PowerPC: strncat optimization for PPC64

The optimization is achieved by following techniques:
1. Doubleword aligned memory access and compares using
   cmpb instruction.
2. Loop unrolling for byte load/store.
3. CPU pre-fetch to avoid cache miss.

1 files changed, 222 insertions, 0 deletions

diff --git a/sysdeps/powerpc/powerpc64/power7/strncat.S b/sysdeps/powerpc/powerpc64/power7/strncat.S
new file mode 100644
index 0000000..1a1a95e
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/power7/strncat.S
@@ -0,0 +1,222 @@
+/* Optimized strncat implementation for PowerPC64/POWER7.
+
+   Copyright (C) 2014 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/>.  */
+
+/* The algorithm is as follows for aligned memory access :
+
+   if address of s2 is divisible by 0x7UL,
+       perform aligned doubleword catenation
+   else
+       perform unaligned catenation
+
+   The aligned comparison are made using cmpb instructions.  */
+
+/* char* [r3] strncat (const char *s1 [r3],
+                       const char *s2 [r4],
+                       size_t size [r5])  */
+
+#include <sysdep.h>
+
+#ifndef STRNCAT
+# undef strncat
+# define STRNCAT  strncat
+#endif
+
+#ifndef STRLEN
+# define STRLEN   __strlen_ppc
+#endif
+
+#define	FRAMESIZE	(FRAME_MIN_SIZE+32)
+
+	.machine  power7
+EALIGN(STRNCAT, 4, 0)
+	CALL_MCOUNT 3
+
+	mflr r0				/* Load link register LR to r0.  */
+
+/* We shall use r29, r30 and r31 non volatile register for retention.
+   Save all the callee registers in the GPR save area.  */
+	std r29, -24(r1)		/* Save callers register r29.  */
+	std r30, -16(r1)		/* Save callers register r30.  */
+	std r31, -8(r1)			/* Save callers register r31.  */
+
+	std r0, 16(r1)			/* Store the link register.  */
+	stdu r1, -FRAMESIZE(r1)		/* Create the stack frame.  */
+
+/* Improve performance with CPU pre-fetch.  */
+	dcbt 0, r3			/* Pre-fetch str to avoid cache
+					   miss.  */
+	dcbt 0, r4			/* Pre-fetch accept to avoid cache
+					   miss.  */
+
+	mr. r29, r5			/* Save "n" in r29.  */
+	mr r30, r3			/* Save "s1" in r30 from r3.  */
+	beq cr0,L(done)
+
+	mr r31, r4			/* Save "s2" in r31 from r4.  */
+	bl STRLEN			/* Call optimized strlen on s1; goto
+					   end of s1.  */
+	nop
+	cmpldi cr7, r29, 7		/* If s2 is <=7 process
+					    byte-by-byte.  */
+	add r3, r30, r3			/* Grab the last character of s1.  */
+	bgt cr7,L(alignment)		/* Process by aligned strings.  */
+
+	cmpldi cr7, r29, 3		/* If n is >= 4, we can
+					   byte-unroll.  */
+	addi r9, r3, -1			/* Make "s1" point before next
+					   character, increment when read.  */
+	bgt cr7, L(bytes_unroll)	/* Process each byte.  */
+
+L(byte_by_byte):
+	lbz r10, 0(r31)
+	addi r8, r9, 1
+	cmpdi cr7, r10, 0		/* Check for NULL in "s2".  */
+	stb r10, 1(r9)
+	beq cr7, L(done)
+	add r9, r9, r29
+	subf r9, r8, r9
+	addi r9, r9, 1
+	mtctr r9
+	b L(branch2)
+	.p2align 4
+L(branch1):
+	lbzu r10, 1(r31)
+	cmpdi cr7, r10, 0
+	stbu r10, 1(r8)
+	beq cr7,L(done)
+L(branch2):
+	mr r9, r8
+	bdnz L(branch1)
+	beq cr7,L(done)
+L(nullTerminate):
+	li r10, 0			/* Load NULL for termination.  */
+	stb r10, 1(r9)			/* Append or terminate s1 with
+					   NULL.  */
+	.p2align 4			/* A small section here.  */
+L(done):				/* We return now.   */
+	addi r1, r1, FRAMESIZE		/* Restore stack pointer.  */
+	mr r3, r30			/* Set the return value length of
+					   string.  */
