iv/linux
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
linux/arch/powerpc/lib/memcmp_64.S

634 lines
11 KiB
ArmAsm
Raw Normal View History

powerpc: Add 64bit optimised memcmp I noticed ksm spending quite a lot of time in memcmp on a large KVM box. The current memcmp loop is very unoptimised - byte at a time compares with no loop unrolling. We can do much much better. Optimise the loop in a few ways: - Unroll the byte at a time loop - For large (at least 32 byte) comparisons that are also 8 byte aligned, use an unrolled modulo scheduled loop using 8 byte loads. This is similar to our glibc memcmp. A simple microbenchmark testing 10000000 iterations of an 8192 byte memcmp was used to measure the performance: baseline: 29.93 s modified: 1.70 s Just over 17x faster. v2: Incorporated some suggestions from Segher: - Use andi. instead of rdlicl. - Convert bdnzt eq, to bdnz. It's just duplicating the earlier compare and was a relic from a previous version. - Don't use cr5, we have plans to use that CR field for fast local atomics. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2015-01-21 12:27:38 +11:00
/*
* Author: Anton Blanchard <anton@au.ibm.com>
* Copyright 2015 IBM Corporation.
*
* This program is free software; you can redistribute it and/or
* modify it under the terms of the GNU General Public License
* as published by the Free Software Foundation; either version
* 2 of the License, or (at your option) any later version.
*/
#include <asm/ppc_asm.h>
ppc: move exports to definitions Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-01-13 23:33:46 -05:00
#include <asm/export.h>
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
#include <asm/ppc-opcode.h>
powerpc: Add 64bit optimised memcmp I noticed ksm spending quite a lot of time in memcmp on a large KVM box. The current memcmp loop is very unoptimised - byte at a time compares with no loop unrolling. We can do much much better. Optimise the loop in a few ways: - Unroll the byte at a time loop - For large (at least 32 byte) comparisons that are also 8 byte aligned, use an unrolled modulo scheduled loop using 8 byte loads. This is similar to our glibc memcmp. A simple microbenchmark testing 10000000 iterations of an 8192 byte memcmp was used to measure the performance: baseline: 29.93 s modified: 1.70 s Just over 17x faster. v2: Incorporated some suggestions from Segher: - Use andi. instead of rdlicl. - Convert bdnzt eq, to bdnz. It's just duplicating the earlier compare and was a relic from a previous version. - Don't use cr5, we have plans to use that CR field for fast local atomics. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2015-01-21 12:27:38 +11:00
#define off8 r6
#define off16 r7
#define off24 r8
#define rA r9
#define rB r10
#define rC r11
#define rD r27
#define rE r28
#define rF r29
#define rG r30
#define rH r31
#ifdef __LITTLE_ENDIAN__
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
#define LH lhbrx
#define LW lwbrx
powerpc: Add 64bit optimised memcmp I noticed ksm spending quite a lot of time in memcmp on a large KVM box. The current memcmp loop is very unoptimised - byte at a time compares with no loop unrolling. We can do much much better. Optimise the loop in a few ways: - Unroll the byte at a time loop - For large (at least 32 byte) comparisons that are also 8 byte aligned, use an unrolled modulo scheduled loop using 8 byte loads. This is similar to our glibc memcmp. A simple microbenchmark testing 10000000 iterations of an 8192 byte memcmp was used to measure the performance: baseline: 29.93 s modified: 1.70 s Just over 17x faster. v2: Incorporated some suggestions from Segher: - Use andi. instead of rdlicl. - Convert bdnzt eq, to bdnz. It's just duplicating the earlier compare and was a relic from a previous version. - Don't use cr5, we have plans to use that CR field for fast local atomics. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2015-01-21 12:27:38 +11:00
#define LD ldbrx
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
#define LVS lvsr
#define VPERM(_VRT,_VRA,_VRB,_VRC) \
vperm _VRT,_VRB,_VRA,_VRC
