From: Paul E Murphy <murphyp@linux.ibm.com>
To: Amrita H S <amritahs@linux.vnet.ibm.com>, libc-alpha@sourceware.org
Subject: Re: [PATCH V4] powerpc: Optimized strcmp for power10
Date: Thu, 30 Nov 2023 10:54:09 -0600 [thread overview]
Message-ID: <54263bc9-c177-4825-8ef5-7c8d8ea22446@linux.ibm.com> (raw)
In-Reply-To: <20231130161159.2742990-1-amritahs@linux.vnet.ibm.com>
On 11/30/23 10:11 AM, Amrita H S wrote:
> This patch is based on __strcmp_power9 and __strlen_power10.
>
> Improvements from __strcmp_power9:
>
> 1. Uses new POWER10 instructions
> - This code uses lxvp to decrease contention on load
> by loading 32 bytes per instruction.
>
> 2. Performance implication
> - This version has around 30% better performance on average.
> - Performance regression is seen for a specific combination
> of sizes and alignments. Some of them is observed without
> changes also, while rest may be induced by the patch.
>
> Signed-off-by: Amrita H S <amritahs@linux.vnet.ibm.com>
> ---
> sysdeps/powerpc/powerpc64/le/power10/strcmp.S | 190 ++++++++++++++++++
> sysdeps/powerpc/powerpc64/multiarch/Makefile | 3 +-
> .../powerpc64/multiarch/ifunc-impl-list.c | 4 +
> .../powerpc64/multiarch/strcmp-power10.S | 26 +++
> sysdeps/powerpc/powerpc64/multiarch/strcmp.c | 4 +
> 5 files changed, 226 insertions(+), 1 deletion(-)
> create mode 100644 sysdeps/powerpc/powerpc64/le/power10/strcmp.S
> create mode 100644 sysdeps/powerpc/powerpc64/multiarch/strcmp-power10.S
>
> diff --git a/sysdeps/powerpc/powerpc64/le/power10/strcmp.S b/sysdeps/powerpc/powerpc64/le/power10/strcmp.S
> new file mode 100644
> index 0000000000..056efae948
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc64/le/power10/strcmp.S
> @@ -0,0 +1,190 @@
> +/* Optimized strcmp implementation for PowerPC64/POWER10.
> + Copyright (C) 2021-2023 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 STRCMP
> +# define STRCMP strcmp
> +#endif
> +
> +/* Implements the function
> + int [r3] strcmp (const char *s1 [r3], const char *s2 [r4]). */
> +
> +/* TODO: Change this to actual instructions when minimum binutils is upgraded
> + to 2.27. Macros are defined below for these newer instructions in order
> + to maintain compatibility. */
> +
> +#define LXVP(xtp,dq,ra) \
> + .long(((6)<<(32-6)) \
> + | ((((xtp)-32)>>1)<<(32-10)) \
> + | ((1)<<(32-11)) \
> + | ((ra)<<(32-16)) \
> + | dq)
> +
> +#define COMPARE_16(vreg1,vreg2,offset) \
> + lxv vreg1+32,offset(r3); \
> + lxv vreg2+32,offset(r4); \
> + vcmpnezb. v7,vreg1,vreg2; \
> + bne cr6,L(different); \
> +
> +#define COMPARE_32(vreg1,vreg2,offset,label1,label2) \
> + LXVP(vreg1+32,offset,r3); \
> + LXVP(vreg2+32,offset,r4); \
> + vcmpnezb. v7,vreg1+1,vreg2+1; \
> + bne cr6,L(label1); \
> + vcmpnezb. v7,vreg1,vreg2; \
> + bne cr6,L(label2); \
> +
> +#define TAIL(vreg1,vreg2) \
> + vctzlsbb r6,v7; \
> + vextubrx r5,r6,vreg1; \
> + vextubrx r4,r6,vreg2; \
> + subf r3,r4,r5; \
> + blr; \
> +
> +#define CHECK_N_BYTES(reg1,reg2,len_reg) \
> + mr r0,len_reg; \
> + sldi r0,r0,56; \
> + lxvl 32+v4,reg1,r0; \
> + lxvl 32+v5,reg2,r0; \
> + add reg1,reg1,len_reg; \
> + add reg2,reg2,len_reg; \
> + vcmpnezb. v7,v4,v5; \
> + vctzlsbb r6,v7; \
> + cmpld cr7,r6,len_reg; \
> + blt cr7,L(different); \
> +
> + /* TODO: change this to .machine power10 when the minimum required
> + binutils allows it. */
> +
> + .machine power9
> +ENTRY_TOCLESS (STRCMP, 4)
> + li r11,16
> + /* r12 is used as swap status flag to indicate if source pointers
> + were swapped. */
> + li r12,0
> + vspltisb v19,-1
> + andi. r7,r3,15
> + sub r7,r11,r7 /* r7(nalign1) = 16 - (str1 & 15). */
> + andi. r9,r4,15
> + sub r5,r11,r9 /* r5(nalign2) = 16 - (str2 & 15). */
> + cmpld cr7,r7,r5
> + beq cr7,L(same_aligned)
> + blt cr7,L(nalign1_min)
> + /* Swap r3 and r4, and r7 and r8 such that r3 and r7 hold the
> + pointer which is closer to the next 16B boundary so that only
> + one CHECK_N_BYTES is needed before entering the loop below. */
> + mr r8,r4
> + mr r4,r3
> + mr r3,r8
> + mr r7,r5
> + li r12,1 /* Set flag on swapping source pointers. */
> +L(nalign1_min):
> + CHECK_N_BYTES(r3,r4,r7)
> +
> +L(s1_aligned):
> + li r10,4096
> + andi. r9,r4,15
> + sub r8,r10,r9 /* r8 is rem_step = page_size - (str2&15). */
We know that s1 (in r3) is aligned to 16B now. We only need to know the
maximum number of bytes we can check without crossing a page when
loading from s2 when the quadword crosses a page. I think that value is
r8-r7 after aligning s1 above. Likewise, the second check is always
16-(r8-r7) bytes.
