|
 |
00db10 |
From 71ae86478edc7b21872464f43fb29ff650c1681a Mon Sep 17 00:00:00 2001
|
|
|
00db10 |
From: Adhemerval Zanella <azanella@linux.vnet.ibm.com>
|
|
|
00db10 |
Date: Tue, 15 Jul 2014 12:19:09 -0400
|
|
|
00db10 |
Subject: [PATCH] PowerPC: memset optimization for POWER8/PPC64
|
|
|
00db10 |
|
|
|
00db10 |
This patch adds an optimized memset implementation for POWER8. For
|
|
|
00db10 |
sizes from 0 to 255 bytes, a word/doubleword algorithm similar to
|
|
|
00db10 |
POWER7 optimized one is used.
|
|
|
00db10 |
|
|
|
00db10 |
For size higher than 255 two strategies are used:
|
|
|
00db10 |
|
|
|
00db10 |
1. If the constant is different than 0, the memory is written with
|
|
|
00db10 |
altivec vector instruction;
|
|
|
00db10 |
|
|
|
00db10 |
2. If constant is 0, dbcz instructions are used. The loop is unrolled
|
|
|
00db10 |
to clear 512 byte at time.
|
|
|
00db10 |
|
|
|
00db10 |
Using vector instructions increases throughput considerable, with a
|
|
|
00db10 |
double performance for sizes larger than 1024. The dcbz loops unrolls
|
|
|
00db10 |
also shows performance improvement, by doubling throughput for sizes
|
|
|
00db10 |
larger than 8192 bytes.
|
|
|
00db10 |
---
|
|
|
00db10 |
ChangeLog | 15 +
|
|
|
00db10 |
benchtests/bench-memset.c | 5 +
|
|
|
00db10 |
sysdeps/powerpc/powerpc64/multiarch/Makefile | 2 +-
|
|
|
00db10 |
sysdeps/powerpc/powerpc64/multiarch/bzero.c | 11 +-
|
|
|
00db10 |
.../powerpc/powerpc64/multiarch/ifunc-impl-list.c | 6 +
|
|
|
00db10 |
.../powerpc/powerpc64/multiarch/memset-power8.S | 43 ++
|
|
|
00db10 |
sysdeps/powerpc/powerpc64/multiarch/memset.c | 11 +-
|
|
|
00db10 |
sysdeps/powerpc/powerpc64/power8/memset.S | 449 +++++++++++++++++++++
|
|
|
00db10 |
8 files changed, 533 insertions(+), 9 deletions(-)
|
|
|
00db10 |
create mode 100644 sysdeps/powerpc/powerpc64/multiarch/memset-power8.S
|
|
|
00db10 |
create mode 100644 sysdeps/powerpc/powerpc64/power8/memset.S
|
|
|
00db10 |
|
|
|
00db10 |
diff --git a/ChangeLog b/ChangeLog
|
|
|
00db10 |
index ddaf70f..dc61c87 100644
|
|
|
00db10 |
diff --git a/sysdeps/powerpc/powerpc64/multiarch/Makefile b/sysdeps/powerpc/powerpc64/multiarch/Makefile
|
|
|
00db10 |
index 0de3804..abc9d2e 100644
|
|
|
00db10 |
--- a/sysdeps/powerpc/powerpc64/multiarch/Makefile
|
|
|
00db10 |
+++ b/sysdeps/powerpc/powerpc64/multiarch/Makefile
|
|
|
00db10 |
@@ -1,7 +1,8 @@ ifeq ($(subdir),string)
|
|
|
00db10 |
ifeq ($(subdir),string)
|
|
|
00db10 |
sysdep_routines += memcpy-power7 memcpy-a2 memcpy-power6 memcpy-cell \
|
|
|
00db10 |
memcpy-power4 memcpy-ppc64 memcmp-power7 memcmp-power4 \
|
|
|
00db10 |
- memcmp-ppc64 memset-power7 memset-power6 memset-power4 \
|
|
|
00db10 |
+ memcmp-ppc64 memset-power8 memset-power7 memset-power6 \
|
|
|
00db10 |
+ memset-power4 \
|
|
|
00db10 |
memset-ppc64 bzero-power4 bzero-power6 bzero-power7 \
|
|
|
00db10 |
mempcpy-power7 mempcpy-ppc64 memchr-power7 memchr-ppc64 \
|
|
|
00db10 |
memrchr-power7 memrchr-ppc64 rawmemchr-power7 \
|
|
|
00db10 |
