# Copyright 2011-2016 The OpenSSL Project Authors. All Rights Reserved.
# SPDX-License-Identifier: Apache-2.0
#
# Generated by admin/import-aws-lc from aws-lc 91c93a8eb860e51482a871a8772cb9b1b310baea:
# generated-src/linux-x86_64/crypto/fipsmodule/x86_64-mont5.S, which aws-lc generates from crypto/fipsmodule/bn/asm/x86_64-mont5.pl.
# Do not edit by hand.
.align 32
.cfi_startproc
leaq 32(%r10),%rbp
leaq (%rsi,%r9,1),%rsi
movq %r9,%rcx
movq -32(%rsi,%rbp,1),%r14
leaq 48+8(%rsp,%r9,2),%rdi
movq -24(%rsi,%rbp,1),%rax
leaq -32(%rdi,%rbp,1),%rdi
movq -16(%rsi,%rbp,1),%rbx
movq %rax,%r15
mulq %r14
movq %rax,%r10
movq %rbx,%rax
movq %rdx,%r11
movq %r10,-24(%rdi,%rbp,1)
mulq %r14
addq %rax,%r11
movq %rbx,%rax
adcq $0,%rdx
movq %r11,-16(%rdi,%rbp,1)
movq %rdx,%r10
movq -8(%rsi,%rbp,1),%rbx
mulq %r15
movq %rax,%r12
movq %rbx,%rax
movq %rdx,%r13
leaq (%rbp),%rcx
mulq %r14
addq %rax,%r10
movq %rbx,%rax
movq %rdx,%r11
adcq $0,%r11
addq %r12,%r10
adcq $0,%r11
movq %r10,-8(%rdi,%rcx,1)
jmp .Lmont5_sqr4x_1st
.align 32
.Lmont5_sqr4x_1st:
movq (%rsi,%rcx,1),%rbx
mulq %r15
addq %rax,%r13
movq %rbx,%rax
movq %rdx,%r12
adcq $0,%r12
mulq %r14
addq %rax,%r11
movq %rbx,%rax
movq 8(%rsi,%rcx,1),%rbx
movq %rdx,%r10
adcq $0,%r10
addq %r13,%r11
adcq $0,%r10
mulq %r15
addq %rax,%r12
movq %rbx,%rax
movq %r11,(%rdi,%rcx,1)
movq %rdx,%r13
adcq $0,%r13
mulq %r14
addq %rax,%r10
movq %rbx,%rax
movq 16(%rsi,%rcx,1),%rbx
movq %rdx,%r11
adcq $0,%r11
addq %r12,%r10
adcq $0,%r11
mulq %r15
addq %rax,%r13
movq %rbx,%rax
movq %r10,8(%rdi,%rcx,1)
movq %rdx,%r12
adcq $0,%r12
mulq %r14
addq %rax,%r11
movq %rbx,%rax
movq 24(%rsi,%rcx,1),%rbx
movq %rdx,%r10
adcq $0,%r10
addq %r13,%r11
adcq $0,%r10
mulq %r15
addq %rax,%r12
movq %rbx,%rax
movq %r11,16(%rdi,%rcx,1)
movq %rdx,%r13
adcq $0,%r13
leaq 32(%rcx),%rcx
mulq %r14
addq %rax,%r10
movq %rbx,%rax
movq %rdx,%r11
adcq $0,%r11
addq %r12,%r10
adcq $0,%r11
movq %r10,-8(%rdi,%rcx,1)
cmpq $0,%rcx
jne .Lmont5_sqr4x_1st
mulq %r15
addq %rax,%r13
leaq 16(%rbp),%rbp
adcq $0,%rdx
addq %r11,%r13
adcq $0,%rdx
movq %r13,(%rdi)
movq %rdx,%r12
movq %rdx,8(%rdi)
