Fix bugs and improve performance of niagara memset/bzero.

	* sysdeps/sparc/sparc64/multiarch/memset-niagara1.S: Unroll main
	loop to 256 bytes instead of 64 bytes and fix test signedness.

diff --git a/ChangeLog b/ChangeLog
index 9069eb3..c24c063 100644 (file)
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,5 +1,8 @@
 2012-03-28  David S. Miller  <davem@davemloft.net>
 
+       * sysdeps/sparc/sparc64/multiarch/memset-niagara1.S: Unroll main
+       loop to 256 bytes instead of 64 bytes and fix test signedness.
+
        * sysdeps/sparc/Makefile: Add -fPIC to ASFLAGS-.os here....
        * sysdeps/sparc/sparc32/Makefile: rather than here...
        * sysdeps/sparc/sparc64/Makefile: and here.

diff --git a/sysdeps/sparc/sparc64/multiarch/memset-niagara1.S b/sysdeps/sparc/sparc64/multiarch/memset-niagara1.S
index 5247bb3..1d1b3cc 100644 (file)
--- a/sysdeps/sparc/sparc64/multiarch/memset-niagara1.S
+++ b/sysdeps/sparc/sparc64/multiarch/memset-niagara1.S
@@ -1,5 +1,5 @@
 /* Set a block of memory to some byte value.  For SUN4V Niagara.
-   Copyright (C) 2006, 2008 Free Software Foundation, Inc.
+   Copyright (C) 2006, 2008, 2012 Free Software Foundation, Inc.
    This file is part of the GNU C Library.
    Contributed by David S. Miller (davem@davemloft.net)
 
@@ -60,7 +60,7 @@ ENTRY(__bzero_niagara1)
        wr              %g0, ASI_P, %asi
 
        cmp             %o1, 15
-       bl,pn           %icc, 70f
+       blu,pn          %XCC, 70f
         andcc          %o0, 0x7, %g1
        be,pt           %XCC, 2f
         mov            8, %g2
@@ -71,7 +71,7 @@ ENTRY(__bzero_niagara1)
        bne,pt          %XCC, 1b
         add            %o0, 1, %o0
 2:     cmp             %o1, 128
-       bl,pn           %icc, 60f
+       blu,pn          %XCC, 60f
         andcc          %o0, (64 - 1), %g1
        be,pt           %XCC, 40f
         mov            64, %g2
@@ -86,6 +86,51 @@ ENTRY(__bzero_niagara1)
        wr              %g0, ASI_BLK_INIT_QUAD_LDD_P, %asi
        andn            %o1, (64 - 1), %g1
        sub             %o1, %g1, %o1
+
+       andn            %g1, (256 - 1), %g2
+       brz,pt          %g2, 50f
+        and            %g1, (256 - 1), %g1
+
+45:
+       stxa            %o2, [%o0 + 0x00] %asi
+       stxa            %o2, [%o0 + 0x08] %asi
+       stxa            %o2, [%o0 + 0x10] %asi
+       stxa            %o2, [%o0 + 0x18] %asi
+       stxa            %o2, [%o0 + 0x20] %asi
+       stxa            %o2, [%o0 + 0x28] %asi
+       stxa            %o2, [%o0 + 0x30] %asi
+       stxa            %o2, [%o0 + 0x38] %asi
+       stxa            %o2, [%o0 + 0x40] %asi
+       stxa            %o2, [%o0 + 0x48] %asi
+       stxa            %o2, [%o0 + 0x50] %asi
+       stxa            %o2, [%o0 + 0x58] %asi
+       stxa            %o2, [%o0 + 0x60] %asi
+       stxa            %o2, [%o0 + 0x68] %asi
+       stxa            %o2, [%o0 + 0x70] %asi
+       stxa            %o2, [%o0 + 0x78] %asi
+       stxa            %o2, [%o0 + 0x80] %asi
+       stxa            %o2, [%o0 + 0x88] %asi
+       stxa            %o2, [%o0 + 0x90] %asi
+       stxa            %o2, [%o0 + 0x98] %asi
+       stxa            %o2, [%o0 + 0xa0] %asi
+       stxa            %o2, [%o0 + 0xa8] %asi
+       stxa            %o2, [%o0 + 0xb0] %asi
+       stxa            %o2, [%o0 + 0xb8] %asi
+       stxa            %o2, [%o0 + 0xc0] %asi
+       stxa            %o2, [%o0 + 0xc8] %asi
+       stxa            %o2, [%o0 + 0xd0] %asi
+       stxa            %o2, [%o0 + 0xd8] %asi
+       stxa            %o2, [%o0 + 0xe0] %asi
+       stxa            %o2, [%o0 + 0xe8] %asi
+       stxa            %o2, [%o0 + 0xf0] %asi
+       stxa            %o2, [%o0 + 0xf8] %asi
+       subcc           %g2, 256, %g2
+       bne,pt          %XCC, 45b
+        add            %o0, 256, %o0
+
+       brz,pn          %g1, 55f
+        nop
+
 50:
        stxa            %o2, [%o0 + 0x00] %asi
        stxa            %o2, [%o0 + 0x08] %asi
@@ -99,6 +144,7 @@ ENTRY(__bzero_niagara1)
        bne,pt          %XCC, 50b
         add            %o0, 64, %o0
 
+55:
        wr              %g0, ASI_P, %asi
        brz,pn          %o1, 80f
 60:
@@ -115,7 +161,7 @@ ENTRY(__bzero_niagara1)
 70:
 1:     stba            %o2, [%o0 + 0x00] %asi
        subcc           %o1, 1, %o1
-       bne,pt          %icc, 1b
+       bne,pt          %XCC, 1b
         add            %o0, 1, %o0
 
        /* fallthrough */