glsl: Inline builtins in a separate pass
[mesa.git] / src / compiler / glsl / ast_function.cpp
1 /*
2 * Copyright © 2010 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
21 * DEALINGS IN THE SOFTWARE.
22 */
23
24 #include "glsl_symbol_table.h"
25 #include "ast.h"
26 #include "compiler/glsl_types.h"
27 #include "ir.h"
28 #include "main/mtypes.h"
29 #include "main/shaderobj.h"
30 #include "builtin_functions.h"
31
32 static ir_rvalue *
33 convert_component(ir_rvalue *src, const glsl_type *desired_type);
34
35 static unsigned
36 process_parameters(exec_list *instructions, exec_list *actual_parameters,
37 exec_list *parameters,
38 struct _mesa_glsl_parse_state *state)
39 {
40 void *mem_ctx = state;
41 unsigned count = 0;
42
43 foreach_list_typed(ast_node, ast, link, parameters) {
44 /* We need to process the parameters first in order to know if we can
45 * raise or not a unitialized warning. Calling set_is_lhs silence the
46 * warning for now. Raising the warning or not will be checked at
47 * verify_parameter_modes.
48 */
49 ast->set_is_lhs(true);
50 ir_rvalue *result = ast->hir(instructions, state);
51
52 ir_constant *const constant =
53 result->constant_expression_value(mem_ctx);
54
55 if (constant != NULL)
56 result = constant;
57
58 actual_parameters->push_tail(result);
59 count++;
60 }
61
62 return count;
63 }
64
65
66 /**
67 * Generate a source prototype for a function signature
68 *
69 * \param return_type Return type of the function. May be \c NULL.
70 * \param name Name of the function.
71 * \param parameters List of \c ir_instruction nodes representing the
72 * parameter list for the function. This may be either a
73 * formal (\c ir_variable) or actual (\c ir_rvalue)
74 * parameter list. Only the type is used.
75 *
76 * \return
77 * A ralloced string representing the prototype of the function.
78 */
79 char *
80 prototype_string(const glsl_type *return_type, const char *name,
81 exec_list *parameters)
82 {
83 char *str = NULL;
84
85 if (return_type != NULL)
86 str = ralloc_asprintf(NULL, "%s ", return_type->name);
87
88 ralloc_asprintf_append(&str, "%s(", name);
89
90 const char *comma = "";
91 foreach_in_list(const ir_variable, param, parameters) {
92 ralloc_asprintf_append(&str, "%s%s", comma, param->type->name);
93 comma = ", ";
94 }
95
96 ralloc_strcat(&str, ")");
97 return str;
98 }
99
100 static bool
101 verify_image_parameter(YYLTYPE *loc, _mesa_glsl_parse_state *state,
102 const ir_variable *formal, const ir_variable *actual)
103 {
104 /**
105 * From the ARB_shader_image_load_store specification:
106 *
107 * "The values of image variables qualified with coherent,
108 * volatile, restrict, readonly, or writeonly may not be passed
109 * to functions whose formal parameters lack such
110 * qualifiers. [...] It is legal to have additional qualifiers
111 * on a formal parameter, but not to have fewer."
112 */
113 if (actual->data.memory_coherent && !formal->data.memory_coherent) {
114 _mesa_glsl_error(loc, state,
115 "function call parameter `%s' drops "
116 "`coherent' qualifier", formal->name);
117 return false;
118 }
119
120 if (actual->data.memory_volatile && !formal->data.memory_volatile) {
121 _mesa_glsl_error(loc, state,
122 "function call parameter `%s' drops "
123 "`volatile' qualifier", formal->name);
124 return false;
125 }
126
127 if (actual->data.memory_restrict && !formal->data.memory_restrict) {
128 _mesa_glsl_error(loc, state,
129 "function call parameter `%s' drops "
130 "`restrict' qualifier", formal->name);
131 return false;
132 }
133
134 if (actual->data.memory_read_only && !formal->data.memory_read_only) {
135 _mesa_glsl_error(loc, state,
136 "function call parameter `%s' drops "
137 "`readonly' qualifier", formal->name);
138 return false;
139 }
140
141 if (actual->data.memory_write_only && !formal->data.memory_write_only) {
142 _mesa_glsl_error(loc, state,
143 "function call parameter `%s' drops "
144 "`writeonly' qualifier", formal->name);
145 return false;
146 }
147
148 return true;
149 }
150
151 static bool
152 verify_first_atomic_parameter(YYLTYPE *loc, _mesa_glsl_parse_state *state,
153 ir_variable *var)
154 {
155 if (!var ||
156 (!var->is_in_shader_storage_block() &&
157 var->data.mode != ir_var_shader_shared)) {
158 _mesa_glsl_error(loc, state, "First argument to atomic function "
159 "must be a buffer or shared variable");
160 return false;
161 }
162 return true;
163 }
164
165 static bool
166 is_atomic_function(const char *func_name)
167 {
168 return !strcmp(func_name, "atomicAdd") ||
169 !strcmp(func_name, "atomicMin") ||
170 !strcmp(func_name, "atomicMax") ||
171 !strcmp(func_name, "atomicAnd") ||
172 !strcmp(func_name, "atomicOr") ||
173 !strcmp(func_name, "atomicXor") ||
174 !strcmp(func_name, "atomicExchange") ||
175 !strcmp(func_name, "atomicCompSwap");
176 }
177
178 /**
179 * Verify that 'out' and 'inout' actual parameters are lvalues. Also, verify
180 * that 'const_in' formal parameters (an extension in our IR) correspond to
181 * ir_constant actual parameters.
182 */
183 static bool
184 verify_parameter_modes(_mesa_glsl_parse_state *state,
185 ir_function_signature *sig,
186 exec_list &actual_ir_parameters,
187 exec_list &actual_ast_parameters)
188 {
189 exec_node *actual_ir_node = actual_ir_parameters.get_head_raw();
190 exec_node *actual_ast_node = actual_ast_parameters.get_head_raw();
191
192 foreach_in_list(const ir_variable, formal, &sig->parameters) {
193 /* The lists must be the same length. */
194 assert(!actual_ir_node->is_tail_sentinel());
195 assert(!actual_ast_node->is_tail_sentinel());
196
197 const ir_rvalue *const actual = (ir_rvalue *) actual_ir_node;
198 const ast_expression *const actual_ast =
199 exec_node_data(ast_expression, actual_ast_node, link);
200
201 /* FIXME: 'loc' is incorrect (as of 2011-01-21). It is always
202 * FIXME: 0:0(0).
203 */
204 YYLTYPE loc = actual_ast->get_location();
205
206 /* Verify that 'const_in' parameters are ir_constants. */
207 if (formal->data.mode == ir_var_const_in &&
208 actual->ir_type != ir_type_constant) {
209 _mesa_glsl_error(&loc, state,
210 "parameter `in %s' must be a constant expression",
211 formal->name);
212 return false;
213 }
214
215 /* Verify that shader_in parameters are shader inputs */
216 if (formal->data.must_be_shader_input) {
217 const ir_rvalue *val = actual;
218
219 /* GLSL 4.40 allows swizzles, while earlier GLSL versions do not. */
220 if (val->ir_type == ir_type_swizzle) {
221 if (!state->is_version(440, 0)) {
222 _mesa_glsl_error(&loc, state,
223 "parameter `%s` must not be swizzled",
224 formal->name);
225 return false;
226 }
227 val = ((ir_swizzle *)val)->val;
228 }
229
230 for (;;) {
231 if (val->ir_type == ir_type_dereference_array) {
232 val = ((ir_dereference_array *)val)->array;
233 } else if (val->ir_type == ir_type_dereference_record &&
234 !state->es_shader) {
235 val = ((ir_dereference_record *)val)->record;
236 } else
237 break;
238 }
239
240 ir_variable *var = NULL;
241 if (const ir_dereference_variable *deref_var = val->as_dereference_variable())
242 var = deref_var->variable_referenced();
243
244 if (!var || var->data.mode != ir_var_shader_in) {
245 _mesa_glsl_error(&loc, state,
246 "parameter `%s` must be a shader input",
247 formal->name);
248 return false;
249 }
250
251 var->data.must_be_shader_input = 1;
252 }
253
254 /* Verify that 'out' and 'inout' actual parameters are lvalues. */
255 if (formal->data.mode == ir_var_function_out
256 || formal->data.mode == ir_var_function_inout) {
257 const char *mode = NULL;
258 switch (formal->data.mode) {
259 case ir_var_function_out: mode = "out"; break;
260 case ir_var_function_inout: mode = "inout"; break;
261 default: assert(false); break;
262 }
263
264 /* This AST-based check catches errors like f(i++). The IR-based
265 * is_lvalue() is insufficient because the actual parameter at the
266 * IR-level is just a temporary value, which is an l-value.
267 */
268 if (actual_ast->non_lvalue_description != NULL) {
269 _mesa_glsl_error(&loc, state,
270 "function parameter '%s %s' references a %s",
271 mode, formal->name,
272 actual_ast->non_lvalue_description);
273 return false;
274 }
275
276 ir_variable *var = actual->variable_referenced();
277
278 if (var && formal->data.mode == ir_var_function_inout) {
279 if ((var->data.mode == ir_var_auto ||
280 var->data.mode == ir_var_shader_out) &&
281 !var->data.assigned &&
282 !is_gl_identifier(var->name)) {
283 _mesa_glsl_warning(&loc, state, "`%s' used uninitialized",
284 var->name);
285 }
286 }
287
288 if (var)
289 var->data.assigned = true;
290
291 if (var && var->data.read_only) {
292 _mesa_glsl_error(&loc, state,
293 "function parameter '%s %s' references the "
294 "read-only variable '%s'",
295 mode, formal->name,
296 actual->variable_referenced()->name);
297 return false;
298 } else if (!actual->is_lvalue(state)) {
299 _mesa_glsl_error(&loc, state,
300 "function parameter '%s %s' is not an lvalue",
301 mode, formal->name);
302 return false;
303 }
304 } else {
305 assert(formal->data.mode == ir_var_function_in ||
306 formal->data.mode == ir_var_const_in);
307 ir_variable *var = actual->variable_referenced();
308 if (var) {
309 if ((var->data.mode == ir_var_auto ||
310 var->data.mode == ir_var_shader_out) &&
311 !var->data.assigned &&
312 !is_gl_identifier(var->name)) {
313 _mesa_glsl_warning(&loc, state, "`%s' used uninitialized",
314 var->name);
315 }
316 }
317 }
318
319 if (formal->type->is_image() &&
320 actual->variable_referenced()) {
321 if (!verify_image_parameter(&loc, state, formal,
322 actual->variable_referenced()))
323 return false;
324 }
325
326 actual_ir_node = actual_ir_node->next;
327 actual_ast_node = actual_ast_node->next;
328 }
329
330 /* The first parameter of atomic functions must be a buffer variable */
331 const char *func_name = sig->function_name();
332 bool is_atomic = is_atomic_function(func_name);
333 if (is_atomic) {
334 const ir_rvalue *const actual =
335 (ir_rvalue *) actual_ir_parameters.get_head_raw();
336
337 const ast_expression *const actual_ast =
338 exec_node_data(ast_expression,
339 actual_ast_parameters.get_head_raw(), link);
340 YYLTYPE loc = actual_ast->get_location();
341
342 if (!verify_first_atomic_parameter(&loc, state,
343 actual->variable_referenced())) {
344 return false;
345 }
346 }
347
348 return true;
349 }
350
351 struct copy_index_deref_data {
352 void *mem_ctx;
353 exec_list *before_instructions;
354 };
355
356 static void
357 copy_index_derefs_to_temps(ir_instruction *ir, void *data)
358 {
359 struct copy_index_deref_data *d = (struct copy_index_deref_data *)data;
360
361 if (ir->ir_type == ir_type_dereference_array) {
362 ir_dereference_array *a = (ir_dereference_array *) ir;
363 ir = a->array->as_dereference();
364
365 ir_rvalue *idx = a->array_index;
366 ir_variable *var = idx->variable_referenced();
367
368 /* If the index is read only it cannot change so there is no need
369 * to copy it.
