1/*
2 * Copyright (C) 2015-2018 Rob Clark <robclark@freedesktop.org>
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors:
24 *    Rob Clark <robclark@freedesktop.org>
25 */
26
27#include "util/u_math.h"
28
29#include "ir3_compiler.h"
30#include "ir3_context.h"
31#include "ir3_image.h"
32#include "ir3_shader.h"
33#include "ir3_nir.h"
34
35struct ir3_context *
36ir3_context_init(struct ir3_compiler *compiler,
37		struct ir3_shader_variant *so)
38{
39	struct ir3_context *ctx = rzalloc(NULL, struct ir3_context);
40
41	if (compiler->gpu_id >= 400) {
42		if (so->type == MESA_SHADER_VERTEX) {
43			ctx->astc_srgb = so->key.vastc_srgb;
44		} else if (so->type == MESA_SHADER_FRAGMENT) {
45			ctx->astc_srgb = so->key.fastc_srgb;
46		}
47
48	} else {
49		if (so->type == MESA_SHADER_VERTEX) {
50			ctx->samples = so->key.vsamples;
51		} else if (so->type == MESA_SHADER_FRAGMENT) {
52			ctx->samples = so->key.fsamples;
53		}
54	}
55
56	if (compiler->gpu_id >= 600) {
57		ctx->funcs = &ir3_a6xx_funcs;
58	} else if (compiler->gpu_id >= 400) {
59		ctx->funcs = &ir3_a4xx_funcs;
60	}
61
62	ctx->compiler = compiler;
63	ctx->so = so;
64	ctx->def_ht = _mesa_hash_table_create(ctx,
65			_mesa_hash_pointer, _mesa_key_pointer_equal);
66	ctx->block_ht = _mesa_hash_table_create(ctx,
67			_mesa_hash_pointer, _mesa_key_pointer_equal);
68
69	/* TODO: maybe generate some sort of bitmask of what key
70	 * lowers vs what shader has (ie. no need to lower
71	 * texture clamp lowering if no texture sample instrs)..
72	 * although should be done further up the stack to avoid
73	 * creating duplicate variants..
74	 */
75
76	if (ir3_key_lowers_nir(&so->key)) {
77		nir_shader *s = nir_shader_clone(ctx, so->shader->nir);
78		ctx->s = ir3_optimize_nir(so->shader, s, &so->key);
79	} else {
80		/* fast-path for shader key that lowers nothing in NIR: */
81		ctx->s = nir_shader_clone(ctx, so->shader->nir);
82	}
83
84	/* this needs to be the last pass run, so do this here instead of
85	 * in ir3_optimize_nir():
86	 */
87	NIR_PASS_V(ctx->s, nir_lower_bool_to_int32);
88	NIR_PASS_V(ctx->s, nir_lower_locals_to_regs);
89	NIR_PASS_V(ctx->s, nir_convert_from_ssa, true);
90
91	if (ir3_shader_debug & IR3_DBG_DISASM) {
92		DBG("dump nir%dv%d: type=%d, k={cts=%u,hp=%u}",
93			so->shader->id, so->id, so->type,
94			so->key.color_two_side, so->key.half_precision);
95		nir_print_shader(ctx->s, stdout);
96	}
97
98	if (shader_debug_enabled(so->type)) {
99		fprintf(stderr, "NIR (final form) for %s shader:\n",
100			_mesa_shader_stage_to_string(so->type));
101		nir_print_shader(ctx->s, stderr);
102	}
103
104	ir3_nir_scan_driver_consts(ctx->s, &so->const_layout);
105
106	so->num_uniforms = ctx->s->num_uniforms;
107	so->num_ubos = ctx->s->info.num_ubos;
108
109	ir3_ibo_mapping_init(&so->image_mapping, ctx->s->info.num_textures);
110
111	/* Layout of constant registers, each section aligned to vec4.  Note
112	 * that pointer size (ubo, etc) changes depending on generation.
