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 if (param_type
->base_type
!= vtn_base_type_pointer
) {
40 assert(param_type
->base_type
== vtn_base_type_image
||
41 param_type
->base_type
== vtn_base_type_sampler
);
42 ptr_type
= rzalloc(b
, struct vtn_type
);
43 ptr_type
->base_type
= vtn_base_type_pointer
;
44 ptr_type
->deref
= param_type
;
45 ptr_type
->storage_class
= SpvStorageClassUniformConstant
;
48 return vtn_pointer_from_ssa(b
, nir_load_param(&b
->nb
, param_idx
), ptr_type
);
52 vtn_type_count_function_params(struct vtn_type
*type
)
54 switch (type
->base_type
) {
55 case vtn_base_type_array
:
56 case vtn_base_type_matrix
:
57 return type
->length
* vtn_type_count_function_params(type
->array_element
);
59 case vtn_base_type_struct
: {
61 for (unsigned i
= 0; i
< type
->length
; i
++)
62 count
+= vtn_type_count_function_params(type
->members
[i
]);
66 case vtn_base_type_sampled_image
:
75 vtn_type_add_to_function_params(struct vtn_type
*type
,
79 static const nir_parameter nir_deref_param
= {
84 switch (type
->base_type
) {
85 case vtn_base_type_array
:
86 case vtn_base_type_matrix
:
87 for (unsigned i
= 0; i
< type
->length
; i
++)
88 vtn_type_add_to_function_params(type
->array_element
, func
, param_idx
);
91 case vtn_base_type_struct
:
92 for (unsigned i
= 0; i
< type
->length
; i
++)
93 vtn_type_add_to_function_params(type
->members
[i
], func
, param_idx
);
96 case vtn_base_type_sampled_image
:
97 func
->params
[(*param_idx
)++] = nir_deref_param
;
98 func
->params
[(*param_idx
)++] = nir_deref_param
;
101 case vtn_base_type_image
:
102 case vtn_base_type_sampler
:
103 func
->params
[(*param_idx
)++] = nir_deref_param
;
106 case vtn_base_type_pointer
:
108 func
->params
[(*param_idx
)++] = (nir_parameter
) {
109 .num_components
= glsl_get_vector_elements(type
->type
),
110 .bit_size
= glsl_get_bit_size(type
->type
),
113 func
->params
[(*param_idx
)++] = nir_deref_param
;
118 func
->params
[(*param_idx
)++] = (nir_parameter
) {
119 .num_components
= glsl_get_vector_elements(type
->type
),
120 .bit_size
= glsl_get_bit_size(type
->type
),
126 vtn_ssa_value_add_to_call_params(struct vtn_builder
*b
,
127 struct vtn_ssa_value
*value
,
128 struct vtn_type
*type
,
129 nir_call_instr
*call
,
132 switch (type
->base_type
) {
133 case vtn_base_type_array
:
134 case vtn_base_type_matrix
:
135 for (unsigned i
= 0; i
< type
->length
; i
++) {
136 vtn_ssa_value_add_to_call_params(b
, value
->elems
[i
],
142 case vtn_base_type_struct
:
143 for (unsigned i
= 0; i
< type
->length
; i
++) {
144 vtn_ssa_value_add_to_call_params(b
, value
->elems
[i
],
151 call
->params
[(*param_idx
)++] = nir_src_for_ssa(value
->def
);
157 vtn_ssa_value_load_function_param(struct vtn_builder
*b
,
158 struct vtn_ssa_value
*value
,
159 struct vtn_type
*type
,
162 switch (type
->base_type
) {
163 case vtn_base_type_array
:
164 case vtn_base_type_matrix
:
165 for (unsigned i
= 0; i
< type
->length
; i
++) {
166 vtn_ssa_value_load_function_param(b
, value
->elems
[i
],
167 type
->array_element
, param_idx
);
171 case vtn_base_type_struct
:
172 for (unsigned i
= 0; i
< type
->length
; i
++) {
173 vtn_ssa_value_load_function_param(b
, value
->elems
[i
],
174 type
->members
[i
], param_idx
);
179 value
->def
= nir_load_param(&b
->nb
, (*param_idx
)++);
185 vtn_handle_function_call(struct vtn_builder
*b
, SpvOp opcode
,
186 const uint32_t *w
, unsigned count
)
188 struct vtn_function
*vtn_callee
=
189 vtn_value(b
, w
[3], vtn_value_type_function
)->func
;
190 struct nir_function
*callee
= vtn_callee
->impl
->function
;
192 vtn_callee
->referenced
= true;
194 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
196 unsigned param_idx
= 0;
198 nir_deref_instr
*ret_deref
= NULL
;
199 struct vtn_type
*ret_type
= vtn_callee
->type
->return_type
;
200 if (ret_type
->base_type
!= vtn_base_type_void
) {
201 nir_variable
*ret_tmp
=
202 nir_local_variable_create(b
->nb
.impl
,
203 glsl_get_bare_type(ret_type
->type
),
205 ret_deref
= nir_build_deref_var(&b
->nb
, ret_tmp
);
206 call
->params
[param_idx
++] = nir_src_for_ssa(&ret_deref
->dest
.ssa
);
209 for (unsigned i
= 0; i
< vtn_callee
->type
->length
; i
++) {
210 struct vtn_type
*arg_type
= vtn_callee
->type
->params
[i
];
211 unsigned arg_id
= w
[4 + i
];
213 if (arg_type
->base_type
== vtn_base_type_sampled_image
) {
214 struct vtn_sampled_image
*sampled_image
=
215 vtn_value(b
, arg_id
, vtn_value_type_sampled_image
)->sampled_image
;
217 call
->params
[param_idx
++] =
218 nir_src_for_ssa(vtn_pointer_to_ssa(b
, sampled_image
->image
));
219 call
->params
[param_idx
++] =
220 nir_src_for_ssa(vtn_pointer_to_ssa(b
, sampled_image
->sampler
));
221 } else if (arg_type
->base_type
== vtn_base_type_pointer
||
222 arg_type
->base_type
== vtn_base_type_image
||
223 arg_type
->base_type
== vtn_base_type_sampler
) {
224 struct vtn_pointer
*pointer
=
225 vtn_value(b
, arg_id
