Disable X32 inline assembly (GH #686, PR #704)

Also use CRYPTOPP_DISABLE_XXX_ASM consistently. The pattern is needed for Clang which still can't compile Intel assembly language. Also see http://llvm.org/bugs/show_bug.cgi?id=24232.

1 files changed, 17 insertions, 67 deletions

diff --git a/vmac.cpp b/vmac.cpp
index c99b2692..0cfc1a87 100644
--- a/vmac.cpp
+++ b/vmac.cpp
@@ -20,26 +20,29 @@
 # pragma warning(disable: 4731)

 #endif

 

-NAMESPACE_BEGIN(CryptoPP)

+ANONYMOUS_NAMESPACE_BEGIN

 

 #if defined(_MSC_VER) && !CRYPTOPP_BOOL_SLOW_WORD64

 #include <intrin.h>

 #endif

 

 #if defined(CRYPTOPP_WORD128_AVAILABLE) && !defined(CRYPTOPP_X64_ASM_AVAILABLE)

+using CryptoPP::word128;

+using CryptoPP::word64;

 # define VMAC_BOOL_WORD128 1

 #else

+using CryptoPP::word64;

 # define VMAC_BOOL_WORD128 0

 #endif

 

 #ifdef __BORLANDC__

 #define const	// Turbo C++ 2006 workaround

 #endif

-static const word64 p64   = W64LIT(0xfffffffffffffeff);  /* 2^64 - 257 prime  */

-static const word64 m62   = W64LIT(0x3fffffffffffffff);  /* 62-bit mask       */

-static const word64 m63   = W64LIT(0x7fffffffffffffff);  /* 63-bit mask       */

-static const word64 m64   = W64LIT(0xffffffffffffffff);  /* 64-bit mask       */

-static const word64 mpoly = W64LIT(0x1fffffff1fffffff);  /* Poly key mask     */

+const word64 p64   = W64LIT(0xfffffffffffffeff);  /* 2^64 - 257 prime  */

+const word64 m62   = W64LIT(0x3fffffffffffffff);  /* 62-bit mask       */

+const word64 m63   = W64LIT(0x7fffffffffffffff);  /* 63-bit mask       */

+const word64 m64   = W64LIT(0xffffffffffffffff);  /* 64-bit mask       */

+const word64 mpoly = W64LIT(0x1fffffff1fffffff);  /* Poly key mask     */

 #ifdef __BORLANDC__

 #undef const

 #endif

@@ -48,10 +51,14 @@ static const word64 mpoly = W64LIT(0x1fffffff1fffffff);  /* Poly key mask     */
 // workaround GCC Bug 31690: ICE with const __uint128_t and C++ front-end

 #define m126				((word128(m62)<<64)|m64)

 #else

-static const word128 m126 = (word128(m62)<<64)|m64;		 /* 126-bit mask      */

+const word128 m126 = (word128(m62)<<64)|m64;		 /* 126-bit mask      */

 #endif

 #endif

 

+ANONYMOUS_NAMESPACE_END

+

+NAMESPACE_BEGIN(CryptoPP)

+

 void VMAC_Base::UncheckedSetKey(const byte *userKey, unsigned int keylength, const NameValuePairs &params)

 {

 	int digestLength = params.GetIntValueWithDefault(Name::DigestSize(), DefaultDigestSize());

@@ -167,7 +174,7 @@ unsigned int VMAC_Base::OptimalDataAlignment() const
 		GetCipher().OptimalDataAlignment();

 }

 

-#if CRYPTOPP_SSE2_ASM_AVAILABLE && (CRYPTOPP_BOOL_X86 || CRYPTOPP_BOOL_X32)

+#if CRYPTOPP_SSE2_ASM_AVAILABLE && (CRYPTOPP_BOOL_X86)

 #if CRYPTOPP_MSC_VERSION

 # pragma warning(disable: 4731)	// frame pointer register 'ebp' modified by inline assembly code

 #endif

@@ -217,13 +224,8 @@ VMAC_Base::VHASH_Update_SSE2(const word64 *data, size_t blocksRemainingInWord64,
 #endif

 

 	AS2(	shr		ebx, 3)

-#if CRYPTOPP_BOOL_X32

-	AS_PUSH_IF86(	bp)

-	AS2(	sub		esp, 24)

