crypton-2.1.3: cbits/s2n/x86_att/bignum_demont_p256_alt.S
// Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
// SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT-0
// ----------------------------------------------------------------------------
// Convert from Montgomery form z := (x / 2^256) mod p_256, assuming x reduced
// Input x[4]; output z[4]
//
// extern void bignum_demont_p256_alt(uint64_t z[static 4],
// const uint64_t x[static 4]);
//
// This assumes the input is < p_256 for correctness. If this is not the case,
// use the variant "bignum_deamont_p256" instead.
//
// Standard x86-64 ABI: RDI = z, RSI = x
// Microsoft x64 ABI: RCX = z, RDX = x
// ----------------------------------------------------------------------------
#include "_internal_s2n_bignum_x86_att.h"
S2N_BN_SYM_VISIBILITY_DIRECTIVE(bignum_demont_p256_alt)
S2N_BN_FUNCTION_TYPE_DIRECTIVE(bignum_demont_p256_alt)
S2N_BN_SYM_PRIVACY_DIRECTIVE(bignum_demont_p256_alt)
.text
#define z %rdi
#define x %rsi
// Add %rdx * m into a register-pair (high,low)
// maintaining consistent double-carrying with adcx and adox,
// using %rax and %rcx as temporaries
#define mulpado(high,low,m) \
mulxq m, %rax, %rcx ; \
adcxq %rax, low ; \
adoxq %rcx, high
// Add %rcx * m into a register-pair (high,low) maintaining consistent
// carry-catching with carry (negated, as bitmask) and using %rax and %rdx
// as temporaries
#define mulpadd(carry,high,low,m) \
movq m, %rax ; \
mulq %rcx; \
subq carry, %rdx ; \
addq %rax, low ; \
adcq %rdx, high ; \
sbbq carry, carry
// Initial version assuming no carry-in
#define mulpadi(carry,high,low,m) \
movq m, %rax ; \
mulq %rcx; \
addq %rax, low ; \
adcq %rdx, high ; \
sbbq carry, carry
// Version with no carry in or out
#define mulpadn(high,low,m) \
movq m, %rax ; \
mulq %rcx; \
addq %rax, low ; \
adcq %rdx, high
S2N_BN_SYMBOL(bignum_demont_p256_alt):
CFI_START
_CET_ENDBR
#if WINDOWS_ABI
CFI_PUSH(%rdi)
CFI_PUSH(%rsi)
movq %rcx, %rdi
movq %rdx, %rsi
#endif
// Set up an initial 4-word window [%r11,%r10,%r9,%r8] = x
movq (x), %r8
movq 8(x), %r9
movq 16(x), %r10
movq 24(x), %r11
// Load constant 2^32; %rcx toggles between this and (1 - %rcx) below
movq $0x0000000100000000, %rcx
// Montgomery reduce windows 0 and 1 together as [%r8;%rsi;%r11;%r10]
mulpadi(%rsi,%r10,%r9,%r8)
mulpadd(%rsi,%r11,%r10,%r9)
negq %rcx
negq %rsi
incq %rcx
mulpadi(%r8,%rsi,%r11,%r8)
negq %r8
mulpadn(%r8,%rsi,%r9)
// Montgomery reduce windows 2 and 3 together as [%r10;%r9;%r8;%rsi]
negq %rcx
incq %rcx
mulpadi(%r9,%rsi,%r11,%r10)
mulpadd(%r9,%r8,%rsi,%r11)
negq %rcx
negq %r9
incq %rcx
mulpadi(%r10,%r9,%r8,%r10)
negq %r10
mulpadn(%r10,%r9,%r11)
// Since the input was assumed reduced modulo, i.e. < p, we actually know that
// 2^256 * [carries; %r10;%r9;%r8;%rsi] is <= (p - 1) + (2^256 - 1) p
// and hence [carries; %r10;%r9;%r8;%rsi] < p. This means in fact carries = 0
// and [%r10;%r9;%r8;%rsi] is already our answer, without further correction.
// Write that back.
movq %rsi, (z)
movq %r8, 8(z)
movq %r9, 16(z)
movq %r10, 24(z)
#if WINDOWS_ABI
CFI_POP(%rsi)
CFI_POP(%rdi)
#endif
CFI_RET
S2N_BN_SIZE_DIRECTIVE(bignum_demont_p256_alt)
#if defined(__linux__) && defined(__ELF__)
.section .note.GNU-stack,"",%progbits
#endif