xref: /linux/arch/riscv/lib/memmove.S (revision 24168c5e6dfbdd5b414f048f47f75d64533296ca)
1/* SPDX-License-Identifier: GPL-2.0-only */
2/*
3 * Copyright (C) 2022 Michael T. Kloos <michael@michaelkloos.com>
4 */
5
6#include <linux/linkage.h>
7#include <asm/asm.h>
8
9SYM_FUNC_START(__memmove)
10	/*
11	 * Returns
12	 *   a0 - dest
13	 *
14	 * Parameters
15	 *   a0 - Inclusive first byte of dest
16	 *   a1 - Inclusive first byte of src
17	 *   a2 - Length of copy n
18	 *
19	 * Because the return matches the parameter register a0,
20	 * we will not clobber or modify that register.
21	 *
22	 * Note: This currently only works on little-endian.
23	 * To port to big-endian, reverse the direction of shifts
24	 * in the 2 misaligned fixup copy loops.
25	 */
26
27	/* Return if nothing to do */
28	beq a0, a1, .Lreturn_from_memmove
29	beqz a2, .Lreturn_from_memmove
30
31	/*
32	 * Register Uses
33	 *      Forward Copy: a1 - Index counter of src
34	 *      Reverse Copy: a4 - Index counter of src
35	 *      Forward Copy: t3 - Index counter of dest
36	 *      Reverse Copy: t4 - Index counter of dest
37	 *   Both Copy Modes: t5 - Inclusive first multibyte/aligned of dest
38	 *   Both Copy Modes: t6 - Non-Inclusive last multibyte/aligned of dest
39	 *   Both Copy Modes: t0 - Link / Temporary for load-store
40	 *   Both Copy Modes: t1 - Temporary for load-store
41	 *   Both Copy Modes: t2 - Temporary for load-store
42	 *   Both Copy Modes: a5 - dest to src alignment offset
43	 *   Both Copy Modes: a6 - Shift ammount
44	 *   Both Copy Modes: a7 - Inverse Shift ammount
45	 *   Both Copy Modes: a2 - Alternate breakpoint for unrolled loops
46	 */
47
48	/*
49	 * Solve for some register values now.
50	 * Byte copy does not need t5 or t6.
51	 */
52	mv   t3, a0
53	add  t4, a0, a2
54	add  a4, a1, a2
55
56	/*
57	 * Byte copy if copying less than (2 * SZREG) bytes. This can
58	 * cause problems with the bulk copy implementation and is
59	 * small enough not to bother.
60	 */
61	andi t0, a2, -(2 * SZREG)
62	beqz t0, .Lbyte_copy
63
64	/*
65	 * Now solve for t5 and t6.
66	 */
67	andi t5, t3, -SZREG
68	andi t6, t4, -SZREG
69	/*
70	 * If dest(Register t3) rounded down to the nearest naturally
71	 * aligned SZREG address, does not equal dest, then add SZREG
72	 * to find the low-bound of SZREG alignment in the dest memory
73	 * region.  Note that this could overshoot the dest memory
74	 * region if n is less than SZREG.  This is one reason why
75	 * we always byte copy if n is less than SZREG.
76	 * Otherwise, dest is already naturally aligned to SZREG.
77	 */
78	beq  t5, t3, 1f
79		addi t5, t5, SZREG
80	1:
81
82	/*
83	 * If the dest and src are co-aligned to SZREG, then there is
84	 * no need for the full rigmarole of a full misaligned fixup copy.
85	 * Instead, do a simpler co-aligned copy.
86	 */
87	xor  t0, a0, a1
88	andi t1, t0, (SZREG - 1)
89	beqz t1, .Lcoaligned_copy
90	/* Fall through to misaligned fixup copy */
91
92.Lmisaligned_fixup_copy:
93	bltu a1, a0, .Lmisaligned_fixup_copy_reverse
94
95.Lmisaligned_fixup_copy_forward:
96	jal  t0, .Lbyte_copy_until_aligned_forward
97
98	andi a5, a1, (SZREG - 1) /* Find the alignment offset of src (a1) */
99	slli a6, a5, 3 /* Multiply by 8 to convert that to bits to shift */
100	sub  a5, a1, t3 /* Find the difference between src and dest */
101	andi a1, a1, -SZREG /* Align the src pointer */
102	addi a2, t6, SZREG /* The other breakpoint for the unrolled loop*/
103
104	/*
105	 * Compute The Inverse Shift
106	 * a7 = XLEN - a6 = XLEN + -a6
107	 * 2s complement negation to find the negative: -a6 = ~a6 + 1
108	 * Add that to XLEN.  XLEN = SZREG * 8.
