crypton-2.1.2: cbits/s2n/x86_att/bignum_deamont_p384_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 almost-Montgomery form, z := (x / 2^384) mod p_384
// Input x[6]; output z[6]
//
// extern void bignum_deamont_p384_alt(uint64_t z[static 6],
// const uint64_t x[static 6]);
//
// Convert a 6-digit bignum x out of its (optionally almost) Montgomery form,
// "almost" meaning any 6-digit input will work, with no range restriction.
//
// 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_deamont_p384_alt)
S2N_BN_FUNCTION_TYPE_DIRECTIVE(bignum_deamont_p384_alt)
S2N_BN_SYM_PRIVACY_DIRECTIVE(bignum_deamont_p384_alt)
.text
#define z %rdi
#define x %rsi
// Additional temps in the correction phase
#define u %rax
#define v %rcx
#define w %rdx
#define vshort %ecx
// Core one-step "short" Montgomery reduction macro. Takes input in
// [d5;d4;d3;d2;d1;d0] and returns result in [d6;d5;d4;d3;d2;d1],
// adding to the existing [d5;d4;d3;d2;d1] and re-using d0 as a
// temporary internally, as well as %rax, %rcx and %rdx.
// It is OK for d6 and d0 to be the same register (they often are)
//
// We want to add (2^384 - 2^128 - 2^96 + 2^32 - 1) * w
// where w = [d0 + (d0<<32)] mod 2^64
//
// montreds(d6,d5,d4,d3,d2,d1,d0)
#define montreds(d6,d5,d4,d3,d2,d1,d0) \
/* Our correction multiplier is w = [d0 + (d0<<32)] mod 2^64 */ \
movq d0, %rcx ; \
shlq $32, %rcx ; \
addq d0, %rcx ; \
/* Construct [%rax;%rdx;d0;-] = (2^384 - p_384) * w */ \
/* We know the lowest word will cancel so we can re-use d0 */ \
/* and %rcx as temps. */ \
movq $0xffffffff00000001, %rax ; \
mulq %rcx; \
movq %rdx, d0 ; \
movq $0x00000000ffffffff, %rax ; \
mulq %rcx; \
addq %rax, d0 ; \
movl $0, %eax ; \
adcq %rcx, %rdx ; \
adcl %eax, %eax ; \
/* Now subtract that and add 2^384 * w */ \
subq d0, d1 ; \
sbbq %rdx, d2 ; \
sbbq %rax, d3 ; \
sbbq $0, d4 ; \
sbbq $0, d5 ; \
movq %rcx, d6 ; \
sbbq $0, d6
S2N_BN_SYMBOL(bignum_deamont_p384_alt):
CFI_START
_CET_ENDBR
#if WINDOWS_ABI
CFI_PUSH(%rdi)
CFI_PUSH(%rsi)
movq %rcx, %rdi
movq %rdx, %rsi
#endif
// Save more registers to play with
CFI_PUSH(%r12)
CFI_PUSH(%r13)
// Set up an initial window [%r13,%r12,%r11,%r10,%r9,%r8] = x
movq (x), %r8
movq 8(x), %r9
movq 16(x), %r10
movq 24(x), %r11
movq 32(x), %r12
movq 40(x), %r13
// Montgomery reduce window 0
montreds(%r8,%r13,%r12,%r11,%r10,%r9,%r8)
// Montgomery reduce window 1
montreds(%r9,%r8,%r13,%r12,%r11,%r10,%r9)
// Montgomery reduce window 2
montreds(%r10,%r9,%r8,%r13,%r12,%r11,%r10)
// Montgomery reduce window 3
montreds(%r11,%r10,%r9,%r8,%r13,%r12,%r11)
// Montgomery reduce window 4
montreds(%r12,%r11,%r10,%r9,%r8,%r13,%r12)
// Montgomery reduce window 5
montreds(%r13,%r12,%r11,%r10,%r9,%r8,%r13)
// Do a test addition of dd = [%r13;%r12;%r11;%r10;%r9;%r8] and
// 2^384 - p_384 = [0;0;0;1;v;u], hence setting CF iff
// dd + (2^384 - p_384) >= 2^384, hence iff dd >= p_384.
movq $0xffffffff00000001, u
movl $0x00000000ffffffff, vshort
movq %r8, w
addq u, w
movq %r9, w
adcq v, w
movq %r10, w
adcq $1, w
movq %r11, w
adcq $0, w
movq %r12, w
adcq $0, w
movq %r13, w
adcq $0, w
// Convert CF to a bitmask in w
sbbq w, w
// Masked addition of 2^384 - p_384, hence subtraction of p_384
andq w, u
andq w, v
andq $1, w
addq u, %r8
adcq v, %r9
adcq w, %r10
adcq $0, %r11
adcq $0, %r12
adcq $0, %r13
// Write back the result
movq %r8, (z)
movq %r9, 8(z)
movq %r10, 16(z)
movq %r11, 24(z)
movq %r12, 32(z)
movq %r13, 40(z)
// Restore registers and return
CFI_POP(%r13)
CFI_POP(%r12)
#if WINDOWS_ABI
CFI_POP(%rsi)
CFI_POP(%rdi)
#endif
CFI_RET
S2N_BN_SIZE_DIRECTIVE(bignum_deamont_p384_alt)
#if defined(__linux__) && defined(__ELF__)
.section .note.GNU-stack,"",%progbits
#endif