Home | History | Annotate | Line # | Download | only in x86_64
      1 /* $NetBSD: fenv.c,v 1.11 2024/02/20 03:53:48 riastradh Exp $ */
      2 
      3 /*-
      4  * Copyright (c) 2004-2005 David Schultz <das (at) FreeBSD.ORG>
      5  * All rights reserved.
      6  *
      7  * Redistribution and use in source and binary forms, with or without
      8  * modification, are permitted provided that the following conditions
      9  * are met:
     10  * 1. Redistributions of source code must retain the above copyright
     11  *    notice, this list of conditions and the following disclaimer.
     12  * 2. Redistributions in binary form must reproduce the above copyright
     13  *    notice, this list of conditions and the following disclaimer in the
     14  *    documentation and/or other materials provided with the distribution.
     15  *
     16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
     17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
     20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     26  * SUCH DAMAGE.
     27  */
     28 
     29 #include <sys/cdefs.h>
     30 __RCSID("$NetBSD: fenv.c,v 1.11 2024/02/20 03:53:48 riastradh Exp $");
     31 
     32 #include "namespace.h"
     33 
     34 #include <assert.h>
     35 #include <fenv.h>
     36 #include <stddef.h>
     37 #include <string.h>
     38 
     39 #ifdef __weak_alias
     40 __weak_alias(feclearexcept,_feclearexcept)
     41 __weak_alias(fedisableexcept,_fedisableexcept)
     42 __weak_alias(feenableexcept,_feenableexcept)
     43 __weak_alias(fegetenv,_fegetenv)
     44 __weak_alias(fegetexcept,_fegetexcept)
     45 __weak_alias(fegetexceptflag,_fegetexceptflag)
     46 __weak_alias(fegetround,_fegetround)
     47 __weak_alias(feholdexcept,_feholdexcept)
     48 __weak_alias(feraiseexcept,_feraiseexcept)
     49 __weak_alias(fesetenv,_fesetenv)
     50 __weak_alias(fesetexceptflag,_fesetexceptflag)
     51 __weak_alias(fesetround,_fesetround)
     52 __weak_alias(fetestexcept,_fetestexcept)
     53 __weak_alias(feupdateenv,_feupdateenv)
     54 #endif
     55 
     56 /* Load x87 Control Word */
     57 #define	__fldcw(__cw)		__asm__ __volatile__ \
     58 	("fldcw %0" : : "m" (__cw))
     59 
     60 /* No-Wait Store Control Word */
     61 #define	__fnstcw(__cw)		__asm__ __volatile__ \
     62 	("fnstcw %0" : "=m" (*(__cw)))
     63 
     64 /* No-Wait Store Status Word */
     65 #define	__fnstsw(__sw)		__asm__ __volatile__ \
     66 	("fnstsw %0" : "=am" (*(__sw)))
     67 
     68 /* No-Wait Clear Exception Flags */
     69 #define	__fnclex()		__asm__ __volatile__ \
     70 	("fnclex")
     71 
     72 /* Load x87 Environment */
     73 #define	__fldenv(__env)		__asm__ __volatile__ \
     74 	("fldenv %0" : : "m" (__env))
     75 
     76 /* No-Wait Store x87 environment */
     77 #define	__fnstenv(__env)	__asm__ __volatile__ \
     78 	("fnstenv %0" : "=m" (*(__env)))
     79 
     80 /* Check for and handle pending unmasked x87 pending FPU exceptions */
     81 #define	__fwait(__env)		__asm__	__volatile__	\
     82 	("fwait")
     83 
     84 /* Load the MXCSR register */
     85 #define	__ldmxcsr(__mxcsr)	__asm__ __volatile__ \
     86 	("ldmxcsr %0" : : "m" (__mxcsr))
     87 
     88 /* Store the MXCSR register state */
     89 #define	__stmxcsr(__mxcsr)	__asm__ __volatile__ \
     90 	("stmxcsr %0" : "=m" (*(__mxcsr)))
     91 
     92 /*
     93  * The following constant represents the default floating-point environment
     94  * (that is, the one installed at program startup) and has type pointer to
     95  * const-qualified fenv_t.