109	 */
110	not  a7, a6
111	addi a7, a7, (SZREG * 8 + 1)
112
113	/*
114	 * Fix Misalignment Copy Loop - Forward
115	 * load_val0 = load_ptr[0];
116	 * do {
117	 * 	load_val1 = load_ptr[1];
118	 * 	store_ptr += 2;
119	 * 	store_ptr[0 - 2] = (load_val0 >> {a6}) | (load_val1 << {a7});
120	 *
121	 * 	if (store_ptr == {a2})
122	 * 		break;
123	 *
124	 * 	load_val0 = load_ptr[2];
125	 * 	load_ptr += 2;
126	 * 	store_ptr[1 - 2] = (load_val1 >> {a6}) | (load_val0 << {a7});
127	 *
128	 * } while (store_ptr != store_ptr_end);
129	 * store_ptr = store_ptr_end;
130	 */
131
132	REG_L t0, (0 * SZREG)(a1)
133	1:
134	REG_L t1, (1 * SZREG)(a1)
135	addi  t3, t3, (2 * SZREG)
136	srl   t0, t0, a6
137	sll   t2, t1, a7
138	or    t2, t0, t2
139	REG_S t2, ((0 * SZREG) - (2 * SZREG))(t3)
140
141	beq   t3, a2, 2f
142
143	REG_L t0, (2 * SZREG)(a1)
144	addi  a1, a1, (2 * SZREG)
145	srl   t1, t1, a6
146	sll   t2, t0, a7
147	or    t2, t1, t2
148	REG_S t2, ((1 * SZREG) - (2 * SZREG))(t3)
149
150	bne   t3, t6, 1b
151	2:
152	mv    t3, t6 /* Fix the dest pointer in case the loop was broken */
153
154	add  a1, t3, a5 /* Restore the src pointer */
155	j .Lbyte_copy_forward /* Copy any remaining bytes */
156
157.Lmisaligned_fixup_copy_reverse:
158	jal  t0, .Lbyte_copy_until_aligned_reverse
159
160	andi a5, a4, (SZREG - 1) /* Find the alignment offset of src (a4) */
161	slli a6, a5, 3 /* Multiply by 8 to convert that to bits to shift */
162	sub  a5, a4, t4 /* Find the difference between src and dest */
163	andi a4, a4, -SZREG /* Align the src pointer */
164	addi a2, t5, -SZREG /* The other breakpoint for the unrolled loop*/
165
166	/*
167	 * Compute The Inverse Shift
168	 * a7 = XLEN - a6 = XLEN + -a6
169	 * 2s complement negation to find the negative: -a6 = ~a6 + 1
170	 * Add that to XLEN.  XLEN = SZREG * 8.
171	 */
172	not  a7, a6
173	addi a7, a7, (SZREG * 8 + 1)
174
175	/*
176	 * Fix Misalignment Copy Loop - Reverse
177	 * load_val1 = load_ptr[0];
178	 * do {
179	 * 	load_val0 = load_ptr[-1];
180	 * 	store_ptr -= 2;
181	 * 	store_ptr[1] = (load_val0 >> {a6}) | (load_val1 << {a7});
182	 *
183	 * 	if (store_ptr == {a2})
184	 * 		break;
185	 *
186	 * 	load_val1 = load_ptr[-2];
187	 * 	load_ptr -= 2;
188	 * 	store_ptr[0] = (load_val1 >> {a6}) | (load_val0 << {a7});
189	 *
190	 * } while (store_ptr != store_ptr_end);
191	 * store_ptr = store_ptr_end;
192	 */
193
194	REG_L t1, ( 0 * SZREG)(a4)
195	1:
196	REG_L t0, (-1 * SZREG)(a4)
197	addi  t4, t4, (-2 * SZREG)
198	sll   t1, t1, a7
199	srl   t2, t0, a6
200	or    t2, t1, t2
201	REG_S t2, ( 1 * SZREG)(t4)
202
203	beq   t4, a2, 2f
204
205	REG_L t1, (-2 * SZREG)(a4)
206	addi  a4, a4, (-2 * SZREG)
207	sll   t0, t0, a7
208	srl   t2, t1, a6
209	or    t2, t0, t2
210	REG_S t2, ( 0 * SZREG)(t4)
211
212	bne   t4, t5, 1b
213	2:
214	mv    t4, t5 /* Fix the dest pointer in case the loop was broken */
215
216	add  a4, t4, a5 /* Restore the src pointer */
217	j .Lbyte_copy_reverse /* Copy any remaining bytes */
218
219/*
220 * Simple copy loops for SZREG co-aligned memory locations.
