[thirdparty/glibc.git] / sysdeps / powerpc / powerpc64 / power6 / memset.S

/* Optimized 64-bit memset implementation for POWER6.
   Copyright (C) 1997-2019 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>

/* void * [r3] memset (void *s [r3], int c [r4], size_t n [r5]));
   Returns 's'.

   The memset is done in three sizes: byte (8 bits), word (32 bits),
   cache line (256 bits). There is a special case for setting cache lines
   to 0, to take advantage of the dcbz instruction.  */

#ifndef MEMSET
# define MEMSET memset
#endif
	.machine power6
ENTRY_TOCLESS (MEMSET, 7)
	CALL_MCOUNT 3

#define rTMP	r0
#define rRTN	r3	/* Initial value of 1st argument.  */
#define rMEMP0	r3	/* Original value of 1st arg.  */
#define rCHR	r4	/* Char to set in each byte.  */
#define rLEN	r5	/* Length of region to set.  */
#define rMEMP	r6	/* Address at which we are storing.  */
#define rALIGN	r7	/* Number of bytes we are setting now (when aligning). */
#define rMEMP2	r8
#define rMEMP3	r9	/* Alt mem pointer.  */
L(_memset):
/* Take care of case for size <= 4.  */
	cmpldi	cr1, rLEN, 8
	andi.	rALIGN, rMEMP0, 7
	mr	rMEMP, rMEMP0
	ble	cr1, L(small)

/* Align to doubleword boundary.  */
	cmpldi	cr5, rLEN, 31
	insrdi	rCHR, rCHR, 8, 48	/* Replicate byte to halfword.  */
	beq+	L(aligned2)
	mtcrf	0x01, rMEMP0
	subfic	rALIGN, rALIGN, 8
	cror	28,30,31		/* Detect odd word aligned.  */
	add	rMEMP, rMEMP, rALIGN
	sub	rLEN, rLEN, rALIGN
	insrdi	rCHR, rCHR, 16, 32	/* Replicate halfword to word.  */
	bt	29, L(g4)
/* Process the even word of doubleword.  */
	bf+	31, L(g2)
	stb	rCHR, 0(rMEMP0)
	bt	30, L(g4x)
L(g2):
	sth	rCHR, -6(rMEMP)
L(g4x):
	stw	rCHR, -4(rMEMP)
	b	L(aligned)
/* Process the odd word of doubleword.  */
L(g4):
	bf	28, L(g4x) /* If false, word aligned on odd word.  */
	bf+	31, L(g0)
	stb	rCHR, 0(rMEMP0)
	bt	30, L(aligned)
L(g0):
	sth	rCHR, -2(rMEMP)

/* Handle the case of size < 31.  */
L(aligned2):
	insrdi	rCHR, rCHR, 16, 32	/* Replicate halfword to word.  */
L(aligned):
	mtcrf	0x01, rLEN
	ble	cr5, L(medium)
/* Align to 32-byte boundary.  */
	andi.	rALIGN, rMEMP, 0x18
	subfic	rALIGN, rALIGN, 0x20
	insrdi	rCHR, rCHR, 32, 0	/* Replicate word to double word. */
	beq	L(caligned)
	mtcrf	0x01, rALIGN
	add	rMEMP, rMEMP, rALIGN
	sub	rLEN, rLEN, rALIGN
	cmplwi	cr1, rALIGN, 0x10
	mr	rMEMP2, rMEMP
	bf	28, L(a1)
	stdu	rCHR, -8(rMEMP2)
L(a1):	blt	cr1, L(a2)
	std	rCHR, -8(rMEMP2)
	stdu	rCHR, -16(rMEMP2)
L(a2):

