projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
Merge branch 'master' of ../mesa into vulkan
[mesa.git] / src / glsl / nir / spirv_to_nir.c
1 /*
2 * Copyright © 2015 Intel Corporation
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
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Jason Ekstrand (jason@jlekstrand.net)
25 *
26 */
27
28 #include "spirv_to_nir_private.h"
29 #include "nir_vla.h"
30 #include "nir_control_flow.h"
31
32 static struct vtn_ssa_value *
33 vtn_const_ssa_value(struct vtn_builder *b, nir_constant *constant,
34 const struct glsl_type *type)
35 {
36 struct hash_entry *entry = _mesa_hash_table_search(b->const_table, constant);
37
38 if (entry)
39 return entry->data;
40
41 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
42 val->type = type;
43
44 switch (glsl_get_base_type(type)) {
45 case GLSL_TYPE_INT:
46 case GLSL_TYPE_UINT:
47 case GLSL_TYPE_BOOL:
48 case GLSL_TYPE_FLOAT:
49 case GLSL_TYPE_DOUBLE:
50 if (glsl_type_is_vector_or_scalar(type)) {
51 unsigned num_components = glsl_get_vector_elements(val->type);
52 nir_load_const_instr *load =
53 nir_load_const_instr_create(b->shader, num_components);
54
55 for (unsigned i = 0; i < num_components; i++)
56 load->value.u[i] = constant->value.u[i];
57
58 nir_instr_insert_before_cf_list(&b->impl->body, &load->instr);
59 val->def = &load->def;
60 } else {
61 assert(glsl_type_is_matrix(type));
62 unsigned rows = glsl_get_vector_elements(val->type);
63 unsigned columns = glsl_get_matrix_columns(val->type);
64 val->elems = ralloc_array(b, struct vtn_ssa_value *, columns);
65
66 for (unsigned i = 0; i < columns; i++) {
67 struct vtn_ssa_value *col_val = rzalloc(b, struct vtn_ssa_value);
68 col_val->type = glsl_get_column_type(val->type);
69 nir_load_const_instr *load =
70 nir_load_const_instr_create(b->shader, rows);
71
72 for (unsigned j = 0; j < rows; j++)
73 load->value.u[j] = constant->value.u[rows * i + j];
74
75 nir_instr_insert_before_cf_list(&b->impl->body, &load->instr);
76 col_val->def = &load->def;
77
78 val->elems[i] = col_val;
79 }
80 }
81 break;
82
83 case GLSL_TYPE_ARRAY: {
84 unsigned elems = glsl_get_length(val->type);
85 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
86 const struct glsl_type *elem_type = glsl_get_array_element(val->type);
87 for (unsigned i = 0; i < elems; i++)
88 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
89 elem_type);
90 break;
91 }
92
93 case GLSL_TYPE_STRUCT: {
94 unsigned elems = glsl_get_length(val->type);
95 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
96 for (unsigned i = 0; i < elems; i++) {
97 const struct glsl_type *elem_type =
98 glsl_get_struct_field(val->type, i);
99 val->elems[i] = vtn_const_ssa_value(b, constant->elements[i],
100 elem_type);
101 }
102 break;
103 }
104
105 default:
106 unreachable("bad constant type");
107 }
108
109 return val;
110 }
111
112 struct vtn_ssa_value *
113 vtn_ssa_value(struct vtn_builder *b, uint32_t value_id)
114 {
115 struct vtn_value *val = vtn_untyped_value(b, value_id);
116 switch (val->value_type) {
117 case vtn_value_type_constant:
118 return vtn_const_ssa_value(b, val->constant, val->const_type);
119
120 case vtn_value_type_ssa:
121 return val->ssa;
122 default:
123 unreachable("Invalid type for an SSA value");
124 }
125 }
126
127 static char *
128 vtn_string_literal(struct vtn_builder *b, const uint32_t *words,
129 unsigned word_count)
130 {
131 return ralloc_strndup(b, (char *)words, word_count * sizeof(*words));
132 }
133
134 static const uint32_t *
135 vtn_foreach_instruction(struct vtn_builder *b, const uint32_t *start,
136 const uint32_t *end, vtn_instruction_handler handler)
137 {
138 const uint32_t *w = start;
139 while (w < end) {
140 SpvOp opcode = w[0] & SpvOpCodeMask;
141 unsigned count = w[0] >> SpvWordCountShift;
142 assert(count >= 1 && w + count <= end);
143
144 if (!handler(b, opcode, w, count))
145 return w;
146
147 w += count;
148 }
149 assert(w == end);
150 return w;
151 }
152
153 static void
154 vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
155 const uint32_t *w, unsigned count)
156 {
157 switch (opcode) {
158 case SpvOpExtInstImport: {
159 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_extension);
160 if (strcmp((const char *)&w[2], "GLSL.std.450") == 0) {
161 val->ext_handler = vtn_handle_glsl450_instruction;
162 } else {
163 assert(!"Unsupported extension");
164 }
165 break;
166 }
167
168 case SpvOpExtInst: {
169 struct vtn_value *val = vtn_value(b, w[3], vtn_value_type_extension);
170 bool handled = val->ext_handler(b, w[4], w, count);
171 (void)handled;
172 assert(handled);
173 break;
174 }
175
176 default:
177 unreachable("Unhandled opcode");
178 }
179 }
180
181 static void
182 _foreach_decoration_helper(struct vtn_builder *b,
183 struct vtn_value *base_value,
184 int member,
185 struct vtn_value *value,
186 vtn_decoration_foreach_cb cb, void *data)
187 {
188 int new_member = member;
189
190 for (struct vtn_decoration *dec = value->decoration; dec; dec = dec->next) {
191 if (dec->member >= 0) {
192 assert(member == -1);
193 new_member = dec->member;
194 }
195
196 if (dec->group) {
197 assert(dec->group->value_type == vtn_value_type_decoration_group);
198 _foreach_decoration_helper(b, base_value, new_member, dec->group,
199 cb, data);
200 } else {
201 cb(b, base_value, new_member, dec, data);
202 }
203 }
204 }
205
206 /** Iterates (recursively if needed) over all of the decorations on a value
207 *
208 * This function iterates over all of the decorations applied to a given
209 * value. If it encounters a decoration group, it recurses into the group
210 * and iterates over all of those decorations as well.
211 */
212 void
213 vtn_foreach_decoration(struct vtn_builder *b, struct vtn_value *value,
214 vtn_decoration_foreach_cb cb, void *data)
215 {
216 _foreach_decoration_helper(b, value, -1, value, cb, data);
217 }
218
219 static void
220 vtn_handle_decoration(struct vtn_builder *b, SpvOp opcode,
221 const uint32_t *w, unsigned count)
222 {
223 const uint32_t *w_end = w + count;
224 const uint32_t target = w[1];
225 w += 2;
226
227 int member = -1;
228 switch (opcode) {
229 case SpvOpDecorationGroup:
230 vtn_push_value(b, target, vtn_value_type_undef);
231 break;
232
233 case SpvOpMemberDecorate:
234 member = *(w++);
235 /* fallthrough */
236 case SpvOpDecorate: {
237 struct vtn_value *val = &b->values[target];
238
239 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
240 dec->member = member;
241 dec->decoration = *(w++);
242 dec->literals = w;
243
244 /* Link into the list */
245 dec->next = val->decoration;
246 val->decoration = dec;
247 break;
248 }
249
250 case SpvOpGroupMemberDecorate:
251 member = *(w++);
252 /* fallthrough */
253 case SpvOpGroupDecorate: {
254 struct vtn_value *group = &b->values[target];
255 assert(group->value_type == vtn_value_type_decoration_group);
256
257 for (; w < w_end; w++) {
258 struct vtn_value *val = &b->values[*w];
259 struct vtn_decoration *dec = rzalloc(b, struct vtn_decoration);
260 dec->member = member;
261 dec->group = group;
262
263 /* Link into the list */
264 dec->next = val->decoration;
265 val->decoration = dec;
266 }
267 break;
268 }
269
270 default:
271 unreachable("Unhandled opcode");
272 }
273 }
274
275 struct member_decoration_ctx {
276 struct glsl_struct_field *fields;
277 struct vtn_type *type;
278 };
279
280 /* does a shallow copy of a vtn_type */
281
282 static struct vtn_type *
283 vtn_type_copy(struct vtn_builder *b, struct vtn_type *src)
284 {
285 struct vtn_type *dest = ralloc(b, struct vtn_type);
286 dest->type = src->type;
287 dest->is_builtin = src->is_builtin;
288 if (src->is_builtin)
289 dest->builtin = src->builtin;
290
291 if (!glsl_type_is_vector_or_scalar(src->type)) {
292 switch (glsl_get_base_type(src->type)) {
293 case GLSL_TYPE_ARRAY:
294 dest->array_element = src->array_element;
295 dest->stride = src->stride;
296 break;
297
298 case GLSL_TYPE_INT:
299 case GLSL_TYPE_UINT:
300 case GLSL_TYPE_BOOL:
301 case GLSL_TYPE_FLOAT:
302 case GLSL_TYPE_DOUBLE:
303 /* matrices */
304 dest->row_major = src->row_major;
305 dest->stride = src->stride;
306 break;
307
308 case GLSL_TYPE_STRUCT: {
309 unsigned elems = glsl_get_length(src->type);
310
311 dest->members = ralloc_array(b, struct vtn_type *, elems);
312 memcpy(dest->members, src->members, elems * sizeof(struct vtn_type *));
313
314 dest->offsets = ralloc_array(b, unsigned, elems);
315 memcpy(dest->offsets, src->offsets, elems * sizeof(unsigned));
316 break;
317 }
318
319 default:
320 unreachable("unhandled type");
321 }
322 }
323
324 return dest;
325 }
326
327 static void
328 struct_member_decoration_cb(struct vtn_builder *b,
329 struct vtn_value *val, int member,
330 const struct vtn_decoration *dec, void *void_ctx)
331 {
332 struct member_decoration_ctx *ctx = void_ctx;
333
334 if (member < 0)
335 return;
336
337 switch (dec->decoration) {
338 case SpvDecorationRelaxedPrecision:
339 break; /* FIXME: Do nothing with this for now. */
340 case SpvDecorationSmooth:
341 ctx->fields[member].interpolation = INTERP_QUALIFIER_SMOOTH;
342 break;
343 case SpvDecorationNoperspective:
344 ctx->fields[member].interpolation = INTERP_QUALIFIER_NOPERSPECTIVE;
345 break;
346 case SpvDecorationFlat:
347 ctx->fields[member].interpolation = INTERP_QUALIFIER_FLAT;
348 break;
349 case SpvDecorationCentroid:
350 ctx->fields[member].centroid = true;
351 break;
352 case SpvDecorationSample:
353 ctx->fields[member].sample = true;
354 break;
355 case SpvDecorationLocation:
356 ctx->fields[member].location = dec->literals[0];
357 break;
358 case SpvDecorationBuiltIn:
359 ctx->type->members[member] = vtn_type_copy(b,
360 ctx->type->members[member]);
361 ctx->type->members[member]->is_builtin = true;
362 ctx->type->members[member]->builtin = dec->literals[0];
363 ctx->type->builtin_block = true;
364 break;
365 case SpvDecorationOffset:
366 ctx->type->offsets[member] = dec->literals[0];
367 break;
368 case SpvDecorationMatrixStride:
369 ctx->type->members[member]->stride = dec->literals[0];
370 break;
371 case SpvDecorationColMajor:
372 break; /* Nothing to do here. Column-major is the default. */
373 default:
374 unreachable("Unhandled member decoration");
375 }
376 }
377
378 static void
379 type_decoration_cb(struct vtn_builder *b,
380 struct vtn_value *val, int member,
381 const struct vtn_decoration *dec, void *ctx)
382 {
383 struct vtn_type *type = val->type;
384
385 if (member != -1)
386 return;
387
388 switch (dec->decoration) {
389 case SpvDecorationArrayStride:
390 type->stride = dec->literals[0];
391 break;
392 case SpvDecorationBlock:
393 type->block = true;
394 break;
395 case SpvDecorationBufferBlock:
396 type->buffer_block = true;
397 break;
398 case SpvDecorationGLSLShared:
399 case SpvDecorationGLSLPacked:
400 /* Ignore these, since we get explicit offsets anyways */
401 break;
402
403 default:
404 unreachable("Unhandled type decoration");
405 }
406 }
407
408 static void
409 vtn_handle_type(struct vtn_builder *b, SpvOp opcode,
410 const uint32_t *w, unsigned count)
411 {
412 struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_type);
413
414 val->type = rzalloc(b, struct vtn_type);
415 val->type->is_builtin = false;
416
417 switch (opcode) {
418 case SpvOpTypeVoid:
419 val->type->type = glsl_void_type();
420 break;
421 case SpvOpTypeBool:
422 val->type->type = glsl_bool_type();
423 break;
424 case SpvOpTypeInt:
425 val->type->type = glsl_int_type();
426 break;
427 case SpvOpTypeFloat:
428 val->type->type = glsl_float_type();
429 break;
430
431 case SpvOpTypeVector: {
432 const struct glsl_type *base =
433 vtn_value(b, w[2], vtn_value_type_type)->type->type;
434 unsigned elems = w[3];
435
436 assert(glsl_type_is_scalar(base));
437 val->type->type = glsl_vector_type(glsl_get_base_type(base), elems);
438 break;
439 }
440
441 case SpvOpTypeMatrix: {
442 struct vtn_type *base =
443 vtn_value(b, w[2], vtn_value_type_type)->type;
444 unsigned columns = w[3];
445
446 assert(glsl_type_is_vector(base->type));
447 val->type->type = glsl_matrix_type(glsl_get_base_type(base->type),
448 glsl_get_vector_elements(base->type),
449 columns);
450 val->type->array_element = base;
451 val->type->row_major = false;
452 val->type->stride = 0;
453 break;
454 }
455
456 case SpvOpTypeArray: {
457 struct vtn_type *array_element =
458 vtn_value(b, w[2], vtn_value_type_type)->type;
459 val->type->type = glsl_array_type(array_element->type, w[3]);
460 val->type->array_element = array_element;
461 val->type->stride = 0;
462 break;
463 }
464
465 case SpvOpTypeStruct: {
466 unsigned num_fields = count - 2;
467 val->type->members = ralloc_array(b, struct vtn_type *, num_fields);
468 val->type->offsets = ralloc_array(b, unsigned, num_fields);
469
470 NIR_VLA(struct glsl_struct_field, fields, count);
471 for (unsigned i = 0; i < num_fields; i++) {
472 /* TODO: Handle decorators */
473 val->type->members[i] =
474 vtn_value(b, w[i + 2], vtn_value_type_type)->type;
475 fields[i].type = val->type->members[i]->type;
476 fields[i].name = ralloc_asprintf(b, "field%d", i);
477 fields[i].location = -1;
478 fields[i].interpolation = 0;
479 fields[i].centroid = 0;
480 fields[i].sample = 0;
481 fields[i].matrix_layout = 2;
482 fields[i].stream = -1;
483 }
484
