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.

diff --git a/ChangeLog b/ChangeLog
index 555c395..23b7290 100644 (file)
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,17 @@
 2012-03-28  David S. Miller  <davem@davemloft.net>
 
+       * 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.
+
        * sysdeps/sparc/sparc64/multiarch/memcpy.S: Provide a hidden def to
        the IFUNC routine in the libc case.
        * sysdeps/sparc/sparc64/multiarch/memcpy.S: Likewise.

diff --git a/sysdeps/sparc/mempcpy.S b/sysdeps/sparc/mempcpy.S
new file mode 100644 (file)
index 0000000..4c98013
--- /dev/null
+++ b/sysdeps/sparc/mempcpy.S
@@ -0,0 +1 @@
+/* mempcpy is in memcpy.S */

diff --git a/sysdeps/sparc/sparc32/memcpy.S b/sysdeps/sparc/sparc32/memcpy.S
index a36f1d6..82fa6d1 100644 (file)
--- a/sysdeps/sparc/sparc32/memcpy.S
+++ b/sysdeps/sparc/sparc32/memcpy.S
@@ -104,11 +104,17 @@
        std     %t2, [%dst + offset + offset2 + 0x08];
 
        .text
-       .align  4
+ENTRY(__mempcpy)
+       add             %o0, %o2, %g1
+       ba              101f
+        st             %g1, [%sp + 64]
+END(__mempcpy)
 
+       .align  4
 ENTRY(memcpy)          /* %o0=dst %o1=src %o2=len */
-       sub             %o0, %o1, %o4
        st              %o0, [%sp + 64]
+101:
+       sub             %o0, %o1, %o4
 9:     andcc           %o4, 3, %o5
 0:     bne             86f
         cmp            %o2, 15
@@ -641,3 +647,7 @@ ENTRY(memcpy)               /* %o0=dst %o1=src %o2=len */
 END(memcpy)
 
 libc_hidden_builtin_def (memcpy)
+
+libc_hidden_def (__mempcpy)
+weak_alias (__mempcpy, mempcpy)
+libc_hidden_builtin_def (mempcpy)

diff --git a/sysdeps/sparc/sparc64/memcpy.S b/sysdeps/sparc/sparc64/memcpy.S
index a77c4e4..668ebec 100644 (file)
--- a/sysdeps/sparc/sparc64/memcpy.S
+++ b/sysdeps/sparc/sparc64/memcpy.S
@@ -374,19 +374,24 @@ ENTRY(__memcpy_large)
         mov            %g4, %o0
 END(__memcpy_large)
 
+ENTRY(__mempcpy)
+       ba,pt           %xcc, 210f
+        add            %o0, %o2, %g4
+END(__mempcpy)
+
        .align          32
 ENTRY(memcpy)
+        mov            %o0, %g4                        /* IEU0         Group           */
 210:
 #ifndef USE_BPR
-       srl             %o2, 0, %o2                     /* IEU1         Group           */
+       srl             %o2, 0, %o2                     /* IEU1                         */
 #endif 
        brz,pn          %o2, 209b                       /* CTI          Group           */
-        mov            %o0, %g4                        /* IEU0                         */
-218:   cmp             %o2, 15                         /* IEU1         Group           */
-       bleu,pn         %xcc, 208b                      /* CTI                          */
-        cmp            %o2, (64 * 6)                   /* IEU1         Group           */
-       bgeu,pn         %xcc, 200b                      /* CTI                          */
-        andcc          %o0, 7, %g2                     /* IEU1         Group           */
+218:    cmp            %o2, 15                         /* IEU1                         */
+       bleu,pn         %xcc, 208b                      /* CTI          Group           */
+        cmp            %o2, (64 * 6)                   /* IEU1                         */
+       bgeu,pn         %xcc, 200b                      /* CTI          Group           */
+        andcc          %o0, 7, %g2                     /* IEU1                         */
        sub             %o0, %o1, %g5                   /* IEU0                         */
        andcc           %g5, 3, %o5                     /* IEU1         Group           */
        bne,pn          %xcc, 212f                      /* CTI                          */
@@ -569,3 +574,7 @@ ENTRY(memcpy)
 END(memcpy)
 
 libc_hidden_builtin_def (memcpy)
+
+libc_hidden_def (__mempcpy)
+weak_alias (__mempcpy, mempcpy)
+libc_hidden_builtin_def (mempcpy)

diff --git a/sysdeps/sparc/sparc64/multiarch/memcpy-niagara1.S b/sysdeps/sparc/sparc64/multiarch/memcpy-niagara1.S
index 8bada0e..6ba1b0c 100644 (file)
--- 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 ccbb025..0e9442d 100644 (file)
--- 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 7e21665..0784ba9 100644 (file)
--- 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 0f3751e..20c72d9 100644 (file)
--- 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
 

diff --git a/sysdeps/sparc/sparc64/rtld-memcpy.c b/sysdeps/sparc/sparc64/rtld-memcpy.c
index 5e50e6e..b1b0647 100644 (file)
--- a/sysdeps/sparc/sparc64/rtld-memcpy.c
+++ b/sysdeps/sparc/sparc64/rtld-memcpy.c
@@ -1 +1,2 @@
 #include <string/memcpy.c>
+#include <string/mempcpy.c>