, vtn_value_type_pointer
)->pointer
;
226 call
->params
[param_idx
++] =
227 nir_src_for_ssa(vtn_pointer_to_ssa(b
, pointer
));
229 vtn_ssa_value_add_to_call_params(b
, vtn_ssa_value(b
, arg_id
),
230 arg_type
, call
, ¶m_idx
);
233 assert(param_idx
== call
->num_params
);
235 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
237 if (ret_type
->base_type
== vtn_base_type_void
) {
238 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
240 vtn_push_ssa_value(b
, w
[2], vtn_local_load(b
, ret_deref
, 0));
245 vtn_cfg_handle_prepass_instruction(struct vtn_builder
*b
, SpvOp opcode
,
246 const uint32_t *w
, unsigned count
)
249 case SpvOpFunction
: {
250 vtn_assert(b
->func
== NULL
);
251 b
->func
= rzalloc(b
, struct vtn_function
);
253 b
->func
->node
.type
= vtn_cf_node_type_function
;
254 b
->func
->node
.parent
= NULL
;
255 list_inithead(&b
->func
->body
);
256 b
->func
->control
= w
[3];
258 UNUSED
const struct glsl_type
*result_type
= vtn_get_type(b
, w
[1])->type
;
259 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_function
);
262 b
->func
->type
= vtn_get_type(b
, w
[4]);
263 const struct vtn_type
*func_type
= b
->func
->type
;
265 vtn_assert(func_type
->return_type
->type
== result_type
);
268 nir_function_create(b
->shader
, ralloc_strdup(b
->shader
, val
->name
));
270 unsigned num_params
= 0;
271 for (unsigned i
= 0; i
< func_type
->length
; i
++)
272 num_params
+= vtn_type_count_function_params(func_type
->params
[i
]);
274 /* Add one parameter for the function return value */
275 if (func_type
->return_type
->base_type
!= vtn_base_type_void
)
278 func
->num_params
= num_params
;
279 func
->params
= ralloc_array(b
->shader
, nir_parameter
, num_params
);
282 if (func_type
->return_type
->base_type
!= vtn_base_type_void
) {
283 nir_address_format addr_format
=
284 vtn_mode_to_address_format(b
, vtn_variable_mode_function
);
285 /* The return value is a regular pointer */
286 func
->params
[idx
++] = (nir_parameter
) {
287 .num_components
= nir_address_format_num_components(addr_format
),
288 .bit_size
= nir_address_format_bit_size(addr_format
),
292 for (unsigned i
= 0; i
< func_type
->length
; i
++)
293 vtn_type_add_to_function_params(func_type
->params
[i
], func
, &idx
);
294 assert(idx
== num_params
);
296 b
->func
->impl
= nir_function_impl_create(func
);
297 nir_builder_init(&b
->nb
, func
->impl
);
298 b
->nb
.cursor
= nir_before_cf_list(&b
->func
->impl
->body
);
299 b
->nb
.exact
= b
->exact
;
301 b
->func_param_idx
= 0;
303 /* The return value is the first parameter */
304 if (func_type
->return_type
->base_type
!= vtn_base_type_void
)
309 case SpvOpFunctionEnd
:
314 case SpvOpFunctionParameter
: {
315 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
317 vtn_assert(b
->func_param_idx
< b
->func
->impl
->function
->num_params
);
319 if (type
->base_type
== vtn_base_type_sampled_image
) {
320 /* Sampled images are actually two parameters. The first is the
321 * image and the second is the sampler.
323 struct vtn_value
*val
=
324 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
326 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
328 struct vtn_type
*image_type
= rzalloc(b
, struct vtn_type
);
329 image_type
->base_type
= vtn_base_type_image
;
330 image_type
->type
= type
->type
;
332 struct vtn_type
*sampler_type
= rzalloc(b
, struct vtn_type
);
333 sampler_type
->base_type
= vtn_base_type_sampler
;
334 sampler_type
->type
= glsl_bare_sampler_type();
336 val
->sampled_image
->image
=
337 vtn_load_param_pointer(b
, image_type
, b
->func_param_idx
++);
338 val
->sampled_image
->sampler
=
339 vtn_load_param_pointer(b
, sampler_type
, b
->func_param_idx
++);
340 } else if (type
->base_type
== vtn_base_type_pointer
&&
341 type
->type
!= NULL
) {
342 /* This is a pointer with an actual storage type */
343 nir_ssa_def
*ssa_ptr
= nir_load_param(&b
->nb
, b
->func_param_idx
++);
344 vtn_push_pointer(b
, w
[2], vtn_pointer_from_ssa(b
, ssa_ptr
, type
));
345 } else if (type
->base_type
== vtn_base_type_pointer
||
346 type
->base_type
== vtn_base_type_image
||
347 type
->base_type
== vtn_base_type_sampler
) {
348 vtn_push_pointer(b
, w
[2], vtn_load_param_pointer(b
, type
, b
->func_param_idx
++));
350 /* We're a regular SSA value. */
351 struct vtn_ssa_value
*value
= vtn_create_ssa_value(b
, type
->type
);
352 vtn_ssa_value_load_function_param(b
, value
, type
, &b
->func_param_idx
);
353 vtn_push_ssa_value(b
, w
[2], value
);
359 vtn_assert(b
->block
== NULL
);
360 b
->block
= rzalloc(b
, struct vtn_block
);
361 b
->block
->node
.type
= vtn_cf_node_type_block
;
363 vtn_push_value(b
, w
[1], vtn_value_type_block
)->block
= b
->block
;
365 if (b
->func
->start_block
== NULL
) {
366 /* This is the first block encountered for this function. In this
367 * case, we set the start block and add it to the list of
368 * implemented functions that we'll walk later.