370 */
371 if (!var || var->data.read_only || var->data.memory_read_only)
372 return;
373
374 ir_variable *tmp = new(d->mem_ctx) ir_variable(idx->type, "idx_tmp",
375 ir_var_temporary);
376 d->before_instructions->push_tail(tmp);
377
378 ir_dereference_variable *const deref_tmp_1 =
379 new(d->mem_ctx) ir_dereference_variable(tmp);
380 ir_assignment *const assignment =
381 new(d->mem_ctx) ir_assignment(deref_tmp_1,
382 idx->clone(d->mem_ctx, NULL));
383 d->before_instructions->push_tail(assignment);
384
385 /* Replace the array index with a dereference of the new temporary */
386 ir_dereference_variable *const deref_tmp_2 =
387 new(d->mem_ctx) ir_dereference_variable(tmp);
388 a->array_index = deref_tmp_2;
389 }
390 }
391
392 static void
393 fix_parameter(void *mem_ctx, ir_rvalue *actual, const glsl_type *formal_type,
394 exec_list *before_instructions, exec_list *after_instructions,
395 bool parameter_is_inout)
396 {
397 ir_expression *const expr = actual->as_expression();
398
399 /* If the types match exactly and the parameter is not a vector-extract,
400 * nothing needs to be done to fix the parameter.
401 */
402 if (formal_type == actual->type
403 && (expr == NULL || expr->operation != ir_binop_vector_extract)
404 && actual->as_dereference_variable())
405 return;
406
407 /* An array index could also be an out variable so we need to make a copy
408 * of them before the function is called.
409 */
410 if (!actual->as_dereference_variable()) {
411 struct copy_index_deref_data data;
412 data.mem_ctx = mem_ctx;
413 data.before_instructions = before_instructions;
414
415 visit_tree(actual, copy_index_derefs_to_temps, &data);
416 }
417
418 /* To convert an out parameter, we need to create a temporary variable to
419 * hold the value before conversion, and then perform the conversion after
420 * the function call returns.
421 *
422 * This has the effect of transforming code like this:
423 *
424 * void f(out int x);
425 * float value;
426 * f(value);
427 *
428 * Into IR that's equivalent to this:
429 *
430 * void f(out int x);
431 * float value;
432 * int out_parameter_conversion;
433 * f(out_parameter_conversion);
434 * value = float(out_parameter_conversion);
435 *
436 * If the parameter is an ir_expression of ir_binop_vector_extract,
437 * additional conversion is needed in the post-call re-write.
438 */
439 ir_variable *tmp =
440 new(mem_ctx) ir_variable(formal_type, "inout_tmp", ir_var_temporary);
441
442 before_instructions->push_tail(tmp);
443
444 /* If the parameter is an inout parameter, copy the value of the actual
445 * parameter to the new temporary. Note that no type conversion is allowed
446 * here because inout parameters must match types exactly.
447 */
448 if (parameter_is_inout) {
449 /* Inout parameters should never require conversion, since that would
450 * require an implicit conversion to exist both to and from the formal
451 * parameter type, and there are no bidirectional implicit conversions.
452 */
453 assert (actual->type == formal_type);
454
455 ir_dereference_variable *const deref_tmp_1 =
456 new(mem_ctx) ir_dereference_variable(tmp);
457 ir_assignment *const assignment =
458 new(mem_ctx) ir_assignment(deref_tmp_1, actual->clone(mem_ctx, NULL));
459 before_instructions->push_tail(assignment);
460 }
461
462 /* Replace the parameter in the call with a dereference of the new
463 * temporary.
464 */
465 ir_dereference_variable *const deref_tmp_2 =
466 new(mem_ctx) ir_dereference_variable(tmp);
467 actual->replace_with(deref_tmp_2);
468
469
470 /* Copy the temporary variable to the actual parameter with optional
471 * type conversion applied.
472 */
473 ir_rvalue *rhs = new(mem_ctx) ir_dereference_variable(tmp);
474 if (actual->type != formal_type)
475 rhs = convert_component(rhs, actual->type);
476
477 ir_rvalue *lhs = actual;
478 if (expr != NULL && expr->operation == ir_binop_vector_extract) {
479 lhs = new(mem_ctx) ir_dereference_array(expr->operands[0]->clone(mem_ctx,
480 NULL),
481 expr->operands[1]->clone(mem_ctx,
482 NULL));
483 }
484
485 ir_assignment *const assignment_2 = new(mem_ctx) ir_assignment(lhs, rhs);
486 after_instructions->push_tail(assignment_2);
487 }
488
489 /**
490 * Generate a function call.
491 *
492 * For non-void functions, this returns a dereference of the temporary
493 * variable which stores the return value for the call. For void functions,
494 * this returns NULL.
495 */
496 static ir_rvalue *
497 generate_call(exec_list *instructions, ir_function_signature *sig,
498 exec_list *actual_parameters,
499 ir_variable *sub_var,
500 ir_rvalue *array_idx,
501 struct _mesa_glsl_parse_state *state)
502 {
503 void *ctx = state;
504 exec_list post_call_conversions;
505
506 /* Perform implicit conversion of arguments. For out parameters, we need
507 * to place them in a temporary variable and do the conversion after the
508 * call takes place. Since we haven't emitted the call yet, we'll place
509 * the post-call conversions in a temporary exec_list, and emit them later.
510 */
511 foreach_two_lists(formal_node, &sig->parameters,
512 actual_node, actual_parameters) {
513 ir_rvalue *actual = (ir_rvalue *) actual_node;
514 ir_variable *formal = (ir_variable *) formal_node;
515
516 if (formal->type->is_numeric() || formal->type->is_boolean()) {
517 switch (formal->data.mode) {
518 case ir_var_const_in:
519 case ir_var_function_in: {
520 ir_rvalue *converted
521 = convert_component(actual, formal->type);
522 actual->replace_with(converted);
523 break;
524 }
525 case ir_var_function_out:
526 case ir_var_function_inout:
527 fix_parameter(ctx, actual, formal->type,
528 instructions, &post_call_conversions,
529 formal->data.mode == ir_var_function_inout);
530 break;
531 default:
532 assert (!"Illegal formal parameter mode");
533 break;
534 }
535 }
536 }
537
538 /* Section 4.3.2 (Const) of the GLSL 1.10.59 spec says:
539 *
540 * "Initializers for const declarations must be formed from literal
541 * values, other const variables (not including function call
542 * paramaters), or expressions of these.
543 *
544 * Constructors may be used in such expressions, but function calls may
545 * not."
546 *
547 * Section 4.3.3 (Constant Expressions) of the GLSL 1.20.8 spec says:
548 *
549 * "A constant expression is one of
550 *
551 * ...
552 *
553 * - a built-in function call whose arguments are all constant
554 * expressions, with the exception of the texture lookup
555 * functions, the noise functions, and ftransform. The built-in
556 * functions dFdx, dFdy, and fwidth must return 0 when evaluated
557 * inside an initializer with an argument that is a constant
558 * expression."
559 *
560 * Section 5.10 (Constant Expressions) of the GLSL ES 1.00.17 spec says:
561 *
562 * "A constant expression is one of
563 *
564 * ...
565 *
566 * - a built-in function call whose arguments are all constant
567 * expressions, with the exception of the texture lookup
568 * functions."
569 *
570 * Section 4.3.3 (Constant Expressions) of the GLSL ES 3.00.4 spec says:
571 *
572 * "A constant expression is one of
573 *
574 * ...
575 *
576 * - a built-in function call whose arguments are all constant
577 * expressions, with the exception of the texture lookup
578 * functions. The built-in functions dFdx, dFdy, and fwidth must
579 * return 0 when evaluated inside an initializer with an argument
580 * that is a constant expression."
581 *
582 * If the function call is a constant expression, don't generate any
583 * instructions; just generate an ir_constant.
584 */
585 if (state->is_version(120, 100) ||
586 state->ctx->Const.AllowGLSLBuiltinConstantExpression) {
587 ir_constant *value = sig->constant_expression_value(ctx,
588 actual_parameters,
589 NULL);
590 if (value != NULL) {
591 return value;
592 }
593 }
594
595 ir_dereference_variable *deref = NULL;
596 if (!sig->return_type->is_void()) {
597 /* Create a new temporary to hold the return value. */
598 char *const name = ir_variable::temporaries_allocate_names
599 ? ralloc_asprintf(ctx, "%s_retval", sig->function_name())
600 : NULL;
601
602 ir_variable *var;
603
604 var = new(ctx) ir_variable(sig->return_type, name, ir_var_temporary);
605 instructions->push_tail(var);
606
607 ralloc_free(name);
608
609 deref = new(ctx) ir_dereference_variable(var);
610 }
611
612 ir_call *call = new(ctx) ir_call(sig, deref,
613 actual_parameters, sub_var, array_idx);
614 instructions->push_tail(call);
615
616 /* Also emit any necessary out-parameter conversions. */
617 instructions->append_list(&post_call_conversions);
618
619 return deref ? deref->clone(ctx, NULL) : NULL;
620 }
621
622 /**
623 * Given a function name and parameter list, find the matching signature.
624 */
625 static ir_function_signature *
626 match_function_by_name(const char *name,
627 exec_list *actual_parameters,
628 struct _mesa_glsl_parse_state *state)
629 {
630 ir_function *f = state->symbols->get_function(name);
631 ir_function_signature *local_sig = NULL;
632 ir_function_signature *sig = NULL;
633
634 /* Is the function hidden by a record type constructor? */
635 if (state->symbols->get_type(name))
636 return sig; /* no match */
637
638 /* Is the function hidden by a variable (impossible in 1.10)? */
639 if (!state->symbols->separate_function_namespace
640 && state->symbols->get_variable(name))
641 return sig; /* no match */
642
643 if (f != NULL) {
644 /* In desktop GL, the presence of a user-defined signature hides any
645 * built-in signatures, so we must ignore them. In contrast, in ES2
646 * user-defined signatures add new overloads, so we must consider them.
647 */
648 bool allow_builtins = state->es_shader || !f->has_user_signature();
649
650 /* Look for a match in the local shader. If exact, we're done. */
651 bool is_exact = false;
652 sig = local_sig = f->matching_signature(state, actual_parameters,
653 allow_builtins, &is_exact);
654 if (is_exact)
655 return sig;
656
657 if (!allow_builtins)
658 return sig;
659 }
660
661 /* Local shader has no exact candidates; check the built-ins. */
662 sig = _mesa_glsl_find_builtin_function(state, name, actual_parameters);
663
664 /* if _mesa_glsl_find_builtin_function failed, fall back to the result
665 * of choose_best_inexact_overload() instead. This should only affect
666 * GLES.
