162306a36Sopenharmony_ci/* SPDX-License-Identifier: GPL-2.0 */
262306a36Sopenharmony_ci/*
362306a36Sopenharmony_ci * arch/alpha/lib/ev67-strchr.S
462306a36Sopenharmony_ci * 21264 version contributed by Rick Gorton <rick.gorton@alpha-processor.com>
562306a36Sopenharmony_ci *
662306a36Sopenharmony_ci * Return the address of a given character within a null-terminated
762306a36Sopenharmony_ci * string, or null if it is not found.
862306a36Sopenharmony_ci *
962306a36Sopenharmony_ci * Much of the information about 21264 scheduling/coding comes from:
1062306a36Sopenharmony_ci *	Compiler Writer's Guide for the Alpha 21264
1162306a36Sopenharmony_ci *	abbreviated as 'CWG' in other comments here
1262306a36Sopenharmony_ci *	ftp.digital.com/pub/Digital/info/semiconductor/literature/dsc-library.html
1362306a36Sopenharmony_ci * Scheduling notation:
1462306a36Sopenharmony_ci *	E	- either cluster
1562306a36Sopenharmony_ci *	U	- upper subcluster; U0 - subcluster U0; U1 - subcluster U1
1662306a36Sopenharmony_ci *	L	- lower subcluster; L0 - subcluster L0; L1 - subcluster L1
1762306a36Sopenharmony_ci * Try not to change the actual algorithm if possible for consistency.
1862306a36Sopenharmony_ci */
1962306a36Sopenharmony_ci#include <linux/export.h>
2062306a36Sopenharmony_ci#include <asm/regdef.h>
2162306a36Sopenharmony_ci
2262306a36Sopenharmony_ci	.set noreorder
2362306a36Sopenharmony_ci	.set noat
2462306a36Sopenharmony_ci
2562306a36Sopenharmony_ci	.align 4
2662306a36Sopenharmony_ci	.globl strchr
2762306a36Sopenharmony_ci	.ent strchr
2862306a36Sopenharmony_cistrchr:
2962306a36Sopenharmony_ci	.frame sp, 0, ra
3062306a36Sopenharmony_ci	.prologue 0
3162306a36Sopenharmony_ci
3262306a36Sopenharmony_ci	ldq_u   t0, 0(a0)	# L : load first quadword Latency=3
3362306a36Sopenharmony_ci	and	a1, 0xff, t3	# E : 00000000000000ch
3462306a36Sopenharmony_ci	insbl	a1, 1, t5	# U : 000000000000ch00
3562306a36Sopenharmony_ci	insbl	a1, 7, a2	# U : ch00000000000000
3662306a36Sopenharmony_ci
3762306a36Sopenharmony_ci	insbl	t3, 6, a3	# U : 00ch000000000000
3862306a36Sopenharmony_ci	or	t5, t3, a1	# E : 000000000000chch
3962306a36Sopenharmony_ci	andnot  a0, 7, v0	# E : align our loop pointer
4062306a36Sopenharmony_ci	lda	t4, -1		# E : build garbage mask
4162306a36Sopenharmony_ci
4262306a36Sopenharmony_ci	mskqh	t4, a0, t4	# U : only want relevant part of first quad
4362306a36Sopenharmony_ci	or	a2, a3, a2	# E : chch000000000000
4462306a36Sopenharmony_ci	inswl	a1, 2, t5	# E : 00000000chch0000
4562306a36Sopenharmony_ci	inswl	a1, 4, a3	# E : 0000chch00000000
4662306a36Sopenharmony_ci
4762306a36Sopenharmony_ci	or	a1, a2, a1	# E : chch00000000chch
4862306a36Sopenharmony_ci	or	a3, t5, t5	# E : 0000chchchch0000
4962306a36Sopenharmony_ci	cmpbge  zero, t0, t2	# E : bits set iff byte == zero
5062306a36Sopenharmony_ci	cmpbge	zero, t4, t4	# E : bits set iff byte is garbage
5162306a36Sopenharmony_ci
5262306a36Sopenharmony_ci	/* This quad is _very_ serialized.  Lots of stalling happens */
5362306a36Sopenharmony_ci	or	t5, a1, a1	# E : chchchchchchchch
5462306a36Sopenharmony_ci	xor	t0, a1, t1	# E : make bytes == c zero
5562306a36Sopenharmony_ci	cmpbge  zero, t1, t3	# E : bits set iff byte == c
5662306a36Sopenharmony_ci	or	t2, t3, t0	# E : bits set iff char match or zero match
5762306a36Sopenharmony_ci
5862306a36Sopenharmony_ci	andnot	t0, t4, t0	# E : clear garbage bits
5962306a36Sopenharmony_ci	cttz	t0, a2		# U0 : speculative (in case we get a match)
6062306a36Sopenharmony_ci	nop			# E :
6162306a36Sopenharmony_ci	bne	t0, $found	# U :
6262306a36Sopenharmony_ci
6362306a36Sopenharmony_ci	/*
6462306a36Sopenharmony_ci	 * Yuk.  This loop is going to stall like crazy waiting for the
6562306a36Sopenharmony_ci	 * data to be loaded.  Not much can be done about it unless it's
6662306a36Sopenharmony_ci	 * unrolled multiple times - is that safe to do in kernel space?
6762306a36Sopenharmony_ci	 * Or would exception handling recovery code do the trick here?
6862306a36Sopenharmony_ci	 */
6962306a36Sopenharmony_ci$loop:	ldq	t0, 8(v0)	# L : Latency=3
7062306a36Sopenharmony_ci	addq	v0, 8, v0	# E :
7162306a36Sopenharmony_ci	xor	t0, a1, t1	# E :
7262306a36Sopenharmony_ci	cmpbge	zero, t0, t2	# E : bits set iff byte == 0
7362306a36Sopenharmony_ci
7462306a36Sopenharmony_ci	cmpbge	zero, t1, t3	# E : bits set iff byte == c
7562306a36Sopenharmony_ci	or	t2, t3, t0	# E :
7662306a36Sopenharmony_ci	cttz	t3, a2		# U0 : speculative (in case we get a match)
7762306a36Sopenharmony_ci	beq	t0, $loop	# U :
7862306a36Sopenharmony_ci
7962306a36Sopenharmony_ci$found:	negq    t0, t1		# E : clear all but least set bit
8062306a36Sopenharmony_ci	and     t0, t1, t0	# E :
8162306a36Sopenharmony_ci	and	t0, t3, t1	# E : bit set iff byte was the char
8262306a36Sopenharmony_ci	addq	v0, a2, v0	# E : Add in the bit number from above
8362306a36Sopenharmony_ci
8462306a36Sopenharmony_ci	cmoveq	t1, $31, v0	# E : Two mapping slots, latency = 2
8562306a36Sopenharmony_ci	nop
8662306a36Sopenharmony_ci	nop
8762306a36Sopenharmony_ci	ret			# L0 :
8862306a36Sopenharmony_ci
8962306a36Sopenharmony_ci	.end strchr
9062306a36Sopenharmony_ci	EXPORT_SYMBOL(strchr)
91