     96  *
     97  * It can be used as an argument to the functions within the <fenv.h> header
     98  * that manage the floating-point environment, namely fesetenv() and
     99  * feupdateenv().
    100  *
    101  * x87 fpu registers are 16bit wide. The upper bits, 31-16, are marked as
    102  * RESERVED. We provide a partial floating-point environment, where we
    103  * define only the lower bits. The reserved bits are extracted and set by
    104  * the consumers of FE_DFL_ENV, during runtime.
    105  */
    106 fenv_t __fe_dfl_env = {
    107 	{
    108 		__NetBSD_NPXCW__,	/* Control word register */
    109 		0x00000000,		/* Status word register */
    110 		0x0000ffff,		/* Tag word register */
    111 		{
    112 			0x00000000,
    113 			0x00000000,
    114 			0x00000000,
    115 			0x00000000,
    116 		},
    117 	},
    118 	__INITIAL_MXCSR__       /* MXCSR register */
    119 };
    120 #define FE_DFL_ENV      ((const fenv_t *) &__fe_dfl_env)
    121 
    122 static void __init_libm(void) __attribute__ ((constructor, used));
    123 
    124 static void __init_libm(void)
    125 {
    126 	uint16_t control;
    127 
    128 	__fnstcw(&control);
    129 	__fe_dfl_env.x87.control = control;
    130 }
    131 
    132 
    133 /*
    134  * The feclearexcept() function clears the supported floating-point exceptions
    135  * represented by `excepts'.
    136  */
    137 int
    138 feclearexcept(int excepts)
    139 {
    140 	fenv_t fenv;
    141 	int ex;
    142 
    143 	_DIAGASSERT((excepts & ~FE_ALL_EXCEPT) == 0);
    144 
    145 	ex = excepts & FE_ALL_EXCEPT;
    146 
    147 	/* Store the current x87 floating-point environment */
    148 	__fnstenv(&fenv);
    149 
    150 	/* Clear the requested floating-point exceptions */
    151 	fenv.x87.status &= ~ex;
    152 
    153 	/* Load the x87 floating-point environment */
    154 	__fldenv(fenv);
    155 
    156 	/* Same for SSE environment */
    157 	__stmxcsr(&fenv.mxcsr);
    158 	fenv.mxcsr &= ~ex;
    159 	__ldmxcsr(fenv.mxcsr);
    160 
    161 	/* Success */
    162 	return (0);
    163 }
    164 
    165 /*
    166  * The fegetexceptflag() function stores an implementation-defined
    167  * representation of the states of the floating-point status flags indicated by
    168  * the argument excepts in the object pointed to by the argument flagp.
    169  */
    170 int
    171 fegetexceptflag(fexcept_t *flagp, int excepts)
    172 {
    173 	uint32_t mxcsr;
    174 	uint16_t x87_status;
    175 	int ex;
    176 
    177 	_DIAGASSERT(flagp != NULL);
    178 	_DIAGASSERT((excepts & ~FE_ALL_EXCEPT) == 0);
    179 
    180 	ex = excepts & FE_ALL_EXCEPT;
    181 
    182 	/* Store the current x87 status register */
    183 	__fnstsw(&x87_status);
    184 
    185 	/* Store the MXCSR register */
    186 	__stmxcsr(&mxcsr);
    187 
    188 	/* Store the results in flagp */
    189 	*flagp = (x87_status | mxcsr) & ex;
    190 
    191 	/* Success */
    192 	return (0);
    193 }
    194 
    195 /*
    196  * The feraiseexcept() function raises the supported floating-point exceptions
    197  * represented by the argument `excepts'.
    198  *
    199  * The standard explicitly allows us to execute an instruction that has the
    200  * exception as a side effect, but we choose to manipulate the status register
    201  * directly.
    202  *
    203  * The validation of input is being deferred to fesetexceptflag().
    204  */
    205 int
    206 feraiseexcept(int excepts)
    207 {
    208 	int ex;
    209 
    210 	_DIAGASSERT((excepts & ~FE_ALL_EXCEPT) == 0);
    211 
    212 	ex = excepts & FE_ALL_EXCEPT;
    213 	fesetexceptflag((unsigned int *)&ex, ex);
    214 	__fwait();
    215 
    216 	/* Success */
    217 	return (0);
    218 }
    219 
    220 /*
    221  * This function sets the floating-point status flags indicated by the argument
    222  * `excepts' to the states stored in the object pointed to by `flagp'. It does
    223  * NOT raise any floating-point exceptions, but only sets the state of the flags.