powerpc: Add 64bit optimised memcmp I noticed ksm spending quite a lot of time in memcmp on a large KVM box. The current memcmp loop is very unoptimised - byte at a time compares with no loop unrolling. We can do much much better. Optimise the loop in a few ways: - Unroll the byte at a time loop - For large (at least 32 byte) comparisons that are also 8 byte aligned, use an unrolled modulo scheduled loop using 8 byte loads. This is similar to our glibc memcmp. A simple microbenchmark testing 10000000 iterations of an 8192 byte memcmp was used to measure the performance: baseline: 29.93 s modified: 1.70 s Just over 17x faster. v2: Incorporated some suggestions from Segher: - Use andi. instead of rdlicl. - Convert bdnzt eq, to bdnz. It's just duplicating the earlier compare and was a relic from a previous version. - Don't use cr5, we have plans to use that CR field for fast local atomics. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2015-01-21 12:27:38 +11:00
#else
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
#define LH lhzx
#define LW lwzx
powerpc: Add 64bit optimised memcmp I noticed ksm spending quite a lot of time in memcmp on a large KVM box. The current memcmp loop is very unoptimised - byte at a time compares with no loop unrolling. We can do much much better. Optimise the loop in a few ways: - Unroll the byte at a time loop - For large (at least 32 byte) comparisons that are also 8 byte aligned, use an unrolled modulo scheduled loop using 8 byte loads. This is similar to our glibc memcmp. A simple microbenchmark testing 10000000 iterations of an 8192 byte memcmp was used to measure the performance: baseline: 29.93 s modified: 1.70 s Just over 17x faster. v2: Incorporated some suggestions from Segher: - Use andi. instead of rdlicl. - Convert bdnzt eq, to bdnz. It's just duplicating the earlier compare and was a relic from a previous version. - Don't use cr5, we have plans to use that CR field for fast local atomics. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2015-01-21 12:27:38 +11:00
#define LD ldx
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
#define LVS lvsl
#define VPERM(_VRT,_VRA,_VRB,_VRC) \
vperm _VRT,_VRA,_VRB,_VRC
powerpc: Add 64bit optimised memcmp I noticed ksm spending quite a lot of time in memcmp on a large KVM box. The current memcmp loop is very unoptimised - byte at a time compares with no loop unrolling. We can do much much better. Optimise the loop in a few ways: - Unroll the byte at a time loop - For large (at least 32 byte) comparisons that are also 8 byte aligned, use an unrolled modulo scheduled loop using 8 byte loads. This is similar to our glibc memcmp. A simple microbenchmark testing 10000000 iterations of an 8192 byte memcmp was used to measure the performance: baseline: 29.93 s modified: 1.70 s Just over 17x faster. v2: Incorporated some suggestions from Segher: - Use andi. instead of rdlicl. - Convert bdnzt eq, to bdnz. It's just duplicating the earlier compare and was a relic from a previous version. - Don't use cr5, we have plans to use that CR field for fast local atomics. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2015-01-21 12:27:38 +11:00
#endif
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
#define VMX_THRESH 4096
#define ENTER_VMX_OPS \
mflr r0; \
std r3,-STACKFRAMESIZE+STK_REG(R31)(r1); \
std r4,-STACKFRAMESIZE+STK_REG(R30)(r1); \
std r5,-STACKFRAMESIZE+STK_REG(R29)(r1); \
std r0,16(r1); \
stdu r1,-STACKFRAMESIZE(r1); \
bl enter_vmx_ops; \
cmpwi cr1,r3,0; \
ld r0,STACKFRAMESIZE+16(r1); \
ld r3,STK_REG(R31)(r1); \
ld r4,STK_REG(R30)(r1); \
ld r5,STK_REG(R29)(r1); \
addi r1,r1,STACKFRAMESIZE; \
mtlr r0
#define EXIT_VMX_OPS \
mflr r0; \
std r3,-STACKFRAMESIZE+STK_REG(R31)(r1); \
std r4,-STACKFRAMESIZE+STK_REG(R30)(r1); \
std r5,-STACKFRAMESIZE+STK_REG(R29)(r1); \
std r0,16(r1); \
stdu r1,-STACKFRAMESIZE(r1); \
bl exit_vmx_ops; \
ld r0,STACKFRAMESIZE+16(r1); \
ld r3,STK_REG(R31)(r1); \
ld r4,STK_REG(R30)(r1); \
ld r5,STK_REG(R29)(r1); \
addi r1,r1,STACKFRAMESIZE; \
mtlr r0
/*
* LD_VSR_CROSS16B load the 2nd 16 bytes for _vaddr which is unaligned with
* 16 bytes boundary and permute the result with the 1st 16 bytes.
* | y y y y y y y y y y y y y 0 1 2 | 3 4 5 6 7 8 9 a b c d e f z z z |
* ^ ^ ^
* 0xbbbb10 0xbbbb20 0xbbb30
* ^
* _vaddr
*
*
* _vmask is the mask generated by LVS
* _v1st_qw is the 1st aligned QW of current addr which is already loaded.
* for example: 0xyyyyyyyyyyyyy012 for big endian
* _v2nd_qw is the 2nd aligned QW of cur _vaddr to be loaded.
* for example: 0x3456789abcdefzzz for big endian
* The permute result is saved in _v_res.
* for example: 0x0123456789abcdef for big endian.
