00db10
    Backport of the following patch as a prerequistite for
00db10
    96d6fd6c4060d739abb1822e7ad633af749532b2:
00db10
    commit 69f13dbf06c6195de0ada8632271d58ca3cf55da
00db10
    Author: Adhemerval Zanella <azanella@linux.vnet.ibm.com>
00db10
    Date:   Thu Sep 26 09:29:19 2013 -0500
00db10
    
00db10
        PowerPC: strcpy/stpcpy optimization for PPC64/POWER7
00db10
    
00db10
        This patch intends to unify both strcpy and stpcpy implementationsi
00db10
        for PPC64 and PPC64/POWER7. The idead default powerpc64 implementation
00db10
        is to provide both doubleword and word aligned memory access.
00db10
    
00db10
        For PPC64/POWER7 is also provide doubleword and word memory access,
00db10
        remove the branch hints, use the cmpb instruction for compare
00db10
        doubleword/words, and add an optimization for inputs of same alignment.
00db10
    
00db10
        ChangeLog:
00db10
    
00db10
            2013-10-04  Adhemerval Zanella  <azanella@linux.vnet.ibm.com>
00db10
    
00db10
            * sysdeps/powerpc/powerpc64/strcpy.S (strcpy): Add word load/store
00db10
            to provide a boost for large inputs with word alignment.
00db10
            * sysdeps/powerpc/powerpc64/stpcpy.S (__stpcpy): Rewrite
00db10
            implementation based on optimized PPC64 strcpy.
00db10
            * sysdeps/powerpc/powerpc64/power7/strcpy.S: New file: optimized
00db10
            strcpy for PPC64/POWER7 based on both doubleword and word load/store.
00db10
            * sysdeps/powerpc/powerpc64/power7/stpcpy.S: New file: optimized
00db10
            stpcpy for PPC64/POWER7 based on PPC64/POWER7 strcpy.
00db10
00db10
diff --git a/sysdeps/powerpc/powerpc64/power7/stpcpy.S b/sysdeps/powerpc/powerpc64/power7/stpcpy.S
00db10
new file mode 100644
00db10
index 0000000..727dd06
00db10
--- /dev/null
00db10
+++ b/sysdeps/powerpc/powerpc64/power7/stpcpy.S
00db10
@@ -0,0 +1,24 @@
00db10
+/* Optimized stpcpy implementation for PowerPC64/POWER7.
00db10
+   Copyright (C) 2013 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
+#define USE_AS_STPCPY
00db10
+#include <sysdeps/powerpc/powerpc64/power7/strcpy.S>
00db10
+
00db10
+weak_alias (__stpcpy, stpcpy)
00db10
+libc_hidden_def (__stpcpy)
00db10
+libc_hidden_builtin_def (stpcpy)
00db10
diff --git a/sysdeps/powerpc/powerpc64/power7/strcpy.S b/sysdeps/powerpc/powerpc64/power7/strcpy.S
00db10
new file mode 100644
00db10
index 0000000..5c341a1
00db10
--- /dev/null
00db10
+++ b/sysdeps/powerpc/powerpc64/power7/strcpy.S
00db10
@@ -0,0 +1,274 @@
00db10
+/* Optimized strcpy/stpcpy implementation for PowerPC64/POWER7.
00db10
+   Copyright (C) 2013 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
+/* Implements the function
00db10
+
00db10
+   char * [r3] strcpy (char *dest [r3], const char *src [r4])
00db10
+
00db10
+   or
00db10
+
00db10
+   char * [r3] strcpy (char *dest [r3], const char *src [r4])
00db10
+
00db10
+   if USE_AS_STPCPY is defined. It tries to use aligned memory accesses
00db10
+   when possible using the following algorithm:
00db10
+
00db10
+   if (((((uintptr_t)dst & 0x7UL) == 0) && ((uintptr_t)src & 0x7UL) == 0))
00db10
+     goto aligned_doubleword_copy;
00db10
+   if (((((uintptr_t)dst & 0x3UL) == 0) && ((uintptr_t)src & 0x3UL) == 0))
00db10
+     goto aligned_word_copy;
00db10
+   if (((uintptr_t)dst & 0x7UL) == ((uintptr_t)src & 0x7UL))
00db10
+     goto same_alignment;
00db10
+   goto unaligned;
00db10
+
00db10
+   The aligned comparison are made using cmpb instructions.  */
00db10
+
00db10
+#ifdef USE_AS_STPCPY
00db10
+# define FUNC_NAME __stpcpy
00db10
+#else
00db10
+# define FUNC_NAME strcpy
