This is the mail archive of the
glibc-cvs@sourceware.org
mailing list for the glibc project.
GNU C Library master sources branch master updated. glibc-2.17-167-ga0b1cd8
- From: mshawcroft at sourceware dot org
- To: glibc-cvs at sourceware dot org
- Date: 24 Jan 2013 14:33:55 -0000
- Subject: GNU C Library master sources branch master updated. glibc-2.17-167-ga0b1cd8
This is an automated email from the git hooks/post-receive script. It was
generated because a ref change was pushed to the repository containing
the project "GNU C Library master sources".
The branch, master has been updated
via a0b1cd8869ffe7d0a602d31a3629c8690987be59 (commit)
via 4499bb3e1b14c5121bcf7334d9b754d20768bb2b (commit)
from 8c53a12c8625f618d6ce8202861dc6f4f70940d3 (commit)
Those revisions listed above that are new to this repository have
not appeared on any other notification email; so we list those
revisions in full, below.
- Log -----------------------------------------------------------------
http://sources.redhat.com/git/gitweb.cgi?p=glibc.git;a=commitdiff;h=a0b1cd8869ffe7d0a602d31a3629c8690987be59
commit a0b1cd8869ffe7d0a602d31a3629c8690987be59
Author: Marcus Shawcroft <marcus.shawcroft@linaro.org>
Date: Wed Jan 23 16:33:14 2013 +0000
AArch64: Adding optimized strncmp implementation.
diff --git a/ports/ChangeLog.aarch64 b/ports/ChangeLog.aarch64
index ba4792f..5a42510 100644
--- a/ports/ChangeLog.aarch64
+++ b/ports/ChangeLog.aarch64
@@ -1,5 +1,9 @@
2013-01-23 Marcus Shawcroft <marcus.shawcroft@linaro.org>
+ * sysdeps/aarch64/strncmp.S: New file.
+
+2013-01-23 Marcus Shawcroft <marcus.shawcroft@linaro.org>
+
* sysdeps/aarch64/sysdep.h (ENTRY_ALIGN_AND_PAD): New.
* sysdeps/aarch64/strnlen.S: New file.
diff --git a/ports/sysdeps/aarch64/strncmp.S b/ports/sysdeps/aarch64/strncmp.S
new file mode 100644
index 0000000..341d5ce
--- /dev/null
+++ b/ports/sysdeps/aarch64/strncmp.S
@@ -0,0 +1,204 @@
+/* Copyright (C) 2013 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/>. */
+
+#include <sysdep.h>
+
+/* Assumptions:
+ *
+ * ARMv8-a, AArch64
+ */
+
+#define REP8_01 0x0101010101010101
+#define REP8_7f 0x7f7f7f7f7f7f7f7f
+#define REP8_80 0x8080808080808080
+
+/* Parameters and result. */
+#define src1 x0
+#define src2 x1
+#define limit x2
+#define result x0
+
+/* Internal variables. */
+#define data1 x3
+#define data1w w3
+#define data2 x4
+#define data2w w4
+#define has_nul x5
+#define diff x6
+#define syndrome x7
+#define tmp1 x8
+#define tmp2 x9
+#define tmp3 x10
+#define zeroones x11
+#define pos x12
+#define limit_wd x13
+#define mask x14
+#define endloop x15
+
+ENTRY_ALIGN_AND_PAD (strncmp, 6, 7)
+ cbz limit, L(ret0)
+ eor tmp1, src1, src2
+ mov zeroones, #REP8_01
+ tst tmp1, #7
+ b.ne L(misaligned8)
+ ands tmp1, src1, #7
+ b.ne L(mutual_align)
+ /* Calculate the number of full and partial words -1. */
+ sub limit_wd, limit, #1 /* limit != 0, so no underflow. */
+ lsr limit_wd, limit_wd, #3 /* Convert to Dwords. */
+
+ /* NUL detection works on the principle that (X - 1) & (~X) & 0x80
+ (=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and
+ can be done in parallel across the entire word. */
+ /* Start of performance-critical section -- one 64B cache line. */
+L(loop_aligned):
+ ldr data1, [src1], #8
+ ldr data2, [src2], #8
+L(start_realigned):
+ subs limit_wd, limit_wd, #1
+ sub tmp1, data1, zeroones
+ orr tmp2, data1, #REP8_7f
+ eor diff, data1, data2 /* Non-zero if differences found. */
+ csinv endloop, diff, xzr, pl /* Last Dword or differences. */
+ bics has_nul, tmp1, tmp2 /* Non-zero if NUL terminator. */
+ ccmp endloop, #0, #0, eq
+ b.eq L(loop_aligned)
+ /* End of performance-critical section -- one 64B cache line. */
+
+ /* Not reached the limit, must have found the end or a diff. */
+ tbz limit_wd, #63, L(not_limit)
+
+ /* Limit % 8 == 0 => all bytes significant. */
+ ands limit, limit, #7
+ b.eq L(not_limit)
+
+ lsl limit, limit, #3 /* Bits -> bytes. */
+ mov mask, #~0
+#ifdef __AARCH64EB__
+ lsr mask, mask, limit
+#else
+ lsl mask, mask, limit
+#endif
+ bic data1, data1, mask
+ bic data2, data2, mask
+
+ /* Make sure that the NUL byte is marked in the syndrome. */
+ orr has_nul, has_nul, mask
+
+L(not_limit):
+ orr syndrome, diff, has_nul
+
+#ifndef __AARCH64EB__
+ rev syndrome, syndrome
+ rev data1, data1
+ /* The MS-non-zero bit of the syndrome marks either the first bit
+ that is different, or the top bit of the first zero byte.
