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"
28 vtn_cfg_handle_prepass_instruction(struct vtn_builder
*b
, SpvOp opcode
,
29 const uint32_t *w
, unsigned count
)
33 vtn_assert(b
->func
== NULL
);
34 b
->func
= rzalloc(b
, struct vtn_function
);
36 list_inithead(&b
->func
->body
);
37 b
->func
->control
= w
[3];
39 MAYBE_UNUSED
const struct glsl_type
*result_type
=
40 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
41 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_function
);
44 const struct vtn_type
*func_type
=
45 vtn_value(b
, w
[4], vtn_value_type_type
)->type
;
47 vtn_assert(func_type
->return_type
->type
== result_type
);
50 nir_function_create(b
->shader
, ralloc_strdup(b
->shader
, val
->name
));
52 func
->num_params
= func_type
->length
;
53 func
->params
= ralloc_array(b
->shader
, nir_parameter
, func
->num_params
);
54 for (unsigned i
= 0; i
< func
->num_params
; i
++) {
55 if (func_type
->params
[i
]->base_type
== vtn_base_type_pointer
&&
56 func_type
->params
[i
]->type
== NULL
) {
57 func
->params
[i
].type
= func_type
->params
[i
]->deref
->type
;
59 func
->params
[i
].type
= func_type
->params
[i
]->type
;
62 /* TODO: We could do something smarter here. */
63 func
->params
[i
].param_type
= nir_parameter_inout
;
66 func
->return_type
= func_type
->return_type
->type
;
68 b
->func
->impl
= nir_function_impl_create(func
);
69 b
->nb
.cursor
= nir_before_cf_list(&b
->func
->impl
->body
);
71 b
->func_param_idx
= 0;
75 case SpvOpFunctionEnd
:
80 case SpvOpFunctionParameter
: {
81 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
83 vtn_assert(b
->func_param_idx
< b
->func
->impl
->num_params
);
84 nir_variable
*param
= b
->func
->impl
->params
[b
->func_param_idx
++];
86 if (type
->base_type
== vtn_base_type_pointer
&& type
->type
== NULL
) {
87 struct vtn_variable
*vtn_var
= rzalloc(b
, struct vtn_variable
);
88 vtn_var
->type
= type
->deref
;
91 vtn_assert(vtn_var
->type
->type
== param
->type
);
93 struct vtn_type
*without_array
= vtn_var
->type
;
94 while(glsl_type_is_array(without_array
->type
))
95 without_array
= without_array
->array_element
;
97 if (glsl_type_is_image(without_array
->type
)) {
98 vtn_var
->mode
= vtn_variable_mode_image
;
99 param
->interface_type
= without_array
->type
;
100 } else if (glsl_type_is_sampler(without_array
->type
)) {
101 vtn_var
->mode
= vtn_variable_mode_sampler
;
102 param
->interface_type
= without_array
->type
;
104 vtn_var
->mode
= vtn_variable_mode_param
;
107 struct vtn_value
*val
=
108 vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
110 /* Name the parameter so it shows up nicely in NIR */
111 param
->name
= ralloc_strdup(param
, val
->name
);
113 val
->pointer
= vtn_pointer_for_variable(b
, vtn_var
, type
);
115 /* We're a regular SSA value. */
116 struct vtn_ssa_value
*param_ssa
=
117 vtn_local_load(b
, nir_deref_var_create(b
, param
));
118 struct vtn_value
*val
= vtn_push_ssa(b
, w
[2], type
, param_ssa
);
120 /* Name the parameter so it shows up nicely in NIR */
121 param
->name
= ralloc_strdup(param
, val
->name
);
127 vtn_assert(b
->block
== NULL
);
128 b
->block
= rzalloc(b
, struct vtn_block
);
129 b
->block
->node
.type
= vtn_cf_node_type_block
;
131 vtn_push_value(b
, w
[1], vtn_value_type_block
)->block
= b
->block
;
133 if (b
->func
->start_block
== NULL
) {
134 /* This is the first block encountered for this function. In this
135 * case, we set the start block and add it to the list of
136 * implemented functions that we'll walk later.
