diff options
| author | polwex <polwex@sortug.com> | 2025-10-05 21:56:51 +0700 |
|---|---|---|
| committer | polwex <polwex@sortug.com> | 2025-10-05 21:56:51 +0700 |
| commit | fcedfddf00b3f994e4f4e40332ac7fc192c63244 (patch) | |
| tree | 51d38e62c7bdfcc5f9a5e9435fe820c93cfc9a3d /vere/ext/openssl/gen/windows-x86_64/crypto/bn | |
claude is gud
Diffstat (limited to 'vere/ext/openssl/gen/windows-x86_64/crypto/bn')
5 files changed, 9981 insertions, 0 deletions
diff --git a/vere/ext/openssl/gen/windows-x86_64/crypto/bn/rsaz-avx2.asm b/vere/ext/openssl/gen/windows-x86_64/crypto/bn/rsaz-avx2.asm new file mode 100644 index 0000000..f854d17 --- /dev/null +++ b/vere/ext/openssl/gen/windows-x86_64/crypto/bn/rsaz-avx2.asm @@ -0,0 +1,1981 @@ +default rel +%define XMMWORD +%define YMMWORD +%define ZMMWORD +section .text code align=64 + + +global rsaz_1024_sqr_avx2 + +ALIGN 64 +rsaz_1024_sqr_avx2: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[16+rsp],rsi + mov rax,rsp +$L$SEH_begin_rsaz_1024_sqr_avx2: + mov rdi,rcx + mov rsi,rdx + mov rdx,r8 + mov rcx,r9 + mov r8,QWORD[40+rsp] + + + + lea rax,[rsp] + + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + + vzeroupper + lea rsp,[((-168))+rsp] + vmovaps XMMWORD[(-216)+rax],xmm6 + vmovaps XMMWORD[(-200)+rax],xmm7 + vmovaps XMMWORD[(-184)+rax],xmm8 + vmovaps XMMWORD[(-168)+rax],xmm9 + vmovaps XMMWORD[(-152)+rax],xmm10 + vmovaps XMMWORD[(-136)+rax],xmm11 + vmovaps XMMWORD[(-120)+rax],xmm12 + vmovaps XMMWORD[(-104)+rax],xmm13 + vmovaps XMMWORD[(-88)+rax],xmm14 + vmovaps XMMWORD[(-72)+rax],xmm15 +$L$sqr_1024_body: + mov rbp,rax + + mov r13,rdx + sub rsp,832 + mov r15,r13 + sub rdi,-128 + sub rsi,-128 + sub r13,-128 + + and r15,4095 + add r15,32*10 + shr r15,12 + vpxor ymm9,ymm9,ymm9 + jz NEAR $L$sqr_1024_no_n_copy + + + + + + sub rsp,32*10 + vmovdqu ymm0,YMMWORD[((0-128))+r13] + and rsp,-2048 + vmovdqu ymm1,YMMWORD[((32-128))+r13] + vmovdqu ymm2,YMMWORD[((64-128))+r13] + vmovdqu ymm3,YMMWORD[((96-128))+r13] + vmovdqu ymm4,YMMWORD[((128-128))+r13] + vmovdqu ymm5,YMMWORD[((160-128))+r13] + vmovdqu ymm6,YMMWORD[((192-128))+r13] + vmovdqu ymm7,YMMWORD[((224-128))+r13] + vmovdqu ymm8,YMMWORD[((256-128))+r13] + lea r13,[((832+128))+rsp] + vmovdqu YMMWORD[(0-128)+r13],ymm0 + vmovdqu YMMWORD[(32-128)+r13],ymm1 + vmovdqu YMMWORD[(64-128)+r13],ymm2 + vmovdqu YMMWORD[(96-128)+r13],ymm3 + vmovdqu YMMWORD[(128-128)+r13],ymm4 + vmovdqu YMMWORD[(160-128)+r13],ymm5 + vmovdqu YMMWORD[(192-128)+r13],ymm6 + vmovdqu YMMWORD[(224-128)+r13],ymm7 + vmovdqu YMMWORD[(256-128)+r13],ymm8 + vmovdqu YMMWORD[(288-128)+r13],ymm9 + +$L$sqr_1024_no_n_copy: + and rsp,-1024 + + vmovdqu ymm1,YMMWORD[((32-128))+rsi] + vmovdqu ymm2,YMMWORD[((64-128))+rsi] + vmovdqu ymm3,YMMWORD[((96-128))+rsi] + vmovdqu ymm4,YMMWORD[((128-128))+rsi] + vmovdqu ymm5,YMMWORD[((160-128))+rsi] + vmovdqu ymm6,YMMWORD[((192-128))+rsi] + vmovdqu ymm7,YMMWORD[((224-128))+rsi] + vmovdqu ymm8,YMMWORD[((256-128))+rsi] + + lea rbx,[192+rsp] + vmovdqu ymm15,YMMWORD[$L$and_mask] + jmp NEAR $L$OOP_GRANDE_SQR_1024 + +ALIGN 32 +$L$OOP_GRANDE_SQR_1024: + lea r9,[((576+128))+rsp] + lea r12,[448+rsp] + + + + + vpaddq ymm1,ymm1,ymm1 + vpbroadcastq ymm10,QWORD[((0-128))+rsi] + vpaddq ymm2,ymm2,ymm2 + vmovdqa YMMWORD[(0-128)+r9],ymm1 + vpaddq ymm3,ymm3,ymm3 + vmovdqa YMMWORD[(32-128)+r9],ymm2 + vpaddq ymm4,ymm4,ymm4 + vmovdqa YMMWORD[(64-128)+r9],ymm3 + vpaddq ymm5,ymm5,ymm5 + vmovdqa YMMWORD[(96-128)+r9],ymm4 + vpaddq ymm6,ymm6,ymm6 + vmovdqa YMMWORD[(128-128)+r9],ymm5 + vpaddq ymm7,ymm7,ymm7 + vmovdqa YMMWORD[(160-128)+r9],ymm6 + vpaddq ymm8,ymm8,ymm8 + vmovdqa YMMWORD[(192-128)+r9],ymm7 + vpxor ymm9,ymm9,ymm9 + vmovdqa YMMWORD[(224-128)+r9],ymm8 + + vpmuludq ymm0,ymm10,YMMWORD[((0-128))+rsi] + vpbroadcastq ymm11,QWORD[((32-128))+rsi] + vmovdqu YMMWORD[(288-192)+rbx],ymm9 + vpmuludq ymm1,ymm1,ymm10 + vmovdqu YMMWORD[(320-448)+r12],ymm9 + vpmuludq ymm2,ymm2,ymm10 + vmovdqu YMMWORD[(352-448)+r12],ymm9 + vpmuludq ymm3,ymm3,ymm10 + vmovdqu YMMWORD[(384-448)+r12],ymm9 + vpmuludq ymm4,ymm4,ymm10 + vmovdqu YMMWORD[(416-448)+r12],ymm9 + vpmuludq ymm5,ymm5,ymm10 + vmovdqu YMMWORD[(448-448)+r12],ymm9 + vpmuludq ymm6,ymm6,ymm10 + vmovdqu YMMWORD[(480-448)+r12],ymm9 + vpmuludq ymm7,ymm7,ymm10 + vmovdqu YMMWORD[(512-448)+r12],ymm9 + vpmuludq ymm8,ymm8,ymm10 + vpbroadcastq ymm10,QWORD[((64-128))+rsi] + vmovdqu YMMWORD[(544-448)+r12],ymm9 + + mov r15,rsi + mov r14d,4 + jmp NEAR $L$sqr_entry_1024 +ALIGN 32 +$L$OOP_SQR_1024: + vpbroadcastq ymm11,QWORD[((32-128))+r15] + vpmuludq ymm0,ymm10,YMMWORD[((0-128))+rsi] + vpaddq ymm0,ymm0,YMMWORD[((0-192))+rbx] + vpmuludq ymm1,ymm10,YMMWORD[((0-128))+r9] + vpaddq ymm1,ymm1,YMMWORD[((32-192))+rbx] + vpmuludq ymm2,ymm10,YMMWORD[((32-128))+r9] + vpaddq ymm2,ymm2,YMMWORD[((64-192))+rbx] + vpmuludq ymm3,ymm10,YMMWORD[((64-128))+r9] + vpaddq ymm3,ymm3,YMMWORD[((96-192))+rbx] + vpmuludq ymm4,ymm10,YMMWORD[((96-128))+r9] + vpaddq ymm4,ymm4,YMMWORD[((128-192))+rbx] + vpmuludq ymm5,ymm10,YMMWORD[((128-128))+r9] + vpaddq ymm5,ymm5,YMMWORD[((160-192))+rbx] + vpmuludq ymm6,ymm10,YMMWORD[((160-128))+r9] + vpaddq ymm6,ymm6,YMMWORD[((192-192))+rbx] + vpmuludq ymm7,ymm10,YMMWORD[((192-128))+r9] + vpaddq ymm7,ymm7,YMMWORD[((224-192))+rbx] + vpmuludq ymm8,ymm10,YMMWORD[((224-128))+r9] + vpbroadcastq ymm10,QWORD[((64-128))+r15] + vpaddq ymm8,ymm8,YMMWORD[((256-192))+rbx] +$L$sqr_entry_1024: + vmovdqu YMMWORD[(0-192)+rbx],ymm0 + vmovdqu YMMWORD[(32-192)+rbx],ymm1 + + vpmuludq ymm12,ymm11,YMMWORD[((32-128))+rsi] + vpaddq ymm2,ymm2,ymm12 + vpmuludq ymm14,ymm11,YMMWORD[((32-128))+r9] + vpaddq ymm3,ymm3,ymm14 + vpmuludq ymm13,ymm11,YMMWORD[((64-128))+r9] + vpaddq ymm4,ymm4,ymm13 + vpmuludq ymm12,ymm11,YMMWORD[((96-128))+r9] + vpaddq ymm5,ymm5,ymm12 + vpmuludq ymm14,ymm11,YMMWORD[((128-128))+r9] + vpaddq ymm6,ymm6,ymm14 + vpmuludq ymm13,ymm11,YMMWORD[((160-128))+r9] + vpaddq ymm7,ymm7,ymm13 + vpmuludq ymm12,ymm11,YMMWORD[((192-128))+r9] + vpaddq ymm8,ymm8,ymm12 + vpmuludq ymm0,ymm11,YMMWORD[((224-128))+r9] + vpbroadcastq ymm11,QWORD[((96-128))+r15] + vpaddq ymm0,ymm0,YMMWORD[((288-192))+rbx] + + vmovdqu YMMWORD[(64-192)+rbx],ymm2 + vmovdqu YMMWORD[(96-192)+rbx],ymm3 + + vpmuludq ymm13,ymm10,YMMWORD[((64-128))+rsi] + vpaddq ymm4,ymm4,ymm13 + vpmuludq ymm12,ymm10,YMMWORD[((64-128))+r9] + vpaddq ymm5,ymm5,ymm12 + vpmuludq ymm14,ymm10,YMMWORD[((96-128))+r9] + vpaddq ymm6,ymm6,ymm14 + vpmuludq ymm13,ymm10,YMMWORD[((128-128))+r9] + vpaddq ymm7,ymm7,ymm13 + vpmuludq ymm12,ymm10,YMMWORD[((160-128))+r9] + vpaddq ymm8,ymm8,ymm12 + vpmuludq ymm14,ymm10,YMMWORD[((192-128))+r9] + vpaddq ymm0,ymm0,ymm14 + vpmuludq ymm1,ymm10,YMMWORD[((224-128))+r9] + vpbroadcastq ymm10,QWORD[((128-128))+r15] + vpaddq ymm1,ymm1,YMMWORD[((320-448))+r12] + + vmovdqu YMMWORD[(128-192)+rbx],ymm4 + vmovdqu YMMWORD[(160-192)+rbx],ymm5 + + vpmuludq ymm12,ymm11,YMMWORD[((96-128))+rsi] + vpaddq ymm6,ymm6,ymm12 + vpmuludq ymm14,ymm11,YMMWORD[((96-128))+r9] + vpaddq ymm7,ymm7,ymm14 + vpmuludq ymm13,ymm11,YMMWORD[((128-128))+r9] + vpaddq ymm8,ymm8,ymm13 + vpmuludq ymm12,ymm11,YMMWORD[((160-128))+r9] + vpaddq ymm0,ymm0,ymm12 + vpmuludq ymm14,ymm11,YMMWORD[((192-128))+r9] + vpaddq ymm1,ymm1,ymm14 + vpmuludq ymm2,ymm11,YMMWORD[((224-128))+r9] + vpbroadcastq ymm11,QWORD[((160-128))+r15] + vpaddq ymm2,ymm2,YMMWORD[((352-448))+r12] + + vmovdqu YMMWORD[(192-192)+rbx],ymm6 + vmovdqu YMMWORD[(224-192)+rbx],ymm7 + + vpmuludq ymm12,ymm10,YMMWORD[((128-128))+rsi] + vpaddq ymm8,ymm8,ymm12 + vpmuludq ymm14,ymm10,YMMWORD[((128-128))+r9] + vpaddq ymm0,ymm0,ymm14 + vpmuludq ymm13,ymm10,YMMWORD[((160-128))+r9] + vpaddq ymm1,ymm1,ymm13 + vpmuludq ymm12,ymm10,YMMWORD[((192-128))+r9] + vpaddq ymm2,ymm2,ymm12 + vpmuludq ymm3,ymm10,YMMWORD[((224-128))+r9] + vpbroadcastq ymm10,QWORD[((192-128))+r15] + vpaddq ymm3,ymm3,YMMWORD[((384-448))+r12] + + vmovdqu YMMWORD[(256-192)+rbx],ymm8 + vmovdqu YMMWORD[(288-192)+rbx],ymm0 + lea rbx,[8+rbx] + + vpmuludq ymm13,ymm11,YMMWORD[((160-128))+rsi] + vpaddq ymm1,ymm1,ymm13 + vpmuludq ymm12,ymm11,YMMWORD[((160-128))+r9] + vpaddq ymm2,ymm2,ymm12 + vpmuludq ymm14,ymm11,YMMWORD[((192-128))+r9] + vpaddq ymm3,ymm3,ymm14 + vpmuludq ymm4,ymm11,YMMWORD[((224-128))+r9] + vpbroadcastq ymm11,QWORD[((224-128))+r15] + vpaddq ymm4,ymm4,YMMWORD[((416-448))+r12] + + vmovdqu YMMWORD[(320-448)+r12],ymm1 + vmovdqu YMMWORD[(352-448)+r12],ymm2 + + vpmuludq ymm12,ymm10,YMMWORD[((192-128))+rsi] + vpaddq ymm3,ymm3,ymm12 + vpmuludq ymm14,ymm10,YMMWORD[((192-128))+r9] + vpbroadcastq ymm0,QWORD[((256-128))+r15] + vpaddq ymm4,ymm4,ymm14 + vpmuludq ymm5,ymm10,YMMWORD[((224-128))+r9] + vpbroadcastq ymm10,QWORD[((0+8-128))+r15] + vpaddq ymm5,ymm5,YMMWORD[((448-448))+r12] + + vmovdqu YMMWORD[(384-448)+r12],ymm3 + vmovdqu YMMWORD[(416-448)+r12],ymm4 + lea r15,[8+r15] + + vpmuludq ymm12,ymm11,YMMWORD[((224-128))+rsi] + vpaddq ymm5,ymm5,ymm12 + vpmuludq ymm6,ymm11,YMMWORD[((224-128))+r9] + vpaddq ymm6,ymm6,YMMWORD[((480-448))+r12] + + vpmuludq ymm7,ymm0,YMMWORD[((256-128))+rsi] + vmovdqu YMMWORD[(448-448)+r12],ymm5 + vpaddq ymm7,ymm7,YMMWORD[((512-448))+r12] + vmovdqu YMMWORD[(480-448)+r12],ymm6 + vmovdqu YMMWORD[(512-448)+r12],ymm7 + lea r12,[8+r12] + + dec r14d + jnz NEAR $L$OOP_SQR_1024 + + vmovdqu ymm8,YMMWORD[256+rsp] + vmovdqu ymm1,YMMWORD[288+rsp] + vmovdqu ymm2,YMMWORD[320+rsp] + lea rbx,[192+rsp] + + vpsrlq ymm14,ymm8,29 + vpand ymm8,ymm8,ymm15 + vpsrlq ymm11,ymm1,29 + vpand ymm1,ymm1,ymm15 + + vpermq ymm14,ymm14,0x93 + vpxor ymm9,ymm9,ymm9 + vpermq ymm11,ymm11,0x93 + + vpblendd ymm10,ymm14,ymm9,3 + vpblendd ymm14,ymm11,ymm14,3 + vpaddq ymm8,ymm8,ymm10 + vpblendd ymm11,ymm9,ymm11,3 + vpaddq ymm1,ymm1,ymm14 + vpaddq ymm2,ymm2,ymm11 + vmovdqu YMMWORD[(288-192)+rbx],ymm1 + vmovdqu YMMWORD[(320-192)+rbx],ymm2 + + mov rax,QWORD[rsp] + mov r10,QWORD[8+rsp] + mov r11,QWORD[16+rsp] + mov r12,QWORD[24+rsp] + vmovdqu ymm1,YMMWORD[32+rsp] + vmovdqu ymm2,YMMWORD[((64-192))+rbx] + vmovdqu ymm3,YMMWORD[((96-192))+rbx] + vmovdqu ymm4,YMMWORD[((128-192))+rbx] + vmovdqu ymm5,YMMWORD[((160-192))+rbx] + vmovdqu ymm6,YMMWORD[((192-192))+rbx] + vmovdqu ymm7,YMMWORD[((224-192))+rbx] + + mov r9,rax + imul eax,ecx + and eax,0x1fffffff + vmovd xmm12,eax + + mov rdx,rax + imul rax,QWORD[((-128))+r13] + vpbroadcastq ymm12,xmm12 + add r9,rax + mov rax,rdx + imul rax,QWORD[((8-128))+r13] + shr r9,29 + add r10,rax + mov rax,rdx + imul rax,QWORD[((16-128))+r13] + add r10,r9 + add r11,rax + imul rdx,QWORD[((24-128))+r13] + add r12,rdx + + mov rax,r10 + imul eax,ecx + and eax,0x1fffffff + + mov r14d,9 + jmp NEAR $L$OOP_REDUCE_1024 + +ALIGN 32 +$L$OOP_REDUCE_1024: + vmovd xmm13,eax + vpbroadcastq ymm13,xmm13 + + vpmuludq ymm10,ymm12,YMMWORD[((32-128))+r13] + mov rdx,rax + imul rax,QWORD[((-128))+r13] + vpaddq ymm1,ymm1,ymm10 + add r10,rax + vpmuludq ymm14,ymm12,YMMWORD[((64-128))+r13] + mov rax,rdx + imul rax,QWORD[((8-128))+r13] + vpaddq ymm2,ymm2,ymm14 + vpmuludq ymm11,ymm12,YMMWORD[((96-128))+r13] +DB 0x67 + add r11,rax +DB 0x67 + mov rax,rdx + imul rax,QWORD[((16-128))+r13] + shr r10,29 + vpaddq ymm3,ymm3,ymm11 + vpmuludq ymm10,ymm12,YMMWORD[((128-128))+r13] + add r12,rax + add r11,r10 + vpaddq ymm4,ymm4,ymm10 + vpmuludq ymm14,ymm12,YMMWORD[((160-128))+r13] + mov rax,r11 + imul eax,ecx + vpaddq ymm5,ymm5,ymm14 + vpmuludq ymm11,ymm12,YMMWORD[((192-128))+r13] + and eax,0x1fffffff + vpaddq ymm6,ymm6,ymm11 + vpmuludq ymm10,ymm12,YMMWORD[((224-128))+r13] + vpaddq ymm7,ymm7,ymm10 + vpmuludq ymm14,ymm12,YMMWORD[((256-128))+r13] + vmovd xmm12,eax + + vpaddq ymm8,ymm8,ymm14 + + vpbroadcastq ymm12,xmm12 + + vpmuludq ymm11,ymm13,YMMWORD[((32-8-128))+r13] + vmovdqu ymm14,YMMWORD[((96-8-128))+r13] + mov rdx,rax + imul rax,QWORD[((-128))+r13] + vpaddq ymm1,ymm1,ymm11 + vpmuludq ymm10,ymm13,YMMWORD[((64-8-128))+r13] + vmovdqu ymm11,YMMWORD[((128-8-128))+r13] + add r11,rax + mov rax,rdx + imul rax,QWORD[((8-128))+r13] + vpaddq ymm2,ymm2,ymm10 + add rax,r12 + shr r11,29 + vpmuludq ymm14,ymm14,ymm13 + vmovdqu ymm10,YMMWORD[((160-8-128))+r13] + add rax,r11 + vpaddq ymm3,ymm3,ymm14 + vpmuludq ymm11,ymm11,ymm13 + vmovdqu ymm14,YMMWORD[((192-8-128))+r13] +DB 0x67 + mov r12,rax + imul eax,ecx + vpaddq ymm4,ymm4,ymm11 + vpmuludq ymm10,ymm10,ymm13 +DB 0xc4,0x41,0x7e,0x6f,0x9d,0x58,0x00,0x00,0x00 + and eax,0x1fffffff + vpaddq ymm5,ymm5,ymm10 + vpmuludq ymm14,ymm14,ymm13 + vmovdqu ymm10,YMMWORD[((256-8-128))+r13] + vpaddq ymm6,ymm6,ymm14 + vpmuludq ymm11,ymm11,ymm13 + vmovdqu ymm9,YMMWORD[((288-8-128))+r13] + vmovd xmm0,eax + imul rax,QWORD[((-128))+r13] + vpaddq ymm7,ymm7,ymm11 + vpmuludq ymm10,ymm10,ymm13 + vmovdqu ymm14,YMMWORD[((32-16-128))+r13] + vpbroadcastq ymm0,xmm0 + vpaddq ymm8,ymm8,ymm10 + vpmuludq ymm9,ymm9,ymm13 + vmovdqu ymm11,YMMWORD[((64-16-128))+r13] + add r12,rax + + vmovdqu ymm13,YMMWORD[((32-24-128))+r13] + vpmuludq ymm14,ymm14,ymm12 + vmovdqu ymm10,YMMWORD[((96-16-128))+r13] + vpaddq ymm1,ymm1,ymm14 + vpmuludq ymm13,ymm13,ymm0 + vpmuludq ymm11,ymm11,ymm12 +DB 0xc4,0x41,0x7e,0x6f,0xb5,0xf0,0xff,0xff,0xff + vpaddq ymm13,ymm13,ymm1 + vpaddq ymm2,ymm2,ymm11 + vpmuludq ymm10,ymm10,ymm12 + vmovdqu ymm11,YMMWORD[((160-16-128))+r13] +DB 0x67 + vmovq rax,xmm13 + vmovdqu YMMWORD[rsp],ymm13 + vpaddq ymm3,ymm3,ymm10 + vpmuludq ymm14,ymm14,ymm12 + vmovdqu ymm10,YMMWORD[((192-16-128))+r13] + vpaddq ymm4,ymm4,ymm14 + vpmuludq ymm11,ymm11,ymm12 + vmovdqu ymm14,YMMWORD[((224-16-128))+r13] + vpaddq ymm5,ymm5,ymm11 + vpmuludq ymm10,ymm10,ymm12 + vmovdqu ymm11,YMMWORD[((256-16-128))+r13] + vpaddq ymm6,ymm6,ymm10 + vpmuludq ymm14,ymm14,ymm12 + shr r12,29 + vmovdqu ymm10,YMMWORD[((288-16-128))+r13] + add rax,r12 + vpaddq ymm7,ymm7,ymm14 + vpmuludq ymm11,ymm11,ymm12 + + mov r9,rax + imul eax,ecx + vpaddq ymm8,ymm8,ymm11 + vpmuludq ymm10,ymm10,ymm12 + and eax,0x1fffffff + vmovd xmm12,eax + vmovdqu ymm11,YMMWORD[((96-24-128))+r13] +DB 0x67 + vpaddq ymm9,ymm9,ymm10 + vpbroadcastq ymm12,xmm12 + + vpmuludq ymm14,ymm0,YMMWORD[((64-24-128))+r13] + vmovdqu ymm10,YMMWORD[((128-24-128))+r13] + mov rdx,rax + imul rax,QWORD[((-128))+r13] + mov r10,QWORD[8+rsp] + vpaddq ymm1,ymm2,ymm14 + vpmuludq ymm11,ymm11,ymm0 + vmovdqu ymm14,YMMWORD[((160-24-128))+r13] + add r9,rax + mov rax,rdx + imul rax,QWORD[((8-128))+r13] +DB 0x67 + shr r9,29 + mov r11,QWORD[16+rsp] + vpaddq ymm2,ymm3,ymm11 + vpmuludq ymm10,ymm10,ymm0 + vmovdqu ymm11,YMMWORD[((192-24-128))+r13] + add r10,rax + mov rax,rdx + imul rax,QWORD[((16-128))+r13] + vpaddq ymm3,ymm4,ymm10 + vpmuludq ymm14,ymm14,ymm0 + vmovdqu ymm10,YMMWORD[((224-24-128))+r13] + imul rdx,QWORD[((24-128))+r13] + add r11,rax + lea rax,[r10*1+r9] + vpaddq ymm4,ymm5,ymm14 + vpmuludq ymm11,ymm11,ymm0 + vmovdqu ymm14,YMMWORD[((256-24-128))+r13] + mov r10,rax + imul eax,ecx + vpmuludq ymm10,ymm10,ymm0 + vpaddq ymm5,ymm6,ymm11 + vmovdqu ymm11,YMMWORD[((288-24-128))+r13] + and eax,0x1fffffff + vpaddq ymm6,ymm7,ymm10 + vpmuludq ymm14,ymm14,ymm0 + add rdx,QWORD[24+rsp] + vpaddq ymm7,ymm8,ymm14 + vpmuludq ymm11,ymm11,ymm0 + vpaddq ymm8,ymm9,ymm11 + vmovq xmm9,r12 + mov r12,rdx + + dec r14d + jnz NEAR $L$OOP_REDUCE_1024 + lea r12,[448+rsp] + vpaddq ymm0,ymm13,ymm9 + vpxor ymm9,ymm9,ymm9 + + vpaddq ymm0,ymm0,YMMWORD[((288-192))+rbx] + vpaddq ymm1,ymm1,YMMWORD[((320-448))+r12] + vpaddq ymm2,ymm2,YMMWORD[((352-448))+r12] + vpaddq ymm3,ymm3,YMMWORD[((384-448))+r12] + vpaddq ymm4,ymm4,YMMWORD[((416-448))+r12] + vpaddq ymm5,ymm5,YMMWORD[((448-448))+r12] + vpaddq ymm6,ymm6,YMMWORD[((480-448))+r12] + vpaddq ymm7,ymm7,YMMWORD[((512-448))+r12] + vpaddq ymm8,ymm8,YMMWORD[((544-448))+r12] + + vpsrlq ymm14,ymm0,29 + vpand ymm0,ymm0,ymm15 + vpsrlq ymm11,ymm1,29 + vpand ymm1,ymm1,ymm15 + vpsrlq ymm12,ymm2,29 + vpermq ymm14,ymm14,0x93 + vpand ymm2,ymm2,ymm15 + vpsrlq ymm13,ymm3,29 + vpermq ymm11,ymm11,0x93 + vpand ymm3,ymm3,ymm15 + vpermq ymm12,ymm12,0x93 + + vpblendd ymm10,ymm14,ymm9,3 + vpermq ymm13,ymm13,0x93 + vpblendd ymm14,ymm11,ymm14,3 + vpaddq ymm0,ymm0,ymm10 + vpblendd ymm11,ymm12,ymm11,3 + vpaddq ymm1,ymm1,ymm14 + vpblendd ymm12,ymm13,ymm12,3 + vpaddq ymm2,ymm2,ymm11 + vpblendd ymm13,ymm9,ymm13,3 + vpaddq ymm3,ymm3,ymm12 + vpaddq ymm4,ymm4,ymm13 + + vpsrlq ymm14,ymm0,29 + vpand ymm0,ymm0,ymm15 + vpsrlq ymm11,ymm1,29 + vpand ymm1,ymm1,ymm15 + vpsrlq ymm12,ymm2,29 + vpermq ymm14,ymm14,0x93 + vpand ymm2,ymm2,ymm15 + vpsrlq ymm13,ymm3,29 + vpermq ymm11,ymm11,0x93 + vpand ymm3,ymm3,ymm15 + vpermq ymm12,ymm12,0x93 + + vpblendd ymm10,ymm14,ymm9,3 + vpermq ymm13,ymm13,0x93 + vpblendd ymm14,ymm11,ymm14,3 + vpaddq ymm0,ymm0,ymm10 + vpblendd ymm11,ymm12,ymm11,3 + vpaddq ymm1,ymm1,ymm14 + vmovdqu YMMWORD[(0-128)+rdi],ymm0 + vpblendd ymm12,ymm13,ymm12,3 + vpaddq ymm2,ymm2,ymm11 + vmovdqu YMMWORD[(32-128)+rdi],ymm1 + vpblendd ymm13,ymm9,ymm13,3 + vpaddq ymm3,ymm3,ymm12 + vmovdqu YMMWORD[(64-128)+rdi],ymm2 + vpaddq ymm4,ymm4,ymm13 + vmovdqu YMMWORD[(96-128)+rdi],ymm3 + vpsrlq ymm14,ymm4,29 + vpand ymm4,ymm4,ymm15 + vpsrlq ymm11,ymm5,29 + vpand ymm5,ymm5,ymm15 + vpsrlq ymm12,ymm6,29 + vpermq ymm14,ymm14,0x93 + vpand ymm6,ymm6,ymm15 + vpsrlq ymm13,ymm7,29 + vpermq ymm11,ymm11,0x93 + vpand ymm7,ymm7,ymm15 + vpsrlq ymm0,ymm8,29 + vpermq ymm12,ymm12,0x93 + vpand ymm8,ymm8,ymm15 + vpermq ymm13,ymm13,0x93 + + vpblendd ymm10,ymm14,ymm9,3 + vpermq ymm0,ymm0,0x93 + vpblendd ymm14,ymm11,ymm14,3 + vpaddq ymm4,ymm4,ymm10 + vpblendd ymm11,ymm12,ymm11,3 + vpaddq ymm5,ymm5,ymm14 + vpblendd ymm12,ymm13,ymm12,3 + vpaddq ymm6,ymm6,ymm11 + vpblendd ymm13,ymm0,ymm13,3 + vpaddq ymm7,ymm7,ymm12 + vpaddq ymm8,ymm8,ymm13 + + vpsrlq ymm14,ymm4,29 + vpand ymm4,ymm4,ymm15 + vpsrlq ymm11,ymm5,29 + vpand ymm5,ymm5,ymm15 + vpsrlq ymm12,ymm6,29 + vpermq ymm14,ymm14,0x93 + vpand ymm6,ymm6,ymm15 + vpsrlq ymm13,ymm7,29 + vpermq ymm11,ymm11,0x93 + vpand ymm7,ymm7,ymm15 + vpsrlq ymm0,ymm8,29 + vpermq ymm12,ymm12,0x93 + vpand ymm8,ymm8,ymm15 + vpermq ymm13,ymm13,0x93 + + vpblendd ymm10,ymm14,ymm9,3 + vpermq ymm0,ymm0,0x93 + vpblendd ymm14,ymm11,ymm14,3 + vpaddq ymm4,ymm4,ymm10 + vpblendd ymm11,ymm12,ymm11,3 + vpaddq ymm5,ymm5,ymm14 + vmovdqu YMMWORD[(128-128)+rdi],ymm4 + vpblendd ymm12,ymm13,ymm12,3 + vpaddq ymm6,ymm6,ymm11 + vmovdqu YMMWORD[(160-128)+rdi],ymm5 + vpblendd ymm13,ymm0,ymm13,3 + vpaddq ymm7,ymm7,ymm12 + vmovdqu YMMWORD[(192-128)+rdi],ymm6 + vpaddq ymm8,ymm8,ymm13 + vmovdqu YMMWORD[(224-128)+rdi],ymm7 + vmovdqu YMMWORD[(256-128)+rdi],ymm8 + + mov rsi,rdi + dec r8d + jne NEAR $L$OOP_GRANDE_SQR_1024 + + vzeroall + mov rax,rbp + +$L$sqr_1024_in_tail: + movaps xmm6,XMMWORD[((-216))+rax] + movaps xmm7,XMMWORD[((-200))+rax] + movaps xmm8,XMMWORD[((-184))+rax] + movaps xmm9,XMMWORD[((-168))+rax] + movaps xmm10,XMMWORD[((-152))+rax] + movaps xmm11,XMMWORD[((-136))+rax] + movaps xmm12,XMMWORD[((-120))+rax] + movaps xmm13,XMMWORD[((-104))+rax] + movaps xmm14,XMMWORD[((-88))+rax] + movaps xmm15,XMMWORD[((-72))+rax] + mov r15,QWORD[((-48))+rax] + + mov r14,QWORD[((-40))+rax] + + mov r13,QWORD[((-32))+rax] + + mov r12,QWORD[((-24))+rax] + + mov rbp,QWORD[((-16))+rax] + + mov rbx,QWORD[((-8))+rax] + + lea rsp,[rax] + +$L$sqr_1024_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_rsaz_1024_sqr_avx2: +global rsaz_1024_mul_avx2 + +ALIGN 64 +rsaz_1024_mul_avx2: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[16+rsp],rsi + mov rax,rsp +$L$SEH_begin_rsaz_1024_mul_avx2: + mov rdi,rcx + mov rsi,rdx + mov rdx,r8 + mov rcx,r9 + mov r8,QWORD[40+rsp] + + + + lea rax,[rsp] + + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + + vzeroupper + lea rsp,[((-168))+rsp] + vmovaps XMMWORD[(-216)+rax],xmm6 + vmovaps XMMWORD[(-200)+rax],xmm7 + vmovaps XMMWORD[(-184)+rax],xmm8 + vmovaps XMMWORD[(-168)+rax],xmm9 + vmovaps XMMWORD[(-152)+rax],xmm10 + vmovaps XMMWORD[(-136)+rax],xmm11 + vmovaps XMMWORD[(-120)+rax],xmm12 + vmovaps XMMWORD[(-104)+rax],xmm13 + vmovaps XMMWORD[(-88)+rax],xmm14 + vmovaps XMMWORD[(-72)+rax],xmm15 +$L$mul_1024_body: + mov rbp,rax + + vzeroall + mov r13,rdx + sub rsp,64 + + + + + + +DB 0x67,0x67 + mov r15,rsi + and r15,4095 + add r15,32*10 + shr r15,12 + mov r15,rsi + cmovnz rsi,r13 + cmovnz r13,r15 + + mov r15,rcx + sub rsi,-128 + sub rcx,-128 + sub rdi,-128 + + and r15,4095 + add r15,32*10 +DB 0x67,0x67 + shr r15,12 + jz NEAR $L$mul_1024_no_n_copy + + + + + + sub rsp,32*10 + vmovdqu ymm0,YMMWORD[((0-128))+rcx] + and rsp,-512 + vmovdqu ymm1,YMMWORD[((32-128))+rcx] + vmovdqu ymm2,YMMWORD[((64-128))+rcx] + vmovdqu ymm3,YMMWORD[((96-128))+rcx] + vmovdqu ymm4,YMMWORD[((128-128))+rcx] + vmovdqu ymm5,YMMWORD[((160-128))+rcx] + vmovdqu ymm6,YMMWORD[((192-128))+rcx] + vmovdqu ymm7,YMMWORD[((224-128))+rcx] + vmovdqu ymm8,YMMWORD[((256-128))+rcx] + lea rcx,[((64+128))+rsp] + vmovdqu YMMWORD[(0-128)+rcx],ymm0 + vpxor ymm0,ymm0,ymm0 + vmovdqu YMMWORD[(32-128)+rcx],ymm1 + vpxor ymm1,ymm1,ymm1 + vmovdqu YMMWORD[(64-128)+rcx],ymm2 + vpxor ymm2,ymm2,ymm2 + vmovdqu YMMWORD[(96-128)+rcx],ymm3 + vpxor ymm3,ymm3,ymm3 + vmovdqu YMMWORD[(128-128)+rcx],ymm4 + vpxor ymm4,ymm4,ymm4 + vmovdqu YMMWORD[(160-128)+rcx],ymm5 + vpxor ymm5,ymm5,ymm5 + vmovdqu YMMWORD[(192-128)+rcx],ymm6 + vpxor ymm6,ymm6,ymm6 + vmovdqu YMMWORD[(224-128)+rcx],ymm7 + vpxor ymm7,ymm7,ymm7 + vmovdqu YMMWORD[(256-128)+rcx],ymm8 + vmovdqa ymm8,ymm0 + vmovdqu YMMWORD[(288-128)+rcx],ymm9 +$L$mul_1024_no_n_copy: + and rsp,-64 + + mov rbx,QWORD[r13] + vpbroadcastq ymm10,QWORD[r13] + vmovdqu YMMWORD[rsp],ymm0 + xor r9,r9 +DB 0x67 + xor r10,r10 + xor r11,r11 + xor r12,r12 + + vmovdqu ymm15,YMMWORD[$L$and_mask] + mov r14d,9 + vmovdqu YMMWORD[(288-128)+rdi],ymm9 + jmp NEAR $L$oop_mul_1024 + +ALIGN 32 +$L$oop_mul_1024: + vpsrlq ymm9,ymm3,29 + mov rax,rbx + imul rax,QWORD[((-128))+rsi] + add rax,r9 + mov r10,rbx + imul r10,QWORD[((8-128))+rsi] + add r10,QWORD[8+rsp] + + mov r9,rax + imul eax,r8d + and eax,0x1fffffff + + mov r11,rbx + imul r11,QWORD[((16-128))+rsi] + add r11,QWORD[16+rsp] + + mov r12,rbx + imul r12,QWORD[((24-128))+rsi] + add r12,QWORD[24+rsp] + vpmuludq ymm0,ymm10,YMMWORD[((32-128))+rsi] + vmovd xmm11,eax + vpaddq ymm1,ymm1,ymm0 + vpmuludq ymm12,ymm10,YMMWORD[((64-128))+rsi] + vpbroadcastq ymm11,xmm11 + vpaddq ymm2,ymm2,ymm12 + vpmuludq ymm13,ymm10,YMMWORD[((96-128))+rsi] + vpand ymm3,ymm3,ymm15 + vpaddq ymm3,ymm3,ymm13 + vpmuludq ymm0,ymm10,YMMWORD[((128-128))+rsi] + vpaddq ymm4,ymm4,ymm0 + vpmuludq ymm12,ymm10,YMMWORD[((160-128))+rsi] + vpaddq ymm5,ymm5,ymm12 + vpmuludq ymm13,ymm10,YMMWORD[((192-128))+rsi] + vpaddq ymm6,ymm6,ymm13 + vpmuludq ymm0,ymm10,YMMWORD[((224-128))+rsi] + vpermq ymm9,ymm9,0x93 + vpaddq ymm7,ymm7,ymm0 + vpmuludq ymm12,ymm10,YMMWORD[((256-128))+rsi] + vpbroadcastq ymm10,QWORD[8+r13] + vpaddq ymm8,ymm8,ymm12 + + mov rdx,rax + imul