (match_operand:SVE_FULL_F 2 "register_operand")]
   8978 	     SVE_COND_FCVTI)
   8979 	   (match_operand:SVE_FULL_HSDI 3 "aarch64_simd_reg_or_zero")]
   8980 	  UNSPEC_SEL))]
   8981   "TARGET_SVE && <SVE_FULL_HSDI:elem_bits> >= <SVE_FULL_F:elem_bits>"
   8982 )
   8983 
   8984 ;; The first alternative doesn't need the earlyclobber, but the only case
   8985 ;; it would help is the uninteresting one in which operands 2 and 3 are
   8986 ;; the same register (despite having different modes).  Making all the
   8987 ;; alternatives earlyclobber makes things more consistent for the
   8988 ;; register allocator.
   8989 (define_insn_and_rewrite "*cond_<optab>_nontrunc<SVE_FULL_F:mode><SVE_FULL_HSDI:mode>_relaxed"
   8990   [(set (match_operand:SVE_FULL_HSDI 0 "register_operand" "=&w, &w, ?&w")
   8991 	(unspec:SVE_FULL_HSDI
   8992 	  [(match_operand:<SVE_FULL_HSDI:VPRED> 1 "register_operand" "Upl, Upl, Upl")
   8993 	   (unspec:SVE_FULL_HSDI
   8994 	     [(match_operand 4)
   8995 	      (const_int SVE_RELAXED_GP)
   8996 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w, w")]
   8997 	     SVE_COND_FCVTI)
   8998 	   (match_operand:SVE_FULL_HSDI 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   8999 	  UNSPEC_SEL))]
   9000   "TARGET_SVE && <SVE_FULL_HSDI:elem_bits> >= <SVE_FULL_F:elem_bits>"
   9001   "@
   9002    fcvtz<su>\t%0.<SVE_FULL_HSDI:Vetype>, %1/m, %2.<SVE_FULL_F:Vetype>
   9003    movprfx\t%0.<SVE_FULL_HSDI:Vetype>, %1/z, %2.<SVE_FULL_HSDI:Vetype>\;fcvtz<su>\t%0.<SVE_FULL_HSDI:Vetype>, %1/m, %2.<SVE_FULL_F:Vetype>
   9004    movprfx\t%0, %3\;fcvtz<su>\t%0.<SVE_FULL_HSDI:Vetype>, %1/m, %2.<SVE_FULL_F:Vetype>"
   9005   "&& !rtx_equal_p (operands[1], operands[4])"
   9006   {
   9007     operands[4] = copy_rtx (operands[1]);
   9008   }
   9009   [(set_attr "movprfx" "*,yes,yes")]
   9010 )
   9011 
   9012 (define_insn "*cond_<optab>_nontrunc<SVE_FULL_F:mode><SVE_FULL_HSDI:mode>_strict"
   9013   [(set (match_operand:SVE_FULL_HSDI 0 "register_operand" "=&w, &w, ?&w")
   9014 	(unspec:SVE_FULL_HSDI
   9015 	  [(match_operand:<SVE_FULL_HSDI:VPRED> 1 "register_operand" "Upl, Upl, Upl")
   9016 	   (unspec:SVE_FULL_HSDI
   9017 	     [(match_dup 1)
   9018 	      (const_int SVE_STRICT_GP)
   9019 	      (match_operand:SVE_FULL_F 2 "register_operand" "w, w, w")]
   9020 	     SVE_COND_FCVTI)
   9021 	   (match_operand:SVE_FULL_HSDI 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   9022 	  UNSPEC_SEL))]
   9023   "TARGET_SVE && <SVE_FULL_HSDI:elem_bits> >= <SVE_FULL_F:elem_bits>"
   9024   "@
   9025    fcvtz<su>\t%0.<SVE_FULL_HSDI:Vetype>, %1/m, %2.<SVE_FULL_F:Vetype>
   9026    movprfx\t%0.<SVE_FULL_HSDI:Vetype>, %1/z, %2.<SVE_FULL_HSDI:Vetype>\;fcvtz<su>\t%0.<SVE_FULL_HSDI:Vetype>, %1/m, %2.<SVE_FULL_F:Vetype>
   9027    movprfx\t%0, %3\;fcvtz<su>\t%0.<SVE_FULL_HSDI:Vetype>, %1/m, %2.<SVE_FULL_F:Vetype>"
   9028   [(set_attr "movprfx" "*,yes,yes")]
   9029 )
   9030 
   9031 ;; Predicated narrowing float-to-integer conversion with merging.
   9032 (define_expand "@cond_<optab>_trunc<VNx2DF_ONLY:mode><VNx4SI_ONLY:mode>"
   9033   [(set (match_operand:VNx4SI_ONLY 0 "register_operand")
   9034 	(unspec:VNx4SI_ONLY
   9035 	  [(match_operand:VNx2BI 1 "register_operand")
   9036 	   (unspec:VNx4SI_ONLY
   9037 	     [(match_dup 1)
   9038 	      (const_int SVE_STRICT_GP)
   9039 	      (match_operand:VNx2DF_ONLY 2 "register_operand")]
   9040 	     SVE_COND_FCVTI)
   9041 	   (match_operand:VNx4SI_ONLY 3 "aarch64_simd_reg_or_zero")]
   9042 	  UNSPEC_SEL))]
   9043   "TARGET_SVE"
   9044 )
   9045 
   9046 (define_insn "*cond_<optab>_trunc<VNx2DF_ONLY:mode><VNx4SI_ONLY:mode>"
   9047   [(set (match_operand:VNx4SI_ONLY 0 "register_operand" "=&w, &w, ?&w")
   9048 	(unspec:VNx4SI_ONLY
   9049 	  [(match_operand:VNx2BI 1 "register_operand" "Upl, Upl, Upl")
   9050 	   (unspec:VNx4SI_ONLY
   9051 	     [(match_dup 1)
   9052 	      (match_operand:SI 4 "aarch64_sve_gp_strictness")
   9053 	      (match_operand:VNx2DF_ONLY 2 "register_operand" "w, w, w")]
   9054 	     SVE_COND_FCVTI)
   9055 	   (match_operand:VNx4SI_ONLY 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   9056 	  UNSPEC_SEL))]
   9057   "TARGET_SVE"
   9058   "@
   9059    fcvtz<su>\t%0.<VNx4SI_ONLY:Vetype>, %1/m, %2.<VNx2DF_ONLY:Vetype>
   9060    movprfx\t%0.<VNx2DF_ONLY:Vetype>, %1/z, %2.<VNx2DF_ONLY:Vetype>\;fcvtz<su>\t%0.<VNx4SI_ONLY:Vetype>, %1/m, %2.<VNx2DF_ONLY:Vetype>
   9061    movprfx\t%0, %3\;fcvtz<su>\t%0.<VNx4SI_ONLY:Vetype>, %1/m, %2.<VNx2DF_ONLY:Vetype>"
   9062   [(set_attr "movprfx" "*,yes,yes")]
   9063 )
   9064 
   9065 ;; -------------------------------------------------------------------------
   9066 ;; ---- [INT<-FP] Packs
   9067 ;; -------------------------------------------------------------------------
   9068 ;; The patterns in this section are synthetic.
   9069 ;; -------------------------------------------------------------------------
   9070 
   9071 ;; Convert two vectors of DF to SI and pack the results into a single vector.
   9072 (define_expand "vec_pack_<su>fix_trunc_vnx2df"
   9073   [(set (match_dup 4)
   9074 	(unspec:VNx4SI
   9075 	  [(match_dup 3)
   9076 	   (const_int SVE_RELAXED_GP)
   9077 	   (match_operand:VNx2DF 1 "register_operand")]
   9078 	  SVE_COND_FCVTI))
   9079    (set (match_dup 5)
   9080 	(unspec:VNx4SI
   9081 	  [(match_dup 3)
   9082 	   (const_int SVE_RELAXED_GP)
   9083 	   (match_operand:VNx2DF 2 "register_operand")]
   9084 	  SVE_COND_FCVTI))
   9085    (set (match_operand:VNx4SI 0 "register_operand")
   9086 	(unspec:VNx4SI [(match_dup 4) (match_dup 5)] UNSPEC_UZP1))]
   9087   "TARGET_SVE"
   9088   {
   9089     operands[3] = aarch64_ptrue_reg (VNx2BImode);
   9090     operands[4] = gen_reg_rtx (VNx4SImode);
   9091     operands[5] = gen_reg_rtx (VNx4SImode);
   9092   }
   9093 )
   9094 
   9095 ;; -------------------------------------------------------------------------
   9096 ;; ---- [INT<-FP] Unpacks
   9097 ;; -------------------------------------------------------------------------
   9098 ;; No patterns here yet!
   9099 ;; -------------------------------------------------------------------------
   9100 
   9101 ;; -------------------------------------------------------------------------
   9102 ;; ---- [FP<-INT] Conversions
   9103 ;; -------------------------------------------------------------------------
   9104 ;; Includes:
   9105 ;; - SCVTF
   9106 ;; - UCVTF
   9107 ;; -------------------------------------------------------------------------
   9108 
   9109 ;; Unpredicated conversion of integers to floats of the same size
   9110 ;; (HI to HF, SI to SF or DI to DF).
   9111 (define_expand "<optab><v_int_equiv><mode>2"
   9112   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   9113 	(unspec:SVE_FULL_F
   9114 	  [(match_dup 2)
   9115 	   (const_int SVE_RELAXED_GP)
   9116 	   (match_operand:<V_INT_EQUIV> 1 "register_operand")]
   9117 	  SVE_COND_ICVTF))]
   9118   "TARGET_SVE"
   9119   {
   9120     operands[2] = aarch64_ptrue_reg (<VPRED>mode);
   9121   }
   9122 )
   9123 
   9124 ;; Predicated integer-to-float conversion, either to the same width or
   9125 ;; narrower.
   9126 (define_insn "@aarch64_sve_<optab>_nonextend<SVE_FULL_HSDI:mode><SVE_FULL_F:mode>"
   9127   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=w, ?&w")
   9128 	(unspec:SVE_FULL_F
   9129 	  [(match_operand:<SVE_FULL_HSDI:VPRED> 1 "register_operand" "Upl, Upl")
   9130 	   (match_operand:SI 3 "aarch64_sve_gp_strictness")
   9131 	   (match_operand:SVE_FULL_HSDI 2 "register_operand" "0, w")]
   9132 	  SVE_COND_ICVTF))]
   9133   "TARGET_SVE && <SVE_FULL_HSDI:elem_bits> >= <SVE_FULL_F:elem_bits>"
   9134   "@
   9135    <su>cvtf\t%0.<SVE_FULL_F:Vetype>, %1/m, %2.<SVE_FULL_HSDI:Vetype>
   9136    movprfx\t%0, %2\;<su>cvtf\t%0.<SVE_FULL_F:Vetype>, %1/m, %2.<SVE_FULL_HSDI:Vetype>"
   9137   [(set_attr "movprfx" "*,yes")]
   9138 )
   9139 
   9140 ;; Predicated widening integer-to-float conversion.
   9141 (define_insn "@aarch64_sve_<optab>_extend<VNx4SI_ONLY:mode><VNx2DF_ONLY:mode>"
   9142   [(set (match_operand:VNx2DF_ONLY 0 "register_operand" "=w, ?&w")
   9143 	(unspec:VNx2DF_ONLY
   9144 	  [(match_operand:VNx2BI 1 "register_operand" "Upl, Upl")
   9145 	   (match_operand:SI 3 "aarch64_sve_gp_strictness")
   9146 	   (match_operand:VNx4SI_ONLY 2 "register_operand" "0, w")]
   9147 	  SVE_COND_ICVTF))]
   9148   "TARGET_SVE"
   9149   "@
   9150    <su>cvtf\t%0.<VNx2DF_ONLY:Vetype>, %1/m, %2.<VNx4SI_ONLY:Vetype>
   9151    movprfx\t%0, %2\;<su>cvtf\t%0.<VNx2DF_ONLY:Vetype>, %1/m, %2.<VNx4SI_ONLY:Vetype>"
   9152   [(set_attr "movprfx" "*,yes")]
   9153 )
   9154 
   9155 ;; Predicated integer-to-float conversion with merging, either to the same
   9156 ;; width or narrower.
