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 list_inithead(&b
->func
->body
);
255 b
->func
->control
= w
[3];
257 UNUSED
const struct glsl_type
*result_type
=
258 vtn_value(b
, w
[1], vtn_value_type_type
)->type
->type
;
259 struct vtn_value
*val
= vtn_push_value(b
, w
[2], vtn_value_type_function
);
262 b
->func
->type
= vtn_value(b
, w
[4], vtn_value_type_type
)->type
;
263 const struct vtn_type
*func_type
= b
->func
->type
;
265 vtn_assert(func_type
->return_type
->type
== result_type
);
268 nir_function_create(b
->shader
, ralloc_strdup(b
->shader
, val
->name
));
270 unsigned num_params
= 0;
271 for (unsigned i
= 0; i
< func_type
->length
; i
++)
272 num_params
+= vtn_type_count_function_params(func_type
->params
[i
]);
274 /* Add one parameter for the function return value */
275 if (func_type
->return_type
->base_type
!= vtn_base_type_void
)
278 func
->num_params
= num_params
;
279 func
->params
= ralloc_array(b
->shader
, nir_parameter
, num_params
);
282 if (func_type
->return_type
->base_type
!= vtn_base_type_void
) {
283 nir_address_format addr_format
=
284 vtn_mode_to_address_format(b
, vtn_variable_mode_function
);
285 /* The return value is a regular pointer */
286 func
->params
[idx
++] = (nir_parameter
) {
287 .num_components
= nir_address_format_num_components(addr_format
),
288 .bit_size
= nir_address_format_bit_size(addr_format
),
292 for (unsigned i
= 0; i
< func_type
->length
; i
++)
293 vtn_type_add_to_function_params(func_type
->params
[i
], func
, &idx
);
294 assert(idx
== num_params
);
296 b
->func
->impl
= nir_function_impl_create(func
);
297 nir_builder_init(&b
->nb
, func
->impl
);
298 b
->nb
.cursor
= nir_before_cf_list(&b
->func
->impl
->body
);
299 b
->nb
.exact
= b
->exact
;
301 b
->func_param_idx
= 0;
303 /* The return value is the first parameter */
304 if (func_type
->return_type
->base_type
!= vtn_base_type_void
)
309 case SpvOpFunctionEnd
:
314 case SpvOpFunctionParameter
: {
315 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
317 vtn_assert(b
->func_param_idx
< b
->func
->impl
->function
->num_params
);
319 if (type
->base_type
== vtn_base_type_sampled_image
) {
320 /* Sampled images are actually two parameters. The first is the
321 * image and the second is the sampler.
323 struct vtn_value
*val
=
324 vtn_push_value(b
, w
[2], vtn_value_type_sampled_image
);
326 val
->sampled_image
= ralloc(b
, struct vtn_sampled_image
);
328 struct vtn_type
*sampler_type
= rzalloc(b
, struct vtn_type
);
329 sampler_type
->base_type
= vtn_base_type_sampler
;
330 sampler_type
->type
= glsl_bare_sampler_type();
332 val
->sampled_image
->image
=
333 vtn_load_param_pointer(b
, type
, b
->func_param_idx
++);
334 val
->sampled_image
->sampler
=
335 vtn_load_param_pointer(b
, sampler_type
, b
->func_param_idx
++);
336 } else if (type
->base_type
== vtn_base_type_pointer
&&
337 type
->type
!= NULL
) {
338 /* This is a pointer with an actual storage type */
339 nir_ssa_def
*ssa_ptr
= nir_load_param(&b
->nb
, b
->func_param_idx
++);
340 vtn_push_value_pointer(b
, w
[2], vtn_pointer_from_ssa(b
, ssa_ptr
, type
));
341 } else if (type
->base_type
== vtn_base_type_pointer
||
342 type
->base_type
== vtn_base_type_image
||
343 type
->base_type
== vtn_base_type_sampler
) {
344 vtn_push_value_pointer(b
, w
[2], vtn_load_param_pointer(b
, type
, b
->func_param_idx
++));
346 /* We're a regular SSA value. */
347 struct vtn_ssa_value
*value
= vtn_create_ssa_value(b
, type
->type
);
348 vtn_ssa_value_load_function_param(b
, value
, type
, &b
->func_param_idx
);
349 vtn_push_ssa(b
, w
[2], type
, value
);
355 vtn_assert(b
->block
== NULL
);
356 b
->block
= rzalloc(b
, struct vtn_block
);
357 b
->block
->node
.type
= vtn_cf_node_type_block
;
359 vtn_push_value(b
, w
[1], vtn_value_type_block
)->block
= b
->block
;
361 if (b
->func
->start_block
== NULL
) {
362 /* This is the first block encountered for this function. In this
363 * case, we set the start block and add it to the list of
364 * implemented functions that we'll walk later.
