* [PATCH] AArch64 optimized memset
@ 2013-01-16 14:09 Marcus Shawcroft
2013-01-16 15:29 ` Joseph S. Myers
0 siblings, 1 reply; 3+ messages in thread
From: Marcus Shawcroft @ 2013-01-16 14:09 UTC (permalink / raw)
To: libc-ports
[-- Attachment #1: Type: text/plain, Size: 73 bytes --]
This patch provides an AArch64 optimzied memset implementation.
/Marcus
[-- Attachment #2: 0003-AArch64-Implement-optimized-memset.patch --]
[-- Type: application/octet-stream, Size: 6643 bytes --]
diff --git a/ports/ChangeLog.aarch64 b/ports/ChangeLog.aarch64
index 5abefa4..9aceffd 100644
--- a/ports/ChangeLog.aarch64
+++ b/ports/ChangeLog.aarch64
@@ -1,5 +1,9 @@
2013-01-16 Marcus Shawcroft <marcus.shawcroft@linaro.org>
+ * sysdeps/aarch64/memset.S: New file.
+
+2013-01-16 Marcus Shawcroft <marcus.shawcroft@linaro.org>
+
* sysdeps/aarch64/sysdep.h (ENTRY_ALIGN): New.
* sysdeps/aarch64/memcmp.S: New file.
diff --git a/ports/sysdeps/aarch64/memset.S b/ports/sysdeps/aarch64/memset.S
new file mode 100644
index 0000000..f96f6a6
--- /dev/null
+++ b/ports/sysdeps/aarch64/memset.S
@@ -0,0 +1,229 @@
+/* Copyright (C) 2012-2013 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/>. */
+
+/* Assumptions:
+ *
+ * ARMv8-a, AArch64
+ * Unaligned accesses
+ *
+ */
+
+#include <sysdep.h>
+
+/* By default we assume that the DC instruction can be used to zero
+ data blocks more efficiently. In some circumstances this might be
+ unsafe, for example in an asymmetric multiprocessor environment with
+ different DC clear lengths (neither the upper nor lower lengths are
+ safe to use). The feature can be disabled by defining DONT_USE_DC.
+
+ If code may be run in a virtualized environment, then define
+ MAYBE_VIRT. This will cause the code to cache the system register
+ values rather than re-reading them each call. */
+
+#define dstin x0
+#define val w1
+#define count x2
+#define tmp1 x3
+#define tmp1w w3
+#define tmp2 x4
+#define tmp2w w4
+#define zva_len_x x5
+#define zva_len w5
+#define zva_bits_x x6
+
+#define A_l x7
+#define A_lw w7
+#define dst x8
+#define tmp3w w9
+
+ENTRY_ALIGN (__memset, 6)
+
+ mov dst, dstin /* Preserve return value. */
+ ands A_lw, val, #255
+#ifndef DONT_USE_DC
+ b.eq L(zero_mem)
+#endif
+ orr A_lw, A_lw, A_lw, lsl #8
+ orr A_lw, A_lw, A_lw, lsl #16
+ orr A_l, A_l, A_l, lsl #32
+L(tail_maybe_long):
+ cmp count, #64
+ b.ge L(not_short)
+L(tail_maybe_tiny):
+ cmp count, #15
+ b.le L(tail15tiny)
+L(tail63):
+ ands tmp1, count, #0x30
+ b.eq L(tail15)
+ add dst, dst, tmp1
+ cmp tmp1w, #0x20
+ b.eq 1f
+ b.lt 2f
+ stp A_l, A_l, [dst, #-48]
+1:
+ stp A_l, A_l, [dst, #-32]
+2:
+ stp A_l, A_l, [dst, #-16]
+
+L(tail15):
+ and count, count, #15
+ add dst, dst, count
+ stp A_l, A_l, [dst, #-16] /* Repeat some/all of last store. */
+ RET
+
+L(tail15tiny):
+ /* Set up to 15 bytes. Does not assume earlier memory
+ being set. */
+ tbz count, #3, 1f
+ str A_l, [dst], #8
+1:
+ tbz count, #2, 1f
+ str A_lw, [dst], #4
+1:
+ tbz count, #1, 1f
+ strh A_lw, [dst], #2
+1:
+ tbz count, #0, 1f
+ strb A_lw, [dst]
+1:
+ RET
+
+ /* Critical loop. Start at a new cache line boundary. Assuming
+ * 64 bytes per line, this ensures the entire loop is in one line. */
+ .p2align 6
+L(not_short):
+ neg tmp2, dst
+ ands tmp2, tmp2, #15
+ b.eq 2f
+ /* Bring DST to 128-bit (16-byte) alignment. We know that there's
+ * more than that to set, so we simply store 16 bytes and advance by
+ * the amount required to reach alignment. */
+ sub count, count, tmp2
+ stp A_l, A_l, [dst]
+ add dst, dst, tmp2
+ /* There may be less than 63 bytes to go now. */
+ cmp count, #63
+ b.le L(tail63)
+2:
+ sub dst, dst, #16 /* Pre-bias. */
+ sub count, count, #64
+1:
+ stp A_l, A_l, [dst, #16]
+ stp A_l, A_l, [dst, #32]
+ stp A_l, A_l, [dst, #48]
+ stp A_l, A_l, [dst, #64]!