/* Now aligned to a 32 byte boundary.  */
        .align 4
L(caligned):
	cmpldi	cr1, rCHR, 0
	clrrdi.	rALIGN, rLEN, 5
	mtcrf	0x01, rLEN
	beq	cr1, L(zloopstart) /* Special case for clearing memory using dcbz.  */
	beq	L(medium)	/* We may not actually get to do a full line.  */
	.align 4
/* Storing a non-zero "c" value. We are aligned at a sector (32-byte)
   boundary may not be at cache line (128-byte) boundary.  */
L(nzloopstart):
/* memset in 32-byte chunks until we get to a cache line boundary.
   If rLEN is less than the distance to the next cache-line boundary use
   cacheAligned1 code to finish the tail.  */
	cmpldi	cr1,rLEN,128

	andi.	rTMP,rMEMP,127
	blt	cr1,L(cacheAligned1)
	addi	rMEMP3,rMEMP,32
	beq	L(nzCacheAligned)
	addi	rLEN,rLEN,-32
	std	rCHR,0(rMEMP)
	std	rCHR,8(rMEMP)
	std	rCHR,16(rMEMP)
	addi	rMEMP,rMEMP,32
	andi.	rTMP,rMEMP3,127
	std	rCHR,-8(rMEMP3)

	beq	L(nzCacheAligned)
	addi	rLEN,rLEN,-32
	std	rCHR,0(rMEMP3)
	addi	rMEMP,rMEMP,32
	std	rCHR,8(rMEMP3)
	andi.	rTMP,rMEMP,127
	std	rCHR,16(rMEMP3)
	std	rCHR,24(rMEMP3)

	beq	L(nzCacheAligned)
	addi	rLEN,rLEN,-32
	std	rCHR,32(rMEMP3)
	addi	rMEMP,rMEMP,32
	cmpldi	cr1,rLEN,128
	std	rCHR,40(rMEMP3)
	cmpldi	cr6,rLEN,256
	li	rMEMP2,128
	std	rCHR,48(rMEMP3)
	std	rCHR,56(rMEMP3)
	blt	cr1,L(cacheAligned1)
	b	L(nzCacheAligned128)

/* Now we are aligned to the cache line and can use dcbtst.  */
        .align 4
L(nzCacheAligned):
	cmpldi	cr1,rLEN,128
	blt	cr1,L(cacheAligned1)
	b	L(nzCacheAligned128)
        .align 5
L(nzCacheAligned128):
	cmpldi	cr1,rLEN,256
	addi	rMEMP3,rMEMP,64
	std	rCHR,0(rMEMP)
	std	rCHR,8(rMEMP)
	std	rCHR,16(rMEMP)
	std	rCHR,24(rMEMP)
	std	rCHR,32(rMEMP)
	std	rCHR,40(rMEMP)
	std	rCHR,48(rMEMP)
	std	rCHR,56(rMEMP)
	addi	rMEMP,rMEMP3,64
	addi	rLEN,rLEN,-128
	std	rCHR,0(rMEMP3)
	std	rCHR,8(rMEMP3)
	std	rCHR,16(rMEMP3)
	std	rCHR,24(rMEMP3)
	std	rCHR,32(rMEMP3)
	std	rCHR,40(rMEMP3)
	std	rCHR,48(rMEMP3)
	std	rCHR,56(rMEMP3)
	bge	cr1,L(nzCacheAligned128)
	dcbtst	0,rMEMP
	b	L(cacheAligned1)
	.align 5
/* Storing a zero "c" value. We are aligned at a sector (32-byte)
   boundary but may not be at cache line (128-byte) boundary.  If the
   remaining length spans a full cache line we can use the Data cache
   block zero instruction. */
L(zloopstart):
/* memset in 32-byte chunks until we get to a cache line boundary.
   If rLEN is less than the distance to the next cache-line boundary use
   cacheAligned1 code to finish the tail.  */
	cmpldi	cr1,rLEN,128
	beq	L(medium)
L(getCacheAligned):
	andi.	rTMP,rMEMP,127
	nop
	blt	cr1,L(cacheAligned1)
	addi	rMEMP3,rMEMP,32
	beq	L(cacheAligned)
	addi	rLEN,rLEN,-32
	std	rCHR,0(rMEMP)
	std	rCHR,8(rMEMP)
	std	rCHR,16(rMEMP)
	addi	rMEMP,rMEMP,32
	andi.	rTMP,rMEMP3,127
	std	rCHR,-8(rMEMP3)
L(getCacheAligned2):
	beq	L(cacheAligned)
	addi	rLEN,rLEN,-32
	std	rCHR,0(rMEMP3)
	std	rCHR,8(rMEMP3)
	addi	rMEMP,rMEMP,32
	andi.	rTMP,rMEMP,127
	std	rCHR,16(rMEMP3)
	std	rCHR,24(rMEMP3)
L(getCacheAligned3):
	beq	L(cacheAligned)
	addi	rLEN,rLEN,-32
	std	rCHR,32(rMEMP3)
	addi	rMEMP,rMEMP,32
	cmpldi	cr1,rLEN,128
	std	rCHR,40(rMEMP3)
	cmpldi	cr6,rLEN,256
	li	rMEMP2,128
	std	rCHR,48(rMEMP3)
	std	rCHR,56(rMEMP3)
	blt	cr1,L(cacheAligned1)
	blt	cr6,L(cacheAligned128)
	b	L(cacheAlignedx)

