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_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
189 struct vtn_function
*vtn_callee
=
190 vtn_value(b
, w
[3], vtn_value_type_function
)->func
;
191 struct nir_function
*callee
= vtn_callee
->impl
->function
;
193 vtn_callee
->referenced
= true;
195 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
197 unsigned param_idx
= 0;
199 nir_deref_instr
*ret_deref
= NULL
;
200 struct vtn_type
*ret_type
= vtn_callee
->type
->return_type
;
201 if (ret_type
->base_type
!= vtn_base_type_void
) {
202 nir_variable
*ret_tmp
=
203 nir_local_variable_create(b
->nb
.impl
,
204 glsl_get_bare_type(ret_type
->type
),
206 ret_deref
= nir_build_deref_var(&b
->nb
, ret_tmp
);
207 call
->params
[param_idx
++] = nir_src_for_ssa(&ret_deref
->dest
.ssa
);
210 for (unsigned i
= 0; i
< vtn_callee
->type
->length
; i
++) {
211 struct vtn_type
*arg_type
= vtn_callee
->type
->params
[i
];
212 unsigned arg_id
= w
[4 + i
];
214 if (arg_type
->base_type
== vtn_base_type_sampled_image
) {
215 struct vtn_sampled_image
*sampled_image
=
216 vtn_value(b
, arg_id
, vtn_value_type_sampled_image
)->sampled_image
;
218 call
->params
[param_idx
++] =
219 nir_src_for_ssa(&sampled_image
->image
->deref
->dest
.ssa
);
220 call
->params
[param_idx
++] =
221 nir_src_for_ssa(&sampled_image
->sampler
->deref
->dest
.ssa
);
222 } else if (arg_type
->base_type
== vtn_base_type_pointer
||
223 arg_type
->base_type
== vtn_base_type_image
||
224 arg_type
->base_type
== vtn_base_type_sampler
) {
225 struct vtn_pointer
*pointer
=
226 vtn_value(b
, arg_id
, vtn_value_type_pointer
)->pointer
;
227 call
->params
[param_idx
++] =
228 nir_src_for_ssa(vtn_pointer_to_ssa(b
, pointer
));
230 vtn_ssa_value_add_to_call_params(b
, vtn_ssa_value(b
, arg_id
),
231 arg_type
, call
, ¶m_idx
);
234 assert(param_idx
== call
->num_params
);
236 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
238 if (ret_type
->base_type
== vtn_base_type_void
) {
239 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
241 vtn_push_ssa(b
, w
[2], res_type
, vtn_local_load(b
, ret_deref
, 0));
246 vtn_cfg_handle_prepass_instruction(struct vtn_builder
*b
, SpvOp opcode
,
247 const uint32_t *w
, unsigned count
)
250 case SpvOpFunction
: {
251 vtn_assert(b
->func
== NULL
);
252 b
->func
= rzalloc(b
, struct vtn_function
);
254 b
->func
->node
.type
= vtn_cf_node_type_function
;
255 b
->func
->node
.parent
= NULL
;
256 list_inithead(&b
->func
->body
);
257 b
->func
->control
= w
[3];
259 UNUSED
const struct glsl_type
*result_type
=
260 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
261 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_function
);
264 b
->func
->type
= vtn_value(b
, w
[4], vtn_value_type_type
)->type
;
265 const struct vtn_type
*func_type
= b
->func
->type
;
267 vtn_assert(func_type
->return_type
->type
== result_type
);
270 nir_function_create(b
->shader
, ralloc_strdup(b
->shader
, val
->name
));
272 unsigned num_params
= 0;
273 for (unsigned i
= 0; i
< func_type
->length
; i
++)
274 num_params
+= vtn_type_count_function_params(func_type
->params
[i
]);
276 /* Add one parameter for the function return value */
277 if (func_type
->return_type
->base_type
!= vtn_base_type_void
)
280 func
->num_params
= num_params
;
281 func
->params
= ralloc_array(b
->shader
, nir_parameter
, num_params
);
284 if (func_type
->return_type
->base_type
!= vtn_base_type_void
) {
285 nir_address_format addr_format
=
286 vtn_mode_to_address_format(b
, vtn_variable_mode_function
);
287 /* The return value is a regular pointer */
288 func
->params
[idx
++] = (nir_parameter
) {
289 .num_components
= nir_address_format_num_components(addr_format
),
290 .bit_size
= nir_address_format_bit_size(addr_format
),
294 for (unsigned i
= 0; i
< func_type
->length
; i
++)
295 vtn_type_add_to_function_params(func_type
->params
[i
], func
, &idx
);
296 assert(idx
== num_params
);
298 b
->func
->impl
= nir_function_impl_create(func
);
299 nir_builder_init(&b
->nb
, func
->impl
);
300 b
->nb
.cursor
= nir_before_cf_list(&b
->func
->impl
->body
);
301 b
->nb
.exact
= b
->exact
;
303 b
->func_param_idx
= 0;
305 /* The return value is the first parameter */
306 if (func_type
->return_type
->base_type
!= vtn_base_type_void
)
311 case SpvOpFunctionEnd
:
316 case SpvOpFunctionParameter
: {
317 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
319 vtn_assert(b
->func_param_idx
< b
->func
->impl
->function
->num_params
);
321 if (type
->base_type
== vtn_base_type_sampled_image
) {
322 /* Sampled images are actually two parameters. The first is the
323 * image and the second is the sampler.