diff --git a/sysdeps/powerpc/powerpc64/multiarch/bzero.c b/sysdeps/powerpc/powerpc64/multiarch/bzero.c
|
|
|
00db10 |
index ed83541..298cf00 100644
|
|
|
00db10 |
--- a/sysdeps/powerpc/powerpc64/multiarch/bzero.c
|
|
|
00db10 |
+++ b/sysdeps/powerpc/powerpc64/multiarch/bzero.c
|
|
|
00db10 |
@@ -26,14 +26,17 @@ extern __typeof (bzero) __bzero_ppc attribute_hidden;
|
|
|
00db10 |
extern __typeof (bzero) __bzero_power4 attribute_hidden;
|
|
|
00db10 |
extern __typeof (bzero) __bzero_power6 attribute_hidden;
|
|
|
00db10 |
extern __typeof (bzero) __bzero_power7 attribute_hidden;
|
|
|
00db10 |
+extern __typeof (bzero) __bzero_power8 attribute_hidden;
|
|
|
00db10 |
|
|
|
00db10 |
libc_ifunc (__bzero,
|
|
|
00db10 |
- (hwcap & PPC_FEATURE_HAS_VSX)
|
|
|
00db10 |
- ? __bzero_power7 :
|
|
|
00db10 |
- (hwcap & PPC_FEATURE_ARCH_2_05)
|
|
|
00db10 |
+ (hwcap2 & PPC_FEATURE2_ARCH_2_07)
|
|
|
00db10 |
+ ? __bzero_power8 :
|
|
|
00db10 |
+ (hwcap & PPC_FEATURE_HAS_VSX)
|
|
|
00db10 |
+ ? __bzero_power7 :
|
|
|
00db10 |
+ (hwcap & PPC_FEATURE_ARCH_2_05)
|
|
|
00db10 |
? __bzero_power6 :
|
|
|
00db10 |
(hwcap & PPC_FEATURE_POWER4)
|
|
|
00db10 |
- ? __bzero_power4
|
|
|
00db10 |
+ ? __bzero_power4
|
|
|
00db10 |
: __bzero_ppc);
|
|
|
00db10 |
|
|
|
00db10 |
weak_alias (__bzero, bzero)
|
|
|
00db10 |
diff --git a/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c b/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
|
|
|
00db10 |
index a574487..06d5be9 100644
|
|
|
00db10 |
--- a/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
|
|
|
00db10 |
+++ b/sysdeps/powerpc/powerpc64/multiarch/ifunc-impl-list.c
|
|
|
00db10 |
@@ -69,6 +71,8 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
|
|
|
00db10 |
|
|
|
00db10 |
/* Support sysdeps/powerpc/powerpc64/multiarch/memset.c. */
|
|
|
00db10 |
IFUNC_IMPL (i, name, memset,
|
|
|
00db10 |
+ IFUNC_IMPL_ADD (array, i, memset, hwcap2 & PPC_FEATURE2_ARCH_2_07,
|
|
|
00db10 |
+ __memset_power8)
|
|
|
00db10 |
IFUNC_IMPL_ADD (array, i, memset, hwcap & PPC_FEATURE_HAS_VSX,
|
|
|
00db10 |
__memset_power7)
|
|
|
00db10 |
IFUNC_IMPL_ADD (array, i, memset, hwcap & PPC_FEATURE_ARCH_2_05,
|
|
|
00db10 |
@@ -134,6 +138,8 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
|
|
|
00db10 |
|
|
|
00db10 |
/* Support sysdeps/powerpc/powerpc64/multiarch/bzero.c. */
|
|
|
00db10 |
IFUNC_IMPL (i, name, bzero,
|
|
|
00db10 |
+ IFUNC_IMPL_ADD (array, i, bzero, hwcap2 & PPC_FEATURE2_ARCH_2_07,
|
|
|
00db10 |
+ __bzero_power8)
|
|
|
00db10 |
IFUNC_IMPL_ADD (array, i, bzero, hwcap & PPC_FEATURE_HAS_VSX,
|
|
|
00db10 |
__bzero_power7)
|
|
|
00db10 |
IFUNC_IMPL_ADD (array, i, bzero, hwcap & PPC_FEATURE_ARCH_2_05,
|
|
|
00db10 |
diff --git a/sysdeps/powerpc/powerpc64/multiarch/memset-power8.S b/sysdeps/powerpc/powerpc64/multiarch/memset-power8.S
|
|
|
00db10 |
new file mode 100644
|
|
|
00db10 |
index 0000000..e8a604b
|
|
|
00db10 |
--- /dev/null
|
|
|
00db10 |
+++ b/sysdeps/powerpc/powerpc64/multiarch/memset-power8.S
|
|
|
00db10 |
@@ -0,0 +1,43 @@
|
|
|
00db10 |
+/* Optimized memset implementation for PowerPC64/POWER8.
