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_pointer
*
28 vtn_load_param_pointer(struct vtn_builder
*b
,
29 struct vtn_type
*param_type
,
32 struct vtn_type
*ptr_type
= param_type
;
33 if (param_type
->base_type
!= vtn_base_type_pointer
) {
34 assert(param_type
->base_type
== vtn_base_type_image
||
35 param_type
->base_type
== vtn_base_type_sampler
);
36 ptr_type
= rzalloc(b
, struct vtn_type
);
37 ptr_type
->base_type
= vtn_base_type_pointer
;
38 ptr_type
->deref
= param_type
;
39 ptr_type
->storage_class
= SpvStorageClassUniformConstant
;
42 return vtn_pointer_from_ssa(b
, nir_load_param(&b
->nb
, param_idx
), ptr_type
);
46 vtn_type_count_function_params(struct vtn_type
*type
)
48 switch (type
->base_type
) {
49 case vtn_base_type_array
:
50 case vtn_base_type_matrix
:
51 return type
->length
* vtn_type_count_function_params(type
->array_element
);
53 case vtn_base_type_struct
: {
55 for (unsigned i
= 0; i
< type
->length
; i
++)
56 count
+= vtn_type_count_function_params(type
->members
[i
]);
60 case vtn_base_type_sampled_image
:
69 vtn_type_add_to_function_params(struct vtn_type
*type
,
73 static const nir_parameter nir_deref_param
= {
78 switch (type
->base_type
) {
79 case vtn_base_type_array
:
80 case vtn_base_type_matrix
:
81 for (unsigned i
= 0; i
< type
->length
; i
++)
82 vtn_type_add_to_function_params(type
->array_element
, func
, param_idx
);
85 case vtn_base_type_struct
:
86 for (unsigned i
= 0; i
< type
->length
; i
++)
87 vtn_type_add_to_function_params(type
->members
[i
], func
, param_idx
);
90 case vtn_base_type_sampled_image
:
91 func
->params
[(*param_idx
)++] = nir_deref_param
;
92 func
->params
[(*param_idx
)++] = nir_deref_param
;
95 case vtn_base_type_image
:
96 case vtn_base_type_sampler
:
97 func
->params
[(*param_idx
)++] = nir_deref_param
;
100 case vtn_base_type_pointer
:
102 func
->params
[(*param_idx
)++] = (nir_parameter
) {
103 .num_components
= glsl_get_vector_elements(type
->type
),
104 .bit_size
= glsl_get_bit_size(type
->type
),
107 func
->params
[(*param_idx
)++] = nir_deref_param
;
112 func
->params
[(*param_idx
)++] = (nir_parameter
) {
113 .num_components
= glsl_get_vector_elements(type
->type
),
114 .bit_size
= glsl_get_bit_size(type
->type
),
120 vtn_ssa_value_add_to_call_params(struct vtn_builder
*b
,
121 struct vtn_ssa_value
*value
,
122 struct vtn_type
*type
,
123 nir_call_instr
*call
,
126 switch (type
->base_type
) {
127 case vtn_base_type_array
:
128 case vtn_base_type_matrix
:
129 for (unsigned i
= 0; i
< type
->length
; i
++) {
130 vtn_ssa_value_add_to_call_params(b
, value
->elems
[i
],
136 case vtn_base_type_struct
:
137 for (unsigned i
= 0; i
< type
->length
; i
++) {
138 vtn_ssa_value_add_to_call_params(b
, value
->elems
[i
],
145 call
->params
[(*param_idx
)++] = nir_src_for_ssa(value
->def
);
151 vtn_ssa_value_load_function_param(struct vtn_builder
*b
,
152 struct vtn_ssa_value
*value
,
153 struct vtn_type
*type
,
156 switch (type
->base_type
) {
157 case vtn_base_type_array
:
158 case vtn_base_type_matrix
:
159 for (unsigned i
= 0; i
< type
->length
; i
++) {
160 vtn_ssa_value_load_function_param(b
, value
->elems
[i
],
161 type
->array_element
, param_idx
);
165 case vtn_base_type_struct
:
166 for (unsigned i
= 0; i
< type
->length
; i
++) {
167 vtn_ssa_value_load_function_param(b
, value
->elems
[i
],
168 type
->members
[i
], param_idx
);
173 value
->def
= nir_load_param(&b
->nb
, (*param_idx
)++);
179 vtn_handle_function_call(struct vtn_builder
*b
, SpvOp opcode
,
180 const uint32_t *w
, unsigned count
)
182 struct vtn_type
*res_type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
183 struct vtn_function
*vtn_callee
=
184 vtn_value(b
, w
[3], vtn_value_type_function
)->func
;
185 struct nir_function
*callee
= vtn_callee
->impl
->function
;
187 vtn_callee
->referenced
= true;
189 nir_call_instr
*call
= nir_call_instr_create(b
->nb
.shader
, callee
);
191 unsigned param_idx
= 0;
193 nir_deref_instr
