xref: /linux/arch/hexagon/lib/memcpy.S (revision 87c9c16317882dd6dbbc07e349bc3223e14f3244)
1/* SPDX-License-Identifier: GPL-2.0-only */
2/*
3 * Copyright (c) 2010-2011, The Linux Foundation. All rights reserved.
4 */
5
6/*
7 * Description
8 *
9 *   library function for memcpy where length bytes are copied from
10 *   ptr_in to ptr_out. ptr_out is returned unchanged.
11 *   Allows any combination of alignment on input and output pointers
12 *   and length from 0 to 2^32-1
13 *
14 * Restrictions
15 *   The arrays should not overlap, the program will produce undefined output
16 *   if they do.
17 *   For blocks less than 16 bytes a byte by byte copy is performed. For
18 *   8byte alignments, and length multiples, a dword copy is performed up to
19 *   96bytes
20 * History
21 *
22 *   DJH  5/15/09 Initial version 1.0
23 *   DJH  6/ 1/09 Version 1.1 modified ABI to inlcude R16-R19
24 *   DJH  7/12/09 Version 1.2 optimized codesize down to 760 was 840
25 *   DJH 10/14/09 Version 1.3 added special loop for aligned case, was
26 *                            overreading bloated codesize back up to 892
27 *   DJH  4/20/10 Version 1.4 fixed Ldword_loop_epilog loop to prevent loads
28 *                            occurring if only 1 left outstanding, fixes bug
29 *                            # 3888, corrected for all alignments. Peeled off
30 *                            1 32byte chunk from kernel loop and extended 8byte
31 *                            loop at end to solve all combinations and prevent
32 *                            over read.  Fixed Ldword_loop_prolog to prevent
33 *                            overread for blocks less than 48bytes. Reduced
34 *                            codesize to 752 bytes
35 *   DJH  4/21/10 version 1.5 1.4 fix broke code for input block ends not
36 *                            aligned to dword boundaries,underwriting by 1
37 *                            byte, added detection for this and fixed. A
38 *                            little bloat.
39 *   DJH  4/23/10 version 1.6 corrected stack error, R20 was not being restored
40 *                            always, fixed the error of R20 being modified
41 *                            before it was being saved
42 * Natural c model
43 * ===============
44 * void * memcpy(char * ptr_out, char * ptr_in, int length) {
45 *   int i;
46 *   if(length) for(i=0; i < length; i++) { ptr_out[i] = ptr_in[i]; }
47 *   return(ptr_out);
48 * }
49 *
50 * Optimized memcpy function
51 * =========================
52 * void * memcpy(char * ptr_out, char * ptr_in, int len) {
53 *   int i, prolog, kernel, epilog, mask;
54 *   u8 offset;
55 *   s64 data0, dataF8, data70;
56 *
57 *   s64 * ptr8_in;
58 *   s64 * ptr8_out;
59 *   s32 * ptr4;
60 *   s16 * ptr2;
61 *
62 *   offset = ((int) ptr_in) & 7;
63 *   ptr8_in = (s64 *) &ptr_in[-offset];   //read in the aligned pointers
64 *
65 *   data70 = *ptr8_in++;
66 *   dataF8 = *ptr8_in++;
67 *
68 *   data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
69 *
70 *   prolog = 32 - ((int) ptr_out);
71 *   mask  = 0x7fffffff >> HEXAGON_R_cl0_R(len);
72 *   prolog = prolog & mask;
73 *   kernel = len - prolog;
74 *   epilog = kernel & 0x1F;
75 *   kernel = kernel>>5;
76 *
77 *   if (prolog & 1) { ptr_out[0] = (u8) data0; data0 >>= 8; ptr_out += 1;}
78 *   ptr2 = (s16 *) &ptr_out[0];
79 *   if (prolog & 2) { ptr2[0] = (u16) data0;  data0 >>= 16; ptr_out += 2;}
