mirror of
https://github.com/ClickHouse/ClickHouse.git
synced 2024-12-05 14:02:21 +00:00
486 lines
8.8 KiB
ArmAsm
486 lines
8.8 KiB
ArmAsm
|
# This file is generated from a similarly-named Perl script in the BoringSSL
|
||
|
# source tree. Do not edit by hand.
|
||
|
|
||
|
#if defined(__i386__)
|
||
|
#if defined(BORINGSSL_PREFIX)
|
||
|
#include <boringssl_prefix_symbols_asm.h>
|
||
|
#endif
|
||
|
.text
|
||
|
.globl _bn_mul_mont
|
||
|
.private_extern _bn_mul_mont
|
||
|
.align 4
|
||
|
_bn_mul_mont:
|
||
|
L_bn_mul_mont_begin:
|
||
|
pushl %ebp
|
||
|
pushl %ebx
|
||
|
pushl %esi
|
||
|
pushl %edi
|
||
|
xorl %eax,%eax
|
||
|
movl 40(%esp),%edi
|
||
|
cmpl $4,%edi
|
||
|
jl L000just_leave
|
||
|
leal 20(%esp),%esi
|
||
|
leal 24(%esp),%edx
|
||
|
addl $2,%edi
|
||
|
negl %edi
|
||
|
leal -32(%esp,%edi,4),%ebp
|
||
|
negl %edi
|
||
|
movl %ebp,%eax
|
||
|
subl %edx,%eax
|
||
|
andl $2047,%eax
|
||
|
subl %eax,%ebp
|
||
|
xorl %ebp,%edx
|
||
|
andl $2048,%edx
|
||
|
xorl $2048,%edx
|
||
|
subl %edx,%ebp
|
||
|
andl $-64,%ebp
|
||
|
movl %esp,%eax
|
||
|
subl %ebp,%eax
|
||
|
andl $-4096,%eax
|
||
|
movl %esp,%edx
|
||
|
leal (%ebp,%eax,1),%esp
|
||
|
movl (%esp),%eax
|
||
|
cmpl %ebp,%esp
|
||
|
ja L001page_walk
|
||
|
jmp L002page_walk_done
|
||
|
.align 4,0x90
|
||
|
L001page_walk:
|
||
|
leal -4096(%esp),%esp
|
||
|
movl (%esp),%eax
|
||
|
cmpl %ebp,%esp
|
||
|
ja L001page_walk
|
||
|
L002page_walk_done:
|
||
|
movl (%esi),%eax
|
||
|
movl 4(%esi),%ebx
|
||
|
movl 8(%esi),%ecx
|
||
|
movl 12(%esi),%ebp
|
||
|
movl 16(%esi),%esi
|
||
|
movl (%esi),%esi
|
||
|
movl %eax,4(%esp)
|
||
|
movl %ebx,8(%esp)
|
||
|
movl %ecx,12(%esp)
|
||
|
movl %ebp,16(%esp)
|
||
|
movl %esi,20(%esp)
|
||
|
leal -3(%edi),%ebx
|
||
|
movl %edx,24(%esp)
|
||
|
call L003PIC_me_up
|
||
|
L003PIC_me_up:
|
||
|
popl %eax
|
||
|
movl L_OPENSSL_ia32cap_P$non_lazy_ptr-L003PIC_me_up(%eax),%eax
|
||
|
btl $26,(%eax)
|
||
|
jnc L004non_sse2
|
||
|
movl $-1,%eax
|
||
|
movd %eax,%mm7
|
||
|
movl 8(%esp),%esi
|
||
|
movl 12(%esp),%edi
|
||
|
movl 16(%esp),%ebp
|
||
|
xorl %edx,%edx
|
||
|
xorl %ecx,%ecx
|
||
|
movd (%edi),%mm4
|
||
|
movd (%esi),%mm5
|
||
|
movd (%ebp),%mm3
|
||
|
pmuludq %mm4,%mm5
|
||
|
movq %mm5,%mm2
|
||
|
movq %mm5,%mm0
|
||
|
pand %mm7,%mm0
|
||
|
pmuludq 20(%esp),%mm5
|
||
|
pmuludq %mm5,%mm3
|
||
|
paddq %mm0,%mm3
|
||
|
movd 4(%ebp),%mm1
|
||
|
movd 4(%esi),%mm0
|
||
|
psrlq $32,%mm2
