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
->variables
);
49 shader
->options
= options
;
52 assert(si
->stage
== stage
);
55 shader
->info
.stage
= stage
;
58 exec_list_make_empty(&shader
->functions
);
60 shader
->num_inputs
= 0;
61 shader
->num_outputs
= 0;
62 shader
->num_uniforms
= 0;
63 shader
->num_shared
= 0;
69 reg_create(void *mem_ctx
, struct exec_list
*list
)
71 nir_register
*reg
= ralloc(mem_ctx
, nir_register
);
73 list_inithead(®
->uses
);
74 list_inithead(®
->defs
);
75 list_inithead(®
->if_uses
);
77 reg
->num_components
= 0;
79 reg
->num_array_elems
= 0;
82 exec_list_push_tail(list
, ®
->node
);
88 nir_local_reg_create(nir_function_impl
*impl
)
90 nir_register
*reg
= reg_create(ralloc_parent(impl
), &impl
->registers
);
91 reg
->index
= impl
->reg_alloc
++;
97 nir_reg_remove(nir_register
*reg
)
99 exec_node_remove(®
->node
);
103 nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
)
105 switch (var
->data
.mode
) {
106 case nir_var_function_temp
:
107 assert(!"nir_shader_add_variable cannot be used for local variables");
110 case nir_var_shader_temp
:
111 case nir_var_shader_in
:
112 case nir_var_shader_out
:
113 case nir_var_uniform
:
114 case nir_var_mem_ubo
:
115 case nir_var_mem_ssbo
:
116 case nir_var_mem_shared
:
117 case nir_var_system_value
:
120 case nir_var_mem_global
:
121 assert(!"nir_shader_add_variable cannot be used for global memory");
124 case nir_var_mem_push_const
:
125 assert(!"nir_var_push_constant is not supposed to be used for variables");
129 assert(!"invalid mode");
133 exec_list_push_tail(&shader
->variables
, &var
->node
);
137 nir_variable_create(nir_shader
*shader
, nir_variable_mode mode
,
138 const struct glsl_type
*type
, const char *name
)
140 nir_variable
*var
= rzalloc(shader
, nir_variable
);
141 var
->name
= ralloc_strdup(var
, name
);
143 var
->data
.mode
= mode
;
144 var
->data
.how_declared
= nir_var_declared_normally
;
146 if ((mode
== nir_var_shader_in
&&
147 shader
->info
.stage
!= MESA_SHADER_VERTEX
) ||
148 (mode
== nir_var_shader_out
&&
149 shader
->info
.stage
!= MESA_SHADER_FRAGMENT
))
150 var
->data
.interpolation
= INTERP_MODE_SMOOTH
;
152 if (mode
== nir_var_shader_in
|| mode
== nir_var_uniform
)
153 var
->data
.read_only
= true;
155 nir_shader_add_variable(shader
, var
);
161 nir_local_variable_create(nir_function_impl
*impl
,
162 const struct glsl_type
*type
, const char *name
)
164 nir_variable
*var
= rzalloc(impl
->function
->shader
, nir_variable
);
165 var
->name
= ralloc_strdup(var
, name
);
167 var
->data
.mode
= nir_var_function_temp
;
169 nir_function_impl_add_variable(impl
, var
);
175 nir_find_variable_with_location(nir_shader
*shader
,
176 nir_variable_mode mode
,
179 assert(util_bitcount(mode
) == 1 && mode
!= nir_var_function_temp
);
180 nir_foreach_variable_with_modes(var
, shader
, mode
) {
181 if (var
->data
.location
== location
)
188 nir_find_variable_with_driver_location(nir_shader
*shader
,
189 nir_variable_mode mode
,
192 assert(util_bitcount(mode
) == 1 && mode
!= nir_var_function_temp
);
193 nir_foreach_variable_with_modes(var
, shader
, mode
) {
194 if (var
->data
.driver_location
== location
)
201 nir_function_create(nir_shader
*shader
, const char *name
)
203 nir_function
*func
= ralloc(shader
, nir_function
);
205 exec_list_push_tail(&shader
->functions
, &func
->node
);
207 func
->name
= ralloc_strdup(func
, name
);
208 func
->shader
= shader
;
209 func
->num_params
= 0;
212 func
->is_entrypoint
= false;
217 /* NOTE: if the instruction you are copying a src to is already added
218 * to the IR, use nir_instr_rewrite_src() instead.
220 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *mem_ctx
)
222 dest
->is_ssa
= src
->is_ssa
;
224 dest
->ssa
= src
->ssa
;
226 dest
->reg
.base_offset
= src
->reg
.base_offset
;
227 dest
->reg
.reg
= src
->reg
.reg
;
228 if (src
->reg
.indirect
) {
229 dest
->reg
.indirect
= ralloc(mem_ctx
, nir_src
);
230 nir_src_copy(dest
->reg
.indirect
, src
->reg
.indirect
, mem_ctx
);
232 dest
->reg
.indirect
= NULL
;
237 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
)
239 /* Copying an SSA definition makes no sense whatsoever. */
240 assert(!src
->is_ssa
);
242 dest
->is_ssa
= false;
244 dest
->reg
.base_offset
= src
->reg
.base_offset
;
245 dest
->reg
.reg
= src
->reg
.reg
;
246 if (src
->reg
.indirect
) {
247 dest
->reg
.indirect
= ralloc(instr
, nir_src
);
248 nir_src_copy(dest
->reg
.indirect
, src
->reg
.indirect
, instr
);
250 dest
->reg
.indirect
= NULL
;
255 nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
256 nir_alu_instr
*instr
)
258 nir_src_copy(&dest
->src
, &src
->src
, &instr
->instr
);
259 dest
->abs
= src
->abs
;
260 dest
->negate
= src
->negate
;
261 for (unsigned i
= 0; i
< NIR_MAX_VEC_COMPONENTS
; i
++)
262 dest
->swizzle
[i
] = src
->swizzle
[i
];
266 nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
267 nir_alu_instr
*instr
)
269 nir_dest_copy(&dest
->dest
, &src
->dest
, &instr
->instr
);
270 dest
->write_mask
= src
->write_mask
;
271 dest
->saturate
= src
->saturate
;
276 cf_init(nir_cf_node
*node
, nir_cf_node_type type
)
278 exec_node_init(&node
->node
);
284 nir_function_impl_create_bare(nir_shader
*shader
)
286 nir_function_impl
*impl
= ralloc(shader
, nir_function_impl
);
288 impl
->function
= NULL
;
290 cf_init(&impl
->cf_node
, nir_cf_node_function
);
292 exec_list_make_empty(&impl
->body
);
293 exec_list_make_empty(&impl
->registers
);
294 exec_list_make_empty(&impl
->locals
);
297 impl
->valid_metadata
= nir_metadata_none
;
298 impl
->structured
= true;
300 /* create start & end blocks */
301 nir_block
*start_block
= nir_block_create(shader
);
302 nir_block
*end_block
= nir_block_create(shader
);
303 start_block
->cf_node
.parent
= &impl
->cf_node
;
304 end_block
->cf_node
.parent
= &impl
->cf_node
;
305 impl
->end_block
= end_block
;
307 exec_list_push_tail(&impl
->body
, &start_block
->cf_node
.node
);
309 start_block
->successors
[0] = end_block
;
310 _mesa_set_add(end_block
->predecessors
, start_block
);
315 nir_function_impl_create(nir_function
*function
)
317 assert(function
->impl
== NULL
);
319 nir_function_impl
*impl
= nir_function_impl_create_bare(function
->shader
);
321 function
->impl
= impl
;
322 impl
->function
= function
;
328 nir_block_create(nir_shader
*shader
)
330 nir_block
*block
= rzalloc(shader
, nir_block
);
332 cf_init(&block
->cf_node
, nir_cf_node_block
);
334 block
->successors
[0] = block
->successors
[1] = NULL
;
335 block
->predecessors
= _mesa_pointer_set_create(block
);
336 block
->imm_dom
= NULL
;
337 /* XXX maybe it would be worth it to defer allocation? This
338 * way it doesn't get allocated for shader refs that never run
339 * nir_calc_dominance? For example, state-tracker creates an
340 * initial IR, clones that, runs appropriate lowering pass, passes
341 * to driver which does common lowering/opt, and then stores ref
342 * which is later used to do state specific lowering and futher
343 * opt. Do any of the references not need dominance metadata?