*/
#define LD_VSR_CROSS16B(_vaddr,_vmask,_v1st_qw,_v2nd_qw,_v_res) \
lvx _v2nd_qw,_vaddr,off16; \
VPERM(_v_res,_v1st_qw,_v2nd_qw,_vmask)
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
/*
* There are 2 categories for memcmp:
* 1) src/dst has the same offset to the 8 bytes boundary. The handlers
* are named like .Lsameoffset_xxxx
* 2) src/dst has different offset to the 8 bytes boundary. The handlers
* are named like .Ldiffoffset_xxxx
*/
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
_GLOBAL_TOC(memcmp)
powerpc: Add 64bit optimised memcmp I noticed ksm spending quite a lot of time in memcmp on a large KVM box. The current memcmp loop is very unoptimised - byte at a time compares with no loop unrolling. We can do much much better. Optimise the loop in a few ways: - Unroll the byte at a time loop - For large (at least 32 byte) comparisons that are also 8 byte aligned, use an unrolled modulo scheduled loop using 8 byte loads. This is similar to our glibc memcmp. A simple microbenchmark testing 10000000 iterations of an 8192 byte memcmp was used to measure the performance: baseline: 29.93 s modified: 1.70 s Just over 17x faster. v2: Incorporated some suggestions from Segher: - Use andi. instead of rdlicl. - Convert bdnzt eq, to bdnz. It's just duplicating the earlier compare and was a relic from a previous version. - Don't use cr5, we have plans to use that CR field for fast local atomics. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2015-01-21 12:27:38 +11:00
cmpdi cr1,r5,0
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
/* Use the short loop if the src/dst addresses are not
* with the same offset of 8 bytes align boundary.
*/
xor r6,r3,r4
powerpc: Add 64bit optimised memcmp I noticed ksm spending quite a lot of time in memcmp on a large KVM box. The current memcmp loop is very unoptimised - byte at a time compares with no loop unrolling. We can do much much better. Optimise the loop in a few ways: - Unroll the byte at a time loop - For large (at least 32 byte) comparisons that are also 8 byte aligned, use an unrolled modulo scheduled loop using 8 byte loads. This is similar to our glibc memcmp. A simple microbenchmark testing 10000000 iterations of an 8192 byte memcmp was used to measure the performance: baseline: 29.93 s modified: 1.70 s Just over 17x faster. v2: Incorporated some suggestions from Segher: - Use andi. instead of rdlicl. - Convert bdnzt eq, to bdnz. It's just duplicating the earlier compare and was a relic from a previous version. - Don't use cr5, we have plans to use that CR field for fast local atomics. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2015-01-21 12:27:38 +11:00
andi. r6,r6,7
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
/* Fall back to short loop if compare at aligned addrs
* with less than 8 bytes.
*/
cmpdi cr6,r5,7
powerpc: Add 64bit optimised memcmp I noticed ksm spending quite a lot of time in memcmp on a large KVM box. The current memcmp loop is very unoptimised - byte at a time compares with no loop unrolling. We can do much much better. Optimise the loop in a few ways: - Unroll the byte at a time loop - For large (at least 32 byte) comparisons that are also 8 byte aligned, use an unrolled modulo scheduled loop using 8 byte loads. This is similar to our glibc memcmp. A simple microbenchmark testing 10000000 iterations of an 8192 byte memcmp was used to measure the performance: baseline: 29.93 s modified: 1.70 s Just over 17x faster. v2: Incorporated some suggestions from Segher: - Use andi. instead of rdlicl. - Convert bdnzt eq, to bdnz. It's just duplicating the earlier compare and was a relic from a previous version. - Don't use cr5, we have plans to use that CR field for fast local atomics. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2015-01-21 12:27:38 +11:00
beq cr1,.Lzero
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
bgt cr6,.Lno_short
powerpc: Add 64bit optimised memcmp I noticed ksm spending quite a lot of time in memcmp on a large KVM box. The current memcmp loop is very unoptimised - byte at a time compares with no loop unrolling. We can do much much better. Optimise the loop in a few ways: - Unroll the byte at a time loop - For large (at least 32 byte) comparisons that are also 8 byte aligned, use an unrolled modulo scheduled loop using 8 byte loads. This is similar to our glibc memcmp. A simple microbenchmark testing 10000000 iterations of an 8192 byte memcmp was used to measure the performance: baseline: 29.93 s modified: 1.70 s Just over 17x faster. v2: Incorporated some suggestions from Segher: - Use andi. instead of rdlicl. - Convert bdnzt eq, to bdnz. It's just duplicating the earlier compare and was a relic from a previous version. - Don't use cr5, we have plans to use that CR field for fast local atomics. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2015-01-21 12:27:38 +11:00
.Lshort:
mtctr r5
1: lbz rA,0(r3)
lbz rB,0(r4)
subf. rC,rB,rA
bne .Lnon_zero
bdz .Lzero
lbz rA,1(r3)
lbz rB,1(r4)
subf. rC,rB,rA
bne .Lnon_zero
bdz .Lzero
lbz rA,2(r3)
lbz rB,2(r4)
subf. rC,rB,rA
bne .Lnon_zero
bdz .Lzero
lbz rA,3(r3)
lbz rB,3(r4)
subf. rC,rB,rA
bne .Lnon_zero
addi r3,r3,4
addi r4,r4,4
bdnz 1b
.Lzero:
li r3,0
blr
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
.Lno_short:
dcbt 0,r3
dcbt 0,r4
bne .Ldiffoffset_8bytes_make_align_start
.Lsameoffset_8bytes_make_align_start:
/* attempt to compare bytes not aligned with 8 bytes so that
* rest comparison can run based on 8 bytes alignment.
*/
andi. r6,r3,7
/* Try to compare the first double word which is not 8 bytes aligned:
* load the first double word at (src & ~7UL) and shift left appropriate
* bits before comparision.
