2 * Copyright © 2014 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 * Connor Abbott (cwabbott0@gmail.com)
29 #include "nir_builder.h"
30 #include "nir_control_flow_private.h"
31 #include "util/half_float.h"
35 #include "util/u_math.h"
37 #include "main/menums.h" /* BITFIELD64_MASK */
40 nir_shader_create(void *mem_ctx
,
41 gl_shader_stage stage
,
42 const nir_shader_compiler_options
*options
,
45 nir_shader
*shader
= rzalloc(mem_ctx
, nir_shader
);
47 exec_list_make_empty(&shader
->uniforms
);
48 exec_list_make_empty(&shader
->inputs
);
49 exec_list_make_empty(&shader
->outputs
);
50 exec_list_make_empty(&shader
->shared
);
52 shader
->options
= options
;
55 assert(si
->stage
== stage
);
58 shader
->info
.stage
= stage
;
61 exec_list_make_empty(&shader
->functions
);
62 exec_list_make_empty(&shader
->globals
);
63 exec_list_make_empty(&shader
->system_values
);
65 shader
->num_inputs
= 0;
66 shader
->num_outputs
= 0;
67 shader
->num_uniforms
= 0;
68 shader
->num_shared
= 0;
74 reg_create(void *mem_ctx
, struct exec_list
*list
)
76 nir_register
*reg
= ralloc(mem_ctx
, nir_register
);
78 list_inithead(®
->uses
);
79 list_inithead(®
->defs
);
80 list_inithead(®
->if_uses
);
82 reg
->num_components
= 0;
84 reg
->num_array_elems
= 0;
87 exec_list_push_tail(list
, ®
->node
);
93 nir_local_reg_create(nir_function_impl
*impl
)
95 nir_register
*reg
= reg_create(ralloc_parent(impl
), &impl
->registers
);
96 reg
->index
= impl
->reg_alloc
++;
102 nir_reg_remove(nir_register
*reg
)
104 exec_node_remove(®
->node
);
108 nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
)
110 switch (var
->data
.mode
) {
111 case nir_num_variable_modes
:
113 assert(!"invalid mode");
116 case nir_var_function_temp
:
117 assert(!"nir_shader_add_variable cannot be used for local variables");
120 case nir_var_shader_temp
:
121 exec_list_push_tail(&shader
->globals
, &var
->node
);
124 case nir_var_shader_in
:
125 exec_list_push_tail(&shader
->inputs
, &var
->node
);
128 case nir_var_shader_out
:
129 exec_list_push_tail(&shader
->outputs
, &var
->node
);
132 case nir_var_uniform
:
133 case nir_var_mem_ubo
:
134 case nir_var_mem_ssbo
:
135 exec_list_push_tail(&shader
->uniforms
, &var
->node
);
138 case nir_var_mem_shared
:
139 assert(gl_shader_stage_is_compute(shader
->info
.stage
));
140 exec_list_push_tail(&shader
->shared
, &var
->node
);
143 case nir_var_mem_global
:
144 assert(!"nir_shader_add_variable cannot be used for global memory");
147 case nir_var_system_value
:
148 exec_list_push_tail(&shader
->system_values
, &var
->node
);
151 case nir_var_mem_push_const
:
152 assert(!"nir_var_push_constant is not supposed to be used for variables");
158 nir_variable_create(nir_shader
*shader
, nir_variable_mode mode
,
159 const struct glsl_type
*type
, const char *name
)
161 nir_variable
*var
= rzalloc(shader
, nir_variable
);
162 var
->name
= ralloc_strdup(var
, name
);
164 var
->data
.mode
= mode
;
165 var
->data
.how_declared
= nir_var_declared_normally
;
167 if ((mode
== nir_var_shader_in
&&
168 shader
->info
.stage
!= MESA_SHADER_VERTEX
) ||
169 (mode
== nir_var_shader_out
&&
170 shader
->info
.stage
!= MESA_SHADER_FRAGMENT
))
171 var
->data
.interpolation
= INTERP_MODE_SMOOTH
;
173 if (mode
== nir_var_shader_in
|| mode
== nir_var_uniform
)
174 var
->data
.read_only
= true;
176 nir_shader_add_variable(shader
, var
);
182 nir_local_variable_create(nir_function_impl
*impl
,
183 const struct glsl_type
*type
, const char *name
)
185 nir_variable
*var
= rzalloc(impl
->function
->shader
, nir_variable
);
186 var
->name
= ralloc_strdup(var
, name
);
188 var
->data
.mode
= nir_var_function_temp
;
190 nir_function_impl_add_variable(impl
, var
);
196 nir_function_create(nir_shader
*shader
, const char *name
)
198 nir_function
*func
= ralloc(shader
, nir_function
);
200 exec_list_push_tail(&shader
->functions
, &func
->node
);
202 func
->name
= ralloc_strdup(func
, name
);
203 func
->shader
= shader
;
204 func
->num_params
= 0;
207 func
->is_entrypoint
= false;
212 /* NOTE: if the instruction you are copying a src to is already added
213 * to the IR, use nir_instr_rewrite_src() instead.
215 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *mem_ctx
)
217 dest
->is_ssa
= src
->is_ssa
;
219 dest
->ssa
= src
->ssa
;
221 dest
->reg
.base_offset
= src
->reg
.base_offset
;
222 dest
->reg
.reg
= src
->reg
.reg
;
223 if (src
->reg
.indirect
) {
224 dest
->reg
.indirect
= ralloc(mem_ctx
, nir_src
);
225 nir_src_copy(dest
->reg
.indirect
, src
->reg
.indirect
, mem_ctx
);
227 dest
->reg
.indirect
= NULL
;
232 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
)
234 /* Copying an SSA definition makes no sense whatsoever. */
235 assert(!src
->is_ssa
);
237 dest
->is_ssa
= false;
239 dest
->reg
.base_offset
= src
->reg
.base_offset
;
240 dest
->reg
.reg
= src
->reg
.reg
;
241 if (src
->reg
.indirect
) {
242 dest
->reg
.indirect
= ralloc(instr
, nir_src
);
243 nir_src_copy(dest
->reg
.indirect
, src
->reg
.indirect
, instr
);
245 dest
->reg
.indirect
= NULL
;
250 nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
251 nir_alu_instr
*instr
)
253 nir_src_copy(&dest
->src
, &src
->src
, &instr
->instr
);
254 dest
->abs
= src
->abs
;
255 dest
->negate
= src
->negate
;
256 for (unsigned i
= 0; i
< NIR_MAX_VEC_COMPONENTS
; i
++)
257 dest
->swizzle
[i
] = src
->swizzle
[i
];
261 nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
262 nir_alu_instr
*instr
)
264 nir_dest_copy(&dest
->dest
, &src
->dest
, &instr
->instr
);
265 dest
->write_mask
= src
->write_mask
;
266 dest
->saturate
= src
->saturate
;
271 cf_init(nir_cf_node
*node
, nir_cf_node_type type
)
273 exec_node_init(&node
->node
);
279 nir_function_impl_create_bare(nir_shader
*shader
)
281 nir_function_impl
*impl
= ralloc(shader
, nir_function_impl
);
283 impl
->function
= NULL
;
285 cf_init(&impl
->cf_node
, nir_cf_node_function
);
287 exec_list_make_empty(&impl
->body
);
288 exec_list_make_empty(&impl
->registers
);
289 exec_list_make_empty(&impl
->locals
);
292 impl
->valid_metadata
= nir_metadata_none
;
294 /* create start & end blocks */
295 nir_block
*start_block
= nir_block_create(shader
);
296 nir_block
*end_block
= nir_block_create(shader
);
297 start_block
->cf_node
.parent
= &impl
->cf_node
;
298 end_block
->cf_node
.parent
= &impl
->cf_node
;
299 impl
->end_block
= end_block
;
301 exec_list_push_tail(&impl
->body
, &start_block
->cf_node
.node
);
303 start_block
->successors
[0] = end_block
;
304 _mesa_set_add(end_block
->predecessors
, start_block
);
309 nir_function_impl_create(nir_function
*function
)
311 assert(function
->impl
== NULL
);
313 nir_function_impl
*impl
= nir_function_impl_create_bare(function
->shader
);
315 function
->impl
= impl
;
316 impl
->function
= function
;
322 nir_block_create(nir_shader
*shader
)
324 nir_block
*block
= rzalloc(shader
, nir_block
);
326 cf_init(&block
->cf_node
, nir_cf_node_block
);
328 block
->successors
[0] = block
->successors
[1] = NULL
;
329 block
->predecessors
= _mesa_pointer_set_create(block
);
330 block
->imm_dom
= NULL
;
331 /* XXX maybe it would be worth it to defer allocation? This
332 * way it doesn't get allocated for shader refs that never run
333 * nir_calc_dominance? For example, state-tracker creates an
334 * initial IR, clones that, runs appropriate lowering pass, passes
335 * to driver which does common lowering/opt, and then stores ref
336 * which is later used to do state specific lowering and futher
337 * opt. Do any of the references not need dominance metadata?
