| Index: third_party/boringssl/win-x86_64/crypto/bn/x86_64-mont.asm
|
| diff --git a/third_party/boringssl/win-x86_64/crypto/bn/x86_64-mont.asm b/third_party/boringssl/win-x86_64/crypto/bn/x86_64-mont.asm
|
| new file mode 100644
|
| index 0000000000000000000000000000000000000000..ec8ccc4712477522d49524a700b066939207dd5d
|
| --- /dev/null
|
| +++ b/third_party/boringssl/win-x86_64/crypto/bn/x86_64-mont.asm
|
| @@ -0,0 +1,946 @@
|
| +OPTION DOTNAME
|
| +.text$ SEGMENT ALIGN(64) 'CODE'
|
| +
|
| +EXTERN OPENSSL_ia32cap_P:NEAR
|
| +
|
| +PUBLIC bn_mul_mont
|
| +
|
| +ALIGN 16
|
| +bn_mul_mont PROC PUBLIC
|
| + mov QWORD PTR[8+rsp],rdi ;WIN64 prologue
|
| + mov QWORD PTR[16+rsp],rsi
|
| + mov rax,rsp
|
| +$L$SEH_begin_bn_mul_mont::
|
| + mov rdi,rcx
|
| + mov rsi,rdx
|
| + mov rdx,r8
|
| + mov rcx,r9
|
| + mov r8,QWORD PTR[40+rsp]
|
| + mov r9,QWORD PTR[48+rsp]
|
| +
|
| +
|
| + test r9d,3
|
| + jnz $L$mul_enter
|
| + cmp r9d,8
|
| + jb $L$mul_enter
|
| + cmp rdx,rsi
|
| + jne $L$mul4x_enter
|
| + test r9d,7
|
| + jz $L$sqr8x_enter
|
| + jmp $L$mul4x_enter
|
| +
|
| +ALIGN 16
|
| +$L$mul_enter::
|
| + push rbx
|
| + push rbp
|
| + push r12
|
| + push r13
|
| + push r14
|
| + push r15
|
| +
|
| + mov r9d,r9d
|
| + lea r10,QWORD PTR[2+r9]
|
| + mov r11,rsp
|
| + neg r10
|
| + lea rsp,QWORD PTR[r10*8+rsp]
|
| + and rsp,-1024
|
| +
|
| + mov QWORD PTR[8+r9*8+rsp],r11
|
| +$L$mul_body::
|
| + mov r12,rdx
|
| + mov r8,QWORD PTR[r8]
|
| + mov rbx,QWORD PTR[r12]
|
| + mov rax,QWORD PTR[rsi]
|
| +
|
| + xor r14,r14
|
| + xor r15,r15
|
| +
|
| + mov rbp,r8
|
| + mul rbx
|
| + mov r10,rax
|
| + mov rax,QWORD PTR[rcx]
|
| +
|
| + imul rbp,r10
|
| + mov r11,rdx
|
| +
|
| + mul rbp
|
| + add r10,rax
|
| + mov rax,QWORD PTR[8+rsi]
|
| + adc rdx,0
|
| + mov r13,rdx
|
| +
|
| + lea r15,QWORD PTR[1+r15]
|
| + jmp $L$1st_enter
|
| +
|
| +ALIGN 16
|
| +$L$1st::
|
| + add r13,rax
|
| + mov rax,QWORD PTR[r15*8+rsi]
|
| + adc rdx,0
|
| + add r13,r11
|
| + mov r11,r10
|
| + adc rdx,0
|
| + mov QWORD PTR[((-16))+r15*8+rsp],r13
|
| + mov r13,rdx
|
| +
|
| +$L$1st_enter::
|
| + mul rbx
|
| + add r11,rax
|
| + mov rax,QWORD PTR[r15*8+rcx]
|
| + adc rdx,0
|
| + lea r15,QWORD PTR[1+r15]
|
| + mov r10,rdx
|
| +
|
| + mul rbp
|
| + cmp r15,r9
|
| + jne $L$1st
|
| +
|
| + add r13,rax
|
| + mov rax,QWORD PTR[rsi]
|
| + adc rdx,0
|
| + add r13,r11
|
| + adc rdx,0
|
| + mov QWORD PTR[((-16))+r15*8+rsp],r13
|
| + mov r13,rdx
|
| + mov r11,r10
|
| +
|
| + xor rdx,rdx
|
| + add r13,r11
|
| + adc rdx,0
|
| + mov QWORD PTR[((-8))+r9*8+rsp],r13
|
| + mov QWORD PTR[r9*8+rsp],rdx
|
| +
|
| + lea r14,QWORD PTR[1+r14]
|
| + jmp $L$outer
|
| +ALIGN 16
|
| +$L$outer::
|
| + mov rbx,QWORD PTR[r14*8+r12]
|