485 struct member_decoration_ctx ctx = {
486 .fields = fields,
487 .type = val->type
488 };
489
490 vtn_foreach_decoration(b, val, struct_member_decoration_cb, &ctx);
491
492 const char *name = val->name ? val->name : "struct";
493
494 val->type->type = glsl_struct_type(fields, num_fields, name);
495 break;
496 }
497
498 case SpvOpTypeFunction: {
499 const struct glsl_type *return_type =
500 vtn_value(b, w[2], vtn_value_type_type)->type->type;
501 NIR_VLA(struct glsl_function_param, params, count - 3);
502 for (unsigned i = 0; i < count - 3; i++) {
503 params[i].type = vtn_value(b, w[i + 3], vtn_value_type_type)->type->type;
504
505 /* FIXME: */
506 params[i].in = true;
507 params[i].out = true;
508 }
509 val->type->type = glsl_function_type(return_type, params, count - 3);
510 break;
511 }
512
513 case SpvOpTypePointer:
514 /* FIXME: For now, we'll just do the really lame thing and return
515 * the same type. The validator should ensure that the proper number
516 * of dereferences happen
517 */
518 val->type = vtn_value(b, w[3], vtn_value_type_type)->type;
519 break;
520
521 case SpvOpTypeImage: {
522 const struct glsl_type *sampled_type =
523 vtn_value(b, w[2], vtn_value_type_type)->type->type;
524
525 assert(glsl_type_is_vector_or_scalar(sampled_type));
526
527 enum glsl_sampler_dim dim;
528 switch ((SpvDim)w[3]) {
529 case SpvDim1D: dim = GLSL_SAMPLER_DIM_1D; break;
530 case SpvDim2D: dim = GLSL_SAMPLER_DIM_2D; break;
531 case SpvDim3D: dim = GLSL_SAMPLER_DIM_3D; break;
532 case SpvDimCube: dim = GLSL_SAMPLER_DIM_CUBE; break;
533 case SpvDimRect: dim = GLSL_SAMPLER_DIM_RECT; break;
534 case SpvDimBuffer: dim = GLSL_SAMPLER_DIM_BUF; break;
535 default:
536 unreachable("Invalid SPIR-V Sampler dimension");
537 }
538
539 bool is_shadow = w[4];
540 bool is_array = w[5];
541
542 assert(w[6] == 0 && "FIXME: Handl multi-sampled textures");
543 assert(w[7] == 1 && "FIXME: Add support for non-sampled images");
544
545 val->type->type = glsl_sampler_type(dim, is_shadow, is_array,
546 glsl_get_base_type(sampled_type));
547 break;
548 }
549
550 case SpvOpTypeSampledImage:
551 val->type = vtn_value(b, w[2], vtn_value_type_type)->type;
552 break;
553
554 case SpvOpTypeRuntimeArray:
555 case SpvOpTypeOpaque:
556 case SpvOpTypeEvent:
557 case SpvOpTypeDeviceEvent:
558 case SpvOpTypeReserveId:
559 case SpvOpTypeQueue:
560 case SpvOpTypePipe:
561 default:
562 unreachable("Unhandled opcode");
563 }
564
565 vtn_foreach_decoration(b, val, type_decoration_cb, NULL);
566 }
567
568 static void
569 vtn_handle_constant(struct vtn_builder *b, SpvOp opcode,
570 const uint32_t *w, unsigned count)
571 {
572 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_constant);
573 val->const_type = vtn_value(b, w[1], vtn_value_type_type)->type->type;
574 val->constant = ralloc(b, nir_constant);
575 switch (opcode) {
576 case SpvOpConstantTrue:
577 assert(val->const_type == glsl_bool_type());
578 val->constant->value.u[0] = NIR_TRUE;
579 break;
580 case SpvOpConstantFalse:
581 assert(val->const_type == glsl_bool_type());
582 val->constant->value.u[0] = NIR_FALSE;
583 break;
584 case SpvOpConstant:
585 assert(glsl_type_is_scalar(val->const_type));
586 val->constant->value.u[0] = w[3];
587 break;
588 case SpvOpConstantComposite: {
589 unsigned elem_count = count - 3;
590 nir_constant **elems = ralloc_array(b, nir_constant *, elem_count);
591 for (unsigned i = 0; i < elem_count; i++)
592 elems[i] = vtn_value(b, w[i + 3], vtn_value_type_constant)->constant;
593
594 switch (glsl_get_base_type(val->const_type)) {
595 case GLSL_TYPE_UINT:
596 case GLSL_TYPE_INT:
597 case GLSL_TYPE_FLOAT:
598 case GLSL_TYPE_BOOL:
599 if (glsl_type_is_matrix(val->const_type)) {
600 unsigned rows = glsl_get_vector_elements(val->const_type);
601 assert(glsl_get_matrix_columns(val->const_type) == elem_count);
602 for (unsigned i = 0; i < elem_count; i++)
603 for (unsigned j = 0; j < rows; j++)
604 val->constant->value.u[rows * i + j] = elems[i]->value.u[j];
605 } else {
606 assert(glsl_type_is_vector(val->const_type));
607 assert(glsl_get_vector_elements(val->const_type) == elem_count);
608 for (unsigned i = 0; i < elem_count; i++)
609 val->constant->value.u[i] = elems[i]->value.u[0];
610 }
611 ralloc_free(elems);
612 break;
613
614 case GLSL_TYPE_STRUCT:
615 case GLSL_TYPE_ARRAY:
616 ralloc_steal(val->constant, elems);
617 val->constant->elements = elems;
618 break;
619
620 default:
621 unreachable("Unsupported type for constants");
622 }
623 break;
624 }
625
626 default:
627 unreachable("Unhandled opcode");
628 }
629 }
630
631 static void
632 vtn_get_builtin_location(SpvBuiltIn builtin, int *location,
633 nir_variable_mode *mode)
634 {
635 switch (builtin) {
636 case SpvBuiltInPosition:
637 *location = VARYING_SLOT_POS;
638 *mode = nir_var_shader_out;
639 break;
640 case SpvBuiltInPointSize:
641 *location = VARYING_SLOT_PSIZ;
642 *mode = nir_var_shader_out;
643 break;
644 case SpvBuiltInClipDistance:
645 *location = VARYING_SLOT_CLIP_DIST0; /* XXX CLIP_DIST1? */
646 *mode = nir_var_shader_in;
647 break;
648 case SpvBuiltInCullDistance:
649 /* XXX figure this out */
650 unreachable("unhandled builtin");
651 case SpvBuiltInVertexId:
652 /* Vulkan defines VertexID to be zero-based and reserves the new
653 * builtin keyword VertexIndex to indicate the non-zero-based value.
654 */
655 *location = SYSTEM_VALUE_VERTEX_ID_ZERO_BASE;
656 *mode = nir_var_system_value;
657 break;
658 case SpvBuiltInInstanceId:
659 *location = SYSTEM_VALUE_INSTANCE_ID;
660 *mode = nir_var_system_value;
661 break;
662 case SpvBuiltInPrimitiveId:
663 *location = VARYING_SLOT_PRIMITIVE_ID;
664 *mode = nir_var_shader_out;
665 break;
666 case SpvBuiltInInvocationId:
667 *location = SYSTEM_VALUE_INVOCATION_ID;
668 *mode = nir_var_system_value;
669 break;
670 case SpvBuiltInLayer:
671 *location = VARYING_SLOT_LAYER;
672 *mode = nir_var_shader_out;
673 break;
674 case SpvBuiltInTessLevelOuter:
675 case SpvBuiltInTessLevelInner:
676 case SpvBuiltInTessCoord:
677 case SpvBuiltInPatchVertices:
678 unreachable("no tessellation support");
679 case SpvBuiltInFragCoord:
680 *location = VARYING_SLOT_POS;
681 *mode = nir_var_shader_in;
682 break;
683 case SpvBuiltInPointCoord:
684 *location = VARYING_SLOT_PNTC;
685 *mode = nir_var_shader_out;
686 break;
687 case SpvBuiltInFrontFacing:
688 *location = VARYING_SLOT_FACE;
689 *mode = nir_var_shader_out;
690 break;
691 case SpvBuiltInSampleId:
692 *location = SYSTEM_VALUE_SAMPLE_ID;
693 *mode = nir_var_shader_in;
694 break;
695 case SpvBuiltInSamplePosition:
696 *location = SYSTEM_VALUE_SAMPLE_POS;
697 *mode = nir_var_shader_in;
698 break;
699 case SpvBuiltInSampleMask:
700 *location = SYSTEM_VALUE_SAMPLE_MASK_IN; /* XXX out? */
701 *mode = nir_var_shader_in;
702 break;
703 case SpvBuiltInFragColor:
704 *location = FRAG_RESULT_COLOR;
705 *mode = nir_var_shader_out;
706 break;
707 case SpvBuiltInFragDepth:
708 *location = FRAG_RESULT_DEPTH;
709 *mode = nir_var_shader_out;
710 break;
711 case SpvBuiltInHelperInvocation:
712 unreachable("unsupported builtin"); /* XXX */
713 break;
714 case SpvBuiltInNumWorkgroups:
715 case SpvBuiltInWorkgroupSize:
716 /* these are constants, need to be handled specially */
717 unreachable("unsupported builtin");
718 case SpvBuiltInWorkgroupId:
719 case SpvBuiltInLocalInvocationId:
720 case SpvBuiltInGlobalInvocationId:
721 case SpvBuiltInLocalInvocationIndex:
722 unreachable("no compute shader support");
723 default:
724 unreachable("unsupported builtin");
725 }
726 }
727
728 static void
729 var_decoration_cb(struct vtn_builder *b, struct vtn_value *val, int member,
730 const struct vtn_decoration *dec, void *void_var)
731 {
732 assert(val->value_type == vtn_value_type_deref);
733 assert(val->deref->deref.child == NULL);
734 assert(val->deref->var == void_var);
735
736 nir_variable *var = void_var;
737 switch (dec->decoration) {
738 case SpvDecorationRelaxedPrecision:
739 break; /* FIXME: Do nothing with this for now. */
740 case SpvDecorationSmooth:
741 var->data.interpolation = INTERP_QUALIFIER_SMOOTH;
742 break;
743 case SpvDecorationNoperspective:
744 var->data.interpolation = INTERP_QUALIFIER_NOPERSPECTIVE;
745 break;
746 case SpvDecorationFlat:
747 var->data.interpolation = INTERP_QUALIFIER_FLAT;
748 break;
749 case SpvDecorationCentroid:
750 var->data.centroid = true;
751 break;
752 case SpvDecorationSample:
753 var->data.sample = true;
754 break;
755 case SpvDecorationInvariant:
756 var->data.invariant = true;
757 break;
758 case SpvDecorationConstant:
759 assert(var->constant_initializer != NULL);
760 var->data.read_only = true;
761 break;
762 case SpvDecorationNonwritable:
763 var->data.read_only = true;
764 break;
765 case SpvDecorationLocation:
766 var->data.location = dec->literals[0];
767 break;
768 case SpvDecorationComponent:
769 var->data.location_frac = dec->literals[0];
770 break;
771 case SpvDecorationIndex:
772 var->data.explicit_index = true;
773 var->data.index = dec->literals[0];
774 break;
775 case SpvDecorationBinding:
776 var->data.explicit_binding = true;
777 var->data.binding = dec->literals[0];
778 break;
779 case SpvDecorationDescriptorSet:
780 var->data.descriptor_set = dec->literals[0];
781 break;
782 case SpvDecorationBuiltIn: {
783 nir_variable_mode mode;
784 vtn_get_builtin_location(dec->literals[0], &var->data.location,
785 &mode);
786 var->data.explicit_location = true;
787 var->data.mode = mode;
788 if (mode == nir_var_shader_in || mode == nir_var_system_value)
789 var->data.read_only = true;
790 b->builtins[dec->literals[0]] = var;
791 break;
792 }
793 case SpvDecorationNoStaticUse:
794 /* This can safely be ignored */
795 break;
796 case SpvDecorationRowMajor:
797 case SpvDecorationColMajor:
798 case SpvDecorationGLSLShared:
799 case SpvDecorationPatch:
800 case SpvDecorationRestrict:
801 case SpvDecorationAliased:
802 case SpvDecorationVolatile:
803 case SpvDecorationCoherent:
804 case SpvDecorationNonreadable:
805 case SpvDecorationUniform:
806 /* This is really nice but we have no use for it right now. */
807 case SpvDecorationCPacked:
808 case SpvDecorationSaturatedConversion:
809 case SpvDecorationStream:
810 case SpvDecorationOffset:
811 case SpvDecorationXfbBuffer:
812 case SpvDecorationFuncParamAttr:
813 case SpvDecorationFPRoundingMode:
814 case SpvDecorationFPFastMathMode:
815 case SpvDecorationLinkageAttributes:
816 case SpvDecorationSpecId:
817 break;
818 default:
819 unreachable("Unhandled variable decoration");
820 }
821 }
822
823 static nir_variable *
824 get_builtin_variable(struct vtn_builder *b,
825 const struct glsl_type *type,
826 SpvBuiltIn builtin)
827 {
828 nir_variable *var = b->builtins[builtin];
829
830 if (!var) {
831 var = ralloc(b->shader, nir_variable);
832 var->type = type;
833
834 nir_variable_mode mode;
835 vtn_get_builtin_location(builtin, &var->data.location, &mode);
836 var->data.explicit_location = true;
837 var->data.mode = mode;
838 var->name = ralloc_strdup(var, "builtin");
839
840 switch (mode) {
841 case nir_var_shader_in:
842 exec_list_push_tail(&b->shader->inputs, &var->node);
843 break;
844 case nir_var_shader_out:
845 exec_list_push_tail(&b->shader->outputs, &var->node);
846 break;
847 case nir_var_system_value:
848 exec_list_push_tail(&b->shader->system_values, &var->node);
849 break;
850 default:
851 unreachable("bad builtin mode");
852 }
853
854 b->builtins[builtin] = var;
855 }
856
857 return var;
858 }
859
860 static void
861 vtn_builtin_load(struct vtn_builder *b,
862 struct vtn_ssa_value *val,
863 SpvBuiltIn builtin)
864 {
865 assert(glsl_type_is_vector_or_scalar(val->type));
866
867 nir_variable *var = get_builtin_variable(b, val->type, builtin);
868
869 nir_intrinsic_instr *load =
870 nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_var);
871 nir_ssa_dest_init(&load->instr, &load->dest,
872 glsl_get_vector_elements(val->type), NULL);
873
874 load->variables[0] = nir_deref_var_create(load, var);
875 load->num_components = glsl_get_vector_elements(val->type);
876 nir_builder_instr_insert(&b->nb, &load->instr);
877 val->def = &load->dest.ssa;
878 }
879
880 static void
881 vtn_builtin_store(struct vtn_builder *b,
882 struct vtn_ssa_value *val,
883 SpvBuiltIn builtin)
884 {
885 assert(glsl_type_is_vector_or_scalar(val->type));
886
887 nir_variable *var = get_builtin_variable(b, val->type, builtin);
888
889 nir_intrinsic_instr *store =
890 nir_intrinsic_instr_create(b->shader, nir_intrinsic_store_var);
891
892 store->variables[0] = nir_deref_var_create(store, var);
893 store->num_components = glsl_get_vector_elements(val->type);
894 store->src[0] = nir_src_for_ssa(val->def);
895 nir_builder_instr_insert(&b->nb, &store->instr);
896 }
897
898 static struct vtn_ssa_value *
899 _vtn_variable_load(struct vtn_builder *b,
900 nir_deref_var *src_deref, struct vtn_type *src_type,
901 nir_deref *src_deref_tail)
902 {
903 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
904 val->type = src_deref_tail->type;
905
906 if (src_type->is_builtin) {
907 vtn_builtin_load(b, val, src_type->builtin);
908 return val;
909 }
910
911 /* The deref tail may contain a deref to select a component of a vector (in
912 * other words, it might not be an actual tail) so we have to save it away
913 * here since we overwrite it later.