370 b
->func
->start_block
= b
->block
;
371 list_addtail(&b
->func
->node
.link
, &b
->functions
);
376 case SpvOpSelectionMerge
:
378 vtn_assert(b
->block
&& b
->block
->merge
== NULL
);
383 case SpvOpBranchConditional
:
387 case SpvOpReturnValue
:
388 case SpvOpUnreachable
:
389 vtn_assert(b
->block
&& b
->block
->branch
== NULL
);
390 b
->block
->branch
= w
;
395 /* Continue on as per normal */
402 /* This function performs a depth-first search of the cases and puts them
403 * in fall-through order.
406 vtn_order_case(struct vtn_switch
*swtch
, struct vtn_case
*cse
)
413 list_del(&cse
->node
.link
);
415 if (cse
->fallthrough
) {
416 vtn_order_case(swtch
, cse
->fallthrough
);
418 /* If we have a fall-through, place this case right before the case it
419 * falls through to. This ensures that fallthroughs come one after
420 * the other. These two can never get separated because that would
421 * imply something else falling through to the same case. Also, this
422 * can't break ordering because the DFS ensures that this case is
423 * visited before anything that falls through to it.
425 list_addtail(&cse
->node
.link
, &cse
->fallthrough
->node
.link
);
427 list_add(&cse
->node
.link
, &swtch
->cases
);
432 vtn_switch_order_cases(struct vtn_switch
*swtch
)
434 struct list_head cases
;
435 list_replace(&swtch
->cases
, &cases
);
436 list_inithead(&swtch
->cases
);
437 while (!list_is_empty(&cases
)) {
438 struct vtn_case
*cse
=
439 list_first_entry(&cases
, struct vtn_case
, node
.link
);
440 vtn_order_case(swtch
, cse
);
445 vtn_block_set_merge_cf_node(struct vtn_builder
*b
, struct vtn_block
*block
,
446 struct vtn_cf_node
*cf_node
)
448 vtn_fail_if(block
->merge_cf_node
!= NULL
,
449 "The merge block declared by a header block cannot be a "
450 "merge block declared by any other header block.");
452 block
->merge_cf_node
= cf_node
;
455 #define VTN_DECL_CF_NODE_FIND(_type) \
456 static inline struct vtn_##_type * \
457 vtn_cf_node_find_##_type(struct vtn_cf_node *node) \
459 while (node && node->type != vtn_cf_node_type_##_type) \
460 node = node->parent; \
461 return (struct vtn_##_type *)node; \
464 VTN_DECL_CF_NODE_FIND(if)
465 VTN_DECL_CF_NODE_FIND(loop
)
466 VTN_DECL_CF_NODE_FIND(case)
467 VTN_DECL_CF_NODE_FIND(switch)
468 VTN_DECL_CF_NODE_FIND(function
)
470 static enum vtn_branch_type
471 vtn_handle_branch(struct vtn_builder
*b
,
472 struct vtn_cf_node
*cf_parent
,
473 struct vtn_block
*target_block
)
475 struct vtn_loop
*loop
= vtn_cf_node_find_loop(cf_parent
);
477 /* Detect a loop back-edge first. That way none of the code below
478 * accidentally operates on a loop back-edge.
480 if (loop
&& target_block
== loop
->header_block
)
481 return vtn_branch_type_loop_back_edge
;
483 /* Try to detect fall-through */
484 if (target_block
->switch_case
) {
485 /* When it comes to handling switch cases, we can break calls to
486 * vtn_handle_branch into two cases: calls from within a case construct
487 * and calls for the jump to each case construct. In the second case,
488 * cf_parent is the vtn_switch itself and vtn_cf_node_find_case() will
489 * return the outer switch case in which this switch is contained. It's
490 * fine if the target block is a switch case from an outer switch as
491 * long as it is also the switch break for this switch.
493 struct vtn_case
*switch_case
= vtn_cf_node_find_case(cf_parent
);
495 /* This doesn't get called for the OpSwitch */
496 vtn_fail_if(switch_case
== NULL
,
497 "A switch case can only be entered through an OpSwitch or "
498 "falling through from another switch case.");
500 /* Because block->switch_case is only set on the entry block for a given
501 * switch case, we only ever get here if we're jumping to the start of a
502 * switch case. It's possible, however, that a switch case could jump
503 * to itself via a back-edge. That *should* get caught by the loop
504 * handling case above but if we have a back edge without a loop merge,
505 * we could en up here.
507 vtn_fail_if(target_block
->switch_case
== switch_case
,
508 "A switch cannot fall-through to itself. Likely, there is "
509 "a back-edge which is not to a loop header.");
511 vtn_fail_if(target_block
->switch_case
->node
.parent
!=
512 switch_case
->node
.parent
,
513 "A switch case fall-through must come from the same "
514 "OpSwitch construct");
516 vtn_fail_if(switch_case
->fallthrough
!= NULL
&&
517 switch_case
->fallthrough
!= target_block
->switch_case
,
518 "Each case construct can have at most one branch to "
519 "another case construct");
521 switch_case
->fallthrough
= target_block
->switch_case
;
523 /* We don't immediately return vtn_branch_type_switch_fallthrough
524 * because it may also be a loop or switch break for an inner loop or
525 * switch and that takes precedence.
529 if (loop
&& target_block
== loop
->cont_block
)
530 return vtn_branch_type_loop_continue
;
532 /* We walk blocks as a breadth-first search on the control-flow construct
533 * tree where, when we find a construct, we add the vtn_cf_node for that
534 * construct and continue iterating at the merge target block (if any).
535 * Therefore, we want merges whose with parent == cf_parent to be treated
536 * as regular branches. We only want to consider merges if they break out
537 * of the current CF construct.
