* [PATCH 0/3] Add ifunc support for str{nlen, cmp, ncmp}
@ 2023-08-22 2:11 dengjianbo
2023-08-22 2:11 ` [PATCH 1/3] Loongarch: Add ifunc support for strnlen{aligned, lsx, lasx} dengjianbo
` (2 more replies)
0 siblings, 3 replies; 9+ messages in thread
From: dengjianbo @ 2023-08-22 2:11 UTC (permalink / raw)
To: libc-alpha
Cc: adhemerval.zanella, xry111, caiyinyu, xuchenghua, huangpei, dengjianbo
This patch add mutiple versions of strnlen, strcmp, strncmp implemented
by Loongarch basic instructions, LSX instructions, LASX instructions.
Even though this implementation experience performance degradation in
few cases, overall, the performace gains are significant.
See:
https://github.com/jiadengx/glibc_test/blob/main/bench/strnlen_compare.out
https://github.com/jiadengx/glibc_test/blob/main/bench/strcmp_compare.out
https://github.com/jiadengx/glibc_test/blob/main/bench/strncmp_compare.out
In the data, positive values in the parentheses indicate that our
implementation took less time, indicating a performance improvement;
negative values in the parentheses mean that our implementation took
more time, indicating a decrease in performance. Following is the
summarise of the performance comparing with the generic version in the
glibc microbenchmark:
name reduce time percent
strnlen-aligned >10%
strnlen-lsx 50%-78%
strnlen-lasx 50%-88%
strcmp-aligned 0%-10% for aligned comparision
10%-20% for unaligned comparision
strcmp-lsx 0%-50%
strncmp-aligned 0%-10% for aligned comparision
10%-25% for unaligned comparision
strncmp-lsx 0%-50%
dengjianbo (3):
Loongarch: Add ifunc support for strnlen{aligned, lsx, lasx}
Loongarch: Add ifunc support for strcmp{aligned, lsx}
Loongarch: Add ifunc support for strncmp{aligned, lsx}
sysdeps/loongarch/lp64/multiarch/Makefile | 7 +
.../lp64/multiarch/ifunc-impl-list.c | 22 ++
.../loongarch/lp64/multiarch/ifunc-strcmp.h | 38 +++
.../loongarch/lp64/multiarch/ifunc-strncmp.h | 38 +++
.../loongarch/lp64/multiarch/ifunc-strnlen.h | 41 ++++
.../loongarch/lp64/multiarch/strcmp-aligned.S | 179 ++++++++++++++
sysdeps/loongarch/lp64/multiarch/strcmp-lsx.S | 162 +++++++++++++
sysdeps/loongarch/lp64/multiarch/strcmp.c | 35 +++
.../lp64/multiarch/strncmp-aligned.S | 218 ++++++++++++++++++
.../loongarch/lp64/multiarch/strncmp-lsx.S | 206 +++++++++++++++++
sysdeps/loongarch/lp64/multiarch/strncmp.c | 35 +++
.../lp64/multiarch/strnlen-aligned.S | 102 ++++++++
.../loongarch/lp64/multiarch/strnlen-lasx.S | 100 ++++++++
.../loongarch/lp64/multiarch/strnlen-lsx.S | 89 +++++++
sysdeps/loongarch/lp64/multiarch/strnlen.c | 39 ++++
15 files changed, 1311 insertions(+)
create mode 100644 sysdeps/loongarch/lp64/multiarch/ifunc-strcmp.h
create mode 100644 sysdeps/loongarch/lp64/multiarch/ifunc-strncmp.h
create mode 100644 sysdeps/loongarch/lp64/multiarch/ifunc-strnlen.h
create mode 100644 sysdeps/loongarch/lp64/multiarch/strcmp-aligned.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strcmp-lsx.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strcmp.c
create mode 100644 sysdeps/loongarch/lp64/multiarch/strncmp-aligned.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strncmp-lsx.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strncmp.c
create mode 100644 sysdeps/loongarch/lp64/multiarch/strnlen-aligned.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strnlen-lasx.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strnlen-lsx.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strnlen.c
--
2.40.0
^ permalink raw reply [flat|nested] 9+ messages in thread
* [PATCH 1/3] Loongarch: Add ifunc support for strnlen{aligned, lsx, lasx}
2023-08-22 2:11 [PATCH 0/3] Add ifunc support for str{nlen, cmp, ncmp} dengjianbo
@ 2023-08-22 2:11 ` dengjianbo
2023-08-22 2:11 ` [PATCH 2/3] Loongarch: Add ifunc support for strcmp{aligned, lsx} dengjianbo
2023-08-22 2:11 ` [PATCH 3/3] Loongarch: Add ifunc support for strncmp{aligned, lsx} dengjianbo
2 siblings, 0 replies; 9+ messages in thread
From: dengjianbo @ 2023-08-22 2:11 UTC (permalink / raw)
To: libc-alpha
Cc: adhemerval.zanella, xry111, caiyinyu, xuchenghua, huangpei, dengjianbo
Based on the glibc microbenchmark, strnlen-aligned implementation could
reduce the runtime more than 10%, strnlen-lsx implementation could reduce
the runtime about 50%-78%, strnlen-lasx implementation could reduce the
runtime about 50%-88%.
---
sysdeps/loongarch/lp64/multiarch/Makefile | 3 +
.../lp64/multiarch/ifunc-impl-list.c | 8 ++
.../loongarch/lp64/multiarch/ifunc-strnlen.h | 41 +++++++
.../lp64/multiarch/strnlen-aligned.S | 102 ++++++++++++++++++
.../loongarch/lp64/multiarch/strnlen-lasx.S | 100 +++++++++++++++++
.../loongarch/lp64/multiarch/strnlen-lsx.S | 89 +++++++++++++++
sysdeps/loongarch/lp64/multiarch/strnlen.c | 39 +++++++
7 files changed, 382 insertions(+)
create mode 100644 sysdeps/loongarch/lp64/multiarch/ifunc-strnlen.h
create mode 100644 sysdeps/loongarch/lp64/multiarch/strnlen-aligned.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strnlen-lasx.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strnlen-lsx.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strnlen.c
diff --git a/sysdeps/loongarch/lp64/multiarch/Makefile b/sysdeps/loongarch/lp64/multiarch/Makefile
index afa5104174..c4dd3143d1 100644
--- a/sysdeps/loongarch/lp64/multiarch/Makefile
+++ b/sysdeps/loongarch/lp64/multiarch/Makefile
@@ -3,6 +3,9 @@ sysdep_routines += \
strlen-aligned \
strlen-lsx \
strlen-lasx \
+ strnlen-aligned \
+ strnlen-lsx \
+ strnlen-lasx \
strchr-aligned \
strchr-lsx \
strchr-lasx \
diff --git a/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c b/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c
index 25eb96b061..7cec0b7724 100644
--- a/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c
+++ b/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c
@@ -38,6 +38,14 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
IFUNC_IMPL_ADD (array, i, strlen, 1, __strlen_aligned)
)
+ IFUNC_IMPL (i, name, strnlen,
+#if !defined __loongarch_soft_float
+ IFUNC_IMPL_ADD (array, i, strnlen, SUPPORT_LASX, __strnlen_lasx)
+ IFUNC_IMPL_ADD (array, i, strnlen, SUPPORT_LSX, __strnlen_lsx)
+#endif
+ IFUNC_IMPL_ADD (array, i, strnlen, 1, __strnlen_aligned)
+ )
+
IFUNC_IMPL (i, name, strchr,
#if !defined __loongarch_soft_float
IFUNC_IMPL_ADD (array, i, strchr, SUPPORT_LASX, __strchr_lasx)
diff --git a/sysdeps/loongarch/lp64/multiarch/ifunc-strnlen.h b/sysdeps/loongarch/lp64/multiarch/ifunc-strnlen.h
new file mode 100644
index 0000000000..5cf8981021
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/ifunc-strnlen.h
@@ -0,0 +1,41 @@
+/* Common definition for strnlen ifunc selections.