345 block
->dom_frontier
= _mesa_pointer_set_create(block
);
347 exec_list_make_empty(&block
->instr_list
);
353 src_init(nir_src
*src
)
357 src
->reg
.indirect
= NULL
;
358 src
->reg
.base_offset
= 0;
362 nir_if_create(nir_shader
*shader
)
364 nir_if
*if_stmt
= ralloc(shader
, nir_if
);
366 if_stmt
->control
= nir_selection_control_none
;
368 cf_init(&if_stmt
->cf_node
, nir_cf_node_if
);
369 src_init(&if_stmt
->condition
);
371 nir_block
*then
= nir_block_create(shader
);
372 exec_list_make_empty(&if_stmt
->then_list
);
373 exec_list_push_tail(&if_stmt
->then_list
, &then
->cf_node
.node
);
374 then
->cf_node
.parent
= &if_stmt
->cf_node
;
376 nir_block
*else_stmt
= nir_block_create(shader
);
377 exec_list_make_empty(&if_stmt
->else_list
);
378 exec_list_push_tail(&if_stmt
->else_list
, &else_stmt
->cf_node
.node
);
379 else_stmt
->cf_node
.parent
= &if_stmt
->cf_node
;
385 nir_loop_create(nir_shader
*shader
)
387 nir_loop
*loop
= rzalloc(shader
, nir_loop
);
389 cf_init(&loop
->cf_node
, nir_cf_node_loop
);
391 nir_block
*body
= nir_block_create(shader
);
392 exec_list_make_empty(&loop
->body
);
393 exec_list_push_tail(&loop
->body
, &body
->cf_node
.node
);
394 body
->cf_node
.parent
= &loop
->cf_node
;
396 body
->successors
[0] = body
;
397 _mesa_set_add(body
->predecessors
, body
);
403 instr_init(nir_instr
*instr
, nir_instr_type type
)
407 exec_node_init(&instr
->node
);
411 dest_init(nir_dest
*dest
)
413 dest
->is_ssa
= false;
414 dest
->reg
.reg
= NULL
;
415 dest
->reg
.indirect
= NULL
;
416 dest
->reg
.base_offset
= 0;
420 alu_dest_init(nir_alu_dest
*dest
)
422 dest_init(&dest
->dest
);
423 dest
->saturate
= false;
424 dest
->write_mask
= 0xf;
428 alu_src_init(nir_alu_src
*src
)
431 src
->abs
= src
->negate
= false;
432 for (int i
= 0; i
< NIR_MAX_VEC_COMPONENTS
; ++i
)
437 nir_alu_instr_create(nir_shader
*shader
, nir_op op
)
439 unsigned num_srcs
= nir_op_infos
[op
].num_inputs
;
440 /* TODO: don't use rzalloc */
441 nir_alu_instr
*instr
=
443 sizeof(nir_alu_instr
) + num_srcs
* sizeof(nir_alu_src
));
445 instr_init(&instr
->instr
, nir_instr_type_alu
);
447 alu_dest_init(&instr
->dest
);
448 for (unsigned i
= 0; i
< num_srcs
; i
++)
449 alu_src_init(&instr
->src
[i
]);
455 nir_deref_instr_create(nir_shader
*shader
, nir_deref_type deref_type
)
457 nir_deref_instr
*instr
=
458 rzalloc_size(shader
, sizeof(nir_deref_instr
));
460 instr_init(&instr
->instr
, nir_instr_type_deref
);
462 instr
->deref_type
= deref_type
;
463 if (deref_type
!= nir_deref_type_var
)
464 src_init(&instr
->parent
);
466 if (deref_type
== nir_deref_type_array
||
467 deref_type
== nir_deref_type_ptr_as_array
)
468 src_init(&instr
->arr
.index
);
470 dest_init(&instr
->dest
);
476 nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
)
478 nir_jump_instr
*instr
= ralloc(shader
, nir_jump_instr
);
479 instr_init(&instr
->instr
, nir_instr_type_jump
);
484 nir_load_const_instr
*
485 nir_load_const_instr_create(nir_shader
*shader
, unsigned num_components
,
488 nir_load_const_instr
*instr
=
489 rzalloc_size(shader
, sizeof(*instr
) + num_components
* sizeof(*instr
->value
));
490 instr_init(&instr
->instr
, nir_instr_type_load_const
);
492 nir_ssa_def_init(&instr
->instr
, &instr
->def
, num_components
, bit_size
, NULL
);
497 nir_intrinsic_instr
*
498 nir_intrinsic_instr_create(nir_shader
*shader
, nir_intrinsic_op op
)
500 unsigned num_srcs
= nir_intrinsic_infos
[op
].num_srcs
;
501 /* TODO: don't use rzalloc */
502 nir_intrinsic_instr
*instr
=
504 sizeof(nir_intrinsic_instr
) + num_srcs
* sizeof(nir_src
));
506 instr_init(&instr
->instr
, nir_instr_type_intrinsic
);
507 instr
->intrinsic
= op
;
509 if (nir_intrinsic_infos
[op
].has_dest
)
510 dest_init(&instr
->dest
);
512 for (unsigned i
= 0; i
< num_srcs
; i
++)
513 src_init(&instr
->src
[i
]);
519 nir_call_instr_create(nir_shader
*shader
, nir_function
*callee
)
521 const unsigned num_params
= callee
->num_params
;
522 nir_call_instr
*instr
=
523 rzalloc_size(shader
, sizeof(*instr
) +
524 num_params
* sizeof(instr
->params
[0]));
526 instr_init(&instr
->instr
, nir_instr_type_call
);
527 instr
->callee
= callee
;
528 instr
->num_params
= num_params
;
529 for (unsigned i
= 0; i
< num_params
; i
++)
530 src_init(&instr
->params
[i
]);
535 static int8_t default_tg4_offsets
[4][2] =
544 nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
)
546 nir_tex_instr
*instr
= rzalloc(shader
, nir_tex_instr
);
547 instr_init(&instr
->instr
, nir_instr_type_tex
);
549 dest_init(&instr
->dest
);
551 instr
->num_srcs
= num_srcs
;
552 instr
->src
= ralloc_array(instr
, nir_tex_src
, num_srcs
);
553 for (unsigned i
= 0; i
< num_srcs
; i
++)
554 src_init(&instr
->src
[i
].src
);
556 instr
->texture_index
= 0;
557 instr
->sampler_index
= 0;
558 memcpy(instr
->tg4_offsets
, default_tg4_offsets
, sizeof(instr
->tg4_offsets
));
564 nir_tex_instr_add_src(nir_tex_instr
*tex
,
565 nir_tex_src_type src_type
,
568 nir_tex_src
*new_srcs
= rzalloc_array(tex
, nir_tex_src
,
571 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
572 new_srcs
[i
].src_type
= tex
->src
[i
].src_type
;
573 nir_instr_move_src(&tex
->instr
, &new_srcs
[i
].src
,
577 ralloc_free(tex
->src
);
580 tex
->src
[tex
->num_srcs
].src_type
= src_type
;
581 nir_instr_rewrite_src(&tex
->instr
, &tex
->src
[tex
->num_srcs
].src
, src
);
586 nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
)
588 assert(src_idx
< tex
->num_srcs
);
590 /* First rewrite the source to NIR_SRC_INIT */
591 nir_instr_rewrite_src(&tex
->instr
, &tex
->src
[src_idx
].src
, NIR_SRC_INIT
);
593 /* Now, move all of the other sources down */
594 for (unsigned i
= src_idx
+ 1; i
< tex
->num_srcs
; i
++) {
595 tex
->src
[i
