1/* Optimized strcpy/stpcpy implementation for PowerPC64/POWER8.
2   Copyright (C) 2015-2021 Free Software Foundation, Inc.
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
16   License along with the GNU C Library; if not, see
17   <https://www.gnu.org/licenses/>.  */
18
19#include <sysdep.h>
20
21#ifdef USE_AS_STPCPY
22# ifndef STPCPY
23#   define FUNC_NAME __stpcpy
24# else
25#   define FUNC_NAME STPCPY
26# endif
27#else
28# ifndef STRCPY
29#  define FUNC_NAME strcpy
30# else
31#  define FUNC_NAME STRCPY
32# endif
33#endif  /* !USE_AS_STPCPY  */
34
35/* Implements the function
36
37   char * [r3] strcpy (char *dest [r3], const char *src [r4])
38
39   or
40
41   char * [r3] stpcpy (char *dest [r3], const char *src [r4])
42
43   if USE_AS_STPCPY is defined.
44
45   The implementation uses unaligned doubleword access to avoid specialized
46   code paths depending of data alignment.  Although recent powerpc64 uses
47   64K as default, the page cross handling assumes minimum page size of
48   4k.  */
49
50	.machine  power8
51ENTRY_TOCLESS (FUNC_NAME, 4)
52        li      r0,0          /* Doubleword with null chars to use
53                                 with cmpb.  */
54
55	/* Check if the [src]+15 will cross a 4K page by checking if the bit
56	   indicating the page size changes.  Basically:
57
58	   uint64_t srcin = (uint64_t)src;
59	   uint64_t ob = srcin & 4096UL;
60	   uint64_t nb = (srcin+15UL) & 4096UL;
61	   if (ob ^ nb)
62	     goto pagecross;  */
63
64	addi	r9,r4,15
65	xor	r9,r9,r4
66	rlwinm.	r9,r9,0,19,19
67	bne	L(pagecross)
68
69	/* For short string (less than 16 bytes), just calculate its size as
70	   strlen and issues a memcpy if null is found.  */
71	mr	r7,r4
72        ld      r12,0(r7)     /* Load doubleword from memory.  */
73        cmpb    r10,r12,r0    /* Check for null bytes in DWORD1.  */
74        cmpdi   cr7,r10,0     /* If r10 == 0, no null's have been found.  */
75        bne     cr7,L(done)
76
77        ldu     r8,8(r7)
78        cmpb    r10,r8,r0
79        cmpdi   cr7,r10,0
80        bne     cr7,L(done)
81
82	b	L(loop_before)
83
84	.align	4
85L(pagecross):
86	clrrdi  r7,r4,3       /* Align the address to doubleword boundary.  */
87	rlwinm  r6,r4,3,26,28 /* Calculate padding.  */
88	li      r5,-1         /* MASK = 0xffffffffffffffff.  */
89        ld      r12,0(r7)     /* Load doubleword from memory.  */
90#ifdef __LITTLE_ENDIAN__
91        sld     r5,r5,r6
92#else
93        srd     r5,r5,r6      /* MASK = MASK >> padding.  */
94#endif
95        orc     r9,r12,r5     /* Mask bits that are not part of the string.  */
96        cmpb    r10,r9,r0     /* Check for null bytes in DWORD1.  */
97        cmpdi   cr7,r10,0     /* If r10 == 0, no null's have been found.  */
98        bne     cr7,L(done)
99
100        ldu     r6,8(r7)
101        cmpb    r10,r6,r0
102        cmpdi   cr7,r10,0
103        bne     cr7,L(done)
104
105        ld      r12,0(r7)
106        cmpb    r10,r12,r0
107        cmpdi   cr7,r10,0
108        bne     cr7,L(done)
109
110        ldu     r6,8(r7)
111        cmpb    r10,r6,r0
112        cmpdi   cr7,r10,0
113        bne     cr7,L(done)
114
115	/* We checked for 24 - x bytes, with x being the source alignment
116	   (0 <= x <= 16), and no zero has been found.  Start the loop
117	   copy with doubleword aligned address.  */
118	mr	r7,r4
119	ld	r12, 0(r7)
120	ldu	r8, 8(r7)
121
122L(loop_before):
123	/* Save the two doublewords read from source and align the source
124	   to 16 bytes for the loop.  */
125	mr	r11,r3
126	std	r12,0(r11)
127	std	r8,8(r11)
128	addi	r11,r11,16
129	rldicl	r9,r4,0,60
130	subf	r7,r9,r7
131	subf	r11,r9,r11
132	/* Source is adjusted to 16B alignment and destination r11 is
133	   also moved based on that adjustment.  Now check if r11 is
134	   also 16B aligned to move to vectorized loop.  */
135	andi.	r6, r11, 0xF
136	bne	L(loop_start)
137
138	/* Prepare for the loop.  */
139	subf	r4, r9, r4	/* Adjust r4 based on alignment.  */
140	li	r7, 16	/* Load required offsets.  */
141	li	r8, 32
142	li	r9, 48
143	vspltisb	v0, 0
144	addi	r4, r4, 16
145	/* Are we 64-byte aligned? If so, jump to the vectorized loop.
