curve25519: eliminate unnecessary "callee save" prologues
The assembly implementations of ladderstep and mul contain register
save prologues that are unnecessary in Go because there are no callee
save registers in the Go ABI. Remove these prologues, update all SP
offsets, and reduce the frame size accordingly.
The SP offsets were updated with:
python -c 'import sys, re; sys.stdout.write(re.sub(r"(\d+)\(SP\)", lambda m: "%d(SP)" % (int(m.group(1))-YYY), sys.stdin.read()))'
where YYY was 64 for mul_amd64.s and 56 for ladderstep_amd64.s.
Change-Id: I728948809f479b1c061cc65167dadad651efab31
Reviewed-on: https://go-review.googlesource.com/31580
Reviewed-by: Adam Langley <agl@golang.org>
Reviewed-by: Minux Ma <minux@golang.org>
diff --git a/curve25519/ladderstep_amd64.s b/curve25519/ladderstep_amd64.s
index af17c0a..ee7b36c 100644
--- a/curve25519/ladderstep_amd64.s
+++ b/curve25519/ladderstep_amd64.s
@@ -8,16 +8,9 @@
// +build amd64,!gccgo,!appengine
// func ladderstep(inout *[5][5]uint64)
-TEXT ·ladderstep(SB),0,$352-8
+TEXT ·ladderstep(SB),0,$296-8
MOVQ inout+0(FP),DI
- MOVQ R11,0(SP)
- MOVQ R12,8(SP)
- MOVQ R13,16(SP)
- MOVQ R14,24(SP)
- MOVQ R15,32(SP)
- MOVQ BX,40(SP)
- MOVQ BP,48(SP)
MOVQ 40(DI),SI
MOVQ 48(DI),DX
MOVQ 56(DI),CX
@@ -43,201 +36,86 @@
SUBQ 96(DI),R11
SUBQ 104(DI),R12
SUBQ 112(DI),R13
- MOVQ SI,56(SP)
- MOVQ DX,64(SP)
- MOVQ CX,72(SP)
- MOVQ R8,80(SP)
- MOVQ R9,88(SP)
- MOVQ AX,96(SP)
- MOVQ R10,104(SP)
- MOVQ R11,112(SP)
- MOVQ R12,120(SP)
- MOVQ R13,128(SP)
- MOVQ 96(SP),AX
- MULQ 96(SP)
+ MOVQ SI,0(SP)
+ MOVQ DX,8(SP)
+ MOVQ CX,16(SP)
+ MOVQ R8,24(SP)
+ MOVQ R9,32(SP)
+ MOVQ AX,40(SP)
+ MOVQ R10,48(SP)
+ MOVQ R11,56(SP)
+ MOVQ R12,64(SP)
+ MOVQ R13,72(SP)
+ MOVQ 40(SP),AX
+ MULQ 40(SP)
MOVQ AX,SI
MOVQ DX,CX
- MOVQ 96(SP),AX
+ MOVQ 40(SP),AX
SHLQ $1,AX
- MULQ 104(SP)
+ MULQ 48(SP)
MOVQ AX,R8
MOVQ DX,R9
- MOVQ 96(SP),AX
+ MOVQ 40(SP),AX
SHLQ $1,AX
- MULQ 112(SP)
+ MULQ 56(SP)
MOVQ AX,R10
MOVQ DX,R11
- MOVQ 96(SP),AX
+ MOVQ 40(SP),AX
SHLQ $1,AX
- MULQ 120(SP)
+ MULQ 64(SP)
MOVQ AX,R12
MOVQ DX,R13
- MOVQ 96(SP),AX
+ MOVQ 40(SP),AX
SHLQ $1,AX
- MULQ 128(SP)
+ MULQ 72(SP)
MOVQ AX,R14
MOVQ DX,R15
- MOVQ 104(SP),AX
- MULQ 104(SP)
+ MOVQ 48(SP),AX
+ MULQ 48(SP)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 104(SP),AX
+ MOVQ 48(SP),AX
SHLQ $1,AX
- MULQ 112(SP)
+ MULQ 56(SP)
ADDQ AX,R12
ADCQ DX,R13
- MOVQ 104(SP),AX
+ MOVQ 48(SP),AX
SHLQ $1,AX
- MULQ 120(SP)
+ MULQ 64(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 104(SP),DX
+ MOVQ 48(SP),DX
IMUL3Q $38,DX,AX
- MULQ 128(SP)
+ MULQ 72(SP)
ADDQ AX,SI