|
|
|
00db10 |
+ Copyright (C) 2014 Free Software Foundation, Inc.
|
|
|
00db10 |
+ This file is part of the GNU C Library.
|
|
|
00db10 |
+
|
|
|
00db10 |
+ The GNU C Library is free software; you can redistribute it and/or
|
|
|
00db10 |
+ modify it under the terms of the GNU Lesser General Public
|
|
|
00db10 |
+ License as published by the Free Software Foundation; either
|
|
|
00db10 |
+ version 2.1 of the License, or (at your option) any later version.
|
|
|
00db10 |
+
|
|
|
00db10 |
+ The GNU C Library is distributed in the hope that it will be useful,
|
|
|
00db10 |
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
00db10 |
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
00db10 |
+ Lesser General Public License for more details.
|
|
|
00db10 |
+
|
|
|
00db10 |
+ You should have received a copy of the GNU Lesser General Public
|
|
|
00db10 |
+ License along with the GNU C Library; if not, see
|
|
|
00db10 |
+ <http://www.gnu.org/licenses/>. */
|
|
|
00db10 |
+
|
|
|
00db10 |
+#include <sysdep.h>
|
|
|
00db10 |
+
|
|
|
00db10 |
+#undef EALIGN
|
|
|
00db10 |
+#define EALIGN(name, alignt, words) \
|
|
|
00db10 |
+ .section ".text"; \
|
|
|
00db10 |
+ ENTRY_2(__memset_power8) \
|
|
|
00db10 |
+ .align ALIGNARG(alignt); \
|
|
|
00db10 |
+ EALIGN_W_##words; \
|
|
|
00db10 |
+ BODY_LABEL(__memset_power8): \
|
|
|
00db10 |
+ cfi_startproc; \
|
|
|
00db10 |
+ LOCALENTRY(__memset_power8)
|
|
|
00db10 |
+
|
|
|
00db10 |
+#undef END_GEN_TB
|
|
|
00db10 |
+#define END_GEN_TB(name, mask) \
|
|
|
00db10 |
+ cfi_endproc; \
|
|
|
00db10 |
+ TRACEBACK_MASK(__memset_power8,mask) \
|
|
|
00db10 |
+ END_2(__memset_power8)
|
|
|
00db10 |
+
|
|
|
00db10 |
+#undef libc_hidden_builtin_def
|
|
|
00db10 |
+#define libc_hidden_builtin_def(name)
|
|
|
00db10 |
+
|
|
|
00db10 |
+#undef __bzero
|
|
|
00db10 |
+#define __bzero __bzero_power8
|
|
|
00db10 |
+
|
|
|
00db10 |
+#include <sysdeps/powerpc/powerpc64/power8/memset.S>
|
|
|
00db10 |
diff --git a/sysdeps/powerpc/powerpc64/multiarch/memset.c b/sysdeps/powerpc/powerpc64/multiarch/memset.c
|
|
|
00db10 |
index aa2ae70..9c7ed10 100644
|
|
|
00db10 |
--- a/sysdeps/powerpc/powerpc64/multiarch/memset.c
|
|
|
00db10 |
+++ b/sysdeps/powerpc/powerpc64/multiarch/memset.c
|
|
|
00db10 |
@@ -32,16 +32,19 @@ extern __typeof (__redirect_memset) __memset_ppc attribute_hidden;
|
|
|
00db10 |
extern __typeof (__redirect_memset) __memset_power4 attribute_hidden;
|
|
|
00db10 |
extern __typeof (__redirect_memset) __memset_power6 attribute_hidden;
|
|
|
00db10 |
extern __typeof (__redirect_memset) __memset_power7 attribute_hidden;
|
|
|
00db10 |
+extern __typeof (__redirect_memset) __memset_power8 attribute_hidden;
|
|
|
00db10 |
|
|
|
00db10 |
/* Avoid DWARF definition DIE on ifunc symbol so that GDB can handle
|
|
|
00db10 |
ifunc symbol properly. */
|
|
|
00db10 |
libc_ifunc (__libc_memset,
|
|
|
00db10 |
- (hwcap & PPC_FEATURE_HAS_VSX)
|
|
|
00db10 |
- ? __memset_power7 :
|
|
|
00db10 |
- (hwcap & PPC_FEATURE_ARCH_2_05)
|
|
|
00db10 |
+ (hwcap2 & PPC_FEATURE2_ARCH_2_07)
|
|
|
00db10 |
+ ? __memset_power8 :
|
|
|
00db10 |
+ (hwcap & PPC_FEATURE_HAS_VSX)
