libbb/sha1: use SSE2 in unrolled x86-64 code. ~10% faster

function                                             old     new   delta
.rodata                                           108241  108305     +64
sha1_process_block64                                3502    3495      -7
------------------------------------------------------------------------------
(add/remove: 5/0 grow/shrink: 1/1 up/down: 64/-7)              Total: 57 bytes

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
This commit is contained in:
Denys Vlasenko 2022-01-23 09:27:30 +01:00
parent 1e825acf8d
commit 39369ff460
2 changed files with 854 additions and 578 deletions

File diff suppressed because it is too large Load Diff

View File

@ -6,33 +6,103 @@
# also contains the diff of the generated file. # also contains the diff of the generated file.
exec >hash_md5_sha_x86-64.S exec >hash_md5_sha_x86-64.S
# There is a way to use XMM registers (which always exist for x86-64!) for W[] # Based on http://arctic.org/~dean/crypto/sha1.html.
# For example, if we load W as follows: # ("This SHA1 implementation is public domain.")
# %xmm0: w[0x0] w[0x1] w[0x2] w[0x3] #
# %xmm4: w[0x4] w[0x5] w[0x6] w[0x7] # x86-64 has at least SSE2 vector insns always available.
# %xmm8: w[0x8] w[0x9] w[0xa] w[0xb] # We can use them without any CPUID checks (and without a need
# %xmm12: w[0xc] w[0xd] w[0xe] w[0xf] # for a fallback code if needed insns are not available).
# then the xor'ing operation to generate next W[0..3] is: # This code uses them to calculate W[] ahead of time.
# movaps %xmm0, %xmmT2 #
# palignr $0x8, %xmm4, %xmmT2 # form (w[0x2],w[0x3],w[0x4],w[0x5]) # Unfortunately, results are passed from vector unit to
# # Right-shifts xmm4:xmmT2 by 8 bytes. Writes shifted result to xmmT2. SSSE3 insn. # integer ALUs on the stack. MOVD/Q insns to move them directly
# movaps %xmm0, %xmmT13 # from vector to integer registers are slower than store-to-load
# palignr $0x4,%xmm0,%xmmT13 # form (w[0xd],w[0xe],w[0xf],w[0x0]) # forwarding in LSU (on Skylake at least).
# xmm0 = xmm0 ^ t2 ^ xmm8 ^ t13 #
# xmm0 = rol32(xmm0,1) # no such insn, have to use pslld+psrld+or # The win against a purely integer code is small on Skylake,
# and then results can be extracted for use: # only about 7-8%. We offload about 1/3 of our operations to the vector unit.
# movd %xmm0, %esi # new W[0] # It can do 4 ops at once in one 128-bit register,
# pextrd $1, %xmm0, %esi # new W[1] # but we have to use x2 of them because of W[0] complication,
# # SSE4.1 insn. Can use EXTRACTPS (also SSE4.1) # SSE2 has no "rotate each word by N bits" insns,
# pextrd $2, %xmm0, %esi # new W[2] # moving data to/from vector unit is clunky, and Skylake
# pextrd $3, %xmm0, %esi # new W[3] # has four integer ALUs unified with three vector ALUs,
# ... but this requires SSE4.1 and SSSE3, which are not universally available on x86-64. # which makes pure integer code rather fast, and makes
# vector ops compete with integer ones.
#
# Zen3, with its separate vector ALUs, wins more, about 12%.
xmmT1="%xmm4"
xmmT2="%xmm5"
xmmRCONST="%xmm6"
