Avoid inline assembly for getting/setting SSE state

1 files changed, 49 insertions, 20 deletions

diff --git a/core/fpu_ctrl.cpp b/core/fpu_ctrl.cpp
index 0cf0d6e7..701f517d 100644
--- a/core/fpu_ctrl.cpp
+++ b/core/fpu_ctrl.cpp
@@ -8,38 +8,69 @@
 #endif
 #ifdef HAVE_SSE_INTRINSICS
 #include <emmintrin.h>
-#ifndef _MM_DENORMALS_ZERO_MASK
+#elif defined(HAVE_SSE)
+#include <xmmintrin.h>
+#endif
+
+#if defined(HAVE_SSE) && !defined(_MM_DENORMALS_ZERO_MASK)
 /* Some headers seem to be missing these? */
 #define _MM_DENORMALS_ZERO_MASK 0x0040u
 #define _MM_DENORMALS_ZERO_ON 0x0040u
 #endif
-#endif
 
 #include "cpu_caps.h"
 
+namespace {
 
-void FPUCtl::enter() noexcept
+#if defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))
+[[gnu::target("sse")]]
+#endif
+void disable_denormals(unsigned int *state [[maybe_unused]])
 {
-    if(this->in_mode) return;
-
 #if defined(HAVE_SSE_INTRINSICS)
-    this->sse_state = _mm_getcsr();
-    unsigned int sseState{this->sse_state};
+    *state = _mm_getcsr();
+    unsigned int sseState{*state};
     sseState &= ~(_MM_FLUSH_ZERO_MASK | _MM_DENORMALS_ZERO_MASK);
     sseState |= _MM_FLUSH_ZERO_ON | _MM_DENORMALS_ZERO_ON;
     _mm_setcsr(sseState);
 
-#elif defined(__GNUC__) && defined(HAVE_SSE)
+#elif defined(HAVE_SSE)
 
-    if((CPUCapFlags&CPU_CAP_SSE))
+    *state = _mm_getcsr();
+    unsigned int sseState{*state};
+    sseState &= ~_MM_FLUSH_ZERO_MASK;
+    sseState |= _MM_FLUSH_ZERO_ON;
+    if((CPUCapFlags&CPU_CAP_SSE2))
     {
-        __asm__ __volatile__("stmxcsr %0" : "=m" (*&this->sse_state));
-        unsigned int sseState{this->sse_state};
-        sseState |= 0x8000; /* set flush-to-zero */
-        if((CPUCapFlags&CPU_CAP_SSE2))
-            sseState |= 0x0040; /* set denormals-are-zero */
-        __asm__ __volatile__("ldmxcsr %0" : : "m" (*&sseState));
+        sseState &= ~_MM_DENORMALS_ZERO_MASK;
+        sseState |= _MM_DENORMALS_ZERO_ON;
     }
+    _mm_setcsr(sseState);
+#endif
+}
+
+#if defined(__GNUC__) && (defined(__i386__) || defined(__x86_64__))
+[[gnu::target("sse")]]
+#endif
+void reset_fpu(unsigned int state [[maybe_unused]])
+{
+#if defined(HAVE_SSE_INTRINSICS) || defined(HAVE_SSE)
+    _mm_setcsr(state);
+#endif
+}
+
+} // namespace
+
+
+void FPUCtl::enter() noexcept
+{
+    if(this->in_mode) return;
+
+#if defined(HAVE_SSE_INTRINSICS)
+    disable_denormals(&this->sse_state);
+#elif defined(HAVE_SSE)
+    if((CPUCapFlags&CPU_CAP_SSE))
+        disable_denormals(&this->sse_state);
 #endif
 
     this->in_mode = true;
@@ -50,12 +81,10 @@ void FPUCtl::leave() noexcept
     if(!this->in_mode) return;
 
 #if defined(HAVE_SSE_INTRINSICS)
-    _mm_setcsr(this->sse_state);
-
-#elif defined(__GNUC__) && defined(HAVE_SSE)
-
+    reset_fpu(this->sse_state);
+#elif defined(HAVE_SSE)
     if((CPUCapFlags&CPU_CAP_SSE))
-        __asm__ __volatile__("ldmxcsr %0" : : "m" (*&this->sse_state));
+        reset_fpu(this->sse_state);
 #endif
     this->in_mode = false;
 }