aboutsummaryrefslogtreecommitdiffstatshomepage
path: root/src/asm
diff options
context:
space:
mode:
authorEgor Tensin <Egor.Tensin@gmail.com>2015-06-10 20:48:53 +0300
committerEgor Tensin <Egor.Tensin@gmail.com>2015-06-10 20:48:53 +0300
commitab42aa88992409eb5db5714df18390387c6e041c (patch)
tree8f5d8adf9e644f192d6d43d79c8db9656e57c90e /src/asm
parentbugfix (diff)
downloadaes-tools-ab42aa88992409eb5db5714df18390387c6e041c.tar.gz
aes-tools-ab42aa88992409eb5db5714df18390387c6e041c.zip
fix register usage in the asm implementation
Apparently, the previously used xmm6/xmm7 registers are exactly the wrong ones to use for storing temp values. The xmm4/xmm5, on the other hand hand, don't need to be preserved, so I switched to those.
Diffstat (limited to 'src/asm')
-rw-r--r--src/asm/aes128.asm96
-rw-r--r--src/asm/aes192.asm112
-rw-r--r--src/asm/aes256.asm146
3 files changed, 177 insertions, 177 deletions
diff --git a/src/asm/aes128.asm b/src/asm/aes128.asm
index b49de0e..874f876 100644
--- a/src/asm/aes128.asm
+++ b/src/asm/aes128.asm
@@ -88,25 +88,25 @@
movdqa [ecx], xmm0 ; sets w[0], w[1], w[2], w[3]
add ecx, 10h ; ecx = &w[4]
- aeskeygenassist xmm7, xmm0, 01h ; xmm7[127:96] = RotWord(SubWord(w[3]))^Rcon
+ aeskeygenassist xmm5, xmm0, 01h ; xmm5[127:96] = RotWord(SubWord(w[3]))^Rcon
call aes128_keygen_assist ; sets w[4], w[5], w[6], w[7]
- aeskeygenassist xmm7, xmm0, 02h ; xmm7[127:96] = RotWord(SubWord(w[7]))^Rcon
+ aeskeygenassist xmm5, xmm0, 02h ; xmm5[127:96] = RotWord(SubWord(w[7]))^Rcon
call aes128_keygen_assist ; sets w[8], w[9], w[10], w[11]
- aeskeygenassist xmm7, xmm0, 04h ; xmm7[127:96] = RotWord(SubWord(w[11]))^Rcon
+ aeskeygenassist xmm5, xmm0, 04h ; xmm5[127:96] = RotWord(SubWord(w[11]))^Rcon
call aes128_keygen_assist ; sets w[12], w[13], w[14], w[15]
- aeskeygenassist xmm7, xmm0, 08h ; xmm7[127:96] = RotWord(SubWord(w[15]))^Rcon
+ aeskeygenassist xmm5, xmm0, 08h ; xmm5[127:96] = RotWord(SubWord(w[15]))^Rcon
call aes128_keygen_assist ; sets w[16], w[17], w[18], w[19]
- aeskeygenassist xmm7, xmm0, 10h ; xmm7[127:96] = RotWord(SubWord(w[19]))^Rcon
+ aeskeygenassist xmm5, xmm0, 10h ; xmm5[127:96] = RotWord(SubWord(w[19]))^Rcon
call aes128_keygen_assist ; sets w[20], w[21], w[22], w[23]
- aeskeygenassist xmm7, xmm0, 20h ; xmm7[127:96] = RotWord(SubWord(w[23]))^Rcon
+ aeskeygenassist xmm5, xmm0, 20h ; xmm5[127:96] = RotWord(SubWord(w[23]))^Rcon
call aes128_keygen_assist ; sets w[24], w[25], w[26], w[27]
- aeskeygenassist xmm7, xmm0, 40h ; xmm7[127:96] = RotWord(SubWord(w[27]))^Rcon
+ aeskeygenassist xmm5, xmm0, 40h ; xmm5[127:96] = RotWord(SubWord(w[27]))^Rcon
call aes128_keygen_assist ; sets w[28], w[29], w[30], w[31]
- aeskeygenassist xmm7, xmm0, 80h ; xmm7[127:96] = RotWord(SubWord(w[31]))^Rcon
+ aeskeygenassist xmm5, xmm0, 80h ; xmm5[127:96] = RotWord(SubWord(w[31]))^Rcon
call aes128_keygen_assist ; sets w[32], w[33], w[34], w[35]
- aeskeygenassist xmm7, xmm0, 1Bh ; xmm7[127:96] = RotWord(SubWord(w[35]))^Rcon
+ aeskeygenassist xmm5, xmm0, 1Bh ; xmm5[127:96] = RotWord(SubWord(w[35]))^Rcon
call aes128_keygen_assist ; sets w[36], w[37], w[38], w[39]
- aeskeygenassist xmm7, xmm0, 36h ; xmm7[127:96] = RotWord(SubWord(w[39]))^Rcon
+ aeskeygenassist xmm5, xmm0, 36h ; xmm5[127:96] = RotWord(SubWord(w[39]))^Rcon
call aes128_keygen_assist ; sets w[40], w[41], w[42], w[43]
ret
@@ -117,7 +117,7 @@ aes128_keygen_assist:
; * xmm0[95:64] == w[i+2],
; * xmm0[63:32] == w[i+1],
; * xmm0[31:0] == w[i],
- ; * xmm7[127:96] == RotWord(SubWord(w[i+3]))^Rcon,
+ ; * xmm5[127:96] == RotWord(SubWord(w[i+3]))^Rcon,
; * ecx == &w[i+4].