|
|
|
00db10 |
+ ? __memset_power7 :
|
|
|
00db10 |
+ (hwcap & PPC_FEATURE_ARCH_2_05)
|
|
|
00db10 |
? __memset_power6 :
|
|
|
00db10 |
(hwcap & PPC_FEATURE_POWER4)
|
|
|
00db10 |
- ? __memset_power4
|
|
|
00db10 |
+ ? __memset_power4
|
|
|
00db10 |
: __memset_ppc);
|
|
|
00db10 |
|
|
|
00db10 |
#undef memset
|
|
|
00db10 |
diff --git a/sysdeps/powerpc/powerpc64/power8/memset.S b/sysdeps/powerpc/powerpc64/power8/memset.S
|
|
|
00db10 |
new file mode 100644
|
|
|
00db10 |
index 0000000..191a4df
|
|
|
00db10 |
--- /dev/null
|
|
|
00db10 |
+++ b/sysdeps/powerpc/powerpc64/power8/memset.S
|
|
|
00db10 |
@@ -0,0 +1,449 @@
|
|
|
00db10 |
+/* Optimized memset implementation for PowerPC64/POWER8.
|
|
|
00db10 |
+ Copyright (C) 2014 Free Software Foundation, Inc.
|
|
|
00db10 |
+ This file is part of the GNU C Library.
|
|
|
00db10 |
+
|
|
|
00db10 |
+ The GNU C Library is free software; you can redistribute it and/or
|
|
|
00db10 |
+ modify it under the terms of the GNU Lesser General Public
|
|
|
00db10 |
+ License as published by the Free Software Foundation; either
|
|
|
00db10 |
+ version 2.1 of the License, or (at your option) any later version.
|
|
|
00db10 |
+
|
|
|
00db10 |
+ The GNU C Library is distributed in the hope that it will be useful,
|
|
|
00db10 |
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
|
00db10 |
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
|
00db10 |
+ Lesser General Public License for more details.
|
|
|
00db10 |
+
|
|
|
00db10 |
+ You should have received a copy of the GNU Lesser General Public
|
|
|
00db10 |
+ License along with the GNU C Library; if not, see
|
|
|
00db10 |
+ <http://www.gnu.org/licenses/>. */
|
|
|
00db10 |
+
|
|
|
00db10 |
+#include <sysdep.h>
|
|
|
00db10 |
+
|
|
|
00db10 |
+/* __ptr_t [r3] memset (__ptr_t s [r3], int c [r4], size_t n [r5]));
|
|
|
00db10 |
+ Returns 's'. */
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .machine power8
|
|
|
00db10 |
+EALIGN (memset, 5, 0)
|
|
|
00db10 |
+ CALL_MCOUNT 3
|
|
|
00db10 |
+
|
|
|
00db10 |
+L(_memset):
|
|
|
00db10 |
+ cmpldi cr7,r5,31
|
|
|
00db10 |
+ neg r0,r3
|
|
|
00db10 |
+ mr r10,r3
|
|
|
00db10 |
+
|
|
|
00db10 |
+ insrdi r4,r4,8,48
|
|
|
00db10 |
+ insrdi r4,r4,16,32 /* Replicate byte to word. */
|
|
|
00db10 |
+ ble cr7,L(write_LT_32)
|
|
|
00db10 |
+
|
|
|
00db10 |
+ andi. r11,r10,15 /* Check alignment of DST. */
|
|
|
00db10 |
+ insrdi r4,r4,32,0 /* Replicate word to double word. */
|
|
|
00db10 |
+
|
|
|
00db10 |
+ beq L(big_aligned)
|
|
|
00db10 |
+
|
|
|
00db10 |
+ mtocrf 0x01,r0
|
|
|
00db10 |
+ clrldi r0,r0,60
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Get DST aligned to 16 bytes. */
|
|
|
00db10 |
+1: bf 31,2f
|
|
|
00db10 |
+ stb r4,0(r10)
|
|
|
00db10 |
+ addi r10,r10,1
|
|
|
00db10 |
+
|
|
|
00db10 |
+2: bf 30,4f
|
|
|
00db10 |
+ sth r4,0(r10)
|
|
|
00db10 |
+ addi r10,r10,2
|
|
|
00db10 |
+
|
|
|
00db10 |
+4: bf 29,8f
|
|
|
00db10 |
+ stw r4,0(r10)
|
|
|
00db10 |
+ addi r10,r10,4
|
|
|
00db10 |
+
|
|
|
00db10 |
+8: bf 28,16f
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ addi r10,r10,8
|
|
|
00db10 |
+