914 */
915 nir_deref *old_child = src_deref_tail->child;
916
917 if (glsl_type_is_vector_or_scalar(val->type)) {
918 nir_intrinsic_instr *load =
919 nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_var);
920 load->variables[0] =
921 nir_deref_as_var(nir_copy_deref(load, &src_deref->deref));
922 load->num_components = glsl_get_vector_elements(val->type);
923 nir_ssa_dest_init(&load->instr, &load->dest, load->num_components, NULL);
924
925 nir_builder_instr_insert(&b->nb, &load->instr);
926
927 if (src_deref->var->data.mode == nir_var_uniform &&
928 glsl_get_base_type(val->type) == GLSL_TYPE_BOOL) {
929 /* Uniform boolean loads need to be fixed up since they're defined
930 * to be zero/nonzero rather than NIR_FALSE/NIR_TRUE.
931 */
932 val->def = nir_ine(&b->nb, &load->dest.ssa, nir_imm_int(&b->nb, 0));
933 } else {
934 val->def = &load->dest.ssa;
935 }
936 } else if (glsl_get_base_type(val->type) == GLSL_TYPE_ARRAY ||
937 glsl_type_is_matrix(val->type)) {
938 unsigned elems = glsl_get_length(val->type);
939 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
940
941 nir_deref_array *deref = nir_deref_array_create(b);
942 deref->deref_array_type = nir_deref_array_type_direct;
943 deref->deref.type = glsl_get_array_element(val->type);
944 src_deref_tail->child = &deref->deref;
945 for (unsigned i = 0; i < elems; i++) {
946 deref->base_offset = i;
947 val->elems[i] = _vtn_variable_load(b, src_deref,
948 src_type->array_element,
949 &deref->deref);
950 }
951 } else {
952 assert(glsl_get_base_type(val->type) == GLSL_TYPE_STRUCT);
953 unsigned elems = glsl_get_length(val->type);
954 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
955
956 nir_deref_struct *deref = nir_deref_struct_create(b, 0);
957 src_deref_tail->child = &deref->deref;
958 for (unsigned i = 0; i < elems; i++) {
959 deref->index = i;
960 deref->deref.type = glsl_get_struct_field(val->type, i);
961 val->elems[i] = _vtn_variable_load(b, src_deref,
962 src_type->members[i],
963 &deref->deref);
964 }
965 }
966
967 src_deref_tail->child = old_child;
968
969 return val;
970 }
971
972 static void
973 _vtn_variable_store(struct vtn_builder *b, struct vtn_type *dest_type,
974 nir_deref_var *dest_deref, nir_deref *dest_deref_tail,
975 struct vtn_ssa_value *src)
976 {
977 if (dest_type->is_builtin) {
978 vtn_builtin_store(b, src, dest_type->builtin);
979 return;
980 }
981
982 nir_deref *old_child = dest_deref_tail->child;
983
984 if (glsl_type_is_vector_or_scalar(src->type)) {
985 nir_intrinsic_instr *store =
986 nir_intrinsic_instr_create(b->shader, nir_intrinsic_store_var);
987 store->variables[0] =
988 nir_deref_as_var(nir_copy_deref(store, &dest_deref->deref));
989 store->num_components = glsl_get_vector_elements(src->type);
990 store->src[0] = nir_src_for_ssa(src->def);
991
992 nir_builder_instr_insert(&b->nb, &store->instr);
993 } else if (glsl_get_base_type(src->type) == GLSL_TYPE_ARRAY ||
994 glsl_type_is_matrix(src->type)) {
995 unsigned elems = glsl_get_length(src->type);
996
997 nir_deref_array *deref = nir_deref_array_create(b);
998 deref->deref_array_type = nir_deref_array_type_direct;
999 deref->deref.type = glsl_get_array_element(src->type);
1000 dest_deref_tail->child = &deref->deref;
1001 for (unsigned i = 0; i < elems; i++) {
1002 deref->base_offset = i;
1003 _vtn_variable_store(b, dest_type->array_element, dest_deref,
1004 &deref->deref, src->elems[i]);
1005 }
1006 } else {
1007 assert(glsl_get_base_type(src->type) == GLSL_TYPE_STRUCT);
1008 unsigned elems = glsl_get_length(src->type);
1009
1010 nir_deref_struct *deref = nir_deref_struct_create(b, 0);
1011 dest_deref_tail->child = &deref->deref;
1012 for (unsigned i = 0; i < elems; i++) {
1013 deref->index = i;
1014 deref->deref.type = glsl_get_struct_field(src->type, i);
1015 _vtn_variable_store(b, dest_type->members[i], dest_deref,
1016 &deref->deref, src->elems[i]);
1017 }
1018 }
1019
1020 dest_deref_tail->child = old_child;
1021 }
1022
1023 static struct vtn_ssa_value *
1024 _vtn_block_load(struct vtn_builder *b, nir_intrinsic_op op,
1025 unsigned set, nir_ssa_def *binding,
1026 unsigned offset, nir_ssa_def *indirect,
1027 struct vtn_type *type)
1028 {
1029 struct vtn_ssa_value *val = ralloc(b, struct vtn_ssa_value);
1030 val->type = type->type;
1031 val->transposed = NULL;
1032 if (glsl_type_is_vector_or_scalar(type->type)) {
1033 nir_intrinsic_instr *load = nir_intrinsic_instr_create(b->shader, op);
1034 load->num_components = glsl_get_vector_elements(type->type);
1035 load->const_index[0] = set;
1036 load->src[0] = nir_src_for_ssa(binding);
1037 load->const_index[1] = offset;
1038 if (indirect)
1039 load->src[1] = nir_src_for_ssa(indirect);
1040 nir_ssa_dest_init(&load->instr, &load->dest, load->num_components, NULL);
1041 nir_builder_instr_insert(&b->nb, &load->instr);
1042 val->def = &load->dest.ssa;
1043 } else {
1044 unsigned elems = glsl_get_length(type->type);
1045 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
1046 if (glsl_type_is_struct(type->type)) {
1047 for (unsigned i = 0; i < elems; i++) {
1048 val->elems[i] = _vtn_block_load(b, op, set, binding,
1049 offset + type->offsets[i],
1050 indirect, type->members[i]);
1051 }
1052 } else {
1053 for (unsigned i = 0; i < elems; i++) {
1054 val->elems[i] = _vtn_block_load(b, op, set, binding,
1055 offset + i * type->stride,
1056 indirect, type->array_element);
1057 }
1058 }
1059 }
1060
1061 return val;
1062 }
1063
1064 static struct vtn_ssa_value *
1065 vtn_block_load(struct vtn_builder *b, nir_deref_var *src,
1066 struct vtn_type *type, nir_deref *src_tail)
1067 {
1068 unsigned set = src->var->data.descriptor_set;
1069
1070 nir_ssa_def *binding = nir_imm_int(&b->nb, src->var->data.binding);
1071 nir_deref *deref = &src->deref;
1072
1073 /* The block variable may be an array, in which case the array index adds
1074 * an offset to the binding. Figure out that index now.
1075 */
1076
1077 if (deref->child->deref_type == nir_deref_type_array) {
1078 deref = deref->child;
1079 type = type->array_element;
1080 nir_deref_array *deref_array = nir_deref_as_array(deref);
1081 if (deref_array->deref_array_type == nir_deref_array_type_direct) {
1082 binding = nir_imm_int(&b->nb, src->var->data.binding +
1083 deref_array->base_offset);
1084 } else {
1085 binding = nir_iadd(&b->nb, binding, deref_array->indirect.ssa);
1086 }
1087 }
1088
1089 unsigned offset = 0;
1090 nir_ssa_def *indirect = NULL;
1091 while (deref != src_tail) {
1092 deref = deref->child;
1093 switch (deref->deref_type) {
1094 case nir_deref_type_array: {
1095 nir_deref_array *deref_array = nir_deref_as_array(deref);
1096 if (deref_array->deref_array_type == nir_deref_array_type_direct) {
1097 offset += type->stride * deref_array->base_offset;
1098 } else {
1099 nir_ssa_def *offset = nir_imul(&b->nb, deref_array->indirect.ssa,
1100 nir_imm_int(&b->nb, type->stride));
1101 indirect = indirect ? nir_iadd(&b->nb, indirect, offset) : offset;
1102 }
1103 type = type->array_element;
1104 break;
1105 }
1106
1107 case nir_deref_type_struct: {
1108 nir_deref_struct *deref_struct = nir_deref_as_struct(deref);
1109 offset += type->offsets[deref_struct->index];
1110 type = type->members[deref_struct->index];
1111 break;
1112 }
1113
1114 default:
1115 unreachable("unknown deref type");
1116 }
1117 }
1118
1119 /* TODO SSBO's */
1120 nir_intrinsic_op op = indirect ? nir_intrinsic_load_ubo_indirect
1121 : nir_intrinsic_load_ubo;
1122
1123 return _vtn_block_load(b, op, set, binding, offset, indirect, type);
1124 }
1125
1126 /*
1127 * Gets the NIR-level deref tail, which may have as a child an array deref
1128 * selecting which component due to OpAccessChain supporting per-component
1129 * indexing in SPIR-V.
1130 */
1131
1132 static nir_deref *
1133 get_deref_tail(nir_deref_var *deref)
1134 {
1135 nir_deref *cur = &deref->deref;
1136 while (!glsl_type_is_vector_or_scalar(cur->type) && cur->child)
1137 cur = cur->child;
1138
1139 return cur;
1140 }
1141
1142 static nir_ssa_def *vtn_vector_extract(struct vtn_builder *b,
1143 nir_ssa_def *src, unsigned index);
1144
1145 static nir_ssa_def *vtn_vector_extract_dynamic(struct vtn_builder *b,
1146 nir_ssa_def *src,
1147 nir_ssa_def *index);
1148
1149 static struct vtn_ssa_value *
1150 vtn_variable_load(struct vtn_builder *b, nir_deref_var *src,
1151 struct vtn_type *src_type)
1152 {
1153 nir_deref *src_tail = get_deref_tail(src);
1154
1155 struct vtn_ssa_value *val;
1156 if (src->var->interface_type && src->var->data.mode == nir_var_uniform)
1157 val = vtn_block_load(b, src, src_type, src_tail);
1158 else
1159 val = _vtn_variable_load(b, src, src_type, src_tail);
1160
1161 if (src_tail->child) {
1162 nir_deref_array *vec_deref = nir_deref_as_array(src_tail->child);
1163 assert(vec_deref->deref.child == NULL);
1164 val->type = vec_deref->deref.type;
1165 if (vec_deref->deref_array_type == nir_deref_array_type_direct)
1166 val->def = vtn_vector_extract(b, val->def, vec_deref->base_offset);
1167 else
1168 val->def = vtn_vector_extract_dynamic(b, val->def,
1169 vec_deref->indirect.ssa);
1170 }
1171
1172 return val;
1173 }
1174
1175 static nir_ssa_def * vtn_vector_insert(struct vtn_builder *b,
1176 nir_ssa_def *src, nir_ssa_def *insert,
1177 unsigned index);
1178
1179 static nir_ssa_def * vtn_vector_insert_dynamic(struct vtn_builder *b,
1180 nir_ssa_def *src,
1181 nir_ssa_def *insert,
1182 nir_ssa_def *index);
1183 static void
1184 vtn_variable_store(struct vtn_builder *b, struct vtn_ssa_value *src,
1185 nir_deref_var *dest, struct vtn_type *dest_type)
1186 {
1187 nir_deref *dest_tail = get_deref_tail(dest);
1188 if (dest_tail->child) {
1189 struct vtn_ssa_value *val = _vtn_variable_load(b, dest, dest_type,
1190 dest_tail);
1191 nir_deref_array *deref = nir_deref_as_array(dest_tail->child);
1192 assert(deref->deref.child == NULL);
1193 if (deref->deref_array_type == nir_deref_array_type_direct)
1194 val->def = vtn_vector_insert(b, val->def, src->def,
1195 deref->base_offset);
1196 else
1197 val->def = vtn_vector_insert_dynamic(b, val->def, src->def,
1198 deref->indirect.ssa);
1199 _vtn_variable_store(b, dest_type, dest, dest_tail, val);
1200 } else {
1201 _vtn_variable_store(b, dest_type, dest, dest_tail, src);
1202 }
1203 }
1204
1205 static void
1206 vtn_variable_copy(struct vtn_builder *b, nir_deref_var *src,
1207 nir_deref_var *dest, struct vtn_type *type)
1208 {
1209 nir_deref *src_tail = get_deref_tail(src);
1210
1211 if (src_tail->child || src->var->interface_type) {
1212 assert(get_deref_tail(dest)->child);
1213 struct vtn_ssa_value *val = vtn_variable_load(b, src, type);
1214 vtn_variable_store(b, val, dest, type);
1215 } else {
1216 nir_intrinsic_instr *copy =
1217 nir_intrinsic_instr_create(b->shader, nir_intrinsic_copy_var);
1218 copy->variables[0] = nir_deref_as_var(nir_copy_deref(copy, &dest->deref));
1219 copy->variables[1] = nir_deref_as_var(nir_copy_deref(copy, &src->deref));
1220
1221 nir_builder_instr_insert(&b->nb, &copy->instr);
1222 }
1223 }
1224
1225 static void
1226 vtn_handle_variables(struct vtn_builder *b, SpvOp opcode,
1227 const uint32_t *w, unsigned count)
1228 {
1229 switch (opcode) {
1230 case SpvOpVariable: {
1231 struct vtn_type *type =
1232 vtn_value(b, w[1], vtn_value_type_type)->type;
1233 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_deref);
1234
1235 nir_variable *var = ralloc(b->shader, nir_variable);
1236
1237 var->type = type->type;
1238 var->name = ralloc_strdup(var, val->name);
1239
1240 bool builtin_block = false;
1241 if (type->block) {
1242 var->interface_type = type->type;
1243 builtin_block = type->builtin_block;
1244 } else if (glsl_type_is_array(type->type) &&
1245 (type->array_element->block ||
1246 type->array_element->buffer_block)) {
1247 var->interface_type = type->array_element->type;
1248 builtin_block = type->array_element->builtin_block;
1249 } else {
1250 var->interface_type = NULL;
1251 }
1252
1253 switch ((SpvStorageClass)w[3]) {
1254 case SpvStorageClassUniform:
1255 case SpvStorageClassUniformConstant:
1256 var->data.mode = nir_var_uniform;
1257 var->data.read_only = true;
1258 break;
1259 case SpvStorageClassInput:
1260 var->data.mode = nir_var_shader_in;
1261 var->data.read_only = true;
1262 break;
1263 case SpvStorageClassOutput:
1264 var->data.mode = nir_var_shader_out;
1265 break;
1266 case SpvStorageClassPrivateGlobal:
1267 var->data.mode = nir_var_global;
1268 break;
1269 case SpvStorageClassFunction:
1270 var->data.mode = nir_var_local;
1271 break;
1272 case SpvStorageClassWorkgroupLocal:
1273 case SpvStorageClassWorkgroupGlobal:
1274 case SpvStorageClassGeneric:
1275 case SpvStorageClassAtomicCounter:
1276 default:
1277 unreachable("Unhandled variable storage class");
1278 }
1279
1280 if (count > 4) {
1281 assert(count == 5);
1282 var->constant_initializer =
1283 vtn_value(b, w[4], vtn_value_type_constant)->constant;
1284 }
1285
1286 val->deref = nir_deref_var_create(b, var);
1287 val->deref_type = type;
1288
1289 /* We handle decorations first because decorations might give us
1290 * location information. We use the data.explicit_location field to
1291 * note that the location provided is the "final" location. If
1292 * data.explicit_location == false, this means that it's relative to
1293 * whatever the base location is.