325 struct vtn_value
*val
=
326 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
328 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
330 struct vtn_type
*sampler_type
= rzalloc(b
, struct vtn_type
);
331 sampler_type
->base_type
= vtn_base_type_sampler
;
332 sampler_type
->type
= glsl_bare_sampler_type();
334 val
->sampled_image
->image
=
335 vtn_load_param_pointer(b
, type
, b
->func_param_idx
++);
336 val
->sampled_image
->sampler
=
337 vtn_load_param_pointer(b
, sampler_type
, b
->func_param_idx
++);
338 } else if (type
->base_type
== vtn_base_type_pointer
&&
339 type
->type
!= NULL
) {
340 /* This is a pointer with an actual storage type */
341 nir_ssa_def
*ssa_ptr
= nir_load_param(&b
->nb
, b
->func_param_idx
++);
342 vtn_push_value_pointer(b
, w
[2], vtn_pointer_from_ssa(b
, ssa_ptr
, type
));
343 } else if (type
->base_type
== vtn_base_type_pointer
||
344 type
->base_type
== vtn_base_type_image
||
345 type
->base_type
== vtn_base_type_sampler
) {
346 vtn_push_value_pointer(b
, w
[2], vtn_load_param_pointer(b
, type
, b
->func_param_idx
++));
348 /* We're a regular SSA value. */
349 struct vtn_ssa_value
*value
= vtn_create_ssa_value(b
, type
->type
);
350 vtn_ssa_value_load_function_param(b
, value
, type
, &b
->func_param_idx
);
351 vtn_push_ssa(b
, w
[2], type
, value
);
357 vtn_assert(b
->block
== NULL
);
358 b
->block
= rzalloc(b
, struct vtn_block
);
359 b
->block
->node
.type
= vtn_cf_node_type_block
;
361 vtn_push_value(b
, w
[1], vtn_value_type_block
)->block
= b
->block
;
363 if (b
->func
->start_block
== NULL
) {
364 /* This is the first block encountered for this function. In this
365 * case, we set the start block and add it to the list of
366 * implemented functions that we'll walk later.
368 b
->func
->start_block
= b
->block
;
369 list_addtail(&b
->func
->node
.link
, &b
->functions
);
374 case SpvOpSelectionMerge
:
376 vtn_assert(b
->block
&& b
->block
->merge
== NULL
);
381 case SpvOpBranchConditional
:
385 case SpvOpReturnValue
:
386 case SpvOpUnreachable
:
387 vtn_assert(b
->block
&& b
->block
->branch
== NULL
);
388 b
->block
->branch
= w
;
393 /* Continue on as per normal */
400 /* This function performs a depth-first search of the cases and puts them
401 * in fall-through order.
404 vtn_order_case(struct vtn_switch
*swtch
, struct vtn_case
*cse
)
411 list_del(&cse
->node
.link
);
413 if (cse
->fallthrough
) {
414 vtn_order_case(swtch
, cse
->fallthrough
);
416 /* If we have a fall-through, place this case right before the case it
417 * falls through to. This ensures that fallthroughs come one after
418 * the other. These two can never get separated because that would
419 * imply something else falling through to the same case. Also, this
420 * can't break ordering because the DFS ensures that this case is
421 * visited before anything that falls through to it.
423 list_addtail(&cse
->node
.link
, &cse
->fallthrough
->node
.link
);
425 list_add(&cse
->node
.link
, &swtch
->cases
);
430 vtn_switch_order_cases(struct vtn_switch
*swtch
)
432 struct list_head cases
;
433 list_replace(&swtch
->cases
, &cases
);
434 list_inithead(&swtch
->cases
);
435 while (!list_is_empty(&cases
)) {
436 struct vtn_case
*cse
=
437 list_first_entry(&cases
, struct vtn_case
, node
.link
);
438 vtn_order_case(swtch
, cse
);
443 vtn_block_set_merge_cf_node(struct vtn_builder
*b
, struct vtn_block
*block
,
444 struct vtn_cf_node
*cf_node
)
446 vtn_fail_if(block
->merge_cf_node
!= NULL
,
447 "The merge block declared by a header block cannot be a "
448 "merge block declared by any other header block.");
450 block
->merge_cf_node
= cf_node
;
453 #define VTN_DECL_CF_NODE_FIND(_type) \
454 static inline struct vtn_##_type * \
455 vtn_cf_node_find_##_type(struct vtn_cf_node *node) \
457 while (node && node->type != vtn_cf_node_type_##_type) \
458 node = node->parent; \
459 return (struct vtn_##_type *)node; \
462 VTN_DECL_CF_NODE_FIND(if)
463 VTN_DECL_CF_NODE_FIND(loop
)
464 VTN_DECL_CF_NODE_FIND(case)
465 VTN_DECL_CF_NODE_FIND(switch)
466 VTN_DECL_CF_NODE_FIND(function
)
468 static enum vtn_branch_type
469 vtn_handle_branch(struct vtn_builder
*b
,
470 struct vtn_cf_node
*cf_parent
,
471 struct vtn_block
*target_block
)
473 struct vtn_loop
*loop
= vtn_cf_node_find_loop(cf_parent
);
475 /* Detect a loop back-edge first. That way none of the code below
476 * accidentally operates on a loop back-edge.