|
|
|
00db10 |
+16: subf r5,r0,r5
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+L(big_aligned):
|
|
|
00db10 |
+ /* For sizes larger than 255 two possible paths:
|
|
|
00db10 |
+ - if constant is '0', zero full cache lines with dcbz
|
|
|
00db10 |
+ - otherwise uses vector instructions. */
|
|
|
00db10 |
+ cmpldi cr5,r5,255
|
|
|
00db10 |
+ dcbtst 0,r10
|
|
|
00db10 |
+ cmpldi cr6,r4,0
|
|
|
00db10 |
+ crand 27,26,21
|
|
|
00db10 |
+ bt 27,L(huge_dcbz)
|
|
|
00db10 |
+ bge cr5,L(huge_vector)
|
|
|
00db10 |
+
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Size between 32 and 255 bytes with constant different than 0, use
|
|
|
00db10 |
+ doubleword store instruction to achieve best throughput. */
|
|
|
00db10 |
+ srdi r8,r5,5
|
|
|
00db10 |
+ clrldi r11,r5,59
|
|
|
00db10 |
+ cmpldi cr6,r11,0
|
|
|
00db10 |
+ cmpdi r8,0
|
|
|
00db10 |
+ beq L(tail_bytes)
|
|
|
00db10 |
+ mtctr r8
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Main aligned write loop, writes 32-bytes at a time. */
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+L(big_loop):
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ std r4,8(r10)
|
|
|
00db10 |
+ std r4,16(r10)
|
|
|
00db10 |
+ std r4,24(r10)
|
|
|
00db10 |
+ addi r10,r10,32
|
|
|
00db10 |
+ bdz L(tail_bytes)
|
|
|
00db10 |
+
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ std r4,8(r10)
|
|
|
00db10 |
+ std r4,16(r10)
|
|
|
00db10 |
+ std r4,24(r10)
|
|
|
00db10 |
+ addi r10,10,32
|
|
|
00db10 |
+ bdnz L(big_loop)
|
|
|
00db10 |
+
|
|
|
00db10 |
+ b L(tail_bytes)
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Write remaining 1~31 bytes. */
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+L(tail_bytes):
|
|
|
00db10 |
+ beqlr cr6
|
|
|
00db10 |
+
|
|
|
00db10 |
+ srdi r7,r11,4
|
|
|
00db10 |
+ clrldi r8,r11,60
|
|
|
00db10 |
+ mtocrf 0x01,r7
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+ bf 31,8f
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ std r4,8(r10)
|
|
|
00db10 |
+ addi r10,r10,16
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+8: mtocrf 0x1,r8
|
|
|
00db10 |
+ bf 28,4f
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ addi r10,r10,8
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+4: bf 29,2f
|
|
|
00db10 |
+ stw 4,0(10)
|
|
|
00db10 |
+ addi 10,10,4
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+2: bf 30,1f
|
|
|
00db10 |
+ sth 4,0(10)
|
|
|
00db10 |
+ addi 10,10,2
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+1: bflr 31
|
|
|
00db10 |
+ stb 4,0(10)
|
|
|
00db10 |
+ blr
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Size larger than 255 bytes with constant different than 0, use
|
|
|
00db10 |
+ vector instruction to achieve best throughput. */
|
|
|
00db10 |
+L(huge_vector):
|
|
|
00db10 |
+ /* Replicate set byte to quadword in VMX register. */
|
|
|
00db10 |
+ mtvsrd v1,r4
|
|
|
00db10 |
+ xxpermdi 32,v0,v1,0
|
|
|
00db10 |
+ vspltb v2,v0,15
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Main aligned write loop: 128 bytes at a time. */
|
|
|
00db10 |
+ li r6,16
|
|
|
00db10 |
+ li r7,32
|
|
|
00db10 |
+ li r8,48
|
|
|
00db10 |