Also, you can use subfic to subtract from a 16 bit constant (the carry
bit is caller-save, and safe to clobber here).
> + rlwinm r7,r4,28,0xFF /* r7 is rem = (4096-(str2&(4096-1)))/16. */
> + subfic r7,r7,255 /* This can be reduced to 255-(r4&4095)>>4. */
> + /* Below check is required only for first iteration. For second
> + iteration and beyond, new loop counter is always 255. */
> + cmpldi r7,0
> + ble L(L3)
> +L(L1):
> + mtctr r7
> +L(L2):
> + COMPARE_16(v4,v5,0) /* Load 16B blocks using lxv. */
> + addi r3,r3,16
> + addi r4,r4,16
> + bdnz L(L2)
> + /* Cross the page boundary of s2, carefully. r9 ans r5 is number of
> + bytes to be read after and before page boundary correspondingly. */
> +L(L3):
> + andi. r9,r4,15
> + sub r5,r11,r9
> + CHECK_N_BYTES(r3,r4,r5)
> + CHECK_N_BYTES(r3,r4,r9)
From my comment above, this can be simplified if the number of bytes to
check pre/post page cross are stored in a register.
> + li r7,255
> + b L(L1)
> +
> +L(same_aligned):
> + CHECK_N_BYTES(r3,r4,r7)
> + /* Align s1 to 32B and adjust s2 address.
> + Use lxvp only if both s1 and s2 are 32B aligned. */
> + COMPARE_16(v4,v5,0)
> + COMPARE_16(v4,v5,16)
> + COMPARE_16(v4,v5,32)
> + COMPARE_16(v4,v5,48)
> + addi r3,r3,64
> + addi r4,r4,64
> + COMPARE_16(v4,v5,0)
> + COMPARE_16(v4,v5,16)
> +
> + clrldi r6,r3,59
> + subfic r5,r6,32
> + add r3,r3,r5
> + add r4,r4,r5
> + andi. r5,r4,0x1F
> + beq cr0,L(32B_aligned_loop)
> +
> +L(16B_aligned_loop):
> + COMPARE_16(v4,v5,0)
> + COMPARE_16(v4,v5,16)
> + COMPARE_16(v4,v5,32)
> + COMPARE_16(v4,v5,48)
> + addi r3,r3,64
> + addi r4,r4,64
> + b L(16B_aligned_loop)
> +
> + /* Calculate and return the difference. */
> +L(different):
> + vctzlsbb r6,v7
> + vextubrx r5,r6,v4
> + vextubrx r4,r6,v5
> + cmpldi r12,1
> + beq L(swapped)
> + subf r3,r4,r5
> + blr
> +
> + /* If src pointers were swapped, then swap the
> + indices and calculate the return value. */
> +L(swapped):
> + subf r3,r5,r4
> + blr
> +
> +L(32B_aligned_loop):
> + COMPARE_32(v14,v16,0,tail1,tail2)
> + COMPARE_32(v18,v20,32,tail3,tail4)
> + COMPARE_32(v22,v24,64,tail5,tail6)
> + COMPARE_32(v26,v28,96,tail7,tail8)
> + addi r3,r3,128
> + addi r4,r4,128
> + b L(32B_aligned_loop)
> +
> +L(tail1): TAIL(v15,v17)
> +L(tail2): TAIL(v14,v16)
> +L(tail3): TAIL(v19,v21)
> +L(tail4): TAIL(v18,v20)
> +L(tail5): TAIL(v23,v25)
> +L(tail6): TAIL(v22,v24)
> +L(tail7): TAIL(v27,v29)
> +L(tail8): TAIL(v26,v28)
> +
> +END (STRCMP)
> +libc_hidden_builtin_def (strcmp)
Otherwise, this LGTM.
prev parent reply other threads:[~2023-11-30 16:54 UTC|newest]
Thread overview: 2+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-11-30 16:11 Amrita H S
2023-11-30 16:54 ` Paul E Murphy [this message]
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=54263bc9-c177-4825-8ef5-7c8d8ea22446@linux.ibm.com \
--to=murphyp@linux.ibm.com \
--cc=amritahs@linux.vnet.ibm.com \
--cc=libc-alpha@sourceware.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).