vector.md revision 1.1.1.3
1 ;;- Instruction patterns for the System z vector facility 2 ;; Copyright (C) 2015-2017 Free Software Foundation, Inc. 3 ;; Contributed by Andreas Krebbel (Andreas.Krebbel (a] de.ibm.com) 4 5 ;; This file is part of GCC. 6 7 ;; GCC is free software; you can redistribute it and/or modify it under 8 ;; the terms of the GNU General Public License as published by the Free 9 ;; Software Foundation; either version 3, or (at your option) any later 10 ;; version. 11 12 ;; GCC is distributed in the hope that it will be useful, but WITHOUT ANY 13 ;; WARRANTY; without even the implied warranty of MERCHANTABILITY or 14 ;; FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 15 ;; 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 ; All vector modes supported in a vector register 22 (define_mode_iterator V 23 [V1QI V2QI V4QI V8QI V16QI V1HI V2HI V4HI V8HI V1SI V2SI V4SI V1DI V2DI V1SF 24 V2SF V4SF V1DF V2DF]) 25 (define_mode_iterator VT 26 [V1QI V2QI V4QI V8QI V16QI V1HI V2HI V4HI V8HI V1SI V2SI V4SI V1DI V2DI V1SF 27 V2SF V4SF V1DF V2DF V1TF V1TI TI]) 28 29 ; All modes directly supported by the hardware having full vector reg size 30 ; V_HW2 is duplicate of V_HW for having two iterators expanding 31 ; independently e.g. vcond 32 (define_mode_iterator V_HW [V16QI V8HI V4SI V2DI V2DF (V4SF "TARGET_VXE") (V1TF "TARGET_VXE")]) 33 (define_mode_iterator V_HW2 [V16QI V8HI V4SI V2DI V2DF (V4SF "TARGET_VXE") (V1TF "TARGET_VXE")]) 34 35 (define_mode_iterator V_HW_64 [V2DI V2DF]) 36 37 ; Including TI for instructions that support it (va, vn, ...) 38 (define_mode_iterator VT_HW [V16QI V8HI V4SI V2DI V2DF V1TI TI (V4SF "TARGET_VXE") (V1TF "TARGET_VXE")]) 39 40 ; All full size integer vector modes supported in a vector register + TImode 41 (define_mode_iterator VIT_HW [V16QI V8HI V4SI V2DI V1TI TI]) 42 (define_mode_iterator VI_HW [V16QI V8HI V4SI V2DI]) 43 (define_mode_iterator VI_HW_QHS [V16QI V8HI V4SI]) 44 (define_mode_iterator VI_HW_HSD [V8HI V4SI V2DI]) 45 (define_mode_iterator VI_HW_HS [V8HI V4SI]) 46 (define_mode_iterator VI_HW_QH [V16QI V8HI]) 47 (define_mode_iterator VI_HW_4 [V4SI V4SF]) 48 49 ; All integer vector modes supported in a vector register + TImode 50 (define_mode_iterator VIT [V1QI V2QI V4QI V8QI V16QI V1HI V2HI V4HI V8HI V1SI V2SI V4SI V1DI V2DI V1TI TI]) 51 (define_mode_iterator VI [V1QI V2QI V4QI V8QI V16QI V1HI V2HI V4HI V8HI V1SI V2SI V4SI V1DI V2DI]) 52 (define_mode_iterator VI_QHS [V1QI V2QI V4QI V8QI V16QI V1HI V2HI V4HI V8HI V1SI V2SI V4SI]) 53 54 (define_mode_iterator VFT [(V1SF "TARGET_VXE") (V2SF "TARGET_VXE") (V4SF "TARGET_VXE") 55 V1DF V2DF 56 (V1TF "TARGET_VXE")]) 57 58 ; FP vector modes directly supported by the HW. This does not include 59 ; vector modes using only part of a vector register and should be used 60 ; for instructions which might trigger IEEE exceptions. 61 (define_mode_iterator VF_HW [(V4SF "TARGET_VXE") V2DF (V1TF "TARGET_VXE")]) 62 63 (define_mode_iterator V_8 [V1QI]) 64 (define_mode_iterator V_16 [V2QI V1HI]) 65 (define_mode_iterator V_32 [V4QI V2HI V1SI V1SF]) 66 (define_mode_iterator V_64 [V8QI V4HI V2SI V2SF V1DI V1DF]) 67 (define_mode_iterator V_128 [V16QI V8HI V4SI V4SF V2DI V2DF V1TI V1TF]) 68 69 (define_mode_iterator V_128_NOSINGLE [V16QI V8HI V4SI V4SF V2DI V2DF]) 70 71 ; Empty string for all but TImode. This is used to hide the TImode 72 ; expander name in case it is defined already. See addti3 for an 73 ; example. 74 (define_mode_attr ti* [(V1QI "") (V2QI "") (V4QI "") (V8QI "") (V16QI "") 75 (V1HI "") (V2HI "") (V4HI "") (V8HI "") 76 (V1SI "") (V2SI "") (V4SI "") 77 (V1DI "") (V2DI "") 78 (V1TI "") (TI "*") 79 (V1SF "") (V2SF "") (V4SF "") 80 (V1DF "") (V2DF "") 81 (V1TF "") (TF "")]) 82 83 ; The element type of the vector. 84 (define_mode_attr non_vec[(V1QI "QI") (V2QI "QI") (V4QI "QI") (V8QI "QI") (V16QI "QI") 85 (V1HI "HI") (V2HI "HI") (V4HI "HI") (V8HI "HI") 86 (V1SI "SI") (V2SI "SI") (V4SI "SI") 87 (V1DI "DI") (V2DI "DI") 88 (V1TI "TI") (TI "TI") 89 (V1SF "SF") (V2SF "SF") (V4SF "SF") 90 (V1DF "DF") (V2DF "DF") 91 (V1TF "TF") (TF "TF")]) 92 93 ; The instruction suffix for integer instructions and instructions 94 ; which do not care about whether it is floating point or integer. 95 (define_mode_attr bhfgq[(V1QI "b") (V2QI "b") (V4QI "b") (V8QI "b") (V16QI "b") 96 (V1HI "h") (V2HI "h") (V4HI "h") (V8HI "h") 97 (V1SI "f") (V2SI "f") (V4SI "f") 98 (V1DI "g") (V2DI "g") 99 (V1TI "q") (TI "q") 100 (V1SF "f") (V2SF "f") (V4SF "f") 101 (V1DF "g") (V2DF "g") 102 (V1TF "q")]) 103 104 ; This is for vmalhw. It gets an 'w' attached to avoid confusion with 105 ; multiply and add logical high vmalh. 106 (define_mode_attr w [(V1QI "") (V2QI "") (V4QI "") (V8QI "") (V16QI "") 107 (V1HI "w") (V2HI "w") (V4HI "w") (V8HI "w") 108 (V1SI "") (V2SI "") (V4SI "") 109 (V1DI "") (V2DI "")]) 110 111 ; Resulting mode of a vector comparison. For floating point modes an 112 ; integer vector mode with the same element size is picked. 113 (define_mode_attr tointvec [(V1QI "V1QI") (V2QI "V2QI") (V4QI "V4QI") (V8QI "V8QI") (V16QI "V16QI") 114 (V1HI "V1HI") (V2HI "V2HI") (V4HI "V4HI") (V8HI "V8HI") 115 (V1SI "V1SI") (V2SI "V2SI") (V4SI "V4SI") 116 (V1DI "V1DI") (V2DI "V2DI") 117 (V1TI "V1TI") 118 (V1SF "V1SI") (V2SF "V2SI") (V4SF "V4SI") 119 (V1DF "V1DI") (V2DF "V2DI") 120 (V1TF "V1TI")]) 121 (define_mode_attr vw [(SF "w") (V1SF "w") (V2SF "v") (V4SF "v") 122 (DF "w") (V1DF "w") (V2DF "v") 123 (TF "w") (V1TF "w")]) 124 125 (define_mode_attr sdx [(SF "s") (V1SF "s") (V2SF "s") (V4SF "s") 126 (DF "d") (V1DF "d") (V2DF "d") 127 (TF "x") (V1TF "x")]) 128 129 ; Vector with doubled element size. 130 (define_mode_attr vec_double [(V1QI "V1HI") (V2QI "V1HI") (V4QI "V2HI") (V8QI "V4HI") (V16QI "V8HI") 131 (V1HI "V1SI") (V2HI "V1SI") (V4HI "V2SI") (V8HI "V4SI") 132 (V1SI "V1DI") (V2SI "V1DI") (V4SI "V2DI") 133 (V1DI "V1TI") (V2DI "V1TI") 134 (V1SF "V1DF") (V2SF "V1DF") (V4SF "V2DF")]) 135 136 ; Vector with half the element size. 137 (define_mode_attr vec_half [(V1HI "V2QI") (V2HI "V4QI") (V4HI "V8QI") (V8HI "V16QI") 138 (V1SI "V2HI") (V2SI "V4HI") (V4SI "V8HI") 139 (V1DI "V2SI") (V2DI "V4SI") 140 (V1TI "V2DI") 141 (V1DF "V2SF") (V2DF "V4SF") 142 (V1TF "V1DF")]) 143 144 ; Vector with half the element size AND half the number of elements. 145 (define_mode_attr vec_halfhalf 146 [(V2HI "V2QI") (V4HI "V4QI") (V8HI "V8QI") 147 (V2SI "V2HI") (V4SI "V4HI") 148 (V2DI "V2SI") 149 (V2DF "V2SF")]) 150 151 (define_mode_attr vec_halfnumelts 152 [(V4SF "V2SF") (V4SI "V2SI")]) 153 154 ; The comparisons not setting CC iterate over the rtx code. 155 (define_code_iterator VFCMP_HW_OP [eq gt ge]) 156 (define_code_attr asm_fcmp_op [(eq "e") (gt "h") (ge "he")]) 157 158 159 160 ; Comparison operators on int and fp compares which are directly 161 ; supported by the HW. 162 (define_code_iterator VICMP_HW_OP [eq gt gtu]) 163 ; For int insn_cmp_op can be used in the insn name as well as in the asm output. 164 (define_code_attr insn_cmp_op [(eq "eq") (gt "h") (gtu "hl") (ge "he")]) 165 166 ; Flags for vector string instructions (vfae all 4, vfee only ZS and CS, vstrc all 4) 167 (define_constants 168 [(VSTRING_FLAG_IN 8) ; invert result 169 (VSTRING_FLAG_RT 4) ; result type 170 (VSTRING_FLAG_ZS 2) ; zero search 171 (VSTRING_FLAG_CS 1)]) ; condition code set 172 173 (include "vx-builtins.md") 174 175 ; Full HW vector size moves 176 177 ; We don't use lm/stm for 128 bit moves since these are slower than 178 ; splitting it into separate moves. 