[src/trunk]: src/common/lib/libc/arch/aarch64/string Re-do previous aarch64eb...

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

details:   https://anonhg.NetBSD.org/src/rev/d15afa57ef7a
branches:  trunk
changeset: 943748:d15afa57ef7a
user:      jakllsch <jakllsch%NetBSD.org@localhost>
date:      Wed Sep 09 14:49:27 2020 +0000

description:
Re-do previous aarch64eb strlen fix more simply and correctly.

diffstat:

 common/lib/libc/arch/aarch64/string/strlen.S |  22 ++++++++--------------
 1 files changed, 8 insertions(+), 14 deletions(-)

diffs (59 lines):

diff -r e8888f928ddf -r d15afa57ef7a common/lib/libc/arch/aarch64/string/strlen.S
--- a/common/lib/libc/arch/aarch64/string/strlen.S      Wed Sep 09 14:38:41 2020 +0000
+++ b/common/lib/libc/arch/aarch64/string/strlen.S      Wed Sep 09 14:49:27 2020 +0000
@@ -1,4 +1,4 @@
-/* $NetBSD: strlen.S,v 1.4 2020/09/05 20:24:43 jakllsch Exp $ */
+/* $NetBSD: strlen.S,v 1.5 2020/09/09 14:49:27 jakllsch Exp $ */
 
 /*-
  * Copyright (c) 2014 The NetBSD Foundation, Inc.
@@ -31,7 +31,7 @@
 
 #include <machine/asm.h>
 
-RCSID("$NetBSD: strlen.S,v 1.4 2020/09/05 20:24:43 jakllsch Exp $")
+RCSID("$NetBSD: strlen.S,v 1.5 2020/09/09 14:49:27 jakllsch Exp $")
 
 #ifdef STRNLEN
 #define FUNCNAME       strnlen
@@ -66,12 +66,11 @@
         */
        add     x4, x4, x0              /* make dword aligned */
        ldr     x7, [x4], #8            /* load dword */
+#ifdef __AARCH64EB__
+       rev     x7, x7                  /* convert to LE */
+#endif
        lsl     x3, x3, #3              /* convert bytes to bits */
-#ifdef __AARCH64EB__
-       lsr     x5, x11, x3             /* make mask for BE */
-#else
        lsl     x5, x11, x3             /* make mask for LE */
-#endif
        eor     x5, x5, x11             /* invert mask */
        orr     x7, x7, x5              /* prevent NULs */
        b       .Lstrlen_dword_loop_noload
@@ -82,6 +81,9 @@
        b.hs    .Lstrlen_done
 #endif
        ldr     x7, [x4], #8            /* load dword */
+#ifdef __AARCH64EB__
+       rev     x7, x7                  /* convert to LE */
+#endif
 .Lstrlen_dword_loop_noload:
        /*
         * Use the formula (X - 1) & ~(X | 0x7f) to find NUL bytes.
@@ -96,14 +98,6 @@
        /*
         * We know there is a NUL in this dword.  Use clz to find it.
         */
-#ifdef __AARCH64EB__
-       /* avoid BE problem due to carry propagation if last non-NUL is \x01 */
-       ldr     x7, [x4, #-8]           /* reload dword */
-       rev     x7, x7                  /* byte swap */
-       sub     x6, x7, x11             /* a = X - 1 */
-       orr     x7, x7, #MASK8_0x7f     /* b = X | 0x7f */
-       bic     x6, x6, x7              /* a & ~b */
-#endif
        rev     x6, x6                  /* convert to BE */
        clz     x6, x6                  /* find null byte */
        add     x0, x0, x6, lsr #3      /* add offset to the length */