2 * Copyright © 2015 Intel Corporation
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:
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
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
24 #include "vtn_private.h"
25 #include "nir/nir_vla.h"
27 static struct vtn_block
*
28 vtn_block(struct vtn_builder
*b
, uint32_t value_id
)
30 return vtn_value(b
, value_id
, vtn_value_type_block
)->block
;
33 static struct vtn_pointer
*
34 vtn_load_param_pointer(struct vtn_builder
*b
,
35 struct vtn_type
*param_type
,
38 struct vtn_type
*ptr_type
= param_type
;
39 assert(param_type
->base_type
== vtn_base_type_image
||
40 param_type
->base_type
== vtn_base_type_sampler
);
41 ptr_type
= rzalloc(b
, struct vtn_type
);
42 ptr_type
->base_type
= vtn_base_type_pointer
;
43 ptr_type
->deref
= param_type
;
44 ptr_type
->storage_class
= SpvStorageClassUniformConstant
;
46 return vtn_pointer_from_ssa(b
, nir_load_param(&b
->nb
, param_idx
), ptr_type
);
50 glsl_type_count_function_params(const struct glsl_type
*type
)
52 if (glsl_type_is_vector_or_scalar(type
)) {
54 } else if (glsl_type_is_array_or_matrix(type
)) {
55 return glsl_get_length(type
) *
56 glsl_type_count_function_params(glsl_get_array_element(type
));
58 assert(glsl_type_is_struct_or_ifc(type
));
60 unsigned elems
= glsl_get_length(type
);
61 for (unsigned i
= 0; i
< elems
; i
++) {
62 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
63 count
+= glsl_type_count_function_params(elem_type
);
70 vtn_type_count_function_params(struct vtn_type
*type
)
72 switch (type
->base_type
) {
73 case vtn_base_type_scalar
:
74 case vtn_base_type_vector
:
75 case vtn_base_type_array
:
76 case vtn_base_type_matrix
:
77 case vtn_base_type_struct
:
78 case vtn_base_type_pointer
:
79 return glsl_type_count_function_params(type
->type
);
81 case vtn_base_type_sampled_image
:
90 glsl_type_add_to_function_params(const struct glsl_type
*type
,
94 if (glsl_type_is_vector_or_scalar(type
)) {
95 func
->params
[(*param_idx
)++] = (nir_parameter
) {
96 .num_components
= glsl_get_vector_elements(type
),
97 .bit_size
= glsl_get_bit_size(type
),
99 } else if (glsl_type_is_array_or_matrix(type
)) {
100 unsigned elems
= glsl_get_length(type
);
101 const struct glsl_type
*elem_type
= glsl_get_array_element(type
);
102 for (unsigned i
= 0; i
< elems
; i
++)
103 glsl_type_add_to_function_params(elem_type
,func
, param_idx
);
105 assert(glsl_type_is_struct_or_ifc(type
));
106 unsigned elems
= glsl_get_length(type
);
107 for (unsigned i
= 0; i
< elems
; i
++) {
108 const struct glsl_type
*elem_type
= glsl_get_struct_field(type
, i
);
109 glsl_type_add_to_function_params(elem_type
, func
, param_idx
);
115 vtn_type_add_to_function_params(struct vtn_type
*type
,
119 static const nir_parameter nir_deref_param
= {
124 switch (type
->base_type
) {
125 case vtn_base_type_scalar
:
126 case vtn_base_type_vector
:
127 case vtn_base_type_array
:
128 case vtn_base_type_matrix
:
129 case vtn_base_type_struct
:
130 case vtn_base_type_pointer
:
131 glsl_type_add_to_function_params(type
->type
, func
, param_idx
);
134 case vtn_base_type_sampled_image
:
135 func
->params
[(*param_idx
)++] = nir_deref_param
;
136 func
->params
[(*param_idx
)++] = nir_deref_param
;
139 case vtn_base_type_image
:
140 case vtn_base_type_sampler
:
141 func
->params
[(*param_idx
)++] = nir_deref_param
;
145 unreachable("Unsupported type");
150 vtn_ssa_value_add_to_call_params(struct vtn_builder
*b
,
151 struct vtn_ssa_value
*value
,
152 nir_call_instr
*call
,
155 if (glsl_type_is_vector_or_scalar(value
->type
)) {
156 call
->params
[(*param_idx
)++] = nir_src_for_ssa(value
->def
);
158 unsigned elems
= glsl_get_length(value
->type
);
159 for (unsigned i
= 0; i
< elems
; i
++) {
160 vtn_ssa_value_add_to_call_params(b
, value
->elems
[i
],
167 vtn_ssa_value_load_function_param(struct vtn_builder
*b
,
168 struct vtn_ssa_value
*value
,
171 if (glsl_type_is_vector_or_scalar(value
->type
)) {
172 value
->def
= nir_load_param(&b
->nb
, (*param_idx
)++);
174 unsigned elems
= glsl_get_length(value
->type
);
175 for (unsigned i
= 0; i
< elems
; i
++)
176 vtn_ssa_value_load_function_param(b
, value
->elems
[i
], param_idx
);
181 vtn_handle_function_call(struct vtn_builder
*b
, SpvOp opcode
,
182 const uint32_t *w
, unsigned count
)
184 struct vtn_function
*vtn_callee
=
185 vtn_value(b
, w
[3], vtn_value_type_function
)->func
;
186 struct nir_function
*callee
= vtn_callee
->impl
->function
;
188 vtn_callee
->referenced
= true;
190 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
192 unsigned param_idx
= 0;
194 nir_deref_instr
*ret_deref
= NULL
;
195 struct vtn_type
*ret_type
= vtn_callee
->type
->return_type
;
196 if (ret_type
->base_type
!= vtn_base_type_void
) {
197 nir_variable
*ret_tmp
=
198 nir_local_variable_create(b
->nb
.impl
,
199 glsl_get_bare_type(ret_type
->type
),
201 ret_deref
= nir_build_deref_var(&b
->nb
, ret_tmp
);
202 call
->params
[param_idx
++] = nir_src_for_ssa(&ret_deref
->dest
.ssa
);
205 for (unsigned i
= 0; i
< vtn_callee
->type
->length
; i
++) {
206 struct vtn_type
*arg_type
= vtn_callee
->type
->params
[i
];
207 unsigned arg_id
= w
[4 + i
];
209 if (arg_type