ADCQ DX,CX
- MOVQ 112(SP),AX
- MULQ 112(SP)
- ADDQ AX,R14
- ADCQ DX,R15
- MOVQ 112(SP),DX
- IMUL3Q $38,DX,AX
- MULQ 120(SP)
- ADDQ AX,SI
- ADCQ DX,CX
- MOVQ 112(SP),DX
- IMUL3Q $38,DX,AX
- MULQ 128(SP)
- ADDQ AX,R8
- ADCQ DX,R9
- MOVQ 120(SP),DX
- IMUL3Q $19,DX,AX
- MULQ 120(SP)
- ADDQ AX,R8
- ADCQ DX,R9
- MOVQ 120(SP),DX
- IMUL3Q $38,DX,AX
- MULQ 128(SP)
- ADDQ AX,R10
- ADCQ DX,R11
- MOVQ 128(SP),DX
- IMUL3Q $19,DX,AX
- MULQ 128(SP)
- ADDQ AX,R12
- ADCQ DX,R13
- MOVQ ·REDMASK51(SB),DX
- SHLQ $13,CX:SI
- ANDQ DX,SI
- SHLQ $13,R9:R8
- ANDQ DX,R8
- ADDQ CX,R8
- SHLQ $13,R11:R10
- ANDQ DX,R10
- ADDQ R9,R10
- SHLQ $13,R13:R12
- ANDQ DX,R12
- ADDQ R11,R12
- SHLQ $13,R15:R14
- ANDQ DX,R14
- ADDQ R13,R14
- IMUL3Q $19,R15,CX
- ADDQ CX,SI
- MOVQ SI,CX
- SHRQ $51,CX
- ADDQ R8,CX
- ANDQ DX,SI
- MOVQ CX,R8
- SHRQ $51,CX
- ADDQ R10,CX
- ANDQ DX,R8
- MOVQ CX,R9
- SHRQ $51,CX
- ADDQ R12,CX
- ANDQ DX,R9
- MOVQ CX,AX
- SHRQ $51,CX
- ADDQ R14,CX
- ANDQ DX,AX
- MOVQ CX,R10
- SHRQ $51,CX
- IMUL3Q $19,CX,CX
- ADDQ CX,SI
- ANDQ DX,R10
- MOVQ SI,136(SP)
- MOVQ R8,144(SP)
- MOVQ R9,152(SP)
- MOVQ AX,160(SP)
- MOVQ R10,168(SP)
MOVQ 56(SP),AX
MULQ 56(SP)
- MOVQ AX,SI
- MOVQ DX,CX
- MOVQ 56(SP),AX
- SHLQ $1,AX
- MULQ 64(SP)
- MOVQ AX,R8
- MOVQ DX,R9
- MOVQ 56(SP),AX
- SHLQ $1,AX
- MULQ 72(SP)
- MOVQ AX,R10
- MOVQ DX,R11
- MOVQ 56(SP),AX
- SHLQ $1,AX
- MULQ 80(SP)
- MOVQ AX,R12
- MOVQ DX,R13
- MOVQ 56(SP),AX
- SHLQ $1,AX
- MULQ 88(SP)
- MOVQ AX,R14
- MOVQ DX,R15
- MOVQ 64(SP),AX
- MULQ 64(SP)
- ADDQ AX,R10
- ADCQ DX,R11
- MOVQ 64(SP),AX
- SHLQ $1,AX
- MULQ 72(SP)
- ADDQ AX,R12
- ADCQ DX,R13
- MOVQ 64(SP),AX
- SHLQ $1,AX
- MULQ 80(SP)
ADDQ AX,R14
ADCQ DX,R15
+ MOVQ 56(SP),DX
+ IMUL3Q $38,DX,AX
+ MULQ 64(SP)
+ ADDQ AX,SI
+ ADCQ DX,CX
+ MOVQ 56(SP),DX
+ IMUL3Q $38,DX,AX
+ MULQ 72(SP)
+ ADDQ AX,R8
+ ADCQ DX,R9
+ MOVQ 64(SP),DX
+ IMUL3Q $19,DX,AX
+ MULQ 64(SP)
+ ADDQ AX,R8
+ ADCQ DX,R9
MOVQ 64(SP),DX
IMUL3Q $38,DX,AX
- MULQ 88(SP)
- ADDQ AX,SI
- ADCQ DX,CX
- MOVQ 72(SP),AX
MULQ 72(SP)
- ADDQ AX,R14
- ADCQ DX,R15
- MOVQ 72(SP),DX
- IMUL3Q $38,DX,AX
- MULQ 80(SP)
- ADDQ AX,SI
- ADCQ DX,CX
- MOVQ 72(SP),DX
- IMUL3Q $38,DX,AX
- MULQ 88(SP)
- ADDQ AX,R8
- ADCQ DX,R9
- MOVQ 80(SP),DX
- IMUL3Q $19,DX,AX
- MULQ 80(SP)
- ADDQ AX,R8
- ADCQ DX,R9
- MOVQ 80(SP),DX
- IMUL3Q $38,DX,AX
- MULQ 88(SP)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 88(SP),DX
+ MOVQ 72(SP),DX
IMUL3Q $19,DX,AX
- MULQ 88(SP)
+ MULQ 72(SP)
ADDQ AX,R12
ADCQ DX,R13
MOVQ ·REDMASK51(SB),DX
@@ -278,11 +156,126 @@
IMUL3Q $19,CX,CX
ADDQ CX,SI
ANDQ DX,R10
- MOVQ SI,176(SP)