rax,QWORD[((-128))+rcx] + add r9,rax + mov rax,rdx + imul rax,QWORD[((8-128))+rcx] + add r10,rax + mov rax,rdx + imul rax,QWORD[((16-128))+rcx] + add r11,rax + shr r9,29 + imul rdx,QWORD[((24-128))+rcx] + add r12,rdx + add r10,r9 + + vpmuludq ymm13,ymm11,YMMWORD[((32-128))+rcx] + vmovq rbx,xmm10 + vpaddq ymm1,ymm1,ymm13 + vpmuludq ymm0,ymm11,YMMWORD[((64-128))+rcx] + vpaddq ymm2,ymm2,ymm0 + vpmuludq ymm12,ymm11,YMMWORD[((96-128))+rcx] + vpaddq ymm3,ymm3,ymm12 + vpmuludq ymm13,ymm11,YMMWORD[((128-128))+rcx] + vpaddq ymm4,ymm4,ymm13 + vpmuludq ymm0,ymm11,YMMWORD[((160-128))+rcx] + vpaddq ymm5,ymm5,ymm0 + vpmuludq ymm12,ymm11,YMMWORD[((192-128))+rcx] + vpaddq ymm6,ymm6,ymm12 + vpmuludq ymm13,ymm11,YMMWORD[((224-128))+rcx] + vpblendd ymm12,ymm9,ymm14,3 + vpaddq ymm7,ymm7,ymm13 + vpmuludq ymm0,ymm11,YMMWORD[((256-128))+rcx] + vpaddq ymm3,ymm3,ymm12 + vpaddq ymm8,ymm8,ymm0 + + mov rax,rbx + imul rax,QWORD[((-128))+rsi] + add r10,rax + vmovdqu ymm12,YMMWORD[((-8+32-128))+rsi] + mov rax,rbx + imul rax,QWORD[((8-128))+rsi] + add r11,rax + vmovdqu ymm13,YMMWORD[((-8+64-128))+rsi] + + mov rax,r10 + vpblendd ymm9,ymm9,ymm14,0xfc + imul eax,r8d + vpaddq ymm4,ymm4,ymm9 + and eax,0x1fffffff + + imul rbx,QWORD[((16-128))+rsi] + add r12,rbx + vpmuludq ymm12,ymm12,ymm10 + vmovd xmm11,eax + vmovdqu ymm0,YMMWORD[((-8+96-128))+rsi] + vpaddq ymm1,ymm1,ymm12 + vpmuludq ymm13,ymm13,ymm10 + vpbroadcastq ymm11,xmm11 + vmovdqu ymm12,YMMWORD[((-8+128-128))+rsi] + vpaddq ymm2,ymm2,ymm13 + vpmuludq ymm0,ymm0,ymm10 + vmovdqu ymm13,YMMWORD[((-8+160-128))+rsi] + vpaddq ymm3,ymm3,ymm0 + vpmuludq ymm12,ymm12,ymm10 + vmovdqu ymm0,YMMWORD[((-8+192-128))+rsi] + vpaddq ymm4,ymm4,ymm12 + vpmuludq ymm13,ymm13,ymm10 + vmovdqu ymm12,YMMWORD[((-8+224-128))+rsi] + vpaddq ymm5,ymm5,ymm13 + vpmuludq ymm0,ymm0,ymm10 + vmovdqu ymm13,YMMWORD[((-8+256-128))+rsi] + vpaddq ymm6,ymm6,ymm0 + vpmuludq ymm12,ymm12,ymm10 + vmovdqu ymm9,YMMWORD[((-8+288-128))+rsi] + vpaddq ymm7,ymm7,ymm12 + vpmuludq ymm13,ymm13,ymm10 + vpaddq ymm8,ymm8,ymm13 + vpmuludq ymm9,ymm9,ymm10 + vpbroadcastq ymm10,QWORD[16+r13] + + mov rdx,rax + imul rax,QWORD[((-128))+rcx] + add r10,rax + vmovdqu ymm0,YMMWORD[((-8+32-128))+rcx] + mov rax,rdx + imul rax,QWORD[((8-128))+rcx] + add r11,rax + vmovdqu ymm12,YMMWORD[((-8+64-128))+rcx] + shr r10,29 + imul rdx,QWORD[((16-128))+rcx] + add r12,rdx + add r11,r10 + + vpmuludq ymm0,ymm0,ymm11 + vmovq rbx,xmm10 + vmovdqu ymm13,YMMWORD[((-8+96-128))+rcx] + vpaddq ymm1,ymm1,ymm0 + vpmuludq ymm12,ymm12,ymm11 + vmovdqu ymm0,YMMWORD[((-8+128-128))+rcx] + vpaddq ymm2,ymm2,ymm12 + vpmuludq ymm13,ymm13,ymm11 + vmovdqu ymm12,YMMWORD[((-8+160-128))+rcx] + vpaddq ymm3,ymm3,ymm13 + vpmuludq ymm0,ymm0,ymm11 + vmovdqu ymm13,YMMWORD[((-8+192-128))+rcx] + vpaddq ymm4,ymm4,ymm0 + vpmuludq ymm12,ymm12,ymm11 + vmovdqu ymm0,YMMWORD[((-8+224-128))+rcx] + vpaddq ymm5,ymm5,ymm12 + vpmuludq ymm13,ymm13,ymm11 + vmovdqu ymm12,YMMWORD[((-8+256-128))+rcx] + vpaddq ymm6,ymm6,ymm13 + vpmuludq ymm0,ymm0,ymm11 + vmovdqu ymm13,YMMWORD[((-8+288-128))+rcx] + vpaddq ymm7,ymm7,ymm0 + vpmuludq ymm12,ymm12,ymm11 + vpaddq ymm8,ymm8,ymm12 + vpmuludq ymm13,ymm13,ymm11 + vpaddq ymm9,ymm9,ymm13 + + vmovdqu ymm0,YMMWORD[((-16+32-128))+rsi] + mov rax,rbx + imul rax,QWORD[((-128))+rsi] + add rax,r11 + + vmovdqu ymm12,YMMWORD[((-16+64-128))+rsi] + mov r11,rax + imul eax,r8d + and eax,0x1fffffff + + imul rbx,QWORD[((8-128))+rsi] + add r12,rbx + vpmuludq ymm0,ymm0,ymm10 + vmovd xmm11,eax + vmovdqu ymm13,YMMWORD[((-16+96-128))+rsi] + vpaddq ymm1,ymm1,ymm0 + vpmuludq ymm12,ymm12,ymm10 + vpbroadcastq ymm11,xmm11 + vmovdqu ymm0,YMMWORD[((-16+128-128))+rsi] + vpaddq ymm2,ymm2,ymm12 + vpmuludq ymm13,ymm13,ymm10 + vmovdqu ymm12,YMMWORD[((-16+160-128))+rsi] + vpaddq ymm3,ymm3,ymm13 + vpmuludq ymm0,ymm0,ymm10 + vmovdqu ymm13,YMMWORD[((-16+192-128))+rsi] + vpaddq ymm4,ymm4,ymm0 + vpmuludq ymm12,ymm12,ymm10 + vmovdqu ymm0,YMMWORD[((-16+224-128))+rsi] + vpaddq ymm5,ymm5,ymm12 + vpmuludq ymm13,ymm13,ymm10 + vmovdqu ymm12,YMMWORD[((-16+256-128))+rsi] + vpaddq ymm6,ymm6,ymm13 + vpmuludq ymm0,ymm0,ymm10 + vmovdqu ymm13,YMMWORD[((-16+288-128))+rsi] + vpaddq ymm7,ymm7,ymm0 + vpmuludq ymm12,ymm12,ymm10 + vpaddq ymm8,ymm8,ymm12 + vpmuludq ymm13,ymm13,ymm10 + vpbroadcastq ymm10,QWORD[24+r13] + vpaddq ymm9,ymm9,ymm13 + + vmovdqu ymm0,YMMWORD[((-16+32-128))+rcx] + mov rdx,rax + imul rax,QWORD[((-128))+rcx] + add r11,rax + vmovdqu ymm12,YMMWORD[((-16+64-128))+rcx] + imul rdx,QWORD[((8-128))+rcx] + add r12,rdx + shr r11,29 + + vpmuludq ymm0,ymm0,ymm11 + vmovq rbx,xmm10 + vmovdqu ymm13,YMMWORD[((-16+96-128))+rcx] + vpaddq ymm1,ymm1,ymm0 + vpmuludq ymm12,ymm12,ymm11 + vmovdqu ymm0,YMMWORD[((-16+128-128))+rcx] + vpaddq ymm2,ymm2,ymm12 + vpmuludq ymm13,ymm13,ymm11 + vmovdqu ymm12,YMMWORD[((-16+160-128))+rcx] + vpaddq ymm3,ymm3,ymm13 + vpmuludq ymm0,ymm0,ymm11 + vmovdqu ymm13,YMMWORD[((-16+192-128))+rcx] + vpaddq ymm4,ymm4,ymm0 + vpmuludq ymm12,ymm12,ymm11 + vmovdqu ymm0,YMMWORD[((-16+224-128))+rcx] + vpaddq ymm5,ymm5,ymm12 + vpmuludq ymm13,ymm13,ymm11 + vmovdqu ymm12,YMMWORD[((-16+256-128))+rcx] + vpaddq ymm6,ymm6,ymm13 + vpmuludq ymm0,ymm0,ymm11 + vmovdqu ymm13,YMMWORD[((-16+288-128))+rcx] + vpaddq ymm7,ymm7,ymm0 + vpmuludq ymm12,ymm12,ymm11 + vmovdqu ymm0,YMMWORD[((-24+32-128))+rsi] + vpaddq ymm8,ymm8,ymm12 + vpmuludq ymm13,ymm13,ymm11 + vmovdqu ymm12,YMMWORD[((-24+64-128))+rsi] + vpaddq ymm9,ymm9,ymm13 + + add r12,r11 + imul rbx,QWORD[((-128))+rsi] + add r12,rbx + + mov rax,r12 + imul eax,r8d + and eax,0x1fffffff + + vpmuludq ymm0,ymm0,ymm10 + vmovd xmm11,eax + vmovdqu ymm13,YMMWORD[((-24+96-128))+rsi] + vpaddq ymm1,ymm1,ymm0 + vpmuludq ymm12,ymm12,ymm10 + vpbroadcastq ymm11,xmm11 + vmovdqu ymm0,YMMWORD[((-24+128-128))+rsi] + vpaddq ymm2,ymm2,ymm12 + vpmuludq ymm13,ymm13,ymm10 + vmovdqu ymm12,YMMWORD[((-24+160-128))+rsi] + vpaddq ymm3,ymm3,ymm13 + vpmuludq ymm0,ymm0,ymm10 + vmovdqu ymm13,YMMWORD[((-24+192-128))+rsi] + vpaddq ymm4,ymm4,ymm0 + vpmuludq ymm12,ymm12,ymm10 + vmovdqu ymm0,YMMWORD[((-24+224-128))+rsi] + vpaddq ymm5,ymm5,ymm12 + vpmuludq ymm13,ymm13,ymm10 + vmovdqu ymm12,YMMWORD[((-24+256-128))+rsi] + vpaddq ymm6,ymm6,ymm13 + vpmuludq ymm0,ymm0,ymm10 + vmovdqu ymm13,YMMWORD[((-24+288-128))+rsi] + vpaddq ymm7,ymm7,ymm0 + vpmuludq ymm12,ymm12,ymm10 + vpaddq ymm8,ymm8,ymm12 + vpmuludq ymm13,ymm13,ymm10 + vpbroadcastq ymm10,QWORD[32+r13] + vpaddq ymm9,ymm9,ymm13 + add r13,32 + + vmovdqu ymm0,YMMWORD[((-24+32-128))+rcx] + imul rax,QWORD[((-128))+rcx] + add r12,rax + shr r12,29 + + vmovdqu ymm12,YMMWORD[((-24+64-128))+rcx] + vpmuludq ymm0,ymm0,ymm11 + vmovq rbx,xmm10 + vmovdqu ymm13,YMMWORD[((-24+96-128))+rcx] + vpaddq ymm0,ymm1,ymm0 + vpmuludq ymm12,ymm12,ymm11 + vmovdqu YMMWORD[rsp],ymm0 + vpaddq ymm1,ymm2,ymm12 + vmovdqu ymm0,YMMWORD[((-24+128-128))+rcx] + vpmuludq ymm13,ymm13,ymm11 + vmovdqu ymm12,YMMWORD[((-24+160-128))+rcx] + vpaddq ymm2,ymm3,ymm13 + vpmuludq ymm0,ymm0,ymm11 + vmovdqu ymm13,YMMWORD[((-24+192-128))+rcx] + vpaddq ymm3,ymm4,ymm0 + vpmuludq ymm12,ymm12,ymm11 + vmovdqu ymm0,YMMWORD[((-24+224-128))+rcx] + vpaddq ymm4,ymm5,ymm12 + vpmuludq ymm13,ymm13,ymm11 + vmovdqu ymm12,YMMWORD[((-24+256-128))+rcx] + vpaddq ymm5,ymm6,ymm13 + vpmuludq ymm0,ymm0,ymm11 + vmovdqu ymm13,YMMWORD[((-24+288-128))+rcx] + mov r9,r12 + vpaddq ymm6,ymm7,ymm0 + vpmuludq ymm12,ymm12,ymm11 + add r9,QWORD[rsp] + vpaddq ymm7,ymm8,ymm12 + vpmuludq ymm13,ymm13,ymm11 + vmovq xmm12,r12 + vpaddq ymm8,ymm9,ymm13 + + dec r14d + jnz NEAR $L$oop_mul_1024 + vpaddq ymm0,ymm12,YMMWORD[rsp] + + vpsrlq ymm12,ymm0,29 + vpand ymm0,ymm0,ymm15 + vpsrlq ymm13,ymm1,29 + vpand ymm1,ymm1,ymm15 + vpsrlq ymm10,ymm2,29 + vpermq ymm12,ymm12,0x93 + vpand ymm2,ymm2,ymm15 + vpsrlq ymm11,ymm3,29 + vpermq ymm13,ymm13,0x93 + vpand ymm3,ymm3,ymm15 + + vpblendd ymm9,ymm12,ymm14,3 + vpermq ymm10,ymm10,0x93 + vpblendd ymm12,ymm13,ymm12,3 + vpermq ymm11,ymm11,0x93 + vpaddq ymm0,ymm0,ymm9 + vpblendd ymm13,ymm10,ymm13,3 + vpaddq ymm1,ymm1,ymm12 + vpblendd ymm10,ymm11,ymm10,3 + vpaddq ymm2,ymm2,ymm13 + vpblendd ymm11,ymm14,ymm11,3 + vpaddq ymm3,ymm3,ymm10 + vpaddq ymm4,ymm4,ymm11 + + vpsrlq ymm12,ymm0,29 + vpand ymm0,ymm0,ymm15 + vpsrlq ymm13,ymm1,29 + vpand ymm1,ymm1,ymm15 + vpsrlq ymm10,ymm2,29 + vpermq ymm12,ymm12,0x93 + vpand ymm2,ymm2,ymm15 + vpsrlq ymm11,ymm3,29 + vpermq ymm13,ymm13,0x93 + vpand ymm3,ymm3,ymm15 + vpermq ymm10,ymm10,0x93 + + vpblendd ymm9,ymm12,ymm14,3 + vpermq ymm11,ymm11,0x93 + vpblendd ymm12,ymm13,ymm12,3 + vpaddq ymm0,ymm0,ymm9 + vpblendd ymm13,ymm10,ymm13,3 + vpaddq ymm1,ymm1,ymm12 + vpblendd ymm10,ymm11,ymm10,3 + vpaddq ymm2,ymm2,ymm13 + vpblendd ymm11,ymm14,ymm11,3 + vpaddq ymm3,ymm3,ymm10 + vpaddq ymm4,ymm4,ymm11 + + vmovdqu YMMWORD[(0-128)+rdi],ymm0 + vmovdqu YMMWORD[(32-128)+rdi],ymm1 + vmovdqu YMMWORD[(64-128)+rdi],ymm2 + vmovdqu YMMWORD[(96-128)+rdi],ymm3 + vpsrlq ymm12,ymm4,29 + vpand ymm4,ymm4,ymm15 + vpsrlq ymm13,ymm5,29 + vpand ymm5,ymm5,ymm15 + vpsrlq ymm10,ymm6,29 + vpermq ymm12,ymm12,0x93 + vpand ymm6,ymm6,ymm15 + vpsrlq ymm11,ymm7,29 + vpermq ymm13,ymm13,0x93 + vpand ymm7,ymm7,ymm15 + vpsrlq ymm0,ymm8,29 + vpermq ymm10,ymm10,0x93 + vpand ymm8,ymm8,ymm15 + vpermq ymm11,ymm11,0x93 + + vpblendd ymm9,ymm12,ymm14,3 + vpermq ymm0,ymm0,0x93 + vpblendd ymm12,ymm13,ymm12,3 + vpaddq ymm4,ymm4,ymm9 + vpblendd ymm13,ymm10,ymm13,3 + vpaddq ymm5,ymm5,ymm12 + vpblendd ymm10,ymm11,ymm10,3 + vpaddq ymm6,ymm6,ymm13 + vpblendd ymm11,ymm0,ymm11,3 + vpaddq ymm7,ymm7,ymm10 + vpaddq ymm8,ymm8,ymm11 + + vpsrlq ymm12,ymm4,29 + vpand ymm4,ymm4,ymm15 + vpsrlq ymm13,ymm5,29 + vpand ymm5,ymm5,ymm15 + vpsrlq ymm10,ymm6,29 + vpermq ymm12,ymm12,0x93 + vpand ymm6,ymm6,ymm15 + vpsrlq ymm11,ymm7,29 + vpermq ymm13,ymm13,0x93 + vpand ymm7,ymm7,ymm15 + vpsrlq ymm0,ymm8,29 + vpermq ymm10,ymm10,0x93 + vpand ymm8,ymm8,ymm15 + vpermq ymm11,ymm11,0x93 + + vpblendd ymm9,ymm12,ymm14,3 + vpermq ymm0,ymm0,0x93 + vpblendd ymm12,ymm13,ymm12,3 + vpaddq ymm4,ymm4,ymm9 + vpblendd ymm13,ymm10,ymm13,3 + vpaddq ymm5,ymm5,ymm12 + vpblendd ymm10,ymm11,ymm10,3 + vpaddq ymm6,ymm6,ymm13 + vpblendd ymm11,ymm0,ymm11,3 + vpaddq ymm7,ymm7,ymm10 + vpaddq ymm8,ymm8,ymm11 + + vmovdqu YMMWORD[(128-128)+rdi],ymm4 + vmovdqu YMMWORD[(160-128)+rdi],ymm5 + vmovdqu YMMWORD[(192-128)+rdi],ymm6 + vmovdqu YMMWORD[(224-128)+rdi],ymm7 + vmovdqu YMMWORD[(256-128)+rdi],ymm8 + vzeroupper + + mov rax,rbp + +$L$mul_1024_in_tail: + movaps xmm6,XMMWORD[((-216))+rax] + movaps xmm7,XMMWORD[((-200))+rax] + movaps xmm8,XMMWORD[((-184))+rax] + movaps xmm9,XMMWORD[((-168))+rax] + movaps xmm10,XMMWORD[((-152))+rax] + movaps xmm11,XMMWORD[((-136))+rax] + movaps xmm12,XMMWORD[((-120))+rax] + movaps xmm13,XMMWORD[((-104))+rax] + movaps xmm14,XMMWORD[((-88))+rax] + movaps xmm15,XMMWORD[((-72))+rax] + mov r15,QWORD[((-48))+rax] + + mov r14,QWORD[((-40))+rax] + + mov r13,QWORD[((-32))+rax] + + mov r12,QWORD[((-24))+rax] + + mov rbp,QWORD[((-16))+rax] + + mov rbx,QWORD[((-8))+rax] + + lea rsp,[rax] + +$L$mul_1024_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_rsaz_1024_mul_avx2: +global rsaz_1024_red2norm_avx2 + +ALIGN 32 +rsaz_1024_red2norm_avx2: + + sub rdx,-128 + xor rax,rax + mov r8,QWORD[((-128))+rdx] + mov r9,QWORD[((-120))+rdx] + mov r10,QWORD[((-112))+rdx] + shl r8,0 + shl r9,29 + mov r11,r10 + shl r10,58 + shr r11,6 + add rax,r8 + add rax,r9 + add rax,r10 + adc r11,0 + mov QWORD[rcx],rax + mov rax,r11 + mov r8,QWORD[((-104))+rdx] + mov r9,QWORD[((-96))+rdx] + shl r8,23 + mov r10,r9 + shl r9,52 + shr r10,12 + add rax,r8 + add rax,r9 + adc r10,0 + mov QWORD[8+rcx],rax + mov rax,r10 + mov r11,QWORD[((-88))+rdx] + mov r8,QWORD[((-80))+rdx] + shl r11,17 + mov r9,r8 + shl r8,46 + shr r9,18 + add rax,r11 + add rax,r8 + adc r9,0 + mov QWORD[16+rcx],rax + mov rax,r9 + mov r10,QWORD[((-72))+rdx] + mov r11,QWORD[((-64))+rdx] + shl r10,11 + mov r8,r11 + shl r11,40 + shr r8,24 + add rax,r10 + add rax,r11 + adc r8,0 + mov QWORD[24+rcx],rax + mov rax,r8 + mov r9,QWORD[((-56))+rdx] + mov r10,QWORD[((-48))+rdx] + mov r11,QWORD[((-40))+rdx] + shl r9,5 + shl r10,34 + mov r8,r11 + shl r11,63 + shr r8,1 + add rax,r9 + add rax,r10 + add rax,r11 + adc r8,0 + mov QWORD[32+rcx],rax + mov rax,r8 + mov r9,QWORD[((-32))+rdx] + mov r10,QWORD[((-24))+rdx] + shl r9,28 + mov r11,r10 + shl r10,57 + shr r11,7 + add rax,r9 + add rax,r10 + adc r11,0 + mov QWORD[40+rcx],rax + mov rax,r11 + mov r8,QWORD[((-16))+rdx] + mov r9,QWORD[((-8))+rdx] + shl r8,22 + mov r10,r9 + shl r9,51 + shr r10,13 + add rax,r8 + add rax,r9 + adc r10,0 + mov QWORD[48+rcx],rax + mov rax,r10 + mov r11,QWORD[rdx] + mov r8,QWORD[8+rdx] + shl r11,16 + mov r9,r8 + shl r8,45 + shr r9,19 + add rax,r11 + add rax,r8 + adc r9,0 + mov QWORD[56+rcx],rax + mov rax,r9 + mov r10,QWORD[16+rdx] + mov r11,QWORD[24+rdx] + shl r10,10 + mov r8,r11 + shl r11,39 + shr r8,25 + add rax,r10 + add rax,r11 + adc r8,0 + mov QWORD[64+rcx],rax + mov rax,r8 + mov r9,QWORD[32+rdx] + mov r10,QWORD[40+rdx] + mov r11,QWORD[48+rdx] + shl r9,4 + shl r10,33 + mov r8,r11 + shl r11,62 + shr r8,2 + add rax,r9 + add rax,r10 + add rax,r11 + adc r8,0 + mov QWORD[72+rcx],rax + mov rax,r8 + mov r9,QWORD[56+rdx] + mov r10,QWORD[64+rdx] + shl r9,27 + mov r11,r10 + shl r10,56 + shr r11,8 + add rax,r9 + add rax,r10 + adc r11,0 + mov QWORD[80+rcx],rax + mov rax,r11 + mov r8,QWORD[72+rdx] + mov r9,QWORD[80+rdx] + shl r8,21 + mov r10,r9 + shl r9,50 + shr r10,14 + add rax,r8 + add rax,r9 + adc r10,0 + mov QWORD[88+rcx],rax + mov rax,r10 + mov r11,QWORD[88+rdx] + mov r8,QWORD[96+rdx] + shl r11,15 + mov r9,r8 + shl r8,44 + shr r9,20 + add rax,r11 + add rax,r8 + adc r9,0 + mov QWORD[96+rcx],rax + mov rax,r9 + mov r10,QWORD[104+rdx] + mov r11,QWORD[112+rdx] + shl r10,9 + mov r8,r11 + shl r11,38 + shr r8,26 + add rax,r10 + add rax,r11 + adc r8,0 + mov QWORD[104+rcx],rax + mov rax,r8 + mov r9,QWORD[120+rdx] + mov r10,QWORD[128+rdx] + mov r11,QWORD[136+rdx] + shl r9,3 + shl r10,32 + mov r8,r11 + shl r11,61 + shr r8,3 + add rax,r9 + add rax,r10 + add rax,r11 + adc r8,0 + mov QWORD[112+rcx],rax + mov rax,r8 + mov r9,QWORD[144+rdx] + mov r10,QWORD[152+rdx] + shl r9,26 + mov r11,r10 + shl r10,55 + shr r11,9 + add rax,r9 + add rax,r10 + adc r11,0 + mov QWORD[120+rcx],rax + mov rax,r11 + DB 0F3h,0C3h ;repret + + + +global rsaz_1024_norm2red_avx2 + +ALIGN 32 +rsaz_1024_norm2red_avx2: + + sub rcx,-128 + mov r8,QWORD[rdx] + mov eax,0x1fffffff + mov r9,QWORD[8+rdx] + mov r11,r8 + shr r11,0 + and r11,rax + mov QWORD[((-128))+rcx],r11 + mov r10,r8 + shr r10,29 + and r10,rax + mov QWORD[((-120))+rcx],r10 + shrd r8,r9,58 + and r8,rax + mov QWORD[((-112))+rcx],r8 + mov r10,QWORD[16+rdx] + mov r8,r9 + shr r8,23 + and r8,rax + mov QWORD[((-104))+rcx],r8 + shrd r9,r10,52 + and r9,rax + mov QWORD[((-96))+rcx],r9 + mov r11,QWORD[24+rdx] + mov r9,r10 + shr r9,17 + and r9,rax + mov QWORD[((-88))+rcx],r9 + shrd r10,r11,46 + and r10,rax + mov QWORD[((-80))+rcx],r10 + mov r8,QWORD[32+rdx] + mov r10,r11 + shr r10,11 + and r10,rax + mov QWORD[((-72))+rcx],r10 + shrd r11,r8,40 + and r11,rax + mov QWORD[((-64))+rcx],r11 + mov r9,QWORD[40+rdx] + mov r11,r8 + shr r11,5 + and r11,rax + mov QWORD[((-56))+rcx],r11 + mov r10,r8 + shr r10,34 + and r10,rax + mov QWORD[((-48))+rcx],r10 + shrd r8,r9,63 + and r8,rax + mov QWORD[((-40))+rcx],r8 + mov r10,QWORD[48+rdx] + mov r8,r9 + shr r8,28 + and r8,rax + mov QWORD[((-32))+rcx],r8 + shrd r9,r10,57 + and r9,rax + mov QWORD[((-24))+rcx],r9 + mov r11,QWORD[56+rdx] + mov r9,r10 + shr r9,22 + and r9,rax + mov QWORD[((-16))+rcx],r9 + shrd r10,r11,51 + and r10,rax + mov QWORD[((-8))+rcx],r10 + mov r8,QWORD[64+rdx] + mov r10,r11 + shr r10,16 + and r10,rax + mov QWORD[rcx],r10 + shrd r11,r8,45 + and r11,rax + mov QWORD[8+rcx],r11 + mov r9,QWORD[72+rdx] + mov r11,r8 + shr r11,10 + and r11,rax + mov QWORD[16+rcx],r11 + shrd r8,r9,39 + and r8,rax + mov QWORD[24+rcx],r8 + mov r10,QWORD[80+rdx] + mov r8,r9 + shr r8,4 + and r8,rax + mov QWORD[32+rcx],r8 + mov r11,r9 + shr r11,33 + and r11,rax + mov QWORD[40+rcx],r11 + shrd r9,r10,62 + and r9,rax + mov QWORD[48+rcx],r9 + mov r11,QWORD[88+rdx] + mov r9,r10 + shr r9,27 + and r9,rax + mov QWORD[56+rcx],r9 + shrd r10,r11,56 + and r10,rax + mov QWORD[64+rcx],r10 + mov r8,QWORD[96+rdx] + mov r10,r11 + shr r10,21 + and r10,rax + mov QWORD[72+rcx],r10 + shrd r11,r8,50 + and r11,rax + mov QWORD[80+rcx],r11 + mov r9,QWORD[104+rdx] + mov r11,r8 + shr r11,15 + and r11,rax + mov QWORD[88+rcx],r11 + shrd r8,r9,44 + and r8,rax + mov QWORD[96+rcx],r8 + mov r10,QWORD[112+rdx] + mov r8,r9 + shr r8,9 + and r8,rax + mov QWORD[104+rcx],r8 + shrd r9,r10,38 + and r9,rax + mov QWORD[112+rcx],r9 + mov r11,QWORD[120+rdx] + mov r9,r10 + shr r9,3 + and r9,rax + mov QWORD[120+rcx],r9 + mov r8,r10 + shr r8,32 + and r8,rax + mov QWORD[128+rcx],r8 + shrd r10,r11,61 + and r10,rax + mov QWORD[136+rcx],r10 + xor r8,r8 + mov r10,r11 + shr r10,26 + and r10,rax + mov QWORD[144+rcx],r10 + shrd r11,r8,55 + and r11,rax + mov QWORD[152+rcx],r11 + mov QWORD[160+rcx],r8 + mov QWORD[168+rcx],r8 + mov QWORD[176+rcx],r8 + mov QWORD[184+rcx],r8 + DB 0F3h,0C3h ;repret + + +global rsaz_1024_scatter5_avx2 + +ALIGN 32 +rsaz_1024_scatter5_avx2: + + vzeroupper + vmovdqu ymm5,YMMWORD[$L$scatter_permd] + shl r8d,4 + lea rcx,[r8*1+rcx] + mov eax,9 + jmp NEAR $L$oop_scatter_1024 + +ALIGN 32 +$L$oop_scatter_1024: + vmovdqu ymm0,YMMWORD[rdx] + lea rdx,[32+rdx] + vpermd ymm0,ymm5,ymm0 + vmovdqu XMMWORD[rcx],xmm0 + lea rcx,[512+rcx] + dec eax + jnz NEAR $L$oop_scatter_1024 + + vzeroupper + DB 0F3h,0C3h ;repret + + + +global rsaz_1024_gather5_avx2 + +ALIGN 32 +rsaz_1024_gather5_avx2: + + vzeroupper + mov r11,rsp + + lea rax,[((-136))+rsp] +$L$SEH_begin_rsaz_1024_gather5: + +DB 0x48,0x8d,0x60,0xe0 +DB 0xc5,0xf8,0x29,0x70,0xe0 +DB 0xc5,0xf8,0x29,0x78,0xf0 +DB 0xc5,0x78,0x29,0x40,0x00 +DB 0xc5,0x78,0x29,0x48,0x10 +DB 0xc5,0x78,0x29,0x50,0x20 +DB 0xc5,0x78,0x29,0x58,0x30 +DB 0xc5,0x78,0x29,0x60,0x40 +DB 0xc5,0x78,0x29,0x68,0x50 +DB 0xc5,0x78,0x29,0x70,0x60 +DB 0xc5,0x78,0x29,0x78,0x70 + lea rsp,[((-256))+rsp] + and rsp,-32 + lea r10,[$L$inc] + lea rax,[((-128))+rsp] + + vmovd xmm4,r8d + vmovdqa ymm0,YMMWORD[r10] + vmovdqa ymm1,YMMWORD[32+r10] + vmovdqa ymm5,YMMWORD[64+r10] + vpbroadcastd ymm4,xmm4 + + vpaddd ymm2,ymm0,ymm5 + vpcmpeqd ymm0,ymm0,ymm4 + vpaddd ymm3,ymm1,ymm5 + vpcmpeqd ymm1,ymm1,ymm4 + vmovdqa YMMWORD[(0+128)+rax],ymm0 + vpaddd ymm0,ymm2,ymm5 + vpcmpeqd ymm2,ymm2,ymm4 + vmovdqa YMMWORD[(32+128)+rax],ymm1 + vpaddd ymm1,ymm3,ymm5 + vpcmpeqd ymm3,ymm3,ymm4 + vmovdqa YMMWORD[(64+128)+rax],ymm2 + vpaddd ymm2,ymm0,ymm5 + vpcmpeqd ymm0,ymm0,ymm4 + vmovdqa YMMWORD[(96+128)+rax],ymm3 + vpaddd ymm3,ymm1,ymm5 + vpcmpeqd ymm1,ymm1,ymm4 + vmovdqa YMMWORD[(128+128)+rax],ymm0 + vpaddd ymm8,ymm2,ymm5 + vpcmpeqd ymm2,ymm2,ymm4 + vmovdqa YMMWORD[(160+128)+rax],ymm1 + vpaddd ymm9,ymm3,ymm5 + vpcmpeqd ymm3,ymm3,ymm4 + vmovdqa YMMWORD[(192+128)+rax],ymm2 + vpaddd ymm10,ymm8,ymm5 + vpcmpeqd ymm8,ymm8,ymm4 + vmovdqa YMMWORD[(224+128)+rax],ymm3 + vpaddd ymm11,ymm9,ymm5 + vpcmpeqd ymm9,ymm9,ymm4 + vpaddd ymm12,ymm10,ymm5 + vpcmpeqd ymm10,ymm10,ymm4 + vpaddd ymm13,ymm11,ymm5 + vpcmpeqd ymm11,ymm11,ymm4 + vpaddd ymm14,ymm12,ymm5 + vpcmpeqd ymm12,ymm12,ymm4 + vpaddd ymm15,ymm13,ymm5 + vpcmpeqd ymm13,ymm13,ymm4 + vpcmpeqd ymm14,ymm14,ymm4 + vpcmpeqd ymm15,ymm15,ymm4 + + vmovdqa ymm7,YMMWORD[((-32))+r10] + lea rdx,[128+rdx] + mov r8d,9 + +$L$oop_gather_1024: + vmovdqa ymm0,YMMWORD[((0-128))+rdx] + vmovdqa ymm1,YMMWORD[((32-128))+rdx] + vmovdqa ymm2,YMMWORD[((64-128))+rdx] + vmovdqa ymm3,YMMWORD[((96-128))+rdx] + vpand ymm0,ymm0,YMMWORD[((0+128))+rax] + vpand ymm1,ymm1,YMMWORD[((32+128))+rax] + vpand ymm2,ymm2,YMMWORD[((64+128))+rax] + vpor ymm4,ymm1,ymm0 + vpand ymm3,ymm3,YMMWORD[((96+128))+rax] + vmovdqa ymm0,YMMWORD[((128-128))+rdx] + vmovdqa ymm1,YMMWORD[((160-128))+rdx] + vpor ymm5,ymm3,ymm2 + vmovdqa ymm2,YMMWORD[((192-128))+rdx] + vmovdqa ymm3,YMMWORD[((224-128))+rdx] + vpand ymm0,ymm0,YMMWORD[((128+128))+rax] + vpand ymm1,ymm1,YMMWORD[((160+128))+rax] + vpand ymm2,ymm2,YMMWORD[((192+128))+rax] + vpor ymm4,ymm4,ymm0 + vpand ymm3,ymm3,YMMWORD[((224+128))+rax] + vpand ymm0,ymm8,YMMWORD[((256-128))+rdx] + vpor ymm5,ymm5,ymm1 + vpand ymm1,ymm9,YMMWORD[((288-128))+rdx] + vpor ymm4,ymm4,ymm2 + vpand ymm2,ymm10,YMMWORD[((320-128))+rdx] + vpor ymm5,ymm5,ymm3 + vpand ymm3,ymm11,YMMWORD[((352-128))+rdx] + vpor ymm4,ymm4,ymm0 + vpand ymm0,ymm12,YMMWORD[((384-128))+rdx] + vpor ymm5,ymm5,ymm1 + vpand ymm1,ymm13,YMMWORD[((416-128))+rdx] + vpor ymm4,ymm4,ymm2 + vpand ymm2,ymm14,YMMWORD[((448-128))+rdx] + vpor ymm5,ymm5,ymm3 + vpand ymm3,ymm15,YMMWORD[((480-128))+rdx] + lea rdx,[512+rdx] + vpor ymm4,ymm4,ymm0 + vpor ymm5,ymm5,ymm1 + vpor ymm4,ymm4,ymm2 + vpor ymm5,ymm5,ymm3 + + vpor ymm4,ymm4,ymm5 + vextracti128 xmm5,ymm4,1 + vpor xmm5,xmm5,xmm4 + vpermd ymm5,ymm7,ymm5 + vmovdqu YMMWORD[rcx],ymm5 + lea rcx,[32+rcx] + dec r8d + jnz NEAR $L$oop_gather_1024 + + vpxor ymm0,ymm0,ymm0 + vmovdqu YMMWORD[rcx],ymm0 + vzeroupper + movaps xmm6,XMMWORD[((-168))+r11] + movaps xmm7,XMMWORD[((-152))+r11] + movaps xmm8,XMMWORD[((-136))+r11] + movaps xmm9,XMMWORD[((-120))+r11] + movaps xmm10,XMMWORD[((-104))+r11] + movaps xmm11,XMMWORD[((-88))+r11] + movaps xmm12,XMMWORD[((-72))+r11] + movaps xmm13,XMMWORD[((-56))+r11] + movaps xmm14,XMMWORD[((-40))+r11] + movaps xmm15,XMMWORD[((-24))+r11] + lea rsp,[r11] + + DB 0F3h,0C3h ;repret + +$L$SEH_end_rsaz_1024_gather5: + +EXTERN OPENSSL_ia32cap_P +global rsaz_avx2_eligible + +ALIGN 32 +rsaz_avx2_eligible: + mov eax,DWORD[((OPENSSL_ia32cap_P+8))] + mov ecx,524544 + mov edx,0 + and ecx,eax + cmp ecx,524544 + cmove eax,edx + and eax,32 + shr eax,5 + DB 0F3h,0C3h ;repret + + +ALIGN 64 +$L$and_mask: + DQ 0x1fffffff,0x1fffffff,0x1fffffff,0x1fffffff +$L$scatter_permd: + DD 0,2,4,6,7,7,7,7 +$L$gather_permd: + DD 0,7,1,7,2,7,3,7 +$L$inc: + DD 0,0,0,0,1,1,1,1 + DD 2,2,2,2,3,3,3,3 + DD 4,4,4,4,4,4,4,4 +ALIGN 64 +EXTERN __imp_RtlVirtualUnwind + +ALIGN 16 +rsaz_se_handler: + push rsi + push rdi + push rbx + push rbp + push r12 + push r13 + push r14 + push r15 + pushfq + sub rsp,64 + + mov rax,QWORD[120+r8] + mov rbx,QWORD[248+r8] + + mov rsi,QWORD[8+r9] + mov r11,QWORD[56+r9] + + mov r10d,DWORD[r11] + lea r10,[r10*1+rsi] + cmp rbx,r10 + jb NEAR $L$common_seh_tail + + mov r10d,DWORD[4+r11] + lea r10,[r10*1+rsi] + cmp rbx,r10 + jae NEAR $L$common_seh_tail + + mov rbp,QWORD[160+r8] + + mov r10d,DWORD[8+r11] + lea r10,[r10*1+rsi] + cmp rbx,r10 + cmovc rax,rbp + + mov r15,QWORD[((-48))+rax] + mov r14,QWORD[((-40))+rax] + mov r13,QWORD[((-32))+rax] + mov r12,QWORD[((-24))+rax] + mov