->base_type
== vtn_base_type_sampled_image
) {
210 struct vtn_sampled_image
*sampled_image
=
211 vtn_value(b
, arg_id
, vtn_value_type_sampled_image
)->sampled_image
;
213 call
->params
[param_idx
++] =
214 nir_src_for_ssa(vtn_pointer_to_ssa(b
, sampled_image
->image
));
215 call
->params
[param_idx
++] =
216 nir_src_for_ssa(vtn_pointer_to_ssa(b
, sampled_image
->sampler
));
217 } else if (arg_type
->base_type
== vtn_base_type_image
||
218 arg_type
->base_type
== vtn_base_type_sampler
) {
219 struct vtn_pointer
*pointer
=
220 vtn_value(b
, arg_id
, vtn_value_type_pointer
)->pointer
;
221 call
->params
[param_idx
++] =
222 nir_src_for_ssa(vtn_pointer_to_ssa(b
, pointer
));
224 struct vtn_ssa_value
*arg
= vtn_ssa_value(b
, arg_id
);
225 vtn_assert(arg
->type
== glsl_get_bare_type(arg_type
->type
));
226 vtn_ssa_value_add_to_call_params(b
, arg
, call
, ¶m_idx
);
229 assert(param_idx
== call
->num_params
);
231 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
233 if (ret_type
->base_type
== vtn_base_type_void
) {
234 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
236 vtn_push_ssa_value(b
, w
[2], vtn_local_load(b
, ret_deref
, 0));
241 vtn_cfg_handle_prepass_instruction(struct vtn_builder
*b
, SpvOp opcode
,
242 const uint32_t *w
, unsigned count
)
245 case SpvOpFunction
: {
246 vtn_assert(b
->func
== NULL
);
247 b
->func
= rzalloc(b
, struct vtn_function
);
249 b
->func
->node
.type
= vtn_cf_node_type_function
;
250 b
->func
->node
.parent
= NULL
;
251 list_inithead(&b
->func
->body
);
252 b
->func
->control
= w
[3];
254 UNUSED
const struct glsl_type
*result_type
= vtn_get_type(b
, w
[1])->type
;
255 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_function
);
258 b
->func
->type
= vtn_get_type(b
, w
[4]);
259 const struct vtn_type
*func_type
= b
->func
->type
;
261 vtn_assert(func_type
->return_type
->type
== result_type
);
264 nir_function_create(b
->shader
, ralloc_strdup(b
->shader
, val
->name
));
266 unsigned num_params
= 0;
267 for (unsigned i
= 0; i
< func_type
->length
; i
++)
268 num_params
+= vtn_type_count_function_params(func_type
->params
[i
]);
270 /* Add one parameter for the function return value */
271 if (func_type
->return_type
->base_type
!= vtn_base_type_void
)
274 func
->num_params
= num_params
;
275 func
->params
= ralloc_array(b
->shader
, nir_parameter
, num_params
);
278 if (func_type
->return_type
->base_type
!= vtn_base_type_void
) {
279 nir_address_format addr_format
=
280 vtn_mode_to_address_format(b
, vtn_variable_mode_function
);
281 /* The return value is a regular pointer */
282 func
->params
[idx
++] = (nir_parameter
) {
283 .num_components
= nir_address_format_num_components(addr_format
),
284 .bit_size
= nir_address_format_bit_size(addr_format
),
288 for (unsigned i
= 0; i
< func_type
->length
; i
++)
289 vtn_type_add_to_function_params(func_type
->params
[i
], func
, &idx
);
290 assert(idx
== num_params
);
292 b
->func
->impl
= nir_function_impl_create(func
);
293 nir_builder_init(&b
->nb
, func
->impl
);
294 b
->nb
.cursor
= nir_before_cf_list(&b
->func
->impl
->body
);
295 b
->nb
.exact
= b
->exact
;
297 b
->func_param_idx
= 0;
299 /* The return value is the first parameter */
300 if (func_type
->return_type
->base_type
!= vtn_base_type_void
)
305 case SpvOpFunctionEnd
:
310 case SpvOpFunctionParameter
: {
311 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
313 vtn_assert(b
->func_param_idx
< b
->func
->impl
->function
->num_params
);
315 if (type
->base_type
== vtn_base_type_sampled_image
) {
316 /* Sampled images are actually two parameters. The first is the
317 * image and the second is the sampler.
319 struct vtn_value
*val
=
320 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
322 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
324 struct vtn_type
*image_type
= rzalloc(b
, struct vtn_type
);
325 image_type
->base_type
= vtn_base_type_image
;
326 image_type
->type
= type
->type
;
328 struct vtn_type
*sampler_type
= rzalloc(b
, struct vtn_type
);
329 sampler_type
->base_type
= vtn_base_type_sampler
;
330 sampler_type
->type
= glsl_bare_sampler_type();
332 val
->sampled_image
->image
=
333 vtn_load_param_pointer(b
, image_type
, b
->func_param_idx
++);
334 val
->sampled_image
->sampler
=
335 vtn_load_param_pointer(b
, sampler_type
, b
->func_param_idx
++);
336 } else if (type
->base_type
== vtn_base_type_image
||
337 type
->base_type
== vtn_base_type_sampler
) {
338 vtn_push_pointer(b
, w
[2], vtn_load_param_pointer(b
, type
, b
->func_param_idx
++));
340 /* We're a regular SSA value. */
341 struct vtn_ssa_value
*value
= vtn_create_ssa_value(b
, type
->type
);
342 vtn_ssa_value_load_function_param(b
, value
, &b
->func_param_idx
);
343 vtn_push_ssa_value(b
, w
[2], value
);
349 vtn_assert(b
->block
== NULL
);
350 b
->block
= rzalloc(b
, struct vtn_block
);
351 b
->block
->node
.type
= vtn_cf_node_type_block
;
353 vtn_push_value(b
, w
[1], vtn_value_type_block
)->block
= b
->block
;
355 if (b
->func
->start_block
== NULL
) {
356 /* This is the first block encountered for this function. In this
357 * case, we set the start block and add it to the list of
358 * implemented functions that we'll walk later.
