ghash-x86_64.pl: add AVX code path.

1 files changed, 45 insertions, 3 deletions

diff --git a/crypto/modes/gcm128.c b/crypto/modes/gcm128.c
index a6e2af1b96..ae5fab1b46 100644
--- a/crypto/modes/gcm128.c
+++ b/crypto/modes/gcm128.c
@@ -658,6 +658,16 @@ void gcm_init_clmul(u128 Htable[16],const u64 Xi[2]);
 void gcm_gmult_clmul(u64 Xi[2],const u128 Htable[16]);
 void gcm_ghash_clmul(u64 Xi[2],const u128 Htable[16],const u8 *inp,size_t len);
 
+#if defined(__i386) || defined(__i386__)
+# define gcm_init_avx	gcm_init_clmul
+# define gcm_gmult_avx	gcm_gmult_clmul
+# define gcm_ghash_avx	gcm_ghash_clmul
+#else
+void gcm_init_avx(u128 Htable[16],const u64 Xi[2]);
+void gcm_gmult_avx(u64 Xi[2],const u128 Htable[16]);
+void gcm_ghash_avx(u64 Xi[2],const u128 Htable[16],const u8 *inp,size_t len);
+#endif
+
 #  if	defined(__i386) || defined(__i386__) || defined(_M_IX86)
 #   define GHASH_ASM_X86
 void gcm_gmult_4bit_mmx(u64 Xi[2],const u128 Htable[16]);
@@ -726,9 +736,15 @@ void CRYPTO_gcm128_init(GCM128_CONTEXT *ctx,void *key,block128_f block)
 #  if	!defined(GHASH_ASM_X86) || defined(OPENSSL_IA32_SSE2)
 	if (OPENSSL_ia32cap_P[0]&(1<<24) &&	/* check FXSR bit */
 	    OPENSSL_ia32cap_P[1]&(1<<1) ) {	/* check PCLMULQDQ bit */
-		gcm_init_clmul(ctx->Htable,ctx->H.u);
-		ctx->gmult = gcm_gmult_clmul;
-		ctx->ghash = gcm_ghash_clmul;
+		if (((OPENSSL_ia32cap_P[1]>>22)&0x41)==0x41) {	/* AVX+MOVBE */
+			gcm_init_avx(ctx->Htable,ctx->H.u);
+			ctx->gmult = gcm_gmult_avx;
+			ctx->ghash = gcm_ghash_avx;
+		} else {
+			gcm_init_clmul(ctx->Htable,ctx->H.u);
+			ctx->gmult = gcm_gmult_clmul;
+			ctx->ghash = gcm_ghash_clmul;
+		}
 		return;
 	}
 #  endif
@@ -1718,6 +1734,31 @@ static const u8 A19[]= {0xd9,0x31,0x32,0x25,0xf8,0x84,0x06,0xe5,0xa5,0x59,0x09,0
 			0xc5,0xf6,0x1e,0x63,0x93,0xba,0x7a,0x0a,0xbc,0xc9,0xf6,0x62,0x89,0x80,0x15,0xad},
 		T19[]= {0x5f,0xea,0x79,0x3a,0x2d,0x6f,0x97,0x4d,0x37,0xe6,0x8e,0x0c,0xb8,0xff,0x94,0x92};
 
+/* Test Case 20 */
+#define K20 K1
+#define A20 A1
+static const u8 IV20[64]={0xff,0xff,0xff,0xff},	/* this results in 0xff in counter LSB */
+		P20[288],
+		C20[]= {0x56,0xb3,0x37,0x3c,0xa9,0xef,0x6e,0x4a,0x2b,0x64,0xfe,0x1e,0x9a,0x17,0xb6,0x14,
+			0x25,0xf1,0x0d,0x47,0xa7,0x5a,0x5f,0xce,0x13,0xef,0xc6,0xbc,0x78,0x4a,0xf2,0x4f,
+			0x41,0x41,0xbd,0xd4,0x8c,0xf7,0xc7,0x70,0x88,0x7a,0xfd,0x57,0x3c,0xca,0x54,0x18,
+			0xa9,0xae,0xff,0xcd,0x7c,0x5c,0xed,0xdf,0xc6,0xa7,0x83,0x97,0xb9,0xa8,0x5b,0x49,
+			0x9d,0xa5,0x58,0x25,0x72,0x67,0xca,0xab,0x2a,0xd0,0xb2,0x3c,0xa4,0x76,0xa5,0x3c,
+			0xb1,0x7f,0xb4,0x1c,0x4b,0x8b,0x47,0x5c,0xb4,0xf3,0xf7,0x16,0x50,0x94,0xc2,0x29,
+			0xc9,0xe8,0xc4,0xdc,0x0a,0x2a,0x5f,0xf1,0x90,0x3e,0x50,0x15,0x11,0x22,0x13,0x76,
+			0xa1,0xcd,0xb8,0x36,0x4c,0x50,0x61,0xa2,0x0c,0xae,0x74,0xbc,0x4a,0xcd,0x76,0xce,
+			0xb0,0xab,0xc9,0xfd,0x32,0x17,0xef,0x9f,0x8c,0x90,0xbe,0x40,0x2d,0xdf,0x6d,0x86,
+			0x97,0xf4,0xf8,0x80,0xdf,0xf1,0x5b,0xfb,0x7a,0x6b,0x28,0x24,0x1e,0xc8,0xfe,0x18,
+			0x3c,0x2d,0x59,0xe3,0xf9,0xdf,0xff,0x65,0x3c,0x71,0x26,0xf0,0xac,0xb9,0xe6,0x42,
+			0x11,0xf4,0x2b,0xae,0x12,0xaf,0x46,0x2b,0x10,0x70,0xbe,0xf1,0xab,0x5e,0x36,0x06,
+			0x87,0x2c,0xa1,0x0d,0xee,0x15,0xb3,0x24,0x9b,0x1a,0x1b,0x95,0x8f,0x23,0x13,0x4c,
+			0x4b,0xcc,0xb7,0xd0,0x32,0x00,0xbc,0xe4,0x20,0xa2,0xf8,0xeb,0x66,0xdc,0xf3,0x64,
+			0x4d,0x14,0x23,0xc1,0xb5,0x69,0x90,0x03,0xc1,0x3e,0xce,0xf4,0xbf,0x38,0xa3,0xb6,
+			0x0e,0xed,0xc3,0x40,0x33,0xba,0xc1,0x90,0x27,0x83,0xdc,0x6d,0x89,0xe2,0xe7,0x74,
+			0x18,0x8a,0x43,0x9c,0x7e,0xbc,0xc0,0x67,0x2d,0xbd,0xa4,0xdd,0xcf,0xb2,0x79,0x46,
+			0x13,0xb0,0xbe,0x41,0x31,0x5e,0xf7,0x78,0x70,0x8a,0x70,0xee,0x7d,0x75,0x16,0x5c},
+		T20[]= {0x8b,0x30,0x7f,0x6b,0x33,0x28,0x6d,0x0a,0xb0,0x26,0xa9,0xed,0x3f,0xe1,0xe8,0x5f};
+
 #define TEST_CASE(n)	do {					\
 	u8 out[sizeof(P##n)];					\
 	AES_set_encrypt_key(K##n,sizeof(K##n)*8,&key);		\
@@ -1763,6 +1804,7 @@ int main()
 	TEST_CASE(17);
 	TEST_CASE(18);
 	TEST_CASE(19);
+	TEST_CASE(20);
 
 #ifdef OPENSSL_CPUID_OBJ
 	{