rbp,QWORD[((-16))+rax] + mov rbx,QWORD[((-8))+rax] + mov QWORD[240+r8],r15 + mov QWORD[232+r8],r14 + mov QWORD[224+r8],r13 + mov QWORD[216+r8],r12 + mov QWORD[160+r8],rbp + mov QWORD[144+r8],rbx + + lea rsi,[((-216))+rax] + lea rdi,[512+r8] + mov ecx,20 + DD 0xa548f3fc + +$L$common_seh_tail: + mov rdi,QWORD[8+rax] + mov rsi,QWORD[16+rax] + mov QWORD[152+r8],rax + mov QWORD[168+r8],rsi + mov QWORD[176+r8],rdi + + mov rdi,QWORD[40+r9] + mov rsi,r8 + mov ecx,154 + DD 0xa548f3fc + + mov rsi,r9 + xor rcx,rcx + mov rdx,QWORD[8+rsi] + mov r8,QWORD[rsi] + mov r9,QWORD[16+rsi] + mov r10,QWORD[40+rsi] + lea r11,[56+rsi] + lea r12,[24+rsi] + mov QWORD[32+rsp],r10 + mov QWORD[40+rsp],r11 + mov QWORD[48+rsp],r12 + mov QWORD[56+rsp],rcx + call QWORD[__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 + + +section .pdata rdata align=4 +ALIGN 4 + DD $L$SEH_begin_rsaz_1024_sqr_avx2 wrt ..imagebase + DD $L$SEH_end_rsaz_1024_sqr_avx2 wrt ..imagebase + DD $L$SEH_info_rsaz_1024_sqr_avx2 wrt ..imagebase + + DD $L$SEH_begin_rsaz_1024_mul_avx2 wrt ..imagebase + DD $L$SEH_end_rsaz_1024_mul_avx2 wrt ..imagebase + DD $L$SEH_info_rsaz_1024_mul_avx2 wrt ..imagebase + + DD $L$SEH_begin_rsaz_1024_gather5 wrt ..imagebase + DD $L$SEH_end_rsaz_1024_gather5 wrt ..imagebase + DD $L$SEH_info_rsaz_1024_gather5 wrt ..imagebase +section .xdata rdata align=8 +ALIGN 8 +$L$SEH_info_rsaz_1024_sqr_avx2: +DB 9,0,0,0 + DD rsaz_se_handler wrt ..imagebase + DD $L$sqr_1024_body wrt ..imagebase,$L$sqr_1024_epilogue wrt ..imagebase,$L$sqr_1024_in_tail wrt ..imagebase + DD 0 +$L$SEH_info_rsaz_1024_mul_avx2: +DB 9,0,0,0 + DD rsaz_se_handler wrt ..imagebase + DD $L$mul_1024_body wrt ..imagebase,$L$mul_1024_epilogue wrt ..imagebase,$L$mul_1024_in_tail wrt ..imagebase + DD 0 +$L$SEH_info_rsaz_1024_gather5: +DB 0x01,0x36,0x17,0x0b +DB 0x36,0xf8,0x09,0x00 +DB 0x31,0xe8,0x08,0x00 +DB 0x2c,0xd8,0x07,0x00 +DB 0x27,0xc8,0x06,0x00 +DB 0x22,0xb8,0x05,0x00 +DB 0x1d,0xa8,0x04,0x00 +DB 0x18,0x98,0x03,0x00 +DB 0x13,0x88,0x02,0x00 +DB 0x0e,0x78,0x01,0x00 +DB 0x09,0x68,0x00,0x00 +DB 0x04,0x01,0x15,0x00 +DB 0x00,0xb3,0x00,0x00 diff --git a/vere/ext/openssl/gen/windows-x86_64/crypto/bn/rsaz-x86_64.asm b/vere/ext/openssl/gen/windows-x86_64/crypto/bn/rsaz-x86_64.asm new file mode 100644 index 0000000..a474c5d --- /dev/null +++ b/vere/ext/openssl/gen/windows-x86_64/crypto/bn/rsaz-x86_64.asm @@ -0,0 +1,2261 @@ +default rel +%define XMMWORD +%define YMMWORD +%define ZMMWORD +section .text code align=64 + + +EXTERN OPENSSL_ia32cap_P + +global rsaz_512_sqr + +ALIGN 32 +rsaz_512_sqr: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[16+rsp],rsi + mov rax,rsp +$L$SEH_begin_rsaz_512_sqr: + mov rdi,rcx + mov rsi,rdx + mov rdx,r8 + mov rcx,r9 + mov r8,QWORD[40+rsp] + + + + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + + + sub rsp,128+24 + +$L$sqr_body: +DB 102,72,15,110,202 + mov rdx,QWORD[rsi] + mov rax,QWORD[8+rsi] + mov QWORD[128+rsp],rcx + mov r11d,0x80100 + and r11d,DWORD[((OPENSSL_ia32cap_P+8))] + cmp r11d,0x80100 + je NEAR $L$oop_sqrx + jmp NEAR $L$oop_sqr + +ALIGN 32 +$L$oop_sqr: + mov DWORD[((128+8))+rsp],r8d + + mov rbx,rdx + mov rbp,rax + mul rdx + mov r8,rax + mov rax,QWORD[16+rsi] + mov r9,rdx + + mul rbx + add r9,rax + mov rax,QWORD[24+rsi] + mov r10,rdx + adc r10,0 + + mul rbx + add r10,rax + mov rax,QWORD[32+rsi] + mov r11,rdx + adc r11,0 + + mul rbx + add r11,rax + mov rax,QWORD[40+rsi] + mov r12,rdx + adc r12,0 + + mul rbx + add r12,rax + mov rax,QWORD[48+rsi] + mov r13,rdx + adc r13,0 + + mul rbx + add r13,rax + mov rax,QWORD[56+rsi] + mov r14,rdx + adc r14,0 + + mul rbx + add r14,rax + mov rax,rbx + adc rdx,0 + + xor rcx,rcx + add r8,r8 + mov r15,rdx + adc rcx,0 + + mul rax + add rdx,r8 + adc rcx,0 + + mov QWORD[rsp],rax + mov QWORD[8+rsp],rdx + + + mov rax,QWORD[16+rsi] + mul rbp + add r10,rax + mov rax,QWORD[24+rsi] + mov rbx,rdx + adc rbx,0 + + mul rbp + add r11,rax + mov rax,QWORD[32+rsi] + adc rdx,0 + add r11,rbx + mov rbx,rdx + adc rbx,0 + + mul rbp + add r12,rax + mov rax,QWORD[40+rsi] + adc rdx,0 + add r12,rbx + mov rbx,rdx + adc rbx,0 + + mul rbp + add r13,rax + mov rax,QWORD[48+rsi] + adc rdx,0 + add r13,rbx + mov rbx,rdx + adc rbx,0 + + mul rbp + add r14,rax + mov rax,QWORD[56+rsi] + adc rdx,0 + add r14,rbx + mov rbx,rdx + adc rbx,0 + + mul rbp + add r15,rax + mov rax,rbp + adc rdx,0 + add r15,rbx + adc rdx,0 + + xor rbx,rbx + add r9,r9 + mov r8,rdx + adc r10,r10 + adc rbx,0 + + mul rax + + add rax,rcx + mov rbp,QWORD[16+rsi] + add r9,rax + mov rax,QWORD[24+rsi] + adc r10,rdx + adc rbx,0 + + mov QWORD[16+rsp],r9 + mov QWORD[24+rsp],r10 + + + mul rbp + add r12,rax + mov rax,QWORD[32+rsi] + mov rcx,rdx + adc rcx,0 + + mul rbp + add r13,rax + mov rax,QWORD[40+rsi] + adc rdx,0 + add r13,rcx + mov rcx,rdx + adc rcx,0 + + mul rbp + add r14,rax + mov rax,QWORD[48+rsi] + adc rdx,0 + add r14,rcx + mov rcx,rdx + adc rcx,0 + + mul rbp + add r15,rax + mov rax,QWORD[56+rsi] + adc rdx,0 + add r15,rcx + mov rcx,rdx + adc rcx,0 + + mul rbp + add r8,rax + mov rax,rbp + adc rdx,0 + add r8,rcx + adc rdx,0 + + xor rcx,rcx + add r11,r11 + mov r9,rdx + adc r12,r12 + adc rcx,0 + + mul rax + + add rax,rbx + mov r10,QWORD[24+rsi] + add r11,rax + mov rax,QWORD[32+rsi] + adc r12,rdx + adc rcx,0 + + mov QWORD[32+rsp],r11 + mov QWORD[40+rsp],r12 + + + mov r11,rax + mul r10 + add r14,rax + mov rax,QWORD[40+rsi] + mov rbx,rdx + adc rbx,0 + + mov r12,rax + mul r10 + add r15,rax + mov rax,QWORD[48+rsi] + adc rdx,0 + add r15,rbx + mov rbx,rdx + adc rbx,0 + + mov rbp,rax + mul r10 + add r8,rax + mov rax,QWORD[56+rsi] + adc rdx,0 + add r8,rbx + mov rbx,rdx + adc rbx,0 + + mul r10 + add r9,rax + mov rax,r10 + adc rdx,0 + add r9,rbx + adc rdx,0 + + xor rbx,rbx + add r13,r13 + mov r10,rdx + adc r14,r14 + adc rbx,0 + + mul rax + + add rax,rcx + add r13,rax + mov rax,r12 + adc r14,rdx + adc rbx,0 + + mov QWORD[48+rsp],r13 + mov QWORD[56+rsp],r14 + + + mul r11 + add r8,rax + mov rax,rbp + mov rcx,rdx + adc rcx,0 + + mul r11 + add r9,rax + mov rax,QWORD[56+rsi] + adc rdx,0 + add r9,rcx + mov rcx,rdx + adc rcx,0 + + mov r14,rax + mul r11 + add r10,rax + mov rax,r11 + adc rdx,0 + add r10,rcx + adc rdx,0 + + xor rcx,rcx + add r15,r15 + mov r11,rdx + adc r8,r8 + adc rcx,0 + + mul rax + + add rax,rbx + add r15,rax + mov rax,rbp + adc r8,rdx + adc rcx,0 + + mov QWORD[64+rsp],r15 + mov QWORD[72+rsp],r8 + + + mul r12 + add r10,rax + mov rax,r14 + mov rbx,rdx + adc rbx,0 + + mul r12 + add r11,rax + mov rax,r12 + adc rdx,0 + add r11,rbx + adc rdx,0 + + xor rbx,rbx + add r9,r9 + mov r12,rdx + adc r10,r10 + adc rbx,0 + + mul rax + + add rax,rcx + add r9,rax + mov rax,r14 + adc r10,rdx + adc rbx,0 + + mov QWORD[80+rsp],r9 + mov QWORD[88+rsp],r10 + + + mul rbp + add r12,rax + mov rax,rbp + adc rdx,0 + + xor rcx,rcx + add r11,r11 + mov r13,rdx + adc r12,r12 + adc rcx,0 + + mul rax + + add rax,rbx + add r11,rax + mov rax,r14 + adc r12,rdx + adc rcx,0 + + mov QWORD[96+rsp],r11 + mov QWORD[104+rsp],r12 + + + xor rbx,rbx + add r13,r13 + adc rbx,0 + + mul rax + + add rax,rcx + add rax,r13 + adc rdx,rbx + + mov r8,QWORD[rsp] + mov r9,QWORD[8+rsp] + mov r10,QWORD[16+rsp] + mov r11,QWORD[24+rsp] + mov r12,QWORD[32+rsp] + mov r13,QWORD[40+rsp] + mov r14,QWORD[48+rsp] + mov r15,QWORD[56+rsp] +DB 102,72,15,126,205 + + mov QWORD[112+rsp],rax + mov QWORD[120+rsp],rdx + + call __rsaz_512_reduce + + add r8,QWORD[64+rsp] + adc r9,QWORD[72+rsp] + adc r10,QWORD[80+rsp] + adc r11,QWORD[88+rsp] + adc r12,QWORD[96+rsp] + adc r13,QWORD[104+rsp] + adc r14,QWORD[112+rsp] + adc r15,QWORD[120+rsp] + sbb rcx,rcx + + call __rsaz_512_subtract + + mov rdx,r8 + mov rax,r9 + mov r8d,DWORD[((128+8))+rsp] + mov rsi,rdi + + dec r8d + jnz NEAR $L$oop_sqr + jmp NEAR $L$sqr_tail + +ALIGN 32 +$L$oop_sqrx: + mov DWORD[((128+8))+rsp],r8d +DB 102,72,15,110,199 + + mulx r9,r8,rax + mov rbx,rax + + mulx r10,rcx,QWORD[16+rsi] + xor rbp,rbp + + mulx r11,rax,QWORD[24+rsi] + adcx r9,rcx + +DB 0xc4,0x62,0xf3,0xf6,0xa6,0x20,0x00,0x00,0x00 + adcx r10,rax + +DB 0xc4,0x62,0xfb,0xf6,0xae,0x28,0x00,0x00,0x00 + adcx r11,rcx + + mulx r14,rcx,QWORD[48+rsi] + adcx r12,rax + adcx r13,rcx + + mulx r15,rax,QWORD[56+rsi] + adcx r14,rax + adcx r15,rbp + + mulx rdi,rax,rdx + mov rdx,rbx + xor rcx,rcx + adox r8,r8 + adcx r8,rdi + adox rcx,rbp + adcx rcx,rbp + + mov QWORD[rsp],rax + mov QWORD[8+rsp],r8 + + +DB 0xc4,0xe2,0xfb,0xf6,0x9e,0x10,0x00,0x00,0x00 + adox r10,rax + adcx r11,rbx + + mulx r8,rdi,QWORD[24+rsi] + adox r11,rdi +DB 0x66 + adcx r12,r8 + + mulx rbx,rax,QWORD[32+rsi] + adox r12,rax + adcx r13,rbx + + mulx r8,rdi,QWORD[40+rsi] + adox r13,rdi + adcx r14,r8 + +DB 0xc4,0xe2,0xfb,0xf6,0x9e,0x30,0x00,0x00,0x00 + adox r14,rax + adcx r15,rbx + +DB 0xc4,0x62,0xc3,0xf6,0x86,0x38,0x00,0x00,0x00 + adox r15,rdi + adcx r8,rbp + mulx rdi,rax,rdx + adox r8,rbp +DB 0x48,0x8b,0x96,0x10,0x00,0x00,0x00 + + xor rbx,rbx + adox r9,r9 + + adcx rax,rcx + adox r10,r10 + adcx r9,rax + adox rbx,rbp + adcx r10,rdi + adcx rbx,rbp + + mov QWORD[16+rsp],r9 +DB 0x4c,0x89,0x94,0x24,0x18,0x00,0x00,0x00 + + + mulx r9,rdi,QWORD[24+rsi] + adox r12,rdi + adcx r13,r9 + + mulx rcx,rax,QWORD[32+rsi] + adox r13,rax + adcx r14,rcx + +DB 0xc4,0x62,0xc3,0xf6,0x8e,0x28,0x00,0x00,0x00 + adox r14,rdi + adcx r15,r9 + +DB 0xc4,0xe2,0xfb,0xf6,0x8e,0x30,0x00,0x00,0x00 + adox r15,rax + adcx r8,rcx + + mulx r9,rdi,QWORD[56+rsi] + adox r8,rdi + adcx r9,rbp + mulx rdi,rax,rdx + adox r9,rbp + mov rdx,QWORD[24+rsi] + + xor rcx,rcx + adox r11,r11 + + adcx rax,rbx + adox r12,r12 + adcx r11,rax + adox rcx,rbp + adcx r12,rdi + adcx rcx,rbp + + mov QWORD[32+rsp],r11 + mov QWORD[40+rsp],r12 + + + mulx rbx,rax,QWORD[32+rsi] + adox r14,rax + adcx r15,rbx + + mulx r10,rdi,QWORD[40+rsi] + adox r15,rdi + adcx r8,r10 + + mulx rbx,rax,QWORD[48+rsi] + adox r8,rax + adcx r9,rbx + + mulx r10,rdi,QWORD[56+rsi] + adox r9,rdi + adcx r10,rbp + mulx rdi,rax,rdx + adox r10,rbp + mov rdx,QWORD[32+rsi] + + xor rbx,rbx + adox r13,r13 + + adcx rax,rcx + adox r14,r14 + adcx r13,rax + adox rbx,rbp + adcx r14,rdi + adcx rbx,rbp + + mov QWORD[48+rsp],r13 + mov QWORD[56+rsp],r14 + + + mulx r11,rdi,QWORD[40+rsi] + adox r8,rdi + adcx r9,r11 + + mulx rcx,rax,QWORD[48+rsi] + adox r9,rax + adcx r10,rcx + + mulx r11,rdi,QWORD[56+rsi] + adox r10,rdi + adcx r11,rbp + mulx rdi,rax,rdx + mov rdx,QWORD[40+rsi] + adox r11,rbp + + xor rcx,rcx + adox r15,r15 + + adcx rax,rbx + adox r8,r8 + adcx r15,rax + adox rcx,rbp + adcx r8,rdi + adcx rcx,rbp + + mov QWORD[64+rsp],r15 + mov QWORD[72+rsp],r8 + + +DB 0xc4,0xe2,0xfb,0xf6,0x9e,0x30,0x00,0x00,0x00 + adox r10,rax + adcx r11,rbx + +DB 0xc4,0x62,0xc3,0xf6,0xa6,0x38,0x00,0x00,0x00 + adox r11,rdi + adcx r12,rbp + mulx rdi,rax,rdx + adox r12,rbp + mov rdx,QWORD[48+rsi] + + xor rbx,rbx + adox r9,r9 + + adcx rax,rcx + adox r10,r10 + adcx r9,rax + adcx r10,rdi + adox rbx,rbp + adcx rbx,rbp + + mov QWORD[80+rsp],r9 + mov QWORD[88+rsp],r10 + + +DB 0xc4,0x62,0xfb,0xf6,0xae,0x38,0x00,0x00,0x00 + adox r12,rax + adox r13,rbp + + mulx rdi,rax,rdx + xor rcx,rcx + mov rdx,QWORD[56+rsi] + adox r11,r11 + + adcx rax,rbx + adox r12,r12 + adcx r11,rax + adox rcx,rbp + adcx r12,rdi + adcx rcx,rbp + +DB 0x4c,0x89,0x9c,0x24,0x60,0x00,0x00,0x00 +DB 0x4c,0x89,0xa4,0x24,0x68,0x00,0x00,0x00 + + + mulx rdx,rax,rdx + xor rbx,rbx + adox r13,r13 + + adcx rax,rcx + adox rbx,rbp + adcx rax,r13 + adcx rbx,rdx + +DB 102,72,15,126,199 +DB 102,72,15,126,205 + + mov rdx,QWORD[128+rsp] + mov r8,QWORD[rsp] + mov r9,QWORD[8+rsp] + mov r10,QWORD[16+rsp] + mov r11,QWORD[24+rsp] + mov r12,QWORD[32+rsp] + mov r13,QWORD[40+rsp] + mov r14,QWORD[48+rsp] + mov r15,QWORD[56+rsp] + + mov QWORD[112+rsp],rax + mov QWORD[120+rsp],rbx + + call __rsaz_512_reducex + + add r8,QWORD[64+rsp] + adc r9,QWORD[72+rsp] + adc r10,QWORD[80+rsp] + adc r11,QWORD[88+rsp] + adc r12,QWORD[96+rsp] + adc r13,QWORD[104+rsp] + adc r14,QWORD[112+rsp] + adc r15,QWORD[120+rsp] + sbb rcx,rcx + + call __rsaz_512_subtract + + mov rdx,r8 + mov rax,r9 + mov r8d,DWORD[((128+8))+rsp] + mov rsi,rdi + + dec r8d + jnz NEAR $L$oop_sqrx + +$L$sqr_tail: + + lea rax,[((128+24+48))+rsp] + + mov r15,QWORD[((-48))+rax] + + mov r14,QWORD[((-40))+rax] + + mov r13,QWORD[((-32))+rax] + + mov r12,QWORD[((-24))+rax] + + mov rbp,QWORD[((-16))+rax] + + mov rbx,QWORD[((-8))+rax] + + lea rsp,[rax] + +$L$sqr_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_rsaz_512_sqr: +global rsaz_512_mul + +ALIGN 32 +rsaz_512_mul: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[16+rsp],rsi + mov rax,rsp +$L$SEH_begin_rsaz_512_mul: + mov rdi,rcx + mov rsi,rdx + mov rdx,r8 + mov rcx,r9 + mov r8,QWORD[40+rsp] + + + + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + + + sub rsp,128+24 + +$L$mul_body: +DB 102,72,15,110,199 +DB 102,72,15,110,201 + mov QWORD[128+rsp],r8 + mov r11d,0x80100 + and r11d,DWORD[((OPENSSL_ia32cap_P+8))] + cmp r11d,0x80100 + je NEAR $L$mulx + mov rbx,QWORD[rdx] + mov rbp,rdx + call __rsaz_512_mul + +DB 102,72,15,126,199 +DB 102,72,15,126,205 + + mov r8,QWORD[rsp] + mov r9,QWORD[8+rsp] + mov r10,QWORD[16+rsp] + mov r11,QWORD[24+rsp] + mov r12,QWORD[32+rsp] + mov r13,QWORD[40+rsp] + mov r14,QWORD[48+rsp] + mov r15,QWORD[56+rsp] + + call __rsaz_512_reduce + jmp NEAR $L$mul_tail + +ALIGN 32 +$L$mulx: + mov rbp,rdx + mov rdx,QWORD[rdx] + call __rsaz_512_mulx + +DB 102,72,15,126,199 +DB 102,72,15,126,205 + + mov rdx,QWORD[128+rsp] + mov r8,QWORD[rsp] + mov r9,QWORD[8+rsp] + mov r10,QWORD[16+rsp] + mov r11,QWORD[24+rsp] + mov r12,QWORD[32+rsp] + mov r13,QWORD[40+rsp] + mov r14,QWORD[48+rsp] + mov r15,QWORD[56+rsp] + + call __rsaz_512_reducex +$L$mul_tail: + add r8,QWORD[64+rsp] + adc r9,QWORD[72+rsp] + adc r10,QWORD[80+rsp] + adc r11,QWORD[88+rsp] + adc r12,QWORD[96+rsp] + adc r13,QWORD[104+rsp] + adc r14,QWORD[112+rsp] + adc r15,QWORD[120+rsp] + sbb rcx,rcx + + call __rsaz_512_subtract + + lea rax,[((128+24+48))+rsp] + + mov r15,QWORD[((-48))+rax] + + mov r14,QWORD[((-40))+rax] + + mov r13,QWORD[((-32))+rax] + + mov r12,QWORD[((-24))+rax] + + mov rbp,QWORD[((-16))+rax] + + mov rbx,QWORD[((-8))+rax] + + lea rsp,[rax] + +$L$mul_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_rsaz_512_mul: +global rsaz_512_mul_gather4 + +ALIGN 32 +rsaz_512_mul_gather4: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[16+rsp],rsi + mov rax,rsp +$L$SEH_begin_rsaz_512_mul_gather4: + mov rdi,rcx + mov rsi,rdx + mov rdx,r8 + mov rcx,r9 + mov r8,QWORD[40+rsp] + mov r9,QWORD[48+rsp] + + + + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + + + sub rsp,328 + + movaps XMMWORD[160+rsp],xmm6 + movaps XMMWORD[176+rsp],xmm7 + movaps XMMWORD[192+rsp],xmm8 + movaps XMMWORD[208+rsp],xmm9 + movaps XMMWORD[224+rsp],xmm10 + movaps XMMWORD[240+rsp],xmm11 + movaps XMMWORD[256+rsp],xmm12 + movaps XMMWORD[272+rsp],xmm13 + movaps XMMWORD[288+rsp],xmm14 + movaps XMMWORD[304+rsp],xmm15 +$L$mul_gather4_body: + movd xmm8,r9d + movdqa xmm1,XMMWORD[(($L$inc+16))] + movdqa xmm0,XMMWORD[$L$inc] + + pshufd xmm8,xmm8,0 + movdqa xmm7,xmm1 + movdqa xmm2,xmm1 + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm8 + movdqa xmm3,xmm7 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm8 + movdqa xmm4,xmm7 + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm8 + movdqa xmm5,xmm7 + paddd xmm4,xmm3 + pcmpeqd xmm3,xmm8 + movdqa xmm6,xmm7 + paddd xmm5,xmm4 + pcmpeqd xmm4,xmm8 + paddd xmm6,xmm5 + pcmpeqd xmm5,xmm8 + paddd xmm7,xmm6 + pcmpeqd xmm6,xmm8 + pcmpeqd xmm7,xmm8 + + movdqa xmm8,XMMWORD[rdx] + movdqa xmm9,XMMWORD[16+rdx] + movdqa xmm10,XMMWORD[32+rdx] + movdqa xmm11,XMMWORD[48+rdx] + pand xmm8,xmm0 + movdqa xmm12,XMMWORD[64+rdx] + pand xmm9,xmm1 + movdqa xmm13,XMMWORD[80+rdx] + pand xmm10,xmm2 + movdqa xmm14,XMMWORD[96+rdx] + pand xmm11,xmm3 + movdqa xmm15,XMMWORD[112+rdx] + lea rbp,[128+rdx] + pand xmm12,xmm4 + pand xmm13,xmm5 + pand xmm14,xmm6 + pand xmm15,xmm7 + por xmm8,xmm10 + por xmm9,xmm11 + por xmm8,xmm12 + por xmm9,xmm13 + por xmm8,xmm14 + por xmm9,xmm15 + + por xmm8,xmm9 + pshufd xmm9,xmm8,0x4e + por xmm8,xmm9 + mov r11d,0x80100 + and r11d,DWORD[((OPENSSL_ia32cap_P+8))] + cmp r11d,0x80100 + je NEAR $L$mulx_gather +DB 102,76,15,126,195 + + mov QWORD[128+rsp],r8 + mov QWORD[((128+8))+rsp],rdi + mov QWORD[((128+16))+rsp],rcx + + mov rax,QWORD[rsi] + mov rcx,QWORD[8+rsi] + mul rbx + mov QWORD[rsp],rax + mov rax,rcx + mov r8,rdx + + mul rbx + add r8,rax + mov rax,QWORD[16+rsi] + mov r9,rdx + adc r9,0 + + mul rbx + add r9,rax + mov rax,QWORD[24+rsi] + mov r10,rdx + adc r10,0 + + mul rbx + add r10,rax + mov rax,QWORD[32+rsi] + mov r11,rdx + adc r11,0 + + mul rbx + add r11,rax + mov rax,QWORD[40+rsi] + mov r12,rdx + adc r12,0 + + mul rbx + add r12,rax + mov rax,QWORD[48+rsi] + mov r13,rdx + adc r13,0 + + mul rbx + add r13,rax + mov rax,QWORD[56+rsi] + mov r14,rdx + adc r14,0 + + mul rbx + add r14,rax + mov rax,QWORD[rsi] + mov r15,rdx + adc r15,0 + + lea rdi,[8+rsp] + mov ecx,7 + jmp NEAR $L$oop_mul_gather + +ALIGN 32 +$L$oop_mul_gather: + movdqa xmm8,XMMWORD[rbp] + movdqa xmm9,XMMWORD[16+rbp] + movdqa xmm10,XMMWORD[32+rbp] + movdqa xmm11,XMMWORD[48+rbp] + pand xmm8,xmm0 + movdqa xmm12,XMMWORD[64+rbp] + pand xmm9,xmm1 + movdqa xmm13,XMMWORD[80+rbp] + pand xmm10,xmm2 + movdqa xmm14,XMMWORD[96+rbp] + pand xmm11,xmm3 + movdqa xmm15,XMMWORD[112+rbp] + lea rbp,[128+rbp] + pand xmm12,xmm4 + pand xmm13,xmm5 + pand xmm14,xmm6 + pand xmm15,xmm7 + por xmm8,xmm10 + por xmm9,xmm11 + por xmm8,xmm12 + por xmm9,xmm13 + por xmm8,xmm14 + por xmm9,xmm15 + + por xmm8,xmm9 + pshufd xmm9,xmm8,0x4e + por xmm8,xmm9 +DB 102,76,15,126,195 + + mul rbx + add r8,rax + mov rax,QWORD[8+rsi] + mov QWORD[rdi],r8 + mov r8,rdx + adc r8,0 + + mul rbx + add r9,rax + mov rax,QWORD[16+rsi] + adc rdx,0 + add r8,r9 + mov r9,rdx + adc r9,0 + + mul rbx + add r10,rax + mov rax,QWORD[24+rsi] + adc rdx,0 + add r9,r10 + mov r10,rdx + adc r10,0 + + mul rbx + add r11,rax + mov rax,QWORD[32+rsi] + adc rdx,0 + add r10,r11 + mov r11,rdx + adc r11,0 + + mul rbx + add r12,rax + mov rax,QWORD[40+rsi] + adc rdx,0 + add r11,r12 + mov r12,rdx + adc r12,0 + + mul rbx + add r13,rax + mov rax,QWORD[48+rsi] + adc rdx,0 + add r12,r13 + mov r13,rdx + adc r13,0 + + mul rbx + add r14,rax + mov rax,QWORD[56+rsi] + adc rdx,0 + add r13,r14 + mov r14,rdx + adc r14,0 + + mul rbx + add r15,rax + mov rax,QWORD[rsi] + adc rdx,0 + add r14,r15 + mov r15,rdx + adc r15,0 + + lea rdi,[8+rdi] + + dec ecx + jnz NEAR $L$oop_mul_gather + + mov QWORD[rdi],r8 + mov QWORD[8+rdi],r9 + mov QWORD[16+rdi],r10 + mov QWORD[24+rdi],r11 + mov QWORD[32+rdi],r12 + mov QWORD[40+rdi],r13 + mov QWORD[48+rdi],r14 + mov QWORD[56+rdi],r15 + + mov rdi,QWORD[((128+8))+rsp] + mov rbp,QWORD[((128+16))+rsp] + + mov r8,QWORD[rsp] + mov r9,QWORD[8+rsp] + mov r10,QWORD[16+rsp] + mov r11,QWORD[24+rsp] + mov r12,QWORD[32+rsp] + mov r13,QWORD[40+rsp] + mov r14,QWORD[48+rsp] + mov r15,QWORD[56+rsp] + + call __rsaz_512_reduce + jmp NEAR $L$mul_gather_tail + +ALIGN 32 +$L$mulx_gather: +DB 102,76,15,126,194 + + mov QWORD[128+rsp],r8 + mov QWORD[((128+8))+rsp],rdi + mov QWORD[((128+16))+rsp],rcx + + mulx r8,rbx,QWORD[rsi] + mov QWORD[rsp],rbx + xor edi,edi + + mulx r9,rax,QWORD[8+rsi] + + mulx r10,rbx,QWORD[16+rsi] + adcx r8,rax + + mulx r11,rax,QWORD[24+rsi] + adcx r9,rbx + + mulx r12,rbx,QWORD[32+rsi] + adcx r10,rax + + mulx r13,rax,QWORD[40+rsi] + adcx r11,rbx + + mulx r14,rbx,QWORD[48+rsi] + adcx r12,rax + + mulx r15,rax,QWORD[56+rsi] + adcx r13,rbx + adcx r14,rax +DB 0x67 + mov rbx,r8 + adcx r15,rdi + + mov rcx,-7 + jmp NEAR $L$oop_mulx_gather + +ALIGN 32 +$L$oop_mulx_gather: + movdqa xmm8,XMMWORD[rbp] + movdqa xmm9,XMMWORD[16+rbp] + movdqa xmm10,XMMWORD[32+rbp] + movdqa xmm11,XMMWORD[48+rbp] + pand xmm8,xmm0 + movdqa xmm12,XMMWORD[64+rbp] + pand xmm9,xmm1 + movdqa xmm13,XMMWORD[80+rbp] + pand xmm10,xmm2 + movdqa xmm14,XMMWORD[96+rbp] + pand xmm11,xmm3 + movdqa xmm15,XMMWORD[112+rbp] + lea rbp,[128+rbp] + pand xmm12,xmm4 + pand xmm13,xmm5 + pand xmm14,xmm6 + pand xmm15,xmm7 + por xmm8,xmm10 + por xmm9,xmm11 + por xmm8,xmm12 + por xmm9,xmm13 + por xmm8,xmm14 + por xmm9,xmm15 + + por xmm8,xmm9 + pshufd xmm9,xmm8,0x4e + por xmm8,xmm9 +DB 102,76,15,126,194 + +DB 0xc4,0x62,0xfb,0xf6,0x86,0x00,0x00,0x00,0x00 + adcx rbx,rax + adox r8,r9 + + mulx r9,rax,QWORD[8+rsi] + adcx r8,rax + adox r9,r10 + + mulx r10,rax,QWORD[16+rsi] + adcx r9,rax + adox r10,r11 + +DB 0xc4,0x62,0xfb,0xf6,0x9e,0x18,0x00,0x00,0x00 + adcx r10,rax + adox r11,r12 + + mulx r12,rax,QWORD[32+rsi] + adcx r11,rax + adox r12,r13 + + mulx r13,rax,QWORD[40+rsi] + adcx r12,rax + adox r13,r14 + +DB 0xc4,0x62,0xfb,0xf6,0xb6,0x30,0x00,0x00,0x00 + adcx r13,rax +DB 0x67 + adox r14,r15 + + mulx r15,rax,QWORD[56+rsi] + mov QWORD[64+rcx*8+rsp],rbx + adcx r14,rax + adox r15,rdi + mov rbx,r8 + adcx r15,rdi + + inc rcx + jnz NEAR $L$oop_mulx_gather + + mov QWORD[64+rsp],r8 + mov QWORD[((64+8))+rsp],r9 + mov QWORD[((64+16))+rsp],r10 + mov QWORD[((64+24))+rsp],r11 + mov QWORD[((64+32))+rsp],r12 + mov QWORD[((64+40))+rsp],r13 + mov QWORD[((64+48))+rsp],r14 + mov QWORD[((64+56))+rsp],r15 + + mov rdx,QWORD[128+rsp] + mov rdi,QWORD[((128+8))+rsp] + mov rbp,QWORD[((128+16))+rsp] + + mov r8,QWORD[rsp] + mov r9,QWORD[8+rsp] + mov r10,QWORD[16+rsp] + mov r11,QWORD[24+rsp] + mov r12,QWORD[32+rsp] + mov r13,QWORD[40+rsp] + mov r14,QWORD[48+rsp] + mov r15,QWORD[56+rsp] + + call __rsaz_512_reducex + +$L$mul_gather_tail: + add r8,QWORD[64+rsp] + adc r9,QWORD[72+rsp] + adc r10,QWORD[80+rsp] + adc r11,QWORD[88+rsp] + adc r12,QWORD[96+rsp] + adc r13,QWORD[104+rsp] + adc r14,QWORD[112+rsp] + adc r15,QWORD[120+rsp] + sbb rcx,rcx + + call __rsaz_512_subtract + + lea rax,[((128+24+48))+rsp] + movaps xmm6,XMMWORD[((160-200))+rax] + movaps xmm7,XMMWORD[((176-200))+rax] + movaps xmm8,XMMWORD[((192-200))+rax] + movaps xmm9,XMMWORD[((208-200))+rax] + movaps xmm10,XMMWORD[((224-200))+rax] + movaps xmm11,XMMWORD[((240-200))+rax] + movaps xmm12,XMMWORD[((256-200))+rax] + movaps xmm13,XMMWORD[((272-200))+rax] + movaps xmm14,XMMWORD[((288-200))+rax] + movaps xmm15,XMMWORD[((304-200))+rax] + lea rax,[176+rax] + + mov r15,QWORD[((-48))+rax] + + mov r14,QWORD[((-40))+rax] + + mov r13,QWORD[((-32))+rax] + + mov r12,QWORD[((-24))+rax] + + mov rbp,QWORD[((-16))+rax] + + mov rbx,QWORD[((-8))+rax] + + lea rsp,[rax] + +$L$mul_gather4_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_rsaz_512_mul_gather4: +global rsaz_512_mul_scatter4 + +ALIGN 32 +rsaz_512_mul_scatter4: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[16+rsp],rsi + mov rax,rsp +$L$SEH_begin_rsaz_512_mul_scatter4: + mov rdi,rcx + mov rsi,rdx + mov rdx,r8 + mov rcx,r9 + mov