;
; Postconditions:
@@ -126,20 +126,20 @@ aes128_keygen_assist:
; * xmm0[63:32] == w[i+5] == RotWord(SubWord(w[i+3]))^Rcon^w[i+1]^w[i],
; * xmm0[31:0] == w[i+4] == RotWord(SubWord(w[i+3]))^Rcon^w[i],
; * ecx == &w[i+8],
- ; * the value in xmm6 is also modified.
+ ; * the value in xmm4 is also modified.
; Calculate
; w[i+3]^w[i+2]^w[i+1]^w[i],
; w[i+2]^w[i+1]^w[i],
; w[i+1]^w[i] and
; w[i].
- movdqa xmm6, xmm0 ; xmm6 = xmm0
- pslldq xmm6, 4 ; xmm6 <<= 32
- pxor xmm0, xmm6 ; xmm0 ^= xmm6
- pslldq xmm6, 4 ; xmm6 <<= 32
- pxor xmm0, xmm6 ; xmm0 ^= xmm6
- pslldq xmm6, 4 ; xmm6 <<= 32
- pxor xmm0, xmm6 ; xmm0 ^= xmm6
+ movdqa xmm4, xmm0 ; xmm4 = xmm0
+ pslldq xmm4, 4 ; xmm4 <<= 32
+ pxor xmm0, xmm4 ; xmm0 ^= xmm4
+ pslldq xmm4, 4 ; xmm4 <<= 32
+ pxor xmm0, xmm4 ; xmm0 ^= xmm4
+ pslldq xmm4, 4 ; xmm4 <<= 32
+ pxor xmm0, xmm4 ; xmm0 ^= xmm4
; xmm0[127:96] == w[i+3]^w[i+2]^w[i+1]^w[i]
; xmm0[95:64] == w[i+2]^w[i+1]^w[i]
; xmm0[63:32] == w[i+1]^w[i]
@@ -150,8 +150,8 @@ aes128_keygen_assist:
; w[i+6] == RotWord(SubWord(w[i+3]))^Rcon^w[i+2]^w[i+1]^w[i],
; w[i+5] == RotWord(SubWord(w[i+3]))^Rcon^w[i+1]^w[i] and
; w[i+4] == RotWord(SubWord(w[i+3]))^Rcon^w[i].
- pshufd xmm6, xmm7, 0FFh ; xmm6[127:96] = xmm6[95:64] = xmm6[63:32] = xmm6[31:0] = xmm7[127:96]
- pxor xmm0, xmm6 ; xmm0 ^= xmm6
+ pshufd xmm4, xmm5, 0FFh ; xmm4[127:96] = xmm4[95:64] = xmm4[63:32] = xmm4[31:0] = xmm5[127:96]
+ pxor xmm0, xmm4 ; xmm0 ^= xmm4
; xmm0[127:96] == w[i+7] == RotWord(SubWord(w[i+3]))^Rcon^w[i+3]^w[i+2]^w[i+1]^w[i]
; xmm0[95:64] == w[i+6] == RotWord(SubWord(w[i+3]))^Rcon^w[i+2]^w[i+1]^w[i]
; xmm0[63:32] == w[i+5] == RotWord(SubWord(w[i+3]))^Rcon^w[i+1]^w[i]
@@ -168,33 +168,33 @@ aes128_keygen_assist:
@raw_aes128_expand_key_schedule@20 endp
@raw_aes128_invert_key_schedule@8 proc
- movdqa xmm7, [ecx]
- movdqa xmm6, [ecx + 0A0h]
- movdqa [edx], xmm6
- movdqa [edx + 0A0h], xmm7
-
- aesimc xmm7, [ecx + 10h]
- aesimc xmm6, [ecx + 90h]
- movdqa [edx + 10h], xmm6
- movdqa [edx + 90h], xmm7
-
- aesimc xmm7, [ecx + 20h]
- aesimc xmm6, [ecx + 80h]
- movdqa [edx + 20h], xmm6
- movdqa [edx + 80h], xmm7
-
- aesimc xmm7, [ecx + 30h]
- aesimc xmm6, [ecx + 70h]
- movdqa [edx + 30h], xmm6
- movdqa [edx + 70h], xmm7
-
- aesimc xmm7, [ecx + 40h]
- aesimc xmm6, [ecx + 60h]
- movdqa [edx + 40h], xmm6
- movdqa [edx + 60h], xmm7
-
- aesimc xmm7, [ecx + 50h]
- movdqa [edx + 50h], xmm7
+ movdqa xmm5, [ecx]
+ movdqa xmm4, [ecx + 0A0h]
+ movdqa [edx], xmm4
+ movdqa [edx + 0A0h], xmm5
+
+ aesimc xmm5, [ecx + 10h]
+ aesimc xmm4, [ecx + 90h]
+ movdqa [edx + 10h], xmm4
+ movdqa [edx + 90h], xmm5
+
+ aesimc xmm5, [ecx + 20h]
+ aesimc xmm4, [ecx + 80h]
+ movdqa [edx + 20h], xmm4