| + xor r15,r15
|
| + mov rbp,r8
|
| + mov r10,QWORD PTR[rsp]
|
| + mul rbx
|
| + add r10,rax
|
| + mov rax,QWORD PTR[rcx]
|
| + adc rdx,0
|
| +
|
| + imul rbp,r10
|
| + mov r11,rdx
|
| +
|
| + mul rbp
|
| + add r10,rax
|
| + mov rax,QWORD PTR[8+rsi]
|
| + adc rdx,0
|
| + mov r10,QWORD PTR[8+rsp]
|
| + mov r13,rdx
|
| +
|
| + lea r15,QWORD PTR[1+r15]
|
| + jmp $L$inner_enter
|
| +
|
| +ALIGN 16
|
| +$L$inner::
|
| + add r13,rax
|
| + mov rax,QWORD PTR[r15*8+rsi]
|
| + adc rdx,0
|
| + add r13,r10
|
| + mov r10,QWORD PTR[r15*8+rsp]
|
| + adc rdx,0
|
| + mov QWORD PTR[((-16))+r15*8+rsp],r13
|
| + mov r13,rdx
|
| +
|
| +$L$inner_enter::
|
| + mul rbx
|
| + add r11,rax
|
| + mov rax,QWORD PTR[r15*8+rcx]
|
| + adc rdx,0
|
| + add r10,r11
|
| + mov r11,rdx
|
| + adc r11,0
|
| + lea r15,QWORD PTR[1+r15]
|
| +
|
| + mul rbp
|
| + cmp r15,r9
|
| + jne $L$inner
|
| +
|
| + add r13,rax
|
| + mov rax,QWORD PTR[rsi]
|
| + adc rdx,0
|
| + add r13,r10
|
| + mov r10,QWORD PTR[r15*8+rsp]
|
| + adc rdx,0
|
| + mov QWORD PTR[((-16))+r15*8+rsp],r13
|
| + mov r13,rdx
|
| +
|
| + xor rdx,rdx
|
| + add r13,r11
|
| + adc rdx,0
|
| + add r13,r10
|
| + adc rdx,0
|
| + mov QWORD PTR[((-8))+r9*8+rsp],r13
|
| + mov QWORD PTR[r9*8+rsp],rdx
|
| +
|
| + lea r14,QWORD PTR[1+r14]
|
| + cmp r14,r9
|
| + jb $L$outer
|
| +
|
| + xor r14,r14
|
| + mov rax,QWORD PTR[rsp]
|
| + lea rsi,QWORD PTR[rsp]
|
| + mov r15,r9
|
| + jmp $L$sub
|
| +ALIGN 16
|
| +$L$sub:: sbb rax,QWORD PTR[r14*8+rcx]
|
| + mov QWORD PTR[r14*8+rdi],rax
|
| + mov rax,QWORD PTR[8+r14*8+rsi]
|
| + lea r14,QWORD PTR[1+r14]
|
| + dec r15
|
| + jnz $L$sub
|
| +
|
| + sbb rax,0
|
| + xor r14,r14
|
| + mov r15,r9
|
| +ALIGN 16
|
| +$L$copy::
|
| + mov rsi,QWORD PTR[r14*8+rsp]
|
| + mov rcx,QWORD PTR[r14*8+rdi]
|
| + xor rsi,rcx
|
| + and rsi,rax
|
| + xor rsi,rcx
|
| + mov QWORD PTR[r14*8+rsp],r14
|
| + mov QWORD PTR[r14*8+rdi],rsi
|
| + lea r14,QWORD PTR[1+r14]
|
| + sub r15,1
|
| + jnz $L$copy
|
| +
|
| + mov rsi,QWORD PTR[8+r9*8+rsp]
|
| + mov rax,1
|
| + mov r15,QWORD PTR[rsi]
|
| + mov r14,QWORD PTR[8+rsi]
|
| + mov r13,QWORD PTR[16+rsi]
|
| + mov r12,QWORD PTR[24+rsi]
|
| + mov rbp,QWORD PTR[32+rsi]
|
| + mov rbx,QWORD PTR[40+rsi]
|
| + lea rsp,QWORD PTR[48+rsi]
|
| +$L$mul_epilogue::
|
| + mov rdi,QWORD PTR[8+rsp] ;WIN64 epilogue
|
| + mov rsi,QWORD PTR[16+rsp]
|
| + DB 0F3h,0C3h ;repret
|
| +$L$SEH_end_bn_mul_mont::
|
| +bn_mul_mont ENDP
|
| +
|
| +ALIGN 16
|
| +bn_mul4x_mont PROC PRIVATE
|
| + mov QWORD PTR[8+rsp],rdi ;WIN64 prologue
|
| + mov QWORD PTR[16+rsp],rsi
|
| + mov rax,rsp
|
| +$L$SEH_begin_bn_mul4x_mont::
|
| + mov rdi,rcx
|
| + mov rsi,rdx
|
| + mov rdx,r8
|
| + mov rcx,r9
|
| + mov r8,QWORD PTR[40+rsp]
|
