powerpc: Add optimized strlen for POWER10
Commit Message
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.
Output of bench-strlen from 'make USE_CLOCK_GETTIME=1 BENCHSET="string-benchset"
using slightly different set of inputs than the default:
$ ./compare_strings.py --functions __strlen_power9,__strlen_power10
-a length,alignment -s benchout_strings.schema.json
-i bench-strlen.out
Function: strlen
Variant:
__strlen_power10 __strlen_power9
================================================================================
length=1, alignment=0: 2.50 2.50 ( 0.00%)
length=1, alignment=1: 2.50 2.50 ( 0.00%)
length=2, alignment=0: 2.50 2.50 ( 0.00%)
length=2, alignment=2: 2.50 2.50 ( 0.00%)
length=3, alignment=0: 2.50 2.50 ( 0.00%)
length=3, alignment=3: 2.50 2.50 ( 0.00%)
length=4, alignment=0: 2.50 2.50 ( 0.00%)
length=4, alignment=4: 2.50 2.50 ( 0.00%)
length=5, alignment=0: 2.50 2.50 ( 0.00%)
length=5, alignment=5: 2.50 2.50 ( 0.00%)
length=6, alignment=0: 2.50 2.50 ( 0.00%)
length=6, alignment=6: 2.50 2.50 ( 0.00%)
length=7, alignment=0: 2.50 2.50 ( 0.00%)
length=7, alignment=7: 2.50 2.50 ( 0.00%)
length=8, alignment=0: 2.50 2.50 ( 0.00%)
length=8, alignment=8: 3.12 3.12 ( 0.00%)
length=9, alignment=0: 2.50 2.50 ( 0.00%)
length=9, alignment=9: 3.12 3.12 ( 0.00%)
length=10, alignment=10: 3.12 3.12 ( 0.00%)
length=16, alignment=0: 3.12 3.40 ( -9.09%)
length=16, alignment=4: 3.12 3.12 ( 0.00%)
length=16, alignment=7: 3.12 3.12 ( 0.00%)
length=21, alignment=0: 3.12 3.40 ( -9.09%)
length=21, alignment=5: 3.12 3.12 ( 0.00%)
length=32, alignment=0: 3.12 3.40 ( -9.09%)
length=32, alignment=7: 3.12 3.40 ( -9.09%)
length=42, alignment=0: 3.12 3.40 ( -9.09%)
length=42, alignment=7: 3.42 3.40 ( 0.51%)
length=48, alignment=0: 3.43 3.74 ( -9.13%)
length=48, alignment=7: 3.40 3.40 ( 0.17%)
length=64, alignment=0: 3.40 5.21 (-53.34%)
length=64, alignment=7: 3.40 3.74 (-10.00%)
length=80, alignment=0: 3.74 5.21 (-39.43%)
length=80, alignment=7: 3.74 4.01 ( -7.14%)
length=85, alignment=0: 3.74 5.21 (-39.42%)
length=85, alignment=7: 3.74 4.01 ( -7.14%)
length=96, alignment=0: 3.74 5.21 (-39.40%)
length=96, alignment=7: 3.74 4.01 ( -7.14%)
length=112, alignment=0: 3.74 5.21 (-39.39%)
length=112, alignment=7: 3.74 4.88 (-30.43%)
length=128, alignment=0: 4.01 5.91 (-47.59%)
length=128, alignment=7: 4.01 6.15 (-53.59%)
length=128, alignment=16: 4.01 6.16 (-53.78%)
length=128, alignment=23: 4.01 5.17 (-29.08%)
length=160, alignment=0: 4.01 5.92 (-47.75%)
length=160, alignment=7: 4.01 6.16 (-53.72%)
length=160, alignment=16: 4.01 6.14 (-53.29%)
length=160, alignment=23: 4.01 6.05 (-50.98%)
length=192, alignment=0: 5.93 6.84 (-15.44%)
length=192, alignment=7: 5.93 6.90 (-16.35%)
length=256, alignment=0: 6.61 7.73 (-17.02%)
length=256, alignment=7: 6.61 7.85 (-18.79%)
