eliminate hardcoded registers

aes.cpp

@@ -26,7 +26,7 @@ namespace crypto
 		"pslldq	$4, %%xmm4 \n" \
 		"pxor %%xmm4, %%xmm1 \n" \
 		"pxor %%xmm2, %%xmm1 \n" \
-		"movups	%%xmm1, (%%rcx) \n" \
+		"movaps	%%xmm1, (%%rcx) \n" \
 		"aeskeygenassist $0, %%xmm1, %%xmm4 \n" \
 		"pshufd	$0xaa, %%xmm4, %%xmm2 \n" \
 		"movaps	%%xmm3, %%xmm4 \n" \
@@ -37,7 +37,7 @@ namespace crypto
 		"pslldq	$4, %%xmm4 \n" \
 		"pxor %%xmm4, %%xmm3 \n" \
 		"pxor %%xmm2, %%xmm3 \n" \
-		"movups	%%xmm3, 16(%%rcx) \n" \
+		"movaps	%%xmm3, 16(%%rcx) \n" \
 		"add $32, %%rcx \n"	
 		
 
@@ -45,11 +45,11 @@ namespace crypto
 	{
 		__asm__
 		(
-			"movups (%%rsi), %%xmm1 \n"
-			"movups 16(%%rsi), %%xmm3 \n"
-			"movups %%xmm1, (%%rdi) \n"
-			"movups %%xmm3, 16(%%rdi) \n"
-			"lea 32(%%rdi), %%rcx \n"
+			"movups (%[key]), %%xmm1 \n"
+			"movups 16(%[key]), %%xmm3 \n"
+			"movaps %%xmm1, (%[shed]) \n"
+			"movaps %%xmm3, 16(%[shed]) \n"
+			"lea 32(%[shed]), %%rcx \n"
 			"aeskeygenassist $1, %%xmm3, %%xmm2 \n"
 			KeyExpansion256
 			"aeskeygenassist $2, %%xmm3, %%xmm2 \n"
@@ -75,8 +75,8 @@ namespace crypto
 			"pxor %%xmm2, %%xmm1 \n"
 			"movups	%%xmm1, (%%rcx) \n"
 			: // output
-			: "S" (key), "D" (m_KeySchedule) // input
-			: "%rcx" // clogged
+			: [key]"r"(key), [shed]"r"(m_KeySchedule) // input
+			: "%rcx", "%xmm1", "%xmm2", "%xmm3", "%xmm4" // clogged
 		);
 	}
 
@@ -84,24 +84,24 @@ namespace crypto
 	{
 		__asm__
 		(
-			"movups	(%%rsi), %%xmm0 \n"
-			"pxor (%%rdx), %%xmm0 \n"
-			"aesenc	16(%%rdx), %%xmm0 \n"
-			"aesenc	32(%%rdx), %%xmm0 \n"
-			"aesenc	48(%%rdx), %%xmm0 \n"
-			"aesenc	64(%%rdx), %%xmm0 \n"
-			"aesenc	80(%%rdx), %%xmm0 \n"
-			"aesenc	96(%%rdx), %%xmm0 \n"
-			"aesenc	112(%%rdx), %%xmm0 \n"
-			"aesenc	128(%%rdx), %%xmm0 \n"
-			"aesenc	144(%%rdx), %%xmm0 \n"
-			"aesenc	160(%%rdx), %%xmm0 \n"
-			"aesenc	176(%%rdx), %%xmm0 \n"
-			"aesenc	192(%%rdx), %%xmm0 \n"
-			"aesenc	208(%%rdx), %%xmm0 \n"
-			"aesenclast	224(%%rdx), %%xmm0 \n"
-			"movups	%%xmm0, (%%rdi) \n"	
-			: : "d" (m_KeySchedule), "S" (in), "D" (out)
+			"movups	(%[in]), %%xmm0 \n"
+			"pxor (%[shed]), %%xmm0 \n"
+			"aesenc	16(%[shed]), %%xmm0 \n"
+			"aesenc	32(%[shed]), %%xmm0 \n"
+			"aesenc	48(%[shed]), %%xmm0 \n"
+			"aesenc	64(%[shed]), %%xmm0 \n"
+			"aesenc	80(%[shed]), %%xmm0 \n"
+			"aesenc	96(%[shed]), %%xmm0 \n"
+			"aesenc	112(%[shed]), %%xmm0 \n"
+			"aesenc	128(%[shed]), %%xmm0 \n"
+			"aesenc	144(%[shed]), %%xmm0 \n"
+			"aesenc	160(%[shed]), %%xmm0 \n"
+			"aesenc	176(%[shed]), %%xmm0 \n"
+			"aesenc	192(%[shed]), %%xmm0 \n"
+			"aesenc	208(%[shed]), %%xmm0 \n"
+			"aesenclast	224(%[shed]), %%xmm0 \n"
+			"movups	%%xmm0, (%[out]) \n"	
+			: : [shed]"r"(m_KeySchedule), [in]"r"(in), [out]"r"(out) : "%xmm0"
 		);
 	}		
 