jmp .Lmont5_sqr4x_outer
.align 32
.Lmont5_sqr4x_outer:
movq -32(%rsi,%rbp,1),%r14
leaq 48+8(%rsp,%r9,2),%rdi
movq -24(%rsi,%rbp,1),%rax
leaq -32(%rdi,%rbp,1),%rdi
movq -16(%rsi,%rbp,1),%rbx
movq %rax,%r15
mulq %r14
movq -24(%rdi,%rbp,1),%r10
addq %rax,%r10
movq %rbx,%rax
adcq $0,%rdx
movq %r10,-24(%rdi,%rbp,1)
movq %rdx,%r11
mulq %r14
addq %rax,%r11
movq %rbx,%rax
adcq $0,%rdx
addq -16(%rdi,%rbp,1),%r11
movq %rdx,%r10
adcq $0,%r10
movq %r11,-16(%rdi,%rbp,1)
xorq %r12,%r12
movq -8(%rsi,%rbp,1),%rbx
mulq %r15
addq %rax,%r12
movq %rbx,%rax
adcq $0,%rdx
addq -8(%rdi,%rbp,1),%r12
movq %rdx,%r13
adcq $0,%r13
mulq %r14
addq %rax,%r10
movq %rbx,%rax
adcq $0,%rdx
addq %r12,%r10
movq %rdx,%r11
adcq $0,%r11
movq %r10,-8(%rdi,%rbp,1)
leaq (%rbp),%rcx
jmp .Lmont5_sqr4x_inner
.align 32
.Lmont5_sqr4x_inner:
movq (%rsi,%rcx,1),%rbx
mulq %r15
addq %rax,%r13
movq %rbx,%rax
movq %rdx,%r12
adcq $0,%r12
addq (%rdi,%rcx,1),%r13
adcq $0,%r12
.byte 0x67
mulq %r14
addq %rax,%r11
movq %rbx,%rax
movq 8(%rsi,%rcx,1),%rbx
movq %rdx,%r10
adcq $0,%r10
addq %r13,%r11
adcq $0,%r10
mulq %r15
addq %rax,%r12
movq %r11,(%rdi,%rcx,1)
movq %rbx,%rax
movq %rdx,%r13
adcq $0,%r13
addq 8(%rdi,%rcx,1),%r12
leaq 16(%rcx),%rcx
adcq $0,%r13
mulq %r14
addq %rax,%r10
movq %rbx,%rax
adcq $0,%rdx
addq %r12,%r10
movq %rdx,%r11
adcq $0,%r11
movq %r10,-8(%rdi,%rcx,1)
cmpq $0,%rcx
jne .Lmont5_sqr4x_inner
.byte 0x67
mulq %r15
addq %rax,%r13
adcq $0,%rdx
addq %r11,%r13
adcq $0,%rdx
movq %r13,(%rdi)
movq %rdx,%r12
movq %rdx,8(%rdi)
addq $16,%rbp
jnz .Lmont5_sqr4x_outer
movq -32(%rsi),%r14
leaq 48+8(%rsp,%r9,2),%rdi
movq -24(%rsi),%rax
leaq -32(%rdi,%rbp,1),%rdi
movq -16(%rsi),%rbx
movq %rax,%r15
mulq %r14
addq %rax,%r10
movq %rbx,%rax
movq %rdx,%r11
adcq $0,%r11
mulq %r14
addq %rax,%r11
movq %rbx,%rax
movq %r10,-24(%rdi)
movq %rdx,%r10
adcq $0,%r10
addq %r13,%r11
movq -8(%rsi),%rbx
adcq $0,%r10
mulq %r15
addq %rax,%r12
movq %rbx,%rax
movq %r11,-16(%rdi)
movq %rdx,%r13
adcq $0,%r13
mulq %r14
addq %rax,%r10
movq %rbx,%rax
movq %rdx,%r11
adcq $0,%r11
addq %r12,%r10
adcq $0,%r11
movq %r10,-8(%rdi)
mulq %r15
addq %rax,%r13