113	 *
114	 *    user consts
115	 *    UBO addresses
116	 *    SSBO sizes
117	 *    if (vertex shader) {
118	 *        driver params (IR3_DP_*)
119	 *        if (stream_output.num_outputs > 0)
120	 *           stream-out addresses
121	 *    }
122	 *    immediates
123	 *
124	 * Immediates go last mostly because they are inserted in the CP pass
125	 * after the nir -> ir3 frontend.
126	 *
127	 * Note UBO size in bytes should be aligned to vec4
128	 */
129	debug_assert((ctx->so->shader->ubo_state.size % 16) == 0);
130	unsigned constoff = align(ctx->so->shader->ubo_state.size / 16, 4);
131	unsigned ptrsz = ir3_pointer_size(ctx);
132
133	memset(&so->constbase, ~0, sizeof(so->constbase));
134
135	if (so->num_ubos > 0) {
136		so->constbase.ubo = constoff;
137		constoff += align(ctx->s->info.num_ubos * ptrsz, 4) / 4;
138	}
139
140	if (so->const_layout.ssbo_size.count > 0) {
141		unsigned cnt = so->const_layout.ssbo_size.count;
142		so->constbase.ssbo_sizes = constoff;
143		constoff += align(cnt, 4) / 4;
144	}
145
146	if (so->const_layout.image_dims.count > 0) {
147		unsigned cnt = so->const_layout.image_dims.count;
148		so->constbase.image_dims = constoff;
149		constoff += align(cnt, 4) / 4;
150	}
151
152	unsigned num_driver_params = 0;
153	if (so->type == MESA_SHADER_VERTEX) {
154		num_driver_params = IR3_DP_VS_COUNT;
155	} else if (so->type == MESA_SHADER_COMPUTE) {
156		num_driver_params = IR3_DP_CS_COUNT;
157	}
158
159	so->constbase.driver_param = constoff;
160	constoff += align(num_driver_params, 4) / 4;
161
162	if ((so->type == MESA_SHADER_VERTEX) &&
163			(compiler->gpu_id < 500) &&
164			so->shader->stream_output.num_outputs > 0) {
165		so->constbase.tfbo = constoff;
166		constoff += align(IR3_MAX_SO_BUFFERS * ptrsz, 4) / 4;
167	}
168
169	so->constbase.immediate = constoff;
170
171	return ctx;
172}
173
174void
175ir3_context_free(struct ir3_context *ctx)
176{
177	ralloc_free(ctx);
178}
179
180/*
181 * Misc helpers
182 */
183
184/* allocate a n element value array (to be populated by caller) and
185 * insert in def_ht
186 */
187struct ir3_instruction **
188ir3_get_dst_ssa(struct ir3_context *ctx, nir_ssa_def *dst, unsigned n)
189{
190	struct ir3_instruction **value =
191		ralloc_array(ctx->def_ht, struct ir3_instruction *, n);
192	_mesa_hash_table_insert(ctx->def_ht, dst, value);
193	return value;
194}
195
196struct ir3_instruction **
197ir3_get_dst(struct ir3_context *ctx, nir_dest *dst, unsigned n)
198{
199	struct ir3_instruction **value;
200
201	if (dst->is_ssa) {
202		value = ir3_get_dst_ssa(ctx, &dst->ssa, n);
203	} else {
204		value = ralloc_array(ctx, struct ir3_instruction *, n);
205	}
206
207	/* NOTE: in non-ssa case, we don't really need to store last_dst
208	 * but this helps us catch cases where put_dst() call is forgotten
209	 */
210	compile_assert(ctx, !ctx->last_dst);
211	ctx->last_dst = value;
212	ctx->last_dst_n = n;
213
214	return value;
215}
216
217struct ir3_instruction * const *
218ir3_get_src(struct ir3_context *ctx, nir_src *src)
219{
220	if (src->is_ssa) {
221		struct hash_entry *entry;
222		entry = _mesa_hash_table_search(ctx->def_ht, src->ssa);