-1].src_type
= tex
->src
[i
].src_type
;
596 nir_instr_move_src(&tex
->instr
, &tex
->src
[i
-1].src
, &tex
->src
[i
].src
);
602 nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
)
604 if (tex
->op
!= nir_texop_tg4
)
606 return memcmp(tex
->tg4_offsets
, default_tg4_offsets
,
607 sizeof(tex
->tg4_offsets
)) != 0;
611 nir_phi_instr_create(nir_shader
*shader
)
613 nir_phi_instr
*instr
= ralloc(shader
, nir_phi_instr
);
614 instr_init(&instr
->instr
, nir_instr_type_phi
);
616 dest_init(&instr
->dest
);
617 exec_list_make_empty(&instr
->srcs
);
621 nir_parallel_copy_instr
*
622 nir_parallel_copy_instr_create(nir_shader
*shader
)
624 nir_parallel_copy_instr
*instr
= ralloc(shader
, nir_parallel_copy_instr
);
625 instr_init(&instr
->instr
, nir_instr_type_parallel_copy
);
627 exec_list_make_empty(&instr
->entries
);
632 nir_ssa_undef_instr
*
633 nir_ssa_undef_instr_create(nir_shader
*shader
,
634 unsigned num_components
,
637 nir_ssa_undef_instr
*instr
= ralloc(shader
, nir_ssa_undef_instr
);
638 instr_init(&instr
->instr
, nir_instr_type_ssa_undef
);
640 nir_ssa_def_init(&instr
->instr
, &instr
->def
, num_components
, bit_size
, NULL
);
645 static nir_const_value
646 const_value_float(double d
, unsigned bit_size
)
649 memset(&v
, 0, sizeof(v
));
651 case 16: v
.u16
= _mesa_float_to_half(d
); break;
652 case 32: v
.f32
= d
; break;
653 case 64: v
.f64
= d
; break;
655 unreachable("Invalid bit size");
660 static nir_const_value
661 const_value_int(int64_t i
, unsigned bit_size
)
664 memset(&v
, 0, sizeof(v
));
666 case 1: v
.b
= i
& 1; break;
667 case 8: v
.i8
= i
; break;
668 case 16: v
.i16
= i
; break;
669 case 32: v
.i32
= i
; break;
670 case 64: v
.i64
= i
; break;
672 unreachable("Invalid bit size");
678 nir_alu_binop_identity(nir_op binop
, unsigned bit_size
)
680 const int64_t max_int
= (1ull << (bit_size
- 1)) - 1;
681 const int64_t min_int
= -max_int
- 1;
684 return const_value_int(0, bit_size
);
686 return const_value_float(0, bit_size
);
688 return const_value_int(1, bit_size
);
690 return const_value_float(1, bit_size
);
692 return const_value_int(max_int
, bit_size
);
694 return const_value_int(~0ull, bit_size
);
696 return const_value_float(INFINITY
, bit_size
);
698 return const_value_int(min_int
, bit_size
);
700 return const_value_int(0, bit_size
);
702 return const_value_float(-INFINITY
, bit_size
);
704 return const_value_int(~0ull, bit_size
);
706 return const_value_int(0, bit_size
);
708 return const_value_int(0, bit_size
);
710 unreachable("Invalid reduction operation");
715 nir_cf_node_get_function(nir_cf_node
*node
)
717 while (node
->type
!= nir_cf_node_function
) {
721 return nir_cf_node_as_function(node
);
724 /* Reduces a cursor by trying to convert everything to after and trying to
725 * go up to block granularity when possible.
728 reduce_cursor(nir_cursor cursor
)
730 switch (cursor
.option
) {
731 case nir_cursor_before_block
:
732 assert(nir_cf_node_prev(&cursor
.block
->cf_node
) == NULL
||
733 nir_cf_node_prev(&cursor
.block
->cf_node
)->type
!= nir_cf_node_block
);
734 if (exec_list_is_empty(&cursor
.block
->instr_list
)) {
735 /* Empty block. After is as good as before. */
736 cursor
.option
= nir_cursor_after_block
;
740 case nir_cursor_after_block
:
743 case nir_cursor_before_instr
: {
744 nir_instr
*prev_instr
= nir_instr_prev(cursor
.instr
);
746 /* Before this instruction is after the previous */
747 cursor
.instr
= prev_instr
;
748 cursor
.option
= nir_cursor_after_instr
;
750 /* No previous instruction. Switch to before block */
751 cursor
.block
= cursor
.instr
->block
;
752 cursor
.option
= nir_cursor_before_block
;
754 return reduce_cursor(cursor
);
757 case nir_cursor_after_instr
:
758 if (nir_instr_next(cursor
.instr
) == NULL
) {
759 /* This is the last instruction, switch to after block */
760 cursor
.option
= nir_cursor_after_block
;
761 cursor
.block
= cursor
.instr
->block
;
766 unreachable("Inavlid cursor option");
771 nir_cursors_equal(nir_cursor a
, nir_cursor b
)
773 /* Reduced cursors should be unique */
774 a
= reduce_cursor(a
);
775 b
= reduce_cursor(b
);
777 return a
.block
== b
.block
&& a
.option
== b
.option
;
781 add_use_cb(nir_src
*src
, void *state
)
783 nir_instr
*instr
= state
;
785 src
->parent_instr
= instr
;
786 list_addtail(&src
->use_link
,
787 src
->is_ssa
? &src
->ssa
->uses
: &src
->reg
.reg
->uses
);
793 add_ssa_def_cb(nir_ssa_def
*def
, void *state
)
795 nir_instr
*instr
= state
;
797 if (instr
->block
&& def
->index
== UINT_MAX
) {
798 nir_function_impl
*impl
=
799 nir_cf_node_get_function(&instr
->block
->cf_node
);
801 def
->index
= impl
->ssa_alloc
++;
808 add_reg_def_cb(nir_dest
*dest
, void *state
)
810 nir_instr
*instr
= state
;
813 dest
->reg
.parent_instr
= instr
;
814 list_addtail(&dest
->reg
.def_link
, &dest
->reg
.reg
->defs
);
821 add_defs_uses(nir_instr
*instr
)
823 nir_foreach_src(instr
, add_use_cb
, instr
);
824 nir_foreach_dest(instr
, add_reg_def_cb
, instr
);
825 nir_foreach_ssa_def(instr
, add_ssa_def_cb
, instr
);
829 nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
)
831 switch (cursor
.option
) {
832 case nir_cursor_before_block
:
833 /* Only allow inserting jumps into empty blocks. */
834 if (instr
->type
== nir_instr_type_jump
)
835 assert(exec_list_is_empty(&cursor
.block
->instr_list
));
837 instr
->block
= cursor
.block
;
838 add_defs_uses(instr
);
839 exec_list_push_head(&cursor
.block
->instr_list
, &instr
->node
);
841 case nir_cursor_after_block
: {
842 /* Inserting instructions after a jump is illegal. */
843 nir_instr
*last
= nir_block_last_instr(cursor
.block
);
844 assert(last
== NULL
|| last
->type
!= nir_instr_type_jump
);
847 instr
->block
= cursor