movq -16(%rsi),%rax
adcq $0,%rdx
addq %r11,%r13
adcq $0,%rdx
movq %r13,(%rdi)
movq %rdx,%r12
movq %rdx,8(%rdi)
mulq %rbx
addq $16,%rbp
xorq %r14,%r14
subq %r9,%rbp
xorq %r15,%r15
addq %r12,%rax
adcq $0,%rdx
movq %rax,8(%rdi)
movq %rdx,16(%rdi)
movq %r15,24(%rdi)
movq -16(%rsi,%rbp,1),%rax
leaq 48+8(%rsp),%rdi
xorq %r10,%r10
movq 8(%rdi),%r11
leaq (%r14,%r10,2),%r12
shrq $63,%r10
leaq (%rcx,%r11,2),%r13
shrq $63,%r11
orq %r10,%r13
movq 16(%rdi),%r10
movq %r11,%r14
mulq %rax
negq %r15
movq 24(%rdi),%r11
adcq %rax,%r12
movq -8(%rsi,%rbp,1),%rax
movq %r12,(%rdi)
adcq %rdx,%r13
leaq (%r14,%r10,2),%rbx
movq %r13,8(%rdi)
sbbq %r15,%r15
shrq $63,%r10
leaq (%rcx,%r11,2),%r8
shrq $63,%r11
orq %r10,%r8
movq 32(%rdi),%r10
movq %r11,%r14
mulq %rax
negq %r15
movq 40(%rdi),%r11
adcq %rax,%rbx
movq 0(%rsi,%rbp,1),%rax
movq %rbx,16(%rdi)
adcq %rdx,%r8
leaq 16(%rbp),%rbp
movq %r8,24(%rdi)
sbbq %r15,%r15
leaq 64(%rdi),%rdi
jmp .Lmont5_sqr4x_shift_n_add
.align 32
.Lmont5_sqr4x_shift_n_add:
leaq (%r14,%r10,2),%r12
shrq $63,%r10
leaq (%rcx,%r11,2),%r13
shrq $63,%r11
orq %r10,%r13
movq -16(%rdi),%r10
movq %r11,%r14
mulq %rax
negq %r15
movq -8(%rdi),%r11
adcq %rax,%r12
movq -8(%rsi,%rbp,1),%rax
movq %r12,-32(%rdi)
adcq %rdx,%r13
leaq (%r14,%r10,2),%rbx
movq %r13,-24(%rdi)
sbbq %r15,%r15
shrq $63,%r10
leaq (%rcx,%r11,2),%r8
shrq $63,%r11
orq %r10,%r8
movq 0(%rdi),%r10
movq %r11,%r14
mulq %rax
negq %r15
movq 8(%rdi),%r11
adcq %rax,%rbx
movq 0(%rsi,%rbp,1),%rax
movq %rbx,-16(%rdi)
adcq %rdx,%r8
leaq (%r14,%r10,2),%r12
movq %r8,-8(%rdi)
sbbq %r15,%r15
shrq $63,%r10
leaq (%rcx,%r11,2),%r13
shrq $63,%r11
orq %r10,%r13
movq 16(%rdi),%r10
movq %r11,%r14
mulq %rax
negq %r15
movq 24(%rdi),%r11
adcq %rax,%r12
movq 8(%rsi,%rbp,1),%rax
movq %r12,0(%rdi)
adcq %rdx,%r13
leaq (%r14,%r10,2),%rbx
movq %r13,8(%rdi)
sbbq %r15,%r15
shrq $63,%r10
leaq (%rcx,%r11,2),%r8
shrq $63,%r11
orq %r10,%r8
movq 32(%rdi),%r10
movq %r11,%r14
mulq %rax
negq %r15
movq 40(%rdi),%r11
adcq %rax,%rbx
movq 16(%rsi,%rbp,1),%rax
movq %rbx,16(%rdi)
adcq %rdx,%r8
movq %r8,24(%rdi)
sbbq %r15,%r15
leaq 64(%rdi),%rdi