223		compile_assert(ctx, entry);
224		return entry->data;
225	} else {
226		nir_register *reg = src->reg.reg;
227		struct ir3_array *arr = ir3_get_array(ctx, reg);
228		unsigned num_components = arr->r->num_components;
229		struct ir3_instruction *addr = NULL;
230		struct ir3_instruction **value =
231			ralloc_array(ctx, struct ir3_instruction *, num_components);
232
233		if (src->reg.indirect)
234			addr = ir3_get_addr(ctx, ir3_get_src(ctx, src->reg.indirect)[0],
235					reg->num_components);
236
237		for (unsigned i = 0; i < num_components; i++) {
238			unsigned n = src->reg.base_offset * reg->num_components + i;
239			compile_assert(ctx, n < arr->length);
240			value[i] = ir3_create_array_load(ctx, arr, n, addr);
241		}
242
243		return value;
244	}
245}
246
247void
248ir3_put_dst(struct ir3_context *ctx, nir_dest *dst)
249{
250	unsigned bit_size = nir_dest_bit_size(*dst);
251
252	/* add extra mov if dst value is HIGH reg.. in some cases not all
253	 * instructions can read from HIGH regs, in cases where they can
254	 * ir3_cp will clean up the extra mov:
255	 */
256	for (unsigned i = 0; i < ctx->last_dst_n; i++) {
257		if (!ctx->last_dst[i])
258			continue;
259		if (ctx->last_dst[i]->regs[0]->flags & IR3_REG_HIGH) {
260			ctx->last_dst[i] = ir3_MOV(ctx->block, ctx->last_dst[i], TYPE_U32);
261		}
262	}
263
264	if (bit_size < 32) {
265		for (unsigned i = 0; i < ctx->last_dst_n; i++) {
266			struct ir3_instruction *dst = ctx->last_dst[i];
267			dst->regs[0]->flags |= IR3_REG_HALF;
268			if (ctx->last_dst[i]->opc == OPC_META_FO)
269				dst->regs[1]->instr->regs[0]->flags |= IR3_REG_HALF;
270		}
271	}
272
273	if (!dst->is_ssa) {
274		nir_register *reg = dst->reg.reg;
275		struct ir3_array *arr = ir3_get_array(ctx, reg);
276		unsigned num_components = ctx->last_dst_n;
277		struct ir3_instruction *addr = NULL;
278
279		if (dst->reg.indirect)
280			addr = ir3_get_addr(ctx, ir3_get_src(ctx, dst->reg.indirect)[0],
281					reg->num_components);
282
283		for (unsigned i = 0; i < num_components; i++) {
284			unsigned n = dst->reg.base_offset * reg->num_components + i;
285			compile_assert(ctx, n < arr->length);
286			if (!ctx->last_dst[i])
287				continue;
288			ir3_create_array_store(ctx, arr, n, ctx->last_dst[i], addr);
289		}
290
291		ralloc_free(ctx->last_dst);
292	}
293
294	ctx->last_dst = NULL;
295	ctx->last_dst_n = 0;
296}
297
298struct ir3_instruction *
299ir3_create_collect(struct ir3_context *ctx, struct ir3_instruction *const *arr,
300		unsigned arrsz)
301{
302	struct ir3_block *block = ctx->block;
303	struct ir3_instruction *collect;
304
305	if (arrsz == 0)
306		return NULL;
307
308	unsigned flags = arr[0]->regs[0]->flags & IR3_REG_HALF;
309
310	collect = ir3_instr_create2(block, OPC_META_FI, 1 + arrsz);
311	ir3_reg_create(collect, 0, flags);     /* dst */
312	for (unsigned i = 0; i < arrsz; i++) {
313		struct ir3_instruction *elem = arr[i];
314
315		/* Since arrays are pre-colored in RA, we can't assume that
316		 * things will end up in the right place.  (Ie. if a collect
317		 * joins elements from two different arrays.)  So insert an
318		 * extra mov.