366 b
->func
->start_block
= b
->block
;
367 exec_list_push_tail(&b
->functions
, &b
->func
->node
);
372 case SpvOpSelectionMerge
:
374 vtn_assert(b
->block
&& b
->block
->merge
== NULL
);
379 case SpvOpBranchConditional
:
383 case SpvOpReturnValue
:
384 case SpvOpUnreachable
:
385 vtn_assert(b
->block
&& b
->block
->branch
== NULL
);
386 b
->block
->branch
= w
;
391 /* Continue on as per normal */
399 vtn_add_case(struct vtn_builder
*b
, struct vtn_switch
*swtch
,
400 struct vtn_block
*break_block
,
401 uint32_t block_id
, uint64_t val
, bool is_default
)
403 struct vtn_block
*case_block
= vtn_block(b
, block_id
);
405 /* Don't create dummy cases that just break */
406 if (case_block
== break_block
)
409 if (case_block
->switch_case
== NULL
) {
410 struct vtn_case
*c
= ralloc(b
, struct vtn_case
);
412 list_inithead(&c
->body
);
413 c
->start_block
= case_block
;
414 c
->fallthrough
= NULL
;
415 util_dynarray_init(&c
->values
, b
);
416 c
->is_default
= false;
419 list_addtail(&c
->link
, &swtch
->cases
);
421 case_block
->switch_case
= c
;
425 case_block
->switch_case
->is_default
= true;
427 util_dynarray_append(&case_block
->switch_case
->values
, uint64_t, val
);
431 /* This function performs a depth-first search of the cases and puts them
432 * in fall-through order.
435 vtn_order_case(struct vtn_switch
*swtch
, struct vtn_case
*cse
)
442 list_del(&cse
->link
);
444 if (cse
->fallthrough
) {
445 vtn_order_case(swtch
, cse
->fallthrough
);
447 /* If we have a fall-through, place this case right before the case it
448 * falls through to. This ensures that fallthroughs come one after
449 * the other. These two can never get separated because that would
450 * imply something else falling through to the same case. Also, this
451 * can't break ordering because the DFS ensures that this case is
452 * visited before anything that falls through to it.
454 list_addtail(&cse
->link
, &cse
->fallthrough
->link
);
456 list_add(&cse
->link
, &swtch
->cases
);
460 static enum vtn_branch_type
461 vtn_get_branch_type(struct vtn_builder
*b
,
462 struct vtn_block
*block
,
463 struct vtn_case
*swcase
, struct vtn_block
*switch_break
,
464 struct vtn_block
*loop_break
, struct vtn_block
*loop_cont
)
466 if (block
->switch_case
) {
467 /* This branch is actually a fallthrough */
468 vtn_assert(swcase
->fallthrough
== NULL
||
469 swcase
->fallthrough
== block
->switch_case
);
470 swcase
->fallthrough
= block
->switch_case
;
471 return vtn_branch_type_switch_fallthrough
;
472 } else if (block
== loop_break
) {
473 return vtn_branch_type_loop_break
;
474 } else if (block
== loop_cont
) {
475 return vtn_branch_type_loop_continue
;
476 } else if (block
== switch_break
) {
477 return vtn_branch_type_switch_break
;
479 return vtn_branch_type_none
;
484 vtn_cfg_walk_blocks(struct vtn_builder
*b
, struct list_head
*cf_list
,
485 struct vtn_block
*start
, struct vtn_case
*switch_case
,
486 struct vtn_block
*switch_break
,
487 struct vtn_block
*loop_break
, struct vtn_block
*loop_cont
,
488 struct vtn_block
*end
)