360 b
->func
->start_block
= b
->block
;
361 list_addtail(&b
->func
->node
.link
, &b
->functions
);
366 case SpvOpSelectionMerge
:
368 vtn_assert(b
->block
&& b
->block
->merge
== NULL
);
373 case SpvOpBranchConditional
:
377 case SpvOpReturnValue
:
378 case SpvOpUnreachable
:
379 vtn_assert(b
->block
&& b
->block
->branch
== NULL
);
380 b
->block
->branch
= w
;
385 /* Continue on as per normal */
392 /* This function performs a depth-first search of the cases and puts them
393 * in fall-through order.
396 vtn_order_case(struct vtn_switch
*swtch
, struct vtn_case
*cse
)
403 list_del(&cse
->node
.link
);
405 if (cse
->fallthrough
) {
406 vtn_order_case(swtch
, cse
->fallthrough
);
408 /* If we have a fall-through, place this case right before the case it
409 * falls through to. This ensures that fallthroughs come one after
410 * the other. These two can never get separated because that would
411 * imply something else falling through to the same case. Also, this
412 * can't break ordering because the DFS ensures that this case is
413 * visited before anything that falls through to it.
415 list_addtail(&cse
->node
.link
, &cse
->fallthrough
->node
.link
);
417 list_add(&cse
->node
.link
, &swtch
->cases
);
422 vtn_switch_order_cases(struct vtn_switch
*swtch
)
424 struct list_head cases
;
425 list_replace(&swtch
->cases
, &cases
);
426 list_inithead(&swtch
->cases
);
427 while (!list_is_empty(&cases
)) {
428 struct vtn_case
*cse
=
429 list_first_entry(&cases
, struct vtn_case
, node
.link
);
430 vtn_order_case(swtch
, cse
);
435 vtn_block_set_merge_cf_node(struct vtn_builder
*b
, struct vtn_block
*block
,
436 struct vtn_cf_node
*cf_node
)
438 vtn_fail_if(block
->merge_cf_node
!= NULL
,
439 "The merge block declared by a header block cannot be a "
440 "merge block declared by any other header block.");
442 block
->merge_cf_node
= cf_node
;
445 #define VTN_DECL_CF_NODE_FIND(_type) \
446 static inline struct vtn_##_type * \
447 vtn_cf_node_find_##_type(struct vtn_cf_node *node) \
449 while (node && node->type != vtn_cf_node_type_##_type) \
450 node = node->parent; \
451 return (struct vtn_##_type *)node; \
454 VTN_DECL_CF_NODE_FIND(if)
455 VTN_DECL_CF_NODE_FIND(loop
)
456 VTN_DECL_CF_NODE_FIND(case)
457 VTN_DECL_CF_NODE_FIND(switch)
458 VTN_DECL_CF_NODE_FIND(function
)
460 static enum vtn_branch_type
461 vtn_handle_branch(struct vtn_builder
*b
,
462 struct vtn_cf_node
*cf_parent
,
463 struct vtn_block
*target_block
)
465 struct vtn_loop
*loop
= vtn_cf_node_find_loop(cf_parent
);
467 /* Detect a loop back-edge first. That way none of the code below
468 * accidentally operates on a loop back-edge.
470 if (loop
&& target_block
== loop
->header_block
)
471 return vtn_branch_type_loop_back_edge
;
473 /* Try to detect fall-through */
474 if (target_block
->switch_case
) {
475 /* When it comes to handling switch cases, we can break calls to
476 * vtn_handle_branch into two cases: calls from within a case construct
477 * and calls for the jump to each case construct. In the second case,
478 * cf_parent is the vtn_switch itself and vtn_cf_node_find_case() will
479 * return the outer switch case in which this switch is contained. It's
480 * fine if the target block is a switch case from an outer switch as
481 * long as it is also the switch break for this switch.
483 struct vtn_case
*switch_case
= vtn_cf_node_find_case(cf_parent
);
485 /* This doesn't get called for the OpSwitch */
486 vtn_fail_if(switch_case
== NULL
,
487 "A switch case can only be entered through an OpSwitch or "
488 "falling through from another switch case.");
490 /* Because block->switch_case is only set on the entry block for a given
491 * switch case, we only ever get here if we're jumping to the start of a
492 * switch case. It's possible, however, that a switch case could jump
493 * to itself via a back-edge. That *should* get caught by the loop
494 * handling case above but if we have a back edge without a loop merge,
495 * we could en up here.
497 vtn_fail_if(target_block
->switch_case
== switch_case
,
498 "A switch cannot fall-through to itself. Likely, there is "
499 "a back-edge which is not to a loop header.");
501 vtn_fail_if(target_block
->switch_case
->node
.parent
!=
502 switch_case
->node
.parent
,
503 "A switch case fall-through must come from the same "
504 "OpSwitch construct");
506 vtn_fail_if(switch_case
->fallthrough
!= NULL
&&
507 switch_case
->fallthrough
!= target_block
->switch_case
,
508 "Each case construct can have at most one branch to "
509 "another case construct");
511 switch_case
->fallthrough
= target_block
->switch_case
;
513 /* We don't immediately return vtn_branch_type_switch_fallthrough
514 * because it may also be a loop or switch break for an inner loop or
515 * switch and that takes precedence.
519 if (loop
&& target_block
== loop
->cont_block
)
520 return vtn_branch_type_loop_continue
;
522 /* We walk blocks as a breadth-first search on the control-flow construct
523 * tree where, when we find a construct, we add the vtn_cf_node for that
524 * construct and continue iterating at the merge target block (if any).
525 * Therefore, we want merges whose with parent == cf_parent to be treated
526 * as regular branches. We only want to consider merges if they break out
527 * of the current CF construct.