539 if (target_block
->merge_cf_node
!= NULL
&&
540 target_block
->merge_cf_node
->parent
!= cf_parent
) {
541 switch (target_block
->merge_cf_node
->type
) {
542 case vtn_cf_node_type_if
:
543 for (struct vtn_cf_node
*node
= cf_parent
;
544 node
!= target_block
->merge_cf_node
; node
= node
->parent
) {
545 vtn_fail_if(node
== NULL
|| node
->type
!= vtn_cf_node_type_if
,
546 "Branching to the merge block of a selection "
547 "construct can only be used to break out of a "
548 "selection construct");
550 struct vtn_if
*if_stmt
= vtn_cf_node_as_if(node
);
552 /* This should be guaranteed by our iteration */
553 assert(if_stmt
->merge_block
!= target_block
);
555 vtn_fail_if(if_stmt
->merge_block
!= NULL
,
556 "Branching to the merge block of a selection "
557 "construct can only be used to break out of the "
558 "inner most nested selection level");
560 return vtn_branch_type_if_merge
;
562 case vtn_cf_node_type_loop
:
563 vtn_fail_if(target_block
->merge_cf_node
!= &loop
->node
,
564 "Loop breaks can only break out of the inner most "
565 "nested loop level");
566 return vtn_branch_type_loop_break
;
568 case vtn_cf_node_type_switch
: {
569 struct vtn_switch
*swtch
= vtn_cf_node_find_switch(cf_parent
);
570 vtn_fail_if(target_block
->merge_cf_node
!= &swtch
->node
,
571 "Switch breaks can only break out of the inner most "
572 "nested switch level");
573 return vtn_branch_type_switch_break
;
577 unreachable("Invalid CF node type for a merge");
581 if (target_block
->switch_case
)
582 return vtn_branch_type_switch_fallthrough
;
584 return vtn_branch_type_none
;
587 struct vtn_cfg_work_item
{
588 struct list_head link
;
590 struct vtn_cf_node
*cf_parent
;
591 struct list_head
*cf_list
;
592 struct vtn_block
*start_block
;
596 vtn_add_cfg_work_item(struct vtn_builder
*b
,
597 struct list_head
*work_list
,
598 struct vtn_cf_node
*cf_parent
,
599 struct list_head
*cf_list
,
600 struct vtn_block
*start_block
)
602 struct vtn_cfg_work_item
*work
= ralloc(b
, struct vtn_cfg_work_item
);
603 work
->cf_parent
= cf_parent
;
604 work
->cf_list
= cf_list
;
605 work
->start_block
= start_block
;
606 list_addtail(&work
->link
, work_list
);
609 /* Processes a block and returns the next block to process or NULL if we've
610 * reached the end of the construct.
612 static struct vtn_block
*
613 vtn_process_block(struct vtn_builder
*b
,
614 struct list_head
*work_list
,
615 struct vtn_cf_node
*cf_parent
,
616 struct list_head
*cf_list
,
617 struct vtn_block
*block
)
619 if (!list_is_empty(cf_list
)) {
620 /* vtn_process_block() acts like an iterator: it processes the given
621 * block and then returns the next block to process. For a given
622 * control-flow construct, vtn_build_cfg() calls vtn_process_block()
623 * repeatedly until it finally returns NULL. Therefore, we know that
624 * the only blocks on which vtn_process_block() can be called are either
625 * the first block in a construct or a block that vtn_process_block()
626 * returned for the current construct. If cf_list is empty then we know
627 * that we're processing the first block in the construct and we have to
628 * add it to the list.
630 * If cf_list is not empty, then it must be the block returned by the
631 * previous call to vtn_process_block(). We know a priori that
632 * vtn_process_block only returns either normal branches
633 * (vtn_branch_type_none) or merge target blocks.
635 switch (vtn_handle_branch(b
, cf_parent
, block
)) {
636 case vtn_branch_type_none
:
637 /* For normal branches, we want to process them and add them to the
638 * current construct. Merge target blocks also look like normal
639 * branches from the perspective of this construct. See also
640 * vtn_handle_branch().
644 case vtn_branch_type_loop_continue
:
645 case vtn_branch_type_switch_fallthrough
:
646 /* The two cases where we can get early exits from a construct that
647 * are not to that construct's merge target are loop continues and
648 * switch fall-throughs. In these cases, we need to break out of the
649 * current construct by returning NULL.
654 /* The only way we can get here is if something was used as two kinds
655 * of merges at the same time and that's illegal.
657 vtn_fail("A block was used as a merge target from two or more "
658 "structured control-flow constructs");
662 /* Once a block has been processed, it is placed into and the list link
663 * will point to something non-null. If we see a node we've already
664 * processed here, it either exists in multiple functions or it's an
667 if (block
->node
.parent
!= NULL
) {
668 vtn_fail_if(vtn_cf_node_find_function(&block
->node
) !=
669 vtn_cf_node_find_function(cf_parent
),
670 "A block cannot exist in two functions at the "
673 vtn_fail("Invalid back or cross-edge in the CFG");
676 if (block
->merge
&& (*block
->merge
& SpvOpCodeMask
) == SpvOpLoopMerge
&&
677 block
->loop
== NULL
) {
678 vtn_fail_if((*block
->branch
& SpvOpCodeMask
) != SpvOpBranch
&&
679 (*block
->branch
& SpvOpCodeMask
) != SpvOpBranchConditional
,
680 "An OpLoopMerge instruction must immediately precede "