667 */
668 return sig ? sig : local_sig;
669 }
670
671 static ir_function_signature *
672 match_subroutine_by_name(const char *name,
673 exec_list *actual_parameters,
674 struct _mesa_glsl_parse_state *state,
675 ir_variable **var_r)
676 {
677 void *ctx = state;
678 ir_function_signature *sig = NULL;
679 ir_function *f, *found = NULL;
680 const char *new_name;
681 ir_variable *var;
682 bool is_exact = false;
683
684 new_name =
685 ralloc_asprintf(ctx, "%s_%s",
686 _mesa_shader_stage_to_subroutine_prefix(state->stage),
687 name);
688 var = state->symbols->get_variable(new_name);
689 if (!var)
690 return NULL;
691
692 for (int i = 0; i < state->num_subroutine_types; i++) {
693 f = state->subroutine_types[i];
694 if (strcmp(f->name, var->type->without_array()->name))
695 continue;
696 found = f;
697 break;
698 }
699
700 if (!found)
701 return NULL;
702 *var_r = var;
703 sig = found->matching_signature(state, actual_parameters,
704 false, &is_exact);
705 return sig;
706 }
707
708 static ir_rvalue *
709 generate_array_index(void *mem_ctx, exec_list *instructions,
710 struct _mesa_glsl_parse_state *state, YYLTYPE loc,
711 const ast_expression *array, ast_expression *idx,
712 const char **function_name, exec_list *actual_parameters)
713 {
714 if (array->oper == ast_array_index) {
715 /* This handles arrays of arrays */
716 ir_rvalue *outer_array = generate_array_index(mem_ctx, instructions,
717 state, loc,
718 array->subexpressions[0],
719 array->subexpressions[1],
720 function_name,
721 actual_parameters);
722 ir_rvalue *outer_array_idx = idx->hir(instructions, state);
723
724 YYLTYPE index_loc = idx->get_location();
725 return _mesa_ast_array_index_to_hir(mem_ctx, state, outer_array,
726 outer_array_idx, loc,
727 index_loc);
728 } else {
729 ir_variable *sub_var = NULL;
730 *function_name = array->primary_expression.identifier;
731
732 if (!match_subroutine_by_name(*function_name, actual_parameters,
733 state, &sub_var)) {
734 _mesa_glsl_error(&loc, state, "Unknown subroutine `%s'",
735 *function_name);
736 *function_name = NULL; /* indicate error condition to caller */
737 return NULL;
738 }
739
740 ir_rvalue *outer_array_idx = idx->hir(instructions, state);
741 return new(mem_ctx) ir_dereference_array(sub_var, outer_array_idx);
742 }
743 }
744
745 static bool
746 function_exists(_mesa_glsl_parse_state *state,
747 struct glsl_symbol_table *symbols, const char *name)
748 {
749 ir_function *f = symbols->get_function(name);
750 if (f != NULL) {
751 foreach_in_list(ir_function_signature, sig, &f->signatures) {
752 if (sig->is_builtin() && !sig->is_builtin_available(state))
753 continue;
754 return true;
755 }
756 }
757 return false;
758 }
759
760 static void
761 print_function_prototypes(_mesa_glsl_parse_state *state, YYLTYPE *loc,
762 ir_function *f)
763 {
764 if (f == NULL)
765 return;
766
767 foreach_in_list(ir_function_signature, sig, &f->signatures) {
768 if (sig->is_builtin() && !sig->is_builtin_available(state))
769 continue;
770
771 char *str = prototype_string(sig->return_type, f->name,
772 &sig->parameters);
773 _mesa_glsl_error(loc, state, " %s", str);
774 ralloc_free(str);
775 }
776 }
777
778 /**
779 * Raise a "no matching function" error, listing all possible overloads the
780 * compiler considered so developers can figure out what went wrong.
781 */
782 static void
783 no_matching_function_error(const char *name,
784 YYLTYPE *loc,
785 exec_list *actual_parameters,
786 _mesa_glsl_parse_state *state)
787 {
788 gl_shader *sh = _mesa_glsl_get_builtin_function_shader();
789
790 if (!function_exists(state, state->symbols, name)
791 && (!state->uses_builtin_functions
792 || !function_exists(state, sh->symbols, name))) {
793 _mesa_glsl_error(loc, state, "no function with name '%s'", name);
794 } else {
795 char *str = prototype_string(NULL, name, actual_parameters);
796 _mesa_glsl_error(loc, state,
797 "no matching function for call to `%s';"
798 " candidates are:",
799 str);
800 ralloc_free(str);
801
802 print_function_prototypes(state, loc,
803 state->symbols->get_function(name));
804
805 if (state->uses_builtin_functions) {
806 print_function_prototypes(state, loc,
807 sh->symbols->get_function(name));
808 }
809 }
810 }
811
812 /**
813 * Perform automatic type conversion of constructor parameters
814 *
815 * This implements the rules in the "Conversion and Scalar Constructors"
816 * section (GLSL 1.10 section 5.4.1), not the "Implicit Conversions" rules.
817 */
818 static ir_rvalue *
819 convert_component(ir_rvalue *src, const glsl_type *desired_type)
820 {
821 void *ctx = ralloc_parent(src);
822 const unsigned a = desired_type->base_type;
823 const unsigned b = src->type->base_type;
824 ir_expression *result = NULL;
825
826 if (src->type->is_error())
827 return src;
828
829 assert(a <= GLSL_TYPE_IMAGE);
830 assert(b <= GLSL_TYPE_IMAGE);
831
832 if (a == b)
833 return src;
834
835 switch (a) {
836 case GLSL_TYPE_UINT:
837 switch (b) {
838 case GLSL_TYPE_INT:
839 result = new(ctx) ir_expression(ir_unop_i2u, src);
840 break;
841 case GLSL_TYPE_FLOAT:
842 result = new(ctx) ir_expression(ir_unop_f2u, src);
843 break;
844 case GLSL_TYPE_BOOL:
845 result = new(ctx) ir_expression(ir_unop_i2u,
846 new(ctx) ir_expression(ir_unop_b2i,
847 src));
848 break;
849 case GLSL_TYPE_DOUBLE:
850 result = new(ctx) ir_expression(ir_unop_d2u, src);
851 break;
852 case GLSL_TYPE_UINT64:
853 result = new(ctx) ir_expression(ir_unop_u642u, src);
854 break;
855 case GLSL_TYPE_INT64:
856 result = new(ctx) ir_expression(ir_unop_i642u, src);
857 break;
858 case GLSL_TYPE_SAMPLER:
859 result = new(ctx) ir_expression(ir_unop_unpack_sampler_2x32, src);
860 break;
861 case GLSL_TYPE_IMAGE:
862 result = new(ctx) ir_expression(ir_unop_unpack_image_2x32, src);
863 break;
864 }
865 break;
866 case GLSL_TYPE_INT:
867 switch (b) {
868 case GLSL_TYPE_UINT:
869 result = new(ctx) ir_expression(ir_unop_u2i, src);
870 break;
871 case GLSL_TYPE_FLOAT:
872 result = new(ctx) ir_expression(ir_unop_f2i, src);
873 break;
874 case GLSL_TYPE_BOOL:
875 result = new(ctx) ir_expression(ir_unop_b2i, src);
876 break;
877 case GLSL_TYPE_DOUBLE:
878 result = new(ctx) ir_expression(ir_unop_d2i, src);
879 break;
880 case GLSL_TYPE_UINT64:
881 result = new(ctx) ir_expression(ir_unop_u642i, src);
882 break;
883 case GLSL_TYPE_INT64:
884 result = new(ctx) ir_expression(ir_unop_i642i, src);
885 break;
886 }
887 break;
888 case GLSL_TYPE_FLOAT:
889 switch (b) {
890 case GLSL_TYPE_UINT:
891 result = new(ctx) ir_expression(ir_unop_u2f, desired_type, src, NULL);
892 break;
893 case GLSL_TYPE_INT:
894 result = new(ctx) ir_expression(ir_unop_i2f, desired_type, src, NULL);
895 break;
896 case GLSL_TYPE_BOOL:
897 result = new(ctx) ir_expression(ir_unop_b2f, desired_type, src, NULL);
898 break;
899 case GLSL_TYPE_DOUBLE:
900 result = new(ctx) ir_expression(ir_unop_d2f, desired_type, src, NULL);
901 break;
902 case GLSL_TYPE_UINT64:
903 result = new(ctx) ir_expression(ir_unop_u642f, desired_type, src, NULL);
904 break;
905 case GLSL_TYPE_INT64:
906 result = new(ctx) ir_expression(ir_unop_i642f, desired_type, src, NULL);
907 break;
908 }
909 break;
910 case GLSL_TYPE_BOOL:
911 switch (b) {
912 case GLSL_TYPE_UINT:
913 result = new(ctx) ir_expression(ir_unop_i2b,
914 new(ctx) ir_expression(ir_unop_u2i,
915 src));
916 break;
917 case GLSL_TYPE_INT:
918 result = new(ctx) ir_expression(ir_unop_i2b, desired_type, src, NULL);
919 break;
920 case GLSL_TYPE_FLOAT:
921 result = new(ctx) ir_expression(ir_unop_f2b, desired_type, src, NULL);
922 break;
923 case GLSL_TYPE_DOUBLE:
924 result = new(ctx) ir_expression(ir_unop_d2b, desired_type, src, NULL);
925 break;
926 case GLSL_TYPE_UINT64:
927 result = new(ctx) ir_expression(ir_unop_i642b,
928 new(ctx) ir_expression(ir_unop_u642i64,
929 src));
930 break;
931 case GLSL_TYPE_INT64:
932 result = new(ctx) ir_expression(ir_unop_i642b, desired_type, src, NULL);
933 break;
934 }
935 break;
936 case GLSL_TYPE_DOUBLE:
937 switch (b) {
938 case GLSL_TYPE_INT:
939 result = new(ctx) ir_expression(ir_unop_i2d, src);
940 break;
941 case GLSL_TYPE_UINT:
942 result = new(ctx) ir_expression(ir_unop_u2d, src);
943 break;
944 case GLSL_TYPE_BOOL:
945 result = new(ctx) ir_expression(ir_unop_f2d,
946 new(ctx) ir_expression(ir_unop_b2f,
947 src));
948 break;
949 case GLSL_TYPE_FLOAT:
950 result = new(ctx) ir_expression(ir_unop_f2d, desired_type, src, NULL);
951 break;
952 case GLSL_TYPE_UINT64:
953 result = new(ctx) ir_expression(ir_unop_u642d, desired_type, src, NULL);
954 break;
955 case GLSL_TYPE_INT64:
956 result = new(ctx) ir_expression(ir_unop_i642d, desired_type, src, NULL);
957 break;
958 }
959 break;
960 case GLSL_TYPE_UINT64:
961 switch (b) {
962 case GLSL_TYPE_INT:
963 result = new(ctx) ir_expression(ir_unop_i2u64, src);
964 break;
965 case GLSL_TYPE_UINT:
966 result = new(ctx) ir_expression(ir_unop_u2u64, src);
967 break;
968 case GLSL_TYPE_BOOL:
969 result = new(ctx) ir_expression(ir_unop_i642u64,
970 new(ctx) ir_expression(ir_unop_b2i64,
971 src));
972 break;
973 case GLSL_TYPE_FLOAT:
974 result = new(ctx) ir_expression(ir_unop_f2u64, src);
975 break;
976 case GLSL_TYPE_DOUBLE:
977 result = new(ctx) ir_expression(ir_unop_d2u64, src);
978 break;
979 case GLSL_TYPE_INT64:
980 result = new(ctx) ir_expression(ir_unop_i642u64, src);
981 break;
982 }
983 break;
984 case GLSL_TYPE_INT64:
985 switch (b) {
986 case GLSL_TYPE_INT:
987 result = new(ctx) ir_expression(ir_unop_i2i64, src);
988 break;
989 case GLSL_TYPE_UINT:
990 result = new(ctx) ir_expression(ir_unop_u2i64, src);
991 break;
992 case GLSL_TYPE_BOOL:
993 result = new(ctx) ir_expression(ir_unop_b2i64, src);
994 break;
995 case GLSL_TYPE_FLOAT:
996 result = new(ctx) ir_expression(ir_unop_f2i64, src);
997 break;
998 case GLSL_TYPE_DOUBLE:
999 result = new(ctx) ir_expression(ir_unop_d2i64, src);
1000 break;
1001 case GLSL_TYPE_UINT64:
1002 result = new(ctx) ir_expression(ir_unop_u642i64, src);
1003 break;
1004 }
1005 break;
1006 case GLSL_TYPE_SAMPLER:
1007 switch (b) {
1008 case GLSL_TYPE_UINT:
1009 result = new(ctx)
1010 ir_expression(ir_unop_pack_sampler_2x32, desired_type, src);
1011 break;
1012 }
1013 break;
1014 case GLSL_TYPE_IMAGE:
1015 switch (b) {
1016 case GLSL_TYPE_UINT:
1017 result = new(ctx)
1018 ir_expression(ir_unop_pack_image_2x32, desired_type, src);
1019 break;
1020 }
1021 break;
1022 }
1023
1024 assert(result != NULL);
1025 assert(result->type == desired_type);
1026
1027 /* Try constant folding; it may fold in the conversion we just added. */
1028 ir_constant *const constant = result->constant_expression_value(ctx);
1029 return (constant != NULL) ? (ir_rvalue *) constant : (ir_rvalue *) result;
1030 }
1031
1032
1033 /**
1034 * Perform automatic type and constant conversion of constructor parameters
1035 *
1036 * This implements the rules in the "Implicit Conversions" rules, not the
1037 * "Conversion and Scalar Constructors".