*ret_deref
= NULL
;
194 struct vtn_type
*ret_type
= vtn_callee
->type
->return_type
;
195 if (ret_type
->base_type
!= vtn_base_type_void
) {
196 nir_variable
*ret_tmp
=
197 nir_local_variable_create(b
->nb
.impl
,
198 glsl_get_bare_type(ret_type
->type
),
200 ret_deref
= nir_build_deref_var(&b
->nb
, ret_tmp
);
201 call
->params
[param_idx
++] = nir_src_for_ssa(&ret_deref
->dest
.ssa
);
204 for (unsigned i
= 0; i
< vtn_callee
->type
->length
; i
++) {
205 struct vtn_type
*arg_type
= vtn_callee
->type
->params
[i
];
206 unsigned arg_id
= w
[4 + i
];
208 if (arg_type
->base_type
== vtn_base_type_sampled_image
) {
209 struct vtn_sampled_image
*sampled_image
=
210 vtn_value(b
, arg_id
, vtn_value_type_sampled_image
)->sampled_image
;
212 call
->params
[param_idx
++] =
213 nir_src_for_ssa(&sampled_image
->image
->deref
->dest
.ssa
);
214 call
->params
[param_idx
++] =
215 nir_src_for_ssa(&sampled_image
->sampler
->deref
->dest
.ssa
);
216 } else if (arg_type
->base_type
== vtn_base_type_pointer
||
217 arg_type
->base_type
== vtn_base_type_image
||
218 arg_type
->base_type
== vtn_base_type_sampler
) {
219 struct vtn_pointer
*pointer
=
220 vtn_value(b
, arg_id
, vtn_value_type_pointer
)->pointer
;
221 call
->params
[param_idx
++] =
222 nir_src_for_ssa(vtn_pointer_to_ssa(b
, pointer
));
224 vtn_ssa_value_add_to_call_params(b
, vtn_ssa_value(b
, arg_id
),
225 arg_type
, call
, ¶m_idx
);
228 assert(param_idx
== call
->num_params
);
230 nir_builder_instr_insert(&b
->nb
, &call
->instr
);
232 if (ret_type
->base_type
== vtn_base_type_void
) {
233 vtn_push_value(b
, w
[2], vtn_value_type_undef
);
235 vtn_push_ssa(b
, w
[2], res_type
, vtn_local_load(b
, ret_deref
, 0));
240 vtn_cfg_handle_prepass_instruction(struct vtn_builder
*b
, SpvOp opcode
,
241 const uint32_t *w
, unsigned count
)
244 case SpvOpFunction
: {
245 vtn_assert(b
->func
== NULL
);
246 b
->func
= rzalloc(b
, struct vtn_function
);
248 list_inithead(&b
->func
->body
);
249 b
->func
->control
= w
[3];
251 MAYBE_UNUSED
const struct glsl_type
*result_type
=
252 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
253 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_function
);
256 b
->func
->type
= vtn_value(b
, w
[4], vtn_value_type_type
)->type
;
257 const struct vtn_type
*func_type
= b
->func
->type
;
259 vtn_assert(func_type
->return_type
->type
== result_type
);
262 nir_function_create(b
->shader
, ralloc_strdup(b
->shader
, val
->name
));
264 unsigned num_params
= 0;
265 for (unsigned i
= 0; i
< func_type
->length
; i
++)
266 num_params
+= vtn_type_count_function_params(func_type
->params
[i
]);
268 /* Add one parameter for the function return value */
269 if (func_type
->return_type
->base_type
!= vtn_base_type_void
)
272 func
->num_params
= num_params
;
273 func
->params
= ralloc_array(b
->shader
, nir_parameter
, num_params
);
276 if (func_type
->return_type
->base_type
!= vtn_base_type_void
) {
277 /* The return value is a regular pointer */
278 func
->params
[idx
++] = (nir_parameter
) {
279 .num_components
= 1, .bit_size
= 32,
283 for (unsigned i
= 0; i
< func_type
->length
; i
++)
284 vtn_type_add_to_function_params(func_type
->params
[i
], func
, &idx
);
285 assert(idx
== num_params
);
287 b
->func
->impl
= nir_function_impl_create(func
);
288 nir_builder_init(&b
->nb
, func
->impl
);
289 b
->nb
.cursor
= nir_before_cf_list(&b
->func
->impl
->body
);
290 b
->nb
.exact
= b
->exact
;
292 b
->func_param_idx
= 0;
294 /* The return value is the first parameter */
295 if (func_type
->return_type
->base_type
!= vtn_base_type_void
)
300 case SpvOpFunctionEnd
:
305 case SpvOpFunctionParameter
: {
306 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
308 vtn_assert(b
->func_param_idx
< b
->func
->impl
->function
->num_params
);
310 if (type
->base_type
== vtn_base_type_sampled_image
) {
311 /* Sampled images are actually two parameters. The first is the
312 * image and the second is the sampler.