   9157 (define_expand "@cond_<optab>_nonextend<SVE_FULL_HSDI:mode><SVE_FULL_F:mode>"
   9158   [(set (match_operand:SVE_FULL_F 0 "register_operand")
   9159 	(unspec:SVE_FULL_F
   9160 	  [(match_operand:<SVE_FULL_HSDI:VPRED> 1 "register_operand")
   9161 	   (unspec:SVE_FULL_F
   9162 	     [(match_dup 1)
   9163 	      (const_int SVE_STRICT_GP)
   9164 	      (match_operand:SVE_FULL_HSDI 2 "register_operand")]
   9165 	     SVE_COND_ICVTF)
   9166 	   (match_operand:SVE_FULL_F 3 "aarch64_simd_reg_or_zero")]
   9167 	  UNSPEC_SEL))]
   9168   "TARGET_SVE && <SVE_FULL_HSDI:elem_bits> >= <SVE_FULL_F:elem_bits>"
   9169 )
   9170 
   9171 ;; The first alternative doesn't need the earlyclobber, but the only case
   9172 ;; it would help is the uninteresting one in which operands 2 and 3 are
   9173 ;; the same register (despite having different modes).  Making all the
   9174 ;; alternatives earlyclobber makes things more consistent for the
   9175 ;; register allocator.
   9176 (define_insn_and_rewrite "*cond_<optab>_nonextend<SVE_FULL_HSDI:mode><SVE_FULL_F:mode>_relaxed"
   9177   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, ?&w")
   9178 	(unspec:SVE_FULL_F
   9179 	  [(match_operand:<SVE_FULL_HSDI:VPRED> 1 "register_operand" "Upl, Upl, Upl")
   9180 	   (unspec:SVE_FULL_F
   9181 	     [(match_operand 4)
   9182 	      (const_int SVE_RELAXED_GP)
   9183 	      (match_operand:SVE_FULL_HSDI 2 "register_operand" "w, w, w")]
   9184 	     SVE_COND_ICVTF)
   9185 	   (match_operand:SVE_FULL_F 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   9186 	  UNSPEC_SEL))]
   9187   "TARGET_SVE && <SVE_FULL_HSDI:elem_bits> >= <SVE_FULL_F:elem_bits>"
   9188   "@
   9189    <su>cvtf\t%0.<SVE_FULL_F:Vetype>, %1/m, %2.<SVE_FULL_HSDI:Vetype>
   9190    movprfx\t%0.<SVE_FULL_HSDI:Vetype>, %1/z, %2.<SVE_FULL_HSDI:Vetype>\;<su>cvtf\t%0.<SVE_FULL_F:Vetype>, %1/m, %2.<SVE_FULL_HSDI:Vetype>
   9191    movprfx\t%0, %3\;<su>cvtf\t%0.<SVE_FULL_F:Vetype>, %1/m, %2.<SVE_FULL_HSDI:Vetype>"
   9192   "&& !rtx_equal_p (operands[1], operands[4])"
   9193   {
   9194     operands[4] = copy_rtx (operands[1]);
   9195   }
   9196   [(set_attr "movprfx" "*,yes,yes")]
   9197 )
   9198 
   9199 (define_insn "*cond_<optab>_nonextend<SVE_FULL_HSDI:mode><SVE_FULL_F:mode>_strict"
   9200   [(set (match_operand:SVE_FULL_F 0 "register_operand" "=&w, &w, ?&w")
   9201 	(unspec:SVE_FULL_F
   9202 	  [(match_operand:<SVE_FULL_HSDI:VPRED> 1 "register_operand" "Upl, Upl, Upl")
   9203 	   (unspec:SVE_FULL_F
   9204 	     [(match_dup 1)
   9205 	      (const_int SVE_STRICT_GP)
   9206 	      (match_operand:SVE_FULL_HSDI 2 "register_operand" "w, w, w")]
   9207 	     SVE_COND_ICVTF)
   9208 	   (match_operand:SVE_FULL_F 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   9209 	  UNSPEC_SEL))]
   9210   "TARGET_SVE && <SVE_FULL_HSDI:elem_bits> >= <SVE_FULL_F:elem_bits>"
   9211   "@
   9212    <su>cvtf\t%0.<SVE_FULL_F:Vetype>, %1/m, %2.<SVE_FULL_HSDI:Vetype>
   9213    movprfx\t%0.<SVE_FULL_HSDI:Vetype>, %1/z, %2.<SVE_FULL_HSDI:Vetype>\;<su>cvtf\t%0.<SVE_FULL_F:Vetype>, %1/m, %2.<SVE_FULL_HSDI:Vetype>
   9214    movprfx\t%0, %3\;<su>cvtf\t%0.<SVE_FULL_F:Vetype>, %1/m, %2.<SVE_FULL_HSDI:Vetype>"
   9215   [(set_attr "movprfx" "*,yes,yes")]
   9216 )
   9217 
   9218 ;; Predicated widening integer-to-float conversion with merging.
   9219 (define_expand "@cond_<optab>_extend<VNx4SI_ONLY:mode><VNx2DF_ONLY:mode>"
   9220   [(set (match_operand:VNx2DF_ONLY 0 "register_operand")
   9221 	(unspec:VNx2DF_ONLY
   9222 	  [(match_operand:VNx2BI 1 "register_operand")
   9223 	   (unspec:VNx2DF_ONLY
   9224 	     [(match_dup 1)
   9225 	      (const_int SVE_STRICT_GP)
   9226 	      (match_operand:VNx4SI_ONLY 2 "register_operand")]
   9227 	     SVE_COND_ICVTF)
   9228 	   (match_operand:VNx2DF_ONLY 3 "aarch64_simd_reg_or_zero")]
   9229 	  UNSPEC_SEL))]
   9230   "TARGET_SVE"
   9231 )
   9232 
   9233 (define_insn "*cond_<optab>_extend<VNx4SI_ONLY:mode><VNx2DF_ONLY:mode>"
   9234   [(set (match_operand:VNx2DF_ONLY 0 "register_operand" "=w, ?&w, ?&w")
   9235 	(unspec:VNx2DF_ONLY
   9236 	  [(match_operand:VNx2BI 1 "register_operand" "Upl, Upl, Upl")
   9237 	   (unspec:VNx2DF_ONLY
   9238 	     [(match_dup 1)
   9239 	      (match_operand:SI 4 "aarch64_sve_gp_strictness")
   9240 	      (match_operand:VNx4SI_ONLY 2 "register_operand" "w, w, w")]
   9241 	     SVE_COND_ICVTF)
   9242 	   (match_operand:VNx2DF_ONLY 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   9243 	  UNSPEC_SEL))]
   9244   "TARGET_SVE"
   9245   "@
   9246    <su>cvtf\t%0.<VNx2DF_ONLY:Vetype>, %1/m, %2.<VNx4SI_ONLY:Vetype>
   9247    movprfx\t%0.<VNx2DF_ONLY:Vetype>, %1/z, %2.<VNx2DF_ONLY:Vetype>\;<su>cvtf\t%0.<VNx2DF_ONLY:Vetype>, %1/m, %2.<VNx4SI_ONLY:Vetype>
   9248    movprfx\t%0, %3\;<su>cvtf\t%0.<VNx2DF_ONLY:Vetype>, %1/m, %2.<VNx4SI_ONLY:Vetype>"
   9249   [(set_attr "movprfx" "*,yes,yes")]
   9250 )
   9251 
   9252 ;; -------------------------------------------------------------------------
   9253 ;; ---- [FP<-INT] Packs
   9254 ;; -------------------------------------------------------------------------
   9255 ;; No patterns here yet!
   9256 ;; -------------------------------------------------------------------------
   9257 
   9258 ;; -------------------------------------------------------------------------
   9259 ;; ---- [FP<-INT] Unpacks
   9260 ;; -------------------------------------------------------------------------
   9261 ;; The patterns in this section are synthetic.
   9262 ;; -------------------------------------------------------------------------
   9263 
   9264 ;; Unpack one half of a VNx4SI to VNx2DF.  First unpack from VNx4SI
   9265 ;; to VNx2DI, reinterpret the VNx2DI as a VNx4SI, then convert the
   9266 ;; unpacked VNx4SI to VNx2DF.
   9267 (define_expand "vec_unpack<su_optab>_float_<perm_hilo>_vnx4si"
   9268   [(match_operand:VNx2DF 0 "register_operand")
   9269    (FLOATUORS:VNx2DF
   9270      (unspec:VNx2DI [(match_operand:VNx4SI 1 "register_operand")]
   9271 		    UNPACK_UNSIGNED))]
   9272   "TARGET_SVE"
   9273   {
   9274     /* Use ZIP to do the unpack, since we don't care about the upper halves
   9275        and since it has the nice property of not needing any subregs.
   9276        If using UUNPK* turns out to be preferable, we could model it as
   9277        a ZIP whose first operand is zero.  */
   9278     rtx temp = gen_reg_rtx (VNx4SImode);
   9279     emit_insn ((<hi_lanes_optab>
   9280 		? gen_aarch64_sve_zip2vnx4si
   9281 		: gen_aarch64_sve_zip1vnx4si)
   9282 	       (temp, operands[1], operands[1]));
   9283     rtx ptrue = aarch64_ptrue_reg (VNx2BImode);
   9284     rtx strictness = gen_int_mode (SVE_RELAXED_GP, SImode);
   9285     emit_insn (gen_aarch64_sve_<FLOATUORS:optab>_extendvnx4sivnx2df
   9286 	       (operands[0], ptrue, temp, strictness));
   9287     DONE;
   9288   }
   9289 )
   9290 
   9291 ;; -------------------------------------------------------------------------
   9292 ;; ---- [FP<-FP] Packs
   9293 ;; -------------------------------------------------------------------------
   9294 ;; Includes:
   9295 ;; - FCVT
   9296 ;; -------------------------------------------------------------------------
   9297 
   9298 ;; Convert two vectors of DF to SF, or two vectors of SF to HF, and pack
   9299 ;; the results into a single vector.
   9300 (define_expand "vec_pack_trunc_<Vwide>"
   9301   [(set (match_dup 4)
   9302 	(unspec:SVE_FULL_HSF
   9303 	  [(match_dup 3)
   9304 	   (const_int SVE_RELAXED_GP)
   9305 	   (match_operand:<VWIDE> 1 "register_operand")]
   9306 	  UNSPEC_COND_FCVT))
   9307    (set (match_dup 5)
   9308 	(unspec:SVE_FULL_HSF
   9309 	  [(match_dup 3)
   9310 	   (const_int SVE_RELAXED_GP)
   9311 	   (match_operand:<VWIDE> 2 "register_operand")]
   9312 	  UNSPEC_COND_FCVT))
   9313    (set (match_operand:SVE_FULL_HSF 0 "register_operand")
   9314 	(unspec:SVE_FULL_HSF [(match_dup 4) (match_dup 5)] UNSPEC_UZP1))]
   9315   "TARGET_SVE"
   9316   {
   9317     operands[3] = aarch64_ptrue_reg (<VWIDE_PRED>mode);
   9318     operands[4] = gen_reg_rtx (<MODE>mode);
   9319     operands[5] = gen_reg_rtx (<MODE>mode);
   9320   }
   9321 )
   9322 
   9323 ;; Predicated float-to-float truncation.
   9324 (define_insn "@aarch64_sve_<optab>_trunc<SVE_FULL_SDF:mode><SVE_FULL_HSF:mode>"
   9325   [(set (match_operand:SVE_FULL_HSF 0 "register_operand" "=w, ?&w")
   9326 	(unspec:SVE_FULL_HSF
   9327 	  [(match_operand:<SVE_FULL_SDF:VPRED> 1 "register_operand" "Upl, Upl")
   9328 	   (match_operand:SI 3 "aarch64_sve_gp_strictness")
   9329 	   (match_operand:SVE_FULL_SDF 2 "register_operand" "0, w")]
   9330 	  SVE_COND_FCVT))]
   9331   "TARGET_SVE && <SVE_FULL_SDF:elem_bits> > <SVE_FULL_HSF:elem_bits>"
   9332   "@
   9333    fcvt\t%0.<SVE_FULL_HSF:Vetype>, %1/m, %2.<SVE_FULL_SDF:Vetype>
   9334    movprfx\t%0, %2\;fcvt\t%0.<SVE_FULL_HSF:Vetype>, %1/m, %2.<SVE_FULL_SDF:Vetype>"
   9335   [(set_attr "movprfx" "*,yes")]
   9336 )
   9337 
   9338 ;; Predicated float-to-float truncation with merging.