478 if (loop
&& target_block
== loop
->header_block
)
479 return vtn_branch_type_loop_back_edge
;
481 /* Try to detect fall-through */
482 if (target_block
->switch_case
) {
483 /* When it comes to handling switch cases, we can break calls to
484 * vtn_handle_branch into two cases: calls from within a case construct
485 * and calls for the jump to each case construct. In the second case,
486 * cf_parent is the vtn_switch itself and vtn_cf_node_find_case() will
487 * return the outer switch case in which this switch is contained. It's
488 * fine if the target block is a switch case from an outer switch as
489 * long as it is also the switch break for this switch.
491 struct vtn_case
*switch_case
= vtn_cf_node_find_case(cf_parent
);
493 /* This doesn't get called for the OpSwitch */
494 vtn_fail_if(switch_case
== NULL
,
495 "A switch case can only be entered through an OpSwitch or "
496 "falling through from another switch case.");
498 /* Because block->switch_case is only set on the entry block for a given
499 * switch case, we only ever get here if we're jumping to the start of a
500 * switch case. It's possible, however, that a switch case could jump
501 * to itself via a back-edge. That *should* get caught by the loop
502 * handling case above but if we have a back edge without a loop merge,
503 * we could en up here.
505 vtn_fail_if(target_block
->switch_case
== switch_case
,
506 "A switch cannot fall-through to itself. Likely, there is "
507 "a back-edge which is not to a loop header.");
509 vtn_fail_if(target_block
->switch_case
->node
.parent
!=
510 switch_case
->node
.parent
,
511 "A switch case fall-through must come from the same "
512 "OpSwitch construct");
514 vtn_fail_if(switch_case
->fallthrough
!= NULL
&&
515 switch_case
->fallthrough
!= target_block
->switch_case
,
516 "Each case construct can have at most one branch to "
517 "another case construct");
519 switch_case
->fallthrough
= target_block
->switch_case
;
521 /* We don't immediately return vtn_branch_type_switch_fallthrough
522 * because it may also be a loop or switch break for an inner loop or
523 * switch and that takes precedence.
527 if (loop
&& target_block
== loop
->cont_block
)
528 return vtn_branch_type_loop_continue
;
530 /* We walk blocks as a breadth-first search on the control-flow construct
531 * tree where, when we find a construct, we add the vtn_cf_node for that
532 * construct and continue iterating at the merge target block (if any).
533 * Therefore, we want merges whose with parent == cf_parent to be treated
534 * as regular branches. We only want to consider merges if they break out
535 * of the current CF construct.
537 if (target_block
->merge_cf_node
!= NULL
&&
538 target_block
->merge_cf_node
->parent
!= cf_parent
) {
539 switch (target_block
->merge_cf_node
->type
) {
540 case vtn_cf_node_type_if
:
541 for (struct vtn_cf_node
*node
= cf_parent
;
542 node
!= target_block
->merge_cf_node
; node
= node
->parent
) {
543 vtn_fail_if(node
== NULL
|| node
->type
!= vtn_cf_node_type_if
,
544 "Branching to the merge block of a selection "
545 "construct can only be used to break out of a "
546 "selection construct");
548 struct vtn_if
*if_stmt
= vtn_cf_node_as_if(node
);
550 /* This should be guaranteed by our iteration */
551 assert(if_stmt
->merge_block
!= target_block
);
553 vtn_fail_if(if_stmt
->merge_block
!= NULL
,
554 "Branching to the merge block of a selection "
555 "construct can only be used to break out of the "
556 "inner most nested selection level");
558 return vtn_branch_type_if_merge
;
560 case vtn_cf_node_type_loop
:
561 vtn_fail_if(target_block
->merge_cf_node
!= &loop
->node
,
562 "Loop breaks can only break out of the inner most "
563 "nested loop level");
564 return vtn_branch_type_loop_break
;
566 case vtn_cf_node_type_switch
: {
567 struct vtn_switch
*swtch
= vtn_cf_node_find_switch(cf_parent
);
568 vtn_fail_if(target_block
->merge_cf_node
!= &swtch
->node
,
569 "Switch breaks can only break out of the inner most "
570 "nested switch level");
571 return vtn_branch_type_switch_break
;
575 unreachable("Invalid CF node type for a merge");
579 if (target_block
->switch_case
)
580 return vtn_branch_type_switch_fallthrough
;
582 return vtn_branch_type_none
;
585 struct vtn_cfg_work_item
{
586 struct list_head link
;
588 struct vtn_cf_node
*cf_parent
;
589 struct list_head
*cf_list
;
590 struct vtn_block
*start_block
;
594 vtn_add_cfg_work_item(struct vtn_builder
*b
,
595 struct list_head
*work_list
,
596 struct vtn_cf_node
*cf_parent
,
597 struct list_head
*cf_list
,
598 struct vtn_block
*start_block
)
600 struct vtn_cfg_work_item
*work
= ralloc(b
, struct vtn_cfg_work_item
);
601 work
->cf_parent
= cf_parent
;
602 work
->cf_list
= cf_list
;
603 work
->start_block
= start_block
;
604 list_addtail(&work
->link
, work_list
);
607 /* Processes a block and returns the next block to process or NULL if we've
608 * reached the end of the construct.