T=`printf '\t'`
# SSE instructions are longer than 4 bytes on average.
# Intel CPUs (up to Tiger Lake at least) can't decode
# more than 16 bytes of code in one cycle.
# By interleaving SSE code and integer code
# we mostly achieve a situation where 16-byte decode fetch window
# contains 4 (or more) insns.
#
# However. On Skylake, there was no observed difference,
# but on Zen3, non-interleaved code is ~3% faster
# (822 Mb/s versus 795 Mb/s hashing speed).
# Off for now:
interleave=false
INTERLEAVE() {
$interleave || \
{
# Generate non-interleaved code
# (it should work correctly too)
echo "$1"
echo "$2"
return
}
(
echo "$1" | grep -v '^$' >"$0.temp1"
echo "$2" | grep -v '^$' >"$0.temp2"
exec 3<"$0.temp1"
exec 4<"$0.temp2"
IFS=''
while :; do
line1=''
line2=''
while :; do
read -r line1 <&3
if test "${line1:0:1}" != "#" && test "${line1:0:2}" != "$T#"; then
break
fi
echo "$line1"
done
while :; do
read -r line2 <&4
if test "${line2:0:4}" = "${T}lea"; then
# We use 7-8 byte long forms of LEA.
# Do not interleave them with SSE insns
# which are also long.
echo "$line2"
read -r line2 <&4
echo "$line2"
continue
fi
if test "${line2:0:1}" != "#" && test "${line2:0:2}" != "$T#"; then
break
fi
echo "$line2"
done
test "$line1$line2" || break
echo "$line1"
echo "$line2"
done
rm "$0.temp1" "$0.temp2"
)
}
echo \ echo \
'### Generated by hash_md5_sha_x86-64.S.sh ### "### Generated by hash_md5_sha_x86-64.S.sh ###
#if CONFIG_SHA1_SMALL == 0 && defined(__GNUC__) && defined(__x86_64__) #if CONFIG_SHA1_SMALL == 0 && defined(__GNUC__) && defined(__x86_64__)
.section .text.sha1_process_block64,"ax",@progbits .section .text.sha1_process_block64,\"ax\",@progbits
.globl sha1_process_block64 .globl sha1_process_block64
.hidden sha1_process_block64 .hidden sha1_process_block64
.type sha1_process_block64, @function .type sha1_process_block64, @function
@ -51,16 +121,10 @@ sha1_process_block64:
# eax..edx: a..d # eax..edx: a..d
# ebp: e # ebp: e
# esi,edi: temps # esi,edi: temps
# -32+4*n(%rsp),r8...r15: W[0..7,8..15] # xmm0..xmm3: W[]
# (TODO: actually W[0..7] are used a bit more often, put _them_ into r8..r15?) # xmm4,xmm5: temps
movl $3, %eax # xmm6: current round constant
1: # -64(%rsp): area for passing RCONST + W[] from vector to integer units
movq (%rdi,%rax,8), %rsi
bswapq %rsi
rolq $32, %rsi
movq %rsi, -32(%rsp,%rax,8)
decl %eax
jns 1b
movl 80(%rdi), %eax # a = ctx->hash[0] movl 80(%rdi), %eax # a = ctx->hash[0]
movl 84(%rdi), %ebx # b = ctx->hash[1] movl 84(%rdi), %ebx # b = ctx->hash[1]
@ -68,32 +132,120 @@ sha1_process_block64:
movl 92(%rdi), %edx # d = ctx->hash[3] movl 92(%rdi), %edx # d = ctx->hash[3]
movl 96(%rdi), %ebp # e = ctx->hash[4] movl 96(%rdi), %ebp # e = ctx->hash[4]