   9339 (define_expand "@cond_<optab>_trunc<SVE_FULL_SDF:mode><SVE_FULL_HSF:mode>"
   9340   [(set (match_operand:SVE_FULL_HSF 0 "register_operand")
   9341 	(unspec:SVE_FULL_HSF
   9342 	  [(match_operand:<SVE_FULL_SDF:VPRED> 1 "register_operand")
   9343 	   (unspec:SVE_FULL_HSF
   9344 	     [(match_dup 1)
   9345 	      (const_int SVE_STRICT_GP)
   9346 	      (match_operand:SVE_FULL_SDF 2 "register_operand")]
   9347 	     SVE_COND_FCVT)
   9348 	   (match_operand:SVE_FULL_HSF 3 "aarch64_simd_reg_or_zero")]
   9349 	  UNSPEC_SEL))]
   9350   "TARGET_SVE && <SVE_FULL_SDF:elem_bits> > <SVE_FULL_HSF:elem_bits>"
   9351 )
   9352 
   9353 (define_insn "*cond_<optab>_trunc<SVE_FULL_SDF:mode><SVE_FULL_HSF:mode>"
   9354   [(set (match_operand:SVE_FULL_HSF 0 "register_operand" "=w, ?&w, ?&w")
   9355 	(unspec:SVE_FULL_HSF
   9356 	  [(match_operand:<SVE_FULL_SDF:VPRED> 1 "register_operand" "Upl, Upl, Upl")
   9357 	   (unspec:SVE_FULL_HSF
   9358 	     [(match_dup 1)
   9359 	      (match_operand:SI 4 "aarch64_sve_gp_strictness")
   9360 	      (match_operand:SVE_FULL_SDF 2 "register_operand" "w, w, w")]
   9361 	     SVE_COND_FCVT)
   9362 	   (match_operand:SVE_FULL_HSF 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   9363 	  UNSPEC_SEL))]
   9364   "TARGET_SVE && <SVE_FULL_SDF:elem_bits> > <SVE_FULL_HSF:elem_bits>"
   9365   "@
   9366    fcvt\t%0.<SVE_FULL_HSF:Vetype>, %1/m, %2.<SVE_FULL_SDF:Vetype>
   9367    movprfx\t%0.<SVE_FULL_SDF:Vetype>, %1/z, %2.<SVE_FULL_SDF:Vetype>\;fcvt\t%0.<SVE_FULL_HSF:Vetype>, %1/m, %2.<SVE_FULL_SDF:Vetype>
   9368    movprfx\t%0, %3\;fcvt\t%0.<SVE_FULL_HSF:Vetype>, %1/m, %2.<SVE_FULL_SDF:Vetype>"
   9369   [(set_attr "movprfx" "*,yes,yes")]
   9370 )
   9371 
   9372 ;; -------------------------------------------------------------------------
   9373 ;; ---- [FP<-FP] Packs (bfloat16)
   9374 ;; -------------------------------------------------------------------------
   9375 ;; Includes:
   9376 ;; - BFCVT (BF16)
   9377 ;; - BFCVTNT (BF16)
   9378 ;; -------------------------------------------------------------------------
   9379 
   9380 ;; Predicated BFCVT.
   9381 (define_insn "@aarch64_sve_<optab>_trunc<VNx4SF_ONLY:mode><VNx8BF_ONLY:mode>"
   9382   [(set (match_operand:VNx8BF_ONLY 0 "register_operand" "=w, ?&w")
   9383 	(unspec:VNx8BF_ONLY
   9384 	  [(match_operand:VNx4BI 1 "register_operand" "Upl, Upl")
   9385 	   (match_operand:SI 3 "aarch64_sve_gp_strictness")
   9386 	   (match_operand:VNx4SF_ONLY 2 "register_operand" "0, w")]
   9387 	  SVE_COND_FCVT))]
   9388   "TARGET_SVE_BF16"
   9389   "@
   9390    bfcvt\t%0.h, %1/m, %2.s
   9391    movprfx\t%0, %2\;bfcvt\t%0.h, %1/m, %2.s"
   9392   [(set_attr "movprfx" "*,yes")]
   9393 )
   9394 
   9395 ;; Predicated BFCVT with merging.
   9396 (define_expand "@cond_<optab>_trunc<VNx4SF_ONLY:mode><VNx8BF_ONLY:mode>"
   9397   [(set (match_operand:VNx8BF_ONLY 0 "register_operand")
   9398 	(unspec:VNx8BF_ONLY
   9399 	  [(match_operand:VNx4BI 1 "register_operand")
   9400 	   (unspec:VNx8BF_ONLY
   9401 	     [(match_dup 1)
   9402 	      (const_int SVE_STRICT_GP)
   9403 	      (match_operand:VNx4SF_ONLY 2 "register_operand")]
   9404 	     SVE_COND_FCVT)
   9405 	   (match_operand:VNx8BF_ONLY 3 "aarch64_simd_reg_or_zero")]
   9406 	  UNSPEC_SEL))]
   9407   "TARGET_SVE_BF16"
   9408 )
   9409 
   9410 (define_insn "*cond_<optab>_trunc<VNx4SF_ONLY:mode><VNx8BF_ONLY:mode>"
   9411   [(set (match_operand:VNx8BF_ONLY 0 "register_operand" "=w, ?&w, ?&w")
   9412 	(unspec:VNx8BF_ONLY
   9413 	  [(match_operand:VNx4BI 1 "register_operand" "Upl, Upl, Upl")
   9414 	   (unspec:VNx8BF_ONLY
   9415 	     [(match_dup 1)
   9416 	      (match_operand:SI 4 "aarch64_sve_gp_strictness")
   9417 	      (match_operand:VNx4SF_ONLY 2 "register_operand" "w, w, w")]
   9418 	     SVE_COND_FCVT)
   9419 	   (match_operand:VNx8BF_ONLY 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   9420 	  UNSPEC_SEL))]
   9421   "TARGET_SVE_BF16"
   9422   "@
   9423    bfcvt\t%0.h, %1/m, %2.s
   9424    movprfx\t%0.s, %1/z, %2.s\;bfcvt\t%0.h, %1/m, %2.s
   9425    movprfx\t%0, %3\;bfcvt\t%0.h, %1/m, %2.s"
   9426   [(set_attr "movprfx" "*,yes,yes")]
   9427 )
   9428 
   9429 ;; Predicated BFCVTNT.  This doesn't give a natural aarch64_pred_*/cond_*
   9430 ;; pair because the even elements always have to be supplied for active
   9431 ;; elements, even if the inactive elements don't matter.
   9432 ;;
   9433 ;; This instructions does not take MOVPRFX.
   9434 (define_insn "@aarch64_sve_cvtnt<mode>"
   9435   [(set (match_operand:VNx8BF_ONLY 0 "register_operand" "=w")
   9436 	(unspec:VNx8BF_ONLY
   9437 	  [(match_operand:VNx4BI 2 "register_operand" "Upl")
   9438 	   (const_int SVE_STRICT_GP)
   9439 	   (match_operand:VNx8BF_ONLY 1 "register_operand" "0")
   9440 	   (match_operand:VNx4SF 3 "register_operand" "w")]
   9441 	  UNSPEC_COND_FCVTNT))]
   9442   "TARGET_SVE_BF16"
   9443   "bfcvtnt\t%0.h, %2/m, %3.s"
   9444 )
   9445 
   9446 ;; -------------------------------------------------------------------------
   9447 ;; ---- [FP<-FP] Unpacks
   9448 ;; -------------------------------------------------------------------------
   9449 ;; Includes:
   9450 ;; - FCVT
   9451 ;; -------------------------------------------------------------------------
   9452 
   9453 ;; Unpack one half of a VNx4SF to VNx2DF, or one half of a VNx8HF to VNx4SF.
   9454 ;; First unpack the source without conversion, then float-convert the
   9455 ;; unpacked source.
   9456 (define_expand "vec_unpacks_<perm_hilo>_<mode>"
   9457   [(match_operand:<VWIDE> 0 "register_operand")
   9458    (unspec:SVE_FULL_HSF
   9459      [(match_operand:SVE_FULL_HSF 1 "register_operand")]
   9460      UNPACK_UNSIGNED)]
   9461   "TARGET_SVE"
   9462   {
   9463     /* Use ZIP to do the unpack, since we don't care about the upper halves
   9464        and since it has the nice property of not needing any subregs.
   9465        If using UUNPK* turns out to be preferable, we could model it as
   9466        a ZIP whose first operand is zero.  */
   9467     rtx temp = gen_reg_rtx (<MODE>mode);
   9468     emit_insn ((<hi_lanes_optab>
   9469 		? gen_aarch64_sve_zip2<mode>
   9470 		: gen_aarch64_sve_zip1<mode>)
   9471 		(temp, operands[1], operands[1]));
   9472     rtx ptrue = aarch64_ptrue_reg (<VWIDE_PRED>mode);
   9473     rtx strictness = gen_int_mode (SVE_RELAXED_GP, SImode);
   9474     emit_insn (gen_aarch64_sve_fcvt_nontrunc<mode><Vwide>
   9475 	       (operands[0], ptrue, temp, strictness));
   9476     DONE;
   9477   }
   9478 )
   9479 
   9480 ;; Predicated float-to-float extension.
   9481 (define_insn "@aarch64_sve_<optab>_nontrunc<SVE_FULL_HSF:mode><SVE_FULL_SDF:mode>"
   9482   [(set (match_operand:SVE_FULL_SDF 0 "register_operand" "=w, ?&w")
   9483 	(unspec:SVE_FULL_SDF
   9484 	  [(match_operand:<SVE_FULL_SDF:VPRED> 1 "register_operand" "Upl, Upl")
   9485 	   (match_operand:SI 3 "aarch64_sve_gp_strictness")
   9486 	   (match_operand:SVE_FULL_HSF 2 "register_operand" "0, w")]
   9487 	  SVE_COND_FCVT))]
   9488   "TARGET_SVE && <SVE_FULL_SDF:elem_bits> > <SVE_FULL_HSF:elem_bits>"
   9489   "@
   9490    fcvt\t%0.<SVE_FULL_SDF:Vetype>, %1/m, %2.<SVE_FULL_HSF:Vetype>
   9491    movprfx\t%0, %2\;fcvt\t%0.<SVE_FULL_SDF:Vetype>, %1/m, %2.<SVE_FULL_HSF:Vetype>"
   9492   [(set_attr "movprfx" "*,yes")]
   9493 )
   9494 
   9495 ;; Predicated float-to-float extension with merging.