138 b
->func
->start_block
= b
->block
;
139 exec_list_push_tail(&b
->functions
, &b
->func
->node
);
144 case SpvOpSelectionMerge
:
146 vtn_assert(b
->block
&& b
->block
->merge
== NULL
);
151 case SpvOpBranchConditional
:
155 case SpvOpReturnValue
:
156 case SpvOpUnreachable
:
157 vtn_assert(b
->block
&& b
->block
->branch
== NULL
);
158 b
->block
->branch
= w
;
163 /* Continue on as per normal */
171 vtn_add_case(struct vtn_builder
*b
, struct vtn_switch
*swtch
,
172 struct vtn_block
*break_block
,
173 uint32_t block_id
, uint32_t val
, bool is_default
)
175 struct vtn_block
*case_block
=
176 vtn_value(b
, block_id
, vtn_value_type_block
)->block
;
178 /* Don't create dummy cases that just break */
179 if (case_block
== break_block
)
182 if (case_block
->switch_case
== NULL
) {
183 struct vtn_case
*c
= ralloc(b
, struct vtn_case
);
185 list_inithead(&c
->body
);
186 c
->start_block
= case_block
;
187 c
->fallthrough
= NULL
;
188 util_dynarray_init(&c
->values
, b
);
189 c
->is_default
= false;
192 list_addtail(&c
->link
, &swtch
->cases
);
194 case_block
->switch_case
= c
;
198 case_block
->switch_case
->is_default
= true;
200 util_dynarray_append(&case_block
->switch_case
->values
, uint32_t, val
);
204 /* This function performs a depth-first search of the cases and puts them
205 * in fall-through order.
208 vtn_order_case(struct vtn_switch
*swtch
, struct vtn_case
*cse
)
215 list_del(&cse
->link
);
217 if (cse
->fallthrough
) {
218 vtn_order_case(swtch
, cse
->fallthrough
);
220 /* If we have a fall-through, place this case right before the case it
221 * falls through to. This ensures that fallthroughs come one after
222 * the other. These two can never get separated because that would
223 * imply something else falling through to the same case. Also, this
224 * can't break ordering because the DFS ensures that this case is
225 * visited before anything that falls through to it.
227 list_addtail(&cse
->link
, &cse
->fallthrough
->link
);
229 list_add(&cse
->link
, &swtch
->cases
);
233 static enum vtn_branch_type
234 vtn_get_branch_type(struct vtn_builder
*b
,
235 struct vtn_block
*block
,
236 struct vtn_case
*swcase
, struct vtn_block
*switch_break
,
237 struct vtn_block
*loop_break
, struct vtn_block
*loop_cont
)
239 if (block
->switch_case
) {
240 /* This branch is actually a fallthrough */
241 vtn_assert(swcase
->fallthrough
== NULL
||
242 swcase
->fallthrough
== block
->switch_case
);
243 swcase
->fallthrough
= block
->switch_case
;
244 return vtn_branch_type_switch_fallthrough
;
245 } else if (block
== loop_break
) {
246 return vtn_branch_type_loop_break
;
247 } else if (block
== loop_cont
) {
248 return vtn_branch_type_loop_continue
;
249 } else if (block
== switch_break
) {
250 return vtn_branch_type_switch_break
;
252 return vtn_branch_type_none
;
257 vtn_cfg_walk_blocks(struct vtn_builder
*b
, struct list_head
*cf_list
,
258 struct vtn_block
*start
, struct vtn_case
*switch_case
,
259 struct vtn_block
*switch_break
,
260 struct vtn_block
*loop_break
, struct vtn_block
*loop_cont
,
261 struct vtn_block
*end
)
263 struct vtn_block
*block
= start
;
264 while (block
!= end
) {
265 if (block
->merge
&& (*block
->merge
& SpvOpCodeMask
) == SpvOpLoopMerge
&&
267 struct vtn_loop
*loop
= ralloc(b
, struct vtn_loop
);
269 loop
->node
.type
= vtn_cf_node_type_loop
;
270 list_inithead(&loop
->body
);
271 list_inithead(&loop
->cont_body
);
272 loop
->control
= block
->merge
[3];
274 list_addtail(&loop
->node
.link
, cf_list
);
277 struct vtn_block
*new_loop_break
=
278 vtn_value(b
, block
->merge
[1], vtn_value_type_block
)->block
;
279 struct vtn_block
*new_loop_cont
=
280 vtn_value(b
, block
->merge
[2], vtn_value_type_block
)->block
;
282 /* Note: This recursive call will start with the current block as
283 * its start block. If we weren't careful, we would get here
284 * again and end up in infinite recursion. This is why we set
285 * block->loop above and check for it before creating one. This
286 * way, we only create the loop once and the second call that
287 * tries to handle this loop goes to the cases below and gets
288 * handled as a regular block.