+ mtocrf 0x02,r5
|
|
|
00db10 |
+ srdi r12,r5,7
|
|
|
00db10 |
+ cmpdi r12,0
|
|
|
00db10 |
+ beq L(aligned_tail)
|
|
|
00db10 |
+ mtctr r12
|
|
|
00db10 |
+ b L(aligned_128loop)
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+L(aligned_128loop):
|
|
|
00db10 |
+ stvx v2,0,r10
|
|
|
00db10 |
+ stvx v2,r10,r6
|
|
|
00db10 |
+ stvx v2,r10,r7
|
|
|
00db10 |
+ stvx v2,r10,r8
|
|
|
00db10 |
+ addi r10,r10,64
|
|
|
00db10 |
+ stvx v2,0,r10
|
|
|
00db10 |
+ stvx v2,r10,r6
|
|
|
00db10 |
+ stvx v2,r10,r7
|
|
|
00db10 |
+ stvx v2,r10,r8
|
|
|
00db10 |
+ addi r10,r10,64
|
|
|
00db10 |
+ bdnz L(aligned_128loop)
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Write remaining 1~127 bytes. */
|
|
|
00db10 |
+L(aligned_tail):
|
|
|
00db10 |
+ mtocrf 0x01,r5
|
|
|
00db10 |
+ bf 25,32f
|
|
|
00db10 |
+ stvx v2,0,r10
|
|
|
00db10 |
+ stvx v2,r10,r6
|
|
|
00db10 |
+ stvx v2,r10,r7
|
|
|
00db10 |
+ stvx v2,r10,r8
|
|
|
00db10 |
+ addi r10,r10,64
|
|
|
00db10 |
+
|
|
|
00db10 |
+32: bf 26,16f
|
|
|
00db10 |
+ stvx v2,0,r10
|
|
|
00db10 |
+ stvx v2,r10,r6
|
|
|
00db10 |
+ addi r10,r10,32
|
|
|
00db10 |
+
|
|
|
00db10 |
+16: bf 27,8f
|
|
|
00db10 |
+ stvx v2,0,r10
|
|
|
00db10 |
+ addi r10,r10,16
|
|
|
00db10 |
+
|
|
|
00db10 |
+8: bf 28,4f
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ addi r10,r10,8
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Copies 4~7 bytes. */
|
|
|
00db10 |
+4: bf 29,L(tail2)
|
|
|
00db10 |
+ stw r4,0(r10)
|
|
|
00db10 |
+ bf 30,L(tail5)
|
|
|
00db10 |
+ sth r4,4(r10)
|
|
|
00db10 |
+ bflr 31
|
|
|
00db10 |
+ stb r4,6(r10)
|
|
|
00db10 |
+ /* Return original DST pointer. */
|
|
|
00db10 |
+ blr
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Special case when value is 0 and we have a long length to deal
|
|
|
00db10 |
+ with. Use dcbz to zero out a full cacheline of 128 bytes at a time.
|
|
|
00db10 |
+ Before using dcbz though, we need to get the destination 128-byte
|
|
|
00db10 |
+ aligned. */
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+L(huge_dcbz):
|
|
|
00db10 |
+ andi. r11,r10,127
|
|
|
00db10 |
+ neg r0,r10
|
|
|
00db10 |
+ beq L(huge_dcbz_aligned)
|
|
|
00db10 |
+
|
|
|
00db10 |
+ clrldi r0,r0,57
|
|
|
00db10 |
+ subf r5,r0,r5
|
|
|
00db10 |
+ srdi r0,r0,3
|
|
|
00db10 |
+ mtocrf 0x01,r0
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Write 1~128 bytes until DST is aligned to 128 bytes. */
|
|
|
00db10 |
+8: bf 28,4f
|
|
|
00db10 |
+
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ std r4,8(r10)
|
|
|
00db10 |
+ std r4,16(r10)
|
|
|
00db10 |
+ std r4,24(r10)
|
|
|
00db10 |
+ std r4,32(r10)
|
|
|
00db10 |
+ std r4,40(r10)
|
|
|
00db10 |
+ std r4,48(r10)
|
|
|
00db10 |
+ std r4,56(r10)
|
|
|
00db10 |
+ addi r10,r10,64
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+4: bf 29,2f
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ std r4,8(r10)
|
|
|
00db10 |
+ std r4,16(r10)
|
|
|
00db10 |
+ std r4,24(r10)
|
|
|
00db10 |
+ addi r10,r10,32
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+2: bf 30,1f
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ std r4,8(r10)
|
|
|
00db10 |
+ addi r10,r10,16