490 struct vtn_block
*block
= start
;
491 while (block
!= end
) {
492 if (block
->merge
&& (*block
->merge
& SpvOpCodeMask
) == SpvOpLoopMerge
&&
494 struct vtn_loop
*loop
= ralloc(b
, struct vtn_loop
);
496 loop
->node
.type
= vtn_cf_node_type_loop
;
497 list_inithead(&loop
->body
);
498 list_inithead(&loop
->cont_body
);
499 loop
->control
= block
->merge
[3];
501 list_addtail(&loop
->node
.link
, cf_list
);
504 struct vtn_block
*new_loop_break
= vtn_block(b
, block
->merge
[1]);
505 struct vtn_block
*new_loop_cont
= vtn_block(b
, block
->merge
[2]);
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
= vtn_block(b
, block
->branch
[1]);
550 block
->branch_type
= vtn_get_branch_type(b
, branch_block
,
551 switch_case
, switch_break
,
552 loop_break
, loop_cont
);
554 if (block
->branch_type
!= vtn_branch_type_none
)
557 block
= branch_block
;
562 case SpvOpReturnValue
:
563 block
->branch_type
= vtn_branch_type_return
;
567 block
->branch_type
= vtn_branch_type_discard
;
570 case SpvOpBranchConditional
: {
571 struct vtn_block
*then_block
= vtn_block(b
, block
->branch
[2]);
572 struct vtn_block
*else_block
= vtn_block(b
, block
->branch
[3]);
574 struct vtn_if
*if_stmt
= ralloc(b
, struct vtn_if
);
576 if_stmt
->node
.type
= vtn_cf_node_type_if
;
577 if_stmt
->condition
= block
->branch
[1];
578 list_inithead(&if_stmt
->then_body
);
579 list_inithead(&if_stmt
->else_body
);
581 list_addtail(&if_stmt
->node
.link
, cf_list
);
584 (*block
->merge
& SpvOpCodeMask
) == SpvOpSelectionMerge
) {
585 if_stmt
->control
= block
->merge
[2];
587 if_stmt
->control
= SpvSelectionControlMaskNone
;
590 if_stmt
->then_type
= vtn_get_branch_type(b
, then_block
,
591 switch_case
, switch_break
,
592 loop_break
, loop_cont
);
593 if_stmt
->else_type
= vtn_get_branch_type(b
, else_block
,
594 switch_case
, switch_break
,
595 loop_break
, loop_cont
);
597 if (then_block
== else_block
) {
598 block
->branch_type
= if_stmt
->then_type
;
599 if (block
->branch_type
== vtn_branch_type_none
) {
605 } else if (if_stmt
->then_type
== vtn_branch_type_none
&&
606 if_stmt
->else_type
== vtn_branch_type_none
) {
607 /* Neither side of the if is something we can short-circuit. */
608 vtn_assert((*block
->merge
& SpvOpCodeMask
) == SpvOpSelectionMerge
);
609 struct vtn_block
*merge_block
= vtn_block(b
, block
->merge
[1]);
611 vtn_cfg_walk_blocks(b
, &if_stmt
->then_body
, then_block
,
612 switch_case
, switch_break
,
613 loop_break
, loop_cont
, merge_block
);
614 vtn_cfg_walk_blocks(b
, &if_stmt
->else_body
, else_block
,
615 switch_case
, switch_break
,
616 loop_break
, loop_cont
, merge_block
);
618 enum vtn_branch_type merge_type
=
619 vtn_get_branch_type(b
, merge_block
, switch_case
, switch_break
,
620 loop_break
, loop_cont
);
621 if (merge_type
== vtn_branch_type_none
) {
627 } else if (if_stmt
->then_type
!= vtn_branch_type_none
&&
628 if_stmt
->else_type
!= vtn_branch_type_none
) {
629 /* Both sides were short-circuited. We're done here. */
632 /* Exeactly one side of the branch could be short-circuited.