529 if (target_block
->merge_cf_node
!= NULL
&&
530 target_block
->merge_cf_node
->parent
!= cf_parent
) {
531 switch (target_block
->merge_cf_node
->type
) {
532 case vtn_cf_node_type_if
:
533 for (struct vtn_cf_node
*node
= cf_parent
;
534 node
!= target_block
->merge_cf_node
; node
= node
->parent
) {
535 vtn_fail_if(node
== NULL
|| node
->type
!= vtn_cf_node_type_if
,
536 "Branching to the merge block of a selection "
537 "construct can only be used to break out of a "
538 "selection construct");
540 struct vtn_if
*if_stmt
= vtn_cf_node_as_if(node
);
542 /* This should be guaranteed by our iteration */
543 assert(if_stmt
->merge_block
!= target_block
);
545 vtn_fail_if(if_stmt
->merge_block
!= NULL
,
546 "Branching to the merge block of a selection "
547 "construct can only be used to break out of the "
548 "inner most nested selection level");
550 return vtn_branch_type_if_merge
;
552 case vtn_cf_node_type_loop
:
553 vtn_fail_if(target_block
->merge_cf_node
!= &loop
->node
,
554 "Loop breaks can only break out of the inner most "
555 "nested loop level");
556 return vtn_branch_type_loop_break
;
558 case vtn_cf_node_type_switch
: {
559 struct vtn_switch
*swtch
= vtn_cf_node_find_switch(cf_parent
);
560 vtn_fail_if(target_block
->merge_cf_node
!= &swtch
->node
,
561 "Switch breaks can only break out of the inner most "
562 "nested switch level");
563 return vtn_branch_type_switch_break
;
567 unreachable("Invalid CF node type for a merge");
571 if (target_block
->switch_case
)
572 return vtn_branch_type_switch_fallthrough
;
574 return vtn_branch_type_none
;
577 struct vtn_cfg_work_item
{
578 struct list_head link
;
580 struct vtn_cf_node
*cf_parent
;
581 struct list_head
*cf_list
;
582 struct vtn_block
*start_block
;
586 vtn_add_cfg_work_item(struct vtn_builder
*b
,
587 struct list_head
*work_list
,
588 struct vtn_cf_node
*cf_parent
,
589 struct list_head
*cf_list
,
590 struct vtn_block
*start_block
)
592 struct vtn_cfg_work_item
*work
= ralloc(b
, struct vtn_cfg_work_item
);
593 work
->cf_parent
= cf_parent
;
594 work
->cf_list
= cf_list
;
595 work
->start_block
= start_block
;
596 list_addtail(&work
->link
, work_list
);
599 /* Processes a block and returns the next block to process or NULL if we've
600 * reached the end of the construct.
602 static struct vtn_block
*
603 vtn_process_block(struct vtn_builder
*b
,
604 struct list_head
*work_list
,
605 struct vtn_cf_node
*cf_parent
,
606 struct list_head
*cf_list
,
607 struct vtn_block
*block
)
609 if (!list_is_empty(cf_list
)) {
610 /* vtn_process_block() acts like an iterator: it processes the given
611 * block and then returns the next block to process. For a given
612 * control-flow construct, vtn_build_cfg() calls vtn_process_block()
613 * repeatedly until it finally returns NULL. Therefore, we know that
614 * the only blocks on which vtn_process_block() can be called are either
615 * the first block in a construct or a block that vtn_process_block()
616 * returned for the current construct. If cf_list is empty then we know
617 * that we're processing the first block in the construct and we have to
618 * add it to the list.
620 * If cf_list is not empty, then it must be the block returned by the
621 * previous call to vtn_process_block(). We know a priori that
622 * vtn_process_block only returns either normal branches
623 * (vtn_branch_type_none) or merge target blocks.
625 switch (vtn_handle_branch(b
, cf_parent
, block
)) {
626 case vtn_branch_type_none
:
627 /* For normal branches, we want to process them and add them to the
628 * current construct. Merge target blocks also look like normal
629 * branches from the perspective of this construct. See also
630 * vtn_handle_branch().
634 case vtn_branch_type_loop_continue
:
635 case vtn_branch_type_switch_fallthrough
:
636 /* The two cases where we can get early exits from a construct that
637 * are not to that construct's merge target are loop continues and
638 * switch fall-throughs. In these cases, we need to break out of the
639 * current construct by returning NULL.
644 /* The only way we can get here is if something was used as two kinds
645 * of merges at the same time and that's illegal.