314 struct vtn_value
*val
=
315 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
317 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
318 val
->sampled_image
->type
= type
;
320 struct vtn_type
*sampler_type
= rzalloc(b
, struct vtn_type
);
321 sampler_type
->base_type
= vtn_base_type_sampler
;
322 sampler_type
->type
= glsl_bare_sampler_type();
324 val
->sampled_image
->image
=
325 vtn_load_param_pointer(b
, type
, b
->func_param_idx
++);
326 val
->sampled_image
->sampler
=
327 vtn_load_param_pointer(b
, sampler_type
, b
->func_param_idx
++);
328 } else if (type
->base_type
== vtn_base_type_pointer
&&
329 type
->type
!= NULL
) {
330 /* This is a pointer with an actual storage type */
331 struct vtn_value
*val
=
332 vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
333 nir_ssa_def
*ssa_ptr
= nir_load_param(&b
->nb
, b
->func_param_idx
++);
334 val
->pointer
= vtn_pointer_from_ssa(b
, ssa_ptr
, type
);
335 } else if (type
->base_type
== vtn_base_type_pointer
||
336 type
->base_type
== vtn_base_type_image
||
337 type
->base_type
== vtn_base_type_sampler
) {
338 struct vtn_value
*val
=
339 vtn_push_value(b
, w
[2], vtn_value_type_pointer
);
341 vtn_load_param_pointer(b
, type
, b
->func_param_idx
++);
343 /* We're a regular SSA value. */
344 struct vtn_ssa_value
*value
= vtn_create_ssa_value(b
, type
->type
);
345 vtn_ssa_value_load_function_param(b
, value
, type
, &b
->func_param_idx
);
346 vtn_push_ssa(b
, w
[2], type
, value
);
352 vtn_assert(b
->block
== NULL
);
353 b
->block
= rzalloc(b
, struct vtn_block
);
354 b
->block
->node
.type
= vtn_cf_node_type_block
;
356 vtn_push_value(b
, w
[1], vtn_value_type_block
)->block
= b
->block
;
358 if (b
->func
->start_block
== NULL
) {
359 /* This is the first block encountered for this function. In this
360 * case, we set the start block and add it to the list of
361 * implemented functions that we'll walk later.
363 b
->func
->start_block
= b
->block
;
364 exec_list_push_tail(&b
->functions
, &b
->func
->node
);
369 case SpvOpSelectionMerge
:
371 vtn_assert(b
->block
&& b
->block
->merge
== NULL
);
376 case SpvOpBranchConditional
:
380 case SpvOpReturnValue
:
381 case SpvOpUnreachable
:
382 vtn_assert(b
->block
&& b
->block
->branch
== NULL
);
383 b
->block
->branch
= w
;
388 /* Continue on as per normal */
396 vtn_add_case(struct vtn_builder
*b
, struct vtn_switch
*swtch
,
397 struct vtn_block
*break_block
,
398 uint32_t block_id
, uint64_t val
, bool is_default
)
400 struct vtn_block
*case_block
=
401 vtn_value(b
, block_id
, vtn_value_type_block
)->block
;
403 /* Don't create dummy cases that just break */
404 if (case_block
== break_block
)
407 if (case_block
->switch_case
== NULL
) {
408 struct vtn_case
*c
= ralloc(b
, struct vtn_case
);
410 list_inithead(&c
->body
);
411 c
->start_block
= case_block
;
412 c
->fallthrough
= NULL
;
413 util_dynarray_init(&c
->values
, b
);
414 c
->is_default
= false;
417 list_addtail(&c
->link
, &swtch
->cases
);
419 case_block
->switch_case
= c
;
423 case_block
->switch_case
->is_default
= true;
425 util_dynarray_append(&case_block
->switch_case
->values
, uint64_t, val
);
429 /* This function performs a depth-first search of the cases and puts them
430 * in fall-through order.
433 vtn_order_case(struct vtn_switch
*swtch
, struct vtn_case
*cse
)
440 list_del(&cse
->link
);
442 if (cse
->fallthrough
) {
443 vtn_order_case(swtch
, cse
->fallthrough
);
445 /* If we have a fall-through, place this case right before the case it
446 * falls through to. This ensures that fallthroughs come one after
447 * the other. These two can never get separated because that would
448 * imply something else falling through to the same case. Also, this
449 * can't break ordering because the DFS ensures that this case is
450 * visited before anything that falls through to it.