+ All versions must be listed in ifunc-impl-list.c.
+ Copyright (C) 2023 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library; if not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include <ldsodefs.h>
+#include <ifunc-init.h>
+
+#if !defined __loongarch_soft_float
+extern __typeof (REDIRECT_NAME) OPTIMIZE (lasx) attribute_hidden;
+extern __typeof (REDIRECT_NAME) OPTIMIZE (lsx) attribute_hidden;
+#endif
+
+extern __typeof (REDIRECT_NAME) OPTIMIZE (aligned) attribute_hidden;
+
+static inline void *
+IFUNC_SELECTOR (void)
+{
+#if !defined __loongarch_soft_float
+ if (SUPPORT_LASX)
+ return OPTIMIZE (lasx);
+ else if (SUPPORT_LSX)
+ return OPTIMIZE (lsx);
+ else
+#endif
+ return OPTIMIZE (aligned);
+}
diff --git a/sysdeps/loongarch/lp64/multiarch/strnlen-aligned.S b/sysdeps/loongarch/lp64/multiarch/strnlen-aligned.S
new file mode 100644
index 0000000000..b900430a5d
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/strnlen-aligned.S
@@ -0,0 +1,102 @@
+/* Optimized strnlen implementation using basic Loongarch instructions.
+ Copyright (C) 2023 Free Software Foundation, Inc.
+
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library. If not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include <sysdep.h>
+#include <sys/regdef.h>
+#include <sys/asm.h>
+
+#if IS_IN (libc)
+# define STRNLEN __strnlen_aligned
+#else
+# define STRNLEN __strnlen
+#endif
+
+LEAF(STRNLEN, 6)
+ beqz a1, L(out)
+ lu12i.w a2, 0x01010
+ andi t1, a0, 0x7
+ move t4, a0
+
+ bstrins.d a0, zero, 2, 0
+ ori a2, a2, 0x101
+ li.w t0, -1
+ ld.d t2, a0, 0
+
+ slli.d t3, t1, 3
+ bstrins.d a2, a2, 63, 32
+ li.w t5, 8
+ slli.d a3, a2, 7
+
+ sub.w t1, t5, t1
+ sll.d t0, t0, t3
+ orn t2, t2, t0
+ sub.d t0, t2, a2
+
+
+ andn t3, a3, t2
+ and t0, t0, t3
+ bnez t0, L(count_pos)
+ sub.d t5, a1, t1
+
+ bgeu t1, a1, L(out)
+ addi.d a0, a0, 8
+L(loop):
+ ld.d t2, a0, 0
+ sub.d t0, t2, a2
+
+ andn t1, a3, t2
+ sltui t6, t5, 9
+ and t0, t0, t1
+ or t7, t0, t6
+
+ bnez t7, L(count_pos)
+ ld.d t2, a0, 8
+ addi.d a0, a0, 16
+ sub.d t0, t2, a2
+
+
+ andn t1, a3, t2
+ sltui t6, t5, 17
+ and t0, t0, t1
+ addi.d t5, t5, -16
+
+ or t7, t0, t6
+ beqz t7, L(loop)
+ addi.d a0, a0, -8
+L(count_pos):
+ ctz.d t1, t0
+
+ sub.d a0, a0, t4
+ srli.d t1, t1, 3
+ add.d a0, t1, a0
+ sltu t0, a0, a1
+
+ masknez t1, a1, t0
+ maskeqz a0, a0, t0
+ or a0, a0, t1
+ jr ra
+
+
+L(out):
+ move a0, a1
+ jr ra
+END(STRNLEN)
+
+weak_alias (STRNLEN, strnlen)
+libc_hidden_builtin_def (STRNLEN)
diff --git a/sysdeps/loongarch/lp64/multiarch/strnlen-lasx.S b/sysdeps/loongarch/lp64/multiarch/strnlen-lasx.S
new file mode 100644
index 0000000000..2c03d3d9b4
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/strnlen-lasx.S
@@ -0,0 +1,100 @@
+/* Optimized strnlen implementation using loongarch LASX instructions
+ Copyright (C) 2023 Free Software Foundation, Inc.
+
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library. If not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include <sysdep.h>
+#include <sys/regdef.h>
+#include <sys/asm.h>
+
+#if IS_IN (libc) && !defined __loongarch_soft_float
+
+# define STRNLEN __strnlen_lasx
+
+LEAF(STRNLEN, 6)
+ beqz a1, L(ret0)
+ andi t1, a0, 0x3f
+ li.d t3, 65
+ sub.d a2, a0, t1
+
+ xvld xr0, a2, 0
+ xvld xr1, a2, 32
+ sub.d t1, t3, t1
+ move a3, a0
+
+ sltu t1, a1, t1
+ xvmsknz.b xr0, xr0
+ xvmsknz.b xr1, xr1
+ xvpickve.w xr2, xr0, 4
+
+ xvpickve.w xr3, xr1, 4
+ vilvl.h vr0, vr2, vr0
+ vilvl.h vr1, vr3, vr1
+ vilvl.w vr0, vr1, vr0
+
+
+ movfr2gr.d t0, fa0
+ sra.d t0, t0, a0
+ orn t1, t1, t0
+ bnez t1, L(end)
+
+ add.d a4, a0, a1
+ move a0, a2
+ addi.d a4, a4, -1
+ bstrins.d a4, zero, 5, 0
+
+L(loop):
+ xvld xr0, a0, 64
+ xvld xr1, a0, 96
+ addi.d a0, a0, 64
+ beq a0, a4, L(out)
+
+ xvmin.bu xr2, xr0, xr1
+ xvsetanyeqz.b fcc0, xr2
+ bceqz fcc0, L(loop)
+L(out):
+ xvmsknz.b xr0, xr0
+
+
+ xvmsknz.b xr1, xr1
+ xvpickve.w xr2, xr0, 4
+ xvpickve.w xr3, xr1, 4
+ vilvl.h vr0, vr2, vr0
+
+ vilvl.h vr1, vr3, vr1
+ vilvl.w vr0, vr1, vr0
+ movfr2gr.d t0, fa0
+L(end):
+ sub.d a0, a0, a3
+
+ cto.d t0, t0
+ add.d a0, a0, t0
+ sltu t1, a0, a1
+ masknez t0, a1, t1
+
+ maskeqz t1, a0, t1
+ or a0, t0, t1
+ jr ra
+L(ret0):
+ move a0, zero
+
+
+ jr ra
+END(STRNLEN)
+
+libc_hidden_def (STRNLEN)
+#endif
diff --git a/sysdeps/loongarch/lp64/multiarch/strnlen-lsx.S b/sysdeps/loongarch/lp64/multiarch/strnlen-lsx.S
new file mode 100644
index 0000000000..b769a89584
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/strnlen-lsx.S
@@ -0,0 +1,89 @@
+/* Optimized strnlen implementation using loongarch LSX instructions
+ Copyright (C) 2023 Free Software Foundation, Inc.