1294 */
1295 vtn_foreach_decoration(b, val, var_decoration_cb, var);
1296
1297 if (!var->data.explicit_location) {
1298 if (b->execution_model == SpvExecutionModelFragment &&
1299 var->data.mode == nir_var_shader_out) {
1300 var->data.location += FRAG_RESULT_DATA0;
1301 } else if (b->execution_model == SpvExecutionModelVertex &&
1302 var->data.mode == nir_var_shader_in) {
1303 var->data.location += VERT_ATTRIB_GENERIC0;
1304 } else if (var->data.mode == nir_var_shader_in ||
1305 var->data.mode == nir_var_shader_out) {
1306 var->data.location += VARYING_SLOT_VAR0;
1307 }
1308 }
1309
1310 /* If this was a uniform block, then we're not going to actually use the
1311 * variable (we're only going to use it to compute offsets), so don't
1312 * declare it in the shader.
1313 */
1314 if (var->data.mode == nir_var_uniform && var->interface_type)
1315 break;
1316
1317 /* Builtin blocks are lowered to individual variables during SPIR-V ->
1318 * NIR, so don't declare them either.
1319 */
1320 if (builtin_block)
1321 break;
1322
1323 switch (var->data.mode) {
1324 case nir_var_shader_in:
1325 exec_list_push_tail(&b->shader->inputs, &var->node);
1326 break;
1327 case nir_var_shader_out:
1328 exec_list_push_tail(&b->shader->outputs, &var->node);
1329 break;
1330 case nir_var_global:
1331 exec_list_push_tail(&b->shader->globals, &var->node);
1332 break;
1333 case nir_var_local:
1334 exec_list_push_tail(&b->impl->locals, &var->node);
1335 break;
1336 case nir_var_uniform:
1337 exec_list_push_tail(&b->shader->uniforms, &var->node);
1338 break;
1339 case nir_var_system_value:
1340 exec_list_push_tail(&b->shader->system_values, &var->node);
1341 break;
1342 }
1343 break;
1344 }
1345
1346 case SpvOpAccessChain:
1347 case SpvOpInBoundsAccessChain: {
1348 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_deref);
1349 nir_deref_var *base = vtn_value(b, w[3], vtn_value_type_deref)->deref;
1350 val->deref = nir_deref_as_var(nir_copy_deref(b, &base->deref));
1351 struct vtn_type *deref_type = vtn_value(b, w[3], vtn_value_type_deref)->deref_type;
1352
1353 nir_deref *tail = &val->deref->deref;
1354 while (tail->child)
1355 tail = tail->child;
1356
1357 for (unsigned i = 0; i < count - 4; i++) {
1358 assert(w[i + 4] < b->value_id_bound);
1359 struct vtn_value *idx_val = &b->values[w[i + 4]];
1360
1361 enum glsl_base_type base_type = glsl_get_base_type(tail->type);
1362 switch (base_type) {
1363 case GLSL_TYPE_UINT:
1364 case GLSL_TYPE_INT:
1365 case GLSL_TYPE_FLOAT:
1366 case GLSL_TYPE_DOUBLE:
1367 case GLSL_TYPE_BOOL:
1368 case GLSL_TYPE_ARRAY: {
1369 nir_deref_array *deref_arr = nir_deref_array_create(b);
1370 if (base_type == GLSL_TYPE_ARRAY ||
1371 glsl_type_is_matrix(tail->type)) {
1372 deref_type = deref_type->array_element;
1373 } else {
1374 assert(glsl_type_is_vector(tail->type));
1375 deref_type = ralloc(b, struct vtn_type);
1376 deref_type->type = glsl_scalar_type(base_type);
1377 }
1378
1379 deref_arr->deref.type = deref_type->type;
1380
1381 if (idx_val->value_type == vtn_value_type_constant) {
1382 unsigned idx = idx_val->constant->value.u[0];
1383 deref_arr->deref_array_type = nir_deref_array_type_direct;
1384 deref_arr->base_offset = idx;
1385 } else {
1386 assert(idx_val->value_type == vtn_value_type_ssa);
1387 deref_arr->deref_array_type = nir_deref_array_type_indirect;
1388 deref_arr->base_offset = 0;
1389 deref_arr->indirect =
1390 nir_src_for_ssa(vtn_ssa_value(b, w[1])->def);
1391 }
1392 tail->child = &deref_arr->deref;
1393 break;
1394 }
1395
1396 case GLSL_TYPE_STRUCT: {
1397 assert(idx_val->value_type == vtn_value_type_constant);
1398 unsigned idx = idx_val->constant->value.u[0];
1399 deref_type = deref_type->members[idx];
1400 nir_deref_struct *deref_struct = nir_deref_struct_create(b, idx);
1401 deref_struct->deref.type = deref_type->type;
1402 tail->child = &deref_struct->deref;
1403 break;
1404 }
1405 default:
1406 unreachable("Invalid type for deref");
1407 }
1408 tail = tail->child;
1409 }
1410
1411 /* For uniform blocks, we don't resolve the access chain until we
1412 * actually access the variable, so we need to keep around the original
1413 * type of the variable.
1414 */
1415 if (base->var->interface_type && base->var->data.mode == nir_var_uniform)
1416 val->deref_type = vtn_value(b, w[3], vtn_value_type_deref)->deref_type;
1417 else
1418 val->deref_type = deref_type;
1419
1420
1421 break;
1422 }
1423
1424 case SpvOpCopyMemory: {
1425 nir_deref_var *dest = vtn_value(b, w[1], vtn_value_type_deref)->deref;
1426 nir_deref_var *src = vtn_value(b, w[2], vtn_value_type_deref)->deref;
1427 struct vtn_type *type =
1428 vtn_value(b, w[1], vtn_value_type_deref)->deref_type;
1429
1430 vtn_variable_copy(b, src, dest, type);
1431 break;
1432 }
1433
1434 case SpvOpLoad: {
1435 nir_deref_var *src = vtn_value(b, w[3], vtn_value_type_deref)->deref;
1436 struct vtn_type *src_type =
1437 vtn_value(b, w[3], vtn_value_type_deref)->deref_type;
1438
1439 if (glsl_get_base_type(src_type->type) == GLSL_TYPE_SAMPLER) {
1440 vtn_push_value(b, w[2], vtn_value_type_deref)->deref = src;
1441 return;
1442 }
1443
1444 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1445 val->ssa = vtn_variable_load(b, src, src_type);
1446 break;
1447 }
1448
1449 case SpvOpStore: {
1450 nir_deref_var *dest = vtn_value(b, w[1], vtn_value_type_deref)->deref;
1451 struct vtn_type *dest_type =
1452 vtn_value(b, w[1], vtn_value_type_deref)->deref_type;
1453 struct vtn_ssa_value *src = vtn_ssa_value(b, w[2]);
1454 vtn_variable_store(b, src, dest, dest_type);
1455 break;
1456 }
1457
1458 case SpvOpCopyMemorySized:
1459 case SpvOpArrayLength:
1460 case SpvOpImageTexelPointer:
1461 default:
1462 unreachable("Unhandled opcode");
1463 }
1464 }
1465
1466 static void
1467 vtn_handle_function_call(struct vtn_builder *b, SpvOp opcode,
1468 const uint32_t *w, unsigned count)
1469 {
1470 unreachable("Unhandled opcode");
1471 }
1472
1473 static struct vtn_ssa_value *
1474 vtn_create_ssa_value(struct vtn_builder *b, const struct glsl_type *type)
1475 {
1476 struct vtn_ssa_value *val = rzalloc(b, struct vtn_ssa_value);
1477 val->type = type;
1478
1479 if (!glsl_type_is_vector_or_scalar(type)) {
1480 unsigned elems = glsl_get_length(type);
1481 val->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
1482 for (unsigned i = 0; i < elems; i++) {
1483 const struct glsl_type *child_type;
1484
1485 switch (glsl_get_base_type(type)) {
1486 case GLSL_TYPE_INT:
1487 case GLSL_TYPE_UINT:
1488 case GLSL_TYPE_BOOL:
1489 case GLSL_TYPE_FLOAT:
1490 case GLSL_TYPE_DOUBLE:
1491 child_type = glsl_get_column_type(type);
1492 break;
1493 case GLSL_TYPE_ARRAY:
1494 child_type = glsl_get_array_element(type);
1495 break;
1496 case GLSL_TYPE_STRUCT:
1497 child_type = glsl_get_struct_field(type, i);
1498 break;
1499 default:
1500 unreachable("unkown base type");
1501 }
1502
1503 val->elems[i] = vtn_create_ssa_value(b, child_type);
1504 }
1505 }
1506
1507 return val;
1508 }
1509
1510 static nir_tex_src
1511 vtn_tex_src(struct vtn_builder *b, unsigned index, nir_tex_src_type type)
1512 {
1513 nir_tex_src src;
1514 src.src = nir_src_for_ssa(vtn_value(b, index, vtn_value_type_ssa)->ssa->def);
1515 src.src_type = type;
1516 return src;
1517 }
1518
1519 static void
1520 vtn_handle_texture(struct vtn_builder *b, SpvOp opcode,
1521 const uint32_t *w, unsigned count)
1522 {
1523 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1524 nir_deref_var *sampler = vtn_value(b, w[3], vtn_value_type_deref)->deref;
1525
1526 nir_tex_src srcs[8]; /* 8 should be enough */
1527 nir_tex_src *p = srcs;
1528
1529 unsigned idx = 4;
1530
1531 unsigned coord_components = 0;
1532 switch (opcode) {
1533 case SpvOpImageSampleImplicitLod:
1534 case SpvOpImageSampleExplicitLod:
1535 case SpvOpImageSampleDrefImplicitLod:
1536 case SpvOpImageSampleDrefExplicitLod:
1537 case SpvOpImageSampleProjImplicitLod:
1538 case SpvOpImageSampleProjExplicitLod:
1539 case SpvOpImageSampleProjDrefImplicitLod:
1540 case SpvOpImageSampleProjDrefExplicitLod:
1541 case SpvOpImageFetch:
1542 case SpvOpImageGather:
1543 case SpvOpImageDrefGather:
1544 case SpvOpImageQueryLod: {
1545 /* All these types have the coordinate as their first real argument */
1546 struct vtn_ssa_value *coord = vtn_ssa_value(b, w[idx++]);
1547 coord_components = glsl_get_vector_elements(coord->type);
1548 p->src = nir_src_for_ssa(coord->def);
1549 p->src_type = nir_tex_src_coord;
1550 p++;
1551 break;
1552 }
1553
1554 default:
1555 break;
1556 }
1557
1558 /* These all have an explicit depth value as their next source */
1559 switch (opcode) {
1560 case SpvOpImageSampleDrefImplicitLod:
1561 case SpvOpImageSampleDrefExplicitLod:
1562 case SpvOpImageSampleProjDrefImplicitLod:
1563 case SpvOpImageSampleProjDrefExplicitLod:
1564 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_comparitor);
1565 break;
1566 default:
1567 break;
1568 }
1569
1570 /* Figure out the base texture operation */
1571 nir_texop texop;
1572 switch (opcode) {
1573 case SpvOpImageSampleImplicitLod:
1574 case SpvOpImageSampleExplicitLod:
1575 case SpvOpImageSampleDrefImplicitLod:
1576 case SpvOpImageSampleDrefExplicitLod:
1577 case SpvOpImageSampleProjImplicitLod:
1578 case SpvOpImageSampleProjExplicitLod:
1579 case SpvOpImageSampleProjDrefImplicitLod:
1580 case SpvOpImageSampleProjDrefExplicitLod:
1581 texop = nir_texop_tex;
1582 break;
1583
1584 case SpvOpImageFetch:
1585 texop = nir_texop_txf;
1586 break;
1587
1588 case SpvOpImageGather:
1589 case SpvOpImageDrefGather:
1590 texop = nir_texop_tg4;
1591 break;
1592
1593 case SpvOpImageQuerySizeLod:
1594 case SpvOpImageQuerySize:
1595 texop = nir_texop_txs;
1596 break;
1597
1598 case SpvOpImageQueryLod:
1599 texop = nir_texop_lod;
1600 break;
1601
1602 case SpvOpImageQueryLevels:
1603 texop = nir_texop_query_levels;
1604 break;
1605
1606 case SpvOpImageQuerySamples:
1607 default:
1608 unreachable("Unhandled opcode");
1609 }
1610
1611 /* Now we need to handle some number of optional arguments */
1612 if (idx < count) {
1613 uint32_t operands = w[idx++];
1614
1615 if (operands & SpvImageOperandsBiasMask) {
1616 assert(texop == nir_texop_tex);
1617 texop = nir_texop_txb;