- MOVQ R8,184(SP)
- MOVQ R9,192(SP)
- MOVQ AX,200(SP)
- MOVQ R10,208(SP)
+ MOVQ SI,80(SP)
+ MOVQ R8,88(SP)
+ MOVQ R9,96(SP)
+ MOVQ AX,104(SP)
+ MOVQ R10,112(SP)
+ MOVQ 0(SP),AX
+ MULQ 0(SP)
+ MOVQ AX,SI
+ MOVQ DX,CX
+ MOVQ 0(SP),AX
+ SHLQ $1,AX
+ MULQ 8(SP)
+ MOVQ AX,R8
+ MOVQ DX,R9
+ MOVQ 0(SP),AX
+ SHLQ $1,AX
+ MULQ 16(SP)
+ MOVQ AX,R10
+ MOVQ DX,R11
+ MOVQ 0(SP),AX
+ SHLQ $1,AX
+ MULQ 24(SP)
+ MOVQ AX,R12
+ MOVQ DX,R13
+ MOVQ 0(SP),AX
+ SHLQ $1,AX
+ MULQ 32(SP)
+ MOVQ AX,R14
+ MOVQ DX,R15
+ MOVQ 8(SP),AX
+ MULQ 8(SP)
+ ADDQ AX,R10
+ ADCQ DX,R11
+ MOVQ 8(SP),AX
+ SHLQ $1,AX
+ MULQ 16(SP)
+ ADDQ AX,R12
+ ADCQ DX,R13
+ MOVQ 8(SP),AX
+ SHLQ $1,AX
+ MULQ 24(SP)
+ ADDQ AX,R14
+ ADCQ DX,R15
+ MOVQ 8(SP),DX
+ IMUL3Q $38,DX,AX
+ MULQ 32(SP)
+ ADDQ AX,SI
+ ADCQ DX,CX
+ MOVQ 16(SP),AX
+ MULQ 16(SP)
+ ADDQ AX,R14
+ ADCQ DX,R15
+ MOVQ 16(SP),DX
+ IMUL3Q $38,DX,AX
+ MULQ 24(SP)
+ ADDQ AX,SI
+ ADCQ DX,CX
+ MOVQ 16(SP),DX
+ IMUL3Q $38,DX,AX
+ MULQ 32(SP)
+ ADDQ AX,R8
+ ADCQ DX,R9
+ MOVQ 24(SP),DX
+ IMUL3Q $19,DX,AX
+ MULQ 24(SP)
+ ADDQ AX,R8
+ ADCQ DX,R9
+ MOVQ 24(SP),DX
+ IMUL3Q $38,DX,AX
+ MULQ 32(SP)
+ ADDQ AX,R10
+ ADCQ DX,R11
+ MOVQ 32(SP),DX
+ IMUL3Q $19,DX,AX
+ MULQ 32(SP)
+ ADDQ AX,R12
+ ADCQ DX,R13
+ MOVQ ·REDMASK51(SB),DX
+ SHLQ $13,CX:SI
+ ANDQ DX,SI
+ SHLQ $13,R9:R8
+ ANDQ DX,R8
+ ADDQ CX,R8
+ SHLQ $13,R11:R10
+ ANDQ DX,R10
+ ADDQ R9,R10
+ SHLQ $13,R13:R12
+ ANDQ DX,R12
+ ADDQ R11,R12
+ SHLQ $13,R15:R14
+ ANDQ DX,R14
+ ADDQ R13,R14
+ IMUL3Q $19,R15,CX
+ ADDQ CX,SI
+ MOVQ SI,CX
+ SHRQ $51,CX
+ ADDQ R8,CX
+ ANDQ DX,SI
+ MOVQ CX,R8
+ SHRQ $51,CX
+ ADDQ R10,CX
+ ANDQ DX,R8
+ MOVQ CX,R9
+ SHRQ $51,CX
+ ADDQ R12,CX
+ ANDQ DX,R9
+ MOVQ CX,AX
+ SHRQ $51,CX
+ ADDQ R14,CX
+ ANDQ DX,AX
+ MOVQ CX,R10
+ SHRQ $51,CX
+ IMUL3Q $19,CX,CX
+ ADDQ CX,SI
+ ANDQ DX,R10
+ MOVQ SI,120(SP)
+ MOVQ R8,128(SP)
+ MOVQ R9,136(SP)
+ MOVQ AX,144(SP)
+ MOVQ R10,152(SP)
MOVQ SI,SI
MOVQ R8,DX
MOVQ R9,CX
@@ -293,16 +286,16 @@
ADDQ ·_2P1234(SB),CX
ADDQ ·_2P1234(SB),R8
ADDQ ·_2P1234(SB),R9
- SUBQ 136(SP),SI
- SUBQ 144(SP),DX
- SUBQ 152(SP),CX
- SUBQ 160(SP),R8
- SUBQ 168(SP),R9
- MOVQ SI,216(SP)
- MOVQ DX,224(SP)
- MOVQ CX,232(SP)
- MOVQ R8,240(SP)
- MOVQ R9,248(SP)
+ SUBQ 80(SP),SI
+ SUBQ 88(SP),DX
+ SUBQ 96(SP),CX
+ SUBQ 104(SP),R8
+ SUBQ 112(SP),R9
+ MOVQ SI,160(SP)
+ MOVQ DX,168(SP)
+ MOVQ CX,176(SP)
+ MOVQ R8,184(SP)
+ MOVQ R9,192(SP)
MOVQ 120(DI),SI
MOVQ 128(DI),DX
MOVQ 136(DI),CX
@@ -328,121 +321,121 @@
SUBQ 176(DI),R11
SUBQ 184(DI),R12
SUBQ 192(DI),R13
- MOVQ SI,256(SP)
- MOVQ DX,264(SP)
- MOVQ CX,272(SP)
- MOVQ R8,280(SP)
- MOVQ R9,288(SP)
- MOVQ AX,296(SP)