80 *   ptr4 = (s32 *) &ptr_out[0];
81 *   if (prolog & 4) { ptr4[0] = (u32) data0;  data0 >>= 32; ptr_out += 4;}
82 *
83 *   offset = offset + (prolog & 7);
84 *   if (offset >= 8) {
85 *     data70 = dataF8;
86 *     dataF8 = *ptr8_in++;
87 *   }
88 *   offset = offset & 0x7;
89 *
90 *   prolog = prolog >> 3;
91 *   if (prolog) for (i=0; i < prolog; i++) {
92 *       data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
93 *       ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
94 *       data70 = dataF8;
95 *       dataF8 = *ptr8_in++;
96 *   }
97 *   if(kernel) { kernel -= 1; epilog += 32; }
98 *   if(kernel) for(i=0; i < kernel; i++) {
99 *       data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
100 *       ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
101 *       data70 = *ptr8_in++;
102 *
103 *       data0 = HEXAGON_P_valignb_PPp(data70, dataF8, offset);
104 *       ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
105 *       dataF8 = *ptr8_in++;
106 *
107 *       data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
108 *       ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
109 *       data70 = *ptr8_in++;
110 *
111 *       data0 = HEXAGON_P_valignb_PPp(data70, dataF8, offset);
112 *       ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
113 *       dataF8 = *ptr8_in++;
114 *   }
115 *   epilogdws = epilog >> 3;
116 *   if (epilogdws) for (i=0; i < epilogdws; i++) {
117 *       data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
118 *       ptr8_out = (s64 *) &ptr_out[0]; *ptr8_out = data0; ptr_out += 8;
119 *       data70 = dataF8;
120 *       dataF8 = *ptr8_in++;
121 *   }
122 *   data0 = HEXAGON_P_valignb_PPp(dataF8, data70, offset);
123 *
124 *   ptr4 = (s32 *) &ptr_out[0];
125 *   if (epilog & 4) { ptr4[0] = (u32) data0; data0 >>= 32; ptr_out += 4;}
126 *   ptr2 = (s16 *) &ptr_out[0];
127 *   if (epilog & 2) { ptr2[0] = (u16) data0; data0 >>= 16; ptr_out += 2;}
128 *   if (epilog & 1) { *ptr_out++ = (u8) data0; }
129 *
130 *   return(ptr_out - length);
131 * }
132 *
133 * Codesize : 784 bytes
134 */
135
136
137#define ptr_out		R0	/*  destination  pounter  */
138#define ptr_in		R1	/*  source pointer  */
139#define len		R2	/*  length of copy in bytes  */
140
141#define data70		R13:12	/*  lo 8 bytes of non-aligned transfer  */
142#define dataF8		R11:10	/*  hi 8 bytes of non-aligned transfer  */
143#define ldata0		R7:6	/*  even 8 bytes chunks  */
144#define ldata1		R25:24	/*  odd 8 bytes chunks  */
145#define data1		R7	/*  lower 8 bytes of ldata1  */
146#define data0		R6	/*  lower 8 bytes of ldata0  */
147
148#define ifbyte		p0	/*  if transfer has bytes in epilog/prolog  */
149#define ifhword		p0	/*  if transfer has shorts in epilog/prolog  */
150#define ifword		p0	/*  if transfer has words in epilog/prolog  */
151#define noprolog	p0	/*  no prolog, xfer starts at 32byte  */
152#define nokernel	p1	/*  no 32byte multiple block in the transfer  */
153#define noepilog	p0	/*  no epilog, xfer ends on 32byte boundary  */
154#define align		p2	/*  alignment of input rel to 8byte boundary  */
155#define kernel1		p0	/*  kernel count == 1  */
156
157#define dalign		R25	/*  rel alignment of input to output data  */
158#define star3		R16	/*  number bytes in prolog - dwords  */
159#define rest		R8	/*  length - prolog bytes  */