647 vtn_fail("A block was used as a merge target from two or more "
648 "structured control-flow constructs");
652 /* Once a block has been processed, it is placed into and the list link
653 * will point to something non-null. If we see a node we've already
654 * processed here, it either exists in multiple functions or it's an
657 if (block
->node
.parent
!= NULL
) {
658 vtn_fail_if(vtn_cf_node_find_function(&block
->node
) !=
659 vtn_cf_node_find_function(cf_parent
),
660 "A block cannot exist in two functions at the "
663 vtn_fail("Invalid back or cross-edge in the CFG");
666 if (block
->merge
&& (*block
->merge
& SpvOpCodeMask
) == SpvOpLoopMerge
&&
667 block
->loop
== NULL
) {
668 vtn_fail_if((*block
->branch
& SpvOpCodeMask
) != SpvOpBranch
&&
669 (*block
->branch
& SpvOpCodeMask
) != SpvOpBranchConditional
,
670 "An OpLoopMerge instruction must immediately precede "
671 "either an OpBranch or OpBranchConditional instruction.");
673 struct vtn_loop
*loop
= rzalloc(b
, struct vtn_loop
);
675 loop
->node
.type
= vtn_cf_node_type_loop
;
676 loop
->node
.parent
= cf_parent
;
677 list_inithead(&loop
->body
);
678 list_inithead(&loop
->cont_body
);
679 loop
->header_block
= block
;
680 loop
->break_block
= vtn_block(b
, block
->merge
[1]);
681 loop
->cont_block
= vtn_block(b
, block
->merge
[2]);
682 loop
->control
= block
->merge
[3];
684 list_addtail(&loop
->node
.link
, cf_list
);
687 /* Note: The work item for the main loop body will start with the
688 * current block as its start block. If we weren't careful, we would
689 * get here again and end up in an infinite loop. This is why we set
690 * block->loop above and check for it before creating one. This way,
691 * we only create the loop once and the second iteration that tries to
692 * handle this loop goes to the cases below and gets handled as a
695 vtn_add_cfg_work_item(b
, work_list
, &loop
->node
,
696 &loop
->body
, loop
->header_block
);
698 /* For continue targets, SPIR-V guarantees the following:
700 * - the Continue Target must dominate the back-edge block
701 * - the back-edge block must post dominate the Continue Target
703 * If the header block is the same as the continue target, this
704 * condition is trivially satisfied and there is no real continue
707 if (loop
->cont_block
!= loop
->header_block
) {
708 vtn_add_cfg_work_item(b
, work_list
, &loop
->node
,
709 &loop
->cont_body
, loop
->cont_block
);
712 vtn_block_set_merge_cf_node(b
, loop
->break_block
, &loop
->node
);
714 return loop
->break_block
;
717 /* Add the block to the CF list */
718 block
->node
.parent
= cf_parent
;
719 list_addtail(&block
->node
.link
, cf_list
);
721 switch (*block
->branch
& SpvOpCodeMask
) {
723 struct vtn_block
*branch_block
= vtn_block(b
, block
->branch
[1]);
725 block
->branch_type
= vtn_handle_branch(b
, cf_parent
, branch_block
);
727 if (block
->branch_type
== vtn_branch_type_none
)
734 case SpvOpReturnValue
:
735 block
->branch_type
= vtn_branch_type_return
;
739 block
->branch_type
= vtn_branch_type_discard
;
742 case SpvOpBranchConditional
: {
743 struct vtn_value
*cond_val
= vtn_untyped_value(b
, block
->branch
[1]);
744 vtn_fail_if(!cond_val
->type
||
745 cond_val
->type
->base_type
!= vtn_base_type_scalar
||
746 cond_val
->type
->type
!= glsl_bool_type(),
747 "Condition must be a Boolean type scalar");
749 struct vtn_block
*then_block
= vtn_block(b
, block
->branch
[2]);
750 struct vtn_block
*else_block
= vtn_block(b
, block
->branch
[3]);
752 if (then_block
== else_block
) {
753 /* This is uncommon but it can happen. We treat this the same way as
754 * an unconditional branch.
756 block
->branch_type
= vtn_handle_branch(b
, cf_parent
, then_block
);
758 if (block
->branch_type
== vtn_branch_type_none
)
764 struct vtn_if
*if_stmt
= rzalloc(b
, struct vtn_if
);
766 if_stmt
->node
.type
= vtn_cf_node_type_if
;
767 if_stmt
->node
.parent
= cf_parent
;
768 if_stmt
->condition
= block
->branch
[1];
769 list_inithead(&if_stmt
->then_body
);
770 list_inithead(&if_stmt
->else_body
);
772 list_addtail(&if_stmt
->node
.link
, cf_list
);
775 (*block
->merge
& SpvOpCodeMask
) == SpvOpSelectionMerge
) {
776 /* We may not always have a merge block and that merge doesn't
777 * technically have to be an OpSelectionMerge. We could have a block
778 * with an OpLoopMerge which ends in an OpBranchConditional.
780 if_stmt
->merge_block
= vtn_block(b
, block
->merge
[1]);
781 vtn_block_set_merge_cf_node(b
, if_stmt
->merge_block
, &if_stmt
->node
);
783 if_stmt
->control
= block
->merge
[2];
786 if_stmt
->then_type
= vtn_handle_branch(b
, &if_stmt
->node
, then_block
);
787 if (if_stmt
->then_type
== vtn_branch_type_none
) {
788 vtn_add_cfg_work_item(b
, work_list
, &if_stmt
->node
,
789 &if_stmt
->then_body
, then_block
);
792 if_stmt
->else_type
= vtn_handle_branch(b
, &if_stmt
->node
, else_block
);
793 if (if_stmt
->else_type
== vtn_branch_type_none
) {
794 vtn_add_cfg_work_item(b
, work_list
, &if_stmt
->node
,
795 &if_stmt
->else_body
, else_block
);
798 return if_stmt
->merge_block
;
802 struct vtn_value
*sel_val
= vtn_untyped_value(b
, block
->branch
[1]);
803 vtn_fail_if(!sel_val
->type
||
804 sel_val
->type
->base_type
!= vtn_base_type_scalar
,
805 "Selector of OpSwitch must have a type of OpTypeInt");
807 nir_alu_type sel_type
=
808 nir_get_nir_type_for_glsl_type(sel_val
->type
->type
);
809 vtn_fail_if(nir_alu_type_get_base_type(sel_type
) != nir_type_int
&&
810 nir_alu_type_get_base_type(sel_type
