mirror of
https://github.com/jart/cosmopolitan.git
synced 2025-01-31 11:37:35 +00:00
b8a6a989c0
This change greatly reduces the number of modules that need to be compiled. The only issue right now is that sometimes when viewing symbol table entries, the aliased symbol is chosen.
177 lines
7.2 KiB
ArmAsm
177 lines
7.2 KiB
ArmAsm
/*-*- mode:unix-assembly; indent-tabs-mode:t; tab-width:8; coding:utf-8 -*-│
|
|
│vi: set et ft=asm ts=8 tw=8 fenc=utf-8 :vi│
|
|
╚──────────────────────────────────────────────────────────────────────────────╝
|
|
│ │
|
|
│ Optimized Routines │
|
|
│ Copyright (c) 1999-2022, Arm Limited. │
|
|
│ │
|
|
│ Permission is hereby granted, free of charge, to any person obtaining │
|
|
│ a copy of this software and associated documentation files (the │
|
|
│ "Software"), to deal in the Software without restriction, including │
|
|
│ without limitation the rights to use, copy, modify, merge, publish, │
|
|
│ distribute, sublicense, and/or sell copies of the Software, and to │
|
|
│ permit persons to whom the Software is furnished to do so, subject to │
|
|
│ the following conditions: │
|
|
│ │
|
|
│ The above copyright notice and this permission notice shall be │
|
|
│ included in all copies or substantial portions of the Software. │
|
|
│ │
|
|
│ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, │
|
|
│ EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF │
|
|
│ MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. │
|
|
│ IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY │
|
|
│ CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, │
|
|
│ TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE │
|
|
│ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. │
|
|
│ │
|
|
╚─────────────────────────────────────────────────────────────────────────────*/
|
|
#include "libc/intrin/aarch64/asmdefs.internal.h"
|
|
|
|
#define __strrchr_aarch64 strrchr
|
|
#define __rindex_aarch64 rindex
|
|
|
|
.ident "\n\n\
|
|
Optimized Routines (MIT License)\n\
|
|
Copyright 2022 ARM Limited\n"
|
|
.include "libc/disclaimer.inc"
|
|
|
|
/* Assumptions:
|
|
*
|
|
* ARMv8-a, AArch64
|
|
* Neon Available.
|
|
*/
|
|
|
|
/* Arguments and results. */
|
|
#define srcin x0
|
|
#define chrin w1
|
|
|
|
#define result x0
|
|
|
|
#define src x2
|
|
#define tmp1 x3
|
|
#define wtmp2 w4
|
|
#define tmp3 x5
|
|
#define src_match x6
|
|
#define src_offset x7
|
|
#define const_m1 x8
|
|
#define tmp4 x9
|
|
#define nul_match x10
|
|
#define chr_match x11
|
|
|
|
#define vrepchr v0
|
|
#define vdata1 v1
|
|
#define vdata2 v2
|
|
#define vhas_nul1 v3
|
|
#define vhas_nul2 v4
|
|
#define vhas_chr1 v5
|
|
#define vhas_chr2 v6
|
|
#define vrepmask_0 v7
|
|
#define vrepmask_c v16
|
|
#define vend1 v17
|
|
#define vend2 v18
|
|
|
|
/* Core algorithm.
|
|
|
|
For each 32-byte hunk we calculate a 64-bit syndrome value, with
|
|
two bits per byte (LSB is always in bits 0 and 1, for both big
|
|
and little-endian systems). For each tuple, bit 0 is set iff
|
|
the relevant byte matched the requested character; bit 1 is set
|
|
iff the relevant byte matched the NUL end of string (we trigger
|
|
off bit0 for the special case of looking for NUL). Since the bits
|
|
in the syndrome reflect exactly the order in which things occur
|
|
in the original string a count_trailing_zeros() operation will
|
|
identify exactly which byte is causing the termination, and why. */
|
|
|
|
ENTRY_ALIAS (__strrchr_aarch64, __rindex_aarch64)
|
|
ENTRY (__strrchr_aarch64)
|
|
PTR_ARG (0)
|
|