1038 *
1039 * After attempting the implicit conversion, an attempt to convert into a
1040 * constant valued expression is also done.
1041 *
1042 * The \c from \c ir_rvalue is converted "in place".
1043 *
1044 * \param from Operand that is being converted
1045 * \param to Base type the operand will be converted to
1046 * \param state GLSL compiler state
1047 *
1048 * \return
1049 * If the attempt to convert into a constant expression succeeds, \c true is
1050 * returned. Otherwise \c false is returned.
1051 */
1052 static bool
1053 implicitly_convert_component(ir_rvalue * &from, const glsl_base_type to,
1054 struct _mesa_glsl_parse_state *state)
1055 {
1056 void *mem_ctx = state;
1057 ir_rvalue *result = from;
1058
1059 if (to != from->type->base_type) {
1060 const glsl_type *desired_type =
1061 glsl_type::get_instance(to,
1062 from->type->vector_elements,
1063 from->type->matrix_columns);
1064
1065 if (from->type->can_implicitly_convert_to(desired_type, state)) {
1066 /* Even though convert_component() implements the constructor
1067 * conversion rules (not the implicit conversion rules), its safe
1068 * to use it here because we already checked that the implicit
1069 * conversion is legal.
1070 */
1071 result = convert_component(from, desired_type);
1072 }
1073 }
1074
1075 ir_rvalue *const constant = result->constant_expression_value(mem_ctx);
1076
1077 if (constant != NULL)
1078 result = constant;
1079
1080 if (from != result) {
1081 from->replace_with(result);
1082 from = result;
1083 }
1084
1085 return constant != NULL;
1086 }
1087
1088
1089 /**
1090 * Dereference a specific component from a scalar, vector, or matrix
1091 */
1092 static ir_rvalue *
1093 dereference_component(ir_rvalue *src, unsigned component)
1094 {
1095 void *ctx = ralloc_parent(src);
1096 assert(component < src->type->components());
1097
1098 /* If the source is a constant, just create a new constant instead of a
1099 * dereference of the existing constant.
1100 */
1101 ir_constant *constant = src->as_constant();
1102 if (constant)
1103 return new(ctx) ir_constant(constant, component);
1104
1105 if (src->type->is_scalar()) {
1106 return src;
1107 } else if (src->type->is_vector()) {
1108 return new(ctx) ir_swizzle(src, component, 0, 0, 0, 1);
1109 } else {
1110 assert(src->type->is_matrix());
1111
1112 /* Dereference a row of the matrix, then call this function again to get
1113 * a specific element from that row.
1114 */
1115 const int c = component / src->type->column_type()->vector_elements;
1116 const int r = component % src->type->column_type()->vector_elements;
1117 ir_constant *const col_index = new(ctx) ir_constant(c);
1118 ir_dereference *const col = new(ctx) ir_dereference_array(src,
1119 col_index);
1120
1121 col->type = src->type->column_type();
1122
1123 return dereference_component(col, r);
1124 }
1125
1126 assert(!"Should not get here.");
1127 return NULL;
1128 }
1129
1130
1131 static ir_rvalue *
1132 process_vec_mat_constructor(exec_list *instructions,
1133 const glsl_type *constructor_type,
1134 YYLTYPE *loc, exec_list *parameters,
1135 struct _mesa_glsl_parse_state *state)
1136 {
1137 void *ctx = state;
1138
1139 /* The ARB_shading_language_420pack spec says:
1140 *
1141 * "If an initializer is a list of initializers enclosed in curly braces,
1142 * the variable being declared must be a vector, a matrix, an array, or a
1143 * structure.
1144 *
1145 * int i = { 1 }; // illegal, i is not an aggregate"
1146 */
1147 if (constructor_type->vector_elements <= 1) {
1148 _mesa_glsl_error(loc, state, "aggregates can only initialize vectors, "
1149 "matrices, arrays, and structs");
1150 return ir_rvalue::error_value(ctx);
1151 }
1152
1153 exec_list actual_parameters;
1154 const unsigned parameter_count =
1155 process_parameters(instructions, &actual_parameters, parameters, state);
1156
1157 if (parameter_count == 0
1158 || (constructor_type->is_vector() &&
1159 constructor_type->vector_elements != parameter_count)
1160 || (constructor_type->is_matrix() &&
1161 constructor_type->matrix_columns != parameter_count)) {
1162 _mesa_glsl_error(loc, state, "%s constructor must have %u parameters",
1163 constructor_type->is_vector() ? "vector" : "matrix",
1164 constructor_type->vector_elements);
1165 return ir_rvalue::error_value(ctx);
1166 }
1167
1168 bool all_parameters_are_constant = true;
1169
1170 /* Type cast each parameter and, if possible, fold constants. */
1171 foreach_in_list_safe(ir_rvalue, ir, &actual_parameters) {
1172 /* Apply implicit conversions (not the scalar constructor rules, see the
1173 * spec quote above!) and attempt to convert the parameter to a constant
1174 * valued expression. After doing so, track whether or not all the
1175 * parameters to the constructor are trivially constant valued
1176 * expressions.
1177 */
1178 all_parameters_are_constant &=
1179 implicitly_convert_component(ir, constructor_type->base_type, state);
1180
1181 if (constructor_type->is_matrix()) {
1182 if (ir->type != constructor_type->column_type()) {
1183 _mesa_glsl_error(loc, state, "type error in matrix constructor: "
1184 "expected: %s, found %s",
1185 constructor_type->column_type()->name,
1186 ir->type->name);
1187 return ir_rvalue::error_value(ctx);
1188 }
1189 } else if (ir->type != constructor_type->get_scalar_type()) {
1190 _mesa_glsl_error(loc, state, "type error in vector constructor: "
1191 "expected: %s, found %s",
1192 constructor_type->get_scalar_type()->name,
1193 ir->type->name);
1194 return ir_rvalue::error_value(ctx);
1195 }
1196 }
1197
1198 if (all_parameters_are_constant)
1199 return new(ctx) ir_constant(constructor_type, &actual_parameters);
1200
1201 ir_variable *var = new(ctx) ir_variable(constructor_type, "vec_mat_ctor",
1202 ir_var_temporary);
1203 instructions->push_tail(var);
1204
1205 int i = 0;
1206
1207 foreach_in_list(ir_rvalue, rhs, &actual_parameters) {
1208 ir_instruction *assignment = NULL;
1209
1210 if (var->type->is_matrix()) {
1211 ir_rvalue *lhs =
1212 new(ctx) ir_dereference_array(var, new(ctx) ir_constant(i));
1213 assignment = new(ctx) ir_assignment(lhs, rhs);
1214 } else {
1215 /* use writemask rather than index for vector */
1216 assert(var->type->is_vector());
1217 assert(i < 4);
1218 ir_dereference *lhs = new(ctx) ir_dereference_variable(var);
1219 assignment = new(ctx) ir_assignment(lhs, rhs, NULL,
1220 (unsigned)(1 << i));
1221 }
1222
1223 instructions->push_tail(assignment);
1224
1225 i++;
1226 }
1227
1228 return new(ctx) ir_dereference_variable(var);
1229 }
1230
1231
1232 static ir_rvalue *
1233 process_array_constructor(exec_list *instructions,
1234 const glsl_type *constructor_type,
1235 YYLTYPE *loc, exec_list *parameters,
1236 struct _mesa_glsl_parse_state *state)
1237 {
1238 void *ctx = state;
1239 /* Array constructors come in two forms: sized and unsized. Sized array
1240 * constructors look like 'vec4[2](a, b)', where 'a' and 'b' are vec4
1241 * variables. In this case the number of parameters must exactly match the
1242 * specified size of the array.
1243 *
1244 * Unsized array constructors look like 'vec4[](a, b)', where 'a' and 'b'
1245 * are vec4 variables. In this case the size of the array being constructed
1246 * is determined by the number of parameters.
1247 *
1248 * From page 52 (page 58 of the PDF) of the GLSL 1.50 spec:
1249 *
1250 * "There must be exactly the same number of arguments as the size of
1251 * the array being constructed. If no size is present in the
1252 * constructor, then the array is explicitly sized to the number of
1253 * arguments provided. The arguments are assigned in order, starting at
1254 * element 0, to the elements of the constructed array. Each argument
1255 * must be the same type as the element type of the array, or be a type
1256 * that can be converted to the element type of the array according to
1257 * Section 4.1.10 "Implicit Conversions.""
1258 */
1259 exec_list actual_parameters;
1260 const unsigned parameter_count =
1261 process_parameters(instructions, &actual_parameters, parameters, state);
1262 bool is_unsized_array = constructor_type->is_unsized_array();
1263
1264 if ((parameter_count == 0) ||
1265 (!is_unsized_array && (constructor_type->length != parameter_count))) {
1266 const unsigned min_param = is_unsized_array
1267 ? 1 : constructor_type->length;
1268
1269 _mesa_glsl_error(loc, state, "array constructor must have %s %u "
1270 "parameter%s",
1271 is_unsized_array ? "at least" : "exactly",
1272 min_param, (min_param <= 1) ? "" : "s");
1273 return ir_rvalue::error_value(ctx);
1274 }
1275
1276 if (is_unsized_array) {
1277 constructor_type =
1278 glsl_type::get_array_instance(constructor_type->fields.array,
1279 parameter_count);
1280 assert(constructor_type != NULL);
1281 assert(constructor_type->length == parameter_count);
1282 }
1283
1284 bool all_parameters_are_constant = true;
1285 const glsl_type *element_type = constructor_type->fields.array;
1286
1287 /* Type cast each parameter and, if possible, fold constants. */
1288 foreach_in_list_safe(ir_rvalue, ir, &actual_parameters) {
1289 /* Apply implicit conversions (not the scalar constructor rules, see the
1290 * spec quote above!) and attempt to convert the parameter to a constant
1291 * valued expression. After doing so, track whether or not all the
1292 * parameters to the constructor are trivially constant valued
1293 * expressions.
1294 */
1295 all_parameters_are_constant &=
1296 implicitly_convert_component(ir, element_type->base_type, state);
1297
1298 if (constructor_type->fields.array->is_unsized_array()) {
1299 /* As the inner parameters of the constructor are created without
1300 * knowledge of each other we need to check to make sure unsized
1301 * parameters of unsized constructors all end up with the same size.