00db10
+#endif
00db10
+
00db10
+	.machine  power7
00db10
+EALIGN (FUNC_NAME, 4, 0)
00db10
+	CALL_MCOUNT 2
00db10
+
00db10
+#define rTMP	r0
00db10
+#ifdef USE_AS_STPCPY
00db10
+#define rRTN	r3	/* pointer to previous word/doubleword in dest */
00db10
+#else
00db10
+#define rRTN	r12	/* pointer to previous word/doubleword in dest */
00db10
+#endif
00db10
+#define rSRC	r4	/* pointer to previous word/doubleword in src */
00db10
+#define rMASK	r5	/* mask 0xffffffff | 0xffffffffffffffff */
00db10
+#define rWORD	r6	/* current word from src */
00db10
+#define rALT	r7	/* alternate word from src */
00db10
+#define rRTNAL	r8	/* alignment of return pointer */
00db10
+#define rSRCAL	r9	/* alignment of source pointer */
00db10
+#define rALCNT	r10	/* bytes to read to reach 8 bytes alignment */
00db10
+#define rSUBAL	r11	/* doubleword minus unaligned displacement */
00db10
+
00db10
+#ifndef USE_AS_STPCPY
00db10
+/* Save the dst pointer to use as return value.  */
00db10
+	mr	rRTN, r3
00db10
+#endif
00db10
+	or	rTMP, rSRC, rRTN
00db10
+	clrldi.	rTMP, rTMP, 61
00db10
+	bne	L(check_word_alignment)
00db10
+	b	L(aligned_doubleword_copy)
00db10
+
00db10
+L(same_alignment):
00db10
+/* Src and dst with same alignment: align both to doubleword.  */
00db10
+	mr	rALCNT, rRTN
00db10
+	lbz	rWORD, 0(rSRC)
00db10
+	subfic	rSUBAL, rRTNAL, 8
00db10
+	addi	rRTN, rRTN, 1
00db10
+	addi	rSRC, rSRC, 1
00db10
+	cmpdi	cr7, rWORD, 0
00db10
+	stb	rWORD, 0(rALCNT)
00db10
+	beq	cr7, L(s2)
00db10
+
00db10
+	add	rALCNT, rALCNT, rSUBAL
00db10
+	subf	rALCNT, rRTN, rALCNT
00db10
+	addi	rALCNT, rALCNT, 1
00db10
+	mtctr	rALCNT
00db10
+	b	L(s1)
00db10
+
00db10
+	.align 4
00db10
+L(s0):
00db10
+	addi	rSRC, rSRC, 1
00db10
+	lbz	rWORD, -1(rSRC)
00db10
+	cmpdi	cr7, rWORD, 0
00db10
+	stb	rWORD, -1(rALCNT)
00db10
+	beqlr	cr7
00db10
+	mr	rRTN, rALCNT
00db10
+L(s1):
00db10
+	addi	rALCNT, rRTN,1
00db10
+	bdnz	L(s0)
00db10
+	b L(aligned_doubleword_copy)
00db10
+	.align 4
00db10
+L(s2):
00db10
+	mr	rRTN, rALCNT
00db10
+	blr
00db10
+
00db10
+/* For doubleword aligned memory, operate using doubleword load and stores.  */
00db10
+	.align 4
00db10
+L(aligned_doubleword_copy):
00db10
+	li	rMASK, 0
00db10
+	addi	rRTN, rRTN, -8
00db10
+	ld	rWORD, 0(rSRC)
00db10
+	b	L(g2)
00db10
+
00db10
+	.align 4
00db10
+L(g0):	ldu	rALT, 8(rSRC)
00db10
+	stdu	rWORD, 8(rRTN)
00db10
+	cmpb	rTMP, rALT, rMASK
00db10
+	cmpdi	rTMP, 0
00db10
+	bne	L(g1)
00db10
+	ldu	rWORD, 8(rSRC)
00db10
+	stdu	rALT, 8(rRTN)
00db10
+L(g2):	cmpb	rTMP, rWORD, rMASK
00db10
+	cmpdi	rTMP, 0		/* If rTMP is 0, no null's have been found.  */
00db10
+	beq	L(g0)
00db10
+
00db10
+	mr	rALT, rWORD
00db10
+/* We've hit the end of the string.  Do the rest byte-by-byte.  */
00db10
+L(g1):
00db10
+#ifdef __LITTLE_ENDIAN__
00db10
+	extrdi.	rTMP, rALT, 8, 56
00db10
+	stbu	rALT, 8(rRTN)
00db10
+	beqlr-
00db10
+	extrdi.	rTMP, rALT, 8, 48
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr-
00db10
+	extrdi.	rTMP, rALT, 8, 40
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr-
00db10
+	extrdi.	rTMP, rALT, 8, 32
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr-
00db10
+	extrdi.	rTMP, rALT, 8, 24
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr-
00db10
+	extrdi.	rTMP, rALT, 8, 16
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr-
00db10
+	extrdi.	rTMP, rALT, 8, 8
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr-
00db10
+	extrdi	rTMP, rALT, 8, 0
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+#else
00db10
+	extrdi.	rTMP, rALT, 8, 0
00db10
+	stbu	rTMP, 8(rRTN)
00db10
+	beqlr
00db10
+	extrdi.	rTMP, rALT, 8, 8
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr
00db10
+	extrdi.	rTMP, rALT, 8, 16