/* Now we are aligned to the cache line and can use dcbz.  */
        .align 5
L(cacheAligned):
	cmpldi	cr1,rLEN,128
	cmpldi	cr6,rLEN,256
	blt	cr1,L(cacheAligned1)
	li	rMEMP2,128
L(cacheAlignedx):
	cmpldi	cr5,rLEN,640
	blt	cr6,L(cacheAligned128)
	bgt	cr5,L(cacheAligned512)
	cmpldi	cr6,rLEN,512
	dcbz	0,rMEMP
	cmpldi	cr1,rLEN,384
	dcbz	rMEMP2,rMEMP
	addi	rMEMP,rMEMP,256
	addi	rLEN,rLEN,-256
	blt	cr1,L(cacheAligned1)
	blt	cr6,L(cacheAligned128)
	b	L(cacheAligned256)
	.align 5
/* A simple loop for the longer (>640 bytes) lengths.  This form limits
   the branch miss-predicted to exactly 1 at loop exit.*/
L(cacheAligned512):
	cmpldi	cr1,rLEN,128
	blt	cr1,L(cacheAligned1)
	dcbz	0,rMEMP
	addi	rLEN,rLEN,-128
	addi	rMEMP,rMEMP,128
	b	L(cacheAligned512)
        .align 5
L(cacheAligned256):

	cmpldi	cr6,rLEN,512

	dcbz	0,rMEMP
	cmpldi	cr1,rLEN,384
	dcbz	rMEMP2,rMEMP
	addi	rMEMP,rMEMP,256
	addi	rLEN,rLEN,-256

	bge	cr6,L(cacheAligned256)

	blt	cr1,L(cacheAligned1)
        .align 4
L(cacheAligned128):
	dcbz	0,rMEMP
	addi	rMEMP,rMEMP,128
	addi	rLEN,rLEN,-128
        nop
L(cacheAligned1):
	cmpldi	cr1,rLEN,32
	blt	cr1,L(handletail32)
	addi	rMEMP3,rMEMP,32
	addi	rLEN,rLEN,-32
	std	rCHR,0(rMEMP)
	std	rCHR,8(rMEMP)
	std	rCHR,16(rMEMP)
	addi	rMEMP,rMEMP,32
	cmpldi	cr1,rLEN,32
	std	rCHR,-8(rMEMP3)
L(cacheAligned2):
	blt	cr1,L(handletail32)
	addi	rLEN,rLEN,-32
	std	rCHR,0(rMEMP3)
	std	rCHR,8(rMEMP3)
	addi	rMEMP,rMEMP,32
	cmpldi	cr1,rLEN,32
	std	rCHR,16(rMEMP3)
	std	rCHR,24(rMEMP3)
	nop
L(cacheAligned3):
	blt	cr1,L(handletail32)
	addi	rMEMP,rMEMP,32
	addi	rLEN,rLEN,-32
	std	rCHR,32(rMEMP3)
	std	rCHR,40(rMEMP3)
	std	rCHR,48(rMEMP3)
	std	rCHR,56(rMEMP3)

/* We are here because the length or remainder (rLEN) is less than the
   cache line/sector size and does not justify aggressive loop unrolling.
   So set up the preconditions for L(medium) and go there.  */
        .align 3
L(handletail32):
	cmpldi	cr1,rLEN,0
	beqlr   cr1
	b	L(medium)