339 block
->dom_frontier
= _mesa_pointer_set_create(block
);
341 exec_list_make_empty(&block
->instr_list
);
347 src_init(nir_src
*src
)
351 src
->reg
.indirect
= NULL
;
352 src
->reg
.base_offset
= 0;
356 nir_if_create(nir_shader
*shader
)
358 nir_if
*if_stmt
= ralloc(shader
, nir_if
);
360 if_stmt
->control
= nir_selection_control_none
;
362 cf_init(&if_stmt
->cf_node
, nir_cf_node_if
);
363 src_init(&if_stmt
->condition
);
365 nir_block
*then
= nir_block_create(shader
);
366 exec_list_make_empty(&if_stmt
->then_list
);
367 exec_list_push_tail(&if_stmt
->then_list
, &then
->cf_node
.node
);
368 then
->cf_node
.parent
= &if_stmt
->cf_node
;
370 nir_block
*else_stmt
= nir_block_create(shader
);
371 exec_list_make_empty(&if_stmt
->else_list
);
372 exec_list_push_tail(&if_stmt
->else_list
, &else_stmt
->cf_node
.node
);
373 else_stmt
->cf_node
.parent
= &if_stmt
->cf_node
;
379 nir_loop_create(nir_shader
*shader
)
381 nir_loop
*loop
= rzalloc(shader
, nir_loop
);
383 cf_init(&loop
->cf_node
, nir_cf_node_loop
);
385 nir_block
*body
= nir_block_create(shader
);
386 exec_list_make_empty(&loop
->body
);
387 exec_list_push_tail(&loop
->body
, &body
->cf_node
.node
);
388 body
->cf_node
.parent
= &loop
->cf_node
;
390 body
->successors
[0] = body
;
391 _mesa_set_add(body
->predecessors
, body
);
397 instr_init(nir_instr
*instr
, nir_instr_type type
)
401 exec_node_init(&instr
->node
);
405 dest_init(nir_dest
*dest
)
407 dest
->is_ssa
= false;
408 dest
->reg
.reg
= NULL
;
409 dest
->reg
.indirect
= NULL
;
410 dest
->reg
.base_offset
= 0;
414 alu_dest_init(nir_alu_dest
*dest
)
416 dest_init(&dest
->dest
);
417 dest
->saturate
= false;
418 dest
->write_mask
= 0xf;
422 alu_src_init(nir_alu_src
*src
)
425 src
->abs
= src
->negate
= false;
426 for (int i
= 0; i
< NIR_MAX_VEC_COMPONENTS
; ++i
)
431 nir_alu_instr_create(nir_shader
*shader
, nir_op op
)
433 unsigned num_srcs
= nir_op_infos
[op
].num_inputs
;
434 /* TODO: don't use rzalloc */
435 nir_alu_instr
*instr
=
437 sizeof(nir_alu_instr
) + num_srcs
* sizeof(nir_alu_src
));
439 instr_init(&instr
->instr
, nir_instr_type_alu
);
441 alu_dest_init(&instr
->dest
);
442 for (unsigned i
= 0; i
< num_srcs
; i
++)
443 alu_src_init(&instr
->src
[i
]);
449 nir_deref_instr_create(nir_shader
*shader
, nir_deref_type deref_type
)
451 nir_deref_instr
*instr
=
452 rzalloc_size(shader
, sizeof(nir_deref_instr
));
454 instr_init(&instr
->instr
, nir_instr_type_deref
);
456 instr
->deref_type
= deref_type
;
457 if (deref_type
!= nir_deref_type_var
)
458 src_init(&instr
->parent
);
460 if (deref_type
== nir_deref_type_array
||
461 deref_type
== nir_deref_type_ptr_as_array
)
462 src_init(&instr
->arr
.index
);
464 dest_init(&instr
->dest
);
470 nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
)
472 nir_jump_instr
*instr
= ralloc(shader
, nir_jump_instr
);
473 instr_init(&instr
->instr
, nir_instr_type_jump
);
478 nir_load_const_instr
*
479 nir_load_const_instr_create(nir_shader
*shader
, unsigned num_components
,
482 nir_load_const_instr
*instr
=
483 rzalloc_size(shader
, sizeof(*instr
) + num_components
* sizeof(*instr
->value
));
484 instr_init(&instr
->instr
, nir_instr_type_load_const
);
486 nir_ssa_def_init(&instr
->instr
, &instr
->def
, num_components
, bit_size
, NULL
);
491 nir_intrinsic_instr
*
492 nir_intrinsic_instr_create(nir_shader
*shader
, nir_intrinsic_op op
)
494 unsigned num_srcs
= nir_intrinsic_infos
[op
].num_srcs
;
495 /* TODO: don't use rzalloc */
496 nir_intrinsic_instr
*instr
=
498 sizeof(nir_intrinsic_instr
) + num_srcs
* sizeof(nir_src
));
500 instr_init(&instr
->instr
, nir_instr_type_intrinsic
);
501 instr
->intrinsic
= op
;
503 if (nir_intrinsic_infos
[op
].has_dest
)
504 dest_init(&instr
->dest
);
506 for (unsigned i
= 0; i
< num_srcs
; i
++)
507 src_init(&instr
->src
[i
]);
513 nir_call_instr_create(nir_shader
*shader
, nir_function
*callee
)
515 const unsigned num_params
= callee
->num_params
;
516 nir_call_instr
*instr
=
517 rzalloc_size(shader
, sizeof(*instr
) +
518 num_params
* sizeof(instr
->params
[0]));
520 instr_init(&instr
->instr
, nir_instr_type_call
);
521 instr
->callee
= callee
;
522 instr
->num_params
= num_params
;
523 for (unsigned i
= 0; i
< num_params
; i
++)
524 src_init(&instr
->params
[i
]);
529 static int8_t default_tg4_offsets
[4][2] =
538 nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
)
540 nir_tex_instr
*instr
= rzalloc(shader
, nir_tex_instr
);
541 instr_init(&instr
->instr
, nir_instr_type_tex
);
543 dest_init(&instr
->dest
);
545 instr
->num_srcs
= num_srcs
;
546 instr
->src
= ralloc_array(instr
, nir_tex_src
, num_srcs
);
547 for (unsigned i
= 0; i
< num_srcs
; i
++)
548 src_init(&instr
->src
[i
].src
);
550 instr
->texture_index
= 0;
551 instr
->texture_array_size
= 0;
552 instr
->sampler_index
= 0;
553 memcpy(instr
->tg4_offsets
, default_tg4_offsets
, sizeof(instr
->tg4_offsets
));
559 nir_tex_instr_add_src(nir_tex_instr
*tex
,
560 nir_tex_src_type src_type
,
563 nir_tex_src
*new_srcs
= rzalloc_array(tex
, nir_tex_src
,
566 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
567 new_srcs
[i
].src_type
= tex
->src
[i
].src_type
;
568 nir_instr_move_src(&tex
->instr
, &new_srcs
[i
].src
,
572 ralloc_free(tex
->src
);
575 tex
->src
[tex
->num_srcs
].src_type
= src_type
;
576 nir_instr_rewrite_src(&tex
->instr
, &tex
->src
[tex
->num_srcs
].src
, src
);
581 nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
)
583 assert(src_idx
< tex
->num_srcs
);
585 /* First rewrite the source to NIR_SRC_INIT */
586 nir_instr_rewrite_src(&tex
->instr
, &tex
->src
[src_idx
].src
, NIR_SRC_INIT
);
588 /* Now, move all of the other sources down */
589 for (unsigned i
= src_idx
+ 1; i
< tex
->num_srcs
; i
++) {
590 tex
->src
[i
-1].src_type
= tex
->src
[i
].src_type
;
591 nir_instr_move_src(&tex
->instr
, &tex
->src
[i
-1].src
, &tex
->src
[i
].src
);
597 nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
)
599 if (tex
->op
!= nir_texop_tg4
)
601 return memcmp(tex
->tg4_offsets
, default_tg4_offsets
,
602 sizeof(tex
->tg4_offsets
)) != 0;
606 nir_phi_instr_create(nir_shader
*shader
)
608 nir_phi_instr
*instr
= ralloc(shader
, nir_phi_instr
);
609 instr_init(&instr
->instr
, nir_instr_type_phi
);
611 dest_init(&instr
->dest
);
612 exec_list_make_empty(&instr
->srcs
);
616 nir_parallel_copy_instr
*
617 nir_parallel_copy_instr_create(nir_shader
*shader
)
619 nir_parallel_copy_instr
*instr
= ralloc(shader
, nir_parallel_copy_instr
);
620 instr_init(&instr
->instr
, nir_instr_type_parallel_copy
);
622 exec_list_make_empty(&instr
->entries
);
627 nir_ssa_undef_instr
*
628 nir_ssa_undef_instr_create(nir_shader
*shader
,
629 unsigned num_components
,
632 nir_ssa_undef_instr
*instr
= ralloc(shader
, nir_ssa_undef_instr
);
633 instr_init(&instr
->instr
, nir_instr_type_ssa_undef
);
635 nir_ssa_def_init(&instr
->instr
, &instr
->def
, num_components
, bit_size
, NULL
);
640 static nir_const_value
641 const_value_float(double d
, unsigned bit_size
)
644 memset(&v
, 0, sizeof(v
));
646 case 16: v
.u16
= _mesa_float_to_half(d
); break;
647 case 32: v
.f32
= d
; break;
648 case 64: v
.f64
= d