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr
00db10
+	extrdi.	rTMP, rALT, 8, 24
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr
00db10
+	extrdi.	rTMP, rALT, 8, 32
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr
00db10
+	extrdi.	rTMP, rALT, 8, 40
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr
00db10
+	extrdi.	rTMP, rALT, 8, 48
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr
00db10
+	stbu	rALT, 1(rRTN)
00db10
+#endif
00db10
+	blr
00db10
+
00db10
+L(check_word_alignment):
00db10
+	clrldi. rTMP, rTMP, 62
00db10
+	beq	L(aligned_word_copy)
00db10
+	rldicl	rRTNAL, rRTN, 0, 61
00db10
+	rldicl	rSRCAL, rSRC, 0, 61
00db10
+	cmpld	cr7, rSRCAL, rRTNAL
00db10
+	beq	cr7, L(same_alignment)
00db10
+	b	L(unaligned)
00db10
+
00db10
+/* For word aligned memory, operate using word load and stores.  */
00db10
+	.align	4
00db10
+L(aligned_word_copy):
00db10
+	li	rMASK, 0
00db10
+	addi	rRTN, rRTN, -4
00db10
+	lwz	rWORD, 0(rSRC)
00db10
+	b	L(g5)
00db10
+
00db10
+	.align	4
00db10
+L(g3):	lwzu	rALT, 4(rSRC)
00db10
+	stwu	rWORD, 4(rRTN)
00db10
+	cmpb	rTMP, rALT, rMASK
00db10
+	cmpwi	rTMP, 0
00db10
+	bne	L(g4)
00db10
+	lwzu	rWORD, 4(rSRC)
00db10
+	stwu	rALT, 4(rRTN)
00db10
+L(g5):	cmpb	rTMP, rWORD, rMASK
00db10
+	cmpwi	rTMP, 0		/* If rTMP is 0, no null in word.  */
00db10
+	beq	L(g3)
00db10
+
00db10
+	mr      rALT, rWORD
00db10
+/* We've hit the end of the string.  Do the rest byte-by-byte.  */
00db10
+L(g4):
00db10
+#ifdef __LITTLE_ENDIAN__
00db10
+	rlwinm.	rTMP, rALT, 0, 24, 31
00db10
+	stbu	rALT, 4(rRTN)
00db10
+	beqlr-
00db10
+	rlwinm.	rTMP, rALT, 24, 24, 31
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr-
00db10
+	rlwinm.	rTMP, rALT, 16, 24, 31
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr-
00db10
+	rlwinm	rTMP, rALT, 8, 24, 31
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+#else
00db10
+	rlwinm. rTMP, rALT, 8, 24, 31
00db10
+	stbu    rTMP, 4(rRTN)
00db10
+	beqlr
00db10
+	rlwinm. rTMP, rALT, 16, 24, 31
00db10
+	stbu    rTMP, 1(rRTN)
00db10
+	beqlr
00db10
+	rlwinm. rTMP, rALT, 24, 24, 31
00db10
+	stbu    rTMP, 1(rRTN)
00db10
+	beqlr
00db10
+	stbu    rALT, 1(rRTN)
00db10
+#endif
00db10
+	blr
00db10
+
00db10
+/* Oh well.  In this case, we just do a byte-by-byte copy.  */
00db10
+	.align	4
00db10
+L(unaligned):
00db10
+	lbz	rWORD, 0(rSRC)
00db10
+	addi	rRTN, rRTN, -1
00db10
+	cmpdi	rWORD, 0
00db10
+	beq	L(u2)
00db10
+
00db10
+	.align 	5
00db10
+L(u0):	lbzu	rALT, 1(rSRC)
00db10
+	stbu	rWORD, 1(rRTN)
00db10
+	cmpdi	rALT, 0
00db10
+	beq	L(u1)
00db10
+	lbzu	rWORD, 1(rSRC)
00db10
+	stbu	rALT, 1(rRTN)
00db10
+	cmpdi	rWORD, 0
00db10
+	beq	L(u2)
00db10
+	lbzu	rALT, 1(rSRC)
00db10
+	stbu	rWORD, 1(rRTN)
00db10
+	cmpdi	rALT, 0
00db10
+	beq	L(u1)
00db10
+	lbzu	rWORD, 1(rSRC)
00db10
+	stbu	rALT, 1(rRTN)
00db10
+	cmpdi	rWORD, 0
00db10
+	bne	L(u0)
00db10
+L(u2):	stbu	rWORD, 1(rRTN)
00db10
+	blr
00db10
+L(u1):	stbu	rALT, 1(rRTN)
00db10
+	blr
00db10
+END (FUNC_NAME)
00db10
+
00db10
+#ifndef USE_AS_STPCPY
00db10
+libc_hidden_builtin_def (strcpy)
00db10
+#endif
00db10
diff --git a/sysdeps/powerpc/powerpc64/stpcpy.S b/sysdeps/powerpc/powerpc64/stpcpy.S
00db10
index d795b61..09aa3be 100644
00db10
--- a/sysdeps/powerpc/powerpc64/stpcpy.S
00db10
+++ b/sysdeps/powerpc/powerpc64/stpcpy.S
00db10
@@ -1,5 +1,5 @@
00db10
 /* Optimized stpcpy implementation for PowerPC64.
00db10
-   Copyright (C) 1997, 1999, 2000, 2002, 2004 Free Software Foundation, Inc.
00db10
+   Copyright (C) 1997-2013 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
@@ -16,123 +16,9 @@
00db10
    License along with the GNU C Library; if not, see
00db10
    <http://www.gnu.org/licenses/>.  */
00db10
 
