1/* Cloned and hacked for uClibc by Paul Mundt, December 2003 */
2/* Modified by SuperH, Inc. September 2003 */
3!
4! Fast SH memset
5!
6! by Toshiyasu Morita (tm@netcom.com)
7!
8! SH5 code by J"orn Rennecke (joern.rennecke@superh.com)
9! Copyright 2002 SuperH Ltd.
10!
11
12#include <features.h>
13#include <endian.h>
14
15#if __BYTE_ORDER == __LITTLE_ENDIAN
16#define SHHI shlld
17#define SHLO shlrd
18#else
19#define SHHI shlrd
20#define SHLO shlld
21#endif
22
23	.section .text..SHmedia32,"ax"
24	.globl	memset
25	.type	memset, @function
26
27	.align 5
28
29memset:
30	pta/l multiquad, tr0
31	andi r2, 7, r22
32	ptabs r18, tr2
33	mshflo.b r3,r3,r3
34	add r4, r22, r23
35	mperm.w r3, r63, r3	/* Fill pattern now in every byte of r3 */
36
37	movi 8, r9
38	bgtu/u r23, r9, tr0 /* multiquad */
39
40	beqi/u r4, 0, tr2       /* Return with size 0 - ensures no mem accesses */
41	ldlo.q r2, 0, r7
42	shlli r4, 2, r4
43	movi -1, r8
44	SHHI r8, r4, r8
45	SHHI r8, r4, r8
46	mcmv r7, r8, r3
47	stlo.q r2, 0, r3
48	blink tr2, r63
49
50multiquad:
51	pta/l lastquad, tr0
52	stlo.q r2, 0, r3
53	shlri r23, 3, r24
54	add r2, r4, r5
55	beqi/u r24, 1, tr0 /* lastquad */
56	pta/l loop, tr1
57	sub r2, r22, r25
58	andi r5, -8, r20   /* calculate end address and */
59	addi r20, -7*8, r8 /* loop end address; This might overflow, so we need
60	                      to use a different test before we start the loop
61	                    */
62	bge/u r24, r9, tr1 /* loop */
63	st.q r25, 8, r3
64	st.q r20, -8, r3
65	shlri r24, 1, r24
66	beqi/u r24, 1, tr0 /* lastquad */
67	st.q r25, 16, r3
68	st.q r20, -16, r3
69	beqi/u r24, 2, tr0 /* lastquad */
70	st.q r25, 24, r3
71	st.q r20, -24, r3
72lastquad:
73	sthi.q r5, -1, r3
74	blink tr2,r63
75
76loop:
77!!!	alloco r25, 32	/* QQQ comment out for short-term fix to SHUK #3895.
78			   QQQ commenting out is locically correct, but sub-optimal
79			   QQQ Sean McGoogan - 4th April 2003. */
80	st.q r25, 8, r3
81	st.q r25, 16, r3
82	st.q r25, 24, r3
83	st.q r25, 32, r3
84	addi r25, 32, r25
85	bgeu/l r8, r25, tr1 /* loop */
86
87	st.q r20, -40, r3
88	st.q r20, -32, r3
89	st.q r20, -24, r3
90	st.q r20, -16, r3
91	st.q r20, -8, r3
92	sthi.q r5, -1, r3
93	blink tr2,r63
94
95	.size	memset,.-memset
96
97libc_hidden_def(memset)
98