1618 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_bias);
1619 }
1620
1621 if (operands & SpvImageOperandsLodMask) {
1622 assert(texop == nir_texop_tex);
1623 texop = nir_texop_txl;
1624 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_lod);
1625 }
1626
1627 if (operands & SpvImageOperandsGradMask) {
1628 assert(texop == nir_texop_tex);
1629 texop = nir_texop_txd;
1630 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddx);
1631 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ddy);
1632 }
1633
1634 if (operands & SpvImageOperandsOffsetMask ||
1635 operands & SpvImageOperandsConstOffsetMask)
1636 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_offset);
1637
1638 if (operands & SpvImageOperandsConstOffsetsMask)
1639 assert(!"Constant offsets to texture gather not yet implemented");
1640
1641 if (operands & SpvImageOperandsSampleMask) {
1642 assert(texop == nir_texop_txf);
1643 texop = nir_texop_txf_ms;
1644 (*p++) = vtn_tex_src(b, w[idx++], nir_tex_src_ms_index);
1645 }
1646 }
1647 /* We should have now consumed exactly all of the arguments */
1648 assert(idx == count);
1649
1650 nir_tex_instr *instr = nir_tex_instr_create(b->shader, p - srcs);
1651
1652 const struct glsl_type *sampler_type = nir_deref_tail(&sampler->deref)->type;
1653 instr->sampler_dim = glsl_get_sampler_dim(sampler_type);
1654
1655 switch (glsl_get_sampler_result_type(sampler_type)) {
1656 case GLSL_TYPE_FLOAT: instr->dest_type = nir_type_float; break;
1657 case GLSL_TYPE_INT: instr->dest_type = nir_type_int; break;
1658 case GLSL_TYPE_UINT: instr->dest_type = nir_type_unsigned; break;
1659 case GLSL_TYPE_BOOL: instr->dest_type = nir_type_bool; break;
1660 default:
1661 unreachable("Invalid base type for sampler result");
1662 }
1663
1664 instr->op = texop;
1665 memcpy(instr->src, srcs, instr->num_srcs * sizeof(*instr->src));
1666 instr->coord_components = coord_components;
1667 instr->is_array = glsl_sampler_type_is_array(sampler_type);
1668 instr->is_shadow = glsl_sampler_type_is_shadow(sampler_type);
1669
1670 instr->sampler = nir_deref_as_var(nir_copy_deref(instr, &sampler->deref));
1671
1672 nir_ssa_dest_init(&instr->instr, &instr->dest, 4, NULL);
1673 val->ssa = vtn_create_ssa_value(b, glsl_vector_type(GLSL_TYPE_FLOAT, 4));
1674 val->ssa->def = &instr->dest.ssa;
1675
1676 nir_builder_instr_insert(&b->nb, &instr->instr);
1677 }
1678
1679
1680 static nir_alu_instr *
1681 create_vec(void *mem_ctx, unsigned num_components)
1682 {
1683 nir_op op;
1684 switch (num_components) {
1685 case 1: op = nir_op_fmov; break;
1686 case 2: op = nir_op_vec2; break;
1687 case 3: op = nir_op_vec3; break;
1688 case 4: op = nir_op_vec4; break;
1689 default: unreachable("bad vector size");
1690 }
1691
1692 nir_alu_instr *vec = nir_alu_instr_create(mem_ctx, op);
1693 nir_ssa_dest_init(&vec->instr, &vec->dest.dest, num_components, NULL);
1694 vec->dest.write_mask = (1 << num_components) - 1;
1695
1696 return vec;
1697 }
1698
1699 static struct vtn_ssa_value *
1700 vtn_transpose(struct vtn_builder *b, struct vtn_ssa_value *src)
1701 {
1702 if (src->transposed)
1703 return src->transposed;
1704
1705 struct vtn_ssa_value *dest =
1706 vtn_create_ssa_value(b, glsl_transposed_type(src->type));
1707
1708 for (unsigned i = 0; i < glsl_get_matrix_columns(dest->type); i++) {
1709 nir_alu_instr *vec = create_vec(b, glsl_get_matrix_columns(src->type));
1710 if (glsl_type_is_vector_or_scalar(src->type)) {
1711 vec->src[0].src = nir_src_for_ssa(src->def);
1712 vec->src[0].swizzle[0] = i;
1713 } else {
1714 for (unsigned j = 0; j < glsl_get_matrix_columns(src->type); j++) {
1715 vec->src[j].src = nir_src_for_ssa(src->elems[j]->def);
1716 vec->src[j].swizzle[0] = i;
1717 }
1718 }
1719 nir_builder_instr_insert(&b->nb, &vec->instr);
1720 dest->elems[i]->def = &vec->dest.dest.ssa;
1721 }
1722
1723 dest->transposed = src;
1724
1725 return dest;
1726 }
1727
1728 /*
1729 * Normally, column vectors in SPIR-V correspond to a single NIR SSA
1730 * definition. But for matrix multiplies, we want to do one routine for
1731 * multiplying a matrix by a matrix and then pretend that vectors are matrices
1732 * with one column. So we "wrap" these things, and unwrap the result before we
1733 * send it off.
1734 */
1735
1736 static struct vtn_ssa_value *
1737 vtn_wrap_matrix(struct vtn_builder *b, struct vtn_ssa_value *val)
1738 {
1739 if (val == NULL)
1740 return NULL;
1741
1742 if (glsl_type_is_matrix(val->type))
1743 return val;
1744
1745 struct vtn_ssa_value *dest = rzalloc(b, struct vtn_ssa_value);
1746 dest->type = val->type;
1747 dest->elems = ralloc_array(b, struct vtn_ssa_value *, 1);
1748 dest->elems[0] = val;
1749
1750 return dest;
1751 }
1752
1753 static struct vtn_ssa_value *
1754 vtn_unwrap_matrix(struct vtn_ssa_value *val)
1755 {
1756 if (glsl_type_is_matrix(val->type))
1757 return val;
1758
1759 return val->elems[0];
1760 }
1761
1762 static struct vtn_ssa_value *
1763 vtn_matrix_multiply(struct vtn_builder *b,
1764 struct vtn_ssa_value *_src0, struct vtn_ssa_value *_src1)
1765 {
1766
1767 struct vtn_ssa_value *src0 = vtn_wrap_matrix(b, _src0);
1768 struct vtn_ssa_value *src1 = vtn_wrap_matrix(b, _src1);
1769 struct vtn_ssa_value *src0_transpose = vtn_wrap_matrix(b, _src0->transposed);
1770 struct vtn_ssa_value *src1_transpose = vtn_wrap_matrix(b, _src1->transposed);
1771
1772 unsigned src0_rows = glsl_get_vector_elements(src0->type);
1773 unsigned src0_columns = glsl_get_matrix_columns(src0->type);
1774 unsigned src1_columns = glsl_get_matrix_columns(src1->type);
1775
1776 struct vtn_ssa_value *dest =
1777 vtn_create_ssa_value(b, glsl_matrix_type(glsl_get_base_type(src0->type),
1778 src0_rows, src1_columns));
1779
1780 dest = vtn_wrap_matrix(b, dest);
1781
1782 bool transpose_result = false;
1783 if (src0_transpose && src1_transpose) {
1784 /* transpose(A) * transpose(B) = transpose(B * A) */
1785 src1 = src0_transpose;
1786 src0 = src1_transpose;
1787 src0_transpose = NULL;
1788 src1_transpose = NULL;
1789 transpose_result = true;
1790 }
1791
1792 if (src0_transpose && !src1_transpose &&
1793 glsl_get_base_type(src0->type) == GLSL_TYPE_FLOAT) {
1794 /* We already have the rows of src0 and the columns of src1 available,
1795 * so we can just take the dot product of each row with each column to
1796 * get the result.
1797 */
1798
1799 for (unsigned i = 0; i < src1_columns; i++) {
1800 nir_alu_instr *vec = create_vec(b, src0_rows);
1801 for (unsigned j = 0; j < src0_rows; j++) {
1802 vec->src[j].src =
1803 nir_src_for_ssa(nir_fdot(&b->nb, src0_transpose->elems[j]->def,
1804 src1->elems[i]->def));
1805 }
1806
1807 nir_builder_instr_insert(&b->nb, &vec->instr);
1808 dest->elems[i]->def = &vec->dest.dest.ssa;
1809 }
1810 } else {
1811 /* We don't handle the case where src1 is transposed but not src0, since
1812 * the general case only uses individual components of src1 so the
1813 * optimizer should chew through the transpose we emitted for src1.
1814 */
1815
1816 for (unsigned i = 0; i < src1_columns; i++) {
1817 /* dest[i] = sum(src0[j] * src1[i][j] for all j) */
1818 dest->elems[i]->def =
1819 nir_fmul(&b->nb, src0->elems[0]->def,
1820 vtn_vector_extract(b, src1->elems[i]->def, 0));
1821 for (unsigned j = 1; j < src0_columns; j++) {
1822 dest->elems[i]->def =
1823 nir_fadd(&b->nb, dest->elems[i]->def,
1824 nir_fmul(&b->nb, src0->elems[j]->def,
1825 vtn_vector_extract(b,
1826 src1->elems[i]->def, j)));
1827 }
1828 }
1829 }
1830
1831 dest = vtn_unwrap_matrix(dest);
1832
1833 if (transpose_result)
1834 dest = vtn_transpose(b, dest);
1835
1836 return dest;
1837 }
1838
1839 static struct vtn_ssa_value *
1840 vtn_mat_times_scalar(struct vtn_builder *b,
1841 struct vtn_ssa_value *mat,
1842 nir_ssa_def *scalar)
1843 {
1844 struct vtn_ssa_value *dest = vtn_create_ssa_value(b, mat->type);
1845 for (unsigned i = 0; i < glsl_get_matrix_columns(mat->type); i++) {
1846 if (glsl_get_base_type(mat->type) == GLSL_TYPE_FLOAT)
1847 dest->elems[i]->def = nir_fmul(&b->nb, mat->elems[i]->def, scalar);
1848 else
1849 dest->elems[i]->def = nir_imul(&b->nb, mat->elems[i]->def, scalar);
1850 }
1851
1852 return dest;
1853 }
1854
1855 static void
1856 vtn_handle_matrix_alu(struct vtn_builder *b, SpvOp opcode,
1857 const uint32_t *w, unsigned count)
1858 {
1859 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1860
1861 switch (opcode) {
1862 case SpvOpTranspose: {
1863 struct vtn_ssa_value *src = vtn_ssa_value(b, w[3]);
1864 val->ssa = vtn_transpose(b, src);
1865 break;
1866 }
1867
1868 case SpvOpOuterProduct: {
1869 struct vtn_ssa_value *src0 = vtn_ssa_value(b, w[3]);
1870 struct vtn_ssa_value *src1 = vtn_ssa_value(b, w[4]);
1871
1872 val->ssa = vtn_matrix_multiply(b, src0, vtn_transpose(b, src1));
1873 break;
1874 }
1875
1876 case SpvOpMatrixTimesScalar: {
1877 struct vtn_ssa_value *mat = vtn_ssa_value(b, w[3]);
1878 struct vtn_ssa_value *scalar = vtn_ssa_value(b, w[4]);
1879
1880 if (mat->transposed) {
1881 val->ssa = vtn_transpose(b, vtn_mat_times_scalar(b, mat->transposed,
1882 scalar->def));
1883 } else {
1884 val->ssa = vtn_mat_times_scalar(b, mat, scalar->def);
1885 }
1886 break;
1887 }
1888
1889 case SpvOpVectorTimesMatrix:
1890 case SpvOpMatrixTimesVector:
1891 case SpvOpMatrixTimesMatrix: {
1892 struct vtn_ssa_value *src0 = vtn_ssa_value(b, w[3]);
1893 struct vtn_ssa_value *src1 = vtn_ssa_value(b, w[4]);
1894
1895 val->ssa = vtn_matrix_multiply(b, src0, src1);
1896 break;
1897 }
1898
1899 default: unreachable("unknown matrix opcode");
1900 }
1901 }
1902
1903 static void
1904 vtn_handle_alu(struct vtn_builder *b, SpvOp opcode,
1905 const uint32_t *w, unsigned count)
1906 {
1907 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
1908 const struct glsl_type *type =
1909 vtn_value(b, w[1], vtn_value_type_type)->type->type;
1910 val->ssa = vtn_create_ssa_value(b, type);
1911
1912 /* Collect the various SSA sources */
1913 unsigned num_inputs = count - 3;
1914 nir_ssa_def *src[4];
1915 for (unsigned i = 0; i < num_inputs; i++)
1916 src[i] = vtn_ssa_value(b, w[i + 3])->def;
1917
1918 /* Indicates that the first two arguments should be swapped. This is
1919 * used for implementing greater-than and less-than-or-equal.