r8,QWORD[40+rsp] + mov r9,QWORD[48+rsp] + + + + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + + + mov r9d,r9d + sub rsp,128+24 + +$L$mul_scatter4_body: + lea r8,[r9*8+r8] +DB 102,72,15,110,199 +DB 102,72,15,110,202 +DB 102,73,15,110,208 + mov QWORD[128+rsp],rcx + + mov rbp,rdi + mov r11d,0x80100 + and r11d,DWORD[((OPENSSL_ia32cap_P+8))] + cmp r11d,0x80100 + je NEAR $L$mulx_scatter + mov rbx,QWORD[rdi] + call __rsaz_512_mul + +DB 102,72,15,126,199 +DB 102,72,15,126,205 + + mov r8,QWORD[rsp] + mov r9,QWORD[8+rsp] + mov r10,QWORD[16+rsp] + mov r11,QWORD[24+rsp] + mov r12,QWORD[32+rsp] + mov r13,QWORD[40+rsp] + mov r14,QWORD[48+rsp] + mov r15,QWORD[56+rsp] + + call __rsaz_512_reduce + jmp NEAR $L$mul_scatter_tail + +ALIGN 32 +$L$mulx_scatter: + mov rdx,QWORD[rdi] + call __rsaz_512_mulx + +DB 102,72,15,126,199 +DB 102,72,15,126,205 + + mov rdx,QWORD[128+rsp] + mov r8,QWORD[rsp] + mov r9,QWORD[8+rsp] + mov r10,QWORD[16+rsp] + mov r11,QWORD[24+rsp] + mov r12,QWORD[32+rsp] + mov r13,QWORD[40+rsp] + mov r14,QWORD[48+rsp] + mov r15,QWORD[56+rsp] + + call __rsaz_512_reducex + +$L$mul_scatter_tail: + add r8,QWORD[64+rsp] + adc r9,QWORD[72+rsp] + adc r10,QWORD[80+rsp] + adc r11,QWORD[88+rsp] + adc r12,QWORD[96+rsp] + adc r13,QWORD[104+rsp] + adc r14,QWORD[112+rsp] + adc r15,QWORD[120+rsp] +DB 102,72,15,126,214 + sbb rcx,rcx + + call __rsaz_512_subtract + + mov QWORD[rsi],r8 + mov QWORD[128+rsi],r9 + mov QWORD[256+rsi],r10 + mov QWORD[384+rsi],r11 + mov QWORD[512+rsi],r12 + mov QWORD[640+rsi],r13 + mov QWORD[768+rsi],r14 + mov QWORD[896+rsi],r15 + + lea rax,[((128+24+48))+rsp] + + mov r15,QWORD[((-48))+rax] + + mov r14,QWORD[((-40))+rax] + + mov r13,QWORD[((-32))+rax] + + mov r12,QWORD[((-24))+rax] + + mov rbp,QWORD[((-16))+rax] + + mov rbx,QWORD[((-8))+rax] + + lea rsp,[rax] + +$L$mul_scatter4_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_rsaz_512_mul_scatter4: +global rsaz_512_mul_by_one + +ALIGN 32 +rsaz_512_mul_by_one: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[16+rsp],rsi + mov rax,rsp +$L$SEH_begin_rsaz_512_mul_by_one: + mov rdi,rcx + mov rsi,rdx + mov rdx,r8 + mov rcx,r9 + + + + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + + + sub rsp,128+24 + +$L$mul_by_one_body: + mov eax,DWORD[((OPENSSL_ia32cap_P+8))] + mov rbp,rdx + mov QWORD[128+rsp],rcx + + mov r8,QWORD[rsi] + pxor xmm0,xmm0 + mov r9,QWORD[8+rsi] + mov r10,QWORD[16+rsi] + mov r11,QWORD[24+rsi] + mov r12,QWORD[32+rsi] + mov r13,QWORD[40+rsi] + mov r14,QWORD[48+rsi] + mov r15,QWORD[56+rsi] + + movdqa XMMWORD[rsp],xmm0 + movdqa XMMWORD[16+rsp],xmm0 + movdqa XMMWORD[32+rsp],xmm0 + movdqa XMMWORD[48+rsp],xmm0 + movdqa XMMWORD[64+rsp],xmm0 + movdqa XMMWORD[80+rsp],xmm0 + movdqa XMMWORD[96+rsp],xmm0 + and eax,0x80100 + cmp eax,0x80100 + je NEAR $L$by_one_callx + call __rsaz_512_reduce + jmp NEAR $L$by_one_tail +ALIGN 32 +$L$by_one_callx: + mov rdx,QWORD[128+rsp] + call __rsaz_512_reducex +$L$by_one_tail: + mov QWORD[rdi],r8 + mov QWORD[8+rdi],r9 + mov QWORD[16+rdi],r10 + mov QWORD[24+rdi],r11 + mov QWORD[32+rdi],r12 + mov QWORD[40+rdi],r13 + mov QWORD[48+rdi],r14 + mov QWORD[56+rdi],r15 + + lea rax,[((128+24+48))+rsp] + + mov r15,QWORD[((-48))+rax] + + mov r14,QWORD[((-40))+rax] + + mov r13,QWORD[((-32))+rax] + + mov r12,QWORD[((-24))+rax] + + mov rbp,QWORD[((-16))+rax] + + mov rbx,QWORD[((-8))+rax] + + lea rsp,[rax] + +$L$mul_by_one_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_rsaz_512_mul_by_one: + +ALIGN 32 +__rsaz_512_reduce: + + mov rbx,r8 + imul rbx,QWORD[((128+8))+rsp] + mov rax,QWORD[rbp] + mov ecx,8 + jmp NEAR $L$reduction_loop + +ALIGN 32 +$L$reduction_loop: + mul rbx + mov rax,QWORD[8+rbp] + neg r8 + mov r8,rdx + adc r8,0 + + mul rbx + add r9,rax + mov rax,QWORD[16+rbp] + adc rdx,0 + add r8,r9 + mov r9,rdx + adc r9,0 + + mul rbx + add r10,rax + mov rax,QWORD[24+rbp] + adc rdx,0 + add r9,r10 + mov r10,rdx + adc r10,0 + + mul rbx + add r11,rax + mov rax,QWORD[32+rbp] + adc rdx,0 + add r10,r11 + mov rsi,QWORD[((128+8))+rsp] + + + adc rdx,0 + mov r11,rdx + + mul rbx + add r12,rax + mov rax,QWORD[40+rbp] + adc rdx,0 + imul rsi,r8 + add r11,r12 + mov r12,rdx + adc r12,0 + + mul rbx + add r13,rax + mov rax,QWORD[48+rbp] + adc rdx,0 + add r12,r13 + mov r13,rdx + adc r13,0 + + mul rbx + add r14,rax + mov rax,QWORD[56+rbp] + adc rdx,0 + add r13,r14 + mov r14,rdx + adc r14,0 + + mul rbx + mov rbx,rsi + add r15,rax + mov rax,QWORD[rbp] + adc rdx,0 + add r14,r15 + mov r15,rdx + adc r15,0 + + dec ecx + jne NEAR $L$reduction_loop + + DB 0F3h,0C3h ;repret + + + +ALIGN 32 +__rsaz_512_reducex: + + + imul rdx,r8 + xor rsi,rsi + mov ecx,8 + jmp NEAR $L$reduction_loopx + +ALIGN 32 +$L$reduction_loopx: + mov rbx,r8 + mulx r8,rax,QWORD[rbp] + adcx rax,rbx + adox r8,r9 + + mulx r9,rax,QWORD[8+rbp] + adcx r8,rax + adox r9,r10 + + mulx r10,rbx,QWORD[16+rbp] + adcx r9,rbx + adox r10,r11 + + mulx r11,rbx,QWORD[24+rbp] + adcx r10,rbx + adox r11,r12 + +DB 0xc4,0x62,0xe3,0xf6,0xa5,0x20,0x00,0x00,0x00 + mov rax,rdx + mov rdx,r8 + adcx r11,rbx + adox r12,r13 + + mulx rdx,rbx,QWORD[((128+8))+rsp] + mov rdx,rax + + mulx r13,rax,QWORD[40+rbp] + adcx r12,rax + adox r13,r14 + +DB 0xc4,0x62,0xfb,0xf6,0xb5,0x30,0x00,0x00,0x00 + adcx r13,rax + adox r14,r15 + + mulx r15,rax,QWORD[56+rbp] + mov rdx,rbx + adcx r14,rax + adox r15,rsi + adcx r15,rsi + + dec ecx + jne NEAR $L$reduction_loopx + + DB 0F3h,0C3h ;repret + + + +ALIGN 32 +__rsaz_512_subtract: + + mov QWORD[rdi],r8 + mov QWORD[8+rdi],r9 + mov QWORD[16+rdi],r10 + mov QWORD[24+rdi],r11 + mov QWORD[32+rdi],r12 + mov QWORD[40+rdi],r13 + mov QWORD[48+rdi],r14 + mov QWORD[56+rdi],r15 + + mov r8,QWORD[rbp] + mov r9,QWORD[8+rbp] + neg r8 + not r9 + and r8,rcx + mov r10,QWORD[16+rbp] + and r9,rcx + not r10 + mov r11,QWORD[24+rbp] + and r10,rcx + not r11 + mov r12,QWORD[32+rbp] + and r11,rcx + not r12 + mov r13,QWORD[40+rbp] + and r12,rcx + not r13 + mov r14,QWORD[48+rbp] + and r13,rcx + not r14 + mov r15,QWORD[56+rbp] + and r14,rcx + not r15 + and r15,rcx + + add r8,QWORD[rdi] + adc r9,QWORD[8+rdi] + adc r10,QWORD[16+rdi] + adc r11,QWORD[24+rdi] + adc r12,QWORD[32+rdi] + adc r13,QWORD[40+rdi] + adc r14,QWORD[48+rdi] + adc r15,QWORD[56+rdi] + + mov QWORD[rdi],r8 + mov QWORD[8+rdi],r9 + mov QWORD[16+rdi],r10 + mov QWORD[24+rdi],r11 + mov QWORD[32+rdi],r12 + mov QWORD[40+rdi],r13 + mov QWORD[48+rdi],r14 + mov QWORD[56+rdi],r15 + + DB 0F3h,0C3h ;repret + + + +ALIGN 32 +__rsaz_512_mul: + + lea rdi,[8+rsp] + + mov rax,QWORD[rsi] + mul rbx + mov QWORD[rdi],rax + mov rax,QWORD[8+rsi] + mov r8,rdx + + mul rbx + add r8,rax + mov rax,QWORD[16+rsi] + mov r9,rdx + adc r9,0 + + mul rbx + add r9,rax + mov rax,QWORD[24+rsi] + mov r10,rdx + adc r10,0 + + mul rbx + add r10,rax + mov rax,QWORD[32+rsi] + mov r11,rdx + adc r11,0 + + mul rbx + add r11,rax + mov rax,QWORD[40+rsi] + mov r12,rdx + adc r12,0 + + mul rbx + add r12,rax + mov rax,QWORD[48+rsi] + mov r13,rdx + adc r13,0 + + mul rbx + add r13,rax + mov rax,QWORD[56+rsi] + mov r14,rdx + adc r14,0 + + mul rbx + add r14,rax + mov rax,QWORD[rsi] + mov r15,rdx + adc r15,0 + + lea rbp,[8+rbp] + lea rdi,[8+rdi] + + mov ecx,7 + jmp NEAR $L$oop_mul + +ALIGN 32 +$L$oop_mul: + mov rbx,QWORD[rbp] + mul rbx + add r8,rax + mov rax,QWORD[8+rsi] + mov QWORD[rdi],r8 + mov r8,rdx + adc r8,0 + + mul rbx + add r9,rax + mov rax,QWORD[16+rsi] + adc rdx,0 + add r8,r9 + mov r9,rdx + adc r9,0 + + mul rbx + add r10,rax + mov rax,QWORD[24+rsi] + adc rdx,0 + add r9,r10 + mov r10,rdx + adc r10,0 + + mul rbx + add r11,rax + mov rax,QWORD[32+rsi] + adc rdx,0 + add r10,r11 + mov r11,rdx + adc r11,0 + + mul rbx + add r12,rax + mov rax,QWORD[40+rsi] + adc rdx,0 + add r11,r12 + mov r12,rdx + adc r12,0 + + mul rbx + add r13,rax + mov rax,QWORD[48+rsi] + adc rdx,0 + add r12,r13 + mov r13,rdx + adc r13,0 + + mul rbx + add r14,rax + mov rax,QWORD[56+rsi] + adc rdx,0 + add r13,r14 + mov r14,rdx + lea rbp,[8+rbp] + adc r14,0 + + mul rbx + add r15,rax + mov rax,QWORD[rsi] + adc rdx,0 + add r14,r15 + mov r15,rdx + adc r15,0 + + lea rdi,[8+rdi] + + dec ecx + jnz NEAR $L$oop_mul + + mov QWORD[rdi],r8 + mov QWORD[8+rdi],r9 + mov QWORD[16+rdi],r10 + mov QWORD[24+rdi],r11 + mov QWORD[32+rdi],r12 + mov QWORD[40+rdi],r13 + mov QWORD[48+rdi],r14 + mov QWORD[56+rdi],r15 + + DB 0F3h,0C3h ;repret + + + +ALIGN 32 +__rsaz_512_mulx: + + mulx r8,rbx,QWORD[rsi] + mov rcx,-6 + + mulx r9,rax,QWORD[8+rsi] + mov QWORD[8+rsp],rbx + + mulx r10,rbx,QWORD[16+rsi] + adc r8,rax + + mulx r11,rax,QWORD[24+rsi] + adc r9,rbx + + mulx r12,rbx,QWORD[32+rsi] + adc r10,rax + + mulx r13,rax,QWORD[40+rsi] + adc r11,rbx + + mulx r14,rbx,QWORD[48+rsi] + adc r12,rax + + mulx r15,rax,QWORD[56+rsi] + mov rdx,QWORD[8+rbp] + adc r13,rbx + adc r14,rax + adc r15,0 + + xor rdi,rdi + jmp NEAR $L$oop_mulx + +ALIGN 32 +$L$oop_mulx: + mov rbx,r8 + mulx r8,rax,QWORD[rsi] + adcx rbx,rax + adox r8,r9 + + mulx r9,rax,QWORD[8+rsi] + adcx r8,rax + adox r9,r10 + + mulx r10,rax,QWORD[16+rsi] + adcx r9,rax + adox r10,r11 + + mulx r11,rax,QWORD[24+rsi] + adcx r10,rax + adox r11,r12 + +DB 0x3e,0xc4,0x62,0xfb,0xf6,0xa6,0x20,0x00,0x00,0x00 + adcx r11,rax + adox r12,r13 + + mulx r13,rax,QWORD[40+rsi] + adcx r12,rax + adox r13,r14 + + mulx r14,rax,QWORD[48+rsi] + adcx r13,rax + adox r14,r15 + + mulx r15,rax,QWORD[56+rsi] + mov rdx,QWORD[64+rcx*8+rbp] + mov QWORD[((8+64-8))+rcx*8+rsp],rbx + adcx r14,rax + adox r15,rdi + adcx r15,rdi + + inc rcx + jnz NEAR $L$oop_mulx + + mov rbx,r8 + mulx r8,rax,QWORD[rsi] + adcx rbx,rax + adox r8,r9 + +DB 0xc4,0x62,0xfb,0xf6,0x8e,0x08,0x00,0x00,0x00 + adcx r8,rax + adox r9,r10 + +DB 0xc4,0x62,0xfb,0xf6,0x96,0x10,0x00,0x00,0x00 + adcx r9,rax + adox r10,r11 + + mulx r11,rax,QWORD[24+rsi] + adcx r10,rax + adox r11,r12 + + mulx r12,rax,QWORD[32+rsi] + adcx r11,rax + adox r12,r13 + + mulx r13,rax,QWORD[40+rsi] + adcx r12,rax + adox r13,r14 + +DB 0xc4,0x62,0xfb,0xf6,0xb6,0x30,0x00,0x00,0x00 + adcx r13,rax + adox r14,r15 + +DB 0xc4,0x62,0xfb,0xf6,0xbe,0x38,0x00,0x00,0x00 + adcx r14,rax + adox r15,rdi + adcx r15,rdi + + mov QWORD[((8+64-8))+rsp],rbx + mov QWORD[((8+64))+rsp],r8 + mov QWORD[((8+64+8))+rsp],r9 + mov QWORD[((8+64+16))+rsp],r10 + mov QWORD[((8+64+24))+rsp],r11 + mov QWORD[((8+64+32))+rsp],r12 + mov QWORD[((8+64+40))+rsp],r13 + mov QWORD[((8+64+48))+rsp],r14 + mov QWORD[((8+64+56))+rsp],r15 + + DB 0F3h,0C3h ;repret + + +global rsaz_512_scatter4 + +ALIGN 16 +rsaz_512_scatter4: + + lea rcx,[r8*8+rcx] + mov r9d,8 + jmp NEAR $L$oop_scatter +ALIGN 16 +$L$oop_scatter: + mov rax,QWORD[rdx] + lea rdx,[8+rdx] + mov QWORD[rcx],rax + lea rcx,[128+rcx] + dec r9d + jnz NEAR $L$oop_scatter + DB 0F3h,0C3h ;repret + + + +global rsaz_512_gather4 + +ALIGN 16 +rsaz_512_gather4: + +$L$SEH_begin_rsaz_512_gather4: +DB 0x48,0x81,0xec,0xa8,0x00,0x00,0x00 +DB 0x0f,0x29,0x34,0x24 +DB 0x0f,0x29,0x7c,0x24,0x10 +DB 0x44,0x0f,0x29,0x44,0x24,0x20 +DB 0x44,0x0f,0x29,0x4c,0x24,0x30 +DB 0x44,0x0f,0x29,0x54,0x24,0x40 +DB 0x44,0x0f,0x29,0x5c,0x24,0x50 +DB 0x44,0x0f,0x29,0x64,0x24,0x60 +DB 0x44,0x0f,0x29,0x6c,0x24,0x70 +DB 0x44,0x0f,0x29,0xb4,0x24,0x80,0,0,0 +DB 0x44,0x0f,0x29,0xbc,0x24,0x90,0,0,0 + movd xmm8,r8d + movdqa xmm1,XMMWORD[(($L$inc+16))] + movdqa xmm0,XMMWORD[$L$inc] + + pshufd xmm8,xmm8,0 + movdqa xmm7,xmm1 + movdqa xmm2,xmm1 + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm8 + movdqa xmm3,xmm7 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm8 + movdqa xmm4,xmm7 + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm8 + movdqa xmm5,xmm7 + paddd xmm4,xmm3 + pcmpeqd xmm3,xmm8 + movdqa xmm6,xmm7 + paddd xmm5,xmm4 + pcmpeqd xmm4,xmm8 + paddd xmm6,xmm5 + pcmpeqd xmm5,xmm8 + paddd xmm7,xmm6 + pcmpeqd xmm6,xmm8 + pcmpeqd xmm7,xmm8 + mov r9d,8 + jmp NEAR $L$oop_gather +ALIGN 16 +$L$oop_gather: + movdqa xmm8,XMMWORD[rdx] + movdqa xmm9,XMMWORD[16+rdx] + movdqa xmm10,XMMWORD[32+rdx] + movdqa xmm11,XMMWORD[48+rdx] + pand xmm8,xmm0 + movdqa xmm12,XMMWORD[64+rdx] + pand xmm9,xmm1 + movdqa xmm13,XMMWORD[80+rdx] + pand xmm10,xmm2 + movdqa xmm14,XMMWORD[96+rdx] + pand xmm11,xmm3 + movdqa xmm15,XMMWORD[112+rdx] + lea rdx,[128+rdx] + pand xmm12,xmm4 + pand xmm13,xmm5 + pand xmm14,xmm6 + pand xmm15,xmm7 + por xmm8,xmm10 + por xmm9,xmm11 + por xmm8,xmm12 + por xmm9,xmm13 + por xmm8,xmm14 + por xmm9,xmm15 + + por xmm8,xmm9 + pshufd xmm9,xmm8,0x4e + por xmm8,xmm9 + movq QWORD[rcx],xmm8 + lea rcx,[8+rcx] + dec r9d + jnz NEAR $L$oop_gather + movaps xmm6,XMMWORD[rsp] + movaps xmm7,XMMWORD[16+rsp] + movaps xmm8,XMMWORD[32+rsp] + movaps xmm9,XMMWORD[48+rsp] + movaps xmm10,XMMWORD[64+rsp] + movaps xmm11,XMMWORD[80+rsp] + movaps xmm12,XMMWORD[96+rsp] + movaps xmm13,XMMWORD[112+rsp] + movaps xmm14,XMMWORD[128+rsp] + movaps xmm15,XMMWORD[144+rsp] + add rsp,0xa8 + DB 0F3h,0C3h ;repret +$L$SEH_end_rsaz_512_gather4: + + + +ALIGN 64 +$L$inc: + DD 0,0,1,1 + DD 2,2,2,2 +EXTERN __imp_RtlVirtualUnwind + +ALIGN 16 +se_handler: + push rsi + push rdi + push rbx + push rbp + push r12 + push r13 + push r14 + push r15 + pushfq + sub rsp,64 + + mov rax,QWORD[120+r8] + mov rbx,QWORD[248+r8] + + mov rsi,QWORD[8+r9] + mov r11,QWORD[56+r9] + + mov r10d,DWORD[r11] + lea r10,[r10*1+rsi] + cmp rbx,r10 + jb NEAR $L$common_seh_tail + + mov rax,QWORD[152+r8] + + mov r10d,DWORD[4+r11] + lea r10,[r10*1+rsi] + cmp rbx,r10 + jae NEAR $L$common_seh_tail + + lea rax,[((128+24+48))+rax] + + lea rbx,[$L$mul_gather4_epilogue] + cmp rbx,r10 + jne NEAR $L$se_not_in_mul_gather4 + + lea rax,[176+rax] + + lea rsi,[((-48-168))+rax] + lea rdi,[512+r8] + mov ecx,20 + DD 0xa548f3fc + +$L$se_not_in_mul_gather4: + mov rbx,QWORD[((-8))+rax] + mov rbp,QWORD[((-16))+rax] + mov r12,QWORD[((-24))+rax] + mov r13,QWORD[((-32))+rax] + mov r14,QWORD[((-40))+rax] + mov r15,QWORD[((-48))+rax] + mov QWORD[144+r8],rbx + mov QWORD[160+r8],rbp + mov QWORD[216+r8],r12 + mov QWORD[224+r8],r13 + mov QWORD[232+r8],r14 + mov QWORD[240+r8],r15 + +$L$common_seh_tail: + mov rdi,QWORD[8+rax] + mov rsi,QWORD[16+rax] + mov QWORD[152+r8],rax + mov QWORD[168+r8],rsi + mov QWORD[176+r8],rdi + + mov rdi,QWORD[40+r9] + mov rsi,r8 + mov ecx,154 + DD 0xa548f3fc + + mov rsi,r9 + xor rcx,rcx + mov rdx,QWORD[8+rsi] + mov r8,QWORD[rsi] + mov r9,QWORD[16+rsi] + mov r10,QWORD[40+rsi] + lea r11,[56+rsi] + lea r12,[24+rsi] + mov QWORD[32+rsp],r10 + mov QWORD[40+rsp],r11 + mov QWORD[48+rsp],r12 + mov QWORD[56+rsp],rcx + call QWORD[__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 + + +section .pdata rdata align=4 +ALIGN 4 + DD $L$SEH_begin_rsaz_512_sqr wrt ..imagebase + DD $L$SEH_end_rsaz_512_sqr wrt ..imagebase + DD $L$SEH_info_rsaz_512_sqr wrt ..imagebase + + DD $L$SEH_begin_rsaz_512_mul wrt ..imagebase + DD $L$SEH_end_rsaz_512_mul wrt ..imagebase + DD $L$SEH_info_rsaz_512_mul wrt ..imagebase + + DD $L$SEH_begin_rsaz_512_mul_gather4 wrt ..imagebase + DD $L$SEH_end_rsaz_512_mul_gather4 wrt ..imagebase + DD $L$SEH_info_rsaz_512_mul_gather4 wrt ..imagebase + + DD $L$SEH_begin_rsaz_512_mul_scatter4 wrt ..imagebase + DD $L$SEH_end_rsaz_512_mul_scatter4 wrt ..imagebase + DD $L$SEH_info_rsaz_512_mul_scatter4 wrt ..imagebase + + DD $L$SEH_begin_rsaz_512_mul_by_one wrt ..imagebase + DD $L$SEH_end_rsaz_512_mul_by_one wrt ..imagebase + DD $L$SEH_info_rsaz_512_mul_by_one wrt ..imagebase + + DD $L$SEH_begin_rsaz_512_gather4 wrt ..imagebase + DD $L$SEH_end_rsaz_512_gather4 wrt ..imagebase + DD $L$SEH_info_rsaz_512_gather4 wrt ..imagebase + +section .xdata rdata align=8 +ALIGN 8 +$L$SEH_info_rsaz_512_sqr: +DB 9,0,0,0 + DD se_handler wrt ..imagebase + DD $L$sqr_body wrt ..imagebase,$L$sqr_epilogue wrt ..imagebase +$L$SEH_info_rsaz_512_mul: +DB 9,0,0,0 + DD se_handler wrt ..imagebase + DD $L$mul_body wrt ..imagebase,$L$mul_epilogue wrt ..imagebase +$L$SEH_info_rsaz_512_mul_gather4: +DB 9,0,0,0 + DD se_handler wrt ..imagebase + DD $L$mul_gather4_body wrt ..imagebase,$L$mul_gather4_epilogue wrt ..imagebase +$L$SEH_info_rsaz_512_mul_scatter4: +DB 9,0,0,0 + DD se_handler wrt ..imagebase + DD $L$mul_scatter4_body wrt ..imagebase,$L$mul_scatter4_epilogue wrt ..imagebase +$L$SEH_info_rsaz_512_mul_by_one: +DB 9,0,0,0 + DD se_handler wrt ..imagebase + DD $L$mul_by_one_body wrt ..imagebase,$L$mul_by_one_epilogue wrt ..imagebase +$L$SEH_info_rsaz_512_gather4: +DB 0x01,0x46,0x16,0x00 +DB 0x46,0xf8,0x09,0x00 +DB 0x3d,0xe8,0x08,0x00 +DB 0x34,0xd8,0x07,0x00 +DB 0x2e,0xc8,0x06,0x00 +DB 0x28,0xb8,0x05,0x00 +DB 0x22,0xa8,0x04,0x00 +DB 0x1c,0x98,0x03,0x00 +DB 0x16,0x88,0x02,0x00 +DB 0x10,0x78,0x01,0x00 +DB 0x0b,0x68,0x00,0x00 +DB 0x07,0x01,0x15,0x00 diff --git a/vere/ext/openssl/gen/windows-x86_64/crypto/bn/x86_64-gf2m.asm b/vere/ext/openssl/gen/windows-x86_64/crypto/bn/x86_64-gf2m.asm new file mode 100644 index 0000000..8123fd1 --- /dev/null +++ b/vere/ext/openssl/gen/windows-x86_64/crypto/bn/x86_64-gf2m.asm @@ -0,0 +1,425 @@ +default rel +%define XMMWORD +%define YMMWORD +%define ZMMWORD +section .text code align=64 + + + +ALIGN 16 +_mul_1x1: + + sub rsp,128+8 + + mov r9,-1 + lea rsi,[rax*1+rax] + shr r9,3 + lea rdi,[rax*4] + and r9,rax + lea r12,[rax*8] + sar rax,63 + lea r10,[r9*1+r9] + sar rsi,63 + lea r11,[r9*4] + and rax,rbp + sar rdi,63 + mov rdx,rax + shl rax,63 + and rsi,rbp + shr rdx,1 + mov rcx,rsi + shl rsi,62 + and rdi,rbp + shr rcx,2 + xor rax,rsi + mov rbx,rdi + shl rdi,61 + xor rdx,rcx + shr rbx,3 + xor rax,rdi + xor rdx,rbx + + mov r13,r9 + mov QWORD[rsp],0 + xor r13,r10 + mov QWORD[8+rsp],r9 + mov r14,r11 + mov QWORD[16+rsp],r10 + xor r14,r12 + mov QWORD[24+rsp],r13 + + xor r9,r11 + mov QWORD[32+rsp],r11 + xor r10,r11 + mov QWORD[40+rsp],r9 + xor r13,r11 + mov QWORD[48+rsp],r10 + xor r9,r14 + mov QWORD[56+rsp],r13 + xor r10,r14 + + mov QWORD[64+rsp],r12 + xor r13,r14 + mov QWORD[72+rsp],r9 + xor r9,r11 + mov QWORD[80+rsp],r10 + xor r10,r11 + mov QWORD[88+rsp],r13 + + xor r13,r11 + mov QWORD[96+rsp],r14 + mov rsi,r8 + mov QWORD[104+rsp],r9 + and rsi,rbp + mov QWORD[112+rsp],r10 + shr rbp,4 + mov QWORD[120+rsp],r13 + mov rdi,r8 + and rdi,rbp + shr rbp,4 + + movq xmm0,QWORD[rsi*8+rsp] + mov rsi,r8 + and rsi,rbp + shr rbp,4 + mov rcx,QWORD[rdi*8+rsp] + mov rdi,r8 + mov rbx,rcx + shl rcx,4 + and rdi,rbp + movq xmm1,QWORD[rsi*8+rsp] + shr rbx,60 + xor rax,rcx + pslldq xmm1,1 + mov rsi,r8 + shr rbp,4 + xor rdx,rbx + and rsi,rbp + shr rbp,4 + pxor xmm0,xmm1 + mov rcx,QWORD[rdi*8+rsp] + mov rdi,r8 + mov rbx,rcx + shl rcx,12 + and rdi,rbp + movq xmm1,QWORD[rsi*8+rsp] + shr rbx,52 + xor rax,rcx + pslldq xmm1,2 + mov rsi,r8 + shr rbp,4 + xor rdx,rbx + and rsi,rbp + shr rbp,4 + pxor xmm0,xmm1 + mov rcx,QWORD[rdi*8+rsp] + mov rdi,r8 + mov rbx,rcx + shl rcx,20 + and rdi,rbp + movq xmm1,QWORD[rsi*8+rsp] + shr rbx,44 + xor rax,rcx + pslldq xmm1,3 + mov rsi,r8 + shr rbp,4 + xor rdx,rbx + and rsi,rbp + shr rbp,4 + pxor xmm0,xmm1 + mov rcx,QWORD[rdi*8+rsp] + mov rdi,r8 + mov rbx,rcx + shl rcx,28 + and rdi,rbp + movq xmm1,QWORD[rsi*8+rsp] + shr rbx,36 + xor rax,rcx + pslldq xmm1,4 + mov rsi,r8 + shr rbp,4 + xor rdx,rbx + and rsi,rbp + shr rbp,4 + pxor xmm0,xmm1 + mov rcx,QWORD[rdi*8+rsp] + mov rdi,r8 + mov rbx,rcx + shl rcx,36 + and rdi,rbp + movq xmm1,QWORD[rsi*8+rsp] + shr rbx,28 + xor rax,rcx + pslldq xmm1,5 + mov rsi,r8 + shr rbp,4 + xor rdx,rbx + and rsi,rbp + shr rbp,4 + pxor xmm0,xmm1 + mov rcx,QWORD[rdi*8+rsp] + mov rdi,r8 + mov rbx,rcx + shl rcx,44 + and rdi,rbp + movq xmm1,QWORD[rsi*8+rsp] + shr rbx,20 + xor rax,rcx + pslldq xmm1,6 + mov rsi,r8 + shr rbp,4 + xor rdx,rbx + and rsi,rbp + shr rbp,4 + pxor xmm0,xmm1 + mov rcx,QWORD[rdi*8+rsp] + mov rdi,r8 + mov rbx,rcx + shl rcx,52 + and rdi,rbp + movq xmm1,QWORD[rsi*8+rsp] + shr rbx,12 + xor rax,rcx + pslldq xmm1,7 + mov rsi,r8 + shr rbp,4 + xor rdx,rbx + and rsi,rbp + shr rbp,4 + pxor xmm0,xmm1 + mov rcx,QWORD[rdi*8+rsp] + mov rbx,rcx + shl rcx,60 +DB 102,72,15,126,198 + shr rbx,4 + xor rax,rcx + psrldq xmm0,8 + xor rdx,rbx +DB 102,72,15,126,199 + xor rax,rsi + xor rdx,rdi + + add rsp,128+8 + + DB 0F3h,0C3h ;repret +$L$end_mul_1x1: + + +EXTERN OPENSSL_ia32cap_P +global bn_GF2m_mul_2x2 + +ALIGN 16 +bn_GF2m_mul_2x2: + + mov rax,rsp + mov r10,QWORD[OPENSSL_ia32cap_P] + bt r10,33 + jnc NEAR $L$vanilla_mul_2x2 + +DB 102,72,15,110,194 +DB 102,73,15,110,201 +DB 102,73,15,110,208 + movq xmm3,QWORD[40+rsp] + movdqa xmm4,xmm0 + movdqa xmm5,xmm1 +DB 102,15,58,68,193,0 + pxor xmm4,xmm2 + pxor xmm5,xmm3 +DB 102,15,58,68,211,0 +DB 102,15,58,68,229,0 + xorps xmm4,xmm0 + xorps xmm4,xmm2 + movdqa xmm5,xmm4 + pslldq xmm4,8 + psrldq xmm5,8 + pxor xmm2,xmm4 + pxor xmm0,xmm5 + movdqu XMMWORD[rcx],xmm2 + movdqu XMMWORD[16+rcx],xmm0 + DB 0F3h,0C3h ;repret + +ALIGN 16 +$L$vanilla_mul_2x2: + lea rsp,[((-136))+rsp] + + mov r10,QWORD[176+rsp] + mov QWORD[120+rsp],rdi + mov QWORD[128+rsp],rsi + mov QWORD[80+rsp],r14 + + mov QWORD[88+rsp],r13 + + mov QWORD[96+rsp],r12 + + mov QWORD[104+rsp],rbp + + mov QWORD[112+rsp],rbx + +$L$body_mul_2x2: + mov QWORD[32+rsp],rcx + mov QWORD[40+rsp],rdx + mov QWORD[48+rsp],r8 + mov QWORD[56+rsp],r9 + mov QWORD[64+rsp],r10 + + mov r8,0xf + mov rax,rdx + mov rbp,r9 + call _mul_1x1 + mov QWORD[16+rsp],rax + mov QWORD[24+rsp],rdx + + mov rax,QWORD[48+rsp] + mov rbp,QWORD[64+rsp] + call _mul_1x1 + mov QWORD[rsp],rax + mov QWORD[8+rsp],rdx + + mov rax,QWORD[40+rsp] + mov rbp,QWORD[56+rsp] + xor rax,QWORD[48+rsp] + xor rbp,QWORD[64+rsp] + call _mul_1x1 + mov rbx,QWORD[rsp] + mov rcx,QWORD[8+rsp] + mov rdi,QWORD[16+rsp] + mov rsi,QWORD[24+rsp] + mov rbp,QWORD[32+rsp] + + xor rax,rdx + xor rdx,rcx + xor rax,rbx + mov QWORD[rbp],rbx + xor rdx,rdi + mov QWORD[24+rbp],rsi + xor rax,rsi + xor rdx,rsi + xor rax,rdx + mov QWORD[16+rbp],rdx + mov QWORD[8+rbp],rax + + mov r14,QWORD[80+rsp] + + mov r13,QWORD[88+rsp] + + mov r12,QWORD[96+rsp] + + mov rbp,QWORD[104+rsp] + + mov rbx,QWORD[112+rsp] + + mov rdi,QWORD[120+rsp] + mov rsi,QWORD[128+rsp] + lea rsp,[136+rsp] + +$L$epilogue_mul_2x2: + DB 0F3h,0C3h ;repret +$L$end_mul_2x2: + + +DB 71,70,40,50,94,109,41,32,77,117,108,116,105,112,108,105 +DB 99,97,116,105,111,110,32,102,111,114,32,120,56,54,95,54 +DB 52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121 +DB 32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46 +DB 111,114,103,62,0 +ALIGN 16 +EXTERN __imp_RtlVirtualUnwind + + +ALIGN 16 +se_handler: + push rsi + push rdi + push rbx + push rbp + push r12 + push r13 + push r14 + push r15 + pushfq + sub rsp,64 + + mov rax,QWORD[120+r8] + mov rbx,QWORD[248+r8] + + lea r10,[$L$body_mul_2x2] + cmp rbx,r10 + jb NEAR $L$in_prologue + + mov rax,QWORD[152+r8] + + lea r10,[$L$epilogue_mul_2x2] + cmp rbx,r10 + jae NEAR $L$in_prologue + + mov r14,QWORD[80+rax] + mov r13,QWORD[88+rax] + mov r12,QWORD[96+rax] + mov