1302 *
1303 * e.g we make sure to fail for a constructor like this:
1304 * vec4[][] a = vec4[][](vec4[](vec4(0.0), vec4(1.0)),
1305 * vec4[](vec4(0.0), vec4(1.0), vec4(1.0)),
1306 * vec4[](vec4(0.0), vec4(1.0)));
1307 */
1308 if (element_type->is_unsized_array()) {
1309 /* This is the first parameter so just get the type */
1310 element_type = ir->type;
1311 } else if (element_type != ir->type) {
1312 _mesa_glsl_error(loc, state, "type error in array constructor: "
1313 "expected: %s, found %s",
1314 element_type->name,
1315 ir->type->name);
1316 return ir_rvalue::error_value(ctx);
1317 }
1318 } else if (ir->type != constructor_type->fields.array) {
1319 _mesa_glsl_error(loc, state, "type error in array constructor: "
1320 "expected: %s, found %s",
1321 constructor_type->fields.array->name,
1322 ir->type->name);
1323 return ir_rvalue::error_value(ctx);
1324 } else {
1325 element_type = ir->type;
1326 }
1327 }
1328
1329 if (constructor_type->fields.array->is_unsized_array()) {
1330 constructor_type =
1331 glsl_type::get_array_instance(element_type,
1332 parameter_count);
1333 assert(constructor_type != NULL);
1334 assert(constructor_type->length == parameter_count);
1335 }
1336
1337 if (all_parameters_are_constant)
1338 return new(ctx) ir_constant(constructor_type, &actual_parameters);
1339
1340 ir_variable *var = new(ctx) ir_variable(constructor_type, "array_ctor",
1341 ir_var_temporary);
1342 instructions->push_tail(var);
1343
1344 int i = 0;
1345 foreach_in_list(ir_rvalue, rhs, &actual_parameters) {
1346 ir_rvalue *lhs = new(ctx) ir_dereference_array(var,
1347 new(ctx) ir_constant(i));
1348
1349 ir_instruction *assignment = new(ctx) ir_assignment(lhs, rhs);
1350 instructions->push_tail(assignment);
1351
1352 i++;
1353 }
1354
1355 return new(ctx) ir_dereference_variable(var);
1356 }
1357
1358
1359 /**
1360 * Determine if a list consists of a single scalar r-value
1361 */
1362 static bool
1363 single_scalar_parameter(exec_list *parameters)
1364 {
1365 const ir_rvalue *const p = (ir_rvalue *) parameters->get_head_raw();
1366 assert(((ir_rvalue *)p)->as_rvalue() != NULL);
1367
1368 return (p->type->is_scalar() && p->next->is_tail_sentinel());
1369 }
1370
1371
1372 /**
1373 * Generate inline code for a vector constructor
1374 *
1375 * The generated constructor code will consist of a temporary variable
1376 * declaration of the same type as the constructor. A sequence of assignments
1377 * from constructor parameters to the temporary will follow.
1378 *
1379 * \return
1380 * An \c ir_dereference_variable of the temprorary generated in the constructor
1381 * body.
1382 */
1383 static ir_rvalue *
1384 emit_inline_vector_constructor(const glsl_type *type,
1385 exec_list *instructions,
1386 exec_list *parameters,
1387 void *ctx)
1388 {
1389 assert(!parameters->is_empty());
1390
1391 ir_variable *var = new(ctx) ir_variable(type, "vec_ctor", ir_var_temporary);
1392 instructions->push_tail(var);
1393
1394 /* There are three kinds of vector constructors.
1395 *
1396 * - Construct a vector from a single scalar by replicating that scalar to
1397 * all components of the vector.
1398 *
1399 * - Construct a vector from at least a matrix. This case should already
1400 * have been taken care of in ast_function_expression::hir by breaking
1401 * down the matrix into a series of column vectors.
1402 *
1403 * - Construct a vector from an arbirary combination of vectors and
1404 * scalars. The components of the constructor parameters are assigned
1405 * to the vector in order until the vector is full.
1406 */
1407 const unsigned lhs_components = type->components();
1408 if (single_scalar_parameter(parameters)) {
1409 ir_rvalue *first_param = (ir_rvalue *)parameters->get_head_raw();
1410 ir_rvalue *rhs = new(ctx) ir_swizzle(first_param, 0, 0, 0, 0,
1411 lhs_components);
1412 ir_dereference_variable *lhs = new(ctx) ir_dereference_variable(var);
1413 const unsigned mask = (1U << lhs_components) - 1;
1414
1415 assert(rhs->type == lhs->type);
1416
1417 ir_instruction *inst = new(ctx) ir_assignment(lhs, rhs, NULL, mask);
1418 instructions->push_tail(inst);
1419 } else {
1420 unsigned base_component = 0;
1421 unsigned base_lhs_component = 0;
1422 ir_constant_data data;
1423 unsigned constant_mask = 0, constant_components = 0;
1424
1425 memset(&data, 0, sizeof(data));
1426
1427 foreach_in_list(ir_rvalue, param, parameters) {
1428 unsigned rhs_components = param->type->components();
1429
1430 /* Do not try to assign more components to the vector than it has! */
1431 if ((rhs_components + base_lhs_component) > lhs_components) {
1432 rhs_components = lhs_components - base_lhs_component;
1433 }
1434
1435 const ir_constant *const c = param->as_constant();
1436 if (c != NULL) {
1437 for (unsigned i = 0; i < rhs_components; i++) {
1438 switch (c->type->base_type) {
1439 case GLSL_TYPE_UINT:
1440 data.u[i + base_component] = c->get_uint_component(i);
1441 break;
1442 case GLSL_TYPE_INT:
1443 data.i[i + base_component] = c->get_int_component(i);
1444 break;
1445 case GLSL_TYPE_FLOAT:
1446 data.f[i + base_component] = c->get_float_component(i);
1447 break;
1448 case GLSL_TYPE_DOUBLE:
1449 data.d[i + base_component] = c->get_double_component(i);
1450 break;
1451 case GLSL_TYPE_BOOL:
1452 data.b[i + base_component] = c->get_bool_component(i);
1453 break;
1454 case GLSL_TYPE_UINT64:
1455 data.u64[i + base_component] = c->get_uint64_component(i);
1456 break;
1457 case GLSL_TYPE_INT64:
1458 data.i64[i + base_component] = c->get_int64_component(i);
1459 break;
1460 default:
1461 assert(!"Should not get here.");
1462 break;
1463 }
1464 }
1465
1466 /* Mask of fields to be written in the assignment. */
1467 constant_mask |= ((1U << rhs_components) - 1) << base_lhs_component;
1468 constant_components += rhs_components;
1469
1470 base_component += rhs_components;
1471 }
1472 /* Advance the component index by the number of components
1473 * that were just assigned.
1474 */
1475 base_lhs_component += rhs_components;
1476 }
1477
1478 if (constant_mask != 0) {
1479 ir_dereference *lhs = new(ctx) ir_dereference_variable(var);
1480 const glsl_type *rhs_type =
1481 glsl_type::get_instance(var->type->base_type,
1482 constant_components,
1483 1);
1484 ir_rvalue *rhs = new(ctx) ir_constant(rhs_type, &data);
1485
1486 ir_instruction *inst =
1487 new(ctx) ir_assignment(lhs, rhs, NULL, constant_mask);
1488 instructions->push_tail(inst);
1489 }
1490
1491 base_component = 0;
1492 foreach_in_list(ir_rvalue, param, parameters) {
1493 unsigned rhs_components = param->type->components();
1494
1495 /* Do not try to assign more components to the vector than it has! */
1496 if ((rhs_components + base_component) > lhs_components) {
1497 rhs_components = lhs_components - base_component;
1498 }
1499
1500 /* If we do not have any components left to copy, break out of the
1501 * loop. This can happen when initializing a vec4 with a mat3 as the
1502 * mat3 would have been broken into a series of column vectors.
1503 */
1504 if (rhs_components == 0) {
1505 break;
1506 }
1507
1508 const ir_constant *const c = param->as_constant();
1509 if (c == NULL) {
1510 /* Mask of fields to be written in the assignment. */
1511 const unsigned write_mask = ((1U << rhs_components) - 1)
1512 << base_component;
1513
1514 ir_dereference *lhs = new(ctx) ir_dereference_variable(var);
1515
1516 /* Generate a swizzle so that LHS and RHS sizes match. */
1517 ir_rvalue *rhs =
1518 new(ctx) ir_swizzle(param, 0, 1, 2, 3, rhs_components);
1519
1520 ir_instruction *inst =
1521 new(ctx) ir_assignment(lhs, rhs, NULL, write_mask);
1522 instructions->push_tail(inst);
1523 }
1524
1525 /* Advance the component index by the number of components that were
1526 * just assigned.
1527 */
1528 base_component += rhs_components;
1529 }
1530 }
1531 return new(ctx) ir_dereference_variable(var);
1532 }
1533
1534
1535 /**
1536 * Generate assignment of a portion of a vector to a portion of a matrix column
1537 *
1538 * \param src_base First component of the source to be used in assignment
1539 * \param column Column of destination to be assiged
1540 * \param row_base First component of the destination column to be assigned
1541 * \param count Number of components to be assigned
1542 *
1543 * \note
1544 * \c src_base + \c count must be less than or equal to the number of
1545 * components in the source vector.
1546 */
1547 static ir_instruction *
1548 assign_to_matrix_column(ir_variable *var, unsigned column, unsigned row_base,
1549 ir_rvalue *src, unsigned src_base, unsigned count,
1550 void *mem_ctx)
1551 {
1552 ir_constant *col_idx = new(mem_ctx) ir_constant(column);
1553 ir_dereference *column_ref = new(mem_ctx) ir_dereference_array(var,
1554 col_idx);
1555
1556 assert(column_ref->type->components() >= (row_base + count));
1557 assert(src->type->components() >= (src_base + count));
1558
1559 /* Generate a swizzle that extracts the number of components from the source
1560 * that are to be assigned to the column of the matrix.
1561 */
1562 if (count < src->type->vector_elements) {
1563 src = new(mem_ctx) ir_swizzle(src,
1564 src_base + 0, src_base + 1,
1565 src_base + 2, src_base + 3,
1566 count);
1567 }
1568
1569 /* Mask of fields to be written in the assignment. */
1570 const unsigned write_mask = ((1U << count) - 1) << row_base;
1571
1572 return new(mem_ctx) ir_assignment(column_ref, src, NULL, write_mask);
1573 }
1574
1575
1576 /**
1577 * Generate inline code for a matrix constructor
1578 *
1579 * The generated constructor code will consist of a temporary variable
1580 * declaration of the same type as the constructor. A sequence of assignments
1581 * from constructor parameters to the temporary will follow.
1582 *
1583 * \return
1584 * An \c ir_dereference_variable of the temprorary generated in the constructor
1585 * body.
1586 */
1587 static ir_rvalue *
1588 emit_inline_matrix_constructor(const glsl_type *type,
1589 exec_list *instructions,
1590 exec_list *parameters,
1591 void *ctx)
1592 {
1593 assert(!parameters->is_empty());
1594
1595 ir_variable *var = new(ctx) ir_variable(type, "mat_ctor", ir_var_temporary);
1596 instructions->push_tail(var);
1597
1598 /* There are three kinds of matrix constructors.
1599 *
1600 * - Construct a matrix from a single scalar by replicating that scalar to
1601 * along the diagonal of the matrix and setting all other components to
1602 * zero.
1603 *
1604 * - Construct a matrix from an arbirary combination of vectors and
1605 * scalars. The components of the constructor parameters are assigned
1606 * to the matrix in column-major order until the matrix is full.
1607 *
1608 * - Construct a matrix from a single matrix. The source matrix is copied
1609 * to the upper left portion of the constructed matrix, and the remaining
1610 * elements take values from the identity matrix.
1611 */
1612 ir_rvalue *const first_param = (ir_rvalue *) parameters->get_head_raw();
1613 if (single_scalar_parameter(parameters)) {
1614 /* Assign the scalar to the X component of a vec4, and fill the remaining
1615 * components with zero.
