powerpc64le: Optimize memset for POWER10

This implementation is based on __memset_power8 and integrates a lot
of suggestions from Anton Blanchard.

The biggest difference is that it makes extensive use of stxvl to
alignment and tail code to avoid branches and small stores.  It has
three main execution paths:

a) "Short lengths" for lengths up to 64 bytes, avoiding as many
   branches as possible.

b) "General case" for larger lengths, it has an alignment section
   using stxvl to avoid branches, a 128 bytes loop and then a tail
   code, again using stxvl with few branches.

c) "Zeroing cache blocks" for lengths from 256 bytes upwards and set
   value being zero.  It is mostly the __memset_power8 code but the
   alignment phase was simplified because, at this point, address is
   already 16-bytes aligned and also changed to use vector stores.
   The tail code was also simplified to reuse the general case tail.

All unaligned stores use stxvl instructions that do not generate
alignment interrupts on POWER10, making it safe to use on
caching-inhibited memory.

On average, this implementation provides something around 30%
improvement when compared to __memset_power8.

Reviewed-by: Matheus Castanho <msc@linux.ibm.com>
Reviewed-by: Tulio Magno Quites Machado Filho <tuliom@linux.ibm.com>

1 files changed, 256 insertions, 0 deletions

diff --git a/sysdeps/powerpc/powerpc64/le/power10/memset.S b/sysdeps/powerpc/powerpc64/le/power10/memset.S
new file mode 100644
index 0000000000..6b8e2cfdaf
--- /dev/null
+++ b/sysdeps/powerpc/powerpc64/le/power10/memset.S
@@ -0,0 +1,256 @@
+/* Optimized memset implementation for POWER10 LE.
+   Copyright (C) 2021 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
+   <https://www.gnu.org/licenses/>.  */
+
+#include <sysdep.h>
+
+/* void * [r3] memset (void *s [r3], int c [r4], size_t n [r5]));
+   Returns 's'.  */
+
+#ifndef MEMSET
+# define MEMSET memset
+#endif
+
+	.machine  power9
+ENTRY_TOCLESS (MEMSET, 5)
+	CALL_MCOUNT 3
+
+L(_memset):
+	/* Assume memset of zero length is uncommon, and just let it go
+	   through the small path below.  */
+	cmpldi	r5,64
+
+	/* Replicate byte to quad word.  */
+	mtvsrd	v0+32,r4
+	vspltb	v0,v0,7
+
+	li	r7,16
+	sldi	r8,r7,56
+
+	bgt	L(large)
+
+	/* For short lengths we want to avoid as many branches as possible.
+	   We use store VSX vector with length instructions to do this.
+	   It takes advantage of the fact that if the length passed to stxvl
+	   is zero nothing is done, effectively a no-op.  */
+	sldi	r5,r5,56
+
+	addi	r10,r3,16
+
+	sub.	r11,r5,r8
+	isellt	r11,0,r11	/* Saturate the subtraction to zero.  */
+
+	stxvl	v0+32,r3,r5
+	stxvl	v0+32,r10,r11
+
+	addi	r9,r3,32
+	addi	r10,r3,48
+
+	sub.	r11,r11,r8
+	isellt	r11,0,r11
+
+	sub.	r5,r11,r8
+	isellt	r5,0,r5
+
+	stxvl	v0+32,r9,r11
+	stxvl	v0+32,r10,r5
+
+	blr
+
+	.balign	16
+L(large):
+	mr	r6,r3	/* Don't modify r3 since we need to return it.  */
+
+	/* Get dest 16B aligned.  */
+	neg	r0,r3
+	clrldi.	r7,r0,(64-4)
+	beq	L(aligned)
+	rldic	r9,r0,56,4	/* (~X & 0xf)<<56 "clrlsldi r9,r0,64-4,56".  */
+
+	stxvl	v0+32,r6,r9	/* Store up to 15B until aligned address.  */