rbp,QWORD[104+rax] + mov rbx,QWORD[112+rax] + mov rdi,QWORD[120+rax] + mov rsi,QWORD[128+rax] + + mov QWORD[144+r8],rbx + mov QWORD[160+r8],rbp + mov QWORD[168+r8],rsi + mov QWORD[176+r8],rdi + mov QWORD[216+r8],r12 + mov QWORD[224+r8],r13 + mov QWORD[232+r8],r14 + + lea rax,[136+rax] + +$L$in_prologue: + mov QWORD[152+r8],rax + + mov rdi,QWORD[40+r9] + mov rsi,r8 + mov ecx,154 + DD 0xa548f3fc + + mov rsi,r9 + xor rcx,rcx + mov rdx,QWORD[8+rsi] + mov r8,QWORD[rsi] + mov r9,QWORD[16+rsi] + mov r10,QWORD[40+rsi] + lea r11,[56+rsi] + lea r12,[24+rsi] + mov QWORD[32+rsp],r10 + mov QWORD[40+rsp],r11 + mov QWORD[48+rsp],r12 + mov QWORD[56+rsp],rcx + call QWORD[__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 + + +section .pdata rdata align=4 +ALIGN 4 + DD _mul_1x1 wrt ..imagebase + DD $L$end_mul_1x1 wrt ..imagebase + DD $L$SEH_info_1x1 wrt ..imagebase + + DD $L$vanilla_mul_2x2 wrt ..imagebase + DD $L$end_mul_2x2 wrt ..imagebase + DD $L$SEH_info_2x2 wrt ..imagebase +section .xdata rdata align=8 +ALIGN 8 +$L$SEH_info_1x1: +DB 0x01,0x07,0x02,0x00 +DB 0x07,0x01,0x11,0x00 +$L$SEH_info_2x2: +DB 9,0,0,0 + DD se_handler wrt ..imagebase diff --git a/vere/ext/openssl/gen/windows-x86_64/crypto/bn/x86_64-mont.asm b/vere/ext/openssl/gen/windows-x86_64/crypto/bn/x86_64-mont.asm new file mode 100644 index 0000000..81c2058 --- /dev/null +++ b/vere/ext/openssl/gen/windows-x86_64/crypto/bn/x86_64-mont.asm @@ -0,0 +1,1472 @@ +default rel +%define XMMWORD +%define YMMWORD +%define ZMMWORD +section .text code align=64 + + +EXTERN OPENSSL_ia32cap_P + +global bn_mul_mont + +ALIGN 16 +bn_mul_mont: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[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[40+rsp] + mov r9,QWORD[48+rsp] + + + + mov r9d,r9d + mov rax,rsp + + test r9d,3 + jnz NEAR $L$mul_enter + cmp r9d,8 + jb NEAR $L$mul_enter + mov r11d,DWORD[((OPENSSL_ia32cap_P+8))] + cmp rdx,rsi + jne NEAR $L$mul4x_enter + test r9d,7 + jz NEAR $L$sqr8x_enter + jmp NEAR $L$mul4x_enter + +ALIGN 16 +$L$mul_enter: + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + + + neg r9 + mov r11,rsp + lea r10,[((-16))+r9*8+rsp] + neg r9 + and r10,-1024 + + + + + + + + + + sub r11,r10 + and r11,-4096 + lea rsp,[r11*1+r10] + mov r11,QWORD[rsp] + cmp rsp,r10 + ja NEAR $L$mul_page_walk + jmp NEAR $L$mul_page_walk_done + +ALIGN 16 +$L$mul_page_walk: + lea rsp,[((-4096))+rsp] + mov r11,QWORD[rsp] + cmp rsp,r10 + ja NEAR $L$mul_page_walk +$L$mul_page_walk_done: + + mov QWORD[8+r9*8+rsp],rax + +$L$mul_body: + mov r12,rdx + mov r8,QWORD[r8] + mov rbx,QWORD[r12] + mov rax,QWORD[rsi] + + xor r14,r14 + xor r15,r15 + + mov rbp,r8 + mul rbx + mov r10,rax + mov rax,QWORD[rcx] + + imul rbp,r10 + mov r11,rdx + + mul rbp + add r10,rax + mov rax,QWORD[8+rsi] + adc rdx,0 + mov r13,rdx + + lea r15,[1+r15] + jmp NEAR $L$1st_enter + +ALIGN 16 +$L$1st: + add r13,rax + mov rax,QWORD[r15*8+rsi] + adc rdx,0 + add r13,r11 + mov r11,r10 + adc rdx,0 + mov QWORD[((-16))+r15*8+rsp],r13 + mov r13,rdx + +$L$1st_enter: + mul rbx + add r11,rax + mov rax,QWORD[r15*8+rcx] + adc rdx,0 + lea r15,[1+r15] + mov r10,rdx + + mul rbp + cmp r15,r9 + jne NEAR $L$1st + + add r13,rax + mov rax,QWORD[rsi] + adc rdx,0 + add r13,r11 + adc rdx,0 + mov QWORD[((-16))+r15*8+rsp],r13 + mov r13,rdx + mov r11,r10 + + xor rdx,rdx + add r13,r11 + adc rdx,0 + mov QWORD[((-8))+r9*8+rsp],r13 + mov QWORD[r9*8+rsp],rdx + + lea r14,[1+r14] + jmp NEAR $L$outer +ALIGN 16 +$L$outer: + mov rbx,QWORD[r14*8+r12] + xor r15,r15 + mov rbp,r8 + mov r10,QWORD[rsp] + mul rbx + add r10,rax + mov rax,QWORD[rcx] + adc rdx,0 + + imul rbp,r10 + mov r11,rdx + + mul rbp + add r10,rax + mov rax,QWORD[8+rsi] + adc rdx,0 + mov r10,QWORD[8+rsp] + mov r13,rdx + + lea r15,[1+r15] + jmp NEAR $L$inner_enter + +ALIGN 16 +$L$inner: + add r13,rax + mov rax,QWORD[r15*8+rsi] + adc rdx,0 + add r13,r10 + mov r10,QWORD[r15*8+rsp] + adc rdx,0 + mov QWORD[((-16))+r15*8+rsp],r13 + mov r13,rdx + +$L$inner_enter: + mul rbx + add r11,rax + mov rax,QWORD[r15*8+rcx] + adc rdx,0 + add r10,r11 + mov r11,rdx + adc r11,0 + lea r15,[1+r15] + + mul rbp + cmp r15,r9 + jne NEAR $L$inner + + add r13,rax + mov rax,QWORD[rsi] + adc rdx,0 + add r13,r10 + mov r10,QWORD[r15*8+rsp] + adc rdx,0 + mov QWORD[((-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[((-8))+r9*8+rsp],r13 + mov QWORD[r9*8+rsp],rdx + + lea r14,[1+r14] + cmp r14,r9 + jb NEAR $L$outer + + xor r14,r14 + mov rax,QWORD[rsp] + mov r15,r9 + +ALIGN 16 +$L$sub: sbb rax,QWORD[r14*8+rcx] + mov QWORD[r14*8+rdi],rax + mov rax,QWORD[8+r14*8+rsp] + lea r14,[1+r14] + dec r15 + jnz NEAR $L$sub + + sbb rax,0 + mov rbx,-1 + xor rbx,rax + xor r14,r14 + mov r15,r9 + +$L$copy: + mov rcx,QWORD[r14*8+rdi] + mov rdx,QWORD[r14*8+rsp] + and rcx,rbx + and rdx,rax + mov QWORD[r14*8+rsp],r9 + or rdx,rcx + mov QWORD[r14*8+rdi],rdx + lea r14,[1+r14] + sub r15,1 + jnz NEAR $L$copy + + mov rsi,QWORD[8+r9*8+rsp] + + mov rax,1 + mov r15,QWORD[((-48))+rsi] + + mov r14,QWORD[((-40))+rsi] + + mov r13,QWORD[((-32))+rsi] + + mov r12,QWORD[((-24))+rsi] + + mov rbp,QWORD[((-16))+rsi] + + mov rbx,QWORD[((-8))+rsi] + + lea rsp,[rsi] + +$L$mul_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_bn_mul_mont: + +ALIGN 16 +bn_mul4x_mont: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[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[40+rsp] + mov r9,QWORD[48+rsp] + + + + mov r9d,r9d + mov rax,rsp + +$L$mul4x_enter: + and r11d,0x80100 + cmp r11d,0x80100 + je NEAR $L$mulx4x_enter + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + + + neg r9 + mov r11,rsp + lea r10,[((-32))+r9*8+rsp] + neg r9 + and r10,-1024 + + sub r11,r10 + and r11,-4096 + lea rsp,[r11*1+r10] + mov r11,QWORD[rsp] + cmp rsp,r10 + ja NEAR $L$mul4x_page_walk + jmp NEAR $L$mul4x_page_walk_done + +$L$mul4x_page_walk: + lea rsp,[((-4096))+rsp] + mov r11,QWORD[rsp] + cmp rsp,r10 + ja NEAR $L$mul4x_page_walk +$L$mul4x_page_walk_done: + + mov QWORD[8+r9*8+rsp],rax + +$L$mul4x_body: + mov QWORD[16+r9*8+rsp],rdi + mov r12,rdx + mov r8,QWORD[r8] + mov rbx,QWORD[r12] + mov rax,QWORD[rsi] + + xor r14,r14 + xor r15,r15 + + mov rbp,r8 + mul rbx + mov r10,rax + mov rax,QWORD[rcx] + + imul rbp,r10 + mov r11,rdx + + mul rbp + add r10,rax + mov rax,QWORD[8+rsi] + adc rdx,0 + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[8+rcx] + adc rdx,0 + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[16+rsi] + adc rdx,0 + add rdi,r11 + lea r15,[4+r15] + adc rdx,0 + mov QWORD[rsp],rdi + mov r13,rdx + jmp NEAR $L$1st4x +ALIGN 16 +$L$1st4x: + mul rbx + add r10,rax + mov rax,QWORD[((-16))+r15*8+rcx] + adc rdx,0 + mov r11,rdx + + mul rbp + add r13,rax + mov rax,QWORD[((-8))+r15*8+rsi] + adc rdx,0 + add r13,r10 + adc rdx,0 + mov QWORD[((-24))+r15*8+rsp],r13 + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[((-8))+r15*8+rcx] + adc rdx,0 + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[r15*8+rsi] + adc rdx,0 + add rdi,r11 + adc rdx,0 + mov QWORD[((-16))+r15*8+rsp],rdi + mov r13,rdx + + mul rbx + add r10,rax + mov rax,QWORD[r15*8+rcx] + adc rdx,0 + mov r11,rdx + + mul rbp + add r13,rax + mov rax,QWORD[8+r15*8+rsi] + adc rdx,0 + add r13,r10 + adc rdx,0 + mov QWORD[((-8))+r15*8+rsp],r13 + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[8+r15*8+rcx] + adc rdx,0 + lea r15,[4+r15] + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[((-16))+r15*8+rsi] + adc rdx,0 + add rdi,r11 + adc rdx,0 + mov QWORD[((-32))+r15*8+rsp],rdi + mov r13,rdx + cmp r15,r9 + jb NEAR $L$1st4x + + mul rbx + add r10,rax + mov rax,QWORD[((-16))+r15*8+rcx] + adc rdx,0 + mov r11,rdx + + mul rbp + add r13,rax + mov rax,QWORD[((-8))+r15*8+rsi] + adc rdx,0 + add r13,r10 + adc rdx,0 + mov QWORD[((-24))+r15*8+rsp],r13 + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[((-8))+r15*8+rcx] + adc rdx,0 + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[rsi] + adc rdx,0 + add rdi,r11 + adc rdx,0 + mov QWORD[((-16))+r15*8+rsp],rdi + mov r13,rdx + + xor rdi,rdi + add r13,r10 + adc rdi,0 + mov QWORD[((-8))+r15*8+rsp],r13 + mov QWORD[r15*8+rsp],rdi + + lea r14,[1+r14] +ALIGN 4 +$L$outer4x: + mov rbx,QWORD[r14*8+r12] + xor r15,r15 + mov r10,QWORD[rsp] + mov rbp,r8 + mul rbx + add r10,rax + mov rax,QWORD[rcx] + adc rdx,0 + + imul rbp,r10 + mov r11,rdx + + mul rbp + add r10,rax + mov rax,QWORD[8+rsi] + adc rdx,0 + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[8+rcx] + adc rdx,0 + add r11,QWORD[8+rsp] + adc rdx,0 + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[16+rsi] + adc rdx,0 + add rdi,r11 + lea r15,[4+r15] + adc rdx,0 + mov QWORD[rsp],rdi + mov r13,rdx + jmp NEAR $L$inner4x +ALIGN 16 +$L$inner4x: + mul rbx + add r10,rax + mov rax,QWORD[((-16))+r15*8+rcx] + adc rdx,0 + add r10,QWORD[((-16))+r15*8+rsp] + adc rdx,0 + mov r11,rdx + + mul rbp + add r13,rax + mov rax,QWORD[((-8))+r15*8+rsi] + adc rdx,0 + add r13,r10 + adc rdx,0 + mov QWORD[((-24))+r15*8+rsp],r13 + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[((-8))+r15*8+rcx] + adc rdx,0 + add r11,QWORD[((-8))+r15*8+rsp] + adc rdx,0 + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[r15*8+rsi] + adc rdx,0 + add rdi,r11 + adc rdx,0 + mov QWORD[((-16))+r15*8+rsp],rdi + mov r13,rdx + + mul rbx + add r10,rax + mov rax,QWORD[r15*8+rcx] + adc rdx,0 + add r10,QWORD[r15*8+rsp] + adc rdx,0 + mov r11,rdx + + mul rbp + add r13,rax + mov rax,QWORD[8+r15*8+rsi] + adc rdx,0 + add r13,r10 + adc rdx,0 + mov QWORD[((-8))+r15*8+rsp],r13 + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[8+r15*8+rcx] + adc rdx,0 + add r11,QWORD[8+r15*8+rsp] + adc rdx,0 + lea r15,[4+r15] + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[((-16))+r15*8+rsi] + adc rdx,0 + add rdi,r11 + adc rdx,0 + mov QWORD[((-32))+r15*8+rsp],rdi + mov r13,rdx + cmp r15,r9 + jb NEAR $L$inner4x + + mul rbx + add r10,rax + mov rax,QWORD[((-16))+r15*8+rcx] + adc rdx,0 + add r10,QWORD[((-16))+r15*8+rsp] + adc rdx,0 + mov r11,rdx + + mul rbp + add r13,rax + mov rax,QWORD[((-8))+r15*8+rsi] + adc rdx,0 + add r13,r10 + adc rdx,0 + mov QWORD[((-24))+r15*8+rsp],r13 + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[((-8))+r15*8+rcx] + adc rdx,0 + add r11,QWORD[((-8))+r15*8+rsp] + adc rdx,0 + lea r14,[1+r14] + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[rsi] + adc rdx,0 + add rdi,r11 + adc rdx,0 + mov QWORD[((-16))+r15*8+rsp],rdi + mov r13,rdx + + xor rdi,rdi + add r13,r10 + adc rdi,0 + add r13,QWORD[r9*8+rsp] + adc rdi,0 + mov QWORD[((-8))+r15*8+rsp],r13 + mov QWORD[r15*8+rsp],rdi + + cmp r14,r9 + jb NEAR $L$outer4x + mov rdi,QWORD[16+r9*8+rsp] + lea r15,[((-4))+r9] + mov rax,QWORD[rsp] + mov rdx,QWORD[8+rsp] + shr r15,2 + lea rsi,[rsp] + xor r14,r14 + + sub rax,QWORD[rcx] + mov rbx,QWORD[16+rsi] + mov rbp,QWORD[24+rsi] + sbb rdx,QWORD[8+rcx] + +$L$sub4x: + mov QWORD[r14*8+rdi],rax + mov QWORD[8+r14*8+rdi],rdx + sbb rbx,QWORD[16+r14*8+rcx] + mov rax,QWORD[32+r14*8+rsi] + mov rdx,QWORD[40+r14*8+rsi] + sbb rbp,QWORD[24+r14*8+rcx] + mov QWORD[16+r14*8+rdi],rbx + mov QWORD[24+r14*8+rdi],rbp + sbb rax,QWORD[32+r14*8+rcx] + mov rbx,QWORD[48+r14*8+rsi] + mov rbp,QWORD[56+r14*8+rsi] + sbb rdx,QWORD[40+r14*8+rcx] + lea r14,[4+r14] + dec r15 + jnz NEAR $L$sub4x + + mov QWORD[r14*8+rdi],rax + mov rax,QWORD[32+r14*8+rsi] + sbb rbx,QWORD[16+r14*8+rcx] + mov QWORD[8+r14*8+rdi],rdx + sbb rbp,QWORD[24+r14*8+rcx] + mov QWORD[16+r14*8+rdi],rbx + + sbb rax,0 + mov QWORD[24+r14*8+rdi],rbp + pxor xmm0,xmm0 +DB 102,72,15,110,224 + pcmpeqd xmm5,xmm5 + pshufd xmm4,xmm4,0 + mov r15,r9 + pxor xmm5,xmm4 + shr r15,2 + xor eax,eax + + jmp NEAR $L$copy4x +ALIGN 16 +$L$copy4x: + movdqa xmm1,XMMWORD[rax*1+rsp] + movdqu xmm2,XMMWORD[rax*1+rdi] + pand xmm1,xmm4 + pand xmm2,xmm5 + movdqa xmm3,XMMWORD[16+rax*1+rsp] + movdqa XMMWORD[rax*1+rsp],xmm0 + por xmm1,xmm2 + movdqu xmm2,XMMWORD[16+rax*1+rdi] + movdqu XMMWORD[rax*1+rdi],xmm1 + pand xmm3,xmm4 + pand xmm2,xmm5 + movdqa XMMWORD[16+rax*1+rsp],xmm0 + por xmm3,xmm2 + movdqu XMMWORD[16+rax*1+rdi],xmm3 + lea rax,[32+rax] + dec r15 + jnz NEAR $L$copy4x + mov rsi,QWORD[8+r9*8+rsp] + + mov rax,1 + mov r15,QWORD[((-48))+rsi] + + mov r14,QWORD[((-40))+rsi] + + mov r13,QWORD[((-32))+rsi] + + mov r12,QWORD[((-24))+rsi] + + mov rbp,QWORD[((-16))+rsi] + + mov rbx,QWORD[((-8))+rsi] + + lea rsp,[rsi] + +$L$mul4x_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_bn_mul4x_mont: +EXTERN bn_sqrx8x_internal +EXTERN bn_sqr8x_internal + + +ALIGN 32 +bn_sqr8x_mont: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[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[40+rsp] + mov r9,QWORD[48+rsp] + + + + mov rax,rsp + +$L$sqr8x_enter: + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + +$L$sqr8x_prologue: + + mov r10d,r9d + shl r9d,3 + shl r10,3+2 + neg r9 + + + + + + + lea r11,[((-64))+r9*2+rsp] + mov rbp,rsp + mov r8,QWORD[r8] + sub r11,rsi + and r11,4095 + cmp r10,r11 + jb NEAR $L$sqr8x_sp_alt + sub rbp,r11 + lea rbp,[((-64))+r9*2+rbp] + jmp NEAR $L$sqr8x_sp_done + +ALIGN 32 +$L$sqr8x_sp_alt: + lea r10,[((4096-64))+r9*2] + lea rbp,[((-64))+r9*2+rbp] + sub r11,r10 + mov r10,0 + cmovc r11,r10 + sub rbp,r11 +$L$sqr8x_sp_done: + and rbp,-64 + mov r11,rsp + sub r11,rbp + and r11,-4096 + lea rsp,[rbp*1+r11] + mov r10,QWORD[rsp] + cmp rsp,rbp + ja NEAR $L$sqr8x_page_walk + jmp NEAR $L$sqr8x_page_walk_done + +ALIGN 16 +$L$sqr8x_page_walk: + lea rsp,[((-4096))+rsp] + mov r10,QWORD[rsp] + cmp rsp,rbp + ja NEAR $L$sqr8x_page_walk +$L$sqr8x_page_walk_done: + + mov r10,r9 + neg r9 + + mov QWORD[32+rsp],r8 + mov QWORD[40+rsp],rax + +$L$sqr8x_body: + +DB 102,72,15,110,209 + pxor xmm0,xmm0 +DB 102,72,15,110,207 +DB 102,73,15,110,218 + mov eax,DWORD[((OPENSSL_ia32cap_P+8))] + and eax,0x80100 + cmp eax,0x80100 + jne NEAR $L$sqr8x_nox + + call bn_sqrx8x_internal + + + + + lea rbx,[rcx*1+r8] + mov r9,rcx + mov rdx,rcx +DB 102,72,15,126,207 + sar rcx,3+2 + jmp NEAR $L$sqr8x_sub + +ALIGN 32 +$L$sqr8x_nox: + call bn_sqr8x_internal + + + + + lea rbx,[r9*1+rdi] + mov rcx,r9 + mov rdx,r9 +DB 102,72,15,126,207 + sar rcx,3+2 + jmp NEAR $L$sqr8x_sub + +ALIGN 32 +$L$sqr8x_sub: + mov r12,QWORD[rbx] + mov r13,QWORD[8+rbx] + mov r14,QWORD[16+rbx] + mov r15,QWORD[24+rbx] + lea rbx,[32+rbx] + sbb r12,QWORD[rbp] + sbb r13,QWORD[8+rbp] + sbb r14,QWORD[16+rbp] + sbb r15,QWORD[24+rbp] + lea rbp,[32+rbp] + mov QWORD[rdi],r12 + mov QWORD[8+rdi],r13 + mov QWORD[16+rdi],r14 + mov QWORD[24+rdi],r15 + lea rdi,[32+rdi] + inc rcx + jnz NEAR $L$sqr8x_sub + + sbb rax,0 + lea rbx,[r9*1+rbx] + lea rdi,[r9*1+rdi] + +DB 102,72,15,110,200 + pxor xmm0,xmm0 + pshufd xmm1,xmm1,0 + mov rsi,QWORD[40+rsp] + + jmp NEAR $L$sqr8x_cond_copy + +ALIGN 32 +$L$sqr8x_cond_copy: + movdqa xmm2,XMMWORD[rbx] + movdqa xmm3,XMMWORD[16+rbx] + lea rbx,[32+rbx] + movdqu xmm4,XMMWORD[rdi] + movdqu xmm5,XMMWORD[16+rdi] + lea rdi,[32+rdi] + movdqa XMMWORD[(-32)+rbx],xmm0 + movdqa XMMWORD[(-16)+rbx],xmm0 + movdqa XMMWORD[(-32)+rdx*1+rbx],xmm0 + movdqa XMMWORD[(-16)+rdx*1+rbx],xmm0 + pcmpeqd xmm0,xmm1 + pand xmm2,xmm1 + pand xmm3,xmm1 + pand xmm4,xmm0 + pand xmm5,xmm0 + pxor xmm0,xmm0 + por xmm4,xmm2 + por xmm5,xmm3 + movdqu XMMWORD[(-32)+rdi],xmm4 + movdqu XMMWORD[(-16)+rdi],xmm5 + add r9,32 + jnz NEAR $L$sqr8x_cond_copy + + mov rax,1 + mov r15,QWORD[((-48))+rsi] + + mov r14,QWORD[((-40))+rsi] + + mov r13,QWORD[((-32))+rsi] + + mov r12,QWORD[((-24))+rsi] + + mov rbp,QWORD[((-16))+rsi] + + mov rbx,QWORD[((-8))+rsi] + + lea rsp,[rsi] + +$L$sqr8x_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_bn_sqr8x_mont: + +ALIGN 32 +bn_mulx4x_mont: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[16+rsp],rsi + mov rax,rsp +$L$SEH_begin_bn_mulx4x_mont: + mov rdi,rcx + mov rsi,rdx + mov rdx,r8 + mov rcx,r9 + mov r8,QWORD[40+rsp] + mov r9,QWORD[48+rsp] + + + + mov rax,rsp + +$L$mulx4x_enter: + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + +$L$mulx4x_prologue: + + shl r9d,3 + xor r10,r10 + sub r10,r9 + mov r8,QWORD[r8] + lea rbp,[((-72))+r10*1+rsp] + and rbp,-128 + mov r11,rsp + sub r11,rbp + and r11,-4096 + lea rsp,[rbp*1+r11] + mov r10,QWORD[rsp] + cmp rsp,rbp + ja NEAR $L$mulx4x_page_walk + jmp NEAR $L$mulx4x_page_walk_done + +ALIGN 16 +$L$mulx4x_page_walk: + lea rsp,[((-4096))+rsp] + mov r10,QWORD[rsp] + cmp rsp,rbp + ja NEAR $L$mulx4x_page_walk +$L$mulx4x_page_walk_done: + + lea r10,[r9*1+rdx] + + + + + + + + + + + + + mov QWORD[rsp],r9 + shr r9,5 + mov QWORD[16+rsp],r10 + sub r9,1 + mov QWORD[24+rsp],r8 + mov QWORD[32+rsp],rdi + mov QWORD[40+rsp],rax + + mov QWORD[48+rsp],r9 + jmp NEAR $L$mulx4x_body + +ALIGN 32 +$L$mulx4x_body: + lea rdi,[8+rdx] + mov rdx,QWORD[rdx] + lea rbx,[((64+32))+rsp] + mov r9,rdx + + mulx rax,r8,QWORD[rsi] + mulx r14,r11,QWORD[8+rsi] + add r11,rax + mov QWORD[8+rsp],rdi + mulx r13,r12,QWORD[16+rsi] + adc r12,r14 + adc r13,0 + + mov rdi,r8 + imul r8,QWORD[24+rsp] + xor rbp,rbp + + mulx r14,rax,QWORD[24+rsi] + mov rdx,r8 + lea rsi,[32+rsi] + adcx r13,rax + adcx r14,rbp + + mulx r10,rax,QWORD[rcx] + adcx rdi,rax + adox r10,r11 + mulx r11,rax,QWORD[8+rcx] + adcx r10,rax + adox r11,r12 +DB 0xc4,0x62,0xfb,0xf6,0xa1,0x10,0x00,0x00,0x00 + mov rdi,QWORD[48+rsp] + mov QWORD[((-32))+rbx],r10 + adcx r11,rax + adox r12,r13 + mulx r15,rax,QWORD[24+rcx] + mov rdx,r9 + mov QWORD[((-24))+rbx],r11 + adcx r12,rax + adox r15,rbp + lea rcx,[32+rcx] + mov QWORD[((-16))+rbx],r12 + + jmp NEAR $L$mulx4x_1st + +ALIGN 32 +$L$mulx4x_1st: + adcx r15,rbp + mulx rax,r10,QWORD[rsi] + adcx r10,r14 + mulx r14,r11,QWORD[8+rsi] + adcx r11,rax + mulx rax,r12,QWORD[16+rsi] + adcx r12,r14 + mulx r14,r13,QWORD[24+rsi] +DB 0x67,0x67 + mov rdx,r8 + adcx r13,rax + adcx r14,rbp + lea rsi,[32+rsi] + lea rbx,[32+rbx] + + adox r10,r15 + mulx r15,rax,QWORD[rcx] + adcx r10,rax + adox r11,r15 + mulx r15,rax,QWORD[8+rcx] + adcx r11,rax + adox r12,r15 + mulx r15,rax,QWORD[16+rcx] + mov QWORD[((-40))+rbx],r10 + adcx r12,rax + mov QWORD[((-32))+rbx],r11 + adox r13,r15 + mulx r15,rax,QWORD[24+rcx] + mov rdx,r9 + mov QWORD[((-24))+rbx],r12 + adcx r13,rax + adox r15,rbp + lea rcx,[32+rcx] + mov QWORD[((-16))+rbx],r13 + + dec rdi + jnz NEAR $L$mulx4x_1st + + mov rax,QWORD[rsp] + mov rdi,QWORD[8+rsp] + adc r15,rbp + add r14,r15 + sbb r15,r15 + mov QWORD[((-8))+rbx],r14 + jmp NEAR $L$mulx4x_outer + +ALIGN 32 +$L$mulx4x_outer: + mov rdx,QWORD[rdi] + lea rdi,[8+rdi] + sub rsi,rax + mov QWORD[rbx],r15 + lea rbx,[((64+32))+rsp] + sub rcx,rax + + mulx r11,r8,QWORD[rsi] + xor ebp,ebp + mov r9,rdx + mulx r12,r14,QWORD[8+rsi] + adox r8,QWORD[((-32))+rbx] + adcx r11,r14 + mulx r13,r15,QWORD[16+rsi] + adox r11,QWORD[((-24))+rbx] + adcx r12,r15 + adox r12,QWORD[((-16))+rbx] + adcx r13,rbp + adox r13,rbp + + mov QWORD[8+rsp],rdi + mov r15,r8 + imul r8,QWORD[24+rsp] + xor ebp,ebp + + mulx r14,rax,QWORD[24+rsi] + mov rdx,r8 + adcx r13,rax + adox r13,QWORD[((-8))+rbx] + adcx r14,rbp + lea rsi,[32+rsi] + adox r14,rbp + + mulx r10,rax,QWORD[rcx] + adcx r15,rax + adox r10,r11 + mulx r11,rax,QWORD[8+rcx] + adcx r10,rax + adox r11,r12 + mulx r12,rax,QWORD[16+rcx] + mov QWORD[((-32))+rbx],r10 + adcx r11,rax + adox r12,r13 + mulx r15,rax,QWORD[24+rcx] + mov rdx,r9 + mov QWORD[((-24))+rbx],r11 + lea rcx,[32+rcx] + adcx r12,rax + adox r15,rbp + mov rdi,QWORD[48+rsp] + mov QWORD[((-16))+rbx],r12 + + jmp NEAR $L$mulx4x_inner + +ALIGN 32 +$L$mulx4x_inner: + mulx rax,r10,QWORD[rsi] + adcx r15,rbp + adox r10,r14 + mulx r14,r11,QWORD[8+rsi] + adcx r10,QWORD[rbx] + adox r11,rax + mulx rax,r12,QWORD[16+rsi] + adcx r11,QWORD[8+rbx] + adox r12,r14 + mulx r14,r13,QWORD[24+rsi] + mov rdx,r8 + adcx r12,QWORD[16+rbx] + adox r13,rax + adcx r13,QWORD[24+rbx] + adox r14,rbp + lea rsi,[32+rsi] + lea rbx,[32+rbx] + adcx r14,rbp + + adox r10,r15 + mulx r15,rax,QWORD[rcx] + adcx r10,rax + adox r11,r15 + mulx r15,rax,QWORD[8+rcx] + adcx r11,rax + adox r12,r15 + mulx r15,rax,QWORD[16+rcx] + mov QWORD[((-40))+rbx],r10 + adcx r12,rax + adox r13,r15 + mulx r15,rax,QWORD[24+rcx] + mov rdx,r9 + mov QWORD[((-32))+rbx],r11 + mov QWORD[((-24))+rbx],r12 + adcx r13,rax + adox r15,rbp + lea rcx,[32+rcx] + mov QWORD[((-16))+rbx],r13 + + dec rdi + jnz NEAR $L$mulx4x_inner + + mov rax,QWORD[rsp] + mov rdi,QWORD[8+rsp] + adc r15,rbp + sub rbp,QWORD[rbx] + adc r14,r15 + sbb r15,r15 + mov QWORD[((-8))+rbx],r14 + + cmp rdi,QWORD[16+rsp] + jne NEAR $L$mulx4x_outer + + lea rbx,[64+rsp] + sub rcx,rax + neg r15 + mov rdx,rax + shr rax,3+2 + mov rdi,QWORD[32+rsp] + jmp NEAR $L$mulx4x_sub + +ALIGN 32 +$L$mulx4x_sub: + mov r11,QWORD[rbx] + mov r12,QWORD[8+rbx] + mov r13,QWORD[16+rbx] + mov r14,QWORD[24+rbx] + lea rbx,[32+rbx] + sbb r11,QWORD[rcx] + sbb r12,QWORD[8+rcx] + sbb r13,QWORD[16+rcx] + sbb r14,QWORD[24+rcx] + lea rcx,[32+rcx] + mov QWORD[rdi],r11 + mov QWORD[8+rdi],r12 + mov QWORD[16+rdi],r13 + mov QWORD[24+rdi],r14 + lea rdi,[32+rdi] + dec rax + jnz NEAR $L$mulx4x_sub + + sbb r15,0 + lea rbx,[64+rsp] + sub rdi,rdx + +DB 102,73,15,110,207 + pxor xmm0,xmm0 + pshufd xmm1,xmm1,0 + mov rsi,QWORD[40+rsp] + + jmp NEAR $L$mulx4x_cond_copy + +ALIGN 32 +$L$mulx4x_cond_copy: + movdqa xmm2,XMMWORD[rbx] + movdqa xmm3,XMMWORD[16+rbx] + lea rbx,[32+rbx] + movdqu xmm4,XMMWORD[rdi] + movdqu xmm5,XMMWORD[16+rdi] + lea rdi,[32+rdi] + movdqa XMMWORD[(-32)+rbx],xmm0 + movdqa XMMWORD[(-16)+rbx],xmm0 + pcmpeqd xmm0,xmm1 + pand xmm2,xmm1 + pand xmm3,xmm1 + pand xmm4,xmm0 + pand xmm5,xmm0 + pxor xmm0,xmm0 + por xmm4,xmm2 + por xmm5,xmm3 + movdqu XMMWORD[(-32)+rdi],xmm4 + movdqu XMMWORD[(-16)+rdi],xmm5 + sub rdx,32 + jnz NEAR $L$mulx4x_cond_copy + + mov QWORD[rbx],rdx + + mov rax,1 + mov r15,QWORD[((-48))+rsi] + + mov r14,QWORD[((-40))+rsi] + + mov r13,QWORD[((-32))+rsi] + + mov r12,QWORD[((-24))+rsi] + + mov rbp,QWORD[((-16))+rsi] + + mov rbx,QWORD[((-8))+rsi] + + lea rsp,[rsi] + +$L$mulx4x_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_bn_mulx4x_mont: +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 + +ALIGN 16 +mul_handler: + push rsi + push rdi + push rbx + push rbp + push r12 + push r13 + push r14 + push r15 + pushfq + sub rsp,64 + + mov rax,QWORD[120+r8] + mov rbx,QWORD[248+r8] + + mov rsi,QWORD[8+r9] + mov r11,QWORD[56+r9] + + mov r10d,DWORD[r11] + lea r10,[r10*1+rsi] + cmp rbx,r10 + jb NEAR $L$common_seh_tail + + mov rax,QWORD[152+r8] + + mov r10d,DWORD[4+r11] + lea r10,[r10*1+rsi] + cmp rbx,r10 + jae NEAR $L$common_seh_tail + + mov r10,QWORD[192+r8] + mov rax,QWORD[8+r10*8+rax] + + jmp NEAR $L$common_pop_regs + + + +ALIGN 16 +sqr_handler: + push rsi + push rdi + push rbx + push rbp + push r12 + push r13 + push r14 + push r15 + pushfq + sub rsp,64 + + mov rax,QWORD[120+r8] + mov rbx,QWORD[248+r8] + + mov rsi,QWORD[8+r9] + mov r11,QWORD[56+r9] + + mov