00db10
-#include <sysdep.h>
00db10
-#include <bp-sym.h>
00db10
-#include <bp-asm.h>
00db10
+#define USE_AS_STPCPY
00db10
+#include <sysdeps/powerpc/powerpc64/strcpy.S>
00db10
 
00db10
-/* See strlen.s for comments on how the end-of-string testing works.  */
00db10
-
00db10
-/* char * [r3] stpcpy (char *dest [r3], const char *src [r4])  */
00db10
-
00db10
-EALIGN (BP_SYM (__stpcpy), 4, 0)
00db10
-	CALL_MCOUNT 2
00db10
-
00db10
-#define rTMP	r0
00db10
-#define rRTN	r3
00db10
-#if __BOUNDED_POINTERS__
00db10
-# define rDEST	r4		/* pointer to previous word in dest */
00db10
-# define rSRC	r5		/* pointer to previous word in src */
00db10
-# define rLOW	r11
00db10
-# define rHIGH	r12
00db10
-#else
00db10
-# define rDEST	r3		/* pointer to previous word in dest */
00db10
-# define rSRC	r4		/* pointer to previous word in src */
00db10
-#endif
00db10
-#define rWORD	r6		/* current word from src */
00db10
-#define rFEFE	r7		/* 0xfefefeff */
00db10
-#define r7F7F	r8		/* 0x7f7f7f7f */
00db10
-#define rNEG	r9		/* ~(word in src | 0x7f7f7f7f) */
00db10
-#define rALT	r10		/* alternate word from src */
00db10
-
00db10
-	CHECK_BOUNDS_LOW (rSRC, rLOW, rHIGH)
00db10
-	CHECK_BOUNDS_LOW (rDEST, rLOW, rHIGH)
00db10
-	STORE_RETURN_BOUNDS (rLOW, rHIGH)
00db10
-
00db10
-	or	rTMP, rSRC, rDEST
00db10
-	clrldi.	rTMP, rTMP, 62
00db10
-	addi	rDEST, rDEST, -4
00db10
-	bne	L(unaligned)
00db10
-
00db10
-	lis	rFEFE, -0x101
00db10
-	lis	r7F7F, 0x7f7f
00db10
-	lwz	rWORD, 0(rSRC)
00db10
-	addi	rFEFE, rFEFE, -0x101
00db10
-	addi	r7F7F, r7F7F, 0x7f7f
00db10
-	b	L(g2)
00db10
-
00db10
-L(g0):	lwzu	rALT, 4(rSRC)
00db10
-	stwu	rWORD, 4(rDEST)
00db10
-	add	rTMP, rFEFE, rALT
00db10
-	nor	rNEG, r7F7F, rALT
00db10
-	and.	rTMP, rTMP, rNEG
00db10
-	bne-	L(g1)
00db10
-	lwzu	rWORD, 4(rSRC)
00db10
-	stwu	rALT, 4(rDEST)
00db10
-L(g2):	add	rTMP, rFEFE, rWORD
00db10
-	nor	rNEG, r7F7F, rWORD
00db10
-	and.	rTMP, rTMP, rNEG
00db10
-	beq+	L(g0)
00db10
-
00db10
-	mr	rALT, rWORD
00db10
-/* We've hit the end of the string.  Do the rest byte-by-byte.  */
00db10
-L(g1):
00db10
-#ifdef __LITTLE_ENDIAN__
00db10
-	rlwinm.	rTMP, rALT, 0, 24, 31
00db10
-	stbu	rALT, 4(rDEST)
00db10
-	beqlr-
00db10