+
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library. If not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include <sysdep.h>
+#include <sys/regdef.h>
+#include <sys/asm.h>
+
+#if IS_IN (libc) && !defined __loongarch_soft_float
+
+# define STRNLEN __strnlen_lsx
+
+LEAF(STRNLEN, 6)
+ beqz a1, L(ret0)
+ andi t1, a0, 0x1f
+ li.d t3, 33
+ sub.d a2, a0, t1
+
+ vld vr0, a2, 0
+ vld vr1, a2, 16
+ sub.d t1, t3, t1
+ move a3, a0
+
+ sltu t1, a1, t1
+ vmsknz.b vr0, vr0
+ vmsknz.b vr1, vr1
+ vilvl.h vr0, vr1, vr0
+
+ movfr2gr.s t0, fa0
+ sra.w t0, t0, a0
+ orn t1, t1, t0
+ bnez t1, L(end)
+
+
+ add.d a4, a0, a1
+ move a0, a2
+ addi.d a4, a4, -1
+ bstrins.d a4, zero, 4, 0
+
+L(loop):
+ vld vr0, a0, 32
+ vld vr1, a0, 48
+ addi.d a0, a0, 32
+ beq a0, a4, L(out)
+
+ vmin.bu vr2, vr0, vr1
+ vsetanyeqz.b fcc0, vr2
+ bceqz fcc0, L(loop)
+L(out):
+ vmsknz.b vr0, vr0
+
+ vmsknz.b vr1, vr1
+ vilvl.h vr0, vr1, vr0
+ movfr2gr.s t0, fa0
+L(end):
+ sub.d a0, a0, a3
+
+
+ cto.w t0, t0
+ add.d a0, a0, t0
+ sltu t1, a0, a1
+ masknez t0, a1, t1
+
+ maskeqz t1, a0, t1
+ or a0, t0, t1
+ jr ra
+L(ret0):
+ move a0, zero
+
+ jr ra
+END(STRNLEN)
+
+libc_hidden_builtin_def (STRNLEN)
+#endif
diff --git a/sysdeps/loongarch/lp64/multiarch/strnlen.c b/sysdeps/loongarch/lp64/multiarch/strnlen.c
new file mode 100644
index 0000000000..38b7a25a7a
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/strnlen.c
@@ -0,0 +1,39 @@
+/* Multiple versions of strnlen.
+ All versions must be listed in ifunc-impl-list.c.
+ Copyright (C) 2023 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library; if not, see
+ <https://www.gnu.org/licenses/>. */
+
+/* Define multiple versions only for the definition in libc. */
+#if IS_IN (libc)
+# define strnlen __redirect_strnlen
+# define __strnlen __redirect___strnlen
+# include <string.h>
+# undef __strnlen
+# undef strnlen
+
+# define SYMBOL_NAME strnlen
+# include "ifunc-strnlen.h"
+
+libc_ifunc_redirected (__redirect_strnlen, __strnlen, IFUNC_SELECTOR ());
+weak_alias (__strnlen, strnlen);
+# ifdef SHARED
+__hidden_ver1 (__strnlen, __GI___strnlen, __redirect___strnlen)
+ __attribute__((visibility ("hidden"))) __attribute_copy__ (strnlen);
+__hidden_ver1 (strnlen, __GI_strnlen, __redirect_strnlen)
+ __attribute__((weak, visibility ("hidden"))) __attribute_copy__ (strnlen);
+# endif
+#endif
--
2.40.0
^ permalink raw reply [flat|nested] 9+ messages in thread
* [PATCH 2/3] Loongarch: Add ifunc support for strcmp{aligned, lsx}
2023-08-22 2:11 [PATCH 0/3] Add ifunc support for str{nlen, cmp, ncmp} dengjianbo
2023-08-22 2:11 ` [PATCH 1/3] Loongarch: Add ifunc support for strnlen{aligned, lsx, lasx} dengjianbo
@ 2023-08-22 2:11 ` dengjianbo
2023-08-22 2:11 ` [PATCH 3/3] Loongarch: Add ifunc support for strncmp{aligned, lsx} dengjianbo
2 siblings, 0 replies; 9+ messages in thread
From: dengjianbo @ 2023-08-22 2:11 UTC (permalink / raw)
To: libc-alpha
Cc: adhemerval.zanella, xry111, caiyinyu, xuchenghua, huangpei, dengjianbo
Based on the glibc microbenchmark, strcmp-aligned implementation could
reduce the runtime 0%-10% for aligned comparison, 10%-20% for unaligned
comparison, strcmp-lsx implemenation could reduce the runtime 0%-50%.
---
sysdeps/loongarch/lp64/multiarch/Makefile | 2 +
.../lp64/multiarch/ifunc-impl-list.c | 7 +
.../loongarch/lp64/multiarch/ifunc-strcmp.h | 38 ++++
.../loongarch/lp64/multiarch/strcmp-aligned.S | 179 ++++++++++++++++++
sysdeps/loongarch/lp64/multiarch/strcmp-lsx.S | 162 ++++++++++++++++
sysdeps/loongarch/lp64/multiarch/strcmp.c | 35 ++++
6 files changed, 423 insertions(+)
create mode 100644 sysdeps/loongarch/lp64/multiarch/ifunc-strcmp.h
create mode 100644 sysdeps/loongarch/lp64/multiarch/strcmp-aligned.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strcmp-lsx.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strcmp.c
diff --git a/sysdeps/loongarch/lp64/multiarch/Makefile b/sysdeps/loongarch/lp64/multiarch/Makefile
index c4dd3143d1..d5a500decd 100644
--- a/sysdeps/loongarch/lp64/multiarch/Makefile
+++ b/sysdeps/loongarch/lp64/multiarch/Makefile
@@ -12,6 +12,8 @@ sysdep_routines += \
strchrnul-aligned \
strchrnul-lsx \
strchrnul-lasx \
+ strcmp-aligned \
+ strcmp-lsx \
memcpy-aligned \
memcpy-unaligned \
memmove-unaligned \
diff --git a/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c b/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c
index 7cec0b7724..9183b7da24 100644
--- a/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c
+++ b/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c
@@ -62,6 +62,13 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
IFUNC_IMPL_ADD (array, i, strchrnul, 1, __strchrnul_aligned)
)
+ IFUNC_IMPL (i, name, strcmp,
+#if !defined __loongarch_soft_float
+ IFUNC_IMPL_ADD (array, i, strcmp, SUPPORT_LSX, __strcmp_lsx)
+#endif
+ IFUNC_IMPL_ADD (array, i, strcmp, 1, __strcmp_aligned)
+ )
+
IFUNC_IMPL (i, name, memcpy,
#if !defined __loongarch_soft_float
IFUNC_IMPL_ADD (array, i, memcpy, SUPPORT_LASX, __memcpy_lasx)
diff --git a/sysdeps/loongarch/lp64/multiarch/ifunc-strcmp.h b/sysdeps/loongarch/lp64/multiarch/ifunc-strcmp.h
new file mode 100644
index 0000000000..ca26352bec
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/ifunc-strcmp.h
@@ -0,0 +1,38 @@
+/* Common definition for strcmp ifunc selection.
+ All versions must be listed in ifunc-impl-list.c.
+ Copyright (C) 2023 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library; if not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include <ldsodefs.h>
+#include <ifunc-init.h>
+
+#if !defined __loongarch_soft_float
+extern __typeof (REDIRECT_NAME) OPTIMIZE (lsx) attribute_hidden;
+#endif
+
+extern __typeof (REDIRECT_NAME) OPTIMIZE (aligned) attribute_hidden;
+
+static inline void *
+IFUNC_SELECTOR (void)
+{
+#if !defined __loongarch_soft_float
+ if (SUPPORT_LSX)
+ return OPTIMIZE (lsx);
+ else
+#endif
+ return OPTIMIZE (aligned);
+}
diff --git a/sysdeps/loongarch/lp64/multiarch/strcmp-aligned.S b/sysdeps/loongarch/lp64/multiarch/strcmp-aligned.S
new file mode 100644
index 0000000000..f5f4f3364e
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/strcmp-aligned.S
@@ -0,0 +1,179 @@
+/* Optimized strcmp implementation using basic Loongarch instructions.
+ Copyright (C) 2023 Free Software Foundation, Inc.
