glibc/sysdeps/powerpc/powerpc64/power8/strspn.S
Raoni Fassina Firmino 066020c5e8 powerpc: Cleanup: use actual power8 assembly mnemonics
Some implementations in sysdeps/powerpc/powerpc64/power8/*.S still had
pre power8 compatible binutils hardcoded macros and were not using
.machine power8.

This patch should not have semantic changes, in fact it should have the
same exact code generated.

Tested that generated stripped shared objects are identical when
using "strip --remove-section=.note.gnu.build-id".

Checked on:
- powerpc64le, power9, build-many-glibcs.py, gcc 6.4.1 20180104, binutils 2.26.2.20160726
- powerpc64le, power8, debian 9, gcc 6.3.0 20170516, binutils 2.28
- powerpc64le, power9, ubuntu 19.04, gcc 8.3.0, binutils 2.32
- powerpc64le, power9, opensuse tumbleweed, gcc 9.1.1 20190527, binutils 2.32
- powerpc64, power9, debian 10, gcc 8.3.0, binutils 2.31.1

Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Reviewed-by: Gabriel F. T. Gomes <gabrielftg@linux.ibm.com>
2019-08-01 15:57:50 -03:00

188 lines
5.1 KiB
ArmAsm

/* Optimized strspn implementation for Power8.
Copyright (C) 2016-2019 Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
The GNU C Library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with the GNU C Library; if not, see
<http://www.gnu.org/licenses/>. */
/* size_t [r3] strspn (const char *string [r3],
const char *needleAccept [r4]) */
/* This takes a novel approach by computing a 256 bit mask whereby
each set bit implies the byte is "accepted". P8 vector hardware
has extremely efficient hardware for selecting bits from a mask.
One might ask "why not use bpermd for short strings"? It is
so slow that its performance about matches the generic PPC64
variant without any fancy masking, with the added expense of
making the mask. That was the first variant of this. */
#include "sysdep.h"
#ifndef USE_AS_STRCSPN
# define USE_AS_STRCSPN 0
# ifndef STRSPN
# define STRSPN strspn
# endif
# define INITIAL_MASK 0
# define UPDATE_MASK(RA, RS, RB) or RA, RS, RB
#else
# ifndef STRSPN
# define STRSPN strcspn
# endif
# define INITIAL_MASK -1
# define UPDATE_MASK(RA, RS, RB) andc RA, RS, RB
#endif
/* Simple macro to use VSX instructions in overlapping VR's. */
#define XXVR(insn, vrt, vra, vrb) \
insn 32+vrt, 32+vra, 32+vrb
.machine power8
ENTRY_TOCLESS (STRSPN, 4)
CALL_MCOUNT 2
/* Generate useful constants for later on. */
vspltisb v1, 7
vspltisb v2, -1
vslb v1, v1, v1 /* 0x80 to swap high bit for vbpermq. */
vspltisb v10, 0
vsldoi v4, v10, v2, 2 /* 0xFFFF into vr4. */
XXVR(xxmrgld, v4, v4, v10) /* Mask for checking matches. */
/* Prepare to compute 256b mask. */
addi r4, r4, -1
li r5, INITIAL_MASK
li r6, INITIAL_MASK
li r7, INITIAL_MASK
li r8, INITIAL_MASK
#if USE_AS_STRCSPN
/* Ensure the null character never matches by clearing ISA bit 0 in
in r5 which is the bit which will check for it in the later usage
of vbpermq. */
srdi r5, r5, 1
#endif
li r11, 1
sldi r11, r11, 63
/* Start interleaved Mask computation.
This will eventually or 1's into ignored bits from vbpermq. */
lvsr v11, 0, r3
vspltb v11, v11, 0 /* Splat shift constant. */
/* Build a 256b mask in r5-r8. */
.align 4
L(next_needle):
lbzu r9, 1(r4)
cmpldi cr0, r9, 0
cmpldi cr1, r9, 128
/* This is a little tricky. srd only uses the first 7 bits,
and if bit 7 is set, value is always 0. So, we can
effectively shift 128b in this case. */
xori r12, r9, 0x40 /* Invert bit 6. */
srd r10, r11, r9 /* Mask for bits 0-63. */
srd r12, r11, r12 /* Mask for bits 64-127. */
beq cr0, L(start_cmp)
/* Now, or the value into the correct GPR. */
bge cr1,L(needle_gt128)
UPDATE_MASK (r5, r5, r10) /* 0 - 63. */
UPDATE_MASK (r6, r6, r12) /* 64 - 127. */
b L(next_needle)
.align 4
L(needle_gt128):
UPDATE_MASK (r7, r7, r10) /* 128 - 191. */
UPDATE_MASK (r8, r8, r12) /* 192 - 255. */
b L(next_needle)
.align 4
L(start_cmp):
/* Move and merge bitmap into 2 VRs. bpermd is slower on P8. */
mr r0, r3 /* Save r3 for final length computation. */
mtvrd v5, r5
mtvrd v6, r6
mtvrd v7, r7
mtvrd v8, r8
/* Continue interleaved mask generation. */
#ifdef __LITTLE_ENDIAN__
vsrw v11, v2, v11 /* Note, shift ignores higher order bits. */
vsplth v11, v11, 0 /* Only care about the high 16 bits of v10. */
#else
vslw v11, v2, v11 /* Note, shift ignores higher order bits. */
vsplth v11, v11, 1 /* Only care about the low 16 bits of v10. */
#endif
lvx v0, 0, r3 /* Note, unaligned load ignores lower bits. */
/* Do the merging of the bitmask. */
XXVR(xxmrghd, v5, v5, v6)
XXVR(xxmrghd, v6, v7, v8)
/* Finish mask generation. */
vand v11, v11, v4 /* Throwaway bits not in the mask. */
/* Compare the first 1-16B, while masking unwanted bytes. */
clrrdi r3, r3, 4 /* Note, counts from qw boundaries. */
vxor v9, v0, v1 /* Swap high bit. */
vbpermq v8, v5, v0
vbpermq v7, v6, v9
vor v7, v7, v8
vor v7, v7, v11 /* Ignore non-participating bytes. */
vcmpequh. v8, v7, v4
bnl cr6, L(done)
addi r3, r3, 16
.align 4
L(vec):
lvx v0, 0, r3
addi r3, r3, 16
vxor v9, v0, v1 /* Swap high bit. */
vbpermq v8, v5, v0
vbpermq v7, v6, v9
vor v7, v7, v8
vcmpequh. v8, v7, v4
blt cr6, L(vec)
addi r3, r3, -16
L(done):
subf r3, r0, r3
mfvrd r10, v7
#ifdef __LITTLE_ENDIAN__
addi r0, r10, 1 /* Count the trailing 1's. */
andc r10, r10, r0
popcntd r10, r10
#else
xori r10, r10, 0xffff /* Count leading 1's by inverting. */
addi r3, r3, -48 /* Account for the extra leading zeros. */
cntlzd r10, r10
#endif
add r3, r3, r10
blr
END(STRSPN)
libc_hidden_builtin_def (STRSPN)