633 * We set the branch up as a predicated break/continue and we
634 * continue on with the other side as if it were what comes
637 if (if_stmt
->then_type
== vtn_branch_type_none
) {
644 vtn_fail("Should have returned or continued");
648 vtn_assert((*block
->merge
& SpvOpCodeMask
) == SpvOpSelectionMerge
);
649 struct vtn_block
*break_block
= vtn_block(b
, block
->merge
[1]);
651 struct vtn_switch
*swtch
= ralloc(b
, struct vtn_switch
);
653 swtch
->node
.type
= vtn_cf_node_type_switch
;
654 swtch
->selector
= block
->branch
[1];
655 list_inithead(&swtch
->cases
);
657 list_addtail(&swtch
->node
.link
, cf_list
);
659 /* First, we go through and record all of the cases. */
660 const uint32_t *branch_end
=
661 block
->branch
+ (block
->branch
[0] >> SpvWordCountShift
);
663 struct vtn_value
*cond_val
= vtn_untyped_value(b
, block
->branch
[1]);
664 vtn_fail_if(!cond_val
->type
||
665 cond_val
->type
->base_type
!= vtn_base_type_scalar
,
666 "Selector of OpSelect must have a type of OpTypeInt");
668 nir_alu_type cond_type
=
669 nir_get_nir_type_for_glsl_type(cond_val
->type
->type
);
670 vtn_fail_if(nir_alu_type_get_base_type(cond_type
) != nir_type_int
&&
671 nir_alu_type_get_base_type(cond_type
) != nir_type_uint
,
672 "Selector of OpSelect must have a type of OpTypeInt");
674 bool is_default
= true;
675 const unsigned bitsize
= nir_alu_type_get_type_size(cond_type
);
676 for (const uint32_t *w
= block
->branch
+ 2; w
< branch_end
;) {
677 uint64_t literal
= 0;
682 assert(bitsize
== 64);
683 literal
= vtn_u64_literal(w
);
688 uint32_t block_id
= *(w
++);
690 vtn_add_case(b
, swtch
, break_block
, block_id
, literal
, is_default
);
694 /* Now, we go through and walk the blocks. While we walk through
695 * the blocks, we also gather the much-needed fall-through
698 list_for_each_entry(struct vtn_case
, cse
, &swtch
->cases
, link
) {
699 vtn_assert(cse
->start_block
!= break_block
);
700 vtn_cfg_walk_blocks(b
, &cse
->body
, cse
->start_block
, cse
,
701 break_block
, loop_break
, loop_cont
, NULL
);
704 /* Finally, we walk over all of the cases one more time and put
705 * them in fall-through order.
707 for (const uint32_t *w
= block
->branch
+ 2; w
< branch_end
;) {
708 struct vtn_block
*case_block
= vtn_block(b
, *w
);
713 assert(bitsize
== 64);
717 if (case_block
== break_block
)
720 vtn_assert(case_block
->switch_case
);
722 vtn_order_case(swtch
, case_block
->switch_case
);
725 enum vtn_branch_type branch_type
=
726 vtn_get_branch_type(b
, break_block
, switch_case
, NULL
,
727 loop_break
, loop_cont
);
729 if (branch_type
!= vtn_branch_type_none
) {
730 /* It is possible that the break is actually the continue block
731 * for the containing loop. In this case, we need to bail and let
732 * the loop parsing code handle the continue properly.
734 vtn_assert(branch_type
== vtn_branch_type_loop_continue
);
742 case SpvOpUnreachable
:
746 vtn_fail("Unhandled opcode");
752 vtn_build_cfg(struct vtn_builder
*b
, const uint32_t *words
, const uint32_t *end
)
754 vtn_foreach_instruction(b
, words
, end
,
755 vtn_cfg_handle_prepass_instruction
);
757 foreach_list_typed(struct vtn_function
, func
, node
, &b
->functions
) {
758 vtn_cfg_walk_blocks(b
, &func
->body
, func
->start_block
,
759 NULL
, NULL
, NULL
, NULL
, NULL
);
764 vtn_handle_phis_first_pass(struct vtn_builder
*b
, SpvOp opcode
,
765 const uint32_t *w
, unsigned count
)
767 if (opcode
== SpvOpLabel
)
768 return true; /* Nothing to do */
770 /* If this isn't a phi node, stop. */
771 if (opcode
!= SpvOpPhi
)
774 /* For handling phi nodes, we do a poor-man's out-of-ssa on the spot.
775 * For each phi, we create a variable with the appropreate type and
776 * do a load from that variable. Then, in a second pass, we add
777 * stores to that variable to each of the predecessor blocks.
779 * We could do something more intelligent here. However, in order to
780 * handle loops and things properly, we really need dominance
781 * information. It would end up basically being the into-SSA
782 * algorithm all over again. It's easier if we just let
783 * lower_vars_to_ssa do that for us instead of repeating it here.
