diff options
Diffstat (limited to 'secure/lib/libcrypto/amd64/x86_64-mont5.S')
| -rw-r--r-- | secure/lib/libcrypto/amd64/x86_64-mont5.S | 1302 |
1 files changed, 1301 insertions, 1 deletions
diff --git a/secure/lib/libcrypto/amd64/x86_64-mont5.S b/secure/lib/libcrypto/amd64/x86_64-mont5.S index b47c677..1724f9c0fb 100644 --- a/secure/lib/libcrypto/amd64/x86_64-mont5.S +++ b/secure/lib/libcrypto/amd64/x86_64-mont5.S @@ -1,4 +1,5 @@ - # $FreeBSD$ +# $FreeBSD$ +# Do not modify. This file is auto-generated from x86_64-mont5.pl. .text @@ -9,6 +10,7 @@ bn_mul_mont_gather5: testl $7,%r9d jnz .Lmul_enter + movl OPENSSL_ia32cap_P+8(%rip),%r11d jmp .Lmul4x_enter .align 16 @@ -415,6 +417,9 @@ bn_mul_mont_gather5: .align 32 bn_mul4x_mont_gather5: .Lmul4x_enter: + andl $0x80108,%r11d + cmpl $0x80108,%r11d + je .Lmulx4x_enter .byte 0x67 movq %rsp,%rax pushq %rbx @@ -1016,6 +1021,10 @@ mul4x_internal: .type bn_power5,@function .align 32 bn_power5: + movl OPENSSL_ia32cap_P+8(%rip),%r11d + andl $0x80108,%r11d + cmpl $0x80108,%r11d + je .Lpowerx5_enter movq %rsp,%rax pushq %rbx pushq %rbp @@ -2058,6 +2067,22 @@ bn_from_mont8x: .byte 0x67 movq %rcx,%rbp .byte 102,73,15,110,218 + movl OPENSSL_ia32cap_P+8(%rip),%r11d + andl $0x80108,%r11d + cmpl $0x80108,%r11d + jne .Lfrom_mont_nox + + leaq (%rax,%r9,1),%rdi + call __bn_sqrx8x_reduction + call __bn_postx4x_internal + + pxor %xmm0,%xmm0 + leaq 48(%rsp),%rax + movq 40(%rsp),%rsi + jmp .Lfrom_mont_zero + +.align 32 +.Lfrom_mont_nox: call __bn_sqr8x_reduction call __bn_post4x_internal @@ -2087,6 +2112,1281 @@ bn_from_mont8x: .Lfrom_epilogue: .byte 0xf3,0xc3 .size bn_from_mont8x,.-bn_from_mont8x +.type bn_mulx4x_mont_gather5,@function +.align 32 +bn_mulx4x_mont_gather5: +.Lmulx4x_enter: + movq %rsp,%rax + pushq %rbx + pushq %rbp + pushq %r12 + pushq %r13 + pushq %r14 + pushq %r15 + + shll $3,%r9d + leaq (%r9,%r9,2),%r10 + negq %r9 + movq (%r8),%r8 + + + + + + + + + + + leaq -320(%rsp,%r9,2),%r11 + subq %rdi,%r11 + andq $4095,%r11 + cmpq %r11,%r10 + jb .Lmulx4xsp_alt + subq %r11,%rsp + leaq -320(%rsp,%r9,2),%rsp + jmp .Lmulx4xsp_done + +.Lmulx4xsp_alt: + leaq 4096-320(,%r9,2),%r10 + leaq -320(%rsp,%r9,2),%rsp + subq %r10,%r11 + movq $0,%r10 + cmovcq %r10,%r11 + subq %r11,%rsp +.Lmulx4xsp_done: + andq $-64,%rsp + movq %rax,%r11 + subq %rsp,%r11 + andq $-4096,%r11 +.Lmulx4x_page_walk: + movq (%rsp,%r11,1),%r10 + subq $4096,%r11 +.byte 0x2e + jnc .Lmulx4x_page_walk + + + + + + + + + + + + + + movq %r8,32(%rsp) + movq %rax,40(%rsp) +.Lmulx4x_body: + call mulx4x_internal + + movq 40(%rsp),%rsi + movq $1,%rax + + movq -48(%rsi),%r15 + movq -40(%rsi),%r14 + movq -32(%rsi),%r13 + movq -24(%rsi),%r12 + movq -16(%rsi),%rbp + movq -8(%rsi),%rbx + leaq (%rsi),%rsp +.Lmulx4x_epilogue: + .byte 0xf3,0xc3 +.size bn_mulx4x_mont_gather5,.