	.align 5
L(small):
/* Memset of 8 bytes or less.  */
	cmpldi	cr6, rLEN, 4
	cmpldi	cr5, rLEN, 1
	ble	cr6,L(le4)
	subi	rLEN, rLEN, 4
	stb	rCHR,0(rMEMP)
	stb	rCHR,1(rMEMP)
	stb	rCHR,2(rMEMP)
	stb	rCHR,3(rMEMP)
	addi	rMEMP,rMEMP, 4
	cmpldi	cr5, rLEN, 1
L(le4):
	cmpldi	cr1, rLEN, 3
	bltlr	cr5
	stb	rCHR, 0(rMEMP)
	beqlr	cr5
	stb	rCHR, 1(rMEMP)
	bltlr	cr1
	stb	rCHR, 2(rMEMP)
	beqlr	cr1
	stb	rCHR, 3(rMEMP)
	blr

/* Memset of 0-31 bytes.  */
	.align 5
L(medium):
	insrdi	rCHR, rCHR, 32, 0	/* Replicate word to double word.  */
	cmpldi	cr1, rLEN, 16
L(medium_tail2):
	add	rMEMP, rMEMP, rLEN
L(medium_tail):
	bt-	31, L(medium_31t)
	bt-	30, L(medium_30t)
L(medium_30f):
	bt	29, L(medium_29t)
L(medium_29f):
	bge	cr1, L(medium_27t)
	bflr	28
	std	rCHR, -8(rMEMP)
	blr

