powerpc: Optimized strcmp for POWER8/PPC64

This patch adds an optimized POWER8 strcmp using unaligned accesses.
The algorithm first check the initial 16 bytes, then align the first
function source and uses unaligned loads on second argument only.
Aditional checks for page boundaries are done for unaligned cases

1 files changed, 257 insertions, 0 deletions

diff --git a/sysdeps/powerpc/powerpc64/power8/strcmp.S b/sysdeps/powerpc/powerpc64/power8/strcmp.S
new file mode 100644
index 0000000000..223d891c2e
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/power8/strcmp.S
@@ -0,0 +1,257 @@
+/* Optimized strcmp implementation for PowerPC64/POWER8.
+   Copyright (C) 2015 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library; if not, see
+   <http://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+
+/* Implements the function
+
+   size_t [r3] strcmp (const char *s1 [r3], const char *s2 [r4])
+
+   The implementation uses unaligned doubleword access to avoid specialized
+   code paths depending of data alignment.  Although recent powerpc64 uses
+   64K as default, the page cross handling assumes minimum page size of
+   4k.  */
+
+EALIGN (strcmp, 4, 0)
+	li	r0,0
+
+	/* Check if [s1]+32 or [s2]+32 will cross a 4K page boundary using
+	   the code:
+
+	    (((size_t) s1) % PAGE_SIZE > (PAGE_SIZE - ITER_SIZE))
+
+	   with PAGE_SIZE being 4096 and ITER_SIZE begin 32.  */
+
+	rldicl	r7,r3,0,52
+	rldicl	r9,r4,0,52
+	cmpldi	cr7,r7,4096-32
+	bgt	cr7,L(pagecross_check)
+	cmpldi	cr5,r9,4096-32
+	bgt	cr5,L(pagecross_check)
+
+	/* For short string up to 32 bytes, load both s1 and s2 using
+	   unaligned dwords and compare.  */
+	ld	r8,0(r3)
+	ld	r10,0(r4)
+	cmpb	r12,r8,r0
+	cmpb	r11,r8,r10
+	orc.	r9,r12,r11
+	bne	cr0,L(different_nocmpb)
+
+	ld	r8,8(r3)
+	ld	r10,8(r4)
+	cmpb	r12,r8,r0
+	cmpb	r11,r8,r10
+	orc.	r9,r12,r11
+	bne	cr0,L(different_nocmpb)
+
+	ld	r8,16(r3)
+	ld	r10,16(r4)
+	cmpb	r12,r8,r0
+	cmpb	r11,r8,r10
+	orc.	r9,r12,r11
+	bne	cr0,L(different_nocmpb)
+
+	ld	r8,24(r3)
+	ld	r10,24(r4)
+	cmpb	r12,r8,r0
+	cmpb	r11,r8,r10
+	orc.	r9,r12,r11
+	bne	cr0,L(different_nocmpb)
+
+	addi	r7,r3,32
+	addi	r4,r4,32
+
+L(align_8b):
+	/* Now it has checked for first 32 bytes, align source1 to doubleword
+	   and adjust source2 address.  */
+	rldicl	r9,r7,0,61	/* source1 alignment to doubleword  */
+	subf	r4,r9,r4	/* Adjust source2 address based on source1
+				   alignment.  */
+	rldicr	r7,r7,0,60	/* Align source1 to doubleword.  */
+
+	/* At this point, source1 alignment is 0 and source2 alignment is
+	   between 0 and 7.  Check is source2 alignment is 0, meaning both
+	   sources have the same alignment.  */
+	andi.	r9,r4,0x7
+	bne	cr0,L(loop_diff_align)
+
+	/* If both source1 and source2 are doubleword aligned, there is no
+	   need for page boundary cross checks.  */
+
+	ld	r8,0(r7)
+	ld	r10,0(r4)
+	cmpb	r12,r8,r0
+	cmpb	r11,r8,r10
+	orc.	r9,r12,r11
+	bne	cr0,L(different_nocmpb)
+
+	.align 4
+L(loop_equal_align):
+	ld	r8,8(r7)
+	ld	r10,8(r4)
+	cmpb	r12,r8,r0
+	cmpb	r11,r8,r10
+	orc.	r9,r12,r11
+	bne	cr0,L(different_nocmpb)
+
+	ld	r8,16(r7)
+	ld	r10,16(r4)