- MOVQ R10,304(SP)
- MOVQ R11,312(SP)
- MOVQ R12,320(SP)
- MOVQ R13,328(SP)
- MOVQ 280(SP),SI
+ MOVQ SI,200(SP)
+ MOVQ DX,208(SP)
+ MOVQ CX,216(SP)
+ MOVQ R8,224(SP)
+ MOVQ R9,232(SP)
+ MOVQ AX,240(SP)
+ MOVQ R10,248(SP)
+ MOVQ R11,256(SP)
+ MOVQ R12,264(SP)
+ MOVQ R13,272(SP)
+ MOVQ 224(SP),SI
IMUL3Q $19,SI,AX
- MOVQ AX,336(SP)
- MULQ 112(SP)
+ MOVQ AX,280(SP)
+ MULQ 56(SP)
MOVQ AX,SI
MOVQ DX,CX
- MOVQ 288(SP),DX
+ MOVQ 232(SP),DX
IMUL3Q $19,DX,AX
- MOVQ AX,344(SP)
- MULQ 104(SP)
+ MOVQ AX,288(SP)
+ MULQ 48(SP)
ADDQ AX,SI
ADCQ DX,CX
- MOVQ 256(SP),AX
- MULQ 96(SP)
+ MOVQ 200(SP),AX
+ MULQ 40(SP)
ADDQ AX,SI
ADCQ DX,CX
- MOVQ 256(SP),AX
- MULQ 104(SP)
+ MOVQ 200(SP),AX
+ MULQ 48(SP)
MOVQ AX,R8
MOVQ DX,R9
- MOVQ 256(SP),AX
- MULQ 112(SP)
+ MOVQ 200(SP),AX
+ MULQ 56(SP)
MOVQ AX,R10
MOVQ DX,R11
- MOVQ 256(SP),AX
- MULQ 120(SP)
+ MOVQ 200(SP),AX
+ MULQ 64(SP)
MOVQ AX,R12
MOVQ DX,R13
- MOVQ 256(SP),AX
- MULQ 128(SP)
+ MOVQ 200(SP),AX
+ MULQ 72(SP)
MOVQ AX,R14
MOVQ DX,R15
- MOVQ 264(SP),AX
- MULQ 96(SP)
+ MOVQ 208(SP),AX
+ MULQ 40(SP)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 264(SP),AX
- MULQ 104(SP)
+ MOVQ 208(SP),AX
+ MULQ 48(SP)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 264(SP),AX
- MULQ 112(SP)
+ MOVQ 208(SP),AX
+ MULQ 56(SP)
ADDQ AX,R12
ADCQ DX,R13
- MOVQ 264(SP),AX
- MULQ 120(SP)
+ MOVQ 208(SP),AX
+ MULQ 64(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 264(SP),DX
+ MOVQ 208(SP),DX
IMUL3Q $19,DX,AX
- MULQ 128(SP)
+ MULQ 72(SP)
ADDQ AX,SI
ADCQ DX,CX
- MOVQ 272(SP),AX
- MULQ 96(SP)
+ MOVQ 216(SP),AX
+ MULQ 40(SP)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 272(SP),AX
- MULQ 104(SP)
+ MOVQ 216(SP),AX
+ MULQ 48(SP)
ADDQ AX,R12
ADCQ DX,R13
- MOVQ 272(SP),AX
- MULQ 112(SP)
+ MOVQ 216(SP),AX
+ MULQ 56(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 272(SP),DX
+ MOVQ 216(SP),DX
IMUL3Q $19,DX,AX
- MULQ 120(SP)
+ MULQ 64(SP)
ADDQ AX,SI
ADCQ DX,CX
- MOVQ 272(SP),DX
+ MOVQ 216(SP),DX
IMUL3Q $19,DX,AX
- MULQ 128(SP)
+ MULQ 72(SP)
+ ADDQ AX,R8
+ ADCQ DX,R9
+ MOVQ 224(SP),AX
+ MULQ 40(SP)
+ ADDQ AX,R12
+ ADCQ DX,R13
+ MOVQ 224(SP),AX
+ MULQ 48(SP)
+ ADDQ AX,R14
+ ADCQ DX,R15
+ MOVQ 280(SP),AX
+ MULQ 64(SP)
ADDQ AX,R8
ADCQ DX,R9
MOVQ 280(SP),AX
- MULQ 96(SP)
- ADDQ AX,R12
- ADCQ DX,R13
- MOVQ 280(SP),AX
- MULQ 104(SP)
+ MULQ 72(SP)
+ ADDQ AX,R10
+ ADCQ DX,R11
+ MOVQ 232(SP),AX
+ MULQ 40(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 336(SP),AX
- MULQ 120(SP)
+ MOVQ 288(SP),AX
+ MULQ 56(SP)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 336(SP),AX
- MULQ 128(SP)