290 * Note: When we make the recursive walk calls, we pass NULL for
291 * the switch break since you have to break out of the loop first.
292 * We do, however, still pass the current switch case because it's
293 * possible that the merge block for the loop is the start of
296 vtn_cfg_walk_blocks(b
, &loop
->body
, block
, switch_case
, NULL
,
297 new_loop_break
, new_loop_cont
, NULL
);
298 vtn_cfg_walk_blocks(b
, &loop
->cont_body
, new_loop_cont
, NULL
, NULL
,
299 new_loop_break
, NULL
, block
);
301 block
= new_loop_break
;
305 vtn_assert(block
->node
.link
.next
== NULL
);
306 list_addtail(&block
->node
.link
, cf_list
);
308 switch (*block
->branch
& SpvOpCodeMask
) {
310 struct vtn_block
*branch_block
=
311 vtn_value(b
, block
->branch
[1], vtn_value_type_block
)->block
;
313 block
->branch_type
= vtn_get_branch_type(b
, branch_block
,
314 switch_case
, switch_break
,
315 loop_break
, loop_cont
);
317 if (block
->branch_type
!= vtn_branch_type_none
)
320 block
= branch_block
;
325 case SpvOpReturnValue
:
326 block
->branch_type
= vtn_branch_type_return
;
330 block
->branch_type
= vtn_branch_type_discard
;
333 case SpvOpBranchConditional
: {
334 struct vtn_block
*then_block
=
335 vtn_value(b
, block
->branch
[2], vtn_value_type_block
)->block
;
336 struct vtn_block
*else_block
=
337 vtn_value(b
, block
->branch
[3], vtn_value_type_block
)->block
;
339 struct vtn_if
*if_stmt
= ralloc(b
, struct vtn_if
);
341 if_stmt
->node
.type
= vtn_cf_node_type_if
;
342 if_stmt
->condition
= block
->branch
[1];
343 list_inithead(&if_stmt
->then_body
);
344 list_inithead(&if_stmt
->else_body
);
346 list_addtail(&if_stmt
->node
.link
, cf_list
);
349 (*block
->merge
& SpvOpCodeMask
) == SpvOpSelectionMerge
) {
350 if_stmt
->control
= block
->merge
[2];
353 if_stmt
->then_type
= vtn_get_branch_type(b
, then_block
,
354 switch_case
, switch_break
,
355 loop_break
, loop_cont
);
356 if_stmt
->else_type
= vtn_get_branch_type(b
, else_block
,
357 switch_case
, switch_break
,
358 loop_break
, loop_cont
);
360 if (then_block
== else_block
) {
361 block
->branch_type
= if_stmt
->then_type
;
362 if (block
->branch_type
== vtn_branch_type_none
) {
368 } else if (if_stmt
->then_type
== vtn_branch_type_none
&&
369 if_stmt
->else_type
== vtn_branch_type_none
) {
370 /* Neither side of the if is something we can short-circuit. */
371 vtn_assert((*block
->merge
& SpvOpCodeMask
) == SpvOpSelectionMerge
);
372 struct vtn_block
*merge_block
=
373 vtn_value(b
, block
->merge
[1], vtn_value_type_block
)->block
;
375 vtn_cfg_walk_blocks(b
, &if_stmt
->then_body
, then_block
,
376 switch_case
, switch_break
,
377 loop_break
, loop_cont
, merge_block
);
378 vtn_cfg_walk_blocks(b
, &if_stmt
->else_body
, else_block
,
379 switch_case
, switch_break
,
380 loop_break
, loop_cont
, merge_block
);
382 enum vtn_branch_type merge_type
=
383 vtn_get_branch_type(b
, merge_block
, switch_case
, switch_break
,
384 loop_break
, loop_cont
);
385 if (merge_type
== vtn_branch_type_none
) {
391 } else if (if_stmt
->then_type
!= vtn_branch_type_none
&&
392 if_stmt
->else_type
!= vtn_branch_type_none
) {
393 /* Both sides were short-circuited. We're done here. */
396 /* Exeactly one side of the branch could be short-circuited.