+ movdqa [edx + 80h], xmm5
+
+ aesimc xmm5, [ecx + 30h]
+ aesimc xmm4, [ecx + 70h]
+ movdqa [edx + 30h], xmm4
+ movdqa [edx + 70h], xmm5
+
+ aesimc xmm5, [ecx + 40h]
+ aesimc xmm4, [ecx + 60h]
+ movdqa [edx + 40h], xmm4
+ movdqa [edx + 60h], xmm5
+
+ aesimc xmm5, [ecx + 50h]
+ movdqa [edx + 50h], xmm5
ret
@raw_aes128_invert_key_schedule@8 endp
diff --git a/src/asm/aes192.asm b/src/asm/aes192.asm
index 5cc8ded..adbc690 100644
--- a/src/asm/aes192.asm
+++ b/src/asm/aes192.asm
@@ -102,35 +102,35 @@
movdqa [ecx], xmm0 ; sets w[0], w[1], w[2], w[3]
movq qword ptr [ecx + 10h], xmm1 ; sets w[4], w[5]
- aeskeygenassist xmm7, xmm1, 1 ; xmm7[63:32] = RotWord(SubWord(w[5]))^Rcon,
+ aeskeygenassist xmm5, xmm1, 1 ; xmm5[63:32] = RotWord(SubWord(w[5]))^Rcon,
call aes192_keygen_assist
movdqu [ecx + 18h], xmm0
movq qword ptr [ecx + 28h], xmm1
- aeskeygenassist xmm7, xmm1, 2 ; xmm7[63:32] = RotWord(SubWord(w[11]))^Rcon
+ aeskeygenassist xmm5, xmm1, 2 ; xmm5[63:32] = RotWord(SubWord(w[11]))^Rcon
call aes192_keygen_assist
movdqa [ecx + 30h], xmm0
movq qword ptr [ecx + 40h], xmm1
- aeskeygenassist xmm7, xmm1, 4 ; xmm7[63:32] = RotWord(SubWord(w[17]))^Rcon
+ aeskeygenassist xmm5, xmm1, 4 ; xmm5[63:32] = RotWord(SubWord(w[17]))^Rcon
call aes192_keygen_assist
movdqu [ecx + 48h], xmm0
movq qword ptr [ecx + 58h], xmm1
- aeskeygenassist xmm7, xmm1, 8 ; xmm7[63:32] = RotWord(SubWord(w[23]))^Rcon
+ aeskeygenassist xmm5, xmm1, 8 ; xmm5[63:32] = RotWord(SubWord(w[23]))^Rcon
call aes192_keygen_assist
movdqa [ecx + 60h], xmm0
movq qword ptr [ecx + 70h], xmm1
- aeskeygenassist xmm7, xmm1, 10h ; xmm7[63:32] = RotWord(SubWord(w[29]))^Rcon
+ aeskeygenassist xmm5, xmm1, 10h ; xmm5[63:32] = RotWord(SubWord(w[29]))^Rcon
call aes192_keygen_assist
movdqu [ecx + 78h], xmm0
movq qword ptr [ecx + 88h], xmm1
- aeskeygenassist xmm7, xmm1, 20h ; xmm7[63:32] = RotWord(SubWord(w[35]))^Rcon
+ aeskeygenassist xmm5, xmm1, 20h ; xmm5[63:32] = RotWord(SubWord(w[35]))^Rcon
call aes192_keygen_assist
movdqa [ecx + 90h], xmm0
movq qword ptr [ecx + 0a0h], xmm1
- aeskeygenassist xmm7, xmm1, 40h ; xmm7[63:32] = RotWord(SubWord(w[41]))^Rcon
+ aeskeygenassist xmm5, xmm1, 40h ; xmm5[63:32] = RotWord(SubWord(w[41]))^Rcon
call aes192_keygen_assist
movdqu [ecx + 0a8h], xmm0
movq qword ptr [ecx + 0b8h], xmm1
- aeskeygenassist xmm7, xmm1, 80h ; xmm7[63:32] = RotWord(SubWord(w[49]))^Rcon
+ aeskeygenassist xmm5, xmm1, 80h ; xmm5[63:32] = RotWord(SubWord(w[49]))^Rcon
call aes192_keygen_assist
movdqa [ecx + 0c0h], xmm0
@@ -146,7 +146,7 @@ aes192_keygen_assist:
; * xmm0[95:64] == w[i+2],
; * xmm0[63:32] == w[i+1],
; * xmm0[31:0] == w[i],
- ; * xmm7[63:32] == RotWord(SubWord(w[i+5]))^Rcon.