785 struct vtn_type
*type
= vtn_value(b
, w
[1], vtn_value_type_type
)->type
;
786 nir_variable
*phi_var
=
787 nir_local_variable_create(b
->nb
.impl
, type
->type
, "phi");
788 _mesa_hash_table_insert(b
->phi_table
, w
, phi_var
);
790 vtn_push_ssa(b
, w
[2], type
,
791 vtn_local_load(b
, nir_build_deref_var(&b
->nb
, phi_var
), 0));
797 vtn_handle_phi_second_pass(struct vtn_builder
*b
, SpvOp opcode
,
798 const uint32_t *w
, unsigned count
)
800 if (opcode
!= SpvOpPhi
)
803 struct hash_entry
*phi_entry
= _mesa_hash_table_search(b
->phi_table
, w
);
804 vtn_assert(phi_entry
);
805 nir_variable
*phi_var
= phi_entry
->data
;
807 for (unsigned i
= 3; i
< count
; i
+= 2) {
808 struct vtn_block
*pred
= vtn_block(b
, w
[i
+ 1]);
810 /* If block does not have end_nop, that is because it is an unreacheable
811 * block, and hence it is not worth to handle it */
815 b
->nb
.cursor
= nir_after_instr(&pred
->end_nop
->instr
);
817 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, w
[i
]);
819 vtn_local_store(b
, src
, nir_build_deref_var(&b
->nb
, phi_var
), 0);
826 vtn_emit_branch(struct vtn_builder
*b
, enum vtn_branch_type branch_type
,
827 nir_variable
*switch_fall_var
, bool *has_switch_break
)
829 switch (branch_type
) {
830 case vtn_branch_type_switch_break
:
831 nir_store_var(&b
->nb
, switch_fall_var
, nir_imm_false(&b
->nb
), 1);
832 *has_switch_break
= true;
834 case vtn_branch_type_switch_fallthrough
:
835 break; /* Nothing to do */
836 case vtn_branch_type_loop_break
:
837 nir_jump(&b
->nb
, nir_jump_break
);
839 case vtn_branch_type_loop_continue
:
840 nir_jump(&b
->nb
, nir_jump_continue
);
842 case vtn_branch_type_return
:
843 nir_jump(&b
->nb
, nir_jump_return
);
845 case vtn_branch_type_discard
: {
846 nir_intrinsic_instr
*discard
=
847 nir_intrinsic_instr_create(b
->nb
.shader
, nir_intrinsic_discard
);
848 nir_builder_instr_insert(&b
->nb
, &discard
->instr
);
852 vtn_fail("Invalid branch type");
857 vtn_switch_case_condition(struct vtn_builder
*b
, struct vtn_switch
*swtch
,
858 nir_ssa_def
*sel
, struct vtn_case
*cse
)
860 if (cse
->is_default
) {
861 nir_ssa_def
*any
= nir_imm_false(&b
->nb
);
862 list_for_each_entry(struct vtn_case
, other
, &swtch
->cases
, link
) {
863 if (other
->is_default
)
866 any
= nir_ior(&b
->nb
, any
,
867 vtn_switch_case_condition(b
, swtch
, sel
, other
));
869 return nir_inot(&b
->nb
, any
);
871 nir_ssa_def
*cond
= nir_imm_false(&b
->nb
);
872 util_dynarray_foreach(&cse
->values
, uint64_t, val
) {
873 nir_ssa_def
*imm
= nir_imm_intN_t(&b
->nb
, *val
, sel
->bit_size
);
874 cond
= nir_ior(&b
->nb
, cond
, nir_ieq(&b
->nb
, sel
, imm
));
880 static nir_loop_control
881 vtn_loop_control(struct vtn_builder
*b
, struct vtn_loop
*vtn_loop
)
883 if (vtn_loop
->control
== SpvLoopControlMaskNone
)
884 return nir_loop_control_none
;
885 else if (vtn_loop
->control
& SpvLoopControlDontUnrollMask
)
886 return nir_loop_control_dont_unroll
;
887 else if (vtn_loop
->control
& SpvLoopControlUnrollMask
)
888 return nir_loop_control_unroll
;
889 else if (vtn_loop
->control
& SpvLoopControlDependencyInfiniteMask
||
890 vtn_loop
->control
& SpvLoopControlDependencyLengthMask
||
891 vtn_loop
->control
& SpvLoopControlMinIterationsMask
||
892 vtn_loop
->control
& SpvLoopControlMaxIterationsMask
||
893 vtn_loop
->control
& SpvLoopControlIterationMultipleMask
||
894 vtn_loop
->control
& SpvLoopControlPeelCountMask
||
895 vtn_loop
->control
& SpvLoopControlPartialCountMask
) {
896 /* We do not do anything special with these yet. */
897 return nir_loop_control_none
;
899 vtn_fail("Invalid loop control");