1920 */
1921 bool swap = false;
1922
1923 nir_op op;
1924 switch (opcode) {
1925 /* Basic ALU operations */
1926 case SpvOpSNegate: op = nir_op_ineg; break;
1927 case SpvOpFNegate: op = nir_op_fneg; break;
1928 case SpvOpNot: op = nir_op_inot; break;
1929
1930 case SpvOpAny:
1931 switch (src[0]->num_components) {
1932 case 1: op = nir_op_imov; break;
1933 case 2: op = nir_op_bany2; break;
1934 case 3: op = nir_op_bany3; break;
1935 case 4: op = nir_op_bany4; break;
1936 }
1937 break;
1938
1939 case SpvOpAll:
1940 switch (src[0]->num_components) {
1941 case 1: op = nir_op_imov; break;
1942 case 2: op = nir_op_ball2; break;
1943 case 3: op = nir_op_ball3; break;
1944 case 4: op = nir_op_ball4; break;
1945 }
1946 break;
1947
1948 case SpvOpIAdd: op = nir_op_iadd; break;
1949 case SpvOpFAdd: op = nir_op_fadd; break;
1950 case SpvOpISub: op = nir_op_isub; break;
1951 case SpvOpFSub: op = nir_op_fsub; break;
1952 case SpvOpIMul: op = nir_op_imul; break;
1953 case SpvOpFMul: op = nir_op_fmul; break;
1954 case SpvOpUDiv: op = nir_op_udiv; break;
1955 case SpvOpSDiv: op = nir_op_idiv; break;
1956 case SpvOpFDiv: op = nir_op_fdiv; break;
1957 case SpvOpUMod: op = nir_op_umod; break;
1958 case SpvOpSMod: op = nir_op_umod; break; /* FIXME? */
1959 case SpvOpFMod: op = nir_op_fmod; break;
1960
1961 case SpvOpDot:
1962 assert(src[0]->num_components == src[1]->num_components);
1963 switch (src[0]->num_components) {
1964 case 1: op = nir_op_fmul; break;
1965 case 2: op = nir_op_fdot2; break;
1966 case 3: op = nir_op_fdot3; break;
1967 case 4: op = nir_op_fdot4; break;
1968 }
1969 break;
1970
1971 case SpvOpShiftRightLogical: op = nir_op_ushr; break;
1972 case SpvOpShiftRightArithmetic: op = nir_op_ishr; break;
1973 case SpvOpShiftLeftLogical: op = nir_op_ishl; break;
1974 case SpvOpLogicalOr: op = nir_op_ior; break;
1975 case SpvOpLogicalEqual: op = nir_op_ieq; break;
1976 case SpvOpLogicalNotEqual: op = nir_op_ine; break;
1977 case SpvOpLogicalAnd: op = nir_op_iand; break;
1978 case SpvOpBitwiseOr: op = nir_op_ior; break;
1979 case SpvOpBitwiseXor: op = nir_op_ixor; break;
1980 case SpvOpBitwiseAnd: op = nir_op_iand; break;
1981 case SpvOpSelect: op = nir_op_bcsel; break;
1982 case SpvOpIEqual: op = nir_op_ieq; break;
1983
1984 /* Comparisons: (TODO: How do we want to handled ordered/unordered?) */
1985 case SpvOpFOrdEqual: op = nir_op_feq; break;
1986 case SpvOpFUnordEqual: op = nir_op_feq; break;
1987 case SpvOpINotEqual: op = nir_op_ine; break;
1988 case SpvOpFOrdNotEqual: op = nir_op_fne; break;
1989 case SpvOpFUnordNotEqual: op = nir_op_fne; break;
1990 case SpvOpULessThan: op = nir_op_ult; break;
1991 case SpvOpSLessThan: op = nir_op_ilt; break;
1992 case SpvOpFOrdLessThan: op = nir_op_flt; break;
1993 case SpvOpFUnordLessThan: op = nir_op_flt; break;
1994 case SpvOpUGreaterThan: op = nir_op_ult; swap = true; break;
1995 case SpvOpSGreaterThan: op = nir_op_ilt; swap = true; break;
1996 case SpvOpFOrdGreaterThan: op = nir_op_flt; swap = true; break;
1997 case SpvOpFUnordGreaterThan: op = nir_op_flt; swap = true; break;
1998 case SpvOpULessThanEqual: op = nir_op_uge; swap = true; break;
1999 case SpvOpSLessThanEqual: op = nir_op_ige; swap = true; break;
2000 case SpvOpFOrdLessThanEqual: op = nir_op_fge; swap = true; break;
2001 case SpvOpFUnordLessThanEqual: op = nir_op_fge; swap = true; break;
2002 case SpvOpUGreaterThanEqual: op = nir_op_uge; break;
2003 case SpvOpSGreaterThanEqual: op = nir_op_ige; break;
2004 case SpvOpFOrdGreaterThanEqual: op = nir_op_fge; break;
2005 case SpvOpFUnordGreaterThanEqual:op = nir_op_fge; break;
2006
2007 /* Conversions: */
2008 case SpvOpConvertFToU: op = nir_op_f2u; break;
2009 case SpvOpConvertFToS: op = nir_op_f2i; break;
2010 case SpvOpConvertSToF: op = nir_op_i2f; break;
2011 case SpvOpConvertUToF: op = nir_op_u2f; break;
2012 case SpvOpBitcast: op = nir_op_imov; break;
2013 case SpvOpUConvert:
2014 case SpvOpSConvert:
2015 op = nir_op_imov; /* TODO: NIR is 32-bit only; these are no-ops. */
2016 break;
2017 case SpvOpFConvert:
2018 op = nir_op_fmov;
2019 break;
2020
2021 /* Derivatives: */
2022 case SpvOpDPdx: op = nir_op_fddx; break;
2023 case SpvOpDPdy: op = nir_op_fddy; break;
2024 case SpvOpDPdxFine: op = nir_op_fddx_fine; break;
2025 case SpvOpDPdyFine: op = nir_op_fddy_fine; break;
2026 case SpvOpDPdxCoarse: op = nir_op_fddx_coarse; break;
2027 case SpvOpDPdyCoarse: op = nir_op_fddy_coarse; break;
2028 case SpvOpFwidth:
2029 val->ssa->def = nir_fadd(&b->nb,
2030 nir_fabs(&b->nb, nir_fddx(&b->nb, src[0])),
2031 nir_fabs(&b->nb, nir_fddx(&b->nb, src[1])));
2032 return;
2033 case SpvOpFwidthFine:
2034 val->ssa->def = nir_fadd(&b->nb,
2035 nir_fabs(&b->nb, nir_fddx_fine(&b->nb, src[0])),
2036 nir_fabs(&b->nb, nir_fddx_fine(&b->nb, src[1])));
2037 return;
2038 case SpvOpFwidthCoarse:
2039 val->ssa->def = nir_fadd(&b->nb,
2040 nir_fabs(&b->nb, nir_fddx_coarse(&b->nb, src[0])),
2041 nir_fabs(&b->nb, nir_fddx_coarse(&b->nb, src[1])));
2042 return;
2043
2044 case SpvOpVectorTimesScalar:
2045 /* The builder will take care of splatting for us. */
2046 val->ssa->def = nir_fmul(&b->nb, src[0], src[1]);
2047 return;
2048
2049 case SpvOpSRem:
2050 case SpvOpFRem:
2051 unreachable("No NIR equivalent");
2052
2053 case SpvOpIsNan:
2054 case SpvOpIsInf:
2055 case SpvOpIsFinite:
2056 case SpvOpIsNormal:
2057 case SpvOpSignBitSet:
2058 case SpvOpLessOrGreater:
2059 case SpvOpOrdered:
2060 case SpvOpUnordered:
2061 default:
2062 unreachable("Unhandled opcode");
2063 }
2064
2065 if (swap) {
2066 nir_ssa_def *tmp = src[0];
2067 src[0] = src[1];
2068 src[1] = tmp;
2069 }
2070
2071 nir_alu_instr *instr = nir_alu_instr_create(b->shader, op);
2072 nir_ssa_dest_init(&instr->instr, &instr->dest.dest,
2073 glsl_get_vector_elements(type), val->name);
2074 instr->dest.write_mask = (1 << glsl_get_vector_elements(type)) - 1;
2075 val->ssa->def = &instr->dest.dest.ssa;
2076
2077 for (unsigned i = 0; i < nir_op_infos[op].num_inputs; i++)
2078 instr->src[i].src = nir_src_for_ssa(src[i]);
2079
2080 nir_builder_instr_insert(&b->nb, &instr->instr);
2081 }
2082
2083 static nir_ssa_def *
2084 vtn_vector_extract(struct vtn_builder *b, nir_ssa_def *src, unsigned index)
2085 {
2086 unsigned swiz[4] = { index };
2087 return nir_swizzle(&b->nb, src, swiz, 1, true);
2088 }
2089
2090
2091 static nir_ssa_def *
2092 vtn_vector_insert(struct vtn_builder *b, nir_ssa_def *src, nir_ssa_def *insert,
2093 unsigned index)
2094 {
2095 nir_alu_instr *vec = create_vec(b->shader, src->num_components);
2096
2097 for (unsigned i = 0; i < src->num_components; i++) {
2098 if (i == index) {
2099 vec->src[i].src = nir_src_for_ssa(insert);
2100 } else {
2101 vec->src[i].src = nir_src_for_ssa(src);
2102 vec->src[i].swizzle[0] = i;
2103 }
2104 }
2105
2106 nir_builder_instr_insert(&b->nb, &vec->instr);
2107
2108 return &vec->dest.dest.ssa;
2109 }
2110
2111 static nir_ssa_def *
2112 vtn_vector_extract_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2113 nir_ssa_def *index)
2114 {
2115 nir_ssa_def *dest = vtn_vector_extract(b, src, 0);
2116 for (unsigned i = 1; i < src->num_components; i++)
2117 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2118 vtn_vector_extract(b, src, i), dest);
2119
2120 return dest;
2121 }
2122
2123 static nir_ssa_def *
2124 vtn_vector_insert_dynamic(struct vtn_builder *b, nir_ssa_def *src,
2125 nir_ssa_def *insert, nir_ssa_def *index)
2126 {
2127 nir_ssa_def *dest = vtn_vector_insert(b, src, insert, 0);
2128 for (unsigned i = 1; i < src->num_components; i++)
2129 dest = nir_bcsel(&b->nb, nir_ieq(&b->nb, index, nir_imm_int(&b->nb, i)),
2130 vtn_vector_insert(b, src, insert, i), dest);
2131
2132 return dest;
2133 }
2134
2135 static nir_ssa_def *
2136 vtn_vector_shuffle(struct vtn_builder *b, unsigned num_components,
2137 nir_ssa_def *src0, nir_ssa_def *src1,
2138 const uint32_t *indices)
2139 {
2140 nir_alu_instr *vec = create_vec(b->shader, num_components);
2141
2142 nir_ssa_undef_instr *undef = nir_ssa_undef_instr_create(b->shader, 1);
2143 nir_builder_instr_insert(&b->nb, &undef->instr);
2144
2145 for (unsigned i = 0; i < num_components; i++) {
2146 uint32_t index = indices[i];
2147 if (index == 0xffffffff) {
2148 vec->src[i].src = nir_src_for_ssa(&undef->def);
2149 } else if (index < src0->num_components) {
2150 vec->src[i].src = nir_src_for_ssa(src0);
2151 vec->src[i].swizzle[0] = index;
2152 } else {
2153 vec->src[i].src = nir_src_for_ssa(src1);
2154 vec->src[i].swizzle[0] = index - src0->num_components;
2155 }
2156 }
2157
2158 nir_builder_instr_insert(&b->nb, &vec->instr);
2159
2160 return &vec->dest.dest.ssa;
2161 }
2162
2163 /*
2164 * Concatentates a number of vectors/scalars together to produce a vector
2165 */
2166 static nir_ssa_def *
2167 vtn_vector_construct(struct vtn_builder *b, unsigned num_components,
2168 unsigned num_srcs, nir_ssa_def **srcs)
2169 {
2170 nir_alu_instr *vec = create_vec(b->shader, num_components);
2171
2172 unsigned dest_idx = 0;
2173 for (unsigned i = 0; i < num_srcs; i++) {
2174 nir_ssa_def *src = srcs[i];
2175 for (unsigned j = 0; j < src->num_components; j++) {
2176 vec->src[dest_idx].src = nir_src_for_ssa(src);
2177 vec->src[dest_idx].swizzle[0] = j;
2178 dest_idx++;
2179 }
2180 }
2181
2182 nir_builder_instr_insert(&b->nb, &vec->instr);
2183
2184 return &vec->dest.dest.ssa;
2185 }
2186
2187 static struct vtn_ssa_value *
2188 vtn_composite_copy(void *mem_ctx, struct vtn_ssa_value *src)
2189 {
2190 struct vtn_ssa_value *dest = rzalloc(mem_ctx, struct vtn_ssa_value);
2191 dest->type = src->type;
2192
2193 if (glsl_type_is_vector_or_scalar(src->type)) {
2194 dest->def = src->def;
2195 } else {
2196 unsigned elems = glsl_get_length(src->type);
2197
2198 dest->elems = ralloc_array(mem_ctx, struct vtn_ssa_value *, elems);
2199 for (unsigned i = 0; i < elems; i++)
2200 dest->elems[i] = vtn_composite_copy(mem_ctx, src->elems[i]);
2201 }
2202
2203 return dest;
2204 }
2205
2206 static struct vtn_ssa_value *
2207 vtn_composite_insert(struct vtn_builder *b, struct vtn_ssa_value *src,
2208 struct vtn_ssa_value *insert, const uint32_t *indices,
2209 unsigned num_indices)
2210 {
2211 struct vtn_ssa_value *dest = vtn_composite_copy(b, src);
2212
2213 struct vtn_ssa_value *cur = dest;
2214 unsigned i;
2215 for (i = 0; i < num_indices - 1; i++) {
2216 cur = cur->elems[indices[i]];
2217 }
2218
2219 if (glsl_type_is_vector_or_scalar(cur->type)) {
2220 /* According to the SPIR-V spec, OpCompositeInsert may work down to
2221 * the component granularity. In that case, the last index will be
2222 * the index to insert the scalar into the vector.
2223 */
2224
2225 cur->def = vtn_vector_insert(b, cur->def, insert->def, indices[i]);
2226 } else {
2227 cur->elems[indices[i]] = insert;
2228 }
2229
2230 return dest;
2231 }
2232
2233 static struct vtn_ssa_value *
2234 vtn_composite_extract(struct vtn_builder *b, struct vtn_ssa_value *src,
2235 const uint32_t *indices, unsigned num_indices)
2236 {
2237 struct vtn_ssa_value *cur = src;
2238 for (unsigned i = 0; i < num_indices; i++) {
2239 if (glsl_type_is_vector_or_scalar(cur->type)) {
2240 assert(i == num_indices - 1);
2241 /* According to the SPIR-V spec, OpCompositeExtract may work down to
2242 * the component granularity. The last index will be the index of the
2243 * vector to extract.
