X-Git-Url: https://main.carlh.net/gitweb/?a=blobdiff_plain;ds=sidebyside;f=libs%2Fpbd%2Ffpu.cc;h=78bcd3d8121f0f71e79cd3ebcc700124ae7f8eb9;hb=659701c59b76bdcbfcf5c1b8ce8962eb3f6a1ca2;hp=9838b4565beb7b354e2d8b18ff054331983a66b8;hpb=22b07e0233a29d9633ffa825a79503befaf2e16e;p=ardour.git

diff --git a/libs/pbd/fpu.cc b/libs/pbd/fpu.cc
index 9838b4565b..78bcd3d812 100644
--- a/libs/pbd/fpu.cc
+++ b/libs/pbd/fpu.cc
@@ -33,7 +33,7 @@
 #include "pbd/fpu.h"
 #include "pbd/error.h"
 
-#include "i18n.h"
+#include "pbd/i18n.h"
 
 using namespace PBD;
 using namespace std;
@@ -48,54 +48,41 @@ FPU* FPU::_instance (0);
 static void
 __cpuid(int regs[4], int cpuid_leaf)
 {
-        int eax, ebx, ecx, edx;
         asm volatile (
 #if defined(__i386__)
 	        "pushl %%ebx;\n\t"
 #endif
-	        "movl %4, %%eax;\n\t"
 	        "cpuid;\n\t"
-	        "movl %%eax, %0;\n\t"
-	        "movl %%ebx, %1;\n\t"
-	        "movl %%ecx, %2;\n\t"
-	        "movl %%edx, %3;\n\t"
+	        "movl %%eax, (%1);\n\t"
+	        "movl %%ebx, 4(%1);\n\t"
+	        "movl %%ecx, 8(%1);\n\t"
+	        "movl %%edx, 12(%1);\n\t"
 #if defined(__i386__)
 	        "popl %%ebx;\n\t"
 #endif
-	        :"=m" (eax), "=m" (ebx), "=m" (ecx), "=m" (edx)
-	        :"r" (cpuid_leaf)
-	        :"%eax",
+	        :"=a" (cpuid_leaf) /* %eax clobbered by CPUID */
+	        :"S" (regs), "a" (cpuid_leaf)
+	        :
 #if !defined(__i386__)
 	         "%ebx",
 #endif
-	         "%ecx", "%edx");
-
-        regs[0] = eax;
-        regs[1] = ebx;
-        regs[2] = ecx;
-        regs[3] = edx;
+	         "%ecx", "%edx", "memory");
 }
 
 #endif /* !PLATFORM_WINDOWS */
 
-#ifndef COMPILER_MSVC
-
-static uint64_t
-_xgetbv (uint32_t xcr)
-{
-#ifdef __APPLE__
-        /* it would be nice to make this work on OS X but as long we use veclib,
-           we don't really need to know about SSE/AVX on that platform.
-        */
-        return 0;
-#else
-	uint32_t eax, edx;
-	__asm__ volatile ("xgetbv" : "=a" (eax), "=d" (edx) : "c" (xcr));
-	return (static_cast<uint64_t>(edx) << 32) | eax;
+#ifndef HAVE_XGETBV // Allow definition by build system
+	#if defined(__MINGW32__) && defined(__MINGW64_VERSION_MAJOR) && __MINGW64_VERSION_MAJOR >= 5
+		#define HAVE_XGETBV
+	#elif defined(_MSC_VER) && _MSC_VER >= 1600
+		// '_xgetbv()' was only available from VC10 onwards
+		#define HAVE_XGETBV
+	#endif
 #endif
-}
 
-#elif _MSC_VER < 1600
+#ifndef HAVE_XGETBV
+
+#ifdef COMPILER_MSVC
 
 // '_xgetbv()' was only available from VC10 onwards
 __declspec(noinline) static uint64_t
@@ -108,11 +95,29 @@ _xgetbv (uint32_t xcr)
 	// to place this function into its own (unoptimized) source file.
 	__asm {
 			 mov ecx, [xcr]
-			 __asm _emit 0x0f __asm _emit 0x01 __asm _emit 0xd0   /*xgetbv*/
-		  }
+			 __asm _emit 0x0f __asm _emit 0x01 __asm _emit 0xd0 /*xgetbv*/
+	}
+}
+
+#else
+
+static uint64_t
+_xgetbv (uint32_t xcr)
+{
+#ifdef __APPLE__
+	/* it would be nice to make this work on OS X but as long we use veclib,
+	   we don't really need to know about SSE/AVX on that platform.
+	*/
+	return 0;
+#else
+	uint32_t eax, edx;
+	__asm__ volatile ("xgetbv" : "=a" (eax), "=d" (edx) : "c" (xcr));
+	return (static_cast<uint64_t>(edx) << 32) | eax;
+#endif
 }
 
 #endif /* !COMPILER_MSVC */
+#endif /* !HAVE_XGETBV */
 #endif /* ARCH_X86 */
 
 #ifndef _XCR_XFEATURE_ENABLED_MASK
@@ -143,6 +148,11 @@ FPU::FPU ()
 		error << _("FPU object instantiated more than once") << endmsg;
 	}
 
+	if (getenv("ARDOUR_FPU_FLAGS")) {
+		_flags = Flags (atoi (getenv("ARDOUR_FPU_FLAGS")));
+		return;
+	}
+
 #if !( (defined __x86_64__) || (defined __i386__) || (defined _M_X64) || (defined _M_IX86) ) // !ARCH_X86
 	/* Non-Intel architecture, nothing to do here */
 	return;