[src/trunk]: src/common/lib/libc/arch/aarch64/string * aarch64/memset.S didn'...

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

details:   https://anonhg.NetBSD.org/src/rev/bdfcb00a4175
branches:  trunk
changeset: 826353:bdfcb00a4175
user:      ryo <ryo%NetBSD.org@localhost>
date:      Tue Aug 29 15:00:23 2017 +0000

description:
* aarch64/memset.S didn't work! fixed some bugs.
* maximum size of DCZID_EL0:BS (2048) supported.

diffstat:

 common/lib/libc/arch/aarch64/string/memset.S |  36 ++++++++++-----------------
 1 files changed, 13 insertions(+), 23 deletions(-)

diffs (74 lines):

diff -r 98aa7f690bb4 -r bdfcb00a4175 common/lib/libc/arch/aarch64/string/memset.S
--- a/common/lib/libc/arch/aarch64/string/memset.S      Tue Aug 29 12:48:50 2017 +0000
+++ b/common/lib/libc/arch/aarch64/string/memset.S      Tue Aug 29 15:00:23 2017 +0000
@@ -1,4 +1,4 @@
-/* $NetBSD: memset.S,v 1.1 2014/08/10 05:47:35 matt Exp $ */
+/* $NetBSD: memset.S,v 1.2 2017/08/29 15:00:23 ryo Exp $ */
 
 /*-
  * Copyright (c) 2014 The NetBSD Foundation, Inc.
@@ -133,7 +133,7 @@
        add     x13, x15, x2    /* get ending address */
        asr     x13, x13, x9    /* "ending" block numebr */
        cmp     x13, x12        /* how many blocks? */
-       b.eq    .Lfilled        /*   none, do it 16 bytes at a time */
+       b.ls    .Lfilled        /*   none, do it 16 bytes at a time */
 
        /*
         * Now we have one or more blocks to deal with.  First now we need
@@ -144,7 +144,7 @@
 
        sub     x7, x10, x7     /* subtract offset from block length */
        sub     x2, x2, x7      /* subtract that from length */
-       asr     x7, x7, #2      /* qword -> word */
+       asr     x7, x7, #4      /* length -> N*16 */
 
        tbz     x15, #0, .Lzero_hword_aligned
        strb    wzr, [x15], #1
@@ -158,28 +158,18 @@
        tbz     x15, #3, .Lzero_qword_aligned
        str     xzr, [x15], #8
 .Lzero_qword_aligned:
-       cbz     x7, .Lblock_aligned /* no qwords? just branch */
-       adr     x6, .Lblock_aligned
-       sub     x6, x6, x7      /* backup to write the last N qwords */
-       br      x6              /* and do it */
+       cbz     x7, .Lblock_aligned     /* less than 16 bytes? just branch */
+       adr     x6, .Lunrolled_end
+       sub     x6, x6, x7, lsl #2      /* backup to write the last N insn */
+       br      x6                      /* and do it */
+
        /*
-        * This is valid for cache lines <= 256 bytes.
+        * The maximum size of DCZID_EL0:BS supported is 2048 bytes.
         */
-       stp     xzr, xzr, [x15], #16
-       stp     xzr, xzr, [x15], #16
-       stp     xzr, xzr, [x15], #16
+       .rept (2048 / 16) - 1
        stp     xzr, xzr, [x15], #16
-       stp     xzr, xzr, [x15], #16
-       stp     xzr, xzr, [x15], #16
-       stp     xzr, xzr, [x15], #16
-       stp     xzr, xzr, [x15], #16
-       stp     xzr, xzr, [x15], #16
-       stp     xzr, xzr, [x15], #16
-       stp     xzr, xzr, [x15], #16
-       stp     xzr, xzr, [x15], #16
-       stp     xzr, xzr, [x15], #16
-       stp     xzr, xzr, [x15], #16
-       stp     xzr, xzr, [x15], #16
+       .endr
+.Lunrolled_end:
 
 /*
  * Now we are block aligned.
@@ -193,7 +183,7 @@
        ret
 
 .Lblock_done:
-       and     x2, x2, x12     /* make positive again */
+       and     x2, x2, x11     /* make positive again */
        mov     x6, xzr         /* fill 2nd xword */
        b       .Lqword_loop    /* and finish filling */