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, 48 insertions, 48 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