|
||
|
psrlq $32,%mm3
|
||
|
incl %ecx
|
||
|
.align 4,0x90
|
||
|
L0051st:
|
||
|
pmuludq %mm4,%mm0
|
||
|
pmuludq %mm5,%mm1
|
||
|
paddq %mm0,%mm2
|
||
|
paddq %mm1,%mm3
|
||
|
movq %mm2,%mm0
|
||
|
pand %mm7,%mm0
|
||
|
movd 4(%ebp,%ecx,4),%mm1
|
||
|
paddq %mm0,%mm3
|
||
|
movd 4(%esi,%ecx,4),%mm0
|
||
|
psrlq $32,%mm2
|
||
|
movd %mm3,28(%esp,%ecx,4)
|
||
|
psrlq $32,%mm3
|
||
|
leal 1(%ecx),%ecx
|
||
|
cmpl %ebx,%ecx
|
||
|
jl L0051st
|
||
|
pmuludq %mm4,%mm0
|
||
|
pmuludq %mm5,%mm1
|
||
|
paddq %mm0,%mm2
|
||
|
paddq %mm1,%mm3
|
||
|
movq %mm2,%mm0
|
||
|
pand %mm7,%mm0
|
||
|
paddq %mm0,%mm3
|
||
|
movd %mm3,28(%esp,%ecx,4)
|
||
|
psrlq $32,%mm2
|
||
|
psrlq $32,%mm3
|
||
|
paddq %mm2,%mm3
|
||
|
movq %mm3,32(%esp,%ebx,4)
|
||
|
incl %edx
|
||
|
L006outer:
|
||
|
xorl %ecx,%ecx
|
||
|
movd (%edi,%edx,4),%mm4
|
||
|
movd (%esi),%mm5
|
||
|
movd 32(%esp),%mm6
|
||
|
movd (%ebp),%mm3
|
||
|
pmuludq %mm4,%mm5
|
||
|
paddq %mm6,%mm5
|
||
|
movq %mm5,%mm0
|
||
|
movq %mm5,%mm2
|
||
|
pand %mm7,%mm0
|
||
|
pmuludq 20(%esp),%mm5
|
||
|
pmuludq %mm5,%mm3
|
||
|
paddq %mm0,%mm3
|
||
|
movd 36(%esp),%mm6
|
||
|
movd 4(%ebp),%mm1
|
||
|
movd 4(%esi),%mm0
|
||
|
psrlq $32,%mm2
|
||
|
psrlq $32,%mm3
|
||
|
paddq %mm6,%mm2
|
||
|
incl %ecx
|
||
|
decl %ebx
|
||
|
L007inner:
|
||
|
pmuludq %mm4,%mm0
|
||
|
pmuludq %mm5,%mm1
|
||
|
paddq %mm0,%mm2
|
||
|
paddq %mm1,%mm3
|
||
|
movq %mm2,%mm0
|
||
|
movd 36(%esp,%ecx,4),%mm6
|
||
|
pand %mm7,%mm0
|
||
|
movd 4(%ebp,%ecx,4),%mm1
|
||
|
paddq %mm0,%mm3
|
||
|
movd 4(%esi,%ecx,4),%mm0
|
||
|
psrlq $32,%mm2
|
||
|
movd %mm3,28(%esp,%ecx,4)
|
||
|
psrlq $32,%mm3
|
||
|
paddq %mm6,%mm2
|
||
|
decl %ebx
|
||
|
leal 1(%ecx),%ecx
|
||
|
jnz L007inner
|
||
|
movl %ecx,%ebx
|
||
|
pmuludq %mm4,%mm0
|
||
|
pmuludq %mm5,%mm1
|
||
|
paddq %mm0,%mm2
|
||
|
paddq %mm1,%mm3
|
||
|
movq %mm2,%mm0
|
||
|
pand %mm7,%mm0
|
||
|
paddq %mm0,%mm3
|
||
|
movd %mm3,28(%esp,%ecx,4)
|
||
|
psrlq $32,%mm2
|
||
|
psrlq $32,%mm3
|
||
|
movd 36(%esp,%ebx,4),%mm6
|
||
|
paddq %mm2,%mm3
|
||
|
paddq %mm6,%mm3
|
||
|
movq %mm3,32(%esp,%ebx,4)
|
||
|
leal 1(%edx),%edx
|
||
|
cmpl %ebx,%edx
|
||
|
jle L006outer
|
||
|
emms
|
||
|
jmp L008common_tail
|
||
|
.align 4,0x90
|
||
|
L004non_sse2:
|
||
|
movl 8(%esp),%esi
|
||
|
leal 1(%ebx),%ebp
|
||
|
movl 12(%esp),%edi
|
||
|
xorl %ecx,%ecx
|
||
|
movl %esi,%edx
|
||
|
andl $1,%ebp
|
||
|
subl %edi,%edx
|
||
|
leal 4(%edi,%ebx,4),%eax