610 static struct vtn_block
*
611 vtn_process_block(struct vtn_builder
*b
,
612 struct list_head
*work_list
,
613 struct vtn_cf_node
*cf_parent
,
614 struct list_head
*cf_list
,
615 struct vtn_block
*block
)
617 if (!list_is_empty(cf_list
)) {
618 /* vtn_process_block() acts like an iterator: it processes the given
619 * block and then returns the next block to process. For a given
620 * control-flow construct, vtn_build_cfg() calls vtn_process_block()
621 * repeatedly until it finally returns NULL. Therefore, we know that
622 * the only blocks on which vtn_process_block() can be called are either
623 * the first block in a construct or a block that vtn_process_block()
624 * returned for the current construct. If cf_list is empty then we know
625 * that we're processing the first block in the construct and we have to
626 * add it to the list.
628 * If cf_list is not empty, then it must be the block returned by the
629 * previous call to vtn_process_block(). We know a priori that
630 * vtn_process_block only returns either normal branches
631 * (vtn_branch_type_none) or merge target blocks.
633 switch (vtn_handle_branch(b
, cf_parent
, block
)) {
634 case vtn_branch_type_none
:
635 /* For normal branches, we want to process them and add them to the
636 * current construct. Merge target blocks also look like normal
637 * branches from the perspective of this construct. See also
638 * vtn_handle_branch().
642 case vtn_branch_type_loop_continue
:
643 case vtn_branch_type_switch_fallthrough
:
644 /* The two cases where we can get early exits from a construct that
645 * are not to that construct's merge target are loop continues and
646 * switch fall-throughs. In these cases, we need to break out of the
647 * current construct by returning NULL.
652 /* The only way we can get here is if something was used as two kinds
653 * of merges at the same time and that's illegal.
655 vtn_fail("A block was used as a merge target from two or more "
656 "structured control-flow constructs");
660 /* Once a block has been processed, it is placed into and the list link
661 * will point to something non-null. If we see a node we've already
662 * processed here, it either exists in multiple functions or it's an
665 if (block
->node
.parent
!= NULL
) {
666 vtn_fail_if(vtn_cf_node_find_function(&block
->node
) !=
667 vtn_cf_node_find_function(cf_parent
),
668 "A block cannot exist in two functions at the "
671 vtn_fail("Invalid back or cross-edge in the CFG");
674 if (block
->merge
&& (*block
->merge
& SpvOpCodeMask
) == SpvOpLoopMerge
&&
675 block
->loop
== NULL
) {
676 vtn_fail_if((*block
->branch
& SpvOpCodeMask
) != SpvOpBranch
&&
677 (*block
->branch
& SpvOpCodeMask
) != SpvOpBranchConditional
,
678 "An OpLoopMerge instruction must immediately precede "
679 "either an OpBranch or OpBranchConditional instruction.");
681 struct vtn_loop
*loop
= rzalloc(b
, struct vtn_loop
);
683 loop
->node
.type
= vtn_cf_node_type_loop
;
684 loop
->node
.parent
= cf_parent
;
685 list_inithead(&loop
->body
);
686 list_inithead(&loop
->cont_body
);
687 loop
->header_block
= block
;
688 loop
->break_block
= vtn_block(b
, block
->merge
[1]);
689 loop
->cont_block
= vtn_block(b
, block
->merge
[2]);
690 loop
->control
= block
->merge
[3];
692 list_addtail(&loop
->node
.link
, cf_list
);
695 /* Note: The work item for the main loop body will start with the
696 * current block as its start block. If we weren't careful, we would
697 * get here again and end up in an infinite loop. This is why we set
698 * block->loop above and check for it before creating one. This way,
699 * we only create the loop once and the second iteration that tries to
700 * handle this loop goes to the cases below and gets handled as a
703 vtn_add_cfg_work_item(b
, work_list
, &loop
->node
,
704 &loop
->body
, loop
->header_block
);
706 /* For continue targets, SPIR-V guarantees the following:
708 * - the Continue Target must dominate the back-edge block
709 * - the back-edge block must post dominate the Continue Target
711 * If the header block is the same as the continue target, this
712 * condition is trivially satisfied and there is no real continue
715 if (loop
->cont_block
!= loop
->header_block
) {
716 vtn_add_cfg_work_item(b
, work_list
, &loop
->node
,
717 &loop
->cont_body
, loop
->cont_block
);
720 vtn_block_set_merge_cf_node(b
, loop
->break_block
, &loop
->node
);
722 return loop
->break_block
;
725 /* Add the block to the CF list */
726 block
->node
.parent
= cf_parent
;
727 list_addtail(&block
->node
.link
, cf_list
);
729 switch (*block
->branch
& SpvOpCodeMask
) {
731 struct vtn_block
*branch_block
= vtn_block(b
, block
->branch
[1]);
733 block
->branch_type
= vtn_handle_branch(b
, cf_parent
, branch_block
);
735 if (block
->branch_type
== vtn_branch_type_none
)
742 case SpvOpReturnValue
:
743 block
->branch_type
= vtn_branch_type_return
;
747 block
->branch_type
= vtn_branch_type_discard
;
750 case SpvOpBranchConditional
: {
751 struct vtn_value
*cond_val
= vtn_untyped_value(b
, block
->branch
[1]);
752 vtn_fail_if(!cond_val
->type
||
753 cond_val
->type
->base_type
!= vtn_base_type_scalar
||
754 cond_val
->type
->type
!= glsl_bool_type(),
755 "Condition must be a Boolean type scalar");
757 struct vtn_block
*then_block
= vtn_block(b
, block
->branch
[2]);
758 struct vtn_block
*else_block
= vtn_block(b
, block
->branch
[3]);
760 if (then_block
== else_block
) {
761 /* This is uncommon but it can happen. We treat this the same way as
762 * an unconditional branch.