L(medium_31t):
	stbu	rCHR, -1(rMEMP)
	bf-	30, L(medium_30f)
L(medium_30t):
	sthu	rCHR, -2(rMEMP)
	bf-	29, L(medium_29f)
L(medium_29t):
	stwu	rCHR, -4(rMEMP)
	blt	cr1, L(medium_27f)
L(medium_27t):
	std	rCHR, -8(rMEMP)
	stdu	rCHR, -16(rMEMP)
L(medium_27f):
	bflr	28
L(medium_28t):
	std	rCHR, -8(rMEMP)
	blr
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 3
	mr	r5,r4
	li	r4,0
	b	L(_memset)
END (__bzero)
#ifndef __bzero
weak_alias (__bzero, bzero)
#endif
Commit	Line	Data
a88f47a7	1	/* Optimized 64-bit memset implementation for POWER6.
04277e02	2	Copyright (C) 1997-2019 Free Software Foundation, Inc.
04067002 UD	3	This file is part of the GNU C Library.
	4
	5	The GNU C Library is free software; you can redistribute it and/or
	6	modify it under the terms of the GNU Lesser General Public
	7	License as published by the Free Software Foundation; either
	8	version 2.1 of the License, or (at your option) any later version.
	9
	10	The GNU C Library is distributed in the hope that it will be useful,
	11	but WITHOUT ANY WARRANTY; without even the implied warranty of
	12	MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
	13	Lesser General Public License for more details.
	14
	15	You should have received a copy of the GNU Lesser General Public
59ba27a6 PE	16	License along with the GNU C Library; if not, see
59ba27a6 PE	17	<http://www.gnu.org/licenses/>. */
04067002 UD	18
04067002 UD	19	#include <sysdep.h>
04067002	20
f17a4233	21	/* void * [r3] memset (void *s [r3], int c [r4], size_t n [r5]));
04067002 UD	22	Returns 's'.
	23
	24	The memset is done in three sizes: byte (8 bits), word (32 bits),
	25	cache line (256 bits). There is a special case for setting cache lines
	26	to 0, to take advantage of the dcbz instruction. */
	27
18e0054b WSM	28	#ifndef MEMSET
	29	# define MEMSET memset
	30	#endif
a88f47a7	31	.machine power6
d5b41185	32	ENTRY_TOCLESS (MEMSET, 7)
04067002 UD	33	CALL_MCOUNT 3
	34
	35	#define rTMP r0
	36	#define rRTN r3 /* Initial value of 1st argument. */
2d67d91a JM	37	#define rMEMP0 r3 /* Original value of 1st arg. */
	38	#define rCHR r4 /* Char to set in each byte. */
	39	#define rLEN r5 /* Length of region to set. */
	40	#define rMEMP r6 /* Address at which we are storing. */
04067002 UD	41	#define rALIGN r7 /* Number of bytes we are setting now (when aligning). */
04067002 UD	42	#define rMEMP2 r8
a88f47a7	43	#define rMEMP3 r9 /* Alt mem pointer. */
04067002	44	L(_memset):
04067002 UD	45	/* Take care of case for size <= 4. */
	46	cmpldi cr1, rLEN, 8
	47	andi. rALIGN, rMEMP0, 7
	48	mr rMEMP, rMEMP0
a88f47a7	49	ble cr1, L(small)
04067002 UD	50
	51	/* Align to doubleword boundary. */
	52	cmpldi cr5, rLEN, 31
3be87c77	53	insrdi rCHR, rCHR, 8, 48 /* Replicate byte to halfword. */
04067002 UD	54	beq+ L(aligned2)
	55	mtcrf 0x01, rMEMP0
	56	subfic rALIGN, rALIGN, 8
	57	cror 28,30,31 /* Detect odd word aligned. */
	58	add rMEMP, rMEMP, rALIGN
	59	sub rLEN, rLEN, rALIGN
3be87c77	60	insrdi rCHR, rCHR, 16, 32 /* Replicate halfword to word. */
04067002 UD	61	bt 29, L(g4)
	62	/* Process the even word of doubleword. */
	63	bf+ 31, L(g2)