.block
;
848 add_defs_uses(instr
);
849 exec_list_push_tail(&cursor
.block
->instr_list
, &instr
->node
);
852 case nir_cursor_before_instr
:
853 assert(instr
->type
!= nir_instr_type_jump
);
854 instr
->block
= cursor
.instr
->block
;
855 add_defs_uses(instr
);
856 exec_node_insert_node_before(&cursor
.instr
->node
, &instr
->node
);
858 case nir_cursor_after_instr
:
859 /* Inserting instructions after a jump is illegal. */
860 assert(cursor
.instr
->type
!= nir_instr_type_jump
);
862 /* Only allow inserting jumps at the end of the block. */
863 if (instr
->type
== nir_instr_type_jump
)
864 assert(cursor
.instr
== nir_block_last_instr(cursor
.instr
->block
));
866 instr
->block
= cursor
.instr
->block
;
867 add_defs_uses(instr
);
868 exec_node_insert_after(&cursor
.instr
->node
, &instr
->node
);
872 if (instr
->type
== nir_instr_type_jump
)
873 nir_handle_add_jump(instr
->block
);
877 src_is_valid(const nir_src
*src
)
879 return src
->is_ssa
? (src
->ssa
!= NULL
) : (src
->reg
.reg
!= NULL
);
883 remove_use_cb(nir_src
*src
, void *state
)
887 if (src_is_valid(src
))
888 list_del(&src
->use_link
);
894 remove_def_cb(nir_dest
*dest
, void *state
)
899 list_del(&dest
->reg
.def_link
);
905 remove_defs_uses(nir_instr
*instr
)
907 nir_foreach_dest(instr
, remove_def_cb
, instr
);
908 nir_foreach_src(instr
, remove_use_cb
, instr
);
911 void nir_instr_remove_v(nir_instr
*instr
)
913 remove_defs_uses(instr
);
914 exec_node_remove(&instr
->node
);
916 if (instr
->type
== nir_instr_type_jump
) {
917 nir_jump_instr
*jump_instr
= nir_instr_as_jump(instr
);
918 nir_handle_remove_jump(instr
->block
, jump_instr
->type
);
925 nir_index_local_regs(nir_function_impl
*impl
)
928 foreach_list_typed(nir_register
, reg
, node
, &impl
->registers
) {
929 reg
->index
= index
++;
931 impl
->reg_alloc
= index
;
935 visit_alu_dest(nir_alu_instr
*instr
, nir_foreach_dest_cb cb
, void *state
)
937 return cb(&instr
->dest
.dest
, state
);
941 visit_deref_dest(nir_deref_instr
*instr
, nir_foreach_dest_cb cb
, void *state
)
943 return cb(&instr
->dest
, state
);
947 visit_intrinsic_dest(nir_intrinsic_instr
*instr
, nir_foreach_dest_cb cb
,
950 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
951 return cb(&instr
->dest
, state
);
957 visit_texture_dest(nir_tex_instr
*instr
, nir_foreach_dest_cb cb
,
960 return cb(&instr
->dest
, state
);
964 visit_phi_dest(nir_phi_instr
*instr
, nir_foreach_dest_cb cb
, void *state
)
966 return cb(&instr
->dest
, state
);
970 visit_parallel_copy_dest(nir_parallel_copy_instr
*instr
,
971 nir_foreach_dest_cb cb
, void *state
)
973 nir_foreach_parallel_copy_entry(entry
, instr
) {
974 if (!cb(&entry
->dest
, state
))
982 nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
)
984 switch (instr
->type
) {
985 case nir_instr_type_alu
:
986 return visit_alu_dest(nir_instr_as_alu(instr
), cb
, state
);
987 case nir_instr_type_deref
:
988 return visit_deref_dest(nir_instr_as_deref(instr
), cb
, state
);
989 case nir_instr_type_intrinsic
:
990 return visit_intrinsic_dest(nir_instr_as_intrinsic(instr
), cb
, state
);
991 case nir_instr_type_tex
:
992 return visit_texture_dest(nir_instr_as_tex(instr
), cb
, state
);
993 case nir_instr_type_phi
:
994 return visit_phi_dest(nir_instr_as_phi(instr
), cb
, state
);
995 case nir_instr_type_parallel_copy
:
996 return visit_parallel_copy_dest(nir_instr_as_parallel_copy(instr
),
999 case nir_instr_type_load_const
:
1000 case nir_instr_type_ssa_undef
:
1001 case nir_instr_type_call
:
1002 case nir_instr_type_jump
:
1006 unreachable("Invalid instruction type");
1013 struct foreach_ssa_def_state
{
1014 nir_foreach_ssa_def_cb cb
;
1019 nir_ssa_def_visitor(nir_dest
*dest
, void *void_state
)
1021 struct foreach_ssa_def_state
*state
= void_state
;
1024 return state
->cb(&dest
->ssa
, state
->client_state
);
1030 nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
, void *state
)
1032 switch (instr
->type
) {
1033 case nir_instr_type_alu
:
1034 case nir_instr_type_deref
:
1035 case nir_instr_type_tex
:
1036 case nir_instr_type_intrinsic
:
1037 case nir_instr_type_phi
:
1038 case nir_instr_type_parallel_copy
: {
1039 struct foreach_ssa_def_state foreach_state
= {cb
, state
};
1040 return nir_foreach_dest(instr
, nir_ssa_def_visitor
, &foreach_state
);
1043 case nir_instr_type_load_const
:
1044 return cb(&nir_instr_as_load_const(instr
)->def
, state
);
1045 case nir_instr_type_ssa_undef
:
1046 return cb(&nir_instr_as_ssa_undef(instr
)->def
, state
);
1047 case nir_instr_type_call
:
1048 case nir_instr_type_jump
:
1051 unreachable("Invalid instruction type");
1056 nir_instr_ssa_def(nir_instr
*instr
)
1058 switch (instr
->type
) {
1059 case nir_instr_type_alu
:
1060 assert(nir_instr_as_alu(instr
)->dest
.dest
.is_ssa
);
1061 return &nir_instr_as_alu(instr
)->dest
.dest
.ssa
;
1063 case nir_instr_type_deref
:
1064 assert(nir_instr_as_deref(instr
)->dest
.is_ssa
);
1065 return &nir_instr_as_deref(instr
)->dest
.ssa
;
1067 case nir_instr_type_tex
:
1068 assert(nir_instr_as_tex(instr
)->dest
.is_ssa
);
1069 return &nir_instr_as_tex(instr
)->dest
.ssa
;
1071 case nir_instr_type_intrinsic
: {
1072 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1073 if (nir_intrinsic_infos
[intrin
->intrinsic
].has_dest
) {
1074 assert(intrin
->dest
.is_ssa
);
1075 return &intrin
->dest
.ssa
;
1081 case nir_instr_type_phi
:
1082 assert(nir_instr_as_phi(instr
)->dest
.is_ssa
);
1083 return &nir_instr_as_phi(instr
)->dest
.ssa
;
1085 case nir_instr_type_parallel_copy
:
1086 unreachable("Parallel copies are unsupported by this function");
1088 case nir_instr_type_load_const
:
1089 return &nir_instr_as_load_const(instr
)->def
;
1091 case nir_instr_type_ssa_undef
:
1092 return &nir_instr_as_ssa_undef(instr
)->def
;
1094 case nir_instr_type_call
:
1095 case nir_instr_type_jump
:
1099 unreachable("Invalid instruction type");
1103 visit_src(nir_src
*src
, nir_foreach_src_cb cb
, void *state
)
1105 if (!cb(src
, state
))
1107 if (!src
->is_ssa
&& src
->reg
.indirect
)
1108 return cb(src
->reg
.indirect
, state
);
1113 visit_alu_src(nir_alu_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1115 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
1116 if (!visit_src(&instr
->src
[i
].src
, cb
, state
))
1123 visit_deref_instr_src(nir_deref_instr
*instr
,
1124 nir_foreach_src_cb cb
, void *state
)
1126 if (instr
->deref_type
!= nir_deref_type_var
) {
1127 if (!visit_src(&instr
->parent
, cb
, state
))
1131 if (instr
->deref_type
== nir_deref_type_array
||
1132 instr
->deref_type
== nir_deref_type_ptr_as_array
) {
1133 if (!visit_src(&instr
->arr
.index
, cb
, state
))
1141 visit_tex_src(nir_tex_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1143 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
1144 if (!visit_src(&instr
->src
[i
].src
, cb
, state
))
1152 visit_intrinsic_src(nir_intrinsic_instr
*instr
, nir_foreach_src_cb cb
,
1155 unsigned num_srcs
= nir_intrinsic_infos
[instr
->intrinsic
].num_srcs
;
1156 for (unsigned i
= 0; i
< num_srcs
; i
++) {
1157 if (!visit_src(&instr
->src
[i
], cb
, state
))
1165 visit_call_src(nir_call_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1167 for (unsigned i
= 0; i
< instr
->num_params
; i
++) {
1168 if (!visit_src(&instr
->params
[i
], cb
, state
))
1176 visit_phi_src(nir_phi_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1178 nir_foreach_phi_src(src
, instr
) {
1179 if (!visit_src(&src
->src
, cb
, state
))
1187 visit_parallel_copy_src(nir_parallel_copy_instr
*instr
,
1188 nir_foreach_src_cb cb
, void *state
)
1190 nir_foreach_parallel_copy_entry(entry
, instr
) {
1191 if (!visit_src(&entry
->src
, cb
, state
))
1200 nir_foreach_src_cb cb
;
1201 } visit_dest_indirect_state
;
1204 visit_dest_indirect(nir_dest
*dest
, void *_state
)
1206 visit_dest_indirect_state
*state
= (visit_dest_indirect_state
*) _state
;
1208 if (!dest
->is_ssa
&& dest
->reg
.indirect
)
1209 return state
->cb(dest
->reg
.indirect
, state
->state
);
1215 nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1217 switch (instr
->type
) {
1218 case nir_instr_type_alu
:
1219 if (!visit_alu_src(nir_instr_as_alu(instr
), cb
, state
))
1222 case nir_instr_type_deref
:
1223 if (!visit_deref_instr_src(nir_instr_as_deref(instr
), cb
, state
))
1226 case nir_instr_type_intrinsic
:
1227 if (!visit_intrinsic_src(nir_instr_as_intrinsic(instr
), cb
, state
))
1230 case nir_instr_type_tex
:
1231 if (!visit_tex_src(nir_instr_as_tex(instr
), cb
, state
))
1234 case nir_instr_type_call
:
1235 if (!visit_call_src(nir_instr_as_call(instr
), cb
, state
))
1238 case nir_instr_type_load_const
:
1239 /* Constant load instructions have no regular sources */
1241 case nir_instr_type_phi
:
1242 if (!visit_phi_src(nir_instr_as_phi(instr
), cb
, state
))
1245 case nir_instr_type_parallel_copy
:
1246 if (!visit_parallel_copy_src(nir_instr_as_parallel_copy(instr
),
1250 case nir_instr_type_jump
:
1251 case nir_instr_type_ssa_undef
:
1255 unreachable("Invalid instruction type");
1259 visit_dest_indirect_state dest_state
;
1260 dest_state
.state
= state
;
1262 return nir_foreach_dest(instr
, visit_dest_indirect
, &dest_state
);
1266 nir_foreach_phi_src_leaving_block(nir_block
*block
,
1267 nir_foreach_src_cb cb
,
1270 for (unsigned i
= 0; i
< ARRAY_SIZE(block
->successors
); i
++) {
1271 if (block
->successors
[i
] == NULL
)
1274 nir_foreach_instr(instr
, block
->successors
[i
]) {
1275 if (instr
->type
!= nir_instr_type_phi
)
1278 nir_phi_instr
*phi
= nir_instr_as_phi(instr
);
1279 nir_foreach_phi_src(phi_src
, phi
) {
1280 if (phi_src
->pred
== block
) {
1281 if (!cb(&phi_src
->src
, state
))
1292 nir_const_value_for_float(double f
, unsigned bit_size
)
1295 memset(&v
, 0, sizeof(v
));
1299 v
.u16
= _mesa_float_to_half(f
);
1308 unreachable("Invalid bit size");
1315 nir_const_value_as_float(nir_const_value value
, unsigned bit_size
)
1318 case 16: return _mesa_half_to_float(value
.u16
);
1319 case 32: return value
.f32
;
1320 case 64: return value
.f64
;
1322 unreachable("Invalid bit size");
1327 nir_src_as_const_value(nir_src src
)
1332 if (src
.ssa
->parent_instr
->type
!= nir_instr_type_load_const
)
1335 nir_load_const_instr
*load
= nir_instr_as_load_const(src
.ssa
->parent_instr
);
1341 * Returns true if the source is known to be dynamically uniform. Otherwise it
1342 * returns false which means it may or may not be dynamically uniform but it
1343 * can't be determined.
1346 nir_src_is_dynamically_uniform(nir_src src
)
1351 /* Constants are trivially dynamically uniform */
1352 if (src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
)
1355 /* As are uniform variables */
1356 if (src
.ssa
->parent_instr
->type
== nir_instr_type_intrinsic
) {
1357 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(src
.ssa
->parent_instr
);
1358 if (intr
->intrinsic
== nir_intrinsic_load_uniform
&&
1359 nir_src_is_dynamically_uniform(intr
->src
[0]))
1363 /* Operating together dynamically uniform expressions produces a
1364 * dynamically uniform result
1366 if (src
.ssa
->parent_instr
->type
== nir_instr_type_alu
) {
1367 nir_alu_instr
*alu
= nir_instr_as_alu(src
.ssa
->parent_instr
);
1368 for (int i
= 0; i
< nir_op_infos
[alu
->op
].num_inputs
; i
++) {
1369 if (!nir_src_is_dynamically_uniform(alu
->src
[i
].src
))
1376 /* XXX: this could have many more tests, such as when a sampler function is
1377 * called with dynamically uniform arguments.
