core/fpu_ctrl.cpp

   1
   2 #include "config.h"
   3 #include "config_simd.h"
   4
   5 #include "fpu_ctrl.h"
   6
   7 #ifdef HAVE_INTRIN_H
   8 #include <intrin.h>
   9 #endif
  10 #if HAVE_SSE_INTRINSICS
  11 #include <emmintrin.h>
  12 #elif HAVE_SSE
  13 #include <xmmintrin.h>
  14 #endif
  15
  16 #if HAVE_SSE && !defined(_MM_DENORMALS_ZERO_MASK)
  17 /* Some headers seem to be missing these? */
  18 #define _MM_DENORMALS_ZERO_MASK 0x0040u
  19 #define _MM_DENORMALS_ZERO_ON 0x0040u
  20 #endif
  21
  22 #if !HAVE_SSE_INTRINSICS && HAVE_SSE
  23 #include "cpu_caps.h"
  24 #endif
  25
  26 namespace {
  27
  28 #if defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))
  29 [[gnu::target("sse")]]
  30 #endif
  31 [[maybe_unused]]
  32 void disable_denormals(unsigned int *state [[maybe_unused]])
  33 {
  34 #if HAVE_SSE_INTRINSICS
  35     *state = _mm_getcsr();
  36     unsigned int sseState{*state};
  37     sseState &= ~(_MM_FLUSH_ZERO_MASK | _MM_DENORMALS_ZERO_MASK);
  38     sseState |= _MM_FLUSH_ZERO_ON | _MM_DENORMALS_ZERO_ON;
  39     _mm_setcsr(sseState);
  40
  41 #elif HAVE_SSE
  42
  43     *state = _mm_getcsr();
  44     unsigned int sseState{*state};
  45     sseState &= ~_MM_FLUSH_ZERO_MASK;
  46     sseState |= _MM_FLUSH_ZERO_ON;
  47     if((CPUCapFlags&CPU_CAP_SSE2))
  48     {
  49         sseState &= ~_MM_DENORMALS_ZERO_MASK;
  50         sseState |= _MM_DENORMALS_ZERO_ON;
  51     }
  52     _mm_setcsr(sseState);
  53 #endif
  54 }
  55
  56 #if defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))
  57 [[gnu::target("sse")]]
  58 #endif
  59 [[maybe_unused]]
  60 void reset_fpu(unsigned int state [[maybe_unused]])
  61 {
  62 #if HAVE_SSE_INTRINSICS || HAVE_SSE
  63     _mm_setcsr(state);
  64 #endif
  65 }
  66
  67 } // namespace
  68
  69
  70 unsigned int FPUCtl::Set() noexcept
  71 {
  72     unsigned int state{};
  73 #if HAVE_SSE_INTRINSICS
  74     disable_denormals(&state);
  75 #elif HAVE_SSE
  76     if((CPUCapFlags&CPU_CAP_SSE))
  77         disable_denormals(&state);
  78 #endif
  79     return state;
  80 }
  81
  82 void FPUCtl::Reset(unsigned int state [[maybe_unused]]) noexcept
  83 {
  84 #if HAVE_SSE_INTRINSICS
  85     reset_fpu(state);
  86 #elif HAVE_SSE
  87     if((CPUCapFlags&CPU_CAP_SSE))
  88         reset_fpu(state);
  89 #endif
  90 }