	64	stb rCHR, 0(rMEMP0)
	65	bt 30, L(g4x)
	66	L(g2):
	67	sth rCHR, -6(rMEMP)
	68	L(g4x):
	69	stw rCHR, -4(rMEMP)
	70	b L(aligned)
	71	/* Process the odd word of doubleword. */
	72	L(g4):
	73	bf 28, L(g4x) /* If false, word aligned on odd word. */
	74	bf+ 31, L(g0)
	75	stb rCHR, 0(rMEMP0)
	76	bt 30, L(aligned)
	77	L(g0):
	78	sth rCHR, -2(rMEMP)
	79
	80	/* Handle the case of size < 31. */
	81	L(aligned2):
3be87c77	82	insrdi rCHR, rCHR, 16, 32 /* Replicate halfword to word. */
04067002 UD	83	L(aligned):
	84	mtcrf 0x01, rLEN
	85	ble cr5, L(medium)
	86	/* Align to 32-byte boundary. */
	87	andi. rALIGN, rMEMP, 0x18
	88	subfic rALIGN, rALIGN, 0x20
3be87c77	89	insrdi rCHR, rCHR, 32, 0 /* Replicate word to double word. */
04067002 UD	90	beq L(caligned)
	91	mtcrf 0x01, rALIGN
	92	add rMEMP, rMEMP, rALIGN
	93	sub rLEN, rLEN, rALIGN
	94	cmplwi cr1, rALIGN, 0x10
	95	mr rMEMP2, rMEMP
	96	bf 28, L(a1)
	97	stdu rCHR, -8(rMEMP2)
	98	L(a1): blt cr1, L(a2)
	99	std rCHR, -8(rMEMP2)
	100	stdu rCHR, -16(rMEMP2)
	101	L(a2):
	102
	103	/* Now aligned to a 32 byte boundary. */
	104	.align 4
	105	L(caligned):
	106	cmpldi cr1, rCHR, 0
	107	clrrdi. rALIGN, rLEN, 5
	108	mtcrf 0x01, rLEN
	109	beq cr1, L(zloopstart) /* Special case for clearing memory using dcbz. */
04067002	110	beq L(medium) /* We may not actually get to do a full line. */
a88f47a7 UD	111	.align 4
	112	/* Storing a non-zero "c" value. We are aligned at a sector (32-byte)
	113	boundary may not be at cache line (128-byte) boundary. */
	114	L(nzloopstart):
	115	/* memset in 32-byte chunks until we get to a cache line boundary.
f24a6d08	116	If rLEN is less than the distance to the next cache-line boundary use
a88f47a7 UD	117	cacheAligned1 code to finish the tail. */
a88f47a7 UD	118	cmpldi cr1,rLEN,128
04067002	119
a88f47a7 UD	120	andi. rTMP,rMEMP,127
	121	blt cr1,L(cacheAligned1)
	122	addi rMEMP3,rMEMP,32
	123	beq L(nzCacheAligned)
	124	addi rLEN,rLEN,-32
	125	std rCHR,0(rMEMP)
	126	std rCHR,8(rMEMP)
	127	std rCHR,16(rMEMP)
	128	addi rMEMP,rMEMP,32
	129	andi. rTMP,rMEMP3,127
	130	std rCHR,-8(rMEMP3)
04067002	131
a88f47a7 UD	132	beq L(nzCacheAligned)
	133	addi rLEN,rLEN,-32
	134	std rCHR,0(rMEMP3)
	135	addi rMEMP,rMEMP,32
	136	std rCHR,8(rMEMP3)
	137	andi. rTMP,rMEMP,127
	138	std rCHR,16(rMEMP3)
	139	std rCHR,24(rMEMP3)
	140
	141	beq L(nzCacheAligned)
	142	addi rLEN,rLEN,-32
	143	std rCHR,32(rMEMP3)
	144	addi rMEMP,rMEMP,32
	145	cmpldi cr1,rLEN,128
	146	std rCHR,40(rMEMP3)
	147	cmpldi cr6,rLEN,256
	148	li rMEMP2,128
	149	std rCHR,48(rMEMP3)
	150	std rCHR,56(rMEMP3)
	151	blt cr1,L(cacheAligned1)
	152	b L(nzCacheAligned128)
	153
	154	/* Now we are aligned to the cache line and can use dcbtst. */
	155	.align 4
	156	L(nzCacheAligned):
	157	cmpldi cr1,rLEN,128
	158	blt cr1,L(cacheAligned1)
	159	b L(nzCacheAligned128)
	160	.align 5
	161	L(nzCacheAligned128):
	162	cmpldi cr1,rLEN,256
	163	addi rMEMP3,rMEMP,64
	164	std rCHR,0(rMEMP)
	165	std rCHR,8(rMEMP)
	166	std rCHR,16(rMEMP)
	167	std rCHR,24(rMEMP)
	168	std rCHR,32(rMEMP)
	169	std rCHR,40(rMEMP)
	170	std rCHR,48(rMEMP)
	171	std rCHR,56(rMEMP)