movaps rconst0x5A827999(%rip), $xmmRCONST
# For round 1, steps 0 and 8..15, we pass W[0,8..15] in esi,r8..r15
# instead of spilling them to stack.
# (We lose parallelized addition of RCONST, but LEA
# can do two additions at once, so...)
movq 4*0(%rdi), %rsi
movq 4*2(%rdi), %r10
bswapq %rsi
bswapq %r10
rolq \$32, %rsi # rsi = W[1]:W[0]
rolq \$32, %r10
movq %rsi, %xmm0
movq %r10, $xmmT1
punpcklqdq $xmmT1, %xmm0 # xmm0 = r10:rsi = (W[0],W[1],W[2],W[3])
movaps %xmm0, $xmmT1
paddd $xmmRCONST, $xmmT1
movups $xmmT1, -64+4*0(%rsp)
movq 4*4(%rdi), %r8
movq 4*6(%rdi), %r10
bswapq %r8
bswapq %r10
rolq \$32, %r8
rolq \$32, %r10
movq %r8, %xmm1
movq %r10, $xmmT1
punpcklqdq $xmmT1, %xmm1 # xmm1 = r10:r8 = (W[4],W[5],W[6],W[7])
movaps %xmm1, $xmmT1
paddd $xmmRCONST, $xmmT1
movups $xmmT1, -64+4*4(%rsp)
movq 4*8(%rdi), %r8 movq 4*8(%rdi), %r8
movq 4*10(%rdi), %r10 movq 4*10(%rdi), %r10
bswapq %r8 bswapq %r8
bswapq %r10 bswapq %r10
movl %r8d, %r9d # r9d = W[9]
rolq \$32, %r8 # r8 = W[9]:W[8]
movl %r10d, %r11d # r11d = W[11]
rolq \$32, %r10 # r10 = W[11]:W[10]
movq %r8, %xmm2
movq %r10, $xmmT1
punpcklqdq $xmmT1, %xmm2 # xmm2 = r10:r8 = (W[8],W[9],W[10],W[11])
movq 4*12(%rdi), %r12 movq 4*12(%rdi), %r12
movq 4*14(%rdi), %r14 movq 4*14(%rdi), %r14
bswapq %r12 bswapq %r12
bswapq %r14 bswapq %r14
movl %r8d, %r9d movl %r12d, %r13d # r13d = W[13]
shrq $32, %r8 rolq \$32, %r12 # r12 = W[13]:W[12]
movl %r10d, %r11d movl %r14d, %r15d # r15d = W[15]
shrq $32, %r10 rolq \$32, %r14 # r14 = W[15]:W[14]
movl %r12d, %r13d movq %r12, %xmm3
shrq $32, %r12 movq %r14, $xmmT1
movl %r14d, %r15d punpcklqdq $xmmT1, %xmm3 # xmm3 = r14:r12 = (W[12],W[13],W[14],W[15])
shrq $32, %r14 "
'
W32() { PREP() {
test "$1" || exit 1 local xmmW0=$1
test "$1" -lt 0 && exit 1 local xmmW4=$2
test "$1" -gt 15 && exit 1 local xmmW8=$3
test "$1" -lt 8 && echo "-32+4*$1(%rsp)" local xmmW12=$4
test "$1" -ge 8 && echo "%r${1}d" # the above must be %xmm0..3 in some permutation
local dstmem=$5
#W[0] = rol(W[13] ^ W[8] ^ W[2] ^ W[0], 1);
#W[1] = rol(W[14] ^ W[9] ^ W[3] ^ W[1], 1);
#W[2] = rol(W[15] ^ W[10] ^ W[4] ^ W[2], 1);
#W[3] = rol( 0 ^ W[11] ^ W[5] ^ W[3], 1);
#W[3] ^= rol(W[0], 1);
echo "# PREP $@
movaps $xmmW12, $xmmT1
psrldq \$4, $xmmT1 # rshift by 4 bytes: T1 = ([13],[14],[15],0)
pshufd \$0x4e, $xmmW0, $xmmT2 # 01001110=2,3,0,1 shuffle, ([2],[3],x,x)
punpcklqdq $xmmW4, $xmmT2 # T2 = W4[0..63]:T2[0..63] = ([2],[3],[4],[5])
xorps $xmmW8, $xmmW0 # ([8],[9],[10],[11]) ^ ([0],[1],[2],[3])
xorps $xmmT1, $xmmT2 # ([13],[14],[15],0) ^ ([2],[3],[4],[5])
xorps $xmmT2, $xmmW0 # ^
# W0 = unrotated (W[0]..W[3]), still needs W[3] fixup
movaps $xmmW0, $xmmT2
xorps $xmmT1, $xmmT1 # rol(W0,1):
pcmpgtd $xmmW0, $xmmT1 # ffffffff for elements <0 (ones with msb bit 1)
paddd $xmmW0, $xmmW0 # shift left by 1
psubd $xmmT1, $xmmW0 # add 1 to those who had msb bit 1
# W0 = rotated (W[0]..W[3]), still needs W[3] fixup
pslldq \$12, $xmmT2 # lshift by 12 bytes: T2 = (0,0,0,unrotW[0])
movaps $xmmT2, $xmmT1
pslld \$2, $xmmT2
psrld \$30, $xmmT1
# xorps $xmmT1, $xmmT2 # rol((0,0,0,unrotW[0]),2)
xorps $xmmT1, $xmmW0 # same result, but does not depend on/does not modify T2
xorps $xmmT2, $xmmW0 # W0 = rol(W[0]..W[3],1) ^ (0,0,0,rol(unrotW[0],2))
"
