From a086ff98ad0bb996241037689188caf394f6c633 Mon Sep 17 00:00:00 2001
From: Max Krummenacher <max.krummenacher@toradex.com>
Date: Fri, 12 Jun 2015 13:27:55 +0200
Subject: [PATCH] memcpy: don't use optimized for VFP/NEON versions

Tests with the tinymembench tool on a Colibri T30 show the performance
of the standard arm based memcpy to be around 2 times faster than
__memcpy_neon or __memcpy_vfp.

Tests on Apalis iMX6 confirm this, although running only around 1.3 times faster.
---
 sysdeps/arm/armv7/multiarch/Makefile          |  3 ---
 sysdeps/arm/armv7/multiarch/ifunc-impl-list.c | 18 ------------------
 sysdeps/arm/armv7/multiarch/memcpy.S          | 17 -----------------
 sysdeps/arm/armv7/multiarch/memcpy_neon.S     |  9 ---------
 sysdeps/arm/armv7/multiarch/memcpy_vfp.S      |  7 -------
 5 files changed, 54 deletions(-)

diff --git a/sysdeps/arm/armv7/multiarch/Makefile b/sysdeps/arm/armv7/multiarch/Makefile
index e834cc9..e69de29 100644
--- a/sysdeps/arm/armv7/multiarch/Makefile
+++ b/sysdeps/arm/armv7/multiarch/Makefile
@@ -1,3 +0,0 @@
-ifeq ($(subdir),string)
-sysdep_routines += memcpy_neon memcpy_vfp
-endif
diff --git a/sysdeps/arm/armv7/multiarch/ifunc-impl-list.c b/sysdeps/arm/armv7/multiarch/ifunc-impl-list.c
index 2515418..322eae6 100644
--- a/sysdeps/arm/armv7/multiarch/ifunc-impl-list.c
+++ b/sysdeps/arm/armv7/multiarch/ifunc-impl-list.c
@@ -31,25 +31,7 @@ __libc_ifunc_impl_list (const char *name, struct libc_ifunc_impl *array,
 {
   size_t i = 0;
 
-  bool use_neon = true;
-#ifdef __ARM_NEON__
-# define __memcpy_neon	memcpy
-#else
-  use_neon = (GLRO(dl_hwcap) & HWCAP_ARM_NEON) != 0;
-#endif
-
-#ifndef __ARM_NEON__
-  bool use_vfp = true;
-# ifdef __SOFTFP__
-  use_vfp = (GLRO(dl_hwcap) & HWCAP_ARM_VFP) != 0;
-# endif
-#endif
-
   IFUNC_IMPL (i, name, memcpy,
-	      IFUNC_IMPL_ADD (array, i, memcpy, use_neon, __memcpy_neon)
-#ifndef __ARM_NEON__
-	      IFUNC_IMPL_ADD (array, i, memcpy, use_vfp, __memcpy_vfp)
-#endif
 	      IFUNC_IMPL_ADD (array, i, memcpy, 1, __memcpy_arm));
 
   return i;
diff --git a/sysdeps/arm/armv7/multiarch/memcpy.S b/sysdeps/arm/armv7/multiarch/memcpy.S
index c4f4e80..9ee4d73 100644
--- a/sysdeps/arm/armv7/multiarch/memcpy.S
+++ b/sysdeps/arm/armv7/multiarch/memcpy.S
@@ -23,37 +23,20 @@
 #include <rtld-global-offsets.h>
 
 #if IS_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__
 	ldr	r1, .Lmemcpy_arm
-	tst	r0, #HWCAP_ARM_VFP
-	ldrne	r1, .Lmemcpy_vfp
-# else
-	ldr	r1, .Lmemcpy_vfp
-# endif
-	tst	r0, #HWCAP_ARM_NEON
-	ldrne	r1, .Lmemcpy_neon
 1:
 	add	r0, r1, pc
 	DO_RET(lr)
 
-# ifdef __SOFTFP__
 .Lmemcpy_arm:
 	.long	C_SYMBOL_NAME(__memcpy_arm) - 1b - PC_OFS
-# endif
-.Lmemcpy_neon:
-	.long	C_SYMBOL_NAME(__memcpy_neon) - 1b - PC_OFS
-.Lmemcpy_vfp:
-	.long	C_SYMBOL_NAME(__memcpy_vfp) - 1b - PC_OFS
 
 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.  */
diff --git a/sysdeps/arm/armv7/multiarch/memcpy_neon.S b/sysdeps/arm/armv7/multiarch/memcpy_neon.S
index e60d1cc..e69de29 100644
--- a/sysdeps/arm/armv7/multiarch/memcpy_neon.S
+++ b/sysdeps/arm/armv7/multiarch/memcpy_neon.S
@@ -1,9 +0,0 @@
-#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
-#include "memcpy_impl.S"
diff --git a/sysdeps/arm/armv7/multiarch/memcpy_vfp.S b/sysdeps/arm/armv7/multiarch/memcpy_vfp.S
index e008c04..e69de29 100644
--- a/sysdeps/arm/armv7/multiarch/memcpy_vfp.S
+++ b/sysdeps/arm/armv7/multiarch/memcpy_vfp.S
@@ -1,7 +0,0 @@
-/* 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
-- 
1.9.3