-rw-r--r-- 29581 lib25519-20240321/crypto_nG/merged25519/amd64-avx512ifma-5l-maax/ge25519_base.S raw
#include "crypto_asm_hidden.h"
// linker define base
// linker use vec19
// linker use vec608
// linker use vecmask47
// linker use vecmask52
// linker use pmask1
// linker use pmask2
// linker use pmask3
// linker use pmask4
// linker use pmask5
// linker use pmask6
// linker use pmask7
// linker use pmask8
// linker use upmask1
// linker use upmask2
// linker use upmask3
// linker use upmask4
// linker use upmask5
// linker use upmask6
// linker use upmask7
// linker use mask63
/* Assembly for fixed base scalar multiplication.
*
* This assembly has been developed after studying the
* amd64-64-24k implementation of the work "High speed
* high security signatures" by Bernstein et al.
*/
#include "consts_namespace.h"
.p2align 5
ASM_HIDDEN _CRYPTO_SHARED_NAMESPACE(base)
.globl _CRYPTO_SHARED_NAMESPACE(base)
ASM_HIDDEN CRYPTO_SHARED_NAMESPACE(base)
.globl CRYPTO_SHARED_NAMESPACE(base)
_CRYPTO_SHARED_NAMESPACE(base):
CRYPTO_SHARED_NAMESPACE(base):
movq %rsp,%r11
andq $-32,%rsp
subq $512,%rsp
movq %r11,0(%rsp)
movq %r12,8(%rsp)
movq %r13,16(%rsp)
movq %r14,24(%rsp)
movq %r15,32(%rsp)
movq %rbx,40(%rsp)
movq %rbp,48(%rsp)
movq %rdi,56(%rsp)
movq %rsi,64(%rsp)
movq %rdx,72(%rsp)
/* choose t and initialize r */
movq %rdx,%rcx
movzbl 0(%rsi),%eax
movsbq %al,%rdx
movq $0,%rsi
imulq $768,%rsi,%rdi
movq %rdx,%rsi
sarq $7,%rsi
movq %rdx,%r8
addq %rsi,%r8
xorq %rsi,%r8
movq $1,%rsi
movq $0,%r9
movq $0,%rax
movq $0,%r10
movq $1,%r11
movq $0,%r12
movq $0,%r13
movq $0,%r14
cmpq $1,%r8
movq 0(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 8(%rcx,%rdi),%r15
cmove %r15,%r9
movq 16(%rcx,%rdi),%r15
cmove %r15,%rax
movq 24(%rcx,%rdi),%r15
cmove %r15,%r10
movq 32(%rcx,%rdi),%r15
cmove %r15,%r11
movq 40(%rcx,%rdi),%r15
cmove %r15,%r12
movq 48(%rcx,%rdi),%r15
cmove %r15,%r13
movq 56(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $2,%r8
movq 96(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 104(%rcx,%rdi),%r15
cmove %r15,%r9
movq 112(%rcx,%rdi),%r15
cmove %r15,%rax
movq 120(%rcx,%rdi),%r15
cmove %r15,%r10
movq 128(%rcx,%rdi),%r15
cmove %r15,%r11
movq 136(%rcx,%rdi),%r15
cmove %r15,%r12
movq 144(%rcx,%rdi),%r15
cmove %r15,%r13
movq 152(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $3,%r8
movq 192(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 200(%rcx,%rdi),%r15
cmove %r15,%r9
movq 208(%rcx,%rdi),%r15
cmove %r15,%rax
movq 216(%rcx,%rdi),%r15
cmove %r15,%r10
movq 224(%rcx,%rdi),%r15
cmove %r15,%r11
movq 232(%rcx,%rdi),%r15
cmove %r15,%r12
movq 240(%rcx,%rdi),%r15
cmove %r15,%r13
movq 248(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $4,%r8
movq 288(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 296(%rcx,%rdi),%r15
cmove %r15,%r9
movq 304(%rcx,%rdi),%r15
cmove %r15,%rax
movq 312(%rcx,%rdi),%r15
cmove %r15,%r10
movq 320(%rcx,%rdi),%r15
cmove %r15,%r11
movq 328(%rcx,%rdi),%r15
cmove %r15,%r12
movq 336(%rcx,%rdi),%r15
cmove %r15,%r13
movq 344(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $5,%r8
movq 384(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 392(%rcx,%rdi),%r15
cmove %r15,%r9
movq 400(%rcx,%rdi),%r15
cmove %r15,%rax
movq 408(%rcx,%rdi),%r15
cmove %r15,%r10
movq 416(%rcx,%rdi),%r15
cmove %r15,%r11
movq 424(%rcx,%rdi),%r15
cmove %r15,%r12
movq 432(%rcx,%rdi),%r15
cmove %r15,%r13