+
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library. If not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include <sysdep.h>
+#include <sys/regdef.h>
+#include <sys/asm.h>
+
+#if IS_IN (libc)
+# define STRCMP_NAME __strcmp_aligned
+#else
+# define STRCMP_NAME strcmp
+#endif
+
+LEAF(STRCMP_NAME, 6)
+ lu12i.w a4, 0x01010
+ andi a2, a0, 0x7
+ ori a4, a4, 0x101
+ andi a3, a1, 0x7
+
+ bstrins.d a4, a4, 63, 32
+ li.d t7, -1
+ li.d t8, 8
+ slli.d a5, a4, 7
+
+ bne a2, a3, L(unaligned)
+ bstrins.d a0, zero, 2, 0
+ bstrins.d a1, zero, 2, 0
+ ld.d t0, a0, 0
+
+ ld.d t1, a1, 0
+ slli.d t3, a2, 3
+ sll.d t2, t7, t3
+ orn t0, t0, t2
+
+
+ orn t1, t1, t2
+ sub.d t2, t0, a4
+ andn t3, a5, t0
+ and t2, t2, t3
+
+ bne t0, t1, L(al_end)
+L(al_loop):
+ bnez t2, L(ret0)
+ ldx.d t0, a0, t8
+ ldx.d t1, a1, t8
+
+ addi.d t8, t8, 8
+ sub.d t2, t0, a4
+ andn t3, a5, t0
+ and t2, t2, t3
+
+ beq t0, t1, L(al_loop)
+L(al_end):
+ xor t3, t0, t1
+ or t2, t2, t3
+ ctz.d t3, t2
+
+
+ bstrins.d t3, zero, 2, 0
+ srl.d t0, t0, t3
+ srl.d t1, t1, t3
+ andi t0, t0, 0xff
+
+ andi t1, t1, 0xff
+ sub.d a0, t0, t1
+ jr ra
+ nop
+
+L(ret0):
+ move a0, zero
+ jr ra
+ nop
+ nop
+
+L(unaligned):
+ slt a6, a3, a2
+ xor t0, a0, a1
+ maskeqz t0, t0, a6
+ xor a0, a0, t0
+
+
+ xor a1, a1, t0
+ andi a2, a0, 0x7
+ andi a3, a1, 0x7
+ bstrins.d a0, zero, 2, 0
+
+ bstrins.d a1, zero, 2, 0
+ ld.d t4, a0, 0
+ ld.d t1, a1, 0
+ slli.d a2, a2, 3
+
+ slli.d a3, a3, 3
+ srl.d t0, t4, a2
+ srl.d t1, t1, a3
+ srl.d t5, t7, a3
+
+ orn t0, t0, t5
+ orn t1, t1, t5
+ bne t0, t1, L(not_equal)
+ sll.d t5, t7, a2
+
+
+ sub.d a3, a2, a3
+ orn t4, t4, t5
+ sub.d a2, zero, a3
+ sub.d t2, t4, a4
+
+ andn t3, a5, t4
+ and t2, t2, t3
+ bnez t2, L(find_zero)
+L(un_loop):
+ srl.d t5, t4, a3
+
+ ldx.d t4, a0, t8
+ ldx.d t1, a1, t8
+ addi.d t8, t8, 8
+ sll.d t0, t4, a2
+
+ or t0, t0, t5
+ bne t0, t1, L(not_equal)
+ sub.d t2, t4, a4
+ andn t3, a5, t4
+
+
+ and t2, t2, t3
+ beqz t2, L(un_loop)
+L(find_zero):
+ sub.d t2, t0, a4
+ andn t3, a5, t0
+
+ and t2, t2, t3
+ bnez t2, L(ret0)
+ ldx.d t1, a1, t8
+ srl.d t0, t4, a3
+
+L(not_equal):
+ sub.d t2, t0, a4
+ andn t3, a5, t0
+ and t2, t2, t3
+ xor t3, t0, t1
+
+ or t2, t2, t3
+L(un_end):
+ ctz.d t3, t2
+ bstrins.d t3, zero, 2, 0
+ srl.d t0, t0, t3
+
+
+ srl.d t1, t1, t3
+ andi t0, t0, 0xff
+ andi t1, t1, 0xff
+ sub.d t2, t0, t1
+
+
+ sub.d t3, t1, t0
+ masknez t0, t2, a6
+ maskeqz t1, t3, a6
+ or a0, t0, t1
+
+ jr ra
+END(STRCMP_NAME)
+
+libc_hidden_builtin_def (STRCMP_NAME)
diff --git a/sysdeps/loongarch/lp64/multiarch/strcmp-lsx.S b/sysdeps/loongarch/lp64/multiarch/strcmp-lsx.S
new file mode 100644
index 0000000000..19776d9cb2
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/strcmp-lsx.S
@@ -0,0 +1,162 @@
+/* Optimized strcmp implementation using Loongarch LSX instructions.
+ Copyright (C) 2023 Free Software Foundation, Inc.
+
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library. If not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include <sysdep.h>
+#include <sys/regdef.h>
+#include <sys/asm.h>
+
+#if IS_IN (libc) && !defined __loongarch_soft_float
+
+# define STRCMP __strcmp_lsx
+
+L(magic_num):
+ .align 6
+ .dword 0x0706050403020100
+ .dword 0x0f0e0d0c0b0a0908
+ENTRY_NO_ALIGN(STRCMP)
+ pcaddi t0, -4
+ andi a2, a0, 0xf
+ vld vr2, t0, 0
+ andi a3, a1, 0xf
+
+ bne a2, a3, L(unaligned)
+ bstrins.d a0, zero, 3, 0
+ bstrins.d a1, zero, 3, 0
+ vld vr0, a0, 0
+
+ vld vr1, a1, 0
+ vreplgr2vr.b vr3, a2
+ vslt.b vr2, vr2, vr3
+ vseq.b vr3, vr0, vr1
+
+
+ vmin.bu vr3, vr0, vr3
+ vor.v vr3, vr3, vr2
+ vsetanyeqz.b fcc0, vr3
+ bcnez fcc0, L(al_out)
+
+L(al_loop):
+ vld vr0, a0, 16
+ vld vr1, a1, 16
+ addi.d a0, a0, 16
+ addi.d a1, a1, 16
+
+ vseq.b vr3, vr0, vr1
+ vmin.bu vr3, vr0, vr3
+ vsetanyeqz.b fcc0, vr3
+ bceqz fcc0, L(al_loop)
+
+L(al_out):
+ vseqi.b vr3, vr3, 0
+ vfrstpi.b vr3, vr3, 0
+ vshuf.b vr0, vr0, vr0, vr3
+ vshuf.b vr1, vr1, vr1, vr3
+
+
+ vpickve2gr.bu t0, vr0, 0
+ vpickve2gr.bu t1, vr1, 0
+ sub.d a0, t0, t1
+ jr ra
+
+L(unaligned):
+ slt a4, a3, a2
+ xor t0, a0, a1
+ maskeqz t0, t0, a4
+ xor a0, a0, t0
+
+ xor a1, a1, t0
+ andi a2, a0, 0xf
+ andi a3, a1, 0xf
+ bstrins.d a0, zero, 3, 0
+
+ bstrins.d a1, zero, 3, 0
+ vld vr3, a0, 0
+ vld vr1, a1, 0
+ vreplgr2vr.b vr4, a2
+
+
+ vreplgr2vr.b vr5, a3
+ vslt.b vr7, vr2, vr5
+ vsub.b vr5, vr5, vr4
+ vaddi.bu vr6, vr2, 16
+
+ vsub.b vr6, vr6, vr5
+ vshuf.b vr0, vr3, vr3, vr6
+ vor.v vr0, vr0, vr7
+ vor.v vr1, vr1, vr7
+
+ vseq.b vr5, vr0, vr1
+ vsetanyeqz.b fcc0, vr5
+ bcnez fcc0, L(not_equal)
+ vslt.b vr4, vr2, vr4
+
+ vor.v vr0, vr3, vr4
+ vsetanyeqz.b fcc0, vr0
+ bcnez fcc0, L(find_zero)
+ nop
+
+L(un_loop):
+ vld vr3, a0, 16
+ vld vr1, a1, 16
+ addi.d a0, a0, 16
+ addi.d a1, a1, 16
+
+ vshuf.b vr0, vr3, vr0, vr6
+ vseq.b vr5, vr0, vr1
+ vsetanyeqz.b fcc0, vr5
+ bcnez fcc0, L(not_equal)
+
+ vsetanyeqz.b fcc0, vr3
+ vor.v vr0, vr3, vr3
+ bceqz fcc0, L(un_loop)
+L(find_zero):
+ vmin.bu vr5, vr1, vr5
+
+ vsetanyeqz.b fcc0, vr5
+ bcnez fcc0, L(ret0)
+ vld vr1, a1, 16
+ vshuf.b vr0, vr3, vr3, vr6
+
+
+ vseq.b vr5, vr0, vr1
+L(not_equal):
+ vmin.bu vr5, vr0, vr5
+L(un_end):
+ vseqi.b vr5, vr5, 0
+ vfrstpi.b vr5, vr5, 0
+
+ vshuf.b vr0, vr0, vr0, vr5