# movq $xmmW0, %r8 # high latency (~6 cycles)
# movaps $xmmW0, $xmmT1
# psrldq \$8, $xmmT1 # rshift by 8 bytes: move upper 64 bits to lower
# movq $xmmT1, %r10 # high latency
# movq %r8, %r9
# movq %r10, %r11
# shrq \$32, %r9
# shrq \$32, %r11
# ^^^ slower than passing the results on stack (!!!)
echo "
movaps $xmmW0, $xmmT2
paddd $xmmRCONST, $xmmT2
movups $xmmT2, $dstmem
"
} }
# It's possible to interleave insns in rounds to mostly eliminate # It's possible to interleave integer insns in rounds to mostly eliminate
# dependency chains, but this likely to only help old Pentium-based # dependency chains, but this likely to only help old Pentium-based
# CPUs (ones without OOO, which can only simultaneously execute a pair # CPUs (ones without OOO, which can only simultaneously execute a pair
# of _adjacent_ insns). # of _adjacent_ insns).
@ -107,21 +259,16 @@ local n0=$(((n+0) & 15))
echo " echo "
# $n # $n
";test $n0 = 0 && echo " ";test $n0 = 0 && echo "
# W[0], already in %esi leal $RCONST(%r$e,%rsi), %e$e # e += RCONST + W[n]
";test $n0 != 0 && test $n0 -lt 8 && echo " ";test $n0 != 0 && test $n0 -lt 8 && echo "
movl `W32 $n0`, %esi # W[n] addl -64+4*$n0(%rsp), %e$e # e += RCONST + W[n]
";test $n0 -ge 8 && echo " ";test $n0 -ge 8 && echo "
# W[n], in %r$n0 leal $RCONST(%r$e,%r$n0), %e$e # e += RCONST + W[n]
";echo " ";echo "
movl %e$c, %edi # c movl %e$c, %edi # c
xorl %e$d, %edi # ^d xorl %e$d, %edi # ^d
andl %e$b, %edi # &b andl %e$b, %edi # &b
xorl %e$d, %edi # (((c ^ d) & b) ^ d) xorl %e$d, %edi # (((c ^ d) & b) ^ d)
";test $n0 -lt 8 && echo "
leal $RCONST(%r$e,%rsi), %e$e # e += RCONST + W[n]
";test $n0 -ge 8 && echo "
leal $RCONST(%r$e,%r$n0), %e$e # e += RCONST + W[n]
";echo "
addl %edi, %e$e # e += (((c ^ d) & b) ^ d) addl %edi, %e$e # e += (((c ^ d) & b) ^ d)
movl %e$a, %esi # movl %e$a, %esi #
roll \$5, %esi # rotl32(a,5) roll \$5, %esi # rotl32(a,5)
@ -138,28 +285,11 @@ local n2=$(((n+2) & 15))
local n0=$(((n+0) & 15)) local n0=$(((n+0) & 15))
echo " echo "
# $n # $n
";test $n0 -lt 8 && echo "
movl `W32 $n13`, %esi # W[(n+13) & 15]
xorl `W32 $n8`, %esi # ^W[(n+8) & 15]
xorl `W32 $n2`, %esi # ^W[(n+2) & 15]
xorl `W32 $n0`, %esi # ^W[n & 15]
roll %esi #
movl %esi, `W32 $n0` # store to W[n & 15]
";test $n0 -ge 8 && echo "
xorl `W32 $n13`, `W32 $n0` # W[n & 15] ^= W[(n+13) & 15]
xorl `W32 $n8`, `W32 $n0` # ^W[(n+8) & 15]
xorl `W32 $n2`, `W32 $n0` # ^W[(n+2) & 15]
roll `W32 $n0` #
";echo "