   9496 (define_expand "@cond_<optab>_nontrunc<SVE_FULL_HSF:mode><SVE_FULL_SDF:mode>"
   9497   [(set (match_operand:SVE_FULL_SDF 0 "register_operand")
   9498 	(unspec:SVE_FULL_SDF
   9499 	  [(match_operand:<SVE_FULL_SDF:VPRED> 1 "register_operand")
   9500 	   (unspec:SVE_FULL_SDF
   9501 	     [(match_dup 1)
   9502 	      (const_int SVE_STRICT_GP)
   9503 	      (match_operand:SVE_FULL_HSF 2 "register_operand")]
   9504 	     SVE_COND_FCVT)
   9505 	   (match_operand:SVE_FULL_SDF 3 "aarch64_simd_reg_or_zero")]
   9506 	  UNSPEC_SEL))]
   9507   "TARGET_SVE && <SVE_FULL_SDF:elem_bits> > <SVE_FULL_HSF:elem_bits>"
   9508 )
   9509 
   9510 (define_insn "*cond_<optab>_nontrunc<SVE_FULL_HSF:mode><SVE_FULL_SDF:mode>"
   9511   [(set (match_operand:SVE_FULL_SDF 0 "register_operand" "=w, ?&w, ?&w")
   9512 	(unspec:SVE_FULL_SDF
   9513 	  [(match_operand:<SVE_FULL_SDF:VPRED> 1 "register_operand" "Upl, Upl, Upl")
   9514 	   (unspec:SVE_FULL_SDF
   9515 	     [(match_dup 1)
   9516 	      (match_operand:SI 4 "aarch64_sve_gp_strictness")
   9517 	      (match_operand:SVE_FULL_HSF 2 "register_operand" "w, w, w")]
   9518 	     SVE_COND_FCVT)
   9519 	   (match_operand:SVE_FULL_SDF 3 "aarch64_simd_reg_or_zero" "0, Dz, w")]
   9520 	  UNSPEC_SEL))]
   9521   "TARGET_SVE && <SVE_FULL_SDF:elem_bits> > <SVE_FULL_HSF:elem_bits>"
   9522   "@
   9523    fcvt\t%0.<SVE_FULL_SDF:Vetype>, %1/m, %2.<SVE_FULL_HSF:Vetype>
   9524    movprfx\t%0.<SVE_FULL_SDF:Vetype>, %1/z, %2.<SVE_FULL_SDF:Vetype>\;fcvt\t%0.<SVE_FULL_SDF:Vetype>, %1/m, %2.<SVE_FULL_HSF:Vetype>
   9525    movprfx\t%0, %3\;fcvt\t%0.<SVE_FULL_SDF:Vetype>, %1/m, %2.<SVE_FULL_HSF:Vetype>"
   9526   [(set_attr "movprfx" "*,yes,yes")]
   9527 )
   9528 
   9529 ;; -------------------------------------------------------------------------
   9530 ;; ---- [PRED<-PRED] Packs
   9531 ;; -------------------------------------------------------------------------
   9532 ;; Includes:
   9533 ;; - UZP1
   9534 ;; -------------------------------------------------------------------------
   9535 
   9536 ;; Predicate pack.  Use UZP1 on the narrower type, which discards
   9537 ;; the high part of each wide element.
   9538 (define_insn "vec_pack_trunc_<Vwide>"
   9539   [(set (match_operand:PRED_BHS 0 "register_operand" "=Upa")
   9540 	(unspec:PRED_BHS
   9541 	  [(match_operand:<VWIDE> 1 "register_operand" "Upa")
   9542 	   (match_operand:<VWIDE> 2 "register_operand" "Upa")]
   9543 	  UNSPEC_PACK))]
   9544   "TARGET_SVE"
   9545   "uzp1\t%0.<Vetype>, %1.<Vetype>, %2.<Vetype>"
   9546 )
   9547 
   9548 ;; -------------------------------------------------------------------------
   9549 ;; ---- [PRED<-PRED] Unpacks
   9550 ;; -------------------------------------------------------------------------
   9551 ;; Includes:
   9552 ;; - PUNPKHI
   9553 ;; - PUNPKLO
   9554 ;; -------------------------------------------------------------------------
   9555 
   9556 ;; Unpack the low or high half of a predicate, where "high" refers to
   9557 ;; the low-numbered lanes for big-endian and the high-numbered lanes
   9558 ;; for little-endian.
   9559 (define_expand "vec_unpack<su>_<perm_hilo>_<mode>"
   9560   [(match_operand:<VWIDE> 0 "register_operand")
   9561    (unspec:<VWIDE> [(match_operand:PRED_BHS 1 "register_operand")]
   9562 		   UNPACK)]
   9563   "TARGET_SVE"
   9564   {
   9565     emit_insn ((<hi_lanes_optab>
   9566 		? gen_aarch64_sve_punpkhi_<PRED_BHS:mode>
   9567 		: gen_aarch64_sve_punpklo_<PRED_BHS:mode>)
   9568 	       (operands[0], operands[1]));
   9569     DONE;
   9570   }
   9571 )
   9572 
   9573 (define_insn "@aarch64_sve_punpk<perm_hilo>_<mode>"
   9574   [(set (match_operand:<VWIDE> 0 "register_operand" "=Upa")
   9575 	(unspec:<VWIDE> [(match_operand:PRED_BHS 1 "register_operand" "Upa")]
   9576 			UNPACK_UNSIGNED))]
   9577   "TARGET_SVE"
   9578   "punpk<perm_hilo>\t%0.h, %1.b"
   9579 )
   9580 
   9581 ;; =========================================================================
   9582 ;; == Vector partitioning
   9583 ;; =========================================================================
   9584 
   9585 ;; -------------------------------------------------------------------------
   9586 ;; ---- [PRED] Unary partitioning
   9587 ;; -------------------------------------------------------------------------
   9588 ;; Includes:
   9589 ;; - BRKA
   9590 ;; - BRKAS
   9591 ;; - BRKB
   9592 ;; - BRKBS
   9593 ;; -------------------------------------------------------------------------
   9594 
   9595 ;; Note that unlike most other instructions that have both merging and
   9596 ;; zeroing forms, these instructions don't operate elementwise and so
   9597 ;; don't fit the IFN_COND model.
   9598 (define_insn "@aarch64_brk<brk_op>"
   9599   [(set (match_operand:VNx16BI 0 "register_operand" "=Upa, Upa")
   9600 	(unspec:VNx16BI
   9601 	  [(match_operand:VNx16BI 1 "register_operand" "Upa, Upa")
   9602 	   (match_operand:VNx16BI 2 "register_operand" "Upa, Upa")
   9603 	   (match_operand:VNx16BI 3 "aarch64_simd_reg_or_zero" "Dz, 0")]
   9604 	  SVE_BRK_UNARY))]
   9605   "TARGET_SVE"
   9606   "@
   9607    brk<brk_op>\t%0.b, %1/z, %2.b
   9608    brk<brk_op>\t%0.b, %1/m, %2.b"
   9609 )
   9610 
   9611 ;; Same, but also producing a flags result.
   9612 (define_insn "*aarch64_brk<brk_op>_cc"
   9613   [(set (reg:CC_NZC CC_REGNUM)
   9614 	(unspec:CC_NZC
   9615 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   9616 	   (match_dup 1)
   9617 	   (match_operand:SI 4 "aarch64_sve_ptrue_flag")
   9618 	   (unspec:VNx16BI
   9619 	     [(match_dup 1)
   9620 	      (match_operand:VNx16BI 2 "register_operand" "Upa")
   9621 	      (match_operand:VNx16BI 3 "aarch64_simd_imm_zero")]
   9622 	     SVE_BRK_UNARY)]
   9623 	  UNSPEC_PTEST))
   9624    (set (match_operand:VNx16BI 0 "register_operand" "=Upa")
   9625 	(unspec:VNx16BI
   9626 	  [(match_dup 1)
   9627 	   (match_dup 2)
   9628 	   (match_dup 3)]
   9629 	  SVE_BRK_UNARY))]
   9630   "TARGET_SVE"
   9631   "brk<brk_op>s\t%0.b, %1/z, %2.b"
   9632 )
   9633 
   9634 ;; Same, but with only the flags result being interesting.
   9635 (define_insn "*aarch64_brk<brk_op>_ptest"
   9636   [(set (reg:CC_NZC CC_REGNUM)
   9637 	(unspec:CC_NZC
   9638 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   9639 	   (match_dup 1)
   9640 	   (match_operand:SI 4 "aarch64_sve_ptrue_flag")
   9641 	   (unspec:VNx16BI
   9642 	     [(match_dup 1)
   9643 	      (match_operand:VNx16BI 2 "register_operand" "Upa")
   9644 	      (match_operand:VNx16BI 3 "aarch64_simd_imm_zero")]
   9645 	     SVE_BRK_UNARY)]
   9646 	  UNSPEC_PTEST))
   9647    (clobber (match_scratch:VNx16BI 0 "=Upa"))]
   9648   "TARGET_SVE"
   9649   "brk<brk_op>s\t%0.b, %1/z, %2.b"
   9650 )
   9651 
   9652 ;; -------------------------------------------------------------------------
   9653 ;; ---- [PRED] Binary partitioning
   9654 ;; -------------------------------------------------------------------------
   9655 ;; Includes:
   9656 ;; - BRKN
   9657 ;; - BRKNS
   9658 ;; - BRKPA
   9659 ;; - BRKPAS
   9660 ;; - BRKPB
   9661 ;; - BRKPBS
   9662 ;; -------------------------------------------------------------------------
   9663 
   9664 ;; Binary BRKs (BRKN, BRKPA, BRKPB).
   9665 (define_insn "@aarch64_brk<brk_op>"
   9666   [(set (match_operand:VNx16BI 0 "register_operand" "=Upa")
   9667 	(unspec:VNx16BI
   9668 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   9669 	   (match_operand:VNx16BI 2 "register_operand" "Upa")
   9670 	   (match_operand:VNx16BI 3 "register_operand" "<brk_reg_con>")]
   9671 	  SVE_BRK_BINARY))]
   9672   "TARGET_SVE"
   9673   "brk<brk_op>\t%0.b, %1/z, %2.b, %<brk_reg_opno>.b"
   9674 )
   9675 
   9676 ;; BRKN, producing both a predicate and a flags result.  Unlike other
   9677 ;; flag-setting instructions, these flags are always set wrt a ptrue.
   9678 (define_insn_and_rewrite "*aarch64_brkn_cc"
   9679   [(set (reg:CC_NZC CC_REGNUM)
   9680 	(unspec:CC_NZC
   9681 	  [(match_operand:VNx16BI 4)
   9682 	   (match_operand:VNx16BI 5)
   9683 	   (const_int SVE_KNOWN_PTRUE)
   9684 	   (unspec:VNx16BI
   9685 	     [(match_operand:VNx16BI 1 "register_operand" "Upa")
   9686 	      (match_operand:VNx16BI 2 "register_operand" "Upa")
   9687 	      (match_operand:VNx16BI 3 "register_operand" "0")]
   9688 	     UNSPEC_BRKN)]
   9689 	  UNSPEC_PTEST))
   9690    (set (match_operand:VNx16BI 0 "register_operand" "=Upa")
   9691 	(unspec:VNx16BI
   9692 	  [(match_dup 1)
   9693 	   (match_dup 2)
   9694 	   (match_dup 3)]
   9695 	  UNSPEC_BRKN))]
   9696   "TARGET_SVE"
   9697   "brkns\t%0.b, %1/z, %2.b, %0.b"
   9698   "&& (operands[4] != CONST0_RTX (VNx16BImode)
   9699        || operands[5] != CONST0_RTX (VNx16BImode))"
   9700   {
   9701     operands[4] = CONST0_RTX (VNx16BImode);
   9702     operands[5] = CONST0_RTX (VNx16BImode);
   9703   }
   9704 )
   9705 
   9706 ;; Same, but with only the flags result being interesting.
   9707 (define_insn_and_rewrite "*aarch64_brkn_ptest"
   9708   [(set (reg:CC_NZC CC_REGNUM)
   9709 	(unspec:CC_NZC
   9710 	  [(match_operand:VNx16BI 4)
   9711 	   (match_operand:VNx16BI 5)
   9712 	   (const_int SVE_KNOWN_PTRUE)
   9713 	   (unspec:VNx16BI
   9714 	     [(match_operand:VNx16BI 1 "register_operand" "Upa")
   9715 	      (match_operand:VNx16BI 2 "register_operand" "Upa")
   9716 	      (match_operand:VNx16BI 3 "register_operand" "0")]
   9717 	     UNSPEC_BRKN)]
   9718 	  UNSPEC_PTEST))
   9719    (clobber (match_scratch:VNx16BI 0 "=Upa"))]
   9720   "TARGET_SVE"
   9721   "brkns\t%0.b, %1/z, %2.b, %0.b"
   9722   "&& (operands[4] != CONST0_RTX (VNx16BImode)
   9723        || operands[5] != CONST0_RTX (VNx16BImode))"
   9724   {
   9725     operands[4] = CONST0_RTX (VNx16BImode);
   9726     operands[5] = CONST0_RTX (VNx16BImode);
   9727   }
   9728 )
   9729 
   9730 ;; BRKPA and BRKPB, producing both a predicate and a flags result.