	172	addi rMEMP,rMEMP3,64
	173	addi rLEN,rLEN,-128
	174	std rCHR,0(rMEMP3)
	175	std rCHR,8(rMEMP3)
	176	std rCHR,16(rMEMP3)
	177	std rCHR,24(rMEMP3)
	178	std rCHR,32(rMEMP3)
	179	std rCHR,40(rMEMP3)
	180	std rCHR,48(rMEMP3)
	181	std rCHR,56(rMEMP3)
	182	bge cr1,L(nzCacheAligned128)
	183	dcbtst 0,rMEMP
	184	b L(cacheAligned1)
04067002	185	.align 5
a88f47a7 UD	186	/* Storing a zero "c" value. We are aligned at a sector (32-byte)
	187	boundary but may not be at cache line (128-byte) boundary. If the
	188	remaining length spans a full cache line we can use the Data cache
	189	block zero instruction. */
04067002	190	L(zloopstart):
a88f47a7	191	/* memset in 32-byte chunks until we get to a cache line boundary.
f24a6d08	192	If rLEN is less than the distance to the next cache-line boundary use
a88f47a7 UD	193	cacheAligned1 code to finish the tail. */
a88f47a7 UD	194	cmpldi cr1,rLEN,128
04067002	195	beq L(medium)
04067002	196	L(getCacheAligned):
04067002	197	andi. rTMP,rMEMP,127
a88f47a7 UD	198	nop
	199	blt cr1,L(cacheAligned1)
	200	addi rMEMP3,rMEMP,32
	201	beq L(cacheAligned)
	202	addi rLEN,rLEN,-32
	203	std rCHR,0(rMEMP)
	204	std rCHR,8(rMEMP)
	205	std rCHR,16(rMEMP)
04067002	206	addi rMEMP,rMEMP,32
a88f47a7 UD	207	andi. rTMP,rMEMP3,127
	208	std rCHR,-8(rMEMP3)
	209	L(getCacheAligned2):
	210	beq L(cacheAligned)
04067002	211	addi rLEN,rLEN,-32
a88f47a7 UD	212	std rCHR,0(rMEMP3)
	213	std rCHR,8(rMEMP3)
	214	addi rMEMP,rMEMP,32
	215	andi. rTMP,rMEMP,127
	216	std rCHR,16(rMEMP3)
	217	std rCHR,24(rMEMP3)
	218	L(getCacheAligned3):
	219	beq L(cacheAligned)
	220	addi rLEN,rLEN,-32
	221	std rCHR,32(rMEMP3)
	222	addi rMEMP,rMEMP,32
	223	cmpldi cr1,rLEN,128
	224	std rCHR,40(rMEMP3)
	225	cmpldi cr6,rLEN,256
	226	li rMEMP2,128
	227	std rCHR,48(rMEMP3)
	228	std rCHR,56(rMEMP3)
	229	blt cr1,L(cacheAligned1)
	230	blt cr6,L(cacheAligned128)
	231	b L(cacheAlignedx)
04067002 UD	232
04067002 UD	233	/* Now we are aligned to the cache line and can use dcbz. */
a88f47a7	234	.align 5
04067002	235	L(cacheAligned):
a88f47a7 UD	236	cmpldi cr1,rLEN,128
	237	cmpldi cr6,rLEN,256
	238	blt cr1,L(cacheAligned1)
	239	li rMEMP2,128
	240	L(cacheAlignedx):
	241	cmpldi cr5,rLEN,640
	242	blt cr6,L(cacheAligned128)
	243	bgt cr5,L(cacheAligned512)
	244	cmpldi cr6,rLEN,512
	245	dcbz 0,rMEMP
	246	cmpldi cr1,rLEN,384
	247	dcbz rMEMP2,rMEMP
	248	addi rMEMP,rMEMP,256
	249	addi rLEN,rLEN,-256
	250	blt cr1,L(cacheAligned1)
	251	blt cr6,L(cacheAligned128)
	252	b L(cacheAligned256)
	253	.align 5
	254	/* A simple loop for the longer (>640 bytes) lengths. This form limits
	255	the branch miss-predicted to exactly 1 at loop exit.*/
	256	L(cacheAligned512):
78b7adba	257	cmpldi cr1,rLEN,128
a88f47a7 UD	258	blt cr1,L(cacheAligned1)
	259	dcbz 0,rMEMP
	260	addi rLEN,rLEN,-128
	261	addi rMEMP,rMEMP,128
	262	b L(cacheAligned512)
	263	.align 5
	264	L(cacheAligned256):
	265
	266	cmpldi cr6,rLEN,512
	267
04067002	268	dcbz 0,rMEMP
a88f47a7 UD	269	cmpldi cr1,rLEN,384
	270	dcbz rMEMP2,rMEMP
	271	addi rMEMP,rMEMP,256
	272	addi rLEN,rLEN,-256
	273
	274	bge cr6,L(cacheAligned256)
	275
	276	blt cr1,L(cacheAligned1)