) != nir_type_uint
,
811 "Selector of OpSwitch must have a type of OpTypeInt");
813 struct vtn_switch
*swtch
= rzalloc(b
, struct vtn_switch
);
815 swtch
->node
.type
= vtn_cf_node_type_switch
;
816 swtch
->node
.parent
= cf_parent
;
817 swtch
->selector
= block
->branch
[1];
818 list_inithead(&swtch
->cases
);
820 list_addtail(&swtch
->node
.link
, cf_list
);
822 /* We may not always have a merge block */
824 vtn_fail_if((*block
->merge
& SpvOpCodeMask
) != SpvOpSelectionMerge
,
825 "An OpLoopMerge instruction must immediately precede "
826 "either an OpBranch or OpBranchConditional "
828 swtch
->break_block
= vtn_block(b
, block
->merge
[1]);
829 vtn_block_set_merge_cf_node(b
, swtch
->break_block
, &swtch
->node
);
832 /* First, we go through and record all of the cases. */
833 const uint32_t *branch_end
=
834 block
->branch
+ (block
->branch
[0] >> SpvWordCountShift
);
836 struct hash_table
*block_to_case
= _mesa_pointer_hash_table_create(b
);
838 bool is_default
= true;
839 const unsigned bitsize
= nir_alu_type_get_type_size(sel_type
);
840 for (const uint32_t *w
= block
->branch
+ 2; w
< branch_end
;) {
841 uint64_t literal
= 0;
846 assert(bitsize
== 64);
847 literal
= vtn_u64_literal(w
);
851 struct vtn_block
*case_block
= vtn_block(b
, *(w
++));
853 struct hash_entry
*case_entry
=
854 _mesa_hash_table_search(block_to_case
, case_block
);
856 struct vtn_case
*cse
;
858 cse
= case_entry
->data
;
860 cse
= rzalloc(b
, struct vtn_case
);
862 cse
->node
.type
= vtn_cf_node_type_case
;
863 cse
->node
.parent
= &swtch
->node
;
864 list_inithead(&cse
->body
);
865 util_dynarray_init(&cse
->values
, b
);
867 cse
->type
= vtn_handle_branch(b
, &swtch
->node
, case_block
);
869 case vtn_branch_type_none
:
870 /* This is a "real" cases which has stuff in it */
871 vtn_fail_if(case_block
->switch_case
!= NULL
,
872 "OpSwitch has a case which is also in another "
873 "OpSwitch construct");
874 case_block
->switch_case
= cse
;
875 vtn_add_cfg_work_item(b
, work_list
, &cse
->node
,
876 &cse
->body
, case_block
);
879 case vtn_branch_type_switch_break
:
880 case vtn_branch_type_loop_break
:
881 case vtn_branch_type_loop_continue
:
882 /* Switch breaks as well as loop breaks and continues can be
883 * used to break out of a switch construct or as direct targets
889 vtn_fail("Target of OpSwitch is not a valid structured exit "
890 "from the switch construct.");
893 list_addtail(&cse
->node
.link
, &swtch
->cases
);
895 _mesa_hash_table_insert(block_to_case
, case_block
, cse
);
899 cse
->is_default
= true;
901 util_dynarray_append(&cse
->values
, uint64_t, literal
);
907 _mesa_hash_table_destroy(block_to_case
, NULL
);
909 return swtch
->break_block
;
912 case SpvOpUnreachable
:
916 vtn_fail("Block did not end with a valid branch instruction");
921 vtn_build_cfg(struct vtn_builder
*b
, const uint32_t *words
, const uint32_t *end
)
923 vtn_foreach_instruction(b
, words
, end
,
924 vtn_cfg_handle_prepass_instruction
);
926 vtn_foreach_cf_node(func_node
, &b
->functions
) {
927 struct vtn_function
*func
= vtn_cf_node_as_function(func_node
);
929 /* We build the CFG for each function by doing a breadth-first search on
930 * the control-flow graph. We keep track of our state using a worklist.
931 * Doing a BFS ensures that we visit each structured control-flow
932 * construct and its merge node before we visit the stuff inside the
935 struct list_head work_list
;
936 list_inithead(&work_list
);
937 vtn_add_cfg_work_item(b
, &work_list
, &func
->node
, &func
->body
,
940 while (!list_is_empty(&work_list
)) {
941 struct vtn_cfg_work_item
*work
=
942 list_first_entry(&work_list
, struct vtn_cfg_work_item
, link
);
943 list_del(&work
->link
);
945 for (struct vtn_block
*block
= work
->start_block
; block
; ) {
946 block
= vtn_process_block(b
, &work_list
, work
->cf_parent
,
947 work
->cf_list
, block
);
954 vtn_handle_phis_first_pass(struct vtn_builder
*b
, SpvOp opcode
,
955 const uint32_t *w
, unsigned count
)
957 if (opcode
== SpvOpLabel
)
958 return true; /* Nothing to do */
960 /* If this isn't a phi node, stop. */
961 if (opcode
!= SpvOpPhi
)
964 /* For handling phi nodes, we do a poor-man's out-of-ssa on the spot.
965 * For each phi, we create a variable with the appropreate type and
966 * do a load from that variable. Then, in a second pass, we add
967 * stores to that variable to each of the predecessor blocks.
969 * We could do something more intelligent here. However, in order to
970 * handle loops and things properly, we really need dominance
971 * information. It would end up basically being the into-SSA
972 * algorithm all over again. It's easier if we just let
973 * lower_vars_to_ssa do that for us instead of repeating it here.
975 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
976 nir_variable
*phi_var
=
977 nir_local_variable_create(b
->nb
.impl
, type
->type
, "phi");
978 _mesa_hash_table_insert(b
->phi_table
, w
, phi_var
);
980 vtn_push_ssa_value(b
, w
[2],
981 vtn_local_load(b
, nir_build_deref_var(&b
->nb
, phi_var
), 0));
987 vtn_handle_phi_second_pass(struct vtn_builder
*b
, SpvOp opcode
,
988 const uint32_t *w
, unsigned count
)
990 if (opcode
!= SpvOpPhi
)
993 struct hash_entry
*phi_entry
= _mesa_hash_table_search(b
->phi_table
, w
);
995 /* It's possible that this phi is in an unreachable block in which case it
996 * may never have been emitted and therefore may not be in the hash table.
997 * In this case, there's no var for it and it's safe to just bail.
