Optimize mempcpy on sparc.

	* sysdeps/sparc/sparc32/memcpy.S: Implement mempcpy using a stub
	that branches into memcpy.
	* sysdeps/sparc/sparc64/memcpy.S: Likewise.
	* sysdeps/sparc/sparc64/multiarch/memcpy-niagara1.S: Likewise.
	* sysdeps/sparc/sparc64/multiarch/memcpy-niagara2.S: Likewise.
	* sysdeps/sparc/sparc64/multiarch/memcpy-ultra3.S: Likewise.
	* sysdeps/sparc/sparc64/multiarch/memcpy.S: Add mempcpy multiarch
	bits.
	* sysdeps/sparc/sparc64/rtld-memcpy.c: Include generic mempcpy
	implementation too.
	* sysdeps/sparc/mempcpy.S: New file.

4 files changed, 82 insertions, 4 deletions

diff --git a/sysdeps/sparc/sparc64/multiarch/memcpy-niagara1.S b/sysdeps/sparc/sparc64/multiarch/memcpy-niagara1.S
index 8bada0eeec..6ba1b0c6e9 100644
--- a/sysdeps/sparc/sparc64/multiarch/memcpy-niagara1.S
+++ b/sysdeps/sparc/sparc64/multiarch/memcpy-niagara1.S
@@ -43,13 +43,19 @@
 
 	.text
 
+ENTRY(__mempcpy_niagara1)
+	ba,pt		%XCC, 101f
+	 add		%o0, %o2, %g5
+END(__mempcpy_niagara1)
+
 	.align		32
 ENTRY(__memcpy_niagara1)
+100:	/* %o0=dst, %o1=src, %o2=len */
+	mov		%o0, %g5
+101:
 # ifndef USE_BPR
 	srl		%o2, 0, %o2
 # endif
-100:	/* %o0=dst, %o1=src, %o2=len */
-	mov		%o0, %g5
 	cmp		%o2, 0
 	be,pn		%XCC, 85f
 218:	 or		%o0, %o1, %o3
diff --git a/sysdeps/sparc/sparc64/multiarch/memcpy-niagara2.S b/sysdeps/sparc/sparc64/multiarch/memcpy-niagara2.S
index ccbb0252f3..0e9442de5f 100644
--- a/sysdeps/sparc/sparc64/multiarch/memcpy-niagara2.S
+++ b/sysdeps/sparc/sparc64/multiarch/memcpy-niagara2.S
@@ -145,13 +145,19 @@
 
 	.text
 
+ENTRY(__mempcpy_niagara2)
+	ba,pt		%XCC, 101f
+	 add		%o0, %o2, %g5
+END(__mempcpy_niagara2)
+
 	.align		32
 ENTRY(__memcpy_niagara2)
+100:	/* %o0=dst, %o1=src, %o2=len */
+	mov		%o0, %g5
+101:
 # ifndef USE_BPR
 	srl		%o2, 0, %o2
 # endif
-100:	/* %o0=dst, %o1=src, %o2=len */
-	mov		%o0, %g5
 	cmp		%o2, 0
 	be,pn		%XCC, 85f
 218:	 or		%o0, %o1, %o3
diff --git a/sysdeps/sparc/sparc64/multiarch/memcpy-ultra3.S b/sysdeps/sparc/sparc64/multiarch/memcpy-ultra3.S
index 7e21665d5f..0784ba9b5d 100644
--- a/sysdeps/sparc/sparc64/multiarch/memcpy-ultra3.S
+++ b/sysdeps/sparc/sparc64/multiarch/memcpy-ultra3.S
@@ -38,6 +38,11 @@
 
 	.text
 
+ENTRY(__mempcpy_ultra3)
+	ba,pt		%XCC, 101f
+	 add		%o0, %o2, %g5
+END(__mempcpy_ultra3)
+
 	/* Special/non-trivial issues of this code:
 	 *
 	 * 1) %o5 is preserved from VISEntryHalf to VISExitHalf
@@ -57,6 +62,7 @@ ENTRY(__memcpy_ultra3)
 
 100: /* %o0=dst, %o1=src, %o2=len */
 	mov		%o0, %g5
+101:
 	cmp		%o2, 0
 	be,pn		%XCC, out
 218:	 or		%o0, %o1, %o3
diff --git a/sysdeps/sparc/sparc64/multiarch/memcpy.S b/sysdeps/sparc/sparc64/multiarch/memcpy.S
index 0f3751e5e7..20c72d9bbd 100644
--- a/sysdeps/sparc/sparc64/multiarch/memcpy.S
+++ b/sysdeps/sparc/sparc64/multiarch/memcpy.S
@@ -72,12 +72,72 @@ ENTRY(memcpy)
 	 mov	%o1, %o0
 END(memcpy)
 
+ENTRY(__mempcpy)
+	.type	__mempcpy, @gnu_indirect_function
+# ifdef SHARED
+	SETUP_PIC_REG_LEAF(o3, o5)
+# endif
+	andcc	%o0, HWCAP_SPARC_N2, %g0
+	be	1f
+	 andcc	%o0, HWCAP_SPARC_BLKINIT, %g0
+# ifdef SHARED
+	sethi	%gdop_hix22(__mempcpy_niagara2), %o1
+	xor	%o1, %gdop_lox10(__mempcpy_niagara2), %o1
+# else
+	set	__mempcpy_niagara2, %o1
+# endif
+	ba	10f
+	 nop
+1:	be	1f
+	 andcc	%o0, HWCAP_SPARC_ULTRA3, %g0
+# ifdef SHARED
+	sethi	%gdop_hix22(__mempcpy_niagara1), %o1
+	xor	%o1, %gdop_lox10(__mempcpy_niagara1), %o1
+# else
+	set	__mempcpy_niagara1, %o1
+# endif
+	ba	10f
+	 nop
+1:	be	9f
+	 nop
+# ifdef SHARED
+	sethi	%gdop_hix22(__mempcpy_ultra3), %o1
+	xor	%o1, %gdop_lox10(__mempcpy_ultra3), %o1
+# else
+	set	__mempcpy_ultra3, %o1
+# endif
+	ba	10f
+	 nop
+9:
+# ifdef SHARED
+	sethi	%gdop_hix22(__mempcpy_ultra1), %o1
+	xor	%o1, %gdop_lox10(__mempcpy_ultra1), %o1
+# else
+	set	__mempcpy_ultra1, %o1
+# endif
+10:
+# ifdef SHARED
+	add	%o3, %o1, %o1
+# endif
+	retl
+	 mov	%o1, %o0
+END(__mempcpy)
+
 libc_hidden_builtin_def (memcpy)
 
+libc_hidden_def (__mempcpy)
+weak_alias (__mempcpy, mempcpy)
+libc_hidden_builtin_def (mempcpy)
+
 #undef libc_hidden_builtin_def
 #define libc_hidden_builtin_def(name)
+#undef weak_alias
+#define weak_alias(x, y)
+#undef libc_hidden_def
+#define libc_hidden_def(name)
 
 #define memcpy __memcpy_ultra1
+#define __mempcpy __mempcpy_ultra1
 
 #endif