addq $32,%rbp
jnz .Lmont5_sqr4x_shift_n_add
leaq (%r14,%r10,2),%r12
.byte 0x67
shrq $63,%r10
leaq (%rcx,%r11,2),%r13
shrq $63,%r11
orq %r10,%r13
movq -16(%rdi),%r10
movq %r11,%r14
mulq %rax
negq %r15
movq -8(%rdi),%r11
adcq %rax,%r12
movq -8(%rsi),%rax
movq %r12,-32(%rdi)
adcq %rdx,%r13
leaq (%r14,%r10,2),%rbx
movq %r13,-24(%rdi)
sbbq %r15,%r15
shrq $63,%r10
leaq (%rcx,%r11,2),%r8
shrq $63,%r11
orq %r10,%r8
mulq %rax
negq %r15
adcq %rax,%rbx
adcq %rdx,%r8
movq %rbx,-16(%rdi)
movq %r8,-8(%rdi)
.byte 102,72,15,126,213
xorq %rax,%rax
leaq (%r9,%rbp,1),%rcx
leaq 48+8(%rsp,%r9,2),%rdx
movq %rcx,0+8(%rsp)
leaq 48+8(%rsp,%r9,1),%rdi
movq %rdx,8+8(%rsp)
negq %r9
jmp .Lmont5_8x_reduction_loop
.align 32
.Lmont5_8x_reduction_loop:
leaq (%rdi,%r9,1),%rdi
.byte 0x66
movq 0(%rdi),%rbx
movq 8(%rdi),%r9
movq 16(%rdi),%r10
movq 24(%rdi),%r11
movq 32(%rdi),%r12
movq 40(%rdi),%r13
movq 48(%rdi),%r14
movq 56(%rdi),%r15
movq %rax,(%rdx)
leaq 64(%rdi),%rdi
.byte 0x67
movq %rbx,%r8
imulq 32+8(%rsp),%rbx
movq 0(%rbp),%rax
movl $8,%ecx
jmp .Lmont5_8x_reduce
.align 32
.Lmont5_8x_reduce:
mulq %rbx
movq 8(%rbp),%rax
negq %r8
movq %rdx,%r8
adcq $0,%r8
mulq %rbx
addq %rax,%r9
movq 16(%rbp),%rax
adcq $0,%rdx
addq %r9,%r8
movq %rbx,48-8+8(%rsp,%rcx,8)
movq %rdx,%r9
adcq $0,%r9
mulq %rbx
addq %rax,%r10
movq 24(%rbp),%rax
adcq $0,%rdx
addq %r10,%r9
movq 32+8(%rsp),%rsi
movq %rdx,%r10
adcq $0,%r10
mulq %rbx
addq %rax,%r11
movq 32(%rbp),%rax
adcq $0,%rdx
imulq %r8,%rsi
addq %r11,%r10
movq %rdx,%r11
adcq $0,%r11
mulq %rbx
addq %rax,%r12
movq 40(%rbp),%rax
adcq $0,%rdx
addq %r12,%r11
movq %rdx,%r12
adcq $0,%r12
mulq %rbx
addq %rax,%r13
movq 48(%rbp),%rax
adcq $0,%rdx
addq %r13,%r12
movq %rdx,%r13
adcq $0,%r13
mulq %rbx
addq %rax,%r14
movq 56(%rbp),%rax
adcq $0,%rdx
addq %r14,%r13
movq %rdx,%r14
adcq $0,%r14
mulq %rbx
movq %rsi,%rbx
addq %rax,%r15
movq 0(%rbp),%rax
adcq $0,%rdx
addq %r15,%r14
movq %rdx,%r15
adcq $0,%r15
decl %ecx
jnz .Lmont5_8x_reduce
leaq 64(%rbp),%rbp
xorq %rax,%rax
movq 8+8(%rsp),%rdx
cmpq 0+8(%rsp),%rbp
jae .Lmont5_8x_no_tail
.byte 0x66
addq 0(%rdi),%r8
adcq 8(%rdi),%r9