movq 440(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $6,%r8
movq 480(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 488(%rcx,%rdi),%r15
cmove %r15,%r9
movq 496(%rcx,%rdi),%r15
cmove %r15,%rax
movq 504(%rcx,%rdi),%r15
cmove %r15,%r10
movq 512(%rcx,%rdi),%r15
cmove %r15,%r11
movq 520(%rcx,%rdi),%r15
cmove %r15,%r12
movq 528(%rcx,%rdi),%r15
cmove %r15,%r13
movq 536(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $7,%r8
movq 576(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 584(%rcx,%rdi),%r15
cmove %r15,%r9
movq 592(%rcx,%rdi),%r15
cmove %r15,%rax
movq 600(%rcx,%rdi),%r15
cmove %r15,%r10
movq 608(%rcx,%rdi),%r15
cmove %r15,%r11
movq 616(%rcx,%rdi),%r15
cmove %r15,%r12
movq 624(%rcx,%rdi),%r15
cmove %r15,%r13
movq 632(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $8,%r8
movq 672(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 680(%rcx,%rdi),%r15
cmove %r15,%r9
movq 688(%rcx,%rdi),%r15
cmove %r15,%rax
movq 696(%rcx,%rdi),%r15
cmove %r15,%r10
movq 704(%rcx,%rdi),%r15
cmove %r15,%r11
movq 712(%rcx,%rdi),%r15
cmove %r15,%r12
movq 720(%rcx,%rdi),%r15
cmove %r15,%r13
movq 728(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $0,%rdx
movq %rsi,%r15
cmovl %r11,%rsi
cmovl %r15,%r11
movq %r9,%r15
cmovl %r12,%r9
cmovl %r15,%r12
movq %rax,%r15
cmovl %r13,%rax
cmovl %r15,%r13
movq %r10,%r15
cmovl %r14,%r10
cmovl %r15,%r14
movq %r11,%r15
movq %r12,80(%rsp)
movq %r13,88(%rsp)
movq %r14,96(%rsp)
// sub
subq %rsi,%r11
sbbq %r9,%r12
sbbq %rax,%r13
sbbq %r10,%r14
movq $0,%rbx
movq $38,%rbp
cmovae %rbx,%rbp
subq %rbp,%r11
sbbq %rbx,%r12
sbbq %rbx,%r13
sbbq %rbx,%r14
cmovc %rbp,%rbx
subq %rbx,%r11
movq %r11,128(%rsp)
movq %r12,136(%rsp)
movq %r13,144(%rsp)
movq %r14,152(%rsp)
// add
addq %r15,%rsi
adcq 80(%rsp),%r9
adcq 88(%rsp),%rax
adcq 96(%rsp),%r10
movq $0,%rbx
movq $38,%rbp
cmovae %rbx,%rbp
addq %rbp,%rsi
adcq %rbx,%r9
adcq %rbx,%rax
adcq %rbx,%r10
cmovc %rbp,%rbx
addq %rbx,%rsi
movq %rsi,160(%rsp)
movq %r9,168(%rsp)
movq %rax,176(%rsp)
movq %r10,184(%rsp)
movq $0,%rsi
movq $0,%r9
movq $0,%rax
movq $0,%r10
cmpq $1,%r8
movq 64(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 72(%rcx,%rdi),%r11
cmove %r11,%r9
movq 80(%rcx,%rdi),%r11
cmove %r11,%rax
movq 88(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $2,%r8
movq 160(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 168(%rcx,%rdi),%r11
cmove %r11,%r9
movq 176(%rcx,%rdi),%r11
cmove %r11,%rax
movq 184(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $3,%r8
movq 256(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 264(%rcx,%rdi),%r11
cmove %r11,%r9
movq 272(%rcx,%rdi),%r11
cmove %r11,%rax
movq 280(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $4,%r8
movq 352(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 360(%rcx,%rdi),%r11
cmove %r11,%r9
movq 368(%rcx,%rdi),%r11
cmove %r11,%rax
movq 376(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $5,%r8
movq 448(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 456(%rcx,%rdi),%r11
cmove %r11,%r9
movq 464(%rcx,%rdi),%r11
cmove %r11,%rax
movq 472(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $6,%r8
movq 544(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 552(%rcx,%rdi),%r11
cmove %r11,%r9
movq 560(%rcx,%rdi),%r11
cmove %r11,%rax
movq 568(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $7,%r8
movq 640(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 648(%rcx,%rdi),%r11