319		 *
320		 * We could possibly skip this if all the collected elements
321		 * are contiguous elements in a single array.. not sure how
322		 * likely that is to happen.
323		 *
324		 * Fixes a problem with glamor shaders, that in effect do
325		 * something like:
326		 *
327		 *   if (foo)
328		 *     texcoord = ..
329		 *   else
330		 *     texcoord = ..
331		 *   color = texture2D(tex, texcoord);
332		 *
333		 * In this case, texcoord will end up as nir registers (which
334		 * translate to ir3 array's of length 1.  And we can't assume
335		 * the two (or more) arrays will get allocated in consecutive
336		 * scalar registers.
337		 *
338		 */
339		if (elem->regs[0]->flags & IR3_REG_ARRAY) {
340			type_t type = (flags & IR3_REG_HALF) ? TYPE_U16 : TYPE_U32;
341			elem = ir3_MOV(block, elem, type);
342		}
343
344		compile_assert(ctx, (elem->regs[0]->flags & IR3_REG_HALF) == flags);
345		ir3_reg_create(collect, 0, IR3_REG_SSA | flags)->instr = elem;
346	}
347
348	collect->regs[0]->wrmask = MASK(arrsz);
349
350	return collect;
351}
352
353/* helper for instructions that produce multiple consecutive scalar
354 * outputs which need to have a split/fanout meta instruction inserted
355 */
356void
357ir3_split_dest(struct ir3_block *block, struct ir3_instruction **dst,
358		struct ir3_instruction *src, unsigned base, unsigned n)
359{
360	struct ir3_instruction *prev = NULL;
361
362	if ((n == 1) && (src->regs[0]->wrmask == 0x1)) {
363		dst[0] = src;
364		return;
365	}
366
367	unsigned flags = src->regs[0]->flags & (IR3_REG_HALF | IR3_REG_HIGH);
368
369	for (int i = 0, j = 0; i < n; i++) {
370		struct ir3_instruction *split = ir3_instr_create(block, OPC_META_FO);
371		ir3_reg_create(split, 0, IR3_REG_SSA | flags);
372		ir3_reg_create(split, 0, IR3_REG_SSA | flags)->instr = src;
373		split->fo.off = i + base;
374
375		if (prev) {
376			split->cp.left = prev;
377			split->cp.left_cnt++;
378			prev->cp.right = split;
379			prev->cp.right_cnt++;
380		}
381		prev = split;
382
383		if (src->regs[0]->wrmask & (1 << (i + base)))
384			dst[j++] = split;
385	}
386}
387
388NORETURN void
389ir3_context_error(struct ir3_context *ctx, const char *format, ...)
390{
391	struct hash_table *errors = NULL;
392	va_list ap;
393	va_start(ap, format);
394	if (ctx->cur_instr) {
395		errors = _mesa_hash_table_create(NULL,
396				_mesa_hash_pointer,
397				_mesa_key_pointer_equal);
398		char *msg = ralloc_vasprintf(errors, format, ap);
399		_mesa_hash_table_insert(errors, ctx->cur_instr, msg);
400	} else {
401		_debug_vprintf(format, ap);
402	}
403	va_end(ap);
404	nir_print_shader_annotated(ctx->s, stdout, errors);
405	ralloc_free(errors);
406	ctx->error = true;
407	unreachable("");
408}
409
410static struct ir3_instruction *
411create_addr(struct ir3_block *block, struct ir3_instruction *src, int align)
412{
413	struct ir3_instruction *instr, *immed;
414
415	/* TODO in at least some cases, the backend could probably be
416	 * made clever enough to propagate IR3_REG_HALF..