+ ; * xmm5[63:32] == RotWord(SubWord(w[i+5]))^Rcon.
;
; Postconditions:
; * xmm1[127:96] == 0,
@@ -157,20 +157,20 @@ aes192_keygen_assist:
; * xmm0[95:64] == w[i+8] == RotWord(SubWord(w[i+5]))^Rcon^w[i+2]^w[i+1]^w[i],
; * xmm0[63:32] == w[i+7] == RotWord(SubWord(w[i+5]))^Rcon^w[i+1]^w[i],
; * xmm0[31:0] == w[i+6] == RotWord(SubWord(w[i+5]))^Rcon^w[i],
- ; * the value in xmm6 is also modified.
+ ; * the value in xmm4 is also modified.
; Calculate
; w[i+3]^w[i+2]^w[i+1]^w[i],
; w[i+2]^w[i+1]^w[i],
; w[i+1]^w[i] and
; w[i].
- movdqa xmm6, xmm0 ; xmm6 = xmm0
- pslldq xmm6, 4 ; xmm6 <<= 32
- pxor xmm0, xmm6 ; xmm0 ^= xmm6
- pslldq xmm6, 4 ; xmm6 <<= 32
- pxor xmm0, xmm6 ; xmm0 ^= xmm6
- pslldq xmm6, 4 ; xmm6 <<= 32
- pxor xmm0, xmm6 ; xmm0 ^= xmm6
+ movdqa xmm4, xmm0 ; xmm4 = xmm0
+ pslldq xmm4, 4 ; xmm4 <<= 32
+ pxor xmm0, xmm4 ; xmm0 ^= xmm4
+ pslldq xmm4, 4 ; xmm4 <<= 32
+ pxor xmm0, xmm4 ; xmm0 ^= xmm4
+ pslldq xmm4, 4 ; xmm4 <<= 32
+ pxor xmm0, xmm4 ; xmm0 ^= xmm4
; xmm0[127:96] == w[i+3]^w[i+2]^w[i+1]^w[i]
; xmm0[95:64] == w[i+2]^w[i+1]^w[i]
; xmm0[63:32] == w[i+1]^w[i]
@@ -181,8 +181,8 @@ aes192_keygen_assist:
; w[i+8] == RotWord(SubWord(w[i+5]))^Rcon^w[i+2]^w[i+1]^w[i],
; w[i+7] == RotWord(SubWord(w[i+5]))^Rcon^w[i+1]^w[i] and
; w[i+6] == RotWord(SubWord(w[i+5]))^Rcon^w[i].
- pshufd xmm6, xmm7, 55h ; xmm6[127:96] = xmm6[95:64] = xmm6[63:32] = xmm6[31:0] = xmm7[63:32]
- pxor xmm0, xmm6 ; xmm0 ^= xmm6
+ pshufd xmm4, xmm5, 55h ; xmm4[127:96] = xmm4[95:64] = xmm4[63:32] = xmm4[31:0] = xmm5[63:32]
+ pxor xmm0, xmm4 ; xmm0 ^= xmm4
; xmm0[127:96] == w[i+9] == RotWord(SubWord(w[i+5]))^Rcon^w[i+3]^w[i+2]^w[i+1]^w[i]
; xmm0[95:64] == w[i+8] == RotWord(SubWord(w[i+5]))^Rcon^w[i+2]^w[i+1]^w[i]
; xmm0[63:32] == w[i+7] == RotWord(SubWord(w[i+5]))^Rcon^w[i+1]^w[i]
@@ -191,17 +191,17 @@ aes192_keygen_assist:
; Calculate
; w[i+5]^w[i+4],
; w[i+4].
- pshufd xmm6, xmm1, 0F3h ; xmm6 = xmm1[31:0] << 32
- pxor xmm1, xmm6 ; xmm1 ^= xmm7
+ pshufd xmm4, xmm1, 0F3h ; xmm4 = xmm1[31:0] << 32
+ pxor xmm1, xmm4 ; xmm1 ^= xmm5
; xmm1[63:32] == w[i+5]^w[i+4]
; xmm1[31:0] == w[i+4]
; Calculate
; w[i+10] == RotWord(SubWord(w[i+5]))^Rcon^w[i+5]^w[i+4]^w[i+3]^w[i+2]^w[i+1]^w[i],
; w[i+11] == RotWord(SubWord(w[i+5]))^Rcon^w[i+4]^w[i+3]^w[i+2]^w[i+1]^w[i].