*/
rlwinm r6,r3,3,26,28
beq .Lsameoffset_8bytes_aligned
clrrdi r3,r3,3
clrrdi r4,r4,3
LD rA,0,r3
LD rB,0,r4
sld rA,rA,r6
sld rB,rB,r6
cmpld cr0,rA,rB
srwi r6,r6,3
bne cr0,.LcmpAB_lightweight
subfic r6,r6,8
subf. r5,r6,r5
addi r3,r3,8
addi r4,r4,8
beq .Lzero
.Lsameoffset_8bytes_aligned:
/* now we are aligned with 8 bytes.
* Use .Llong loop if left cmp bytes are equal or greater than 32B.
*/
cmpdi cr6,r5,31
bgt cr6,.Llong
.Lcmp_lt32bytes:
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
/* compare 1 ~ 31 bytes, at least r3 addr is 8 bytes aligned now */
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
cmpdi cr5,r5,7
srdi r0,r5,3
ble cr5,.Lcmp_rest_lt8bytes
/* handle 8 ~ 31 bytes */
clrldi r5,r5,61
mtctr r0
2:
LD rA,0,r3
LD rB,0,r4
cmpld cr0,rA,rB
addi r3,r3,8
addi r4,r4,8
bne cr0,.LcmpAB_lightweight
bdnz 2b
cmpwi r5,0
beq .Lzero
.Lcmp_rest_lt8bytes:
/* Here we have only less than 8 bytes to compare with. at least s1
* Address is aligned with 8 bytes.
* The next double words are load and shift right with appropriate
* bits.
*/
subfic r6,r5,8
slwi r6,r6,3
LD rA,0,r3
LD rB,0,r4
srd rA,rA,r6
srd rB,rB,r6
cmpld cr0,rA,rB
bne cr0,.LcmpAB_lightweight
b .Lzero
powerpc: Add 64bit optimised memcmp I noticed ksm spending quite a lot of time in memcmp on a large KVM box. The current memcmp loop is very unoptimised - byte at a time compares with no loop unrolling. We can do much much better. Optimise the loop in a few ways: - Unroll the byte at a time loop - For large (at least 32 byte) comparisons that are also 8 byte aligned, use an unrolled modulo scheduled loop using 8 byte loads. This is similar to our glibc memcmp. A simple microbenchmark testing 10000000 iterations of an 8192 byte memcmp was used to measure the performance: baseline: 29.93 s modified: 1.70 s Just over 17x faster. v2: Incorporated some suggestions from Segher: - Use andi. instead of rdlicl. - Convert bdnzt eq, to bdnz. It's just duplicating the earlier compare and was a relic from a previous version. - Don't use cr5, we have plans to use that CR field for fast local atomics. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2015-01-21 12:27:38 +11:00
.Lnon_zero:
mr r3,rC
blr
.Llong:
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
#ifdef CONFIG_ALTIVEC
BEGIN_FTR_SECTION
/* Try to use vmx loop if length is equal or greater than 4K */
cmpldi cr6,r5,VMX_THRESH
bge cr6,.Lsameoffset_vmx_cmp
END_FTR_SECTION_IFSET(CPU_FTR_ARCH_207S)
.Llong_novmx_cmp:
#endif
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
/* At least s1 addr is aligned with 8 bytes */
powerpc: Add 64bit optimised memcmp I noticed ksm spending quite a lot of time in memcmp on a large KVM box. The current memcmp loop is very unoptimised - byte at a time compares with no loop unrolling. We can do much much better. Optimise the loop in a few ways: - Unroll the byte at a time loop - For large (at least 32 byte) comparisons that are also 8 byte aligned, use an unrolled modulo scheduled loop using 8 byte loads. This is similar to our glibc memcmp. A simple microbenchmark testing 10000000 iterations of an 8192 byte memcmp was used to measure the performance: baseline: 29.93 s modified: 1.70 s Just over 17x faster. v2: Incorporated some suggestions from Segher: - Use andi. instead of rdlicl. - Convert bdnzt eq, to bdnz. It's just duplicating the earlier compare and was a relic from a previous version. - Don't use cr5, we have plans to use that CR field for fast local atomics. Signed-off-by: Anton Blanchard <anton@samba.org> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2015-01-21 12:27:38 +11:00
li off8,8
li off16,16
li off24,24
std r31,-8(r1)
std r30,-16(r1)
std r29,-24(r1)
std r28,-32(r1)
std r27,-40(r1)
srdi r0,r5,5
mtctr r0
andi. r5,r5,31
LD rA,0,r3
LD rB,0,r4
LD rC,off8,r3
LD rD,off8,r4