417	 */
418	instr = ir3_COV(block, src, TYPE_U32, TYPE_S16);
419	instr->regs[0]->flags |= IR3_REG_HALF;
420
421	switch(align){
422	case 1:
423		/* src *= 1: */
424		break;
425	case 2:
426		/* src *= 2	=> src <<= 1: */
427		immed = create_immed(block, 1);
428		immed->regs[0]->flags |= IR3_REG_HALF;
429
430		instr = ir3_SHL_B(block, instr, 0, immed, 0);
431		instr->regs[0]->flags |= IR3_REG_HALF;
432		instr->regs[1]->flags |= IR3_REG_HALF;
433		break;
434	case 3:
435		/* src *= 3: */
436		immed = create_immed(block, 3);
437		immed->regs[0]->flags |= IR3_REG_HALF;
438
439		instr = ir3_MULL_U(block, instr, 0, immed, 0);
440		instr->regs[0]->flags |= IR3_REG_HALF;
441		instr->regs[1]->flags |= IR3_REG_HALF;
442		break;
443	case 4:
444		/* src *= 4 => src <<= 2: */
445		immed = create_immed(block, 2);
446		immed->regs[0]->flags |= IR3_REG_HALF;
447
448		instr = ir3_SHL_B(block, instr, 0, immed, 0);
449		instr->regs[0]->flags |= IR3_REG_HALF;
450		instr->regs[1]->flags |= IR3_REG_HALF;
451		break;
452	default:
453		unreachable("bad align");
454		return NULL;
455	}
456
457	instr = ir3_MOV(block, instr, TYPE_S16);
458	instr->regs[0]->num = regid(REG_A0, 0);
459	instr->regs[0]->flags |= IR3_REG_HALF;
460	instr->regs[1]->flags |= IR3_REG_HALF;
461
462	return instr;
463}
464
465/* caches addr values to avoid generating multiple cov/shl/mova
466 * sequences for each use of a given NIR level src as address
467 */
468struct ir3_instruction *
469ir3_get_addr(struct ir3_context *ctx, struct ir3_instruction *src, int align)
470{
471	struct ir3_instruction *addr;
472	unsigned idx = align - 1;
473
474	compile_assert(ctx, idx < ARRAY_SIZE(ctx->addr_ht));
475
476	if (!ctx->addr_ht[idx]) {
477		ctx->addr_ht[idx] = _mesa_hash_table_create(ctx,
478				_mesa_hash_pointer, _mesa_key_pointer_equal);
479	} else {
480		struct hash_entry *entry;
481		entry = _mesa_hash_table_search(ctx->addr_ht[idx], src);
482		if (entry)
483			return entry->data;
484	}
485
486	addr = create_addr(ctx->block, src, align);
487	_mesa_hash_table_insert(ctx->addr_ht[idx], src, addr);
488
489	return addr;
490}
491
492struct ir3_instruction *
493ir3_get_predicate(struct ir3_context *ctx, struct ir3_instruction *src)
494{
495	struct ir3_block *b = ctx->block;
496	struct ir3_instruction *cond;
497
498	/* NOTE: only cmps.*.* can write p0.x: */
499	cond = ir3_CMPS_S(b, src, 0, create_immed(b, 0), 0);
500	cond->cat2.condition = IR3_COND_NE;
501
502	/* condition always goes in predicate register: */
503	cond->regs[0]->num = regid(REG_P0, 0);
504
505	return cond;
506}
507
508/*
509 * Array helpers
510 */
511
512void
513ir3_declare_array(struct ir3_context *ctx, nir_register *reg)
514{
515	struct ir3_array *arr = rzalloc(ctx, struct ir3_array);
516	arr->id = ++ctx->num_arrays;
517	/* NOTE: sometimes we get non array regs, for example for arrays of
518	 * length 1.  See fs-const-array-of-struct-of-array.shader_test.  So
519	 * treat a non-array as if it was an array of length 1.
520	 *
521	 * It would be nice if there was a nir pass to convert arrays of
522	 * length 1 to ssa.