+ MOVQ 288(SP),AX
+ MULQ 64(SP)
ADDQ AX,R10
ADCQ DX,R11
MOVQ 288(SP),AX
- MULQ 96(SP)
- ADDQ AX,R14
- ADCQ DX,R15
- MOVQ 344(SP),AX
- MULQ 112(SP)
- ADDQ AX,R8
- ADCQ DX,R9
- MOVQ 344(SP),AX
- MULQ 120(SP)
- ADDQ AX,R10
- ADCQ DX,R11
- MOVQ 344(SP),AX
- MULQ 128(SP)
+ MULQ 72(SP)
ADDQ AX,R12
ADCQ DX,R13
MOVQ ·REDMASK51(SB),DX
@@ -483,116 +476,116 @@
IMUL3Q $19,CX,CX
ADDQ CX,SI
ANDQ DX,R10
- MOVQ SI,96(SP)
- MOVQ R8,104(SP)
- MOVQ R9,112(SP)
- MOVQ AX,120(SP)
- MOVQ R10,128(SP)
- MOVQ 320(SP),SI
+ MOVQ SI,40(SP)
+ MOVQ R8,48(SP)
+ MOVQ R9,56(SP)
+ MOVQ AX,64(SP)
+ MOVQ R10,72(SP)
+ MOVQ 264(SP),SI
IMUL3Q $19,SI,AX
- MOVQ AX,256(SP)
- MULQ 72(SP)
+ MOVQ AX,200(SP)
+ MULQ 16(SP)
MOVQ AX,SI
MOVQ DX,CX
- MOVQ 328(SP),DX
+ MOVQ 272(SP),DX
IMUL3Q $19,DX,AX
- MOVQ AX,264(SP)
- MULQ 64(SP)
+ MOVQ AX,208(SP)
+ MULQ 8(SP)
ADDQ AX,SI
ADCQ DX,CX
- MOVQ 296(SP),AX
- MULQ 56(SP)
+ MOVQ 240(SP),AX
+ MULQ 0(SP)
ADDQ AX,SI
ADCQ DX,CX
- MOVQ 296(SP),AX
- MULQ 64(SP)
+ MOVQ 240(SP),AX
+ MULQ 8(SP)
MOVQ AX,R8
MOVQ DX,R9
- MOVQ 296(SP),AX
- MULQ 72(SP)
+ MOVQ 240(SP),AX
+ MULQ 16(SP)
MOVQ AX,R10
MOVQ DX,R11
- MOVQ 296(SP),AX
- MULQ 80(SP)
+ MOVQ 240(SP),AX
+ MULQ 24(SP)
MOVQ AX,R12
MOVQ DX,R13
- MOVQ 296(SP),AX
- MULQ 88(SP)
+ MOVQ 240(SP),AX
+ MULQ 32(SP)
MOVQ AX,R14
MOVQ DX,R15
- MOVQ 304(SP),AX
- MULQ 56(SP)
+ MOVQ 248(SP),AX
+ MULQ 0(SP)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 304(SP),AX
- MULQ 64(SP)
+ MOVQ 248(SP),AX
+ MULQ 8(SP)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 304(SP),AX
- MULQ 72(SP)
+ MOVQ 248(SP),AX
+ MULQ 16(SP)
ADDQ AX,R12
ADCQ DX,R13
- MOVQ 304(SP),AX
- MULQ 80(SP)
+ MOVQ 248(SP),AX
+ MULQ 24(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 304(SP),DX
+ MOVQ 248(SP),DX
IMUL3Q $19,DX,AX
- MULQ 88(SP)
+ MULQ 32(SP)
ADDQ AX,SI
ADCQ DX,CX
- MOVQ 312(SP),AX
- MULQ 56(SP)
+ MOVQ 256(SP),AX
+ MULQ 0(SP)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 312(SP),AX
- MULQ 64(SP)
+ MOVQ 256(SP),AX
+ MULQ 8(SP)
ADDQ AX,R12
ADCQ DX,R13
- MOVQ 312(SP),AX
- MULQ 72(SP)
+ MOVQ 256(SP),AX
+ MULQ 16(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 312(SP),DX
+ MOVQ 256(SP),DX
IMUL3Q $19,DX,AX
- MULQ 80(SP)
+ MULQ 24(SP)
ADDQ AX,SI
ADCQ DX,CX
- MOVQ 312(SP),DX
+ MOVQ 256(SP),DX
IMUL3Q $19,DX,AX
- MULQ 88(SP)
+ MULQ 32(SP)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 320(SP),AX
- MULQ 56(SP)
+ MOVQ 264(SP),AX
+ MULQ 0(SP)
ADDQ AX,R12
ADCQ DX,R13
- MOVQ 320(SP),AX
- MULQ 64(SP)
+ MOVQ 264(SP),AX
+ MULQ 8(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 256(SP),AX
- MULQ 80(SP)
+ MOVQ 200(SP),AX