681 "either an OpBranch or OpBranchConditional instruction.");
683 struct vtn_loop
*loop
= rzalloc(b
, struct vtn_loop
);
685 loop
->node
.type
= vtn_cf_node_type_loop
;
686 loop
->node
.parent
= cf_parent
;
687 list_inithead(&loop
->body
);
688 list_inithead(&loop
->cont_body
);
689 loop
->header_block
= block
;
690 loop
->break_block
= vtn_block(b
, block
->merge
[1]);
691 loop
->cont_block
= vtn_block(b
, block
->merge
[2]);
692 loop
->control
= block
->merge
[3];
694 list_addtail(&loop
->node
.link
, cf_list
);
697 /* Note: The work item for the main loop body will start with the
698 * current block as its start block. If we weren't careful, we would
699 * get here again and end up in an infinite loop. This is why we set
700 * block->loop above and check for it before creating one. This way,
701 * we only create the loop once and the second iteration that tries to
702 * handle this loop goes to the cases below and gets handled as a
705 vtn_add_cfg_work_item(b
, work_list
, &loop
->node
,
706 &loop
->body
, loop
->header_block
);
708 /* For continue targets, SPIR-V guarantees the following:
710 * - the Continue Target must dominate the back-edge block
711 * - the back-edge block must post dominate the Continue Target
713 * If the header block is the same as the continue target, this
714 * condition is trivially satisfied and there is no real continue
717 if (loop
->cont_block
!= loop
->header_block
) {
718 vtn_add_cfg_work_item(b
, work_list
, &loop
->node
,
719 &loop
->cont_body
, loop
->cont_block
);
722 vtn_block_set_merge_cf_node(b
, loop
->break_block
, &loop
->node
);
724 return loop
->break_block
;
727 /* Add the block to the CF list */
728 block
->node
.parent
= cf_parent
;
729 list_addtail(&block
->node
.link
, cf_list
);
731 switch (*block
->branch
& SpvOpCodeMask
) {
733 struct vtn_block
*branch_block
= vtn_block(b
, block
->branch
[1]);
735 block
->branch_type
= vtn_handle_branch(b
, cf_parent
, branch_block
);
737 if (block
->branch_type
== vtn_branch_type_none
)
744 case SpvOpReturnValue
:
745 block
->branch_type
= vtn_branch_type_return
;
749 block
->branch_type
= vtn_branch_type_discard
;
752 case SpvOpBranchConditional
: {
753 struct vtn_value
*cond_val
= vtn_untyped_value(b
, block
->branch
[1]);
754 vtn_fail_if(!cond_val
->type
||
755 cond_val
->type
->base_type
!= vtn_base_type_scalar
||
756 cond_val
->type
->type
!= glsl_bool_type(),
757 "Condition must be a Boolean type scalar");
759 struct vtn_block
*then_block
= vtn_block(b
, block
->branch
[2]);
760 struct vtn_block
*else_block
= vtn_block(b
, block
->branch
[3]);
762 if (then_block
== else_block
) {
763 /* This is uncommon but it can happen. We treat this the same way as
764 * an unconditional branch.
766 block
->branch_type
= vtn_handle_branch(b
, cf_parent
, then_block
);
768 if (block
->branch_type
== vtn_branch_type_none
)
774 struct vtn_if
*if_stmt
= rzalloc(b
, struct vtn_if
);
776 if_stmt
->node
.type
= vtn_cf_node_type_if
;
777 if_stmt
->node
.parent
= cf_parent
;
778 if_stmt
->condition
= block
->branch
[1];
779 list_inithead(&if_stmt
->then_body
);
780 list_inithead(&if_stmt
->else_body
);
782 list_addtail(&if_stmt
->node
.link
, cf_list
);
785 (*block
->merge
& SpvOpCodeMask
) == SpvOpSelectionMerge
) {
786 /* We may not always have a merge block and that merge doesn't
787 * technically have to be an OpSelectionMerge. We could have a block
788 * with an OpLoopMerge which ends in an OpBranchConditional.
790 if_stmt
->merge_block
= vtn_block(b
, block
->merge
[1]);
791 vtn_block_set_merge_cf_node(b
, if_stmt
->merge_block
, &if_stmt
->node
);
793 if_stmt
->control
= block
->merge
[2];
796 if_stmt
->then_type
= vtn_handle_branch(b
, &if_stmt
->node
, then_block
);
797 if (if_stmt
->then_type
== vtn_branch_type_none
) {
798 vtn_add_cfg_work_item(b
, work_list
, &if_stmt
->node
,
799 &if_stmt
->then_body
, then_block
);
802 if_stmt
->else_type
= vtn_handle_branch(b
, &if_stmt
->node
, else_block
);
803 if (if_stmt
->else_type
== vtn_branch_type_none
) {
804 vtn_add_cfg_work_item(b
, work_list
, &if_stmt
->node
,
805 &if_stmt
->else_body
, else_block
);
808 return if_stmt
->merge_block
;
812 struct vtn_value
*sel_val
= vtn_untyped_value(b
, block
->branch
[1]);
813 vtn_fail_if(!sel_val
->type
||
814 sel_val
->type
->base_type
!= vtn_base_type_scalar
,
815 "Selector of OpSwitch must have a type of OpTypeInt");
817 nir_alu_type sel_type
=
818 nir_get_nir_type_for_glsl_type(sel_val
->type
->type
);
819 vtn_fail_if(nir_alu_type_get_base_type(sel_type
) != nir_type_int
&&
820 nir_alu_type_get_base_type(sel_type
) != nir_type_uint
,
821 "Selector of OpSwitch must have a type of OpTypeInt");
823 struct vtn_switch
*swtch
= rzalloc(b
, struct vtn_switch
);
825 swtch
->node
.type
= vtn_cf_node_type_switch
;
826 swtch
->node
.parent
= cf_parent
;
827 swtch
->selector
= block
->branch
[1];
828 list_inithead(&swtch
->cases
);
830 list_addtail(&swtch
->node
.link
, cf_list
);
832 /* We may not always have a merge block */
834 vtn_fail_if((*block