| + mov r9,QWORD PTR[48+rsp]
|
| +
|
| +
|
| +$L$mul4x_enter::
|
| + push rbx
|
| + push rbp
|
| + push r12
|
| + push r13
|
| + push r14
|
| + push r15
|
| +
|
| + mov r9d,r9d
|
| + lea r10,QWORD PTR[4+r9]
|
| + mov r11,rsp
|
| + neg r10
|
| + lea rsp,QWORD PTR[r10*8+rsp]
|
| + and rsp,-1024
|
| +
|
| + mov QWORD PTR[8+r9*8+rsp],r11
|
| +$L$mul4x_body::
|
| + mov QWORD PTR[16+r9*8+rsp],rdi
|
| + mov r12,rdx
|
| + mov r8,QWORD PTR[r8]
|
| + mov rbx,QWORD PTR[r12]
|
| + mov rax,QWORD PTR[rsi]
|
| +
|
| + xor r14,r14
|
| + xor r15,r15
|
| +
|
| + mov rbp,r8
|
| + mul rbx
|
| + mov r10,rax
|
| + mov rax,QWORD PTR[rcx]
|
| +
|
| + imul rbp,r10
|
| + mov r11,rdx
|
| +
|
| + mul rbp
|
| + add r10,rax
|
| + mov rax,QWORD PTR[8+rsi]
|
| + adc rdx,0
|
| + mov rdi,rdx
|
| +
|
| + mul rbx
|
| + add r11,rax
|
| + mov rax,QWORD PTR[8+rcx]
|
| + adc rdx,0
|
| + mov r10,rdx
|
| +
|
| + mul rbp
|
| + add rdi,rax
|
| + mov rax,QWORD PTR[16+rsi]
|
| + adc rdx,0
|
| + add rdi,r11
|
| + lea r15,QWORD PTR[4+r15]
|
| + adc rdx,0
|
| + mov QWORD PTR[rsp],rdi
|
| + mov r13,rdx
|
| + jmp $L$1st4x
|
| +ALIGN 16
|
| +$L$1st4x::
|
| + mul rbx
|
| + add r10,rax
|
| + mov rax,QWORD PTR[((-16))+r15*8+rcx]
|
| + adc rdx,0
|
| + mov r11,rdx
|
| +
|
| + mul rbp
|
| + add r13,rax
|
| + mov rax,QWORD PTR[((-8))+r15*8+rsi]
|
| + adc rdx,0
|
| + add r13,r10
|
| + adc rdx,0
|
| + mov QWORD PTR[((-24))+r15*8+rsp],r13
|
| + mov rdi,rdx
|
| +
|
| + mul rbx
|
| + add r11,rax
|
| + mov rax,QWORD PTR[((-8))+r15*8+rcx]
|
| + adc rdx,0
|
| + mov r10,rdx
|
| +
|
| + mul rbp
|
| + add rdi,rax
|
| + mov rax,QWORD PTR[r15*8+rsi]
|
| + adc rdx,0
|
| + add rdi,r11
|
| + adc rdx,0
|
| + mov QWORD PTR[((-16))+r15*8+rsp],rdi
|
| + mov r13,rdx
|
| +
|
| + mul rbx
|
| + add r10,rax
|
| + mov rax,QWORD PTR[r15*8+rcx]
|
| + adc rdx,0
|
| + mov r11,rdx
|
| +
|
| + mul rbp
|
| + add r13,rax
|
| + mov rax,QWORD PTR[8+r15*8+rsi]
|
| + adc rdx,0
|
| + add r13,r10
|
| + adc rdx,0
|
| + mov QWORD PTR[((-8))+r15*8+rsp],r13
|
| + mov rdi,rdx
|
| +
|
| + mul rbx
|
| + add r11,rax
|
| + mov rax,QWORD PTR[8+r15*8+rcx]
|
| + adc rdx,0
|
| + lea r15,QWORD PTR[4+r15]
|
| + mov r10,rdx
|
| +
|
| + mul rbp
|
| + add rdi,rax
|
| + mov rax,QWORD PTR[((-16))+r15*8+rsi]
|
| + adc rdx,0
|
| + add rdi,r11
|
| + adc rdx,0
|
| + mov QWORD PTR[((-32))+r15*8+rsp],rdi
|
| + mov r13,rdx
|
| + cmp r15,r9
|
| + jb $L$1st4x
|
| +
|
| + mul rbx
|
| + add r10,rax
|
| + mov rax,QWORD PTR[((-16))+r15*8+rcx]
|
| + adc rdx,0
|
| + mov r11,rdx
|
| +
|
| + mul rbp
|
| + add r13,rax
|
| + mov rax,QWORD PTR[((-8))+r15*8+rsi]
|
| + adc rdx,0
|
| + add r13,r10
|
| + adc rdx,0
|
| + mov QWORD PTR[((-24))+r15*8+rsp],r13
|
| + mov rdi,rdx
|
| +
|
| + mul rbx
|
| + add r11,rax