-	rlwinm.	rTMP, rALT, 24, 24, 31
00db10
-	stbu	rTMP, 1(rDEST)
00db10
-	beqlr-
00db10
-	rlwinm.	rTMP, rALT, 16, 24, 31
00db10
-	stbu	rTMP, 1(rDEST)
00db10
-	beqlr-
00db10
-	rlwinm	rTMP, rALT, 8, 24, 31
00db10
-	stbu	rTMP, 1(rDEST)
00db10
-	blr
00db10
-#else
00db10
-	rlwinm.	rTMP, rALT, 8, 24, 31
00db10
-	stbu	rTMP, 4(rDEST)
00db10
-	beqlr-
00db10
-	rlwinm.	rTMP, rALT, 16, 24, 31
00db10
-	stbu	rTMP, 1(rDEST)
00db10
-	beqlr-
00db10
-	rlwinm.	rTMP, rALT, 24, 24, 31
00db10
-	stbu	rTMP, 1(rDEST)
00db10
-	beqlr-
00db10
-	stbu	rALT, 1(rDEST)
00db10
-	CHECK_BOUNDS_HIGH (rDEST, rHIGH, twlgt)
00db10
-	STORE_RETURN_VALUE (rDEST)
00db10
-	blr
00db10
-#endif
00db10
-
00db10
-/* Oh well.  In this case, we just do a byte-by-byte copy.  */
00db10
-	.align 4
00db10
-	nop
00db10
-L(unaligned):
00db10
-	lbz	rWORD, 0(rSRC)
00db10
-	addi	rDEST, rDEST, 3
00db10
-	cmpwi	rWORD, 0
00db10
-	beq-	L(u2)
00db10
-
00db10
-L(u0):	lbzu	rALT, 1(rSRC)
00db10
-	stbu	rWORD, 1(rDEST)
00db10
-	cmpwi	rALT, 0
00db10
-	beq-	L(u1)
00db10
-	nop		/* Let 601 load start of loop.  */
00db10
-	lbzu	rWORD, 1(rSRC)
00db10
-	stbu	rALT, 1(rDEST)
00db10
-	cmpwi	rWORD, 0
00db10
-	bne+	L(u0)
00db10
-L(u2):	stbu	rWORD, 1(rDEST)
00db10
-	CHECK_BOUNDS_HIGH (rDEST, rHIGH, twlgt)
00db10
-	STORE_RETURN_VALUE (rDEST)
00db10
-	blr
00db10
-L(u1):	stbu	rALT, 1(rDEST)
00db10
-	CHECK_BOUNDS_HIGH (rDEST, rHIGH, twlgt)
00db10
-	STORE_RETURN_VALUE (rDEST)
00db10
-	blr
00db10
-END (BP_SYM (__stpcpy))
00db10
-
00db10
-weak_alias (BP_SYM (__stpcpy), BP_SYM (stpcpy))
00db10
+weak_alias (__stpcpy, stpcpy)
00db10
 libc_hidden_def (__stpcpy)
00db10
 libc_hidden_builtin_def (stpcpy)
00db10
diff --git a/sysdeps/powerpc/powerpc64/strcpy.S b/sysdeps/powerpc/powerpc64/strcpy.S
00db10
index 9434c27..793325d 100644
00db10
--- a/sysdeps/powerpc/powerpc64/strcpy.S
00db10
+++ b/sysdeps/powerpc/powerpc64/strcpy.S
00db10
@@ -1,5 +1,5 @@
00db10
 /* Optimized strcpy implementation for PowerPC64.
00db10
-   Copyright (C) 1997, 1999, 2000, 2002, 2003, 2011 Free Software Foundation, Inc.
00db10
+   Copyright (C) 1997-2013 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
@@ -17,52 +17,43 @@
00db10
    <http://www.gnu.org/licenses/>.  */
00db10
 