764 block
->branch_type
= vtn_handle_branch(b
, cf_parent
, then_block
);
766 if (block
->branch_type
== vtn_branch_type_none
)
772 struct vtn_if
*if_stmt
= rzalloc(b
, struct vtn_if
);
774 if_stmt
->node
.type
= vtn_cf_node_type_if
;
775 if_stmt
->node
.parent
= cf_parent
;
776 if_stmt
->condition
= block
->branch
[1];
777 list_inithead(&if_stmt
->then_body
);
778 list_inithead(&if_stmt
->else_body
);
780 list_addtail(&if_stmt
->node
.link
, cf_list
);
783 (*block
->merge
& SpvOpCodeMask
) == SpvOpSelectionMerge
) {
784 /* We may not always have a merge block and that merge doesn't
785 * technically have to be an OpSelectionMerge. We could have a block
786 * with an OpLoopMerge which ends in an OpBranchConditional.
788 if_stmt
->merge_block
= vtn_block(b
, block
->merge
[1]);
789 vtn_block_set_merge_cf_node(b
, if_stmt
->merge_block
, &if_stmt
->node
);
791 if_stmt
->control
= block
->merge
[2];
794 if_stmt
->then_type
= vtn_handle_branch(b
, &if_stmt
->node
, then_block
);
795 if (if_stmt
->then_type
== vtn_branch_type_none
) {
796 vtn_add_cfg_work_item(b
, work_list
, &if_stmt
->node
,
797 &if_stmt
->then_body
, then_block
);
800 if_stmt
->else_type
= vtn_handle_branch(b
, &if_stmt
->node
, else_block
);
801 if (if_stmt
->else_type
== vtn_branch_type_none
) {
802 vtn_add_cfg_work_item(b
, work_list
, &if_stmt
->node
,
803 &if_stmt
->else_body
, else_block
);
806 return if_stmt
->merge_block
;
810 struct vtn_value
*sel_val
= vtn_untyped_value(b
, block
->branch
[1]);
811 vtn_fail_if(!sel_val
->type
||
812 sel_val
->type
->base_type
!= vtn_base_type_scalar
,
813 "Selector of OpSwitch must have a type of OpTypeInt");
815 nir_alu_type sel_type
=
816 nir_get_nir_type_for_glsl_type(sel_val
->type
->type
);
817 vtn_fail_if(nir_alu_type_get_base_type(sel_type
) != nir_type_int
&&
818 nir_alu_type_get_base_type(sel_type
) != nir_type_uint
,
819 "Selector of OpSwitch must have a type of OpTypeInt");
821 struct vtn_switch
*swtch
= rzalloc(b
, struct vtn_switch
);
823 swtch
->node
.type
= vtn_cf_node_type_switch
;
824 swtch
->node
.parent
= cf_parent
;
825 swtch
->selector
= block
->branch
[1];
826 list_inithead(&swtch
->cases
);
828 list_addtail(&swtch
->node
.link
, cf_list
);
830 /* We may not always have a merge block */
832 vtn_fail_if((*block
->merge
& SpvOpCodeMask
) != SpvOpSelectionMerge
,
833 "An OpLoopMerge instruction must immediately precede "
834 "either an OpBranch or OpBranchConditional "
836 swtch
->break_block
= vtn_block(b
, block
->merge
[1]);
837 vtn_block_set_merge_cf_node(b
, swtch
->break_block
, &swtch
->node
);
840 /* First, we go through and record all of the cases. */
841 const uint32_t *branch_end
=
842 block
->branch
+ (block
->branch
[0] >> SpvWordCountShift
);
844 struct hash_table
*block_to_case
= _mesa_pointer_hash_table_create(b
);
846 bool is_default
= true;
847 const unsigned bitsize
= nir_alu_type_get_type_size(sel_type
);
848 for (const uint32_t *w
= block
->branch
+ 2; w
< branch_end
;) {
849 uint64_t literal
= 0;
854 assert(bitsize
== 64);
855 literal
= vtn_u64_literal(w
);
859 struct vtn_block
*case_block
= vtn_block(b
, *(w
++));
861 struct hash_entry
*case_entry
=
862 _mesa_hash_table_search(block_to_case
, case_block
);
864 struct vtn_case
*cse
;
866 cse
= case_entry
->data
;
868 cse
= rzalloc(b
, struct vtn_case
);
870 cse
->node
.type
= vtn_cf_node_type_case
;
871 cse
->node
.parent
= &swtch
->node
;
872 list_inithead(&cse
->body
);
873 util_dynarray_init(&cse
->values
, b
);
875 cse
->type
= vtn_handle_branch(b
, &swtch
->node
, case_block
);
877 case vtn_branch_type_none
:
878 /* This is a "real" cases which has stuff in it */
879 vtn_fail_if(case_block
->switch_case
!= NULL
,
880 "OpSwitch has a case which is also in another "
881 "OpSwitch construct");
882 case_block
->switch_case
= cse
;
883 vtn_add_cfg_work_item(b
, work_list
, &cse
->node
,
884 &cse
->body
, case_block
);
887 case vtn_branch_type_switch_break
:
888 case vtn_branch_type_loop_break
:
889 case vtn_branch_type_loop_continue
:
890 /* Switch breaks as well as loop breaks and continues can be
891 * used to break out of a switch construct or as direct targets
897 vtn_fail("Target of OpSwitch is not a valid structured exit "
898 "from the switch construct.");
901 list_addtail(&cse
->node
.link
, &swtch
->cases
);
903 _mesa_hash_table_insert(block_to_case
, case_block
, cse
);
907 cse
->is_default
= true;
909 util_dynarray_append(&cse
->values
, uint64_t, literal
);
915 _mesa_hash_table_destroy(block_to_case
, NULL
);
917 return swtch
->break_block
;
920 case SpvOpUnreachable
:
924 vtn_fail("Block did not end with a valid branch instruction");
929 vtn_build_cfg(struct vtn_builder
*b
, const uint32_t *words
, const uint32_t *end
)
931 vtn_foreach_instruction(b
, words
, end
,
932 vtn_cfg_handle_prepass_instruction
);
934 vtn_foreach_cf_node(func_node
, &b
->functions
) {
935 struct vtn_function
*func
= vtn_cf_node_as_function(func_node
);
937 /* We build the CFG for each function by doing a breadth-first search on
938 * the control-flow graph. We keep track of our state using a worklist.