397 * We set the branch up as a predicated break/continue and we
398 * continue on with the other side as if it were what comes
401 if (if_stmt
->then_type
== vtn_branch_type_none
) {
408 vtn_fail("Should have returned or continued");
412 vtn_assert((*block
->merge
& SpvOpCodeMask
) == SpvOpSelectionMerge
);
413 struct vtn_block
*break_block
=
414 vtn_value(b
, block
->merge
[1], vtn_value_type_block
)->block
;
416 struct vtn_switch
*swtch
= ralloc(b
, struct vtn_switch
);
418 swtch
->node
.type
= vtn_cf_node_type_switch
;
419 swtch
->selector
= block
->branch
[1];
420 list_inithead(&swtch
->cases
);
422 list_addtail(&swtch
->node
.link
, cf_list
);
424 /* First, we go through and record all of the cases. */
425 const uint32_t *branch_end
=
426 block
->branch
+ (block
->branch
[0] >> SpvWordCountShift
);
428 vtn_add_case(b
, swtch
, break_block
, block
->branch
[2], 0, true);
429 for (const uint32_t *w
= block
->branch
+ 3; w
< branch_end
; w
+= 2)
430 vtn_add_case(b
, swtch
, break_block
, w
[1], w
[0], false);
432 /* Now, we go through and walk the blocks. While we walk through
433 * the blocks, we also gather the much-needed fall-through
436 list_for_each_entry(struct vtn_case
, cse
, &swtch
->cases
, link
) {
437 vtn_assert(cse
->start_block
!= break_block
);
438 vtn_cfg_walk_blocks(b
, &cse
->body
, cse
->start_block
, cse
,
439 break_block
, loop_break
, loop_cont
, NULL
);
442 /* Finally, we walk over all of the cases one more time and put
443 * them in fall-through order.
445 for (const uint32_t *w
= block
->branch
+ 2; w
< branch_end
; w
+= 2) {
446 struct vtn_block
*case_block
=
447 vtn_value(b
, *w
, vtn_value_type_block
)->block
;
449 if (case_block
== break_block
)
452 vtn_assert(case_block
->switch_case
);
454 vtn_order_case(swtch
, case_block
->switch_case
);
457 enum vtn_branch_type branch_type
=
458 vtn_get_branch_type(b
, break_block
, switch_case
, NULL
,
459 loop_break
, loop_cont
);
461 if (branch_type
!= vtn_branch_type_none
) {
462 /* It is possible that the break is actually the continue block
463 * for the containing loop. In this case, we need to bail and let
464 * the loop parsing code handle the continue properly.