160#define back		R7	/*  nr bytes > dword boundary in src block  */
161#define epilog		R3	/*  bytes in epilog  */
162#define inc		R15:14	/*  inc kernel by -1 and defetch ptr by 32  */
163#define kernel		R4	/*  number of 32byte chunks in kernel  */
164#define ptr_in_p_128	R5	/*  pointer for prefetch of input data  */
165#define mask		R8	/*  mask used to determine prolog size  */
166#define shift		R8	/*  used to work a shifter to extract bytes  */
167#define shift2		R5	/*  in epilog to workshifter to extract bytes */
168#define prolog		R15	/*  bytes in  prolog  */
169#define epilogdws	R15	/*  number dwords in epilog  */
170#define shiftb		R14	/*  used to extract bytes  */
171#define offset		R9	/*  same as align in reg  */
172#define ptr_out_p_32	R17	/*  pointer to output dczero  */
173#define align888	R14	/*  if simple dword loop can be used  */
174#define len8		R9	/*  number of dwords in length  */
175#define over		R20	/*  nr of bytes > last inp buf dword boundary */
176
177#define ptr_in_p_128kernel	R5:4	/*  packed fetch pointer & kernel cnt */
178
179	.section .text
180	.p2align 4
181        .global memcpy
182        .type memcpy, @function
183memcpy:
184{
185	p2 = cmp.eq(len, #0);		/*  =0 */
186	align888 = or(ptr_in, ptr_out);	/*  %8 < 97 */
187	p0 = cmp.gtu(len, #23);		/*  %1, <24 */
188	p1 = cmp.eq(ptr_in, ptr_out);	/*  attempt to overwrite self */
189}
190{
191	p1 = or(p2, p1);
192	p3 = cmp.gtu(len, #95);		/*  %8 < 97 */
193	align888 = or(align888, len);	/*  %8 < 97 */
194	len8 = lsr(len, #3);		/*  %8 < 97 */
195}
196{
197	dcfetch(ptr_in);		/*  zero/ptrin=ptrout causes fetch */
198	p2 = bitsclr(align888, #7);	/*  %8 < 97  */
199	if(p1) jumpr r31;		/*  =0  */
200}
201{
202	p2 = and(p2,!p3);			/*  %8 < 97  */
203	if (p2.new) len = add(len, #-8);	/*  %8 < 97  */
204	if (p2.new) jump:NT .Ldwordaligned; 	/*  %8 < 97  */
205}
206{
207	if(!p0) jump .Lbytes23orless;	/*  %1, <24  */
208	mask.l = #LO(0x7fffffff);
209	/*  all bytes before line multiples of data  */
210	prolog = sub(#0, ptr_out);
211}
212{
213	/*  save r31 on stack, decrement sp by 16  */
214	allocframe(#24);
215	mask.h = #HI(0x7fffffff);
216	ptr_in_p_128 = add(ptr_in, #32);
217	back = cl0(len);
218}
219{
220	memd(sp+#0) = R17:16;		/*  save r16,r17 on stack6  */
221	r31.l = #LO(.Lmemcpy_return);	/*  set up final return pointer  */
222	prolog &= lsr(mask, back);
223	offset = and(ptr_in, #7);
224}
225{
226	memd(sp+#8) = R25:24;		/*  save r25,r24 on stack  */
227	dalign = sub(ptr_out, ptr_in);
228	r31.h = #HI(.Lmemcpy_return);	/*  set up final return pointer  */
229}
230{
231	/*  see if there if input buffer end if aligned  */
232	over = add(len, ptr_in);
233	back = add(len, offset);
234	memd(sp+#16) = R21:20;		/*  save r20,r21 on stack  */
235}
236{
237	noprolog = bitsclr(prolog, #7);
238	prolog = and(prolog, #31);
239	dcfetch(ptr_in_p_128);
240	ptr_in_p_128 = add(ptr_in_p_128, #32);
241}
242{
243	kernel = sub(len, prolog);
244	shift = asl(prolog, #3);
245	star3 = and(prolog, #7);
246	ptr_in = and(ptr_in, #-8);
247}
248{
249	prolog = lsr(prolog, #3);
250	epilog = and(kernel, #31);
251	ptr_out_p_32 = add(ptr_out, prolog);
252	over = and(over, #7);
253}
254{
255	p3 = cmp.gtu(back, #8);
256	kernel = lsr(kernel, #5);
257	dcfetch(ptr_in_p_128);
258	ptr_in_p_128 = add(ptr_in_p_128, #32);
259}
260{
261	p1 = cmp.eq(prolog, #0);
262	if(!p1.new) prolog = add(prolog, #1);
