qiurui
/
glibc
mirror of git://sourceware.org/git/glibc.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
glibc/sysdeps/powerpc/powerpc64/le/power10/strlen.S

222 lines
4.9 KiB
ArmAsm
Raw Normal View History

powerpc: Add optimized strlen for POWER10 Improvements compared to POWER9 version: 1. Take into account first 16B comparison for aligned strings The previous version compares the first 16B and increments r4 by the number of bytes until the address is 16B-aligned, then starts doing aligned loads at that address. For aligned strings, this causes the first 16B to be compared twice, because the increment is 0. Here we calculate the next 16B-aligned address differently, which avoids that issue. 2. Use simple comparisons for the first ~192 bytes The main loop is good for big strings, but comparing 16B each time is better for smaller strings. So after aligning the address to 16 Bytes, we check more 176B in 16B chunks. There may be some overlaps with the main loop for unaligned strings, but we avoid using the more aggressive strategy too soon, and also allow the loop to start at a 64B-aligned address. This greatly benefits smaller strings and avoids overlapping checks if the string is already aligned at a 64B boundary. 3. Reduce dependencies between load blocks caused by address calculation on loop Doing a precise time tracing on the code showed many loads in the loop were stalled waiting for updates to r4 from previous code blocks. This implementation avoids that as much as possible by using 2 registers (r4 and r5) to hold addresses to be used by different parts of the code. Also, the previous code aligned the address to 16B, then to 64B by doing a few 48B loops (if needed) until the address was aligned. The main loop could not start until that 48B loop had finished and r4 was updated with the current address. Here we calculate the address used by the loop very early, so it can start sooner. The main loop now uses 2 pointers 128B apart to make pointer updates less frequent, and also unrolls 1 iteration to guarantee there is enough time between iterations to update the pointers, reducing stalled cycles. 4. Use new P10 instructions lxvp is used to load 32B with a single instruction, reducing contention in the load queue. vextractbm allows simplifying the tail code for the loop, replacing vbpermq and avoiding having to generate a permute control vector. Reviewed-by: Paul E Murphy <murphyp@linux.ibm.com> Reviewed-by: Raphael M Zinsly <rzinsly@linux.ibm.com> Reviewed-by: Lucas A. M. Magalhaes <lamm@linux.ibm.com>
2020-09-29 18:40:08 +00:00
/* Optimized strlen implementation for POWER10 LE.
Copyright (C) 2021 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
<https://www.gnu.org/licenses/>. */
#include <sysdep.h>
#ifndef STRLEN
# define STRLEN __strlen
# define DEFINE_STRLEN_HIDDEN_DEF 1
#endif
/* TODO: Replace macros by the actual instructions when minimum binutils becomes
>= 2.35. This is used to keep compatibility with older versions. */
#define VEXTRACTBM(rt,vrb) \
.long(((4)<<(32-6)) \
| ((rt)<<(32-11)) \
| ((8)<<(32-16)) \
| ((vrb)<<(32-21)) \
| 1602)
#define LXVP(xtp,dq,ra) \
.long(((6)<<(32-6)) \
| ((((xtp)-32)>>1)<<(32-10)) \
| ((1)<<(32-11)) \
| ((ra)<<(32-16)) \
| dq)
#define CHECK16(vreg,offset,addr,label) \