   9731 (define_insn "*aarch64_brk<brk_op>_cc"
   9732   [(set (reg:CC_NZC CC_REGNUM)
   9733 	(unspec:CC_NZC
   9734 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   9735 	   (match_dup 1)
   9736 	   (match_operand:SI 4 "aarch64_sve_ptrue_flag")
   9737 	   (unspec:VNx16BI
   9738 	     [(match_dup 1)
   9739 	      (match_operand:VNx16BI 2 "register_operand" "Upa")
   9740 	      (match_operand:VNx16BI 3 "register_operand" "Upa")]
   9741 	     SVE_BRKP)]
   9742 	  UNSPEC_PTEST))
   9743    (set (match_operand:VNx16BI 0 "register_operand" "=Upa")
   9744 	(unspec:VNx16BI
   9745 	  [(match_dup 1)
   9746 	   (match_dup 2)
   9747 	   (match_dup 3)]
   9748 	  SVE_BRKP))]
   9749   "TARGET_SVE"
   9750   "brk<brk_op>s\t%0.b, %1/z, %2.b, %3.b"
   9751 )
   9752 
   9753 ;; Same, but with only the flags result being interesting.
   9754 (define_insn "*aarch64_brk<brk_op>_ptest"
   9755   [(set (reg:CC_NZC CC_REGNUM)
   9756 	(unspec:CC_NZC
   9757 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   9758 	   (match_dup 1)
   9759 	   (match_operand:SI 4 "aarch64_sve_ptrue_flag")
   9760 	   (unspec:VNx16BI
   9761 	     [(match_dup 1)
   9762 	      (match_operand:VNx16BI 2 "register_operand" "Upa")
   9763 	      (match_operand:VNx16BI 3 "register_operand" "Upa")]
   9764 	     SVE_BRKP)]
   9765 	  UNSPEC_PTEST))
   9766    (clobber (match_scratch:VNx16BI 0 "=Upa"))]
   9767   "TARGET_SVE"
   9768   "brk<brk_op>s\t%0.b, %1/z, %2.b, %3.b"
   9769 )
   9770 
   9771 ;; -------------------------------------------------------------------------
   9772 ;; ---- [PRED] Scalarization
   9773 ;; -------------------------------------------------------------------------
   9774 ;; Includes:
   9775 ;; - PFIRST
   9776 ;; - PNEXT
   9777 ;; -------------------------------------------------------------------------
   9778 
   9779 (define_insn "@aarch64_sve_<sve_pred_op><mode>"
   9780   [(set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   9781 	(unspec:PRED_ALL
   9782 	  [(match_operand:PRED_ALL 1 "register_operand" "Upa")
   9783 	   (match_operand:SI 2 "aarch64_sve_ptrue_flag")
   9784 	   (match_operand:PRED_ALL 3 "register_operand" "0")]
   9785 	  SVE_PITER))
   9786    (clobber (reg:CC_NZC CC_REGNUM))]
   9787   "TARGET_SVE && <max_elem_bits> >= <elem_bits>"
   9788   "<sve_pred_op>\t%0.<Vetype>, %1, %0.<Vetype>"
   9789 )
   9790 
   9791 ;; Same, but also producing a flags result.
   9792 (define_insn_and_rewrite "*aarch64_sve_<sve_pred_op><mode>_cc"
   9793   [(set (reg:CC_NZC CC_REGNUM)
   9794 	(unspec:CC_NZC
   9795 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   9796 	   (match_operand 2)
   9797 	   (match_operand:SI 3 "aarch64_sve_ptrue_flag")
   9798 	   (unspec:PRED_ALL
   9799 	     [(match_operand 4)
   9800 	      (match_operand:SI 5 "aarch64_sve_ptrue_flag")
   9801 	      (match_operand:PRED_ALL 6 "register_operand" "0")]
   9802 	     SVE_PITER)]
   9803 	  UNSPEC_PTEST))
   9804    (set (match_operand:PRED_ALL 0 "register_operand" "=Upa")
   9805 	(unspec:PRED_ALL
   9806 	  [(match_dup 4)
   9807 	   (match_dup 5)
   9808 	   (match_dup 6)]
   9809 	  SVE_PITER))]
   9810   "TARGET_SVE
   9811    && <max_elem_bits> >= <elem_bits>
   9812    && aarch64_sve_same_pred_for_ptest_p (&operands[2], &operands[4])"
   9813   "<sve_pred_op>\t%0.<Vetype>, %1, %0.<Vetype>"
   9814   "&& !rtx_equal_p (operands[2], operands[4])"
   9815   {
   9816     operands[4] = operands[2];
   9817     operands[5] = operands[3];
   9818   }
   9819 )
   9820 
   9821 ;; Same, but with only the flags result being interesting.
   9822 (define_insn_and_rewrite "*aarch64_sve_<sve_pred_op><mode>_ptest"
   9823   [(set (reg:CC_NZC CC_REGNUM)
   9824 	(unspec:CC_NZC
   9825 	  [(match_operand:VNx16BI 1 "register_operand" "Upa")
   9826 	   (match_operand 2)
   9827 	   (match_operand:SI 3 "aarch64_sve_ptrue_flag")
   9828 	   (unspec:PRED_ALL
   9829 	     [(match_operand 4)
   9830 	      (match_operand:SI 5 "aarch64_sve_ptrue_flag")
   9831 	      (match_operand:PRED_ALL 6 "register_operand" "0")]
   9832 	     SVE_PITER)]
   9833 	  UNSPEC_PTEST))
   9834    (clobber (match_scratch:PRED_ALL 0 "=Upa"))]
   9835   "TARGET_SVE
   9836    && <max_elem_bits> >= <elem_bits>
   9837    && aarch64_sve_same_pred_for_ptest_p (&operands[2], &operands[4])"
   9838   "<sve_pred_op>\t%0.<Vetype>, %1, %0.<Vetype>"
   9839   "&& !rtx_equal_p (operands[2], operands[4])"
   9840   {
   9841     operands[4] = operands[2];
   9842     operands[5] = operands[3];
   9843   }
   9844 )
   9845 
   9846 ;; =========================================================================
   9847 ;; == Counting elements
   9848 ;; =========================================================================
   9849 
   9850 ;; -------------------------------------------------------------------------
   9851 ;; ---- [INT] Count elements in a pattern (scalar)
   9852 ;; -------------------------------------------------------------------------
   9853 ;; Includes:
   9854 ;; - CNTB
   9855 ;; - CNTD
   9856 ;; - CNTH
   9857 ;; - CNTW
   9858 ;; -------------------------------------------------------------------------
   9859 
   9860 ;; Count the number of elements in an svpattern.  Operand 1 is the pattern,
   9861 ;; operand 2 is the number of elements that fit in a 128-bit block, and
   9862 ;; operand 3 is a multiplier in the range [1, 16].
   9863 ;;
   9864 ;; Note that this pattern isn't used for SV_ALL (but would work for that too).
   9865 (define_insn "aarch64_sve_cnt_pat"
   9866   [(set (match_operand:DI 0 "register_operand" "=r")
   9867 	(zero_extend:DI
   9868 	  (unspec:SI [(match_operand:DI 1 "const_int_operand")
   9869 		      (match_operand:DI 2 "const_int_operand")
   9870 		      (match_operand:DI 3 "const_int_operand")]
   9871 		     UNSPEC_SVE_CNT_PAT)))]
   9872   "TARGET_SVE"
   9873   {
   9874     return aarch64_output_sve_cnt_pat_immediate ("cnt", "%x0", operands + 1);
   9875   }
   9876 )
   9877 
   9878 ;; -------------------------------------------------------------------------
   9879 ;; ---- [INT] Increment by the number of elements in a pattern (scalar)
   9880 ;; -------------------------------------------------------------------------
   9881 ;; Includes:
   9882 ;; - INC
   9883 ;; - SQINC
   9884 ;; - UQINC
   9885 ;; -------------------------------------------------------------------------
   9886 
   9887 ;; Increment a DImode register by the number of elements in an svpattern.
   9888 ;; See aarch64_sve_cnt_pat for the counting behavior.
   9889 (define_insn "@aarch64_sve_<inc_dec><mode>_pat"
   9890   [(set (match_operand:DI 0 "register_operand" "=r")
   9891 	(ANY_PLUS:DI (zero_extend:DI
   9892 		       (unspec:SI [(match_operand:DI 2 "const_int_operand")
   9893 				   (match_operand:DI 3 "const_int_operand")
   9894 				   (match_operand:DI 4 "const_int_operand")]
   9895 				  UNSPEC_SVE_CNT_PAT))
   9896 		     (match_operand:DI_ONLY 1 "register_operand" "0")))]
   9897   "TARGET_SVE"
   9898   {
   9899     return aarch64_output_sve_cnt_pat_immediate ("<inc_dec>", "%x0",
   9900 						 operands + 2);
   9901   }
   9902 )
   9903 
   9904 ;; Increment an SImode register by the number of elements in an svpattern
   9905 ;; using modular arithmetic.  See aarch64_sve_cnt_pat for the counting
   9906 ;; behavior.
   9907 (define_insn "*aarch64_sve_incsi_pat"
   9908   [(set (match_operand:SI 0 "register_operand" "=r")
   9909 	(plus:SI (unspec:SI [(match_operand:DI 2 "const_int_operand")
   9910 			     (match_operand:DI 3 "const_int_operand")
   9911 			     (match_operand:DI 4 "const_int_operand")]
   9912 			    UNSPEC_SVE_CNT_PAT)
   9913 		 (match_operand:SI 1 "register_operand" "0")))]
   9914   "TARGET_SVE"
   9915   {
   9916     return aarch64_output_sve_cnt_pat_immediate ("inc", "%x0", operands + 2);
   9917   }
   9918 )
   9919 
   9920 ;; Increment an SImode register by the number of elements in an svpattern
   9921 ;; using saturating arithmetic, extending the result to 64 bits.
   9922 ;;
   9923 ;; See aarch64_sve_cnt_pat for the counting behavior.
   9924 (define_insn "@aarch64_sve_<inc_dec><mode>_pat"
   9925   [(set (match_operand:DI 0 "register_operand" "=r")
   9926 	(<paired_extend>:DI
   9927 	  (SAT_PLUS:SI
   9928 	    (unspec:SI [(match_operand:DI 2 "const_int_operand")
   9929 			(match_operand:DI 3 "const_int_operand")
   9930 			(match_operand:DI 4 "const_int_operand")]
   9931 		       UNSPEC_SVE_CNT_PAT)
   9932 	    (match_operand:SI_ONLY 1 "register_operand" "0"))))]
   9933   "TARGET_SVE"
   9934   {
   9935     const char *registers = (<CODE> == SS_PLUS ? "%x0, %w0" : "%w0");
   9936     return aarch64_output_sve_cnt_pat_immediate ("<inc_dec>", registers,
   9937 						 operands + 2);
   9938   }
   9939 )
   9940 
   9941 ;; -------------------------------------------------------------------------
   9942 ;; ---- [INT] Increment by the number of elements in a pattern (vector)
   9943 ;; -------------------------------------------------------------------------
   9944 ;; Includes:
   9945 ;; - INC
   9946 ;; - SQINC
   9947 ;; - UQINC
   9948 ;; -------------------------------------------------------------------------
   9949 
   9950 ;; Increment a vector of DIs by the number of elements in an svpattern.
   9951 ;; See aarch64_sve_cnt_pat for the counting behavior.
   9952 (define_insn "@aarch64_sve_<inc_dec><mode>_pat"
   9953   [(set (match_operand:VNx2DI 0 "register_operand" "=w, ?&w")
   9954 	(ANY_PLUS:VNx2DI
   9955 	  (vec_duplicate:VNx2DI
   9956 	    (zero_extend:DI
   9957 	      (unspec:SI [(match_operand:DI 2 "const_int_operand")
   9958 			  (match_operand:DI 3 "const_int_operand")
   9959 			  (match_operand:DI 4 "const_int_operand")]
   9960 			 UNSPEC_SVE_CNT_PAT)))
   9961 	  (match_operand:VNx2DI_ONLY 1 "register_operand" "0, w")))]
   9962   "TARGET_SVE"
   9963   {
   9964     if (which_alternative == 1)
   9965       output_asm_insn ("movprfx\t%0, %1", operands);
   9966     return aarch64_output_sve_cnt_pat_immediate ("<inc_dec>", "%0.<Vetype>",
   9967 						 operands + 2);
   9968   }
   9969   [(set_attr "movprfx" "*,yes")]
   9970 )
   9971 
   9972 ;; Increment a vector of SIs by the number of elements in an svpattern.
   9973 ;; See aarch64_sve_cnt_pat for the counting behavior.