|
| + mov rax,QWORD PTR[((-8))+r15*8+rcx]
|
| + adc rdx,0
|
| + mov r10,rdx
|
| +
|
| + mul rbp
|
| + add rdi,rax
|
| + mov rax,QWORD PTR[rsi]
|
| + adc rdx,0
|
| + add rdi,r11
|
| + adc rdx,0
|
| + mov QWORD PTR[((-16))+r15*8+rsp],rdi
|
| + mov r13,rdx
|
| +
|
| + xor rdi,rdi
|
| + add r13,r10
|
| + adc rdi,0
|
| + mov QWORD PTR[((-8))+r15*8+rsp],r13
|
| + mov QWORD PTR[r15*8+rsp],rdi
|
| +
|
| + lea r14,QWORD PTR[1+r14]
|
| +ALIGN 4
|
| +$L$outer4x::
|
| + mov rbx,QWORD PTR[r14*8+r12]
|
| + xor r15,r15
|
| + mov r10,QWORD PTR[rsp]
|
| + mov rbp,r8
|
| + mul rbx
|
| + add r10,rax
|
| + mov rax,QWORD PTR[rcx]
|
| + adc rdx,0
|
| +
|
| + imul rbp,r10
|
| + mov r11,rdx
|
| +
|
| + mul rbp
|
| + add r10,rax
|
| + mov rax,QWORD PTR[8+rsi]
|
| + adc rdx,0
|
| + mov rdi,rdx
|
| +
|
| + mul rbx
|
| + add r11,rax
|
| + mov rax,QWORD PTR[8+rcx]
|
| + adc rdx,0
|
| + add r11,QWORD PTR[8+rsp]
|
| + adc rdx,0
|
| + mov r10,rdx
|
| +
|
| + mul rbp
|
| + add rdi,rax
|
| + mov rax,QWORD PTR[16+rsi]
|
| + adc rdx,0
|
| + add rdi,r11
|
| + lea r15,QWORD PTR[4+r15]
|
| + adc rdx,0
|
| + mov QWORD PTR[rsp],rdi
|
| + mov r13,rdx
|
| + jmp $L$inner4x
|
| +ALIGN 16
|
| +$L$inner4x::
|
| + mul rbx
|
| + add r10,rax
|
| + mov rax,QWORD PTR[((-16))+r15*8+rcx]
|
| + adc rdx,0
|
| + add r10,QWORD PTR[((-16))+r15*8+rsp]
|
| + adc rdx,0
|
| + mov r11,rdx
|
| +
|
| + mul rbp
|
| + add r13,rax
|
| + mov rax,QWORD PTR[((-8))+r15*8+rsi]
|
| + adc rdx,0
|
| + add r13,r10
|
| + adc rdx,0
|
| + mov QWORD PTR[((-24))+r15*8+rsp],r13
|
| + mov rdi,rdx
|
| +
|
| + mul rbx
|
| + add r11,rax
|
| + mov rax,QWORD PTR[((-8))+r15*8+rcx]
|
| + adc rdx,0
|
| + add r11,QWORD PTR[((-8))+r15*8+rsp]
|
| + adc rdx,0
|
| + mov r10,rdx
|
| +
|
| + mul rbp
|
| + add rdi,rax
|
| + mov rax,QWORD PTR[r15*8+rsi]
|
| + adc rdx,0
|
| + add rdi,r11
|
| + adc rdx,0
|
| + mov QWORD PTR[((-16))+r15*8+rsp],rdi
|
| + mov r13,rdx
|
| +
|
| + mul rbx
|
| + add r10,rax
|
| + mov rax,QWORD PTR[r15*8+rcx]
|
| + adc rdx,0
|
| + add r10,QWORD PTR[r15*8+rsp]
|
| + adc rdx,0
|
| + mov r11,rdx
|
| +
|
| + mul rbp
|
| + add r13,rax
|
| + mov rax,QWORD PTR[8+r15*8+rsi]
|
| + adc rdx,0
|
| + add r13,r10
|
| + adc rdx,0
|
| + mov QWORD PTR[((-8))+r15*8+rsp],r13
|
| + mov rdi,rdx
|
| +
|
| + mul rbx
|
| + add r11,rax
|
| + mov rax,QWORD PTR[8+r15*8+rcx]
|
| + adc rdx,0
|
| + add r11,QWORD PTR[8+r15*8+rsp]
|
| + adc rdx,0
|
| + lea r15,QWORD PTR[4+r15]
|
| + mov r10,rdx
|
| +
|
| + mul rbp
|
| + add rdi,rax
|
| + mov rax,QWORD PTR[((-16))+r15*8+rsi]
|
| + adc rdx,0
|
| + add rdi,r11
|
| + adc rdx,0
|
| + mov QWORD PTR[((-32))+r15*8+rsp],rdi
|
| + mov r13,rdx
|
| + cmp r15,r9
|
| + jb $L$inner4x
|
| +
|
| + mul rbx
|
| + add r10,rax
|