+ vshuf.b vr1, vr1, vr1, vr5
+ vpickve2gr.bu t0, vr0, 0
+ vpickve2gr.bu t1, vr1, 0
+
+ sub.d t3, t0, t1
+ sub.d t4, t1, t0
+ masknez t0, t3, a4
+ maskeqz t1, t4, a4
+
+ or a0, t0, t1
+ jr ra
+L(ret0):
+ move a0, zero
+ jr ra
+END(STRCMP)
+
+libc_hidden_builtin_def (STRCMP)
+#endif
diff --git a/sysdeps/loongarch/lp64/multiarch/strcmp.c b/sysdeps/loongarch/lp64/multiarch/strcmp.c
new file mode 100644
index 0000000000..6f249c0b59
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/strcmp.c
@@ -0,0 +1,35 @@
+/* Multiple versions of strcmp.
+ All versions must be listed in ifunc-impl-list.c.
+ Copyright (C) 2023 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library; if not, see
+ <http://www.gnu.org/licenses/>. */
+
+/* Define multiple versions only for the definition in libc. */
+#if IS_IN (libc)
+# define strcmp __redirect_strcmp
+# include <string.h>
+# undef strcmp
+
+# define SYMBOL_NAME strcmp
+# include "ifunc-strcmp.h"
+
+libc_ifunc_redirected (__redirect_strcmp, strcmp, IFUNC_SELECTOR ());
+
+# ifdef SHARED
+__hidden_ver1 (strcmp, __GI_strcmp, __redirect_strcmp)
+ __attribute__ ((visibility ("hidden"))) __attribute_copy__ (strcmp);
+# endif
+#endif
--
2.40.0
^ permalink raw reply [flat|nested] 9+ messages in thread
* [PATCH 3/3] Loongarch: Add ifunc support for strncmp{aligned, lsx}
2023-08-22 2:11 [PATCH 0/3] Add ifunc support for str{nlen, cmp, ncmp} dengjianbo
2023-08-22 2:11 ` [PATCH 1/3] Loongarch: Add ifunc support for strnlen{aligned, lsx, lasx} dengjianbo
2023-08-22 2:11 ` [PATCH 2/3] Loongarch: Add ifunc support for strcmp{aligned, lsx} dengjianbo
@ 2023-08-22 2:11 ` dengjianbo
2023-08-22 3:56 ` Richard Henderson
2 siblings, 1 reply; 9+ messages in thread
From: dengjianbo @ 2023-08-22 2:11 UTC (permalink / raw)
To: libc-alpha
Cc: adhemerval.zanella, xry111, caiyinyu, xuchenghua, huangpei, dengjianbo
Based on the glibc microbenchmark, only a few short inputs with this
strncmp-aligned and strncmp-lsx implementation experience performance
degradation, overall, strncmp-aligned could reduce the runtime 0%-10%
for aligned comparision, 10%-25% for unaligend comparision, strncmp-lsx
could reduce the runtime about 0%-60%.
---
sysdeps/loongarch/lp64/multiarch/Makefile | 2 +
.../lp64/multiarch/ifunc-impl-list.c | 7 +
.../loongarch/lp64/multiarch/ifunc-strncmp.h | 38 +++
.../lp64/multiarch/strncmp-aligned.S | 218 ++++++++++++++++++
.../loongarch/lp64/multiarch/strncmp-lsx.S | 206 +++++++++++++++++
sysdeps/loongarch/lp64/multiarch/strncmp.c | 35 +++
6 files changed, 506 insertions(+)
create mode 100644 sysdeps/loongarch/lp64/multiarch/ifunc-strncmp.h
create mode 100644 sysdeps/loongarch/lp64/multiarch/strncmp-aligned.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strncmp-lsx.S
create mode 100644 sysdeps/loongarch/lp64/multiarch/strncmp.c
diff --git a/sysdeps/loongarch/lp64/multiarch/Makefile b/sysdeps/loongarch/lp64/multiarch/Makefile
index d5a500decd..5d7ae7ae73 100644
--- a/sysdeps/loongarch/lp64/multiarch/Makefile
+++ b/sysdeps/loongarch/lp64/multiarch/Makefile
@@ -14,6 +14,8 @@ sysdep_routines += \
strchrnul-lasx \
strcmp-aligned \
strcmp-lsx \
+ strncmp-aligned \
+ strncmp-lsx \
memcpy-aligned \
memcpy-unaligned \
memmove-unaligned \
diff --git a/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c b/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c
index 9183b7da24..c8ba87bd81 100644
--- a/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c
+++ b/sysdeps/loongarch/lp64/multiarch/ifunc-impl-list.c
@@ -69,6 +69,13 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
IFUNC_IMPL_ADD (array, i, strcmp, 1, __strcmp_aligned)
)
+ IFUNC_IMPL (i, name, strncmp,
+#if !defined __loongarch_soft_float
+ IFUNC_IMPL_ADD (array, i, strncmp, SUPPORT_LSX, __strncmp_lsx)
+#endif
+ IFUNC_IMPL_ADD (array, i, strncmp, 1, __strncmp_aligned)
+ )
+
IFUNC_IMPL (i, name, memcpy,
#if !defined __loongarch_soft_float
IFUNC_IMPL_ADD (array, i, memcpy, SUPPORT_LASX, __memcpy_lasx)
diff --git a/sysdeps/loongarch/lp64/multiarch/ifunc-strncmp.h b/sysdeps/loongarch/lp64/multiarch/ifunc-strncmp.h
new file mode 100644
index 0000000000..1a7dc36ba6
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/ifunc-strncmp.h
@@ -0,0 +1,38 @@
+/* Common definition for strncmp ifunc selection.
+ All versions must be listed in ifunc-impl-list.c.
+ Copyright (C) 2023 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library; if not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include <ldsodefs.h>
+#include <ifunc-init.h>
+
+#if !defined __loongarch_soft_float
+extern __typeof (REDIRECT_NAME) OPTIMIZE (lsx) attribute_hidden;
+#endif
+
+extern __typeof (REDIRECT_NAME) OPTIMIZE (aligned) attribute_hidden;
+
+static inline void *
+IFUNC_SELECTOR (void)
+{
+#if !defined __loongarch_soft_float
+ if (SUPPORT_LSX)
+ return OPTIMIZE (lsx);
+ else
+#endif
+ return OPTIMIZE (aligned);
+}
diff --git a/sysdeps/loongarch/lp64/multiarch/strncmp-aligned.S b/sysdeps/loongarch/lp64/multiarch/strncmp-aligned.S
new file mode 100644
index 0000000000..e2687fa770
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/strncmp-aligned.S
@@ -0,0 +1,218 @@
+/* Optimized strncmp implementation using basic Loongarch instructions.