146	   Else copy 16B till r4 is 64B aligned.  */
147	andi.	r6, r4, 63
148	beq	L(qw_loop)
149
150	lvx	v6, 0, r4	/* Load 16 bytes from memory.  */
151	vcmpequb.	v5, v0, v6	/* Check for null.  */
152	bne	cr6, L(qw_done)
153	stvx	v6, 0, r11	/* Store 16 bytes.  */
154	addi	r4, r4, 16	/* Increment the address.  */
155	addi	r11, r11, 16
156	andi.	r6, r4, 63
157	beq	L(qw_loop)
158
159	lvx	v6, 0, r4
160	vcmpequb.	v5, v0, v6
161	bne	cr6, L(qw_done)
162	stvx	v6, 0, r11
163	addi	r4, r4, 16
164	addi	r11, r11, 16
165	andi.	r6, r4, 63
166	beq	L(qw_loop)
167
168	lvx	v6, 0, r4
169	vcmpequb.	v5, v0, v6
170	bne	cr6, L(qw_done)
171	stvx	v6, 0, r11
172	addi	r4, r4, 16
173	addi	r11, r11, 16
174
175	.align	4
176L(qw_loop):
177	lvx	v1, r4, r0  /* Load 4 quadwords.  */
178	lvx	v2, r4, r7
179	lvx	v3, r4, r8
180	lvx	v4, r4, r9
181	vminub	v5, v1, v2  /* Compare and merge into one VR for speed.  */
182	vminub	v8, v3, v4
183	vminub	v7, v5, v8
184	vcmpequb.	v7, v7, v0  /* Check for NULLs.  */
185	bne	cr6, L(qw_loop_done)
186	stvx	v1, r11, r0	/* Store 4 quadwords.  */
187	stvx	v2, r11, r7
188	stvx	v3, r11, r8
189	stvx	v4, r11, r9
190	addi	r4, r4, 64  /* Adjust address for the next iteration.  */
191	addi	r11, r11, 64	/* Adjust address for the next iteration.  */
192
193	lvx	v1, r4, r0  /* Load 4 quadwords.  */
194	lvx	v2, r4, r7
195	lvx	v3, r4, r8
196	lvx	v4, r4, r9
197	vminub	v5, v1, v2  /* Compare and merge into one VR for speed.  */
198	vminub	v8, v3, v4
199	vminub	v7, v5, v8
200	vcmpequb.	v7, v7, v0  /* Check for NULLs.  */
201	bne	cr6, L(qw_loop_done)
202	stvx	v1, r11, r0	/* Store 4 quadwords.  */
203	stvx	v2, r11, r7
204	stvx	v3, r11, r8
205	stvx	v4, r11, r9
206	addi	r4, r4, 64  /* Adjust address for the next iteration.  */
207	addi	r11, r11, 64	/* Adjust address for the next iteration.  */
208
209	lvx	v1, r4, r0  /* Load 4 quadwords.  */
210	lvx	v2, r4, r7
211	lvx	v3, r4, r8
212	lvx	v4, r4, r9
213	vminub	v5, v1, v2  /* Compare and merge into one VR for speed.  */
214	vminub	v8, v3, v4
215	vminub	v7, v5, v8
216	vcmpequb.	v7, v7, v0  /* Check for NULLs.  */
217	bne	cr6, L(qw_loop_done)
218	stvx	v1, r11, r0	/* Store 4 quadwords.  */
219	stvx	v2, r11, r7
220	stvx	v3, r11, r8
221	stvx	v4, r11, r9
222	addi	r4, r4, 64  /* Adjust address for the next iteration.  */
223	addi	r11, r11, 64	/* Adjust address for the next iteration.  */
224	b	L(qw_loop)
225
226	.align	4
227L(qw_loop_done):
228	/* Null found in one of the 4 loads.  */
229	vcmpequb.	v7, v1, v0
230	vor	v6, v1, v1
231	bne	cr6, L(qw_done)
232	/* Not on the first 16B, So store it.  */
233	stvx	v1, r11, r0
234	addi	r4, r4, 16
235	addi	r11, r11, 16
236	vcmpequb.	v7, v2, v0
237	vor	v6, v2, v2
238	bne	cr6, L(qw_done)
239	/* Not on the second 16B, So store it.  */
240	stvx	v2, r11, r0
241	addi	r4, r4, 16
242	addi	r11, r11, 16
243	vcmpequb.	v7, v3, v0
244	vor	v6, v3, v3
245	bne	cr6, L(qw_done)
246	/* Not on the third 16B, So store it.  */
247	stvx	v6, r11, r0
248	addi	r4, r4, 16
249	addi	r11, r11, 16
250	vor	v6, v4, v4
251
252	.align	4
253L(qw_done):
254	mr	r7, r4
255	/* Move the result to GPR.  */
256#ifdef __LITTLE_ENDIAN__
257	vsldoi	v4, v6, v0, 8
258	mfvrd	r12, v4
259#else
260	mfvrd	r12, v6
261#endif
262	/* Check for null in the first 8 bytes.  */
263	cmpb	r10, r12, r0
264	cmpdi	cr6, r10, 0
265	bne	cr6, L(done2)
266	/* Null found in second doubleword.  */
267#ifdef __LITTLE_ENDIAN__
268	mfvrd	r6, v6
269#else
270	vsldoi	v6, v6, v0, 8
271	mfvrd	r6, v6
272#endif