; break;
650 unreachable("Invalid bit size");
655 static nir_const_value
656 const_value_int(int64_t i
, unsigned bit_size
)
659 memset(&v
, 0, sizeof(v
));
661 case 1: v
.b
= i
& 1; break;
662 case 8: v
.i8
= i
; break;
663 case 16: v
.i16
= i
; break;
664 case 32: v
.i32
= i
; break;
665 case 64: v
.i64
= i
; break;
667 unreachable("Invalid bit size");
673 nir_alu_binop_identity(nir_op binop
, unsigned bit_size
)
675 const int64_t max_int
= (1ull << (bit_size
- 1)) - 1;
676 const int64_t min_int
= -max_int
- 1;
679 return const_value_int(0, bit_size
);
681 return const_value_float(0, bit_size
);
683 return const_value_int(1, bit_size
);
685 return const_value_float(1, bit_size
);
687 return const_value_int(max_int
, bit_size
);
689 return const_value_int(~0ull, bit_size
);
691 return const_value_float(INFINITY
, bit_size
);
693 return const_value_int(min_int
, bit_size
);
695 return const_value_int(0, bit_size
);
697 return const_value_float(-INFINITY
, bit_size
);
699 return const_value_int(~0ull, bit_size
);
701 return const_value_int(0, bit_size
);
703 return const_value_int(0, bit_size
);
705 unreachable("Invalid reduction operation");
710 nir_cf_node_get_function(nir_cf_node
*node
)
712 while (node
->type
!= nir_cf_node_function
) {
716 return nir_cf_node_as_function(node
);
719 /* Reduces a cursor by trying to convert everything to after and trying to
720 * go up to block granularity when possible.
723 reduce_cursor(nir_cursor cursor
)
725 switch (cursor
.option
) {
726 case nir_cursor_before_block
:
727 assert(nir_cf_node_prev(&cursor
.block
->cf_node
) == NULL
||
728 nir_cf_node_prev(&cursor
.block
->cf_node
)->type
!= nir_cf_node_block
);
729 if (exec_list_is_empty(&cursor
.block
->instr_list
)) {
730 /* Empty block. After is as good as before. */
731 cursor
.option
= nir_cursor_after_block
;
735 case nir_cursor_after_block
:
738 case nir_cursor_before_instr
: {
739 nir_instr
*prev_instr
= nir_instr_prev(cursor
.instr
);
741 /* Before this instruction is after the previous */
742 cursor
.instr
= prev_instr
;
743 cursor
.option
= nir_cursor_after_instr
;
745 /* No previous instruction. Switch to before block */
746 cursor
.block
= cursor
.instr
->block
;
747 cursor
.option
= nir_cursor_before_block
;
749 return reduce_cursor(cursor
);
752 case nir_cursor_after_instr
:
753 if (nir_instr_next(cursor
.instr
) == NULL
) {
754 /* This is the last instruction, switch to after block */
755 cursor
.option
= nir_cursor_after_block
;
756 cursor
.block
= cursor
.instr
->block
;
761 unreachable("Inavlid cursor option");
766 nir_cursors_equal(nir_cursor a
, nir_cursor b
)
768 /* Reduced cursors should be unique */
769 a
= reduce_cursor(a
);
770 b
= reduce_cursor(b
);
772 return a
.block
== b
.block
&& a
.option
== b
.option
;
776 add_use_cb(nir_src
*src
, void *state
)
778 nir_instr
*instr
= state
;
780 src
->parent_instr
= instr
;
781 list_addtail(&src
->use_link
,
782 src
->is_ssa
? &src
->ssa
->uses
: &src
->reg
.reg
->uses
);
788 add_ssa_def_cb(nir_ssa_def
*def
, void *state
)
790 nir_instr
*instr
= state
;
792 if (instr
->block
&& def
->index
== UINT_MAX
) {
793 nir_function_impl
*impl
=
794 nir_cf_node_get_function(&instr
->block
->cf_node
);
796 def
->index
= impl
->ssa_alloc
++;
803 add_reg_def_cb(nir_dest
*dest
, void *state
)
805 nir_instr
*instr
= state
;
808 dest
->reg
.parent_instr
= instr
;
809 list_addtail(&dest
->reg
.def_link
, &dest
->reg
.reg
->defs
);
816 add_defs_uses(nir_instr
*instr
)
818 nir_foreach_src(instr
, add_use_cb
, instr
);
819 nir_foreach_dest(instr
, add_reg_def_cb
, instr
);
820 nir_foreach_ssa_def(instr
, add_ssa_def_cb
, instr
);
824 nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
)
826 switch (cursor
.option
) {
827 case nir_cursor_before_block
:
828 /* Only allow inserting jumps into empty blocks. */
829 if (instr
->type
== nir_instr_type_jump
)
830 assert(exec_list_is_empty(&cursor
.block
->instr_list
));
832 instr
->block
= cursor
.block
;
833 add_defs_uses(instr
);
834 exec_list_push_head(&cursor
.block
->instr_list
, &instr
->node
);
836 case nir_cursor_after_block
: {
837 /* Inserting instructions after a jump is illegal. */
838 nir_instr
*last
= nir_block_last_instr(cursor
.block
);
839 assert(last
== NULL
|| last
->type
!= nir_instr_type_jump
);
842 instr
->block
= cursor
.block
;
843 add_defs_uses(instr
);
844 exec_list_push_tail(&cursor
.block
->instr_list
, &instr
->node
);
847 case nir_cursor_before_instr
:
848 assert(instr
->type
!= nir_instr_type_jump
);
849 instr
->block
= cursor
.instr
->block
;
850 add_defs_uses(instr
);
851 exec_node_insert_node_before(&cursor
.instr
->node
, &instr
->node
);
853 case nir_cursor_after_instr
:
854 /* Inserting instructions after a jump is illegal. */
855 assert(cursor
.instr
->type
!= nir_instr_type_jump
);
857 /* Only allow inserting jumps at the end of the block. */
858 if (instr
->type
== nir_instr_type_jump
)
859 assert(cursor
.instr
== nir_block_last_instr(cursor
.instr
->block
));
861 instr
->block
= cursor
.instr
->block
;
862 add_defs_uses(instr
);
863 exec_node_insert_after(&cursor
.instr
->node
, &instr
->node
);
867 if (instr
->type
== nir_instr_type_jump
)
868 nir_handle_add_jump(instr
->block
);
872 src_is_valid(const nir_src
*src
)
874 return src
->is_ssa
? (src
->ssa
!= NULL
) : (src
->reg
.reg
!= NULL
);
878 remove_use_cb(nir_src
*src
, void *state
)
882 if (src_is_valid(src
))
883 list_del(&src
->use_link
);
889 remove_def_cb(nir_dest
*dest
, void *state
)
894 list_del(&dest
->reg
.def_link
);
900 remove_defs_uses(nir_instr
*instr
)
902 nir_foreach_dest(instr
, remove_def_cb
, instr
);
903 nir_foreach_src(instr
, remove_use_cb
, instr
);
906 void nir_instr_remove_v(nir_instr
*instr
)
908 remove_defs_uses(instr
);
909 exec_node_remove(&instr
->node
);
911 if (instr
->type
== nir_instr_type_jump
) {
912 nir_jump_instr
*jump_instr
= nir_instr_as_jump(instr
);
913 nir_handle_remove_jump(instr
->block
, jump_instr
->type
);
920 nir_index_local_regs(nir_function_impl
*impl
)
923 foreach_list_typed(nir_register
, reg
, node
, &impl
->registers
) {
924 reg
->index
= index
++;
926 impl
->reg_alloc
= index
;
930 visit_alu_dest(nir_alu_instr
*instr
, nir_foreach_dest_cb cb
, void *state
)
932 return cb(&instr
->dest
.dest
, state
);
936 visit_deref_dest(nir_deref_instr
*instr
, nir_foreach_dest_cb cb
, void *state
)
938 return cb(&instr
->dest
, state
);
942 visit_intrinsic_dest(nir_intrinsic_instr
*instr
, nir_foreach_dest_cb cb
,
945 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
946 return cb(&instr
->dest
, state
);
952 visit_texture_dest(nir_tex_instr
*instr
, nir_foreach_dest_cb cb
,
955 return cb(&instr
->dest
, state
);
959 visit_phi_dest(nir_phi_instr
*instr
, nir_foreach_dest_cb cb
, void *state
)
961 return cb(&instr
->dest
, state
);
965 visit_parallel_copy_dest(nir_parallel_copy_instr
*instr
,
966 nir_foreach_dest_cb cb
, void *state
)
968 nir_foreach_parallel_copy_entry(entry
, instr
) {
969 if (!cb(&entry
->dest
, state
))
977 nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
)
979 switch (instr
->type
) {
980 case nir_instr_type_alu
:
981 return visit_alu_dest(nir_instr_as_alu(instr
), cb
, state
);
982 case nir_instr_type_deref
:
983 return visit_deref_dest(nir_instr_as_deref(instr
), cb
, state
);
984 case nir_instr_type_intrinsic
:
985 return visit_intrinsic_dest(nir_instr_as_intrinsic(instr
), cb
, state
);
986 case nir_instr_type_tex
:
987 return visit_texture_dest(nir_instr_as_tex(instr
), cb
, state
);
988 case nir_instr_type_phi
:
989 return visit_phi_dest(nir_instr_as_phi(instr
), cb
, state
);
990 case nir_instr_type_parallel_copy
:
991 return visit_parallel_copy_dest(nir_instr_as_parallel_copy(instr
),
994 case nir_instr_type_load_const
:
995 case nir_instr_type_ssa_undef
:
996 case nir_instr_type_call
:
997 case nir_instr_type_jump
:
1001 unreachable("Invalid instruction type");
1008 struct foreach_ssa_def_state
{
1009 nir_foreach_ssa_def_cb cb
;
1014 nir_ssa_def_visitor(nir_dest
*dest
, void *void_state
)
1016 struct foreach_ssa_def_state
*state
= void_state
;
1019 return state
->cb(&dest
->ssa
, state
->client_state
);
1025 nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
, void *state
)
1027 switch (instr
->type
) {
1028 case nir_instr_type_alu
:
1029 case nir_instr_type_deref
:
1030 case nir_instr_type_tex
:
1031 case nir_instr_type_intrinsic
:
1032 case nir_instr_type_phi
:
1033 case nir_instr_type_parallel_copy
: {
1034 struct foreach_ssa_def_state foreach_state
= {cb
, state
};
1035 return nir_foreach_dest(instr
, nir_ssa_def_visitor
, &foreach_state
);
1038 case nir_instr_type_load_const
:
1039 return cb(&nir_instr_as_load_const(instr
)->def
, state
);
1040 case nir_instr_type_ssa_undef
:
1041 return cb(&nir_instr_as_ssa_undef(instr
)->def
, state
);
1042 case nir_instr_type_call
:
1043 case nir_instr_type_jump
:
1046 unreachable("Invalid instruction type");
1051 nir_instr_ssa_def(nir_instr
*instr
)
1053 switch (instr
->type
) {
1054 case nir_instr_type_alu
:
1055 assert(nir_instr_as_alu(instr
)->dest
.dest
.is_ssa
);
1056 return &nir_instr_as_alu(instr
)->dest
.dest
.ssa
;
1058 case nir_instr_type_deref
:
1059 assert(nir_instr_as_deref(instr
)->dest
.is_ssa
);
1060 return &nir_instr_as_deref(instr
)->dest
.ssa
;
1062 case nir_instr_type_tex
:
1063 assert(nir_instr_as_tex(instr
)->dest
.is_ssa
);
1064 return &nir_instr_as_tex(instr
)->dest
.ssa
;
1066 case nir_instr_type_intrinsic
: {
1067 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1068 if (nir_intrinsic_infos
[intrin
->intrinsic
].has_dest
) {
1069 assert(intrin
->dest
.is_ssa
);
1070 return &intrin
->dest
.ssa
;
1076 case nir_instr_type_phi
:
1077 assert(nir_instr_as_phi(instr
)->dest
.is_ssa
);
1078 return &nir_instr_as_phi(instr
)->dest
.ssa
;
1080 case nir_instr_type_parallel_copy
:
1081 unreachable("Parallel copies are unsupported by this function");
1083 case nir_instr_type_load_const
:
1084 return &nir_instr_as_load_const(instr
)->def
;
1086 case nir_instr_type_ssa_undef
:
1087 return &nir_instr_as_ssa_undef(instr
)->def
;
1089 case nir_instr_type_call
:
1090 case nir_instr_type_jump
:
1094 unreachable("Invalid instruction type");
1098 visit_src(nir_src
*src
, nir_foreach_src_cb cb
, void *state
)
1100 if (!cb(src
, state
))
1102 if (!src
->is_ssa
&& src
->reg
.indirect
)
1103 return cb(src
->reg
.indirect
, state
);
1108 visit_alu_src(nir_alu_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1110 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
1111 if (!visit_src(&instr
->src
[i
].src
, cb
, state
))
1118 visit_deref_instr_src(nir_deref_instr
*instr
,
1119 nir_foreach_src_cb cb
, void *state
)
1121 if (instr
->deref_type
!= nir_deref_type_var
) {
1122 if (!visit_src(&instr
->parent
, cb
, state
))
1126 if (instr
->deref_type
== nir_deref_type_array
||
1127 instr
->deref_type
== nir_deref_type_ptr_as_array
) {
1128 if (!visit_src(&instr
->arr
.index
, cb
, state
))
1136 visit_tex_src(nir_tex_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1138 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
1139 if (!visit_src(&instr
->src
[i
].src
, cb
, state
))
1147 visit_intrinsic_src(nir_intrinsic_instr
*instr
, nir_foreach_src_cb cb
,
1150 unsigned num_srcs
= nir_intrinsic_infos
[instr
->intrinsic
].num_srcs
;
1151 for (unsigned i
= 0; i
< num_srcs
; i
++) {
1152 if (!visit_src(&instr
->src
[i
], cb
, state
))
1160 visit_call_src(nir_call_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1162 for (unsigned i
= 0; i
< instr
->num_params
; i
++) {
1163 if (!visit_src(&instr
->params
[i
], cb
, state
))
1171 visit_phi_src(nir_phi_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1173 nir_foreach_phi_src(src
, instr
) {
1174 if (!visit_src(&src
->src
, cb
, state
))
1182 visit_parallel_copy_src(nir_parallel_copy_instr
*instr
,
1183 nir_foreach_src_cb cb
, void *state
)
1185 nir_foreach_parallel_copy_entry(entry
, instr
) {
1186 if (!visit_src(&entry
->src
, cb
, state
))
1195 nir_foreach_src_cb cb
;
1196 } visit_dest_indirect_state
;
1199 visit_dest_indirect(nir_dest
*dest
, void *_state
)
1201 visit_dest_indirect_state
*state
= (visit_dest_indirect_state
*) _state
;
1203 if (!dest
->is_ssa
&& dest
->reg
.indirect
)
1204 return state
->cb(dest
->reg
.indirect
, state
->state
);
1210 nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1212 switch (instr
->type
) {
1213 case nir_instr_type_alu
:
1214 if (!visit_alu_src(nir_instr_as_alu(instr
), cb
, state
))
1217 case nir_instr_type_deref
:
1218 if (!visit_deref_instr_src(nir_instr_as_deref(instr
), cb
, state
))
1221 case nir_instr_type_intrinsic
:
1222 if (!visit_intrinsic_src(nir_instr_as_intrinsic(instr
), cb
, state
))
1225 case nir_instr_type_tex
:
1226 if (!visit_tex_src(nir_instr_as_tex(instr
), cb
, state
))
1229 case nir_instr_type_call
:
1230 if (!visit_call_src(nir_instr_as_call(instr
), cb
, state
))
1233 case nir_instr_type_load_const
:
1234 /* Constant load instructions have no regular sources */
1236 case nir_instr_type_phi
:
1237 if (!visit_phi_src(nir_instr_as_phi(instr
), cb
, state
))
1240 case nir_instr_type_parallel_copy
:
1241 if (!visit_parallel_copy_src(nir_instr_as_parallel_copy(instr
),
1245 case nir_instr_type_jump
:
1246 case nir_instr_type_ssa_undef
:
1250 unreachable("Invalid instruction type");
1254 visit_dest_indirect_state dest_state
;
1255 dest_state
.state
= state
;
1257 return nir_foreach_dest(instr
, visit_dest_indirect
, &dest_state
);
1261 nir_const_value_for_float(double f
, unsigned bit_size
)
1264 memset(&v
, 0, sizeof(v
));
1268 v
.u16
= _mesa_float_to_half(f
);
1277 unreachable("Invalid bit size");
1284 nir_const_value_as_float(nir_const_value value
, unsigned bit_size
)
1287 case 16: return _mesa_half_to_float(value
.u16
);
1288 case 32: return value
.f32
;
1289 case 64: return value
.f64
;
1291 unreachable("Invalid bit size");
1296 nir_src_as_const_value(nir_src src
)
1301 if (src
.ssa
->parent_instr
->type
!= nir_instr_type_load_const
)
1304 nir_load_const_instr
*load
= nir_instr_as_load_const(src
.ssa
->parent_instr
);
1310 * Returns true if the source is known to be dynamically uniform. Otherwise it
1311 * returns false which means it may or may not be dynamically uniform but it
1312 * can't be determined.