length=320, alignment=0: 7.26 8.65 (-19.12%)
length=320, alignment=7: 7.26 8.76 (-20.70%)
length=384, alignment=0: 7.95 9.62 (-20.98%)
length=384, alignment=7: 7.95 9.49 (-19.37%)
length=448, alignment=0: 8.73 10.39 (-19.06%)
length=448, alignment=7: 8.73 10.51 (-20.40%)
length=512, alignment=0: 9.44 11.13 (-17.87%)
length=512, alignment=7: 9.45 11.32 (-19.85%)
length=576, alignment=0: 10.10 11.93 (-18.05%)
length=576, alignment=7: 10.10 12.02 (-18.97%)
length=640, alignment=0: 10.71 12.73 (-18.86%)
length=640, alignment=7: 10.67 12.89 (-20.76%)
length=704, alignment=0: 11.59 13.39 (-15.61%)
length=704, alignment=7: 11.59 13.61 (-17.45%)
length=768, alignment=0: 12.27 14.22 (-15.90%)
length=768, alignment=7: 12.27 14.44 (-17.72%)
length=896, alignment=0: 13.48 15.70 (-16.47%)
length=896, alignment=7: 13.47 15.97 (-18.56%)
length=960, alignment=0: 14.22 16.63 (-16.92%)
length=960, alignment=7: 14.19 16.70 (-17.66%)
length=1024, alignment=0: 14.85 17.46 (-17.54%)
length=1024, alignment=7: 14.87 17.68 (-18.94%)
length=1280, alignment=0: 17.58 20.91 (-18.94%)
length=1280, alignment=7: 17.62 21.35 (-21.13%)
length=1536, alignment=0: 20.61 24.54 (-19.07%)
length=1536, alignment=7: 20.61 24.21 (-17.48%)
length=1792, alignment=0: 23.02 27.94 (-21.39%)
length=1792, alignment=7: 23.02 27.83 (-20.90%)
length=2048, alignment=0: 25.98 30.71 (-18.23%)
length=2048, alignment=7: 25.96 31.26 (-20.45%)
length=2560, alignment=0: 31.37 37.82 (-20.57%)
length=2560, alignment=7: 31.34 37.69 (-20.26%)
length=3008, alignment=0: 35.61 43.29 (-21.56%)
length=3008, alignment=7: 35.55 43.84 (-23.31%)
length=3520, alignment=0: 41.08 50.48 (-22.90%)
length=3520, alignment=7: 41.12 50.63 (-23.13%)
length=4096, alignment=0: 47.80 57.96 (-21.25%)
length=4096, alignment=7: 47.79 57.66 (-20.66%)
---
sysdeps/powerpc/powerpc64/le/power10/strlen.S | 219 ++++++++++++++++++
sysdeps/powerpc/powerpc64/multiarch/Makefile | 3 +-
.../powerpc64/multiarch/ifunc-impl-list.c | 2 +
.../powerpc64/multiarch/strlen-power10.S | 2 +
sysdeps/powerpc/powerpc64/multiarch/strlen.c | 3 +
5 files changed, 228 insertions(+), 1 deletion(-)
create mode 100644 sysdeps/powerpc/powerpc64/le/power10/strlen.S
create mode 100644 sysdeps/powerpc/powerpc64/multiarch/strlen-power10.S
Comments
On 4/16/21 4:24 PM, Matheus Castanho via Libc-alpha wrote:
> diff --git a/sysdeps/powerpc/powerpc64/le/power10/strlen.S b/sysdeps/powerpc/powerpc64/le/power10/strlen.S
> new file mode 100644
> index 0000000000..5218a2277c
> --- /dev/null
> +++ b/sysdeps/powerpc/powerpc64/le/power10/strlen.S
> @@ -0,0 +1,219 @@
> +/* 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
> +
> +#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)
Can you add a comment to these noting the minimum version of binutils
which supports these instructions?
