libgcc/config/i386/crtfastmath.c

   1 /*
   2  * Copyright (C) 2005-2013 Free Software Foundation, Inc.
   3  *
   4  * This file is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License as published by the
   6  * Free Software Foundation; either version 3, or (at your option) any
   7  * later version.
   8  *
   9  * This file is distributed in the hope that it will be useful, but
  10  * WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  12  * General Public License for more details.
  13  *
  14  * Under Section 7 of GPL version 3, you are granted additional
  15  * permissions described in the GCC Runtime Library Exception, version
  16  * 3.1, as published by the Free Software Foundation.
  17  *
  18  * You should have received a copy of the GNU General Public License and
  19  * a copy of the GCC Runtime Library Exception along with this program;
  20  * see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
  21  * <http://www.gnu.org/licenses/>.
  22  */
  23
  24 #define MXCSR_DAZ (1 << 6)      /* Enable denormals are zero mode */
  25 #define MXCSR_FTZ (1 << 15)     /* Enable flush to zero mode */
  26
  27 #ifndef __x86_64__
  28 /* All 64-bit targets have SSE and DAZ;
  29    only check them explicitly for 32-bit ones. */
  30 #include "cpuid.h"
  31 #endif
  32
  33 #if !defined __x86_64__ && defined __sun__ && defined __svr4__
  34 #include <signal.h>
  35 #include <ucontext.h>
  36
  37 static volatile sig_atomic_t sigill_caught;
  38
  39 static void
  40 sigill_hdlr (int sig __attribute((unused)),
  41              siginfo_t *sip __attribute__((unused)),
  42              ucontext_t *ucp)
  43 {
  44   sigill_caught = 1;
  45   /* Set PC to the instruction after the faulting one to skip over it,
  46      otherwise we enter an infinite loop.  3 is the size of the movaps
  47      instruction.  */
  48   ucp->uc_mcontext.gregs[EIP] += 3;
  49   setcontext (ucp);
  50 }
  51 #endif
  52
  53 static void __attribute__((constructor))
  54 #ifndef __x86_64__
  55 /* The i386 ABI only requires 4-byte stack alignment, so this is necessary
  56    to make sure the fxsave struct gets correct alignment.
  57    See PR27537 and PR28621.  */
  58 __attribute__ ((force_align_arg_pointer))
  59 #endif
  60 set_fast_math (void)
  61 {
  62 #ifndef __x86_64__
  63   unsigned int eax, ebx, ecx, edx;
  64
  65   if (!__get_cpuid (1, &eax, &ebx, &ecx, &edx))
  66     return;
  67
  68   if (edx & bit_SSE)
  69     {
  70       unsigned int mxcsr;
  71
  72 #if defined __sun__ && defined __svr4__
  73       /* Solaris 2 before Solaris 9 4/04 cannot execute SSE instructions even
  74          if the CPU supports them.  Programs receive SIGILL instead, so check
  75          for that at runtime.  */
  76       struct sigaction act, oact;
  77
  78       act.sa_handler = sigill_hdlr;
  79       sigemptyset (&act.sa_mask);
  80       /* Need to set SA_SIGINFO so a ucontext_t * is passed to the handler.  */
  81       act.sa_flags = SA_SIGINFO;
  82       sigaction (SIGILL, &act, &oact);
  83
  84       /* We need a single SSE instruction here so the handler can safely skip
  85          over it.  */
  86       __asm__ volatile ("movaps %xmm0,%xmm0");
  87
  88       sigaction (SIGILL, &oact, NULL);
  89
  90       if (sigill_caught)
  91         return;
  92 #endif /* __sun__ && __svr4__ */
  93
  94       if (edx & bit_FXSAVE)
  95         {
  96           /* Check if DAZ is available.  */
  97           struct
  98             {
  99               unsigned short cwd;
 100               unsigned short swd;
 101               unsigned short twd;
 102               unsigned short fop;
 103               unsigned int fip;
 104               unsigned int fcs;
 105               unsigned int foo;
 106               unsigned int fos;
 107               unsigned int mxcsr;
 108               unsigned int mxcsr_mask;
 109               unsigned int st_space[32];
 110               unsigned int xmm_space[32];
 111               unsigned int padding[56];
 112             } __attribute__ ((aligned (16))) fxsave;
 113
 114           /* This is necessary since some implementations of FXSAVE
 115              do not modify reserved areas within the image.  */
 116           fxsave.mxcsr_mask = 0;
 117
 118           __builtin_ia32_fxsave (&fxsave);
 119
 120           mxcsr = fxsave.mxcsr;
 121
 122           if (fxsave.mxcsr_mask & MXCSR_DAZ)
 123             mxcsr |= MXCSR_DAZ;
 124         }
 125       else
 126         mxcsr = __builtin_ia32_stmxcsr ();
 127
 128       mxcsr |= MXCSR_FTZ;
 129       __builtin_ia32_ldmxcsr (mxcsr);
 130     }
 131 #else
 132   unsigned int mxcsr = __builtin_ia32_stmxcsr ();
 133   mxcsr |= MXCSR_DAZ | MXCSR_FTZ;
 134   __builtin_ia32_ldmxcsr (mxcsr);
 135 #endif
 136 }