221 * These also make calls to do byte copies for any unaligned
222 * data at their terminations.
223 */
224.Lcoaligned_copy:
225	bltu a1, a0, .Lcoaligned_copy_reverse
226
227.Lcoaligned_copy_forward:
228	jal t0, .Lbyte_copy_until_aligned_forward
229
230	1:
231	REG_L t1, ( 0 * SZREG)(a1)
232	addi  a1, a1, SZREG
233	addi  t3, t3, SZREG
234	REG_S t1, (-1 * SZREG)(t3)
235	bne   t3, t6, 1b
236
237	j .Lbyte_copy_forward /* Copy any remaining bytes */
238
239.Lcoaligned_copy_reverse:
240	jal t0, .Lbyte_copy_until_aligned_reverse
241
242	1:
243	REG_L t1, (-1 * SZREG)(a4)
244	addi  a4, a4, -SZREG
245	addi  t4, t4, -SZREG
246	REG_S t1, ( 0 * SZREG)(t4)
247	bne   t4, t5, 1b
248
249	j .Lbyte_copy_reverse /* Copy any remaining bytes */
250
251/*
252 * These are basically sub-functions within the function.  They
253 * are used to byte copy until the dest pointer is in alignment.
254 * At which point, a bulk copy method can be used by the
255 * calling code.  These work on the same registers as the bulk
256 * copy loops.  Therefore, the register values can be picked
257 * up from where they were left and we avoid code duplication
258 * without any overhead except the call in and return jumps.
259 */
260.Lbyte_copy_until_aligned_forward:
261	beq  t3, t5, 2f
262	1:
263	lb   t1,  0(a1)
264	addi a1, a1, 1
265	addi t3, t3, 1
266	sb   t1, -1(t3)
267	bne  t3, t5, 1b
268	2:
269	jalr zero, 0x0(t0) /* Return to multibyte copy loop */
270
271.Lbyte_copy_until_aligned_reverse:
272	beq  t4, t6, 2f
273	1:
274	lb   t1, -1(a4)
275	addi a4, a4, -1
276	addi t4, t4, -1
277	sb   t1,  0(t4)
278	bne  t4, t6, 1b
279	2:
280	jalr zero, 0x0(t0) /* Return to multibyte copy loop */
281
282/*
283 * Simple byte copy loops.
284 * These will byte copy until they reach the end of data to copy.
285 * At that point, they will call to return from memmove.
286 */
287.Lbyte_copy:
288	bltu a1, a0, .Lbyte_copy_reverse
289
290.Lbyte_copy_forward:
291	beq  t3, t4, 2f
292	1:
293	lb   t1,  0(a1)
294	addi a1, a1, 1
295	addi t3, t3, 1
296	sb   t1, -1(t3)
297	bne  t3, t4, 1b
298	2:
299	ret
300
301.Lbyte_copy_reverse:
302	beq  t4, t3, 2f
303	1:
304	lb   t1, -1(a4)
305	addi a4, a4, -1
306	addi t4, t4, -1
307	sb   t1,  0(t4)
308	bne  t4, t3, 1b
309	2:
310
311.Lreturn_from_memmove:
312	ret
313
314SYM_FUNC_END(__memmove)
315SYM_FUNC_ALIAS_WEAK(memmove, __memmove)
316SYM_FUNC_ALIAS(__pi_memmove, __memmove)
317SYM_FUNC_ALIAS(__pi___memmove, __memmove)
318