aboutsummaryrefslogtreecommitdiffstatshomepage
path: root/src
diff options
context:
space:
mode:
Diffstat (limited to 'src')
-rw-r--r--src/aes192.asm247
1 files changed, 155 insertions, 92 deletions
diff --git a/src/aes192.asm b/src/aes192.asm
index 9695961..9cd4cfa 100644
--- a/src/aes192.asm
+++ b/src/aes192.asm
@@ -53,107 +53,170 @@ inverted_key_schedule oword 13 dup(0)
@raw_aes192ecb_decrypt@48 endp
expand_keys_192ecb proc
- ; key = k0 k1 k2 k3 k4 k5
- ; xmm1 = k0 k1 k2 k3
- ; xmm2 = 0 0 k5 k4
-
- ; w[0] = k0 k1 k2 k3
- ; w[1] = k4 k5 - -
-
+ ; A "word" (in terms of the FIPS 187 standard) is a 32-bit block.
+ ; Words are denoted by `w[N]`.
+ ;
+ ; A key schedule is composed of 12 "regular" keys and a dumb key for
+ ; the "whitening" step.
+ ; It's stored in `key_schedule`.
+ ;
+ ; A key schedule is thus composed of 52 "words".
+ ; The FIPS standard includes an algorithm to calculate these words via
+ ; a simple loop:
+ ;
; i = 6
- ; while (i < 52):
+ ; while i < 52:
; temp = w[i - 1]
- ; if (i % 6 == 0):
- ; temp = SubWord(RotWord(w[i - 1])) * Rcon
- ; w[i] = w[i - 6] * temp
+ ; if i % 6 == 0:
+ ; temp = SubWord(RotWord(temp))^Rcon
+ ; w[i] = w[i - 6]^temp
; i = i + 1
+ ;
+ ; The loop above may be unrolled like this:
+ ;
+ ; w[6] = SubWord(RotWord(w[5]))^Rcon^w[0]
+ ; w[7] = w[6]^w[1]
+ ; = SubWord(RotWord(w[5]))^Rcon^w[0]^w[1]
+ ; w[8] = w[7]^w[2]
+ ; = SubWord(RotWord(w[5]))^Rcon^w[0]^w[1]^w[2]
+ ; w[9] = w[8]^w[3]
+ ; = SubWord(RotWord(w[5]))^Rcon^w[0]^w[1]^w[2]^w[3]
+ ; w[10] = w[9]^w[4]
+ ; = SubWord(RotWord(w[5]))^Rcon^w[0]^w[1]^w[2]^w[3]^w[4]
+ ; w[11] = w[10]^w[5]
+ ; = SubWord(RotWord(w[5]))^Rcon^w[0]^w[1]^w[2]^w[3]^w[4]^w[5]
+ ; w[12] = SubWord(RotWord(w[11]))^Rcon^w[6]
+ ; w[13] = w[12]^w[7]
+ ; = SubWord(RotWord(w[11]))^Rcon^w[6]^w[7]
+ ; w[14] = w[13]^w[8]
+ ; = SubWord(RotWord(w[11]))^Rcon^w[6]^w[7]^w[8]
+ ; w[15] = w[14]^w[9]
+ ; = SubWord(RotWord(w[11]))^Rcon^w[6]^w[7]^w[8]^w[9]
+ ; w[16] = w[15]^w[10]
+ ; = SubWord(RotWord(w[11]))^Rcon^w[6]^w[7]^w[8]^w[9]^w[10]
+ ; w[17] = w[16]^w[11]
+ ; = SubWort(RotWord(w[11]))^Rcon^w[6]^w[7]^w[8]^w[9]^w[10]^w[11]
+ ;
+ ; ... and so on.
+ ;
+ ; The Intel AES-NI instruction set facilitates calculating SubWord
+ ; and RotWord using `aeskeygenassist`, which is used in this routine.