+ MULQ 24(SP)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 256(SP),AX
- MULQ 88(SP)
+ MOVQ 200(SP),AX
+ MULQ 32(SP)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 328(SP),AX
- MULQ 56(SP)
+ MOVQ 272(SP),AX
+ MULQ 0(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 264(SP),AX
- MULQ 72(SP)
+ MOVQ 208(SP),AX
+ MULQ 16(SP)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 264(SP),AX
- MULQ 80(SP)
+ MOVQ 208(SP),AX
+ MULQ 24(SP)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 264(SP),AX
- MULQ 88(SP)
+ MOVQ 208(SP),AX
+ MULQ 32(SP)
ADDQ AX,R12
ADCQ DX,R13
MOVQ ·REDMASK51(SB),DX
@@ -643,16 +636,16 @@
ADDQ ·_2P1234(SB),R11
ADDQ ·_2P1234(SB),R12
ADDQ ·_2P1234(SB),R13
- ADDQ 96(SP),SI
- ADDQ 104(SP),R8
- ADDQ 112(SP),R9
- ADDQ 120(SP),AX
- ADDQ 128(SP),R10
- SUBQ 96(SP),DX
- SUBQ 104(SP),CX
- SUBQ 112(SP),R11
- SUBQ 120(SP),R12
- SUBQ 128(SP),R13
+ ADDQ 40(SP),SI
+ ADDQ 48(SP),R8
+ ADDQ 56(SP),R9
+ ADDQ 64(SP),AX
+ ADDQ 72(SP),R10
+ SUBQ 40(SP),DX
+ SUBQ 48(SP),CX
+ SUBQ 56(SP),R11
+ SUBQ 64(SP),R12
+ SUBQ 72(SP),R13
MOVQ SI,120(DI)
MOVQ R8,128(DI)
MOVQ R9,136(DI)
@@ -895,13 +888,13 @@
MOVQ R10,192(DI)
MOVQ 184(DI),SI
IMUL3Q $19,SI,AX
- MOVQ AX,56(SP)
+ MOVQ AX,0(SP)
MULQ 16(DI)
MOVQ AX,SI
MOVQ DX,CX
MOVQ 192(DI),DX
IMUL3Q $19,DX,AX
- MOVQ AX,64(SP)
+ MOVQ AX,8(SP)
MULQ 8(DI)
ADDQ AX,SI
ADCQ DX,CX
@@ -976,11 +969,11 @@
MULQ 8(DI)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 56(SP),AX
+ MOVQ 0(SP),AX
MULQ 24(DI)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 56(SP),AX
+ MOVQ 0(SP),AX
MULQ 32(DI)
ADDQ AX,R10
ADCQ DX,R11
@@ -988,15 +981,15 @@
MULQ 0(DI)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 64(SP),AX
+ MOVQ 8(SP),AX
MULQ 16(DI)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 64(SP),AX
+ MOVQ 8(SP),AX
MULQ 24(DI)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 64(SP),AX
+ MOVQ 8(SP),AX
MULQ 32(DI)
ADDQ AX,R12
ADCQ DX,R13
@@ -1043,111 +1036,111 @@
MOVQ R9,176(DI)
MOVQ AX,184(DI)
MOVQ R10,192(DI)
- MOVQ 200(SP),SI
+ MOVQ 144(SP),SI
IMUL3Q $19,SI,AX
- MOVQ AX,56(SP)
- MULQ 152(SP)
+ MOVQ AX,0(SP)
+ MULQ 96(SP)
MOVQ AX,SI
MOVQ DX,CX
- MOVQ 208(SP),DX
+ MOVQ 152(SP),DX
IMUL3Q $19,DX,AX
- MOVQ AX,64(SP)
- MULQ 144(SP)
+ MOVQ AX,8(SP)
+ MULQ 88(SP)
ADDQ AX,SI
ADCQ DX,CX
- MOVQ 176(SP),AX
- MULQ 136(SP)
+ MOVQ 120(SP),AX
+ MULQ 80(SP)
ADDQ AX,SI
ADCQ DX,CX
- MOVQ 176(SP),AX
- MULQ 144(SP)
+ MOVQ 120(SP),AX
+ MULQ 88(SP)
MOVQ AX,R8
MOVQ DX,R9
- MOVQ 176(SP),AX
- MULQ 152(SP)
+ MOVQ 120(SP),AX
+ MULQ 96(SP)
MOVQ AX,R10
MOVQ DX,R11
- MOVQ 176(SP),AX
- MULQ 160(SP)