r10d,DWORD[r11] + lea r10,[r10*1+rsi] + cmp rbx,r10 + jb NEAR $L$common_seh_tail + + mov r10d,DWORD[4+r11] + lea r10,[r10*1+rsi] + cmp rbx,r10 + jb NEAR $L$common_pop_regs + + mov rax,QWORD[152+r8] + + mov r10d,DWORD[8+r11] + lea r10,[r10*1+rsi] + cmp rbx,r10 + jae NEAR $L$common_seh_tail + + mov rax,QWORD[40+rax] + +$L$common_pop_regs: + mov rbx,QWORD[((-8))+rax] + mov rbp,QWORD[((-16))+rax] + mov r12,QWORD[((-24))+rax] + mov r13,QWORD[((-32))+rax] + mov r14,QWORD[((-40))+rax] + mov r15,QWORD[((-48))+rax] + mov QWORD[144+r8],rbx + mov QWORD[160+r8],rbp + mov QWORD[216+r8],r12 + mov QWORD[224+r8],r13 + mov QWORD[232+r8],r14 + mov QWORD[240+r8],r15 + +$L$common_seh_tail: + mov rdi,QWORD[8+rax] + mov rsi,QWORD[16+rax] + mov QWORD[152+r8],rax + mov QWORD[168+r8],rsi + mov QWORD[176+r8],rdi + + mov rdi,QWORD[40+r9] + mov rsi,r8 + mov ecx,154 + DD 0xa548f3fc + + mov rsi,r9 + xor rcx,rcx + mov rdx,QWORD[8+rsi] + mov r8,QWORD[rsi] + mov r9,QWORD[16+rsi] + mov r10,QWORD[40+rsi] + lea r11,[56+rsi] + lea r12,[24+rsi] + mov QWORD[32+rsp],r10 + mov QWORD[40+rsp],r11 + mov QWORD[48+rsp],r12 + mov QWORD[56+rsp],rcx + call QWORD[__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 + + +section .pdata rdata align=4 +ALIGN 4 + DD $L$SEH_begin_bn_mul_mont wrt ..imagebase + DD $L$SEH_end_bn_mul_mont wrt ..imagebase + DD $L$SEH_info_bn_mul_mont wrt ..imagebase + + DD $L$SEH_begin_bn_mul4x_mont wrt ..imagebase + DD $L$SEH_end_bn_mul4x_mont wrt ..imagebase + DD $L$SEH_info_bn_mul4x_mont wrt ..imagebase + + DD $L$SEH_begin_bn_sqr8x_mont wrt ..imagebase + DD $L$SEH_end_bn_sqr8x_mont wrt ..imagebase + DD $L$SEH_info_bn_sqr8x_mont wrt ..imagebase + DD $L$SEH_begin_bn_mulx4x_mont wrt ..imagebase + DD $L$SEH_end_bn_mulx4x_mont wrt ..imagebase + DD $L$SEH_info_bn_mulx4x_mont wrt ..imagebase +section .xdata rdata align=8 +ALIGN 8 +$L$SEH_info_bn_mul_mont: +DB 9,0,0,0 + DD mul_handler wrt ..imagebase + DD $L$mul_body wrt ..imagebase,$L$mul_epilogue wrt ..imagebase +$L$SEH_info_bn_mul4x_mont: +DB 9,0,0,0 + DD mul_handler wrt ..imagebase + DD $L$mul4x_body wrt ..imagebase,$L$mul4x_epilogue wrt ..imagebase +$L$SEH_info_bn_sqr8x_mont: +DB 9,0,0,0 + DD sqr_handler wrt ..imagebase + DD $L$sqr8x_prologue wrt ..imagebase,$L$sqr8x_body wrt ..imagebase,$L$sqr8x_epilogue wrt ..imagebase +ALIGN 8 +$L$SEH_info_bn_mulx4x_mont: +DB 9,0,0,0 + DD sqr_handler wrt ..imagebase + DD $L$mulx4x_prologue wrt ..imagebase,$L$mulx4x_body wrt ..imagebase,$L$mulx4x_epilogue wrt ..imagebase +ALIGN 8 diff --git a/vere/ext/openssl/gen/windows-x86_64/crypto/bn/x86_64-mont5.asm b/vere/ext/openssl/gen/windows-x86_64/crypto/bn/x86_64-mont5.asm new file mode 100644 index 0000000..1186764 --- /dev/null +++ b/vere/ext/openssl/gen/windows-x86_64/crypto/bn/x86_64-mont5.asm @@ -0,0 +1,3842 @@ +default rel +%define XMMWORD +%define YMMWORD +%define ZMMWORD +section .text code align=64 + + +EXTERN OPENSSL_ia32cap_P + +global bn_mul_mont_gather5 + +ALIGN 64 +bn_mul_mont_gather5: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[16+rsp],rsi + mov rax,rsp +$L$SEH_begin_bn_mul_mont_gather5: + mov rdi,rcx + mov rsi,rdx + mov rdx,r8 + mov rcx,r9 + mov r8,QWORD[40+rsp] + mov r9,QWORD[48+rsp] + + + + mov r9d,r9d + mov rax,rsp + + test r9d,7 + jnz NEAR $L$mul_enter + mov r11d,DWORD[((OPENSSL_ia32cap_P+8))] + jmp NEAR $L$mul4x_enter + +ALIGN 16 +$L$mul_enter: + movd xmm5,DWORD[56+rsp] + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + + + neg r9 + mov r11,rsp + lea r10,[((-280))+r9*8+rsp] + neg r9 + and r10,-1024 + + + + + + + + + + sub r11,r10 + and r11,-4096 + lea rsp,[r11*1+r10] + mov r11,QWORD[rsp] + cmp rsp,r10 + ja NEAR $L$mul_page_walk + jmp NEAR $L$mul_page_walk_done + +$L$mul_page_walk: + lea rsp,[((-4096))+rsp] + mov r11,QWORD[rsp] + cmp rsp,r10 + ja NEAR $L$mul_page_walk +$L$mul_page_walk_done: + + lea r10,[$L$inc] + mov QWORD[8+r9*8+rsp],rax + +$L$mul_body: + + lea r12,[128+rdx] + movdqa xmm0,XMMWORD[r10] + movdqa xmm1,XMMWORD[16+r10] + lea r10,[((24-112))+r9*8+rsp] + and r10,-16 + + pshufd xmm5,xmm5,0 + movdqa xmm4,xmm1 + movdqa xmm2,xmm1 + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 +DB 0x67 + movdqa xmm3,xmm4 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[112+r10],xmm0 + movdqa xmm0,xmm4 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[128+r10],xmm1 + movdqa xmm1,xmm4 + + paddd xmm0,xmm3 + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[144+r10],xmm2 + movdqa xmm2,xmm4 + + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa XMMWORD[160+r10],xmm3 + movdqa xmm3,xmm4 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[176+r10],xmm0 + movdqa xmm0,xmm4 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[192+r10],xmm1 + movdqa xmm1,xmm4 + + paddd xmm0,xmm3 + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[208+r10],xmm2 + movdqa xmm2,xmm4 + + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa XMMWORD[224+r10],xmm3 + movdqa xmm3,xmm4 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[240+r10],xmm0 + movdqa xmm0,xmm4 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[256+r10],xmm1 + movdqa xmm1,xmm4 + + paddd xmm0,xmm3 + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[272+r10],xmm2 + movdqa xmm2,xmm4 + + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa XMMWORD[288+r10],xmm3 + movdqa xmm3,xmm4 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[304+r10],xmm0 + + paddd xmm3,xmm2 +DB 0x67 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[320+r10],xmm1 + + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[336+r10],xmm2 + pand xmm0,XMMWORD[64+r12] + + pand xmm1,XMMWORD[80+r12] + pand xmm2,XMMWORD[96+r12] + movdqa XMMWORD[352+r10],xmm3 + pand xmm3,XMMWORD[112+r12] + por xmm0,xmm2 + por xmm1,xmm3 + movdqa xmm4,XMMWORD[((-128))+r12] + movdqa xmm5,XMMWORD[((-112))+r12] + movdqa xmm2,XMMWORD[((-96))+r12] + pand xmm4,XMMWORD[112+r10] + movdqa xmm3,XMMWORD[((-80))+r12] + pand xmm5,XMMWORD[128+r10] + por xmm0,xmm4 + pand xmm2,XMMWORD[144+r10] + por xmm1,xmm5 + pand xmm3,XMMWORD[160+r10] + por xmm0,xmm2 + por xmm1,xmm3 + movdqa xmm4,XMMWORD[((-64))+r12] + movdqa xmm5,XMMWORD[((-48))+r12] + movdqa xmm2,XMMWORD[((-32))+r12] + pand xmm4,XMMWORD[176+r10] + movdqa xmm3,XMMWORD[((-16))+r12] + pand xmm5,XMMWORD[192+r10] + por xmm0,xmm4 + pand xmm2,XMMWORD[208+r10] + por xmm1,xmm5 + pand xmm3,XMMWORD[224+r10] + por xmm0,xmm2 + por xmm1,xmm3 + movdqa xmm4,XMMWORD[r12] + movdqa xmm5,XMMWORD[16+r12] + movdqa xmm2,XMMWORD[32+r12] + pand xmm4,XMMWORD[240+r10] + movdqa xmm3,XMMWORD[48+r12] + pand xmm5,XMMWORD[256+r10] + por xmm0,xmm4 + pand xmm2,XMMWORD[272+r10] + por xmm1,xmm5 + pand xmm3,XMMWORD[288+r10] + por xmm0,xmm2 + por xmm1,xmm3 + por xmm0,xmm1 + pshufd xmm1,xmm0,0x4e + por xmm0,xmm1 + lea r12,[256+r12] +DB 102,72,15,126,195 + + mov r8,QWORD[r8] + mov rax,QWORD[rsi] + + xor r14,r14 + xor r15,r15 + + mov rbp,r8 + mul rbx + mov r10,rax + mov rax,QWORD[rcx] + + imul rbp,r10 + mov r11,rdx + + mul rbp + add r10,rax + mov rax,QWORD[8+rsi] + adc rdx,0 + mov r13,rdx + + lea r15,[1+r15] + jmp NEAR $L$1st_enter + +ALIGN 16 +$L$1st: + add r13,rax + mov rax,QWORD[r15*8+rsi] + adc rdx,0 + add r13,r11 + mov r11,r10 + adc rdx,0 + mov QWORD[((-16))+r15*8+rsp],r13 + mov r13,rdx + +$L$1st_enter: + mul rbx + add r11,rax + mov rax,QWORD[r15*8+rcx] + adc rdx,0 + lea r15,[1+r15] + mov r10,rdx + + mul rbp + cmp r15,r9 + jne NEAR $L$1st + + + add r13,rax + adc rdx,0 + add r13,r11 + adc rdx,0 + mov QWORD[((-16))+r9*8+rsp],r13 + mov r13,rdx + mov r11,r10 + + xor rdx,rdx + add r13,r11 + adc rdx,0 + mov QWORD[((-8))+r9*8+rsp],r13 + mov QWORD[r9*8+rsp],rdx + + lea r14,[1+r14] + jmp NEAR $L$outer +ALIGN 16 +$L$outer: + lea rdx,[((24+128))+r9*8+rsp] + and rdx,-16 + pxor xmm4,xmm4 + pxor xmm5,xmm5 + movdqa xmm0,XMMWORD[((-128))+r12] + movdqa xmm1,XMMWORD[((-112))+r12] + movdqa xmm2,XMMWORD[((-96))+r12] + movdqa xmm3,XMMWORD[((-80))+r12] + pand xmm0,XMMWORD[((-128))+rdx] + pand xmm1,XMMWORD[((-112))+rdx] + por xmm4,xmm0 + pand xmm2,XMMWORD[((-96))+rdx] + por xmm5,xmm1 + pand xmm3,XMMWORD[((-80))+rdx] + por xmm4,xmm2 + por xmm5,xmm3 + movdqa xmm0,XMMWORD[((-64))+r12] + movdqa xmm1,XMMWORD[((-48))+r12] + movdqa xmm2,XMMWORD[((-32))+r12] + movdqa xmm3,XMMWORD[((-16))+r12] + pand xmm0,XMMWORD[((-64))+rdx] + pand xmm1,XMMWORD[((-48))+rdx] + por xmm4,xmm0 + pand xmm2,XMMWORD[((-32))+rdx] + por xmm5,xmm1 + pand xmm3,XMMWORD[((-16))+rdx] + por xmm4,xmm2 + por xmm5,xmm3 + movdqa xmm0,XMMWORD[r12] + movdqa xmm1,XMMWORD[16+r12] + movdqa xmm2,XMMWORD[32+r12] + movdqa xmm3,XMMWORD[48+r12] + pand xmm0,XMMWORD[rdx] + pand xmm1,XMMWORD[16+rdx] + por xmm4,xmm0 + pand xmm2,XMMWORD[32+rdx] + por xmm5,xmm1 + pand xmm3,XMMWORD[48+rdx] + por xmm4,xmm2 + por xmm5,xmm3 + movdqa xmm0,XMMWORD[64+r12] + movdqa xmm1,XMMWORD[80+r12] + movdqa xmm2,XMMWORD[96+r12] + movdqa xmm3,XMMWORD[112+r12] + pand xmm0,XMMWORD[64+rdx] + pand xmm1,XMMWORD[80+rdx] + por xmm4,xmm0 + pand xmm2,XMMWORD[96+rdx] + por xmm5,xmm1 + pand xmm3,XMMWORD[112+rdx] + por xmm4,xmm2 + por xmm5,xmm3 + por xmm4,xmm5 + pshufd xmm0,xmm4,0x4e + por xmm0,xmm4 + lea r12,[256+r12] + + mov rax,QWORD[rsi] +DB 102,72,15,126,195 + + xor r15,r15 + mov rbp,r8 + mov r10,QWORD[rsp] + + mul rbx + add r10,rax + mov rax,QWORD[rcx] + adc rdx,0 + + imul rbp,r10 + mov r11,rdx + + mul rbp + add r10,rax + mov rax,QWORD[8+rsi] + adc rdx,0 + mov r10,QWORD[8+rsp] + mov r13,rdx + + lea r15,[1+r15] + jmp NEAR $L$inner_enter + +ALIGN 16 +$L$inner: + add r13,rax + mov rax,QWORD[r15*8+rsi] + adc rdx,0 + add r13,r10 + mov r10,QWORD[r15*8+rsp] + adc rdx,0 + mov QWORD[((-16))+r15*8+rsp],r13 + mov r13,rdx + +$L$inner_enter: + mul rbx + add r11,rax + mov rax,QWORD[r15*8+rcx] + adc rdx,0 + add r10,r11 + mov r11,rdx + adc r11,0 + lea r15,[1+r15] + + mul rbp + cmp r15,r9 + jne NEAR $L$inner + + add r13,rax + adc rdx,0 + add r13,r10 + mov r10,QWORD[r9*8+rsp] + adc rdx,0 + mov QWORD[((-16))+r9*8+rsp],r13 + mov r13,rdx + + xor rdx,rdx + add r13,r11 + adc rdx,0 + add r13,r10 + adc rdx,0 + mov QWORD[((-8))+r9*8+rsp],r13 + mov QWORD[r9*8+rsp],rdx + + lea r14,[1+r14] + cmp r14,r9 + jb NEAR $L$outer + + xor r14,r14 + mov rax,QWORD[rsp] + lea rsi,[rsp] + mov r15,r9 + jmp NEAR $L$sub +ALIGN 16 +$L$sub: sbb rax,QWORD[r14*8+rcx] + mov QWORD[r14*8+rdi],rax + mov rax,QWORD[8+r14*8+rsi] + lea r14,[1+r14] + dec r15 + jnz NEAR $L$sub + + sbb rax,0 + mov rbx,-1 + xor rbx,rax + xor r14,r14 + mov r15,r9 + +$L$copy: + mov rcx,QWORD[r14*8+rdi] + mov rdx,QWORD[r14*8+rsp] + and rcx,rbx + and rdx,rax + mov QWORD[r14*8+rsp],r14 + or rdx,rcx + mov QWORD[r14*8+rdi],rdx + lea r14,[1+r14] + sub r15,1 + jnz NEAR $L$copy + + mov rsi,QWORD[8+r9*8+rsp] + + mov rax,1 + + mov r15,QWORD[((-48))+rsi] + + mov r14,QWORD[((-40))+rsi] + + mov r13,QWORD[((-32))+rsi] + + mov r12,QWORD[((-24))+rsi] + + mov rbp,QWORD[((-16))+rsi] + + mov rbx,QWORD[((-8))+rsi] + + lea rsp,[rsi] + +$L$mul_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_bn_mul_mont_gather5: + +ALIGN 32 +bn_mul4x_mont_gather5: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[16+rsp],rsi + mov rax,rsp +$L$SEH_begin_bn_mul4x_mont_gather5: + mov rdi,rcx + mov rsi,rdx + mov rdx,r8 + mov rcx,r9 + mov r8,QWORD[40+rsp] + mov r9,QWORD[48+rsp] + + + +DB 0x67 + mov rax,rsp + +$L$mul4x_enter: + and r11d,0x80108 + cmp r11d,0x80108 + je NEAR $L$mulx4x_enter + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + +$L$mul4x_prologue: + +DB 0x67 + shl r9d,3 + lea r10,[r9*2+r9] + neg r9 + + + + + + + + + + + lea r11,[((-320))+r9*2+rsp] + mov rbp,rsp + sub r11,rdi + and r11,4095 + cmp r10,r11 + jb NEAR $L$mul4xsp_alt + sub rbp,r11 + lea rbp,[((-320))+r9*2+rbp] + jmp NEAR $L$mul4xsp_done + +ALIGN 32 +$L$mul4xsp_alt: + lea r10,[((4096-320))+r9*2] + lea rbp,[((-320))+r9*2+rbp] + sub r11,r10 + mov r10,0 + cmovc r11,r10 + sub rbp,r11 +$L$mul4xsp_done: + and rbp,-64 + mov r11,rsp + sub r11,rbp + and r11,-4096 + lea rsp,[rbp*1+r11] + mov r10,QWORD[rsp] + cmp rsp,rbp + ja NEAR $L$mul4x_page_walk + jmp NEAR $L$mul4x_page_walk_done + +$L$mul4x_page_walk: + lea rsp,[((-4096))+rsp] + mov r10,QWORD[rsp] + cmp rsp,rbp + ja NEAR $L$mul4x_page_walk +$L$mul4x_page_walk_done: + + neg r9 + + mov QWORD[40+rsp],rax + +$L$mul4x_body: + + call mul4x_internal + + mov rsi,QWORD[40+rsp] + + mov rax,1 + + mov r15,QWORD[((-48))+rsi] + + mov r14,QWORD[((-40))+rsi] + + mov r13,QWORD[((-32))+rsi] + + mov r12,QWORD[((-24))+rsi] + + mov rbp,QWORD[((-16))+rsi] + + mov rbx,QWORD[((-8))+rsi] + + lea rsp,[rsi] + +$L$mul4x_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_bn_mul4x_mont_gather5: + + +ALIGN 32 +mul4x_internal: + + shl r9,5 + movd xmm5,DWORD[56+rax] + lea rax,[$L$inc] + lea r13,[128+r9*1+rdx] + shr r9,5 + movdqa xmm0,XMMWORD[rax] + movdqa xmm1,XMMWORD[16+rax] + lea r10,[((88-112))+r9*1+rsp] + lea r12,[128+rdx] + + pshufd xmm5,xmm5,0 + movdqa xmm4,xmm1 +DB 0x67,0x67 + movdqa xmm2,xmm1 + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 +DB 0x67 + movdqa xmm3,xmm4 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[112+r10],xmm0 + movdqa xmm0,xmm4 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[128+r10],xmm1 + movdqa xmm1,xmm4 + + paddd xmm0,xmm3 + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[144+r10],xmm2 + movdqa xmm2,xmm4 + + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa XMMWORD[160+r10],xmm3 + movdqa xmm3,xmm4 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[176+r10],xmm0 + movdqa xmm0,xmm4 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[192+r10],xmm1 + movdqa xmm1,xmm4 + + paddd xmm0,xmm3 + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[208+r10],xmm2 + movdqa xmm2,xmm4 + + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa XMMWORD[224+r10],xmm3 + movdqa xmm3,xmm4 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[240+r10],xmm0 + movdqa xmm0,xmm4 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[256+r10],xmm1 + movdqa xmm1,xmm4 + + paddd xmm0,xmm3 + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[272+r10],xmm2 + movdqa xmm2,xmm4 + + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa XMMWORD[288+r10],xmm3 + movdqa xmm3,xmm4 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[304+r10],xmm0 + + paddd xmm3,xmm2 +DB 0x67 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[320+r10],xmm1 + + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[336+r10],xmm2 + pand xmm0,XMMWORD[64+r12] + + pand xmm1,XMMWORD[80+r12] + pand xmm2,XMMWORD[96+r12] + movdqa XMMWORD[352+r10],xmm3 + pand xmm3,XMMWORD[112+r12] + por xmm0,xmm2 + por xmm1,xmm3 + movdqa xmm4,XMMWORD[((-128))+r12] + movdqa xmm5,XMMWORD[((-112))+r12] + movdqa xmm2,XMMWORD[((-96))+r12] + pand xmm4,XMMWORD[112+r10] + movdqa xmm3,XMMWORD[((-80))+r12] + pand xmm5,XMMWORD[128+r10] + por xmm0,xmm4 + pand xmm2,XMMWORD[144+r10] + por xmm1,xmm5 + pand xmm3,XMMWORD[160+r10] + por xmm0,xmm2 + por xmm1,xmm3 + movdqa xmm4,XMMWORD[((-64))+r12] + movdqa xmm5,XMMWORD[((-48))+r12] + movdqa xmm2,XMMWORD[((-32))+r12] + pand xmm4,XMMWORD[176+r10] + movdqa xmm3,XMMWORD[((-16))+r12] + pand xmm5,XMMWORD[192+r10] + por xmm0,xmm4 + pand xmm2,XMMWORD[208+r10] + por xmm1,xmm5 + pand xmm3,XMMWORD[224+r10] + por xmm0,xmm2 + por xmm1,xmm3 + movdqa xmm4,XMMWORD[r12] + movdqa xmm5,XMMWORD[16+r12] + movdqa xmm2,XMMWORD[32+r12] + pand xmm4,XMMWORD[240+r10] + movdqa xmm3,XMMWORD[48+r12] + pand xmm5,XMMWORD[256+r10] + por xmm0,xmm4 + pand xmm2,XMMWORD[272+r10] + por xmm1,xmm5 + pand xmm3,XMMWORD[288+r10] + por xmm0,xmm2 + por xmm1,xmm3 + por xmm0,xmm1 + pshufd xmm1,xmm0,0x4e + por xmm0,xmm1 + lea r12,[256+r12] +DB 102,72,15,126,195 + + mov QWORD[((16+8))+rsp],r13 + mov QWORD[((56+8))+rsp],rdi + + mov r8,QWORD[r8] + mov rax,QWORD[rsi] + lea rsi,[r9*1+rsi] + neg r9 + + mov rbp,r8 + mul rbx + mov r10,rax + mov rax,QWORD[rcx] + + imul rbp,r10 + lea r14,[((64+8))+rsp] + mov r11,rdx + + mul rbp + add r10,rax + mov rax,QWORD[8+r9*1+rsi] + adc rdx,0 + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[8+rcx] + adc rdx,0 + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[16+r9*1+rsi] + adc rdx,0 + add rdi,r11 + lea r15,[32+r9] + lea rcx,[32+rcx] + adc rdx,0 + mov QWORD[r14],rdi + mov r13,rdx + jmp NEAR $L$1st4x + +ALIGN 32 +$L$1st4x: + mul rbx + add r10,rax + mov rax,QWORD[((-16))+rcx] + lea r14,[32+r14] + adc rdx,0 + mov r11,rdx + + mul rbp + add r13,rax + mov rax,QWORD[((-8))+r15*1+rsi] + adc rdx,0 + add r13,r10 + adc rdx,0 + mov QWORD[((-24))+r14],r13 + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[((-8))+rcx] + adc rdx,0 + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[r15*1+rsi] + adc rdx,0 + add rdi,r11 + adc rdx,0 + mov QWORD[((-16))+r14],rdi + mov r13,rdx + + mul rbx + add r10,rax + mov rax,QWORD[rcx] + adc rdx,0 + mov r11,rdx + + mul rbp + add r13,rax + mov rax,QWORD[8+r15*1+rsi] + adc rdx,0 + add r13,r10 + adc rdx,0 + mov QWORD[((-8))+r14],r13 + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[8+rcx] + adc rdx,0 + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[16+r15*1+rsi] + adc rdx,0 + add rdi,r11 + lea rcx,[32+rcx] + adc rdx,0 + mov QWORD[r14],rdi + mov r13,rdx + + add r15,32 + jnz NEAR $L$1st4x + + mul rbx + add r10,rax + mov rax,QWORD[((-16))+rcx] + lea r14,[32+r14] + adc rdx,0 + mov r11,rdx + + mul rbp + add r13,rax + mov rax,QWORD[((-8))+rsi] + adc rdx,0 + add r13,r10 + adc rdx,0 + mov QWORD[((-24))+r14],r13 + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[((-8))+rcx] + adc rdx,0 + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[r9*1+rsi] + adc rdx,0 + add rdi,r11 + adc rdx,0 + mov QWORD[((-16))+r14],rdi + mov r13,rdx + + lea rcx,[r9*1+rcx] + + xor rdi,rdi + add r13,r10 + adc rdi,0 + mov QWORD[((-8))+r14],r13 + + jmp NEAR $L$outer4x + +ALIGN 32 +$L$outer4x: + lea rdx,[((16+128))+r14] + pxor xmm4,xmm4 + pxor xmm5,xmm5 + movdqa xmm0,XMMWORD[((-128))+r12] + movdqa xmm1,XMMWORD[((-112))+r12] + movdqa xmm2,XMMWORD[((-96))+r12] + movdqa xmm3,XMMWORD[((-80))+r12] + pand xmm0,XMMWORD[((-128))+rdx] + pand xmm1,XMMWORD[((-112))+rdx] + por xmm4,xmm0 + pand xmm2,XMMWORD[((-96))+rdx] + por xmm5,xmm1 + pand xmm3,XMMWORD[((-80))+rdx] + por xmm4,xmm2 + por xmm5,xmm3 + movdqa xmm0,XMMWORD[((-64))+r12] + movdqa xmm1,XMMWORD[((-48))+r12] + movdqa xmm2,XMMWORD[((-32))+r12] + movdqa xmm3,XMMWORD[((-16))+r12] + pand xmm0,XMMWORD[((-64))+rdx] + pand xmm1,XMMWORD[((-48))+rdx] + por xmm4,xmm0 + pand xmm2,XMMWORD[((-32))+rdx] + por xmm5,xmm1 + pand xmm3,XMMWORD[((-16))+rdx] + por xmm4,xmm2 + por xmm5,xmm3 + movdqa xmm0,XMMWORD[r12] + movdqa xmm1,XMMWORD[16+r12] + movdqa xmm2,XMMWORD[32+r12] + movdqa xmm3,XMMWORD[48+r12] + pand xmm0,XMMWORD[rdx] + pand xmm1,XMMWORD[16+rdx] + por xmm4,xmm0 + pand xmm2,XMMWORD[32+rdx] + por xmm5,xmm1 + pand xmm3,XMMWORD[48+rdx] + por xmm4,xmm2 + por xmm5,xmm3 + movdqa xmm0,XMMWORD[64+r12] + movdqa xmm1,XMMWORD[80+r12] + movdqa xmm2,XMMWORD[96+r12] + movdqa xmm3,XMMWORD[112+r12] + pand xmm0,XMMWORD[64+rdx] + pand xmm1,XMMWORD[80+rdx] + por xmm4,xmm0 + pand xmm2,XMMWORD[96+rdx] + por xmm5,xmm1 + pand xmm3,XMMWORD[112+rdx] + por xmm4,xmm2 + por xmm5,xmm3 + por xmm4,xmm5 + pshufd xmm0,xmm4,0x4e + por xmm0,xmm4 + lea r12,[256+r12] +DB 102,72,15,126,195 + + mov r10,QWORD[r9*1+r14] + mov rbp,r8 + mul rbx + add r10,rax + mov rax,QWORD[rcx] + adc rdx,0 + + imul rbp,r10 + mov r11,rdx + mov QWORD[r14],rdi + + lea r14,[r9*1+r14] + + mul rbp + add r10,rax + mov rax,QWORD[8+r9*1+rsi] + adc rdx,0 + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[8+rcx] + adc rdx,0 + add r11,QWORD[8+r14] + adc rdx,0 + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[16+r9*1+rsi] + adc rdx,0 + add rdi,r11 + lea r15,[32+r9] + lea rcx,[32+rcx] + adc rdx,0 + mov r13,rdx + jmp NEAR $L$inner4x + +ALIGN 32 +$L$inner4x: + mul rbx + add r10,rax + mov rax,QWORD[((-16))+rcx] + adc rdx,0 + add r10,QWORD[16+r14] + lea r14,[32+r14] + adc rdx,0 + mov r11,rdx + + mul rbp + add r13,rax + mov rax,QWORD[((-8))+r15*1+rsi] + adc rdx,0 + add r13,r10 + adc rdx,0 + mov QWORD[((-32))+r14],rdi + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[((-8))+rcx] + adc rdx,0 + add r11,QWORD[((-8))+r14] + adc rdx,0 + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[r15*1+rsi] + adc rdx,0 + add rdi,r11 + adc rdx,0 + mov QWORD[((-24))+r14],r13 + mov r13,rdx + + mul rbx + add r10,rax + mov rax,QWORD[rcx] + adc rdx,0 + add r10,QWORD[r14] + adc rdx,0 + mov r11,rdx + + mul rbp + add r13,rax + mov rax,QWORD[8+r15*1+rsi] + adc rdx,0 + add r13,r10 + adc rdx,0 + mov QWORD[((-16))+r14],rdi + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,QWORD[8+rcx] + adc rdx,0 + add r11,QWORD[8+r14] + adc rdx,0 + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[16+r15*1+rsi] + adc rdx,0 + add rdi,r11 + lea rcx,[32+rcx] + adc rdx,0 + mov QWORD[((-8))+r14],r13 + mov r13,rdx + + add r15,32 + jnz NEAR $L$inner4x + + mul rbx + add r10,rax + mov rax,QWORD[((-16))+rcx] + adc rdx,0 + add r10,QWORD[16+r14] + lea r14,[32+r14] + adc rdx,0 + mov r11,rdx + + mul rbp + add r13,rax + mov rax,QWORD[((-8))+rsi] + adc rdx,0 + add r13,r10 + adc rdx,0 + mov QWORD[((-32))+r14],rdi + mov rdi,rdx + + mul rbx + add r11,rax + mov rax,rbp + mov rbp,QWORD[((-8))+rcx] + adc rdx,0 + add r11,QWORD[((-8))+r14] + adc rdx,0 + mov r10,rdx + + mul rbp + add rdi,rax + mov rax,QWORD[r9*1+rsi] + adc rdx,0 + add rdi,r11 + adc rdx,0 + mov QWORD[((-24))+r14],r13 + mov r13,rdx + + mov QWORD[((-16))+r14],rdi + lea rcx,[r9*1+rcx] + + xor rdi,rdi + add r13,r10 + adc rdi,0 + add r13,QWORD[r14] + adc rdi,0 + mov QWORD[((-8))+r14],r13 + + cmp r12,QWORD[((16+8))+rsp] + jb NEAR $L$outer4x + xor rax,rax + sub rbp,r13 + adc r15,r15 + or rdi,r15 + sub rax,rdi + lea rbx,[r9*1+r14] + mov r12,QWORD[rcx] + lea rbp,[rcx] + mov rcx,r9 + sar rcx,3+2 + mov rdi,QWORD[((56+8))+rsp] + dec r12 + xor r10,r10 + mov r13,QWORD[8+rbp] + mov r14,QWORD[16+rbp] + mov r15,QWORD[24+rbp] + jmp NEAR $L$sqr4x_sub_entry + + +global bn_power5 + +ALIGN 32 +bn_power5: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[16+rsp],rsi + mov rax,rsp +$L$SEH_begin_bn_power5: + mov rdi,rcx + mov rsi,rdx + mov rdx,r8 + mov rcx,r9 + mov r8,QWORD[40+rsp] + mov r9,QWORD[48+rsp] + + + + mov rax,rsp + + mov r11d,DWORD[((OPENSSL_ia32cap_P+8))] + and r11d,0x80108 + cmp r11d,0x80108 + je NEAR $L$powerx5_enter + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + +$L$power5_prologue: + + shl r9d,3 + lea r10d,[r9*2+r9] + neg r9 + mov r8,QWORD[r8] + + + + + + + + + lea r11,[((-320))+r9*2+rsp] + mov rbp,rsp + sub r11,rdi + and r11,4095 + cmp r10,r11 + jb NEAR $L$pwr_sp_alt + sub rbp,r11 + lea rbp,[((-320))+r9*2+rbp] + jmp NEAR $L$pwr_sp_done + +ALIGN 32 +$L$pwr_sp_alt: + lea r10,[((4096-320))+r9*2] + lea