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+1: bf 31,L(huge_dcbz_aligned)
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ addi r10,r10,8
|
|
|
00db10 |
+
|
|
|
00db10 |
+L(huge_dcbz_aligned):
|
|
|
00db10 |
+ /* Setup dcbz unroll offsets and count numbers. */
|
|
|
00db10 |
+ srdi r8,r5,9
|
|
|
00db10 |
+ clrldi r11,r5,55
|
|
|
00db10 |
+ cmpldi cr6,r11,0
|
|
|
00db10 |
+ li r9,128
|
|
|
00db10 |
+ cmpdi r8,0
|
|
|
00db10 |
+ beq L(huge_tail)
|
|
|
00db10 |
+ li r7,256
|
|
|
00db10 |
+ li r6,384
|
|
|
00db10 |
+ mtctr r8
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+L(huge_loop):
|
|
|
00db10 |
+ /* Sets 512 bytes to zero in each iteration, the loop unrolling shows
|
|
|
00db10 |
+ a throughput boost for large sizes (2048 bytes or higher). */
|
|
|
00db10 |
+ dcbz 0,r10
|
|
|
00db10 |
+ dcbz r9,r10
|
|
|
00db10 |
+ dcbz r7,r10
|
|
|
00db10 |
+ dcbz r6,r10
|
|
|
00db10 |
+ addi r10,r10,512
|
|
|
00db10 |
+ bdnz L(huge_loop)
|
|
|
00db10 |
+
|
|
|
00db10 |
+ beqlr cr6
|
|
|
00db10 |
+
|
|
|
00db10 |
+L(huge_tail):
|
|
|
00db10 |
+ srdi r6,r11,8
|
|
|
00db10 |
+ srdi r7,r11,4
|
|
|
00db10 |
+ clrldi r8,r11,4
|
|
|
00db10 |
+ cmpldi cr6,r8,0
|
|
|
00db10 |
+ mtocrf 0x01,r6
|
|
|
00db10 |
+
|
|
|
00db10 |
+ beq cr6,L(tail)
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* We have 1~511 bytes remaining. */
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+32: bf 31,16f
|
|
|
00db10 |
+ dcbz 0,r10
|
|
|
00db10 |
+ dcbz r9,r10
|
|
|
00db10 |
+ addi r10,r10,256
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+16: mtocrf 0x01,r7
|
|
|
00db10 |
+ bf 28,8f
|
|
|
00db10 |
+ dcbz 0,r10
|
|
|
00db10 |
+ addi r10,r10,128
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+8: bf 29,4f
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ std r4,8(r10)
|
|
|
00db10 |
+ std r4,16(r10)
|
|
|
00db10 |
+ std r4,24(r10)
|
|
|
00db10 |
+ std r4,32(r10)
|
|
|
00db10 |
+ std r4,40(r10)
|
|
|
00db10 |
+ std r4,48(r10)
|
|
|
00db10 |
+ std r4,56(r10)
|
|
|
00db10 |
+ addi r10,r10,64
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+4: bf 30,2f
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ std r4,8(r10)
|
|
|
00db10 |
+ std r4,16(r10)
|
|
|
00db10 |
+ std r4,24(r10)
|
|
|
00db10 |
+ addi r10,r10,32
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+2: bf 31,L(tail)
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ std r4,8(r10)
|
|
|
00db10 |
+ addi r10,r10,16
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Remaining 1~15 bytes. */
|
|
|
00db10 |
+L(tail):
|
|
|
00db10 |
+ mtocrf 0x01,r8
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align
|
|
|
00db10 |
+8: bf 28,4f
|
|
|
00db10 |
+ std r4,0(r10)
|
|
|
00db10 |
+ addi r10,r10,8
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+4: bf 29,2f
|
|
|
00db10 |
+ stw r4,0(r10)
|
|
|
00db10 |
+ addi r10,r10,4
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+2: bf 30,1f
|
|
|
00db10 |
+ sth r4,0(r10)
|
|
|
00db10 |
+ addi r10,r10,2
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+1: bflr 31
|
|
|
00db10 |
+ stb r4,0(r10)
|
|
|
00db10 |
+ blr