2244 */
2245
2246 struct vtn_ssa_value *ret = rzalloc(b, struct vtn_ssa_value);
2247 ret->type = glsl_scalar_type(glsl_get_base_type(cur->type));
2248 ret->def = vtn_vector_extract(b, cur->def, indices[i]);
2249 return ret;
2250 }
2251 }
2252
2253 return cur;
2254 }
2255
2256 static void
2257 vtn_handle_composite(struct vtn_builder *b, SpvOp opcode,
2258 const uint32_t *w, unsigned count)
2259 {
2260 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2261 const struct glsl_type *type =
2262 vtn_value(b, w[1], vtn_value_type_type)->type->type;
2263 val->ssa = vtn_create_ssa_value(b, type);
2264
2265 switch (opcode) {
2266 case SpvOpVectorExtractDynamic:
2267 val->ssa->def = vtn_vector_extract_dynamic(b, vtn_ssa_value(b, w[3])->def,
2268 vtn_ssa_value(b, w[4])->def);
2269 break;
2270
2271 case SpvOpVectorInsertDynamic:
2272 val->ssa->def = vtn_vector_insert_dynamic(b, vtn_ssa_value(b, w[3])->def,
2273 vtn_ssa_value(b, w[4])->def,
2274 vtn_ssa_value(b, w[5])->def);
2275 break;
2276
2277 case SpvOpVectorShuffle:
2278 val->ssa->def = vtn_vector_shuffle(b, glsl_get_vector_elements(type),
2279 vtn_ssa_value(b, w[3])->def,
2280 vtn_ssa_value(b, w[4])->def,
2281 w + 5);
2282 break;
2283
2284 case SpvOpCompositeConstruct: {
2285 unsigned elems = count - 3;
2286 if (glsl_type_is_vector_or_scalar(type)) {
2287 nir_ssa_def *srcs[4];
2288 for (unsigned i = 0; i < elems; i++)
2289 srcs[i] = vtn_ssa_value(b, w[3 + i])->def;
2290 val->ssa->def =
2291 vtn_vector_construct(b, glsl_get_vector_elements(type),
2292 elems, srcs);
2293 } else {
2294 val->ssa->elems = ralloc_array(b, struct vtn_ssa_value *, elems);
2295 for (unsigned i = 0; i < elems; i++)
2296 val->ssa->elems[i] = vtn_ssa_value(b, w[3 + i]);
2297 }
2298 break;
2299 }
2300 case SpvOpCompositeExtract:
2301 val->ssa = vtn_composite_extract(b, vtn_ssa_value(b, w[3]),
2302 w + 4, count - 4);
2303 break;
2304
2305 case SpvOpCompositeInsert:
2306 val->ssa = vtn_composite_insert(b, vtn_ssa_value(b, w[4]),
2307 vtn_ssa_value(b, w[3]),
2308 w + 5, count - 5);
2309 break;
2310
2311 case SpvOpCopyObject:
2312 val->ssa = vtn_composite_copy(b, vtn_ssa_value(b, w[3]));
2313 break;
2314
2315 default:
2316 unreachable("unknown composite operation");
2317 }
2318 }
2319
2320 static void
2321 vtn_phi_node_init(struct vtn_builder *b, struct vtn_ssa_value *val)
2322 {
2323 if (glsl_type_is_vector_or_scalar(val->type)) {
2324 nir_phi_instr *phi = nir_phi_instr_create(b->shader);
2325 nir_ssa_dest_init(&phi->instr, &phi->dest,
2326 glsl_get_vector_elements(val->type), NULL);
2327 exec_list_make_empty(&phi->srcs);
2328 nir_builder_instr_insert(&b->nb, &phi->instr);
2329 val->def = &phi->dest.ssa;
2330 } else {
2331 unsigned elems = glsl_get_length(val->type);
2332 for (unsigned i = 0; i < elems; i++)
2333 vtn_phi_node_init(b, val->elems[i]);
2334 }
2335 }
2336
2337 static struct vtn_ssa_value *
2338 vtn_phi_node_create(struct vtn_builder *b, const struct glsl_type *type)
2339 {
2340 struct vtn_ssa_value *val = vtn_create_ssa_value(b, type);
2341 vtn_phi_node_init(b, val);
2342 return val;
2343 }
2344
2345 static void
2346 vtn_handle_phi_first_pass(struct vtn_builder *b, const uint32_t *w)
2347 {
2348 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
2349 const struct glsl_type *type =
2350 vtn_value(b, w[1], vtn_value_type_type)->type->type;
2351 val->ssa = vtn_phi_node_create(b, type);
2352 }
2353
2354 static void
2355 vtn_phi_node_add_src(struct vtn_ssa_value *phi, const nir_block *pred,
2356 struct vtn_ssa_value *val)
2357 {
2358 assert(phi->type == val->type);
2359 if (glsl_type_is_vector_or_scalar(phi->type)) {
2360 nir_phi_instr *phi_instr = nir_instr_as_phi(phi->def->parent_instr);
2361 nir_phi_src *src = ralloc(phi_instr, nir_phi_src);
2362 src->pred = (nir_block *) pred;
2363 src->src = nir_src_for_ssa(val->def);
2364 exec_list_push_tail(&phi_instr->srcs, &src->node);
2365 } else {
2366 unsigned elems = glsl_get_length(phi->type);
2367 for (unsigned i = 0; i < elems; i++)
2368 vtn_phi_node_add_src(phi->elems[i], pred, val->elems[i]);
2369 }
2370 }
2371
2372 static struct vtn_ssa_value *
2373 vtn_get_phi_node_src(struct vtn_builder *b, nir_block *block,
2374 const struct glsl_type *type, const uint32_t *w,
2375 unsigned count)
2376 {
2377 struct hash_entry *entry = _mesa_hash_table_search(b->block_table, block);
2378 if (entry) {
2379 struct vtn_block *spv_block = entry->data;
2380 for (unsigned off = 4; off < count; off += 2) {
2381 if (spv_block == vtn_value(b, w[off], vtn_value_type_block)->block) {
2382 return vtn_ssa_value(b, w[off - 1]);
2383 }
2384 }
2385 }
2386
2387 b->nb.cursor = nir_before_block(block);
2388 struct vtn_ssa_value *phi = vtn_phi_node_create(b, type);
2389
2390 struct set_entry *entry2;
2391 set_foreach(block->predecessors, entry2) {
2392 nir_block *pred = (nir_block *) entry2->key;
2393 struct vtn_ssa_value *val = vtn_get_phi_node_src(b, pred, type, w,
2394 count);
2395 vtn_phi_node_add_src(phi, pred, val);
2396 }
2397
2398 return phi;
2399 }
2400
2401 static bool
2402 vtn_handle_phi_second_pass(struct vtn_builder *b, SpvOp opcode,
2403 const uint32_t *w, unsigned count)
2404 {
2405 if (opcode == SpvOpLabel) {
2406 b->block = vtn_value(b, w[1], vtn_value_type_block)->block;
2407 return true;
2408 }
2409
2410 if (opcode != SpvOpPhi)
2411 return true;
2412
2413 struct vtn_ssa_value *phi = vtn_value(b, w[2], vtn_value_type_ssa)->ssa;
2414
2415 struct set_entry *entry;
2416 set_foreach(b->block->block->predecessors, entry) {
2417 nir_block *pred = (nir_block *) entry->key;
2418
2419 struct vtn_ssa_value *val = vtn_get_phi_node_src(b, pred, phi->type, w,
2420 count);
2421 vtn_phi_node_add_src(phi, pred, val);
2422 }
2423
2424 return true;
2425 }
2426
2427 static bool
2428 vtn_handle_preamble_instruction(struct vtn_builder *b, SpvOp opcode,
2429 const uint32_t *w, unsigned count)
2430 {
2431 switch (opcode) {
2432 case SpvOpSource:
2433 case SpvOpSourceExtension:
2434 case SpvOpExtension:
2435 /* Unhandled, but these are for debug so that's ok. */
2436 break;
2437
2438 case SpvOpCapability:
2439 /*
2440 * TODO properly handle these and give a real error if asking for too
2441 * much.
2442 */
2443 assert(w[1] == SpvCapabilityMatrix ||
2444 w[1] == SpvCapabilityShader);
2445 break;
2446
2447 case SpvOpExtInstImport:
2448 vtn_handle_extension(b, opcode, w, count);
2449 break;
2450
2451 case SpvOpMemoryModel:
2452 assert(w[1] == SpvAddressingModelLogical);
2453 assert(w[2] == SpvMemoryModelGLSL450);
2454 break;
2455
2456 case SpvOpEntryPoint:
2457 assert(b->entry_point == NULL);
2458 b->entry_point = &b->values[w[2]];
2459 b->execution_model = w[1];
2460 break;
2461
2462 case SpvOpExecutionMode:
2463 /* TODO */
2464 break;
2465
2466 case SpvOpString:
2467 vtn_push_value(b, w[1], vtn_value_type_string)->str =
2468 vtn_string_literal(b, &w[2], count - 2);
2469 break;
2470
2471 case SpvOpName:
2472 b->values[w[1]].name = vtn_string_literal(b, &w[2], count - 2);
2473 break;
2474
2475 case SpvOpMemberName:
2476 /* TODO */
2477 break;
2478
2479 case SpvOpLine:
2480 break; /* Ignored for now */
2481
2482 case SpvOpDecorationGroup:
2483 case SpvOpDecorate:
2484 case SpvOpMemberDecorate:
2485 case SpvOpGroupDecorate:
2486 case SpvOpGroupMemberDecorate:
2487 vtn_handle_decoration(b, opcode, w, count);
2488 break;
2489
2490 case SpvOpTypeVoid:
2491 case SpvOpTypeBool:
2492 case SpvOpTypeInt:
2493 case SpvOpTypeFloat:
2494 case SpvOpTypeVector:
2495 case SpvOpTypeMatrix:
2496 case SpvOpTypeImage:
2497 case SpvOpTypeSampler:
2498 case SpvOpTypeSampledImage:
2499 case SpvOpTypeArray:
2500 case SpvOpTypeRuntimeArray:
2501 case SpvOpTypeStruct:
2502 case SpvOpTypeOpaque:
2503 case SpvOpTypePointer:
2504 case SpvOpTypeFunction:
2505 case SpvOpTypeEvent:
2506 case SpvOpTypeDeviceEvent:
2507 case SpvOpTypeReserveId:
2508 case SpvOpTypeQueue:
2509 case SpvOpTypePipe:
2510 vtn_handle_type(b, opcode, w, count);
2511 break;
2512
2513 case SpvOpConstantTrue:
2514 case SpvOpConstantFalse:
2515 case SpvOpConstant:
2516 case SpvOpConstantComposite:
2517 case SpvOpConstantSampler:
2518 case SpvOpSpecConstantTrue:
2519 case SpvOpSpecConstantFalse:
2520 case SpvOpSpecConstant:
2521 case SpvOpSpecConstantComposite:
2522 vtn_handle_constant(b, opcode, w, count);
2523 break;
2524
2525 case SpvOpVariable:
2526 vtn_handle_variables(b, opcode, w, count);
2527 break;
2528
2529 default:
2530 return false; /* End of preamble */
2531 }
2532
2533 return true;
2534 }
2535
2536 static bool
2537 vtn_handle_first_cfg_pass_instruction(struct vtn_builder *b, SpvOp opcode,
2538 const uint32_t *w, unsigned count)
2539 {
2540 switch (opcode) {
2541 case SpvOpFunction: {
2542 assert(b->func == NULL);
2543 b->func = rzalloc(b, struct vtn_function);
2544
2545 const struct glsl_type *result_type =
2546 vtn_value(b, w[1], vtn_value_type_type)->type->type;
2547 struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_function);
2548 const struct glsl_type *func_type =
2549 vtn_value(b, w[4], vtn_value_type_type)->type->type;
2550
2551 assert(glsl_get_function_return_type(func_type) == result_type);
2552
2553 nir_function *func =
2554 nir_function_create(b->shader, ralloc_strdup(b->shader, val->name));
2555
2556 nir_function_overload *overload = nir_function_overload_create(func);
2557 overload->num_params = glsl_get_length(func_type);
2558 overload->params = ralloc_array(overload, nir_parameter,
2559 overload->num_params);
2560 for (unsigned i = 0; i < overload->num_params; i++) {
2561 const struct glsl_function_param *param =
2562 glsl_get_function_param(func_type, i);
2563 overload->params[i].type = param->type;
2564 if (param->in) {
2565 if (param->out) {
2566 overload->params[i].param_type = nir_parameter_inout;
2567 } else {
2568 overload->params[i].param_type = nir_parameter_in;
2569 }
2570 } else {
2571 if (param->out) {
2572 overload->params[i].param_type = nir_parameter_out;
2573 } else {
2574 assert(!"Parameter is neither in nor out");
2575 }
2576 }
2577 }
2578 b->func->overload = overload;
2579 break;
2580 }
2581
2582 case SpvOpFunctionEnd:
2583 b->func->end = w;
2584 b->func = NULL;
2585 break;
2586
2587 case SpvOpFunctionParameter:
2588 break; /* Does nothing */
2589
2590 case SpvOpLabel: {
2591 assert(b->block == NULL);
2592 b->block = rzalloc(b, struct vtn_block);
2593 b->block->label = w;
2594 vtn_push_value(b, w[1], vtn_value_type_block)->block = b->block;
2595
2596 if (b->func->start_block == NULL) {
2597 /* This is the first block encountered for this function. In this
2598 * case, we set the start block and add it to the list of
2599 * implemented functions that we'll walk later.