   9974 (define_insn "@aarch64_sve_<inc_dec><mode>_pat"
   9975   [(set (match_operand:VNx4SI 0 "register_operand" "=w, ?&w")
   9976 	(ANY_PLUS:VNx4SI
   9977 	  (vec_duplicate:VNx4SI
   9978 	    (unspec:SI [(match_operand:DI 2 "const_int_operand")
   9979 			(match_operand:DI 3 "const_int_operand")
   9980 			(match_operand:DI 4 "const_int_operand")]
   9981 		       UNSPEC_SVE_CNT_PAT))
   9982 	  (match_operand:VNx4SI_ONLY 1 "register_operand" "0, w")))]
   9983   "TARGET_SVE"
   9984   {
   9985     if (which_alternative == 1)
   9986       output_asm_insn ("movprfx\t%0, %1", operands);
   9987     return aarch64_output_sve_cnt_pat_immediate ("<inc_dec>", "%0.<Vetype>",
   9988 						 operands + 2);
   9989   }
   9990   [(set_attr "movprfx" "*,yes")]
   9991 )
   9992 
   9993 ;; Increment a vector of HIs by the number of elements in an svpattern.
   9994 ;; See aarch64_sve_cnt_pat for the counting behavior.
   9995 (define_expand "@aarch64_sve_<inc_dec><mode>_pat"
   9996   [(set (match_operand:VNx8HI 0 "register_operand")
   9997 	(ANY_PLUS:VNx8HI
   9998 	  (vec_duplicate:VNx8HI
   9999 	    (truncate:HI
   10000 	      (unspec:SI [(match_operand:DI 2 "const_int_operand")
   10001 			  (match_operand:DI 3 "const_int_operand")
   10002 			  (match_operand:DI 4 "const_int_operand")]
   10003 			 UNSPEC_SVE_CNT_PAT)))
   10004 	  (match_operand:VNx8HI_ONLY 1 "register_operand")))]
   10005   "TARGET_SVE"
   10006 )
   10007 
   10008 (define_insn "*aarch64_sve_<inc_dec><mode>_pat"
   10009   [(set (match_operand:VNx8HI 0 "register_operand" "=w, ?&w")
   10010 	(ANY_PLUS:VNx8HI
   10011 	  (vec_duplicate:VNx8HI
   10012 	    (match_operator:HI 5 "subreg_lowpart_operator"
   10013 	      [(unspec:SI [(match_operand:DI 2 "const_int_operand")
   10014 			   (match_operand:DI 3 "const_int_operand")
   10015 			   (match_operand:DI 4 "const_int_operand")]
   10016 			  UNSPEC_SVE_CNT_PAT)]))
   10017 	  (match_operand:VNx8HI_ONLY 1 "register_operand" "0, w")))]
   10018   "TARGET_SVE"
   10019   {
   10020     if (which_alternative == 1)
   10021       output_asm_insn ("movprfx\t%0, %1", operands);
   10022     return aarch64_output_sve_cnt_pat_immediate ("<inc_dec>", "%0.<Vetype>",
   10023 						 operands + 2);
   10024   }
   10025   [(set_attr "movprfx" "*,yes")]
   10026 )
   10027 
   10028 ;; -------------------------------------------------------------------------
   10029 ;; ---- [INT] Decrement by the number of elements in a pattern (scalar)
   10030 ;; -------------------------------------------------------------------------
   10031 ;; Includes:
   10032 ;; - DEC
   10033 ;; - SQDEC
   10034 ;; - UQDEC
   10035 ;; -------------------------------------------------------------------------
   10036 
   10037 ;; Decrement a DImode register by the number of elements in an svpattern.
   10038 ;; See aarch64_sve_cnt_pat for the counting behavior.
   10039 (define_insn "@aarch64_sve_<inc_dec><mode>_pat"
   10040   [(set (match_operand:DI 0 "register_operand" "=r")
   10041 	(ANY_MINUS:DI (match_operand:DI_ONLY 1 "register_operand" "0")
   10042 		      (zero_extend:DI
   10043 			(unspec:SI [(match_operand:DI 2 "const_int_operand")
   10044 				    (match_operand:DI 3 "const_int_operand")
   10045 				    (match_operand:DI 4 "const_int_operand")]
   10046 				   UNSPEC_SVE_CNT_PAT))))]
   10047   "TARGET_SVE"
   10048   {
   10049     return aarch64_output_sve_cnt_pat_immediate ("<inc_dec>", "%x0",
   10050 						 operands + 2);
   10051   }
   10052 )
   10053 
   10054 ;; Decrement an SImode register by the number of elements in an svpattern
   10055 ;; using modular arithmetic.  See aarch64_sve_cnt_pat for the counting
   10056 ;; behavior.
   10057 (define_insn "*aarch64_sve_decsi_pat"
   10058   [(set (match_operand:SI 0 "register_operand" "=r")
   10059 	(minus:SI (match_operand:SI 1 "register_operand" "0")
   10060 		  (unspec:SI [(match_operand:DI 2 "const_int_operand")
   10061 			      (match_operand:DI 3 "const_int_operand")
   10062 			      (match_operand:DI 4 "const_int_operand")]
   10063 			     UNSPEC_SVE_CNT_PAT)))]
   10064   "TARGET_SVE"
   10065   {
   10066     return aarch64_output_sve_cnt_pat_immediate ("dec", "%x0", operands + 2);
   10067   }
   10068 )
   10069 
   10070 ;; Decrement an SImode register by the number of elements in an svpattern
   10071 ;; using saturating arithmetic, extending the result to 64 bits.
   10072 ;;
   10073 ;; See aarch64_sve_cnt_pat for the counting behavior.
   10074 (define_insn "@aarch64_sve_<inc_dec><mode>_pat"
   10075   [(set (match_operand:DI 0 "register_operand" "=r")
   10076 	(<paired_extend>:DI
   10077 	  (SAT_MINUS:SI
   10078 	    (match_operand:SI_ONLY 1 "register_operand" "0")
   10079 	    (unspec:SI [(match_operand:DI 2 "const_int_operand")
   10080 			(match_operand:DI 3 "const_int_operand")
   10081 			(match_operand:DI 4 "const_int_operand")]
   10082 		       UNSPEC_SVE_CNT_PAT))))]
   10083   "TARGET_SVE"
   10084   {
   10085     const char *registers = (<CODE> == SS_MINUS ? "%x0, %w0" : "%w0");
   10086     return aarch64_output_sve_cnt_pat_immediate ("<inc_dec>", registers,
   10087 						 operands + 2);
   10088   }
   10089 )
   10090 
   10091 ;; -------------------------------------------------------------------------
   10092 ;; ---- [INT] Decrement by the number of elements in a pattern (vector)
   10093 ;; -------------------------------------------------------------------------
   10094 ;; Includes:
   10095 ;; - DEC
   10096 ;; - SQDEC
   10097 ;; - UQDEC
   10098 ;; -------------------------------------------------------------------------
   10099 
   10100 ;; Decrement a vector of DIs by the number of elements in an svpattern.
   10101 ;; See aarch64_sve_cnt_pat for the counting behavior.
   10102 (define_insn "@aarch64_sve_<inc_dec><mode>_pat"
   10103   [(set (match_operand:VNx2DI 0 "register_operand" "=w, ?&w")
   10104 	(ANY_MINUS:VNx2DI
   10105 	  (match_operand:VNx2DI_ONLY 1 "register_operand" "0, w")
   10106 	  (vec_duplicate:VNx2DI
   10107 	    (zero_extend:DI
   10108 	      (unspec:SI [(match_operand:DI 2 "const_int_operand")
   10109 			  (match_operand:DI 3 "const_int_operand")
   10110 			  (match_operand:DI 4 "const_int_operand")]
   10111 			 UNSPEC_SVE_CNT_PAT)))))]
   10112   "TARGET_SVE"
   10113   {
   10114     if (which_alternative == 1)
   10115       output_asm_insn ("movprfx\t%0, %1", operands);
   10116     return aarch64_output_sve_cnt_pat_immediate ("<inc_dec>", "%0.<Vetype>",
   10117 						 operands + 2);
   10118   }
   10119   [(set_attr "movprfx" "*,yes")]
   10120 )
   10121 
   10122 ;; Decrement a vector of SIs by the number of elements in an svpattern.
   10123 ;; See aarch64_sve_cnt_pat for the counting behavior.
   10124 (define_insn "@aarch64_sve_<inc_dec><mode>_pat"
   10125   [(set (match_operand:VNx4SI 0 "register_operand" "=w, ?&w")
   10126 	(ANY_MINUS:VNx4SI
   10127 	  (match_operand:VNx4SI_ONLY 1 "register_operand" "0, w")
   10128 	  (vec_duplicate:VNx4SI
   10129 	    (unspec:SI [(match_operand:DI 2 "const_int_operand")
   10130 			(match_operand:DI 3 "const_int_operand")
   10131 			(match_operand:DI 4 "const_int_operand")]
   10132 		       UNSPEC_SVE_CNT_PAT))))]
   10133   "TARGET_SVE"
   10134   {
   10135     if (which_alternative == 1)
   10136       output_asm_insn ("movprfx\t%0, %1", operands);
   10137     return aarch64_output_sve_cnt_pat_immediate ("<inc_dec>", "%0.<Vetype>",
   10138 						 operands + 2);
   10139   }
   10140   [(set_attr "movprfx" "*,yes")]
   10141 )
   10142 
   10143 ;; Decrement a vector of HIs by the number of elements in an svpattern.
   10144 ;; See aarch64_sve_cnt_pat for the counting behavior.
   10145 (define_expand "@aarch64_sve_<inc_dec><mode>_pat"
   10146   [(set (match_operand:VNx8HI 0 "register_operand")
   10147 	(ANY_MINUS:VNx8HI
   10148 	  (match_operand:VNx8HI_ONLY 1 "register_operand")
   10149 	  (vec_duplicate:VNx8HI
   10150 	    (truncate:HI
   10151 	      (unspec:SI [(match_operand:DI 2 "const_int_operand")
   10152 			  (match_operand:DI 3 "const_int_operand")
   10153 			  (match_operand:DI 4 "const_int_operand")]
   10154 			 UNSPEC_SVE_CNT_PAT)))))]
   10155   "TARGET_SVE"
   10156 )
   10157 
   10158 (define_insn "*aarch64_sve_<inc_dec><mode>_pat"
   10159   [(set (match_operand:VNx8HI 0 "register_operand" "=w, ?&w")
   10160 	(ANY_MINUS:VNx8HI
   10161 	  (match_operand:VNx8HI_ONLY 1 "register_operand" "0, w")
   10162 	  (vec_duplicate:VNx8HI
   10163 	    (match_operator:HI 5 "subreg_lowpart_operator"
   10164 	      [(unspec:SI [(match_operand:DI 2 "const_int_operand")
   10165 			   (match_operand:DI 3 "const_int_operand")
   10166 			   (match_operand:DI 4 "const_int_operand")]
   10167 			  UNSPEC_SVE_CNT_PAT)]))))]
   10168   "TARGET_SVE"
   10169   {
   10170     if (which_alternative == 1)
   10171       output_asm_insn ("movprfx\t%0, %1", operands);
   10172     return aarch64_output_sve_cnt_pat_immediate ("<inc_dec>", "%0.<Vetype>",
   10173 						 operands + 2);
   10174   }
   10175   [(set_attr "movprfx" "*,yes")]
   10176 )
   10177 
   10178 ;; -------------------------------------------------------------------------
   10179 ;; ---- [INT] Count elements in a predicate (scalar)
   10180 ;; -------------------------------------------------------------------------
   10181 ;; Includes:
   10182 ;; - CNTP
   10183 ;; -------------------------------------------------------------------------
   10184 
   10185 ;; Count the number of set bits in a predicate.  Operand 3 is true if
   10186 ;; operand 1 is known to be all-true.