+ Copyright (C) 2023 Free Software Foundation, Inc.
+
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library. If not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include <sysdep.h>
+#include <sys/regdef.h>
+#include <sys/asm.h>
+
+#if IS_IN (libc)
+# define STRNCMP __strncmp_aligned
+#else
+# define STRNCMP strncmp
+#endif
+
+LEAF(STRNCMP, 6)
+ beqz a2, L(ret0)
+ lu12i.w a5, 0x01010
+ andi a3, a0, 0x7
+ ori a5, a5, 0x101
+
+ andi a4, a1, 0x7
+ bstrins.d a5, a5, 63, 32
+ li.d t7, -1
+ li.d t8, 8
+
+ addi.d a2, a2, -1
+ slli.d a6, a5, 7
+ bne a3, a4, L(unaligned)
+ bstrins.d a0, zero, 2, 0
+
+ bstrins.d a1, zero, 2, 0
+ ld.d t0, a0, 0
+ ld.d t1, a1, 0
+ slli.d t2, a3, 3
+
+
+ sub.d t5, t8, a3
+ srl.d t3, t7, t2
+ srl.d t0, t0, t2
+ srl.d t1, t1, t2
+
+ orn t0, t0, t3
+ orn t1, t1, t3
+ sub.d t2, t0, a5
+ andn t3, a6, t0
+
+ and t2, t2, t3
+ bne t0, t1, L(al_end)
+ sltu t4, a2, t5
+ sub.d a2, a2, t5
+
+L(al_loop):
+ or t4, t2, t4
+ bnez t4, L(ret0)
+ ldx.d t0, a0, t8
+ ldx.d t1, a1, t8
+
+
+ addi.d t8, t8, 8
+ sltui t4, a2, 8
+ addi.d a2, a2, -8
+ sub.d t2, t0, a5
+
+ andn t3, a6, t0
+ and t2, t2, t3
+ beq t0, t1, L(al_loop)
+ addi.d a2, a2, 8
+
+L(al_end):
+ xor t3, t0, t1
+ or t2, t2, t3
+ ctz.d t2, t2
+ srli.d t4, t2, 3
+
+ bstrins.d t2, zero, 2, 0
+ srl.d t0, t0, t2
+ srl.d t1, t1, t2
+ andi t0, t0, 0xff
+
+
+ andi t1, t1, 0xff
+ sltu t2, a2, t4
+ sub.d a0, t0, t1
+ masknez a0, a0, t2
+
+ jr ra
+L(ret0):
+ move a0, zero
+ jr ra
+ nop
+
+L(unaligned):
+ slt a7, a4, a3
+ xor t0, a0, a1
+ maskeqz t0, t0, a7
+ xor a0, a0, t0
+
+ xor a1, a1, t0
+ andi a3, a0, 0x7
+ andi a4, a1, 0x7
+ bstrins.d a0, zero, 2, 0
+
+
+ bstrins.d a1, zero, 2, 0
+ ld.d t4, a0, 0
+ ld.d t1, a1, 0
+ slli.d t2, a3, 3
+
+ slli.d t3, a4, 3
+ srl.d t5, t7, t3
+ srl.d t0, t4, t2
+ srl.d t1, t1, t3
+
+ orn t0, t0, t5
+ orn t1, t1, t5
+ bne t0, t1, L(not_equal)
+ sub.d t6, t8, a4
+
+ sub.d a4, t2, t3
+ sll.d t2, t7, t2
+ sub.d t5, t8, a3
+ orn t4, t4, t2
+
+
+ sub.d t2, t4, a5
+ andn t3, a6, t4
+ sltu t7, a2, t5
+ and t2, t2, t3
+
+ sub.d a3, zero, a4
+ or t2, t2, t7
+ bnez t2, L(un_end)
+ sub.d t7, t5, t6
+
+ sub.d a2, a2, t5
+ sub.d t6, t8, t7
+L(un_loop):
+ srl.d t5, t4, a4
+ ldx.d t4, a0, t8
+
+ ldx.d t1, a1, t8
+ addi.d t8, t8, 8
+ sll.d t0, t4, a3
+ or t0, t0, t5
+
+
+ bne t0, t1, L(loop_not_equal)
+ sub.d t2, t4, a5
+ andn t3, a6, t4
+ sltui t5, a2, 8
+
+ and t2, t2, t3
+ addi.d a2, a2, -8
+ or t3, t2, t5
+ beqz t3, L(un_loop)
+
+ addi.d a2, a2, 8
+L(un_end):
+ sub.d t2, t0, a5
+ andn t3, a6, t0
+ sltu t5, a2, t6
+
+ and t2, t2, t3
+ or t2, t2, t5
+ bnez t2, L(ret0)
+ ldx.d t1, a1, t8
+
+
+ srl.d t0, t4, a4
+ sub.d a2, a2, t6
+L(not_equal):
+ sub.d t2, t0, a5
+ andn t3, a6, t0
+
+ xor t4, t0, t1
+ and t2, t2, t3
+ or t2, t2, t4
+ ctz.d t2, t2
+
+ bstrins.d t2, zero, 2, 0
+ srli.d t4, t2, 3
+ srl.d t0, t0, t2
+ srl.d t1, t1, t2
+
+ andi t0, t0, 0xff
+ andi t1, t1, 0xff
+ sub.d t2, t0, t1
+ sub.d t3, t1, t0
+
+
+ masknez t0, t2, a7
+ maskeqz t1, t3, a7
+ sltu t2, a2, t4
+ or a0, t0, t1
+
+ masknez a0, a0, t2
+ jr ra
+L(loop_not_equal):
+ add.d a2, a2, t7
+ b L(not_equal)
+END(STRNCMP)
+
+libc_hidden_builtin_def (STRNCMP)
diff --git a/sysdeps/loongarch/lp64/multiarch/strncmp-lsx.S b/sysdeps/loongarch/lp64/multiarch/strncmp-lsx.S
new file mode 100644
index 0000000000..595472fcda
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/strncmp-lsx.S
@@ -0,0 +1,206 @@
+/* Optimized strncmp implementation using Loongarch LSX instructions.
+ Copyright (C) 2023 Free Software Foundation, Inc.