    224  */
    225 int
    226 fesetexceptflag(const fexcept_t *flagp, int excepts)
    227 {
    228 	fenv_t fenv;
    229 	int ex;
    230 
    231 	_DIAGASSERT(flagp != NULL);
    232 	_DIAGASSERT((excepts & ~FE_ALL_EXCEPT) == 0);
    233 
    234 	ex = excepts & FE_ALL_EXCEPT;
    235 
    236 	/* Store the current x87 floating-point environment */
    237 	__fnstenv(&fenv);
    238 
    239 	/* Set the requested status flags */
    240 	fenv.x87.status |= *flagp & ex;
    241 
    242 	/* Load the x87 floating-point environment */
    243 	__fldenv(fenv);
    244 
    245 	/* Same for SSE environment */
    246 	__stmxcsr(&fenv.mxcsr);
    247 	fenv.mxcsr |= *flagp & ex;
    248 	__ldmxcsr(fenv.mxcsr);
    249 
    250 	/* Success */
    251 	return (0);
    252 }
    253 
    254 /*
    255  * The fetestexcept() function determines which of a specified subset of the
    256  * floating-point exception flags are currently set. The `excepts' argument
    257  * specifies the floating-point status flags to be queried.
    258  */
    259 int
    260 fetestexcept(int excepts)
    261 {
    262 	uint32_t mxcsr;
    263 	uint16_t status;
    264 	int ex;
    265 
    266 	_DIAGASSERT((excepts & ~FE_ALL_EXCEPT) == 0);
    267 
    268 	ex = excepts & FE_ALL_EXCEPT;
    269 
    270 	__fnstsw(&status);
    271 	__stmxcsr(&mxcsr);
    272 
    273 	return ((status | mxcsr) & ex);
    274 }
    275 
    276 /*
    277  * The fegetround() function gets the current rounding direction.
    278  */
    279 int
    280 fegetround(void)
    281 {
    282 	uint32_t mxcsr;
    283 	uint16_t control;
    284 
    285 	/*
    286 	 * We check both the x87 floating-point unit _and_ the SSE unit.
    287 	 * Normally, those two must agree with respect to each other. If they
    288 	 * don't, it's not our fault and the result is non-determinable, in
    289 	 * which case POSIX says that a negative value should be returned.
    290 	 */
    291 	__fnstcw(&control);
    292 	__stmxcsr(&mxcsr);
    293 
    294 	if ((control & _X87_ROUNDING_MASK)
    295 	    != ((mxcsr & _SSE_ROUNDING_MASK) >> 3)) {
    296 		return (-1);
    297 	}
    298 
    299 	return (control & _X87_ROUNDING_MASK);
    300 }
    301 
    302 /*
    303  * The fesetround() function establishes the rounding direction represented by
    304  * its argument `round'. If the argument is not equal to the value of a rounding
    305  * direction macro, the rounding direction is not changed.
    306  */
    307 int
    308 fesetround(int round)
    309 {
    310 	uint32_t  mxcsr;
    311 	uint16_t control;
    312 
    313 	/* Check whether requested rounding direction is supported */
    314 	if (round & (~_X87_ROUNDING_MASK))
    315 		return (-1);
    316 
    317 	/* Store the current x87 control word register  */
    318 	__fnstcw(&control);
    319 
    320 	/*
    321 	 * Set the rounding direction
    322 	 * Rounding Control is bits 10-11, so shift appropriately
    323 	 */
    324 	control &= ~_X87_ROUNDING_MASK;
    325 	control |= round;
    326 
    327 	/* Load the x87 control word register */
    328 	__fldcw(control);
    329 
    330 	/*
    331 	 * Same for the SSE environment
    332 	 * Rounding Control is bits 13-14, so shift appropriately
    333 	 */
    334 	__stmxcsr(&mxcsr);
    335 	mxcsr &= ~_SSE_ROUNDING_MASK;
    336 	mxcsr |= (round << _SSE_ROUND_SHIFT);
    337 	__ldmxcsr(mxcsr);
    338 
    339 	/* Success */
    340 	return (0);
    341 }
    342 
    343 /*
    344  * The fegetenv() function attempts to store the current floating-point
    345  * environment in the object pointed to by envp.