cmove %r11,%r9
movq 656(%rcx,%rdi),%r11
cmove %r11,%rax
movq 664(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $8,%r8
movq 736(%rcx,%rdi),%r8
cmove %r8,%rsi
movq 744(%rcx,%rdi),%r8
cmove %r8,%r9
movq 752(%rcx,%rdi),%r8
cmove %r8,%rax
movq 760(%rcx,%rdi),%rdi
cmove %rdi,%r10
movq $0,%rdi
movq $0,%rcx
movq $0,%r8
movq $0,%r11
subq %rsi,%rdi
sbbq %r9,%rcx
sbbq %rax,%r8
sbbq %r10,%r11
movq $0,%r12
movq $38,%r13
cmovae %r12,%r13
subq %r13,%rdi
sbbq %r12,%rcx
sbbq %r12,%r8
sbbq %r12,%r11
cmovc %r13,%r12
subq %r12,%rdi
cmpq $0,%rdx
cmovl %rdi,%rsi
cmovl %rcx,%r9
cmovl %r8,%rax
cmovl %r11,%r10
movq %rsi,224(%rsp)
movq %r9,232(%rsp)
movq %rax,240(%rsp)
movq %r10,248(%rsp)
movq $2,192(%rsp)
movq $0,200(%rsp)
movq $0,208(%rsp)
movq $0,216(%rsp)
/* loop: i=1,i<64,i=i+1 */
movq $1,80(%rsp)
.L:
/* choose t */
movq 72(%rsp),%rcx
movq 80(%rsp),%rsi
movq 64(%rsp),%rdx
movzbl 0(%rdx,%rsi),%eax
movsbq %al,%rdx
imulq $768,%rsi,%rdi
movq %rdx,%rsi
sarq $7,%rsi
movq %rdx,%r8
addq %rsi,%r8
xorq %rsi,%r8
movq $1,%rsi
movq $0,%r9
movq $0,%rax
movq $0,%r10
movq $1,%r11
movq $0,%r12
movq $0,%r13
movq $0,%r14
cmpq $1,%r8
movq 0(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 8(%rcx,%rdi),%r15
cmove %r15,%r9
movq 16(%rcx,%rdi),%r15
cmove %r15,%rax
movq 24(%rcx,%rdi),%r15
cmove %r15,%r10
movq 32(%rcx,%rdi),%r15
cmove %r15,%r11
movq 40(%rcx,%rdi),%r15
cmove %r15,%r12
movq 48(%rcx,%rdi),%r15
cmove %r15,%r13
movq 56(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $2,%r8
movq 96(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 104(%rcx,%rdi),%r15
cmove %r15,%r9
movq 112(%rcx,%rdi),%r15
cmove %r15,%rax
movq 120(%rcx,%rdi),%r15
cmove %r15,%r10
movq 128(%rcx,%rdi),%r15
cmove %r15,%r11
movq 136(%rcx,%rdi),%r15
cmove %r15,%r12
movq 144(%rcx,%rdi),%r15
cmove %r15,%r13
movq 152(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $3,%r8
movq 192(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 200(%rcx,%rdi),%r15
cmove %r15,%r9
movq 208(%rcx,%rdi),%r15
cmove %r15,%rax
movq 216(%rcx,%rdi),%r15
cmove %r15,%r10
movq 224(%rcx,%rdi),%r15
cmove %r15,%r11
movq 232(%rcx,%rdi),%r15
cmove %r15,%r12
movq 240(%rcx,%rdi),%r15
cmove %r15,%r13
movq 248(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $4,%r8
movq 288(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 296(%rcx,%rdi),%r15
cmove %r15,%r9
movq 304(%rcx,%rdi),%r15
cmove %r15,%rax
movq 312(%rcx,%rdi),%r15
cmove %r15,%r10
movq 320(%rcx,%rdi),%r15
cmove %r15,%r11
movq 328(%rcx,%rdi),%r15
cmove %r15,%r12
movq 336(%rcx,%rdi),%r15
cmove %r15,%r13
movq 344(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $5,%r8
movq 384(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 392(%rcx,%rdi),%r15
cmove %r15,%r9
movq 400(%rcx,%rdi),%r15
cmove %r15,%rax
movq 408(%rcx,%rdi),%r15
cmove %r15,%r10
movq 416(%rcx,%rdi),%r15
cmove %r15,%r11
movq 424(%rcx,%rdi),%r15
cmove %r15,%r12
movq 432(%rcx,%rdi),%r15
cmove %r15,%r13
movq 440(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $6,%r8
movq 480(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 488(%rcx,%rdi),%r15
cmove %r15,%r9
movq 496(%rcx,%rdi),%r15
cmove %r15,%rax
movq 504(%rcx,%rdi),%r15
cmove %r15,%r10
movq 512(%rcx,%rdi),%r15
cmove %r15,%r11
movq 520(%rcx,%rdi),%r15
cmove %r15,%r12
movq 528(%rcx,%rdi),%r15
cmove %r15,%r13
movq 536(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $7,%r8
movq 576(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 584(%rcx,%rdi),%r15