00db10
 #include <sysdep.h>
00db10
-#include <bp-sym.h>
00db10
-#include <bp-asm.h>
00db10
 
00db10
 /* See strlen.s for comments on how the end-of-string testing works.  */
00db10
 
00db10
 /* char * [r3] strcpy (char *dest [r3], const char *src [r4])  */
00db10
 
00db10
-EALIGN (BP_SYM (strcpy), 4, 0)
00db10
+#ifdef USE_AS_STPCPY
00db10
+# define FUNC_NAME __stpcpy
00db10
+#else
00db10
+# define FUNC_NAME strcpy
00db10
+#endif
00db10
+
00db10
+EALIGN (FUNC_NAME, 4, 0)
00db10
 	CALL_MCOUNT 2
00db10
 
00db10
 #define rTMP	r0
00db10
-#define rRTN	r3	/* incoming DEST arg preserved as result */
00db10
-/* Note.  The Bounded pointer support in this code is broken.  This code
00db10
-   was inherited from PPC32 and that support was never completed.
00db10
-   Current PPC gcc does not support -fbounds-check or -fbounded-pointers.
00db10
-   These artifacts are left in the code as a reminder in case we need
00db10
-   bounded pointer support in the future.  */
00db10
-#if __BOUNDED_POINTERS__
00db10
-# define rDEST	r4	/* pointer to previous word in dest */
00db10
-# define rSRC	r5	/* pointer to previous word in src */
00db10
-# define rLOW	r11
00db10
-# define rHIGH	r12
00db10
+#ifdef USE_AS_STPCPY
00db10
+#define rRTN    r3      /* pointer to previous word/doubleword in dest */
00db10
 #else
00db10
-# define rSRC	r4	/* pointer to previous word in src */
00db10
-# define rDEST	r5	/* pointer to previous word in dest */
00db10
+#define rRTN    r12     /* pointer to previous word/doubleword in dest */
00db10
 #endif
00db10
+#define rSRC	r4	/* pointer to previous word/doubleword in src */
00db10
 #define rWORD	r6	/* current word from src */
00db10
-#define rFEFE	r7	/* constant 0xfefefefefefefeff (-0x0101010101010101) */
00db10
-#define r7F7F	r8	/* constant 0x7f7f7f7f7f7f7f7f */
00db10
-#define rNEG	r9	/* ~(word in s1 | 0x7f7f7f7f7f7f7f7f) */
00db10
+#define rFEFE	r7	/* constant 0xfefefeff | 0xfefefefefefefeff */
00db10
+#define r7F7F	r8	/* constant 0x7f7f7f7f | 0x7f7f7f7f7f7f7f7f */
00db10
+#define rNEG	r9	/* ~(word in s1 | r7F7F) */
00db10
 #define rALT	r10	/* alternate word from src */
00db10
 
00db10
-	CHECK_BOUNDS_LOW (rSRC, rLOW, rHIGH)
00db10
-	CHECK_BOUNDS_LOW (rDEST, rLOW, rHIGH)
00db10
-	STORE_RETURN_BOUNDS (rLOW, rHIGH)
00db10
-
00db10
-	dcbt	0,rSRC
00db10
+#ifndef USE_AS_STPCPY
00db10
+/* Save the dst pointer to use as return value.  */
00db10
+	mr      rRTN, r3
00db10
+#endif
00db10
 	or	rTMP, rSRC, rRTN
00db10
 	clrldi.	rTMP, rTMP, 61
00db10
-#if __BOUNDED_POINTERS__
00db10
-	addi	rDEST, rDEST, -8
00db10
-#else
00db10
-	addi	rDEST, rRTN, -8
00db10
-#endif
00db10
-	dcbtst	0,rRTN
00db10
-	bne	L(unaligned)
00db10
+	bne	L(check_word_alignment)
00db10
+
00db10
+/* For doubleword aligned memory, operate using doubleword load and stores.  */
00db10
+	addi	rRTN, rRTN, -8
00db10
 
