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.

1 files changed, 73 insertions, 73 deletions

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