179 180 ; FIXME: More constants are possible by enabling jxx, jyy constraints 181 ; for TImode (use double-int for the calculations) 182 183 ; vgmb, vgmh, vgmf, vgmg, vrepib, vrepih, vrepif, vrepig 184 (define_insn "mov<mode>" 185 [(set (match_operand:V_128 0 "nonimmediate_operand" "=v,v,R, v, v, v, v, v,v,*d,*d,?o") 186 (match_operand:V_128 1 "general_operand" " v,R,v,j00,jm1,jyy,jxx,jKK,d, v,dT,*d"))] 187 "" 188 "@ 189 vlr\t%v0,%v1 190 vl\t%v0,%1 191 vst\t%v1,%0 192 vzero\t%v0 193 vone\t%v0 194 vgbm\t%v0,%t1 195 vgm<bhfgq>\t%v0,%s1,%e1 196 vrepi<bhfgq>\t%v0,%h1 197 vlvgp\t%v0,%1,%N1 198 # 199 # 200 #" 201 [(set_attr "cpu_facility" "vx,vx,vx,vx,vx,vx,vx,vx,vx,vx,*,*") 202 (set_attr "op_type" "VRR,VRX,VRX,VRI,VRI,VRI,VRI,VRI,VRR,*,*,*")]) 203 204 ; VR -> GPR, no instruction so split it into 64 element sets. 205 (define_split 206 [(set (match_operand:V_128 0 "register_operand" "") 207 (match_operand:V_128 1 "register_operand" ""))] 208 "TARGET_VX && GENERAL_REG_P (operands[0]) && VECTOR_REG_P (operands[1])" 209 [(set (match_dup 2) 210 (unspec:DI [(subreg:V2DI (match_dup 1) 0) 211 (const_int 0)] UNSPEC_VEC_EXTRACT)) 212 (set (match_dup 3) 213 (unspec:DI [(subreg:V2DI (match_dup 1) 0) 214 (const_int 1)] UNSPEC_VEC_EXTRACT))] 215 { 216 operands[2] = operand_subword (operands[0], 0, 0, <MODE>mode); 217 operands[3] = operand_subword (operands[0], 1, 0, <MODE>mode); 218 }) 219 220 ; Split the 128 bit GPR move into two word mode moves 221 ; s390_split_ok_p decides which part needs to be moved first. 222 223 (define_split 224 [(set (match_operand:V_128 0 "nonimmediate_operand" "") 225 (match_operand:V_128 1 "general_operand" ""))] 226 "reload_completed 227 && s390_split_ok_p (operands[0], operands[1], <MODE>mode, 0)" 228 [(set (match_dup 2) (match_dup 4)) 229 (set (match_dup 3) (match_dup 5))] 230 { 231 operands[2] = operand_subword (operands[0], 0, 0, <MODE>mode); 232 operands[3] = operand_subword (operands[0], 1, 0, <MODE>mode); 233 operands[4] = operand_subword (operands[1], 0, 0, <MODE>mode); 234 operands[5] = operand_subword (operands[1], 1, 0, <MODE>mode); 235 }) 236 237 (define_split 238 [(set (match_operand:V_128 0 "nonimmediate_operand" "") 239 (match_operand:V_128 1 "general_operand" ""))] 240 "reload_completed 241 && s390_split_ok_p (operands[0], operands[1], <MODE>mode, 1)" 242 [(set (match_dup 2) (match_dup 4)) 243 (set (match_dup 3) (match_dup 5))] 244 { 245 operands[2] = operand_subword (operands[0], 1, 0, <MODE>mode); 246 operands[3] = operand_subword (operands[0], 0, 0, <MODE>mode); 247 operands[4] = operand_subword (operands[1], 1, 0, <MODE>mode); 248 operands[5] = operand_subword (operands[1], 0, 0, <MODE>mode); 249 }) 250 251 ; This is the vector equivalent to the TImode splitter in s390.md. It 252 ; is required if both target GPRs occur in the source address operand. 253 254 ; For non-s_operands at least one of the target GPRs does not conflict 255 ; with the address operand and one of the splitters above will take 256 ; over. 257 (define_split 258 [(set (match_operand:V_128 0 "register_operand" "") 259 (match_operand:V_128 1 "memory_operand" ""))] 260 "TARGET_ZARCH && reload_completed 261 && !VECTOR_REG_P (operands[0]) 262 && !s_operand (operands[1], VOIDmode)" 263 [(set (match_dup 0) (match_dup 1))] 264 { 265 rtx addr = operand_subword (operands[0], 1, 0, <MODE>mode); 266 addr = gen_lowpart (Pmode, addr); 267 s390_load_address (addr, XEXP (operands[1], 0)); 268 operands[1] = replace_equiv_address (operands[1], addr); 269 }) 270 271 ; Moves for smaller vector modes. 272 273 ; In these patterns only the vlr, vone, and vzero instructions write 274 ; VR bytes outside the mode. This should be ok since we disallow 275 ; formerly bigger modes being accessed with smaller modes via 276 ; subreg. Note: The vone, vzero instructions could easily be replaced 277 ; with vlei which would only access the bytes belonging to the mode. 278 ; However, this would probably be slower. 279 280 (define_insn "mov<mode>" 281 [(set (match_operand:V_8 0 "nonimmediate_operand" "=v,v,d,v,R, v, v, v, v,d, Q, S, Q, S, d, d,d,d,d,R,T") 282 (match_operand:V_8 1 "general_operand" " v,d,v,R,v,j00,jm1,jyy,jxx,d,j00,j00,jm1,jm1,j00,jm1,R,T,b,d,d"))] 283 "" 284 "@ 285 vlr\t%v0,%v1 286 vlvgb\t%v0,%1,0 287 vlgvb\t%0,%v1,0 288 vleb\t%v0,%1,0 289 vsteb\t%v1,%0,0 290 vzero\t%v0 291 vone\t%v0 292 vgbm\t%v0,%t1 293 vgm\t%v0,%s1,%e1 294 lr\t%0,%1 295 mvi\t%0,0 296 mviy\t%0,0 297 mvi\t%0,-1 298 mviy\t%0,-1 299 lhi\t%0,0 300 lhi\t%0,-1 301 lh\t%0,%1 302 lhy\t%0,%1 303 lhrl\t%0,%1 304 stc\t%1,%0 305 stcy\t%1,%0" 306 [(set_attr "op_type" "VRR,VRS,VRS,VRX,VRX,VRI,VRI,VRI,VRI,RR,SI,SIY,SI,SIY,RI,RI,RX,RXY,RIL,RX,RXY")]) 307 308 (define_insn "mov<mode>" 309 [(set (match_operand:V_16 0 "nonimmediate_operand" "=v,v,d,v,R, v, v, v, v,d, Q, Q, d, d,d,d,d,R,T,b") 310 (match_operand:V_16 1 "general_operand" " v,d,v,R,v,j00,jm1,jyy,jxx,d,j00,jm1,j00,jm1,R,T,b,d,d,d"))] 311 "" 312 "@ 313 vlr\t%v0,%v1 314 vlvgh\t%v0,%1,0 315 vlgvh\t%0,%v1,0 316 vleh\t%v0,%1,0 317 vsteh\t%v1,%0,0 318 vzero\t%v0 319 vone\t%v0 320 vgbm\t%v0,%t1 321 vgm\t%v0,%s1,%e1 322 lr\t%0,%1 323 mvhhi\t%0,0 324 mvhhi\t%0,-1 325 lhi\t%0,0 326 lhi\t%0,-1 327 lh\t%0,%1 328 lhy\t%0,%1 329 lhrl\t%0,%1 330 sth\t%1,%0 331 sthy\t%1,%0 332 sthrl\t%1,%0" 333 [(set_attr "op_type" "VRR,VRS,VRS,VRX,VRX,VRI,VRI,VRI,VRI,RR,SIL,SIL,RI,RI,RX,RXY,RIL,RX,RXY,RIL")]) 334 335 (define_insn "mov<mode>" 336 [(set (match_operand:V_32 0 "nonimmediate_operand" "=f,f,f,R,T,v,v,d,v,R, f, v, v, v, v, Q, Q, d, d,d,d,d,d,R,T,b") 337 (match_operand:V_32 1 "general_operand" " f,R,T,f,f,v,d,v,R,v,j00,j00,jm1,jyy,jxx,j00,jm1,j00,jm1,b,d,R,T,d,d,d"))] 338 "TARGET_VX" 339 "@ 340 ldr\t%v0,%v1 341 lde\t%0,%1 342 ley\t%0,%1 343 ste\t%1,%0 344 stey\t%1,%0 345 vlr\t%v0,%v1 346 vlvgf\t%v0,%1,0 347 vlgvf\t%0,%v1,0 348 vlef\t%v0,%1,0 349 vstef\t%1,%0,0 350 lzer\t%v0 351 vzero\t%v0 352 vone\t%v0 353 vgbm\t%v0,%t1 354 vgm\t%v0,%s1,%e1 355 mvhi\t%0,0 356 mvhi\t%0,-1 357 lhi\t%0,0 358 lhi\t%0,-1 359 lrl\t%0,%1 360 lr\t%0,%1 361 l\t%0,%1 362 ly\t%0,%1 363 st\t%1,%0 364 sty\t%1,%0 365 strl\t%1,%0" 366 [(set_attr "op_type" "RR,RXE,RXY,RX,RXY,VRR,VRS,VRS,VRX,VRX,RRE,VRI,VRI,VRI,VRI,SIL,SIL,RI,RI, 367 RIL,RR,RX,RXY,RX,RXY,RIL")]) 368 369 (define_insn "mov<mode>" 370 [(set (match_operand:V_64 0 "nonimmediate_operand" 371 "=f,f,f,R,T,v,v,d,v,R, f, v, v, v, v, Q, Q, d, d,f,d,d,d,d,T,b") 372 (match_operand:V_64 1 "general_operand" 373 " f,R,T,f,f,v,d,v,R,v,j00,j00,jm1,jyy,jxx,j00,jm1,j00,jm1,d,f,b,d,T,d,d"))] 374 "TARGET_ZARCH" 375 "@ 376 ldr\t%0,%1 377 ld\t%0,%1 378 ldy\t%0,%1 379 std\t%1,%0 380 stdy\t%1,%0 381 vlr\t%v0,%v1 382 vlvgg\t%v0,%1,0 383 vlgvg\t%0,%v1,0 384 vleg\t%v0,%1,0 385 vsteg\t%v1,%0,0 386 lzdr\t%0 387 vzero\t%v0 388 vone\t%v0 389 vgbm\t%v0,%t1 390 vgm\t%v0,%s1,%e1 391 mvghi\t%0,0 392 mvghi\t%0,-1 393 lghi\t%0,0 394 lghi\t%0,-1 395 ldgr\t%0,%1 396 lgdr\t%0,%1 397 lgrl\t%0,%1 398 lgr\t%0,%1 399 lg\t%0,%1 400 stg\t%1,%0 401 stgrl\t%1,%0" 402 [(set_attr "op_type" "RRE,RX,RXY,RX,RXY,VRR,VRS,VRS,VRX,VRX,RRE,VRI,VRI,VRI,VRI, 403 SIL,SIL,RI,RI,RRE,RRE,RIL,RR,RXY,RXY,RIL")]) 404 405 406 ; vec_load_lanes? 407 408 ; vec_store_lanes? 409 410 ; vec_set is supposed to *modify* an existing vector so operand 0 is 411 ; duplicated as input operand. 