- pshufd xmm6, xmm0, 0FFh ; xmm6[127:96] = xmm6[95:64] = xmm6[63:32] = xmm6[31:0] = xmm0[127:96]
- psrldq xmm6, 8 ; xmm6 >>= 64
- pxor xmm1, xmm6 ; xmm1 ^= xmm6
+ pshufd xmm4, xmm0, 0FFh ; xmm4[127:96] = xmm4[95:64] = xmm4[63:32] = xmm4[31:0] = xmm0[127:96]
+ psrldq xmm4, 8 ; xmm4 >>= 64
+ pxor xmm1, xmm4 ; xmm1 ^= xmm4
; xmm1[63:32] == w[i+11] == RotWord(SubWord(w[i+5]))^Rcon^w[i+5]^w[i+4]^w[i+3]^w[i+2]^w[i+1]^w[i]
; xmm1[31:0] == w[i+10] == RotWord(SubWord(w[i+5]))^Rcon^w[i+4]^w[i+3]^w[i+2]^w[i+1]^w[i]
@@ -209,38 +209,38 @@ aes192_keygen_assist:
@raw_aes192_expand_key_schedule@36 endp
@raw_aes192_invert_key_schedule@8 proc
- movdqa xmm7, [ecx]
- movdqa xmm6, [ecx + 0C0h]
- movdqa [edx], xmm6
- movdqa [edx + 0C0h], xmm7
-
- aesimc xmm7, [ecx + 10h]
- aesimc xmm6, [ecx + 0B0h]
- movdqa [edx + 10h], xmm6
- movdqa [edx + 0B0h], xmm7
-
- aesimc xmm7, [ecx + 20h]
- aesimc xmm6, [ecx + 0A0h]
- movdqa [edx + 20h], xmm6
- movdqa [edx + 0A0h], xmm7
-
- aesimc xmm7, [ecx + 30h]
- aesimc xmm6, [ecx + 90h]
- movdqa [edx + 30h], xmm6
- movdqa [edx + 90h], xmm7
-
- aesimc xmm7, [ecx + 40h]
- aesimc xmm6, [ecx + 80h]
- movdqa [edx + 40h], xmm6
- movdqa [edx + 80h], xmm7
-
- aesimc xmm7, [ecx + 50h]
- aesimc xmm6, [ecx + 70h]
- movdqa [edx + 50h], xmm6
- movdqa [edx + 70h], xmm7
-
- aesimc xmm7, [ecx + 60h]
- movdqa [edx + 60h], xmm7
+ movdqa xmm5, [ecx]
+ movdqa xmm4, [ecx + 0C0h]
+ movdqa [edx], xmm4
+ movdqa [edx + 0C0h], xmm5
+
+ aesimc xmm5, [ecx + 10h]
+ aesimc xmm4, [ecx + 0B0h]
+ movdqa [edx + 10h], xmm4
+ movdqa [edx + 0B0h], xmm5
+
+ aesimc xmm5, [ecx + 20h]
+ aesimc xmm4, [ecx + 0A0h]
+ movdqa [edx + 20h], xmm4
+ movdqa [edx + 0A0h], xmm5
+
+ aesimc xmm5, [ecx + 30h]
+ aesimc xmm4, [ecx + 90h]
+ movdqa [edx + 30h], xmm4
+ movdqa [edx + 90h], xmm5
+
+ aesimc xmm5, [ecx + 40h]
+ aesimc xmm4, [ecx + 80h]
+ movdqa [edx + 40h], xmm4
+ movdqa [edx + 80h], xmm5
+
+ aesimc xmm5, [ecx + 50h]
+ aesimc xmm4, [ecx + 70h]
+ movdqa [edx + 50h], xmm4
+ movdqa [edx + 70h], xmm5
+
+ aesimc xmm5, [ecx + 60h]
+ movdqa [edx + 60h], xmm5
ret
@raw_aes192_invert_key_schedule@8 endp
diff --git a/src/asm/aes256.asm b/src/asm/aes256.asm
index 413e67b..479b0f5 100644
--- a/src/asm/aes256.asm
+++ b/src/asm/aes256.asm
@@ -117,56 +117,56 @@
movdqa [ecx + 10h], xmm1 ; sets w[4], w[5], w[6], w[7]
lea ecx, [ecx + 20h] ; ecx = &w[8]
- aeskeygenassist xmm7, xmm1, 1h ; xmm7[127:96] = RotWord(SubWord(w[7]))^Rcon
- pshufd xmm7, xmm7, 0FFh ; xmm7[95:64] = xmm7[63:32] = xmm7[31:0] = xmm7[127:96]
+ aeskeygenassist xmm5, xmm1, 1h ; xmm5[127:96] = RotWord(SubWord(w[7]))^Rcon