1383 src_remove_all_uses(nir_src
*src
)
1385 for (; src
; src
= src
->is_ssa
? NULL
: src
->reg
.indirect
) {
1386 if (!src_is_valid(src
))
1389 list_del(&src
->use_link
);
1394 src_add_all_uses(nir_src
*src
, nir_instr
*parent_instr
, nir_if
*parent_if
)
1396 for (; src
; src
= src
->is_ssa
? NULL
: src
->reg
.indirect
) {
1397 if (!src_is_valid(src
))
1401 src
->parent_instr
= parent_instr
;
1403 list_addtail(&src
->use_link
, &src
->ssa
->uses
);
1405 list_addtail(&src
->use_link
, &src
->reg
.reg
->uses
);
1408 src
->parent_if
= parent_if
;
1410 list_addtail(&src
->use_link
, &src
->ssa
->if_uses
);
1412 list_addtail(&src
->use_link
, &src
->reg
.reg
->if_uses
);
1418 nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
)
1420 assert(!src_is_valid(src
) || src
->parent_instr
== instr
);
1422 src_remove_all_uses(src
);
1424 src_add_all_uses(src
, instr
, NULL
);
1428 nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
)
1430 assert(!src_is_valid(dest
) || dest
->parent_instr
== dest_instr
);
1432 src_remove_all_uses(dest
);
1433 src_remove_all_uses(src
);
1435 *src
= NIR_SRC_INIT
;
1436 src_add_all_uses(dest
, dest_instr
, NULL
);
1440 nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
)
1442 nir_src
*src
= &if_stmt
->condition
;
1443 assert(!src_is_valid(src
) || src
->parent_if
== if_stmt
);
1445 src_remove_all_uses(src
);
1447 src_add_all_uses(src
, NULL
, if_stmt
);
1451 nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
, nir_dest new_dest
)
1454 /* We can only overwrite an SSA destination if it has no uses. */
1455 assert(list_is_empty(&dest
->ssa
.uses
) && list_is_empty(&dest
->ssa
.if_uses
));
1457 list_del(&dest
->reg
.def_link
);
1458 if (dest
->reg
.indirect
)
1459 src_remove_all_uses(dest
->reg
.indirect
);
1462 /* We can't re-write with an SSA def */
1463 assert(!new_dest
.is_ssa
);
1465 nir_dest_copy(dest
, &new_dest
, instr
);
1467 dest
->reg
.parent_instr
= instr
;
1468 list_addtail(&dest
->reg
.def_link
, &new_dest
.reg
.reg
->defs
);
1470 if (dest
->reg
.indirect
)
1471 src_add_all_uses(dest
->reg
.indirect
, instr
, NULL
);
1474 /* note: does *not* take ownership of 'name' */
1476 nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
1477 unsigned num_components
,
1478 unsigned bit_size
, const char *name
)
1480 def
->name
= ralloc_strdup(instr
, name
);
1481 def
->parent_instr
= instr
;
1482 list_inithead(&def
->uses
);
1483 list_inithead(&def
->if_uses
);
1484 def
->num_components
= num_components
;
1485 def
->bit_size
= bit_size
;
1486 def
->divergent
= true; /* This is the safer default */
1489 nir_function_impl
*impl
=
1490 nir_cf_node_get_function(&instr
->block
->cf_node
);
1492 def
->index
= impl
->ssa_alloc
++;
1494 def
->index
= UINT_MAX
;
1498 /* note: does *not* take ownership of 'name' */
1500 nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
1501 unsigned num_components
, unsigned bit_size
,
1504 dest
->is_ssa
= true;
1505 nir_ssa_def_init(instr
, &dest
->ssa
, num_components
, bit_size
, name
);
1509 nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
)
1511 assert(!new_src
.is_ssa
|| def
!= new_src
.ssa
);
1513 nir_foreach_use_safe(use_src
, def
)
1514 nir_instr_rewrite_src(use_src
->parent_instr
, use_src
, new_src
);
1516 nir_foreach_if_use_safe(use_src
, def
)
1517 nir_if_rewrite_condition(use_src
->parent_if
, new_src
);
1521 is_instr_between(nir_instr
*start
, nir_instr
*end
, nir_instr
*between
)
1523 assert(start
->block
== end
->block
);
1525 if (between
->block
!= start
->block
)
1528 /* Search backwards looking for "between" */
1529 while (start
!= end
) {
1533 end
= nir_instr_prev(end
);
1540 /* Replaces all uses of the given SSA def with the given source but only if
1541 * the use comes after the after_me instruction. This can be useful if you
1542 * are emitting code to fix up the result of some instruction: you can freely
1543 * use the result in that code and then call rewrite_uses_after and pass the
1544 * last fixup instruction as after_me and it will replace all of the uses you
1545 * want without touching the fixup code.
1547 * This function assumes that after_me is in the same block as
1548 * def->parent_instr and that after_me comes after def->parent_instr.
1551 nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
1552 nir_instr
*after_me
)
1554 if (new_src
.is_ssa
&& def
== new_src
.ssa
)
1557 nir_foreach_use_safe(use_src
, def
) {
1558 assert(use_src
->parent_instr
!= def
->parent_instr
);
1559 /* Since def already dominates all of its uses, the only way a use can
1560 * not be dominated by after_me is if it is between def and after_me in
1561 * the instruction list.
1563 if (!is_instr_between(def
->parent_instr
, after_me
, use_src
->parent_instr
))
1564 nir_instr_rewrite_src(use_src
->parent_instr
, use_src
, new_src
);
1567 nir_foreach_if_use_safe(use_src
, def
)
1568 nir_if_rewrite_condition(use_src
->parent_if
, new_src
);
1571 nir_component_mask_t
1572 nir_ssa_def_components_read(const nir_ssa_def
*def
)
1574 nir_component_mask_t read_mask
= 0;
1575 nir_foreach_use(use
, def
) {
1576 if (use
->parent_instr
->type
== nir_instr_type_alu
) {
1577 nir_alu_instr
*alu
= nir_instr_as_alu(use
->parent_instr
);
1578 nir_alu_src
*alu_src
= exec_node_data(nir_alu_src
, use
, src
);
1579 int src_idx
= alu_src
- &alu
->src
[0];
1580 assert(src_idx
>= 0 && src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
1581 read_mask
|= nir_alu_instr_src_read_mask(alu
, src_idx
);
1583 return (1 << def
->num_components
) - 1;
1587 if (!list_is_empty(&def
->if_uses
))
1594 nir_block_cf_tree_next(nir_block
*block
)
1596 if (block
== NULL
) {
1597 /* nir_foreach_block_safe() will call this function on a NULL block
1598 * after the last iteration, but it won't use the result so just return
1604 assert(nir_cf_node_get_function(&block
->cf_node
)->structured
);
1606 nir_cf_node
*cf_next
= nir_cf_node_next(&block
->cf_node
);
1608 return nir_cf_node_cf_tree_first(cf_next
);
1610 nir_cf_node
*parent
= block
->cf_node
.parent
;
1612 switch (parent
->type
) {
1613 case nir_cf_node_if
: {
1614 /* Are we at the end of the if? Go to the beginning of the else */
1615 nir_if
*if_stmt
= nir_cf_node_as_if(parent
);
1616 if (block
== nir_if_last_then_block(if_stmt
))
1617 return nir_if_first_else_block(if_stmt
);
1619 assert(block
== nir_if_last_else_block(if_stmt
));
1623 case nir_cf_node_loop
:
1624 return nir_cf_node_as_block(nir_cf_node_next(parent
));
1626 case nir_cf_node_function
:
1630 unreachable("unknown cf node type");
1635 nir_block_cf_tree_prev(nir_block
*block
)
1637 if (block
== NULL
) {
1638 /* do this for consistency with nir_block_cf_tree_next() */
1642 assert(nir_cf_node_get_function(&block
->cf_node
)->structured
);
1644 nir_cf_node
*cf_prev
= nir_cf_node_prev(&block
->cf_node
);
1646 return nir_cf_node_cf_tree_last(cf_prev
);
1648 nir_cf_node
*parent
= block
->cf_node
.parent
;
1650 switch (parent
->type
) {
1651 case nir_cf_node_if
: {
1652 /* Are we at the beginning of the else? Go to the end of the if */
1653 nir_if
*if_stmt
= nir_cf_node_as_if(parent
);
1654 if (block
== nir_if_first_else_block(if_stmt
))
1655 return nir_if_last_then_block(if_stmt
);
1657 assert(block
== nir_if_first_then_block(if_stmt
));
1661 case nir_cf_node_loop
:
1662 return nir_cf_node_as_block(nir_cf_node_prev(parent
));
1664 case nir_cf_node_function
:
1668 unreachable("unknown cf node type");
1672 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
)
1674 switch (node
->type
) {
1675 case nir_cf_node_function
: {
1676 nir_function_impl
*impl
= nir_cf_node_as_function(node
);
1677 return nir_start_block(impl
);
1680 case nir_cf_node_if
: {
1681 nir_if
*if_stmt
= nir_cf_node_as_if(node
);
1682 return nir_if_first_then_block(if_stmt
);
1685 case nir_cf_node_loop
: {
1686 nir_loop
*loop
= nir_cf_node_as_loop(node
);
1687 return nir_loop_first_block(loop
);
1690 case nir_cf_node_block
: {
1691 return nir_cf_node_as_block(node
);
1695 unreachable("unknown node type");
1699 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
)
1701 switch (node
->type
) {
1702 case nir_cf_node_function
: {
1703 nir_function_impl
*impl
= nir_cf_node_as_function(node
);
1704 return nir_impl_last_block(impl
);
1707 case nir_cf_node_if
: {
1708 nir_if
*if_stmt
= nir_cf_node_as_if(node
);
1709 return nir_if_last_else_block(if_stmt
);
1712 case nir_cf_node_loop
: {
1713 nir_loop
*loop
= nir_cf_node_as_loop(node
);
1714 return nir_loop_last_block(loop
);
1717 case nir_cf_node_block
: {
1718 return nir_cf_node_as_block(node
);
1722 unreachable("unknown node type");
1726 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
)
1728 if (node
->type
== nir_cf_node_block
)
1729 return nir_block_cf_tree_next(nir_cf_node_as_block(node
));
1730 else if (node
->type
== nir_cf_node_function
)
1733 return nir_cf_node_as_block(nir_cf_node_next(node
));
1737 nir_block_get_following_if(nir_block
*block
)
1739 if (exec_node_is_tail_sentinel(&block
->cf_node
.node
))
1742 if (nir_cf_node_is_last(&block
->cf_node
))
1745 nir_cf_node
*next_node
= nir_cf_node_next(&block
->cf_node
);
1747 if (next_node
->type
!= nir_cf_node_if
)
1750 return nir_cf_node_as_if(next_node
);
1754 nir_block_get_following_loop(nir_block
*block
)
1756 if (exec_node_is_tail_sentinel(&block
->cf_node
.node
))
1759 if (nir_cf_node_is_last(&block
->cf_node
))
1762 nir_cf_node
*next_node
= nir_cf_node_next(&block
->cf_node
);
1764 if (next_node
->type
!= nir_cf_node_loop
)
1767 return nir_cf_node_as_loop(next_node
);
1771 nir_index_blocks(nir_function_impl
*impl
)
1775 if (impl
->valid_metadata
& nir_metadata_block_index
)
1778 nir_foreach_block(block
, impl
) {
1779 block
->index
= index
++;
1782 /* The end_block isn't really part of the program, which is why its index
1785 impl
->num_blocks
= impl
->end_block
->index
= index
;
1789 index_ssa_def_cb(nir_ssa_def
*def
, void *state
)
1791 unsigned *index
= (unsigned *) state
;
1792 def
->index
= (*index
)++;
1798 * The indices are applied top-to-bottom which has the very nice property
1799 * that, if A dominates B, then A->index <= B->index.
