* [PATCH roland/arm-memcpy] ARM: Make multiarch memcpy always use NEON when compiler does
@ 2013-05-13 22:47 Roland McGrath
2013-05-13 23:12 ` Joseph S. Myers
` (2 more replies)
0 siblings, 3 replies; 6+ messages in thread
From: Roland McGrath @ 2013-05-13 22:47 UTC (permalink / raw)
To: libc-ports
When the compiler is emitting NEON instructions anyway, there is no point
in using IFUNC when we can just use the NEON memcpy unconditionally.
Tested on armv7l-linux-gnueabihf with CC='gcc -mfpu=neon',
no check-abi failures, no regressions in 'make check subdirs=string'.
Thanks,
Roland
ports/ChangeLog.arm
2013-05-13 Roland McGrath <roland@hack.frob.com>
* sysdeps/arm/armv7/multiarch/memcpy.S [__ARM_NEON__]: Don't define
memcpy here, just __memcpy_arm and __aeabi_memcpy*.
* sysdeps/arm/armv7/multiarch/memcpy_neon.S [__ARM_NEON__]:
Define memcpy here, not __memcpy_neon.
* sysdeps/arm/armv7/multiarch/memcpy_vfp.S [__ARM_NEON__]:
Define nothing here.
* sysdeps/arm/armv7/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list) [__ARM_NEON__]: Don't list __memcpy_vfp;
use memcpy name for NEON implementation.
--- a/ports/sysdeps/arm/armv7/multiarch/ifunc-impl-list.c
+++ b/ports/sysdeps/arm/armv7/multiarch/ifunc-impl-list.c
@@ -35,9 +35,16 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
IFUNC_IMPL (i, name, memcpy,
IFUNC_IMPL_ADD (array, i, memcpy, hwcap & HWCAP_ARM_NEON,
- __memcpy_neon)
+#ifdef __ARM_NEON__
+ memcpy
+#else
+ __memcpy_neon
+#endif
+ )
+#ifndef __ARM_NEON__
IFUNC_IMPL_ADD (array, i, memcpy, hwcap & HWCAP_ARM_VFP,
__memcpy_vfp)
+#endif
IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_arm));
return i;
--- a/ports/sysdeps/arm/armv7/multiarch/memcpy.S
+++ b/ports/sysdeps/arm/armv7/multiarch/memcpy.S
@@ -22,27 +22,29 @@
#include <sysdep.h>
#include <rtld-global-offsets.h>
-#if !defined NOT_IN_libc
+#ifndef NOT_IN_libc
+/* Under __ARM_NEON__, memcpy_neon.S defines the name memcpy. */
+# ifndef __ARM_NEON__
.text
ENTRY(memcpy)
.type memcpy, %gnu_indirect_function
-#ifdef __SOFTFP__
+# ifdef __SOFTFP__
ldr r1, .Lmemcpy_arm
tst r0, #HWCAP_ARM_VFP
ldrne r1, .Lmemcpy_vfp
-#else
+# else
ldr r1, .Lmemcpy_vfp
-#endif
+# endif
tst r0, #HWCAP_ARM_NEON
ldrne r1, .Lmemcpy_neon
1:
add r0, r1, pc
DO_RET(lr)
-#ifdef __SOFTFP__
+# ifdef __SOFTFP__
.Lmemcpy_arm:
.long C_SYMBOL_NAME(__memcpy_arm) - 1b - PC_OFS
-#endif
+# endif
.Lmemcpy_neon:
.long C_SYMBOL_NAME(__memcpy_neon) - 1b - PC_OFS
.Lmemcpy_vfp:
@@ -51,6 +53,7 @@ ENTRY(memcpy)
END(memcpy)
libc_hidden_builtin_def (memcpy)
+#endif /* Not __ARM_NEON__. */
/* These versions of memcpy are defined not to clobber any VFP or NEON
registers so they must always call the ARM variant of the memcpy code. */
--- a/ports/sysdeps/arm/armv7/multiarch/memcpy_neon.S
+++ b/ports/sysdeps/arm/armv7/multiarch/memcpy_neon.S
@@ -1,3 +1,9 @@
+#ifdef __ARM_NEON__
+/* Under __ARM_NEON__, this file defines memcpy directly. */
+libc_hidden_builtin_def (memcpy)
+#else
+# define memcpy __memcpy_neon
+#endif
+
#define MEMCPY_NEON
-#define memcpy __memcpy_neon
#include "memcpy_impl.S"
--- a/ports/sysdeps/arm/armv7/multiarch/memcpy_vfp.S
+++ b/ports/sysdeps/arm/armv7/multiarch/memcpy_vfp.S
@@ -1,3 +1,7 @@
-#define MEMCPY_VFP
-#define memcpy __memcpy_vfp
-#include "memcpy_impl.S"
+/* Under __ARM_NEON__, memcpy_neon.S defines memcpy directly
+ and the __memcpy_vfp code will never be used. */
+#ifndef __ARM_NEON__
+# define MEMCPY_VFP
+# define memcpy __memcpy_vfp
+# include "memcpy_impl.S"
+#endif
^ permalink raw reply [flat|nested] 6+ messages in thread* Re: [PATCH roland/arm-memcpy] ARM: Make multiarch memcpy always use NEON when compiler does
2013-05-13 22:47 [PATCH roland/arm-memcpy] ARM: Make multiarch memcpy always use NEON when compiler does Roland McGrath
@ 2013-05-13 23:12 ` Joseph S. Myers
2013-05-14 7:47 ` Andreas Schwab
2013-05-14 15:47 ` Richard Henderson
2 siblings, 0 replies; 6+ messages in thread
From: Joseph S. Myers @ 2013-05-13 23:12 UTC (permalink / raw)
To: Roland McGrath; +Cc: libc-ports
On Mon, 13 May 2013, Roland McGrath wrote:
> When the compiler is emitting NEON instructions anyway, there is no point
> in using IFUNC when we can just use the NEON memcpy unconditionally.