+ pshufd xmm5, xmm5, 0FFh ; xmm5[95:64] = xmm5[63:32] = xmm5[31:0] = xmm5[127:96]
call aes256_keygen_assist ; sets w[8], w[9], w[10], w[11]
- aeskeygenassist xmm7, xmm1, 0 ; xmm7[95:64] = SubWord(w[11])
- pshufd xmm7, xmm7, 0AAh ; xmm7[127:96] = xmm7[63:32] = xmm7[31:0] = xmm7[95:64]
+ aeskeygenassist xmm5, xmm1, 0 ; xmm5[95:64] = SubWord(w[11])
+ pshufd xmm5, xmm5, 0AAh ; xmm5[127:96] = xmm5[63:32] = xmm5[31:0] = xmm5[95:64]
call aes256_keygen_assist ; sets w[12], w[13], w[14], w[15]
- aeskeygenassist xmm7, xmm1, 2h ; xmm7[127:96] = RotWord(SubWord(w[15]))^Rcon
- pshufd xmm7, xmm7, 0FFh ; xmm7[95:64] = xmm7[63:32] = xmm7[31:0] = xmm7[127:96]
+ aeskeygenassist xmm5, xmm1, 2h ; xmm5[127:96] = RotWord(SubWord(w[15]))^Rcon
+ pshufd xmm5, xmm5, 0FFh ; xmm5[95:64] = xmm5[63:32] = xmm5[31:0] = xmm5[127:96]
call aes256_keygen_assist ; sets w[16], w[17], w[18], w[19]
- aeskeygenassist xmm7, xmm1, 0 ; xmm7[95:64] = SubWord(w[19])
- pshufd xmm7, xmm7, 0AAh ; xmm7[127:96] = xmm7[63:32] = xmm7[31:0] = xmm7[95:64]
+ aeskeygenassist xmm5, xmm1, 0 ; xmm5[95:64] = SubWord(w[19])
+ pshufd xmm5, xmm5, 0AAh ; xmm5[127:96] = xmm5[63:32] = xmm5[31:0] = xmm5[95:64]
call aes256_keygen_assist ; sets w[20], w[21], w[22], w[23]
- aeskeygenassist xmm7, xmm1, 4h ; xmm7[127:96] = RotWord(SubWord(w[23]))^Rcon
- pshufd xmm7, xmm7, 0FFh ; xmm7[95:64] = xmm7[63:32] = xmm7[31:0] = xmm7[127:96]
+ aeskeygenassist xmm5, xmm1, 4h ; xmm5[127:96] = RotWord(SubWord(w[23]))^Rcon
+ pshufd xmm5, xmm5, 0FFh ; xmm5[95:64] = xmm5[63:32] = xmm5[31:0] = xmm5[127:96]
call aes256_keygen_assist ; sets w[24], w[25], w[26], w[27]
- aeskeygenassist xmm7, xmm1, 0 ; xmm7[95:64] = SubWord(w[27])
- pshufd xmm7, xmm7, 0AAh ; xmm7[127:96] = xmm7[63:32] = xmm7[31:0] = xmm7[95:64]
+ aeskeygenassist xmm5, xmm1, 0 ; xmm5[95:64] = SubWord(w[27])
+ pshufd xmm5, xmm5, 0AAh ; xmm5[127:96] = xmm5[63:32] = xmm5[31:0] = xmm5[95:64]
call aes256_keygen_assist ; sets w[28], w[29], w[30], w[31]
- aeskeygenassist xmm7, xmm1, 8h ; xmm7[127:96] = RotWord(SubWord(w[31]))^Rcon
- pshufd xmm7, xmm7, 0FFh ; xmm7[95:64] = xmm7[63:32] = xmm7[31:0] = xmm7[127:96]
+ aeskeygenassist xmm5, xmm1, 8h ; xmm5[127:96] = RotWord(SubWord(w[31]))^Rcon
+ pshufd xmm5, xmm5, 0FFh ; xmm5[95:64] = xmm5[63:32] = xmm5[31:0] = xmm5[127:96]
call aes256_keygen_assist ; sets w[32], w[33], w[34], w[35]
- aeskeygenassist xmm7, xmm1, 0 ; xmm7[95:64] = SubWord(w[35])
- pshufd xmm7, xmm7, 0AAh ; xmm7[127:96] = xmm7[63:32] = xmm7[31:0] = xmm7[95:64]
+ aeskeygenassist xmm5, xmm1, 0 ; xmm5[95:64] = SubWord(w[35])
+ pshufd xmm5, xmm5, 0AAh ; xmm5[127:96] = xmm5[63:32] = xmm5[31:0] = xmm5[95:64]