452 list_addtail(&cse
->link
, &cse
->fallthrough
->link
);
454 list_add(&cse
->link
, &swtch
->cases
);
458 static enum vtn_branch_type
459 vtn_get_branch_type(struct vtn_builder
*b
,
460 struct vtn_block
*block
,
461 struct vtn_case
*swcase
, struct vtn_block
*switch_break
,
462 struct vtn_block
*loop_break
, struct vtn_block
*loop_cont
)
464 if (block
->switch_case
) {
465 /* This branch is actually a fallthrough */
466 vtn_assert(swcase
->fallthrough
== NULL
||
467 swcase
->fallthrough
== block
->switch_case
);
468 swcase
->fallthrough
= block
->switch_case
;
469 return vtn_branch_type_switch_fallthrough
;
470 } else if (block
== loop_break
) {
471 return vtn_branch_type_loop_break
;
472 } else if (block
== loop_cont
) {
473 return vtn_branch_type_loop_continue
;
474 } else if (block
== switch_break
) {
475 return vtn_branch_type_switch_break
;
477 return vtn_branch_type_none
;
482 vtn_cfg_walk_blocks(struct vtn_builder
*b
, struct list_head
*cf_list
,
483 struct vtn_block
*start
, struct vtn_case
*switch_case
,
484 struct vtn_block
*switch_break
,
485 struct vtn_block
*loop_break
, struct vtn_block
*loop_cont
,
486 struct vtn_block
*end
)
488 struct vtn_block
*block
= start
;
489 while (block
!= end
) {
490 if (block
->merge
&& (*block
->merge
& SpvOpCodeMask
) == SpvOpLoopMerge
&&
492 struct vtn_loop
*loop
= ralloc(b
, struct vtn_loop
);
494 loop
->node
.type
= vtn_cf_node_type_loop
;
495 list_inithead(&loop
->body
);
496 list_inithead(&loop
->cont_body
);
497 loop
->control
= block
->merge
[3];
499 list_addtail(&loop
->node
.link
, cf_list
);
502 struct vtn_block
*new_loop_break
=
503 vtn_value(b
, block
->merge
[1], vtn_value_type_block
)->block
;
504 struct vtn_block
*new_loop_cont
=
505 vtn_value(b
, block
->merge
[2], vtn_value_type_block
)->block
;
507 /* Note: This recursive call will start with the current block as
508 * its start block. If we weren't careful, we would get here
509 * again and end up in infinite recursion. This is why we set
510 * block->loop above and check for it before creating one. This
511 * way, we only create the loop once and the second call that
512 * tries to handle this loop goes to the cases below and gets
513 * handled as a regular block.
515 * Note: When we make the recursive walk calls, we pass NULL for
516 * the switch break since you have to break out of the loop first.
517 * We do, however, still pass the current switch case because it's
518 * possible that the merge block for the loop is the start of
521 vtn_cfg_walk_blocks(b
, &loop
->body
, block
, switch_case
, NULL
,
522 new_loop_break
, new_loop_cont
, NULL
);
523 vtn_cfg_walk_blocks(b
, &loop
->cont_body
, new_loop_cont
, NULL
, NULL
,
524 new_loop_break
, NULL
, block
);
526 enum vtn_branch_type branch_type
=
527 vtn_get_branch_type(b
, new_loop_break
, switch_case
, switch_break
,
528 loop_break
, loop_cont
);
530 if (branch_type
!= vtn_branch_type_none
) {
531 /* Stop walking through the CFG when this inner loop's break block
532 * ends up as the same block as the outer loop's continue block
533 * because we are already going to visit it.
535 vtn_assert(branch_type
== vtn_branch_type_loop_continue
);
539 block
= new_loop_break
;
543 vtn_assert(block
->node
.link
.next
== NULL
);
544 list_addtail(&block
->node
.link
, cf_list
);
546 switch (*block
->branch
& SpvOpCodeMask
) {
548 struct vtn_block
*branch_block
=
549 vtn_value(b
, block
->branch
[1], vtn_value_type_block
)->block
;
551 block
->branch_type
= vtn_get_branch_type(b
, branch_block
,
552 switch_case
, switch_break
,
553 loop_break
, loop_cont
);
555 if (block
->branch_type
!= vtn_branch_type_none
)
558 block
= branch_block
;
563 case SpvOpReturnValue
:
564 block
->branch_type
= vtn_branch_type_return
;
568 block
->branch_type
= vtn_branch_type_discard
;
571 case SpvOpBranchConditional
: {
572 struct vtn_block
*then_block
=
573 vtn_value(b
, block
->branch
[2], vtn_value_type_block
)->block
;
574 struct vtn_block
*else_block
=
575 vtn_value(b
, block
->branch
[3], vtn_value_type_block
)->block
;
577 struct vtn_if
*if_stmt
