powerpc: strcmp optimization for power9

Vectorized loops are used for strings > 32B when compared
to power8 optimization.

Tested on power9 ppc64le simulator.

1 files changed, 278 insertions, 0 deletions

diff --git a/sysdeps/powerpc/powerpc64/power9/strcmp.S b/sysdeps/powerpc/powerpc64/power9/strcmp.S
new file mode 100644
index 0000000000..754d508804
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/power9/strcmp.S
@@ -0,0 +1,278 @@
+/* Optimized strcmp implementation for PowerPC64/POWER9.
+   Copyright (C) 2016 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/>.  */
+#ifdef __LITTLE_ENDIAN__
+#include <sysdep.h>
+
+/* Implements the function
+
+   int [r3] strcmp (const char *s1 [r3], const char *s2 [r4])
+
+   The implementation uses unaligned doubleword access for first 32 bytes
+   as in POWER8 patch and uses vectorised loops after that.  */
+
+/* TODO: Change this to actual instructions when minimum binutils is upgraded
+   to 2.27. Macros are defined below for these newer instructions in order
+   to maintain compatibility.  */
+# define VCTZLSBB(r,v) .long (0x10010602 | ((r)<<(32-11)) | ((v)<<(32-21)))
+
+# define VEXTUBRX(t,a,b) .long (0x1000070d \
+				| ((t)<<(32-11))  \
+				| ((a)<<(32-16))  \
+				| ((b)<<(32-21)) )
+
+# define VCMPNEZB(t,a,b) .long (0x10000507 \
+				| ((t)<<(32-11))  \
+				| ((a)<<(32-16))  \
+				| ((b)<<(32-21)) )
+
+/* Get 16 bytes for unaligned case.
+   reg1: Vector to hold next 16 bytes.
+   reg2: Address to read from.
+   reg3: Permute control vector.  */
+# define GET16BYTES(reg1, reg2, reg3) \
+	lvx	reg1, 0, reg2; \
+	vperm	v8, v2, reg1, reg3; \
+	vcmpequb.	v8, v0, v8; \
+	beq	cr6, 1f; \
+	vspltisb	v9, 0; \
+	b	2f; \
+	.align 4; \
+1: \
+	addi    r6, reg2, 16; \
+	lvx     v9, 0, r6; \
+2: \
+	vperm   reg1, v9, reg1, reg3;
+
+/* TODO: change this to .machine power9 when the minimum required binutils
+   allows it.  */
+
+	.machine  power7
+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 strings 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):
+	/* Now it has checked for first 32 bytes.  */
+	vspltisb	v0, 0
+	vspltisb	v2, -1
+	lvsr	v6, 0, r4   /* Compute mask.  */
+	or	r5, r4, r7
+	andi.	r5, r5, 0xF
+	beq	cr0, L(aligned)
+	andi.	r5, r7, 0xF
+	beq	cr0, L(s1_align)
+	lvsr	v10, 0, r7   /* Compute mask.  */
+
+	/* Both s1 and s2 are unaligned.  */
+	GET16BYTES(v4, r7, v10)
+	GET16BYTES(v5, r4, v6)
+	VCMPNEZB(v7, v5, v4)
+	beq	cr6, L(match)
+	b	L(different)
+
+	/* Align s1 to qw and adjust s2 address.  */
+	.align  4
+L(match):
+	clrldi	r6, r7, 60
+	subfic	r5, r6, 16
+	add	r7, r7, r5
+	add	r4, r4, r5
+	andi.	r5, r4, 0xF
+	beq	cr0, L(aligned)
+	lvsr	v6, 0, r4
+	/* There are 2 loops depending on the input alignment.
+	   Each loop gets 16 bytes from s1 and s2 and compares.
+	   Loop until a mismatch or null occurs.  */
+L(s1_align):
+	lvx	v4, r7, r0
+	GET16BYTES(v5, r4, v6)
+	VCMPNEZB(v7, v5, v4)
+	addi	r7, r7, 16
+	addi	r4, r4, 16
+	bne	cr6, L(different)
+
+	lvx	v4, r7, r0
+	GET16BYTES(v5, r4, v6)
+	VCMPNEZB(v7, v5, v4)
+	addi	r7, r7, 16
+	addi	r4, r4, 16
+	bne	cr6, L(different)
+
+	lvx	v4, r7, r0
+	GET16BYTES(v5, r4, v6)
+	VCMPNEZB(v7, v5, v4)
+	addi	r7, r7, 16
+	addi	r4, r4, 16
+	bne	cr6, L(different)
+
+	lvx	v4, r7, r0
+	GET16BYTES(v5, r4, v6)
+	VCMPNEZB(v7, v5, v4)
+	addi	r7, r7, 16
+	addi	r4, r4, 16
+	beq	cr6, L(s1_align)
+	b	L(different)
+
+	.align  4
+L(aligned):
+	lvx	v4, 0, r7
+	lvx	v5, 0, r4
+	VCMPNEZB(v7, v5, v4)
+	addi	r7, r7, 16
+	addi	r4, r4, 16
+	bne	cr6, L(different)
+
+	lvx	v4, 0, r7
+	lvx	v5, 0, r4
+	VCMPNEZB(v7, v5, v4)
+	addi	r7, r7, 16
+	addi	r4, r4, 16
+	bne	cr6, L(different)
+
+	lvx	v4, 0, r7
+	lvx	v5, 0, r4
+	VCMPNEZB(v7, v5, v4)
+	addi	r7, r7, 16
+	addi	r4, r4, 16
+	bne	cr6, L(different)
+
+	lvx	v4, 0, r7
+	lvx	v5, 0, r4
+	VCMPNEZB(v7, v5, v4)
+	addi	r7, r7, 16
+	addi	r4, r4, 16
+	beq	cr6, L(aligned)
+
+	/* Calculate and return the difference.  */
+L(different):
+	VCTZLSBB(r6, v7)
+	VEXTUBRX(r5, r6, v4)
+	VEXTUBRX(r4, r6, v5)
+	subf	r3, r4, r5
+	extsw	r3, r3
+	blr
+
+	.align  4
+L(different_nocmpb):
+	neg	r3, r9
+	and	r9, r9, r3
+	cntlzd	r9, r9
+	subfic	r9, r9, 63
+	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)
+
+	.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)
+#else
+#include <sysdeps/powerpc/powerpc64/power8/strcmp.S>
+#endif