1802 nir_index_ssa_defs(nir_function_impl
*impl
)
1806 nir_foreach_block(block
, impl
) {
1807 nir_foreach_instr(instr
, block
)
1808 nir_foreach_ssa_def(instr
, index_ssa_def_cb
, &index
);
1811 impl
->ssa_alloc
= index
;
1815 * The indices are applied top-to-bottom which has the very nice property
1816 * that, if A dominates B, then A->index <= B->index.
1819 nir_index_instrs(nir_function_impl
*impl
)
1823 nir_foreach_block(block
, impl
) {
1824 nir_foreach_instr(instr
, block
)
1825 instr
->index
= index
++;
1832 nir_shader_index_vars(nir_shader
*shader
, nir_variable_mode modes
)
1835 nir_foreach_variable_with_modes(var
, shader
, modes
)
1836 var
->index
= count
++;
1841 nir_function_impl_index_vars(nir_function_impl
*impl
)
1844 nir_foreach_function_temp_variable(var
, impl
)
1845 var
->index
= count
++;
1850 cursor_next_instr(nir_cursor cursor
)
1852 switch (cursor
.option
) {
1853 case nir_cursor_before_block
:
1854 for (nir_block
*block
= cursor
.block
; block
;
1855 block
= nir_block_cf_tree_next(block
)) {
1856 nir_instr
*instr
= nir_block_first_instr(block
);
1862 case nir_cursor_after_block
:
1863 cursor
.block
= nir_block_cf_tree_next(cursor
.block
);
1864 if (cursor
.block
== NULL
)
1867 cursor
.option
= nir_cursor_before_block
;
1868 return cursor_next_instr(cursor
);
1870 case nir_cursor_before_instr
:
1871 return cursor
.instr
;
1873 case nir_cursor_after_instr
:
1874 if (nir_instr_next(cursor
.instr
))
1875 return nir_instr_next(cursor
.instr
);
1877 cursor
.option
= nir_cursor_after_block
;
1878 cursor
.block
= cursor
.instr
->block
;
1879 return cursor_next_instr(cursor
);
1882 unreachable("Inavlid cursor option");
1885 ASSERTED
static bool
1886 dest_is_ssa(nir_dest
*dest
, void *_state
)
1889 return dest
->is_ssa
;
1893 nir_function_impl_lower_instructions(nir_function_impl
*impl
,
1894 nir_instr_filter_cb filter
,
1895 nir_lower_instr_cb lower
,
1899 nir_builder_init(&b
, impl
);
1901 nir_metadata preserved
= nir_metadata_block_index
|
1902 nir_metadata_dominance
;
1904 bool progress
= false;
1905 nir_cursor iter
= nir_before_cf_list(&impl
->body
);
1907 while ((instr
= cursor_next_instr(iter
)) != NULL
) {
1908 if (filter
&& !filter(instr
, cb_data
)) {
1909 iter
= nir_after_instr(instr
);
1913 assert(nir_foreach_dest(instr
, dest_is_ssa
, NULL
));
1914 nir_ssa_def
*old_def
= nir_instr_ssa_def(instr
);
1915 if (old_def
== NULL
) {
1916 iter
= nir_after_instr(instr
);
1920 /* We're about to ask the callback to generate a replacement for instr.
1921 * Save off the uses from instr's SSA def so we know what uses to
1922 * rewrite later. If we use nir_ssa_def_rewrite_uses, it fails in the
1923 * case where the generated replacement code uses the result of instr
1924 * itself. If we use nir_ssa_def_rewrite_uses_after (which is the
1925 * normal solution to this problem), it doesn't work well if control-
1926 * flow is inserted as part of the replacement, doesn't handle cases
1927 * where the replacement is something consumed by instr, and suffers
1928 * from performance issues. This is the only way to 100% guarantee
1929 * that we rewrite the correct set efficiently.
