Miroslav's patch to add run-time detection of operating system support for SSE/SSE2 instructions

src/libFLAC/cpu.c

@@ -21,6 +21,12 @@
 #include<stdlib.h>
 #include<stdio.h>
 
+#if !defined(FLAC__NO_ASM) && defined(FLAC__CPU_IA32) && defined(FLAC__HAS_NASM) && !defined(FLAC__SSE_OS) && !defined(NO_VFORK)
+#include <unistd.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#endif
+
 const unsigned FLAC__CPUINFO_IA32_CPUID_CMOV = 0x00008000;
 const unsigned FLAC__CPUINFO_IA32_CPUID_MMX = 0x00800000;
 const unsigned FLAC__CPUINFO_IA32_CPUID_FXSR = 0x01000000;
@@ -44,17 +50,36 @@ void FLAC__cpu_info(FLAC__CPUInfo *info)
 		info->data.ia32.mmx = (cpuid & FLAC__CPUINFO_IA32_CPUID_MMX)? true : false;
 		info->data.ia32.fxsr = (cpuid & FLAC__CPUINFO_IA32_CPUID_FXSR)? true : false;
 		info->data.ia32.sse = (cpuid & FLAC__CPUINFO_IA32_CPUID_SSE)? true : false;
-		info->data.ia32.sse2 = (cpuid & FLAC__CPUINFO_IA32_CPUID_SSE2)? true : false; /* @@@ also need to check for operating system support? */
+		info->data.ia32.sse2 = (cpuid & FLAC__CPUINFO_IA32_CPUID_SSE2)? true : false;
+
+#ifndef FLAC__SSE_OS
+#ifndef NO_VFORK
+		if(info->data.ia32.sse == true || info->data.ia32.sse2 == true) {
+			int pid, status, sse;
+			pid = vfork();
+			if(!pid) {
+				FLAC__cpu_info_sse_test_asm_ia32();
+				exit(0);
+			}
+			sse = 0;
+			if(pid > 0) {
+				waitpid(pid, &status, 0);
+				if(WIFEXITED(status) && WEXITSTATUS(status) == 0)
+					sse = 1;        /* there was normal exit, no SIGILL */
+			}
+			if(!sse)
+				info->data.ia32.sse = info->data.ia32.sse2 = false;
+		}
+#else
+		/* we are assuming OS isn't supporting SSE */
+		info->data.ia32.sse = info->data.ia32.sse2 = false;
+#endif
+#endif
 
 		cpuid = FLAC__cpu_info_extended_amd_asm_ia32();
 		info->data.ia32._3dnow = (cpuid & FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_3DNOW)? true : false;
 		info->data.ia32.ext3dnow = (cpuid & FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_EXT3DNOW)? true : false;
 		info->data.ia32.extmmx = (cpuid & FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_EXTMMX)? true : false;
-
-#ifndef FLAC__SSE_OS
-		if(!FLAC__cpu_info_sse_os_asm_ia32()) /* this function currently always returns false */
-			info->data.ia32.fxsr = info->data.ia32.sse = info->data.ia32.sse2 = false;
-#endif
 	}
 #else
 	info->use_asm = false;

src/libFLAC/ia32/cpu_asm.nasm

@@ -22,7 +22,7 @@
 
 cglobal FLAC__cpu_info_asm_ia32
 cglobal FLAC__cpu_info_extended_amd_asm_ia32
-cglobal FLAC__cpu_info_sse_os_asm_ia32
+cglobal FLAC__cpu_info_sse_test_asm_ia32
 
 	code_section
 
@@ -81,20 +81,8 @@ cident FLAC__cpu_info_extended_amd_asm_ia32
 	pop	ebx
 	ret
 
-;WATCHOUT - DO NOT call this function until you have verified CPU support of
-;           SSE by inspecting the return value from FLAC__cpu_info_asm_ia32
-;NOTE - Since we're not in priv level 0 we can't just check CR4 bits 9 & 10,
-;       so right now we just assume there is no OS support.  If you know
-;       how to write code to trap a #UD exception in nasm so we can implement
-;       this function correctly, let us know!
-cident FLAC__cpu_info_sse_os_asm_ia32
-	push	ebx
-	mov	eax, 1
-	cpuid
-	mov	eax, 0		;we would like to 'move eax, cr4'
-	shr	eax, 9
-	and	eax, 3
+cident FLAC__cpu_info_sse_test_asm_ia32
+	xorps	xmm0, xmm0
 	pop	ebx
-	ret
 
 end

src/libFLAC/include/private/cpu.h

@@ -63,7 +63,7 @@ void FLAC__cpu_info(FLAC__CPUInfo *info);
 #ifdef FLAC__HAS_NASM
 unsigned FLAC__cpu_info_asm_ia32();
 unsigned FLAC__cpu_info_extended_amd_asm_ia32();
-unsigned FLAC__cpu_info_sse_os_asm_ia32();
+unsigned FLAC__cpu_info_sse_test_asm_ia32();
 #endif
 #endif
 #endif