1/* Optimized version of the standard memcmp() function.
2   This file is part of the GNU C Library.
3   Copyright (C) 2000-2022 Free Software Foundation, Inc.
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
16   License along with the GNU C Library; if not, see
17   <https://www.gnu.org/licenses/>.  */
18
19/* Return: the result of the comparison
20
21   Inputs:
22        in0:    dest (aka s1)
23        in1:    src  (aka s2)
24        in2:    byte count
25
26   In this form, it assumes little endian mode.  For big endian mode,
27   the two shifts in .l2 must be inverted:
28
29	shl   	tmp1[0] = r[1 + MEMLAT], sh1   // tmp1 = w0 << sh1
30	shr.u   tmp2[0] = r[0 + MEMLAT], sh2   // tmp2 = w1 >> sh2
31
32   and all the mux1 instructions should be replaced by plain mov's.  */
33
34#include <sysdep.h>
35#undef ret
36
37#define OP_T_THRES 	16
38#define OPSIZ 		8
39#define MEMLAT		2
40
41#define start		r15
42#define saved_pr	r17
43#define saved_lc	r18
44#define dest		r19
45#define src		r20
46#define len		r21
47#define asrc		r22
48#define tmp		r23
49#define value1		r24
50#define value2		r25
51#define sh2		r28
52#define	sh1		r29
53#define loopcnt		r30
54
55ENTRY(memcmp)
56	.prologue
57	alloc 	r2 = ar.pfs, 3, 37, 0, 40
58
59	.rotr	r[MEMLAT + 2], q[MEMLAT + 5], tmp1[4], tmp2[4], val[2]
60	.rotp	p[MEMLAT + 4 + 1]
61
62	mov	ret0 = r0		// by default return value = 0
63	.save pr, saved_pr
64	mov	saved_pr = pr		// save the predicate registers
65	.save ar.lc, saved_lc
66        mov 	saved_lc = ar.lc	// save the loop counter
67	.body
68	mov 	dest = in0		// dest
69	mov 	src = in1		// src
70	mov	len = in2		// len
71	sub	tmp = r0, in0		// tmp = -dest
72	;;
73	and	loopcnt = 7, tmp		// loopcnt = -dest % 8
74	cmp.ge	p6, p0 = OP_T_THRES, len	// is len <= OP_T_THRES
75(p6)	br.cond.spnt	.cmpfew			// compare byte by byte
76	;;
77	cmp.eq	p6, p0 = loopcnt, r0
78(p6)	br.cond.sptk .dest_aligned
79	sub	len = len, loopcnt	// len -= -dest % 8
80	adds	loopcnt = -1, loopcnt	// --loopcnt
81	;;
82	mov	ar.lc = loopcnt
83.l1:					// copy -dest % 8 bytes
84	ld1	value1 = [src], 1	// value = *src++
85	ld1	value2 = [dest], 1
86	;;
87	cmp.ne	p6, p0 = value1, value2
88(p6)	br.cond.spnt .done
89	br.cloop.dptk .l1
90.dest_aligned:
91	and	sh1 = 7, src 		// sh1 = src % 8
92	and	tmp = -8, len   	// tmp = len & -OPSIZ
93	and	asrc = -8, src		// asrc = src & -OPSIZ  -- align src
94	shr.u	loopcnt = len, 3	// loopcnt = len / 8
95	and	len = 7, len ;;		// len = len % 8
96	shl	sh1 = sh1, 3		// sh1 = 8 * (src % 8)
97	adds	loopcnt = -1, loopcnt	// --loopcnt
98	mov     pr.rot = 1 << 16 ;;	// set rotating predicates
99	sub	sh2 = 64, sh1		// sh2 = 64 - sh1
100	mov	ar.lc = loopcnt		// set LC
101	cmp.eq  p6, p0 = sh1, r0 	// is the src aligned?
102(p6)    br.cond.sptk .src_aligned
103	add	src = src, tmp		// src += len & -OPSIZ
104	mov	ar.ec = MEMLAT + 4 + 1 	// four more passes needed
105	ld8	r[1] = [asrc], 8 ;;	// r[1] = w0
106	.align	32
107
108// We enter this loop with p6 cleared by the above comparison
109
110.l2:
111(p[0])		ld8	r[0] = [asrc], 8		// r[0] = w1
112(p[0])		ld8	q[0] = [dest], 8
113(p[MEMLAT])	shr.u	tmp1[0] = r[1 + MEMLAT], sh1	// tmp1 = w0 >> sh1
114(p[MEMLAT])	shl	tmp2[0] = r[0 + MEMLAT], sh2  	// tmp2 = w1 << sh2
115(p[MEMLAT+4])	cmp.ne	p6, p0 = q[MEMLAT + 4], val[1]
116(p[MEMLAT+3])	or	val[0] = tmp1[3], tmp2[3] 	// val = tmp1 | tmp2
117(p6)		br.cond.spnt .l2exit
118		br.ctop.sptk    .l2
119		br.cond.sptk .cmpfew
120.l3exit:
121	mux1	value1 = r[MEMLAT], @rev
122	mux1	value2 = q[MEMLAT], @rev
123	cmp.ne	p6, p0 = r0, r0	;;	// clear p6
124.l2exit:
125(p6)	mux1	value1 = val[1], @rev
126(p6)	mux1	value2 = q[MEMLAT + 4], @rev ;;
127	cmp.ltu	p6, p7 = value2, value1 ;;
128(p6)	mov	ret0 = -1
129(p7)	mov	ret0 = 1
130	mov     pr = saved_pr, -1    	// restore the predicate registers
131	mov 	ar.lc = saved_lc	// restore the loop counter
132	br.ret.sptk.many b0
133.src_aligned:
134	cmp.ne	p6, p0 = r0, r0		// clear p6
135	mov     ar.ec = MEMLAT + 1 ;;	// set EC
136.l3:
137(p[0])		ld8	r[0] = [src], 8
138(p[0])		ld8	q[0] = [dest], 8
139(p[MEMLAT])	cmp.ne	p6, p0 = r[MEMLAT], q[MEMLAT]
140(p6)		br.cond.spnt .l3exit
141		br.ctop.dptk .l3 ;;
142.cmpfew:
143	cmp.eq	p6, p0 = len, r0	// is len == 0 ?
144	adds	len = -1, len		// --len;
145(p6)	br.cond.spnt	.restore_and_exit ;;
146	mov	ar.lc = len
147.l4:
148	ld1	value1 = [src], 1
149	ld1	value2 = [dest], 1
150	;;
151	cmp.ne	p6, p0 = value1, value2
152(p6)	br.cond.spnt	.done
153	br.cloop.dptk	.l4 ;;
154.done:
155(p6)	sub	ret0 = value2, value1	// don't execute it if falling thru
156.restore_and_exit:
157	mov     pr = saved_pr, -1    	// restore the predicate registers
158	mov 	ar.lc = saved_lc	// restore the loop counter
159	br.ret.sptk.many b0
160END(memcmp)
161
162weak_alias (memcmp, bcmp)
163strong_alias (memcmp, __memcmpeq)
164libc_hidden_builtin_def (memcmp)
165libc_hidden_def (__memcmpeq)
166