cmove %r15,%r9
movq 592(%rcx,%rdi),%r15
cmove %r15,%rax
movq 600(%rcx,%rdi),%r15
cmove %r15,%r10
movq 608(%rcx,%rdi),%r15
cmove %r15,%r11
movq 616(%rcx,%rdi),%r15
cmove %r15,%r12
movq 624(%rcx,%rdi),%r15
cmove %r15,%r13
movq 632(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $8,%r8
movq 672(%rcx,%rdi),%r15
cmove %r15,%rsi
movq 680(%rcx,%rdi),%r15
cmove %r15,%r9
movq 688(%rcx,%rdi),%r15
cmove %r15,%rax
movq 696(%rcx,%rdi),%r15
cmove %r15,%r10
movq 704(%rcx,%rdi),%r15
cmove %r15,%r11
movq 712(%rcx,%rdi),%r15
cmove %r15,%r12
movq 720(%rcx,%rdi),%r15
cmove %r15,%r13
movq 728(%rcx,%rdi),%r15
cmove %r15,%r14
cmpq $0,%rdx
movq %rsi,%r15
cmovl %r11,%rsi
cmovl %r15,%r11
movq %r9,%r15
cmovl %r12,%r9
cmovl %r15,%r12
movq %rax,%r15
cmovl %r13,%rax
cmovl %r15,%r13
movq %r10,%r15
cmovl %r14,%r10
cmovl %r15,%r14
movq %rsi,256(%rsp)
movq %r9,264(%rsp)
movq %rax,272(%rsp)
movq %r10,280(%rsp)
movq %r11,288(%rsp)
movq %r12,296(%rsp)
movq %r13,304(%rsp)
movq %r14,312(%rsp)
movq $0,%rsi
movq $0,%r9
movq $0,%rax
movq $0,%r10
cmpq $1,%r8
movq 64(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 72(%rcx,%rdi),%r11
cmove %r11,%r9
movq 80(%rcx,%rdi),%r11
cmove %r11,%rax
movq 88(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $2,%r8
movq 160(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 168(%rcx,%rdi),%r11
cmove %r11,%r9
movq 176(%rcx,%rdi),%r11
cmove %r11,%rax
movq 184(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $3,%r8
movq 256(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 264(%rcx,%rdi),%r11
cmove %r11,%r9
movq 272(%rcx,%rdi),%r11
cmove %r11,%rax
movq 280(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $4,%r8
movq 352(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 360(%rcx,%rdi),%r11
cmove %r11,%r9
movq 368(%rcx,%rdi),%r11
cmove %r11,%rax
movq 376(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $5,%r8
movq 448(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 456(%rcx,%rdi),%r11
cmove %r11,%r9
movq 464(%rcx,%rdi),%r11
cmove %r11,%rax
movq 472(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $6,%r8
movq 544(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 552(%rcx,%rdi),%r11
cmove %r11,%r9
movq 560(%rcx,%rdi),%r11
cmove %r11,%rax
movq 568(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $7,%r8
movq 640(%rcx,%rdi),%r11
cmove %r11,%rsi
movq 648(%rcx,%rdi),%r11
cmove %r11,%r9
movq 656(%rcx,%rdi),%r11
cmove %r11,%rax
movq 664(%rcx,%rdi),%r11
cmove %r11,%r10
cmpq $8,%r8
movq 736(%rcx,%rdi),%r8
cmove %r8,%rsi
movq 744(%rcx,%rdi),%r8
cmove %r8,%r9
movq 752(%rcx,%rdi),%r8
cmove %r8,%rax
movq 760(%rcx,%rdi),%rdi
cmove %rdi,%r10
movq $0,%rdi
movq $0,%rcx
movq $0,%r8
movq $0,%r11
subq %rsi,%rdi
sbbq %r9,%rcx
sbbq %rax,%r8
sbbq %r10,%r11
movq $0,%r12
movq $38,%r13
cmovae %r12,%r13
subq %r13,%rdi
sbbq %r12,%rcx
sbbq %r12,%r8
sbbq %r12,%r11
cmovc %r13,%r12
subq %r12,%rdi
cmpq $0,%rdx
cmovl %rdi,%rsi
cmovl %rcx,%r9
cmovl %r8,%rax
cmovl %r11,%r10
movq %rsi,320(%rsp)
movq %r9,328(%rsp)
movq %rax,336(%rsp)
movq %r10,344(%rsp)
/* nielsadd2 */
// load
movq 160(%rsp),%rdx
movq 168(%rsp),%rcx
movq 176(%rsp),%r8
movq 184(%rsp),%r9
// copy
movq %rdx,%rax
movq %rcx,%r10
movq %r8,%r11
movq %r9,%r12
// sub
subq 128(%rsp),%rdx
sbbq 136(%rsp),%rcx
sbbq 144(%rsp),%r8
sbbq 152(%rsp),%r9
movq $0,%r13
movq $38,%r14
cmovae %r13,%r14
subq %r14,%rdx
sbbq %r13,%rcx
sbbq %r13,%r8
sbbq %r13,%r9