-#else

 	AS_PUSH_IF86(	bp)

 	AS2(	sub		esp, 12)

-#endif

 	ASL(4)

 	AS2(	mov		ebp, ebx)

 	AS2(	cmp		ecx, ebx)

@@ -246,11 +248,7 @@ VMAC_Base::VHASH_Update_SSE2(const word64 *data, size_t blocksRemainingInWord64,
 	AS2(	pxor	mm7, mm7)

 	AS2(	movd	[esp], mm6)

 	AS2(	psrlq	mm6, 32)

-#if CRYPTOPP_BOOL_X32

-	AS2(	movd	[esp+8], mm5)

-#else

 	AS2(	movd	[esp+4], mm5)

-#endif

 	AS2(	psrlq	mm5, 32)

 	AS2(	cmp		edi, ebp)

 	ASJ(	je,		1, f)

@@ -265,11 +263,7 @@ VMAC_Base::VHASH_Update_SSE2(const word64 *data, size_t blocksRemainingInWord64,
 	AS2(	paddq	mm5, mm2)

 	ASS(	pshufw	mm2, mm0, 1, 0, 3, 2)

 	AS2(	pmuludq	mm0, mm1)

-#if CRYPTOPP_BOOL_X32

-	AS2(	movd	[esp+16], mm3)

-#else

 	AS2(	movd	[esp+8], mm3)

-#endif

 	AS2(	psrlq	mm3, 32)

 	AS2(	paddq	mm5, mm3)

 	ASS(	pshufw	mm3, mm1, 1, 0, 3, 2)

@@ -278,48 +272,28 @@ VMAC_Base::VHASH_Update_SSE2(const word64 *data, size_t blocksRemainingInWord64,
 	AS2(	pmuludq	mm3, mm4)

 	AS2(	movd	mm4, [esp])

 	AS2(	paddq	mm7, mm4)

-#if CRYPTOPP_BOOL_X32

-	AS2(	movd	mm4, [esp+8])

-	AS2(	paddq	mm6, mm4)

-	AS2(	movd	mm4, [esp+16])

-#else

 	AS2(	movd	mm4, [esp+4])

 	AS2(	paddq	mm6, mm4)

 	AS2(	movd	mm4, [esp+8])

-#endif

 	AS2(	paddq	mm6, mm4)

 	AS2(	movd	[esp], mm0)

 	AS2(	psrlq	mm0, 32)

 	AS2(	paddq	mm6, mm0)

-#if CRYPTOPP_BOOL_X32

-	AS2(	movd	[esp+8], mm1)

-#else

 	AS2(	movd	[esp+4], mm1)

-#endif

 	AS2(	psrlq	mm1, 32)

 	AS2(	paddq	mm5, mm1)

 	AS2(	cmp		edi, ebp)

 	ASJ(	jne,	0, b)

 	ASL(1)

 	AS2(	paddq	mm5, mm2)

-#if CRYPTOPP_BOOL_X32

-	AS2(	movd	[esp+16], mm3)

-#else

 	AS2(	movd	[esp+8], mm3)

-#endif

 	AS2(	psrlq	mm3, 32)

 	AS2(	paddq	mm5, mm3)

 	AS2(	movd	mm4, [esp])

 	AS2(	paddq	mm7, mm4)

-#if CRYPTOPP_BOOL_X32

-	AS2(	movd	mm4, [esp+8])

-	AS2(	paddq	mm6, mm4)

-	AS2(	movd	mm4, [esp+16])

-#else

 	AS2(	movd	mm4, [esp+4])

 	AS2(	paddq	mm6, mm4)

 	AS2(	movd	mm4, [esp+8])

-#endif

 	AS2(	paddq	mm6, mm4)

 	AS2(	lea		ebp, [8*ebx])

 	AS2(	sub		edi, ebp)		// reset edi to start of nhK

@@ -327,11 +301,7 @@ VMAC_Base::VHASH_Update_SSE2(const word64 *data, size_t blocksRemainingInWord64,
 	AS2(	movd	[esp], mm7)

 	AS2(	psrlq	mm7, 32)

 	AS2(	paddq	mm6, mm7)

-#if CRYPTOPP_BOOL_X32

-	AS2(	movd	[esp+8], mm6)