movl %e$c, %edi # c movl %e$c, %edi # c
xorl %e$d, %edi # ^d xorl %e$d, %edi # ^d
andl %e$b, %edi # &b andl %e$b, %edi # &b
xorl %e$d, %edi # (((c ^ d) & b) ^ d) xorl %e$d, %edi # (((c ^ d) & b) ^ d)
";test $n0 -lt 8 && echo " addl -64+4*$n0(%rsp), %e$e # e += RCONST + W[n & 15]
leal $RCONST(%r$e,%rsi), %e$e # e += RCONST + W[n & 15]
";test $n0 -ge 8 && echo "
leal $RCONST(%r$e,%r$n0), %e$e # e += RCONST + W[n & 15]
";echo "
addl %edi, %e$e # e += (((c ^ d) & b) ^ d) addl %edi, %e$e # e += (((c ^ d) & b) ^ d)
movl %e$a, %esi # movl %e$a, %esi #
roll \$5, %esi # rotl32(a,5) roll \$5, %esi # rotl32(a,5)
@ -167,13 +297,6 @@ echo "
rorl \$2, %e$b # b = rotl32(b,30) rorl \$2, %e$b # b = rotl32(b,30)
" "
} }
{
RCONST=0x5A827999
RD1A ax bx cx dx bp 0; RD1A bp ax bx cx dx 1; RD1A dx bp ax bx cx 2; RD1A cx dx bp ax bx 3; RD1A bx cx dx bp ax 4
RD1A ax bx cx dx bp 5; RD1A bp ax bx cx dx 6; RD1A dx bp ax bx cx 7; RD1A cx dx bp ax bx 8; RD1A bx cx dx bp ax 9
RD1A ax bx cx dx bp 10; RD1A bp ax bx cx dx 11; RD1A dx bp ax bx cx 12; RD1A cx dx bp ax bx 13; RD1A bx cx dx bp ax 14
RD1A ax bx cx dx bp 15; RD1B bp ax bx cx dx 16; RD1B dx bp ax bx cx 17; RD1B cx dx bp ax bx 18; RD1B bx cx dx bp ax 19
} | grep -v '^$'
RD2() { RD2() {
local a=$1;local b=$2;local c=$3;local d=$4;local e=$5 local a=$1;local b=$2;local c=$3;local d=$4;local e=$5
@ -184,27 +307,10 @@ local n2=$(((n+2) & 15))
local n0=$(((n+0) & 15)) local n0=$(((n+0) & 15))
echo " echo "
# $n # $n
";test $n0 -lt 8 && echo "
movl `W32 $n13`, %esi # W[(n+13) & 15]
xorl `W32 $n8`, %esi # ^W[(n+8) & 15]
xorl `W32 $n2`, %esi # ^W[(n+2) & 15]
xorl `W32 $n0`, %esi # ^W[n & 15]
roll %esi #
movl %esi, `W32 $n0` # store to W[n & 15]
";test $n0 -ge 8 && echo "
xorl `W32 $n13`, `W32 $n0` # W[n & 15] ^= W[(n+13) & 15]
xorl `W32 $n8`, `W32 $n0` # ^W[(n+8) & 15]
xorl `W32 $n2`, `W32 $n0` # ^W[(n+2) & 15]
roll `W32 $n0` #
";echo "
movl %e$c, %edi # c movl %e$c, %edi # c
xorl %e$d, %edi # ^d xorl %e$d, %edi # ^d
xorl %e$b, %edi # ^b xorl %e$b, %edi # ^b
";test $n0 -lt 8 && echo " addl -64+4*$n0(%rsp), %e$e # e += RCONST + W[n & 15]
leal $RCONST(%r$e,%rsi), %e$e # e += RCONST + W[n & 15]
";test $n0 -ge 8 && echo "
leal $RCONST(%r$e,%r$n0), %e$e # e += RCONST + W[n & 15]
";echo "
addl %edi, %e$e # e += (c ^ d ^ b) addl %edi, %e$e # e += (c ^ d ^ b)
movl %e$a, %esi # movl %e$a, %esi #
roll \$5, %esi # rotl32(a,5) roll \$5, %esi # rotl32(a,5)
@ -212,13 +318,6 @@ echo "
rorl \$2, %e$b # b = rotl32(b,30) rorl \$2, %e$b # b = rotl32(b,30)
