# define RCON ecx // round constants table
# define LIM edx // limit pointer
# define LIMn edx // ... as integer offset from base
+# define CYIX edi // index in shift-register cycle
# define NR ecx // number of rounds
# define LRK eax // distance to last key
# define RCON rdi // round constants table
# define LIMn ecx // limit pointer
# define LIM rcx // ... as integer offset from base
+# define CYIX r11d // index in shift-register cycle
# define NR ecx // number of rounds
# define LRK eax // distance to last key
# define RCON rdi // round constants table
# define LIMn ecx // limit pointer
# define LIM rcx // ... as integer offset from base
+# define CYIX r11d // index in shift-register cycle
# define NR ecx // number of rounds
# define LRK eax // distance to last key
// We'll need the index registers, which belong to the caller in this
// ABI.
push rsi
+ .seh_pushreg rsi
push rdi
+ .seh_pushreg rdi
+ .seh_endprologue
// Move arguments to more useful places.
mov SI, r8 // key material
lea SI, [CTX + w]
mov eax, [SI + 4*KSZo - 4] // most recent key word
lea LIM, [SI + 4*LIM] // limit, offset by one key expansion
+ xor CYIX, CYIX // start of new cycle
// Main key expansion loop. The first word of each key-length chunk
// needs special treatment.
// as an immediate, so it's kind of annoying if you're not
// open-coding the whole thing. It's much easier to leave that as
// zero and XOR in the round constant by hand.
-0: movd xmm0, eax
+0: cmp CYIX, 0 // first word of the cycle?
+ je 1f
+ cmp CYIX, 4 // fourth word of the cycle?
+ jne 2f
+ cmp KSZ, 7 // and a large key?
+ jb 2f
+
+ // Fourth word of the cycle, and seven or eight words of key. Do a
+ // byte substitution.
+ movd xmm0, eax
+ pshufd xmm0, xmm0, ROTL
+ aeskeygenassist xmm1, xmm0, 0
+ movd eax, xmm1
+ jmp 2f
+
+ // First word of the cycle. This is the complicated piece.
+1: movd xmm0, eax
pshufd xmm0, xmm0, ROTR
aeskeygenassist xmm1, xmm0, 0
pshufd xmm1, xmm1, ROTL
movd eax, xmm1
- xor eax, [SI]
xor al, [RCON]
inc RCON
- mov [SI + 4*KSZo], eax
- add SI, 4
- cmp SI, LIM
- jae 9f
-
- // The next three words are simple...
- xor eax, [SI]
- mov [SI + 4*KSZo], eax
- add SI, 4
- cmp SI, LIM
- jae 9f
-
- // (Word 2...)
- xor eax, [SI]
- mov [SI + 4*KSZo], eax
- add SI, 4
- cmp SI, LIM
- jae 9f
-
- // (Word 3...)
- xor eax, [SI]
- mov [SI + 4*KSZo], eax
- add SI, 4
- cmp SI, LIM
- jae 9f
-
- // Word 4. If the key is /more/ than 6 words long, then we must
- // apply a substitution here.
- cmp KSZ, 5
- jb 0b
- cmp KSZ, 7
- jb 1f
- movd xmm0, eax
- pshufd xmm0, xmm0, ROTL
- aeskeygenassist xmm1, xmm0, 0
- movd eax, xmm1
-1: xor eax, [SI]
- mov [SI + 4*KSZo], eax
- add SI, 4
- cmp SI, LIM
- jae 9f
-
- // (Word 5...)
- cmp KSZ, 6
- jb 0b
- xor eax, [SI]
- mov [SI + 4*KSZo], eax
- add SI, 4
- cmp SI, LIM
- jae 9f
- // (Word 6...)
- cmp KSZ, 7
- jb 0b
- xor eax, [SI]
+ // Common tail. Mix in the corresponding word from the previous
+ // cycle and prepare for the next loop.
+2: xor eax, [SI]
mov [SI + 4*KSZo], eax
add SI, 4
+ inc CYIX
cmp SI, LIM
jae 9f
-
- // (Word 7...)
- cmp KSZ, 8
+ cmp CYIX, KSZ
jb 0b
- xor eax, [SI]
- mov [SI + 4*KSZo], eax
- add SI, 4
- cmp SI, LIM
- jae 9f
-
- // Must be done by now.
+ xor CYIX, CYIX
jmp 0b
// Next job is to construct the decryption keys. The keys for the
.macro encdec op, aes, koff
FUNC(rijndael_\op\()_x86ish_aesni)
- // Find the magic endianness-swapping table.
- ldgot ecx
- movdqa xmm5, [INTADDR(endswap_tab, ecx)]
-
#if CPUFAM_X86
// Arguments come in on the stack, and need to be collected. We
// don't have a shortage of registers.
# define SRC rdx
# define DST r8
# define NR eax
+ .seh_endprologue
#endif
+ // Find the magic endianness-swapping table.
+ ldgot ecx
+ movdqa xmm5, [INTADDR(endswap_tab, ecx)]
+
// Initial setup.
movdqu xmm0, [SRC]
pshufb xmm0, xmm5
0: hlt
jmp 0b
- gotaux ecx
-
///--------------------------------------------------------------------------
/// Data tables.
~mdw / catacomb / blobdiff