| + mov rax,QWORD PTR[((-16))+r15*8+rcx]
|
| + adc rdx,0
|
| + add r10,QWORD PTR[((-16))+r15*8+rsp]
|
| + adc rdx,0
|
| + mov r11,rdx
|
| +
|
| + mul rbp
|
| + add r13,rax
|
| + mov rax,QWORD PTR[((-8))+r15*8+rsi]
|
| + adc rdx,0
|
| + add r13,r10
|
| + adc rdx,0
|
| + mov QWORD PTR[((-24))+r15*8+rsp],r13
|
| + mov rdi,rdx
|
| +
|
| + mul rbx
|
| + add r11,rax
|
| + mov rax,QWORD PTR[((-8))+r15*8+rcx]
|
| + adc rdx,0
|
| + add r11,QWORD PTR[((-8))+r15*8+rsp]
|
| + adc rdx,0
|
| + lea r14,QWORD PTR[1+r14]
|
| + mov r10,rdx
|
| +
|
| + mul rbp
|
| + add rdi,rax
|
| + mov rax,QWORD PTR[rsi]
|
| + adc rdx,0
|
| + add rdi,r11
|
| + adc rdx,0
|
| + mov QWORD PTR[((-16))+r15*8+rsp],rdi
|
| + mov r13,rdx
|
| +
|
| + xor rdi,rdi
|
| + add r13,r10
|
| + adc rdi,0
|
| + add r13,QWORD PTR[r9*8+rsp]
|
| + adc rdi,0
|
| + mov QWORD PTR[((-8))+r15*8+rsp],r13
|
| + mov QWORD PTR[r15*8+rsp],rdi
|
| +
|
| + cmp r14,r9
|
| + jb $L$outer4x
|
| + mov rdi,QWORD PTR[16+r9*8+rsp]
|
| + mov rax,QWORD PTR[rsp]
|
| + mov rdx,QWORD PTR[8+rsp]
|
| + shr r9,2
|
| + lea rsi,QWORD PTR[rsp]
|
| + xor r14,r14
|
| +
|
| + sub rax,QWORD PTR[rcx]
|
| + mov rbx,QWORD PTR[16+rsi]
|
| + mov rbp,QWORD PTR[24+rsi]
|
| + sbb rdx,QWORD PTR[8+rcx]
|
| + lea r15,QWORD PTR[((-1))+r9]
|
| + jmp $L$sub4x
|
| +ALIGN 16
|
| +$L$sub4x::
|
| + mov QWORD PTR[r14*8+rdi],rax
|
| + mov QWORD PTR[8+r14*8+rdi],rdx
|
| + sbb rbx,QWORD PTR[16+r14*8+rcx]
|
| + mov rax,QWORD PTR[32+r14*8+rsi]
|
| + mov rdx,QWORD PTR[40+r14*8+rsi]
|
| + sbb rbp,QWORD PTR[24+r14*8+rcx]
|
| + mov QWORD PTR[16+r14*8+rdi],rbx
|
| + mov QWORD PTR[24+r14*8+rdi],rbp
|
| + sbb rax,QWORD PTR[32+r14*8+rcx]
|
| + mov rbx,QWORD PTR[48+r14*8+rsi]
|
| + mov rbp,QWORD PTR[56+r14*8+rsi]
|
| + sbb rdx,QWORD PTR[40+r14*8+rcx]
|
| + lea r14,QWORD PTR[4+r14]
|
| + dec r15
|
| + jnz $L$sub4x
|
| +
|
| + mov QWORD PTR[r14*8+rdi],rax
|
| + mov rax,QWORD PTR[32+r14*8+rsi]
|
| + sbb rbx,QWORD PTR[16+r14*8+rcx]
|
| + mov QWORD PTR[8+r14*8+rdi],rdx
|
| + sbb rbp,QWORD PTR[24+r14*8+rcx]
|
| + mov QWORD PTR[16+r14*8+rdi],rbx
|
| +
|
| + sbb rax,0
|
| +DB 66h, 48h, 0fh, 6eh, 0c0h
|
| + punpcklqdq xmm0,xmm0
|
| + mov QWORD PTR[24+r14*8+rdi],rbp
|
| + xor r14,r14
|
| +
|
| + mov r15,r9
|
| + pxor xmm5,xmm5
|
| + jmp $L$copy4x
|
| +ALIGN 16
|
| +$L$copy4x::
|
| + movdqu xmm2,XMMWORD PTR[r14*1+rsp]
|
| + movdqu xmm4,XMMWORD PTR[16+r14*1+rsp]
|
| + movdqu xmm1,XMMWORD PTR[r14*1+rdi]
|
| + movdqu xmm3,XMMWORD PTR[16+r14*1+rdi]
|
| + pxor xmm2,xmm1
|
| + pxor xmm4,xmm3
|
| + pand xmm2,xmm0
|
| + pand xmm4,xmm0
|
| + pxor xmm2,xmm1
|
| + pxor xmm4,xmm3
|
| + movdqu XMMWORD PTR[r14*1+rdi],xmm2
|
| + movdqu XMMWORD PTR[16+r14*1+rdi],xmm4
|
| + movdqa XMMWORD PTR[r14*1+rsp],xmm5
|