/* Magic constant 0x40100401 to allow us to identify which lane
|
|
matches the requested byte. Magic constant 0x80200802 used
|
|
similarly for NUL termination. */
|
|
mov wtmp2, #0x0401
|
|
movk wtmp2, #0x4010, lsl #16
|
|
dup vrepchr.16b, chrin
|
|
bic src, srcin, #31 /* Work with aligned 32-byte hunks. */
|
|
dup vrepmask_c.4s, wtmp2
|
|
mov src_offset, #0
|
|
ands tmp1, srcin, #31
|
|
add vrepmask_0.4s, vrepmask_c.4s, vrepmask_c.4s /* equiv: lsl #1 */
|
|
b.eq L(aligned)
|
|
|
|
/* Input string is not 32-byte aligned. Rather than forcing
|
|
the padding bytes to a safe value, we calculate the syndrome
|
|
for all the bytes, but then mask off those bits of the
|
|
syndrome that are related to the padding. */
|
|
ld1 {vdata1.16b, vdata2.16b}, [src], #32
|
|
neg tmp1, tmp1
|
|
cmeq vhas_nul1.16b, vdata1.16b, #0
|
|
cmeq vhas_chr1.16b, vdata1.16b, vrepchr.16b
|
|
cmeq vhas_nul2.16b, vdata2.16b, #0
|
|
cmeq vhas_chr2.16b, vdata2.16b, vrepchr.16b
|
|
and vhas_nul1.16b, vhas_nul1.16b, vrepmask_0.16b
|
|
and vhas_chr1.16b, vhas_chr1.16b, vrepmask_c.16b
|
|
and vhas_nul2.16b, vhas_nul2.16b, vrepmask_0.16b
|
|
and vhas_chr2.16b, vhas_chr2.16b, vrepmask_c.16b
|
|
addp vhas_nul1.16b, vhas_nul1.16b, vhas_nul2.16b // 256->128
|
|
addp vhas_chr1.16b, vhas_chr1.16b, vhas_chr2.16b // 256->128
|
|
addp vend1.16b, vhas_nul1.16b, vhas_chr1.16b // 128->64
|
|
mov nul_match, vend1.d[0]
|
|
lsl tmp1, tmp1, #1
|
|
mov const_m1, #~0
|
|
lsr tmp3, const_m1, tmp1
|
|
mov chr_match, vend1.d[1]
|
|
|
|
bic nul_match, nul_match, tmp3 // Mask padding bits.
|
|
bic chr_match, chr_match, tmp3 // Mask padding bits.
|
|
cbnz nul_match, L(tail)
|
|
|
|
.p2align 4
|
|
L(loop):
|
|
cmp chr_match, #0
|
|
csel src_match, src, src_match, ne
|
|
csel src_offset, chr_match, src_offset, ne
|
|
L(aligned):
|
|
ld1 {vdata1.16b, vdata2.16b}, [src], #32
|
|
cmeq vhas_chr1.16b, vdata1.16b, vrepchr.16b
|
|
cmeq vhas_chr2.16b, vdata2.16b, vrepchr.16b
|
|
uminp vend1.16b, vdata1.16b, vdata2.16b
|
|
and vhas_chr1.16b, vhas_chr1.16b, vrepmask_c.16b
|
|
and vhas_chr2.16b, vhas_chr2.16b, vrepmask_c.16b
|
|
cmeq vend1.16b, vend1.16b, 0
|
|
addp vhas_chr1.16b, vhas_chr1.16b, vhas_chr2.16b // 256->128
|
|
addp vend1.16b, vend1.16b, vhas_chr1.16b // 128->64
|
|
mov nul_match, vend1.d[0]
|
|
mov chr_match, vend1.d[1]
|
|
cbz nul_match, L(loop)
|
|
|
|
cmeq vhas_nul1.16b, vdata1.16b, #0
|
|
cmeq vhas_nul2.16b, vdata2.16b, #0
|
|
and vhas_nul1.16b, vhas_nul1.16b, vrepmask_0.16b
|
|
and vhas_nul2.16b, vhas_nul2.16b, vrepmask_0.16b
|
|
addp vhas_nul1.16b, vhas_nul1.16b, vhas_nul2.16b
|
|
addp vhas_nul1.16b, vhas_nul1.16b, vhas_nul1.16b
|
|
mov nul_match, vhas_nul1.d[0]
|
|
|
|
L(tail):
|
|
/* Work out exactly where the string ends. */
|
|
sub tmp4, nul_match, #1
|
|
eor tmp4, tmp4, nul_match
|
|
ands chr_match, chr_match, tmp4
|
|
/* And pick the values corresponding to the last match. */
|
|
csel src_match, src, src_match, ne
|
|
csel src_offset, chr_match, src_offset, ne
|
|
|
|
/* Count down from the top of the syndrome to find the last match. */
|
|
clz tmp3, src_offset
|
|
/* Src_match points beyond the word containing the match, so we can
|
|
simply subtract half the bit-offset into the syndrome. Because
|
|
we are counting down, we need to go back one more character. */
|
|
add tmp3, tmp3, #2
|
|
sub result, src_match, tmp3, lsr #1
|
|
/* But if the syndrome shows no match was found, then return NULL. */
|
|
cmp src_offset, #0
|
|
csel result, result, xzr, ne
|
|
|
|
ret
|
|
|
|
END (__strrchr_aarch64)
|