+ ;
+ ; Preconditions:
+ ; * xmm2[63:32] == w[5],
+ ; * xmm2[31:0] == w[4],
+ ; * xmm1[127:96] == w[3],
+ ; * xmm1[95:64] == w[2],
+ ; * xmm1[63:32] == w[1],
+ ; * xmm1[31:0] == w[0].
+
+ movdqa [key_schedule], xmm1 ; sets w[0], w[1], w[2], w[3]
+ movdqa [key_schedule + 10h], xmm2 ; sets w[4], w[5]
- ; w[6] = SubWord(RotWord(w[5])) * Rcon * w[0]
- ; w[7] = w[6] * w[1]
- ; = SubWord(RotWord(w[5])) * Rcon * w[0] * w[1]
- ; w[8] = w[7] * w[2]
- ; = SubWord(RotWord(w[5])) * Rcon * w[0] * w[1] * w[2]
- ; w[9] = w[8] * w[3]
- ; = SubWord(RotWord(w[5])) * Rcon * w[0] * w[1] * w[2] * w[3]
- ; w[10] = w[9] * w[4]
- ; = SubWord(RotWord(w[5])) * Rcon * w[0] * w[1] * w[2] * w[3] * w[4]
- ; w[11] = w[10] * w[5]
- ; = SubWord(RotWord(w[5])) * Rcon * w[0] * w[1] * w[2] * w[3] * w[4] * w[5]
-
- movdqa [key_schedule], xmm1
- movdqa [key_schedule + 10h], xmm2
-
- lea ecx, [key_schedule + 18h]
- aeskeygenassist xmm7, xmm2, 1
- call gen_round_key
- aeskeygenassist xmm7, xmm2, 2
- call gen_round_key
- aeskeygenassist xmm7, xmm2, 4
- call gen_round_key
- aeskeygenassist xmm7, xmm2, 8
- call gen_round_key
- aeskeygenassist xmm7, xmm2, 10h
- call gen_round_key
- aeskeygenassist xmm7, xmm2, 20h
- call gen_round_key
- aeskeygenassist xmm7, xmm2, 40h
- call gen_round_key
- aeskeygenassist xmm7, xmm2, 80h
- call gen_round_key
+ lea ecx, [key_schedule + 18h] ; ecx = &w[6]
+ aeskeygenassist xmm7, xmm2, 1 ; xmm7[63:32] = RotWord(SubWord(w[5]))^Rcon,
+ call gen_round_key ; sets w[6], w[7], w[8], w[9], w[10], w[11]
+ aeskeygenassist xmm7, xmm2, 2 ; xmm7[63:32] = RotWord(SubWord(w[11]))^Rcon
+ call gen_round_key ; sets w[12], w[13], w[14], w[15], w[16], w[17]
+ aeskeygenassist xmm7, xmm2, 4 ; xmm7[63:32] = RotWord(SubWord(w[17]))^Rcon
+ call gen_round_key ; sets w[18], w[19], w[20], w[21], w[22], w[23]
+ aeskeygenassist xmm7, xmm2, 8 ; xmm7[63:32] = RotWord(SubWord(w[23]))^Rcon
+ call gen_round_key ; sets w[24], w[25], w[26], w[27], w[28], w[29]
+ aeskeygenassist xmm7, xmm2, 10h ; xmm7[63:32] = RotWord(SubWord(w[29]))^Rcon
+ call gen_round_key ; sets w[30], w[31], w[32], w[33], w[34], w[35]
+ aeskeygenassist xmm7, xmm2, 20h ; xmm7[63:32] = RotWord(SubWord(w[35]))^Rcon