+
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library. If not, see
+ <https://www.gnu.org/licenses/>. */
+
+#include <sysdep.h>
+#include <sys/regdef.h>
+#include <sys/asm.h>
+
+#if IS_IN (libc) && !defined __loongarch_soft_float
+
+# define STRNCMP __strncmp_lsx
+
+L(magic_num):
+ .align 6
+ .dword 0x0706050403020100
+ .dword 0x0f0e0d0c0b0a0908
+ENTRY_NO_ALIGN(STRNCMP)
+ beqz a2, L(ret0)
+ pcaddi t0, -5
+ andi a3, a0, 0xf
+ vld vr2, t0, 0
+
+ andi a4, a1, 0xf
+ li.d t2, 16
+ bne a3, a4, L(unaligned)
+ xor t0, a0, a3
+
+ xor t1, a1, a4
+ vld vr0, t0, 0
+ vld vr1, t1, 0
+ vreplgr2vr.b vr3, a3
+
+
+ sub.d t2, t2, a3
+ vadd.b vr3, vr3, vr2
+ vshuf.b vr0, vr3, vr0, vr3
+ vshuf.b vr1, vr3, vr1, vr3
+
+ vseq.b vr3, vr0, vr1
+ vmin.bu vr3, vr0, vr3
+ bgeu t2, a2, L(al_early_end)
+ vsetanyeqz.b fcc0, vr3
+
+ bcnez fcc0, L(al_end)
+ add.d a3, a0, a2
+ addi.d a4, a3, -1
+ bstrins.d a4, zero, 3, 0
+
+ sub.d a2, a3, a4
+L(al_loop):
+ vld vr0, t0, 16
+ vld vr1, t1, 16
+ addi.d t0, t0, 16
+
+
+ addi.d t1, t1, 16
+ vseq.b vr3, vr0, vr1
+ vmin.bu vr3, vr0, vr3
+ beq t0, a4, L(al_early_end)
+
+ vsetanyeqz.b fcc0, vr3
+ bceqz fcc0, L(al_loop)
+L(al_end):
+ vseqi.b vr3, vr3, 0
+ vfrstpi.b vr3, vr3, 0
+
+ vshuf.b vr0, vr0, vr0, vr3
+ vshuf.b vr1, vr1, vr1, vr3
+ vpickve2gr.bu t0, vr0, 0
+ vpickve2gr.bu t1, vr1, 0
+
+ sub.d a0, t0, t1
+ jr ra
+L(al_early_end):
+ vreplgr2vr.b vr4, a2
+ vslt.b vr4, vr2, vr4
+
+
+ vorn.v vr3, vr3, vr4
+ b L(al_end)
+L(unaligned):
+ slt a5, a3, a4
+ xor t0, a0, a1
+
+ maskeqz t0, t0, a5
+ xor a0, a0, t0
+ xor a1, a1, t0
+ andi a3, a0, 0xf
+
+ andi a4, a1, 0xf
+ xor t0, a0, a3
+ xor t1, a1, a4
+ vld vr0, t0, 0
+
+ vld vr3, t1, 0
+ sub.d t2, t2, a3
+ vreplgr2vr.b vr4, a3
+ vreplgr2vr.b vr5, a4
+
+
+ vaddi.bu vr6, vr2, 16
+ vsub.b vr7, vr4, vr5
+ vsub.b vr6, vr6, vr7
+ vadd.b vr4, vr2, vr4
+
+ vshuf.b vr1, vr3, vr3, vr6
+ vshuf.b vr0, vr7, vr0, vr4
+ vshuf.b vr1, vr7, vr1, vr4
+ vseq.b vr4, vr0, vr1
+
+ vmin.bu vr4, vr0, vr4
+ bgeu t2, a2, L(un_early_end)
+ vsetanyeqz.b fcc0, vr4
+ bcnez fcc0, L(un_end)
+
+ add.d a6, a0, a2
+ vslt.b vr5, vr2, vr5
+ addi.d a7, a6, -1
+ vor.v vr3, vr3, vr5
+
+
+ bstrins.d a7, zero, 3, 0
+ sub.d a2, a6, a7
+L(un_loop):
+ vld vr0, t0, 16
+ addi.d t0, t0, 16
+
+ vsetanyeqz.b fcc0, vr3
+ bcnez fcc0, L(has_zero)
+ beq t0, a7, L(end_with_len)
+ vor.v vr1, vr3, vr3
+
+ vld vr3, t1, 16
+ addi.d t1, t1, 16
+ vshuf.b vr1, vr3, vr1, vr6
+ vseq.b vr4, vr0, vr1
+
+ vmin.bu vr4, vr0, vr4
+ vsetanyeqz.b fcc0, vr4
+ bceqz fcc0, L(un_loop)
+L(un_end):
+ vseqi.b vr4, vr4, 0
+
+
+ vfrstpi.b vr4, vr4, 0
+ vshuf.b vr0, vr0, vr0, vr4
+ vshuf.b vr1, vr1, vr1, vr4
+ vpickve2gr.bu t0, vr0, 0
+
+ vpickve2gr.bu t1, vr1, 0
+ sub.d t2, t0, t1
+ sub.d t3, t1, t0
+ masknez t0, t2, a5
+
+ maskeqz t1, t3, a5
+ or a0, t0, t1
+ jr ra
+L(has_zero):
+ vshuf.b vr1, vr3, vr3, vr6
+
+ vseq.b vr4, vr0, vr1
+ vmin.bu vr4, vr0, vr4
+ bne t0, a7, L(un_end)
+L(un_early_end):
+ vreplgr2vr.b vr5, a2
+
+ vslt.b vr5, vr2, vr5
+ vorn.v vr4, vr4, vr5
+ b L(un_end)
+L(end_with_len):
+ sub.d a6, a3, a4
+
+ bgeu a6, a2, 1f
+ vld vr4, t1, 16
+1:
+ vshuf.b vr1, vr4, vr3, vr6
+ vseq.b vr4, vr0, vr1
+
+ vmin.bu vr4, vr0, vr4
+ vreplgr2vr.b vr5, a2
+ vslt.b vr5, vr2, vr5
+ vorn.v vr4, vr4, vr5
+
+ b L(un_end)
+L(ret0):
+ move a0, zero
+ jr ra
+END(STRNCMP)
+
+libc_hidden_builtin_def (STRNCMP)
+#endif
diff --git a/sysdeps/loongarch/lp64/multiarch/strncmp.c b/sysdeps/loongarch/lp64/multiarch/strncmp.c
new file mode 100644
index 0000000000..af6d0bc4a7
--- /dev/null
+++ b/sysdeps/loongarch/lp64/multiarch/strncmp.c
@@ -0,0 +1,35 @@
+/* Multiple versions of strncmp.
+ All versions must be listed in ifunc-impl-list.c.
+ Copyright (C) 2023 Free Software Foundation, Inc.
+ This file is part of the GNU C Library.
+
+ The GNU C Library is free software; you can redistribute it and/or
+ modify it under the terms of the GNU Lesser General Public
+ License as published by the Free Software Foundation; either
+ version 2.1 of the License, or (at your option) any later version.
+
+ The GNU C Library is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ Lesser General Public License for more details.
+
+ You should have received a copy of the GNU Lesser General Public
+ License along with the GNU C Library; if not, see
+ <http://www.gnu.org/licenses/>. */
+
+/* Define multiple versions only for the definition in libc. */
+#if IS_IN (libc)
+# define strncmp __redirect_strncmp
+# include <string.h>
+# undef strncmp
+
+# define SYMBOL_NAME strncmp
+# include "ifunc-strncmp.h"
+
+libc_ifunc_redirected (__redirect_strncmp, strncmp, IFUNC_SELECTOR ());
+
+# ifdef SHARED
+__hidden_ver1 (strncmp, __GI_strncmp, __redirect_strncmp)
+ __attribute__ ((visibility ("hidden"))) __attribute_copy__ (strncmp);
+# endif
+#endif
--
2.40.0
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH 3/3] Loongarch: Add ifunc support for strncmp{aligned, lsx}
2023-08-22 2:11 ` [PATCH 3/3] Loongarch: Add ifunc support for strncmp{aligned, lsx} dengjianbo
@ 2023-08-22 3:56 ` Richard Henderson
2023-08-22 6:37 ` dengjianbo
0 siblings, 1 reply; 9+ messages in thread
From: Richard Henderson @ 2023-08-22 3:56 UTC (permalink / raw)
To: dengjianbo, libc-alpha
Cc: adhemerval.zanella, xry111, caiyinyu, xuchenghua, huangpei
On 8/21/23 19:11, dengjianbo wrote:
> +L(magic_num):
> + .align 6
> + .dword 0x0706050403020100
> + .dword 0x0f0e0d0c0b0a0908
> +ENTRY_NO_ALIGN(STRNCMP)
> + beqz a2, L(ret0)
> + pcaddi t0, -5
> + andi a3, a0, 0xf
> + vld vr2, t0, 0
Why is the data not in .rodata or a mergable constant section?