->merge
& SpvOpCodeMask
) != SpvOpSelectionMerge
,
835 "An OpLoopMerge instruction must immediately precede "
836 "either an OpBranch or OpBranchConditional "
838 swtch
->break_block
= vtn_block(b
, block
->merge
[1]);
839 vtn_block_set_merge_cf_node(b
, swtch
->break_block
, &swtch
->node
);
842 /* First, we go through and record all of the cases. */
843 const uint32_t *branch_end
=
844 block
->branch
+ (block
->branch
[0] >> SpvWordCountShift
);
846 struct hash_table
*block_to_case
= _mesa_pointer_hash_table_create(b
);
848 bool is_default
= true;
849 const unsigned bitsize
= nir_alu_type_get_type_size(sel_type
);
850 for (const uint32_t *w
= block
->branch
+ 2; w
< branch_end
;) {
851 uint64_t literal
= 0;
856 assert(bitsize
== 64);
857 literal
= vtn_u64_literal(w
);
861 struct vtn_block
*case_block
= vtn_block(b
, *(w
++));
863 struct hash_entry
*case_entry
=
864 _mesa_hash_table_search(block_to_case
, case_block
);
866 struct vtn_case
*cse
;
868 cse
= case_entry
->data
;
870 cse
= rzalloc(b
, struct vtn_case
);
872 cse
->node
.type
= vtn_cf_node_type_case
;
873 cse
->node
.parent
= &swtch
->node
;
874 list_inithead(&cse
->body
);
875 util_dynarray_init(&cse
->values
, b
);
877 cse
->type
= vtn_handle_branch(b
, &swtch
->node
, case_block
);
879 case vtn_branch_type_none
:
880 /* This is a "real" cases which has stuff in it */
881 vtn_fail_if(case_block
->switch_case
!= NULL
,
882 "OpSwitch has a case which is also in another "
883 "OpSwitch construct");
884 case_block
->switch_case
= cse
;
885 vtn_add_cfg_work_item(b
, work_list
, &cse
->node
,
886 &cse
->body
, case_block
);
889 case vtn_branch_type_switch_break
:
890 case vtn_branch_type_loop_break
:
891 case vtn_branch_type_loop_continue
:
892 /* Switch breaks as well as loop breaks and continues can be
893 * used to break out of a switch construct or as direct targets
899 vtn_fail("Target of OpSwitch is not a valid structured exit "
900 "from the switch construct.");
903 list_addtail(&cse
->node
.link
, &swtch
->cases
);
905 _mesa_hash_table_insert(block_to_case
, case_block
, cse
);
909 cse
->is_default
= true;
911 util_dynarray_append(&cse
->values
, uint64_t, literal
);
917 _mesa_hash_table_destroy(block_to_case
, NULL
);
919 return swtch
->break_block
;
922 case SpvOpUnreachable
:
926 vtn_fail("Block did not end with a valid branch instruction");
931 vtn_build_cfg(struct vtn_builder
*b
, const uint32_t *words
, const uint32_t *end
)
933 vtn_foreach_instruction(b
, words
, end
,
934 vtn_cfg_handle_prepass_instruction
);
936 vtn_foreach_cf_node(func_node
, &b
->functions
) {
937 struct vtn_function
*func
= vtn_cf_node_as_function(func_node
);
939 /* We build the CFG for each function by doing a breadth-first search on
940 * the control-flow graph. We keep track of our state using a worklist.
941 * Doing a BFS ensures that we visit each structured control-flow
942 * construct and its merge node before we visit the stuff inside the
945 struct list_head work_list
;
946 list_inithead(&work_list
);
947 vtn_add_cfg_work_item(b
, &work_list
, &func
->node
, &func
->body
,
950 while (!list_is_empty(&work_list
)) {
951 struct vtn_cfg_work_item
*work
=
952 list_first_entry(&work_list
, struct vtn_cfg_work_item
, link
);
953 list_del(&work
->link
);
955 for (struct vtn_block
*block
= work
->start_block
; block
; ) {
956 block
= vtn_process_block(b
, &work_list
, work
->cf_parent
,
957 work
->cf_list
, block
);
964 vtn_handle_phis_first_pass(struct vtn_builder
*b
, SpvOp opcode
,
965 const uint32_t *w
, unsigned count
)
967 if (opcode
== SpvOpLabel
)
968 return true; /* Nothing to do */
970 /* If this isn't a phi node, stop. */
971 if (opcode
!= SpvOpPhi
)
974 /* For handling phi nodes, we do a poor-man's out-of-ssa on the spot.
975 * For each phi, we create a variable with the appropreate type and
976 * do a load from that variable. Then, in a second pass, we add
977 * stores to that variable to each of the predecessor blocks.
979 * We could do something more intelligent here. However, in order to
980 * handle loops and things properly, we really need dominance
981 * information. It would end up basically being the into-SSA
982 * algorithm all over again. It's easier if we just let
983 * lower_vars_to_ssa do that for us instead of repeating it here.
985 struct vtn_type
*type
= vtn_get_type(b
, w
[1]);
986 nir_variable
*phi_var
=
987 nir_local_variable_create(b
->nb
.impl
, type
->type
, "phi");
988 _mesa_hash_table_insert(b
->phi_table
, w
, phi_var
);
990 vtn_push_ssa_value(b
, w
[2],
991 vtn_local_load(b
, nir_build_deref_var(&b
->nb
, phi_var
), 0));
997 vtn_handle_phi_second_pass(struct vtn_builder
*b
, SpvOp opcode
,
998 const uint32_t *w
, unsigned count
)
1000 if (opcode
!= SpvOpPhi
)
1003 struct hash_entry
*phi_entry
= _mesa_hash_table_search(b
->phi_table
, w
);
1005 /* It's possible that this phi is in an unreachable block in which case it
1006 * may never have been emitted and therefore may not be in the hash table.
1007 * In this case, there's no var for it and it's safe to just bail.
