crypto/aes: Made faster by eliminating some indirection

Made te and td arrays into variables te0-3 and td0-3,
which improves performance from 7000ns/op to 5800.

R=rsc, rogpeppe, agl
CC=golang-dev
https://golang.org/cl/5449077

src/pkg/crypto/aes/aes_test.go

@@ -91,6 +91,7 @@ func TestTe(t *testing.T) {
 		s2 := mul(s, 2)
 		s3 := mul(s, 3)
 		w := s2<<24 | s<<16 | s<<8 | s3
+		te := [][256]uint32{te0, te1, te2, te3}
 		for j := 0; j < 4; j++ {
 			if x := te[j][i]; x != w {
 				t.Fatalf("te[%d][%d] = %#x, want %#x", j, i, x, w)
@@ -110,6 +111,7 @@ func TestTd(t *testing.T) {
 		sd := mul(s, 0xd)
 		se := mul(s, 0xe)
 		w := se<<24 | s9<<16 | sd<<8 | sb
+		td := [][256]uint32{td0, td1, td2, td3}
 		for j := 0; j < 4; j++ {
 			if x := td[j][i]; x != w {
 				t.Fatalf("td[%d][%d] = %#x, want %#x", j, i, x, w)

src/pkg/crypto/aes/block.go

@@ -56,10 +56,10 @@ func encryptBlock(xk []uint32, dst, src []byte) {
 	nr := len(xk)/4 - 2 // - 2: one above, one more below
 	k := 4
 	for r := 0; r < nr; r++ {
-		t0 = xk[k+0] ^ te[0][uint8(s0>>24)] ^ te[1][uint8(s1>>16)] ^ te[2][uint8(s2>>8)] ^ te[3][uint8(s3)]
-		t1 = xk[k+1] ^ te[0][uint8(s1>>24)] ^ te[1][uint8(s2>>16)] ^ te[2][uint8(s3>>8)] ^ te[3][uint8(s0)]
-		t2 = xk[k+2] ^ te[0][uint8(s2>>24)] ^ te[1][uint8(s3>>16)] ^ te[2][uint8(s0>>8)] ^ te[3][uint8(s1)]
-		t3 = xk[k+3] ^ te[0][uint8(s3>>24)] ^ te[1][uint8(s0>>16)] ^ te[2][uint8(s1>>8)] ^ te[3][uint8(s2)]
+		t0 = xk[k+0] ^ te0[uint8(s0>>24)] ^ te1[uint8(s1>>16)] ^ te2[uint8(s2>>8)] ^ te3[uint8(s3)]
+		t1 = xk[k+1] ^ te0[uint8(s1>>24)] ^ te1[uint8(s2>>16)] ^ te2[uint8(s3>>8)] ^ te3[uint8(s0)]
+		t2 = xk[k+2] ^ te0[uint8(s2>>24)] ^ te1[uint8(s3>>16)] ^ te2[uint8(s0>>8)] ^ te3[uint8(s1)]
+		t3 = xk[k+3] ^ te0[uint8(s3>>24)] ^ te1[uint8(s0>>16)] ^ te2[uint8(s1>>8)] ^ te3[uint8(s2)]
 		k += 4
 		s0, s1, s2, s3 = t0, t1, t2, t3
 	}
@@ -101,10 +101,10 @@ func decryptBlock(xk []uint32, dst, src []byte) {
 	nr := len(xk)/4 - 2 // - 2: one above, one more below
 	k := 4
 	for r := 0; r < nr; r++ {
-		t0 = xk[k+0] ^ td[0][uint8(s0>>24)] ^ td[1][uint8(s3>>16)] ^ td[2][uint8(s2>>8)] ^ td[3][uint8(s1)]
-		t1 = xk[k+1] ^ td[0][uint8(s1>>24)] ^ td[1][uint8(s0>>16)] ^ td[2][uint8(s3>>8)] ^ td[3][uint8(s2)]
-		t2 = xk[k+2] ^ td[0][uint8(s2>>24)] ^ td[1][uint8(s1>>16)] ^ td[2][uint8(s0>>8)] ^ td[3][uint8(s3)]
-		t3 = xk[k+3] ^ td[0][uint8(s3>>24)] ^ td[1][uint8(s2>>16)] ^ td[2][uint8(s1>>8)] ^ td[3][uint8(s0)]
+		t0 = xk[k+0] ^ td0[uint8(s0>>24)] ^ td1[uint8(s3>>16)] ^ td2[uint8(s2>>8)] ^ td3[uint8(s1)]
+		t1 = xk[k+1] ^ td0[uint8(s1>>24)] ^ td1[uint8(s0>>16)] ^ td2[uint8(s3>>8)] ^ td3[uint8(s2)]
+		t2 = xk[k+2] ^ td0[uint8(s2>>24)] ^ td1[uint8(s1>>16)] ^ td2[uint8(s0>>8)] ^ td3[uint8(s3)]
+		t3 = xk[k+3] ^ td0[uint8(s3>>24)] ^ td1[uint8(s2>>16)] ^ td2[uint8(s1>>8)] ^ td3[uint8(s0)]
 		k += 4
 		s0, s1, s2, s3 = t0, t1, t2, t3
 	}
@@ -168,7 +168,7 @@ func expandKey(key []byte, enc, dec []uint32) {
 		for j := 0; j < 4; j++ {
 			x := enc[ei+j]
 			if i > 0 && i+4 < n {
-				x = td[0][sbox0[x>>24]] ^ td[1][sbox0[x>>16&0xff]] ^ td[2][sbox0[x>>8&0xff]] ^ td[3][sbox0[x&0xff]]
+				x = td0[sbox0[x>>24]] ^ td1[sbox0[x>>16&0xff]] ^ td2[sbox0[x>>8&0xff]] ^ td3[sbox0[x&0xff]]
 			}
 			dec[i+j] = x
 		}

