diff options
| -rw-r--r-- | src/Crypto/sha512-x64-nayuki.S | 33 |
1 files changed, 19 insertions, 14 deletions
diff --git a/src/Crypto/sha512-x64-nayuki.S b/src/Crypto/sha512-x64-nayuki.S index 167b31f5..96ffb1a3 100644 --- a/src/Crypto/sha512-x64-nayuki.S +++ b/src/Crypto/sha512-x64-nayuki.S @@ -22,8 +22,9 @@ */ # Adapted for VeraCrypt -# Adapt to Windows calling convention when building on Windows. -# avoid using xmm6 register since it must be preserved on Windows. We use MMX registers instead. +# Adapt to Windows build: +# - specific calling convention +# - avoid using xmm6 register since it must be preserved. We use the stack to save RBX, RDI and RSI /* void sha512_compress_nayuki(uint64_t state[8], const uint8_t block[128]) */ @@ -64,15 +65,19 @@ _sha512_compress_nayuki: movq %r13, %xmm3 movq %r14, %xmm4 movq %r15, %xmm5 - movq %rbx, %mm0 + .ifdef WINABI - movq %rdi, %mm1 - movq %rsi, %mm2 + subq $152, %rsp + movq %rbx, (0*8 + 128)(%rsp) + movq %rdi, (1*8 + 128)(%rsp) + movq %rsi, (2*8 + 128)(%rsp) movq %rcx, %rdi movq %rdx, %rsi -.endif + .else + movq %rbx, %xmm6 subq $128, %rsp - +.endif + movq 0(%rdi), %r8 movq 8(%rdi), %r9 @@ -182,16 +187,16 @@ _sha512_compress_nayuki: movq %xmm3, %r13 movq %xmm4, %r14 movq %xmm5, %r15 - movq %mm0, %rbx .ifdef WINABI - movq %mm1, %rdi - movq %mm2, %rsi -.endif - - emms - + movq (0*8 + 128)(%rsp), %rbx + movq (1*8 + 128)(%rsp), %rdi + movq (2*8 + 128)(%rsp), %rsi + addq $152, %rsp + .else + movq %xmm6, %rbx addq $128, %rsp +.endif retq |