+
+	add	r6,r6,r7
+	sub	r5,r5,r7
+
+	/* Go to tail if there is less than 64B left after alignment.  */
+	cmpldi	r5,64
+	blt	L(tail_64)
+
+	.balign	16
+L(aligned):
+	/* Go to tail if there is less than 128B left after alignment.  */
+	srdi.	r0,r5,7
+	beq	L(tail_128)
+
+	/* If c == 0 && n >= 256 use dcbz to zero out full cache blocks.  */
+	cmpldi	cr5,r5,255
+	cmpldi	cr6,r4,0
+	crand	27,26,21
+	bt	27,L(dcbz)
+
+	mtctr	r0
+
+	.balign	32
+L(loop):
+	stxv	v0+32,0(r6)
+	stxv	v0+32,16(r6)
+	stxv	v0+32,32(r6)
+	stxv	v0+32,48(r6)
+	stxv	v0+32,64(r6)
+	stxv	v0+32,80(r6)
+	stxv	v0+32,96(r6)
+	stxv	v0+32,112(r6)
+	addi	r6,r6,128
+	bdnz	L(loop)
+
+	.balign	16
+L(tail):
+	/* 127B or less left, finish the tail or return.  */
+	andi.	r5,r5,127
+	beqlr
+
+	cmpldi	r5,64
+	blt	L(tail_64)
+
+	.balign	16
+L(tail_128):
+	/* Stores a minimum of 64B and up to 128B and return.  */
+	stxv	v0+32,0(r6)
+	stxv	v0+32,16(r6)
+	stxv	v0+32,32(r6)
+	stxv	v0+32,48(r6)
+	addi	r6,r6,64
+	andi.	r5,r5,63
+	beqlr
+
+	.balign	16
+L(tail_64):
+	/* Stores up to 64B and return.  */
+	sldi	r5,r5,56
+
+	addi	r10,r6,16
+
+	sub.	r11,r5,r8
+	isellt	r11,0,r11
+
+	stxvl	v0+32,r6,r5
+	stxvl	v0+32,r10,r11
+
+	sub.	r11,r11,r8
+	blelr
+
+	addi	r9,r6,32
+	addi	r10,r6,48
+
+	isellt	r11,0,r11
+
+	sub.	r5,r11,r8
+	isellt	r5,0,r5
+
+	stxvl	v0+32,r9,r11
+	stxvl	v0+32,r10,r5
+
+	blr
+
+	.balign	16
+L(dcbz):
+	/* Special case when value is 0 and we have a long length to deal
+	   with.  Use dcbz to zero out a full cacheline of 128 bytes at a time.
+	   Before using dcbz though, we need to get the destination 128-byte
+	   aligned.  */
+	neg	r0,r6
+	clrldi.	r0,r0,(64-7)
+	beq	L(dcbz_aligned)
+
+	sub	r5,r5,r0
+	mtocrf	0x2,r0	/* copying bits 57..59 to cr6. The ones for sizes 64,
+			   32 and 16 which need to be checked.  */
+
+	/* Write 16-128 bytes until DST is aligned to 128 bytes.  */
+64:	bf	25,32f
+	stxv	v0+32,0(r6)
+	stxv	v0+32,16(r6)
+	stxv	v0+32,32(r6)
+	stxv	v0+32,48(r6)
+	addi	r6,r6,64
+
+32:	bf	26,16f
+	stxv	v0+32,0(r6)
+	stxv	v0+32,16(r6)
+	addi	r6,r6,32
+
+16:	bf	27,L(dcbz_aligned)
+	stxv	v0+32,0(r6)
+	addi	r6,r6,16
+
+	.balign	16
+L(dcbz_aligned):
+	/* Setup dcbz unroll offsets and count numbers.  */
+	srdi.	r0,r5,9
+	li	r9,128
+	beq	L(bcdz_tail)
+	li	r10,256
+	li	r11,384
+	mtctr	r0
+
+	.balign	16
+L(dcbz_loop):
+	/* Sets 512 bytes to zero in each iteration, the loop unrolling shows
+	   a throughput boost for large sizes (2048 bytes or higher).  */
+	dcbz	0,r6
+	dcbz	r9,r6
+	dcbz	r10,r6
+	dcbz	r11,r6
+	addi	r6,r6,512
+	bdnz	L(dcbz_loop)
+
+	andi.	r5,r5,511
+	beqlr
+
+	.balign	16
+L(bcdz_tail):
+	/* We have 1-511 bytes remaining.  */
+	srdi.	r0,r5,7
+	beq	L(tail)
+
+	mtocrf	0x1,r0
+
+256:	bf	30,128f
+	dcbz	0,r6
+	dcbz	r9,r6
+	addi	r6,r6,256
+
+128:	bf	31,L(tail)
+	dcbz	0,r6
+	addi	r6,r6,128
+
+	b	L(tail)
+
+END_GEN_TB (MEMSET,TB_TOCLESS)
+libc_hidden_builtin_def (memset)
+
+/* Copied from bzero.S to prevent the linker from inserting a stub
+   between bzero and memset.  */
+ENTRY_TOCLESS (__bzero)
+	CALL_MCOUNT 2
+	mr	r5,r4
+	li	r4,0
+	b	L(_memset)
+END (__bzero)
+#ifndef __bzero
+weak_alias (__bzero, bzero)
+#endif