   10187 (define_insn "@aarch64_pred_cntp<mode>"
   10188   [(set (match_operand:DI 0 "register_operand" "=r")
   10189 	(zero_extend:DI
   10190 	  (unspec:SI [(match_operand:PRED_ALL 1 "register_operand" "Upl")
   10191 		      (match_operand:SI 2 "aarch64_sve_ptrue_flag")
   10192 		      (match_operand:PRED_ALL 3 "register_operand" "Upa")]
   10193 		     UNSPEC_CNTP)))]
   10194   "TARGET_SVE"
   10195   "cntp\t%x0, %1, %3.<Vetype>")
   10196 
   10197 ;; -------------------------------------------------------------------------
   10198 ;; ---- [INT] Increment by the number of elements in a predicate (scalar)
   10199 ;; -------------------------------------------------------------------------
   10200 ;; Includes:
   10201 ;; - INCP
   10202 ;; - SQINCP
   10203 ;; - UQINCP
   10204 ;; -------------------------------------------------------------------------
   10205 
   10206 ;; Increment a DImode register by the number of set bits in a predicate.
   10207 ;; See aarch64_sve_cntp for a description of the operands.
   10208 (define_expand "@aarch64_sve_<inc_dec><DI_ONLY:mode><PRED_ALL:mode>_cntp"
   10209   [(set (match_operand:DI 0 "register_operand")
   10210 	(ANY_PLUS:DI
   10211 	  (zero_extend:DI
   10212 	    (unspec:SI [(match_dup 3)
   10213 			(const_int SVE_KNOWN_PTRUE)
   10214 			(match_operand:PRED_ALL 2 "register_operand")]
   10215 		       UNSPEC_CNTP))
   10216 	  (match_operand:DI_ONLY 1 "register_operand")))]
   10217   "TARGET_SVE"
   10218   {
   10219     operands[3] = CONSTM1_RTX (<PRED_ALL:MODE>mode);
   10220   }
   10221 )
   10222 
   10223 (define_insn_and_rewrite "*aarch64_sve_<inc_dec><DI_ONLY:mode><PRED_ALL:mode>_cntp"
   10224   [(set (match_operand:DI 0 "register_operand" "=r")
   10225 	(ANY_PLUS:DI
   10226 	  (zero_extend:DI
   10227 	    (unspec:SI [(match_operand 3)
   10228 			(const_int SVE_KNOWN_PTRUE)
   10229 			(match_operand:PRED_ALL 2 "register_operand" "Upa")]
   10230 		       UNSPEC_CNTP))
   10231 	  (match_operand:DI_ONLY 1 "register_operand" "0")))]
   10232   "TARGET_SVE"
   10233   "<inc_dec>p\t%x0, %2.<PRED_ALL:Vetype>"
   10234   "&& !CONSTANT_P (operands[3])"
   10235   {
   10236     operands[3] = CONSTM1_RTX (<PRED_ALL:MODE>mode);
   10237   }
   10238 )
   10239 
   10240 ;; Increment an SImode register by the number of set bits in a predicate
   10241 ;; using modular arithmetic.  See aarch64_sve_cntp for a description of
   10242 ;; the operands.
   10243 (define_insn_and_rewrite "*aarch64_incsi<mode>_cntp"
   10244   [(set (match_operand:SI 0 "register_operand" "=r")
   10245 	(plus:SI
   10246 	  (unspec:SI [(match_operand 3)
   10247 		      (const_int SVE_KNOWN_PTRUE)
   10248 		      (match_operand:PRED_ALL 2 "register_operand" "Upa")]
   10249 		     UNSPEC_CNTP)
   10250 	  (match_operand:SI 1 "register_operand" "0")))]
   10251   "TARGET_SVE"
   10252   "incp\t%x0, %2.<Vetype>"
   10253   "&& !CONSTANT_P (operands[3])"
   10254   {
   10255     operands[3] = CONSTM1_RTX (<MODE>mode);
   10256   }
   10257 )
   10258 
   10259 ;; Increment an SImode register by the number of set bits in a predicate
   10260 ;; using saturating arithmetic, extending the result to 64 bits.
   10261 ;;
   10262 ;; See aarch64_sve_cntp for a description of the operands.
   10263 (define_expand "@aarch64_sve_<inc_dec><SI_ONLY:mode><PRED_ALL:mode>_cntp"
   10264   [(set (match_operand:DI 0 "register_operand")
   10265 	(<paired_extend>:DI
   10266 	  (SAT_PLUS:SI
   10267 	    (unspec:SI [(match_dup 3)
   10268 			(const_int SVE_KNOWN_PTRUE)
   10269 			(match_operand:PRED_ALL 2 "register_operand")]
   10270 		       UNSPEC_CNTP)
   10271 	    (match_operand:SI_ONLY 1 "register_operand"))))]
   10272   "TARGET_SVE"
   10273   {
   10274     operands[3] = CONSTM1_RTX (<PRED_ALL:MODE>mode);
   10275   }
   10276 )
   10277 
   10278 (define_insn_and_rewrite "*aarch64_sve_<inc_dec><SI_ONLY:mode><PRED_ALL:mode>_cntp"
   10279   [(set (match_operand:DI 0 "register_operand" "=r")
   10280 	(<paired_extend>:DI
   10281 	  (SAT_PLUS:SI
   10282 	    (unspec:SI [(match_operand 3)
   10283 			(const_int SVE_KNOWN_PTRUE)
   10284 			(match_operand:PRED_ALL 2 "register_operand" "Upa")]
   10285 		       UNSPEC_CNTP)
   10286 	    (match_operand:SI_ONLY 1 "register_operand" "0"))))]
   10287   "TARGET_SVE"
   10288   {
   10289     if (<CODE> == SS_PLUS)
   10290       return "<inc_dec>p\t%x0, %2.<PRED_ALL:Vetype>, %w0";
   10291     else
   10292       return "<inc_dec>p\t%w0, %2.<PRED_ALL:Vetype>";
   10293   }
   10294   "&& !CONSTANT_P (operands[3])"
   10295   {
   10296     operands[3] = CONSTM1_RTX (<PRED_ALL:MODE>mode);
   10297   }
   10298 )
   10299 
   10300 ;; -------------------------------------------------------------------------
   10301 ;; ---- [INT] Increment by the number of elements in a predicate (vector)
   10302 ;; -------------------------------------------------------------------------
   10303 ;; Includes:
   10304 ;; - INCP
   10305 ;; - SQINCP
   10306 ;; - UQINCP
   10307 ;; -------------------------------------------------------------------------
   10308 
   10309 ;; Increment a vector of DIs by the number of set bits in a predicate.
   10310 ;; See aarch64_sve_cntp for a description of the operands.
   10311 (define_expand "@aarch64_sve_<inc_dec><mode>_cntp"
   10312   [(set (match_operand:VNx2DI 0 "register_operand")
   10313 	(ANY_PLUS:VNx2DI
   10314 	  (vec_duplicate:VNx2DI
   10315 	    (zero_extend:DI
   10316 	      (unspec:SI
   10317 		[(match_dup 3)
   10318 		 (const_int SVE_KNOWN_PTRUE)
   10319 		 (match_operand:<VPRED> 2 "register_operand")]
   10320 		UNSPEC_CNTP)))
   10321 	  (match_operand:VNx2DI_ONLY 1 "register_operand")))]
   10322   "TARGET_SVE"
   10323   {
   10324     operands[3] = CONSTM1_RTX (<VPRED>mode);
   10325   }
   10326 )
   10327 
   10328 (define_insn_and_rewrite "*aarch64_sve_<inc_dec><mode>_cntp"
   10329   [(set (match_operand:VNx2DI 0 "register_operand" "=w, ?&w")
   10330 	(ANY_PLUS:VNx2DI
   10331 	  (vec_duplicate:VNx2DI
   10332 	    (zero_extend:DI
   10333 	      (unspec:SI
   10334 		[(match_operand 3)
   10335 		 (const_int SVE_KNOWN_PTRUE)
   10336 		 (match_operand:<VPRED> 2 "register_operand" "Upa, Upa")]
   10337 		UNSPEC_CNTP)))
   10338 	  (match_operand:VNx2DI_ONLY 1 "register_operand" "0, w")))]
   10339   "TARGET_SVE"
   10340   "@
   10341    <inc_dec>p\t%0.d, %2
   10342    movprfx\t%0, %1\;<inc_dec>p\t%0.d, %2"
   10343   "&& !CONSTANT_P (operands[3])"
   10344   {
   10345     operands[3] = CONSTM1_RTX (<VPRED>mode);
   10346   }
   10347   [(set_attr "movprfx" "*,yes")]
   10348 )
   10349 
   10350 ;; Increment a vector of SIs by the number of set bits in a predicate.
   10351 ;; See aarch64_sve_cntp for a description of the operands.
   10352 (define_expand "@aarch64_sve_<inc_dec><mode>_cntp"
   10353   [(set (match_operand:VNx4SI 0 "register_operand")
   10354 	(ANY_PLUS:VNx4SI
   10355 	  (vec_duplicate:VNx4SI
   10356 	    (unspec:SI
   10357 	      [(match_dup 3)
   10358 	       (const_int SVE_KNOWN_PTRUE)
   10359 	       (match_operand:<VPRED> 2 "register_operand")]
   10360 	      UNSPEC_CNTP))
   10361 	  (match_operand:VNx4SI_ONLY 1 "register_operand")))]
   10362   "TARGET_SVE"
   10363   {
   10364     operands[3] = CONSTM1_RTX (<VPRED>mode);
   10365   }
   10366 )
   10367 
   10368 (define_insn_and_rewrite "*aarch64_sve_<inc_dec><mode>_cntp"
   10369   [(set (match_operand:VNx4SI 0 "register_operand" "=w, ?&w")
   10370 	(ANY_PLUS:VNx4SI
   10371 	  (vec_duplicate:VNx4SI
   10372 	    (unspec:SI
   10373 	      [(match_operand 3)
   10374 	       (const_int SVE_KNOWN_PTRUE)
   10375 	       (match_operand:<VPRED> 2 "register_operand" "Upa, Upa")]
   10376 	      UNSPEC_CNTP))
   10377 	  (match_operand:VNx4SI_ONLY 1 "register_operand" "0, w")))]
   10378   "TARGET_SVE"
   10379   "@
   10380    <inc_dec>p\t%0.s, %2
   10381    movprfx\t%0, %1\;<inc_dec>p\t%0.s, %2"
   10382   "&& !CONSTANT_P (operands[3])"
   10383   {
   10384     operands[3] = CONSTM1_RTX (<VPRED>mode);
   10385   }
   10386   [(set_attr "movprfx" "*,yes")]
   10387 )
   10388 
   10389 ;; Increment a vector of HIs by the number of set bits in a predicate.
   10390 ;; See aarch64_sve_cntp for a description of the operands.
   10391 (define_expand "@aarch64_sve_<inc_dec><mode>_cntp"
   10392   [(set (match_operand:VNx8HI 0 "register_operand")
   10393 	(ANY_PLUS:VNx8HI
   10394 	  (vec_duplicate:VNx8HI
   10395 	    (truncate:HI
   10396 	      (unspec:SI
   10397 		[(match_dup 3)
   10398 		 (const_int SVE_KNOWN_PTRUE)
   10399 		 (match_operand:<VPRED> 2 "register_operand")]
   10400 		UNSPEC_CNTP)))
   10401 	  (match_operand:VNx8HI_ONLY 1 "register_operand")))]
   10402   "TARGET_SVE"
   10403   {
   10404     operands[3] = CONSTM1_RTX (<VPRED>mode);
   10405   }
   10406 )
   10407 
   10408 (define_insn_and_rewrite "*aarch64_sve_<inc_dec><mode>_cntp"
   10409   [(set (match_operand:VNx8HI 0 "register_operand" "=w, ?&w")
   10410 	(ANY_PLUS:VNx8HI
   10411 	  (vec_duplicate:VNx8HI
   10412 	    (match_operator:HI 3 "subreg_lowpart_operator"
   10413 	      [(unspec:SI
   10414 		 [(match_operand 4)
   10415 		  (const_int SVE_KNOWN_PTRUE)
   10416 		  (match_operand:<VPRED> 2 "register_operand" "Upa, Upa")]
   10417 		 UNSPEC_CNTP)]))
   10418 	  (match_operand:VNx8HI_ONLY 1 "register_operand" "0, w")))]
   10419   "TARGET_SVE"
   10420   "@
   10421    <inc_dec>p\t%0.h, %2
   10422    movprfx\t%0, %1\;<inc_dec>p\t%0.h, %2"
   10423   "&& !CONSTANT_P (operands[4])"
   10424   {
   10425     operands[4] = CONSTM1_RTX (<VPRED>mode);
   10426   }
   10427   [(set_attr "movprfx" "*,yes")]
   10428 )
   10429 
   10430 ;; -------------------------------------------------------------------------
   10431 ;; ---- [INT] Decrement by the number of elements in a predicate (scalar)
   10432 ;; -------------------------------------------------------------------------
   10433 ;; Includes:
   10434 ;; - DECP
   10435 ;; - SQDECP
   10436 ;; - UQDECP
   10437 ;; -------------------------------------------------------------------------
   10438 
   10439 ;; Decrement a DImode register by the number of set bits in a predicate.