> +
> +#define CHECK16(reg,offset,addr,label) \
> + lxv reg+32,offset(addr); \
> + vcmpequb. reg,reg,v18; \
> + bne cr6,L(label);
Maybe rename reg to vreg for clarity?
With the trivial comments above addressed, I think this patch is OK.
Paul E Murphy <murphyp@linux.ibm.com> writes:
> On 4/16/21 4:24 PM, Matheus Castanho via Libc-alpha wrote:
>
>> diff --git a/sysdeps/powerpc/powerpc64/le/power10/strlen.S b/sysdeps/powerpc/powerpc64/le/power10/strlen.S
>> new file mode 100644
>> index 0000000000..5218a2277c
>> --- /dev/null
>> +++ b/sysdeps/powerpc/powerpc64/le/power10/strlen.S
>> @@ -0,0 +1,219 @@
>> +/* 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
>> +
>> +#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)
>
> Can you add a comment to these noting the minimum version of binutils
> which supports these instructions?
>
>> +
>> +#define CHECK16(reg,offset,addr,label) \
>> + lxv reg+32,offset(addr); \
>> + vcmpequb. reg,reg,v18; \
>> + bne cr6,L(label);
>
> Maybe rename reg to vreg for clarity?
>
> With the trivial comments above addressed, I think this patch is OK.
Thanks for the review. Just posted v2 [1] with the suggested fixes.
[1] https://sourceware.org/pipermail/libc-alpha/2021-April/125307.html
--
Matheus Castanho
new file mode 100644
@@ -0,0 +1,219 @@
+/* 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
+
+#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(reg,offset,addr,label) \
+ lxv reg+32,offset(addr); \
+ vcmpequb. reg,reg,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 more 112B, 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
@@ -33,7 +33,8 @@ sysdep_routines += memcpy-power8-cached memcpy-power7 memcpy-a2 memcpy-power6 \
ifneq (,$(filter %le,$(config-machine)))
sysdep_routines += strcmp-power9 strncmp-power9 strcpy-power9 stpcpy-power9 \
- rawmemchr-power9 strlen-power9 strncpy-power9 stpncpy-power9
+ rawmemchr-power9 strlen-power9 strncpy-power9 stpncpy-power9 \
+ strlen-power10
endif
CFLAGS-strncase-power7.c += -mcpu=power7 -funroll-loops
CFLAGS-strncase_l-power7.c += -mcpu=power7 -funroll-loops
@@ -112,6 +112,8 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
/* Support sysdeps/powerpc/powerpc64/multiarch/strlen.c. */
IFUNC_IMPL (i, name, strlen,
#ifdef __LITTLE_ENDIAN__
+ IFUNC_IMPL_ADD (array, i, strlen, hwcap2 & PPC_FEATURE2_ARCH_3_1,
+ __strlen_power10)
IFUNC_IMPL_ADD (array, i, strlen, hwcap2 & PPC_FEATURE2_ARCH_3_00,
__strlen_power9)
#endif
new file mode 100644
@@ -0,0 +1,2 @@
+#define STRLEN __strlen_power10
+#include <sysdeps/powerpc/powerpc64/le/power10/strlen.S>
@@ -31,9 +31,12 @@ extern __typeof (__redirect_strlen) __strlen_ppc attribute_hidden;
extern __typeof (__redirect_strlen) __strlen_power7 attribute_hidden;
extern __typeof (__redirect_strlen) __strlen_power8 attribute_hidden;
extern __typeof (__redirect_strlen) __strlen_power9 attribute_hidden;
+extern __typeof (__redirect_strlen) __strlen_power10 attribute_hidden;
libc_ifunc (__libc_strlen,
# ifdef __LITTLE_ENDIAN__
+ (hwcap2 & PPC_FEATURE2_ARCH_3_1)
+ ? __strlen_power10 :
(hwcap2 & PPC_FEATURE2_ARCH_3_00)
? __strlen_power9 :
# endif