+	cmpb	r12,r8,r0
+	cmpb	r11,r8,r10
+	orc.	r9,r12,r11
+	bne	cr0,L(different_nocmpb)
+
+	ldu	r8,24(r7)
+	ldu	r10,24(r4)
+	cmpb	r12,r8,r0
+	cmpb	r11,r8,r10
+	orc.	r9,r12,r11
+	bne	cr0,L(different_nocmpb)
+
+	b	L(loop_equal_align)
+
+	/* A zero byte was found in r8 (s1 dword), r9 contains the cmpb
+	   result and r10 the dword from s2.  To code isolate the byte
+	   up to end (including the '\0'), masking with 0xFF the remaining
+	   ones:
+
+           #if __LITTLE_ENDIAN__
+	     (__builtin_ffsl (x) - 1) = counting trailing zero bits
+	     r9 = (__builtin_ffsl (r9) - 1) + 8;
+	     r9 = -1UL << r9
+	   #else
+	     r9  = __builtin_clzl (r9) + 8;
+	     r9  = -1UL >> r9
+	   #endif
+	     r8  = r8  | r9
+	     r10 = r10 | r9  */
+
+#ifdef __LITTLE_ENDIAN__
+	nor 	r9,r9,r9
+L(different_nocmpb):
+	neg	r3,r9
+	and	r9,r9,r3
+	cntlzd	r9,r9
+	subfic	r9,r9,63
+#else
+	not	r9,r9
+L(different_nocmpb):
+	cntlzd	r9,r9
+	subfic	r9,r9,56
+#endif
+	srd	r3,r8,r9
+	srd	r10,r10,r9
+	rldicl	r10,r10,0,56
+	rldicl	r3,r3,0,56
+	subf	r3,r10,r3
+	extsw	r3,r3
+	blr
+
+	.align	4
+L(pagecross_check):
+	subfic	r9,r9,4096
+	subfic	r7,r7,4096
+	cmpld	cr7,r7,r9
+	bge	cr7,L(pagecross)
+	mr	r7,r9
+
+	/* If unaligned 16 bytes reads across a 4K page boundary, it uses
+	   a simple byte a byte comparison until the page alignment for s1
+	   is reached.  */
+L(pagecross):
+	add	r7,r3,r7
+	subf	r9,r3,r7
+	mtctr	r9
+
+	.align	4
+L(pagecross_loop):
+	/* Loads a byte from s1 and s2, compare if *s1 is equal to *s2
+	   and if *s1 is '\0'.  */
+	lbz	r9,0(r3)
+	lbz	r10,0(r4)
+	addi	r3,r3,1
+	addi	r4,r4,1
+	cmplw	cr7,r9,r10
+	cmpdi	cr5,r9,r0
+	bne	cr7,L(pagecross_ne)
+	beq	cr5,L(pagecross_nullfound)
+	bdnz	L(pagecross_loop)
+	b	L(align_8b)
+
+	.align	4
+	/* The unaligned read of source2 will cross a 4K page boundary,
+	   and the different byte or NULL maybe be in the remaining page
+	   bytes. Since it can not use the unaligned load, the algorithm
+	   reads and compares 8 bytes to keep source1 doubleword aligned.  */
+L(check_source2_byte):
+	li	r9,8
+	mtctr	r9
+
+	.align	4
+L(check_source2_byte_loop):
+	lbz	r9,0(r7)
+	lbz	r10,0(r4)
+	addi	r7,r7,1
+	addi	r4,r4,1
+	cmplw	cr7,r9,10
+	cmpdi	r5,r9,0
+	bne	cr7,L(pagecross_ne)
+	beq	cr5,L(pagecross_nullfound)
+	bdnz	L(check_source2_byte_loop)
+
+	/* If source2 is unaligned to doubleword, the code needs to check
+	   on each interation if the unaligned doubleword access will cross
+	   a 4k page boundary.  */
+	.align	5
+L(loop_unaligned):
+	ld	r8,0(r7)
+	ld	r10,0(r4)
+	cmpb	r12,r8,r0
+	cmpb	r11,r8,r10
+	orc.	r9,r12,r11
+	bne	cr0,L(different_nocmpb)
+	addi	r7,r7,8
+	addi	r4,r4,8
+
+L(loop_diff_align):
+	/* Check if [src2]+8 cross a 4k page boundary:
+
+	     srcin2 % PAGE_SIZE > (PAGE_SIZE - 8)
+
+	     with PAGE_SIZE being 4096.  */
+	rldicl	r9,r4,0,52
+	cmpldi	cr7,r9,4088
+	ble	cr7,L(loop_unaligned)
+	b	L(check_source2_byte)
+
+	.align	4
+L(pagecross_ne):
+	extsw	r3,r9
+	mr	r9,r10
+L(pagecross_retdiff):
+	subf	r9,r9,r3
+	extsw	r3,r9
+	blr
+
+	.align	4
+L(pagecross_nullfound):
+	li	r3,0
+	b	L(pagecross_retdiff)
+END (strcmp)
+libc_hidden_builtin_def (strcmp)