LD rE,off16,r3
LD rF,off16,r4
LD rG,off24,r3
LD rH,off24,r4
cmpld cr0,rA,rB
addi r3,r3,32
addi r4,r4,32
bdz .Lfirst32
LD rA,0,r3
LD rB,0,r4
cmpld cr1,rC,rD
LD rC,off8,r3
LD rD,off8,r4
cmpld cr6,rE,rF
LD rE,off16,r3
LD rF,off16,r4
cmpld cr7,rG,rH
bne cr0,.LcmpAB
LD rG,off24,r3
LD rH,off24,r4
cmpld cr0,rA,rB
bne cr1,.LcmpCD
addi r3,r3,32
addi r4,r4,32
bdz .Lsecond32
.balign 16
1: LD rA,0,r3
LD rB,0,r4
cmpld cr1,rC,rD
bne cr6,.LcmpEF
LD rC,off8,r3
LD rD,off8,r4
cmpld cr6,rE,rF
bne cr7,.LcmpGH
LD rE,off16,r3
LD rF,off16,r4
cmpld cr7,rG,rH
bne cr0,.LcmpAB
LD rG,off24,r3
LD rH,off24,r4
cmpld cr0,rA,rB
bne cr1,.LcmpCD
addi r3,r3,32
addi r4,r4,32
bdnz 1b
.Lsecond32:
cmpld cr1,rC,rD
bne cr6,.LcmpEF
cmpld cr6,rE,rF
bne cr7,.LcmpGH
cmpld cr7,rG,rH
bne cr0,.LcmpAB
bne cr1,.LcmpCD
bne cr6,.LcmpEF
bne cr7,.LcmpGH
.Ltail:
ld r31,-8(r1)
ld r30,-16(r1)
ld r29,-24(r1)
ld r28,-32(r1)
ld r27,-40(r1)
cmpdi r5,0
beq .Lzero
b .Lshort
.Lfirst32:
cmpld cr1,rC,rD
cmpld cr6,rE,rF
cmpld cr7,rG,rH
bne cr0,.LcmpAB
bne cr1,.LcmpCD
bne cr6,.LcmpEF
bne cr7,.LcmpGH
b .Ltail
.LcmpAB:
li r3,1
bgt cr0,.Lout
li r3,-1
b .Lout
.LcmpCD:
li r3,1
bgt cr1,.Lout
li r3,-1
b .Lout
.LcmpEF:
li r3,1
bgt cr6,.Lout
li r3,-1
b .Lout
.LcmpGH:
li r3,1
bgt cr7,.Lout
li r3,-1
.Lout:
ld r31,-8(r1)
ld r30,-16(r1)
ld r29,-24(r1)
ld r28,-32(r1)
ld r27,-40(r1)
blr
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
.LcmpAB_lightweight: /* skip NV GPRS restore */
li r3,1
bgtlr
li r3,-1
blr
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
#ifdef CONFIG_ALTIVEC
.Lsameoffset_vmx_cmp:
/* Enter with src/dst addrs has the same offset with 8 bytes
powerpc/64: add 32 bytes prechecking before using VMX optimization on memcmp() This patch is based on the previous VMX patch on memcmp(). To optimize ppc64 memcmp() with VMX instruction, we need to think about the VMX penalty brought with: If kernel uses VMX instruction, it needs to save/restore current thread's VMX registers. There are 32 x 128 bits VMX registers in PPC, which means 32 x 16 = 512 bytes for load and store. The major concern regarding the memcmp() performance in kernel is KSM, who will use memcmp() frequently to merge identical pages. So it will make sense to take some measures/enhancement on KSM to see whether any improvement can be done here. Cyril Bur indicates that the memcmp() for KSM has a higher possibility to fail (unmatch) early in previous bytes in following mail. https://patchwork.ozlabs.org/patch/817322/#1773629 And I am taking a follow-up on this with this patch. Per some testing, it shows KSM memcmp() will fail early at previous 32 bytes. More specifically: - 76% cases will fail/unmatch before 16 bytes; - 83% cases will fail/unmatch before 32 bytes; - 84% cases will fail/unmatch before 64 bytes; So 32 bytes looks a better choice than other bytes for pre-checking. The early failure is also true for memcmp() for non-KSM case. With a non-typical call load, it shows ~73% cases fail before first 32 bytes. This patch adds a 32 bytes pre-checking firstly before jumping into VMX operations, to avoid the unnecessary VMX penalty. It is not limited to KSM case. And the testing shows ~20% improvement on memcmp() average execution time with this patch. And note the 32B pre-checking is only performed when the compare size is long enough (>=4K currently) to allow VMX operation. The detail data and analysis is at: https://github.com/justdoitqd/publicFiles/blob/master/memcmp/README.md Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:54 +08:00
* align boundary.
*
* There is an optimization based on following fact: memcmp()
* prones to fail early at the first 32 bytes.
* Before applying VMX instructions which will lead to 32x128bits
* VMX regs load/restore penalty, we compare the first 32 bytes
* so that we can catch the ~80% fail cases.