523	 */
524	arr->length = reg->num_components * MAX2(1, reg->num_array_elems);
525	compile_assert(ctx, arr->length > 0);
526	arr->r = reg;
527	list_addtail(&arr->node, &ctx->ir->array_list);
528}
529
530struct ir3_array *
531ir3_get_array(struct ir3_context *ctx, nir_register *reg)
532{
533	list_for_each_entry (struct ir3_array, arr, &ctx->ir->array_list, node) {
534		if (arr->r == reg)
535			return arr;
536	}
537	ir3_context_error(ctx, "bogus reg: %s\n", reg->name);
538	return NULL;
539}
540
541/* relative (indirect) if address!=NULL */
542struct ir3_instruction *
543ir3_create_array_load(struct ir3_context *ctx, struct ir3_array *arr, int n,
544		struct ir3_instruction *address)
545{
546	struct ir3_block *block = ctx->block;
547	struct ir3_instruction *mov;
548	struct ir3_register *src;
549
550	mov = ir3_instr_create(block, OPC_MOV);
551	mov->cat1.src_type = TYPE_U32;
552	mov->cat1.dst_type = TYPE_U32;
553	mov->barrier_class = IR3_BARRIER_ARRAY_R;
554	mov->barrier_conflict = IR3_BARRIER_ARRAY_W;
555	ir3_reg_create(mov, 0, 0);
556	src = ir3_reg_create(mov, 0, IR3_REG_ARRAY |
557			COND(address, IR3_REG_RELATIV));
558	src->instr = arr->last_write;
559	src->size  = arr->length;
560	src->array.id = arr->id;
561	src->array.offset = n;
562
563	if (address)
564		ir3_instr_set_address(mov, address);
565
566	return mov;
567}
568
569/* relative (indirect) if address!=NULL */
570void
571ir3_create_array_store(struct ir3_context *ctx, struct ir3_array *arr, int n,
572		struct ir3_instruction *src, struct ir3_instruction *address)
573{
574	struct ir3_block *block = ctx->block;
575	struct ir3_instruction *mov;
576	struct ir3_register *dst;
577
578	/* if not relative store, don't create an extra mov, since that
579	 * ends up being difficult for cp to remove.
580	 */
581	if (!address) {
582		dst = src->regs[0];
583
584		src->barrier_class |= IR3_BARRIER_ARRAY_W;
585		src->barrier_conflict |= IR3_BARRIER_ARRAY_R | IR3_BARRIER_ARRAY_W;
586
587		dst->flags |= IR3_REG_ARRAY;
588		dst->instr = arr->last_write;
589		dst->size = arr->length;
590		dst->array.id = arr->id;
591		dst->array.offset = n;
592
593		arr->last_write = src;
594
595		array_insert(block, block->keeps, src);
596
597		return;
598	}
599
600	mov = ir3_instr_create(block, OPC_MOV);
601	mov->cat1.src_type = TYPE_U32;
602	mov->cat1.dst_type = TYPE_U32;
603	mov->barrier_class = IR3_BARRIER_ARRAY_W;
604	mov->barrier_conflict = IR3_BARRIER_ARRAY_R | IR3_BARRIER_ARRAY_W;
605	dst = ir3_reg_create(mov, 0, IR3_REG_ARRAY |
606			COND(address, IR3_REG_RELATIV));
607	dst->instr = arr->last_write;
608	dst->size  = arr->length;
609	dst->array.id = arr->id;
610	dst->array.offset = n;
611	ir3_reg_create(mov, 0, IR3_REG_SSA)->instr = src;
612
613	if (address)
614		ir3_instr_set_address(mov, address);
615
616	arr->last_write = mov;
617
618	/* the array store may only matter to something in an earlier
619	 * block (ie. loops), but since arrays are not in SSA, depth
620	 * pass won't know this.. so keep all array stores:
621	 */
622	array_insert(block, block->keeps, mov);
623}
624