1 files changed, 58 insertions, 25 deletions
diff --git a/src/Crypto/cpu.c b/src/Crypto/cpu.c
index 6896240c..766edea8 100644..100755
--- a/src/Crypto/cpu.c
+++ b/src/Crypto/cpu.c
@@ -3,6 +3,10 @@
 #include "cpu.h"
 #include "misc.h"
 
+#ifndef EXCEPTION_EXECUTE_HANDLER
+#define EXCEPTION_EXECUTE_HANDLER 1
+#endif
+
 #ifndef CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY
 #include <signal.h>
 #include <setjmp.h>
@@ -16,7 +20,7 @@
 
 #if _MSC_VER >= 1400 && CRYPTOPP_BOOL_X64
 
-int CpuId(uint32 input, uint32 *output)
+int CpuId(uint32 input, uint32 output[4])
 {
 	__cpuid((int *)output, input);
 	return 1;
@@ -45,12 +49,13 @@ static void SigIllHandlerSSE2(int p)
 	longjmp(s_jmpNoSSE2, 1);
 }
 #endif
+
 #if defined(__cplusplus)
 }
 #endif
 #endif
 
-int CpuId(uint32 input, uint32 *output)
+int CpuId(uint32 input, uint32 output[4])
 {
 #ifdef CRYPTOPP_MS_STYLE_INLINE_ASSEMBLY
     __try
@@ -58,6 +63,7 @@ int CpuId(uint32 input, uint32 *output)
 		__asm
 		{
 			mov eax, input
+            mov ecx, 0
 			cpuid
 			mov edi, output
 			mov [edi], eax
@@ -66,32 +72,42 @@ int CpuId(uint32 input, uint32 *output)
 			mov [edi+12], edx
 		}
 	}
-    __except (1)
+    __except (EXCEPTION_EXECUTE_HANDLER)
 	{
 		return 0;
     }
+    
+	// function 0 returns the highest basic function understood in EAX
+	if(input == 0)
+        return !!output[0]? 1 : 0;
+
 	return 1;
 #else
+	// longjmp and clobber warnings. Volatile is required.
+	// http://github.com/weidai11/cryptopp/issues/24
+	// http://stackoverflow.com/q/7721854
+	volatile int result = 1;
+
 	SigHandler oldHandler = signal(SIGILL, SigIllHandlerCPUID);
 	if (oldHandler == SIG_ERR)
-		return 0;
+		result = 0;
 
-	int result = 1;
 	if (setjmp(s_jmpNoCPUID))
 		result = 0;
 	else
 	{
-		asm
+		asm volatile
 		(
-			// save ebx in case -fPIC is being used
-#if CRYPTOPP_BOOL_X86
-			"push %%ebx; cpuid; mov %%ebx, %%edi; pop %%ebx"
+            // save ebx in case -fPIC is being used
+            // TODO: this might need an early clobber on EDI.
+#if CRYPTOPP_BOOL_X32 || CRYPTOPP_BOOL_X64
+            "pushq %%rbx; cpuid; mov %%ebx, %%edi; popq %%rbx"
 #else
-			"pushq %%rbx; cpuid; mov %%ebx, %%edi; popq %%rbx"
+            "push %%ebx; cpuid; mov %%ebx, %%edi; pop %%ebx"
 #endif
-			: "=a" (output[0]), "=D" (output[1]), "=c" (output[2]), "=d" (output[3])
-			: "a" (input)
-		);
+            : "=a" (output[0]), "=D" (output[1]), "=c" (output[2]), "=d" (output[3])
+            : "a" (input), "c" (0)
+         );
 	}
 
 	signal(SIGILL, oldHandler);
@@ -112,29 +128,33 @@ static int TrySSE2()
         AS2(por xmm0, xmm0)        // executing SSE2 instruction
 #elif CRYPTOPP_BOOL_SSE2_INTRINSICS_AVAILABLE
 		__m128i x = _mm_setzero_si128();
-		return _mm_cvtsi128_si32(x) == 0;
+		return _mm_cvtsi128_si32(x) == 0 ? 1 : 0;
 #endif
 	}
-    __except (1)
+    __except (EXCEPTION_EXECUTE_HANDLER)
 	{
 		return 0;
     }
 	return 1;
 #else
+	// longjmp and clobber warnings. Volatile is required.
+	// http://github.com/weidai11/cryptopp/issues/24
+	// http://stackoverflow.com/q/7721854
+	volatile int result = 1;
+
 	SigHandler oldHandler = signal(SIGILL, SigIllHandlerSSE2);
 	if (oldHandler == SIG_ERR)
 		return 0;
 
-	int result = 1;
 	if (setjmp(s_jmpNoSSE2))
-		result = 0;
+		result = 1;
 	else
 	{
 #if CRYPTOPP_BOOL_SSE2_ASM_AVAILABLE
 		__asm __volatile ("por %xmm0, %xmm0");
 #elif CRYPTOPP_BOOL_SSE2_INTRINSICS_AVAILABLE
 		__m128i x = _mm_setzero_si128();
-		result = _mm_cvtsi128_si32(x) == 0;
+		result = _mm_cvtsi128_si32(x) == 0? 1 : 0;
 #endif
 	}
 
@@ -147,9 +167,25 @@ int g_x86DetectionDone = 0;
 int g_hasISSE = 0, g_hasSSE2 = 0, g_hasSSSE3 = 0, g_hasMMX = 0, g_hasAESNI = 0, g_hasCLMUL = 0, g_isP4 = 0;
 uint32 g_cacheLineSize = CRYPTOPP_L1_CACHE_LINE_SIZE;
 
+static inline int IsIntel(const uint32 output[4])
+{
+	// This is the "GenuineIntel" string
+	return (output[1] /*EBX*/ == 0x756e6547) &&
+    (output[2] /*ECX*/ == 0x6c65746e) &&
+    (output[3] /*EDX*/ == 0x49656e69);
+}
+
+static inline int IsAMD(const uint32 output[4])
+{
+	// This is the "AuthenticAMD" string
+	return (output[1] /*EBX*/ == 0x68747541) &&
+    (output[2] /*ECX*/ == 0x69746E65) &&
+    (output[3] /*EDX*/ == 0x444D4163);
+}
+
 void DetectX86Features()
 {
-	uint32 cpuid[4], cpuid1[4], tmp;
+	uint32 cpuid[4], cpuid1[4];
 	if (!CpuId(0, cpuid))
 		return;
 	if (!CpuId(1, cpuid1))
@@ -175,15 +211,12 @@ void DetectX86Features()
 		}
 	}
 
-	tmp = cpuid[2];
-	cpuid[2] = cpuid[3];
-	cpuid[3] = tmp;
-	if (memcmp(cpuid+1, "GenuineIntel", 12) == 0)
+	if (IsIntel(cpuid))
 	{
 		g_isP4 = ((cpuid1[0] >> 8) & 0xf) == 0xf;
 		g_cacheLineSize = 8 * GETBYTE(cpuid1[1], 1);
 	}
-	else if (memcmp(cpuid+1, "AuthenticAMD", 12) == 0)
+	else if (IsAMD(cpuid))
 	{
 		CpuId(0x80000005, cpuid);
 		g_cacheLineSize = GETBYTE(cpuid[2], 0);
@@ -192,7 +225,7 @@ void DetectX86Features()
 	if (!g_cacheLineSize)
 		g_cacheLineSize = CRYPTOPP_L1_CACHE_LINE_SIZE;
 
-	g_x86DetectionDone = 1;
+	*((volatile int*)&g_x86DetectionDone) = 1;
 }
 
 #endif