    346  */
    347 int
    348 fegetenv(fenv_t *envp)
    349 {
    350 	_DIAGASSERT(envp != NULL);
    351 
    352 	/* Store the current x87 floating-point environment */
    353 	__fnstenv(envp);
    354 
    355 	/* Store the MXCSR register state */
    356 	__stmxcsr(&envp->mxcsr);
    357 
    358      /*
    359       * When an FNSTENV instruction is executed, all pending exceptions are
    360       * essentially lost (either the x87 FPU status register is cleared or all
    361       * exceptions are masked).
    362       *
    363       * 8.6 X87 FPU EXCEPTION SYNCHRONIZATION -
    364       * Intel(R) 64 and IA-32 Architectures Software Developer's Manual - Vol 1
    365       *
    366       */
    367 	__fldcw(envp->x87.control);
    368 
    369 	/* Success */
    370 	return (0);
    371 }
    372 
    373 /*
    374  * The feholdexcept() function saves the current floating-point environment
    375  * in the object pointed to by envp, clears the floating-point status flags, and
    376  * then installs a non-stop (continue on floating-point exceptions) mode, if
    377  * available, for all floating-point exceptions.
    378  */
    379 int
    380 feholdexcept(fenv_t *envp)
    381 {
    382 	uint32_t mxcsr;
    383 
    384 	_DIAGASSERT(envp != NULL);
    385 
    386 	/* Store the current x87 floating-point environment */
    387 	__fnstenv(envp);
    388 
    389 	/* Clear all exception flags in FPU */
    390 	__fnclex();
    391 
    392 	/* Store the MXCSR register state */
    393 	__stmxcsr(&envp->mxcsr);
    394 
    395 	/* Clear exception flags in MXCSR XXX */
    396 	mxcsr = envp->mxcsr;
    397 	mxcsr &= ~FE_ALL_EXCEPT;
    398 
    399 	/* Mask all exceptions */
    400 	mxcsr |= FE_ALL_EXCEPT << _SSE_EMASK_SHIFT;
    401 
    402 	__ldmxcsr(mxcsr);
    403 
    404 	/* Success */
    405 	return (0);
    406 }
    407 
    408 /*
    409  * The fesetenv() function attempts to establish the floating-point environment
    410  * represented by the object pointed to by envp. The argument `envp' points
    411  * to an object set by a call to fegetenv() or feholdexcept(), or equal a
    412  * floating-point environment macro. The fesetenv() function does not raise
    413  * floating-point exceptions, but only installs the state of the floating-point
    414  * status flags represented through its argument.
    415  */
    416 int
    417 fesetenv(const fenv_t *envp)
    418 {
    419 	fenv_t fenv;
    420 
    421 	_DIAGASSERT(envp != NULL);
    422 
    423 	/* Store the x87 floating-point environment */
    424 	memset(&fenv, 0, sizeof fenv);
    425 	__fnstenv(&fenv);
    426 
    427 	__fe_dfl_env.x87.control = (fenv.x87.control & 0xffff0000)
    428 	    | (__fe_dfl_env.x87.control & 0x0000ffff);
    429 	__fe_dfl_env.x87.status = (fenv.x87.status & 0xffff0000)
    430 	    | (__fe_dfl_env.x87.status & 0x0000ffff);
    431 	__fe_dfl_env.x87.tag = (fenv.x87.tag & 0xffff0000)
    432 	    | (__fe_dfl_env.x87.tag & 0x0000ffff);
    433 	__fe_dfl_env.x87.others[3] = (fenv.x87.others[3] & 0xffff0000)
    434 	    | (__fe_dfl_env.x87.others[3] & 0x0000ffff);
    435 	__fldenv(*envp);
    436 
    437 	/* Store the MXCSR register */
    438 	__ldmxcsr(envp->mxcsr);
    439 
    440 	/* Success */
    441 	return (0);
    442 }
    443 
    444 /*
    445  * The feupdateenv() function saves the currently raised floating-point
    446  * exceptions in its automatic storage, installs the floating-point environment
    447  * represented by the object pointed to by `envp', and then raises the saved
    448  * floating-point exceptions. The argument `envp' shall point to an object set
    449  * by a call to feholdexcept() or fegetenv(), or equal a floating-point
    450  * environment macro.