903 static nir_selection_control
904 vtn_selection_control(struct vtn_builder
*b
, struct vtn_if
*vtn_if
)
906 if (vtn_if
->control
== SpvSelectionControlMaskNone
)
907 return nir_selection_control_none
;
908 else if (vtn_if
->control
& SpvSelectionControlDontFlattenMask
)
909 return nir_selection_control_dont_flatten
;
910 else if (vtn_if
->control
& SpvSelectionControlFlattenMask
)
911 return nir_selection_control_flatten
;
913 vtn_fail("Invalid selection control");
917 vtn_emit_cf_list(struct vtn_builder
*b
, struct list_head
*cf_list
,
918 nir_variable
*switch_fall_var
, bool *has_switch_break
,
919 vtn_instruction_handler handler
)
921 list_for_each_entry(struct vtn_cf_node
, node
, cf_list
, link
) {
922 switch (node
->type
) {
923 case vtn_cf_node_type_block
: {
924 struct vtn_block
*block
= (struct vtn_block
*)node
;
926 const uint32_t *block_start
= block
->label
;
927 const uint32_t *block_end
= block
->merge
? block
->merge
:
930 block_start
= vtn_foreach_instruction(b
, block_start
, block_end
,
931 vtn_handle_phis_first_pass
);
933 vtn_foreach_instruction(b
, block_start
, block_end
, handler
);
935 block
->end_nop
= nir_intrinsic_instr_create(b
->nb
.shader
,
937 nir_builder_instr_insert(&b
->nb
, &block
->end_nop
->instr
);
939 if ((*block
->branch
& SpvOpCodeMask
) == SpvOpReturnValue
) {
940 vtn_fail_if(b
->func
->type
->return_type
->base_type
==
942 "Return with a value from a function returning void");
943 struct vtn_ssa_value
*src
= vtn_ssa_value(b
, block
->branch
[1]);
944 const struct glsl_type
*ret_type
=
945 glsl_get_bare_type(b
->func
->type
->return_type
->type
);
946 nir_deref_instr
*ret_deref
=
947 nir_build_deref_cast(&b
->nb
, nir_load_param(&b
->nb
, 0),
948 nir_var_function_temp
, ret_type
, 0);
949 vtn_local_store(b
, src
, ret_deref
, 0);
952 if (block
->branch_type
!= vtn_branch_type_none
) {
953 vtn_emit_branch(b
, block
->branch_type
,
954 switch_fall_var
, has_switch_break
);
961 case vtn_cf_node_type_if
: {
962 struct vtn_if
*vtn_if
= (struct vtn_if
*)node
;
963 bool sw_break
= false;
966 nir_push_if(&b
->nb
, vtn_ssa_value(b
, vtn_if
->condition
)->def
);
968 nif
->control
= vtn_selection_control(b
, vtn_if
);
970 if (vtn_if
->then_type
== vtn_branch_type_none
) {
971 vtn_emit_cf_list(b
, &vtn_if
->then_body
,
972 switch_fall_var
, &sw_break
, handler
);
974 vtn_emit_branch(b
, vtn_if
->then_type
, switch_fall_var
, &sw_break
);
977 nir_push_else(&b
->nb
, nif
);
978 if (vtn_if
->else_type
== vtn_branch_type_none
) {
979 vtn_emit_cf_list(b
, &vtn_if
->else_body
,
980 switch_fall_var
, &sw_break
, handler
);
982 vtn_emit_branch(b
, vtn_if
->else_type
, switch_fall_var
, &sw_break
);
985 nir_pop_if(&b
->nb
, nif
);
987 /* If we encountered a switch break somewhere inside of the if,
988 * then it would have been handled correctly by calling
989 * emit_cf_list or emit_branch for the interrior. However, we
990 * need to predicate everything following on wether or not we're
994 *has_switch_break
= true;
995 nir_push_if(&b
->nb
, nir_load_var(&b
->nb
, switch_fall_var
));
1000 case vtn_cf_node_type_loop
: {
1001 struct vtn_loop
*vtn_loop
= (struct vtn_loop
*)node
;
1003 nir_loop
*loop
= nir_push_loop(&b
->nb
);
1004 loop
->control
= vtn_loop_control(b
, vtn_loop
);
1006 vtn_emit_cf_list(b
, &vtn_loop
->body
, NULL
, NULL
, handler
);
1008 if (!list_is_empty(&vtn_loop
->cont_body
)) {
1009 /* If we have a non-trivial continue body then we need to put
1010 * it at the beginning of the loop with a flag to ensure that
1011 * it doesn't get executed in the first iteration.