= ralloc(b
, struct vtn_if
);
579 if_stmt
->node
.type
= vtn_cf_node_type_if
;
580 if_stmt
->condition
= block
->branch
[1];
581 list_inithead(&if_stmt
->then_body
);
582 list_inithead(&if_stmt
->else_body
);
584 list_addtail(&if_stmt
->node
.link
, cf_list
);
587 (*block
->merge
& SpvOpCodeMask
) == SpvOpSelectionMerge
) {
588 if_stmt
->control
= block
->merge
[2];
590 if_stmt
->control
= SpvSelectionControlMaskNone
;
593 if_stmt
->then_type
= vtn_get_branch_type(b
, then_block
,
594 switch_case
, switch_break
,
595 loop_break
, loop_cont
);
596 if_stmt
->else_type
= vtn_get_branch_type(b
, else_block
,
597 switch_case
, switch_break
,
598 loop_break
, loop_cont
);
600 if (then_block
== else_block
) {
601 block
->branch_type
= if_stmt
->then_type
;
602 if (block
->branch_type
== vtn_branch_type_none
) {
608 } else if (if_stmt
->then_type
== vtn_branch_type_none
&&
609 if_stmt
->else_type
== vtn_branch_type_none
) {
610 /* Neither side of the if is something we can short-circuit. */
611 vtn_assert((*block
->merge
& SpvOpCodeMask
) == SpvOpSelectionMerge
);
612 struct vtn_block
*merge_block
=
613 vtn_value(b
, block
->merge
[1], vtn_value_type_block
)->block
;
615 vtn_cfg_walk_blocks(b
, &if_stmt
->then_body
, then_block
,
616 switch_case
, switch_break
,
617 loop_break
, loop_cont
, merge_block
);
618 vtn_cfg_walk_blocks(b
, &if_stmt
->else_body
, else_block
,
619 switch_case
, switch_break
,
620 loop_break
, loop_cont
, merge_block
);
622 enum vtn_branch_type merge_type
=
623 vtn_get_branch_type(b
, merge_block
, switch_case
, switch_break
,
624 loop_break
, loop_cont
);
625 if (merge_type
== vtn_branch_type_none
) {
631 } else if (if_stmt
->then_type
!= vtn_branch_type_none
&&
632 if_stmt
->else_type
!= vtn_branch_type_none
) {
633 /* Both sides were short-circuited. We're done here. */
636 /* Exeactly one side of the branch could be short-circuited.
637 * We set the branch up as a predicated break/continue and we
638 * continue on with the other side as if it were what comes
641 if (if_stmt
->then_type
== vtn_branch_type_none
) {
648 vtn_fail("Should have returned or continued");
652 vtn_assert((*block
->merge
& SpvOpCodeMask
) == SpvOpSelectionMerge
);
653 struct vtn_block
*break_block
=
654 vtn_value(b
, block
->merge
[1], vtn_value_type_block
)->block
;
656 struct vtn_switch
*swtch
= ralloc(b
, struct vtn_switch
);
658 swtch
->node
.type
= vtn_cf_node_type_switch
;
659 swtch
->selector
= block
->branch
[1];
660 list_inithead(&swtch
->cases
);
662 list_addtail(&swtch
->node
.link
, cf_list
);
664 /* First, we go through and record all of the cases. */
665 const uint32_t *branch_end
=
666 block
->branch
+ (block
->branch
[0] >> SpvWordCountShift
);
668 struct vtn_value
*cond_val
= vtn_untyped_value(b
, block
->branch
[1]);
669 vtn_fail_if(!cond_val
->type
||
670 cond_val
->type
->base_type
!= vtn_base_type_scalar
,
671 "Selector of OpSelect must have a type of OpTypeInt");
673 nir_alu_type cond_type
=
674 nir_get_nir_type_for_glsl_type(cond_val
->type
->type
);
675 vtn_fail_if(nir_alu_type_get_base_type(cond_type
) != nir_type_int
&&
676 nir_alu_type_get_base_type(cond_type
) != nir_type_uint
,
677 "Selector of OpSelect must have a type of OpTypeInt");
679 bool is_default
= true;
680 const unsigned bitsize
= nir_alu_type_get_type_size(cond_type
);
681 for (const uint32_t *w
= block
->branch
+ 2; w
< branch_end
;) {
682 uint64_t literal
= 0;
687 assert(bitsize
== 64);
688 literal
= vtn_u64_literal(w
);
693 uint32_t block_id
= *(w
++);
695 vtn_add_case(b
, swtch
, break_block
, block_id
, literal
, is_default
);
699 /* Now, we go through and walk the blocks. While we walk through
700 * the blocks, we also gather the much-needed fall-through
703 list_for_each_entry(struct vtn_case
, cse
, &swtch
->cases
, link
) {
704 vtn_assert(cse
->start_block
!= break_block
);
705 vtn_cfg_walk_blocks(b
, &cse
->body
, cse
->start_block
, cse
,
706 break_block
, loop_break
, loop_cont
, NULL
);
709 /* Finally, we walk over all of the cases one more time and put
710 * them in fall-through order.