src/pkg/crypto/aes/const.go

@@ -80,8 +80,7 @@ var sbox1 = [256]byte{
 // Lookup tables for encryption.
 // These can be recomputed by adapting the tests in aes_test.go.
 
-var te = [4][256]uint32{
-	{
+var te0 = [256]uint32{
 	0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d, 0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554,
 	0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d, 0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a,
 	0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87, 0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b,
@@ -114,8 +113,8 @@ var te = [4][256]uint32{
 	0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920, 0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a,
 	0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17, 0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8,
 	0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11, 0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a,
-	},
-	{
+}
+var te1 = [256]uint32{
 	0xa5c66363, 0x84f87c7c, 0x99ee7777, 0x8df67b7b, 0x0dfff2f2, 0xbdd66b6b, 0xb1de6f6f, 0x5491c5c5,
 	0x50603030, 0x03020101, 0xa9ce6767, 0x7d562b2b, 0x19e7fefe, 0x62b5d7d7, 0xe64dabab, 0x9aec7676,
 	0x458fcaca, 0x9d1f8282, 0x4089c9c9, 0x87fa7d7d, 0x15effafa, 0xebb25959, 0xc98e4747, 0x0bfbf0f0,
@@ -148,8 +147,8 @@ var te = [4][256]uint32{
 	0xb62d9b9b, 0x223c1e1e, 0x92158787, 0x20c9e9e9, 0x4987cece, 0xffaa5555, 0x78502828, 0x7aa5dfdf,
 	0x8f038c8c, 0xf859a1a1, 0x80098989, 0x171a0d0d, 0xda65bfbf, 0x31d7e6e6, 0xc6844242, 0xb8d06868,
 	0xc3824141, 0xb0299999, 0x775a2d2d, 0x111e0f0f, 0xcb7bb0b0, 0xfca85454, 0xd66dbbbb, 0x3a2c1616,
-	},
-	{
+}
+var te2 = [256]uint32{
 	0x63a5c663, 0x7c84f87c, 0x7799ee77, 0x7b8df67b, 0xf20dfff2, 0x6bbdd66b, 0x6fb1de6f, 0xc55491c5,
 	0x30506030, 0x01030201, 0x67a9ce67, 0x2b7d562b, 0xfe19e7fe, 0xd762b5d7, 0xabe64dab, 0x769aec76,
 	0xca458fca, 0x829d1f82, 0xc94089c9, 0x7d87fa7d, 0xfa15effa, 0x59ebb259, 0x47c98e47, 0xf00bfbf0,
@@ -182,8 +181,8 @@ var te = [4][256]uint32{
 	0x9bb62d9b, 0x1e223c1e, 0x87921587, 0xe920c9e9, 0xce4987ce, 0x55ffaa55, 0x28785028, 0xdf7aa5df,
 	0x8c8f038c, 0xa1f859a1, 0x89800989, 0x0d171a0d, 0xbfda65bf, 0xe631d7e6, 0x42c68442, 0x68b8d068,
 	0x41c38241, 0x99b02999, 0x2d775a2d, 0x0f111e0f, 0xb0cb7bb0, 0x54fca854, 0xbbd66dbb, 0x163a2c16,
-	},
-	{
+}
+var te3 = [256]uint32{
 	0x6363a5c6, 0x7c7c84f8, 0x777799ee, 0x7b7b8df6, 0xf2f20dff, 0x6b6bbdd6, 0x6f6fb1de, 0xc5c55491,
 	0x30305060, 0x01010302, 0x6767a9ce, 0x2b2b7d56, 0xfefe19e7, 0xd7d762b5, 0xababe64d, 0x76769aec,
 	0xcaca458f, 0x82829d1f, 0xc9c94089, 0x7d7d87fa, 0xfafa15ef, 0x5959ebb2, 0x4747c98e, 0xf0f00bfb,
@@ -216,14 +215,12 @@ var te = [4][256]uint32{
 	0x9b9bb62d, 0x1e1e223c, 0x87879215, 0xe9e920c9, 0xcece4987, 0x5555ffaa, 0x28287850, 0xdfdf7aa5,
 	0x8c8c8f03, 0xa1a1f859, 0x89898009, 0x0d0d171a, 0xbfbfda65, 0xe6e631d7, 0x4242c684, 0x6868b8d0,
 	0x4141c382, 0x9999b029, 0x2d2d775a, 0x0f0f111e, 0xb0b0cb7b, 0x5454fca8, 0xbbbbd66d, 0x16163a2c,
-	},
 }
 
 // Lookup tables for decryption.
 // These can be recomputed by adapting the tests in aes_test.go.
 