1315 nir_src_is_dynamically_uniform(nir_src src
)
1320 /* Constants are trivially dynamically uniform */
1321 if (src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
)
1324 /* As are uniform variables */
1325 if (src
.ssa
->parent_instr
->type
== nir_instr_type_intrinsic
) {
1326 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(src
.ssa
->parent_instr
);
1328 if (intr
->intrinsic
== nir_intrinsic_load_uniform
)
1332 /* Operating together dynamically uniform expressions produces a
1333 * dynamically uniform result
1335 if (src
.ssa
->parent_instr
->type
== nir_instr_type_alu
) {
1336 nir_alu_instr
*alu
= nir_instr_as_alu(src
.ssa
->parent_instr
);
1337 for (int i
= 0; i
< nir_op_infos
[alu
->op
].num_inputs
; i
++) {
1338 if (!nir_src_is_dynamically_uniform(alu
->src
[i
].src
))
1345 /* XXX: this could have many more tests, such as when a sampler function is
1346 * called with dynamically uniform arguments.
1352 src_remove_all_uses(nir_src
*src
)
1354 for (; src
; src
= src
->is_ssa
? NULL
: src
->reg
.indirect
) {
1355 if (!src_is_valid(src
))
1358 list_del(&src
->use_link
);
1363 src_add_all_uses(nir_src
*src
, nir_instr
*parent_instr
, nir_if
*parent_if
)
1365 for (; src
; src
= src
->is_ssa
? NULL
: src
->reg
.indirect
) {
1366 if (!src_is_valid(src
))
1370 src
->parent_instr
= parent_instr
;
1372 list_addtail(&src
->use_link
, &src
->ssa
->uses
);
1374 list_addtail(&src
->use_link
, &src
->reg
.reg
->uses
);
1377 src
->parent_if
= parent_if
;
1379 list_addtail(&src
->use_link
, &src
->ssa
->if_uses
);
1381 list_addtail(&src
->use_link
, &src
->reg
.reg
->if_uses
);
1387 nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
)
1389 assert(!src_is_valid(src
) || src
->parent_instr
== instr
);
1391 src_remove_all_uses(src
);
1393 src_add_all_uses(src
, instr
, NULL
);
1397 nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
)
1399 assert(!src_is_valid(dest
) || dest
->parent_instr
== dest_instr
);
1401 src_remove_all_uses(dest
);
1402 src_remove_all_uses(src
);
1404 *src
= NIR_SRC_INIT
;
1405 src_add_all_uses(dest
, dest_instr
, NULL
);
1409 nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
)
1411 nir_src
*src
= &if_stmt
->condition
;
1412 assert(!src_is_valid(src
) || src
->parent_if
== if_stmt
);
1414 src_remove_all_uses(src
);
1416 src_add_all_uses(src
, NULL
, if_stmt
);
1420 nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
, nir_dest new_dest
)
1423 /* We can only overwrite an SSA destination if it has no uses. */
1424 assert(list_is_empty(&dest
->ssa
.uses
) && list_is_empty(&dest
->ssa
.if_uses
));
1426 list_del(&dest
->reg
.def_link
);
1427 if (dest
->reg
.indirect
)
1428 src_remove_all_uses(dest
->reg
.indirect
);
1431 /* We can't re-write with an SSA def */
1432 assert(!new_dest
.is_ssa
);
1434 nir_dest_copy(dest
, &new_dest
, instr
);
1436 dest
->reg
.parent_instr
= instr
;
1437 list_addtail(&dest
->reg
.def_link
, &new_dest
.reg
.reg
->defs
);
1439 if (dest
->reg
.indirect
)
1440 src_add_all_uses(dest
->reg
.indirect
, instr
, NULL
);
1443 /* note: does *not* take ownership of 'name' */
1445 nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
1446 unsigned num_components
,
1447 unsigned bit_size
, const char *name
)
1449 def
->name
= ralloc_strdup(instr
, name
);
1450 def
->parent_instr
= instr
;
1451 list_inithead(&def
->uses
);
1452 list_inithead(&def
->if_uses
);
1453 def
->num_components
= num_components
;
1454 def
->bit_size
= bit_size
;
1457 nir_function_impl
*impl
=
1458 nir_cf_node_get_function(&instr
->block
->cf_node
);
1460 def
->index
= impl
->ssa_alloc
++;
1462 def
->index
= UINT_MAX
;
1466 /* note: does *not* take ownership of 'name' */
1468 nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
1469 unsigned num_components
, unsigned bit_size
,
1472 dest
->is_ssa
= true;
1473 nir_ssa_def_init(instr
, &dest
->ssa
, num_components
, bit_size
, name
);
1477 nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
)
1479 assert(!new_src
.is_ssa
|| def
!= new_src
.ssa
);
1481 nir_foreach_use_safe(use_src
, def
)
1482 nir_instr_rewrite_src(use_src
->parent_instr
, use_src
, new_src
);
1484 nir_foreach_if_use_safe(use_src
, def
)
1485 nir_if_rewrite_condition(use_src
->parent_if
, new_src
);
1489 is_instr_between(nir_instr
*start
, nir_instr
*end
, nir_instr
*between
)
1491 assert(start
->block
== end
->block
);
1493 if (between
->block
!= start
->block
)
1496 /* Search backwards looking for "between" */
1497 while (start
!= end
) {
1501 end
= nir_instr_prev(end
);
1508 /* Replaces all uses of the given SSA def with the given source but only if
1509 * the use comes after the after_me instruction. This can be useful if you
1510 * are emitting code to fix up the result of some instruction: you can freely
1511 * use the result in that code and then call rewrite_uses_after and pass the
1512 * last fixup instruction as after_me and it will replace all of the uses you
1513 * want without touching the fixup code.
1515 * This function assumes that after_me is in the same block as
1516 * def->parent_instr and that after_me comes after def->parent_instr.
1519 nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
1520 nir_instr
*after_me
)
1522 if (new_src
.is_ssa
&& def
== new_src
.ssa
)
1525 nir_foreach_use_safe(use_src
, def
) {
1526 assert(use_src
->parent_instr
!= def
->parent_instr
);
1527 /* Since def already dominates all of its uses, the only way a use can
1528 * not be dominated by after_me is if it is between def and after_me in
1529 * the instruction list.