" "
} }
{
RCONST=0x6ED9EBA1
RD2 ax bx cx dx bp 20; RD2 bp ax bx cx dx 21; RD2 dx bp ax bx cx 22; RD2 cx dx bp ax bx 23; RD2 bx cx dx bp ax 24
RD2 ax bx cx dx bp 25; RD2 bp ax bx cx dx 26; RD2 dx bp ax bx cx 27; RD2 cx dx bp ax bx 28; RD2 bx cx dx bp ax 29
RD2 ax bx cx dx bp 30; RD2 bp ax bx cx dx 31; RD2 dx bp ax bx cx 32; RD2 cx dx bp ax bx 33; RD2 bx cx dx bp ax 34
RD2 ax bx cx dx bp 35; RD2 bp ax bx cx dx 36; RD2 dx bp ax bx cx 37; RD2 cx dx bp ax bx 38; RD2 bx cx dx bp ax 39
} | grep -v '^$'
RD3() { RD3() {
local a=$1;local b=$2;local c=$3;local d=$4;local e=$5 local a=$1;local b=$2;local c=$3;local d=$4;local e=$5
@ -235,53 +334,82 @@ echo "
andl %e$c, %esi # si: b & c andl %e$c, %esi # si: b & c
andl %e$d, %edi # di: (b | c) & d andl %e$d, %edi # di: (b | c) & d
orl %esi, %edi # ((b | c) & d) | (b & c) orl %esi, %edi # ((b | c) & d) | (b & c)
";test $n0 -lt 8 && echo "
movl `W32 $n13`, %esi # W[(n+13) & 15]
xorl `W32 $n8`, %esi # ^W[(n+8) & 15]
xorl `W32 $n2`, %esi # ^W[(n+2) & 15]
xorl `W32 $n0`, %esi # ^W[n & 15]
roll %esi #
movl %esi, `W32 $n0` # store to W[n & 15]
";test $n0 -ge 8 && echo "
xorl `W32 $n13`, `W32 $n0` # W[n & 15] ^= W[(n+13) & 15]
xorl `W32 $n8`, `W32 $n0` # ^W[(n+8) & 15]
xorl `W32 $n2`, `W32 $n0` # ^W[(n+2) & 15]
roll `W32 $n0` #
";echo "
addl %edi, %e$e # += ((b | c) & d) | (b & c) addl %edi, %e$e # += ((b | c) & d) | (b & c)
";test $n0 -lt 8 && echo " addl -64+4*$n0(%rsp), %e$e # e += RCONST + W[n & 15]
leal $RCONST(%r$e,%rsi), %e$e # e += RCONST + W[n & 15]
";test $n0 -ge 8 && echo "
leal $RCONST(%r$e,%r$n0), %e$e # e += RCONST + W[n & 15]
";echo "
movl %e$a, %esi # movl %e$a, %esi #
roll \$5, %esi # rotl32(a,5) roll \$5, %esi # rotl32(a,5)
addl %esi, %e$e # e += rotl32(a,5) addl %esi, %e$e # e += rotl32(a,5)
rorl \$2, %e$b # b = rotl32(b,30) rorl \$2, %e$b # b = rotl32(b,30)
" "
} }
{ {
#RCONST=0x8F1BBCDC "out of range for signed 32bit displacement" # Round 1
RCONST=-0x70E44324 RCONST=0x5A827999
RD3 ax bx cx dx bp 40; RD3 bp ax bx cx dx 41; RD3 dx bp ax bx cx 42; RD3 cx dx bp ax bx 43; RD3 bx cx dx bp ax 44 RD1A ax bx cx dx bp 0; RD1A bp ax bx cx dx 1; RD1A dx bp ax bx cx 2; RD1A cx dx bp ax bx 3;
RD3 ax bx cx dx bp 45; RD3 bp ax bx cx dx 46; RD3 dx bp ax bx cx 47; RD3 cx dx bp ax bx 48; RD3 bx cx dx bp ax 49 RD1A bx cx dx bp ax 4; RD1A ax bx cx dx bp 5; RD1A bp ax bx cx dx 6; RD1A dx bp ax bx cx 7;