939 * Doing a BFS ensures that we visit each structured control-flow
940 * construct and its merge node before we visit the stuff inside the
943 struct list_head work_list
;
944 list_inithead(&work_list
);
945 vtn_add_cfg_work_item(b
, &work_list
, &func
->node
, &func
->body
,
948 while (!list_is_empty(&work_list
)) {
949 struct vtn_cfg_work_item
*work
=
950 list_first_entry(&work_list
, struct vtn_cfg_work_item
, link
);
951 list_del(&work
->link
);
953 for (struct vtn_block
*block
= work
->start_block
; block
; ) {
954 block
= vtn_process_block(b
, &work_list
, work
->cf_parent
,
955 work
->cf_list
, block
);
962 vtn_handle_phis_first_pass(struct vtn_builder
*b
, SpvOp opcode
,
963 const uint32_t *w
, unsigned count
)
965 if (opcode
== SpvOpLabel
)
966 return true; /* Nothing to do */
968 /* If this isn't a phi node, stop. */
969 if (opcode
!= SpvOpPhi
)
972 /* For handling phi nodes, we do a poor-man's out-of-ssa on the spot.
973 * For each phi, we create a variable with the appropreate type and
974 * do a load from that variable. Then, in a second pass, we add
975 * stores to that variable to each of the predecessor blocks.
977 * We could do something more intelligent here. However, in order to
978 * handle loops and things properly, we really need dominance
979 * information. It would end up basically being the into-SSA
980 * algorithm all over again. It's easier if we just let
981 * lower_vars_to_ssa do that for us instead of repeating it here.
983 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
984 nir_variable
*phi_var
=
985 nir_local_variable_create(b
->nb
.impl
, type
->type
, "phi");
986 _mesa_hash_table_insert(b
->phi_table
, w
, phi_var
);
988 vtn_push_ssa(b
, w
[2], type
,
989 vtn_local_load(b
, nir_build_deref_var(&b
->nb
, phi_var
), 0));
995 vtn_handle_phi_second_pass(struct vtn_builder
*b
, SpvOp opcode
,
996 const uint32_t *w
, unsigned count
)
998 if (opcode
!= SpvOpPhi
)
1001 struct hash_entry
*phi_entry
= _mesa_hash_table_search(b
->phi_table
, w
);
1002 vtn_assert(phi_entry
);
1003 nir_variable
*phi_var
= phi_entry
->data
;
1005 for (unsigned i
= 3; i
< count
; i
+= 2) {
1006 struct vtn_block
*pred
= vtn_block(b
, w
[i
+ 1]);
1008 /* If block does not have end_nop, that is because it is an unreacheable
1009 * block, and hence it is not worth to handle it */
1013 b
->nb
.cursor
= nir_after_instr(&pred
->end_nop
->instr
);
1015 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[i
]);
1017 vtn_local_store(b
, src
, nir_build_deref_var(&b
->nb
, phi_var
), 0);
1024 vtn_emit_branch(struct vtn_builder
*b
, enum vtn_branch_type branch_type
,
1025 nir_variable
*switch_fall_var
, bool *has_switch_break
)
1027 switch (branch_type
) {
1028 case vtn_branch_type_if_merge
:
1029 break; /* Nothing to do */
1030 case vtn_branch_type_switch_break
:
1031 nir_store_var(&b
->nb
, switch_fall_var
, nir_imm_false(&b
->nb
), 1);
1032 *has_switch_break
= true;
1034 case vtn_branch_type_switch_fallthrough
:
1035 break; /* Nothing to do */
1036 case vtn_branch_type_loop_break
:
1037 nir_jump(&b
->nb
, nir_jump_break
);
1039 case vtn_branch_type_loop_continue
:
1040 nir_jump(&b
->nb
, nir_jump_continue
);
1042 case vtn_branch_type_loop_back_edge
:
1044 case vtn_branch_type_return
:
1045 nir_jump(&b
->nb
, nir_jump_return
);
1047 case vtn_branch_type_discard
: {
1048 nir_intrinsic_instr
*discard
=
1049 nir_intrinsic_instr_create(b
->nb
.shader
, nir_intrinsic_discard
);
1050 nir_builder_instr_insert(&b
->nb
, &discard
->instr
);
1054 vtn_fail("Invalid branch type");
1058 static nir_ssa_def
*
1059 vtn_switch_case_condition(struct vtn_builder
*b
, struct vtn_switch
*swtch
,
1060 nir_ssa_def
*sel
, struct vtn_case
*cse
)
1062 if (cse
->is_default
) {
1063 nir_ssa_def
*any
= nir_imm_false(&b
->nb
);
1064 vtn_foreach_cf_node(other_node
, &swtch
->cases
) {
1065 struct vtn_case
*other
= vtn_cf_node_as_case(other_node
);
1066 if (other
->is_default
)
1069 any
= nir_ior(&b
->nb
, any
,
1070 vtn_switch_case_condition(b
, swtch
, sel
, other
));
1072 return nir_inot(&b
->nb
, any
);
1074 nir_ssa_def
*cond
= nir_imm_false(&b
->nb
);
1075 util_dynarray_foreach(&cse
->values
, uint64_t, val
) {
1076 nir_ssa_def
*imm
= nir_imm_intN_t(&b
->nb
, *val
, sel
->bit_size
);
1077 cond
= nir_ior(&b
->nb
, cond
, nir_ieq(&b
->nb
, sel
, imm
));
1083 static nir_loop_control
1084 vtn_loop_control(struct vtn_builder
*b
, struct vtn_loop
*vtn_loop
)
1086 if (vtn_loop
->control