712 for (const uint32_t *w
= block
->branch
+ 2; w
< branch_end
;) {
713 struct vtn_block
*case_block
=
714 vtn_value(b
, *w
, vtn_value_type_block
)->block
;
719 assert(bitsize
== 64);
723 if (case_block
== break_block
)
726 vtn_assert(case_block
->switch_case
);
728 vtn_order_case(swtch
, case_block
->switch_case
);
731 enum vtn_branch_type branch_type
=
732 vtn_get_branch_type(b
, break_block
, switch_case
, NULL
,
733 loop_break
, loop_cont
);
735 if (branch_type
!= vtn_branch_type_none
) {
736 /* It is possible that the break is actually the continue block
737 * for the containing loop. In this case, we need to bail and let
738 * the loop parsing code handle the continue properly.
740 vtn_assert(branch_type
== vtn_branch_type_loop_continue
);
748 case SpvOpUnreachable
:
752 vtn_fail("Unhandled opcode");
758 vtn_build_cfg(struct vtn_builder
*b
, const uint32_t *words
, const uint32_t *end
)
760 vtn_foreach_instruction(b
, words
, end
,
761 vtn_cfg_handle_prepass_instruction
);
763 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
764 vtn_cfg_walk_blocks(b
, &func
->body
, func
->start_block
,
765 NULL
, NULL
, NULL
, NULL
, NULL
);
770 vtn_handle_phis_first_pass(struct vtn_builder
*b
, SpvOp opcode
,
771 const uint32_t *w
, unsigned count
)
773 if (opcode
== SpvOpLabel
)
774 return true; /* Nothing to do */
776 /* If this isn't a phi node, stop. */
777 if (opcode
!= SpvOpPhi
)
780 /* For handling phi nodes, we do a poor-man's out-of-ssa on the spot.
781 * For each phi, we create a variable with the appropreate type and
782 * do a load from that variable. Then, in a second pass, we add
783 * stores to that variable to each of the predecessor blocks.
785 * We could do something more intelligent here. However, in order to
786 * handle loops and things properly, we really need dominance
787 * information. It would end up basically being the into-SSA
788 * algorithm all over again. It's easier if we just let
789 * lower_vars_to_ssa do that for us instead of repeating it here.
791 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
792 nir_variable
*phi_var
=
793 nir_local_variable_create(b
->nb
.impl
, type
->type
, "phi");
794 _mesa_hash_table_insert(b
->phi_table
, w
, phi_var
);
796 vtn_push_ssa(b
, w
[2], type
,
797 vtn_local_load(b
, nir_build_deref_var(&b
->nb
, phi_var
), 0));
803 vtn_handle_phi_second_pass(struct vtn_builder
*b
, SpvOp opcode
,
804 const uint32_t *w
, unsigned count
)
806 if (opcode
!= SpvOpPhi
)
809 struct hash_entry
*phi_entry
= _mesa_hash_table_search(b
->phi_table
, w
);
810 vtn_assert(phi_entry
);
811 nir_variable
*phi_var
= phi_entry
->data
;
813 for (unsigned i
= 3; i
< count
; i
+= 2) {
814 struct vtn_block
*pred
=
815 vtn_value(b
, w
[i
+ 1], vtn_value_type_block
)->block
;
817 b
->nb
.cursor
= nir_after_instr(&pred
->end_nop
->instr
);
819 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[i
]);
821 vtn_local_store(b
, src
, nir_build_deref_var(&b
->nb
, phi_var
), 0);
828 vtn_emit_branch(struct vtn_builder
*b
, enum vtn_branch_type branch_type
,
829 nir_variable
*switch_fall_var
, bool *has_switch_break
)
831 switch (branch_type
) {
832 case vtn_branch_type_switch_break
:
833 nir_store_var(&b
->nb
, switch_fall_var
, nir_imm_false(&b
->nb
), 1);
834 *has_switch_break
= true;
836 case vtn_branch_type_switch_fallthrough
:
837 break; /* Nothing to do */
838 case vtn_branch_type_loop_break
:
839 nir_jump(&b
->nb
, nir_jump_break
);
841 case vtn_branch_type_loop_continue
:
842 nir_jump(&b
->nb
, nir_jump_continue
);
844 case vtn_branch_type_return
:
845 nir_jump(&b
->nb
, nir_jump_return
);
847 case vtn_branch_type_discard
: {
848 nir_intrinsic_instr
*discard
=
849 nir_intrinsic_instr_create(b
->nb
.shader
, nir_intrinsic_discard
);
850 nir_builder_instr_insert(&b
->nb
, &discard
->instr
);
854 vtn_fail("Invalid branch type");
859 vtn_switch_case_condition(struct vtn_builder
*b
, struct vtn_switch
*swtch
,
860 nir_ssa_def
*sel
, struct vtn_case
*cse
)
862 if (cse
->is_default
) {
863 nir_ssa_def
*any
= nir_imm_false(&b
->nb
);
864 list_for_each_entry(struct vtn_case
, other
, &swtch
->cases
, link
) {
865 if (other
->is_default
)
868 any
= nir_ior(&b
->nb
, any
,
869 vtn_switch_case_condition(b
, swtch
, sel
, other
));
871 return nir_inot(&b
->nb
, any
);
873 nir_ssa_def
*cond
= nir_imm_false(&b
->nb
);
874 util_dynarray_foreach(&cse
->values
, uint64_t, val