1616 */
1617 glsl_base_type param_base_type = first_param->type->base_type;
1618 assert(first_param->type->is_float() || first_param->type->is_double());
1619 ir_variable *rhs_var =
1620 new(ctx) ir_variable(glsl_type::get_instance(param_base_type, 4, 1),
1621 "mat_ctor_vec",
1622 ir_var_temporary);
1623 instructions->push_tail(rhs_var);
1624
1625 ir_constant_data zero;
1626 for (unsigned i = 0; i < 4; i++)
1627 if (first_param->type->is_float())
1628 zero.f[i] = 0.0;
1629 else
1630 zero.d[i] = 0.0;
1631
1632 ir_instruction *inst =
1633 new(ctx) ir_assignment(new(ctx) ir_dereference_variable(rhs_var),
1634 new(ctx) ir_constant(rhs_var->type, &zero));
1635 instructions->push_tail(inst);
1636
1637 ir_dereference *const rhs_ref =
1638 new(ctx) ir_dereference_variable(rhs_var);
1639
1640 inst = new(ctx) ir_assignment(rhs_ref, first_param, NULL, 0x01);
1641 instructions->push_tail(inst);
1642
1643 /* Assign the temporary vector to each column of the destination matrix
1644 * with a swizzle that puts the X component on the diagonal of the
1645 * matrix. In some cases this may mean that the X component does not
1646 * get assigned into the column at all (i.e., when the matrix has more
1647 * columns than rows).
1648 */
1649 static const unsigned rhs_swiz[4][4] = {
1650 { 0, 1, 1, 1 },
1651 { 1, 0, 1, 1 },
1652 { 1, 1, 0, 1 },
1653 { 1, 1, 1, 0 }
1654 };
1655
1656 const unsigned cols_to_init = MIN2(type->matrix_columns,
1657 type->vector_elements);
1658 for (unsigned i = 0; i < cols_to_init; i++) {
1659 ir_constant *const col_idx = new(ctx) ir_constant(i);
1660 ir_rvalue *const col_ref = new(ctx) ir_dereference_array(var,
1661 col_idx);
1662
1663 ir_rvalue *const rhs_ref = new(ctx) ir_dereference_variable(rhs_var);
1664 ir_rvalue *const rhs = new(ctx) ir_swizzle(rhs_ref, rhs_swiz[i],
1665 type->vector_elements);
1666
1667 inst = new(ctx) ir_assignment(col_ref, rhs);
1668 instructions->push_tail(inst);
1669 }
1670
1671 for (unsigned i = cols_to_init; i < type->matrix_columns; i++) {
1672 ir_constant *const col_idx = new(ctx) ir_constant(i);
1673 ir_rvalue *const col_ref = new(ctx) ir_dereference_array(var,
1674 col_idx);
1675
1676 ir_rvalue *const rhs_ref = new(ctx) ir_dereference_variable(rhs_var);
1677 ir_rvalue *const rhs = new(ctx) ir_swizzle(rhs_ref, 1, 1, 1, 1,
1678 type->vector_elements);
1679
1680 inst = new(ctx) ir_assignment(col_ref, rhs);
1681 instructions->push_tail(inst);
1682 }
1683 } else if (first_param->type->is_matrix()) {
1684 /* From page 50 (56 of the PDF) of the GLSL 1.50 spec:
1685 *
1686 * "If a matrix is constructed from a matrix, then each component
1687 * (column i, row j) in the result that has a corresponding
1688 * component (column i, row j) in the argument will be initialized
1689 * from there. All other components will be initialized to the
1690 * identity matrix. If a matrix argument is given to a matrix
1691 * constructor, it is an error to have any other arguments."
1692 */
1693 assert(first_param->next->is_tail_sentinel());
1694 ir_rvalue *const src_matrix = first_param;
1695
1696 /* If the source matrix is smaller, pre-initialize the relavent parts of
1697 * the destination matrix to the identity matrix.
1698 */
1699 if ((src_matrix->type->matrix_columns < var->type->matrix_columns) ||
1700 (src_matrix->type->vector_elements < var->type->vector_elements)) {
1701
1702 /* If the source matrix has fewer rows, every column of the
1703 * destination must be initialized. Otherwise only the columns in
1704 * the destination that do not exist in the source must be
1705 * initialized.
1706 */
1707 unsigned col =
1708 (src_matrix->type->vector_elements < var->type->vector_elements)
1709 ? 0 : src_matrix->type->matrix_columns;
1710
1711 const glsl_type *const col_type = var->type->column_type();
1712 for (/* empty */; col < var->type->matrix_columns; col++) {
1713 ir_constant_data ident;
1714
1715 if (!col_type->is_double()) {
1716 ident.f[0] = 0.0f;
1717 ident.f[1] = 0.0f;
1718 ident.f[2] = 0.0f;
1719 ident.f[3] = 0.0f;
1720 ident.f[col] = 1.0f;
1721 } else {
1722 ident.d[0] = 0.0;
1723 ident.d[1] = 0.0;
1724 ident.d[2] = 0.0;
1725 ident.d[3] = 0.0;
1726 ident.d[col] = 1.0;
1727 }
1728
1729 ir_rvalue *const rhs = new(ctx) ir_constant(col_type, &ident);
1730
1731 ir_rvalue *const lhs =
1732 new(ctx) ir_dereference_array(var, new(ctx) ir_constant(col));
1733
1734 ir_instruction *inst = new(ctx) ir_assignment(lhs, rhs);
1735 instructions->push_tail(inst);
1736 }
1737 }
1738
1739 /* Assign columns from the source matrix to the destination matrix.
1740 *
1741 * Since the parameter will be used in the RHS of multiple assignments,
1742 * generate a temporary and copy the paramter there.
1743 */
1744 ir_variable *const rhs_var =
1745 new(ctx) ir_variable(first_param->type, "mat_ctor_mat",
1746 ir_var_temporary);
1747 instructions->push_tail(rhs_var);
1748
1749 ir_dereference *const rhs_var_ref =
1750 new(ctx) ir_dereference_variable(rhs_var);
1751 ir_instruction *const inst =
1752 new(ctx) ir_assignment(rhs_var_ref, first_param);
1753 instructions->push_tail(inst);
1754
1755 const unsigned last_row = MIN2(src_matrix->type->vector_elements,
1756 var->type->vector_elements);
1757 const unsigned last_col = MIN2(src_matrix->type->matrix_columns,
1758 var->type->matrix_columns);
1759
1760 unsigned swiz[4] = { 0, 0, 0, 0 };
1761 for (unsigned i = 1; i < last_row; i++)
1762 swiz[i] = i;
1763
1764 const unsigned write_mask = (1U << last_row) - 1;
1765
1766 for (unsigned i = 0; i < last_col; i++) {
1767 ir_dereference *const lhs =
1768 new(ctx) ir_dereference_array(var, new(ctx) ir_constant(i));
1769 ir_rvalue *const rhs_col =
1770 new(ctx) ir_dereference_array(rhs_var, new(ctx) ir_constant(i));
1771
1772 /* If one matrix has columns that are smaller than the columns of the
1773 * other matrix, wrap the column access of the larger with a swizzle
1774 * so that the LHS and RHS of the assignment have the same size (and
1775 * therefore have the same type).
1776 *
1777 * It would be perfectly valid to unconditionally generate the
1778 * swizzles, this this will typically result in a more compact IR
1779 * tree.
1780 */
1781 ir_rvalue *rhs;
1782 if (lhs->type->vector_elements != rhs_col->type->vector_elements) {
1783 rhs = new(ctx) ir_swizzle(rhs_col, swiz, last_row);
1784 } else {
1785 rhs = rhs_col;
1786 }
1787
1788 ir_instruction *inst =
1789 new(ctx) ir_assignment(lhs, rhs, NULL, write_mask);
1790 instructions->push_tail(inst);
1791 }
1792 } else {
1793 const unsigned cols = type->matrix_columns;
1794 const unsigned rows = type->vector_elements;
1795 unsigned remaining_slots = rows * cols;
1796 unsigned col_idx = 0;
1797 unsigned row_idx = 0;
1798
1799 foreach_in_list(ir_rvalue, rhs, parameters) {
1800 unsigned rhs_components = rhs->type->components();
1801 unsigned rhs_base = 0;
1802
1803 if (remaining_slots == 0)
1804 break;
1805
1806 /* Since the parameter might be used in the RHS of two assignments,
1807 * generate a temporary and copy the paramter there.
1808 */
1809 ir_variable *rhs_var =
1810 new(ctx) ir_variable(rhs->type, "mat_ctor_vec", ir_var_temporary);
1811 instructions->push_tail(rhs_var);
1812
1813 ir_dereference *rhs_var_ref =
1814 new(ctx) ir_dereference_variable(rhs_var);
1815 ir_instruction *inst = new(ctx) ir_assignment(rhs_var_ref, rhs);
1816 instructions->push_tail(inst);
1817
1818 do {
1819 /* Assign the current parameter to as many components of the matrix
1820 * as it will fill.
1821 *
1822 * NOTE: A single vector parameter can span two matrix columns. A
1823 * single vec4, for example, can completely fill a mat2.
1824 */
1825 unsigned count = MIN2(rows - row_idx,
1826 rhs_components - rhs_base);
1827
1828 rhs_var_ref = new(ctx) ir_dereference_variable(rhs_var);
1829 ir_instruction *inst = assign_to_matrix_column(var, col_idx,
1830 row_idx,
1831 rhs_var_ref,
1832 rhs_base,
1833 count, ctx);
1834 instructions->push_tail(inst);
1835 rhs_base += count;
1836 row_idx += count;
1837 remaining_slots -= count;
1838
1839 /* Sometimes, there is still data left in the parameters and
1840 * components left to be set in the destination but in other
1841 * column.
1842 */
1843 if (row_idx >= rows) {
1844 row_idx = 0;
1845 col_idx++;
1846 }
1847 } while(remaining_slots > 0 && rhs_base < rhs_components);
1848 }
1849 }
1850
1851 return new(ctx) ir_dereference_variable(var);
1852 }
1853
1854
1855 static ir_rvalue *
1856 emit_inline_record_constructor(const glsl_type *type,
1857 exec_list *instructions,
1858 exec_list *parameters,
1859 void *mem_ctx)
1860 {
1861 ir_variable *const var =
1862 new(mem_ctx) ir_variable(type, "record_ctor", ir_var_temporary);
1863 ir_dereference_variable *const d =
1864 new(mem_ctx) ir_dereference_variable(var);
1865
1866 instructions->push_tail(var);
1867
1868 exec_node *node = parameters->get_head_raw();
1869 for (unsigned i = 0; i < type->length; i++) {
1870 assert(!node->is_tail_sentinel());
1871
1872 ir_dereference *const lhs =
1873 new(mem_ctx) ir_dereference_record(d->clone(mem_ctx, NULL),
1874 type->fields.structure[i].name);
1875
1876 ir_rvalue *const rhs = ((ir_instruction *) node)->as_rvalue();
1877 assert(rhs != NULL);
1878
1879 ir_instruction *const assign = new(mem_ctx) ir_assignment(lhs, rhs);
1880
1881 instructions->push_tail(assign);
1882 node = node->next;
1883 }
1884
1885 return d;
1886 }
1887
1888
1889 static ir_rvalue *
1890 process_record_constructor(exec_list *instructions,
1891 const glsl_type *constructor_type,
1892 YYLTYPE *loc, exec_list *parameters,
1893 struct _mesa_glsl_parse_state *state)
1894 {
1895 void *ctx = state;
1896 /* From page 32 (page 38 of the PDF) of the GLSL 1.20 spec:
1897 *
1898 * "The arguments to the constructor will be used to set the structure's
1899 * fields, in order, using one argument per field. Each argument must
1900 * be the same type as the field it sets, or be a type that can be
1901 * converted to the field's type according to Section 4.1.10 “Implicit
1902 * Conversions.”"