call aes256_keygen_assist ; sets w[36], w[37], w[38], w[39]
- aeskeygenassist xmm7, xmm1, 10h ; xmm7[127:96] = RotWord(SubWord(w[39]))^Rcon
- pshufd xmm7, xmm7, 0FFh ; xmm7[95:64] = xmm7[63:32] = xmm7[31:0] = xmm7[127:96]
+ aeskeygenassist xmm5, xmm1, 10h ; xmm5[127:96] = RotWord(SubWord(w[39]))^Rcon
+ pshufd xmm5, xmm5, 0FFh ; xmm5[95:64] = xmm5[63:32] = xmm5[31:0] = xmm5[127:96]
call aes256_keygen_assist ; sets w[40], w[41], w[42], w[43]
- aeskeygenassist xmm7, xmm1, 0 ; xmm7[95:64] = SubWord(w[43])
- pshufd xmm7, xmm7, 0AAh ; xmm7[127:96] = xmm7[63:32] = xmm7[31:0] = xmm7[95:64]
+ aeskeygenassist xmm5, xmm1, 0 ; xmm5[95:64] = SubWord(w[43])
+ pshufd xmm5, xmm5, 0AAh ; xmm5[127:96] = xmm5[63:32] = xmm5[31:0] = xmm5[95:64]
call aes256_keygen_assist ; sets w[44], w[45], w[46], w[47]
- aeskeygenassist xmm7, xmm1, 20h ; xmm7[127:96] = RotWord(SubWord(w[47]))^Rcon
- pshufd xmm7, xmm7, 0FFh ; xmm7[95:64] = xmm7[63:32] = xmm7[31:0] = xmm7[127:96]
+ aeskeygenassist xmm5, xmm1, 20h ; xmm5[127:96] = RotWord(SubWord(w[47]))^Rcon
+ pshufd xmm5, xmm5, 0FFh ; xmm5[95:64] = xmm5[63:32] = xmm5[31:0] = xmm5[127:96]
call aes256_keygen_assist ; sets w[48], w[49], w[50], w[51]
- aeskeygenassist xmm7, xmm1, 0 ; xmm7[95:64] = SubWord(w[51])
- pshufd xmm7, xmm7, 0AAh ; xmm7[127:96] = xmm7[63:32] = xmm7[31:0] = xmm7[95:64]
+ aeskeygenassist xmm5, xmm1, 0 ; xmm5[95:64] = SubWord(w[51])
+ pshufd xmm5, xmm5, 0AAh ; xmm5[127:96] = xmm5[63:32] = xmm5[31:0] = xmm5[95:64]
call aes256_keygen_assist ; sets w[52], w[53], w[54], w[55]
- aeskeygenassist xmm7, xmm1, 40h ; xmm7[127:96] = RotWord(SubWord(w[55]))^Rcon
- pshufd xmm7, xmm7, 0FFh ; xmm7[95:64] = xmm7[63:32] = xmm7[31:0] = xmm7[127:96]
+ aeskeygenassist xmm5, xmm1, 40h ; xmm5[127:96] = RotWord(SubWord(w[55]))^Rcon
+ pshufd xmm5, xmm5, 0FFh ; xmm5[95:64] = xmm5[63:32] = xmm5[31:0] = xmm5[127:96]
call aes256_keygen_assist ; sets w[56], w[57], w[58], w[59]
ret
@@ -181,7 +181,7 @@ aes256_keygen_assist:
; * xmm0[95:64] == w[i+2],
; * xmm0[63:32] == w[i+1],
; * xmm0[31:0] == w[i],
- ; * xmm7[127:96] == xmm7[95:64] == xmm7[63:32] == xmm7[31:0] == HWGEN,
+ ; * xmm5[127:96] == xmm5[95:64] == xmm5[63:32] == xmm5[31:0] == HWGEN,
; where HWGEN is either RotWord(SubWord(w[i+7]))^Rcon or SubWord(w[i+7]),
; depending on the number of the round being processed,
; * ecx == &w[i+8].
@@ -196,20 +196,20 @@ aes256_keygen_assist:
; * xmm0[63:32] == w[i+5],
; * xmm0[31:0] == w[i+4],
; * ecx == &w[i+12],
- ; * the value in xmm6 is also modified.
+ ; * the value in xmm4 is also modified.
; Calculate
; w[i+3]^w[i+2]^w[i+1]^w[i],
; w[i+2]^w[i+1]^w[i],
; w[i+1]^w[i] and
; w[i].