263	dcfetch(ptr_in_p_128);	/*  reserve the line 64bytes on  */
264	ptr_in_p_128 = add(ptr_in_p_128, #32);
265}
266{
267	nokernel = cmp.eq(kernel,#0);
268	dcfetch(ptr_in_p_128);	/* reserve the line 64bytes on  */
269	ptr_in_p_128 = add(ptr_in_p_128, #32);
270	shiftb = and(shift, #8);
271}
272{
273	dcfetch(ptr_in_p_128);		/*  reserve the line 64bytes on  */
274	ptr_in_p_128 = add(ptr_in_p_128, #32);
275	if(nokernel) jump .Lskip64;
276	p2 = cmp.eq(kernel, #1);	/*  skip ovr if kernel == 0  */
277}
278{
279	dczeroa(ptr_out_p_32);
280	/*  don't advance pointer  */
281	if(!p2) ptr_out_p_32 = add(ptr_out_p_32, #32);
282}
283{
284	dalign = and(dalign, #31);
285	dczeroa(ptr_out_p_32);
286}
287.Lskip64:
288{
289	data70 = memd(ptr_in++#16);
290	if(p3) dataF8 = memd(ptr_in+#8);
291	if(noprolog) jump .Lnoprolog32;
292	align = offset;
293}
294/*  upto initial 7 bytes  */
295{
296	ldata0 = valignb(dataF8, data70, align);
297	ifbyte = tstbit(shift,#3);
298	offset = add(offset, star3);
299}
300{
301	if(ifbyte) memb(ptr_out++#1) = data0;
302	ldata0 = lsr(ldata0, shiftb);
303	shiftb = and(shift, #16);
304	ifhword = tstbit(shift,#4);
305}
306{
307	if(ifhword) memh(ptr_out++#2) = data0;
308	ldata0 = lsr(ldata0, shiftb);
309	ifword = tstbit(shift,#5);
310	p2 = cmp.gtu(offset, #7);
311}
312{
313	if(ifword) memw(ptr_out++#4) = data0;
314	if(p2) data70 = dataF8;
315	if(p2) dataF8 = memd(ptr_in++#8);	/*  another 8 bytes  */
316	align = offset;
317}
318.Lnoprolog32:
319{
320	p3 = sp1loop0(.Ldword_loop_prolog, prolog)
321	rest = sub(len, star3);	/*  whats left after the loop  */
322	p0 = cmp.gt(over, #0);
323}
324	if(p0) rest = add(rest, #16);
325.Ldword_loop_prolog:
326{
327	if(p3) memd(ptr_out++#8) = ldata0;
328	ldata0 = valignb(dataF8, data70, align);
329	p0 = cmp.gt(rest, #16);
330}
331{
332	data70 = dataF8;
333	if(p0) dataF8 = memd(ptr_in++#8);
334	rest = add(rest, #-8);
335}:endloop0
336.Lkernel:
337{
338	/*  kernel is at least 32bytes  */
339	p3 = cmp.gtu(kernel, #0);
340	/*  last itn. remove edge effects  */
341	if(p3.new) kernel = add(kernel, #-1);
342	/*  dealt with in last dword loop  */
343	if(p3.new) epilog = add(epilog, #32);
344}
345{
346	nokernel = cmp.eq(kernel, #0);		/*  after adjustment, recheck */
347	if(nokernel.new) jump:NT .Lepilog;	/*  likely not taken  */
348	inc = combine(#32, #-1);
349	p3 = cmp.gtu(dalign, #24);
350}
351{
352	if(p3) jump .Lodd_alignment;
353}
354{
355	loop0(.Loword_loop_25to31, kernel);
356	kernel1 = cmp.gtu(kernel, #1);
357	rest = kernel;
358}
359	.falign
360.Loword_loop_25to31:
361{
362	dcfetch(ptr_in_p_128);	/*  prefetch 4 lines ahead  */
363	if(kernel1) ptr_out_p_32 = add(ptr_out_p_32, #32);
364}
365{
366	dczeroa(ptr_out_p_32);	/*  reserve the next 32bytes in cache  */
367	p3 = cmp.eq(kernel, rest);
368}
369{
370	/*  kernel -= 1  */
371	ptr_in_p_128kernel = vaddw(ptr_in_p_128kernel, inc);
372	/*  kill write on first iteration  */
373	if(!p3) memd(ptr_out++#8) = ldata1;
374	ldata1 = valignb(dataF8, data70, align);
375	data70 = memd(ptr_in++#8);
376}
377{
378	memd(ptr_out++#8) = ldata0;
379	ldata0 = valignb(data70, dataF8, align);
380	dataF8 = memd(ptr_in++#8);
381}
382{
383	memd(ptr_out++#8) = ldata1;
384	ldata1 = valignb(dataF8, data70, align);
385	data70 = memd(ptr_in++#8);
386}
387{
388	memd(ptr_out++#8) = ldata0;
389	ldata0 = valignb(data70, dataF8, align);
390	dataF8 = memd(ptr_in++#8);
391	kernel1 = cmp.gtu(kernel, #1);