1009 if (phi_entry
== NULL
)
1012 nir_variable
*phi_var
= phi_entry
->data
;
1014 for (unsigned i
= 3; i
< count
; i
+= 2) {
1015 struct vtn_block
*pred
= vtn_block(b
, w
[i
+ 1]);
1017 /* If block does not have end_nop, that is because it is an unreacheable
1018 * block, and hence it is not worth to handle it */
1022 b
->nb
.cursor
= nir_after_instr(&pred
->end_nop
->instr
);
1024 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[i
]);
1026 vtn_local_store(b
, src
, nir_build_deref_var(&b
->nb
, phi_var
), 0);
1033 vtn_emit_branch(struct vtn_builder
*b
, enum vtn_branch_type branch_type
,
1034 nir_variable
*switch_fall_var
, bool *has_switch_break
)
1036 switch (branch_type
) {
1037 case vtn_branch_type_if_merge
:
1038 break; /* Nothing to do */
1039 case vtn_branch_type_switch_break
:
1040 nir_store_var(&b
->nb
, switch_fall_var
, nir_imm_false(&b
->nb
), 1);
1041 *has_switch_break
= true;
1043 case vtn_branch_type_switch_fallthrough
:
1044 break; /* Nothing to do */
1045 case vtn_branch_type_loop_break
:
1046 nir_jump(&b
->nb
, nir_jump_break
);
1048 case vtn_branch_type_loop_continue
:
1049 nir_jump(&b
->nb
, nir_jump_continue
);
1051 case vtn_branch_type_loop_back_edge
:
1053 case vtn_branch_type_return
:
1054 nir_jump(&b
->nb
, nir_jump_return
);
1056 case vtn_branch_type_discard
: {
1057 nir_intrinsic_instr
*discard
=
1058 nir_intrinsic_instr_create(b
->nb
.shader
, nir_intrinsic_discard
);
1059 nir_builder_instr_insert(&b
->nb
, &discard
->instr
);
1063 vtn_fail("Invalid branch type");
1067 static nir_ssa_def
*
1068 vtn_switch_case_condition(struct vtn_builder
*b
, struct vtn_switch
*swtch
,
1069 nir_ssa_def
*sel
, struct vtn_case
*cse
)
1071 if (cse
->is_default
) {
1072 nir_ssa_def
*any
= nir_imm_false(&b
->nb
);
1073 vtn_foreach_cf_node(other_node
, &swtch
->cases
) {
1074 struct vtn_case
*other
= vtn_cf_node_as_case(other_node
);
1075 if (other
->is_default
)
1078 any
= nir_ior(&b
->nb
, any
,
1079 vtn_switch_case_condition(b
, swtch
, sel
, other
));
1081 return nir_inot(&b
->nb
, any
);
1083 nir_ssa_def
*cond
= nir_imm_false(&b
->nb
);
1084 util_dynarray_foreach(&cse
->values
, uint64_t, val
) {
1085 nir_ssa_def
*imm
= nir_imm_intN_t(&b
->nb
, *val
, sel
->bit_size
);
1086 cond
= nir_ior(&b
->nb
, cond
, nir_ieq(&b
->nb
, sel
, imm
));
1092 static nir_loop_control
1093 vtn_loop_control(struct vtn_builder
*b
, struct vtn_loop
*vtn_loop
)
1095 if (vtn_loop
->control
== SpvLoopControlMaskNone
)
1096 return nir_loop_control_none
;
1097 else if (vtn_loop
->control
& SpvLoopControlDontUnrollMask
)
1098 return nir_loop_control_dont_unroll
;
1099 else if (vtn_loop
->control
& SpvLoopControlUnrollMask
)
1100 return nir_loop_control_unroll
;
1101 else if (vtn_loop
->control
& SpvLoopControlDependencyInfiniteMask
||
1102 vtn_loop
->control
& SpvLoopControlDependencyLengthMask
||
1103 vtn_loop
->control
& SpvLoopControlMinIterationsMask
||
1104 vtn_loop
->control
& SpvLoopControlMaxIterationsMask
||
1105 vtn_loop
->control
& SpvLoopControlIterationMultipleMask
||
1106 vtn_loop
->control
& SpvLoopControlPeelCountMask
||
1107 vtn_loop
->control
& SpvLoopControlPartialCountMask
) {
1108 /* We do not do anything special with these yet. */
1109 return nir_loop_control_none
;
1111 vtn_fail("Invalid loop control");
1115 static nir_selection_control
1116 vtn_selection_control(struct vtn_builder
*b
, struct vtn_if
*vtn_if
)
1118 if (vtn_if
->control
== SpvSelectionControlMaskNone
)
1119 return nir_selection_control_none
;
1120 else if (vtn_if
->control
& SpvSelectionControlDontFlattenMask
)
1121 return nir_selection_control_dont_flatten
;
1122 else if (vtn_if
->control
& SpvSelectionControlFlattenMask
)
1123 return nir_selection_control_flatten
;
1125 vtn_fail("Invalid selection control");
1129 vtn_emit_cf_list(struct vtn_builder
*b
, struct list_head
*cf_list
,
1130 nir_variable
*switch_fall_var
, bool *has_switch_break
,
1131 vtn_instruction_handler handler
)
1133 vtn_foreach_cf_node(node
, cf_list
) {
1134 switch (node
->type
) {
1135 case vtn_cf_node_type_block
: {
1136 struct vtn_block
*block
= vtn_cf_node_as_block(node
);
1138 const uint32_t *block_start
= block
->label
;
1139 const uint32_t *block_end
= block
->merge
? block
->merge
:
1142 block_start
= vtn_foreach_instruction(b
, block_start
, block_end
,
1143 vtn_handle_phis_first_pass
);
1145 vtn_foreach_instruction(b
, block_start
, block_end
, handler
);
1147 block
->end_nop
= nir_intrinsic_instr_create(b
->nb
.shader
,
1149 nir_builder_instr_insert(&b
->nb
, &block
->end_nop
->instr
);
1151 if ((*block
->branch
& SpvOpCodeMask
) == SpvOpReturnValue
) {
1152 vtn_fail_if(b
->func
->type
->return_type
->base_type
==
1154 "Return with a value from a function returning void");
1155 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, block
->branch
[1]);
1156 const struct glsl_type
*ret_type
=
1157 glsl_get_bare_type(b
->func
->type
->return_type
->type
);