adcq 16(%rdi),%r10
adcq 24(%rdi),%r11
adcq 32(%rdi),%r12
adcq 40(%rdi),%r13
adcq 48(%rdi),%r14
adcq 56(%rdi),%r15
sbbq %rsi,%rsi
movq 48+56+8(%rsp),%rbx
movl $8,%ecx
movq 0(%rbp),%rax
jmp .Lmont5_8x_tail
.align 32
.Lmont5_8x_tail:
mulq %rbx
addq %rax,%r8
movq 8(%rbp),%rax
movq %r8,(%rdi)
movq %rdx,%r8
adcq $0,%r8
mulq %rbx
addq %rax,%r9
movq 16(%rbp),%rax
adcq $0,%rdx
addq %r9,%r8
leaq 8(%rdi),%rdi
movq %rdx,%r9
adcq $0,%r9
mulq %rbx
addq %rax,%r10
movq 24(%rbp),%rax
adcq $0,%rdx
addq %r10,%r9
movq %rdx,%r10
adcq $0,%r10
mulq %rbx
addq %rax,%r11
movq 32(%rbp),%rax
adcq $0,%rdx
addq %r11,%r10
movq %rdx,%r11
adcq $0,%r11
mulq %rbx
addq %rax,%r12
movq 40(%rbp),%rax
adcq $0,%rdx
addq %r12,%r11
movq %rdx,%r12
adcq $0,%r12
mulq %rbx
addq %rax,%r13
movq 48(%rbp),%rax
adcq $0,%rdx
addq %r13,%r12
movq %rdx,%r13
adcq $0,%r13
mulq %rbx
addq %rax,%r14
movq 56(%rbp),%rax
adcq $0,%rdx
addq %r14,%r13
movq %rdx,%r14
adcq $0,%r14
mulq %rbx
movq 48-16+8(%rsp,%rcx,8),%rbx
addq %rax,%r15
adcq $0,%rdx
addq %r15,%r14
movq 0(%rbp),%rax
movq %rdx,%r15
adcq $0,%r15
decl %ecx
jnz .Lmont5_8x_tail
leaq 64(%rbp),%rbp
movq 8+8(%rsp),%rdx
cmpq 0+8(%rsp),%rbp
jae .Lmont5_8x_tail_done
movq 48+56+8(%rsp),%rbx
negq %rsi
movq 0(%rbp),%rax
adcq 0(%rdi),%r8
adcq 8(%rdi),%r9
adcq 16(%rdi),%r10
adcq 24(%rdi),%r11
adcq 32(%rdi),%r12
adcq 40(%rdi),%r13
adcq 48(%rdi),%r14
adcq 56(%rdi),%r15
sbbq %rsi,%rsi
movl $8,%ecx
jmp .Lmont5_8x_tail
.align 32
.Lmont5_8x_tail_done:
xorq %rax,%rax
addq (%rdx),%r8
adcq $0,%r9
adcq $0,%r10
adcq $0,%r11
adcq $0,%r12
adcq $0,%r13
adcq $0,%r14
adcq $0,%r15
adcq $0,%rax
negq %rsi
.Lmont5_8x_no_tail:
adcq 0(%rdi),%r8
adcq 8(%rdi),%r9
adcq 16(%rdi),%r10
adcq 24(%rdi),%r11
adcq 32(%rdi),%r12
adcq 40(%rdi),%r13
adcq 48(%rdi),%r14
adcq 56(%rdi),%r15
adcq $0,%rax
movq -8(%rbp),%rcx
xorq %rsi,%rsi
.byte 102,72,15,126,213
movq %r8,0(%rdi)
movq %r9,8(%rdi)
.byte 102,73,15,126,217
movq %r10,16(%rdi)
movq %r11,24(%rdi)
movq %r12,32(%rdi)
movq %r13,40(%rdi)
movq %r14,48(%rdi)
movq %r15,56(%rdi)
leaq 64(%rdi),%rdi
cmpq %rdx,%rdi
jb .Lmont5_8x_reduction_loop
.byte 0xf3,0xc3
.cfi_endproc