1531 if (!is_instr_between(def
->parent_instr
, after_me
, use_src
->parent_instr
))
1532 nir_instr_rewrite_src(use_src
->parent_instr
, use_src
, new_src
);
1535 nir_foreach_if_use_safe(use_src
, def
)
1536 nir_if_rewrite_condition(use_src
->parent_if
, new_src
);
1539 nir_component_mask_t
1540 nir_ssa_def_components_read(const nir_ssa_def
*def
)
1542 nir_component_mask_t read_mask
= 0;
1543 nir_foreach_use(use
, def
) {
1544 if (use
->parent_instr
->type
== nir_instr_type_alu
) {
1545 nir_alu_instr
*alu
= nir_instr_as_alu(use
->parent_instr
);
1546 nir_alu_src
*alu_src
= exec_node_data(nir_alu_src
, use
, src
);
1547 int src_idx
= alu_src
- &alu
->src
[0];
1548 assert(src_idx
>= 0 && src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
1549 read_mask
|= nir_alu_instr_src_read_mask(alu
, src_idx
);
1551 return (1 << def
->num_components
) - 1;
1555 if (!list_is_empty(&def
->if_uses
))
1562 nir_block_cf_tree_next(nir_block
*block
)
1564 if (block
== NULL
) {
1565 /* nir_foreach_block_safe() will call this function on a NULL block
1566 * after the last iteration, but it won't use the result so just return
1572 nir_cf_node
*cf_next
= nir_cf_node_next(&block
->cf_node
);
1574 return nir_cf_node_cf_tree_first(cf_next
);
1576 nir_cf_node
*parent
= block
->cf_node
.parent
;
1578 switch (parent
->type
) {
1579 case nir_cf_node_if
: {
1580 /* Are we at the end of the if? Go to the beginning of the else */
1581 nir_if
*if_stmt
= nir_cf_node_as_if(parent
);
1582 if (block
== nir_if_last_then_block(if_stmt
))
1583 return nir_if_first_else_block(if_stmt
);
1585 assert(block
== nir_if_last_else_block(if_stmt
));
1589 case nir_cf_node_loop
:
1590 return nir_cf_node_as_block(nir_cf_node_next(parent
));
1592 case nir_cf_node_function
:
1596 unreachable("unknown cf node type");
1601 nir_block_cf_tree_prev(nir_block
*block
)
1603 if (block
== NULL
) {
1604 /* do this for consistency with nir_block_cf_tree_next() */
1608 nir_cf_node
*cf_prev
= nir_cf_node_prev(&block
->cf_node
);
1610 return nir_cf_node_cf_tree_last(cf_prev
);
1612 nir_cf_node
*parent
= block
->cf_node
.parent
;
1614 switch (parent
->type
) {
1615 case nir_cf_node_if
: {
1616 /* Are we at the beginning of the else? Go to the end of the if */
1617 nir_if
*if_stmt
= nir_cf_node_as_if(parent
);
1618 if (block
== nir_if_first_else_block(if_stmt
))
1619 return nir_if_last_then_block(if_stmt
);
1621 assert(block
== nir_if_first_then_block(if_stmt
));
1625 case nir_cf_node_loop
:
1626 return nir_cf_node_as_block(nir_cf_node_prev(parent
));
1628 case nir_cf_node_function
:
1632 unreachable("unknown cf node type");
1636 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
)
1638 switch (node
->type
) {
1639 case nir_cf_node_function
: {
1640 nir_function_impl
*impl
= nir_cf_node_as_function(node
);
1641 return nir_start_block(impl
);
1644 case nir_cf_node_if
: {
1645 nir_if
*if_stmt
= nir_cf_node_as_if(node
);
1646 return nir_if_first_then_block(if_stmt
);
1649 case nir_cf_node_loop
: {
1650 nir_loop
*loop
= nir_cf_node_as_loop(node
);
1651 return nir_loop_first_block(loop
);
1654 case nir_cf_node_block
: {
1655 return nir_cf_node_as_block(node
);
1659 unreachable("unknown node type");
1663 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
)
1665 switch (node
->type
) {
1666 case nir_cf_node_function
: {
1667 nir_function_impl
*impl
= nir_cf_node_as_function(node
);
1668 return nir_impl_last_block(impl
);
1671 case nir_cf_node_if
: {
1672 nir_if
*if_stmt
= nir_cf_node_as_if(node
);
1673 return nir_if_last_else_block(if_stmt
);
1676 case nir_cf_node_loop
: {
1677 nir_loop
*loop
= nir_cf_node_as_loop(node
);
1678 return nir_loop_last_block(loop
);
1681 case nir_cf_node_block
: {
1682 return nir_cf_node_as_block(node
);
1686 unreachable("unknown node type");
1690 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
)
1692 if (node
->type
== nir_cf_node_block
)
1693 return nir_block_cf_tree_next(nir_cf_node_as_block(node
));
1694 else if (node
->type
== nir_cf_node_function
)
1697 return nir_cf_node_as_block(nir_cf_node_next(node
));
1701 nir_block_get_following_if(nir_block
*block
)
1703 if (exec_node_is_tail_sentinel(&block
->cf_node
.node
))
1706 if (nir_cf_node_is_last(&block
->cf_node
))
1709 nir_cf_node
*next_node
= nir_cf_node_next(&block
->cf_node
);
1711 if (next_node
->type
!= nir_cf_node_if
)
1714 return nir_cf_node_as_if(next_node
);
1718 nir_block_get_following_loop(nir_block
*block
)
1720 if (exec_node_is_tail_sentinel(&block
->cf_node
.node
))
1723 if (nir_cf_node_is_last(&block
->cf_node
))
1726 nir_cf_node
*next_node
= nir_cf_node_next(&block
->cf_node
);
1728 if (next_node
->type
!= nir_cf_node_loop
)
1731 return nir_cf_node_as_loop(next_node
);
1735 nir_index_blocks(nir_function_impl
*impl
)
1739 if (impl
->valid_metadata
& nir_metadata_block_index
)
1742 nir_foreach_block(block
, impl
) {
1743 block
->index
= index
++;
1746 /* The end_block isn't really part of the program, which is why its index
1749 impl
->num_blocks
= impl
->end_block
->index
= index
;
1753 index_ssa_def_cb(nir_ssa_def
*def
, void *state
)
1755 unsigned *index
= (unsigned *) state
;
1756 def
->index
= (*index
)++;
1762 * The indices are applied top-to-bottom which has the very nice property
1763 * that, if A dominates B, then A->index <= B->index.
1766 nir_index_ssa_defs(nir_function_impl
*impl
)
1770 nir_foreach_block(block
, impl
) {
1771 nir_foreach_instr(instr
, block
)
1772 nir_foreach_ssa_def(instr
, index_ssa_def_cb
, &index
);
1775 impl
->ssa_alloc
= index
;
1779 * The indices are applied top-to-bottom which has the very nice property
1780 * that, if A dominates B, then A->index <= B->index.
1783 nir_index_instrs(nir_function_impl
*impl
)
1787 nir_foreach_block(block
, impl
) {
1788 nir_foreach_instr(instr
, block
)
1789 instr
->index
= index
++;
1796 index_var_list(struct exec_list
*list
)
1798 unsigned next_index
= 0;
1799 nir_foreach_variable(var
, list
)
1800 var
->index
= next_index
++;
1804 nir_index_vars(nir_shader
*shader
, nir_function_impl
*impl
, nir_variable_mode modes
)
1806 if ((modes
& nir_var_function_temp
) && impl
)
1807 index_var_list(&impl
->locals
);
1809 if (modes
& nir_var_shader_temp
)
1810 index_var_list(&shader
->globals
);
1812 if (modes
& nir_var_shader_in
)
1813 index_var_list(&shader
->inputs
);
1815 if (modes
& nir_var_shader_out
)
1816 index_var_list(&shader
->outputs
);
1818 if (modes
& (nir_var_uniform
| nir_var_mem_ubo
| nir_var_mem_ssbo
))
1819 index_var_list(&shader
->uniforms
);
1821 if (modes
& nir_var_mem_shared
)
1822 index_var_list(&shader
->shared
);
1824 if (modes
& nir_var_system_value
)
1825 index_var_list(&shader
->system_values
);
1829 cursor_next_instr(nir_cursor cursor
)
1831 switch (cursor
.option
) {
1832 case nir_cursor_before_block
:
1833 for (nir_block
*block
= cursor
.block
; block
;
1834 block
= nir_block_cf_tree_next(block
)) {
1835 nir_instr
*instr
= nir_block_first_instr(block
);
1841 case nir_cursor_after_block
:
1842 cursor
.block
= nir_block_cf_tree_next(cursor
.block
);
1843 if (cursor
.block
== NULL
)
1846 cursor
.option
= nir_cursor_before_block
;
1847 return cursor_next_instr(cursor
);
1849 case nir_cursor_before_instr
:
1850 return cursor
.instr
;
1852 case nir_cursor_after_instr
:
1853 if (nir_instr_next(cursor
.instr
))
1854 return nir_instr_next(cursor
.instr
);
1856 cursor
.option
= nir_cursor_after_block
;
1857 cursor
.block
= cursor
.instr
->block
;
1858 return cursor_next_instr(cursor
);
1861 unreachable("Inavlid cursor option");
1864 ASSERTED
static bool
1865 dest_is_ssa(nir_dest
*dest
, void *_state
)
1868 return dest
->is_ssa
;
1872 nir_function_impl_lower_instructions(nir_function_impl
*impl
,
1873 nir_instr_filter_cb filter
,
1874 nir_lower_instr_cb lower
,
1878 nir_builder_init(&b
, impl
);
1880 nir_metadata preserved
= nir_metadata_block_index
|
1881 nir_metadata_dominance
;
1883 bool progress
= false;
1884 nir_cursor iter
= nir_before_cf_list(&impl
->body
);
1886 while ((instr
= cursor_next_instr(iter
)) != NULL
) {
1887 if (filter
&& !filter(instr
, cb_data
)) {
1888 iter
= nir_after_instr(instr
);
1892 assert(nir_foreach_dest(instr
, dest_is_ssa
, NULL
));
1893 nir_ssa_def
*old_def
= nir_instr_ssa_def(instr
);
1894 if (old_def
== NULL
) {
1895 iter
= nir_after_instr(instr
);
1899 /* We're about to ask the callback to generate a replacement for instr.