cmovc %r14,%r13
subq %r13,%rdx
movq %rdx,352(%rsp)
movq %rcx,360(%rsp)
movq %r8,368(%rsp)
movq %r9,376(%rsp)
// add
addq 128(%rsp),%rax
adcq 136(%rsp),%r10
adcq 144(%rsp),%r11
adcq 152(%rsp),%r12
movq $0,%r13
movq $38,%r14
cmovae %r13,%r14
addq %r14,%rax
adcq %r13,%r10
adcq %r13,%r11
adcq %r13,%r12
cmovc %r14,%r13
addq %r13,%rax
movq %rax,384(%rsp)
movq %r10,392(%rsp)
movq %r11,400(%rsp)
movq %r12,408(%rsp)
// mul
xorq %r13,%r13
movq 256(%rsp),%rdx
mulx 352(%rsp),%r8,%r9
mulx 360(%rsp),%rcx,%r10
adcx %rcx,%r9
mulx 368(%rsp),%rcx,%r11
adcx %rcx,%r10
mulx 376(%rsp),%rcx,%r12
adcx %rcx,%r11
adcx %r13,%r12
xorq %r14,%r14
movq 264(%rsp),%rdx
mulx 352(%rsp),%rcx,%rbp
adcx %rcx,%r9
adox %rbp,%r10
mulx 360(%rsp),%rcx,%rbp
adcx %rcx,%r10
adox %rbp,%r11
mulx 368(%rsp),%rcx,%rbp
adcx %rcx,%r11
adox %rbp,%r12
mulx 376(%rsp),%rcx,%rbp
adcx %rcx,%r12
adox %rbp,%r13
adcx %r14,%r13
xorq %r15,%r15
movq 272(%rsp),%rdx
mulx 352(%rsp),%rcx,%rbp
adcx %rcx,%r10
adox %rbp,%r11
mulx 360(%rsp),%rcx,%rbp
adcx %rcx,%r11
adox %rbp,%r12
mulx 368(%rsp),%rcx,%rbp
adcx %rcx,%r12
adox %rbp,%r13
mulx 376(%rsp),%rcx,%rbp
adcx %rcx,%r13
adox %rbp,%r14
adcx %r15,%r14
xorq %rax,%rax
movq 280(%rsp),%rdx
mulx 352(%rsp),%rcx,%rbp
adcx %rcx,%r11
adox %rbp,%r12
mulx 360(%rsp),%rcx,%rbp
adcx %rcx,%r12
adox %rbp,%r13
mulx 368(%rsp),%rcx,%rbp
adcx %rcx,%r13
adox %rbp,%r14
mulx 376(%rsp),%rcx,%rbp
adcx %rcx,%r14
adox %rbp,%r15
adcx %rax,%r15
xorq %rbp,%rbp
movq $38,%rdx
mulx %r12,%rax,%r12
adcx %rax,%r8
adox %r12,%r9
mulx %r13,%rcx,%r13
adcx %rcx,%r9
adox %r13,%r10
mulx %r14,%rcx,%r14
adcx %rcx,%r10
adox %r14,%r11
mulx %r15,%rcx,%r15
adcx %rcx,%r11
adox %rbp,%r15
adcx %rbp,%r15
shld $1,%r11,%r15
andq mask63(%rip),%r11
imul $19,%r15,%r15
addq %r15,%r8
adcq $0,%r9
adcq $0,%r10
adcq $0,%r11
movq %r8,352(%rsp)
movq %r9,360(%rsp)
movq %r10,368(%rsp)
movq %r11,376(%rsp)
// mul
xorq %r13,%r13
movq 288(%rsp),%rdx
mulx 384(%rsp),%r8,%r9
mulx 392(%rsp),%rcx,%r10
adcx %rcx,%r9
mulx 400(%rsp),%rcx,%r11
adcx %rcx,%r10
mulx 408(%rsp),%rcx,%r12
adcx %rcx,%r11
adcx %r13,%r12
xorq %r14,%r14
movq 296(%rsp),%rdx
mulx 384(%rsp),%rcx,%rbp
adcx %rcx,%r9
adox %rbp,%r10
mulx 392(%rsp),%rcx,%rbp
adcx %rcx,%r10
adox %rbp,%r11
mulx 400(%rsp),%rcx,%rbp
adcx %rcx,%r11
adox %rbp,%r12
mulx 408(%rsp),%rcx,%rbp
adcx %rcx,%r12
adox %rbp,%r13
adcx %r14,%r13
xorq %r15,%r15
movq 304(%rsp),%rdx
mulx 384(%rsp),%rcx,%rbp
adcx %rcx,%r10
adox %rbp,%r11
mulx 392(%rsp),%rcx,%rbp
adcx %rcx,%r11
adox %rbp,%r12
mulx 400(%rsp),%rcx,%rbp
adcx %rcx,%r12
adox %rbp,%r13
mulx 408(%rsp),%rcx,%rbp
adcx %rcx,%r13
adox %rbp,%r14
adcx %r15,%r14
xorq %rax,%rax
movq 312(%rsp),%rdx
mulx 384(%rsp),%rcx,%rbp
adcx %rcx,%r11
adox %rbp,%r12
mulx 392(%rsp),%rcx,%rbp
adcx %rcx,%r12
adox %rbp,%r13
mulx 400(%rsp),%rcx,%rbp
adcx %rcx,%r13
adox %rbp,%r14
mulx 408(%rsp),%rcx,%rbp
adcx %rcx,%r14
adox %rbp,%r15
adcx %rax,%r15
xorq %rbp,%rbp
movq $38,%rdx
mulx %r12,%rax,%r12
adcx %rax,%r8
adox %r12,%r9
mulx %r13,%rcx,%r13
adcx %rcx,%r9
adox %r13,%r10
mulx %r14,%rcx,%r14
adcx %rcx,%r10
adox %r14,%r11
mulx %r15,%rcx,%r15
adcx %rcx,%r11
adox %rbp,%r15
adcx %rbp,%r15
shld $1,%r11,%r15
andq mask63(%rip),%r11
imul $19,%r15,%r15
addq %r15,%r8
adcq $0,%r9
adcq $0,%r10
adcq $0,%r11
// copy
movq %r8,%r12
movq %r9,%r13
movq %r10,%r14
movq %r11,%r15
// sub
subq 352(%rsp),%r8
sbbq 360(%rsp),%r9
sbbq 368(%rsp),%r10
sbbq 376(%rsp),%r11
movq $0,%rax
movq $38,%rbx