RD3 ax bx cx dx bp 50; RD3 bp ax bx cx dx 51; RD3 dx bp ax bx cx 52; RD3 cx dx bp ax bx 53; RD3 bx cx dx bp ax 54 a=`PREP %xmm0 %xmm1 %xmm2 %xmm3 "-64+16*0(%rsp)"`
RD3 ax bx cx dx bp 55; RD3 bp ax bx cx dx 56; RD3 dx bp ax bx cx 57; RD3 cx dx bp ax bx 58; RD3 bx cx dx bp ax 59 b=`RD1A cx dx bp ax bx 8; RD1A bx cx dx bp ax 9; RD1A ax bx cx dx bp 10; RD1A bp ax bx cx dx 11;`
} | grep -v '^$' INTERLEAVE "$a" "$b"
a=`echo " movaps rconst0x6ED9EBA1(%rip), $xmmRCONST"
PREP %xmm1 %xmm2 %xmm3 %xmm0 "-64+16*1(%rsp)"`
b=`RD1A dx bp ax bx cx 12; RD1A cx dx bp ax bx 13; RD1A bx cx dx bp ax 14; RD1A ax bx cx dx bp 15;`
INTERLEAVE "$a" "$b"
a=`PREP %xmm2 %xmm3 %xmm0 %xmm1 "-64+16*2(%rsp)"`
b=`RD1B bp ax bx cx dx 16; RD1B dx bp ax bx cx 17; RD1B cx dx bp ax bx 18; RD1B bx cx dx bp ax 19;`
INTERLEAVE "$a" "$b"
# Round 2
RCONST=0x6ED9EBA1
a=`PREP %xmm3 %xmm0 %xmm1 %xmm2 "-64+16*3(%rsp)"`
b=`RD2 ax bx cx dx bp 20; RD2 bp ax bx cx dx 21; RD2 dx bp ax bx cx 22; RD2 cx dx bp ax bx 23;`
INTERLEAVE "$a" "$b"
a=`PREP %xmm0 %xmm1 %xmm2 %xmm3 "-64+16*0(%rsp)"`
b=`RD2 bx cx dx bp ax 24; RD2 ax bx cx dx bp 25; RD2 bp ax bx cx dx 26; RD2 dx bp ax bx cx 27;`
INTERLEAVE "$a" "$b"
a=`PREP %xmm1 %xmm2 %xmm3 %xmm0 "-64+16*1(%rsp)"`
b=`RD2 cx dx bp ax bx 28; RD2 bx cx dx bp ax 29; RD2 ax bx cx dx bp 30; RD2 bp ax bx cx dx 31;`
INTERLEAVE "$a" "$b"
a=`echo " movaps rconst0x8F1BBCDC(%rip), $xmmRCONST"
PREP %xmm2 %xmm3 %xmm0 %xmm1 "-64+16*2(%rsp)"`
b=`RD2 dx bp ax bx cx 32; RD2 cx dx bp ax bx 33; RD2 bx cx dx bp ax 34; RD2 ax bx cx dx bp 35;`
INTERLEAVE "$a" "$b"
a=`PREP %xmm3 %xmm0 %xmm1 %xmm2 "-64+16*3(%rsp)"`
b=`RD2 bp ax bx cx dx 36; RD2 dx bp ax bx cx 37; RD2 cx dx bp ax bx 38; RD2 bx cx dx bp ax 39;`
INTERLEAVE "$a" "$b"
# Round 3
RCONST=0x8F1BBCDC
a=`PREP %xmm0 %xmm1 %xmm2 %xmm3 "-64+16*0(%rsp)"`
b=`RD3 ax bx cx dx bp 40; RD3 bp ax bx cx dx 41; RD3 dx bp ax bx cx 42; RD3 cx dx bp ax bx 43;`
INTERLEAVE "$a" "$b"
a=`PREP %xmm1 %xmm2 %xmm3 %xmm0 "-64+16*1(%rsp)"`
b=`RD3 bx cx dx bp ax 44; RD3 ax bx cx dx bp 45; RD3 bp ax bx cx dx 46; RD3 dx bp ax bx cx 47;`
INTERLEAVE "$a" "$b"
a=`PREP %xmm2 %xmm3 %xmm0 %xmm1 "-64+16*2(%rsp)"`
b=`RD3 cx dx bp ax bx 48; RD3 bx cx dx bp ax 49; RD3 ax bx cx dx bp 50; RD3 bp ax bx cx dx 51;`
INTERLEAVE "$a" "$b"
a=`echo " movaps rconst0xCA62C1D6(%rip), $xmmRCONST"
PREP %xmm3 %xmm0 %xmm1 %xmm2 "-64+16*3(%rsp)"`
b=`RD3 dx bp ax bx cx 52; RD3 cx dx bp ax bx 53; RD3 bx cx dx bp ax 54; RD3 ax bx cx dx bp 55;`
INTERLEAVE "$a" "$b"
a=`PREP %xmm0 %xmm1 %xmm2 %xmm3 "-64+16*0(%rsp)"`