== SpvLoopControlMaskNone
)
1087 return nir_loop_control_none
;
1088 else if (vtn_loop
->control
& SpvLoopControlDontUnrollMask
)
1089 return nir_loop_control_dont_unroll
;
1090 else if (vtn_loop
->control
& SpvLoopControlUnrollMask
)
1091 return nir_loop_control_unroll
;
1092 else if (vtn_loop
->control
& SpvLoopControlDependencyInfiniteMask
||
1093 vtn_loop
->control
& SpvLoopControlDependencyLengthMask
||
1094 vtn_loop
->control
& SpvLoopControlMinIterationsMask
||
1095 vtn_loop
->control
& SpvLoopControlMaxIterationsMask
||
1096 vtn_loop
->control
& SpvLoopControlIterationMultipleMask
||
1097 vtn_loop
->control
& SpvLoopControlPeelCountMask
||
1098 vtn_loop
->control
& SpvLoopControlPartialCountMask
) {
1099 /* We do not do anything special with these yet. */
1100 return nir_loop_control_none
;
1102 vtn_fail("Invalid loop control");
1106 static nir_selection_control
1107 vtn_selection_control(struct vtn_builder
*b
, struct vtn_if
*vtn_if
)
1109 if (vtn_if
->control
== SpvSelectionControlMaskNone
)
1110 return nir_selection_control_none
;
1111 else if (vtn_if
->control
& SpvSelectionControlDontFlattenMask
)
1112 return nir_selection_control_dont_flatten
;
1113 else if (vtn_if
->control
& SpvSelectionControlFlattenMask
)
1114 return nir_selection_control_flatten
;
1116 vtn_fail("Invalid selection control");
1120 vtn_emit_cf_list(struct vtn_builder
*b
, struct list_head
*cf_list
,
1121 nir_variable
*switch_fall_var
, bool *has_switch_break
,
1122 vtn_instruction_handler handler
)
1124 vtn_foreach_cf_node(node
, cf_list
) {
1125 switch (node
->type
) {
1126 case vtn_cf_node_type_block
: {
1127 struct vtn_block
*block
= vtn_cf_node_as_block(node
);
1129 const uint32_t *block_start
= block
->label
;
1130 const uint32_t *block_end
= block
->merge
? block
->merge
:
1133 block_start
= vtn_foreach_instruction(b
, block_start
, block_end
,
1134 vtn_handle_phis_first_pass
);
1136 vtn_foreach_instruction(b
, block_start
, block_end
, handler
);
1138 block
->end_nop
= nir_intrinsic_instr_create(b
->nb
.shader
,
1140 nir_builder_instr_insert(&b
->nb
, &block
->end_nop
->instr
);
1142 if ((*block
->branch
& SpvOpCodeMask
) == SpvOpReturnValue
) {
1143 vtn_fail_if(b
->func
->type
->return_type
->base_type
==
1145 "Return with a value from a function returning void");
1146 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, block
->branch
[1]);
1147 const struct glsl_type
*ret_type
=
1148 glsl_get_bare_type(b
->func
->type
->return_type
->type
);
1149 nir_deref_instr
*ret_deref
=
1150 nir_build_deref_cast(&b
->nb
, nir_load_param(&b
->nb
, 0),
1151 nir_var_function_temp
, ret_type
, 0);
1152 vtn_local_store(b
, src
, ret_deref
, 0);
1155 if (block
->branch_type
!= vtn_branch_type_none
) {
1156 vtn_emit_branch(b
, block
->branch_type
,
1157 switch_fall_var
, has_switch_break
);
1164 case vtn_cf_node_type_if
: {
1165 struct vtn_if
*vtn_if
= vtn_cf_node_as_if(node
);
1166 bool sw_break
= false;
1169 nir_push_if(&b
->nb
, vtn_ssa_value(b
, vtn_if
->condition
)->def
);
1171 nif
->control
= vtn_selection_control(b
, vtn_if
);
1173 if (vtn_if
->then_type
== vtn_branch_type_none
) {
1174 vtn_emit_cf_list(b
, &vtn_if
->then_body
,
1175 switch_fall_var
, &sw_break
, handler
);
1177 vtn_emit_branch(b
, vtn_if
->then_type
, switch_fall_var
, &sw_break
);
1180 nir_push_else(&b
->nb
, nif
);
1181 if (vtn_if
->else_type
== vtn_branch_type_none
) {
1182 vtn_emit_cf_list(b
, &vtn_if
->else_body
,
1183 switch_fall_var
, &sw_break
, handler
);
1185 vtn_emit_branch(b
, vtn_if
->else_type
, switch_fall_var
, &sw_break
);
1188 nir_pop_if(&b
->nb
, nif
);
1190 /* If we encountered a switch break somewhere inside of the if,
1191 * then it would have been handled correctly by calling
1192 * emit_cf_list or emit_branch for the interrior. However, we
1193 * need to predicate everything following on wether or not we're
1197 *has_switch_break
= true;
1198 nir_push_if(&b
->nb
, nir_load_var(&b
->nb
, switch_fall_var
));
1203 case vtn_cf_node_type_loop
: {
1204 struct vtn_loop
*vtn_loop
= vtn_cf_node_as_loop(node
);
1206 nir_loop
*loop
= nir_push_loop(&b
->nb
);
1207 loop
->control
= vtn_loop_control(b
, vtn_loop
);
1209 vtn_emit_cf_list(b
, &vtn_loop
->body
, NULL
, NULL
, handler
);
1211 if (!list_is_empty(&vtn_loop
->cont_body
)) {
1212 /* If we have a non-trivial continue body then we need to put
1213 * it at the beginning of the loop with a flag to ensure that
1214 * it doesn't get executed in the first iteration.