412 (define_expand "vec_set<mode>" 413 [(set (match_operand:V 0 "register_operand" "") 414 (unspec:V [(match_operand:<non_vec> 1 "general_operand" "") 415 (match_operand:SI 2 "nonmemory_operand" "") 416 (match_dup 0)] 417 UNSPEC_VEC_SET))] 418 "TARGET_VX") 419 420 ; FIXME: Support also vector mode operands for 1 421 ; FIXME: A target memory operand seems to be useful otherwise we end 422 ; up with vl vlvgg vst. Shouldn't the middle-end be able to handle 423 ; that itself? 424 ; vlvgb, vlvgh, vlvgf, vlvgg, vleb, vleh, vlef, vleg, vleib, vleih, vleif, vleig 425 (define_insn "*vec_set<mode>" 426 [(set (match_operand:V 0 "register_operand" "=v,v,v") 427 (unspec:V [(match_operand:<non_vec> 1 "general_operand" "d,R,K") 428 (match_operand:SI 2 "nonmemory_operand" "an,I,I") 429 (match_operand:V 3 "register_operand" "0,0,0")] 430 UNSPEC_VEC_SET))] 431 "TARGET_VX 432 && (!CONST_INT_P (operands[2]) 433 || UINTVAL (operands[2]) < GET_MODE_NUNITS (<V:MODE>mode))" 434 "@ 435 vlvg<bhfgq>\t%v0,%1,%Y2 436 vle<bhfgq>\t%v0,%1,%2 437 vlei<bhfgq>\t%v0,%1,%2" 438 [(set_attr "op_type" "VRS,VRX,VRI")]) 439 440 ; vlvgb, vlvgh, vlvgf, vlvgg 441 (define_insn "*vec_set<mode>_plus" 442 [(set (match_operand:V 0 "register_operand" "=v") 443 (unspec:V [(match_operand:<non_vec> 1 "general_operand" "d") 444 (plus:SI (match_operand:SI 2 "register_operand" "a") 445 (match_operand:SI 4 "const_int_operand" "n")) 446 (match_operand:V 3 "register_operand" "0")] 447 UNSPEC_VEC_SET))] 448 "TARGET_VX" 449 "vlvg<bhfgq>\t%v0,%1,%Y4(%2)" 450 [(set_attr "op_type" "VRS")]) 451 452 453 ; FIXME: Support also vector mode operands for 0 454 ; FIXME: This should be (vec_select ..) or something but it does only allow constant selectors :( 455 ; This is used via RTL standard name as well as for expanding the builtin 456 (define_expand "vec_extract<mode>" 457 [(set (match_operand:<non_vec> 0 "nonimmediate_operand" "") 458 (unspec:<non_vec> [(match_operand:V 1 "register_operand" "") 459 (match_operand:SI 2 "nonmemory_operand" "")] 460 UNSPEC_VEC_EXTRACT))] 461 "TARGET_VX") 462 463 ; vlgvb, vlgvh, vlgvf, vlgvg, vsteb, vsteh, vstef, vsteg 464 (define_insn "*vec_extract<mode>" 465 [(set (match_operand:<non_vec> 0 "nonimmediate_operand" "=d,R") 466 (unspec:<non_vec> [(match_operand:V 1 "register_operand" "v,v") 467 (match_operand:SI 2 "nonmemory_operand" "an,I")] 468 UNSPEC_VEC_EXTRACT))] 469 "TARGET_VX 470 && (!CONST_INT_P (operands[2]) 471 || UINTVAL (operands[2]) < GET_MODE_NUNITS (<V:MODE>mode))" 472 "@ 473 vlgv<bhfgq>\t%0,%v1,%Y2 474 vste<bhfgq>\t%v1,%0,%2" 475 [(set_attr "op_type" "VRS,VRX")]) 476 477 ; vlgvb, vlgvh, vlgvf, vlgvg 478 (define_insn "*vec_extract<mode>_plus" 479 [(set (match_operand:<non_vec> 0 "nonimmediate_operand" "=d") 480 (unspec:<non_vec> [(match_operand:V 1 "register_operand" "v") 481 (plus:SI (match_operand:SI 2 "nonmemory_operand" "a") 482 (match_operand:SI 3 "const_int_operand" "n"))] 483 UNSPEC_VEC_EXTRACT))] 484 "TARGET_VX" 485 "vlgv<bhfgq>\t%0,%v1,%Y3(%2)" 486 [(set_attr "op_type" "VRS")]) 487 488 (define_expand "vec_init<mode>" 489 [(match_operand:V_128 0 "register_operand" "") 490 (match_operand:V_128 1 "nonmemory_operand" "")] 491 "TARGET_VX" 492 { 493 s390_expand_vec_init (operands[0], operands[1]); 494 DONE; 495 }) 496 497 (define_insn "*vec_vllezlf<mode>" 498 [(set (match_operand:VI_HW_4 0 "register_operand" "=v") 499 (vec_concat:VI_HW_4 500 (vec_concat:<vec_halfnumelts> 501 (match_operand:<non_vec> 1 "memory_operand" "R") 502 (const_int 0)) 503 (vec_concat:<vec_halfnumelts> 504 (const_int 0) 505 (const_int 0))))] 506 "TARGET_VXE" 507 "vllezlf\t%v0,%1" 508 [(set_attr "op_type" "VRX")]) 509 510 ; Replicate from vector element 511 ; vrepb, vreph, vrepf, vrepg 512 (define_insn "*vec_splat<mode>" 513 [(set (match_operand:V_128_NOSINGLE 0 "register_operand" "=v") 514 (vec_duplicate:V_128_NOSINGLE 515 (vec_select:<non_vec> 516 (match_operand:V_128_NOSINGLE 1 "register_operand" "v") 517 (parallel 518 [(match_operand:QI 2 "const_mask_operand" "C")]))))] 519 "TARGET_VX && UINTVAL (operands[2]) < GET_MODE_NUNITS (<MODE>mode)" 520 "vrep<bhfgq>\t%v0,%v1,%2" 521 [(set_attr "op_type" "VRI")]) 522 523 ; vlrepb, vlreph, vlrepf, vlrepg, vrepib, vrepih, vrepif, vrepig, vrepb, vreph, vrepf, vrepg 524 (define_insn "*vec_splats<mode>" 525 [(set (match_operand:V_128_NOSINGLE 0 "register_operand" "=v,v,v,v") 526 (vec_duplicate:V_128_NOSINGLE (match_operand:<non_vec> 1 "general_operand" " R,K,v,d")))] 527 "TARGET_VX" 528 "@ 529 vlrep<bhfgq>\t%v0,%1 530 vrepi<bhfgq>\t%v0,%h1 531 vrep<bhfgq>\t%v0,%v1,0 532 #" 533 [(set_attr "op_type" "VRX,VRI,VRI,*")]) 534 535 ; A TFmode operand resides in FPR register pairs while V1TF is in a 536 ; single vector register. 537 (define_insn "*vec_tf_to_v1tf" 538 [(set (match_operand:V1TF 0 "nonimmediate_operand" "=v,v,R,v,v") 539 (vec_duplicate:V1TF (match_operand:TF 1 "general_operand" "v,R,v,G,d")))] 540 "TARGET_VX" 541 "@ 542 vmrhg\t%v0,%1,%N1 543 vl\t%v0,%1 544 vst\t%v1,%0 545 vzero\t%v0 546 vlvgp\t%v0,%1,%N1" 547 [(set_attr "op_type" "VRR,VRX,VRX,VRI,VRR")]) 548 549 (define_insn "*vec_ti_to_v1ti" 550 [(set (match_operand:V1TI 0 "nonimmediate_operand" "=v,v,R, v, v,v") 551 (vec_duplicate:V1TI (match_operand:TI 1 "general_operand" "v,R,v,j00,jm1,d")))] 552 "TARGET_VX" 553 "@ 554 vlr\t%v0,%v1 555 vl\t%v0,%1 556 vst\t%v1,%0 557 vzero\t%v0 558 vone\t%v0 559 vlvgp\t%v0,%1,%N1" 560 [(set_attr "op_type" "VRR,VRX,VRX,VRI,VRI,VRR")]) 561 562 ; vec_splats is supposed to replicate op1 into all elements of op0 563 ; This splitter first sets the rightmost element of op0 to op1 and 564 ; then does a vec_splat to replicate that element into all other 565 ; elements. 566 (define_split 567 [(set (match_operand:V_128_NOSINGLE 0 "register_operand" "") 568 (vec_duplicate:V_128_NOSINGLE (match_operand:<non_vec> 1 "register_operand" "")))] 569 "TARGET_VX && GENERAL_REG_P (operands[1])" 570 [(set (match_dup 0) 571 (unspec:V_128_NOSINGLE [(match_dup 1) (match_dup 2) (match_dup 0)] UNSPEC_VEC_SET)) 572 (set (match_dup 0) 573 (vec_duplicate:V_128_NOSINGLE 574 (vec_select:<non_vec> 575 (match_dup 0) (parallel [(match_dup 2)]))))] 576 { 577 operands[2] = GEN_INT (GET_MODE_NUNITS (<MODE>mode) - 1); 578 }) 579 580 (define_expand "vcond<V_HW:mode><V_HW2:mode>" 581 [(set (match_operand:V_HW 0 "register_operand" "") 582 (if_then_else:V_HW 583 (match_operator 3 "comparison_operator" 584 [(match_operand:V_HW2 4 "register_operand" "") 585 (match_operand:V_HW2 5 "nonmemory_operand" "")]) 586 (match_operand:V_HW 1 "nonmemory_operand" "") 587 (match_operand:V_HW 2 "nonmemory_operand" "")))] 588 "TARGET_VX && GET_MODE_NUNITS (<V_HW:MODE>mode) == GET_MODE_NUNITS (<V_HW2:MODE>mode)" 589 { 590 s390_expand_vcond (operands[0], operands[1], operands[2], 591 GET_CODE (operands[3]), operands[4], operands[5]); 592 DONE; 593 }) 594 595 (define_expand "vcondu<V_HW:mode><V_HW2:mode>" 596 [(set (match_operand:V_HW 0 "register_operand" "") 597 (if_then_else:V_HW 598 (match_operator 3 "comparison_operator" 599 [(match_operand:V_HW2 4 "register_operand" "") 600 (match_operand:V_HW2 5 "nonmemory_operand" "")]) 601 (match_operand:V_HW 1 "nonmemory_operand" "") 602 (match_operand:V_HW 2 "nonmemory_operand" "")))] 603 "TARGET_VX && GET_MODE_NUNITS (<V_HW:MODE>mode) == GET_MODE_NUNITS (<V_HW2:MODE>mode)" 604 { 605 s390_expand_vcond (operands[0], operands[1], operands[2], 606 GET_CODE (operands[3]), operands[4], operands[5]); 607 DONE; 608 }) 609 610 ; We only have HW support for byte vectors. The middle-end is 611 ; supposed to lower the mode if required. 