b=`RD3 bp ax bx cx dx 56; RD3 dx bp ax bx cx 57; RD3 cx dx bp ax bx 58; RD3 bx cx dx bp ax 59;`
INTERLEAVE "$a" "$b"
# Round 4 has the same logic as round 2, only n and RCONST are different # Round 4 has the same logic as round 2, only n and RCONST are different
{ RCONST=0xCA62C1D6
#RCONST=0xCA62C1D6 "out of range for signed 32bit displacement" a=`PREP %xmm1 %xmm2 %xmm3 %xmm0 "-64+16*1(%rsp)"`
RCONST=-0x359D3E2A b=`RD2 ax bx cx dx bp 60; RD2 bp ax bx cx dx 61; RD2 dx bp ax bx cx 62; RD2 cx dx bp ax bx 63;`
RD2 ax bx cx dx bp 60; RD2 bp ax bx cx dx 61; RD2 dx bp ax bx cx 62; RD2 cx dx bp ax bx 63; RD2 bx cx dx bp ax 64 INTERLEAVE "$a" "$b"
RD2 ax bx cx dx bp 65; RD2 bp ax bx cx dx 66; RD2 dx bp ax bx cx 67; RD2 cx dx bp ax bx 68; RD2 bx cx dx bp ax 69 a=`PREP %xmm2 %xmm3 %xmm0 %xmm1 "-64+16*2(%rsp)"`
RD2 ax bx cx dx bp 70; RD2 bp ax bx cx dx 71; RD2 dx bp ax bx cx 72; RD2 cx dx bp ax bx 73; RD2 bx cx dx bp ax 74 b=`RD2 bx cx dx bp ax 64; RD2 ax bx cx dx bp 65; RD2 bp ax bx cx dx 66; RD2 dx bp ax bx cx 67;`
RD2 ax bx cx dx bp 75; RD2 bp ax bx cx dx 76; RD2 dx bp ax bx cx 77; RD2 cx dx bp ax bx 78; RD2 bx cx dx bp ax 79 INTERLEAVE "$a" "$b"
# Note: new W[n&15] values generated in last 3 iterations a=`PREP %xmm3 %xmm0 %xmm1 %xmm2 "-64+16*3(%rsp)"`
# (W[13,14,15]) are unused after each of these iterations. b=`RD2 cx dx bp ax bx 68; RD2 bx cx dx bp ax 69; RD2 ax bx cx dx bp 70; RD2 bp ax bx cx dx 71;`
# Since we use r8..r15 for W[8..15], this does not matter. INTERLEAVE "$a" "$b"
# If we switch to e.g. using r8..r15 for W[0..7], then saving of W[13,14,15] RD2 dx bp ax bx cx 72; RD2 cx dx bp ax bx 73; RD2 bx cx dx bp ax 74; RD2 ax bx cx dx bp 75;
# (the "movl %esi, `W32 $n0`" insn) is a dead store and can be removed. RD2 bp ax bx cx dx 76; RD2 dx bp ax bx cx 77; RD2 cx dx bp ax bx 78; RD2 bx cx dx bp ax 79;
} | grep -v '^$' } | grep -v '^$'
echo " echo "
@ -300,4 +428,28 @@ echo "
ret ret
.size sha1_process_block64, .-sha1_process_block64 .size sha1_process_block64, .-sha1_process_block64
.section .rodata.cst16.sha1const, \"aM\", @progbits, 16
.align 16
rconst0x5A827999:
.long 0x5A827999
.long 0x5A827999
.long 0x5A827999
.long 0x5A827999
rconst0x6ED9EBA1:
.long 0x6ED9EBA1
.long 0x6ED9EBA1
.long 0x6ED9EBA1
.long 0x6ED9EBA1
rconst0x8F1BBCDC:
.long 0x8F1BBCDC
.long 0x8F1BBCDC
.long 0x8F1BBCDC
.long 0x8F1BBCDC
rconst0xCA62C1D6:
.long 0xCA62C1D6
.long 0xCA62C1D6
.long 0xCA62C1D6
.long 0xCA62C1D6
#endif" #endif"