forked from TrueCloudLab/tzhash
*: format assembly code with asmfmt
Signed-off-by: Evgenii Stratonikov <evgeniy@nspcc.ru>
This commit is contained in:
parent
1520cde665
commit
0fa6b1314e
4 changed files with 251 additions and 252 deletions
|
@ -56,11 +56,11 @@ TEXT ·Mul(SB),NOSPLIT,$0
|
||||||
PCLMULQDQ $0x10, X3, X3 // X3 = (a0 + a1) * (b0 + b1)
|
PCLMULQDQ $0x10, X3, X3 // X3 = (a0 + a1) * (b0 + b1)
|
||||||
VPCLMULQDQ $0x00, X0, X1, X4 // X4 = a0 * b0
|
VPCLMULQDQ $0x00, X0, X1, X4 // X4 = a0 * b0
|
||||||
VPCLMULQDQ $0x11, X0, X1, X5 // X5 = a1 * b1
|
VPCLMULQDQ $0x11, X0, X1, X5 // X5 = a1 * b1
|
||||||
XORPD X4, X3 //
|
XORPD X4, X3
|
||||||
XORPD X5, X3 // X3 = a0 * b1 + a1 * b0
|
XORPD X5, X3 // X3 = a0 * b1 + a1 * b0
|
||||||
VPSLLDQ $8, X3, X2 //
|
VPSLLDQ $8, X3, X2
|
||||||
XORPD X2, X4 // X4 = a0 * b0 + lo(X3)
|
XORPD X2, X4 // X4 = a0 * b0 + lo(X3)
|
||||||
VPSRLDQ $8, X3, X6 //
|
VPSRLDQ $8, X3, X6
|
||||||
XORPD X6, X5 // X5 = a1 * b1 + hi(X3)
|
XORPD X6, X5 // X5 = a1 * b1 + hi(X3)
|
||||||
|
|
||||||
// at this point, a * b = X4 . X5 (as 256-bit number)
|
// at this point, a * b = X4 . X5 (as 256-bit number)
|
||||||
|
|
|
@ -16,7 +16,6 @@ TEXT ·Mul10x2(SB),NOSPLIT,$0
|
||||||
VMOVDQA Y3, (AX)
|
VMOVDQA Y3, (AX)
|
||||||
RET
|
RET
|
||||||
|
|
||||||
|
|
||||||
// func Mul11x2(a, b) *[4]uint64
|
// func Mul11x2(a, b) *[4]uint64
|
||||||
TEXT ·Mul11x2(SB), NOSPLIT, $0
|
TEXT ·Mul11x2(SB), NOSPLIT, $0
|
||||||
MOVQ a+0(FP), AX
|
MOVQ a+0(FP), AX
|
||||||
|
|
Loading…
Add table
Reference in a new issue