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, 56 insertions, 56 deletions

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