+	ld r0, 16(r1)			/* Read the saved link register.  */
+	ld r29, -24(r1)			/* Restore save register r29.  */
+	ld r30, -16(r1)			/* Restore save register r30.  */
+	ld r31, -8(r1)			/* Restore save register r31.  */
+	mtlr r0				/* Restore link register.  */
+	blr				/* Branch to link register.  */
+
+	.p2align 4
+L(alignment):
+	rldicl. r9, r31, 0, 61		/* Check if s2 is 8byte aligned  */
+	beq cr0,L(dwordAligned)
+
+	.p2align 4
+/* Unaligned bytes in string, so process byte by byte.
+   POWER7 has performance gains over loop unroll.  */
+L(bytes_unroll):
+	addi r9, r3, -1
+	srdi r10, r29, 2
+	mtctr r10
+	b L(L10)
+	.p2align 4
+L(L44):
+	lbz r10, 1(r31)			/* Load byte.  */
+	cmpdi cr7, r10, 0		/* Compare ; if byte not zero,
+					   continue.  */
+	stb r10, 2(r9)			/* Store byte  */
+	beq cr7, L(done)
+	addi r31, r31, 4
+
+	lbz r10, -2(r31)		/* Perform loop unroll here on byte
+					   load and store.  */
+	cmpdi cr7, r10, 0
+	stb r10, 3(r9)
+	beq cr7, L(done)
+
+	lbz r10, -1(r31)		/* Loop unroll here.  */
+	cmpdi cr7, r10, 0
+	stbu r10, 4(r9)
+	beq cr7, L(done)
+
+	bdz L(leftNbytes)
+
+L(L10):
+	lbz r10, 0(r31)			/* Loop unroll here.  */
+	cmpdi cr7, r10, 0
+	stb r10, 1(r9)
+	bne cr7,L(L44)
+	b L(done)
+	.p2align 4
+/* If s2 is double word aligned, we load and store double word.  */
+L(dwordAligned):
+/* read, write 8 bytes at a time  */
+	srdi r8, r29, 3			/* Compute count for CTR to loop;
+					   count = n/8.  */
+	li r7, 0			/* Load r7 with NULL.  */
+	li r10, 0			/* Load r10 with MASK '0'.  */
+
+	mtctr r8			/* Move count to CTR.  */
+L(loop8):
+	ld r9, 0(r31)			/* Read double word from s2.  */
+	cmpb r6, r9, r10		/* Compare bytes in s2 we read
+					   just now.  */
+	cmpdi r6, 0			/* If cmpb returned NULL,
+					   we continue.  */
+	bne+ L(a8)
+	std r9, 0(r3)			/* Append double word from s2
+					   with s1.  */
+	addi r3, r3, 8			/* Increment s1.  */
+	addi r31, r31, 8		/* Increment s2.  */
+	subi r29, r29, 8		/* Decrement count by 8.  */
+	bdnz L(loop8)			/* Continue until "count" is
+					   non zero.  */
+
+L(a8):
+	cmpdi r29, 0			/* If "n" is already zero, we skip. */
+	beq+ L(align8align)
+
+	mtctr r29			/* Process left over bytes in "n".  */
+L(unaligned0):
+	lbz r9, 0(r31)			/* Read a byte from s2.  */
+	cmpw r9, r7			/* If byte is NULL, we stop here . */
+	beq+ L(align8align)		/* Skip processing further if NULL.  */
+	stb  r9, 0(r3)			/* If not NULL, store byte into s1.  */
+	addi r3, r3, 1			/* Increment s1 by 1.  */
+	addi r31, r31, 1		/* Increment s2 by 1.  */
+	bdnz L(unaligned0)		/* Decrement counter "n" and loop
+					   until non zero.  */
+L(align8align):
+	stb r7, 0(r3)			/* Terminate s1 with NULL.  */
+
+	addi r1, r1, FRAMESIZE		/* Restore stack pointer.  */
+	mr r3, r30			/* Set the return value, length of
+					   string.  */
+	ld r0, 16(r1)			/* Read the saved link register.  */
+	ld r29, -24(r1)			/* Restore save register r29.  */
+	ld r30, -16(r1)			/* Restore save register r30.  */
+	ld r31, -8(r1)			/* Restore save register r31.  */
+	mtlr r0				/* Restore link register.  */
+	blr				/* Branch to link register  */
+
+	.p2align 4
+L(leftNbytes):
+	rldicl. r29, r29, 0, 62		/* Check if n>0 and n < 4 bytes.  */
+	bne cr0,L(byte_by_byte)		/* Process bytes one by one. */
+	b L(nullTerminate)		/* Now, finish catenation with
+					   NULL termination.  */
+END(STRNCAT)