1216 nir_variable
*do_cont
=
1217 nir_local_variable_create(b
->nb
.impl
, glsl_bool_type(), "cont");
1219 b
->nb
.cursor
= nir_before_cf_node(&loop
->cf_node
);
1220 nir_store_var(&b
->nb
, do_cont
, nir_imm_false(&b
->nb
), 1);
1222 b
->nb
.cursor
= nir_before_cf_list(&loop
->body
);
1225 nir_push_if(&b
->nb
, nir_load_var(&b
->nb
, do_cont
));
1227 vtn_emit_cf_list(b
, &vtn_loop
->cont_body
, NULL
, NULL
, handler
);
1229 nir_pop_if(&b
->nb
, cont_if
);
1231 nir_store_var(&b
->nb
, do_cont
, nir_imm_true(&b
->nb
), 1);
1233 b
->has_loop_continue
= true;
1236 nir_pop_loop(&b
->nb
, loop
);
1240 case vtn_cf_node_type_switch
: {
1241 struct vtn_switch
*vtn_switch
= vtn_cf_node_as_switch(node
);
1243 /* Before we can emit anything, we need to sort the list of cases in
1244 * fall-through order.
1246 vtn_switch_order_cases(vtn_switch
);
1248 /* First, we create a variable to keep track of whether or not the
1249 * switch is still going at any given point. Any switch breaks
1250 * will set this variable to false.
1252 nir_variable
*fall_var
=
1253 nir_local_variable_create(b
->nb
.impl
, glsl_bool_type(), "fall");
1254 nir_store_var(&b
->nb
, fall_var
, nir_imm_false(&b
->nb
), 1);
1256 nir_ssa_def
*sel
= vtn_ssa_value(b
, vtn_switch
->selector
)->def
;
1258 /* Now we can walk the list of cases and actually emit code */
1259 vtn_foreach_cf_node(case_node
, &vtn_switch
->cases
) {
1260 struct vtn_case
*cse
= vtn_cf_node_as_case(case_node
);
1262 /* Figure out the condition */
1264 vtn_switch_case_condition(b
, vtn_switch
, sel
, cse
);
1265 /* Take fallthrough into account */
1266 cond
= nir_ior(&b
->nb
, cond
, nir_load_var(&b
->nb
, fall_var
));
1268 nir_if
*case_if
= nir_push_if(&b
->nb
, cond
);
1270 bool has_break
= false;
1271 nir_store_var(&b
->nb
, fall_var
, nir_imm_true(&b
->nb
), 1);
1272 vtn_emit_cf_list(b
, &cse
->body
, fall_var
, &has_break
, handler
);
1273 (void)has_break
; /* We don't care */
1275 nir_pop_if(&b
->nb
, case_if
);
1282 vtn_fail("Invalid CF node type");
1288 vtn_function_emit(struct vtn_builder
*b
, struct vtn_function
*func
,
1289 vtn_instruction_handler instruction_handler
)
1291 nir_builder_init(&b
->nb
, func
->impl
);
1293 b
->nb
.cursor
= nir_after_cf_list(&func
->impl
->body
);
1294 b
->nb
.exact
= b
->exact
;
1295 b
->has_loop_continue
= false;
1296 b
->phi_table
= _mesa_pointer_hash_table_create(b
);
1298 vtn_emit_cf_list(b
, &func
->body
, NULL
, NULL
, instruction_handler
);
1300 vtn_foreach_instruction(b
, func
->start_block
->label
, func
->end
,
1301 vtn_handle_phi_second_pass
);
1303 nir_rematerialize_derefs_in_use_blocks_impl(func
->impl
);
1305 /* Continue blocks for loops get inserted before the body of the loop
1306 * but instructions in the continue may use SSA defs in the loop body.
1307 * Therefore, we need to repair SSA to insert the needed phi nodes.
1309 if (b
->has_loop_continue
)
1310 nir_repair_ssa_impl(func
->impl
);
1312 func
->emitted
= true;