-var td = [4][256]uint32{
-	{
+var td0 = [256]uint32{
 	0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96, 0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393,
 	0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25, 0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f,
 	0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1, 0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6,
@@ -256,8 +253,8 @@ var td = [4][256]uint32{
 	0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf, 0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86,
 	0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f, 0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541,
 	0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190, 0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742,
-	},
-	{
+}
+var td1 = [256]uint32{
 	0x5051f4a7, 0x537e4165, 0xc31a17a4, 0x963a275e, 0xcb3bab6b, 0xf11f9d45, 0xabacfa58, 0x934be303,
 	0x552030fa, 0xf6ad766d, 0x9188cc76, 0x25f5024c, 0xfc4fe5d7, 0xd7c52acb, 0x80263544, 0x8fb562a3,
 	0x49deb15a, 0x6725ba1b, 0x9845ea0e, 0xe15dfec0, 0x02c32f75, 0x12814cf0, 0xa38d4697, 0xc66bd3f9,
@@ -290,8 +287,8 @@ var td = [4][256]uint32{
 	0x599cd2df, 0x3f55f273, 0x791814ce, 0xbf73c737, 0xea53f7cd, 0x5b5ffdaa, 0x14df3d6f, 0x867844db,
 	0x81caaff3, 0x3eb968c4, 0x2c382434, 0x5fc2a340, 0x72161dc3, 0x0cbce225, 0x8b283c49, 0x41ff0d95,
 	0x7139a801, 0xde080cb3, 0x9cd8b4e4, 0x906456c1, 0x617bcb84, 0x70d532b6, 0x74486c5c, 0x42d0b857,
-	},
-	{
+}
+var td2 = [256]uint32{
 	0xa75051f4, 0x65537e41, 0xa4c31a17, 0x5e963a27, 0x6bcb3bab, 0x45f11f9d, 0x58abacfa, 0x03934be3,
 	0xfa552030, 0x6df6ad76, 0x769188cc, 0x4c25f502, 0xd7fc4fe5, 0xcbd7c52a, 0x44802635, 0xa38fb562,
 	0x5a49deb1, 0x1b6725ba, 0x0e9845ea, 0xc0e15dfe, 0x7502c32f, 0xf012814c, 0x97a38d46, 0xf9c66bd3,
@@ -324,8 +321,8 @@ var td = [4][256]uint32{
 	0xdf599cd2, 0x733f55f2, 0xce791814, 0x37bf73c7, 0xcdea53f7, 0xaa5b5ffd, 0x6f14df3d, 0xdb867844,
 	0xf381caaf, 0xc43eb968, 0x342c3824, 0x405fc2a3, 0xc372161d, 0x250cbce2, 0x498b283c, 0x9541ff0d,
 	0x017139a8, 0xb3de080c, 0xe49cd8b4, 0xc1906456, 0x84617bcb, 0xb670d532, 0x5c74486c, 0x5742d0b8,
-	},
-	{
+}
+var td3 = [256]uint32{
 	0xf4a75051, 0x4165537e, 0x17a4c31a, 0x275e963a, 0xab6bcb3b, 0x9d45f11f, 0xfa58abac, 0xe303934b,
 	0x30fa5520, 0x766df6ad, 0xcc769188, 0x024c25f5, 0xe5d7fc4f, 0x2acbd7c5, 0x35448026, 0x62a38fb5,
 	0xb15a49de, 0xba1b6725, 0xea0e9845, 0xfec0e15d, 0x2f7502c3, 0x4cf01281, 0x4697a38d, 0xd3f9c66b,
@@ -358,5 +355,4 @@ var td = [4][256]uint32{
 	0xd2df599c, 0xf2733f55, 0x14ce7918, 0xc737bf73, 0xf7cdea53, 0xfdaa5b5f, 0x3d6f14df, 0x44db8678,
 	0xaff381ca, 0x68c43eb9, 0x24342c38, 0xa3405fc2, 0x1dc37216, 0xe2250cbc, 0x3c498b28, 0x0d9541ff,
 	0xa8017139, 0x0cb3de08, 0xb4e49cd8, 0x56c19064, 0xcb84617b, 0x32b670d5, 0x6c5c7448, 0xb85742d0,
-	},
 }