|
||
|
orl %edx,%ebp
|
||
|
movl (%edi),%edi
|
||
|
jz L009bn_sqr_mont
|
||
|
movl %eax,28(%esp)
|
||
|
movl (%esi),%eax
|
||
|
xorl %edx,%edx
|
||
|
.align 4,0x90
|
||
|
L010mull:
|
||
|
movl %edx,%ebp
|
||
|
mull %edi
|
||
|
addl %eax,%ebp
|
||
|
leal 1(%ecx),%ecx
|
||
|
adcl $0,%edx
|
||
|
movl (%esi,%ecx,4),%eax
|
||
|
cmpl %ebx,%ecx
|
||
|
movl %ebp,28(%esp,%ecx,4)
|
||
|
jl L010mull
|
||
|
movl %edx,%ebp
|
||
|
mull %edi
|
||
|
movl 20(%esp),%edi
|
||
|
addl %ebp,%eax
|
||
|
movl 16(%esp),%esi
|
||
|
adcl $0,%edx
|
||
|
imull 32(%esp),%edi
|
||
|
movl %eax,32(%esp,%ebx,4)
|
||
|
xorl %ecx,%ecx
|
||
|
movl %edx,36(%esp,%ebx,4)
|
||
|
movl %ecx,40(%esp,%ebx,4)
|
||
|
movl (%esi),%eax
|
||
|
mull %edi
|
||
|
addl 32(%esp),%eax
|
||
|
movl 4(%esi),%eax
|
||
|
adcl $0,%edx
|
||
|
incl %ecx
|
||
|
jmp L0112ndmadd
|
||
|
.align 4,0x90
|
||
|
L0121stmadd:
|
||
|
movl %edx,%ebp
|
||
|
mull %edi
|
||
|
addl 32(%esp,%ecx,4),%ebp
|
||
|
leal 1(%ecx),%ecx
|
||
|
adcl $0,%edx
|
||
|
addl %eax,%ebp
|
||
|
movl (%esi,%ecx,4),%eax
|
||
|
adcl $0,%edx
|
||
|
cmpl %ebx,%ecx
|
||
|
movl %ebp,28(%esp,%ecx,4)
|
||
|
jl L0121stmadd
|
||
|
movl %edx,%ebp
|
||
|
mull %edi
|
||
|
addl 32(%esp,%ebx,4),%eax
|
||
|
movl 20(%esp),%edi
|
||
|
adcl $0,%edx
|
||
|
movl 16(%esp),%esi
|
||
|
addl %eax,%ebp
|
||
|
adcl $0,%edx
|
||
|
imull 32(%esp),%edi
|
||
|
xorl %ecx,%ecx
|
||
|
addl 36(%esp,%ebx,4),%edx
|
||
|
movl %ebp,32(%esp,%ebx,4)
|
||
|
adcl $0,%ecx
|
||
|
movl (%esi),%eax
|
||
|
movl %edx,36(%esp,%ebx,4)
|
||
|
movl %ecx,40(%esp,%ebx,4)
|
||
|
mull %edi
|
||
|
addl 32(%esp),%eax
|
||
|
movl 4(%esi),%eax
|
||
|
adcl $0,%edx
|
||
|
movl $1,%ecx
|
||
|
.align 4,0x90
|
||
|
L0112ndmadd:
|
||
|
movl %edx,%ebp
|
||
|
mull %edi
|
||
|
addl 32(%esp,%ecx,4),%ebp
|
||
|
leal 1(%ecx),%ecx
|
||
|
adcl $0,%edx
|
||
|
addl %eax,%ebp
|
||
|
movl (%esi,%ecx,4),%eax
|
||
|
adcl $0,%edx
|
||
|
cmpl %ebx,%ecx
|
||
|
movl %ebp,24(%esp,%ecx,4)
|
||
|
jl L0112ndmadd
|
||
|
movl %edx,%ebp
|
||
|
mull %edi
|
||
|
addl 32(%esp,%ebx,4),%ebp
|
||
|
adcl $0,%edx
|
||
|
addl %eax,%ebp
|
||
|
adcl $0,%edx
|
||
|
movl %ebp,28(%esp,%ebx,4)
|
||
|
xorl %eax,%eax
|
||
|
movl 12(%esp),%ecx
|
||
|
addl 36(%esp,%ebx,4),%edx
|
||
|
adcl 40(%esp,%ebx,4),%eax
|
||
|
leal 4(%ecx),%ecx
|
||
|
movl %edx,32(%esp,%ebx,4)
|
||
|
cmpl 28(%esp),%ecx
|
||
|
movl %eax,36(%esp,%ebx,4)
|
||
|
je L008common_tail
|
||
|
movl (%ecx),%edi