612 (define_insn "vec_permv16qi" 613 [(set (match_operand:V16QI 0 "register_operand" "=v") 614 (unspec:V16QI [(match_operand:V16QI 1 "register_operand" "v") 615 (match_operand:V16QI 2 "register_operand" "v") 616 (match_operand:V16QI 3 "register_operand" "v")] 617 UNSPEC_VEC_PERM))] 618 "TARGET_VX" 619 "vperm\t%v0,%v1,%v2,%v3" 620 [(set_attr "op_type" "VRR")]) 621 622 ; vec_perm_const for V2DI using vpdi? 623 624 ;; 625 ;; Vector integer arithmetic instructions 626 ;; 627 628 ; vab, vah, vaf, vag, vaq 629 630 ; We use nonimmediate_operand instead of register_operand since it is 631 ; better to have the reloads into VRs instead of splitting the 632 ; operation into two DImode ADDs. 633 (define_insn "<ti*>add<mode>3" 634 [(set (match_operand:VIT 0 "nonimmediate_operand" "=v") 635 (plus:VIT (match_operand:VIT 1 "nonimmediate_operand" "%v") 636 (match_operand:VIT 2 "general_operand" "v")))] 637 "TARGET_VX" 638 "va<bhfgq>\t%v0,%v1,%v2" 639 [(set_attr "op_type" "VRR")]) 640 641 ; vsb, vsh, vsf, vsg, vsq 642 (define_insn "<ti*>sub<mode>3" 643 [(set (match_operand:VIT 0 "nonimmediate_operand" "=v") 644 (minus:VIT (match_operand:VIT 1 "nonimmediate_operand" "v") 645 (match_operand:VIT 2 "general_operand" "v")))] 646 "TARGET_VX" 647 "vs<bhfgq>\t%v0,%v1,%v2" 648 [(set_attr "op_type" "VRR")]) 649 650 ; vmlb, vmlhw, vmlf 651 (define_insn "mul<mode>3" 652 [(set (match_operand:VI_QHS 0 "register_operand" "=v") 653 (mult:VI_QHS (match_operand:VI_QHS 1 "register_operand" "%v") 654 (match_operand:VI_QHS 2 "register_operand" "v")))] 655 "TARGET_VX" 656 "vml<bhfgq><w>\t%v0,%v1,%v2" 657 [(set_attr "op_type" "VRR")]) 658 659 ; vlcb, vlch, vlcf, vlcg 660 (define_insn "neg<mode>2" 661 [(set (match_operand:VI 0 "register_operand" "=v") 662 (neg:VI (match_operand:VI 1 "register_operand" "v")))] 663 "TARGET_VX" 664 "vlc<bhfgq>\t%v0,%v1" 665 [(set_attr "op_type" "VRR")]) 666 667 ; vlpb, vlph, vlpf, vlpg 668 (define_insn "abs<mode>2" 669 [(set (match_operand:VI 0 "register_operand" "=v") 670 (abs:VI (match_operand:VI 1 "register_operand" "v")))] 671 "TARGET_VX" 672 "vlp<bhfgq>\t%v0,%v1" 673 [(set_attr "op_type" "VRR")]) 674 675 676 ; Vector sum across 677 678 ; Sum across DImode parts of the 1st operand and add the rightmost 679 ; element of 2nd operand 680 ; vsumgh, vsumgf 681 (define_insn "*vec_sum2<mode>" 682 [(set (match_operand:V2DI 0 "register_operand" "=v") 683 (unspec:V2DI [(match_operand:VI_HW_HS 1 "register_operand" "v") 684 (match_operand:VI_HW_HS 2 "register_operand" "v")] 685 UNSPEC_VEC_VSUMG))] 686 "TARGET_VX" 687 "vsumg<bhfgq>\t%v0,%v1,%v2" 688 [(set_attr "op_type" "VRR")]) 689 690 ; vsumb, vsumh 691 (define_insn "*vec_sum4<mode>" 692 [(set (match_operand:V4SI 0 "register_operand" "=v") 693 (unspec:V4SI [(match_operand:VI_HW_QH 1 "register_operand" "v") 694 (match_operand:VI_HW_QH 2 "register_operand" "v")] 695 UNSPEC_VEC_VSUM))] 696 "TARGET_VX" 697 "vsum<bhfgq>\t%v0,%v1,%v2" 698 [(set_attr "op_type" "VRR")]) 699 700 ;; 701 ;; Vector bit instructions (int + fp) 702 ;; 703 704 ; Vector and 705 706 (define_insn "and<mode>3" 707 [(set (match_operand:VT 0 "register_operand" "=v") 708 (and:VT (match_operand:VT 1 "register_operand" "%v") 709 (match_operand:VT 2 "register_operand" "v")))] 710 "TARGET_VX" 711 "vn\t%v0,%v1,%v2" 712 [(set_attr "op_type" "VRR")]) 713 714 ; Vector not and 715 716 (define_insn "notand<mode>3" 717 [(set (match_operand:VT 0 "register_operand" "=v") 718 (ior:VT (not:VT (match_operand:VT 1 "register_operand" "%v")) 719 (not:VT (match_operand:VT 2 "register_operand" "v"))))] 720 "TARGET_VXE" 721 "vnn\t%v0,%v1,%v2" 722 [(set_attr "op_type" "VRR")]) 723 724 ; Vector or 725 726 (define_insn "ior<mode>3" 727 [(set (match_operand:VT 0 "register_operand" "=v") 728 (ior:VT (match_operand:VT 1 "register_operand" "%v") 729 (match_operand:VT 2 "register_operand" "v")))] 730 "TARGET_VX" 731 "vo\t%v0,%v1,%v2" 732 [(set_attr "op_type" "VRR")]) 733 734 ; Vector or with complement 735 736 (define_insn "ior_not<mode>3" 737 [(set (match_operand:VT 0 "register_operand" "=v") 738 (ior:VT (not:VT (match_operand:VT 2 "register_operand" "v")) 739 (match_operand:VT 1 "register_operand" "%v")))] 740 "TARGET_VXE" 741 "voc\t%v0,%v1,%v2" 742 [(set_attr "op_type" "VRR")]) 743 744 ; Vector xor 745 746 (define_insn "xor<mode>3" 747 [(set (match_operand:VT 0 "register_operand" "=v") 748 (xor:VT (match_operand:VT 1 "register_operand" "%v") 749 (match_operand:VT 2 "register_operand" "v")))] 750 "TARGET_VX" 751 "vx\t%v0,%v1,%v2" 752 [(set_attr "op_type" "VRR")]) 753 754 ; Vector not xor 755 756 (define_insn "notxor<mode>3" 757 [(set (match_operand:VT 0 "register_operand" "=v") 758 (not:VT (xor:VT (match_operand:VT 1 "register_operand" "%v") 759 (match_operand:VT 2 "register_operand" "v"))))] 760 "TARGET_VXE" 761 "vnx\t%v0,%v1,%v2" 762 [(set_attr "op_type" "VRR")]) 763 764 ; Bitwise inversion of a vector 765 (define_insn "one_cmpl<mode>2" 766 [(set (match_operand:VT 0 "register_operand" "=v") 767 (not:VT (match_operand:VT 1 "register_operand" "v")))] 768 "TARGET_VX" 769 "vnot\t%v0,%v1" 770 [(set_attr "op_type" "VRR")]) 771 772 ; Vector population count 773 774 (define_expand "popcount<mode>2" 775 [(set (match_operand:VI_HW 0 "register_operand" "=v") 776 (unspec:VI_HW [(match_operand:VI_HW 1 "register_operand" "v")] 777 UNSPEC_POPCNT))] 778 "TARGET_VX" 779 { 780 if (TARGET_VXE) 781 emit_insn (gen_popcount<mode>2_vxe (operands[0], operands[1])); 782 else 783 emit_insn (gen_popcount<mode>2_vx (operands[0], operands[1])); 784 DONE; 785 }) 786 787 ; vpopctb, vpopcth, vpopctf, vpopctg 788 (define_insn "popcount<mode>2_vxe" 789 [(set (match_operand:VI_HW 0 "register_operand" "=v") 790 (unspec:VI_HW [(match_operand:VI_HW 1 "register_operand" "v")] 791 UNSPEC_POPCNT))] 792 "TARGET_VXE" 793 "vpopct<bhfgq>\t%v0,%v1" 794 [(set_attr "op_type" "VRR")]) 795 796 (define_insn "popcountv16qi2_vx" 797 [(set (match_operand:V16QI 0 "register_operand" "=v") 798 (unspec:V16QI [(match_operand:V16QI 1 "register_operand" "v")] 799 UNSPEC_POPCNT))] 800 "TARGET_VX && !TARGET_VXE" 801 "vpopct\t%v0,%v1,0" 802 [(set_attr "op_type" "VRR")]) 803 804 ; vpopct only counts bits in byte elements. Bigger element sizes need 805 ; to be emulated. Word and doubleword elements can use the sum across 806 ; instructions. For halfword sized elements we do a shift of a copy 807 ; of the result, add it to the result and extend it to halfword 808 ; element size (unpack). 809 810 (define_expand "popcountv8hi2_vx" 811 [(set (match_dup 2) 812 (unspec:V16QI [(subreg:V16QI (match_operand:V8HI 1 "register_operand" "v") 0)] 813 UNSPEC_POPCNT)) 814 ; Make a copy of the result 815 (set (match_dup 3) (match_dup 2)) 816 ; Generate the shift count operand in a VR (8->byte 7) 817 (set (match_dup 4) (match_dup 5)) 818 (set (match_dup 4) (unspec:V16QI [(const_int 8) 819 (const_int 7) 820 (match_dup 4)] UNSPEC_VEC_SET)) 821 ; Vector shift right logical by one byte 822 (set (match_dup 3) 823 (unspec:V16QI [(match_dup 3) (match_dup 4)] UNSPEC_VEC_SRLB)) 824 ; Add the shifted and the original result 825 (set (match_dup 2) 826 (plus:V16QI (match_dup 2) (match_dup 3))) 827 ; Generate mask for the odd numbered byte elements 828 (set (match_dup 3) 829 (const_vector:V16QI [(const_int 0) (const_int 255) 830 (const_int 0) (const_int 255) 831 (const_int 0) (const_int 255) 832 (const_int 0) (const_int 255) 833 (const_int 0) (const_int 255) 834 (const_int 0) (const_int 255) 835 (const_int 0) (const_int 255) 836 (const_int 0) (const_int 255)])) 837 ; Zero out the even indexed bytes 838 (set (match_operand:V8HI 0 "register_operand" "=v") 