466 vtn_assert(branch_type
== vtn_branch_type_loop_continue
);
474 case SpvOpUnreachable
:
478 vtn_fail("Unhandled opcode");
484 vtn_build_cfg(struct vtn_builder
*b
, const uint32_t *words
, const uint32_t *end
)
486 vtn_foreach_instruction(b
, words
, end
,
487 vtn_cfg_handle_prepass_instruction
);
489 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
490 vtn_cfg_walk_blocks(b
, &func
->body
, func
->start_block
,
491 NULL
, NULL
, NULL
, NULL
, NULL
);
496 vtn_handle_phis_first_pass(struct vtn_builder
*b
, SpvOp opcode
,
497 const uint32_t *w
, unsigned count
)
499 if (opcode
== SpvOpLabel
)
500 return true; /* Nothing to do */
502 /* If this isn't a phi node, stop. */
503 if (opcode
!= SpvOpPhi
)
506 /* For handling phi nodes, we do a poor-man's out-of-ssa on the spot.
507 * For each phi, we create a variable with the appropreate type and
508 * do a load from that variable. Then, in a second pass, we add
509 * stores to that variable to each of the predecessor blocks.
511 * We could do something more intelligent here. However, in order to
512 * handle loops and things properly, we really need dominance
513 * information. It would end up basically being the into-SSA
514 * algorithm all over again. It's easier if we just let
515 * lower_vars_to_ssa do that for us instead of repeating it here.
517 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
518 nir_variable
*phi_var
=
519 nir_local_variable_create(b
->nb
.impl
, type
->type
, "phi");
520 _mesa_hash_table_insert(b
->phi_table
, w
, phi_var
);
522 vtn_push_ssa(b
, w
[2], type
,
523 vtn_local_load(b
, nir_deref_var_create(b
, phi_var
)));
529 vtn_handle_phi_second_pass(struct vtn_builder
*b
, SpvOp opcode
,
530 const uint32_t *w
, unsigned count
)
532 if (opcode
!= SpvOpPhi
)
535 struct hash_entry
*phi_entry
= _mesa_hash_table_search(b
->phi_table
, w
);
536 vtn_assert(phi_entry
);
537 nir_variable
*phi_var
= phi_entry
->data
;
539 for (unsigned i
= 3; i
< count
; i
+= 2) {
540 struct vtn_block
*pred
=
541 vtn_value(b
, w
[i
+ 1], vtn_value_type_block
)->block
;
543 b
->nb
.cursor
= nir_after_instr(&pred
->end_nop
->instr
);
545 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[i
]);
547 vtn_local_store(b
, src
, nir_deref_var_create(b
, phi_var
));
554 vtn_emit_branch(struct vtn_builder
*b
, enum vtn_branch_type branch_type
,
555 nir_variable
*switch_fall_var
, bool *has_switch_break
)
557 switch (branch_type
) {
558 case vtn_branch_type_switch_break
:
559 nir_store_var(&b
->nb
, switch_fall_var
, nir_imm_int(&b
->nb
, NIR_FALSE
), 1);
560 *has_switch_break
= true;
562 case vtn_branch_type_switch_fallthrough
:
563 break; /* Nothing to do */
564 case vtn_branch_type_loop_break
:
565 nir_jump(&b
->nb
, nir_jump_break
);
567 case vtn_branch_type_loop_continue
:
568 nir_jump(&b
->nb
, nir_jump_continue
);
570 case vtn_branch_type_return
:
571 nir_jump(&b
->nb
, nir_jump_return
);
573 case vtn_branch_type_discard
: {
574 nir_intrinsic_instr
*discard
=
575 nir_intrinsic_instr_create(b
->nb
.shader
, nir_intrinsic_discard
);
576 nir_builder_instr_insert(&b
->nb
, &discard
->instr
);
580 vtn_fail("Invalid branch type");
585 vtn_emit_cf_list(struct vtn_builder
*b
, struct list_head
*cf_list
,
586 nir_variable
*switch_fall_var
, bool *has_switch_break
,
587 vtn_instruction_handler handler
)
589 list_for_each_entry(struct vtn_cf_node