+ call gen_round_key ; sets w[36], w[37], w[38], w[39], w[40], w[41]
+ aeskeygenassist xmm7, xmm2, 40h ; xmm7[63:32] = RotWord(SubWord(w[41]))^Rcon
+ call gen_round_key ; sets w[42], w[43], w[44], w[45], w[46], w[47]
+ aeskeygenassist xmm7, xmm2, 80h ; xmm7[63:32] = RotWord(SubWord(w[49]))^Rcon
+ call gen_round_key ; sets w[48], w[49], w[50], w[51], w[52], w[53] // FIXME
call invert_key_schedule
ret
gen_round_key:
- ; xmm1 = x3 x2 x1 x0
- ; xmm2 = - - x5 x4
- ; xmm7 = RotWord(SubWord(-)) xor Rcon
- ; SubWord(-)
- ; RotWord(SubWord(x5)) xor Rcon
- ; SubWord(x5)
- movdqa xmm6, xmm1
-
- pslldq xmm6, 4 ; xmm6 = x2 x1 x0 0
- pxor xmm1, xmm6 ; xmm1 = (x3 * x2) (x1 * x2) (x1 * x0) x0
- pslldq xmm6, 4 ; xmm6 = x1 x0 0 0
- pxor xmm1, xmm6 ; xmm1 = (x3 * x2 * x1) (x1 * x2 * x0) (x1 * x0) x0
- pslldq xmm6, 4 ; xmm6 = x0 0 0 0
- pxor xmm1, xmm6 ; xmm1 = (x3 * x2 * x1 * x0) (x1 * x2 * x0) (x1 * x0) x0
-
- pshufd xmm7, xmm7, 55h ; xmm7 = RotWord(SubWord(x5)) * Rcon
- ; RotWord(SubWord(x5)) * Rcon
- ; RotWord(SubWord(x5)) * Rcon
- ; RotWord(SubWord(x5)) * Rcon
-
- pxor xmm1, xmm7 ; xmm1 = RotWord(SubWord(x5)) * Rcon * x3 * x2 * x1 * x0
- ; RotWord(SubWord(x5)) * Rcon * x2 * x1 * x0
- ; RotWord(SubWord(x5)) * Rcon * x1 * x0
- ; RotWord(SubWord(x5)) * Rcon * x0
-
- movq qword ptr [ecx], xmm1
- add ecx, 8
-
- pshufd xmm7, xmm1, 0FFh ; xmm7 = -
- ; -
- ; RotWord(SubWord(x5)) * Rcon * x3 * x2 * x1 * x0
- ; RotWord(SubWord(x5)) * Rcon * x3 * x2 * x1 * x0
- pxor xmm7, xmm2 ; xmm7 = -
- ; -
- ; RotWord(SubWord(x5)) * Rcon * x5 * x3 * x2 * x1 * x0
- ; RotWord(SubWord(x5)) * Rcon * x4 * x3 * x2 * x1 * x0
- pslldq xmm2, 4 ; xmm2 = - k5 k4 0
- pxor xmm7, xmm2 ; xmm7 = -
- ; -
- ; RotWord(SubWord(x5)) * Rcon * x5 * x4 * x3 * x2 * x1 * x0
- ; RotWord(SubWord(x5)) * Rcon * x4 * x3 * x2 * x1 * x0
-
- movq xmm2, xmm7
- pslldq xmm7, 8
- movdqa xmm6, xmm1
- psrldq xmm6, 8
- por xmm7, xmm6
- movdqu [ecx], xmm7
- add ecx, 10h
+ ; Preconditions:
+ ; * xmm2[127:96] == 0,
+ ; * xmm2[95:64] == 0,
+ ; * xmm2[63:32] == w[i+5],
+ ; * xmm2[31:0] == w[i+4],
+ ; * xmm1[127:96] == w[i+3],
+ ; * xmm1[95:64] == w[i+2],
+ ; * xmm1[63:32] == w[i+1],
+ ; * xmm1[31:0] == w[i],
+ ; * xmm7[63:32] == RotWord(SubWord(w[i+5]))^Rcon,
+ ; * ecx == &w[i+6].