   10440 ;; See aarch64_sve_cntp for a description of the operands.
   10441 (define_expand "@aarch64_sve_<inc_dec><DI_ONLY:mode><PRED_ALL:mode>_cntp"
   10442   [(set (match_operand:DI 0 "register_operand")
   10443 	(ANY_MINUS:DI
   10444 	  (match_operand:DI_ONLY 1 "register_operand")
   10445 	  (zero_extend:DI
   10446 	    (unspec:SI [(match_dup 3)
   10447 			(const_int SVE_KNOWN_PTRUE)
   10448 			(match_operand:PRED_ALL 2 "register_operand")]
   10449 		       UNSPEC_CNTP))))]
   10450   "TARGET_SVE"
   10451   {
   10452     operands[3] = CONSTM1_RTX (<PRED_ALL:MODE>mode);
   10453   }
   10454 )
   10455 
   10456 (define_insn_and_rewrite "*aarch64_sve_<inc_dec><DI_ONLY:mode><PRED_ALL:mode>_cntp"
   10457   [(set (match_operand:DI 0 "register_operand" "=r")
   10458 	(ANY_MINUS:DI
   10459 	  (match_operand:DI_ONLY 1 "register_operand" "0")
   10460 	  (zero_extend:DI
   10461 	    (unspec:SI [(match_operand 3)
   10462 			(const_int SVE_KNOWN_PTRUE)
   10463 			(match_operand:PRED_ALL 2 "register_operand" "Upa")]
   10464 		       UNSPEC_CNTP))))]
   10465   "TARGET_SVE"
   10466   "<inc_dec>p\t%x0, %2.<PRED_ALL:Vetype>"
   10467   "&& !CONSTANT_P (operands[3])"
   10468   {
   10469     operands[3] = CONSTM1_RTX (<PRED_ALL:MODE>mode);
   10470   }
   10471 )
   10472 
   10473 ;; Decrement an SImode register by the number of set bits in a predicate
   10474 ;; using modular arithmetic.  See aarch64_sve_cntp for a description of the
   10475 ;; operands.
   10476 (define_insn_and_rewrite "*aarch64_decsi<mode>_cntp"
   10477   [(set (match_operand:SI 0 "register_operand" "=r")
   10478 	(minus:SI
   10479 	  (match_operand:SI 1 "register_operand" "0")
   10480 	  (unspec:SI [(match_operand 3)
   10481 		      (const_int SVE_KNOWN_PTRUE)
   10482 		      (match_operand:PRED_ALL 2 "register_operand" "Upa")]
   10483 		     UNSPEC_CNTP)))]
   10484   "TARGET_SVE"
   10485   "decp\t%x0, %2.<Vetype>"
   10486   "&& !CONSTANT_P (operands[3])"
   10487   {
   10488     operands[3] = CONSTM1_RTX (<MODE>mode);
   10489   }
   10490 )
   10491 
   10492 ;; Decrement an SImode register by the number of set bits in a predicate
   10493 ;; using saturating arithmetic, extending the result to 64 bits.
   10494 ;;
   10495 ;; See aarch64_sve_cntp for a description of the operands.
   10496 (define_expand "@aarch64_sve_<inc_dec><SI_ONLY:mode><PRED_ALL:mode>_cntp"
   10497   [(set (match_operand:DI 0 "register_operand")
   10498 	(<paired_extend>:DI
   10499 	  (SAT_MINUS:SI
   10500 	    (match_operand:SI_ONLY 1 "register_operand")
   10501 	    (unspec:SI [(match_dup 3)
   10502 			(const_int SVE_KNOWN_PTRUE)
   10503 			(match_operand:PRED_ALL 2 "register_operand")]
   10504 		       UNSPEC_CNTP))))]
   10505   "TARGET_SVE"
   10506   {
   10507     operands[3] = CONSTM1_RTX (<PRED_ALL:MODE>mode);
   10508   }
   10509 )
   10510 
   10511 (define_insn_and_rewrite "*aarch64_sve_<inc_dec><SI_ONLY:mode><PRED_ALL:mode>_cntp"
   10512   [(set (match_operand:DI 0 "register_operand" "=r")
   10513 	(<paired_extend>:DI
   10514 	  (SAT_MINUS:SI
   10515 	    (match_operand:SI_ONLY 1 "register_operand" "0")
   10516 	    (unspec:SI [(match_operand 3)
   10517 			(const_int SVE_KNOWN_PTRUE)
   10518 			(match_operand:PRED_ALL 2 "register_operand" "Upa")]
   10519 		       UNSPEC_CNTP))))]
   10520   "TARGET_SVE"
   10521   {
   10522     if (<CODE> == SS_MINUS)
   10523       return "<inc_dec>p\t%x0, %2.<PRED_ALL:Vetype>, %w0";
   10524     else
   10525       return "<inc_dec>p\t%w0, %2.<PRED_ALL:Vetype>";
   10526   }
   10527   "&& !CONSTANT_P (operands[3])"
   10528   {
   10529     operands[3] = CONSTM1_RTX (<PRED_ALL:MODE>mode);
   10530   }
   10531 )
   10532 
   10533 ;; -------------------------------------------------------------------------
   10534 ;; ---- [INT] Decrement by the number of elements in a predicate (vector)
   10535 ;; -------------------------------------------------------------------------
   10536 ;; Includes:
   10537 ;; - DECP
   10538 ;; - SQDECP
   10539 ;; - UQDECP
   10540 ;; -------------------------------------------------------------------------
   10541 
   10542 ;; Decrement a vector of DIs by the number of set bits in a predicate.
   10543 ;; See aarch64_sve_cntp for a description of the operands.
   10544 (define_expand "@aarch64_sve_<inc_dec><mode>_cntp"
   10545   [(set (match_operand:VNx2DI 0 "register_operand")
   10546 	(ANY_MINUS:VNx2DI
   10547 	  (match_operand:VNx2DI_ONLY 1 "register_operand")
   10548 	  (vec_duplicate:VNx2DI
   10549 	    (zero_extend:DI
   10550 	      (unspec:SI
   10551 		[(match_dup 3)
   10552 		 (const_int SVE_KNOWN_PTRUE)
   10553 		 (match_operand:<VPRED> 2 "register_operand")]
   10554 		UNSPEC_CNTP)))))]
   10555   "TARGET_SVE"
   10556   {
   10557     operands[3] = CONSTM1_RTX (<VPRED>mode);
   10558   }
   10559 )
   10560 
   10561 (define_insn_and_rewrite "*aarch64_sve_<inc_dec><mode>_cntp"
   10562   [(set (match_operand:VNx2DI 0 "register_operand" "=w, ?&w")
   10563 	(ANY_MINUS:VNx2DI
   10564 	  (match_operand:VNx2DI_ONLY 1 "register_operand" "0, w")
   10565 	  (vec_duplicate:VNx2DI
   10566 	    (zero_extend:DI
   10567 	      (unspec:SI
   10568 		[(match_operand 3)
   10569 		 (const_int SVE_KNOWN_PTRUE)
   10570 		 (match_operand:<VPRED> 2 "register_operand" "Upa, Upa")]
   10571 		UNSPEC_CNTP)))))]
   10572   "TARGET_SVE"
   10573   "@
   10574    <inc_dec>p\t%0.d, %2
   10575    movprfx\t%0, %1\;<inc_dec>p\t%0.d, %2"
   10576   "&& !CONSTANT_P (operands[3])"
   10577   {
   10578     operands[3] = CONSTM1_RTX (<VPRED>mode);
   10579   }
   10580   [(set_attr "movprfx" "*,yes")]
   10581 )
   10582 
   10583 ;; Decrement a vector of SIs by the number of set bits in a predicate.
   10584 ;; See aarch64_sve_cntp for a description of the operands.
   10585 (define_expand "@aarch64_sve_<inc_dec><mode>_cntp"
   10586   [(set (match_operand:VNx4SI 0 "register_operand")
   10587 	(ANY_MINUS:VNx4SI
   10588 	  (match_operand:VNx4SI_ONLY 1 "register_operand")
   10589 	  (vec_duplicate:VNx4SI
   10590 	    (unspec:SI
   10591 	      [(match_dup 3)
   10592 	       (const_int SVE_KNOWN_PTRUE)
   10593 	       (match_operand:<VPRED> 2 "register_operand")]
   10594 	      UNSPEC_CNTP))))]
   10595   "TARGET_SVE"
   10596   {
   10597     operands[3] = CONSTM1_RTX (<VPRED>mode);
   10598   }
   10599 )
   10600 
   10601 (define_insn_and_rewrite "*aarch64_sve_<inc_dec><mode>_cntp"
   10602   [(set (match_operand:VNx4SI 0 "register_operand" "=w, ?&w")
   10603 	(ANY_MINUS:VNx4SI
   10604 	  (match_operand:VNx4SI_ONLY 1 "register_operand" "0, w")
   10605 	  (vec_duplicate:VNx4SI
   10606 	    (unspec:SI
   10607 	      [(match_operand 3)
   10608 	       (const_int SVE_KNOWN_PTRUE)
   10609 	       (match_operand:<VPRED> 2 "register_operand" "Upa, Upa")]
   10610 	      UNSPEC_CNTP))))]
   10611   "TARGET_SVE"
   10612   "@
   10613    <inc_dec>p\t%0.s, %2
   10614    movprfx\t%0, %1\;<inc_dec>p\t%0.s, %2"
   10615   "&& !CONSTANT_P (operands[3])"
   10616   {
   10617     operands[3] = CONSTM1_RTX (<VPRED>mode);
   10618   }
   10619   [(set_attr "movprfx" "*,yes")]
   10620 )
   10621 
   10622 ;; Decrement a vector of HIs by the number of set bits in a predicate.
   10623 ;; See aarch64_sve_cntp for a description of the operands.
   10624 (define_expand "@aarch64_sve_<inc_dec><mode>_cntp"
   10625   [(set (match_operand:VNx8HI 0 "register_operand")
   10626 	(ANY_MINUS:VNx8HI
   10627 	  (match_operand:VNx8HI_ONLY 1 "register_operand")
   10628 	  (vec_duplicate:VNx8HI
   10629 	    (truncate:HI
   10630 	      (unspec:SI
   10631 		[(match_dup 3)
   10632 		 (const_int SVE_KNOWN_PTRUE)
   10633 		 (match_operand:<VPRED> 2 "register_operand")]
   10634 		UNSPEC_CNTP)))))]
   10635   "TARGET_SVE"
   10636   {
   10637     operands[3] = CONSTM1_RTX (<VPRED>mode);
   10638   }
   10639 )
   10640 
   10641 (define_insn_and_rewrite "*aarch64_sve_<inc_dec><mode>_cntp"
   10642   [(set (match_operand:VNx8HI 0 "register_operand" "=w, ?&w")
   10643 	(ANY_MINUS:VNx8HI
   10644 	  (match_operand:VNx8HI_ONLY 1 "register_operand" "0, w")
   10645 	  (vec_duplicate:VNx8HI
   10646 	    (match_operator:HI 3 "subreg_lowpart_operator"
   10647 	      [(unspec:SI
   10648 		 [(match_operand 4)
   10649 		  (const_int SVE_KNOWN_PTRUE)
   10650 		  (match_operand:<VPRED> 2 "register_operand" "Upa, Upa")]
   10651 		 UNSPEC_CNTP)]))))]
   10652   "TARGET_SVE"
   10653   "@
   10654    <inc_dec>p\t%0.h, %2
   10655    movprfx\t%0, %1\;<inc_dec>p\t%0.h, %2"
   10656   "&& !CONSTANT_P (operands[4])"
   10657   {
   10658     operands[4] = CONSTM1_RTX (<VPRED>mode);
   10659   }
   10660   [(set_attr "movprfx" "*,yes")]
   10661 )
   10662