839 (and:V8HI (subreg:V8HI (match_dup 2) 0) 840 (subreg:V8HI (match_dup 3) 0))) 841 ] 842 "TARGET_VX && !TARGET_VXE" 843 { 844 operands[2] = gen_reg_rtx (V16QImode); 845 operands[3] = gen_reg_rtx (V16QImode); 846 operands[4] = gen_reg_rtx (V16QImode); 847 operands[5] = CONST0_RTX (V16QImode); 848 }) 849 850 (define_expand "popcountv4si2_vx" 851 [(set (match_dup 2) 852 (unspec:V16QI [(subreg:V16QI (match_operand:V4SI 1 "register_operand" "v") 0)] 853 UNSPEC_POPCNT)) 854 (set (match_operand:V4SI 0 "register_operand" "=v") 855 (unspec:V4SI [(match_dup 2) (match_dup 3)] 856 UNSPEC_VEC_VSUM))] 857 "TARGET_VX && !TARGET_VXE" 858 { 859 operands[2] = gen_reg_rtx (V16QImode); 860 operands[3] = force_reg (V16QImode, CONST0_RTX (V16QImode)); 861 }) 862 863 (define_expand "popcountv2di2_vx" 864 [(set (match_dup 2) 865 (unspec:V16QI [(subreg:V16QI (match_operand:V2DI 1 "register_operand" "v") 0)] 866 UNSPEC_POPCNT)) 867 (set (match_dup 3) 868 (unspec:V4SI [(match_dup 2) (match_dup 4)] 869 UNSPEC_VEC_VSUM)) 870 (set (match_operand:V2DI 0 "register_operand" "=v") 871 (unspec:V2DI [(match_dup 3) (match_dup 5)] 872 UNSPEC_VEC_VSUMG))] 873 "TARGET_VX && !TARGET_VXE" 874 { 875 operands[2] = gen_reg_rtx (V16QImode); 876 operands[3] = gen_reg_rtx (V4SImode); 877 operands[4] = force_reg (V16QImode, CONST0_RTX (V16QImode)); 878 operands[5] = force_reg (V4SImode, CONST0_RTX (V4SImode)); 879 }) 880 881 ; Count leading zeros 882 ; vclzb, vclzh, vclzf, vclzg 883 (define_insn "clz<mode>2" 884 [(set (match_operand:V 0 "register_operand" "=v") 885 (clz:V (match_operand:V 1 "register_operand" "v")))] 886 "TARGET_VX" 887 "vclz<bhfgq>\t%v0,%v1" 888 [(set_attr "op_type" "VRR")]) 889 890 ; Count trailing zeros 891 ; vctzb, vctzh, vctzf, vctzg 892 (define_insn "ctz<mode>2" 893 [(set (match_operand:V 0 "register_operand" "=v") 894 (ctz:V (match_operand:V 1 "register_operand" "v")))] 895 "TARGET_VX" 896 "vctz<bhfgq>\t%v0,%v1" 897 [(set_attr "op_type" "VRR")]) 898 899 900 901 ; Each vector element rotated by the corresponding vector element 902 ; verllvb, verllvh, verllvf, verllvg 903 (define_insn "vrotl<mode>3" 904 [(set (match_operand:VI 0 "register_operand" "=v") 905 (rotate:VI (match_operand:VI 1 "register_operand" "v") 906 (match_operand:VI 2 "register_operand" "v")))] 907 "TARGET_VX" 908 "verllv<bhfgq>\t%v0,%v1,%v2" 909 [(set_attr "op_type" "VRR")]) 910 911 912 ; Vector rotate and shift by scalar instructions 913 914 (define_code_iterator VEC_SHIFTS [ashift ashiftrt lshiftrt rotate]) 915 (define_code_attr vec_shifts_name [(ashift "ashl") (ashiftrt "ashr") 916 (lshiftrt "lshr") (rotate "rotl")]) 917 (define_code_attr vec_shifts_mnem [(ashift "vesl") (ashiftrt "vesra") 918 (lshiftrt "vesrl") (rotate "verll")]) 919 920 ; Each vector element rotated by a scalar 921 (define_expand "<vec_shifts_name><mode>3" 922 [(set (match_operand:VI 0 "register_operand" "") 923 (VEC_SHIFTS:VI (match_operand:VI 1 "register_operand" "") 924 (match_operand:SI 2 "nonmemory_operand" "")))] 925 "TARGET_VX") 926 927 ; verllb, verllh, verllf, verllg 928 ; veslb, veslh, veslf, veslg 929 ; vesrab, vesrah, vesraf, vesrag 930 ; vesrlb, vesrlh, vesrlf, vesrlg 931 (define_insn "*<vec_shifts_name><mode>3<addr_style_op>" 932 [(set (match_operand:VI 0 "register_operand" "=v") 933 (VEC_SHIFTS:VI (match_operand:VI 1 "register_operand" "v") 934 (match_operand:SI 2 "nonmemory_operand" "an")))] 935 "TARGET_VX" 936 "<vec_shifts_mnem><bhfgq>\t%v0,%v1,%Y2" 937 [(set_attr "op_type" "VRS")]) 938 939 ; Shift each element by corresponding vector element 940 941 ; veslvb, veslvh, veslvf, veslvg 942 (define_insn "vashl<mode>3" 943 [(set (match_operand:VI 0 "register_operand" "=v") 944 (ashift:VI (match_operand:VI 1 "register_operand" "v") 945 (match_operand:VI 2 "register_operand" "v")))] 946 "TARGET_VX" 947 "veslv<bhfgq>\t%v0,%v1,%v2" 948 [(set_attr "op_type" "VRR")]) 949 950 ; vesravb, vesravh, vesravf, vesravg 951 (define_insn "vashr<mode>3" 952 [(set (match_operand:VI 0 "register_operand" "=v") 953 (ashiftrt:VI (match_operand:VI 1 "register_operand" "v") 954 (match_operand:VI 2 "register_operand" "v")))] 955 "TARGET_VX" 956 "vesrav<bhfgq>\t%v0,%v1,%v2" 957 [(set_attr "op_type" "VRR")]) 958 959 ; vesrlvb, vesrlvh, vesrlvf, vesrlvg 960 (define_insn "vlshr<mode>3" 961 [(set (match_operand:VI 0 "register_operand" "=v") 962 (lshiftrt:VI (match_operand:VI 1 "register_operand" "v") 963 (match_operand:VI 2 "register_operand" "v")))] 964 "TARGET_VX" 965 "vesrlv<bhfgq>\t%v0,%v1,%v2" 966 [(set_attr "op_type" "VRR")]) 967 968 ; Vector shift right logical by byte 969 970 ; Pattern used by e.g. popcount 971 (define_insn "*vec_srb<mode>" 972 [(set (match_operand:V_HW 0 "register_operand" "=v") 973 (unspec:V_HW [(match_operand:V_HW 1 "register_operand" "v") 974 (match_operand:<tointvec> 2 "register_operand" "v")] 975 UNSPEC_VEC_SRLB))] 976 "TARGET_VX" 977 "vsrlb\t%v0,%v1,%v2" 978 [(set_attr "op_type" "VRR")]) 979 980 981 ; vmnb, vmnh, vmnf, vmng 982 (define_insn "smin<mode>3" 983 [(set (match_operand:VI 0 "register_operand" "=v") 984 (smin:VI (match_operand:VI 1 "register_operand" "%v") 985 (match_operand:VI 2 "register_operand" "v")))] 986 "TARGET_VX" 987 "vmn<bhfgq>\t%v0,%v1,%v2" 988 [(set_attr "op_type" "VRR")]) 989 990 ; vmxb, vmxh, vmxf, vmxg 991 (define_insn "smax<mode>3" 992 [(set (match_operand:VI 0 "register_operand" "=v") 993 (smax:VI (match_operand:VI 1 "register_operand" "%v") 994 (match_operand:VI 2 "register_operand" "v")))] 995 "TARGET_VX" 996 "vmx<bhfgq>\t%v0,%v1,%v2" 997 [(set_attr "op_type" "VRR")]) 998 999 ; vmnlb, vmnlh, vmnlf, vmnlg 1000 (define_insn "umin<mode>3" 1001 [(set (match_operand:VI 0 "register_operand" "=v") 1002 (umin:VI (match_operand:VI 1 "register_operand" "%v") 1003 (match_operand:VI 2 "register_operand" "v")))] 1004 "TARGET_VX" 1005 "vmnl<bhfgq>\t%v0,%v1,%v2" 1006 [(set_attr "op_type" "VRR")]) 1007 1008 ; vmxlb, vmxlh, vmxlf, vmxlg 1009 (define_insn "umax<mode>3" 1010 [(set (match_operand:VI 0 "register_operand" "=v") 1011 (umax:VI (match_operand:VI 1 "register_operand" "%v") 1012 (match_operand:VI 2 "register_operand" "v")))] 1013 "TARGET_VX" 1014 "vmxl<bhfgq>\t%v0,%v1,%v2" 1015 [(set_attr "op_type" "VRR")]) 1016 1017 ; vmeb, vmeh, vmef 1018 (define_insn "vec_widen_smult_even_<mode>" 1019 [(set (match_operand:<vec_double> 0 "register_operand" "=v") 1020 (unspec:<vec_double> [(match_operand:VI_QHS 1 "register_operand" "%v") 1021 (match_operand:VI_QHS 2 "register_operand" "v")] 1022 UNSPEC_VEC_SMULT_EVEN))] 1023 "TARGET_VX" 1024 "vme<bhfgq>\t%v0,%v1,%v2" 1025 [(set_attr "op_type" "VRR")]) 1026 1027 ; vmleb, vmleh, vmlef 1028 (define_insn "vec_widen_umult_even_<mode>" 1029 [(set (match_operand:<vec_double> 0 "register_operand" "=v") 1030 (unspec:<vec_double> [(match_operand:VI_QHS 1 "register_operand" "%v") 1031 (match_operand:VI_QHS 2 "register_operand" "v")] 1032 UNSPEC_VEC_UMULT_EVEN))] 1033 "TARGET_VX" 1034 "vmle<bhfgq>\t%v0,%v1,%v2" 1035 [(set_attr "op_type" "VRR")]) 1036 1037 ; vmob, vmoh, vmof 1038 (define_insn "vec_widen_smult_odd_<mode>" 1039 [(set (match_operand:<vec_double> 0 "register_operand" "=v") 1040 (unspec:<vec_double> [(match_operand:VI_QHS 1 "register_operand" "%v") 1041 (match_operand:VI_QHS 2 "register_operand" "v")] 1042 UNSPEC_VEC_SMULT_ODD))] 1043 "TARGET_VX" 1044 "vmo<bhfgq>\t%v0,%v1,%v2" 1045 [(set_attr "op_type" "VRR")]) 1046 1047 ; vmlob, vmloh, vmlof 1048 (define_insn "vec_widen_umult_odd_<mode>" 1049 [(set (match_operand:<vec_double> 0 "register_operand" "=v") 1050 (unspec:<vec_double> [(match_operand:VI_QHS 1 "register_operand" "%v") 1051 (match_operand:VI_QHS 2 "register_operand" "v")] 1052 UNSPEC_VEC_UMULT_ODD))] 1053 "TARGET_VX" 1054 "vmlo<bhfgq>\t%v0,%v1,%v2" 1055 [(set_attr "op_type" "VRR")]) 1056 1057 ; vec_widen_umult_hi 1058 ; vec_widen_umult_lo 1059 ; vec_widen_smult_hi 1060 ; vec_widen_smult_lo 1061 1062 ; vec_widen_ushiftl_hi 1063 ; vec_widen_ushiftl_lo 1064 ; vec_widen_sshiftl_hi 1065 ; vec_widen_sshiftl_lo 1066 1067 ;; 1068 ;; Vector floating point arithmetic instructions 1069 ;; 1070 1071 ; vfasb, vfadb, wfasb, wfadb, wfaxb 1072 (define_insn "add<mode>3" 1073 [(set (match_operand:VF_HW 0 "register_operand" "=v") 1074 (plus:VF_HW (match_operand:VF_HW 1 "register_operand" "%v") 1075 (match_operand:VF_HW 2 "register_operand" "v")))] 1076 "TARGET_VX" 1077 "<vw>fa<sdx>b\t%v0,%v1,%v2" 1078 [(set_attr "op_type" "VRR")]) 1079 1080 ; vfssb, vfsdb, wfssb, wfsdb, wfsxb 1081 (define_insn "sub<mode>3" 1082 [(set (match_operand:VF_HW 0 "register_operand" "=v") 1083 (minus:VF_HW (match_operand:VF_HW 1 "register_operand" "%v") 1084 (match_operand:VF_HW 2 "register_operand" "v")))] 1085 "TARGET_VX" 1086 "<vw>fs<sdx>b\t%v0,%v1,%v2" 1087 [(set_attr "op_type" "VRR")]) 1088 1089 ; vfmsb, vfmdb, wfmsb, wfmdb, wfmxb 1090 (define_insn "mul<mode>3" 1091 [(set (match_operand:VF_HW 0 "register_operand" "=v") 1092 (mult:VF_HW (match_operand:VF_HW 1 "register_operand" "%v") 1093 (match_operand:VF_HW 2 "register_operand" "v")))] 1094 "TARGET_VX" 1095 "<vw>fm<sdx>b\t%v0,%v1,%v2" 1096 [(set_attr "op_type" "VRR")]) 1097 1098 ; vfdsb, vfddb, wfdsb, wfddb, wfdxb 1099 (define_insn "div<mode>3" 1100 [(set (match_operand:VF_HW 0 "register_operand" "=v") 1101 (div:VF_HW (match_operand:VF_HW 1 "register_operand" "v") 1102 (match_operand:VF_HW 2 "register_operand" "v")))] 1103 "TARGET_VX" 1104 "<vw>fd<sdx>b\t%v0,%v1,%v2" 1105 [(set_attr "op_type" "VRR")]) 1106 1107 ; vfsqsb, vfsqdb, wfsqsb, wfsqdb, wfsqxb 1108 (define_insn "sqrt<mode>2" 1109 [(set (match_operand:VF_HW 0 "register_operand" "=v") 1110 (sqrt:VF_HW (match_operand:VF_HW 1 "register_operand" "v")))] 1111 "TARGET_VX" 1112 "<vw>fsq<sdx>b\t%v0,%v1" 1113 [(set_attr "op_type" "VRR")]) 1114 1115 ; vfmasb, vfmadb, wfmasb, wfmadb, wfmaxb 1116 (define_insn "fma<mode>4" 1117 [(set (match_operand:VF_HW 0 "register_operand" "=v") 1118 (fma:VF_HW (match_operand:VF_HW 1 "register_operand" "%v") 1119 (match_operand:VF_HW 2 "register_operand" "v") 1120 (match_operand:VF_HW 3 "register_operand" "v")))] 1121 "TARGET_VX" 1122 "<vw>fma<sdx>b\t%v0,%v1,%v2,%v3" 1123 [(set_attr "op_type" "VRR")]) 1124 1125 ; vfmssb, vfmsdb, wfmssb, wfmsdb, wfmsxb 1126 (define_insn "fms<mode>4" 1127 [(set (match_operand:VF_HW 0 "register_operand" "=v") 1128 (fma:VF_HW (match_operand:VF_HW 1 "register_operand" "%v") 1129 (match_operand:VF_HW 2 "register_operand" "v") 1130 (neg:VF_HW (match_operand:VF_HW 3 "register_operand" "v"))))] 1131 "TARGET_VX" 1132 "<vw>fms<sdx>b\t%v0,%v1,%v2,%v3" 1133 [(set_attr "op_type" "VRR")]) 1134 1135 ; vfnmasb, vfnmadb, wfnmasb, wfnmadb, wfnmaxb 1136 (define_insn "neg_fma<mode>4" 1137 [(set (match_operand:VF_HW 0 "register_operand" "=v") 1138 (neg:VF_HW 1139 (fma:VF_HW (match_operand:VF_HW 1 "register_operand" "%v") 1140 (match_operand:VF_HW 2 "register_operand" "v") 1141 (match_operand:VF_HW 3 "register_operand" "v"))))] 1142 "TARGET_VXE" 1143 "<vw>fnma<sdx>b\t%v0,%v1,%v2,%v3" 1144 [(set_attr "op_type" "VRR")]) 1145 1146 ; vfnmssb, vfnmsdb, wfnmssb, wfnmsdb, wfnmsxb 1147 (define_insn "neg_fms<mode>4" 1148 [(set (match_operand:VF_HW 0 "register_operand" "=v") 1149 (neg:VF_HW 1150 (fma:VF_HW (match_operand:VF_HW 1 "register_operand" "%v") 1151 (match_operand:VF_HW 2 "register_operand" "v") 1152 (neg:VF_HW (match_operand:VF_HW 3 "register_operand" "v")))))] 1153 "TARGET_VXE" 1154 "<vw>fnms<sdx>b\t%v0,%v1,%v2,%v3" 1155 [(set_attr "op_type" "VRR")]) 1156 1157 ; vflcsb, vflcdb, wflcsb, wflcdb, wflcxb 1158 (define_insn "neg<mode>2" 1159 [(set (match_operand:VFT 0 "register_operand" "=v") 1160 (neg:VFT (match_operand:VFT 1 "register_operand" "v")))] 1161 "TARGET_VX" 1162 "<vw>flc<sdx>b\t%v0,%v1" 1163 [(set_attr "op_type" "VRR")]) 1164 1165 ; vflpsb, vflpdb, wflpsb, wflpdb, wflpxb 1166 (define_insn "abs<mode>2" 1167 [(set (match_operand:VFT 0 "register_operand" "=v") 1168 (abs:VFT (match_operand:VFT 1 "register_operand" "v")))] 1169 "TARGET_VX" 1170 "<vw>flp<sdx>b\t%v0,%v1" 1171 [(set_attr "op_type" "VRR")]) 1172 1173 ; vflnsb, vflndb, wflnsb, wflndb, wflnxb 1174 (define_insn "negabs<mode>2" 1175 [(set (match_operand:VFT 0 "register_operand" "=v") 1176 (neg:VFT (abs:VFT (match_operand:VFT 1 "register_operand" "v"))))] 1177 "TARGET_VX" 1178 "<vw>fln<sdx>b\t%v0,%v1" 1179 [(set_attr "op_type" "VRR")]) 1180 1181 (define_expand "smax<mode>3" 1182 [(set (match_operand:VF_HW 0 "register_operand") 1183 (smax:VF_HW (match_operand:VF_HW 1 "register_operand") 1184 (match_operand:VF_HW 2 "register_operand")))] 1185 "TARGET_VX") 1186 1187 ; vfmaxsb, vfmaxdb, wfmaxsb, wfmaxdb, wfmaxxb 1188 (define_insn "*smax<mode>3_vxe" 1189 [(set (match_operand:VF_HW 0 "register_operand" "=v") 1190 (smax:VF_HW (match_operand:VF_HW 1 "register_operand" "%v") 1191 (match_operand:VF_HW 2 "register_operand" "v")))] 1192 "TARGET_VXE" 1193 "<vw>fmax<sdx>b\t%v0,%v1,%v2,4" 1194 [(set_attr "op_type" "VRR")]) 1195 1196 ; Emulate with compare + select 1197 (define_insn_and_split "*smaxv2df3_vx" 1198 [(set (match_operand:V2DF 0 "register_operand" "=v") 1199 (smax:V2DF (match_operand:V2DF 1 "register_operand" "%v") 1200 (match_operand:V2DF 2 "register_operand" "v")))] 1201 "TARGET_VX && !TARGET_VXE" 1202 "#" 1203 "&& 1" 1204 [(set (match_dup 3) 1205 (gt:V2DI (match_dup 1) (match_dup 2))) 1206 (set (match_dup 0) 1207 (if_then_else:V2DF 1208 (eq (match_dup 3) (match_dup 4)) 1209 (match_dup 2) 1210 (match_dup 1)))] 1211 { 1212 operands[3] = gen_reg_rtx (V2DImode); 1213 operands[4] = CONST0_RTX (V2DImode); 1214 }) 1215 1216 (define_expand "smin<mode>3" 1217 [(set (match_operand:VF_HW 0 "register_operand") 1218 (smin:VF_HW (match_operand:VF_HW 1 "register_operand") 1219 (match_operand:VF_HW 2 "register_operand")))] 1220 "TARGET_VX") 1221 1222 ; vfminsb, vfmindb, wfminsb, wfmindb, wfminxb 1223 (define_insn "*smin<mode>3_vxe" 1224 [(set (match_operand:VF_HW 0 "register_operand" "=v") 1225 (smin:VF_HW (match_operand:VF_HW 1 "register_operand" "%v") 1226 (match_operand:VF_HW 2 "register_operand" "v")))] 1227 "TARGET_VXE" 1228 "<vw>fmin<sdx>b\t%v0,%v1,%v2,4" 1229 [(set_attr "op_type" "VRR")]) 1230 1231 ; Emulate with compare + select 1232 (define_insn_and_split "*sminv2df3_vx" 1233 [(set (match_operand:V2DF 0 "register_operand" "=v") 1234 (smin:V2DF (match_operand:V2DF 1 "register_operand" "%v") 1235 (match_operand:V2DF 2 "register_operand" "v")))] 1236 "TARGET_VX && !TARGET_VXE" 1237 "#" 1238 "&& 1" 1239 [(set (match_dup 3) 1240 (gt:V2DI (match_dup 1) (match_dup 2))) 1241 (set (match_dup 0) 1242 (if_then_else:V2DF 1243 (eq (match_dup 3) (match_dup 4)) 1244 (match_dup 1) 1245 (match_dup 2)))] 1246 { 1247 operands[3] = gen_reg_rtx (V2DImode); 1248 operands[4] = CONST0_RTX (V2DImode); 1249 }) 1250 1251 1252 ;; 1253 ;; Integer compares 1254 ;; 1255 1256 (define_insn "*vec_cmp<VICMP_HW_OP:code><VI:mode>_nocc" 1257 [(set (match_operand:VI 2 "register_operand" "=v") 1258 (VICMP_HW_OP:VI (match_operand:VI 0 "register_operand" "v") 1259 (match_operand:VI 1 "register_operand" "v")))] 1260 "TARGET_VX" 1261 "vc<VICMP_HW_OP:insn_cmp_op><VI:bhfgq>\t%v2,%v0,%v1" 1262 [(set_attr "op_type" "VRR")]) 1263 1264 1265 ;; 1266 ;; Floating point compares 1267 ;; 1268 1269 ; EQ, GT, GE 1270 ; vfcesb, vfcedb, wfcexb, vfchsb, vfchdb, wfchxb, vfchesb, vfchedb, wfchexb 1271 (define_insn "*vec_cmp<VFCMP_HW_OP:code><mode>_nocc" 1272 [(set (match_operand:<tointvec> 0 "register_operand" "=v") 1273 (VFCMP_HW_OP:<tointvec> (match_operand:VFT 1 "register_operand" "v") 1274 (match_operand:VFT 2 "register_operand" "v")))] 1275 "TARGET_VX" 1276 "<vw>fc<VFCMP_HW_OP:asm_fcmp_op><sdx>b\t%v0,%v1,%v2" 1277 [(set_attr "op_type" "VRR")]) 1278 1279 ; Expanders for not directly supported comparisons 1280 1281 ; UNEQ a u== b -> !