- movdqa xmm6, xmm0 ; xmm6 = xmm0
- pslldq xmm6, 4 ; xmm6 <<= 32
- pxor xmm0, xmm6 ; xmm0 ^= xmm6
- pslldq xmm6, 4 ; xmm6 <<= 32
- pxor xmm0, xmm6 ; xmm0 ^= xmm6
- pslldq xmm6, 4 ; xmm6 <<= 32
- pxor xmm0, xmm6 ; xmm0 ^= xmm6
+ movdqa xmm4, xmm0 ; xmm4 = xmm0
+ pslldq xmm4, 4 ; xmm4 <<= 32
+ pxor xmm0, xmm4 ; xmm0 ^= xmm4
+ pslldq xmm4, 4 ; xmm4 <<= 32
+ pxor xmm0, xmm4 ; xmm0 ^= xmm4
+ pslldq xmm4, 4 ; xmm4 <<= 32
+ pxor xmm0, xmm4 ; xmm0 ^= xmm4
; xmm0[127:96] == w[i+3]^w[i+2]^w[i+1]^w[i]
; xmm0[95:64] == w[i+2]^w[i+1]^w[i]
; xmm0[63:32] == w[i+1]^w[i]
@@ -220,7 +220,7 @@ aes256_keygen_assist:
; HWGEN^w[i+2]^w[i+1]^w[i],
; HWGEN^w[i+1]^w[i] and
; HWGEN^w[i].
- pxor xmm0, xmm7 ; xmm0 ^= xmm7
+ pxor xmm0, xmm5 ; xmm0 ^= xmm5
; xmm0[127:96] == w[i+11] == HWGEN^w[i+3]^w[i+2]^w[i+1]^w[i]
; xmm0[95:64] == w[i+10] == HWGEN^w[i+2]^w[i+1]^w[i]
; xmm0[63:32] == w[i+9] == HWGEN^w[i+1]^w[i]
@@ -242,43 +242,43 @@ aes256_keygen_assist:
@raw_aes256_expand_key_schedule@36 endp
@raw_aes256_invert_key_schedule@8 proc
- movdqa xmm7, [ecx]
- movdqa xmm6, [ecx + 0E0h]
- movdqa [edx], xmm6
- movdqa [edx + 0E0h], xmm7
-
- aesimc xmm7, [ecx + 10h]
- aesimc xmm6, [ecx + 0D0h]
- movdqa [edx + 10h], xmm6
- movdqa [edx + 0D0h], xmm7
-
- aesimc xmm7, [ecx + 20h]
- aesimc xmm6, [ecx + 0C0h]
- movdqa [edx + 20h], xmm6
- movdqa [edx + 0C0h], xmm7
-
- aesimc xmm7, [ecx + 30h]
- aesimc xmm6, [ecx + 0B0h]
- movdqa [edx + 30h], xmm6
- movdqa [edx + 0B0h], xmm7
-
- aesimc xmm7, [ecx + 40h]
- aesimc xmm6, [ecx + 0A0h]
- movdqa [edx + 40h], xmm6
- movdqa [edx + 0A0h], xmm7
-
- aesimc xmm7, [ecx + 50h]
- aesimc xmm6, [ecx + 90h]
- movdqa [edx + 50h], xmm6
- movdqa [edx + 90h], xmm7
-
- aesimc xmm7, [ecx + 60h]
- aesimc xmm6, [ecx + 80h]
- movdqa [edx + 60h], xmm6
- movdqa [edx + 80h], xmm7
-
- aesimc xmm7, [ecx + 70h]
- movdqa [edx + 70h], xmm7
+ movdqa xmm5, [ecx]
+ movdqa xmm4, [ecx + 0E0h]
+ movdqa [edx], xmm4
+ movdqa [edx + 0E0h], xmm5
+
+ aesimc xmm5, [ecx + 10h]
+ aesimc xmm4, [ecx + 0D0h]
+ movdqa [edx + 10h], xmm4
+ movdqa [edx + 0D0h], xmm5
+
+ aesimc xmm5, [ecx + 20h]
+ aesimc xmm4, [ecx + 0C0h]
+ movdqa [edx + 20h], xmm4
+ movdqa [edx + 0C0h], xmm5
+
+ aesimc xmm5, [ecx + 30h]
+ aesimc xmm4, [ecx + 0B0h]
+ movdqa [edx + 30h], xmm4
+ movdqa [edx + 0B0h], xmm5
+
+ aesimc xmm5, [ecx + 40h]
+ aesimc xmm4, [ecx + 0A0h]
+ movdqa [edx + 40h], xmm4
+ movdqa [edx + 0A0h], xmm5
+
+ aesimc xmm5, [ecx + 50h]
+ aesimc xmm4, [ecx + 90h]
+ movdqa [edx + 50h], xmm4
+ movdqa [edx + 90h], xmm5
+
+ aesimc xmm5, [ecx + 60h]
+ aesimc xmm4, [ecx + 80h]
+ movdqa [edx + 60h], xmm4
+ movdqa [edx + 80h], xmm5
+
+ aesimc xmm5, [ecx + 70h]
+ movdqa [edx + 70h], xmm5
ret
@raw_aes256_invert_key_schedule@8 endp