999 if (phi_entry
== NULL
)
1002 nir_variable
*phi_var
= phi_entry
->data
;
1004 for (unsigned i
= 3; i
< count
; i
+= 2) {
1005 struct vtn_block
*pred
= vtn_block(b
, w
[i
+ 1]);
1007 /* If block does not have end_nop, that is because it is an unreacheable
1008 * block, and hence it is not worth to handle it */
1012 b
->nb
.cursor
= nir_after_instr(&pred
->end_nop
->instr
);
1014 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[i
]);
1016 vtn_local_store(b
, src
, nir_build_deref_var(&b
->nb
, phi_var
), 0);
1023 vtn_emit_branch(struct vtn_builder
*b
, enum vtn_branch_type branch_type
,
1024 nir_variable
*switch_fall_var
, bool *has_switch_break
)
1026 switch (branch_type
) {
1027 case vtn_branch_type_if_merge
:
1028 break; /* Nothing to do */
1029 case vtn_branch_type_switch_break
:
1030 nir_store_var(&b
->nb
, switch_fall_var
, nir_imm_false(&b
->nb
), 1);
1031 *has_switch_break
= true;
1033 case vtn_branch_type_switch_fallthrough
:
1034 break; /* Nothing to do */
1035 case vtn_branch_type_loop_break
:
1036 nir_jump(&b
->nb
, nir_jump_break
);
1038 case vtn_branch_type_loop_continue
:
1039 nir_jump(&b
->nb
, nir_jump_continue
);
1041 case vtn_branch_type_loop_back_edge
:
1043 case vtn_branch_type_return
:
1044 nir_jump(&b
->nb
, nir_jump_return
);
1046 case vtn_branch_type_discard
: {
1047 nir_intrinsic_instr
*discard
=
1048 nir_intrinsic_instr_create(b
->nb
.shader
, nir_intrinsic_discard
);
1049 nir_builder_instr_insert(&b
->nb
, &discard
->instr
);
1053 vtn_fail("Invalid branch type");
1057 static nir_ssa_def
*
1058 vtn_switch_case_condition(struct vtn_builder
*b
, struct vtn_switch
*swtch
,
1059 nir_ssa_def
*sel
, struct vtn_case
*cse
)
1061 if (cse
->is_default
) {
1062 nir_ssa_def
*any
= nir_imm_false(&b
->nb
);
1063 vtn_foreach_cf_node(other_node
, &swtch
->cases
) {
1064 struct vtn_case
*other
= vtn_cf_node_as_case(other_node
);
1065 if (other
->is_default
)
1068 any
= nir_ior(&b
->nb
, any
,
1069 vtn_switch_case_condition(b
, swtch
, sel
, other
));
1071 return nir_inot(&b
->nb
, any
);
1073 nir_ssa_def
*cond
= nir_imm_false(&b
->nb
);
1074 util_dynarray_foreach(&cse
->values
, uint64_t, val
) {
1075 nir_ssa_def
*imm
= nir_imm_intN_t(&b
->nb
, *val
, sel
->bit_size
);
1076 cond
= nir_ior(&b
->nb
, cond
, nir_ieq(&b
->nb
, sel
, imm
));
1082 static nir_loop_control
1083 vtn_loop_control(struct vtn_builder
*b
, struct vtn_loop
*vtn_loop
)
1085 if (vtn_loop
->control
== SpvLoopControlMaskNone
)
1086 return nir_loop_control_none
;
1087 else if (vtn_loop
->control
& SpvLoopControlDontUnrollMask
)
1088 return nir_loop_control_dont_unroll
;
1089 else if (vtn_loop
->control
& SpvLoopControlUnrollMask
)
1090 return nir_loop_control_unroll
;
1091 else if (vtn_loop
->control
& SpvLoopControlDependencyInfiniteMask
||
1092 vtn_loop
->control
& SpvLoopControlDependencyLengthMask
||
1093 vtn_loop
->control
& SpvLoopControlMinIterationsMask
||
1094 vtn_loop
->control
& SpvLoopControlMaxIterationsMask
||
1095 vtn_loop
->control
& SpvLoopControlIterationMultipleMask
||
1096 vtn_loop
->control
& SpvLoopControlPeelCountMask
||
1097 vtn_loop
->control
& SpvLoopControlPartialCountMask
) {
1098 /* We do not do anything special with these yet. */
1099 return nir_loop_control_none
;
1101 vtn_fail("Invalid loop control");
1105 static nir_selection_control
1106 vtn_selection_control(struct vtn_builder
*b
, struct vtn_if
*vtn_if
)
1108 if (vtn_if
->control
== SpvSelectionControlMaskNone
)
1109 return nir_selection_control_none
;
1110 else if (vtn_if
->control
& SpvSelectionControlDontFlattenMask
)
1111 return nir_selection_control_dont_flatten
;
1112 else if (vtn_if
->control
& SpvSelectionControlFlattenMask
)
1113 return nir_selection_control_flatten
;
1115 vtn_fail("Invalid selection control");
1119 vtn_emit_cf_list(struct vtn_builder
*b
, struct list_head
*cf_list
,
1120 nir_variable
*switch_fall_var
, bool *has_switch_break
,
1121 vtn_instruction_handler handler
)
1123 vtn_foreach_cf_node(node
, cf_list
) {
1124 switch (node
->type
) {
1125 case vtn_cf_node_type_block
: {
1126 struct vtn_block
*block
= vtn_cf_node_as_block(node
);
1128 const uint32_t *block_start
= block
->label
;
1129 const uint32_t *block_end
= block
->merge
? block
->merge
:
1132 block_start
= vtn_foreach_instruction(b
, block_start
, block_end
,
1133 vtn_handle_phis_first_pass
);
1135 vtn_foreach_instruction(b
, block_start
, block_end
, handler
);
1137 block
->end_nop
= nir_intrinsic_instr_create(b
->nb
.shader
,
1139 nir_builder_instr_insert(&b
->nb
, &block
->end_nop
->instr
);
1141 if ((*block
->branch
& SpvOpCodeMask
) == SpvOpReturnValue
) {
1142 vtn_fail_if(b
->func
->type
->return_type
->base_type
==
1144 "Return with a value from a function returning void");
1145 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, block
->branch
[1]);
1146 const struct glsl_type
*ret_type
=
1147 glsl_get_bare_type(b
->func
->type
->return_type
->type
);
1148 nir_deref_instr