+ MOVQ 120(SP),AX
+ MULQ 104(SP)
MOVQ AX,R12
MOVQ DX,R13
- MOVQ 176(SP),AX
- MULQ 168(SP)
+ MOVQ 120(SP),AX
+ MULQ 112(SP)
MOVQ AX,R14
MOVQ DX,R15
- MOVQ 184(SP),AX
- MULQ 136(SP)
+ MOVQ 128(SP),AX
+ MULQ 80(SP)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 184(SP),AX
- MULQ 144(SP)
+ MOVQ 128(SP),AX
+ MULQ 88(SP)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 184(SP),AX
- MULQ 152(SP)
+ MOVQ 128(SP),AX
+ MULQ 96(SP)
ADDQ AX,R12
ADCQ DX,R13
- MOVQ 184(SP),AX
- MULQ 160(SP)
+ MOVQ 128(SP),AX
+ MULQ 104(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 184(SP),DX
+ MOVQ 128(SP),DX
IMUL3Q $19,DX,AX
- MULQ 168(SP)
+ MULQ 112(SP)
ADDQ AX,SI
ADCQ DX,CX
- MOVQ 192(SP),AX
- MULQ 136(SP)
+ MOVQ 136(SP),AX
+ MULQ 80(SP)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 192(SP),AX
- MULQ 144(SP)
+ MOVQ 136(SP),AX
+ MULQ 88(SP)
ADDQ AX,R12
ADCQ DX,R13
- MOVQ 192(SP),AX
- MULQ 152(SP)
+ MOVQ 136(SP),AX
+ MULQ 96(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 192(SP),DX
+ MOVQ 136(SP),DX
IMUL3Q $19,DX,AX
- MULQ 160(SP)
+ MULQ 104(SP)
ADDQ AX,SI
ADCQ DX,CX
- MOVQ 192(SP),DX
+ MOVQ 136(SP),DX
IMUL3Q $19,DX,AX
- MULQ 168(SP)
+ MULQ 112(SP)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 200(SP),AX
- MULQ 136(SP)
+ MOVQ 144(SP),AX
+ MULQ 80(SP)
ADDQ AX,R12
ADCQ DX,R13
- MOVQ 200(SP),AX
- MULQ 144(SP)
+ MOVQ 144(SP),AX
+ MULQ 88(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 56(SP),AX
- MULQ 160(SP)
+ MOVQ 0(SP),AX
+ MULQ 104(SP)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 56(SP),AX
- MULQ 168(SP)
+ MOVQ 0(SP),AX
+ MULQ 112(SP)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 208(SP),AX
- MULQ 136(SP)
+ MOVQ 152(SP),AX
+ MULQ 80(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 64(SP),AX
- MULQ 152(SP)
+ MOVQ 8(SP),AX
+ MULQ 96(SP)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 64(SP),AX
- MULQ 160(SP)
+ MOVQ 8(SP),AX
+ MULQ 104(SP)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 64(SP),AX
- MULQ 168(SP)
+ MOVQ 8(SP),AX
+ MULQ 112(SP)
ADDQ AX,R12
ADCQ DX,R13
MOVQ ·REDMASK51(SB),DX
@@ -1193,37 +1186,37 @@
MOVQ R9,56(DI)
MOVQ AX,64(DI)
MOVQ R10,72(DI)
- MOVQ 216(SP),AX
+ MOVQ 160(SP),AX
MULQ ·_121666_213(SB)
SHRQ $13,AX
MOVQ AX,SI
MOVQ DX,CX
- MOVQ 224(SP),AX
+ MOVQ 168(SP),AX
MULQ ·_121666_213(SB)
SHRQ $13,AX
ADDQ AX,CX
MOVQ DX,R8
- MOVQ 232(SP),AX
+ MOVQ 176(SP),AX
MULQ ·_121666_213(SB)
SHRQ $13,AX
ADDQ AX,R8
MOVQ DX,R9
- MOVQ 240(SP),AX
+ MOVQ 184(SP),AX
MULQ ·_121666_213(SB)
SHRQ $13,AX
ADDQ AX,R9
MOVQ DX,R10
- MOVQ 248(SP),AX
+ MOVQ 192(SP),AX
MULQ ·_121666_213(SB)