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
*/
powerpc/64: add 32 bytes prechecking before using VMX optimization on memcmp() This patch is based on the previous VMX patch on memcmp(). To optimize ppc64 memcmp() with VMX instruction, we need to think about the VMX penalty brought with: If kernel uses VMX instruction, it needs to save/restore current thread's VMX registers. There are 32 x 128 bits VMX registers in PPC, which means 32 x 16 = 512 bytes for load and store. The major concern regarding the memcmp() performance in kernel is KSM, who will use memcmp() frequently to merge identical pages. So it will make sense to take some measures/enhancement on KSM to see whether any improvement can be done here. Cyril Bur indicates that the memcmp() for KSM has a higher possibility to fail (unmatch) early in previous bytes in following mail. https://patchwork.ozlabs.org/patch/817322/#1773629 And I am taking a follow-up on this with this patch. Per some testing, it shows KSM memcmp() will fail early at previous 32 bytes. More specifically: - 76% cases will fail/unmatch before 16 bytes; - 83% cases will fail/unmatch before 32 bytes; - 84% cases will fail/unmatch before 64 bytes; So 32 bytes looks a better choice than other bytes for pre-checking. The early failure is also true for memcmp() for non-KSM case. With a non-typical call load, it shows ~73% cases fail before first 32 bytes. This patch adds a 32 bytes pre-checking firstly before jumping into VMX operations, to avoid the unnecessary VMX penalty. It is not limited to KSM case. And the testing shows ~20% improvement on memcmp() average execution time with this patch. And note the 32B pre-checking is only performed when the compare size is long enough (>=4K currently) to allow VMX operation. The detail data and analysis is at: https://github.com/justdoitqd/publicFiles/blob/master/memcmp/README.md Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:54 +08:00
li r0,4
mtctr r0
.Lsameoffset_prechk_32B_loop:
LD rA,0,r3
LD rB,0,r4
cmpld cr0,rA,rB
addi r3,r3,8
addi r4,r4,8
bne cr0,.LcmpAB_lightweight
addi r5,r5,-8
bdnz .Lsameoffset_prechk_32B_loop
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
ENTER_VMX_OPS
beq cr1,.Llong_novmx_cmp
3:
/* need to check whether r4 has the same offset with r3
* for 16 bytes boundary.
*/
xor r0,r3,r4
andi. r0,r0,0xf
bne .Ldiffoffset_vmx_cmp_start
/* len is no less than 4KB. Need to align with 16 bytes further.
*/
andi. rA,r3,8
LD rA,0,r3
beq 4f
LD rB,0,r4
cmpld cr0,rA,rB
addi r3,r3,8
addi r4,r4,8
addi r5,r5,-8
beq cr0,4f
/* save and restore cr0 */
mfocrf r5,128
EXIT_VMX_OPS
mtocrf 128,r5
b .LcmpAB_lightweight
4:
/* compare 32 bytes for each loop */
srdi r0,r5,5
mtctr r0
clrldi r5,r5,59
li off16,16
.balign 16
5:
lvx v0,0,r3
lvx v1,0,r4
VCMPEQUD_RC(v0,v0,v1)
bnl cr6,7f
lvx v0,off16,r3
lvx v1,off16,r4
VCMPEQUD_RC(v0,v0,v1)
bnl cr6,6f
addi r3,r3,32
addi r4,r4,32
bdnz 5b
EXIT_VMX_OPS
cmpdi r5,0
beq .Lzero
b .Lcmp_lt32bytes
6:
addi r3,r3,16
addi r4,r4,16
7:
/* diff the last 16 bytes */
EXIT_VMX_OPS
LD rA,0,r3
LD rB,0,r4
cmpld cr0,rA,rB
li off8,8
bne cr0,.LcmpAB_lightweight
LD rA,off8,r3
LD rB,off8,r4
cmpld cr0,rA,rB
bne cr0,.LcmpAB_lightweight
b .Lzero
#endif
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
.Ldiffoffset_8bytes_make_align_start:
/* now try to align s1 with 8 bytes */
rlwinm r6,r3,3,26,28
beq .Ldiffoffset_align_s1_8bytes
clrrdi r3,r3,3
LD rA,0,r3
LD rB,0,r4 /* unaligned load */
sld rA,rA,r6
srd rA,rA,r6
srd rB,rB,r6
cmpld cr0,rA,rB
srwi r6,r6,3
bne cr0,.LcmpAB_lightweight
subfic r6,r6,8
subf. r5,r6,r5
addi r3,r3,8
add r4,r4,r6
beq .Lzero
.Ldiffoffset_align_s1_8bytes:
/* now s1 is aligned with 8 bytes. */
powerpc/64: add 32 bytes prechecking before using VMX optimization on memcmp() This patch is based on the previous VMX patch on memcmp(). To optimize ppc64 memcmp() with VMX instruction, we need to think about the VMX penalty brought with: If kernel uses VMX instruction, it needs to save/restore current thread's VMX registers. There are 32 x 128 bits VMX registers in PPC, which means 32 x 16 = 512 bytes for load and store. The major concern regarding the memcmp() performance in kernel is KSM, who will use memcmp() frequently to merge identical pages. So it will make sense to take some measures/enhancement on KSM to see whether any improvement can be done here. Cyril Bur indicates that the memcmp() for KSM has a higher possibility to fail (unmatch) early in previous bytes in following