+ ;
+ ; Postconditions:
+ ; * xmm2[127:96] == 0,
+ ; * xmm2[95:64] == 0,
+ ; * xmm2[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],
+ ; * xmm2[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],
+ ; * xmm1[127:96] == w[i+9] == RotWord(SubWord(w[i+5]))^Rcon^w[i+3]^w[i+2]^w[i+1]^w[i],
+ ; * xmm1[95:64] == w[i+8] == RotWord(SubWord(w[i+5]))^Rcon^w[i+2]^w[i+1]^w[i],
+ ; * xmm1[63:32] == w[i+7] == RotWord(SubWord(w[i+5]))^Rcon^w[i+1]^w[i],
+ ; * xmm1[31:0] == w[i+6] == RotWord(SubWord(w[i+5]))^Rcon^w[i],
+ ; * ecx == &w[i+12],
+ ; * the value in xmm6 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, xmm1 ; xmm6 = xmm1
+ pslldq xmm6, 4 ; xmm6 <<= 4
+ pxor xmm1, xmm6 ; xmm1 ^= xmm6
+ pslldq xmm6, 4 ; xmm6 <<= 4
+ pxor xmm1, xmm6 ; xmm1 ^= xmm6
+ pslldq xmm6, 4 ; xmm6 <<= 4
+ pxor xmm1, xmm6 ; xmm1 ^= xmm6
+ ; xmm1[127:96] == w[i+3]^w[i+2]^w[i+1]^w[i]
+ ; xmm1[95:64] == w[i+2]^w[i+1]^w[i]
+ ; xmm1[63:32] == w[i+1]^w[i]
+ ; xmm1[31:0] == w[i]
+
+ ; Calculate
+ ; w[i+9] == RotWord(SubWord(w[i+5]))^Rcon^w[i+3]^w[i+2]^w[i+1]^w[i],
+ ; 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 xmm1, xmm6 ; xmm1 ^= xmm6
+ ; xmm1[127:96] == w[i+9] == RotWord(SubWord(w[i+5]))^Rcon^w[i+3]^w[i+2]^w[i+1]^w[i]
+ ; xmm1[95:64] == w[i+8] == RotWord(SubWord(w[i+5]))^Rcon^w[i+2]^w[i+1]^w[i]
+ ; xmm1[63:32] == w[i+7] == RotWord(SubWord(w[i+5]))^Rcon^w[i+1]^w[i]
+ ; xmm1[31:0] == w[i+6] == RotWord(SubWord(w[i+5]))^Rcon^w[i]
+
+ ; Set w[i+6], w[i+7], w[i+8] and w[i+9].
+ movdqu [ecx], xmm1 ; w[i+6] = RotWord(SubWord(w[i+5]))^Rcon^w[i]
+ ; w[i+7] = RotWord(SubWord(w[i+5]))^Rcon^w[i+1]^w[i]
+ ; w[i+8] = RotWord(SubWord(w[i+5]))^Rcon^w[i+2]^w[i+1]^w[i]
+ ; w[i+9] = RotWord(SubWord(w[i+5]))^Rcon^w[i+3]^w[i+2]^w[i+1]^w[i]
+ add ecx, 10h ; ecx = &w[i+10]
+
+ ; Calculate
+ ; w[i+5]^w[i+4],
+ ; w[i+4].
+ pshufd xmm6, xmm2, 0F3h ; xmm6 = xmm2[31:0] << 4
+ pxor xmm2, xmm6 ; xmm2 ^= xmm7
+ ; xmm2[63:32] == w[i+5]^w[i+4]
+ ; xmm2[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, xmm1, 0FFh ; xmm6[127:96] = xmm6[95:64] = xmm6[63:32] = xmm6[31:0] = xmm1[127:96]
+ psrldq xmm6, 8 ; xmm6 >>= 8
+ pxor xmm2, xmm6 ; xmm2 ^= xmm6
+ ; xmm2[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]
+ ; xmm2[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]
+
+ ; Set w[i+10] and w[i+11].
+ movq qword ptr [ecx], xmm2 ; w[i+10] = RotWord(SubWord(w[i+5]))^Rcon^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+5]^w[i+4]^w[i+3]^w[i+2]^w[i+1]^w[i]
+ add ecx, 8 ; ecx = &w[i+12]
+
ret
invert_key_schedule:
generated by cgit v1.2.3 (git 2.25.1) at 2024-12-02 01:30:48 +0000