(a > b | b > a) 1282 (define_expand "vec_cmpuneq<mode>" 1283 [(set (match_operand:<tointvec> 0 "register_operand" "=v") 1284 (gt:<tointvec> (match_operand:VFT 1 "register_operand" "v") 1285 (match_operand:VFT 2 "register_operand" "v"))) 1286 (set (match_dup 3) 1287 (gt:<tointvec> (match_dup 2) (match_dup 1))) 1288 (set (match_dup 0) (ior:<tointvec> (match_dup 0) (match_dup 3))) 1289 (set (match_dup 0) (not:<tointvec> (match_dup 0)))] 1290 "TARGET_VX" 1291 { 1292 operands[3] = gen_reg_rtx (<tointvec>mode); 1293 }) 1294 1295 ; LTGT a <> b -> a > b | b > a 1296 (define_expand "vec_cmpltgt<mode>" 1297 [(set (match_operand:<tointvec> 0 "register_operand" "=v") 1298 (gt:<tointvec> (match_operand:VFT 1 "register_operand" "v") 1299 (match_operand:VFT 2 "register_operand" "v"))) 1300 (set (match_dup 3) (gt:<tointvec> (match_dup 2) (match_dup 1))) 1301 (set (match_dup 0) (ior:<tointvec> (match_dup 0) (match_dup 3)))] 1302 "TARGET_VX" 1303 { 1304 operands[3] = gen_reg_rtx (<tointvec>mode); 1305 }) 1306 1307 ; ORDERED (a, b): a >= b | b > a 1308 (define_expand "vec_ordered<mode>" 1309 [(set (match_operand:<tointvec> 0 "register_operand" "=v") 1310 (ge:<tointvec> (match_operand:VFT 1 "register_operand" "v") 1311 (match_operand:VFT 2 "register_operand" "v"))) 1312 (set (match_dup 3) (gt:<tointvec> (match_dup 2) (match_dup 1))) 1313 (set (match_dup 0) (ior:<tointvec> (match_dup 0) (match_dup 3)))] 1314 "TARGET_VX" 1315 { 1316 operands[3] = gen_reg_rtx (<tointvec>mode); 1317 }) 1318 1319 ; UNORDERED (a, b): !ORDERED (a, b) 1320 (define_expand "vec_unordered<mode>" 1321 [(set (match_operand:<tointvec> 0 "register_operand" "=v") 1322 (ge:<tointvec> (match_operand:VFT 1 "register_operand" "v") 1323 (match_operand:VFT 2 "register_operand" "v"))) 1324 (set (match_dup 3) (gt:<tointvec> (match_dup 2) (match_dup 1))) 1325 (set (match_dup 0) (ior:<tointvec> (match_dup 0) (match_dup 3))) 1326 (set (match_dup 0) (not:<tointvec> (match_dup 0)))] 1327 "TARGET_VX" 1328 { 1329 operands[3] = gen_reg_rtx (<tointvec>mode); 1330 }) 1331 1332 (define_insn "*vec_load_pair<mode>" 1333 [(set (match_operand:V_HW_64 0 "register_operand" "=v,v") 1334 (vec_concat:V_HW_64 (match_operand:<non_vec> 1 "register_operand" "d,v") 1335 (match_operand:<non_vec> 2 "register_operand" "d,v")))] 1336 "TARGET_VX" 1337 "@ 1338 vlvgp\t%v0,%1,%2 1339 vmrhg\t%v0,%v1,%v2" 1340 [(set_attr "op_type" "VRR,VRR")]) 1341 1342 (define_insn "vllv16qi" 1343 [(set (match_operand:V16QI 0 "register_operand" "=v") 1344 (unspec:V16QI [(match_operand:SI 1 "register_operand" "d") 1345 (match_operand:BLK 2 "memory_operand" "Q")] 1346 UNSPEC_VEC_LOAD_LEN))] 1347 "TARGET_VX" 1348 "vll\t%v0,%1,%2" 1349 [(set_attr "op_type" "VRS")]) 1350 1351 ; vfenebs, vfenehs, vfenefs 1352 ; vfenezbs, vfenezhs, vfenezfs 1353 (define_insn "vec_vfenes<mode>" 1354 [(set (match_operand:VI_HW_QHS 0 "register_operand" "=v") 1355 (unspec:VI_HW_QHS [(match_operand:VI_HW_QHS 1 "register_operand" "v") 1356 (match_operand:VI_HW_QHS 2 "register_operand" "v") 1357 (match_operand:QI 3 "const_mask_operand" "C")] 1358 UNSPEC_VEC_VFENE)) 1359 (set (reg:CCRAW CC_REGNUM) 1360 (unspec:CCRAW [(match_dup 1) 1361 (match_dup 2) 1362 (match_dup 3)] 1363 UNSPEC_VEC_VFENECC))] 1364 "TARGET_VX" 1365 { 1366 unsigned HOST_WIDE_INT flags = UINTVAL (operands[3]); 1367 1368 gcc_assert (!(flags & ~(VSTRING_FLAG_ZS | VSTRING_FLAG_CS))); 1369 flags &= ~VSTRING_FLAG_CS; 1370 1371 if (flags == VSTRING_FLAG_ZS) 1372 return "vfenez<bhfgq>s\t%v0,%v1,%v2"; 1373 return "vfene<bhfgq>s\t%v0,%v1,%v2"; 1374 } 1375 [(set_attr "op_type" "VRR")]) 1376 1377 1378 ; Vector select 1379 1380 ; The following splitters simplify vec_sel for constant 0 or -1 1381 ; selection sources. This is required to generate efficient code for 1382 ; vcond. 1383 1384 ; a = b == c; 1385 (define_split 1386 [(set (match_operand:V 0 "register_operand" "") 1387 (if_then_else:V 1388 (eq (match_operand:<tointvec> 3 "register_operand" "") 1389 (match_operand:V 4 "const0_operand" "")) 1390 (match_operand:V 1 "const0_operand" "") 1391 (match_operand:V 2 "all_ones_operand" "")))] 1392 "TARGET_VX" 1393 [(set (match_dup 0) (match_dup 3))] 1394 { 1395 PUT_MODE (operands[3], <V:MODE>mode); 1396 }) 1397 1398 ; a = ~(b == c) 1399 (define_split 1400 [(set (match_operand:V 0 "register_operand" "") 1401 (if_then_else:V 1402 (eq (match_operand:<tointvec> 3 "register_operand" "") 1403 (match_operand:V 4 "const0_operand" "")) 1404 (match_operand:V 1 "all_ones_operand" "") 1405 (match_operand:V 2 "const0_operand" "")))] 1406 "TARGET_VX" 1407 [(set (match_dup 0) (not:V (match_dup 3)))] 1408 { 1409 PUT_MODE (operands[3], <V:MODE>mode); 1410 }) 1411 1412 ; a = b != c 1413 (define_split 1414 [(set (match_operand:V 0 "register_operand" "") 1415 (if_then_else:V 1416 (ne (match_operand:<tointvec> 3 "register_operand" "") 1417 (match_operand:V 4 "const0_operand" "")) 1418 (match_operand:V 1 "all_ones_operand" "") 1419 (match_operand:V 2 "const0_operand" "")))] 1420 "TARGET_VX" 1421 [(set (match_dup 0) (match_dup 3))] 1422 { 1423 PUT_MODE (operands[3], <V:MODE>mode); 1424 }) 1425 1426 ; a = ~(b != c) 1427 (define_split 1428 [(set (match_operand:V 0 "register_operand" "") 1429 (if_then_else:V 1430 (ne (match_operand:<tointvec> 3 "register_operand" "") 1431 (match_operand:V 4 "const0_operand" "")) 1432 (match_operand:V 1 "const0_operand" "") 1433 (match_operand:V 2 "all_ones_operand" "")))] 1434 "TARGET_VX" 1435 [(set (match_dup 0) (not:V (match_dup 3)))] 1436 { 1437 PUT_MODE (operands[3], <V:MODE>mode); 1438 }) 1439 1440 ; op0 = op3 == 0 ? op1 : op2 1441 (define_insn "*vec_sel0<mode>" 1442 [(set (match_operand:V 0 "register_operand" "=v") 1443 (if_then_else:V 1444 (eq (match_operand:<tointvec> 3 "register_operand" "v") 1445 (match_operand:<tointvec> 4 "const0_operand" "")) 1446 (match_operand:V 1 "register_operand" "v") 1447 (match_operand:V 2 "register_operand" "v")))] 1448 "TARGET_VX" 1449 "vsel\t%v0,%2,%1,%3" 1450 [(set_attr "op_type" "VRR")]) 1451 1452 ; op0 = !op3 == 0 ? op1 : op2 1453 (define_insn "*vec_sel0<mode>" 1454 [(set (match_operand:V 0 "register_operand" "=v") 1455 (if_then_else:V 1456 (eq (not:<tointvec> (match_operand:<tointvec> 3 "register_operand" "v")) 1457 (match_operand:<tointvec> 4 "const0_operand" "")) 1458 (match_operand:V 1 "register_operand" "v") 1459 (match_operand:V 2 "register_operand" "v")))] 1460 "TARGET_VX" 1461 "vsel\t%v0,%1,%2,%3" 1462 [(set_attr "op_type" "VRR")]) 1463 1464 ; op0 = op3 == -1 ? op1 : op2 1465 (define_insn "*vec_sel1<mode>" 1466 [(set (match_operand:V 0 "register_operand" "=v") 1467 (if_then_else:V 1468 (eq (match_operand:<tointvec> 3 "register_operand" "v") 1469 (match_operand:<tointvec> 4 "all_ones_operand" "")) 1470 (match_operand:V 1 "register_operand" "v") 1471 (match_operand:V 2 "register_operand" "v")))] 1472 "TARGET_VX" 1473 "vsel\t%v0,%1,%2,%3" 1474 [(set_attr "op_type" "VRR")]) 1475 1476 ; op0 = !op3 == -1 ? op1 : op2 1477 (define_insn "*vec_sel1<mode>" 1478 [(set (match_operand:V 0 "register_operand" "=v") 1479 (if_then_else:V 1480 (eq (not:<tointvec> (match_operand:<tointvec> 3 "register_operand" "v")) 1481 (match_operand:<tointvec> 4 "all_ones_operand" "")) 1482 (match_operand:V 1 "register_operand" "v") 1483 (match_operand:V 2 "register_operand" "v")))] 1484 "TARGET_VX" 1485 "vsel\t%v0,%2,%1,%3" 1486 [(set_attr "op_type" "VRR")]) 1487 1488 ; vec_pack_trunc 1489 1490 ; vpkh, vpkf, vpkg 1491 (define_insn "vec_pack_trunc_<mode>" 1492 [(set (match_operand:<vec_half> 0 "register_operand" "=v") 1493 (vec_concat:<vec_half> 1494 (truncate:<vec_halfhalf> 1495 (match_operand:VI_HW_HSD 1 "register_operand" "v")) 1496 (truncate:<vec_halfhalf> 1497 (match_operand:VI_HW_HSD 2 "register_operand" "v"))))] 1498 "TARGET_VX" 1499 "vpk<bhfgq>\t%0,%1,%2" 1500 [(set_attr "op_type" "VRR")]) 1501 1502 ; vpksh, vpksf, vpksg 1503 (define_insn "vec_pack_ssat_<mode>" 1504 [(set (match_operand:<vec_half> 0 "register_operand" "=v") 1505 (vec_concat:<vec_half> 1506 (ss_truncate:<vec_halfhalf> 1507 (match_operand:VI_HW_HSD 1 "register_operand" "v")) 1508 (ss_truncate:<vec_halfhalf> 1509 (match_operand:VI_HW_HSD 2 "register_operand" "v"))))] 1510 "TARGET_VX" 1511 "vpks<bhfgq>\t%0,%1,%2" 1512 [(set_attr "op_type" "VRR")]) 1513 1514 ; vpklsh, vpklsf, vpklsg 1515 (define_insn "vec_pack_usat_<mode>" 1516 [(set (match_operand:<vec_half> 0 "register_operand" "=v") 1517 (vec_concat:<vec_half> 1518 (us_truncate:<vec_halfhalf> 1519 (match_operand:VI_HW_HSD 1 "register_operand" "v")) 1520 (us_truncate:<vec_halfhalf> 1521 (match_operand:VI_HW_HSD 2 "register_operand" "v"))))] 1522 "TARGET_VX" 1523 "vpkls<bhfgq>\t%0,%1,%2" 1524 [(set_attr "op_type" "VRR")]) 1525 1526 ;; vector unpack v16qi 1527 1528 ; signed 1529 1530 (define_insn "vec_unpacks_hi_v16qi" 1531 [(set (match_operand:V8HI 0 "register_operand" "=v") 1532 (sign_extend:V8HI 1533 (vec_select:V8QI 1534 (match_operand:V16QI 1 "register_operand" "v") 1535 (parallel [(const_int 0)(const_int 1)(const_int 2)(const_int 3) 1536 (const_int 4)(const_int 5)(const_int 6)(const_int 7)]))))] 1537 "TARGET_VX" 1538 "vuphb\t%0,%1" 1539 [(set_attr "op_type" "VRR")]) 1540 1541 (define_insn "vec_unpacks_low_v16qi" 1542 [(set (match_operand:V8HI 0 "register_operand" "=v") 1543 (sign_extend:V8HI 1544 (vec_select:V8QI 1545 (match_operand:V16QI 1 "register_operand" "v") 1546 (parallel [(const_int 8) (const_int 9) (const_int 10)(const_int 11) 1547 (const_int 12)(const_int 13)(const_int 14)(const_int 15)]))))] 1548 "TARGET_VX" 1549 "vuplb\t%0,%1" 1550 [(set_attr "op_type" "VRR")]) 1551 1552 ; unsigned 1553 1554 (define_insn "vec_unpacku_hi_v16qi" 1555 [(set (match_operand:V8HI 0 "register_operand" "=v") 1556 (zero_extend:V8HI 1557 (vec_select:V8QI 1558 (match_operand:V16QI 1 "register_operand" "v") 1559 (parallel [(const_int 0)(const_int 1)(const_int 2)(const_int 3) 1560 (const_int 4)(const_int 5)(const_int 6)(const_int 7)]))))] 1561 "TARGET_VX" 1562 "vuplhb\t%0,%1" 1563 [(set_attr "op_type" "VRR")]) 1564 1565 (define_insn "vec_unpacku_low_v16qi" 1566 [(set (match_operand:V8HI 0 "register_operand" "=v") 1567 (zero_extend:V8HI 1568 (vec_select:V8QI 1569 (match_operand:V16QI 1 "register_operand" "v") 1570 (parallel [(const_int 8) (const_int 9) (const_int 10)(const_int 11) 1571 (const_int 12)(const_int 13)(const_int 14)(const_int 15)]))))] 1572 "TARGET_VX" 1573 "vupllb\t%0,%1" 1574 [(set_attr "op_type" "VRR")]) 1575 1576 ;; vector unpack v8hi 1577 1578 ; signed 1579 1580 (define_insn "vec_unpacks_hi_v8hi" 1581 [(set (match_operand:V4SI 0 "register_operand" "=v") 1582 (sign_extend:V4SI 1583 (vec_select:V4HI 1584 (match_operand:V8HI 1 "register_operand" "v") 1585 (parallel [(const_int 0)(const_int 1)(const_int 2)(const_int 3)]))))] 1586 "TARGET_VX" 1587 "vuphh\t%0,%1" 1588 [(set_attr "op_type" "VRR")]) 1589 1590 (define_insn "vec_unpacks_lo_v8hi" 1591 [(set (match_operand:V4SI 0 "register_operand" "=v") 1592 (sign_extend:V4SI 1593 (vec_select:V4HI 1594 (match_operand:V8HI 1 "register_operand" "v") 1595 (parallel [(const_int 4)(const_int 5)(const_int 6)(const_int 7)]))))] 1596 "TARGET_VX" 1597 "vuplhw\t%0,%1" 1598 [(set_attr "op_type" "VRR")]) 1599 1600 ; unsigned 1601 1602 (define_insn "vec_unpacku_hi_v8hi" 1603 [(set (match_operand:V4SI 0 "register_operand" "=v") 1604 (zero_extend:V4SI 1605 (vec_select:V4HI 1606 (match_operand:V8HI 1 "register_operand" "v") 1607 (parallel [(const_int 0)(const_int 1)(const_int 2)(const_int 3)]))))] 1608 "TARGET_VX" 1609 "vuplhh\t%0,%1" 1610 [(set_attr "op_type" "VRR")]) 1611 1612 (define_insn "vec_unpacku_lo_v8hi" 1613 [(set (match_operand:V4SI 0 "register_operand" "=v") 1614 (zero_extend:V4SI 1615 (vec_select:V4HI 1616 (match_operand:V8HI 1 "register_operand" "v") 1617 (parallel [(const_int 4)(const_int 5)(const_int 6)(const_int 7)]))))] 1618 "TARGET_VX" 1619 "vupllh\t%0,%1" 1620 [(set_attr "op_type" "VRR")]) 1621 1622 ;; vector unpack v4si 1623 1624 ; signed 1625 1626 (define_insn "vec_unpacks_hi_v4si" 1627 [(set (match_operand:V2DI 0 "register_operand" "=v") 1628 (sign_extend:V2DI 1629 (vec_select:V2SI 1630 (match_operand:V4SI 1 "register_operand" "v") 1631 (parallel [(const_int 0)(const_int 1)]))))] 1632 "TARGET_VX" 1633 "vuphf\t%0,%1" 1634 [(set_attr "op_type" "VRR")]) 1635 1636 (define_insn "vec_unpacks_lo_v4si" 1637 [(set (match_operand:V2DI 0 "register_operand" "=v") 1638 (sign_extend:V2DI 1639 (vec_select:V2SI 1640 (match_operand:V4SI 1 "register_operand" "v") 1641 (parallel [(const_int 2)(const_int 3)]))))] 1642 "TARGET_VX" 1643 "vuplf\t%0,%1" 1644 [(set_attr "op_type" "VRR")]) 1645 1646 ; unsigned 1647 1648 (define_insn "vec_unpacku_hi_v4si" 1649 [(set (match_operand:V2DI 0 "register_operand" "=v") 1650 (zero_extend:V2DI 1651 (vec_select:V2SI 1652 (match_operand:V4SI 1 "register_operand" "v") 1653 (parallel [(const_int 0)(const_int 1)]))))] 1654 "TARGET_VX" 1655 "vuplhf\t%0,%1" 1656 [(set_attr "op_type" "VRR")]) 1657 1658 (define_insn "vec_unpacku_lo_v4si" 1659 [(set (match_operand:V2DI 0 "register_operand" "=v") 1660 (zero_extend:V2DI 1661 (vec_select:V2SI 1662 (match_operand:V4SI 1 "register_operand" "v") 1663 (parallel [(const_int 2)(const_int 3)]))))] 1664 "TARGET_VX" 1665 "vupllf\t%0,%1" 1666 [(set_attr "op_type" "VRR")]) 1667 1668 ;; vector load lengthened 1669 1670 ; vflls 1671 (define_insn "*vec_extendv4sf" 1672 [(set (match_operand:V2DF 0 "register_operand" "=v") 1673 (float_extend:V2DF 1674 (vec_select:V2SF 1675 (match_operand:V4SF 1 "register_operand" "v") 1676 (parallel [(const_int 0) (const_int 2)]))))] 1677 "TARGET_VX" 1678 "vldeb\t%v0,%v1" 1679 [(set_attr "op_type" "VRR")]) 1680 1681 (define_insn "*vec_extendv2df" 1682 [(set (match_operand:V1TF 0 "register_operand" "=v") 1683 (float_extend:V1TF 1684 (vec_select:V1DF 1685 (match_operand:V2DF 1 "register_operand" "v") 1686 (parallel [(const_int 0)]))))] 1687 "TARGET_VXE" 1688 "wflld\t%v0,%v1" 1689 [(set_attr "op_type" "VRR")]) 1690 1691 ; reduc_smin 1692 ; reduc_smax 1693 ; reduc_umin 1694 ; reduc_umax 1695 1696 ; vec_shl vrep + vsl 1697 ; vec_shr 1698 1699 ; vec_pack_sfix_trunc: convert + pack ? 1700 ; vec_pack_ufix_trunc 1701 ; vec_unpacks_float_hi 1702 ; vec_unpacks_float_lo 1703 ; vec_unpacku_float_hi 1704 ; vec_unpacku_float_lo 1705
Indexes created Wed Apr 15 00:22:22 UTC 2026