) {
875 nir_ssa_def
*imm
= nir_imm_intN_t(&b
->nb
, *val
, sel
->bit_size
);
876 cond
= nir_ior(&b
->nb
, cond
, nir_ieq(&b
->nb
, sel
, imm
));
882 static nir_loop_control
883 vtn_loop_control(struct vtn_builder
*b
, struct vtn_loop
*vtn_loop
)
885 if (vtn_loop
->control
== SpvLoopControlMaskNone
)
886 return nir_loop_control_none
;
887 else if (vtn_loop
->control
& SpvLoopControlDontUnrollMask
)
888 return nir_loop_control_dont_unroll
;
889 else if (vtn_loop
->control
& SpvLoopControlUnrollMask
)
890 return nir_loop_control_unroll
;
891 else if (vtn_loop
->control
& SpvLoopControlDependencyInfiniteMask
||
892 vtn_loop
->control
& SpvLoopControlDependencyLengthMask
) {
893 /* We do not do anything special with these yet. */
894 return nir_loop_control_none
;
896 vtn_fail("Invalid loop control");
900 static nir_selection_control
901 vtn_selection_control(struct vtn_builder
*b
, struct vtn_if
*vtn_if
)
903 if (vtn_if
->control
== SpvSelectionControlMaskNone
)
904 return nir_selection_control_none
;
905 else if (vtn_if
->control
& SpvSelectionControlDontFlattenMask
)
906 return nir_selection_control_dont_flatten
;
907 else if (vtn_if
->control
& SpvSelectionControlFlattenMask
)
908 return nir_selection_control_flatten
;
910 vtn_fail("Invalid selection control");
914 vtn_emit_cf_list(struct vtn_builder
*b
, struct list_head
*cf_list
,
915 nir_variable
*switch_fall_var
, bool *has_switch_break
,
916 vtn_instruction_handler handler
)
918 list_for_each_entry(struct vtn_cf_node
, node
, cf_list
, link
) {
919 switch (node
->type
) {
920 case vtn_cf_node_type_block
: {
921 struct vtn_block
*block
= (struct vtn_block
*)node
;
923 const uint32_t *block_start
= block
->label
;
924 const uint32_t *block_end
= block
->merge
? block
->merge
:
927 block_start
= vtn_foreach_instruction(b
, block_start
, block_end
,
928 vtn_handle_phis_first_pass
);
930 vtn_foreach_instruction(b
, block_start
, block_end
, handler
);
932 block
->end_nop
= nir_intrinsic_instr_create(b
->nb
.shader
,
934 nir_builder_instr_insert(&b
->nb
, &block
->end_nop
->instr
);
936 if ((*block
->branch
& SpvOpCodeMask
) == SpvOpReturnValue
) {
937 vtn_fail_if(b
->func
->type
->return_type
->base_type
==
939 "Return with a value from a function returning void");
940 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, block
->branch
[1]);
941 const struct glsl_type
*ret_type
=
942 glsl_get_bare_type(b
->func
->type
->return_type
->type
);
943 nir_deref_instr
*ret_deref
=
944 nir_build_deref_cast(&b
->nb
, nir_load_param(&b
->nb
, 0),
945 nir_var_function_temp
, ret_type
, 0);
946 vtn_local_store(b
, src
, ret_deref
, 0);
949 if (block
->branch_type
!= vtn_branch_type_none
) {
950 vtn_emit_branch(b
, block
->branch_type
,
951 switch_fall_var
, has_switch_break
);
958 case vtn_cf_node_type_if
: {
959 struct vtn_if
*vtn_if
= (struct vtn_if
*)node
;
960 bool sw_break
= false;
963 nir_push_if(&b
->nb
, vtn_ssa_value(b
, vtn_if
->condition
)->def
);
965 nif
->control
= vtn_selection_control(b
, vtn_if
);
967 if (vtn_if
->then_type
== vtn_branch_type_none
) {
968 vtn_emit_cf_list(b
, &vtn_if
->then_body
,
969 switch_fall_var
, &sw_break
, handler
);
971 vtn_emit_branch(b
, vtn_if
->then_type
, switch_fall_var
, &sw_break
);
974 nir_push_else(&b
->nb
, nif
);
975 if (vtn_if
->else_type
== vtn_branch_type_none
) {
976 vtn_emit_cf_list(b
, &vtn_if
->else_body
,
977 switch_fall_var
, &sw_break
, handler
);
979 vtn_emit_branch(b
, vtn_if
->else_type
, switch_fall_var
, &sw_break
);
982 nir_pop_if(&b
->nb
, nif
);
984 /* If we encountered a switch break somewhere inside of the if,
985 * then it would have been handled correctly by calling
986 * emit_cf_list or emit_branch for the interrior. However, we
987 * need to predicate everything following on wether or not we're
991 *has_switch_break
= true;
992 nir_push_if(&b
->nb
, nir_load_var(&b
->nb
, switch_fall_var
));
997 case vtn_cf_node_type_loop
: {
998 struct vtn_loop
*vtn_loop
= (struct vtn_loop
*)node
;
1000 nir_loop
*loop
= nir_push_loop(&b
->nb
);
1001 loop
->control
= vtn_loop_control(b
, vtn_loop
);
1003 vtn_emit_cf_list(b
, &vtn_loop
->body
, NULL
, NULL
, handler
);
1005 if (!list_empty(&vtn_loop
->cont_body
)) {
1006 /* If we have a non-trivial continue body then we need to put
1007 * it at the beginning of the loop with a flag to ensure that
1008 * it doesn't get executed in the first iteration.