|
||
|
movl 8(%esp),%esi
|
||
|
movl %ecx,12(%esp)
|
||
|
xorl %ecx,%ecx
|
||
|
xorl %edx,%edx
|
||
|
movl (%esi),%eax
|
||
|
jmp L0121stmadd
|
||
|
.align 4,0x90
|
||
|
L009bn_sqr_mont:
|
||
|
movl %ebx,(%esp)
|
||
|
movl %ecx,12(%esp)
|
||
|
movl %edi,%eax
|
||
|
mull %edi
|
||
|
movl %eax,32(%esp)
|
||
|
movl %edx,%ebx
|
||
|
shrl $1,%edx
|
||
|
andl $1,%ebx
|
||
|
incl %ecx
|
||
|
.align 4,0x90
|
||
|
L013sqr:
|
||
|
movl (%esi,%ecx,4),%eax
|
||
|
movl %edx,%ebp
|
||
|
mull %edi
|
||
|
addl %ebp,%eax
|
||
|
leal 1(%ecx),%ecx
|
||
|
adcl $0,%edx
|
||
|
leal (%ebx,%eax,2),%ebp
|
||
|
shrl $31,%eax
|
||
|
cmpl (%esp),%ecx
|
||
|
movl %eax,%ebx
|
||
|
movl %ebp,28(%esp,%ecx,4)
|
||
|
jl L013sqr
|
||
|
movl (%esi,%ecx,4),%eax
|
||
|
movl %edx,%ebp
|
||
|
mull %edi
|
||
|
addl %ebp,%eax
|
||
|
movl 20(%esp),%edi
|
||
|
adcl $0,%edx
|
||
|
movl 16(%esp),%esi
|
||
|
leal (%ebx,%eax,2),%ebp
|
||
|
imull 32(%esp),%edi
|
||
|
shrl $31,%eax
|
||
|
movl %ebp,32(%esp,%ecx,4)
|
||
|
leal (%eax,%edx,2),%ebp
|
||
|
movl (%esi),%eax
|
||
|
shrl $31,%edx
|
||
|
movl %ebp,36(%esp,%ecx,4)
|
||
|
movl %edx,40(%esp,%ecx,4)
|
||
|
mull %edi
|
||
|
addl 32(%esp),%eax
|
||
|
movl %ecx,%ebx
|
||
|
adcl $0,%edx
|
||
|
movl 4(%esi),%eax
|
||
|
movl $1,%ecx
|
||
|
.align 4,0x90
|
||
|
L0143rdmadd:
|
||
|
movl %edx,%ebp
|
||
|
mull %edi
|
||
|
addl 32(%esp,%ecx,4),%ebp
|
||
|
adcl $0,%edx
|
||
|
addl %eax,%ebp
|
||
|
movl 4(%esi,%ecx,4),%eax
|
||
|
adcl $0,%edx
|
||
|
movl %ebp,28(%esp,%ecx,4)
|
||
|
movl %edx,%ebp
|
||
|
mull %edi
|
||
|
addl 36(%esp,%ecx,4),%ebp
|
||
|
leal 2(%ecx),%ecx
|
||
|
adcl $0,%edx
|
||
|
addl %eax,%ebp
|
||
|
movl (%esi,%ecx,4),%eax
|
||
|
adcl $0,%edx
|
||
|
cmpl %ebx,%ecx
|
||
|
movl %ebp,24(%esp,%ecx,4)
|
||
|
jl L0143rdmadd
|
||
|
movl %edx,%ebp
|
||
|
mull %edi
|
||
|
addl 32(%esp,%ebx,4),%ebp
|
||
|
adcl $0,%edx
|
||
|
addl %eax,%ebp
|
||
|
adcl $0,%edx
|
||
|
movl %ebp,28(%esp,%ebx,4)
|
||
|
movl 12(%esp),%ecx
|
||
|
xorl %eax,%eax
|
||
|
movl 8(%esp),%esi
|
||
|
addl 36(%esp,%ebx,4),%edx
|
||
|
adcl 40(%esp,%ebx,4),%eax
|
||
|
movl %edx,32(%esp,%ebx,4)
|
||
|
cmpl %ebx,%ecx
|
||
|
movl %eax,36(%esp,%ebx,4)
|
||
|
je L008common_tail
|
||
|
movl 4(%esi,%ecx,4),%edi
|
||
|
leal 1(%ecx),%ecx
|
||
|
movl %edi,%eax
|
||
|
movl %ecx,12(%esp)
|
||
|
mull %edi
|
||
|
addl 32(%esp,%ecx,4),%eax
|
||
|
adcl $0,%edx
|
||
|
movl %eax,32(%esp,%ecx,4)
|
||
|
xorl %ebp,%ebp
|
||
|