+ Shifting left now will bring the critical information into the
+ top bits. */
+ clz pos, syndrome
+ rev data2, data2
+ lsl data1, data1, pos
+ lsl data2, data2, pos
+ /* But we need to zero-extend (char is unsigned) the value and then
+ perform a signed 32-bit subtraction. */
+ lsr data1, data1, #56
+ sub result, data1, data2, lsr #56
+ RET
+#else
+ /* For big-endian we cannot use the trick with the syndrome value
+ as carry-propagation can corrupt the upper bits if the trailing
+ bytes in the string contain 0x01. */
+ /* However, if there is no NUL byte in the dword, we can generate
+ the result directly. We can't just subtract the bytes as the
+ MSB might be significant. */
+ cbnz has_nul, 1f
+ cmp data1, data2
+ cset result, ne
+ cneg result, result, lo
+ RET
+1:
+ /* Re-compute the NUL-byte detection, using a byte-reversed value. */
+ rev tmp3, data1
+ sub tmp1, tmp3, zeroones
+ orr tmp2, tmp3, #REP8_7f
+ bic has_nul, tmp1, tmp2
+ rev has_nul, has_nul
+ orr syndrome, diff, has_nul
+ clz pos, syndrome
+ /* The MS-non-zero bit of the syndrome marks either the first bit
+ that is different, or the top bit of the first zero byte.
+ Shifting left now will bring the critical information into the
+ top bits. */
+ lsl data1, data1, pos
+ lsl data2, data2, pos
+ /* But we need to zero-extend (char is unsigned) the value and then
+ perform a signed 32-bit subtraction. */
+ lsr data1, data1, #56
+ sub result, data1, data2, lsr #56
+ RET
+#endif
+
+L(mutual_align):
+ /* Sources are mutually aligned, but are not currently at an
+ alignment boundary. Round down the addresses and then mask off
+ the bytes that precede the start point.
+ We also need to adjust the limit calculations, but without
+ overflowing if the limit is near ULONG_MAX. */
+ bic src1, src1, #7
+ bic src2, src2, #7
+ ldr data1, [src1], #8
+ neg tmp3, tmp1, lsl #3 /* 64 - bits(bytes beyond align). */
+ ldr data2, [src2], #8
+ mov tmp2, #~0
+ sub limit_wd, limit, #1 /* limit != 0, so no underflow. */
+#ifdef __AARCH64EB__
+ /* Big-endian. Early bytes are at MSB. */
+ lsl tmp2, tmp2, tmp3 /* Shift (tmp1 & 63). */
+#else
+ /* Little-endian. Early bytes are at LSB. */
+ lsr tmp2, tmp2, tmp3 /* Shift (tmp1 & 63). */
+#endif
+ and tmp3, limit_wd, #7
+ lsr limit_wd, limit_wd, #3
+ /* Adjust the limit. Only low 3 bits used, so overflow irrelevant. */
+ add limit, limit, tmp1
+ add tmp3, tmp3, tmp1
+ orr data1, data1, tmp2
+ orr data2, data2, tmp2
+ add limit_wd, limit_wd, tmp3, lsr #3
+ b L(start_realigned)
+
+L(ret0):
+ mov result, #0
+ RET
+
+ .p2align 6
+L(misaligned8):
+ sub limit, limit, #1
+1:
+ /* Perhaps we can do better than this. */
+ ldrb data1w, [src1], #1
+ ldrb data2w, [src2], #1
+ subs limit, limit, #1
+ ccmp data1w, #1, #0, cs /* NZCV = 0b0000. */
+ ccmp data1w, data2w, #0, cs /* NZCV = 0b0000. */
+ b.eq 1b
+ sub result, data1, data2
+ RET
+END (strncmp)
+libc_hidden_builtin_def (strncmp)
http://sources.redhat.com/git/gitweb.cgi?p=glibc.git;a=commitdiff;h=4499bb3e1b14c5121bcf7334d9b754d20768bb2b
commit 4499bb3e1b14c5121bcf7334d9b754d20768bb2b
Author: Marcus Shawcroft <marcus.shawcroft@linaro.org>
Date: Wed Jan 23 16:32:33 2013 +0000
AArch64: Adding optimized strnlen implementation.
