[src/trunk]: src/lib/libc/arch/sparc64/string Just use a store/load pair to g...

[mycroft <mycroft%NetBSD.org@localhost>]

details:   https://anonhg.NetBSD.org/src/rev/da41f9a5cb06
branches:  trunk
changeset: 495165:da41f9a5cb06
user:      mycroft <mycroft%NetBSD.org@localhost>
date:      Sun Jul 23 20:34:07 2000 +0000

description:
Just use a store/load pair to get the pattern into the FP reg.  This eliminates
a bunch of cruft and avoids using a v9a instruction.
In addition, eliminate 8 of the fmovda's, which we are not using the result of
anyway.
Net result is that this should be faster in all cases.

diffstat:

 lib/libc/arch/sparc64/string/bzero.S |  28 ++++------------------------
 1 files changed, 4 insertions(+), 24 deletions(-)

diffs (66 lines):

diff -r 053215b00b57 -r da41f9a5cb06 lib/libc/arch/sparc64/string/bzero.S
--- a/lib/libc/arch/sparc64/string/bzero.S      Sun Jul 23 20:22:30 2000 +0000
+++ b/lib/libc/arch/sparc64/string/bzero.S      Sun Jul 23 20:34:07 2000 +0000
@@ -1,4 +1,4 @@
-/*     $NetBSD: bzero.S,v 1.3 2000/07/08 16:02:15 eeh Exp $    */
+/*     $NetBSD: bzero.S,v 1.4 2000/07/23 20:34:07 mycroft Exp $        */
 
 /*
  * Copyright (c) 1992, 1993, 1999
@@ -51,7 +51,7 @@
 #if 0
        .asciz "@(#)bzero.s     8.1 (Berkeley) 6/4/93"
 #else
-       RCSID("$NetBSD: bzero.S,v 1.3 2000/07/08 16:02:15 eeh Exp $")
+       RCSID("$NetBSD: bzero.S,v 1.4 2000/07/23 20:34:07 mycroft Exp $")
 #endif
 #endif  /* LIBC_SCCS and not lint */
 
@@ -132,7 +132,7 @@
        bl,a,pn %icc, Lbzero_cleanup    ! Less than 8 left
         dec    8, %o1                  ! Fixup count -8
 2:
-       !! Now we're 64-bit aligned
+       !! Now we are 64-bit aligned
        cmp     %o1, 256                ! Use block clear if len > 256
        bge,pt  %xcc, Lbzero_block      ! use block store insns
         deccc  8, %o1
@@ -223,22 +223,10 @@
         dec    8, %i1
 
 2:
-       brz,pt  %i2, 4f                                 ! Do we have a pattern to load?
-        fzero  %f0                                     ! Set up FPU
-
-       btst    1, %fp
-       bnz,pt  %icc, 3f                                ! 64-bit stack?
-        nop
-       stw     %i2, [%fp + 0x28]                       ! Flush this puppy to RAM
-       membar  #StoreLoad
-       ld      [%fp + 0x28], %f0
-       ba,pt   %icc, 4f
-        fmovsa %icc, %f0, %f1
-3:     
        stx     %i2, [%fp + BIAS + 0x50]                ! Flush this puppy to RAM
        membar  #StoreLoad
        ldd     [%fp + BIAS + 0x50], %f0
-4:     
+
        fmovda  %icc, %f0, %f2                          ! Duplicate the pattern
        fmovda  %icc, %f0, %f4
        fmovda  %icc, %f0, %f6
@@ -246,14 +234,6 @@
        fmovda  %icc, %f0, %f10
        fmovda  %icc, %f0, %f12
        fmovda  %icc, %f0, %f14
-       fmovda  %icc, %f0, %f16                         ! And second bank
-       fmovda  %icc, %f0, %f18
-       fmovda  %icc, %f0, %f20
-       fmovda  %icc, %f0, %f22
-       fmovda  %icc, %f0, %f24
-       fmovda  %icc, %f0, %f26
-       fmovda  %icc, %f0, %f28
-       fmovda  %icc, %f0, %f30
        
        !! Remember: we were 8 bytes too far
        dec     56, %i1                 ! Go one iteration too far