| + movdqa XMMWORD PTR[16+r14*1+rsp],xmm5
|
| +
|
| + lea r14,QWORD PTR[32+r14]
|
| + dec r15
|
| + jnz $L$copy4x
|
| +
|
| + shl r9,2
|
| + mov rsi,QWORD PTR[8+r9*8+rsp]
|
| + mov rax,1
|
| + mov r15,QWORD PTR[rsi]
|
| + mov r14,QWORD PTR[8+rsi]
|
| + mov r13,QWORD PTR[16+rsi]
|
| + mov r12,QWORD PTR[24+rsi]
|
| + mov rbp,QWORD PTR[32+rsi]
|
| + mov rbx,QWORD PTR[40+rsi]
|
| + lea rsp,QWORD PTR[48+rsi]
|
| +$L$mul4x_epilogue::
|
| + mov rdi,QWORD PTR[8+rsp] ;WIN64 epilogue
|
| + mov rsi,QWORD PTR[16+rsp]
|
| + DB 0F3h,0C3h ;repret
|
| +$L$SEH_end_bn_mul4x_mont::
|
| +bn_mul4x_mont ENDP
|
| +EXTERN bn_sqr8x_internal:NEAR
|
| +EXTERN bn_sqrx8x_internal:NEAR
|
| +
|
| +
|
| +ALIGN 32
|
| +bn_sqr8x_mont PROC PRIVATE
|
| + mov QWORD PTR[8+rsp],rdi ;WIN64 prologue
|
| + mov QWORD PTR[16+rsp],rsi
|
| + mov rax,rsp
|
| +$L$SEH_begin_bn_sqr8x_mont::
|
| + mov rdi,rcx
|
| + mov rsi,rdx
|
| + mov rdx,r8
|
| + mov rcx,r9
|
| + mov r8,QWORD PTR[40+rsp]
|
| + mov r9,QWORD PTR[48+rsp]
|
| +
|
| +
|
| +$L$sqr8x_enter::
|
| + mov rax,rsp
|
| + push rbx
|
| + push rbp
|
| + push r12
|
| + push r13
|
| + push r14
|
| + push r15
|
| +
|
| + mov r10d,r9d
|
| + shl r9d,3
|
| + shl r10,3+2
|
| + neg r9
|
| +
|
| +
|
| +
|
| +
|
| +
|
| +
|
| + lea r11,QWORD PTR[((-64))+r9*4+rsp]
|
| + mov r8,QWORD PTR[r8]
|
| + sub r11,rsi
|
| + and r11,4095
|
| + cmp r10,r11
|
| + jb $L$sqr8x_sp_alt
|
| + sub rsp,r11
|
| + lea rsp,QWORD PTR[((-64))+r9*4+rsp]
|
| + jmp $L$sqr8x_sp_done
|
| +
|
| +ALIGN 32
|
| +$L$sqr8x_sp_alt::
|
| + lea r10,QWORD PTR[((4096-64))+r9*4]
|
| + lea rsp,QWORD PTR[((-64))+r9*4+rsp]
|
| + sub r11,r10
|
| + mov r10,0
|
| + cmovc r11,r10
|
| + sub rsp,r11
|
| +$L$sqr8x_sp_done::
|
| + and rsp,-64
|
| + mov r10,r9
|
| + neg r9
|
| +
|
| + lea r11,QWORD PTR[64+r9*2+rsp]
|
| + mov QWORD PTR[32+rsp],r8
|
| + mov QWORD PTR[40+rsp],rax
|
| +$L$sqr8x_body::
|
| +
|
| + mov rbp,r9
|
| +DB 102,73,15,110,211
|
| + shr rbp,3+2
|
| + mov eax,DWORD PTR[((OPENSSL_ia32cap_P+8))]
|
| + jmp $L$sqr8x_copy_n
|
| +
|
| +ALIGN 32
|
| +$L$sqr8x_copy_n::
|
| + movq xmm0,QWORD PTR[rcx]
|
| + movq xmm1,QWORD PTR[8+rcx]
|
| + movq xmm3,QWORD PTR[16+rcx]
|
| + movq xmm4,QWORD PTR[24+rcx]
|
| + lea rcx,QWORD PTR[32+rcx]
|
| + movdqa XMMWORD PTR[r11],xmm0
|
| + movdqa XMMWORD PTR[16+r11],xmm1
|
| + movdqa XMMWORD PTR[32+r11],xmm3
|
| + movdqa XMMWORD PTR[48+r11],xmm4
|
| + lea r11,QWORD PTR[64+r11]
|
| + dec rbp
|
| + jnz $L$sqr8x_copy_n
|
| +
|
| + pxor xmm0,xmm0
|
| +DB 102,72,15,110,207
|
| +DB 102,73,15,110,218
|
| + call bn_sqr8x_internal
|
| +
|
| + pxor xmm0,xmm0
|
| + lea rax,QWORD PTR[48+rsp]
|
| + lea rdx,QWORD PTR[64+r9*2+rsp]
|
| + shr r9,3+2
|
| + mov rsi,QWORD PTR[40+rsp]
|
| + jmp $L$sqr8x_zero