+ subs count, count, #64
+ b.ge 1b
+ tst count, #0x3f
+ add dst, dst, #16
+ b.ne L(tail63)
+ RET
+
+#ifndef DONT_USE_DC
+ /* For zeroing memory, check to see if we can use the ZVA feature to
+ * zero entire 'cache' lines. */
+L(zero_mem):
+ mov A_l, #0
+ cmp count, #63
+ b.le L(tail_maybe_tiny)
+ neg tmp2, dst
+ ands tmp2, tmp2, #15
+ b.eq 1f
+ sub count, count, tmp2
+ stp A_l, A_l, [dst]
+ add dst, dst, tmp2
+ cmp count, #63
+ b.le L(tail63)
+1:
+ /* For zeroing small amounts of memory, it's not worth setting up
+ * the line-clear code. */
+ cmp count, #128
+ b.lt L(not_short)
+#ifdef MAYBE_VIRT
+ /* For efficiency when virtualized, we cache the ZVA capability. */
+ adrp tmp2, L(cache_clear)
+ ldr zva_len, [tmp2, #:lo12:L(cache_clear)]
+ tbnz zva_len, #31, L(not_short)
+ cbnz zva_len, L(zero_by_line)
+ mrs tmp1, dczid_el0
+ tbz tmp1, #4, 1f
+ /* ZVA not available. Remember this for next time. */
+ mov zva_len, #~0
+ str zva_len, [tmp2, #:lo12:L(cache_clear)]
+ b L(not_short)
+1:
+ mov tmp3w, #4
+ and zva_len, tmp1w, #15 /* Safety: other bits reserved. */
+ lsl zva_len, tmp3w, zva_len
+ str zva_len, [tmp2, #:lo12:L(cache_clear)]
+#else
+ mrs tmp1, dczid_el0
+ tbnz tmp1, #4, L(not_short)
+ mov tmp3w, #4
+ and zva_len, tmp1w, #15 /* Safety: other bits reserved. */
+ lsl zva_len, tmp3w, zva_len
+#endif
+
+L(zero_by_line):
+ /* Compute how far we need to go to become suitably aligned. We're
+ * already at quad-word alignment. */
+ cmp count, zva_len_x
+ b.lt L(not_short) /* Not enough to reach alignment. */
+ sub zva_bits_x, zva_len_x, #1
+ neg tmp2, dst
+ ands tmp2, tmp2, zva_bits_x
+ b.eq 1f /* Already aligned. */
+ /* Not aligned, check that there's enough to copy after alignment. */
+ sub tmp1, count, tmp2
+ cmp tmp1, #64
+ ccmp tmp1, zva_len_x, #8, ge /* NZCV=0b1000 */
+ b.lt L(not_short)
+ /* We know that there's at least 64 bytes to zero and that it's safe
+ * to overrun by 64 bytes. */
+ mov count, tmp1
+2:
+ stp A_l, A_l, [dst]
+ stp A_l, A_l, [dst, #16]
+ stp A_l, A_l, [dst, #32]
+ subs tmp2, tmp2, #64
+ stp A_l, A_l, [dst, #48]
+ add dst, dst, #64
+ b.ge 2b
+ /* We've overrun a bit, so adjust dst downwards. */
+ add dst, dst, tmp2
+1:
+ sub count, count, zva_len_x
+3:
+ dc zva, dst
+ add dst, dst, zva_len_x
+ subs count, count, zva_len_x
+ b.ge 3b
+ ands count, count, zva_bits_x
+ b.ne L(tail_maybe_long)
+ RET
+#ifdef MAYBE_VIRT
+ .bss
+ .p2align 2
+L(cache_clear):
+ .space 4
+#endif
+#endif /* DONT_USE_DC */
+
+END (__memset)
+weak_alias (__memset, memset)
+libc_hidden_builtin_def (memset)
--
1.7.11.5
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] AArch64 optimized memset
2013-01-16 14:09 [PATCH] AArch64 optimized memset Marcus Shawcroft
@ 2013-01-16 15:29 ` Joseph S. Myers
2013-01-16 16:32 ` Marcus Shawcroft
0 siblings, 1 reply; 3+ messages in thread
From: Joseph S. Myers @ 2013-01-16 15:29 UTC (permalink / raw)
To: Marcus Shawcroft; +Cc: libc-ports
On Wed, 16 Jan 2013, Marcus Shawcroft wrote:
> This patch provides an AArch64 optimzied memset implementation.
