git: 9f50aa45be18 - main - [PowerPC64] Port optimized strcpy to PPC64LE

[Leandro Lupori]

The branch main has been updated by luporl:

URL: https://cgit.FreeBSD.org/src/commit/?id=9f50aa45be18b9b11b14345fe0a137d1ac51a391

commit 9f50aa45be18b9b11b14345fe0a137d1ac51a391
Author:     Leandro Lupori <luporl at FreeBSD.org>
AuthorDate: 2021-03-25 16:14:00 +0000
Commit:     Leandro Lupori <luporl at FreeBSD.org>
CommitDate: 2021-03-25 16:20:12 +0000

    [PowerPC64] Port optimized strcpy to PPC64LE
    
    Submitted by:           Bruno Larsen <bruno.larsen at eldorado.org.br>
    Reviewed by:            luporl, bdragon (IRC)
    MFC after:              1 week
    Sponsored by:           Eldorado Research Institute (eldorado.org.br)
    Differential Revision:  https://reviews.freebsd.org/D29067
---
 lib/libc/powerpc64/string/Makefile.inc       |  7 +--
 lib/libc/powerpc64/string/strcpy_arch_2_05.S | 66 ++++++++++++++++++++++++++--
 2 files changed, 64 insertions(+), 9 deletions(-)

diff --git a/lib/libc/powerpc64/string/Makefile.inc b/lib/libc/powerpc64/string/Makefile.inc
index 14f0845595c9..486ca47a44be 100644
--- a/lib/libc/powerpc64/string/Makefile.inc
+++ b/lib/libc/powerpc64/string/Makefile.inc
@@ -12,12 +12,7 @@ MDSRCS+= \
 	memmove_resolver.c \
 	strncpy_arch_2_05.S \
 	strncpy.c \
-	strncpy_resolver.c
-
-# XXX Port strcpy to LE.
-.if ${MACHINE_ARCH} == "powerpc64"
-MDSRCS+= \
+	strncpy_resolver.c \
 	strcpy_arch_2_05.S \
 	strcpy.c \
 	strcpy_resolver.c
-.endif
diff --git a/lib/libc/powerpc64/string/strcpy_arch_2_05.S b/lib/libc/powerpc64/string/strcpy_arch_2_05.S
index 23ffb5c2a75e..7f42fd813395 100644
--- a/lib/libc/powerpc64/string/strcpy_arch_2_05.S
+++ b/lib/libc/powerpc64/string/strcpy_arch_2_05.S
@@ -26,9 +26,6 @@
  * SUCH DAMAGE.
  */
 
-#if !defined(__BIG_ENDIAN__)
-#error "Optimized SRTCPY is only supported by big-endian architecture!"
-#endif
 
 #include <machine/asm.h>
 __FBSDID("$FreeBSD$");
@@ -71,6 +68,7 @@ ENTRY(__strcpy_arch_2_05)
 	beq	cr7,.Lcopy_dw_loop
 
 	addi	%r8,%r8,8   /* Forward r8 to use std instruction. */
+#if defined(__BIG_ENDIAN__)
 /* Find where the zero is located. */
 .Lcheck_zero:
 	rldicr.	%r5,%r0,0,7
@@ -136,6 +134,68 @@ ENTRY(__strcpy_arch_2_05)
 .Lfound_on_byte_0:
 	srdi	%r6,%r0,56
 	stb	%r6,0(%r8)
+#elif defined(__LITTLE_ENDIAN__)
+/* Find where the zero is located. */
+.Lcheck_zero:
+	andi.	%r7,%r0,0xff
+	beq	.Lfound_on_byte_0
+	andi.	%r7,%r0,0xff00
+	beq	.Lfound_on_byte_1
+	andis.	%r7,%r0,0xff
+	beq	.Lfound_on_byte_2
+	andis.	%r7,%r0,0xff00
+	beq	.Lfound_on_byte_3
+	rldicr.	%r7,%r0,24,7
+	beq	.Lfound_on_byte_4
+	rldicr.	%r7,%r0,16,7
+	beq	.Lfound_on_byte_5
+	rldicr.	%r7,%r0,8,7
+	beq	.Lfound_on_byte_6
+
+/* Copy the last string bytes according to the string end position. */
+.Lfound_on_byte_7:
+	std	%r0,0(%r8)
+	b	.Lexit
+
+.Lfound_on_byte_6:
+	stw	%r0,0(%r8)
+	srdi	%r6,%r0,32
+	sth	%r6,4(%r8)
+	srdi	%r6,%r0,48
+	stb	%r6,6(%r8)
+	b	.Lexit
+
+.Lfound_on_byte_5:
+	stw	%r0,0(%r8)
+	srdi	%r6,%r0,32
+	sth	%r6,4(%r8)
+	b	.Lexit
+
+.Lfound_on_byte_4:
+	stw	%r0,0(%r8)
+	srdi	%r6,%r0,32
+	stb	%r6,4(%r8)
+	b	.Lexit
+
+.Lfound_on_byte_3:
+	stw	%r0,0(%r8)
+	b	.Lexit
+
+.Lfound_on_byte_2:
+	sth	%r0,0(%r8)
+	srdi	%r6,%r0,16
+	stb	%r6,2(%r8)
+	b	.Lexit
+
+.Lfound_on_byte_1:
+	sth	%r0,0(%r8)
+	b	.Lexit
+
+.Lfound_on_byte_0:
+	stb	%r0,0(%r8)
+#else
+#error "Unable to determine Endianness"
+#endif
 .Lexit:
 	blr