cmovae %rax,%rbx
subq %rbx,%r8
sbbq %rax,%r9
sbbq %rax,%r10
sbbq %rax,%r11
cmovc %rbx,%rax
subq %rax,%r8
movq %r8,384(%rsp)
movq %r9,392(%rsp)
movq %r10,400(%rsp)
movq %r11,408(%rsp)
// add
addq 352(%rsp),%r12
adcq 360(%rsp),%r13
adcq 368(%rsp),%r14
adcq 376(%rsp),%r15
movq $0,%rax
movq $38,%rbx
cmovae %rax,%rbx
addq %rbx,%r12
adcq %rax,%r13
adcq %rax,%r14
adcq %rax,%r15
cmovc %rbx,%rax
addq %rax,%r12
movq %r12,352(%rsp)
movq %r13,360(%rsp)
movq %r14,368(%rsp)
movq %r15,376(%rsp)
// mul
xorq %r13,%r13
movq 320(%rsp),%rdx
mulx 224(%rsp),%r8,%r9
mulx 232(%rsp),%rcx,%r10
adcx %rcx,%r9
mulx 240(%rsp),%rcx,%r11
adcx %rcx,%r10
mulx 248(%rsp),%rcx,%r12
adcx %rcx,%r11
adcx %r13,%r12
xorq %r14,%r14
movq 328(%rsp),%rdx
mulx 224(%rsp),%rcx,%rbp
adcx %rcx,%r9
adox %rbp,%r10
mulx 232(%rsp),%rcx,%rbp
adcx %rcx,%r10
adox %rbp,%r11
mulx 240(%rsp),%rcx,%rbp
adcx %rcx,%r11
adox %rbp,%r12
mulx 248(%rsp),%rcx,%rbp
adcx %rcx,%r12
adox %rbp,%r13
adcx %r14,%r13
xorq %r15,%r15
movq 336(%rsp),%rdx
mulx 224(%rsp),%rcx,%rbp
adcx %rcx,%r10
adox %rbp,%r11
mulx 232(%rsp),%rcx,%rbp
adcx %rcx,%r11
adox %rbp,%r12
mulx 240(%rsp),%rcx,%rbp
adcx %rcx,%r12
adox %rbp,%r13
mulx 248(%rsp),%rcx,%rbp
adcx %rcx,%r13
adox %rbp,%r14
adcx %r15,%r14
xorq %rax,%rax
movq 344(%rsp),%rdx
mulx 224(%rsp),%rcx,%rbp
adcx %rcx,%r11
adox %rbp,%r12
mulx 232(%rsp),%rcx,%rbp
adcx %rcx,%r12
adox %rbp,%r13
mulx 240(%rsp),%rcx,%rbp
adcx %rcx,%r13
adox %rbp,%r14
mulx 248(%rsp),%rcx,%rbp
adcx %rcx,%r14
adox %rbp,%r15
adcx %rax,%r15
xorq %rbp,%rbp
movq $38,%rdx
mulx %r12,%rax,%r12
adcx %rax,%r8
adox %r12,%r9
mulx %r13,%rcx,%r13
adcx %rcx,%r9
adox %r13,%r10
mulx %r14,%rcx,%r14
adcx %rcx,%r10
adox %r14,%r11
mulx %r15,%rcx,%r15
adcx %rcx,%r11
adox %rbp,%r15
adcx %rbp,%r15
shld $1,%r11,%r15
andq mask63(%rip),%r11
imul $19,%r15,%r15
addq %r15,%r8
adcq $0,%r9
adcq $0,%r10
adcq $0,%r11
movq 192(%rsp),%r12
movq 200(%rsp),%r13
movq 208(%rsp),%r14
movq 216(%rsp),%r15
// double
addq %r12,%r12
adcq %r13,%r13
adcq %r14,%r14
adcq %r15,%r15
movq $0,%rbx
movq $38,%rax
cmovae %rbx,%rax
addq %rax,%r12
adcq %rbx,%r13
adcq %rbx,%r14
adcq %rbx,%r15
cmovc %rax,%rbx
addq %rbx,%r12
// copy
movq %r12,%rsi
movq %r13,%rax
movq %r14,%rcx
movq %r15,%rdx
// sub
subq %r8,%r12
sbbq %r9,%r13
sbbq %r10,%r14
sbbq %r11,%r15
movq $0,%rbx
movq $38,%rbp
cmovae %rbx,%rbp
subq %rbp,%r12
sbbq %rbx,%r13
sbbq %rbx,%r14
sbbq %rbx,%r15
cmovc %rbp,%rbx
subq %rbx,%r12
movq %r12,448(%rsp)
movq %r13,456(%rsp)
movq %r14,464(%rsp)
movq %r15,472(%rsp)
// add
addq %r8,%rsi
adcq %r9,%rax
adcq %r10,%rcx
adcq %r11,%rdx
movq $0,%rbx
movq $38,%rbp
cmovae %rbx,%rbp
addq %rbp,%rsi
adcq %rbx,%rax
adcq %rbx,%rcx
adcq %rbx,%rdx
cmovc %rbp,%rbx
addq %rbx,%rsi
movq %rsi,416(%rsp)
movq %rax,424(%rsp)
movq %rcx,432(%rsp)
movq %rdx,440(%rsp)
/* p1p1 to p3 */
// convert to 5x4 form
vmovdqa 384(%rsp),%ymm8
vmovdqa 352(%rsp),%ymm9
vmovdqa 416(%rsp),%ymm10
vmovdqa 352(%rsp),%ymm11
vpunpcklqdq %ymm9,%ymm8,%ymm12
vpunpckhqdq %ymm9,%ymm8,%ymm13
vpunpcklqdq %ymm11,%ymm10,%ymm14
vpunpckhqdq %ymm11,%ymm10,%ymm15
vpermq $68,%ymm14,%ymm7
vpblendd $240,%ymm7,%ymm12,%ymm10
vpermq $68,%ymm15,%ymm7
vpblendd $240,%ymm7,%ymm13,%ymm11
vpermq $238,%ymm12,%ymm7
vpblendd $240,%ymm14,%ymm7,%ymm12
vpermq $238,%ymm13,%ymm7
vpblendd $240,%ymm15,%ymm7,%ymm13
vpand pmask1(%rip),%ymm10,%ymm5
vpand pmask2(%rip),%ymm10,%ymm6
vpsrlq $52,%ymm6,%ymm6
vpand pmask3(%rip),%ymm11,%ymm7
vpsllq $12,%ymm7,%ymm7