@@ -109,31 +109,31 @@ namespace crypto
 	{
 		__asm__
 		(
-			"movups	(%%rsi), %%xmm0 \n"
-			"pxor 224(%%rdx), %%xmm0 \n"
-			"aesdec	208(%%rdx), %%xmm0 \n"
-			"aesdec	192(%%rdx), %%xmm0 \n"
-			"aesdec	176(%%rdx), %%xmm0 \n"
-			"aesdec	160(%%rdx), %%xmm0 \n"
-			"aesdec	144(%%rdx), %%xmm0 \n"
-			"aesdec	128(%%rdx), %%xmm0 \n"
-			"aesdec	112(%%rdx), %%xmm0 \n"
-			"aesdec	96(%%rdx), %%xmm0 \n"
-			"aesdec	80(%%rdx), %%xmm0 \n"
-			"aesdec	64(%%rdx), %%xmm0 \n"
-			"aesdec	48(%%rdx), %%xmm0 \n"
-			"aesdec	32(%%rdx), %%xmm0 \n"
-			"aesdec	16(%%rdx), %%xmm0 \n"
-			"aesdeclast (%%rdx), %%xmm0 \n"
-			"movups	%%xmm0, (%%rdi) \n"	
-			: : "d" (m_KeySchedule), "S" (in), "D" (out)
+			"movups	(%[in]), %%xmm0 \n"
+			"pxor 224(%[shed]), %%xmm0 \n"
+			"aesdec	208(%[shed]), %%xmm0 \n"
+			"aesdec	192(%[shed]), %%xmm0 \n"
+			"aesdec	176(%[shed]), %%xmm0 \n"
+			"aesdec	160(%[shed]), %%xmm0 \n"
+			"aesdec	144(%[shed]), %%xmm0 \n"
+			"aesdec	128(%[shed]), %%xmm0 \n"
+			"aesdec	112(%[shed]), %%xmm0 \n"
+			"aesdec	96(%[shed]), %%xmm0 \n"
+			"aesdec	80(%[shed]), %%xmm0 \n"
+			"aesdec	64(%[shed]), %%xmm0 \n"
+			"aesdec	48(%[shed]), %%xmm0 \n"
+			"aesdec	32(%[shed]), %%xmm0 \n"
+			"aesdec	16(%[shed]), %%xmm0 \n"
+			"aesdeclast (%[shed]), %%xmm0 \n"
+			"movups	%%xmm0, (%[out]) \n"	
+			: : [shed]"r"(m_KeySchedule), [in]"r"(in), [out]"r"(out) : "%xmm0"
 		);		
 	}
 
 	#define CallAESIMC(offset) \
-		"movups "#offset"(%%rdx), %%xmm0 \n"	\
+		"movaps "#offset"(%[shed]), %%xmm0 \n"	\
 		"aesimc %%xmm0, %%xmm0 \n" \
-		"movups %%xmm0, "#offset"(%%rdx) \n" 
+		"movaps %%xmm0, "#offset"(%[shed]) \n" 
 
 	void ECBDecryptionAESNI::SetKey (const uint8_t * key)
 	{
@@ -154,7 +154,7 @@ namespace crypto
 			CallAESIMC(176)
 			CallAESIMC(192)
 			CallAESIMC(208)
-			: : "d" (m_KeySchedule)
+			: : [shed]"r"(m_KeySchedule) : "%xmm0"
 		);
 	}