[src/trunk]: src/common/lib/libc/arch/arm/string A few slight speedups (remov...

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

details:   https://anonhg.NetBSD.org/src/rev/d7f1625e9559
branches:  trunk
changeset: 783558:d7f1625e9559
user:      matt <matt%NetBSD.org@localhost>
date:      Sat Dec 29 05:36:57 2012 +0000

description:
A few slight speedups (remove one instruction from the main loop).

diffstat:

 common/lib/libc/arch/arm/string/strlen_armv6.S |  27 +++++++++++++++----------
 1 files changed, 16 insertions(+), 11 deletions(-)

diffs (67 lines):

diff -r 6ddd93fe234d -r d7f1625e9559 common/lib/libc/arch/arm/string/strlen_armv6.S
--- a/common/lib/libc/arch/arm/string/strlen_armv6.S    Sat Dec 29 05:00:36 2012 +0000
+++ b/common/lib/libc/arch/arm/string/strlen_armv6.S    Sat Dec 29 05:36:57 2012 +0000
@@ -29,27 +29,27 @@
 
 #include <machine/asm.h>
 
-RCSID("$NetBSD: strlen_armv6.S,v 1.1 2012/12/28 07:10:41 matt Exp $")
+RCSID("$NetBSD: strlen_armv6.S,v 1.2 2012/12/29 05:36:57 matt Exp $")
 
        .text
 ENTRY(strlen)
+       add     ip, r0, #4              /* for the final post-inc */
        ands    r1, r0, #3              /* get misalignment */
-       bic     ip, r0, #3              /* align to word boundary */
-       ldr     r3, [ip], #4            /* load first word */
-       neg     r0, r1                  /* subtract misalignment from length */
+       bic     r0, r0, #3              /* align to word boundary */
+       ldr     r3, [r0], #4            /* load first word */
        beq     .Lpre_main_loop         /*   misaligned?  no, go to loop */
        /*
         * For misaligned string, we need to make sure that the bytes before
         * the start of the string will not cause a false match to a NUL.
         */
        mvn     r2, #0                  /* create a mask */
-       and     r1, r0, #3              /* find out how many bytes to clear */
        mov     r1, r1, lsl #3          /* bytes -> bits */
 #ifdef __ARMEL__
-       mov     r2, r2, lsr r1          /* clear relavent bytes */
+       mov     r2, r2, lsl r1          /* clear relavent bytes */
 #else
-       mov     r2, r2, lsl r1          /* clear relavent bytes */
+       mov     r2, r2, lsr r1          /* clear relavent bytes */
 #endif
+       mvn     r2, r2                  /* invert mask */
        orr     r3, r3, r2              /* orr in mask for leading bytes */
 .Lpre_main_loop:
 #ifdef _ARM_ARCH_7
@@ -68,19 +68,24 @@
         */
        uqadd8  r3, r3, r1              /* magic happens here */
        mvns    r3, r3                  /* is the complemented result 0? */
-       bne     .Lreturn                /*    no, return # of bytes */
-       add     r0, r0, #4              /* add 4 to the count */
-       ldr     r3, [ip], #4            /* load next word */
+       bne     .Lreturn                /*    no, then we encountered a NUL */
+       ldr     r3, [r0], #4            /* load next word */
        b       .Lmain_loop             /* and go */
-.Lreturn:
        /*
         * We encountered a NUL.  Find out where by doing a CLZ and then
         * shifting right by 3.  That will be the number of non-NUL bytes.
         */
+.Lreturn:
 #ifdef __ARMEL__
        rev     r3, r3                  /* we want this in BE for the CLZ */
 #endif
        clz     r3, r3                  /* count how many leading zeros */
        add     r0, r0, r3, lsr #3      /* divide that by 8 and add to count */
+       /*
+        * r0 now points to 4 past the NUL due to the post-inc.  Subtract
+        * the start of the string (which also has 4 added to it to compensate
+        * for the post-inc.
+        */
+       sub     r0, r0, ip              /* subtract start to get length */
        RET
 END(strlen)