vpor %ymm7,%ymm6,%ymm6
vpand pmask4(%rip),%ymm11,%ymm7
vpsrlq $40,%ymm7,%ymm7
vpand pmask5(%rip),%ymm12,%ymm8
vpsllq $24,%ymm8,%ymm8
vpor %ymm8,%ymm7,%ymm7
vpand pmask6(%rip),%ymm12,%ymm8
vpsrlq $28,%ymm8,%ymm8
vpand pmask7(%rip),%ymm13,%ymm9
vpsllq $36,%ymm9,%ymm9
vpor %ymm9,%ymm8,%ymm8
vpand pmask8(%rip),%ymm13,%ymm9
vpsrlq $16,%ymm9,%ymm9
// convert to 5x4 form
vmovdqa 448(%rsp),%ymm0
vmovdqa 416(%rsp),%ymm1
vmovdqa 448(%rsp),%ymm2
vmovdqa 384(%rsp),%ymm3
vpunpcklqdq %ymm1,%ymm0,%ymm12
vpunpckhqdq %ymm1,%ymm0,%ymm13
vpunpcklqdq %ymm3,%ymm2,%ymm14
vpunpckhqdq %ymm3,%ymm2,%ymm15
vpermq $68,%ymm14,%ymm0
vpblendd $240,%ymm0,%ymm12,%ymm10
vpermq $68,%ymm15,%ymm0
vpblendd $240,%ymm0,%ymm13,%ymm11
vpermq $238,%ymm12,%ymm0
vpblendd $240,%ymm14,%ymm0,%ymm12
vpermq $238,%ymm13,%ymm0
vpblendd $240,%ymm15,%ymm0,%ymm13
vpand pmask1(%rip),%ymm10,%ymm0
vpand pmask2(%rip),%ymm10,%ymm1
vpsrlq $52,%ymm1,%ymm1
vpand pmask3(%rip),%ymm11,%ymm2
vpsllq $12,%ymm2,%ymm2
vpor %ymm2,%ymm1,%ymm1
vpand pmask4(%rip),%ymm11,%ymm2
vpsrlq $40,%ymm2,%ymm2
vpand pmask5(%rip),%ymm12,%ymm3
vpsllq $24,%ymm3,%ymm3
vpor %ymm3,%ymm2,%ymm2
vpand pmask6(%rip),%ymm12,%ymm3
vpsrlq $28,%ymm3,%ymm3
vpand pmask7(%rip),%ymm13,%ymm4
vpsllq $36,%ymm4,%ymm4
vpor %ymm4,%ymm3,%ymm3
vpand pmask8(%rip),%ymm13,%ymm4
vpsrlq $16,%ymm4,%ymm4
// mul4x1
vpxorq %ymm10,%ymm10,%ymm10
vpxorq %ymm11,%ymm11,%ymm11
vpxorq %ymm12,%ymm12,%ymm12
vpxorq %ymm13,%ymm13,%ymm13
vpxorq %ymm14,%ymm14,%ymm14
vpxorq %ymm15,%ymm15,%ymm15
vpxorq %ymm16,%ymm16,%ymm16
vpxorq %ymm17,%ymm17,%ymm17
vpxorq %ymm18,%ymm18,%ymm18
vpxorq %ymm19,%ymm19,%ymm19
vpxorq %ymm25,%ymm25,%ymm25
vpxorq %ymm26,%ymm26,%ymm26
vpxorq %ymm27,%ymm27,%ymm27
vpxorq %ymm28,%ymm28,%ymm28
vpxorq %ymm29,%ymm29,%ymm29
vpxorq %ymm30,%ymm30,%ymm30
vpxorq %ymm31,%ymm31,%ymm31
vpmadd52luq %ymm0,%ymm5,%ymm10
vpmadd52huq %ymm0,%ymm5,%ymm11
vpmadd52luq %ymm0,%ymm6,%ymm25
vpmadd52huq %ymm0,%ymm6,%ymm12
vpmadd52luq %ymm1,%ymm5,%ymm25
vpmadd52huq %ymm1,%ymm5,%ymm12
vpaddq %ymm25,%ymm11,%ymm11
vpmadd52luq %ymm0,%ymm7,%ymm26
vpmadd52huq %ymm0,%ymm7,%ymm13
vpmadd52luq %ymm1,%ymm6,%ymm26
vpmadd52huq %ymm1,%ymm6,%ymm13
vpmadd52luq %ymm2,%ymm5,%ymm26
vpmadd52huq %ymm2,%ymm5,%ymm13
vpaddq %ymm26,%ymm12,%ymm12
vpmadd52luq %ymm0,%ymm8,%ymm27
vpmadd52huq %ymm0,%ymm8,%ymm14
vpmadd52luq %ymm1,%ymm7,%ymm27
vpmadd52huq %ymm1,%ymm7,%ymm14
vpmadd52luq %ymm2,%ymm6,%ymm27
vpmadd52huq %ymm2,%ymm6,%ymm14
vpmadd52luq %ymm3,%ymm5,%ymm27
vpmadd52huq %ymm3,%ymm5,%ymm14
vpaddq %ymm27,%ymm13,%ymm13
vpmadd52luq %ymm0,%ymm9,%ymm28
vpmadd52huq %ymm0,%ymm9,%ymm15
vpmadd52luq %ymm1,%ymm8,%ymm28
vpmadd52huq %ymm1,%ymm8,%ymm15
vpmadd52luq %ymm2,%ymm7,%ymm28
vpmadd52huq %ymm2,%ymm7,%ymm15
vpmadd52luq %ymm3,%ymm6,%ymm28
vpmadd52huq %ymm3,%ymm6,%ymm15
vpmadd52luq %ymm4,%ymm5,%ymm28
vpmadd52huq %ymm4,%ymm5,%ymm15
vpaddq %ymm28,%ymm14,%ymm14
vpmadd52luq %ymm1,%ymm9,%ymm29
vpmadd52huq %ymm1,%ymm9,%ymm16
vpmadd52luq %ymm2,%ymm8,%ymm29
vpmadd52huq %ymm2,%ymm8,%ymm16
vpmadd52luq %ymm3,%ymm7,%ymm29
vpmadd52huq %ymm3,%ymm7,%ymm16
vpmadd52luq %ymm4,%ymm6,%ymm29
vpmadd52huq %ymm4,%ymm6,%ymm16
vpaddq %ymm29,%ymm15,%ymm15
vpmadd52luq %ymm2,%ymm9,%ymm30
vpmadd52huq %ymm2,%ymm9,%ymm17
vpmadd52luq %ymm3,%ymm8,%ymm30
vpmadd52huq %ymm3,%ymm8,%ymm17
vpmadd52luq %ymm4,%ymm7,%ymm30
vpmadd52huq %ymm4,%ymm7,%ymm17
vpaddq %ymm30,%ymm16,%ymm16
vpmadd52luq %ymm3,%ymm9,%ymm31
vpmadd52huq %ymm3,%ymm9,%ymm18