lxv vreg+32,offset(addr); \
vcmpequb. vreg,vreg,v18; \
bne cr6,L(label);
/* Load 4 quadwords, merge into one VR for speed and check for NULLs. r6 has #
of bytes already checked. */
#define CHECK64(offset,addr,label) \
li r6,offset; \
LXVP(v4+32,offset,addr); \
LXVP(v6+32,offset+32,addr); \
vminub v14,v4,v5; \
vminub v15,v6,v7; \
vminub v16,v14,v15; \
vcmpequb. v0,v16,v18; \
bne cr6,L(label)
#define TAIL(vreg,increment) \
vctzlsbb r4,vreg; \
subf r3,r3,r5; \
addi r4,r4,increment; \
add r3,r3,r4; \
blr
/* Implements the function
int [r3] strlen (const void *s [r3])
The implementation can load bytes past a matching byte, but only
up to the next 64B boundary, so it never crosses a page. */
.machine power9
ENTRY_TOCLESS (STRLEN, 4)
CALL_MCOUNT 1
vspltisb v18,0
vspltisb v19,-1
/* Next 16B-aligned address. Prepare address for L(aligned). */
addi r5,r3,16
clrrdi r5,r5,4
/* Align data and fill bytes not loaded with non matching char. */
lvx v0,0,r3
lvsr v1,0,r3
vperm v0,v19,v0,v1
vcmpequb. v6,v0,v18
beq cr6,L(aligned)
vctzlsbb r3,v6
blr
/* Test next 176B, 16B at a time. The main loop is optimized for longer
strings, so checking the first bytes in 16B chunks benefits a lot
small strings. */
.p2align 5
L(aligned):
/* Prepare address for the loop. */
addi r4,r3,192
clrrdi r4,r4,6
CHECK16(v0,0,r5,tail1)
CHECK16(v1,16,r5,tail2)
CHECK16(v2,32,r5,tail3)
CHECK16(v3,48,r5,tail4)
CHECK16(v4,64,r5,tail5)
CHECK16(v5,80,r5,tail6)
CHECK16(v6,96,r5,tail7)
CHECK16(v7,112,r5,tail8)
CHECK16(v8,128,r5,tail9)
CHECK16(v9,144,r5,tail10)
CHECK16(v10,160,r5,tail11)
addi r5,r4,128
/* Switch to a more aggressive approach checking 64B each time. Use 2
pointers 128B apart and unroll the loop once to make the pointer
updates and usages separated enough to avoid stalls waiting for
address calculation. */
.p2align 5
L(loop):
CHECK64(0,r4,pre_tail_64b)
CHECK64(64,r4,pre_tail_64b)
addi r4,r4,256
CHECK64(0,r5,tail_64b)
CHECK64(64,r5,tail_64b)
addi r5,r5,256
b L(loop)
.p2align 5
L(pre_tail_64b):
mr r5,r4
L(tail_64b):
/* OK, we found a null byte. Let's look for it in the current 64-byte
block and mark it in its corresponding VR. lxvp vx,0(ry) puts the
low 16B bytes into vx+1, and the high into vx, so the order here is
v5, v4, v7, v6. */
vcmpequb v1,v5,v18
vcmpequb v2,v4,v18
vcmpequb v3,v7,v18
vcmpequb v4,v6,v18
/* Take into account the other 64B blocks we had already checked. */
add r5,r5,r6
/* Extract first bit of each byte. */
VEXTRACTBM(r7,v1)
VEXTRACTBM(r8,v2)
VEXTRACTBM(r9,v3)
VEXTRACTBM(r10,v4)
/* Shift each value into their corresponding position. */
sldi r8,r8,16
sldi r9,r9,32
sldi r10,r10,48
/* Merge the results. */
or r7,r7,r8
or r8,r9,r10
or r10,r8,r7
cnttzd r0,r10 /* Count trailing zeros before the match. */
subf r5,r3,r5
add r3,r5,r0 /* Compute final length. */
blr
.p2align 5
L(tail1):
TAIL(v0,0)
.p2align 5
L(tail2):
TAIL(v1,16)
.p2align 5
L(tail3):
TAIL(v2,32)
.p2align 5
L(tail4):
TAIL(v3,48)
.p2align 5
L(tail5):
TAIL(v4,64)
.p2align 5
L(tail6):
TAIL(v5,80)
.p2align 5
L(tail7):
TAIL(v6,96)
.p2align 5
L(tail8):
TAIL(v7,112)
.p2align 5
L(tail9):
TAIL(v8,128)
.p2align 5
L(tail10):
TAIL(v9,144)
.p2align 5
L(tail11):
TAIL(v10,160)
END (STRLEN)
#ifdef DEFINE_STRLEN_HIDDEN_DEF
weak_alias (__strlen, strlen)
libc_hidden_builtin_def (strlen)
#endif