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Handle short copies of 0~31 bytes. Best throughput is achieved
|
|
|
00db10 |
+ by just unrolling all operations. */
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+L(write_LT_32):
|
|
|
00db10 |
+ cmpldi cr6,5,8
|
|
|
00db10 |
+ mtocrf 0x01,r5
|
|
|
00db10 |
+ ble cr6,L(write_LE_8)
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* At least 9 bytes to go. */
|
|
|
00db10 |
+ neg r8,r4
|
|
|
00db10 |
+ andi. r0,r8,3
|
|
|
00db10 |
+ cmpldi cr1,r5,16
|
|
|
00db10 |
+ beq L(write_LT_32_aligned)
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Force 4-byte alignment for SRC. */
|
|
|
00db10 |
+ mtocrf 0x01,r0
|
|
|
00db10 |
+ subf r5,r0,r5
|
|
|
00db10 |
+
|
|
|
00db10 |
+2: bf 30,1f
|
|
|
00db10 |
+ sth r4,0(r10)
|
|
|
00db10 |
+ addi r10,r10,2
|
|
|
00db10 |
+
|
|
|
00db10 |
+1: bf 31,L(end_4bytes_alignment)
|
|
|
00db10 |
+ stb r4,0(r10)
|
|
|
00db10 |
+ addi r10,r10,1
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+L(end_4bytes_alignment):
|
|
|
00db10 |
+ cmpldi cr1,r5,16
|
|
|
00db10 |
+ mtocrf 0x01,r5
|
|
|
00db10 |
+
|
|
|
00db10 |
+L(write_LT_32_aligned):
|
|
|
00db10 |
+ blt cr1,8f
|
|
|
00db10 |
+
|
|
|
00db10 |
+ stw r4,0(r10)
|
|
|
00db10 |
+ stw r4,4(r10)
|
|
|
00db10 |
+ stw r4,8(r10)
|
|
|
00db10 |
+ stw r4,12(r10)
|
|
|
00db10 |
+ addi r10,r10,16
|
|
|
00db10 |
+
|
|
|
00db10 |
+8: bf 28,L(tail4)
|
|
|
00db10 |
+ stw r4,0(r10)
|
|
|
00db10 |
+ stw r4,4(r10)
|
|
|
00db10 |
+ addi r10,r10,8
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+ /* Copies 4~7 bytes. */
|
|
|
00db10 |
+L(tail4):
|
|
|
00db10 |
+ bf 29,L(tail2)
|
|
|
00db10 |
+ stw r4,0(r10)
|
|
|
00db10 |
+ bf 30,L(tail5)
|
|
|
00db10 |
+ sth r4,4(r10)
|
|
|
00db10 |
+ bflr 31
|
|
|
00db10 |
+ stb r4,6(r10)
|
|
|
00db10 |
+ blr
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+ /* Copies 2~3 bytes. */
|
|
|
00db10 |
+L(tail2):
|
|
|
00db10 |
+ bf 30,1f
|
|
|
00db10 |
+ sth r4,0(r10)
|
|
|
00db10 |
+ bflr 31
|
|
|
00db10 |
+ stb r4,2(r10)
|
|
|
00db10 |
+ blr
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+L(tail5):
|
|
|
00db10 |
+ bflr 31
|
|
|
00db10 |
+ stb r4,4(r10)
|
|
|
00db10 |
+ blr
|
|
|
00db10 |
+
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+1: bflr 31
|
|
|
00db10 |
+ stb r4,0(r10)
|
|
|
00db10 |
+ blr
|
|
|
00db10 |
+
|
|
|
00db10 |
+ /* Handles copies of 0~8 bytes. */
|
|
|
00db10 |
+ .align 4
|
|
|
00db10 |
+L(write_LE_8):
|
|
|
00db10 |
+ bne cr6,L(tail4)
|
|
|
00db10 |
+
|
|
|
00db10 |
+ stw r4,0(r10)
|
|
|
00db10 |
+ stw r4,4(r10)
|
|
|
00db10 |
+ blr
|
|
|
00db10 |
+END_GEN_TB (memset,TB_TOCLESS)
|
|
|
00db10 |
+libc_hidden_builtin_def (memset)
|
|
|
00db10 |
+
|
|
|
00db10 |
+/* Copied from bzero.S to prevent the linker from inserting a stub
|
|
|
00db10 |
+ between bzero and memset. */
|
|
|
00db10 |
+ENTRY (__bzero)
|
|
|
00db10 |
+ CALL_MCOUNT 3
|
|
|
00db10 |
+ mr r5,r4
|
|
|
00db10 |
+ li r4,0
|
|
|
00db10 |
+ b L(_memset)
|
|
|
00db10 |
+END (__bzero)
|
|
|
00db10 |
+#ifndef __bzero
|
|
|
00db10 |
+weak_alias (__bzero, bzero)
|
|
|
00db10 |
+#endif
|
|
|
00db10 |
--
|
|
|
00db10 |
2.1.0
|
|
|
00db10 |
|