diff options
Diffstat (limited to 'arch/x86/crypto')
| -rw-r--r-- | arch/x86/crypto/ghash-clmulni-intel_asm.S | 29 |
1 files changed, 10 insertions, 19 deletions
diff --git a/arch/x86/crypto/ghash-clmulni-intel_asm.S b/arch/x86/crypto/ghash-clmulni-intel_asm.S index 59584982fb75..1528dc4886cf 100644 --- a/arch/x86/crypto/ghash-clmulni-intel_asm.S +++ b/arch/x86/crypto/ghash-clmulni-intel_asm.S | |||
| @@ -17,7 +17,7 @@ | |||
| 17 | */ | 17 | */ |
| 18 | 18 | ||
| 19 | #include <linux/linkage.h> | 19 | #include <linux/linkage.h> |
| 20 | #include <asm/i387.h> | 20 | #include <asm/inst.h> |
| 21 | 21 | ||
| 22 | .align 16 | 22 | .align 16 |
| 23 | .Lbswap_mask: | 23 | .Lbswap_mask: |
| @@ -56,12 +56,9 @@ __clmul_gf128mul_ble: | |||
| 56 | pxor DATA, T2 | 56 | pxor DATA, T2 |
| 57 | pxor SHASH, T3 | 57 | pxor SHASH, T3 |
| 58 | 58 | ||
| 59 | # pclmulqdq $0x00, SHASH, DATA # DATA = a0 * b0 | 59 | PCLMULQDQ 0x00 SHASH DATA # DATA = a0 * b0 |
| 60 | .byte 0x66, 0x0f, 0x3a, 0x44, 0xc1, 0x00 | 60 | PCLMULQDQ 0x11 SHASH T1 # T1 = a1 * b1 |
| 61 | # pclmulqdq $0x11, SHASH, T1 # T1 = a1 * b1 | 61 | PCLMULQDQ 0x00 T3 T2 # T2 = (a1 + a0) * (b1 + b0) |
| 62 | .byte 0x66, 0x0f, 0x3a, 0x44, 0xd1, 0x11 | ||
| 63 | # pclmulqdq $0x00, T3, T2 # T2 = (a1 + a0) * (b1 + b0) | ||
| 64 | .byte 0x66, 0x0f, 0x3a, 0x44, 0xdc, 0x00 | ||
| 65 | pxor DATA, T2 | 62 | pxor DATA, T2 |
| 66 | pxor T1, T2 # T2 = a0 * b1 + a1 * b0 | 63 | pxor T1, T2 # T2 = a0 * b1 + a1 * b0 |
| 67 | 64 | ||
| @@ -101,11 +98,9 @@ ENTRY(clmul_ghash_mul) | |||
| 101 | movups (%rdi), DATA | 98 | movups (%rdi), DATA |
| 102 | movups (%rsi), SHASH | 99 | movups (%rsi), SHASH |
| 103 | movaps .Lbswap_mask, BSWAP | 100 | movaps .Lbswap_mask, BSWAP |
| 104 | # pshufb BSWAP, DATA | 101 | PSHUFB_XMM BSWAP DATA |
| 105 | PSHUFB_XMM5_XMM0 | ||
| 106 | call __clmul_gf128mul_ble | 102 | call __clmul_gf128mul_ble |
| 107 | # pshufb BSWAP, DATA | 103 | PSHUFB_XMM BSWAP DATA |
| 108 | .byte 0x66, 0x0f, 0x38, 0x00, 0xc5 | ||
| 109 | movups DATA, (%rdi) | 104 | movups DATA, (%rdi) |
| 110 | ret | 105 | ret |
| 111 | 106 | ||
| @@ -119,21 +114,18 @@ ENTRY(clmul_ghash_update) | |||
| 119 | movaps .Lbswap_mask, BSWAP | 114 | movaps .Lbswap_mask, BSWAP |
| 120 | movups (%rdi), DATA | 115 | movups (%rdi), DATA |
| 121 | movups (%rcx), SHASH | 116 | movups (%rcx), SHASH |
| 122 | # pshufb BSWAP, DATA | 117 | PSHUFB_XMM BSWAP DATA |
| 123 | PSHUFB_XMM5_XMM0 | ||
| 124 | .align 4 | 118 | .align 4 |
| 125 | .Lupdate_loop: | 119 | .Lupdate_loop: |
| 126 | movups (%rsi), IN1 | 120 | movups (%rsi), IN1 |
| 127 | # pshufb BSWAP, IN1 | 121 | PSHUFB_XMM BSWAP IN1 |
| 128 | PSHUFB_XMM5_XMM6 | ||
| 129 | pxor IN1, DATA | 122 | pxor IN1, DATA |
| 130 | call __clmul_gf128mul_ble | 123 | call __clmul_gf128mul_ble |
| 131 | sub $16, %rdx | 124 | sub $16, %rdx |
| 132 | add $16, %rsi | 125 | add $16, %rsi |
| 133 | cmp $16, %rdx | 126 | cmp $16, %rdx |
| 134 | jge .Lupdate_loop | 127 | jge .Lupdate_loop |
| 135 | # pshufb BSWAP, DATA | 128 | PSHUFB_XMM BSWAP DATA |
| 136 | PSHUFB_XMM5_XMM0 | ||
| 137 | movups DATA, (%rdi) | 129 | movups DATA, (%rdi) |
| 138 | .Lupdate_just_ret: | 130 | .Lupdate_just_ret: |
| 139 | ret | 131 | ret |
| @@ -146,8 +138,7 @@ ENTRY(clmul_ghash_update) | |||
| 146 | ENTRY(clmul_ghash_setkey) | 138 | ENTRY(clmul_ghash_setkey) |
| 147 | movaps .Lbswap_mask, BSWAP | 139 | movaps .Lbswap_mask, BSWAP |
| 148 | movups (%rsi), %xmm0 | 140 | movups (%rsi), %xmm0 |
| 149 | # pshufb BSWAP, %xmm0 | 141 | PSHUFB_XMM BSWAP %xmm0 |
| 150 | PSHUFB_XMM5_XMM0 | ||
| 151 | movaps %xmm0, %xmm1 | 142 | movaps %xmm0, %xmm1 |
| 152 | psllq $1, %xmm0 | 143 | psllq $1, %xmm0 |
| 153 | psrlq $63, %xmm1 | 144 | psrlq $63, %xmm1 |