1010 nir_variable
*do_cont
=
1011 nir_local_variable_create(b
->nb
.impl
, glsl_bool_type(), "cont");
1013 b
->nb
.cursor
= nir_before_cf_node(&loop
->cf_node
);
1014 nir_store_var(&b
->nb
, do_cont
, nir_imm_false(&b
->nb
), 1);
1016 b
->nb
.cursor
= nir_before_cf_list(&loop
->body
);
1019 nir_push_if(&b
->nb
, nir_load_var(&b
->nb
, do_cont
));
1021 vtn_emit_cf_list(b
, &vtn_loop
->cont_body
, NULL
, NULL
, handler
);
1023 nir_pop_if(&b
->nb
, cont_if
);
1025 nir_store_var(&b
->nb
, do_cont
, nir_imm_true(&b
->nb
), 1);
1027 b
->has_loop_continue
= true;
1030 nir_pop_loop(&b
->nb
, loop
);
1034 case vtn_cf_node_type_switch
: {
1035 struct vtn_switch
*vtn_switch
= (struct vtn_switch
*)node
;
1037 /* First, we create a variable to keep track of whether or not the
1038 * switch is still going at any given point. Any switch breaks
1039 * will set this variable to false.
1041 nir_variable
*fall_var
=
1042 nir_local_variable_create(b
->nb
.impl
, glsl_bool_type(), "fall");
1043 nir_store_var(&b
->nb
, fall_var
, nir_imm_false(&b
->nb
), 1);
1045 nir_ssa_def
*sel
= vtn_ssa_value(b
, vtn_switch
->selector
)->def
;
1047 /* Now we can walk the list of cases and actually emit code */
1048 list_for_each_entry(struct vtn_case
, cse
, &vtn_switch
->cases
, link
) {
1049 /* Figure out the condition */
1051 vtn_switch_case_condition(b
, vtn_switch
, sel
, cse
);
1052 /* Take fallthrough into account */
1053 cond
= nir_ior(&b
->nb
, cond
, nir_load_var(&b
->nb
, fall_var
));
1055 nir_if
*case_if
= nir_push_if(&b
->nb
, cond
);
1057 bool has_break
= false;
1058 nir_store_var(&b
->nb
, fall_var
, nir_imm_true(&b
->nb
), 1);
1059 vtn_emit_cf_list(b
, &cse
->body
, fall_var
, &has_break
, handler
);
1060 (void)has_break
; /* We don't care */
1062 nir_pop_if(&b
->nb
, case_if
);
1069 vtn_fail("Invalid CF node type");
1075 vtn_function_emit(struct vtn_builder
*b
, struct vtn_function
*func
,
1076 vtn_instruction_handler instruction_handler
)
1078 nir_builder_init(&b
->nb
, func
->impl
);
1080 b
->nb
.cursor
= nir_after_cf_list(&func
->impl
->body
);
1081 b
->nb
.exact
= b
->exact
;
1082 b
->has_loop_continue
= false;
1083 b
->phi_table
= _mesa_pointer_hash_table_create(b
);
1085 vtn_emit_cf_list(b
, &func
->body
, NULL
, NULL
, instruction_handler
);
1087 vtn_foreach_instruction(b
, func
->start_block
->label
, func
->end
,
1088 vtn_handle_phi_second_pass
);
1090 nir_rematerialize_derefs_in_use_blocks_impl(func
->impl
);
1092 /* Continue blocks for loops get inserted before the body of the loop
1093 * but instructions in the continue may use SSA defs in the loop body.
1094 * Therefore, we need to repair SSA to insert the needed phi nodes.
1096 if (b
->has_loop_continue
)
1097 nir_repair_ssa_impl(func
->impl
);
1099 func
->emitted
= true;