273	cmpb	r10, r6, r0
274	addi	r7, r7, 8
275	b	L(done2)
276
277        .align  5
278L(loop):
279        std     r12, 0(r11)
280        std     r6, 8(r11)
281	addi	r11,r11,16
282L(loop_start):
283        /* Load two doublewords, compare and merge in a
284           single register for speed.  This is an attempt
285           to speed up the null-checking process for bigger strings.  */
286
287        ld      r12, 8(r7)
288        ldu     r6, 16(r7)
289        cmpb    r10,r12,r0
290        cmpb    r9,r6,r0
291        or      r8,r9,r10     /* Merge everything in one doubleword.  */
292        cmpdi   cr7,r8,0
293        beq     cr7,L(loop)
294
295
296        /* OK, one (or both) of the doublewords contains a null byte.  Check
297           the first doubleword and decrement the address in case the first
298           doubleword really contains a null byte.  */
299
300	addi	r4,r7,-8
301        cmpdi   cr6,r10,0
302        addi    r7,r7,-8
303        bne     cr6,L(done2)
304
305        /* The null byte must be in the second doubleword.  Adjust the address
306           again and move the result of cmpb to r10 so we can calculate the
307           length.  */
308
309        mr      r10,r9
310        addi    r7,r7,8
311	b	L(done2)
312
313        /* r10 has the output of the cmpb instruction, that is, it contains
314           0xff in the same position as the null byte in the original
315           doubleword from the string.  Use that to calculate the length.  */
316L(done):
317	mr	r11,r3
318L(done2):
319#ifdef __LITTLE_ENDIAN__
320        addi    r9, r10, -1   /* Form a mask from trailing zeros.  */
321        andc    r9, r9, r10
322        popcntd r6, r9        /* Count the bits in the mask.  */
323#else
324        cntlzd  r6,r10        /* Count leading zeros before the match.  */
325#endif
326        subf    r5,r4,r7
327        srdi    r6,r6,3       /* Convert leading/trailing zeros to bytes.  */
328        add     r8,r5,r6      /* Compute final length.  */
329#ifdef USE_AS_STPCPY
330	/* stpcpy returns the dest address plus the size not counting the
331	   final '\0'.  */
332	add	r3,r11,r8
333#endif
334	addi	r8,r8,1       /* Final '/0'.  */
335
336	cmpldi	cr6,r8,8
337	mtocrf	0x01,r8
338	ble	cr6,L(copy_LE_8)
339
340	cmpldi	cr1,r8,16
341	blt	cr1,8f
342
343	/* Handle copies of 0~31 bytes.  */
344	.align	4
345L(copy_LT_32):
346	/* At least 6 bytes to go.  */
347	blt	cr1,8f
348
349	/* Copy 16 bytes.  */
350	ld	r6,0(r4)
351	ld	r8,8(r4)
352	addi	r4,r4,16
353	std	r6,0(r11)
354	std	r8,8(r11)
355	addi	r11,r11,16
3568:	/* Copy 8 bytes.  */
357	bf	28,L(tail4)
358	ld	r6,0(r4)
359	addi	r4,r4,8
360	std	r6,0(r11)
361	addi	r11,r11,8
362
363	.align	4
364/* Copies 4~7 bytes.  */
365L(tail4):
366	bf	29,L(tail2)
367	lwz	r6,0(r4)
368	stw	r6,0(r11)
369	bf	30,L(tail5)
370	lhz	r7,4(r4)
371	sth	r7,4(r11)
372	bflr	31
373	lbz	r8,6(r4)
374	stb	r8,6(r11)
375	blr
376
377	.align	4
378/* Copies 2~3 bytes.  */
379L(tail2):
380	bf	30,1f
381	lhz	r6,0(r4)
382	sth	r6,0(r11)
383	bflr	31
384	lbz	r7,2(r4)
385	stb	r7,2(r11)
386	blr
387
388	.align	4
389L(tail5):
390	bf	31,1f
391	lbz	r6,4(r4)
392	stb	r6,4(r11)
393	blr
394
395	.align	4
3961:
397	bflr	31
398	lbz	r6,0(r4)
399	stb	r6,0(r11)
400	blr
401
402/* Handles copies of 0~8 bytes.  */
403	.align	4
404L(copy_LE_8):
405	bne	cr6,L(tail4)
406	ld	r6,0(r4)
407	std	r6,0(r11)
408	blr
409END (FUNC_NAME)
410
411#ifndef USE_AS_STPCPY
412libc_hidden_builtin_def (strcpy)
413#endif
414