|
| +
|
| +ALIGN 32
|
| +$L$sqr8x_zero::
|
| + movdqa XMMWORD PTR[rax],xmm0
|
| + movdqa XMMWORD PTR[16+rax],xmm0
|
| + movdqa XMMWORD PTR[32+rax],xmm0
|
| + movdqa XMMWORD PTR[48+rax],xmm0
|
| + lea rax,QWORD PTR[64+rax]
|
| + movdqa XMMWORD PTR[rdx],xmm0
|
| + movdqa XMMWORD PTR[16+rdx],xmm0
|
| + movdqa XMMWORD PTR[32+rdx],xmm0
|
| + movdqa XMMWORD PTR[48+rdx],xmm0
|
| + lea rdx,QWORD PTR[64+rdx]
|
| + dec r9
|
| + jnz $L$sqr8x_zero
|
| +
|
| + mov rax,1
|
| + mov r15,QWORD PTR[((-48))+rsi]
|
| + mov r14,QWORD PTR[((-40))+rsi]
|
| + mov r13,QWORD PTR[((-32))+rsi]
|
| + mov r12,QWORD PTR[((-24))+rsi]
|
| + mov rbp,QWORD PTR[((-16))+rsi]
|
| + mov rbx,QWORD PTR[((-8))+rsi]
|
| + lea rsp,QWORD PTR[rsi]
|
| +$L$sqr8x_epilogue::
|
| + mov rdi,QWORD PTR[8+rsp] ;WIN64 epilogue
|
| + mov rsi,QWORD PTR[16+rsp]
|
| + DB 0F3h,0C3h ;repret
|
| +$L$SEH_end_bn_sqr8x_mont::
|
| +bn_sqr8x_mont ENDP
|
| +DB 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
|
| +DB 112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
|
| +DB 54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83
|
| +DB 32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115
|
| +DB 115,108,46,111,114,103,62,0
|
| +ALIGN 16
|
| +EXTERN __imp_RtlVirtualUnwind:NEAR
|
| +
|
| +ALIGN 16
|
| +mul_handler PROC PRIVATE
|
| + push rsi
|
| + push rdi
|
| + push rbx
|
| + push rbp
|
| + push r12
|
| + push r13
|
| + push r14
|
| + push r15
|
| + pushfq
|
| + sub rsp,64
|
| +
|
| + mov rax,QWORD PTR[120+r8]
|
| + mov rbx,QWORD PTR[248+r8]
|
| +
|
| + mov rsi,QWORD PTR[8+r9]
|
| + mov r11,QWORD PTR[56+r9]
|
| +
|
| + mov r10d,DWORD PTR[r11]
|
| + lea r10,QWORD PTR[r10*1+rsi]
|
| + cmp rbx,r10
|
| + jb $L$common_seh_tail
|
| +
|
| + mov rax,QWORD PTR[152+r8]
|
| +
|
| + mov r10d,DWORD PTR[4+r11]
|
| + lea r10,QWORD PTR[r10*1+rsi]
|
| + cmp rbx,r10
|
| + jae $L$common_seh_tail
|
| +
|
| + mov r10,QWORD PTR[192+r8]
|
| + mov rax,QWORD PTR[8+r10*8+rax]
|
| + lea rax,QWORD PTR[48+rax]
|
| +
|
| + mov rbx,QWORD PTR[((-8))+rax]
|
| + mov rbp,QWORD PTR[((-16))+rax]
|
| + mov r12,QWORD PTR[((-24))+rax]
|
| + mov r13,QWORD PTR[((-32))+rax]
|
| + mov r14,QWORD PTR[((-40))+rax]
|
| + mov r15,QWORD PTR[((-48))+rax]
|
| + mov QWORD PTR[144+r8],rbx
|
| + mov QWORD PTR[160+r8],rbp
|
| + mov QWORD PTR[216+r8],r12
|
| + mov QWORD PTR[224+r8],r13
|
| + mov QWORD PTR[232+r8],r14
|
| + mov QWORD PTR[240+r8],r15
|
| +
|
| + jmp $L$common_seh_tail
|
| +mul_handler ENDP
|
| +
|
| +
|
| +ALIGN 16
|
| +sqr_handler PROC PRIVATE
|
| + push rsi
|
| + push rdi
|
| + push rbx
|
| + push rbp
|
| + push r12
|
| + push r13
|
| + push r14
|
| + push r15
|
| + pushfq
|
| + sub rsp,64
|
| +
|
| + mov rax,QWORD PTR[120+r8]
|