    451  */
    452 int
    453 feupdateenv(const fenv_t *envp)
    454 {
    455 	fenv_t fenv;
    456 	uint32_t mxcsr;
    457 	uint16_t sw;
    458 
    459 	_DIAGASSERT(envp != NULL);
    460 
    461 	/* Store the x87 floating-point environment */
    462 	memset(&fenv, 0, sizeof(fenv));
    463 	__fnstenv(&fenv);
    464 
    465 	__fe_dfl_env.x87.control = (fenv.x87.control & 0xffff0000)
    466 	    | (__fe_dfl_env.x87.control & 0x0000ffff);
    467 	__fe_dfl_env.x87.status = (fenv.x87.status & 0xffff0000)
    468 	    | (__fe_dfl_env.x87.status & 0x0000ffff);
    469 	__fe_dfl_env.x87.tag = (fenv.x87.tag & 0xffff0000)
    470 	    | (__fe_dfl_env.x87.tag & 0x0000ffff);
    471 	__fe_dfl_env.x87.others[3] = (fenv.x87.others[3] & 0xffff0000)
    472 	    | (__fe_dfl_env.x87.others[3] & 0x0000ffff);
    473 
    474 	/* Store the x87 status register */
    475 	__fnstsw(&sw);
    476 
    477 	/* Store the MXCSR register */
    478 	__stmxcsr(&mxcsr);
    479 
    480 	/* Install new floating-point environment */
    481 	fesetenv(envp);
    482 
    483 	/* Raise any previously accumulated exceptions */
    484 	feraiseexcept((sw | mxcsr) & FE_ALL_EXCEPT);
    485 
    486 	/* Success */
    487 	return (0);
    488 }
    489 
    490 /*
    491  * The following functions are extensions to the standard
    492  */
    493 int
    494 feenableexcept(int mask)
    495 {
    496 	uint32_t mxcsr, omask;
    497 	uint16_t control;
    498 
    499 	_DIAGASSERT((mask & ~FE_ALL_EXCEPT) == 0);
    500 	mask &= FE_ALL_EXCEPT;
    501 
    502 	__fnstcw(&control);
    503 	__stmxcsr(&mxcsr);
    504 
    505 	omask = (control | mxcsr >> _SSE_EMASK_SHIFT) & FE_ALL_EXCEPT;
    506 	control &= ~mask;
    507 	__fldcw(control);
    508 
    509 	mxcsr &= ~(mask << _SSE_EMASK_SHIFT);
    510 	__ldmxcsr(mxcsr);
    511 
    512 	return (FE_ALL_EXCEPT & ~omask);
    513 
    514 }
    515 
    516 int
    517 fedisableexcept(int mask)
    518 {
    519 	uint32_t mxcsr, omask;
    520 	uint16_t control;
    521 
    522 	_DIAGASSERT((mask & ~FE_ALL_EXCEPT) == 0);
    523 
    524 	__fnstcw(&control);
    525 	__stmxcsr(&mxcsr);
    526 
    527 	omask = (control | mxcsr >> _SSE_EMASK_SHIFT) & FE_ALL_EXCEPT;
    528 	control |= mask;
    529 	__fldcw(control);
    530 
    531 	mxcsr |= mask << _SSE_EMASK_SHIFT;
    532 	__ldmxcsr(mxcsr);
    533 
    534 	return (FE_ALL_EXCEPT & ~omask);
    535 }
    536 
    537 int
    538 fegetexcept(void)
    539 {
    540 	uint16_t control;
    541 
    542 	/*
    543 	 * We assume that the masks for the x87 and the SSE unit are
    544 	 * the same.
    545 	 */
    546 	__fnstcw(&control);
    547 
    548 	return (~control & FE_ALL_EXCEPT);
    549 }
    550 
    551