rbp,[((-320))+r9*2+rbp] + sub r11,r10 + mov r10,0 + cmovc r11,r10 + sub rbp,r11 +$L$pwr_sp_done: + and rbp,-64 + mov r11,rsp + sub r11,rbp + and r11,-4096 + lea rsp,[rbp*1+r11] + mov r10,QWORD[rsp] + cmp rsp,rbp + ja NEAR $L$pwr_page_walk + jmp NEAR $L$pwr_page_walk_done + +$L$pwr_page_walk: + lea rsp,[((-4096))+rsp] + mov r10,QWORD[rsp] + cmp rsp,rbp + ja NEAR $L$pwr_page_walk +$L$pwr_page_walk_done: + + mov r10,r9 + neg r9 + + + + + + + + + + + mov QWORD[32+rsp],r8 + mov QWORD[40+rsp],rax + +$L$power5_body: +DB 102,72,15,110,207 +DB 102,72,15,110,209 +DB 102,73,15,110,218 +DB 102,72,15,110,226 + + call __bn_sqr8x_internal + call __bn_post4x_internal + call __bn_sqr8x_internal + call __bn_post4x_internal + call __bn_sqr8x_internal + call __bn_post4x_internal + call __bn_sqr8x_internal + call __bn_post4x_internal + call __bn_sqr8x_internal + call __bn_post4x_internal + +DB 102,72,15,126,209 +DB 102,72,15,126,226 + mov rdi,rsi + mov rax,QWORD[40+rsp] + lea r8,[32+rsp] + + call mul4x_internal + + mov rsi,QWORD[40+rsp] + + mov rax,1 + mov r15,QWORD[((-48))+rsi] + + mov r14,QWORD[((-40))+rsi] + + mov r13,QWORD[((-32))+rsi] + + mov r12,QWORD[((-24))+rsi] + + mov rbp,QWORD[((-16))+rsi] + + mov rbx,QWORD[((-8))+rsi] + + lea rsp,[rsi] + +$L$power5_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_bn_power5: + +global bn_sqr8x_internal + + +ALIGN 32 +bn_sqr8x_internal: +__bn_sqr8x_internal: + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + lea rbp,[32+r10] + lea rsi,[r9*1+rsi] + + mov rcx,r9 + + + mov r14,QWORD[((-32))+rbp*1+rsi] + lea rdi,[((48+8))+r9*2+rsp] + mov rax,QWORD[((-24))+rbp*1+rsi] + lea rdi,[((-32))+rbp*1+rdi] + mov rbx,QWORD[((-16))+rbp*1+rsi] + mov r15,rax + + mul r14 + mov r10,rax + mov rax,rbx + mov r11,rdx + mov QWORD[((-24))+rbp*1+rdi],r10 + + mul r14 + add r11,rax + mov rax,rbx + adc rdx,0 + mov QWORD[((-16))+rbp*1+rdi],r11 + mov r10,rdx + + + mov rbx,QWORD[((-8))+rbp*1+rsi] + mul r15 + mov r12,rax + mov rax,rbx + mov r13,rdx + + lea rcx,[rbp] + mul r14 + add r10,rax + mov rax,rbx + mov r11,rdx + adc r11,0 + add r10,r12 + adc r11,0 + mov QWORD[((-8))+rcx*1+rdi],r10 + jmp NEAR $L$sqr4x_1st + +ALIGN 32 +$L$sqr4x_1st: + mov rbx,QWORD[rcx*1+rsi] + mul r15 + add r13,rax + mov rax,rbx + mov r12,rdx + adc r12,0 + + mul r14 + add r11,rax + mov rax,rbx + mov rbx,QWORD[8+rcx*1+rsi] + mov r10,rdx + adc r10,0 + add r11,r13 + adc r10,0 + + + mul r15 + add r12,rax + mov rax,rbx + mov QWORD[rcx*1+rdi],r11 + mov r13,rdx + adc r13,0 + + mul r14 + add r10,rax + mov rax,rbx + mov rbx,QWORD[16+rcx*1+rsi] + mov r11,rdx + adc r11,0 + add r10,r12 + adc r11,0 + + mul r15 + add r13,rax + mov rax,rbx + mov QWORD[8+rcx*1+rdi],r10 + mov r12,rdx + adc r12,0 + + mul r14 + add r11,rax + mov rax,rbx + mov rbx,QWORD[24+rcx*1+rsi] + mov r10,rdx + adc r10,0 + add r11,r13 + adc r10,0 + + + mul r15 + add r12,rax + mov rax,rbx + mov QWORD[16+rcx*1+rdi],r11 + mov r13,rdx + adc r13,0 + lea rcx,[32+rcx] + + mul r14 + add r10,rax + mov rax,rbx + mov r11,rdx + adc r11,0 + add r10,r12 + adc r11,0 + mov QWORD[((-8))+rcx*1+rdi],r10 + + cmp rcx,0 + jne NEAR $L$sqr4x_1st + + mul r15 + add r13,rax + lea rbp,[16+rbp] + adc rdx,0 + add r13,r11 + adc rdx,0 + + mov QWORD[rdi],r13 + mov r12,rdx + mov QWORD[8+rdi],rdx + jmp NEAR $L$sqr4x_outer + +ALIGN 32 +$L$sqr4x_outer: + mov r14,QWORD[((-32))+rbp*1+rsi] + lea rdi,[((48+8))+r9*2+rsp] + mov rax,QWORD[((-24))+rbp*1+rsi] + lea rdi,[((-32))+rbp*1+rdi] + mov rbx,QWORD[((-16))+rbp*1+rsi] + mov r15,rax + + mul r14 + mov r10,QWORD[((-24))+rbp*1+rdi] + add r10,rax + mov rax,rbx + adc rdx,0 + mov QWORD[((-24))+rbp*1+rdi],r10 + mov r11,rdx + + mul r14 + add r11,rax + mov rax,rbx + adc rdx,0 + add r11,QWORD[((-16))+rbp*1+rdi] + mov r10,rdx + adc r10,0 + mov QWORD[((-16))+rbp*1+rdi],r11 + + xor r12,r12 + + mov rbx,QWORD[((-8))+rbp*1+rsi] + mul r15 + add r12,rax + mov rax,rbx + adc rdx,0 + add r12,QWORD[((-8))+rbp*1+rdi] + mov r13,rdx + adc r13,0 + + mul r14 + add r10,rax + mov rax,rbx + adc rdx,0 + add r10,r12 + mov r11,rdx + adc r11,0 + mov QWORD[((-8))+rbp*1+rdi],r10 + + lea rcx,[rbp] + jmp NEAR $L$sqr4x_inner + +ALIGN 32 +$L$sqr4x_inner: + mov rbx,QWORD[rcx*1+rsi] + mul r15 + add r13,rax + mov rax,rbx + mov r12,rdx + adc r12,0 + add r13,QWORD[rcx*1+rdi] + adc r12,0 + +DB 0x67 + mul r14 + add r11,rax + mov rax,rbx + mov rbx,QWORD[8+rcx*1+rsi] + mov r10,rdx + adc r10,0 + add r11,r13 + adc r10,0 + + mul r15 + add r12,rax + mov QWORD[rcx*1+rdi],r11 + mov rax,rbx + mov r13,rdx + adc r13,0 + add r12,QWORD[8+rcx*1+rdi] + lea rcx,[16+rcx] + adc r13,0 + + mul r14 + add r10,rax + mov rax,rbx + adc rdx,0 + add r10,r12 + mov r11,rdx + adc r11,0 + mov QWORD[((-8))+rcx*1+rdi],r10 + + cmp rcx,0 + jne NEAR $L$sqr4x_inner + +DB 0x67 + mul r15 + add r13,rax + adc rdx,0 + add r13,r11 + adc rdx,0 + + mov QWORD[rdi],r13 + mov r12,rdx + mov QWORD[8+rdi],rdx + + add rbp,16 + jnz NEAR $L$sqr4x_outer + + + mov r14,QWORD[((-32))+rsi] + lea rdi,[((48+8))+r9*2+rsp] + mov rax,QWORD[((-24))+rsi] + lea rdi,[((-32))+rbp*1+rdi] + mov rbx,QWORD[((-16))+rsi] + mov r15,rax + + mul r14 + add r10,rax + mov rax,rbx + mov r11,rdx + adc r11,0 + + mul r14 + add r11,rax + mov rax,rbx + mov QWORD[((-24))+rdi],r10 + mov r10,rdx + adc r10,0 + add r11,r13 + mov rbx,QWORD[((-8))+rsi] + adc r10,0 + + mul r15 + add r12,rax + mov rax,rbx + mov QWORD[((-16))+rdi],r11 + mov r13,rdx + adc r13,0 + + mul r14 + add r10,rax + mov rax,rbx + mov r11,rdx + adc r11,0 + add r10,r12 + adc r11,0 + mov QWORD[((-8))+rdi],r10 + + mul r15 + add r13,rax + mov rax,QWORD[((-16))+rsi] + adc rdx,0 + add r13,r11 + adc rdx,0 + + mov QWORD[rdi],r13 + mov r12,rdx + mov QWORD[8+rdi],rdx + + mul rbx + add rbp,16 + xor r14,r14 + sub rbp,r9 + xor r15,r15 + + add rax,r12 + adc rdx,0 + mov QWORD[8+rdi],rax + mov QWORD[16+rdi],rdx + mov QWORD[24+rdi],r15 + + mov rax,QWORD[((-16))+rbp*1+rsi] + lea rdi,[((48+8))+rsp] + xor r10,r10 + mov r11,QWORD[8+rdi] + + lea r12,[r10*2+r14] + shr r10,63 + lea r13,[r11*2+rcx] + shr r11,63 + or r13,r10 + mov r10,QWORD[16+rdi] + mov r14,r11 + mul rax + neg r15 + mov r11,QWORD[24+rdi] + adc r12,rax + mov rax,QWORD[((-8))+rbp*1+rsi] + mov QWORD[rdi],r12 + adc r13,rdx + + lea rbx,[r10*2+r14] + mov QWORD[8+rdi],r13 + sbb r15,r15 + shr r10,63 + lea r8,[r11*2+rcx] + shr r11,63 + or r8,r10 + mov r10,QWORD[32+rdi] + mov r14,r11 + mul rax + neg r15 + mov r11,QWORD[40+rdi] + adc rbx,rax + mov rax,QWORD[rbp*1+rsi] + mov QWORD[16+rdi],rbx + adc r8,rdx + lea rbp,[16+rbp] + mov QWORD[24+rdi],r8 + sbb r15,r15 + lea rdi,[64+rdi] + jmp NEAR $L$sqr4x_shift_n_add + +ALIGN 32 +$L$sqr4x_shift_n_add: + lea r12,[r10*2+r14] + shr r10,63 + lea r13,[r11*2+rcx] + shr r11,63 + or r13,r10 + mov r10,QWORD[((-16))+rdi] + mov r14,r11 + mul rax + neg r15 + mov r11,QWORD[((-8))+rdi] + adc r12,rax + mov rax,QWORD[((-8))+rbp*1+rsi] + mov QWORD[((-32))+rdi],r12 + adc r13,rdx + + lea rbx,[r10*2+r14] + mov QWORD[((-24))+rdi],r13 + sbb r15,r15 + shr r10,63 + lea r8,[r11*2+rcx] + shr r11,63 + or r8,r10 + mov r10,QWORD[rdi] + mov r14,r11 + mul rax + neg r15 + mov r11,QWORD[8+rdi] + adc rbx,rax + mov rax,QWORD[rbp*1+rsi] + mov QWORD[((-16))+rdi],rbx + adc r8,rdx + + lea r12,[r10*2+r14] + mov QWORD[((-8))+rdi],r8 + sbb r15,r15 + shr r10,63 + lea r13,[r11*2+rcx] + shr r11,63 + or r13,r10 + mov r10,QWORD[16+rdi] + mov r14,r11 + mul rax + neg r15 + mov r11,QWORD[24+rdi] + adc r12,rax + mov rax,QWORD[8+rbp*1+rsi] + mov QWORD[rdi],r12 + adc r13,rdx + + lea rbx,[r10*2+r14] + mov QWORD[8+rdi],r13 + sbb r15,r15 + shr r10,63 + lea r8,[r11*2+rcx] + shr r11,63 + or r8,r10 + mov r10,QWORD[32+rdi] + mov r14,r11 + mul rax + neg r15 + mov r11,QWORD[40+rdi] + adc rbx,rax + mov rax,QWORD[16+rbp*1+rsi] + mov QWORD[16+rdi],rbx + adc r8,rdx + mov QWORD[24+rdi],r8 + sbb r15,r15 + lea rdi,[64+rdi] + add rbp,32 + jnz NEAR $L$sqr4x_shift_n_add + + lea r12,[r10*2+r14] +DB 0x67 + shr r10,63 + lea r13,[r11*2+rcx] + shr r11,63 + or r13,r10 + mov r10,QWORD[((-16))+rdi] + mov r14,r11 + mul rax + neg r15 + mov r11,QWORD[((-8))+rdi] + adc r12,rax + mov rax,QWORD[((-8))+rsi] + mov QWORD[((-32))+rdi],r12 + adc r13,rdx + + lea rbx,[r10*2+r14] + mov QWORD[((-24))+rdi],r13 + sbb r15,r15 + shr r10,63 + lea r8,[r11*2+rcx] + shr r11,63 + or r8,r10 + mul rax + neg r15 + adc rbx,rax + adc r8,rdx + mov QWORD[((-16))+rdi],rbx + mov QWORD[((-8))+rdi],r8 +DB 102,72,15,126,213 +__bn_sqr8x_reduction: + xor rax,rax + lea rcx,[rbp*1+r9] + lea rdx,[((48+8))+r9*2+rsp] + mov QWORD[((0+8))+rsp],rcx + lea rdi,[((48+8))+r9*1+rsp] + mov QWORD[((8+8))+rsp],rdx + neg r9 + jmp NEAR $L$8x_reduction_loop + +ALIGN 32 +$L$8x_reduction_loop: + lea rdi,[r9*1+rdi] +DB 0x66 + mov rbx,QWORD[rdi] + mov r9,QWORD[8+rdi] + mov r10,QWORD[16+rdi] + mov r11,QWORD[24+rdi] + mov r12,QWORD[32+rdi] + mov r13,QWORD[40+rdi] + mov r14,QWORD[48+rdi] + mov r15,QWORD[56+rdi] + mov QWORD[rdx],rax + lea rdi,[64+rdi] + +DB 0x67 + mov r8,rbx + imul rbx,QWORD[((32+8))+rsp] + mov rax,QWORD[rbp] + mov ecx,8 + jmp NEAR $L$8x_reduce + +ALIGN 32 +$L$8x_reduce: + mul rbx + mov rax,QWORD[8+rbp] + neg r8 + mov r8,rdx + adc r8,0 + + mul rbx + add r9,rax + mov rax,QWORD[16+rbp] + adc rdx,0 + add r8,r9 + mov QWORD[((48-8+8))+rcx*8+rsp],rbx + mov r9,rdx + adc r9,0 + + mul rbx + add r10,rax + mov rax,QWORD[24+rbp] + adc rdx,0 + add r9,r10 + mov rsi,QWORD[((32+8))+rsp] + mov r10,rdx + adc r10,0 + + mul rbx + add r11,rax + mov rax,QWORD[32+rbp] + adc rdx,0 + imul rsi,r8 + add r10,r11 + mov r11,rdx + adc r11,0 + + mul rbx + add r12,rax + mov rax,QWORD[40+rbp] + adc rdx,0 + add r11,r12 + mov r12,rdx + adc r12,0 + + mul rbx + add r13,rax + mov rax,QWORD[48+rbp] + adc rdx,0 + add r12,r13 + mov r13,rdx + adc r13,0 + + mul rbx + add r14,rax + mov rax,QWORD[56+rbp] + adc rdx,0 + add r13,r14 + mov r14,rdx + adc r14,0 + + mul rbx + mov rbx,rsi + add r15,rax + mov rax,QWORD[rbp] + adc rdx,0 + add r14,r15 + mov r15,rdx + adc r15,0 + + dec ecx + jnz NEAR $L$8x_reduce + + lea rbp,[64+rbp] + xor rax,rax + mov rdx,QWORD[((8+8))+rsp] + cmp rbp,QWORD[((0+8))+rsp] + jae NEAR $L$8x_no_tail + +DB 0x66 + add r8,QWORD[rdi] + adc r9,QWORD[8+rdi] + adc r10,QWORD[16+rdi] + adc r11,QWORD[24+rdi] + adc r12,QWORD[32+rdi] + adc r13,QWORD[40+rdi] + adc r14,QWORD[48+rdi] + adc r15,QWORD[56+rdi] + sbb rsi,rsi + + mov rbx,QWORD[((48+56+8))+rsp] + mov ecx,8 + mov rax,QWORD[rbp] + jmp NEAR $L$8x_tail + +ALIGN 32 +$L$8x_tail: + mul rbx + add r8,rax + mov rax,QWORD[8+rbp] + mov QWORD[rdi],r8 + mov r8,rdx + adc r8,0 + + mul rbx + add r9,rax + mov rax,QWORD[16+rbp] + adc rdx,0 + add r8,r9 + lea rdi,[8+rdi] + mov r9,rdx + adc r9,0 + + mul rbx + add r10,rax + mov rax,QWORD[24+rbp] + adc rdx,0 + add r9,r10 + mov r10,rdx + adc r10,0 + + mul rbx + add r11,rax + mov rax,QWORD[32+rbp] + adc rdx,0 + add r10,r11 + mov r11,rdx + adc r11,0 + + mul rbx + add r12,rax + mov rax,QWORD[40+rbp] + adc rdx,0 + add r11,r12 + mov r12,rdx + adc r12,0 + + mul rbx + add r13,rax + mov rax,QWORD[48+rbp] + adc rdx,0 + add r12,r13 + mov r13,rdx + adc r13,0 + + mul rbx + add r14,rax + mov rax,QWORD[56+rbp] + adc rdx,0 + add r13,r14 + mov r14,rdx + adc r14,0 + + mul rbx + mov rbx,QWORD[((48-16+8))+rcx*8+rsp] + add r15,rax + adc rdx,0 + add r14,r15 + mov rax,QWORD[rbp] + mov r15,rdx + adc r15,0 + + dec ecx + jnz NEAR $L$8x_tail + + lea rbp,[64+rbp] + mov rdx,QWORD[((8+8))+rsp] + cmp rbp,QWORD[((0+8))+rsp] + jae NEAR $L$8x_tail_done + + mov rbx,QWORD[((48+56+8))+rsp] + neg rsi + mov rax,QWORD[rbp] + adc r8,QWORD[rdi] + adc r9,QWORD[8+rdi] + adc r10,QWORD[16+rdi] + adc r11,QWORD[24+rdi] + adc r12,QWORD[32+rdi] + adc r13,QWORD[40+rdi] + adc r14,QWORD[48+rdi] + adc r15,QWORD[56+rdi] + sbb rsi,rsi + + mov ecx,8 + jmp NEAR $L$8x_tail + +ALIGN 32 +$L$8x_tail_done: + xor rax,rax + add r8,QWORD[rdx] + adc r9,0 + adc r10,0 + adc r11,0 + adc r12,0 + adc r13,0 + adc r14,0 + adc r15,0 + adc rax,0 + + neg rsi +$L$8x_no_tail: + adc r8,QWORD[rdi] + adc r9,QWORD[8+rdi] + adc r10,QWORD[16+rdi] + adc r11,QWORD[24+rdi] + adc r12,QWORD[32+rdi] + adc r13,QWORD[40+rdi] + adc r14,QWORD[48+rdi] + adc r15,QWORD[56+rdi] + adc rax,0 + mov rcx,QWORD[((-8))+rbp] + xor rsi,rsi + +DB 102,72,15,126,213 + + mov QWORD[rdi],r8 + mov QWORD[8+rdi],r9 +DB 102,73,15,126,217 + mov QWORD[16+rdi],r10 + mov QWORD[24+rdi],r11 + mov QWORD[32+rdi],r12 + mov QWORD[40+rdi],r13 + mov QWORD[48+rdi],r14 + mov QWORD[56+rdi],r15 + lea rdi,[64+rdi] + + cmp rdi,rdx + jb NEAR $L$8x_reduction_loop + DB 0F3h,0C3h ;repret + + + +ALIGN 32 +__bn_post4x_internal: + + mov r12,QWORD[rbp] + lea rbx,[r9*1+rdi] + mov rcx,r9 +DB 102,72,15,126,207 + neg rax +DB 102,72,15,126,206 + sar rcx,3+2 + dec r12 + xor r10,r10 + mov r13,QWORD[8+rbp] + mov r14,QWORD[16+rbp] + mov r15,QWORD[24+rbp] + jmp NEAR $L$sqr4x_sub_entry + +ALIGN 16 +$L$sqr4x_sub: + mov r12,QWORD[rbp] + mov r13,QWORD[8+rbp] + mov r14,QWORD[16+rbp] + mov r15,QWORD[24+rbp] +$L$sqr4x_sub_entry: + lea rbp,[32+rbp] + not r12 + not r13 + not r14 + not r15 + and r12,rax + and r13,rax + and r14,rax + and r15,rax + + neg r10 + adc r12,QWORD[rbx] + adc r13,QWORD[8+rbx] + adc r14,QWORD[16+rbx] + adc r15,QWORD[24+rbx] + mov QWORD[rdi],r12 + lea rbx,[32+rbx] + mov QWORD[8+rdi],r13 + sbb r10,r10 + mov QWORD[16+rdi],r14 + mov QWORD[24+rdi],r15 + lea rdi,[32+rdi] + + inc rcx + jnz NEAR $L$sqr4x_sub + + mov r10,r9 + neg r9 + DB 0F3h,0C3h ;repret + + + +ALIGN 32 +bn_mulx4x_mont_gather5: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[16+rsp],rsi + mov rax,rsp +$L$SEH_begin_bn_mulx4x_mont_gather5: + mov rdi,rcx + mov rsi,rdx + mov rdx,r8 + mov rcx,r9 + mov r8,QWORD[40+rsp] + mov r9,QWORD[48+rsp] + + + + mov rax,rsp + +$L$mulx4x_enter: + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + +$L$mulx4x_prologue: + + shl r9d,3 + lea r10,[r9*2+r9] + neg r9 + mov r8,QWORD[r8] + + + + + + + + + + + lea r11,[((-320))+r9*2+rsp] + mov rbp,rsp + sub r11,rdi + and r11,4095 + cmp r10,r11 + jb NEAR $L$mulx4xsp_alt + sub rbp,r11 + lea rbp,[((-320))+r9*2+rbp] + jmp NEAR $L$mulx4xsp_done + +$L$mulx4xsp_alt: + lea r10,[((4096-320))+r9*2] + lea rbp,[((-320))+r9*2+rbp] + sub r11,r10 + mov r10,0 + cmovc r11,r10 + sub rbp,r11 +$L$mulx4xsp_done: + and rbp,-64 + mov r11,rsp + sub r11,rbp + and r11,-4096 + lea rsp,[rbp*1+r11] + mov r10,QWORD[rsp] + cmp rsp,rbp + ja NEAR $L$mulx4x_page_walk + jmp NEAR $L$mulx4x_page_walk_done + +$L$mulx4x_page_walk: + lea rsp,[((-4096))+rsp] + mov r10,QWORD[rsp] + cmp rsp,rbp + ja NEAR $L$mulx4x_page_walk +$L$mulx4x_page_walk_done: + + + + + + + + + + + + + + mov QWORD[32+rsp],r8 + mov QWORD[40+rsp],rax + +$L$mulx4x_body: + call mulx4x_internal + + mov rsi,QWORD[40+rsp] + + mov rax,1 + + mov r15,QWORD[((-48))+rsi] + + mov r14,QWORD[((-40))+rsi] + + mov r13,QWORD[((-32))+rsi] + + mov r12,QWORD[((-24))+rsi] + + mov rbp,QWORD[((-16))+rsi] + + mov rbx,QWORD[((-8))+rsi] + + lea rsp,[rsi] + +$L$mulx4x_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_bn_mulx4x_mont_gather5: + + +ALIGN 32 +mulx4x_internal: + + mov QWORD[8+rsp],r9 + mov r10,r9 + neg r9 + shl r9,5 + neg r10 + lea r13,[128+r9*1+rdx] + shr r9,5+5 + movd xmm5,DWORD[56+rax] + sub r9,1 + lea rax,[$L$inc] + mov QWORD[((16+8))+rsp],r13 + mov QWORD[((24+8))+rsp],r9 + mov QWORD[((56+8))+rsp],rdi + movdqa xmm0,XMMWORD[rax] + movdqa xmm1,XMMWORD[16+rax] + lea r10,[((88-112))+r10*1+rsp] + lea rdi,[128+rdx] + + pshufd xmm5,xmm5,0 + movdqa xmm4,xmm1 +DB 0x67 + movdqa xmm2,xmm1 +DB 0x67 + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa xmm3,xmm4 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[112+r10],xmm0 + movdqa xmm0,xmm4 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[128+r10],xmm1 + movdqa xmm1,xmm4 + + paddd xmm0,xmm3 + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[144+r10],xmm2 + movdqa xmm2,xmm4 + + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa XMMWORD[160+r10],xmm3 + movdqa xmm3,xmm4 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[176+r10],xmm0 + movdqa xmm0,xmm4 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[192+r10],xmm1 + movdqa xmm1,xmm4 + + paddd xmm0,xmm3 + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[208+r10],xmm2 + movdqa xmm2,xmm4 + + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa XMMWORD[224+r10],xmm3 + movdqa xmm3,xmm4 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[240+r10],xmm0 + movdqa xmm0,xmm4 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[256+r10],xmm1 + movdqa xmm1,xmm4 + + paddd xmm0,xmm3 + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[272+r10],xmm2 + movdqa xmm2,xmm4 + + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa XMMWORD[288+r10],xmm3 + movdqa xmm3,xmm4 +DB 0x67 + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[304+r10],xmm0 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[320+r10],xmm1 + + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[336+r10],xmm2 + + pand xmm0,XMMWORD[64+rdi] + pand xmm1,XMMWORD[80+rdi] + pand xmm2,XMMWORD[96+rdi] + movdqa XMMWORD[352+r10],xmm3 + pand xmm3,XMMWORD[112+rdi] + por xmm0,xmm2 + por xmm1,xmm3 + movdqa xmm4,XMMWORD[((-128))+rdi] + movdqa xmm5,XMMWORD[((-112))+rdi] + movdqa xmm2,XMMWORD[((-96))+rdi] + pand xmm4,XMMWORD[112+r10] + movdqa xmm3,XMMWORD[((-80))+rdi] + pand xmm5,XMMWORD[128+r10] + por xmm0,xmm4 + pand xmm2,XMMWORD[144+r10] + por xmm1,xmm5 + pand xmm3,XMMWORD[160+r10] + por xmm0,xmm2 + por xmm1,xmm3 + movdqa xmm4,XMMWORD[((-64))+rdi] + movdqa xmm5,XMMWORD[((-48))+rdi] + movdqa xmm2,XMMWORD[((-32))+rdi] + pand xmm4,XMMWORD[176+r10] + movdqa xmm3,XMMWORD[((-16))+rdi] + pand xmm5,XMMWORD[192+r10] + por xmm0,xmm4 + pand xmm2,XMMWORD[208+r10] + por xmm1,xmm5 + pand xmm3,XMMWORD[224+r10] + por xmm0,xmm2 + por xmm1,xmm3 + movdqa xmm4,XMMWORD[rdi] + movdqa xmm5,XMMWORD[16+rdi] + movdqa xmm2,XMMWORD[32+rdi] + pand xmm4,XMMWORD[240+r10] + movdqa xmm3,XMMWORD[48+rdi] + pand xmm5,XMMWORD[256+r10] + por xmm0,xmm4 + pand xmm2,XMMWORD[272+r10] + por xmm1,xmm5 + pand xmm3,XMMWORD[288+r10] + por xmm0,xmm2 + por xmm1,xmm3 + pxor xmm0,xmm1 + pshufd xmm1,xmm0,0x4e + por xmm0,xmm1 + lea rdi,[256+rdi] +DB 102,72,15,126,194 + lea rbx,[((64+32+8))+rsp] + + mov r9,rdx + mulx rax,r8,QWORD[rsi] + mulx r12,r11,QWORD[8+rsi] + add r11,rax + mulx r13,rax,QWORD[16+rsi] + adc r12,rax + adc r13,0 + mulx r14,rax,QWORD[24+rsi] + + mov r15,r8 + imul r8,QWORD[((32+8))+rsp] + xor rbp,rbp + mov rdx,r8 + + mov QWORD[((8+8))+rsp],rdi + + lea rsi,[32+rsi] + adcx r13,rax + adcx r14,rbp + + mulx r10,rax,QWORD[rcx] + adcx r15,rax + adox r10,r11 + mulx r11,rax,QWORD[8+rcx] + adcx r10,rax + adox r11,r12 + mulx r12,rax,QWORD[16+rcx] + mov rdi,QWORD[((24+8))+rsp] + mov QWORD[((-32))+rbx],r10 + adcx r11,rax + adox r12,r13 + mulx r15,rax,QWORD[24+rcx] + mov rdx,r9 + mov QWORD[((-24))+rbx],r11 + adcx r12,rax + adox r15,rbp + lea rcx,[32+rcx] + mov QWORD[((-16))+rbx],r12 + jmp NEAR $L$mulx4x_1st + +ALIGN 32 +$L$mulx4x_1st: + adcx r15,rbp + mulx rax,r10,QWORD[rsi] + adcx r10,r14 + mulx r14,r11,QWORD[8+rsi] + adcx r11,rax + mulx rax,r12,QWORD[16+rsi] + adcx r12,r14 + mulx r14,r13,QWORD[24+rsi] +DB 0x67,0x67 + mov rdx,r8 + adcx r13,rax + adcx r14,rbp + lea rsi,[32+rsi] + lea rbx,[32+rbx] + + adox r10,r15 + mulx r15,rax,QWORD[rcx] + adcx r10,rax + adox r11,r15 + mulx r15,rax,QWORD[8+rcx] + adcx r11,rax + adox r12,r15 + mulx r15,rax,QWORD[16+rcx] + mov QWORD[((-40))+rbx],r10 + adcx r12,rax + mov QWORD[((-32))+rbx],r11 + adox r13,r15 + mulx r15,rax,QWORD[24+rcx] + mov rdx,r9 + mov QWORD[((-24))+rbx],r12 + adcx r13,rax + adox r15,rbp + lea rcx,[32+rcx] + mov QWORD[((-16))+rbx],r13 + + dec rdi + jnz NEAR $L$mulx4x_1st + + mov rax,QWORD[8+rsp] + adc r15,rbp + lea rsi,[rax*1+rsi] + add r14,r15 + mov rdi,QWORD[((8+8))+rsp] + adc rbp,rbp + mov QWORD[((-8))+rbx],r14 + jmp NEAR $L$mulx4x_outer + +ALIGN 32 +$L$mulx4x_outer: + lea r10,[((16-256))+rbx] + pxor xmm4,xmm4 +DB 0x67,0x67 + pxor xmm5,xmm5 + movdqa xmm0,XMMWORD[((-128))+rdi] + movdqa xmm1,XMMWORD[((-112))+rdi] + movdqa xmm2,XMMWORD[((-96))+rdi] + pand xmm0,XMMWORD[256+r10] + movdqa xmm3,XMMWORD[((-80))+rdi] + pand xmm1,XMMWORD[272+r10] + por xmm4,xmm0 + pand xmm2,XMMWORD[288+r10] + por xmm5,xmm1 + pand xmm3,XMMWORD[304+r10] + por xmm4,xmm2 + por xmm5,xmm3 + movdqa xmm0,XMMWORD[((-64))+rdi] + movdqa xmm1,XMMWORD[((-48))+rdi] + movdqa xmm2,XMMWORD[((-32))+rdi] + pand xmm0,XMMWORD[320+r10] + movdqa xmm3,XMMWORD[((-16))+rdi] + pand xmm1,XMMWORD[336+r10] + por xmm4,xmm0 + pand xmm2,XMMWORD[352+r10] + por xmm5,xmm1 + pand xmm3,XMMWORD[368+r10] + por xmm4,xmm2 + por xmm5,xmm3 + movdqa xmm0,XMMWORD[rdi] + movdqa xmm1,XMMWORD[16+rdi] + movdqa xmm2,XMMWORD[32+rdi] + pand xmm0,XMMWORD[384+r10] + movdqa xmm3,XMMWORD[48+rdi] + pand xmm1,XMMWORD[400+r10] + por xmm4,xmm0 + pand xmm2,XMMWORD[416+r10] + por xmm5,xmm1 + pand xmm3,XMMWORD[432+r10] + por xmm4,xmm2 + por xmm5,xmm3 + movdqa xmm0,XMMWORD[64+rdi] + movdqa xmm1,XMMWORD[80+rdi] + movdqa xmm2,XMMWORD[96+rdi] + pand xmm0,XMMWORD[448+r10] + movdqa xmm3,XMMWORD[112+rdi] + pand xmm1,XMMWORD[464+r10] + por xmm4,xmm0 + pand xmm2,XMMWORD[480+r10] + por xmm5,xmm1 + pand xmm3,XMMWORD[496+r10] + por xmm4,xmm2 + por xmm5,xmm3 + por xmm4,xmm5 + pshufd xmm0,xmm4,0x4e + por xmm0,xmm4 + lea rdi,[256+rdi] +DB 102,72,15,126,194 + + mov QWORD[rbx],rbp + lea rbx,[32+rax*1+rbx] + mulx r11,r8,QWORD[rsi] + xor rbp,rbp + mov r9,rdx + mulx r12,r14,QWORD[8+rsi] + adox r8,QWORD[((-32))+rbx] + adcx r11,r14 + mulx r13,r15,QWORD[16+rsi] + adox r11,QWORD[((-24))+rbx] + adcx r12,r15 + mulx r14,rdx,QWORD[24+rsi] + adox r12,QWORD[((-16))+rbx] + adcx r13,rdx + lea rcx,[rax*1+rcx] + lea rsi,[32+rsi] + adox r13,QWORD[((-8))+rbx] + adcx r14,rbp + adox r14,rbp + + mov r15,r8 + imul r8,QWORD[((32+8))+rsp] + + mov rdx,r8 + xor rbp,rbp + mov QWORD[((8+8))+rsp],rdi + + mulx r10,rax,QWORD[rcx] + adcx r15,rax + adox r10,r11 + mulx r11,rax,QWORD[8+rcx] + adcx r10,rax + adox r11,r12 + mulx r12,rax,QWORD[16+rcx] + adcx r11,rax + adox r12,r13 + mulx r15,rax,QWORD[24+rcx] + mov rdx,r9 + mov rdi,QWORD[((24+8))+rsp] + mov QWORD[((-32))+rbx],r10 + adcx r12,rax + mov QWORD[((-24))+rbx],r11 + adox r15,rbp + mov QWORD[((-16))+rbx],r12 + lea rcx,[32+rcx] + jmp NEAR $L$mulx4x_inner + +ALIGN 32 +$L$mulx4x_inner: + mulx rax,r10,QWORD[rsi] + adcx r15,rbp + adox r10,r14 + mulx r14,r11,QWORD[8+rsi] + adcx r10,QWORD[rbx] + adox r11,rax + mulx rax,r12,QWORD[16+rsi] + adcx r11,QWORD[8+rbx] + adox r12,r14 + mulx r14,r13,QWORD[24+rsi] + mov rdx,r8 + adcx r12,QWORD[16+rbx] + adox r13,rax + adcx r13,QWORD[24+rbx] + adox r14,rbp + lea rsi,[32+rsi] + lea rbx,[32+rbx] + adcx r14,rbp + + adox r10,r15 + mulx r15,rax,QWORD[rcx] + adcx r10,rax + adox r11,r15 + mulx r15,rax,QWORD[8+rcx] + adcx r11,rax + adox r12,r15 + mulx r15,rax,QWORD[16+rcx] + mov QWORD[((-40))+rbx],r10 + adcx r12,rax + adox r13,r15 + mov QWORD[((-32))+rbx],r11 + mulx r15,rax,QWORD[24+rcx] + mov rdx,r9 + lea rcx,[32+rcx] + mov QWORD[((-24))+rbx],r12 + adcx r13,rax + adox r15,rbp + mov QWORD[((-16))+rbx],r13 + + dec rdi + jnz NEAR $L$mulx4x_inner + + mov rax,QWORD[((0+8))+rsp] + adc r15,rbp + sub rdi,QWORD[rbx] + mov rdi,QWORD[((8+8))+rsp] + mov r10,QWORD[((16+8))+rsp] + adc r14,r15 + lea rsi,[rax*1+rsi] + adc rbp,rbp + mov QWORD[((-8))+rbx],r14 + + cmp rdi,r10 + jb NEAR $L$mulx4x_outer + + mov r10,QWORD[((-8))+rcx] + mov r8,rbp + mov r12,QWORD[rax*1+rcx] + lea rbp,[rax*1+rcx] + mov rcx,rax + lea rdi,[rax*1+rbx] + xor eax,eax + xor r15,r15 + sub r10,r14 + adc r15,r15 + or r8,r15 + sar rcx,3+2 + sub rax,r8 + mov rdx,QWORD[((56+8))+rsp] + dec r12 + mov r13,QWORD[8+rbp] + xor r8,r8 + mov r14,QWORD[16+rbp] + mov r15,QWORD[24+rbp] + jmp NEAR $L$sqrx4x_sub_entry + + + +ALIGN 32 +bn_powerx5: + mov QWORD[8+rsp],rdi ;WIN64 prologue + mov QWORD[16+rsp],rsi + mov rax,rsp +$L$SEH_begin_bn_powerx5: + mov rdi,rcx + mov rsi,rdx + mov rdx,r8 + mov rcx,r9 + mov r8,QWORD[40+rsp] + mov r9,QWORD[48+rsp] + + + + mov rax,rsp + +$L$powerx5_enter: + push rbx + + push rbp + + push r12 + + push r13 + + push r14 + + push r15 + +$L$powerx5_prologue: + + shl r9d,3 + lea r10,[r9*2+r9] + neg r9 + mov r8,QWORD[r8] + + + + + + + + + lea r11,[((-320))+r9*2+rsp] + mov rbp,rsp + sub r11,rdi + and r11,4095 + cmp r10,r11 + jb NEAR $L$pwrx_sp_alt + sub rbp,r11 + lea rbp,[((-320))+r9*2+rbp] + jmp NEAR $L$pwrx_sp_done + +ALIGN 32 +$L$pwrx_sp_alt: + lea r10,[((4096-320))+r9*2] + lea rbp,[((-320))+r9*2+rbp] + sub r11,r10 + mov r10,0 + cmovc r11,r10 + sub rbp,r11 +$L$pwrx_sp_done: + and rbp,-64 + mov r11,rsp + sub r11,rbp + and r11,-4096 + lea rsp,[rbp*1+r11] + mov r10,QWORD[rsp] + cmp rsp,rbp + ja NEAR $L$pwrx_page_walk + jmp NEAR $L$pwrx_page_walk_done + +$L$pwrx_page_walk: + lea rsp,[((-4096))+rsp] + mov r10,QWORD[rsp] + cmp rsp,rbp + ja NEAR $L$pwrx_page_walk +$L$pwrx_page_walk_done: + + mov r10,r9 + neg r9 + + + + + + + + + + + + + pxor xmm0,xmm0 +DB 102,72,15,110,207 +DB 102,72,15,110,209 +DB 102,73,15,110,218 +DB 102,72,15,110,226 + mov QWORD[32+rsp],r8 + mov QWORD[40+rsp],rax + +$L$powerx5_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 + + mov r9,r10 + mov rdi,rsi +DB 102,72,15,126,209 +DB 102,72,15,126,226 + mov rax,QWORD[40+rsp] + + call mulx4x_internal + + mov rsi,QWORD[40+rsp] + + mov rax,1 + + mov r15,QWORD[((-48))+rsi] + + mov r14,QWORD[((-40))+rsi] + + mov r13,QWORD[((-32))+rsi] + + mov r12,QWORD[((-24))+rsi] + + mov rbp,QWORD[((-16))+rsi] + + mov rbx,QWORD[((-8))+rsi] + + lea rsp,[rsi] + +$L$powerx5_epilogue: + mov rdi,QWORD[8+rsp] ;WIN64 epilogue + mov rsi,QWORD[16+rsp] + DB 0F3h,0C3h ;repret + +$L$SEH_end_bn_powerx5: + +global bn_sqrx8x_internal + + +ALIGN 32 +bn_sqrx8x_internal: +__bn_sqrx8x_internal: + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + lea rdi,[((48+8))+rsp] + lea rbp,[r9*1+rsi] + mov QWORD[((0+8))+rsp],r9 + mov QWORD[((8+8))+rsp],rbp + jmp NEAR $L$sqr8x_zero_start + +ALIGN 32 +DB 0x66,0x66,0x66,0x2e,0x0f,0x1f,0x84,0x00,0x00,0x00,0x00,0x00 +$L$sqrx8x_zero: +DB 0x3e + movdqa XMMWORD[rdi],xmm0 + movdqa XMMWORD[16+rdi],xmm0 + movdqa XMMWORD[32+rdi],xmm0 + movdqa XMMWORD[48+rdi],xmm0 +$L$sqr8x_zero_start: + movdqa XMMWORD[64+rdi],xmm0 + movdqa XMMWORD[80+rdi],xmm0 + movdqa XMMWORD[96+rdi],xmm0 + movdqa XMMWORD[112+rdi],xmm0 + lea rdi,[128+rdi] + sub r9,64 + jnz NEAR $L$sqrx8x_zero + + mov rdx,QWORD[rsi] + + xor r10,r10 + xor r11,r11 + xor r12,r12 + xor r13,r13 + xor r14,r14 + xor r15,r15 + lea rdi,[((48+8))+rsp] + xor rbp,rbp + jmp NEAR $L$sqrx8x_outer_loop + +ALIGN 32 +$L$sqrx8x_outer_loop: + mulx rax,r8,QWORD[8+rsi] + adcx r8,r9 + adox r10,rax + mulx rax,r9,QWORD[16+rsi] + adcx r9,r10 + adox r11,rax +DB 0xc4,0xe2,0xab,0xf6,0x86,0x18,0x00,0x00,0x00 + adcx r10,r11 + adox r12,rax +DB 0xc4,0xe2,0xa3,0xf6,0x86,0x20,0x00,0x00,0x00 + adcx r11,r12 + adox r13,rax + mulx rax,r12,QWORD[40+rsi] + adcx r12,r13 + adox r14,rax + mulx rax,r13,QWORD[48+rsi] + adcx r13,r14 + adox rax,r15 + mulx r15,r14,QWORD[56+rsi] + mov rdx,QWORD[8+rsi] + adcx r14,rax + adox r15,rbp + adc r15,QWORD[64+rdi] + mov QWORD[8+rdi],r8 + mov QWORD[16+rdi],r9 + sbb rcx,rcx + xor rbp,rbp + + + mulx rbx,r8,QWORD[16+rsi] + mulx rax,r9,QWORD[24+rsi] + adcx r8,r10 + adox r9,rbx + mulx rbx,r10,QWORD[32+rsi] + adcx r9,r11 + adox r10,rax +DB 0xc4,0xe2,0xa3,0xf6,0x86,0x28,0x00,0x00,0x00 + adcx r10,r12 + adox r11,rbx +DB 0xc4,0xe2,0x9b,0xf6,0x9e,0x30,0x00,0x00,0x00 + adcx r11,r13 + adox r12,r14 +DB 0xc4,0x62,0x93,0xf6,0xb6,0x38,0x00,0x00,0x00 + mov rdx,QWORD[16+rsi] + adcx r12,rax + adox r13,rbx + adcx r13,r15 + adox r14,rbp + adcx r14,rbp + + mov QWORD[24+rdi],r8 + mov QWORD[32+rdi],r9 + + mulx rbx,r8,QWORD[24+rsi] + mulx rax,r9,QWORD[32+rsi] + adcx r8,r10 + adox r9,rbx + mulx rbx,r10,QWORD[40+rsi] + adcx r9,r11 + adox r10,rax +DB 0xc4,0xe2,0xa3,0xf6,0x86,0x30,0x00,0x00,0x00 + adcx r10,r12 + adox r11,r13 +DB 0xc4,0x62,0x9b,0xf6,0xae,0x38,0x00,0x00,0x00 +DB 0x3e + mov rdx,QWORD[24+rsi] + adcx r11,rbx + adox r12,rax + adcx r12,r14 + mov QWORD[40+rdi],r8 + mov QWORD[48+rdi],r9 + mulx rax,r8,QWORD[32+rsi] + adox r13,rbp + adcx r13,rbp + + mulx rbx,r9,QWORD[40+rsi] + adcx r8,r10 + adox r9,rax + mulx rax,r10,QWORD[48+rsi] + adcx r9,r11 + adox r10,r12 + mulx r12,r11,QWORD[56+rsi] + mov rdx,QWORD[32+rsi] + mov r14,QWORD[40+rsi] + adcx r10,rbx + adox r11,rax + mov r15,QWORD[48+rsi] + adcx r11,r13 + adox r12,rbp + adcx r12,rbp + + mov QWORD[56+rdi],r8 + mov QWORD[64+rdi],r9 + + mulx rax,r9,r14 + mov r8,QWORD[56+rsi] + adcx r9,r10 + mulx rbx,r10,r15 + adox r10,rax + adcx r10,r11 + mulx rax,r11,r8 + mov rdx,r14 + adox r11,rbx + adcx r11,r12 + + adcx rax,rbp + + mulx rbx,r14,r15 + mulx r13,r12,r8 + mov rdx,r15 + lea rsi,[64+rsi] + adcx r11,r14 + adox r12,rbx + adcx r12,rax + adox r13,rbp + +DB 0x67,0x67 + mulx r14,r8,r8 + adcx r13,r8 + adcx r14,rbp + + cmp rsi,QWORD[((8+8))+rsp] + je NEAR $L$sqrx8x_outer_break + + neg rcx + mov rcx,-8 + mov r15,rbp + mov r8,QWORD[64+rdi] + adcx r9,QWORD[72+rdi] + adcx r10,QWORD[80+rdi] + adcx r11,QWORD[88+rdi] + adc r12,QWORD[96+rdi] + adc r13,QWORD[104+rdi] + adc r14,QWORD[112+rdi] + adc r15,QWORD[120+rdi] + lea rbp,[rsi] + lea rdi,[128+rdi] + sbb rax,rax + + mov rdx,QWORD[((-64))+rsi] + mov QWORD[((16+8))+rsp],rax + mov QWORD[((24+8))+rsp],rdi + + + xor eax,eax + jmp NEAR $L$sqrx8x_loop + +ALIGN 32 +$L$sqrx8x_loop: + mov rbx,r8 + mulx r8,rax,QWORD[rbp] + adcx rbx,rax + adox r8,r9 + + mulx r9,rax,QWORD[8+rbp] + adcx r8,rax + adox r9,r10 + + mulx r10,rax,QWORD[16+rbp] + adcx r9,rax + adox r10,r11 + + mulx r11,rax,QWORD[24+rbp] + adcx r10,rax + adox r11,r12 + +DB 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00 + adcx r11,rax + adox r12,r13 + + mulx r13,rax,QWORD[40+rbp] + adcx r12,rax + adox r13,r14 + + mulx r14,rax,QWORD[48+rbp] + mov QWORD[rcx*8+rdi],rbx + mov ebx,0 + adcx r13,rax + adox r14,r15 + +DB 0xc4,0x62,0xfb,0xf6,0xbd,0x38,0x00,0x00,0x00 + mov rdx,QWORD[8+rcx*8+rsi] + adcx r14,rax + adox r15,rbx + adcx r15,rbx + +DB 0x67 + inc rcx + jnz NEAR $L$sqrx8x_loop + + lea rbp,[64+rbp] + mov rcx,-8 + cmp rbp,QWORD[((8+8))+rsp] + je NEAR $L$sqrx8x_break + + sub rbx,QWORD[((16+8))+rsp] +DB 0x66 + mov rdx,QWORD[((-64))+rsi] + adcx r8,QWORD[rdi] + adcx r9,QWORD[8+rdi] + adc r10,QWORD[16+rdi] + adc r11,QWORD[24+rdi] + adc r12,QWORD[32+rdi] + adc r13,QWORD[40+rdi] + adc r14,QWORD[48+rdi] + adc r15,QWORD[56+rdi] + lea rdi,[64+rdi] +DB 0x67 + sbb rax,rax + xor ebx,ebx + mov QWORD[((16+8))+rsp],rax + jmp NEAR $L$sqrx8x_loop + +ALIGN 32 +$L$sqrx8x_break: + xor rbp,rbp + sub rbx,QWORD[((16+8))+rsp] + adcx r8,rbp + mov rcx,QWORD[((24+8))+rsp] + adcx r9,rbp + mov rdx,QWORD[rsi] + adc r10,0 + mov QWORD[rdi],r8 + adc r11,0 + adc r12,0 + adc r13,0 + adc r14,0 + adc r15,0 + cmp rdi,rcx + je NEAR $L$sqrx8x_outer_loop + + mov QWORD[8+rdi],r9 + mov r9,QWORD[8+rcx] + mov QWORD[16+rdi],r10 + mov r10,QWORD[16+rcx] + mov QWORD[24+rdi],r11 + mov r11,QWORD[24+rcx] + mov QWORD[32+rdi],r12 + mov r12,QWORD[32+rcx] + mov QWORD[40+rdi],r13 + mov r13,QWORD[40+rcx] + mov QWORD[48+rdi],r14 + mov r14,QWORD[48+rcx] + mov QWORD[56+rdi],r15 + mov r15,QWORD[56+rcx] + mov rdi,rcx + jmp NEAR $L$sqrx8x_outer_loop + +ALIGN 32 +$L$sqrx8x_outer_break: + mov QWORD[72+rdi],r9 +DB 102,72,15,126,217 + mov QWORD[80+rdi],r10 + mov QWORD[88+rdi],r11 + mov QWORD[96+rdi],r12 + mov QWORD[104+rdi],r13 + mov QWORD[112+rdi],r14 + lea rdi,[((48+8))+rsp] + mov rdx,QWORD[rcx*1+rsi] + + mov r11,QWORD[8+rdi] + xor r10,r10 + mov r9,QWORD[((0+8))+rsp] + adox r11,r11 + mov r12,QWORD[16+rdi] + mov r13,QWORD[24+rdi] + + +ALIGN 32 +$L$sqrx4x_shift_n_add: + mulx rbx,rax,rdx + adox r12,r12 + adcx rax,r10 +DB 0x48,0x8b,0x94,0x0e,0x08,0x00,0x00,0x00 +DB 0x4c,0x8b,0x97,0x20,0x00,0x00,0x00 + adox r13,r13 + adcx rbx,r11 + mov r11,QWORD[40+rdi] + mov QWORD[rdi],rax + mov QWORD[8+rdi],rbx + + mulx rbx,rax,rdx + adox r10,r10 + adcx rax,r12 + mov rdx,QWORD[16+rcx*1+rsi] + mov r12,QWORD[48+rdi] + adox r11,r11 + adcx rbx,r13 + mov r13,QWORD[56+rdi] + mov QWORD[16+rdi],rax + mov QWORD[24+rdi],rbx + + mulx rbx,rax,rdx + adox r12,r12 + adcx rax,r10 + mov rdx,QWORD[24+rcx*1+rsi] + lea rcx,[32+rcx] + mov r10,QWORD[64+rdi] + adox r13,r13 + adcx rbx,r11 + mov r11,QWORD[72+rdi] + mov QWORD[32+rdi],rax + mov QWORD[40+rdi],rbx + + mulx rbx,rax,rdx + adox r10,r10 + adcx rax,r12 + jrcxz $L$sqrx4x_shift_n_add_break +DB 0x48,0x8b,0x94,0x0e,0x00,0x00,0x00,0x00 + adox r11,r11 + adcx rbx,r13 + mov r12,QWORD[80+rdi] + mov r13,QWORD[88+rdi] + mov QWORD[48+rdi],rax + mov QWORD[56+rdi],rbx + lea rdi,[64+rdi] + nop + jmp NEAR $L$sqrx4x_shift_n_add + +ALIGN 32 +$L$sqrx4x_shift_n_add_break: + adcx rbx,r13 + mov QWORD[48+rdi],rax + mov QWORD[56+rdi],rbx + lea rdi,[64+rdi] +DB 102,72,15,126,213 +__bn_sqrx8x_reduction: + xor eax,eax + mov rbx,QWORD[((32+8))+rsp] + mov rdx,QWORD[((48+8))+rsp] + lea rcx,[((-64))+r9*1+rbp] + + mov QWORD[((0+8))+rsp],rcx + mov QWORD[((8+8))+rsp],rdi + + lea rdi,[((48+8))+rsp] + jmp NEAR $L$sqrx8x_reduction_loop + +ALIGN 32 +$L$sqrx8x_reduction_loop: + mov r9,QWORD[8+rdi] + mov r10,QWORD[16+rdi] + mov r11,QWORD[24+rdi] + mov r12,QWORD[32+rdi] + mov r8,rdx + imul rdx,rbx + mov r13,QWORD[40+rdi] + mov r14,QWORD[48+rdi] + mov r15,QWORD[56+rdi] + mov QWORD[((24+8))+rsp],rax + + lea rdi,[64+rdi] + xor rsi,rsi + mov rcx,-8 + jmp NEAR $L$sqrx8x_reduce + +ALIGN 32 +$L$sqrx8x_reduce: + mov rbx,r8 + mulx r8,rax,QWORD[rbp] + adcx rax,rbx + adox r8,r9 + + mulx r9,rbx,QWORD[8+rbp] + adcx r8,rbx + adox r9,r10 + + mulx r10,rbx,QWORD[16+rbp] + adcx r9,rbx + adox r10,r11 + + mulx r11,rbx,QWORD[24+rbp] + adcx r10,rbx + adox r11,r12 + +DB 0xc4,0x62,0xe3,0xf6,0xa5,0x20,0x00,0x00,0x00 + mov rax,rdx + mov rdx,r8 + adcx r11,rbx + adox r12,r13 + + mulx rdx,rbx,QWORD[((32+8))+rsp] + mov rdx,rax + mov QWORD[((64+48+8))+rcx*8+rsp],rax + + mulx r13,rax,QWORD[40+rbp] + adcx r12,rax + adox r13,r14 + + mulx r14,rax,QWORD[48+rbp] + adcx r13,rax + adox r14,r15 + + mulx r15,rax,QWORD[56+rbp] + mov rdx,rbx + adcx r14,rax + adox r15,rsi + adcx r15,rsi + +DB 0x67,0x67,0x67 + inc rcx + jnz NEAR $L$sqrx8x_reduce + + mov rax,rsi + cmp rbp,QWORD[((0+8))+rsp] + jae NEAR $L$sqrx8x_no_tail + + mov rdx,QWORD[((48+8))+rsp] + add r8,QWORD[rdi] + lea rbp,[64+rbp] + mov rcx,-8 + adcx r9,QWORD[8+rdi] + adcx r10,QWORD[16+rdi] + adc r11,QWORD[24+rdi] + adc r12,QWORD[32+rdi] + adc r13,QWORD[40+rdi] + adc r14,QWORD[48+rdi] + adc r15,QWORD[56+rdi] + lea rdi,[64+rdi] + sbb rax,rax + + xor rsi,rsi + mov QWORD[((16+8))+rsp],rax + jmp NEAR $L$sqrx8x_tail + +ALIGN 32 +$L$sqrx8x_tail: + mov rbx,r8 + mulx r8,rax,QWORD[rbp] + adcx rbx,rax + adox r8,r9 + + mulx r9,rax,QWORD[8+rbp] + adcx r8,rax + adox r9,r10 + + mulx r10,rax,QWORD[16+rbp] + adcx r9,rax + adox r10,r11 + + mulx r11,rax,QWORD[24+rbp] + adcx r10,rax + adox r11,r12 + +DB 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00 + adcx r11,rax + adox r12,r13 + + mulx r13,rax,QWORD[40+rbp] + adcx r12,rax + adox r13,r14 + + mulx r14,rax,QWORD[48+rbp] + adcx r13,rax + adox r14,r15 + + mulx r15,rax,QWORD[56+rbp] + mov rdx,QWORD[((72+48+8))+rcx*8+rsp] + adcx r14,rax + adox r15,rsi + mov QWORD[rcx*8+rdi],rbx + mov rbx,r8 + adcx r15,rsi + + inc rcx + jnz NEAR $L$sqrx8x_tail + + cmp rbp,QWORD[((0+8))+rsp] + jae NEAR $L$sqrx8x_tail_done + + sub rsi,QWORD[((16+8))+rsp] + mov rdx,QWORD[((48+8))+rsp] + lea rbp,[64+rbp] + adc r8,QWORD[rdi] + adc r9,QWORD[8+rdi] + adc r10,QWORD[16+rdi] + adc r11,QWORD[24+rdi] + adc r12,QWORD[32+rdi] + adc r13,QWORD[40+rdi] + adc r14,QWORD[48+rdi] + adc r15,QWORD[56+rdi] + lea rdi,[64+rdi] + sbb rax,rax + sub rcx,8 + + xor rsi,rsi + mov QWORD[((16+8))+rsp],rax + jmp NEAR $L$sqrx8x_tail + +ALIGN 32 +$L$sqrx8x_tail_done: + xor rax,rax + add r8,QWORD[((24+8))+rsp] + adc r9,0 + adc r10,0 + adc r11,0 + adc r12,0 + adc r13,0 + adc r14,0 + adc r15,0 + adc rax,0 + + sub rsi,QWORD[((16+8))+rsp] +$L$sqrx8x_no_tail: + adc r8,QWORD[rdi] +DB 102,72,15,126,217 + adc r9,QWORD[8+rdi] + mov rsi,QWORD[56+rbp] +DB 102,72,15,126,213 + adc r10,QWORD[16+rdi] + adc r11,QWORD[24+rdi] + adc r12,QWORD[32+rdi] + adc r13,QWORD[40+rdi] + adc r14,QWORD[48+rdi] + adc r15,QWORD[56+rdi] + adc rax,0 + + mov rbx,QWORD[((32+8))+rsp] + mov rdx,QWORD[64+rcx*1+rdi] + + mov QWORD[rdi],r8 + lea r8,[64+rdi] + mov QWORD[8+rdi],r9 + mov QWORD[16+rdi],r10 + mov QWORD[24+rdi],r11 + mov QWORD[32+rdi],r12 + mov QWORD[40+rdi],r13 + mov QWORD[48+rdi],r14 + mov QWORD[56+rdi],r15 + + lea rdi,[64+rcx*1+rdi] + cmp r8,QWORD[((8+8))+rsp] + jb NEAR $L$sqrx8x_reduction_loop + DB 0F3h,0C3h ;repret + + +ALIGN 32 +__bn_postx4x_internal: + + mov r12,QWORD[rbp] + mov r10,rcx + mov r9,rcx + neg rax + sar rcx,3+2 + +DB 102,72,15,126,202 +DB 102,72,15,126,206 + dec r12 + mov r13,QWORD[8+rbp] + xor r8,r8 + mov r14,QWORD[16+rbp] + mov r15,QWORD[24+rbp] + jmp NEAR $L$sqrx4x_sub_entry + +ALIGN 16 +$L$sqrx4x_sub: + mov r12,QWORD[rbp] + mov r13,QWORD[8+rbp] + mov r14,QWORD[16+rbp] + mov r15,QWORD[24+rbp] +$L$sqrx4x_sub_entry: + andn r12,r12,rax + lea rbp,[32+rbp] + andn r13,r13,rax + andn r14,r14,rax + andn r15,r15,rax + + neg r8 + adc r12,QWORD[rdi] + adc r13,QWORD[8+rdi] + adc r14,QWORD[16+rdi] + adc r15,QWORD[24+rdi] + mov QWORD[rdx],r12 + lea rdi,[32+rdi] + mov QWORD[8+rdx],r13 + sbb r8,r8 + mov QWORD[16+rdx],r14 + mov QWORD[24+rdx],r15 + lea rdx,[32+rdx] + + inc rcx + jnz NEAR $L$sqrx4x_sub + + neg r9 + + DB 0F3h,0C3h ;repret + + +global bn_get_bits5 + +ALIGN 16 +bn_get_bits5: + + lea r10,[rcx] + lea r11,[1+rcx] + mov ecx,edx + shr edx,4 + and ecx,15 + lea eax,[((-8))+rcx] + cmp ecx,11 + cmova r10,r11 + cmova ecx,eax + movzx eax,WORD[rdx*2+r10] + shr eax,cl + and eax,31 + DB 0F3h,0C3h ;repret + + + +global bn_scatter5 + +ALIGN 16 +bn_scatter5: + + cmp edx,0 + jz NEAR $L$scatter_epilogue + lea r8,[r9*8+r8] +$L$scatter: + mov rax,QWORD[rcx] + lea rcx,[8+rcx] + mov QWORD[r8],rax + lea r8,[256+r8] + sub edx,1 + jnz NEAR $L$scatter +$L$scatter_epilogue: + DB 0F3h,0C3h ;repret + + + +global bn_gather5 + +ALIGN 32 +bn_gather5: +$L$SEH_begin_bn_gather5: + + +DB 0x4c,0x8d,0x14,0x24 +DB 0x48,0x81,0xec,0x08,0x01,0x00,0x00 + lea rax,[$L$inc] + and rsp,-16 + + movd xmm5,r9d + movdqa xmm0,XMMWORD[rax] + movdqa xmm1,XMMWORD[16+rax] + lea r11,[128+r8] + lea rax,[128+rsp] + + pshufd xmm5,xmm5,0 + movdqa xmm4,xmm1 + movdqa xmm2,xmm1 + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa xmm3,xmm4 + + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[(-128)+rax],xmm0 + movdqa xmm0,xmm4 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[(-112)+rax],xmm1 + movdqa xmm1,xmm4 + + paddd xmm0,xmm3 + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[(-96)+rax],xmm2 + movdqa xmm2,xmm4 + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa XMMWORD[(-80)+rax],xmm3 + movdqa xmm3,xmm4 + + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[(-64)+rax],xmm0 + movdqa xmm0,xmm4 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[(-48)+rax],xmm1 + movdqa xmm1,xmm4 + + paddd xmm0,xmm3 + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[(-32)+rax],xmm2 + movdqa xmm2,xmm4 + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa XMMWORD[(-16)+rax],xmm3 + movdqa xmm3,xmm4 + + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[rax],xmm0 + movdqa xmm0,xmm4 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[16+rax],xmm1 + movdqa xmm1,xmm4 + + paddd xmm0,xmm3 + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[32+rax],xmm2 + movdqa xmm2,xmm4 + paddd xmm1,xmm0 + pcmpeqd xmm0,xmm5 + movdqa XMMWORD[48+rax],xmm3 + movdqa xmm3,xmm4 + + paddd xmm2,xmm1 + pcmpeqd xmm1,xmm5 + movdqa XMMWORD[64+rax],xmm0 + movdqa xmm0,xmm4 + + paddd xmm3,xmm2 + pcmpeqd xmm2,xmm5 + movdqa XMMWORD[80+rax],xmm1 + movdqa xmm1,xmm4 + + paddd xmm0,xmm3 + pcmpeqd xmm3,xmm5 + movdqa XMMWORD[96+rax],xmm2 + movdqa xmm2,xmm4 + movdqa XMMWORD[112+rax],xmm3 + jmp NEAR $L$gather + +ALIGN 32 +$L$gather: + pxor xmm4,xmm4 + pxor xmm5,xmm5 + movdqa xmm0,XMMWORD[((-128))+r11] + movdqa xmm1,XMMWORD[((-112))+r11] + movdqa xmm2,XMMWORD[((-96))+r11] + pand xmm0,XMMWORD[((-128))+rax] + movdqa xmm3,XMMWORD[((-80))+r11] + pand xmm1,XMMWORD[((-112))+rax] + por xmm4,xmm0 + pand xmm2,XMMWORD[((-96))+rax] + por xmm5,xmm1 + pand xmm3,XMMWORD[((-80))+rax] + por xmm4,xmm2 + por xmm5,xmm3 + movdqa xmm0,XMMWORD[((-64))+r11] + movdqa xmm1,XMMWORD[((-48))+r11] + movdqa xmm2,XMMWORD[((-32))+r11] + pand xmm0,XMMWORD[((-64))+rax] + movdqa xmm3,XMMWORD[((-16))+r11] + pand xmm1,XMMWORD[((-48))+rax] + por xmm4,xmm0 + pand xmm2,XMMWORD[((-32))+rax] + por xmm5,xmm1 + pand xmm3,XMMWORD[((-16))+rax] + por xmm4,xmm2 + por xmm5,xmm3 + movdqa xmm0,XMMWORD[r11] + movdqa xmm1,XMMWORD[16+r11] + movdqa xmm2,XMMWORD[32+r11] + pand xmm0,XMMWORD[rax] + movdqa xmm3,XMMWORD[48+r11] + pand xmm1,XMMWORD[16+rax] + por xmm4,xmm0 + pand xmm2,XMMWORD[32+rax] + por xmm5,xmm1 + pand xmm3,XMMWORD[48+rax] + por xmm4,xmm2 + por xmm5,xmm3 + movdqa xmm0,XMMWORD[64+r11] + movdqa xmm1,XMMWORD[80+r11] + movdqa xmm2,XMMWORD[96+r11] + pand xmm0,XMMWORD[64+rax] + movdqa xmm3,XMMWORD[112+r11] + pand xmm1,XMMWORD[80+rax] + por xmm4,xmm0 + pand xmm2,XMMWORD[96+rax] + por xmm5,xmm1 + pand xmm3,XMMWORD[112+rax] + por xmm4,xmm2 + por xmm5,xmm3 + por xmm4,xmm5 + lea r11,[256+r11] + pshufd xmm0,xmm4,0x4e + por xmm0,xmm4 + movq QWORD[rcx],xmm0 + lea rcx,[8+rcx] + sub edx,1 + jnz NEAR $L$gather + + lea rsp,[r10] + DB 0F3h,0C3h ;repret +$L$SEH_end_bn_gather5: + + +ALIGN 64 +$L$inc: + DD 0,0,1,1 + DD 2,2,2,2 +DB 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105 +DB 112,108,105,99,97,116,105,111,110,32,119,105,116,104,32,115 +DB 99,97,116,116,101,114,47,103,97,116,104,101,114,32,102,111 +DB 114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79 +DB 71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111 +DB 112,101,110,115,115,108,46,111,114,103,62,0 +EXTERN __imp_RtlVirtualUnwind + +ALIGN 16 +mul_handler: + push rsi + push rdi + push rbx + push rbp + push r12 + push r13 + push r14 + push r15 + pushfq + sub rsp,64 + + mov rax,QWORD[120+r8] + mov rbx,QWORD[248+r8] + + mov rsi,QWORD[8+r9] + mov r11,QWORD[56+r9] + + mov r10d,DWORD[r11] + lea r10,[r10*1+rsi] + cmp rbx,r10 + jb NEAR $L$common_seh_tail + + mov r10d,DWORD[4+r11] + lea r10,[r10*1+rsi] + cmp rbx,r10 + jb NEAR $L$common_pop_regs + + mov rax,QWORD[152+r8] + + mov r10d,DWORD[8+r11] + lea r10,[r10*1+rsi] + cmp rbx,r10 + jae NEAR $L$common_seh_tail + + lea r10,[$L$mul_epilogue] + cmp rbx,r10 + ja NEAR $L$body_40 + + mov r10,QWORD[192+r8] + mov rax,QWORD[8+r10*8+rax] + + jmp NEAR $L$common_pop_regs + +$L$body_40: + mov rax,QWORD[40+rax] +$L$common_pop_regs: + mov rbx,QWORD[((-8))+rax] + mov rbp,QWORD[((-16))+rax] + mov r12,QWORD[((-24))+rax] + mov r13,QWORD[((-32))+rax] + mov r14,QWORD[((-40))+rax] + mov r15,QWORD[((-48))+rax] + mov QWORD[144+r8],rbx + mov QWORD[160+r8],rbp + mov QWORD[216+r8],r12 + mov QWORD[224+r8],r13 + mov QWORD[232+r8],r14 + mov QWORD[240+r8],r15 + +$L$common_seh_tail: + mov rdi,QWORD[8+rax] + mov rsi,QWORD[16+rax] + mov QWORD[152+r8],rax + mov QWORD[168+r8],rsi + mov QWORD[176+r8],rdi + + mov rdi,QWORD[40+r9] + mov rsi,r8 + mov ecx,154 + DD 0xa548f3fc + + mov rsi,r9 + xor rcx,rcx + mov rdx,QWORD[8+rsi] + mov r8,QWORD[rsi] + mov r9,QWORD[16+rsi] + mov r10,QWORD[40+rsi] + lea r11,[56+rsi] + lea r12,[24+rsi] + mov QWORD[32+rsp],r10 + mov QWORD[40+rsp],r11 + mov QWORD[48+rsp],r12 + mov QWORD[56+rsp],rcx + call QWORD[__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 + + +section .pdata rdata align=4 +ALIGN 4 + DD $L$SEH_begin_bn_mul_mont_gather5 wrt ..imagebase + DD $L$SEH_end_bn_mul_mont_gather5 wrt ..imagebase + DD $L$SEH_info_bn_mul_mont_gather5 wrt ..imagebase + + DD $L$SEH_begin_bn_mul4x_mont_gather5 wrt ..imagebase + DD $L$SEH_end_bn_mul4x_mont_gather5 wrt ..imagebase + DD $L$SEH_info_bn_mul4x_mont_gather5 wrt ..imagebase + + DD $L$SEH_begin_bn_power5 wrt ..imagebase + DD $L$SEH_end_bn_power5 wrt ..imagebase + DD $L$SEH_info_bn_power5 wrt ..imagebase + DD $L$SEH_begin_bn_mulx4x_mont_gather5 wrt ..imagebase + DD $L$SEH_end_bn_mulx4x_mont_gather5 wrt ..imagebase + DD $L$SEH_info_bn_mulx4x_mont_gather5 wrt ..imagebase + + DD $L$SEH_begin_bn_powerx5 wrt ..imagebase + DD $L$SEH_end_bn_powerx5 wrt ..imagebase + DD $L$SEH_info_bn_powerx5 wrt ..imagebase + DD $L$SEH_begin_bn_gather5 wrt ..imagebase + DD $L$SEH_end_bn_gather5 wrt ..imagebase + DD $L$SEH_info_bn_gather5 wrt ..imagebase + +section .xdata rdata align=8 +ALIGN 8 +$L$SEH_info_bn_mul_mont_gather5: +DB 9,0,0,0 + DD mul_handler wrt ..imagebase + DD $L$mul_body wrt ..imagebase,$L$mul_body wrt ..imagebase,$L$mul_epilogue wrt ..imagebase +ALIGN 8 +$L$SEH_info_bn_mul4x_mont_gather5: +DB 9,0,0,0 + DD mul_handler wrt ..imagebase + DD $L$mul4x_prologue wrt ..imagebase,$L$mul4x_body wrt ..imagebase,$L$mul4x_epilogue wrt ..imagebase +ALIGN 8 +$L$SEH_info_bn_power5: +DB 9,0,0,0 + DD mul_handler wrt ..imagebase + DD $L$power5_prologue wrt ..imagebase,$L$power5_body wrt ..imagebase,$L$power5_epilogue wrt ..imagebase +ALIGN 8 +$L$SEH_info_bn_mulx4x_mont_gather5: +DB 9,0,0,0 + DD mul_handler wrt ..imagebase + DD $L$mulx4x_prologue wrt ..imagebase,$L$mulx4x_body wrt ..imagebase,$L$mulx4x_epilogue wrt ..imagebase +ALIGN 8 +$L$SEH_info_bn_powerx5: +DB 9,0,0,0 + DD mul_handler wrt ..imagebase + DD $L$powerx5_prologue wrt ..imagebase,$L$powerx5_body wrt ..imagebase,$L$powerx5_epilogue wrt ..imagebase +ALIGN 8 +$L$SEH_info_bn_gather5: +DB 0x01,0x0b,0x03,0x0a +DB 0x0b,0x01,0x21,0x00 +DB 0x04,0xa3,0x00,0x00 +ALIGN 8 |