1013 nir_variable
*do_cont
=
1014 nir_local_variable_create(b
->nb
.impl
, glsl_bool_type(), "cont");
1016 b
->nb
.cursor
= nir_before_cf_node(&loop
->cf_node
);
1017 nir_store_var(&b
->nb
, do_cont
, nir_imm_false(&b
->nb
), 1);
1019 b
->nb
.cursor
= nir_before_cf_list(&loop
->body
);
1022 nir_push_if(&b
->nb
, nir_load_var(&b
->nb
, do_cont
));
1024 vtn_emit_cf_list(b
, &vtn_loop
->cont_body
, NULL
, NULL
, handler
);
1026 nir_pop_if(&b
->nb
, cont_if
);
1028 nir_store_var(&b
->nb
, do_cont
, nir_imm_true(&b
->nb
), 1);
1030 b
->has_loop_continue
= true;
1033 nir_pop_loop(&b
->nb
, loop
);
1037 case vtn_cf_node_type_switch
: {
1038 struct vtn_switch
*vtn_switch
= (struct vtn_switch
*)node
;
1040 /* First, we create a variable to keep track of whether or not the
1041 * switch is still going at any given point. Any switch breaks
1042 * will set this variable to false.
1044 nir_variable
*fall_var
=
1045 nir_local_variable_create(b
->nb
.impl
, glsl_bool_type(), "fall");
1046 nir_store_var(&b
->nb
, fall_var
, nir_imm_false(&b
->nb
), 1);
1048 nir_ssa_def
*sel
= vtn_ssa_value(b
, vtn_switch
->selector
)->def
;
1050 /* Now we can walk the list of cases and actually emit code */
1051 list_for_each_entry(struct vtn_case
, cse
, &vtn_switch
->cases
, link
) {
1052 /* Figure out the condition */
1054 vtn_switch_case_condition(b
, vtn_switch
, sel
, cse
);
1055 /* Take fallthrough into account */
1056 cond
= nir_ior(&b
->nb
, cond
, nir_load_var(&b
->nb
, fall_var
));
1058 nir_if
*case_if
= nir_push_if(&b
->nb
, cond
);
1060 bool has_break
= false;
1061 nir_store_var(&b
->nb
, fall_var
, nir_imm_true(&b
->nb
), 1);
1062 vtn_emit_cf_list(b
, &cse
->body
, fall_var
, &has_break
, handler
);
1063 (void)has_break
; /* We don't care */
1065 nir_pop_if(&b
->nb
, case_if
);
1072 vtn_fail("Invalid CF node type");
1078 vtn_function_emit(struct vtn_builder
*b
, struct vtn_function
*func
,
1079 vtn_instruction_handler instruction_handler
)
1081 nir_builder_init(&b
->nb
, func
->impl
);
1083 b
->nb
.cursor
= nir_after_cf_list(&func
->impl
->body
);
1084 b
->nb
.exact
= b
->exact
;
1085 b
->has_loop_continue
= false;
1086 b
->phi_table
= _mesa_pointer_hash_table_create(b
);
1088 vtn_emit_cf_list(b
, &func
->body
, NULL
, NULL
, instruction_handler
);
1090 vtn_foreach_instruction(b
, func
->start_block
->label
, func
->end
,
1091 vtn_handle_phi_second_pass
);
1093 nir_rematerialize_derefs_in_use_blocks_impl(func
->impl
);
1095 /* Continue blocks for loops get inserted before the body of the loop
1096 * but instructions in the continue may use SSA defs in the loop body.
1097 * Therefore, we need to repair SSA to insert the needed phi nodes.
1099 if (b
->has_loop_continue
)
1100 nir_repair_ssa_impl(func
->impl
);
1102 func
->emitted
= true;