diff --git a/ports/ChangeLog.aarch64 b/ports/ChangeLog.aarch64
index c7487f5..ba4792f 100644
--- a/ports/ChangeLog.aarch64
+++ b/ports/ChangeLog.aarch64
@@ -1,3 +1,8 @@
+2013-01-23 Marcus Shawcroft <marcus.shawcroft@linaro.org>
+
+ * sysdeps/aarch64/sysdep.h (ENTRY_ALIGN_AND_PAD): New.
+ * sysdeps/aarch64/strnlen.S: New file.
+
2013-01-17 Marcus Shawcroft <marcus.shawcroft@linaro.org>
* sysdeps/aarch64/strlen.S: New file.
diff --git a/ports/sysdeps/aarch64/strnlen.S b/ports/sysdeps/aarch64/strnlen.S
new file mode 100644
index 0000000..e582e8a
--- /dev/null
+++ b/ports/sysdeps/aarch64/strnlen.S
@@ -0,0 +1,161 @@
+/* strnlen - calculate the length of a string with limit.
+
+ Copyright (C) 2013 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/>. */
+
+#include <sysdep.h>
+
+/* Assumptions:
+ *
+ * ARMv8-a, AArch64
+ */
+
+/* Arguments and results. */
+#define srcin x0
+#define len x0
+#define limit x1
+
+/* Locals and temporaries. */
+#define src x2
+#define data1 x3
+#define data2 x4
+#define data2a x5
+#define has_nul1 x6
+#define has_nul2 x7
+#define tmp1 x8
+#define tmp2 x9
+#define tmp3 x10
+#define tmp4 x11
+#define zeroones x12
+#define pos x13
+#define limit_wd x14
+
+#define REP8_01 0x0101010101010101
+#define REP8_7f 0x7f7f7f7f7f7f7f7f
+#define REP8_80 0x8080808080808080
+
+ENTRY_ALIGN_AND_PAD (__strnlen, 6, 9)
+ cbz limit, L(hit_limit)
+ mov zeroones, #REP8_01
+ bic src, srcin, #15
+ ands tmp1, srcin, #15
+ b.ne L(misaligned)
+ /* Calculate the number of full and partial words -1. */
+ sub limit_wd, limit, #1 /* Limit != 0, so no underflow. */
+ lsr limit_wd, limit_wd, #4 /* Convert to Qwords. */
+
+ /* NUL detection works on the principle that (X - 1) & (~X) & 0x80
+ (=> (X - 1) & ~(X | 0x7f)) is non-zero iff a byte is zero, and
+ can be done in parallel across the entire word. */
+ /* The inner loop deals with two Dwords at a time. This has a
+ slightly higher start-up cost, but we should win quite quickly,
+ especially on cores with a high number of issue slots per
+ cycle, as we get much better parallelism out of the operations. */
+
+ /* Start of critial section -- keep to one 64Byte cache line. */
+L(loop):
+ ldp data1, data2, [src], #16
+L(realigned):
+ sub tmp1, data1, zeroones
+ orr tmp2, data1, #REP8_7f
+ sub tmp3, data2, zeroones
+ orr tmp4, data2, #REP8_7f
+ bic has_nul1, tmp1, tmp2
+ bic has_nul2, tmp3, tmp4
+ subs limit_wd, limit_wd, #1
+ orr tmp1, has_nul1, has_nul2
+ ccmp tmp1, #0, #0, pl /* NZCV = 0000 */
+ b.eq L(loop)
+ /* End of critical section -- keep to one 64Byte cache line. */
+
+ orr tmp1, has_nul1, has_nul2
+ cbz tmp1, L(hit_limit) /* No null in final Qword. */
+
+ /* We know there's a null in the final Qword. The easiest thing
+ to do now is work out the length of the string and return
+ MIN (len, limit). */
+
+ sub len, src, srcin
+ cbz has_nul1, L(nul_in_data2)
+#ifdef __AARCH64EB__
+ mov data2, data1
+#endif
+ sub len, len, #8
+ mov has_nul2, has_nul1
+L(nul_in_data2):
+#ifdef __AARCH64EB__
+ /* For big-endian, carry propagation (if the final byte in the
+ string is 0x01) means we cannot use has_nul directly. The
+ easiest way to get the correct byte is to byte-swap the data
+ and calculate the syndrome a second time. */
+ rev data2, data2
+ sub tmp1, data2, zeroones
+ orr tmp2, data2, #REP8_7f
+ bic has_nul2, tmp1, tmp2
+#endif
+ sub len, len, #8
+ rev has_nul2, has_nul2
+ clz pos, has_nul2
+ add len, len, pos, lsr #3 /* Bits to bytes. */
+ cmp len, limit
+ csel len, len, limit, ls /* Return the lower value. */
+ RET
+
+L(misaligned):
+ /* Deal with a partial first word.