	277	.align 4
	278	L(cacheAligned128):
	279	dcbz 0,rMEMP
	280	addi rMEMP,rMEMP,128
	281	addi rLEN,rLEN,-128
	282	nop
	283	L(cacheAligned1):
	284	cmpldi cr1,rLEN,32
	285	blt cr1,L(handletail32)
	286	addi rMEMP3,rMEMP,32
	287	addi rLEN,rLEN,-32
	288	std rCHR,0(rMEMP)
	289	std rCHR,8(rMEMP)
	290	std rCHR,16(rMEMP)
	291	addi rMEMP,rMEMP,32
	292	cmpldi cr1,rLEN,32
	293	std rCHR,-8(rMEMP3)
	294	L(cacheAligned2):
	295	blt cr1,L(handletail32)
	296	addi rLEN,rLEN,-32
	297	std rCHR,0(rMEMP3)
	298	std rCHR,8(rMEMP3)
	299	addi rMEMP,rMEMP,32
	300	cmpldi cr1,rLEN,32
	301	std rCHR,16(rMEMP3)
	302	std rCHR,24(rMEMP3)
	303	nop
	304	L(cacheAligned3):
	305	blt cr1,L(handletail32)
	306	addi rMEMP,rMEMP,32
	307	addi rLEN,rLEN,-32
	308	std rCHR,32(rMEMP3)
	309	std rCHR,40(rMEMP3)
	310	std rCHR,48(rMEMP3)
	311	std rCHR,56(rMEMP3)
04067002	312
a88f47a7 UD	313	/* We are here because the length or remainder (rLEN) is less than the
	314	cache line/sector size and does not justify aggressive loop unrolling.
	315	So set up the preconditions for L(medium) and go there. */
04067002 UD	316	.align 3
04067002 UD	317	L(handletail32):
a88f47a7 UD	318	cmpldi cr1,rLEN,0
	319	beqlr cr1
	320	b L(medium)
04067002 UD	321
	322	.align 5
	323	L(small):
	324	/* Memset of 8 bytes or less. */
	325	cmpldi cr6, rLEN, 4
	326	cmpldi cr5, rLEN, 1
	327	ble cr6,L(le4)
	328	subi rLEN, rLEN, 4
	329	stb rCHR,0(rMEMP)
	330	stb rCHR,1(rMEMP)
	331	stb rCHR,2(rMEMP)
	332	stb rCHR,3(rMEMP)
	333	addi rMEMP,rMEMP, 4
	334	cmpldi cr5, rLEN, 1
	335	L(le4):
	336	cmpldi cr1, rLEN, 3
	337	bltlr cr5
	338	stb rCHR, 0(rMEMP)
	339	beqlr cr5
	340	stb rCHR, 1(rMEMP)
	341	bltlr cr1
	342	stb rCHR, 2(rMEMP)
	343	beqlr cr1
	344	stb rCHR, 3(rMEMP)
	345	blr
	346
	347	/* Memset of 0-31 bytes. */
	348	.align 5
	349	L(medium):
3be87c77	350	insrdi rCHR, rCHR, 32, 0 /* Replicate word to double word. */
04067002 UD	351	cmpldi cr1, rLEN, 16
	352	L(medium_tail2):
	353	add rMEMP, rMEMP, rLEN
	354	L(medium_tail):
	355	bt- 31, L(medium_31t)
	356	bt- 30, L(medium_30t)
	357	L(medium_30f):
a88f47a7	358	bt 29, L(medium_29t)
04067002	359	L(medium_29f):
a88f47a7 UD	360	bge cr1, L(medium_27t)
a88f47a7 UD	361	bflr 28
04067002 UD	362	std rCHR, -8(rMEMP)
	363	blr
	364
	365	L(medium_31t):
	366	stbu rCHR, -1(rMEMP)
	367	bf- 30, L(medium_30f)
	368	L(medium_30t):
	369	sthu rCHR, -2(rMEMP)
	370	bf- 29, L(medium_29f)
	371	L(medium_29t):
	372	stwu rCHR, -4(rMEMP)
a88f47a7	373	blt cr1, L(medium_27f)
04067002 UD	374	L(medium_27t):
	375	std rCHR, -8(rMEMP)
	376	stdu rCHR, -16(rMEMP)
	377	L(medium_27f):
a88f47a7	378	bflr 28
04067002 UD	379	L(medium_28t):
	380	std rCHR, -8(rMEMP)
	381	blr
18e0054b	382	END_GEN_TB (MEMSET,TB_TOCLESS)
04067002 UD	383	libc_hidden_builtin_def (memset)
	384
	385	/* Copied from bzero.S to prevent the linker from inserting a stub
	386	between bzero and memset. */
d5b41185	387	ENTRY_TOCLESS (__bzero)
04067002	388	CALL_MCOUNT 3
04067002 UD	389	mr r5,r4
	390	li r4,0
	391	b L(_memset)
3b473fec AZ	392	END (__bzero)
3b473fec AZ	393	#ifndef __bzero
2d67d91a	394	weak_alias (__bzero, bzero)
8a29a3d0	395	#endif