2600 */
2601 b->func->start_block = b->block;
2602 exec_list_push_tail(&b->functions, &b->func->node);
2603 }
2604 break;
2605 }
2606
2607 case SpvOpBranch:
2608 case SpvOpBranchConditional:
2609 case SpvOpSwitch:
2610 case SpvOpKill:
2611 case SpvOpReturn:
2612 case SpvOpReturnValue:
2613 case SpvOpUnreachable:
2614 assert(b->block);
2615 b->block->branch = w;
2616 b->block = NULL;
2617 break;
2618
2619 case SpvOpSelectionMerge:
2620 case SpvOpLoopMerge:
2621 assert(b->block && b->block->merge_op == SpvOpNop);
2622 b->block->merge_op = opcode;
2623 b->block->merge_block_id = w[1];
2624 break;
2625
2626 default:
2627 /* Continue on as per normal */
2628 return true;
2629 }
2630
2631 return true;
2632 }
2633
2634 static bool
2635 vtn_handle_body_instruction(struct vtn_builder *b, SpvOp opcode,
2636 const uint32_t *w, unsigned count)
2637 {
2638 switch (opcode) {
2639 case SpvOpLabel: {
2640 struct vtn_block *block = vtn_value(b, w[1], vtn_value_type_block)->block;
2641 assert(block->block == NULL);
2642
2643 block->block = nir_cursor_current_block(b->nb.cursor);
2644 break;
2645 }
2646
2647 case SpvOpLoopMerge:
2648 case SpvOpSelectionMerge:
2649 /* This is handled by cfg pre-pass and walk_blocks */
2650 break;
2651
2652 case SpvOpUndef:
2653 vtn_push_value(b, w[2], vtn_value_type_undef);
2654 break;
2655
2656 case SpvOpExtInst:
2657 vtn_handle_extension(b, opcode, w, count);
2658 break;
2659
2660 case SpvOpVariable:
2661 case SpvOpLoad:
2662 case SpvOpStore:
2663 case SpvOpCopyMemory:
2664 case SpvOpCopyMemorySized:
2665 case SpvOpAccessChain:
2666 case SpvOpInBoundsAccessChain:
2667 case SpvOpArrayLength:
2668 case SpvOpImageTexelPointer:
2669 vtn_handle_variables(b, opcode, w, count);
2670 break;
2671
2672 case SpvOpFunctionCall:
2673 vtn_handle_function_call(b, opcode, w, count);
2674 break;
2675
2676 case SpvOpImageSampleImplicitLod:
2677 case SpvOpImageSampleExplicitLod:
2678 case SpvOpImageSampleDrefImplicitLod:
2679 case SpvOpImageSampleDrefExplicitLod:
2680 case SpvOpImageSampleProjImplicitLod:
2681 case SpvOpImageSampleProjExplicitLod:
2682 case SpvOpImageSampleProjDrefImplicitLod:
2683 case SpvOpImageSampleProjDrefExplicitLod:
2684 case SpvOpImageFetch:
2685 case SpvOpImageGather:
2686 case SpvOpImageDrefGather:
2687 case SpvOpImageQuerySizeLod:
2688 case SpvOpImageQuerySize:
2689 case SpvOpImageQueryLod:
2690 case SpvOpImageQueryLevels:
2691 case SpvOpImageQuerySamples:
2692 vtn_handle_texture(b, opcode, w, count);
2693 break;
2694
2695 case SpvOpSNegate:
2696 case SpvOpFNegate:
2697 case SpvOpNot:
2698 case SpvOpAny:
2699 case SpvOpAll:
2700 case SpvOpConvertFToU:
2701 case SpvOpConvertFToS:
2702 case SpvOpConvertSToF:
2703 case SpvOpConvertUToF:
2704 case SpvOpUConvert:
2705 case SpvOpSConvert:
2706 case SpvOpFConvert:
2707 case SpvOpConvertPtrToU:
2708 case SpvOpConvertUToPtr:
2709 case SpvOpPtrCastToGeneric:
2710 case SpvOpGenericCastToPtr:
2711 case SpvOpBitcast:
2712 case SpvOpIsNan:
2713 case SpvOpIsInf:
2714 case SpvOpIsFinite:
2715 case SpvOpIsNormal:
2716 case SpvOpSignBitSet:
2717 case SpvOpLessOrGreater:
2718 case SpvOpOrdered:
2719 case SpvOpUnordered:
2720 case SpvOpIAdd:
2721 case SpvOpFAdd:
2722 case SpvOpISub:
2723 case SpvOpFSub:
2724 case SpvOpIMul:
2725 case SpvOpFMul:
2726 case SpvOpUDiv:
2727 case SpvOpSDiv:
2728 case SpvOpFDiv:
2729 case SpvOpUMod:
2730 case SpvOpSRem:
2731 case SpvOpSMod:
2732 case SpvOpFRem:
2733 case SpvOpFMod:
2734 case SpvOpVectorTimesScalar:
2735 case SpvOpDot:
2736 case SpvOpShiftRightLogical:
2737 case SpvOpShiftRightArithmetic:
2738 case SpvOpShiftLeftLogical:
2739 case SpvOpLogicalOr:
2740 case SpvOpLogicalEqual:
2741 case SpvOpLogicalNotEqual:
2742 case SpvOpLogicalAnd:
2743 case SpvOpBitwiseOr:
2744 case SpvOpBitwiseXor:
2745 case SpvOpBitwiseAnd:
2746 case SpvOpSelect:
2747 case SpvOpIEqual:
2748 case SpvOpFOrdEqual:
2749 case SpvOpFUnordEqual:
2750 case SpvOpINotEqual:
2751 case SpvOpFOrdNotEqual:
2752 case SpvOpFUnordNotEqual:
2753 case SpvOpULessThan:
2754 case SpvOpSLessThan:
2755 case SpvOpFOrdLessThan:
2756 case SpvOpFUnordLessThan:
2757 case SpvOpUGreaterThan:
2758 case SpvOpSGreaterThan:
2759 case SpvOpFOrdGreaterThan:
2760 case SpvOpFUnordGreaterThan:
2761 case SpvOpULessThanEqual:
2762 case SpvOpSLessThanEqual:
2763 case SpvOpFOrdLessThanEqual:
2764 case SpvOpFUnordLessThanEqual:
2765 case SpvOpUGreaterThanEqual:
2766 case SpvOpSGreaterThanEqual:
2767 case SpvOpFOrdGreaterThanEqual:
2768 case SpvOpFUnordGreaterThanEqual:
2769 case SpvOpDPdx:
2770 case SpvOpDPdy:
2771 case SpvOpFwidth:
2772 case SpvOpDPdxFine:
2773 case SpvOpDPdyFine:
2774 case SpvOpFwidthFine:
2775 case SpvOpDPdxCoarse:
2776 case SpvOpDPdyCoarse:
2777 case SpvOpFwidthCoarse:
2778 vtn_handle_alu(b, opcode, w, count);
2779 break;
2780
2781 case SpvOpTranspose:
2782 case SpvOpOuterProduct:
2783 case SpvOpMatrixTimesScalar:
2784 case SpvOpVectorTimesMatrix:
2785 case SpvOpMatrixTimesVector:
2786 case SpvOpMatrixTimesMatrix:
2787 vtn_handle_matrix_alu(b, opcode, w, count);
2788 break;
2789
2790 case SpvOpVectorExtractDynamic:
2791 case SpvOpVectorInsertDynamic:
2792 case SpvOpVectorShuffle:
2793 case SpvOpCompositeConstruct:
2794 case SpvOpCompositeExtract:
2795 case SpvOpCompositeInsert:
2796 case SpvOpCopyObject:
2797 vtn_handle_composite(b, opcode, w, count);
2798 break;
2799
2800 case SpvOpPhi:
2801 vtn_handle_phi_first_pass(b, w);
2802 break;
2803
2804 default:
2805 unreachable("Unhandled opcode");
2806 }
2807
2808 return true;
2809 }
2810
2811 static void
2812 vtn_walk_blocks(struct vtn_builder *b, struct vtn_block *start,
2813 struct vtn_block *break_block, struct vtn_block *cont_block,
2814 struct vtn_block *end_block)
2815 {
2816 struct vtn_block *block = start;
2817 while (block != end_block) {
2818 if (block->merge_op == SpvOpLoopMerge) {
2819 /* This is the jump into a loop. */
2820 struct vtn_block *new_cont_block = block;
2821 struct vtn_block *new_break_block =
2822 vtn_value(b, block->merge_block_id, vtn_value_type_block)->block;
2823
2824 nir_loop *loop = nir_loop_create(b->shader);
2825 nir_cf_node_insert(b->nb.cursor, &loop->cf_node);
2826
2827 /* Reset the merge_op to prerevent infinite recursion */
2828 block->merge_op = SpvOpNop;
2829
2830 b->nb.cursor = nir_after_cf_list(&loop->body);
2831 vtn_walk_blocks(b, block, new_break_block, new_cont_block, NULL);
2832
2833 b->nb.cursor = nir_after_cf_node(&loop->cf_node);
2834 block = new_break_block;
2835 continue;
2836 }
2837
2838 const uint32_t *w = block->branch;
2839 SpvOp branch_op = w[0] & SpvOpCodeMask;
2840
2841 b->block = block;
2842 vtn_foreach_instruction(b, block->label, block->branch,
2843 vtn_handle_body_instruction);
2844
2845 nir_block *cur_block = nir_cursor_current_block(b->nb.cursor);
2846 assert(cur_block == block->block);
2847 _mesa_hash_table_insert(b->block_table, cur_block, block);
2848
2849 switch (branch_op) {
2850 case SpvOpBranch: {
2851 struct vtn_block *branch_block =
2852 vtn_value(b, w[1], vtn_value_type_block)->block;
2853
2854 if (branch_block == break_block) {
2855 nir_jump_instr *jump = nir_jump_instr_create(b->shader,
2856 nir_jump_break);
2857 nir_builder_instr_insert(&b->nb, &jump->instr);
2858
2859 return;
2860 } else if (branch_block == cont_block) {
2861 nir_jump_instr *jump = nir_jump_instr_create(b->shader,
2862 nir_jump_continue);
2863 nir_builder_instr_insert(&b->nb, &jump->instr);
2864
2865 return;
2866 } else if (branch_block == end_block) {
2867 /* We're branching to the merge block of an if, since for loops
2868 * and functions end_block == NULL, so we're done here.
2869 */
2870 return;
2871 } else {
2872 /* We're branching to another block, and according to the rules,
2873 * we can only branch to another block with one predecessor (so
2874 * we're the only one jumping to it) so we can just process it
2875 * next.
2876 */
2877 block = branch_block;
2878 continue;
2879 }
2880 }
2881
2882 case SpvOpBranchConditional: {
2883 /* Gather up the branch blocks */
2884 struct vtn_block *then_block =
2885 vtn_value(b, w[2], vtn_value_type_block)->block;
2886 struct vtn_block *else_block =
2887 vtn_value(b, w[3], vtn_value_type_block)->block;
2888
2889 nir_if *if_stmt = nir_if_create(b->shader);
2890 if_stmt->condition = nir_src_for_ssa(vtn_ssa_value(b, w[1])->def);
2891 nir_cf_node_insert(b->nb.cursor, &if_stmt->cf_node);
2892
2893 if (then_block == break_block) {
2894 nir_jump_instr *jump = nir_jump_instr_create(b->shader,
2895 nir_jump_break);
2896 nir_instr_insert_after_cf_list(&if_stmt->then_list,
2897 &jump->instr);
2898 block = else_block;
2899 } else if (else_block == break_block) {
2900 nir_jump_instr *jump = nir_jump_instr_create(b->shader,
2901 nir_jump_break);
2902 nir_instr_insert_after_cf_list(&if_stmt->else_list,
2903 &jump->instr);
2904 block = then_block;
2905 } else if (then_block == cont_block) {
2906 nir_jump_instr *jump = nir_jump_instr_create(b->shader,
2907 nir_jump_continue);
2908 nir_instr_insert_after_cf_list(&if_stmt->then_list,
2909 &jump->instr);
2910 block = else_block;
2911 } else if (else_block == cont_block) {
2912 nir_jump_instr *jump = nir_jump_instr_create(b->shader,
2913 nir_jump_continue);
2914 nir_instr_insert_after_cf_list(&if_stmt->else_list,
2915 &jump->instr);
2916 block = then_block;
2917 } else {
2918 /* According to the rules we're branching to two blocks that don't
2919 * have any other predecessors, so we can handle this as a
2920 * conventional if.
2921 */
2922 assert(block->merge_op == SpvOpSelectionMerge);
2923 struct vtn_block *merge_block =
2924 vtn_value(b, block->merge_block_id, vtn_value_type_block)->block;
2925
2926 b->nb.cursor = nir_after_cf_list(&if_stmt->then_list);
2927 vtn_walk_blocks(b, then_block, break_block, cont_block, merge_block);
2928
2929 b->nb.cursor = nir_after_cf_list(&if_stmt->else_list);
2930 vtn_walk_blocks(b, else_block, break_block, cont_block, merge_block);
2931
2932 b->nb.cursor = nir_after_cf_node(&if_stmt->cf_node);
2933 block = merge_block;
2934 continue;
2935 }
2936
2937 /* If we got here then we inserted a predicated break or continue
2938 * above and we need to handle the other case. We already set
2939 * `block` above to indicate what block to visit after the
2940 * predicated break.
2941 */
2942
2943 /* It's possible that the other branch is also a break/continue.
2944 * If it is, we handle that here.
2945 */
2946 if (block == break_block) {
2947 nir_jump_instr *jump = nir_jump_instr_create(b->shader,
2948 nir_jump_break);
2949 nir_builder_instr_insert(&b->nb, &jump->instr);
2950
2951 return;
2952 } else if (block == cont_block) {
2953 nir_jump_instr *jump = nir_jump_instr_create(b->shader,
2954 nir_jump_continue);
2955 nir_builder_instr_insert(&b->nb, &jump->instr);
2956
2957 return;
2958 }
2959
2960 /* If we got here then there was a predicated break/continue but
2961 * the other half of the if has stuff in it. `block` was already
2962 * set above so there is nothing left for us to do.
2963 */
2964 continue;
2965 }
2966
2967 case SpvOpReturn: {
2968 nir_jump_instr *jump = nir_jump_instr_create(b->shader,
2969 nir_jump_return);
2970 nir_builder_instr_insert(&b->nb, &jump->instr);
2971 return;
2972 }
2973
2974 case SpvOpKill: {
2975 nir_intrinsic_instr *discard =
2976 nir_intrinsic_instr_create(b->shader, nir_intrinsic_discard);
2977 nir_builder_instr_insert(&b->nb, &discard->instr);
2978 return;
2979 }
2980
2981 case SpvOpSwitch:
2982 case SpvOpReturnValue:
2983 case SpvOpUnreachable:
2984 default:
2985 unreachable("Unhandled opcode");
2986 }
2987 }
2988 }
2989
2990 nir_shader *
2991 spirv_to_nir(const uint32_t *words, size_t word_count,
2992 gl_shader_stage stage,
2993 const nir_shader_compiler_options *options)
2994 {
2995 const uint32_t *word_end = words + word_count;
2996
2997 /* Handle the SPIR-V header (first 4 dwords) */
2998 assert(word_count > 5);
2999
3000 assert(words[0] == SpvMagicNumber);
3001 assert(words[1] == 99);
3002 /* words[2] == generator magic */
3003 unsigned value_id_bound = words[3];
3004 assert(words[4] == 0);
3005
3006 words+= 5;
3007
3008 nir_shader *shader = nir_shader_create(NULL, stage, options);
3009
3010 /* Initialize the stn_builder object */
3011 struct vtn_builder *b = rzalloc(NULL, struct vtn_builder);
3012 b->shader = shader;
3013 b->value_id_bound = value_id_bound;
3014 b->values = rzalloc_array(b, struct vtn_value, value_id_bound);
3015 exec_list_make_empty(&b->functions);
3016
3017 /* Handle all the preamble instructions */
3018 words = vtn_foreach_instruction(b, words, word_end,
3019 vtn_handle_preamble_instruction);
3020
3021 /* Do a very quick CFG analysis pass */
3022 vtn_foreach_instruction(b, words, word_end,
3023 vtn_handle_first_cfg_pass_instruction);
3024
3025 foreach_list_typed(struct vtn_function, func, node, &b->functions) {
3026 b->impl = nir_function_impl_create(func->overload);
3027 b->const_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
3028 _mesa_key_pointer_equal);
3029 b->block_table = _mesa_hash_table_create(b, _mesa_hash_pointer,
3030 _mesa_key_pointer_equal);
3031 nir_builder_init(&b->nb, b->impl);
3032 b->nb.cursor = nir_after_cf_list(&b->impl->body);
3033 vtn_walk_blocks(b, func->start_block, NULL, NULL, NULL);
3034 vtn_foreach_instruction(b, func->start_block->label, func->end,
3035 vtn_handle_phi_second_pass);
3036 }
3037
3038 /* Because we can still have output reads in NIR, we need to lower
3039 * outputs to temporaries before we are truely finished.
3040 */
3041 nir_lower_outputs_to_temporaries(shader);
3042
3043 ralloc_free(b);
3044
3045 return shader;
3046 }