00db10
 	lis	rFEFE, -0x101
00db10
 	lis	r7F7F, 0x7f7f
00db10
@@ -75,13 +66,13 @@ EALIGN (BP_SYM (strcpy), 4, 0)
00db10
 	b	L(g2)
00db10
 
00db10
 L(g0):	ldu	rALT, 8(rSRC)
00db10
-	stdu	rWORD, 8(rDEST)
00db10
+	stdu	rWORD, 8(rRTN)
00db10
 	add	rTMP, rFEFE, rALT
00db10
 	nor	rNEG, r7F7F, rALT
00db10
 	and.	rTMP, rTMP, rNEG
00db10
 	bne-	L(g1)
00db10
 	ldu	rWORD, 8(rSRC)
00db10
-	stdu	rALT, 8(rDEST)
00db10
+	stdu	rALT, 8(rRTN)
00db10
 L(g2):	add	rTMP, rFEFE, rWORD
00db10
 	nor	rNEG, r7F7F, rWORD
00db10
 	and.	rTMP, rTMP, rNEG
00db10
@@ -92,80 +83,134 @@ L(g2):	add	rTMP, rFEFE, rWORD
00db10
 L(g1):
00db10
 #ifdef __LITTLE_ENDIAN__
00db10
 	extrdi.	rTMP, rALT, 8, 56
00db10
-	stb	rALT, 8(rDEST)
00db10
+	stbu	rALT, 8(rRTN)
00db10
 	beqlr-
00db10
 	extrdi.	rTMP, rALT, 8, 48
00db10
-	stb	rTMP, 9(rDEST)
00db10
+	stbu	rTMP, 1(rRTN)
00db10
 	beqlr-
00db10
 	extrdi.	rTMP, rALT, 8, 40
00db10
-	stb	rTMP, 10(rDEST)
00db10
+	stbu	rTMP, 1(rRTN)
00db10
 	beqlr-
00db10
 	extrdi.	rTMP, rALT, 8, 32
00db10
-	stb	rTMP, 11(rDEST)
00db10
+	stbu	rTMP, 1(rRTN)
00db10
 	beqlr-
00db10
 	extrdi.	rTMP, rALT, 8, 24
00db10
-	stb	rTMP, 12(rDEST)
00db10
+	stbu	rTMP, 1(rRTN)
00db10
 	beqlr-
00db10
 	extrdi.	rTMP, rALT, 8, 16
00db10
-	stb	rTMP, 13(rDEST)
00db10
+	stbu	rTMP, 1(rRTN)
00db10
 	beqlr-
00db10
 	extrdi.	rTMP, rALT, 8, 8
00db10
-	stb	rTMP, 14(rDEST)
00db10
+	stbu	rTMP, 1(rRTN)
00db10
 	beqlr-
00db10
 	extrdi	rTMP, rALT, 8, 0
00db10
-	stb	rTMP, 15(rDEST)
00db10
-	blr
00db10
+	stbu	rTMP, 1(rRTN)
00db10
 #else
00db10
 	extrdi.	rTMP, rALT, 8, 0
00db10
-	stb	rTMP, 8(rDEST)
00db10
+	stbu	rTMP, 8(rRTN)
00db10
 	beqlr-
00db10
 	extrdi.	rTMP, rALT, 8, 8
00db10
-	stb	rTMP, 9(rDEST)
00db10
+	stbu	rTMP, 1(rRTN)
00db10
 	beqlr-
00db10
 	extrdi.	rTMP, rALT, 8, 16
00db10
-	stb	rTMP, 10(rDEST)
00db10
+	stbu	rTMP, 1(rRTN)
00db10
 	beqlr-
00db10
 	extrdi.	rTMP, rALT, 8, 24
00db10
-	stb	rTMP, 11(rDEST)
00db10
+	stbu	rTMP, 1(rRTN)
00db10
 	beqlr-
00db10
 	extrdi.	rTMP, rALT, 8, 32
00db10
-	stb	rTMP, 12(rDEST)
00db10
-	beqlr-
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr
00db10
 	extrdi.	rTMP, rALT, 8, 40
00db10
-	stb	rTMP, 13(rDEST)
00db10
+	stbu	rTMP, 1(rRTN)
00db10
 	beqlr-
00db10
 	extrdi.	rTMP, rALT, 8, 48
00db10
-	stb	rTMP, 14(rDEST)
00db10
+	stbu	rTMP, 1(rRTN)
00db10
 	beqlr-
00db10
-	stb	rALT, 15(rDEST)
00db10
-	/* GKM FIXME: check high bound.  */
00db10
+	stbu	rALT, 1(rRTN)
00db10
+#endif
00db10
 	blr
00db10
+
00db10
+L(check_word_alignment):
00db10
+	clrldi. rTMP, rTMP, 62
00db10
+	bne     L(unaligned)
00db10
+
00db10
+/* For word aligned memory, operate using word load and stores.  */
00db10
+	addi	rRTN, rRTN, -4
00db10
+
00db10
+	lis	rFEFE, -0x101
00db10
+	lis	r7F7F, 0x7f7f
00db10
+	lwz	rWORD, 0(rSRC)
00db10
+	addi	rFEFE, rFEFE, -0x101
00db10
+	addi	r7F7F, r7F7F, 0x7f7f
00db10
+	b	L(g5)
00db10
+
00db10
+L(g3):	lwzu	rALT, 4(rSRC)
00db10
+	stwu	rWORD, 4(rRTN)
00db10
+	add	rTMP, rFEFE, rALT
00db10
+	nor	rNEG, r7F7F, rALT
00db10
+	and.	rTMP, rTMP, rNEG
00db10
+	bne-	L(g4)
00db10
+	lwzu	rWORD, 4(rSRC)
00db10
+	stwu	rALT, 4(rRTN)
00db10
+L(g5):	add	rTMP, rFEFE, rWORD
00db10
+	nor	rNEG, r7F7F, rWORD
00db10
+	and.	rTMP, rTMP, rNEG
00db10
+	beq+	L(g3)
00db10
+
00db10
+	mr	rALT, rWORD
00db10
+/* We've hit the end of the string.  Do the rest byte-by-byte.  */
00db10
+L(g4):
00db10
+#ifdef __LITTLE_ENDIAN__
00db10
+	rlwinm.	rTMP, rALT, 0, 24, 31
00db10
+	stbu	rALT, 4(rRTN)
00db10
+	beqlr-
00db10
+	rlwinm.	rTMP, rALT, 24, 24, 31
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr-
00db10
+	rlwinm.	rTMP, rALT, 16, 24, 31
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr-
00db10
+	rlwinm	rTMP, rALT, 8, 24, 31
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+#else
00db10
+	rlwinm.	rTMP, rALT, 8, 24, 31
00db10
+	stbu	rTMP, 4(rRTN)
00db10
+	beqlr-
00db10
+	rlwinm.	rTMP, rALT, 16, 24, 31
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr-
00db10
+	rlwinm.	rTMP, rALT, 24, 24, 31
00db10
+	stbu	rTMP, 1(rRTN)
00db10
+	beqlr-
00db10
+	stbu	rALT, 1(rRTN)
00db10
 #endif
00db10
+	blr
00db10
 