>
> Tested on armv7l-linux-gnueabihf with CC='gcc -mfpu=neon',
> no check-abi failures, no regressions in 'make check subdirs=string'.
OK.
--
Joseph S. Myers
joseph@codesourcery.com
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH roland/arm-memcpy] ARM: Make multiarch memcpy always use NEON when compiler does
2013-05-13 22:47 [PATCH roland/arm-memcpy] ARM: Make multiarch memcpy always use NEON when compiler does Roland McGrath
2013-05-13 23:12 ` Joseph S. Myers
@ 2013-05-14 7:47 ` Andreas Schwab
2013-05-17 0:31 ` Roland McGrath
2013-05-14 15:47 ` Richard Henderson
2 siblings, 1 reply; 6+ messages in thread
From: Andreas Schwab @ 2013-05-14 7:47 UTC (permalink / raw)
To: Roland McGrath; +Cc: libc-ports
Roland McGrath <roland@hack.frob.com> writes:
> IFUNC_IMPL (i, name, memcpy,
> IFUNC_IMPL_ADD (array, i, memcpy, hwcap & HWCAP_ARM_NEON,
> - __memcpy_neon)
> +#ifdef __ARM_NEON__
> + memcpy
> +#else
> + __memcpy_neon
> +#endif
> + )
Preprocessor directives in macro arguments give undefined behaviour.
Andreas.
--
Andreas Schwab, SUSE Labs, schwab@suse.de
GPG Key fingerprint = 0196 BAD8 1CE9 1970 F4BE 1748 E4D4 88E3 0EEA B9D7
"And now for something completely different."
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH roland/arm-memcpy] ARM: Make multiarch memcpy always use NEON when compiler does
2013-05-13 22:47 [PATCH roland/arm-memcpy] ARM: Make multiarch memcpy always use NEON when compiler does Roland McGrath
2013-05-13 23:12 ` Joseph S. Myers
2013-05-14 7:47 ` Andreas Schwab
@ 2013-05-14 15:47 ` Richard Henderson
2013-05-17 0:32 ` Roland McGrath
2 siblings, 1 reply; 6+ messages in thread
From: Richard Henderson @ 2013-05-14 15:47 UTC (permalink / raw)
To: Roland McGrath; +Cc: libc-ports
On 05/13/2013 03:47 PM, Roland McGrath wrote:
> IFUNC_IMPL (i, name, memcpy,
> IFUNC_IMPL_ADD (array, i, memcpy, hwcap & HWCAP_ARM_NEON,
> - __memcpy_neon)
> +#ifdef __ARM_NEON__
> + memcpy
> +#else
> + __memcpy_neon
> +#endif
> + )
> +#ifndef __ARM_NEON__
> IFUNC_IMPL_ADD (array, i, memcpy, hwcap & HWCAP_ARM_VFP,
> __memcpy_vfp)
> +#endif
> IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_arm));
I think you're better off leaving the __memcpy_neon symbol in all cases.
Having memcpy be an alias to __memcpy_neon would make the rest of the patch
simpler.
r~
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: [PATCH roland/arm-memcpy] ARM: Make multiarch memcpy always use NEON when compiler does
2013-05-14 15:47 ` Richard Henderson
@ 2013-05-17 0:32 ` Roland McGrath
0 siblings, 0 replies; 6+ messages in thread
From: Roland McGrath @ 2013-05-17 0:32 UTC (permalink / raw)
To: Richard Henderson; +Cc: libc-ports
> I think you're better off leaving the __memcpy_neon symbol in all cases.
> Having memcpy be an alias to __memcpy_neon would make the rest of the patch
> simpler.
I don't see what could be any simpler (or different) other than this one
#ifdef in the test case.
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2013-05-17 0:32 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2013-05-13 22:47 [PATCH roland/arm-memcpy] ARM: Make multiarch memcpy always use NEON when compiler does Roland McGrath
2013-05-13 23:12 ` Joseph S. Myers
2013-05-14 7:47 ` Andreas Schwab
2013-05-17 0:31 ` Roland McGrath
2013-05-14 15:47 ` Richard Henderson
2013-05-17 0:32 ` Roland McGrath
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).