392}:endloop0
393{
394	memd(ptr_out++#8) = ldata1;
395	jump .Lepilog;
396}
397.Lodd_alignment:
398{
399	loop0(.Loword_loop_00to24, kernel);
400	kernel1 = cmp.gtu(kernel, #1);
401	rest = add(kernel, #-1);
402}
403	.falign
404.Loword_loop_00to24:
405{
406	dcfetch(ptr_in_p_128);	/*  prefetch 4 lines ahead  */
407	ptr_in_p_128kernel = vaddw(ptr_in_p_128kernel, inc);
408	if(kernel1) ptr_out_p_32 = add(ptr_out_p_32, #32);
409}
410{
411	dczeroa(ptr_out_p_32);	/*  reserve the next 32bytes in cache  */
412}
413{
414	memd(ptr_out++#8) = ldata0;
415	ldata0 = valignb(dataF8, data70, align);
416	data70 = memd(ptr_in++#8);
417}
418{
419	memd(ptr_out++#8) = ldata0;
420	ldata0 = valignb(data70, dataF8, align);
421	dataF8 = memd(ptr_in++#8);
422}
423{
424	memd(ptr_out++#8) = ldata0;
425	ldata0 = valignb(dataF8, data70, align);
426	data70 = memd(ptr_in++#8);
427}
428{
429	memd(ptr_out++#8) = ldata0;
430	ldata0 = valignb(data70, dataF8, align);
431	dataF8 = memd(ptr_in++#8);
432	kernel1 = cmp.gtu(kernel, #1);
433}:endloop0
434.Lepilog:
435{
436	noepilog = cmp.eq(epilog,#0);
437	epilogdws = lsr(epilog, #3);
438	kernel = and(epilog, #7);
439}
440{
441	if(noepilog) jumpr r31;
442	if(noepilog) ptr_out = sub(ptr_out, len);
443	p3 = cmp.eq(epilogdws, #0);
444	shift2 = asl(epilog, #3);
445}
446{
447	shiftb = and(shift2, #32);
448	ifword = tstbit(epilog,#2);
449	if(p3) jump .Lepilog60;
450	if(!p3) epilog = add(epilog, #-16);
451}
452{
453	loop0(.Ldword_loop_epilog, epilogdws);
454	/*  stop criteria is lsbs unless = 0 then its 8  */
455	p3 = cmp.eq(kernel, #0);
456	if(p3.new) kernel= #8;
457	p1 = cmp.gt(over, #0);
458}
459	/*  if not aligned to end of buffer execute 1 more iteration  */
460	if(p1) kernel= #0;
461.Ldword_loop_epilog:
462{
463	memd(ptr_out++#8) = ldata0;
464	ldata0 = valignb(dataF8, data70, align);
465	p3 = cmp.gt(epilog, kernel);
466}
467{
468	data70 = dataF8;
469	if(p3) dataF8 = memd(ptr_in++#8);
470	epilog = add(epilog, #-8);
471}:endloop0
472/* copy last 7 bytes */
473.Lepilog60:
474{
475	if(ifword) memw(ptr_out++#4) = data0;
476	ldata0 = lsr(ldata0, shiftb);
477	ifhword = tstbit(epilog,#1);
478	shiftb = and(shift2, #16);
479}
480{
481	if(ifhword) memh(ptr_out++#2) = data0;
482	ldata0 = lsr(ldata0, shiftb);
483	ifbyte = tstbit(epilog,#0);
484	if(ifbyte.new) len = add(len, #-1);
485}
486{
487	if(ifbyte) memb(ptr_out) = data0;
488	ptr_out = sub(ptr_out, len);	/*  return dest pointer  */
489        jumpr r31;
490}
491/*  do byte copy for small n  */
492.Lbytes23orless:
493{
494	p3 = sp1loop0(.Lbyte_copy, len);
495	len = add(len, #-1);
496}
497.Lbyte_copy:
498{
499	data0 = memb(ptr_in++#1);
500	if(p3) memb(ptr_out++#1) = data0;
501}:endloop0
502{
503	memb(ptr_out) = data0;
504	ptr_out = sub(ptr_out, len);
505	jumpr r31;
506}
507/*  do dword copies for aligned in, out and length  */
508.Ldwordaligned:
509{
510	p3 = sp1loop0(.Ldword_copy, len8);
511}
512.Ldword_copy:
513{
514	if(p3) memd(ptr_out++#8) = ldata0;
515	ldata0 = memd(ptr_in++#8);
516}:endloop0
517{
518	memd(ptr_out) = ldata0;
519	ptr_out = sub(ptr_out, len);
520	jumpr r31;	/*  return to function caller  */
521}
522.Lmemcpy_return:
523	r21:20 = memd(sp+#16);	/*  restore r20+r21  */
524{
525	r25:24 = memd(sp+#8);	/*  restore r24+r25  */
526	r17:16 = memd(sp+#0);	/*  restore r16+r17  */
527}
528	deallocframe;	/*  restore r31 and incrment stack by 16  */
529	jumpr r31
530