-bn_mulx4x_mont_gather5 + +.type mulx4x_internal,@function +.align 32 +mulx4x_internal: + movq %r9,8(%rsp) + movq %r9,%r10 + negq %r9 + shlq $5,%r9 + negq %r10 + leaq 128(%rdx,%r9,1),%r13 + shrq $5+5,%r9 + movd 8(%rax),%xmm5 + subq $1,%r9 + leaq .Linc(%rip),%rax + movq %r13,16+8(%rsp) + movq %r9,24+8(%rsp) + movq %rdi,56+8(%rsp) + movdqa 0(%rax),%xmm0 + movdqa 16(%rax),%xmm1 + leaq 88-112(%rsp,%r10,1),%r10 + leaq 128(%rdx),%rdi + + pshufd $0,%xmm5,%xmm5 + movdqa %xmm1,%xmm4 +.byte 0x67 + movdqa %xmm1,%xmm2 +.byte 0x67 + paddd %xmm0,%xmm1 + pcmpeqd %xmm5,%xmm0 + movdqa %xmm4,%xmm3 + paddd %xmm1,%xmm2 + pcmpeqd %xmm5,%xmm1 + movdqa %xmm0,112(%r10) + movdqa %xmm4,%xmm0 + + paddd %xmm2,%xmm3 + pcmpeqd %xmm5,%xmm2 + movdqa %xmm1,128(%r10) + movdqa %xmm4,%xmm1 + + paddd %xmm3,%xmm0 + pcmpeqd %xmm5,%xmm3 + movdqa %xmm2,144(%r10) + movdqa %xmm4,%xmm2 + + paddd %xmm0,%xmm1 + pcmpeqd %xmm5,%xmm0 + movdqa %xmm3,160(%r10) + movdqa %xmm4,%xmm3 + paddd %xmm1,%xmm2 + pcmpeqd %xmm5,%xmm1 + movdqa %xmm0,176(%r10) + movdqa %xmm4,%xmm0 + + paddd %xmm2,%xmm3 + pcmpeqd %xmm5,%xmm2 + movdqa %xmm1,192(%r10) + movdqa %xmm4,%xmm1 + + paddd %xmm3,%xmm0 + pcmpeqd %xmm5,%xmm3 + movdqa %xmm2,208(%r10) + movdqa %xmm4,%xmm2 + + paddd %xmm0,%xmm1 + pcmpeqd %xmm5,%xmm0 + movdqa %xmm3,224(%r10) + movdqa %xmm4,%xmm3 + paddd %xmm1,%xmm2 + pcmpeqd %xmm5,%xmm1 + movdqa %xmm0,240(%r10) + movdqa %xmm4,%xmm0 + + paddd %xmm2,%xmm3 + pcmpeqd %xmm5,%xmm2 + movdqa %xmm1,256(%r10) + movdqa %xmm4,%xmm1 + + paddd %xmm3,%xmm0 + pcmpeqd %xmm5,%xmm3 + movdqa %xmm2,272(%r10) + movdqa %xmm4,%xmm2 + + paddd %xmm0,%xmm1 + pcmpeqd %xmm5,%xmm0 + movdqa %xmm3,288(%r10) + movdqa %xmm4,%xmm3 +.byte 0x67 + paddd %xmm1,%xmm2 + pcmpeqd %xmm5,%xmm1 + movdqa %xmm0,304(%r10) + + paddd %xmm2,%xmm3 + pcmpeqd %xmm5,%xmm2 + movdqa %xmm1,320(%r10) + + pcmpeqd %xmm5,%xmm3 + movdqa %xmm2,336(%r10) + + pand 64(%rdi),%xmm0 + pand 80(%rdi),%xmm1 + pand 96(%rdi),%xmm2 + movdqa %xmm3,352(%r10) + pand 112(%rdi),%xmm3 + por %xmm2,%xmm0 + por %xmm3,%xmm1 + movdqa -128(%rdi),%xmm4 + movdqa -112(%rdi),%xmm5 + movdqa -96(%rdi),%xmm2 + pand 112(%r10),%xmm4 + movdqa -80(%rdi),%xmm3 + pand 128(%r10),%xmm5 + por %xmm4,%xmm0 + pand 144(%r10),%xmm2 + por %xmm5,%xmm1 + pand 160(%r10),%xmm3 + por %xmm2,%xmm0 + por %xmm3,%xmm1 + movdqa -64(%rdi),%xmm4 + movdqa -48(%rdi),%xmm5 + movdqa -32(%rdi),%xmm2 + pand 176(%r10),%xmm4 + movdqa -16(%rdi),%xmm3 + pand 192(%r10),%xmm5 + por %xmm4,%xmm0 + pand 208(%r10),%xmm2 + por %xmm5,%xmm1 + pand 224(%r10),%xmm3 + por %xmm2,%xmm0 + por %xmm3,%xmm1 + movdqa 0(%rdi),%xmm4 + movdqa 16(%rdi),%xmm5 + movdqa 32(%rdi),%xmm2 + pand 240(%r10),%xmm4 + movdqa 48(%rdi),%xmm3 + pand 256(%r10),%xmm5 + por %xmm4,%xmm0 + pand 272(%r10),%xmm2 + por %xmm5,%xmm1 + pand 288(%r10),%xmm3 + por %xmm2,%xmm0 + por %xmm3,%xmm1 + pxor %xmm1,%xmm0 + pshufd $0x4e,%xmm0,%xmm1 + por %xmm1,%xmm0 + leaq 256(%rdi),%rdi +.byte 102,72,15,126,194 + leaq 64+32+8(%rsp),%rbx + + movq %rdx,%r9 + mulxq 0(%rsi),%r8,%rax + mulxq 8(%rsi),%r11,%r12 + addq %rax,%r11 + mulxq 16(%rsi),%rax,%r13 + adcq %rax,%r12 + adcq $0,%r13 + mulxq 24(%rsi),%rax,%r14 + + movq %r8,%r15 + imulq 32+8(%rsp),%r8 + xorq %rbp,%rbp + movq %r8,%rdx + + movq %rdi,8+8(%rsp) + + leaq 32(%rsi),%rsi + adcxq %rax,%r13 + adcxq %rbp,%r14 + + mulxq 0(%rcx),%rax,%r10 + adcxq %rax,%r15 + adoxq %r11,%r10 + mulxq 8(%rcx),%rax,%r11 + adcxq %rax,%r10 + adoxq %r12,%r11 + mulxq 16(%rcx),%rax,%r12 + movq 24+8(%rsp),%rdi + movq %r10,-32(%rbx) + adcxq %rax,%r11 + adoxq %r13,%r12 + mulxq 24(%rcx),%rax,%r15 + movq %r9,%rdx + movq %r11,-24(%rbx) + adcxq %rax,%r12 + adoxq %rbp,%r15 + leaq 32(%rcx),%rcx + movq %r12,-16(%rbx) + jmp .Lmulx4x_1st + +.align 32 +.Lmulx4x_1st: + adcxq %rbp,%r15 + mulxq 0(%rsi),%r10,%rax + adcxq %r14,%r10 + mulxq 8(%rsi),%r11,%r14 + adcxq %rax,%r11 + mulxq 16(%rsi),%r12,%rax + adcxq %r14,%r12 + mulxq 24(%rsi),%r13,%r14 +.byte 0x67,0x67 + movq %r8,%rdx + adcxq %rax,%r13 + adcxq %rbp,%r14 + leaq 32(%rsi),%rsi + leaq 32(%rbx),%rbx + + adoxq %r15,%r10 + mulxq 0(%rcx),%rax,%r15 + adcxq %rax,%r10 + adoxq %r15,%r11 + mulxq 8(%rcx),%rax,%r15 + adcxq %rax,%r11 + adoxq %r15,%r12 + mulxq 16(%rcx),%rax,%r15 + movq %r10,-40(%rbx) + adcxq %rax,%r12 + movq %r11,-32(%rbx) + adoxq %r15,%r13 + mulxq 24(%rcx),%rax,%r15 + movq %r9,%rdx + movq %r12,-24(%rbx) + adcxq %rax,%r13 + adoxq %rbp,%r15 + leaq 32(%rcx),%rcx + movq %r13,-16(%rbx) + + decq %rdi + jnz .Lmulx4x_1st + + movq 8(%rsp),%rax + adcq %rbp,%r15 + leaq (%rsi,%rax,1),%rsi + addq %r15,%r14 + movq 8+8(%rsp),%rdi + adcq %rbp,%rbp + movq %r14,-8(%rbx) + jmp .Lmulx4x_outer + +.align 32 +.Lmulx4x_outer: + leaq 16-256(%rbx),%r10 + pxor %xmm4,%xmm4 +.byte 0x67,0x67 + pxor %xmm5,%xmm5 + movdqa -128(%rdi),%xmm0 + movdqa -112(%rdi),%xmm1 + movdqa -96(%rdi),%xmm2 + pand 256(%r10),%xmm0 + movdqa -80(%rdi),%xmm3 + pand 272(%r10),%xmm1 + por %xmm0,%xmm4 + pand 288(%r10),%xmm2 + por %xmm1,%xmm5 + pand 304(%r10),%xmm3 + por %xmm2,%xmm4 + por %xmm3,%xmm5 + movdqa -64(%rdi),%xmm0 + movdqa -48(%rdi),%xmm1 + movdqa -32(%rdi),%xmm2 + pand 320(%r10),%xmm0 + movdqa -16(%rdi),%xmm3 + pand 336(%r10),%xmm1 + por %xmm0,%xmm4 + pand 352(%r10),%xmm2 + por %xmm1,%xmm5 + pand 368(%r10),%xmm3 + por %xmm2,%xmm4 + por %xmm3,%xmm5 + movdqa 0(%rdi),%xmm0 + movdqa 16(%rdi),%xmm1 + movdqa 32(%rdi),%xmm2 + pand 384(%r10),%xmm0 + movdqa 48(%rdi),%xmm3 + pand 400(%r10),%xmm1 + por %xmm0,%xmm4 + pand 416(%r10),%xmm2 + por %xmm1,%xmm5 + pand 432(%r10),%xmm3 + por %xmm2,%xmm4 + por %xmm3,%xmm5 + movdqa 64(%rdi),%xmm0 + movdqa 80(%rdi),%xmm1 + movdqa 96(%rdi),%xmm2 + pand 448(%r10),%xmm0 + movdqa 112(%rdi),%xmm3 + pand 464(%r10),%xmm1 + por %xmm0,%xmm4 + pand 480(%r10),%xmm2 + por %xmm1,%xmm5 + pand 496(%r10),%xmm3 + por %xmm2,%xmm4 + por %xmm3,%xmm5 + por %xmm5,%xmm4 + pshufd $0x4e,%xmm4,%xmm0 + por %xmm4,%xmm0 + leaq 256(%rdi),%rdi +.byte 102,72,15,126,194 + + movq %rbp,(%rbx) + leaq 32(%rbx,%rax,1),%rbx + mulxq 0(%rsi),%r8,%r11 + xorq %rbp,%rbp + movq %rdx,%r9 + mulxq 8(%rsi),%r14,%r12 + adoxq -32(%rbx),%r8 + adcxq %r14,%r11 + mulxq 16(%rsi),%r15,%r13 + adoxq -24(%rbx),%r11 + adcxq %r15,%r12 + mulxq 24(%rsi),%rdx,%r14 + adoxq -16(%rbx),%r12 + adcxq %rdx,%r13 + leaq (%rcx,%rax,1),%rcx + leaq 32(%rsi),%rsi + adoxq -8(%rbx),%r13 + adcxq %rbp,%r14 + adoxq %rbp,%r14 + + movq %r8,%r15 + imulq 32+8(%rsp),%r8 + + movq %r8,%rdx + xorq %rbp,%rbp + movq %rdi,8+8(%rsp) + + mulxq 0(%rcx),%rax,%r10 + adcxq %rax,%r15 + adoxq %r11,%r10 + mulxq 8(%rcx),%rax,%r11 + adcxq %rax,%r10 + adoxq %r12,%r11 + mulxq 16(%rcx),%rax,%r12 + adcxq %rax,%r11 + adoxq %r13,%r12 + mulxq 24(%rcx),%rax,%r15 + movq %r9,%rdx + movq 24+8(%rsp),%rdi + movq %r10,-32(%rbx) + adcxq %rax,%r12 + movq %r11,-24(%rbx) + adoxq %rbp,%r15 + movq %r12,-16(%rbx) + leaq 32(%rcx),%rcx + jmp .Lmulx4x_inner + +.align 32 +.Lmulx4x_inner: + mulxq 0(%rsi),%r10,%rax + adcxq %rbp,%r15 + adoxq %r14,%r10 + mulxq 8(%rsi),%r11,%r14 + adcxq 0(%rbx),%r10 + adoxq %rax,%r11 + mulxq 16(%rsi),%r12,%rax + adcxq 8(%rbx),%r11 + adoxq %r14,%r12 + mulxq 24(%rsi),%r13,%r14 + movq %r8,%rdx + adcxq 16(%rbx),%r12 + adoxq %rax,%r13 + adcxq 24(%rbx),%r13 + adoxq %rbp,%r14 + leaq 32(%rsi),%rsi + leaq 32(%rbx),%rbx + adcxq %rbp,%r14 + + adoxq %r15,%r10 + mulxq 0(%rcx),%rax,%r15 + adcxq %rax,%r10 + adoxq %r15,%r11 + mulxq 8(%rcx),%rax,%r15 + adcxq %rax,%r11 + adoxq %r15,%r12 + mulxq 16(%rcx),%rax,%r15 + movq %r10,-40(%rbx) + adcxq %rax,%r12 + adoxq %r15,%r13 + movq %r11,-32(%rbx) + mulxq 24(%rcx),%rax,%r15 + movq %r9,%rdx + leaq 32(%rcx),%rcx + movq %r12,-24(%rbx) + adcxq %rax,%r13 + adoxq %rbp,%r15 + movq %r13,-16(%rbx) + + decq %rdi + jnz .Lmulx4x_inner + + movq 0+8(%rsp),%rax + adcq %rbp,%r15 + subq 0(%rbx),%rdi + movq 8+8(%rsp),%rdi + movq 16+8(%rsp),%r10 + adcq %r15,%r14 + leaq (%rsi,%rax,1),%rsi + adcq %rbp,%rbp + movq %r14,-8(%rbx) + + cmpq %r10,%rdi + jb .Lmulx4x_outer + + movq -8(%rcx),%r10 + movq %rbp,%r8 + movq (%rcx,%rax,1),%r12 + leaq (%rcx,%rax,1),%rbp + movq %rax,%rcx + leaq (%rbx,%rax,1),%rdi + xorl %eax,%eax + xorq %r15,%r15 + subq %r14,%r10 + adcq %r15,%r15 + orq %r15,%r8 + sarq $3+2,%rcx + subq %r8,%rax + movq 56+8(%rsp),%rdx + decq %r12 + movq 8(%rbp),%r13 + xorq %r8,%r8 + movq 16(%rbp),%r14 + movq 24(%rbp),%r15 + jmp .Lsqrx4x_sub_entry +.size mulx4x_internal,.-mulx4x_internal +.type bn_powerx5,@function +.align 32 +bn_powerx5: +.Lpowerx5_enter: + movq %rsp,%rax + pushq %rbx + pushq %rbp + pushq %r12 + pushq %r13 + pushq %r14 + pushq %r15 + + shll $3,%r9d + leaq (%r9,%r9,2),%r10 + negq %r9 + movq (%r8),%r8 + + + + + + + + + leaq -320(%rsp,%r9,2),%r11 + subq %rdi,%r11 + andq $4095,%r11 + cmpq %r11,%r10 + jb .Lpwrx_sp_alt + subq %r11,%rsp + leaq -320(%rsp,%r9,2),%rsp + jmp .Lpwrx_sp_done + +.align 32 +.Lpwrx_sp_alt: + leaq 4096-320(,%r9,2),%r10 + leaq -320(%rsp,%r9,2),%rsp + subq %r10,%r11 + movq $0,%r10 + cmovcq %r10,%r11 + subq %r11,%rsp +.Lpwrx_sp_done: + andq $-64,%rsp + movq %rax,%r11 + subq %rsp,%r11 + andq $-4096,%r11 +.Lpwrx_page_walk: + movq (%rsp,%r11,1),%r10 + subq $4096,%r11 +.byte 0x2e + jnc .Lpwrx_page_walk + + movq %r9,%r10 + negq %r9 + + + + + + + + + + + + + pxor %xmm0,%xmm0 +.byte 102,72,15,110,207 +.byte 102,72,15,110,209 +.byte 102,73,15,110,218 +.byte 102,72,15,110,226 + movq %r8,32(%rsp) + movq %rax,40(%rsp) +.Lpowerx5_body: + + call __bn_sqrx8x_internal + call __bn_postx4x_internal + call __bn_sqrx8x_internal + call __bn_postx4x_internal + call __bn_sqrx8x_internal + call __bn_postx4x_internal + call __bn_sqrx8x_internal + call __bn_postx4x_internal + call __bn_sqrx8x_internal + call __bn_postx4x_internal + + movq %r10,%r9 + movq %rsi,%rdi +.byte 102,72,15,126,209 +.byte 102,72,15,126,226 + movq 40(%rsp),%rax + + call mulx4x_internal + + movq 40(%rsp),%rsi + movq $1,%rax + + movq -48(%rsi),%r15 + movq -40(%rsi),%r14 + movq -32(%rsi),%r13 + movq -24(%rsi),%r12 + movq -16(%rsi),%rbp + movq -8(%rsi),%rbx + leaq (%rsi),%rsp +.Lpowerx5_epilogue: + .byte 0xf3,0xc3 +.size bn_powerx5,.-bn_powerx5 + +.globl bn_sqrx8x_internal +.hidden bn_sqrx8x_internal +.type bn_sqrx8x_internal,@function +.align 32 +bn_sqrx8x_internal: +__bn_sqrx8x_internal: + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + leaq 48+8(%rsp),%rdi + leaq (%rsi,%r9,1),%rbp + movq %r9,0+8(%rsp) + movq %rbp,8+8(%rsp) + jmp .Lsqr8x_zero_start + +.align 32 +.byte 0x66,0x66,0x66,0x2e,0x0f,0x1f,0x84,0x00,0x00,0x00,0x00,0x00 +.Lsqrx8x_zero: +.byte 0x3e + movdqa %xmm0,0(%rdi) + movdqa %xmm0,16(%rdi) + movdqa %xmm0,32(%rdi) + movdqa %xmm0,48(%rdi) +.Lsqr8x_zero_start: + movdqa %xmm0,64(%rdi) + movdqa %xmm0,80(%rdi) + movdqa %xmm0,96(%rdi) + movdqa %xmm0,112(%rdi) + leaq 128(%rdi),%rdi + subq $64,%r9 + jnz .Lsqrx8x_zero + + movq 0(%rsi),%rdx + + xorq %r10,%r10 + xorq %r11,%r11 + xorq %r12,%r12 + xorq %r13,%r13 + xorq %r14,%r14 + xorq %r15,%r15 + leaq 48+8(%rsp),%rdi + xorq %rbp,%rbp + jmp .Lsqrx8x_outer_loop + +.align 32 +.Lsqrx8x_outer_loop: + mulxq 8(%rsi),%r8,%rax + adcxq %r9,%r8 + adoxq %rax,%r10 + mulxq 16(%rsi),%r9,%rax + adcxq %r10,%r9 + adoxq %rax,%r11 +.byte 0xc4,0xe2,0xab,0xf6,0x86,0x18,0x00,0x00,0x00 + adcxq %r11,%r10 + adoxq %rax,%r12 +.byte 0xc4,0xe2,0xa3,0xf6,0x86,0x20,0x00,0x00,0x00 + adcxq %r12,%r11 + adoxq %rax,%r13 + mulxq 40(%rsi),%r12,%rax + adcxq %r13,%r12 + adoxq %rax,%r14 + mulxq 48(%rsi),%r13,%rax + adcxq %r14,%r13 + adoxq %r15,%rax + mulxq 56(%rsi),%r14,%r15 + movq 8(%rsi),%rdx + adcxq %rax,%r14 + adoxq %rbp,%r15 + adcq 64(%rdi),%r15 + movq %r8,8(%rdi) + movq %r9,16(%rdi) + sbbq %rcx,%rcx + xorq %rbp,%rbp + + + mulxq 16(%rsi),%r8,%rbx + mulxq 24(%rsi),%r9,%rax + adcxq %r10,%r8 + adoxq %rbx,%r9 + mulxq 32(%rsi),%r10,%rbx + adcxq %r11,%r9 + adoxq %rax,%r10 +.byte 0xc4,0xe2,0xa3,0xf6,0x86,0x28,0x00,0x00,0x00 + adcxq %r12,%r10 + adoxq %rbx,%r11 +.byte 0xc4,0xe2,0x9b,0xf6,0x9e,0x30,0x00,0x00,0x00 + adcxq %r13,%r11 + adoxq %r14,%r12 +.byte 0xc4,0x62,0x93,0xf6,0xb6,0x38,0x00,0x00,0x00 + movq 16(%rsi),%rdx + adcxq %rax,%r12 + adoxq %rbx,%r13 + adcxq %r15,%r13 + adoxq %rbp,%r14 + adcxq %rbp,%r14 + + movq %r8,24(%rdi) + movq %r9,32(%rdi) + + mulxq 24(%rsi),%r8,%rbx + mulxq 32(%rsi),%r9,%rax + adcxq %r10,%r8 + adoxq %rbx,%r9 + mulxq 40(%rsi),%r10,%rbx + adcxq %r11,%r9 + adoxq %rax,%r10 +.byte 0xc4,0xe2,0xa3,0xf6,0x86,0x30,0x00,0x00,0x00 + adcxq %r12,%r10 + adoxq %r13,%r11 +.byte 0xc4,0x62,0x9b,0xf6,0xae,0x38,0x00,0x00,0x00 +.byte 0x3e + movq 24(%rsi),%rdx + adcxq %rbx,%r11 + adoxq %rax,%r12 + adcxq %r14,%r12 + movq %r8,40(%rdi) + movq %r9,48(%rdi) + mulxq 32(%rsi),%r8,%rax + adoxq %rbp,%r13 + adcxq %rbp,%r13 + + mulxq 40(%rsi),%r9,%rbx + adcxq %r10,%r8 + adoxq %rax,%r9 + mulxq 48(%rsi),%r10,%rax + adcxq %r11,%r9 + adoxq %r12,%r10 + mulxq 56(%rsi),%r11,%r12 + movq 32(%rsi),%rdx + movq 40(%rsi),%r14 + adcxq %rbx,%r10 + adoxq %rax,%r11 + movq 48(%rsi),%r15 + adcxq %r13,%r11 + adoxq %rbp,%r12 + adcxq %rbp,%r12 + + movq %r8,56(%rdi) + movq %r9,64(%rdi) + + mulxq %r14,%r9,%rax + movq 56(%rsi),%r8 + adcxq %r10,%r9 + mulxq %r15,%r10,%rbx + adoxq %rax,%r10 + adcxq %r11,%r10 + mulxq %r8,%r11,%rax + movq %r14,%rdx + adoxq %rbx,%r11 + adcxq %r12,%r11 + + adcxq %rbp,%rax + + mulxq %r15,%r14,%rbx + mulxq %r8,%r12,%r13 + movq %r15,%rdx + leaq 64(%rsi),%rsi + adcxq %r14,%r11 + adoxq %rbx,%r12 + adcxq %rax,%r12 + adoxq %rbp,%r13 + +.byte 0x67,0x67 + mulxq %r8,%r8,%r14 + adcxq %r8,%r13 + adcxq %rbp,%r14 + + cmpq 8+8(%rsp),%rsi + je .Lsqrx8x_outer_break + + negq %rcx + movq $-8,%rcx + movq %rbp,%r15 + movq 64(%rdi),%r8 + adcxq 72(%rdi),%r9 + adcxq 80(%rdi),%r10 + adcxq 88(%rdi),%r11 + adcq 96(%rdi),%r12 + adcq 104(%rdi),%r13 + adcq 112(%rdi),%r14 + adcq 120(%rdi),%r15 + leaq (%rsi),%rbp + leaq 128(%rdi),%rdi + sbbq %rax,%rax + + movq -64(%rsi),%rdx + movq %rax,16+8(%rsp) + movq %rdi,24+8(%rsp) + + + xorl %eax,%eax + jmp .Lsqrx8x_loop + +.align 32 +.Lsqrx8x_loop: + movq %r8,%rbx + mulxq 0(%rbp),%rax,%r8 + adcxq %rax,%rbx + adoxq %r9,%r8 + + mulxq 8(%rbp),%rax,%r9 + adcxq %rax,%r8 + adoxq %r10,%r9 + + mulxq 16(%rbp),%rax,%r10 + adcxq %rax,%r9 + adoxq %r11,%r10 + + mulxq 24(%rbp),%rax,%r11 + adcxq %rax,%r10 + adoxq %r12,%r11 + +.byte 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00 + adcxq %rax,%r11 + adoxq %r13,%r12 + + mulxq 40(%rbp),%rax,%r13 + adcxq %rax,%r12 + adoxq %r14,%r13 + + mulxq 48(%rbp),%rax,%r14 + movq %rbx,(%rdi,%rcx,8) + movl $0,%ebx + adcxq %rax,%r13 + adoxq %r15,%r14 + +.byte 0xc4,0x62,0xfb,0xf6,0xbd,0x38,0x00,0x00,0x00 + movq 8(%rsi,%rcx,8),%rdx + adcxq %rax,%r14 + adoxq %rbx,%r15 + adcxq %rbx,%r15 + +.byte 0x67 + incq %rcx + jnz .Lsqrx8x_loop + + leaq 64(%rbp),%rbp + movq $-8,%rcx + cmpq 8+8(%rsp),%rbp + je .Lsqrx8x_break + + subq 16+8(%rsp),%rbx +.byte 0x66 + movq -64(%rsi),%rdx + adcxq 0(%rdi),%r8 + adcxq 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 + leaq 64(%rdi),%rdi +.byte 0x67 + sbbq %rax,%rax + xorl %ebx,%ebx + movq %rax,16+8(%rsp) + jmp .Lsqrx8x_loop + +.align 32 +.Lsqrx8x_break: + subq 16+8(%rsp),%r8 + movq 24+8(%rsp),%rcx + movq 0(%rsi),%rdx + xorl %ebp,%ebp + movq %r8,0(%rdi) + cmpq %rcx,%rdi + je .Lsqrx8x_outer_loop + + movq %r9,8(%rdi) + movq 8(%rcx),%r9 + movq %r10,16(%rdi) + movq 16(%rcx),%r10 + movq %r11,24(%rdi) + movq 24(%rcx),%r11 + movq %r12,32(%rdi) + movq 32(%rcx),%r12 + movq %r13,40(%rdi) + movq 40(%rcx),%r13 + movq %r14,48(%rdi) + movq 48(%rcx),%r14 + movq %r15,56(%rdi) + movq 56(%rcx),%r15 + movq %rcx,%rdi + jmp .Lsqrx8x_outer_loop + +.align 32 +.Lsqrx8x_outer_break: + movq %r9,72(%rdi) +.byte 102,72,15,126,217 + movq %r10,80(%rdi) + movq %r11,88(%rdi) + movq %r12,96(%rdi) + movq %r13,104(%rdi) + movq %r14,112(%rdi) + leaq 48+8(%rsp),%rdi + movq (%rsi,%rcx,1),%rdx + + movq 8(%rdi),%r11 + xorq %r10,%r10 + movq 0+8(%rsp),%r9 + adoxq %r11,%r11 + movq 16(%rdi),%r12 + movq 24(%rdi),%r13 + + +.align 32 +.Lsqrx4x_shift_n_add: + mulxq %rdx,%rax,%rbx + adoxq %r12,%r12 + adcxq %r10,%rax +.byte 0x48,0x8b,0x94,0x0e,0x08,0x00,0x00,0x00 +.byte 0x4c,0x8b,0x97,0x20,0x00,0x00,0x00 + adoxq %r13,%r13 + adcxq %r11,%rbx + movq 40(%rdi),%r11 + movq %rax,0(%rdi) + movq %rbx,8(%rdi) + + mulxq %rdx,%rax,%rbx + adoxq %r10,%r10 + adcxq %r12,%rax + movq 16(%rsi,%rcx,1),%rdx + movq 48(%rdi),%r12 + adoxq %r11,%r11 + adcxq %r13,%rbx + movq 56(%rdi),%r13 + movq %rax,16(%rdi) + movq %rbx,24(%rdi) + + mulxq %rdx,%rax,%rbx + adoxq %r12,%r12 + adcxq %r10,%rax + movq 24(%rsi,%rcx,1),%rdx + leaq 32(%rcx),%rcx + movq 64(%rdi),%r10 + adoxq %r13,%r13 + adcxq %r11,%rbx + movq 72(%rdi),%r11 + movq %rax,32(%rdi) + movq %rbx,40(%rdi) + + mulxq %rdx,%rax,%rbx + adoxq %r10,%r10 + adcxq %r12,%rax + jrcxz .Lsqrx4x_shift_n_add_break +.byte 0x48,0x8b,0x94,0x0e,0x00,0x00,0x00,0x00 + adoxq %r11,%r11 + adcxq %r13,%rbx + movq 80(%rdi),%r12 + movq 88(%rdi),%r13 + movq %rax,48(%rdi) + movq %rbx,56(%rdi) + leaq 64(%rdi),%rdi + nop + jmp .Lsqrx4x_shift_n_add + +.align 32 +.Lsqrx4x_shift_n_add_break: + adcxq %r13,%rbx + movq %rax,48(%rdi) + movq %rbx,56(%rdi) + leaq 64(%rdi),%rdi +.byte 102,72,15,126,213 +__bn_sqrx8x_reduction: + xorl %eax,%eax + movq 32+8(%rsp),%rbx + movq 48+8(%rsp),%rdx + leaq -64(%rbp,%r9,1),%rcx + + movq %rcx,0+8(%rsp) + movq %rdi,8+8(%rsp) + + leaq 48+8(%rsp),%rdi + jmp .Lsqrx8x_reduction_loop + +.align 32 +.Lsqrx8x_reduction_loop: + movq 8(%rdi),%r9 + movq 16(%rdi),%r10 + movq 24(%rdi),%r11 + movq 32(%rdi),%r12 + movq %rdx,%r8 + imulq %rbx,%rdx + movq 40(%rdi),%r13 + movq 48(%rdi),%r14 + movq 56(%rdi),%r15 + movq %rax,24+8(%rsp) + + leaq 64(%rdi),%rdi + xorq %rsi,%rsi + movq $-8,%rcx + jmp .Lsqrx8x_reduce + +.align 32 +.Lsqrx8x_reduce: + movq %r8,%rbx + mulxq 0(%rbp),%rax,%r8 + adcxq %rbx,%rax + adoxq %r9,%r8 + + mulxq 8(%rbp),%rbx,%r9 + adcxq %rbx,%r8 + adoxq %r10,%r9 + + mulxq 16(%rbp),%rbx,%r10 + adcxq %rbx,%r9 + adoxq %r11,%r10 + + mulxq 24(%rbp),%rbx,%r11 + adcxq %rbx,%r10 + adoxq %r12,%r11 + +.byte 0xc4,0x62,0xe3,0xf6,0xa5,0x20,0x00,0x00,0x00 + movq %rdx,%rax + movq %r8,%rdx + adcxq %rbx,%r11 + adoxq %r13,%r12 + + mulxq 32+8(%rsp),%rbx,%rdx + movq %rax,%rdx + movq %rax,64+48+8(%rsp,%rcx,8) + + mulxq 40(%rbp),%rax,%r13 + adcxq %rax,%r12 + adoxq %r14,%r13 + + mulxq 48(%rbp),%rax,%r14 + adcxq %rax,%r13 + adoxq %r15,%r14 + + mulxq 56(%rbp),%rax,%r15 + movq %rbx,%rdx + adcxq %rax,%r14 + adoxq %rsi,%r15 + adcxq %rsi,%r15 + +.byte 0x67,0x67,0x67 + incq %rcx + jnz .Lsqrx8x_reduce + + movq %rsi,%rax + cmpq 0+8(%rsp),%rbp + jae .Lsqrx8x_no_tail + + movq 48+8(%rsp),%rdx + addq 0(%rdi),%r8 + leaq 64(%rbp),%rbp + movq $-8,%rcx + adcxq 8(%rdi),%r9 + adcxq 16(%rdi),%r10 + adcq 24(%rdi),%r11 + adcq 32(%rdi),%r12 + adcq 40(%rdi),%r13 + adcq 48(%rdi),%r14 + adcq 56(%rdi),%r15 + leaq 64(%rdi),%rdi + sbbq %rax,%rax + + xorq %rsi,%rsi + movq %rax,16+8(%rsp) + jmp .Lsqrx8x_tail + +.align 32 +.Lsqrx8x_tail: + movq %r8,%rbx + mulxq 0(%rbp),%rax,%r8 + adcxq %rax,%rbx + adoxq %r9,%r8 + + mulxq 8(%rbp),%rax,%r9 + adcxq %rax,%r8 + adoxq %r10,%r9 + + mulxq 16(%rbp),%rax,%r10 + adcxq %rax,%r9 + adoxq %r11,%r10 + + mulxq 24(%rbp),%rax,%r11 + adcxq %rax,%r10 + adoxq %r12,%r11 + +.byte 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00 + adcxq %rax,%r11 + adoxq %r13,%r12 + + mulxq 40(%rbp),%rax,%r13 + adcxq %rax,%r12 + adoxq %r14,%r13 + + mulxq 48(%rbp),%rax,%r14 + adcxq %rax,%r13 + adoxq %r15,%r14 + + mulxq 56(%rbp),%rax,%r15 + movq 72+48+8(%rsp,%rcx,8),%rdx + adcxq %rax,%r14 + adoxq %rsi,%r15 + movq %rbx,(%rdi,%rcx,8) + movq %r8,%rbx + adcxq %rsi,%r15 + + incq %rcx + jnz .Lsqrx8x_tail + + cmpq 0+8(%rsp),%rbp + jae .Lsqrx8x_tail_done + + subq 16+8(%rsp),%rsi + movq 48+8(%rsp),%rdx + leaq 64(%rbp),%rbp + 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 + leaq 64(%rdi),%rdi + sbbq %rax,%rax + subq $8,%rcx + + xorq %rsi,%rsi + movq %rax,16+8(%rsp) + jmp .Lsqrx8x_tail + +.align 32 +.Lsqrx8x_tail_done: + addq 24+8(%rsp),%r8 + adcq $0,%r9 + adcq $0,%r10 + adcq $0,%r11 + adcq $0,%r12 + adcq $0,%r13 + adcq $0,%r14 + adcq $0,%r15 + + + movq %rsi,%rax + + subq 16+8(%rsp),%rsi +.Lsqrx8x_no_tail: + adcq 0(%rdi),%r8 +.byte 102,72,15,126,217 + adcq 8(%rdi),%r9 + movq 56(%rbp),%rsi +.byte 102,72,15,126,213 + 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 %rax,%rax + + movq 32+8(%rsp),%rbx + movq 64(%rdi,%rcx,1),%rdx + + movq %r8,0(%rdi) + leaq 64(%rdi),%r8 + movq %r9,8(%rdi) + 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,%rcx,1),%rdi + cmpq 8+8(%rsp),%r8 + jb .Lsqrx8x_reduction_loop + .byte 0xf3,0xc3 +.size bn_sqrx8x_internal,.-bn_sqrx8x_internal +.align 32 +__bn_postx4x_internal: + movq 0(%rbp),%r12 + movq %rcx,%r10 + movq %rcx,%r9 + negq %rax + sarq $3+2,%rcx + +.byte 102,72,15,126,202 +.byte 102,72,15,126,206 + decq %r12 + movq 8(%rbp),%r13 + xorq %r8,%r8 + movq 16(%rbp),%r14 + movq 24(%rbp),%r15 + jmp .Lsqrx4x_sub_entry + +.align 16 +.Lsqrx4x_sub: + movq 0(%rbp),%r12 + movq 8(%rbp),%r13 + movq 16(%rbp),%r14 + movq 24(%rbp),%r15 +.Lsqrx4x_sub_entry: + andnq %rax,%r12,%r12 + leaq 32(%rbp),%rbp + andnq %rax,%r13,%r13 + andnq %rax,%r14,%r14 + andnq %rax,%r15,%r15 + + negq %r8 + adcq 0(%rdi),%r12 + adcq 8(%rdi),%r13 + adcq 16(%rdi),%r14 + adcq 24(%rdi),%r15 + movq %r12,0(%rdx) + leaq 32(%rdi),%rdi + movq %r13,8(%rdx) + sbbq %r8,%r8 + movq %r14,16(%rdx) + movq %r15,24(%rdx) + leaq 32(%rdx),%rdx + + incq %rcx + jnz .Lsqrx4x_sub + + negq %r9 + + .byte 0xf3,0xc3 +.size __bn_postx4x_internal,.-__bn_postx4x_internal .globl bn_get_bits5 .type bn_get_bits5,@function .align 16 |