1158 nir_deref_instr
*ret_deref
=
1159 nir_build_deref_cast(&b
->nb
, nir_load_param(&b
->nb
, 0),
1160 nir_var_function_temp
, ret_type
, 0);
1161 vtn_local_store(b
, src
, ret_deref
, 0);
1164 if (block
->branch_type
!= vtn_branch_type_none
) {
1165 vtn_emit_branch(b
, block
->branch_type
,
1166 switch_fall_var
, has_switch_break
);
1173 case vtn_cf_node_type_if
: {
1174 struct vtn_if
*vtn_if
= vtn_cf_node_as_if(node
);
1175 bool sw_break
= false;
1178 nir_push_if(&b
->nb
, vtn_get_nir_ssa(b
, vtn_if
->condition
));
1180 nif
->control
= vtn_selection_control(b
, vtn_if
);
1182 if (vtn_if
->then_type
== vtn_branch_type_none
) {
1183 vtn_emit_cf_list(b
, &vtn_if
->then_body
,
1184 switch_fall_var
, &sw_break
, handler
);
1186 vtn_emit_branch(b
, vtn_if
->then_type
, switch_fall_var
, &sw_break
);
1189 nir_push_else(&b
->nb
, nif
);
1190 if (vtn_if
->else_type
== vtn_branch_type_none
) {
1191 vtn_emit_cf_list(b
, &vtn_if
->else_body
,
1192 switch_fall_var
, &sw_break
, handler
);
1194 vtn_emit_branch(b
, vtn_if
->else_type
, switch_fall_var
, &sw_break
);
1197 nir_pop_if(&b
->nb
, nif
);
1199 /* If we encountered a switch break somewhere inside of the if,
1200 * then it would have been handled correctly by calling
1201 * emit_cf_list or emit_branch for the interrior. However, we
1202 * need to predicate everything following on wether or not we're
1206 *has_switch_break
= true;
1207 nir_push_if(&b
->nb
, nir_load_var(&b
->nb
, switch_fall_var
));
1212 case vtn_cf_node_type_loop
: {
1213 struct vtn_loop
*vtn_loop
= vtn_cf_node_as_loop(node
);
1215 nir_loop
*loop
= nir_push_loop(&b
->nb
);
1216 loop
->control
= vtn_loop_control(b
, vtn_loop
);
1218 vtn_emit_cf_list(b
, &vtn_loop
->body
, NULL
, NULL
, handler
);
1220 if (!list_is_empty(&vtn_loop
->cont_body
)) {
1221 /* If we have a non-trivial continue body then we need to put
1222 * it at the beginning of the loop with a flag to ensure that
1223 * it doesn't get executed in the first iteration.
1225 nir_variable
*do_cont
=
1226 nir_local_variable_create(b
->nb
.impl
, glsl_bool_type(), "cont");
1228 b
->nb
.cursor
= nir_before_cf_node(&loop
->cf_node
);
1229 nir_store_var(&b
->nb
, do_cont
, nir_imm_false(&b
->nb
), 1);
1231 b
->nb
.cursor
= nir_before_cf_list(&loop
->body
);
1234 nir_push_if(&b
->nb
, nir_load_var(&b
->nb
, do_cont
));
1236 vtn_emit_cf_list(b
, &vtn_loop
->cont_body
, NULL
, NULL
, handler
);
1238 nir_pop_if(&b
->nb
, cont_if
);
1240 nir_store_var(&b
->nb
, do_cont
, nir_imm_true(&b
->nb
), 1);
1242 b
->has_loop_continue
= true;
1245 nir_pop_loop(&b
->nb
, loop
);
1249 case vtn_cf_node_type_switch
: {
1250 struct vtn_switch
*vtn_switch
= vtn_cf_node_as_switch(node
);
1252 /* Before we can emit anything, we need to sort the list of cases in
1253 * fall-through order.
1255 vtn_switch_order_cases(vtn_switch
);
1257 /* First, we create a variable to keep track of whether or not the
1258 * switch is still going at any given point. Any switch breaks
1259 * will set this variable to false.
1261 nir_variable
*fall_var
=
1262 nir_local_variable_create(b
->nb
.impl
, glsl_bool_type(), "fall");
1263 nir_store_var(&b
->nb
, fall_var
, nir_imm_false(&b
->nb
), 1);
1265 nir_ssa_def
*sel
= vtn_get_nir_ssa(b
, vtn_switch
->selector
);
1267 /* Now we can walk the list of cases and actually emit code */
1268 vtn_foreach_cf_node(case_node
, &vtn_switch
->cases
) {
1269 struct vtn_case
*cse
= vtn_cf_node_as_case(case_node
);
1271 /* Figure out the condition */
1273 vtn_switch_case_condition(b
, vtn_switch
, sel
, cse
);
1274 /* Take fallthrough into account */
1275 cond
= nir_ior(&b
->nb
, cond
, nir_load_var(&b
->nb
, fall_var
));
1277 nir_if
*case_if
= nir_push_if(&b
->nb
, cond
);
1279 bool has_break
= false;
1280 nir_store_var(&b
->nb
, fall_var
, nir_imm_true(&b
->nb
), 1);
1281 vtn_emit_cf_list(b
, &cse
->body
, fall_var
, &has_break
, handler
);
1282 (void)has_break
; /* We don't care */
1284 nir_pop_if(&b
->nb
, case_if
);
1291 vtn_fail("Invalid CF node type");
1297 vtn_function_emit(struct vtn_builder
*b
, struct vtn_function
*func
,
1298 vtn_instruction_handler instruction_handler
)
1300 nir_builder_init(&b
->nb
, func
->impl
);
1302 b
->nb
.cursor
= nir_after_cf_list(&func
->impl
->body
);
1303 b
->nb
.exact
= b
->exact
;
1304 b
->has_loop_continue
= false;
1305 b
->phi_table
= _mesa_pointer_hash_table_create(b
);
1307 vtn_emit_cf_list(b
, &func
->body
, NULL
, NULL
, instruction_handler
);
1309 vtn_foreach_instruction(b
, func
->start_block
->label
, func
->end
,
1310 vtn_handle_phi_second_pass
);
1312 nir_rematerialize_derefs_in_use_blocks_impl(func
->impl
);
1314 /* Continue blocks for loops get inserted before the body of the loop
1315 * but instructions in the continue may use SSA defs in the loop body.
1316 * Therefore, we need to repair SSA to insert the needed phi nodes.
1318 if (b
->has_loop_continue
)
1319 nir_repair_ssa_impl(func
->impl
);
1321 func
->emitted
= true;