1900 * Save off the uses from instr's SSA def so we know what uses to
1901 * rewrite later. If we use nir_ssa_def_rewrite_uses, it fails in the
1902 * case where the generated replacement code uses the result of instr
1903 * itself. If we use nir_ssa_def_rewrite_uses_after (which is the
1904 * normal solution to this problem), it doesn't work well if control-
1905 * flow is inserted as part of the replacement, doesn't handle cases
1906 * where the replacement is something consumed by instr, and suffers
1907 * from performance issues. This is the only way to 100% guarantee
1908 * that we rewrite the correct set efficiently.
1910 struct list_head old_uses
, old_if_uses
;
1911 list_replace(&old_def
->uses
, &old_uses
);
1912 list_inithead(&old_def
->uses
);
1913 list_replace(&old_def
->if_uses
, &old_if_uses
);
1914 list_inithead(&old_def
->if_uses
);
1916 b
.cursor
= nir_after_instr(instr
);
1917 nir_ssa_def
*new_def
= lower(&b
, instr
, cb_data
);
1918 if (new_def
&& new_def
!= NIR_LOWER_INSTR_PROGRESS
) {
1919 assert(old_def
!= NULL
);
1920 if (new_def
->parent_instr
->block
!= instr
->block
)
1921 preserved
= nir_metadata_none
;
1923 nir_src new_src
= nir_src_for_ssa(new_def
);
1924 list_for_each_entry_safe(nir_src
, use_src
, &old_uses
, use_link
)
1925 nir_instr_rewrite_src(use_src
->parent_instr
, use_src
, new_src
);
1927 list_for_each_entry_safe(nir_src
, use_src
, &old_if_uses
, use_link
)
1928 nir_if_rewrite_condition(use_src
->parent_if
, new_src
);
1930 if (list_is_empty(&old_def
->uses
) && list_is_empty(&old_def
->if_uses
)) {
1931 iter
= nir_instr_remove(instr
);
1933 iter
= nir_after_instr(instr
);
1937 /* We didn't end up lowering after all. Put the uses back */
1939 list_replace(&old_uses
, &old_def
->uses
);
1940 list_replace(&old_if_uses
, &old_def
->if_uses
);
1942 iter
= nir_after_instr(instr
);
1944 if (new_def
== NIR_LOWER_INSTR_PROGRESS
)
1950 nir_metadata_preserve(impl
, preserved
);
1953 impl
->valid_metadata
&= ~nir_metadata_not_properly_reset
;
1961 nir_shader_lower_instructions(nir_shader
*shader
,
1962 nir_instr_filter_cb filter
,
1963 nir_lower_instr_cb lower
,
1966 bool progress
= false;
1968 nir_foreach_function(function
, shader
) {
1969 if (function
->impl
&&
1970 nir_function_impl_lower_instructions(function
->impl
,
1971 filter
, lower
, cb_data
))
1979 nir_intrinsic_from_system_value(gl_system_value val
)
1982 case SYSTEM_VALUE_VERTEX_ID
:
1983 return nir_intrinsic_load_vertex_id
;
1984 case SYSTEM_VALUE_INSTANCE_ID
:
1985 return nir_intrinsic_load_instance_id
;
1986 case SYSTEM_VALUE_DRAW_ID
:
1987 return nir_intrinsic_load_draw_id
;
1988 case SYSTEM_VALUE_BASE_INSTANCE
:
1989 return nir_intrinsic_load_base_instance
;
1990 case SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
:
1991 return nir_intrinsic_load_vertex_id_zero_base
;
1992 case SYSTEM_VALUE_IS_INDEXED_DRAW
:
1993 return nir_intrinsic_load_is_indexed_draw
;
1994 case SYSTEM_VALUE_FIRST_VERTEX
:
1995 return nir_intrinsic_load_first_vertex
;
1996 case SYSTEM_VALUE_BASE_VERTEX
:
1997 return nir_intrinsic_load_base_vertex
;
1998 case SYSTEM_VALUE_INVOCATION_ID
:
1999 return nir_intrinsic_load_invocation_id
;
2000 case SYSTEM_VALUE_FRAG_COORD
:
2001 return nir_intrinsic_load_frag_coord
;
2002 case SYSTEM_VALUE_POINT_COORD
:
2003 return nir_intrinsic_load_point_coord
;
2004 case SYSTEM_VALUE_FRONT_FACE
:
2005 return nir_intrinsic_load_front_face
;
2006 case SYSTEM_VALUE_SAMPLE_ID
:
2007 return nir_intrinsic_load_sample_id
;
2008 case SYSTEM_VALUE_SAMPLE_POS
:
2009 return nir_intrinsic_load_sample_pos
;
2010 case SYSTEM_VALUE_SAMPLE_MASK_IN
:
2011 return nir_intrinsic_load_sample_mask_in
;
2012 case SYSTEM_VALUE_LOCAL_INVOCATION_ID
:
2013 return nir_intrinsic_load_local_invocation_id
;
2014 case SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
:
2015 return nir_intrinsic_load_local_invocation_index
;
2016 case SYSTEM_VALUE_WORK_GROUP_ID
:
2017 return nir_intrinsic_load_work_group_id
;
2018 case SYSTEM_VALUE_NUM_WORK_GROUPS
:
2019 return nir_intrinsic_load_num_work_groups
;
2020 case SYSTEM_VALUE_PRIMITIVE_ID
:
2021 return nir_intrinsic_load_primitive_id
;
2022 case SYSTEM_VALUE_TESS_COORD
:
2023 return nir_intrinsic_load_tess_coord
;
2024 case SYSTEM_VALUE_TESS_LEVEL_OUTER
:
2025 return nir_intrinsic_load_tess_level_outer
;
2026 case SYSTEM_VALUE_TESS_LEVEL_INNER
:
2027 return nir_intrinsic_load_tess_level_inner
;
2028 case SYSTEM_VALUE_TESS_LEVEL_OUTER_DEFAULT
:
2029 return nir_intrinsic_load_tess_level_outer_default
;
2030 case SYSTEM_VALUE_TESS_LEVEL_INNER_DEFAULT
:
2031 return nir_intrinsic_load_tess_level_inner_default
;
2032 case SYSTEM_VALUE_VERTICES_IN
:
2033 return nir_intrinsic_load_patch_vertices_in
;
2034 case SYSTEM_VALUE_HELPER_INVOCATION
:
2035 return nir_intrinsic_load_helper_invocation
;
2036 case SYSTEM_VALUE_COLOR0
:
2037 return nir_intrinsic_load_color0
;
2038 case SYSTEM_VALUE_COLOR1
:
2039 return nir_intrinsic_load_color1
;
2040 case SYSTEM_VALUE_VIEW_INDEX
:
2041 return nir_intrinsic_load_view_index
;
2042 case SYSTEM_VALUE_SUBGROUP_SIZE
:
2043 return nir_intrinsic_load_subgroup_size
;
2044 case SYSTEM_VALUE_SUBGROUP_INVOCATION
:
2045 return nir_intrinsic_load_subgroup_invocation
;
2046 case SYSTEM_VALUE_SUBGROUP_EQ_MASK
:
2047 return nir_intrinsic_load_subgroup_eq_mask
;
2048 case SYSTEM_VALUE_SUBGROUP_GE_MASK
:
2049 return nir_intrinsic_load_subgroup_ge_mask
;
2050 case SYSTEM_VALUE_SUBGROUP_GT_MASK
:
2051 return nir_intrinsic_load_subgroup_gt_mask
;
2052 case SYSTEM_VALUE_SUBGROUP_LE_MASK
:
2053 return nir_intrinsic_load_subgroup_le_mask
;
2054 case SYSTEM_VALUE_SUBGROUP_LT_MASK
:
2055 return nir_intrinsic_load_subgroup_lt_mask
;
2056 case SYSTEM_VALUE_NUM_SUBGROUPS
:
2057 return nir_intrinsic_load_num_subgroups
;
2058 case SYSTEM_VALUE_SUBGROUP_ID
:
2059 return nir_intrinsic_load_subgroup_id
;
2060 case SYSTEM_VALUE_LOCAL_GROUP_SIZE
:
2061 return nir_intrinsic_load_local_group_size
;
2062 case SYSTEM_VALUE_GLOBAL_INVOCATION_ID
:
2063 return nir_intrinsic_load_global_invocation_id
;
2064 case SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX
:
2065 return nir_intrinsic_load_global_invocation_index
;
2066 case SYSTEM_VALUE_WORK_DIM
:
2067 return nir_intrinsic_load_work_dim
;
2068 case SYSTEM_VALUE_USER_DATA_AMD
:
2069 return nir_intrinsic_load_user_data_amd
;
