162306a36Sopenharmony_ci/* SPDX-License-Identifier: GPL-2.0 */
262306a36Sopenharmony_ci/*
362306a36Sopenharmony_ci * arch/alpha/lib/strrchr.S
462306a36Sopenharmony_ci * Contributed by Richard Henderson (rth@tamu.edu)
562306a36Sopenharmony_ci *
662306a36Sopenharmony_ci * Return the address of the last occurrence of a given character
762306a36Sopenharmony_ci * within a null-terminated string, or null if it is not found.
862306a36Sopenharmony_ci */
962306a36Sopenharmony_ci#include <linux/export.h>
1062306a36Sopenharmony_ci#include <asm/regdef.h>
1162306a36Sopenharmony_ci
1262306a36Sopenharmony_ci	.set noreorder
1362306a36Sopenharmony_ci	.set noat
1462306a36Sopenharmony_ci
1562306a36Sopenharmony_ci	.align 3
1662306a36Sopenharmony_ci	.ent strrchr
1762306a36Sopenharmony_ci	.globl strrchr
1862306a36Sopenharmony_cistrrchr:
1962306a36Sopenharmony_ci	.frame sp, 0, ra
2062306a36Sopenharmony_ci	.prologue 0
2162306a36Sopenharmony_ci
2262306a36Sopenharmony_ci	zapnot	a1, 1, a1	# e0    : zero extend our test character
2362306a36Sopenharmony_ci	mov	zero, t6	# .. e1 : t6 is last match aligned addr
2462306a36Sopenharmony_ci	sll	a1, 8, t5	# e0    : replicate our test character
2562306a36Sopenharmony_ci	mov	zero, t8	# .. e1 : t8 is last match byte compare mask
2662306a36Sopenharmony_ci	or	t5, a1, a1	# e0    :
2762306a36Sopenharmony_ci	ldq_u   t0, 0(a0)	# .. e1 : load first quadword
2862306a36Sopenharmony_ci	sll	a1, 16, t5	# e0    :
2962306a36Sopenharmony_ci	andnot  a0, 7, v0	# .. e1 : align source addr
3062306a36Sopenharmony_ci	or	t5, a1, a1	# e0    :
3162306a36Sopenharmony_ci	lda	t4, -1		# .. e1 : build garbage mask
3262306a36Sopenharmony_ci	sll	a1, 32, t5	# e0    :
3362306a36Sopenharmony_ci	cmpbge  zero, t0, t1	# .. e1 : bits set iff byte == zero
3462306a36Sopenharmony_ci	mskqh	t4, a0, t4	# e0    :
3562306a36Sopenharmony_ci	or	t5, a1, a1	# .. e1 : character replication complete
3662306a36Sopenharmony_ci	xor	t0, a1, t2	# e0    : make bytes == c zero
3762306a36Sopenharmony_ci	cmpbge	zero, t4, t4	# .. e1 : bits set iff byte is garbage
3862306a36Sopenharmony_ci	cmpbge  zero, t2, t3	# e0    : bits set iff byte == c
3962306a36Sopenharmony_ci	andnot	t1, t4, t1	# .. e1 : clear garbage from null test
4062306a36Sopenharmony_ci	andnot	t3, t4, t3	# e0    : clear garbage from char test
4162306a36Sopenharmony_ci	bne	t1, $eos	# .. e1 : did we already hit the terminator?
4262306a36Sopenharmony_ci
4362306a36Sopenharmony_ci	/* Character search main loop */
4462306a36Sopenharmony_ci$loop:
4562306a36Sopenharmony_ci	ldq	t0, 8(v0)	# e0    : load next quadword
4662306a36Sopenharmony_ci	cmovne	t3, v0, t6	# .. e1 : save previous comparisons match
4762306a36Sopenharmony_ci	cmovne	t3, t3, t8	# e0    :
4862306a36Sopenharmony_ci	addq	v0, 8, v0	# .. e1 :
4962306a36Sopenharmony_ci	xor	t0, a1, t2	# e0    :
5062306a36Sopenharmony_ci	cmpbge	zero, t0, t1	# .. e1 : bits set iff byte == zero
5162306a36Sopenharmony_ci	cmpbge	zero, t2, t3	# e0    : bits set iff byte == c
5262306a36Sopenharmony_ci	beq	t1, $loop	# .. e1 : if we havnt seen a null, loop
5362306a36Sopenharmony_ci
5462306a36Sopenharmony_ci	/* Mask out character matches after terminator */
5562306a36Sopenharmony_ci$eos:
5662306a36Sopenharmony_ci	negq	t1, t4		# e0    : isolate first null byte match
5762306a36Sopenharmony_ci	and	t1, t4, t4	# e1    :
5862306a36Sopenharmony_ci	subq	t4, 1, t5	# e0    : build a mask of the bytes up to...
5962306a36Sopenharmony_ci	or	t4, t5, t4	# e1    : ... and including the null
6062306a36Sopenharmony_ci
6162306a36Sopenharmony_ci	and	t3, t4, t3	# e0    : mask out char matches after null
6262306a36Sopenharmony_ci	cmovne	t3, t3, t8	# .. e1 : save it, if match found
6362306a36Sopenharmony_ci	cmovne	t3, v0, t6	# e0    :
6462306a36Sopenharmony_ci
6562306a36Sopenharmony_ci	/* Locate the address of the last matched character */
6662306a36Sopenharmony_ci
6762306a36Sopenharmony_ci	/* Retain the early exit for the ev4 -- the ev5 mispredict penalty
6862306a36Sopenharmony_ci	   is 5 cycles -- the same as just falling through.  */
6962306a36Sopenharmony_ci	beq	t8, $retnull	# .. e1 :
7062306a36Sopenharmony_ci
7162306a36Sopenharmony_ci	and	t8, 0xf0, t2	# e0    : binary search for the high bit set
7262306a36Sopenharmony_ci	cmovne	t2, t2, t8	# .. e1 (zdb)
7362306a36Sopenharmony_ci	cmovne	t2, 4, t2	# e0    :
7462306a36Sopenharmony_ci	and	t8, 0xcc, t1	# .. e1 :
7562306a36Sopenharmony_ci	cmovne	t1, t1, t8	# e0    :
7662306a36Sopenharmony_ci	cmovne	t1, 2, t1	# .. e1 :
7762306a36Sopenharmony_ci	and	t8, 0xaa, t0	# e0    :
7862306a36Sopenharmony_ci	cmovne	t0, 1, t0	# .. e1 (zdb)
7962306a36Sopenharmony_ci	addq	t2, t1, t1	# e0    :
8062306a36Sopenharmony_ci	addq	t6, t0, v0	# .. e1 : add our aligned base ptr to the mix
8162306a36Sopenharmony_ci	addq	v0, t1, v0	# e0    :
8262306a36Sopenharmony_ci	ret			# .. e1 :
8362306a36Sopenharmony_ci
8462306a36Sopenharmony_ci$retnull:
8562306a36Sopenharmony_ci	mov	zero, v0	# e0    :
8662306a36Sopenharmony_ci	ret			# .. e1 :
8762306a36Sopenharmony_ci
8862306a36Sopenharmony_ci	.end strrchr
8962306a36Sopenharmony_ci	EXPORT_SYMBOL(strrchr)
90