cheng/libsodium
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Align loops

Browse Source
This commit is contained in:
Frank Denis 2016-05-16 12:25:35 +02:00
parent c303c1f709
commit 593599a11a
1 changed files with 8 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
src/libsodium/crypto_stream/salsa20/amd64_xmm6/stream_salsa20_amd64_xmm6.S
View File

@ -136,6 +136,7 @@ movdqa %xmm1,304(%rsp)
movdqa %xmm2,320(%rsp) movdqa %xmm2,320(%rsp)
movdqa %xmm0,336(%rsp) movdqa %xmm0,336(%rsp)
.p2align 4
._bytesatleast256: ._bytesatleast256:
movq 472(%rsp),%rdx movq 472(%rsp),%rdx
mov %rdx,%rcx mov %rdx,%rcx
@ -182,6 +183,7 @@ movdqa 224(%rsp),%xmm13
movdqa 304(%rsp),%xmm14 movdqa 304(%rsp),%xmm14
movdqa 352(%rsp),%xmm15 movdqa 352(%rsp),%xmm15
.p2align 4
._mainloop1: ._mainloop1:
movdqa %xmm1,384(%rsp) movdqa %xmm1,384(%rsp)
movdqa %xmm2,400(%rsp) movdqa %xmm2,400(%rsp)
@ -425,6 +427,7 @@ movdqa 384(%rsp),%xmm12
movdqa 400(%rsp),%xmm0 movdqa 400(%rsp),%xmm0
sub $2,%rdx sub $2,%rdx
ja ._mainloop1 ja ._mainloop1
paddd 176(%rsp),%xmm12 paddd 176(%rsp),%xmm12
paddd 240(%rsp),%xmm7 paddd 240(%rsp),%xmm7
paddd 288(%rsp),%xmm10 paddd 288(%rsp),%xmm10
@ -687,12 +690,14 @@ add $256,%rsi
add $256,%rdi add $256,%rdi
cmp $256,%r9 cmp $256,%r9
jae ._bytesatleast256 jae ._bytesatleast256
cmp $0,%r9 cmp $0,%r9
jbe ._done jbe ._done
._bytesbetween1and255: ._bytesbetween1and255:
cmp $64,%r9 cmp $64,%r9
jae ._nocopy jae ._nocopy
mov %rdi,%rdx mov %rdi,%rdx
leaq 0(%rsp),%rdi leaq 0(%rsp),%rdi
mov %r9,%rcx mov %r9,%rcx
@ -709,6 +714,7 @@ movdqa 96(%rsp),%xmm3
movdqa %xmm1,%xmm4 movdqa %xmm1,%xmm4
mov $20,%rcx mov $20,%rcx
.p2align 4
._mainloop2: ._mainloop2:
paddd %xmm0,%xmm4 paddd %xmm0,%xmm4
movdqa %xmm0,%xmm5 movdqa %xmm0,%xmm5
@ -837,6 +843,7 @@ pxor %xmm5,%xmm0
pshufd $0x39,%xmm3,%xmm3 pshufd $0x39,%xmm3,%xmm3
pxor %xmm6,%xmm0 pxor %xmm6,%xmm0
ja ._mainloop2 ja ._mainloop2
paddd 112(%rsp),%xmm0 paddd 112(%rsp),%xmm0
paddd 64(%rsp),%xmm1 paddd 64(%rsp),%xmm1
paddd 80(%rsp),%xmm2 paddd 80(%rsp),%xmm2
@ -912,6 +919,7 @@ movq %rcx,472(%rsp)
cmp $64,%r9 cmp $64,%r9
ja ._bytesatleast65 ja ._bytesatleast65
jae ._bytesatleast64 jae ._bytesatleast64
mov %rdi,%rsi mov %rdi,%rsi
mov %rdx,%rdi mov %rdx,%rdi
mov %r9,%rcx mov %r9,%rcx