*ret_deref
=
1149 nir_build_deref_cast(&b
->nb
, nir_load_param(&b
->nb
, 0),
1150 nir_var_function_temp
, ret_type
, 0);
1151 vtn_local_store(b
, src
, ret_deref
, 0);
1154 if (block
->branch_type
!= vtn_branch_type_none
) {
1155 vtn_emit_branch(b
, block
->branch_type
,
1156 switch_fall_var
, has_switch_break
);
1163 case vtn_cf_node_type_if
: {
1164 struct vtn_if
*vtn_if
= vtn_cf_node_as_if(node
);
1165 bool sw_break
= false;
1168 nir_push_if(&b
->nb
, vtn_get_nir_ssa(b
, vtn_if
->condition
));
1170 nif
->control
= vtn_selection_control(b
, vtn_if
);
1172 if (vtn_if
->then_type
== vtn_branch_type_none
) {
1173 vtn_emit_cf_list(b
, &vtn_if
->then_body
,
1174 switch_fall_var
, &sw_break
, handler
);
1176 vtn_emit_branch(b
, vtn_if
->then_type
, switch_fall_var
, &sw_break
);
1179 nir_push_else(&b
->nb
, nif
);
1180 if (vtn_if
->else_type
== vtn_branch_type_none
) {
1181 vtn_emit_cf_list(b
, &vtn_if
->else_body
,
1182 switch_fall_var
, &sw_break
, handler
);
1184 vtn_emit_branch(b
, vtn_if
->else_type
, switch_fall_var
, &sw_break
);
1187 nir_pop_if(&b
->nb
, nif
);
1189 /* If we encountered a switch break somewhere inside of the if,
1190 * then it would have been handled correctly by calling
1191 * emit_cf_list or emit_branch for the interrior. However, we
1192 * need to predicate everything following on wether or not we're
1196 *has_switch_break
= true;
1197 nir_push_if(&b
->nb
, nir_load_var(&b
->nb
, switch_fall_var
));
1202 case vtn_cf_node_type_loop
: {
1203 struct vtn_loop
*vtn_loop
= vtn_cf_node_as_loop(node
);
1205 nir_loop
*loop
= nir_push_loop(&b
->nb
);
1206 loop
->control
= vtn_loop_control(b
, vtn_loop
);
1208 vtn_emit_cf_list(b
, &vtn_loop
->body
, NULL
, NULL
, handler
);
1210 if (!list_is_empty(&vtn_loop
->cont_body
)) {
1211 /* If we have a non-trivial continue body then we need to put
1212 * it at the beginning of the loop with a flag to ensure that
1213 * it doesn't get executed in the first iteration.
1215 nir_variable
*do_cont
=
1216 nir_local_variable_create(b
->nb
.impl
, glsl_bool_type(), "cont");
1218 b
->nb
.cursor
= nir_before_cf_node(&loop
->cf_node
);
1219 nir_store_var(&b
->nb
, do_cont
, nir_imm_false(&b
->nb
), 1);
1221 b
->nb
.cursor
= nir_before_cf_list(&loop
->body
);
1224 nir_push_if(&b
->nb
, nir_load_var(&b
->nb
, do_cont
));
1226 vtn_emit_cf_list(b
, &vtn_loop
->cont_body
, NULL
, NULL
, handler
);
1228 nir_pop_if(&b
->nb
, cont_if
);
1230 nir_store_var(&b
->nb
, do_cont
, nir_imm_true(&b
->nb
), 1);
1232 b
->has_loop_continue
= true;
1235 nir_pop_loop(&b
->nb
, loop
);
1239 case vtn_cf_node_type_switch
: {
1240 struct vtn_switch
*vtn_switch
= vtn_cf_node_as_switch(node
);
1242 /* Before we can emit anything, we need to sort the list of cases in
1243 * fall-through order.
1245 vtn_switch_order_cases(vtn_switch
);
1247 /* First, we create a variable to keep track of whether or not the
1248 * switch is still going at any given point. Any switch breaks
1249 * will set this variable to false.
1251 nir_variable
*fall_var
=
1252 nir_local_variable_create(b
->nb
.impl
, glsl_bool_type(), "fall");
1253 nir_store_var(&b
->nb
, fall_var
, nir_imm_false(&b
->nb
), 1);
1255 nir_ssa_def
*sel
= vtn_get_nir_ssa(b
, vtn_switch
->selector
);
1257 /* Now we can walk the list of cases and actually emit code */
1258 vtn_foreach_cf_node(case_node
, &vtn_switch
->cases
) {
1259 struct vtn_case
*cse
= vtn_cf_node_as_case(case_node
);
1261 /* Figure out the condition */
1263 vtn_switch_case_condition(b
, vtn_switch
, sel
, cse
);
1264 /* Take fallthrough into account */
1265 cond
= nir_ior(&b
->nb
, cond
, nir_load_var(&b
->nb
, fall_var
));
1267 nir_if
*case_if
= nir_push_if(&b
->nb
, cond
);
1269 bool has_break
= false;
1270 nir_store_var(&b
->nb
, fall_var
, nir_imm_true(&b
->nb
), 1);
1271 vtn_emit_cf_list(b
, &cse
->body
, fall_var
, &has_break
, handler
);
1272 (void)has_break
; /* We don't care */
1274 nir_pop_if(&b
->nb
, case_if
);
1281 vtn_fail("Invalid CF node type");
1287 vtn_function_emit(struct vtn_builder
*b
, struct vtn_function
*func
,
1288 vtn_instruction_handler instruction_handler
)
1290 nir_builder_init(&b
->nb
, func
->impl
);
1292 b
->nb
.cursor
= nir_after_cf_list(&func
->impl
->body
);
1293 b
->nb
.exact
= b
->exact
;
1294 b
->has_loop_continue
= false;
1295 b
->phi_table
= _mesa_pointer_hash_table_create(b
);
1297 vtn_emit_cf_list(b
, &func
->body
, NULL
, NULL
, instruction_handler
);
1299 vtn_foreach_instruction(b
, func
->start_block
->label
, func
->end
,
1300 vtn_handle_phi_second_pass
);
1302 nir_rematerialize_derefs_in_use_blocks_impl(func
->impl
);
1304 /* Continue blocks for loops get inserted before the body of the loop
1305 * but instructions in the continue may use SSA defs in the loop body.
1306 * Therefore, we need to repair SSA to insert the needed phi nodes.
1308 if (b
->has_loop_continue
)
1309 nir_repair_ssa_impl(func
->impl
);
1311 func
->emitted
= true;