| + mov rbx,QWORD PTR[248+r8]
|
| +
|
| + mov rsi,QWORD PTR[8+r9]
|
| + mov r11,QWORD PTR[56+r9]
|
| +
|
| + mov r10d,DWORD PTR[r11]
|
| + lea r10,QWORD PTR[r10*1+rsi]
|
| + cmp rbx,r10
|
| + jb $L$common_seh_tail
|
| +
|
| + mov rax,QWORD PTR[152+r8]
|
| +
|
| + mov r10d,DWORD PTR[4+r11]
|
| + lea r10,QWORD PTR[r10*1+rsi]
|
| + cmp rbx,r10
|
| + jae $L$common_seh_tail
|
| +
|
| + mov rax,QWORD PTR[40+rax]
|
| +
|
| + mov rbx,QWORD PTR[((-8))+rax]
|
| + mov rbp,QWORD PTR[((-16))+rax]
|
| + mov r12,QWORD PTR[((-24))+rax]
|
| + mov r13,QWORD PTR[((-32))+rax]
|
| + mov r14,QWORD PTR[((-40))+rax]
|
| + mov r15,QWORD PTR[((-48))+rax]
|
| + mov QWORD PTR[144+r8],rbx
|
| + mov QWORD PTR[160+r8],rbp
|
| + mov QWORD PTR[216+r8],r12
|
| + mov QWORD PTR[224+r8],r13
|
| + mov QWORD PTR[232+r8],r14
|
| + mov QWORD PTR[240+r8],r15
|
| +
|
| +$L$common_seh_tail::
|
| + mov rdi,QWORD PTR[8+rax]
|
| + mov rsi,QWORD PTR[16+rax]
|
| + mov QWORD PTR[152+r8],rax
|
| + mov QWORD PTR[168+r8],rsi
|
| + mov QWORD PTR[176+r8],rdi
|
| +
|
| + mov rdi,QWORD PTR[40+r9]
|
| + mov rsi,r8
|
| + mov ecx,154
|
| + DD 0a548f3fch
|
| +
|
| + mov rsi,r9
|
| + xor rcx,rcx
|
| + mov rdx,QWORD PTR[8+rsi]
|
| + mov r8,QWORD PTR[rsi]
|
| + mov r9,QWORD PTR[16+rsi]
|
| + mov r10,QWORD PTR[40+rsi]
|
| + lea r11,QWORD PTR[56+rsi]
|
| + lea r12,QWORD PTR[24+rsi]
|
| + mov QWORD PTR[32+rsp],r10
|
| + mov QWORD PTR[40+rsp],r11
|
| + mov QWORD PTR[48+rsp],r12
|
| + mov QWORD PTR[56+rsp],rcx
|
| + call QWORD PTR[__imp_RtlVirtualUnwind]
|
| +
|
| + mov eax,1
|
| + add rsp,64
|
| + popfq
|
| + pop r15
|
| + pop r14
|
| + pop r13
|
| + pop r12
|
| + pop rbp
|
| + pop rbx
|
| + pop rdi
|
| + pop rsi
|
| + DB 0F3h,0C3h ;repret
|
| +sqr_handler ENDP
|
| +
|
| +.text$ ENDS
|
| +.pdata SEGMENT READONLY ALIGN(4)
|
| +ALIGN 4
|
| + DD imagerel $L$SEH_begin_bn_mul_mont
|
| + DD imagerel $L$SEH_end_bn_mul_mont
|
| + DD imagerel $L$SEH_info_bn_mul_mont
|
| +
|
| + DD imagerel $L$SEH_begin_bn_mul4x_mont
|
| + DD imagerel $L$SEH_end_bn_mul4x_mont
|
| + DD imagerel $L$SEH_info_bn_mul4x_mont
|
| +
|
| + DD imagerel $L$SEH_begin_bn_sqr8x_mont
|
| + DD imagerel $L$SEH_end_bn_sqr8x_mont
|
| + DD imagerel $L$SEH_info_bn_sqr8x_mont
|
| +.pdata ENDS
|
| +.xdata SEGMENT READONLY ALIGN(8)
|
| +ALIGN 8
|
| +$L$SEH_info_bn_mul_mont::
|
| +DB 9,0,0,0
|
| + DD imagerel mul_handler
|
| + DD imagerel $L$mul_body,imagerel $L$mul_epilogue
|
| +$L$SEH_info_bn_mul4x_mont::
|
| +DB 9,0,0,0
|
| + DD imagerel mul_handler
|
| + DD imagerel $L$mul4x_body,imagerel $L$mul4x_epilogue
|
| +$L$SEH_info_bn_sqr8x_mont::
|
| +DB 9,0,0,0
|
| + DD imagerel sqr_handler
|
| + DD imagerel $L$sqr8x_body,imagerel $L$sqr8x_epilogue
|
| +
|
| +.xdata ENDS
|
| +END
|
|
|