-#else

 	AS2(	movd	[esp+4], mm6)

-#endif

 	AS2(	psrlq	mm6, 32)

 	AS2(	paddq	mm5, mm6)

 	AS2(	psllq	mm5, 2)

@@ -353,11 +323,7 @@ VMAC_Base::VHASH_Update_SSE2(const word64 *data, size_t blocksRemainingInWord64,
 	AS2(	movd	a0, mm0)

 	AS2(	psrlq	mm0, 32)

 	AS2(	movd	mm1, k1)

-#if CRYPTOPP_BOOL_X32

-	AS2(	movd	mm2, [esp+8])

-#else

 	AS2(	movd	mm2, [esp+4])

-#endif

 	AS2(	paddq	mm1, mm2)

 	AS2(	paddq	mm0, mm1)

 	AS2(	movd	a1, mm0)

@@ -395,11 +361,7 @@ VMAC_Base::VHASH_Update_SSE2(const word64 *data, size_t blocksRemainingInWord64,
 	AS2(	movq	mm3, mm2)

 	AS2(	pmuludq	mm2, k3)		// a0*k3

 	AS2(	pmuludq	mm3, mm7)		// a0*k0

-#if CRYPTOPP_BOOL_X32

-	AS2(	movd	[esp+16], mm0)

-#else

 	AS2(	movd	[esp+8], mm0)

-#endif

 	AS2(	psrlq	mm0, 32)

 	AS2(	pmuludq	mm7, mm5)		// a1*k0

 	AS2(	pmuludq	mm5, k3)		// a1*k3

@@ -422,22 +384,14 @@ VMAC_Base::VHASH_Update_SSE2(const word64 *data, size_t blocksRemainingInWord64,
 	AS2(	movd	mm1, a3)

 	AS2(	pmuludq	mm1, k2)		// a3*k2

 	AS2(	paddq	mm5, mm2)

-#if CRYPTOPP_BOOL_X32

-	AS2(	movd	mm2, [esp+8])

-#else

 	AS2(	movd	mm2, [esp+4])

-#endif

 	AS2(	psllq	mm5, 1)

 	AS2(	paddq	mm0, mm5)

 	AS2(	psllq	mm4, 33)

 	AS2(	movd	a0, mm0)

 	AS2(	psrlq	mm0, 32)

 	AS2(	paddq	mm6, mm7)

-#if CRYPTOPP_BOOL_X32

-	AS2(	movd	mm7, [esp+16])

-#else

 	AS2(	movd	mm7, [esp+8])

-#endif

 	AS2(	paddq	mm0, mm6)

 	AS2(	paddq	mm0, mm2)

 	AS2(	paddq	mm3, mm1)

@@ -461,11 +415,7 @@ VMAC_Base::VHASH_Update_SSE2(const word64 *data, size_t blocksRemainingInWord64,
 	ASL(3)

 	AS2(	test	ecx, ecx)

 	ASJ(	jnz,	4, b)

-#if CRYPTOPP_BOOL_X32

-	AS2(	add		esp, 24)

-#else

 	AS2(	add		esp, 12)

-#endif

 	AS_POP_IF86(	bp)

 	AS1(	emms)

 #ifdef __GNUC__

@@ -800,7 +750,7 @@ void VMAC_Base::VHASH_Update_Template(const word64 *data, size_t blocksRemaining
 

 inline void VMAC_Base::VHASH_Update(const word64 *data, size_t blocksRemainingInWord64)

 {

-#if CRYPTOPP_SSE2_ASM_AVAILABLE && (CRYPTOPP_BOOL_X86 || CRYPTOPP_BOOL_X32)

+#if CRYPTOPP_SSE2_ASM_AVAILABLE && (CRYPTOPP_BOOL_X86)

 	if (HasSSE2())

 	{

 		VHASH_Update_SSE2(data, blocksRemainingInWord64, 0);

@@ -825,7 +775,7 @@ size_t VMAC_Base::HashMultipleBlocks(const word64 *data, size_t length)
 	return remaining;

 }

 

-static word64 L3Hash(const word64 *input, const word64 *l3Key, size_t len)

+word64 L3Hash(const word64 *input, const word64 *l3Key, size_t len)

 {

     word64 rh, rl, t, z=0;

 	word64 p1 = input[0], p2 = input[1];