2071 unreachable("system value does not directly correspond to intrinsic");
2076 nir_system_value_from_intrinsic(nir_intrinsic_op intrin
)
2079 case nir_intrinsic_load_vertex_id
:
2080 return SYSTEM_VALUE_VERTEX_ID
;
2081 case nir_intrinsic_load_instance_id
:
2082 return SYSTEM_VALUE_INSTANCE_ID
;
2083 case nir_intrinsic_load_draw_id
:
2084 return SYSTEM_VALUE_DRAW_ID
;
2085 case nir_intrinsic_load_base_instance
:
2086 return SYSTEM_VALUE_BASE_INSTANCE
;
2087 case nir_intrinsic_load_vertex_id_zero_base
:
2088 return SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
2089 case nir_intrinsic_load_first_vertex
:
2090 return SYSTEM_VALUE_FIRST_VERTEX
;
2091 case nir_intrinsic_load_is_indexed_draw
:
2092 return SYSTEM_VALUE_IS_INDEXED_DRAW
;
2093 case nir_intrinsic_load_base_vertex
:
2094 return SYSTEM_VALUE_BASE_VERTEX
;
2095 case nir_intrinsic_load_invocation_id
:
2096 return SYSTEM_VALUE_INVOCATION_ID
;
2097 case nir_intrinsic_load_frag_coord
:
2098 return SYSTEM_VALUE_FRAG_COORD
;
2099 case nir_intrinsic_load_point_coord
:
2100 return SYSTEM_VALUE_POINT_COORD
;
2101 case nir_intrinsic_load_front_face
:
2102 return SYSTEM_VALUE_FRONT_FACE
;
2103 case nir_intrinsic_load_sample_id
:
2104 return SYSTEM_VALUE_SAMPLE_ID
;
2105 case nir_intrinsic_load_sample_pos
:
2106 return SYSTEM_VALUE_SAMPLE_POS
;
2107 case nir_intrinsic_load_sample_mask_in
:
2108 return SYSTEM_VALUE_SAMPLE_MASK_IN
;
2109 case nir_intrinsic_load_local_invocation_id
:
2110 return SYSTEM_VALUE_LOCAL_INVOCATION_ID
;
2111 case nir_intrinsic_load_local_invocation_index
:
2112 return SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
;
2113 case nir_intrinsic_load_num_work_groups
:
2114 return SYSTEM_VALUE_NUM_WORK_GROUPS
;
2115 case nir_intrinsic_load_work_group_id
:
2116 return SYSTEM_VALUE_WORK_GROUP_ID
;
2117 case nir_intrinsic_load_primitive_id
:
2118 return SYSTEM_VALUE_PRIMITIVE_ID
;
2119 case nir_intrinsic_load_tess_coord
:
2120 return SYSTEM_VALUE_TESS_COORD
;
2121 case nir_intrinsic_load_tess_level_outer
:
2122 return SYSTEM_VALUE_TESS_LEVEL_OUTER
;
2123 case nir_intrinsic_load_tess_level_inner
:
2124 return SYSTEM_VALUE_TESS_LEVEL_INNER
;
2125 case nir_intrinsic_load_tess_level_outer_default
:
2126 return SYSTEM_VALUE_TESS_LEVEL_OUTER_DEFAULT
;
2127 case nir_intrinsic_load_tess_level_inner_default
:
2128 return SYSTEM_VALUE_TESS_LEVEL_INNER_DEFAULT
;
2129 case nir_intrinsic_load_patch_vertices_in
:
2130 return SYSTEM_VALUE_VERTICES_IN
;
2131 case nir_intrinsic_load_helper_invocation
:
2132 return SYSTEM_VALUE_HELPER_INVOCATION
;
2133 case nir_intrinsic_load_color0
:
2134 return SYSTEM_VALUE_COLOR0
;
2135 case nir_intrinsic_load_color1
:
2136 return SYSTEM_VALUE_COLOR1
;
2137 case nir_intrinsic_load_view_index
:
2138 return SYSTEM_VALUE_VIEW_INDEX
;
2139 case nir_intrinsic_load_subgroup_size
:
2140 return SYSTEM_VALUE_SUBGROUP_SIZE
;
2141 case nir_intrinsic_load_subgroup_invocation
:
2142 return SYSTEM_VALUE_SUBGROUP_INVOCATION
;
2143 case nir_intrinsic_load_subgroup_eq_mask
:
2144 return SYSTEM_VALUE_SUBGROUP_EQ_MASK
;
2145 case nir_intrinsic_load_subgroup_ge_mask
:
2146 return SYSTEM_VALUE_SUBGROUP_GE_MASK
;
2147 case nir_intrinsic_load_subgroup_gt_mask
:
2148 return SYSTEM_VALUE_SUBGROUP_GT_MASK
;
2149 case nir_intrinsic_load_subgroup_le_mask
:
2150 return SYSTEM_VALUE_SUBGROUP_LE_MASK
;
2151 case nir_intrinsic_load_subgroup_lt_mask
:
2152 return SYSTEM_VALUE_SUBGROUP_LT_MASK
;
2153 case nir_intrinsic_load_num_subgroups
:
2154 return SYSTEM_VALUE_NUM_SUBGROUPS
;
2155 case nir_intrinsic_load_subgroup_id
:
2156 return SYSTEM_VALUE_SUBGROUP_ID
;
2157 case nir_intrinsic_load_local_group_size
:
2158 return SYSTEM_VALUE_LOCAL_GROUP_SIZE
;
2159 case nir_intrinsic_load_global_invocation_id
:
2160 return SYSTEM_VALUE_GLOBAL_INVOCATION_ID
;
2161 case nir_intrinsic_load_user_data_amd
:
2162 return SYSTEM_VALUE_USER_DATA_AMD
;
2164 unreachable("intrinsic doesn't produce a system value");
2168 /* OpenGL utility method that remaps the location attributes if they are
2169 * doubles. Not needed for vulkan due the differences on the input location
2170 * count for doubles on vulkan vs OpenGL
2172 * The bitfield returned in dual_slot is one bit for each double input slot in
2173 * the original OpenGL single-slot input numbering. The mapping from old
2174 * locations to new locations is as follows:
2176 * new_loc = loc + util_bitcount(dual_slot & BITFIELD64_MASK(loc))
2179 nir_remap_dual_slot_attributes(nir_shader
*shader
, uint64_t *dual_slot
)
2181 assert(shader
->info
.stage
== MESA_SHADER_VERTEX
);
2184 nir_foreach_variable(var
, &shader
->inputs
) {
2185 if (glsl_type_is_dual_slot(glsl_without_array(var
->type
))) {
2186 unsigned slots
= glsl_count_attribute_slots(var
->type
, true);
2187 *dual_slot
|= BITFIELD64_MASK(slots
) << var
->data
.location
;
2191 nir_foreach_variable(var
, &shader
->inputs
) {
2192 var
->data
.location
+=
2193 util_bitcount64(*dual_slot
& BITFIELD64_MASK(var
->data
.location
));
2197 /* Returns an attribute mask that has been re-compacted using the given
2201 nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
)
2204 unsigned loc
= u_bit_scan64(&dual_slot
);
2205 /* mask of all bits up to and including loc */
2206 uint64_t mask
= BITFIELD64_MASK(loc
+ 1);
2207 attribs
= (attribs
& mask
) | ((attribs
& ~mask
) >> 1);
2213 nir_rewrite_image_intrinsic(nir_intrinsic_instr
*intrin
, nir_ssa_def
*src
,
2216 enum gl_access_qualifier access
= nir_intrinsic_access(intrin
);
2218 switch (intrin
->intrinsic
) {
2220 case nir_intrinsic_image_deref_##op: \
2221 intrin->intrinsic = bindless ? nir_intrinsic_bindless_image_##op \
2222 : nir_intrinsic_image_##op; \
2234 CASE(atomic_exchange
)
2235 CASE(atomic_comp_swap
)
2239 CASE(load_raw_intel
)
2240 CASE(store_raw_intel
)
2243 unreachable("Unhanded image intrinsic");
2246 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
2247 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
2249 nir_intrinsic_set_image_dim(intrin
, glsl_get_sampler_dim(deref
->type
));
2250 nir_intrinsic_set_image_array(intrin
, glsl_sampler_type_is_array(deref
->type
));
2251 nir_intrinsic_set_access(intrin
, access
| var
->data
.access
);
2252 nir_intrinsic_set_format(intrin
, var
->data
.image
.format
);
2254 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[0],
2255 nir_src_for_ssa(src
));