vpmadd52luq %ymm4,%ymm8,%ymm31
vpmadd52huq %ymm4,%ymm8,%ymm18
vpaddq %ymm31,%ymm17,%ymm17
vpmadd52luq %ymm4,%ymm9,%ymm18
vpmadd52huq %ymm4,%ymm9,%ymm19
vpsrlq $52,%ymm15,%ymm22
vpaddq %ymm22,%ymm16,%ymm16
vpandq vecmask52(%rip),%ymm15,%ymm15
vpmadd52luq vec608(%rip),%ymm15,%ymm10
vpmadd52huq vec608(%rip),%ymm15,%ymm11
vpsrlq $52,%ymm16,%ymm22
vpaddq %ymm22,%ymm17,%ymm17
vpandq vecmask52(%rip),%ymm16,%ymm16
vpmadd52luq vec608(%rip),%ymm16,%ymm11
vpmadd52huq vec608(%rip),%ymm16,%ymm12
vpsrlq $52,%ymm17,%ymm22
vpaddq %ymm22,%ymm18,%ymm18
vpandq vecmask52(%rip),%ymm17,%ymm17
vpmadd52luq vec608(%rip),%ymm17,%ymm12
vpmadd52huq vec608(%rip),%ymm17,%ymm13
vpsrlq $52,%ymm18,%ymm22
vpaddq %ymm22,%ymm19,%ymm19
vpandq vecmask52(%rip),%ymm18,%ymm18
vpmadd52luq vec608(%rip),%ymm18,%ymm13
vpmadd52huq vec608(%rip),%ymm18,%ymm14
vpxorq %ymm15,%ymm15,%ymm15
vpmadd52luq vec608(%rip),%ymm19,%ymm14
vpmadd52huq vec608(%rip),%ymm19,%ymm15
vpmadd52luq vec608(%rip),%ymm15,%ymm10
vpsrlq $52,%ymm13,%ymm22
vpaddq %ymm22,%ymm14,%ymm14
vpandq vecmask52(%rip),%ymm13,%ymm13
vpsrlq $47,%ymm14,%ymm22
vpandq vecmask47(%rip),%ymm14,%ymm14
vpmadd52luq vec19(%rip),%ymm22,%ymm10
vpsrlq $52,%ymm10,%ymm22
vpaddq %ymm22,%ymm11,%ymm11
vpandq vecmask52(%rip),%ymm10,%ymm10
vpsrlq $52,%ymm11,%ymm22
vpaddq %ymm22,%ymm12,%ymm12
vpandq vecmask52(%rip),%ymm11,%ymm11
vpsrlq $52,%ymm12,%ymm22
vpaddq %ymm22,%ymm13,%ymm13
vpandq vecmask52(%rip),%ymm12,%ymm12
vpsrlq $52,%ymm13,%ymm22
vpaddq %ymm22,%ymm14,%ymm14
vpandq vecmask52(%rip),%ymm13,%ymm13
// get back to 4x4 form
vpand upmask1(%rip),%ymm10,%ymm0
vpand upmask2(%rip),%ymm11,%ymm1
vpsllq $52,%ymm1,%ymm1
vpor %ymm1,%ymm0,%ymm0
vpand upmask3(%rip),%ymm11,%ymm1
vpsrlq $12,%ymm1,%ymm1
vpand upmask4(%rip),%ymm12,%ymm2
vpsllq $40,%ymm2,%ymm2
vpor %ymm2,%ymm1,%ymm1
vpand upmask5(%rip),%ymm12,%ymm2
vpsrlq $24,%ymm2,%ymm2
vpand upmask6(%rip),%ymm13,%ymm3
vpsllq $28,%ymm3,%ymm3
vpor %ymm3,%ymm2,%ymm2
vpand upmask7(%rip),%ymm13,%ymm3
vpsrlq $36,%ymm3,%ymm3
vpand upmask1(%rip),%ymm14,%ymm4
vpsllq $16,%ymm4,%ymm4
vpor %ymm4,%ymm3,%ymm3
vpunpcklqdq %ymm1,%ymm0,%ymm12
vpunpckhqdq %ymm1,%ymm0,%ymm13
vpunpcklqdq %ymm3,%ymm2,%ymm14
vpunpckhqdq %ymm3,%ymm2,%ymm15
vpermq $68,%ymm14,%ymm7
vpblendd $240,%ymm7,%ymm12,%ymm0
vpermq $68,%ymm15,%ymm7
vpblendd $240,%ymm7,%ymm13,%ymm1
vpermq $238,%ymm12,%ymm7
vpblendd $240,%ymm14,%ymm7,%ymm2
vpermq $238,%ymm13,%ymm7
vpblendd $240,%ymm15,%ymm7,%ymm3
vmovdqa %ymm0,128(%rsp)
vmovdqa %ymm1,160(%rsp)
vmovdqa %ymm2,192(%rsp)
vmovdqa %ymm3,224(%rsp)
movq 80(%rsp),%r15
addq $1,%r15
movq %r15,80(%rsp)
cmpq $63,%r15
jle .L
/* store final value of r */
movq 56(%rsp),%rdi
movq 128(%rsp),%r8
movq 136(%rsp),%r9
movq 144(%rsp),%r10
movq 152(%rsp),%r11
movq %r8,0(%rdi)
movq %r9,8(%rdi)
movq %r10,16(%rdi)
movq %r11,24(%rdi)
movq 160(%rsp),%r8
movq 168(%rsp),%r9
movq 176(%rsp),%r10
movq 184(%rsp),%r11
movq %r8,32(%rdi)
movq %r9,40(%rdi)
movq %r10,48(%rdi)
movq %r11,56(%rdi)
movq 192(%rsp),%r8
movq 200(%rsp),%r9
movq 208(%rsp),%r10
movq 216(%rsp),%r11
movq %r8,64(%rdi)
movq %r9,72(%rdi)
movq %r10,80(%rdi)
movq %r11,88(%rdi)
movq 224(%rsp),%r8
movq 232(%rsp),%r9
movq 240(%rsp),%r10
movq 248(%rsp),%r11
movq %r8,96(%rdi)
movq %r9,104(%rdi)
movq %r10,112(%rdi)
movq %r11,120(%rdi)
movq 0(%rsp),%r11
movq 8(%rsp),%r12
movq 16(%rsp),%r13
movq 24(%rsp),%r14
movq 32(%rsp),%r15
movq 40(%rsp),%rbx
movq 48(%rsp),%rbp
movq %r11,%rsp
ret