1903 *
1904 * From page 35 (page 41 of the PDF) of the GLSL 4.20 spec:
1905 *
1906 * "In all cases, the innermost initializer (i.e., not a list of
1907 * initializers enclosed in curly braces) applied to an object must
1908 * have the same type as the object being initialized or be a type that
1909 * can be converted to the object's type according to section 4.1.10
1910 * "Implicit Conversions". In the latter case, an implicit conversion
1911 * will be done on the initializer before the assignment is done."
1912 */
1913 exec_list actual_parameters;
1914
1915 const unsigned parameter_count =
1916 process_parameters(instructions, &actual_parameters, parameters,
1917 state);
1918
1919 if (parameter_count != constructor_type->length) {
1920 _mesa_glsl_error(loc, state,
1921 "%s parameters in constructor for `%s'",
1922 parameter_count > constructor_type->length
1923 ? "too many": "insufficient",
1924 constructor_type->name);
1925 return ir_rvalue::error_value(ctx);
1926 }
1927
1928 bool all_parameters_are_constant = true;
1929
1930 int i = 0;
1931 /* Type cast each parameter and, if possible, fold constants. */
1932 foreach_in_list_safe(ir_rvalue, ir, &actual_parameters) {
1933
1934 const glsl_struct_field *struct_field =
1935 &constructor_type->fields.structure[i];
1936
1937 /* Apply implicit conversions (not the scalar constructor rules, see the
1938 * spec quote above!) and attempt to convert the parameter to a constant
1939 * valued expression. After doing so, track whether or not all the
1940 * parameters to the constructor are trivially constant valued
1941 * expressions.
1942 */
1943 all_parameters_are_constant &=
1944 implicitly_convert_component(ir, struct_field->type->base_type,
1945 state);
1946
1947 if (ir->type != struct_field->type) {
1948 _mesa_glsl_error(loc, state,
1949 "parameter type mismatch in constructor for `%s.%s' "
1950 "(%s vs %s)",
1951 constructor_type->name,
1952 struct_field->name,
1953 ir->type->name,
1954 struct_field->type->name);
1955 return ir_rvalue::error_value(ctx);
1956 }
1957
1958 i++;
1959 }
1960
1961 if (all_parameters_are_constant) {
1962 return new(ctx) ir_constant(constructor_type, &actual_parameters);
1963 } else {
1964 return emit_inline_record_constructor(constructor_type, instructions,
1965 &actual_parameters, state);
1966 }
1967 }
1968
1969 ir_rvalue *
1970 ast_function_expression::handle_method(exec_list *instructions,
1971 struct _mesa_glsl_parse_state *state)
1972 {
1973 const ast_expression *field = subexpressions[0];
1974 ir_rvalue *op;
1975 ir_rvalue *result;
1976 void *ctx = state;
1977 /* Handle "method calls" in GLSL 1.20 - namely, array.length() */
1978 YYLTYPE loc = get_location();
1979 state->check_version(120, 300, &loc, "methods not supported");
1980
1981 const char *method;
1982 method = field->primary_expression.identifier;
1983
1984 /* This would prevent to raise "uninitialized variable" warnings when
1985 * calling array.length.
1986 */
1987 field->subexpressions[0]->set_is_lhs(true);
1988 op = field->subexpressions[0]->hir(instructions, state);
1989 if (strcmp(method, "length") == 0) {
1990 if (!this->expressions.is_empty()) {
1991 _mesa_glsl_error(&loc, state, "length method takes no arguments");
1992 goto fail;
1993 }
1994
1995 if (op->type->is_array()) {
1996 if (op->type->is_unsized_array()) {
1997 if (!state->has_shader_storage_buffer_objects()) {
1998 _mesa_glsl_error(&loc, state,
1999 "length called on unsized array"
2000 " only available with"
2001 " ARB_shader_storage_buffer_object");
2002 }
2003 /* Calculate length of an unsized array in run-time */
2004 result = new(ctx) ir_expression(ir_unop_ssbo_unsized_array_length,
2005 op);
2006 } else {
2007 result = new(ctx) ir_constant(op->type->array_size());
2008 }
2009 } else if (op->type->is_vector()) {
2010 if (state->has_420pack()) {
2011 /* .length() returns int. */
2012 result = new(ctx) ir_constant((int) op->type->vector_elements);
2013 } else {
2014 _mesa_glsl_error(&loc, state, "length method on matrix only"
2015 " available with ARB_shading_language_420pack");
2016 goto fail;
2017 }
2018 } else if (op->type->is_matrix()) {
2019 if (state->has_420pack()) {
2020 /* .length() returns int. */
2021 result = new(ctx) ir_constant((int) op->type->matrix_columns);
2022 } else {
2023 _mesa_glsl_error(&loc, state, "length method on matrix only"
2024 " available with ARB_shading_language_420pack");
2025 goto fail;
2026 }
2027 } else {
2028 _mesa_glsl_error(&loc, state, "length called on scalar.");
2029 goto fail;
2030 }
2031 } else {
2032 _mesa_glsl_error(&loc, state, "unknown method: `%s'", method);
2033 goto fail;
2034 }
2035 return result;
2036 fail:
2037 return ir_rvalue::error_value(ctx);
2038 }
2039
2040 static inline bool is_valid_constructor(const glsl_type *type,
2041 struct _mesa_glsl_parse_state *state)
2042 {
2043 return type->is_numeric() || type->is_boolean() ||
2044 (state->has_bindless() && (type->is_sampler() || type->is_image()));
2045 }
2046
2047 ir_rvalue *
2048 ast_function_expression::hir(exec_list *instructions,
2049 struct _mesa_glsl_parse_state *state)
2050 {
2051 void *ctx = state;
2052 /* There are three sorts of function calls.
2053 *
2054 * 1. constructors - The first subexpression is an ast_type_specifier.
2055 * 2. methods - Only the .length() method of array types.
2056 * 3. functions - Calls to regular old functions.
2057 *
2058 */
2059 if (is_constructor()) {
2060 const ast_type_specifier *type =
2061 (ast_type_specifier *) subexpressions[0];
2062 YYLTYPE loc = type->get_location();
2063 const char *name;
2064
2065 const glsl_type *const constructor_type = type->glsl_type(& name, state);
2066
2067 /* constructor_type can be NULL if a variable with the same name as the
2068 * structure has come into scope.
2069 */
2070 if (constructor_type == NULL) {
2071 _mesa_glsl_error(& loc, state, "unknown type `%s' (structure name "
2072 "may be shadowed by a variable with the same name)",
2073 type->type_name);
2074 return ir_rvalue::error_value(ctx);
2075 }
2076
2077
2078 /* Constructors for opaque types are illegal.
2079 *
2080 * From section 4.1.7 of the ARB_bindless_texture spec:
2081 *
2082 * "Samplers are represented using 64-bit integer handles, and may be "
2083 * converted to and from 64-bit integers using constructors."
2084 *
2085 * From section 4.1.X of the ARB_bindless_texture spec:
2086 *
2087 * "Images are represented using 64-bit integer handles, and may be
2088 * converted to and from 64-bit integers using constructors."
2089 */
2090 if (constructor_type->contains_atomic() ||
2091 (!state->has_bindless() && constructor_type->contains_opaque())) {
2092 _mesa_glsl_error(& loc, state, "cannot construct %s type `%s'",
2093 state->has_bindless() ? "atomic" : "opaque",
2094 constructor_type->name);
2095 return ir_rvalue::error_value(ctx);
2096 }
2097
2098 if (constructor_type->is_subroutine()) {
2099 _mesa_glsl_error(& loc, state,
2100 "subroutine name cannot be a constructor `%s'",
2101 constructor_type->name);
2102 return ir_rvalue::error_value(ctx);
2103 }
2104
2105 if (constructor_type->is_array()) {
2106 if (!state->check_version(120, 300, &loc,
2107 "array constructors forbidden")) {
2108 return ir_rvalue::error_value(ctx);
2109 }
2110
2111 return process_array_constructor(instructions, constructor_type,
2112 & loc, &this->expressions, state);
2113 }
2114
2115
2116 /* There are two kinds of constructor calls. Constructors for arrays and
2117 * structures must have the exact number of arguments with matching types
2118 * in the correct order. These constructors follow essentially the same
2119 * type matching rules as functions.
2120 *
2121 * Constructors for built-in language types, such as mat4 and vec2, are
2122 * free form. The only requirements are that the parameters must provide
2123 * enough values of the correct scalar type and that no arguments are
2124 * given past the last used argument.
2125 *
2126 * When using the C-style initializer syntax from GLSL 4.20, constructors
2127 * must have the exact number of arguments with matching types in the
2128 * correct order.
2129 */
2130 if (constructor_type->is_struct()) {
2131 return process_record_constructor(instructions, constructor_type,
2132 &loc, &this->expressions,
2133 state);
2134 }
2135
2136 if (!is_valid_constructor(constructor_type, state))
2137 return ir_rvalue::error_value(ctx);
2138
2139 /* Total number of components of the type being constructed. */
2140 const unsigned type_components = constructor_type->components();
2141
2142 /* Number of components from parameters that have actually been
2143 * consumed. This is used to perform several kinds of error checking.
2144 */
2145 unsigned components_used = 0;
2146
2147 unsigned matrix_parameters = 0;
2148 unsigned nonmatrix_parameters = 0;
2149 exec_list actual_parameters;
2150
2151 foreach_list_typed(ast_node, ast, link, &this->expressions) {
2152 ir_rvalue *result = ast->hir(instructions, state);
2153
2154 /* From page 50 (page 56 of the PDF) of the GLSL 1.50 spec:
2155 *
2156 * "It is an error to provide extra arguments beyond this
2157 * last used argument."
2158 */
2159 if (components_used >= type_components) {
2160 _mesa_glsl_error(& loc, state, "too many parameters to `%s' "
2161 "constructor",
2162 constructor_type->name);
2163 return ir_rvalue::error_value(ctx);
2164 }
2165
2166 if (!is_valid_constructor(result->type, state)) {
2167 _mesa_glsl_error(& loc, state, "cannot construct `%s' from a "
2168 "non-numeric data type",
2169 constructor_type->name);
2170 return ir_rvalue::error_value(ctx);
2171 }
2172
2173 /* Count the number of matrix and nonmatrix parameters. This
2174 * is used below to enforce some of the constructor rules.
2175 */
2176 if (result->type->is_matrix())
2177 matrix_parameters++;
2178 else
2179 nonmatrix_parameters++;
2180
2181 actual_parameters.push_tail(result);
2182 components_used += result->type->components();
2183 }
2184
2185 /* From page 28 (page 34 of the PDF) of the GLSL 1.10 spec:
2186 *
2187 * "It is an error to construct matrices from other matrices. This
2188 * is reserved for future use."
2189 */
2190 if (matrix_parameters > 0
2191 && constructor_type->is_matrix()
2192 && !state->check_version(120, 100, &loc,
2193 "cannot construct `%s' from a matrix",
2194 constructor_type->name)) {
2195 return ir_rvalue::error_value(ctx);
2196 }
2197
2198 /* From page 50 (page 56 of the PDF) of the GLSL 1.50 spec:
2199 *
2200 * "If a matrix argument is given to a matrix constructor, it is
2201 * an error to have any other arguments."