+ We're doing two things in parallel here;
+ 1) Calculate the number of words (but avoiding overflow if
+ limit is near ULONG_MAX) - to do this we need to work out
+ limit + tmp1 - 1 as a 65-bit value before shifting it;
+ 2) Load and mask the initial data words - we force the bytes
+ before the ones we are interested in to 0xff - this ensures
+ early bytes will not hit any zero detection. */
+ sub limit_wd, limit, #1
+ neg tmp4, tmp1
+ cmp tmp1, #8
+
+ and tmp3, limit_wd, #15
+ lsr limit_wd, limit_wd, #4
+ mov tmp2, #~0
+
+ ldp data1, data2, [src], #16
+ lsl tmp4, tmp4, #3 /* Bytes beyond alignment -> bits. */
+ add tmp3, tmp3, tmp1
+
+#ifdef __AARCH64EB__
+ /* Big-endian. Early bytes are at MSB. */
+ lsl tmp2, tmp2, tmp4 /* Shift (tmp1 & 63). */
+#else
+ /* Little-endian. Early bytes are at LSB. */
+ lsr tmp2, tmp2, tmp4 /* Shift (tmp1 & 63). */
+#endif
+ add limit_wd, limit_wd, tmp3, lsr #4
+
+ orr data1, data1, tmp2
+ orr data2a, data2, tmp2
+
+ csinv data1, data1, xzr, le
+ csel data2, data2, data2a, le
+ b L(realigned)
+
+L(hit_limit):
+ mov len, limit
+ RET
+END (__strnlen)
+weak_alias (__strnlen, strnlen)
+libc_hidden_def (strnlen)
diff --git a/ports/sysdeps/aarch64/sysdep.h b/ports/sysdeps/aarch64/sysdep.h
index 6b75ada..9349471 100644
--- a/ports/sysdeps/aarch64/sysdep.h
+++ b/ports/sysdeps/aarch64/sysdep.h
@@ -42,6 +42,22 @@
cfi_startproc; \
CALL_MCOUNT
+/* Define an entry point visible from C with a specified alignment and
+ pre-padding with NOPs. This can be used to ensure that a critical
+ loop within a function is cache line aligned. Note this version
+ does not adjust the padding if CALL_MCOUNT is defined. */
+
+#define ENTRY_ALIGN_AND_PAD(name, align, padding) \
+ .globl C_SYMBOL_NAME(name); \
+ .type C_SYMBOL_NAME(name),%function; \
+ .p2align align; \
+ .rep padding; \
+ nop; \
+ .endr; \
+ C_LABEL(name) \
+ cfi_startproc; \
+ CALL_MCOUNT
+
#undef END
#define END(name) \
cfi_endproc; \
-----------------------------------------------------------------------
Summary of changes:
ports/ChangeLog.aarch64 | 9 ++
ports/sysdeps/aarch64/{strcmp.S => strncmp.S} | 111 ++++++++++++++++++-------
ports/sysdeps/aarch64/{strlen.S => strnlen.S} | 92 +++++++++++++++------
ports/sysdeps/aarch64/sysdep.h | 16 ++++
4 files changed, 173 insertions(+), 55 deletions(-)
copy ports/sysdeps/aarch64/{strcmp.S => strncmp.S} (65%)
copy ports/sysdeps/aarch64/{strlen.S => strnlen.S} (59%)
hooks/post-receive
--
GNU C Library master sources