, node
, cf_list
, link
) {
590 switch (node
->type
) {
591 case vtn_cf_node_type_block
: {
592 struct vtn_block
*block
= (struct vtn_block
*)node
;
594 const uint32_t *block_start
= block
->label
;
595 const uint32_t *block_end
= block
->merge
? block
->merge
:
598 block_start
= vtn_foreach_instruction(b
, block_start
, block_end
,
599 vtn_handle_phis_first_pass
);
601 vtn_foreach_instruction(b
, block_start
, block_end
, handler
);
603 block
->end_nop
= nir_intrinsic_instr_create(b
->nb
.shader
,
605 nir_builder_instr_insert(&b
->nb
, &block
->end_nop
->instr
);
607 if ((*block
->branch
& SpvOpCodeMask
) == SpvOpReturnValue
) {
608 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, block
->branch
[1]);
609 vtn_local_store(b
, src
,
610 nir_deref_var_create(b
, b
->nb
.impl
->return_var
));
613 if (block
->branch_type
!= vtn_branch_type_none
) {
614 vtn_emit_branch(b
, block
->branch_type
,
615 switch_fall_var
, has_switch_break
);
621 case vtn_cf_node_type_if
: {
622 struct vtn_if
*vtn_if
= (struct vtn_if
*)node
;
623 bool sw_break
= false;
626 nir_push_if(&b
->nb
, vtn_ssa_value(b
, vtn_if
->condition
)->def
);
627 if (vtn_if
->then_type
== vtn_branch_type_none
) {
628 vtn_emit_cf_list(b
, &vtn_if
->then_body
,
629 switch_fall_var
, &sw_break
, handler
);
631 vtn_emit_branch(b
, vtn_if
->then_type
, switch_fall_var
, &sw_break
);
634 nir_push_else(&b
->nb
, nif
);
635 if (vtn_if
->else_type
== vtn_branch_type_none
) {
636 vtn_emit_cf_list(b
, &vtn_if
->else_body
,
637 switch_fall_var
, &sw_break
, handler
);
639 vtn_emit_branch(b
, vtn_if
->else_type
, switch_fall_var
, &sw_break
);
642 nir_pop_if(&b
->nb
, nif
);
644 /* If we encountered a switch break somewhere inside of the if,
645 * then it would have been handled correctly by calling
646 * emit_cf_list or emit_branch for the interrior. However, we
647 * need to predicate everything following on wether or not we're
651 *has_switch_break
= true;
652 nir_push_if(&b
->nb
, nir_load_var(&b
->nb
, switch_fall_var
));
657 case vtn_cf_node_type_loop
: {
658 struct vtn_loop
*vtn_loop
= (struct vtn_loop
*)node
;
660 nir_loop
*loop
= nir_push_loop(&b
->nb
);
661 vtn_emit_cf_list(b
, &vtn_loop
->body
, NULL
, NULL
, handler
);
663 if (!list_empty(&vtn_loop
->cont_body
)) {
664 /* If we have a non-trivial continue body then we need to put
665 * it at the beginning of the loop with a flag to ensure that
666 * it doesn't get executed in the first iteration.
668 nir_variable
*do_cont
=
669 nir_local_variable_create(b
->nb
.impl
, glsl_bool_type(), "cont");
671 b
->nb
.cursor
= nir_before_cf_node(&loop
->cf_node
);
672 nir_store_var(&b
->nb
, do_cont
, nir_imm_int(&b
->nb
, NIR_FALSE
), 1);
674 b
->nb
.cursor
= nir_before_cf_list(&loop
->body
);
677 nir_push_if(&b
->nb
, nir_load_var(&b
->nb
, do_cont
));
679 vtn_emit_cf_list(b
, &vtn_loop
->cont_body
, NULL
, NULL
, handler
);
681 nir_pop_if(&b
->nb
, cont_if
);
683 nir_store_var(&b
->nb
, do_cont
, nir_imm_int(&b
->nb
, NIR_TRUE
), 1);
685 b
->has_loop_continue
= true;
688 nir_pop_loop(&b
->nb
, loop
);
692 case vtn_cf_node_type_switch
: {
693 struct vtn_switch
*vtn_switch
= (struct vtn_switch
*)node
;
695 /* First, we create a variable to keep track of whether or not the
696 * switch is still going at any given point. Any switch breaks
697 * will set this variable to false.