2202 */
2203 if ((matrix_parameters > 0)
2204 && ((matrix_parameters + nonmatrix_parameters) > 1)
2205 && constructor_type->is_matrix()) {
2206 _mesa_glsl_error(& loc, state, "for matrix `%s' constructor, "
2207 "matrix must be only parameter",
2208 constructor_type->name);
2209 return ir_rvalue::error_value(ctx);
2210 }
2211
2212 /* From page 28 (page 34 of the PDF) of the GLSL 1.10 spec:
2213 *
2214 * "In these cases, there must be enough components provided in the
2215 * arguments to provide an initializer for every component in the
2216 * constructed value."
2217 */
2218 if (components_used < type_components && components_used != 1
2219 && matrix_parameters == 0) {
2220 _mesa_glsl_error(& loc, state, "too few components to construct "
2221 "`%s'",
2222 constructor_type->name);
2223 return ir_rvalue::error_value(ctx);
2224 }
2225
2226 /* Matrices can never be consumed as is by any constructor but matrix
2227 * constructors. If the constructor type is not matrix, always break the
2228 * matrix up into a series of column vectors.
2229 */
2230 if (!constructor_type->is_matrix()) {
2231 foreach_in_list_safe(ir_rvalue, matrix, &actual_parameters) {
2232 if (!matrix->type->is_matrix())
2233 continue;
2234
2235 /* Create a temporary containing the matrix. */
2236 ir_variable *var = new(ctx) ir_variable(matrix->type, "matrix_tmp",
2237 ir_var_temporary);
2238 instructions->push_tail(var);
2239 instructions->push_tail(
2240 new(ctx) ir_assignment(new(ctx) ir_dereference_variable(var),
2241 matrix));
2242 var->constant_value = matrix->constant_expression_value(ctx);
2243
2244 /* Replace the matrix with dereferences of its columns. */
2245 for (int i = 0; i < matrix->type->matrix_columns; i++) {
2246 matrix->insert_before(
2247 new (ctx) ir_dereference_array(var,
2248 new(ctx) ir_constant(i)));
2249 }
2250 matrix->remove();
2251 }
2252 }
2253
2254 bool all_parameters_are_constant = true;
2255
2256 /* Type cast each parameter and, if possible, fold constants.*/
2257 foreach_in_list_safe(ir_rvalue, ir, &actual_parameters) {
2258 const glsl_type *desired_type;
2259
2260 /* From section 5.4.1 of the ARB_bindless_texture spec:
2261 *
2262 * "In the following four constructors, the low 32 bits of the sampler
2263 * type correspond to the .x component of the uvec2 and the high 32
2264 * bits correspond to the .y component."
2265 *
2266 * uvec2(any sampler type) // Converts a sampler type to a
2267 * // pair of 32-bit unsigned integers
2268 * any sampler type(uvec2) // Converts a pair of 32-bit unsigned integers to
2269 * // a sampler type
2270 * uvec2(any image type) // Converts an image type to a
2271 * // pair of 32-bit unsigned integers
2272 * any image type(uvec2) // Converts a pair of 32-bit unsigned integers to
2273 * // an image type
2274 */
2275 if (ir->type->is_sampler() || ir->type->is_image()) {
2276 /* Convert a sampler/image type to a pair of 32-bit unsigned
2277 * integers as defined by ARB_bindless_texture.
2278 */
2279 if (constructor_type != glsl_type::uvec2_type) {
2280 _mesa_glsl_error(&loc, state, "sampler and image types can only "
2281 "be converted to a pair of 32-bit unsigned "
2282 "integers");
2283 }
2284 desired_type = glsl_type::uvec2_type;
2285 } else if (constructor_type->is_sampler() ||
2286 constructor_type->is_image()) {
2287 /* Convert a pair of 32-bit unsigned integers to a sampler or image
2288 * type as defined by ARB_bindless_texture.
2289 */
2290 if (ir->type != glsl_type::uvec2_type) {
2291 _mesa_glsl_error(&loc, state, "sampler and image types can only "
2292 "be converted from a pair of 32-bit unsigned "
2293 "integers");
2294 }
2295 desired_type = constructor_type;
2296 } else {
2297 desired_type =
2298 glsl_type::get_instance(constructor_type->base_type,
2299 ir->type->vector_elements,
2300 ir->type->matrix_columns);
2301 }
2302
2303 ir_rvalue *result = convert_component(ir, desired_type);
2304
2305 /* Attempt to convert the parameter to a constant valued expression.
2306 * After doing so, track whether or not all the parameters to the
2307 * constructor are trivially constant valued expressions.
2308 */
2309 ir_rvalue *const constant = result->constant_expression_value(ctx);
2310
2311 if (constant != NULL)
2312 result = constant;
2313 else
2314 all_parameters_are_constant = false;
2315
2316 if (result != ir) {
2317 ir->replace_with(result);
2318 }
2319 }
2320
2321 /* If all of the parameters are trivially constant, create a
2322 * constant representing the complete collection of parameters.
2323 */
2324 if (all_parameters_are_constant) {
2325 return new(ctx) ir_constant(constructor_type, &actual_parameters);
2326 } else if (constructor_type->is_scalar()) {
2327 return dereference_component((ir_rvalue *)
2328 actual_parameters.get_head_raw(),
2329 0);
2330 } else if (constructor_type->is_vector()) {
2331 return emit_inline_vector_constructor(constructor_type,
2332 instructions,
2333 &actual_parameters,
2334 ctx);
2335 } else {
2336 assert(constructor_type->is_matrix());
2337 return emit_inline_matrix_constructor(constructor_type,
2338 instructions,
2339 &actual_parameters,
2340 ctx);
2341 }
2342 } else if (subexpressions[0]->oper == ast_field_selection) {
2343 return handle_method(instructions, state);
2344 } else {
2345 const ast_expression *id = subexpressions[0];
2346 const char *func_name = NULL;
2347 YYLTYPE loc = get_location();
2348 exec_list actual_parameters;
2349 ir_variable *sub_var = NULL;
2350 ir_rvalue *array_idx = NULL;
2351
2352 process_parameters(instructions, &actual_parameters, &this->expressions,
2353 state);
2354
2355 if (id->oper == ast_array_index) {
2356 array_idx = generate_array_index(ctx, instructions, state, loc,
2357 id->subexpressions[0],
2358 id->subexpressions[1], &func_name,
2359 &actual_parameters);
2360 } else if (id->oper == ast_identifier) {
2361 func_name = id->primary_expression.identifier;
2362 } else {
2363 _mesa_glsl_error(&loc, state, "function name is not an identifier");
2364 }
2365
2366 /* an error was emitted earlier */
2367 if (!func_name)
2368 return ir_rvalue::error_value(ctx);
2369
2370 ir_function_signature *sig =
2371 match_function_by_name(func_name, &actual_parameters, state);
2372
2373 ir_rvalue *value = NULL;
2374 if (sig == NULL) {
2375 sig = match_subroutine_by_name(func_name, &actual_parameters,
2376 state, &sub_var);
2377 }
2378
2379 if (sig == NULL) {
2380 no_matching_function_error(func_name, &loc,
2381 &actual_parameters, state);
2382 value = ir_rvalue::error_value(ctx);
2383 } else if (!verify_parameter_modes(state, sig,
2384 actual_parameters,
2385 this->expressions)) {
2386 /* an error has already been emitted */
2387 value = ir_rvalue::error_value(ctx);
2388 } else if (sig->is_builtin() && strcmp(func_name, "ftransform") == 0) {
2389 /* ftransform refers to global variables, and we don't have any code
2390 * for remapping the variable references in the built-in shader.
2391 */
2392 ir_variable *mvp =
2393 state->symbols->get_variable("gl_ModelViewProjectionMatrix");
2394 ir_variable *vtx = state->symbols->get_variable("gl_Vertex");
2395 value = new(ctx) ir_expression(ir_binop_mul, glsl_type::vec4_type,
2396 new(ctx) ir_dereference_variable(mvp),
2397 new(ctx) ir_dereference_variable(vtx));
2398 } else {
2399 bool is_begin_interlock = false;
2400 bool is_end_interlock = false;
2401 if (sig->is_builtin() &&
2402 state->stage == MESA_SHADER_FRAGMENT &&
2403 state->ARB_fragment_shader_interlock_enable) {
2404 is_begin_interlock = strcmp(func_name, "beginInvocationInterlockARB") == 0;
2405 is_end_interlock = strcmp(func_name, "endInvocationInterlockARB") == 0;
2406 }
2407
2408 if (sig->is_builtin() &&
2409 ((state->stage == MESA_SHADER_TESS_CTRL &&
2410 strcmp(func_name, "barrier") == 0) ||
2411 is_begin_interlock || is_end_interlock)) {
2412 if (state->current_function == NULL ||
2413 strcmp(state->current_function->function_name(), "main") != 0) {
2414 _mesa_glsl_error(&loc, state,
2415 "%s() may only be used in main()", func_name);
2416 }
2417
2418 if (state->found_return) {
2419 _mesa_glsl_error(&loc, state,
2420 "%s() may not be used after return", func_name);
2421 }
2422
2423 if (instructions != &state->current_function->body) {
2424 _mesa_glsl_error(&loc, state,
2425 "%s() may not be used in control flow", func_name);
2426 }
2427 }
2428
2429 /* There can be only one begin/end interlock pair in the function. */
2430 if (is_begin_interlock) {
2431 if (state->found_begin_interlock)
2432 _mesa_glsl_error(&loc, state,
2433 "beginInvocationInterlockARB may not be used twice");
2434 state->found_begin_interlock = true;
2435 } else if (is_end_interlock) {
2436 if (!state->found_begin_interlock)
2437 _mesa_glsl_error(&loc, state,
2438 "endInvocationInterlockARB may not be used "
2439 "before beginInvocationInterlockARB");
2440 if (state->found_end_interlock)
2441 _mesa_glsl_error(&loc, state,
2442 "endInvocationInterlockARB may not be used twice");
2443 state->found_end_interlock = true;
2444 }
2445
2446 value = generate_call(instructions, sig, &actual_parameters, sub_var,
2447 array_idx, state);
2448 if (!value) {
2449 ir_variable *const tmp = new(ctx) ir_variable(glsl_type::void_type,
2450 "void_var",
2451 ir_var_temporary);
2452 instructions->push_tail(tmp);
2453 value = new(ctx) ir_dereference_variable(tmp);
2454 }
2455 }
2456
2457 return value;
2458 }
2459
2460 unreachable("not reached");
2461 }
2462
2463 bool
2464 ast_function_expression::has_sequence_subexpression() const
2465 {
2466 foreach_list_typed(const ast_node, ast, link, &this->expressions) {
2467 if (ast->has_sequence_subexpression())
2468 return true;
2469 }
2470
2471 return false;
2472 }
2473
2474 ir_rvalue *
2475 ast_aggregate_initializer::hir(exec_list *instructions,
2476 struct _mesa_glsl_parse_state *state)
2477 {
2478 void *ctx = state;
2479 YYLTYPE loc = this->get_location();
2480
2481 if (!this->constructor_type) {
2482 _mesa_glsl_error(&loc, state, "type of C-style initializer unknown");
2483 return ir_rvalue::error_value(ctx);
2484 }
2485 const glsl_type *const constructor_type = this->constructor_type;
2486
2487 if (!state->has_420pack()) {
2488 _mesa_glsl_error(&loc, state, "C-style initialization requires the "
2489 "GL_ARB_shading_language_420pack extension");
2490 return ir_rvalue::error_value(ctx);
2491 }
2492
2493 if (constructor_type->is_array()) {
2494 return process_array_constructor(instructions, constructor_type, &loc,
2495 &this->expressions, state);
2496 }
2497
2498 if (constructor_type->is_struct()) {
2499 return process_record_constructor(instructions, constructor_type, &loc,
2500 &this->expressions, state);
2501 }
2502
2503 return process_vec_mat_constructor(instructions, constructor_type, &loc,
2504 &this->expressions, state);
2505 }