00db10
 /* Oh well.  In this case, we just do a byte-by-byte copy.  */
00db10
 	.align 4
00db10
 	nop
00db10
 L(unaligned):
00db10
 	lbz	rWORD, 0(rSRC)
00db10
-	addi	rDEST, rRTN, -1
00db10
+	addi	rRTN, rRTN, -1
00db10
 	cmpwi	rWORD, 0
00db10
 	beq-	L(u2)
00db10
 
00db10
 L(u0):	lbzu	rALT, 1(rSRC)
00db10
-	stbu	rWORD, 1(rDEST)
00db10
+	stbu	rWORD, 1(rRTN)
00db10
 	cmpwi	rALT, 0
00db10
 	beq-	L(u1)
00db10
 	nop		/* Let 601 load start of loop.  */
00db10
 	lbzu	rWORD, 1(rSRC)
00db10
-	stbu	rALT, 1(rDEST)
00db10
+	stbu	rALT, 1(rRTN)
00db10
 	cmpwi	rWORD, 0
00db10
 	bne+	L(u0)
00db10
-L(u2):	stb	rWORD, 1(rDEST)
00db10
-	/* GKM FIXME: check high bound.  */
00db10
+L(u2):	stbu	rWORD, 1(rRTN)
00db10
 	blr
00db10
-L(u1):	stb	rALT, 1(rDEST)
00db10
-	/* GKM FIXME: check high bound.  */
00db10
+L(u1):	stbu	rALT, 1(rRTN)
00db10
 	blr
00db10
+END (FUNC_NAME)
00db10
 
00db10
-END (BP_SYM (strcpy))
00db10
+#ifndef USE_AS_STPCPY
00db10
 libc_hidden_builtin_def (strcpy)
00db10
+#endif