You can use pcalau12i and %pc_lo12 with vld to place this data anywhere.
r~
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH 3/3] Loongarch: Add ifunc support for strncmp{aligned, lsx}
2023-08-22 3:56 ` Richard Henderson
@ 2023-08-22 6:37 ` dengjianbo
2023-08-22 11:13 ` Xi Ruoyao
0 siblings, 1 reply; 9+ messages in thread
From: dengjianbo @ 2023-08-22 6:37 UTC (permalink / raw)
To: Richard Henderson, libc-alpha
Cc: adhemerval.zanella, xry111, caiyinyu, xuchenghua, huangpei
[-- Attachment #1: Type: text/plain, Size: 742 bytes --]
On 2023-08-22 11:56, Richard Henderson wrote:
>> +L(magic_num):
>> + .align 6
>> + .dword 0x0706050403020100
>> + .dword 0x0f0e0d0c0b0a0908
>> +ENTRY_NO_ALIGN(STRNCMP)
>> + beqz a2, L(ret0)
>> + pcaddi t0, -5
>> + andi a3, a0, 0xf
>> + vld vr2, t0, 0
>
> Why is the data not in .rodata or a mergable constant section?
>
> You can use pcalau12i and %pc_lo12 with vld to place this data anywhere.
>
>
> r~
Putting the data here is due to the performance. When the vld
instruction is executed, the data will be in the cache, it can
speed up the data loading.
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH 3/3] Loongarch: Add ifunc support for strncmp{aligned, lsx}
2023-08-22 6:37 ` dengjianbo
@ 2023-08-22 11:13 ` Xi Ruoyao
2023-08-22 11:23 ` Xi Ruoyao
0 siblings, 1 reply; 9+ messages in thread
From: Xi Ruoyao @ 2023-08-22 11:13 UTC (permalink / raw)
To: dengjianbo, Richard Henderson, libc-alpha
Cc: adhemerval.zanella, caiyinyu, xuchenghua, huangpei
On Tue, 2023-08-22 at 14:37 +0800, dengjianbo wrote:
> Putting the data here is due to the performance. When the vld
> instruction is executed, the data will be in the cache, it can
> speed up the data loading.
AFAIK LoongArch CPUs have separate icache and dcache like all modern
CPUs, so this is not valid to me.
--
Xi Ruoyao <xry111@xry111.site>
School of Aerospace Science and Technology, Xidian University
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH 3/3] Loongarch: Add ifunc support for strncmp{aligned, lsx}
2023-08-22 11:13 ` Xi Ruoyao
@ 2023-08-22 11:23 ` Xi Ruoyao
2023-08-23 7:25 ` dengjianbo
0 siblings, 1 reply; 9+ messages in thread
From: Xi Ruoyao @ 2023-08-22 11:23 UTC (permalink / raw)
To: dengjianbo, Richard Henderson, libc-alpha
Cc: adhemerval.zanella, caiyinyu, xuchenghua, huangpei
On Tue, 2023-08-22 at 19:13 +0800, Xi Ruoyao via Libc-alpha wrote:
> On Tue, 2023-08-22 at 14:37 +0800, dengjianbo wrote:
>
> > Putting the data here is due to the performance. When the vld
> > instruction is executed, the data will be in the cache, it can
> > speed up the data loading.
>
> AFAIK LoongArch CPUs have separate icache and dcache like all modern
> CPUs, so this is not valid to me.
And even if it can really improve the performance, this is not on the
hot path of the algorithm so we should not use bizarre optimizations
here for marginal improvement.
--
Xi Ruoyao <xry111@xry111.site>
School of Aerospace Science and Technology, Xidian University
^ permalink raw reply [flat|nested] 9+ messages in thread
* Re: [PATCH 3/3] Loongarch: Add ifunc support for strncmp{aligned, lsx}
2023-08-22 11:23 ` Xi Ruoyao
@ 2023-08-23 7:25 ` dengjianbo
0 siblings, 0 replies; 9+ messages in thread
From: dengjianbo @ 2023-08-23 7:25 UTC (permalink / raw)
To: Xi Ruoyao, Richard Henderson, libc-alpha
Cc: adhemerval.zanella, caiyinyu, xuchenghua, huangpei
[-- Attachment #1: Type: text/plain, Size: 2170 bytes --]
On 2023-08-22 19:23, Xi Ruoyao wrote:
> On Tue, 2023-08-22 at 19:13 +0800, Xi Ruoyao via Libc-alpha wrote:
>> On Tue, 2023-08-22 at 14:37 +0800, dengjianbo wrote:
>>
>>> Putting the data here is due to the performance. When the vld
>>> instruction is executed, the data will be in the cache, it can
>>> speed up the data loading.
>> AFAIK LoongArch CPUs have separate icache and dcache like all modern
>> CPUs, so this is not valid to me.
> And even if it can really improve the performance, this is not on the
> hot path of the algorithm so we should not use bizarre optimizations
> here for marginal improvement.
>
> -- Xi Ruoyao <xry111@xry111.site> School of Aerospace Science and Technology, Xidian University
Thanks for your suggestion. We have changed strcmp and strncmp to put
the data in the rodata section with mergeable flags, and also use pcalau12i
and %pc_lo12 with the vld to get the data.
diff --git a/sysdeps/loongarch/lp64/multiarch/strncmp-lsx.S b/sysdeps/loongarch/lp64/multiarch/strncmp-lsx.S
index 595472fcda..0b4eee2a98 100644
--- a/sysdeps/loongarch/lp64/multiarch/strncmp-lsx.S
+++ b/sysdeps/loongarch/lp64/multiarch/strncmp-lsx.S
@@ -25,15 +25,11 @@
# define STRNCMP __strncmp_lsx
-L(magic_num):
- .align 6
- .dword 0x0706050403020100
- .dword 0x0f0e0d0c0b0a0908
-ENTRY_NO_ALIGN(STRNCMP)
+LEAF(STRNCMP, 6)
beqz a2, L(ret0)
- pcaddi t0, -5
+ pcalau12i t0, %pc_hi20(L(INDEX))
andi a3, a0, 0xf
- vld vr2, t0, 0
+ vld vr2, t0, %pc_lo12(L(INDEX))
andi a4, a1, 0xf
li.d t2, 16
@@ -202,5 +198,11 @@ L(ret0):
jr ra
END(STRNCMP)
+ .section .rodata.cst16,"M",@progbits,16
+ .align 4
+L(INDEX):
+ .dword 0x0706050403020100
+ .dword 0x0f0e0d0c0b0a0908
+
libc_hidden_builtin_def (STRNCMP)
#endif
^ permalink raw reply [flat|nested] 9+ messages in thread
end of thread, other threads:[~2023-08-23 7:25 UTC | newest]
Thread overview: 9+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-08-22 2:11 [PATCH 0/3] Add ifunc support for str{nlen, cmp, ncmp} dengjianbo
2023-08-22 2:11 ` [PATCH 1/3] Loongarch: Add ifunc support for strnlen{aligned, lsx, lasx} dengjianbo
2023-08-22 2:11 ` [PATCH 2/3] Loongarch: Add ifunc support for strcmp{aligned, lsx} dengjianbo
2023-08-22 2:11 ` [PATCH 3/3] Loongarch: Add ifunc support for strncmp{aligned, lsx} dengjianbo
2023-08-22 3:56 ` Richard Henderson
2023-08-22 6:37 ` dengjianbo
2023-08-22 11:13 ` Xi Ruoyao
2023-08-22 11:23 ` Xi Ruoyao
2023-08-23 7:25 ` dengjianbo
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).