cmpl %ebx,%ecx
|
||
|
leal 1(%ecx),%ecx
|
||
|
je L015sqrlast
|
||
|
movl %edx,%ebx
|
||
|
shrl $1,%edx
|
||
|
andl $1,%ebx
|
||
|
.align 4,0x90
|
||
|
L016sqradd:
|
||
|
movl (%esi,%ecx,4),%eax
|
||
|
movl %edx,%ebp
|
||
|
mull %edi
|
||
|
addl %ebp,%eax
|
||
|
leal (%eax,%eax,1),%ebp
|
||
|
adcl $0,%edx
|
||
|
shrl $31,%eax
|
||
|
addl 32(%esp,%ecx,4),%ebp
|
||
|
leal 1(%ecx),%ecx
|
||
|
adcl $0,%eax
|
||
|
addl %ebx,%ebp
|
||
|
adcl $0,%eax
|
||
|
cmpl (%esp),%ecx
|
||
|
movl %ebp,28(%esp,%ecx,4)
|
||
|
movl %eax,%ebx
|
||
|
jle L016sqradd
|
||
|
movl %edx,%ebp
|
||
|
addl %edx,%edx
|
||
|
shrl $31,%ebp
|
||
|
addl %ebx,%edx
|
||
|
adcl $0,%ebp
|
||
|
L015sqrlast:
|
||
|
movl 20(%esp),%edi
|
||
|
movl 16(%esp),%esi
|
||
|
imull 32(%esp),%edi
|
||
|
addl 32(%esp,%ecx,4),%edx
|
||
|
movl (%esi),%eax
|
||
|
adcl $0,%ebp
|
||
|
movl %edx,32(%esp,%ecx,4)
|
||
|
movl %ebp,36(%esp,%ecx,4)
|
||
|
mull %edi
|
||
|
addl 32(%esp),%eax
|
||
|
leal -1(%ecx),%ebx
|
||
|
adcl $0,%edx
|
||
|
movl $1,%ecx
|
||
|
movl 4(%esi),%eax
|
||
|
jmp L0143rdmadd
|
||
|
.align 4,0x90
|
||
|
L008common_tail:
|
||
|
movl 16(%esp),%ebp
|
||
|
movl 4(%esp),%edi
|
||
|
leal 32(%esp),%esi
|
||
|
movl (%esi),%eax
|
||
|
movl %ebx,%ecx
|
||
|
xorl %edx,%edx
|
||
|
.align 4,0x90
|
||
|
L017sub:
|
||
|
sbbl (%ebp,%edx,4),%eax
|
||
|
movl %eax,(%edi,%edx,4)
|
||
|
decl %ecx
|
||
|
movl 4(%esi,%edx,4),%eax
|
||
|
leal 1(%edx),%edx
|
||
|
jge L017sub
|
||
|
sbbl $0,%eax
|
||
|
movl $-1,%edx
|
||
|
xorl %eax,%edx
|
||
|
jmp L018copy
|
||
|
.align 4,0x90
|
||
|
L018copy:
|
||
|
movl 32(%esp,%ebx,4),%esi
|
||
|
movl (%edi,%ebx,4),%ebp
|
||
|
movl %ecx,32(%esp,%ebx,4)
|
||
|
andl %eax,%esi
|
||
|
andl %edx,%ebp
|
||
|
orl %esi,%ebp
|
||
|
movl %ebp,(%edi,%ebx,4)
|
||
|
decl %ebx
|
||
|
jge L018copy
|
||
|
movl 24(%esp),%esp
|
||
|
movl $1,%eax
|
||
|
L000just_leave:
|
||
|
popl %edi
|
||
|
popl %esi
|
||
|
popl %ebx
|
||
|
popl %ebp
|
||
|
ret
|
||
|
.byte 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
|
||
|
.byte 112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
|
||
|
.byte 54,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121
|
||
|
.byte 32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46
|
||
|
.byte 111,114,103,62,0
|
||
|
.section __IMPORT,__pointers,non_lazy_symbol_pointers
|
||
|
L_OPENSSL_ia32cap_P$non_lazy_ptr:
|
||
|
.indirect_symbol _OPENSSL_ia32cap_P
|
||
|
.long 0
|
||
|
#endif
|