SHRQ $13,AX
ADDQ AX,R10
IMUL3Q $19,DX,DX
ADDQ DX,SI
- ADDQ 136(SP),SI
- ADDQ 144(SP),CX
- ADDQ 152(SP),R8
- ADDQ 160(SP),R9
- ADDQ 168(SP),R10
+ ADDQ 80(SP),SI
+ ADDQ 88(SP),CX
+ ADDQ 96(SP),R8
+ ADDQ 104(SP),R9
+ ADDQ 112(SP),R10
MOVQ SI,80(DI)
MOVQ CX,88(DI)
MOVQ R8,96(DI)
@@ -1231,109 +1224,109 @@
MOVQ R10,112(DI)
MOVQ 104(DI),SI
IMUL3Q $19,SI,AX
- MOVQ AX,56(SP)
- MULQ 232(SP)
+ MOVQ AX,0(SP)
+ MULQ 176(SP)
MOVQ AX,SI
MOVQ DX,CX
MOVQ 112(DI),DX
IMUL3Q $19,DX,AX
- MOVQ AX,64(SP)
- MULQ 224(SP)
+ MOVQ AX,8(SP)
+ MULQ 168(SP)
ADDQ AX,SI
ADCQ DX,CX
MOVQ 80(DI),AX
- MULQ 216(SP)
+ MULQ 160(SP)
ADDQ AX,SI
ADCQ DX,CX
MOVQ 80(DI),AX
- MULQ 224(SP)
+ MULQ 168(SP)
MOVQ AX,R8
MOVQ DX,R9
MOVQ 80(DI),AX
- MULQ 232(SP)
+ MULQ 176(SP)
MOVQ AX,R10
MOVQ DX,R11
MOVQ 80(DI),AX
- MULQ 240(SP)
+ MULQ 184(SP)
MOVQ AX,R12
MOVQ DX,R13
MOVQ 80(DI),AX
- MULQ 248(SP)
+ MULQ 192(SP)
MOVQ AX,R14
MOVQ DX,R15
MOVQ 88(DI),AX
- MULQ 216(SP)
+ MULQ 160(SP)
ADDQ AX,R8
ADCQ DX,R9
MOVQ 88(DI),AX
- MULQ 224(SP)
+ MULQ 168(SP)
ADDQ AX,R10
ADCQ DX,R11
MOVQ 88(DI),AX
- MULQ 232(SP)
+ MULQ 176(SP)
ADDQ AX,R12
ADCQ DX,R13
MOVQ 88(DI),AX
- MULQ 240(SP)
+ MULQ 184(SP)
ADDQ AX,R14
ADCQ DX,R15
MOVQ 88(DI),DX
IMUL3Q $19,DX,AX
- MULQ 248(SP)
+ MULQ 192(SP)
ADDQ AX,SI
ADCQ DX,CX
MOVQ 96(DI),AX
- MULQ 216(SP)
+ MULQ 160(SP)
ADDQ AX,R10
ADCQ DX,R11
MOVQ 96(DI),AX
- MULQ 224(SP)
+ MULQ 168(SP)
ADDQ AX,R12
ADCQ DX,R13
MOVQ 96(DI),AX
- MULQ 232(SP)
+ MULQ 176(SP)
ADDQ AX,R14
ADCQ DX,R15
MOVQ 96(DI),DX
IMUL3Q $19,DX,AX
- MULQ 240(SP)
+ MULQ 184(SP)
ADDQ AX,SI
ADCQ DX,CX
MOVQ 96(DI),DX
IMUL3Q $19,DX,AX
- MULQ 248(SP)
+ MULQ 192(SP)
ADDQ AX,R8
ADCQ DX,R9
MOVQ 104(DI),AX
- MULQ 216(SP)
+ MULQ 160(SP)
ADDQ AX,R12
ADCQ DX,R13
MOVQ 104(DI),AX
- MULQ 224(SP)
+ MULQ 168(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 56(SP),AX
- MULQ 240(SP)
+ MOVQ 0(SP),AX
+ MULQ 184(SP)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 56(SP),AX
- MULQ 248(SP)
+ MOVQ 0(SP),AX
+ MULQ 192(SP)
ADDQ AX,R10
ADCQ DX,R11
MOVQ 112(DI),AX
- MULQ 216(SP)
+ MULQ 160(SP)
ADDQ AX,R14
ADCQ DX,R15
- MOVQ 64(SP),AX
- MULQ 232(SP)
+ MOVQ 8(SP),AX
+ MULQ 176(SP)
ADDQ AX,R8
ADCQ DX,R9
- MOVQ 64(SP),AX
- MULQ 240(SP)
+ MOVQ 8(SP),AX
+ MULQ 184(SP)
ADDQ AX,R10
ADCQ DX,R11
- MOVQ 64(SP),AX
- MULQ 248(SP)
+ MOVQ 8(SP),AX
+ MULQ 192(SP)
ADDQ AX,R12
ADCQ DX,R13
MOVQ ·REDMASK51(SB),DX
@@ -1379,13 +1372,4 @@
MOVQ R9,96(DI)
MOVQ AX,104(DI)
MOVQ R10,112(DI)
- MOVQ 0(SP),R11
- MOVQ 8(SP),R12
- MOVQ 16(SP),R13
- MOVQ 24(SP),R14
- MOVQ 32(SP),R15
- MOVQ 40(SP),BX
- MOVQ 48(SP),BP
- MOVQ DI,AX
- MOVQ SI,DX
RET