699 nir_variable
*fall_var
=
700 nir_local_variable_create(b
->nb
.impl
, glsl_bool_type(), "fall");
701 nir_store_var(&b
->nb
, fall_var
, nir_imm_int(&b
->nb
, NIR_FALSE
), 1);
703 /* Next, we gather up all of the conditions. We have to do this
704 * up-front because we also need to build an "any" condition so
705 * that we can use !any for default.
707 const int num_cases
= list_length(&vtn_switch
->cases
);
708 NIR_VLA(nir_ssa_def
*, conditions
, num_cases
);
710 nir_ssa_def
*sel
= vtn_ssa_value(b
, vtn_switch
->selector
)->def
;
711 /* An accumulation of all conditions. Used for the default */
712 nir_ssa_def
*any
= NULL
;
715 list_for_each_entry(struct vtn_case
, cse
, &vtn_switch
->cases
, link
) {
716 if (cse
->is_default
) {
717 conditions
[i
++] = NULL
;
721 nir_ssa_def
*cond
= NULL
;
722 util_dynarray_foreach(&cse
->values
, uint32_t, val
) {
723 nir_ssa_def
*is_val
=
724 nir_ieq(&b
->nb
, sel
, nir_imm_int(&b
->nb
, *val
));
726 cond
= cond
? nir_ior(&b
->nb
, cond
, is_val
) : is_val
;
729 any
= any
? nir_ior(&b
->nb
, any
, cond
) : cond
;
730 conditions
[i
++] = cond
;
732 vtn_assert(i
== num_cases
);
734 /* Now we can walk the list of cases and actually emit code */
736 list_for_each_entry(struct vtn_case
, cse
, &vtn_switch
->cases
, link
) {
737 /* Figure out the condition */
738 nir_ssa_def
*cond
= conditions
[i
++];
739 if (cse
->is_default
) {
740 vtn_assert(cond
== NULL
);
741 cond
= nir_inot(&b
->nb
, any
);
743 /* Take fallthrough into account */
744 cond
= nir_ior(&b
->nb
, cond
, nir_load_var(&b
->nb
, fall_var
));
746 nir_if
*case_if
= nir_push_if(&b
->nb
, cond
);
748 bool has_break
= false;
749 nir_store_var(&b
->nb
, fall_var
, nir_imm_int(&b
->nb
, NIR_TRUE
), 1);
750 vtn_emit_cf_list(b
, &cse
->body
, fall_var
, &has_break
, handler
);
751 (void)has_break
; /* We don't care */
753 nir_pop_if(&b
->nb
, case_if
);
755 vtn_assert(i
== num_cases
);
761 vtn_fail("Invalid CF node type");
767 vtn_function_emit(struct vtn_builder
*b
, struct vtn_function
*func
,
768 vtn_instruction_handler instruction_handler
)
770 nir_builder_init(&b
->nb
, func
->impl
);
771 b
->nb
.cursor
= nir_after_cf_list(&func
->impl
->body
);
772 b
->has_loop_continue
= false;
773 b
->phi_table
= _mesa_hash_table_create(b
, _mesa_hash_pointer
,
774 _mesa_key_pointer_equal
);
776 vtn_emit_cf_list(b
, &func
->body
, NULL
, NULL
, instruction_handler
);
778 vtn_foreach_instruction(b
, func
->start_block
->label
, func
->end
,
779 vtn_handle_phi_second_pass
);
781 /* Continue blocks for loops get inserted before the body of the loop
782 * but instructions in the continue may use SSA defs in the loop body.
783 * Therefore, we need to repair SSA to insert the needed phi nodes.
785 if (b
->has_loop_continue
)
786 nir_repair_ssa_impl(func
->impl
);
788 func
->emitted
= true;