mail. https://patchwork.ozlabs.org/patch/817322/#1773629 And I am taking a follow-up on this with this patch. Per some testing, it shows KSM memcmp() will fail early at previous 32 bytes. More specifically: - 76% cases will fail/unmatch before 16 bytes; - 83% cases will fail/unmatch before 32 bytes; - 84% cases will fail/unmatch before 64 bytes; So 32 bytes looks a better choice than other bytes for pre-checking. The early failure is also true for memcmp() for non-KSM case. With a non-typical call load, it shows ~73% cases fail before first 32 bytes. This patch adds a 32 bytes pre-checking firstly before jumping into VMX operations, to avoid the unnecessary VMX penalty. It is not limited to KSM case. And the testing shows ~20% improvement on memcmp() average execution time with this patch. And note the 32B pre-checking is only performed when the compare size is long enough (>=4K currently) to allow VMX operation. The detail data and analysis is at: https://github.com/justdoitqd/publicFiles/blob/master/memcmp/README.md Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:54 +08:00
#ifdef CONFIG_ALTIVEC
BEGIN_FTR_SECTION
/* only do vmx ops when the size equal or greater than 4K bytes */
cmpdi cr5,r5,VMX_THRESH
bge cr5,.Ldiffoffset_vmx_cmp
END_FTR_SECTION_IFSET(CPU_FTR_ARCH_207S)
.Ldiffoffset_novmx_cmp:
#endif
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
cmpdi cr5,r5,31
ble cr5,.Lcmp_lt32bytes
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
#ifdef CONFIG_ALTIVEC
b .Llong_novmx_cmp
#else
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
b .Llong
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
#endif
#ifdef CONFIG_ALTIVEC
.Ldiffoffset_vmx_cmp:
powerpc/64: add 32 bytes prechecking before using VMX optimization on memcmp() This patch is based on the previous VMX patch on memcmp(). To optimize ppc64 memcmp() with VMX instruction, we need to think about the VMX penalty brought with: If kernel uses VMX instruction, it needs to save/restore current thread's VMX registers. There are 32 x 128 bits VMX registers in PPC, which means 32 x 16 = 512 bytes for load and store. The major concern regarding the memcmp() performance in kernel is KSM, who will use memcmp() frequently to merge identical pages. So it will make sense to take some measures/enhancement on KSM to see whether any improvement can be done here. Cyril Bur indicates that the memcmp() for KSM has a higher possibility to fail (unmatch) early in previous bytes in following mail. https://patchwork.ozlabs.org/patch/817322/#1773629 And I am taking a follow-up on this with this patch. Per some testing, it shows KSM memcmp() will fail early at previous 32 bytes. More specifically: - 76% cases will fail/unmatch before 16 bytes; - 83% cases will fail/unmatch before 32 bytes; - 84% cases will fail/unmatch before 64 bytes; So 32 bytes looks a better choice than other bytes for pre-checking. The early failure is also true for memcmp() for non-KSM case. With a non-typical call load, it shows ~73% cases fail before first 32 bytes. This patch adds a 32 bytes pre-checking firstly before jumping into VMX operations, to avoid the unnecessary VMX penalty. It is not limited to KSM case. And the testing shows ~20% improvement on memcmp() average execution time with this patch. And note the 32B pre-checking is only performed when the compare size is long enough (>=4K currently) to allow VMX operation. The detail data and analysis is at: https://github.com/justdoitqd/publicFiles/blob/master/memcmp/README.md Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:54 +08:00
/* perform a 32 bytes pre-checking before
* enable VMX operations.
*/
li r0,4
mtctr r0
.Ldiffoffset_prechk_32B_loop:
LD rA,0,r3
LD rB,0,r4
cmpld cr0,rA,rB
addi r3,r3,8
addi r4,r4,8
bne cr0,.LcmpAB_lightweight
addi r5,r5,-8
bdnz .Ldiffoffset_prechk_32B_loop
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
ENTER_VMX_OPS
beq cr1,.Ldiffoffset_novmx_cmp
.Ldiffoffset_vmx_cmp_start:
/* Firstly try to align r3 with 16 bytes */
andi. r6,r3,0xf
li off16,16
beq .Ldiffoffset_vmx_s1_16bytes_align
powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp() Currently memcmp() 64bytes version in powerpc will fall back to .Lshort (compare per byte mode) if either src or dst address is not 8 bytes aligned. It can be opmitized in 2 situations: 1) if both addresses are with the same offset with 8 bytes boundary: memcmp() can compare the unaligned bytes within 8 bytes boundary firstly and then compare the rest 8-bytes-aligned content with .Llong mode. 2) If src/dst addrs are not with the same offset of 8 bytes boundary: memcmp() can align src addr with 8 bytes, increment dst addr accordingly, then load src with aligned mode and load dst with unaligned mode. This patch optmizes memcmp() behavior in the above 2 situations. Tested with both little/big endian. Performance result below is based on little endian. Following is the test result with src/dst having the same offset case: (a similar result was observed when src/dst having different offset): (1) 256 bytes Test with the existing tools/testing/selftests/powerpc/stringloops/memcmp: - without patch 29.773018302 seconds time elapsed ( +- 0.09% ) - with patch 16.485568173 seconds time elapsed ( +- 0.02% ) -> There is ~+80% percent improvement (2) 32 bytes To observe performance impact on < 32 bytes, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 +#define SIZE 32 #define ITERATIONS 10000 int test_memcmp(const void *s1, const void *s2, size_t n); -------- - Without patch 0.244746482 seconds time elapsed ( +- 0.36%) - with patch 0.215069477 seconds time elapsed ( +- 0.51%) -> There is ～+13% improvement (3) 0~8 bytes To observe <8 bytes performance impact, modify tools/testing/selftests/powerpc/stringloops/memcmp.c with following: ------- #include <string.h> #include "utils.h" -#define SIZE 256 -#define ITERATIONS 10000 +#define SIZE 8 +#define ITERATIONS 1000000 int test_memcmp(const void *s1, const void *s2, size_t n); ------- - Without patch 1.845642503 seconds time elapsed ( +- 0.12% ) - With patch 1.849767135 seconds time elapsed ( +- 0.26% ) -> They are nearly the same. (-0.2%) Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:51 +08:00
powerpc/64: enhance memcmp() with VMX instruction for long bytes comparision This patch add VMX primitives to do memcmp() in case the compare size is equal or greater than 4K bytes. KSM feature can benefit from this. Test result with following test program(replace the "^>" with ""): ------ ># cat tools/testing/selftests/powerpc/stringloops/memcmp.c >#include <malloc.h> >#include <stdlib.h> >#include <string.h> >#include <time.h> >#include "utils.h" >#define SIZE (1024 * 1024 * 900) >#define ITERATIONS 40 int test_memcmp(const void *s1, const void *s2, size_t n); static int testcase(void) { char *s1; char *s2; unsigned long i; s1 = memalign(128, SIZE); if (!s1) { perror("memalign"); exit(1); } s2 = memalign(128, SIZE); if (!s2) { perror("memalign"); exit(1); } for (i = 0; i < SIZE; i++) { s1[i] = i & 0xff; s2[i] = i & 0xff; } for (i = 0; i < ITERATIONS; i++) { int ret = test_memcmp(s1, s2, SIZE); if (ret) { printf("return %d at[%ld]! should have returned zero\n", ret, i); abort(); } } return 0; } int main(void) { return test_harness(testcase, "memcmp"); } ------ Without this patch (but with the first patch "powerpc/64: Align bytes before fall back to .Lshort in powerpc64 memcmp()." in the series): 4.726728762 seconds time elapsed ( +- 3.54%) With VMX patch: 4.234335473 seconds time elapsed ( +- 2.63%) There is ~+10% improvement. Testing with unaligned and different offset version (make s1 and s2 shift random offset within 16 bytes) can archieve higher improvement than 10%.. Signed-off-by: Simon Guo <wei.guo.simon@gmail.com> Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
2018-06-07 09:57:53 +08:00
LVS v3,0,r3
LVS v4,0,r4
lvx v5,0,r3
lvx v6,0,r4
LD_VSR_CROSS16B(r3,v3,v5,v7,v9)
LD_VSR_CROSS16B(r4,v4,v6,v8,v10)
VCMPEQUB_RC(v7,v9,v10)
bnl cr6,.Ldiffoffset_vmx_diff_found
subfic r6,r6,16
subf r5,r6,r5
add r3,r3,r6
add r4,r4,r6
.Ldiffoffset_vmx_s1_16bytes_align:
/* now s1 is aligned with 16 bytes */
lvx v6,0,r4
LVS v4,0,r4
srdi r6,r5,5 /* loop for 32 bytes each */
clrldi r5,r5,59
mtctr r6
.balign 16
.Ldiffoffset_vmx_32bytesloop:
/* the first qw of r4 was saved in v6 */
lvx v9,0,r3
LD_VSR_CROSS16B(r4,v4,v6,v8,v10)
VCMPEQUB_RC(v7,v9,v10)
vor v6,v8,v8
bnl cr6,.Ldiffoffset_vmx_diff_found
addi r3,r3,16
addi r4,r4,16
lvx v9,0,r3
LD_VSR_CROSS16B(r4,v4,v6,v8,v10)
VCMPEQUB_RC(v7,v9,v10)
vor v6,v8,v8
bnl cr6,.Ldiffoffset_vmx_diff_found
addi r3,r3,16
addi r4,r4,16
bdnz .Ldiffoffset_vmx_32bytesloop
EXIT_VMX_OPS
cmpdi r5,0
beq .Lzero
b .Lcmp_lt32bytes
.Ldiffoffset_vmx_diff_found:
EXIT_VMX_OPS
/* anyway, the diff will appear in next 16 bytes */
li r5,16
b .Lcmp_lt32bytes
#endif
ppc: move exports to definitions Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2016-01-13 23:33:46 -05:00
EXPORT_SYMBOL(memcmp)
Reference in New Issue Copy Permalink