You have two macros DONT_USE_DC and MAYBE_VIRT in this code, but no
configure options to set them. It seems unfortunate if a libc configured
with the default configuration is unsuitable for virtualized environments,
for example. Shouldn't you have appropriate ways to configure these
things, documented in the installation manual, with safe defaults if the
user doesn't use any special options? (The default could use IFUNC to
select between different versions based on the system being used, I
suppose.)
Is the caching definitely safe if multiple threads on different processors
are executing the code using the cache at the same time?
--
Joseph S. Myers
joseph@codesourcery.com
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH] AArch64 optimized memset
2013-01-16 15:29 ` Joseph S. Myers
@ 2013-01-16 16:32 ` Marcus Shawcroft
0 siblings, 0 replies; 3+ messages in thread
From: Marcus Shawcroft @ 2013-01-16 16:32 UTC (permalink / raw)
To: Joseph S. Myers; +Cc: libc-ports
On 16 January 2013 15:29, Joseph S. Myers <joseph@codesourcery.com> wrote:
> On Wed, 16 Jan 2013, Marcus Shawcroft wrote:
>
>> This patch provides an AArch64 optimzied memset implementation.
>
> You have two macros DONT_USE_DC and MAYBE_VIRT in this code, but no
> configure options to set them. It seems unfortunate if a libc configured
> with the default configuration is unsuitable for virtualized environments,
> for example. Shouldn't you have appropriate ways to configure these
> things, documented in the installation manual, with safe defaults if the
> user doesn't use any special options? (The default could use IFUNC to
> select between different versions based on the system being used, I
> suppose.)
This implementation, as configured, will function correctly on a
virtualized environment.
The implementation has been written to provide a default
implementation that has reasonable behavior across a range of ARMv8
uArchs and a range of environments. The code is derived from and is
functionally identical to the code linaro has published in their
cortex strings library.
I expect that in due course we will gain IFUNC support for AArch64
glibc and that we will see IFUNC'd memset implementations specifically
optimized to specific AArch64 uArch implementations.
The two macros were provided in the original cortex strings
implementation to provide flexibility to users of that library. In
the context of glibc I believe the correct configuration is to enable
the use of DC, we really do want to exploit that mechanism as widely
as possible. We should expect that any sane multiprocessor
environment will implement the same DC clear length on all cores, or
if they really must have different DC clear lengths on each core, then
the dczid_el0 and dc zca will need to be virtualized.
Defining MAYBE_VIRT has marginal value in a virtualized environment
where virtualization has disabled the clear cache mechanism. In this
situation we would avoid repeatedly reading the virtualized dczid_el0
register in order to determine that DC ZVA is disabled.
When porting the code from cortex strings to glibc I did contemplate
striping these macros out along with the code that they gate. I
decided to leave them in, despite not being used in glibc in order to
keep the implementation as close as possible to the original.
> Is the caching definitely safe if multiple threads on different processors
> are executing the code using the cache at the same time?
Yes.
Cheers
/Marcus
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2013-01-16 16:32 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2013-01-16 14:09 [PATCH] AArch64 optimized memset Marcus Shawcroft
2013-01-16 15:29 ` Joseph S. Myers
2013-01-16 16:32 ` Marcus Shawcroft
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).