1931 struct list_head old_uses
, old_if_uses
;
1932 list_replace(&old_def
->uses
, &old_uses
);
1933 list_inithead(&old_def
->uses
);
1934 list_replace(&old_def
->if_uses
, &old_if_uses
);
1935 list_inithead(&old_def
->if_uses
);
1937 b
.cursor
= nir_after_instr(instr
);
1938 nir_ssa_def
*new_def
= lower(&b
, instr
, cb_data
);
1939 if (new_def
&& new_def
!= NIR_LOWER_INSTR_PROGRESS
) {
1940 assert(old_def
!= NULL
);
1941 if (new_def
->parent_instr
->block
!= instr
->block
)
1942 preserved
= nir_metadata_none
;
1944 nir_src new_src
= nir_src_for_ssa(new_def
);
1945 list_for_each_entry_safe(nir_src
, use_src
, &old_uses
, use_link
)
1946 nir_instr_rewrite_src(use_src
->parent_instr
, use_src
, new_src
);
1948 list_for_each_entry_safe(nir_src
, use_src
, &old_if_uses
, use_link
)
1949 nir_if_rewrite_condition(use_src
->parent_if
, new_src
);
1951 if (list_is_empty(&old_def
->uses
) && list_is_empty(&old_def
->if_uses
)) {
1952 iter
= nir_instr_remove(instr
);
1954 iter
= nir_after_instr(instr
);
1958 /* We didn't end up lowering after all. Put the uses back */
1960 list_replace(&old_uses
, &old_def
->uses
);
1961 list_replace(&old_if_uses
, &old_def
->if_uses
);
1963 iter
= nir_after_instr(instr
);
1965 if (new_def
== NIR_LOWER_INSTR_PROGRESS
)
1971 nir_metadata_preserve(impl
, preserved
);
1973 nir_metadata_preserve(impl
, nir_metadata_all
);
1980 nir_shader_lower_instructions(nir_shader
*shader
,
1981 nir_instr_filter_cb filter
,
1982 nir_lower_instr_cb lower
,
1985 bool progress
= false;
1987 nir_foreach_function(function
, shader
) {
1988 if (function
->impl
&&
1989 nir_function_impl_lower_instructions(function
->impl
,
1990 filter
, lower
, cb_data
))
1998 nir_intrinsic_from_system_value(gl_system_value val
)
2001 case SYSTEM_VALUE_VERTEX_ID
:
2002 return nir_intrinsic_load_vertex_id
;
2003 case SYSTEM_VALUE_INSTANCE_ID
:
2004 return nir_intrinsic_load_instance_id
;
2005 case SYSTEM_VALUE_DRAW_ID
:
2006 return nir_intrinsic_load_draw_id
;
2007 case SYSTEM_VALUE_BASE_INSTANCE
:
2008 return nir_intrinsic_load_base_instance
;
2009 case SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
:
2010 return nir_intrinsic_load_vertex_id_zero_base
;
2011 case SYSTEM_VALUE_IS_INDEXED_DRAW
:
2012 return nir_intrinsic_load_is_indexed_draw
;
2013 case SYSTEM_VALUE_FIRST_VERTEX
:
2014 return nir_intrinsic_load_first_vertex
;
2015 case SYSTEM_VALUE_BASE_VERTEX
:
2016 return nir_intrinsic_load_base_vertex
;
2017 case SYSTEM_VALUE_INVOCATION_ID
:
2018 return nir_intrinsic_load_invocation_id
;
2019 case SYSTEM_VALUE_FRAG_COORD
:
2020 return nir_intrinsic_load_frag_coord
;
2021 case SYSTEM_VALUE_POINT_COORD
:
2022 return nir_intrinsic_load_point_coord
;
2023 case SYSTEM_VALUE_LINE_COORD
:
2024 return nir_intrinsic_load_line_coord
;
2025 case SYSTEM_VALUE_FRONT_FACE
:
2026 return nir_intrinsic_load_front_face
;
2027 case SYSTEM_VALUE_SAMPLE_ID
:
2028 return nir_intrinsic_load_sample_id
;
2029 case SYSTEM_VALUE_SAMPLE_POS
:
2030 return nir_intrinsic_load_sample_pos
;
2031 case SYSTEM_VALUE_SAMPLE_MASK_IN
:
2032 return nir_intrinsic_load_sample_mask_in
;
2033 case SYSTEM_VALUE_LOCAL_INVOCATION_ID
:
2034 return nir_intrinsic_load_local_invocation_id
;
2035 case SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
:
2036 return nir_intrinsic_load_local_invocation_index
;
2037 case SYSTEM_VALUE_WORK_GROUP_ID
:
2038 return nir_intrinsic_load_work_group_id
;
2039 case SYSTEM_VALUE_NUM_WORK_GROUPS
:
2040 return nir_intrinsic_load_num_work_groups
;
2041 case SYSTEM_VALUE_PRIMITIVE_ID
:
2042 return nir_intrinsic_load_primitive_id
;
2043 case SYSTEM_VALUE_TESS_COORD
:
2044 return nir_intrinsic_load_tess_coord
;
2045 case SYSTEM_VALUE_TESS_LEVEL_OUTER
:
2046 return nir_intrinsic_load_tess_level_outer
;
2047 case SYSTEM_VALUE_TESS_LEVEL_INNER
:
2048 return nir_intrinsic_load_tess_level_inner
;
2049 case SYSTEM_VALUE_TESS_LEVEL_OUTER_DEFAULT
:
2050 return nir_intrinsic_load_tess_level_outer_default
;
2051 case SYSTEM_VALUE_TESS_LEVEL_INNER_DEFAULT
:
2052 return nir_intrinsic_load_tess_level_inner_default
;
2053 case SYSTEM_VALUE_VERTICES_IN
:
2054 return nir_intrinsic_load_patch_vertices_in
;
2055 case SYSTEM_VALUE_HELPER_INVOCATION
:
2056 return nir_intrinsic_load_helper_invocation
;
2057 case SYSTEM_VALUE_COLOR0
:
2058 return nir_intrinsic_load_color0
;
2059 case SYSTEM_VALUE_COLOR1
:
2060 return nir_intrinsic_load_color1
;
2061 case SYSTEM_VALUE_VIEW_INDEX
:
2062 return nir_intrinsic_load_view_index
;
2063 case SYSTEM_VALUE_SUBGROUP_SIZE
:
2064 return nir_intrinsic_load_subgroup_size
;
2065 case SYSTEM_VALUE_SUBGROUP_INVOCATION
:
2066 return nir_intrinsic_load_subgroup_invocation
;
2067 case SYSTEM_VALUE_SUBGROUP_EQ_MASK
:
2068 return nir_intrinsic_load_subgroup_eq_mask
;
2069 case SYSTEM_VALUE_SUBGROUP_GE_MASK
:
2070 return nir_intrinsic_load_subgroup_ge_mask
;
2071 case SYSTEM_VALUE_SUBGROUP_GT_MASK
:
2072 return nir_intrinsic_load_subgroup_gt_mask
;
2073 case SYSTEM_VALUE_SUBGROUP_LE_MASK
:
2074 return nir_intrinsic_load_subgroup_le_mask
;
2075 case SYSTEM_VALUE_SUBGROUP_LT_MASK
:
2076 return nir_intrinsic_load_subgroup_lt_mask
;
2077 case SYSTEM_VALUE_NUM_SUBGROUPS
:
2078 return nir_intrinsic_load_num_subgroups
;
2079 case SYSTEM_VALUE_SUBGROUP_ID
:
2080 return nir_intrinsic_load_subgroup_id
;
2081 case SYSTEM_VALUE_LOCAL_GROUP_SIZE
:
2082 return nir_intrinsic_load_local_group_size
;
2083 case SYSTEM_VALUE_GLOBAL_INVOCATION_ID
:
2084 return nir_intrinsic_load_global_invocation_id
;
2085 case SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX
:
2086 return nir_intrinsic_load_global_invocation_index
;
2087 case SYSTEM_VALUE_WORK_DIM
:
2088 return nir_intrinsic_load_work_dim
;
2089 case SYSTEM_VALUE_USER_DATA_AMD
:
2090 return nir_intrinsic_load_user_data_amd
;
2092 unreachable("system value does not directly correspond to intrinsic");
2097 nir_system_value_from_intrinsic(nir_intrinsic_op intrin
)
2100 case nir_intrinsic_load_vertex_id
:
2101 return SYSTEM_VALUE_VERTEX_ID
;
2102 case nir_intrinsic_load_instance_id
:
2103 return SYSTEM_VALUE_INSTANCE_ID
;
2104 case nir_intrinsic_load_draw_id
:
2105 return SYSTEM_VALUE_DRAW_ID
;
2106 case nir_intrinsic_load_base_instance
:
2107 return SYSTEM_VALUE_BASE_INSTANCE
;
2108 case nir_intrinsic_load_vertex_id_zero_base
:
2109 return SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
2110 case nir_intrinsic_load_first_vertex
:
2111 return SYSTEM_VALUE_FIRST_VERTEX
;
2112 case nir_intrinsic_load_is_indexed_draw
:
2113 return SYSTEM_VALUE_IS_INDEXED_DRAW
;
2114 case nir_intrinsic_load_base_vertex
:
2115 return SYSTEM_VALUE_BASE_VERTEX
;
2116 case nir_intrinsic_load_invocation_id
:
2117 return SYSTEM_VALUE_INVOCATION_ID
;
2118 case nir_intrinsic_load_frag_coord
:
2119 return SYSTEM_VALUE_FRAG_COORD
;
2120 case nir_intrinsic_load_point_coord
:
2121 return SYSTEM_VALUE_POINT_COORD
;
2122 case nir_intrinsic_load_line_coord
:
2123 return SYSTEM_VALUE_LINE_COORD
;
2124 case nir_intrinsic_load_front_face
:
2125 return SYSTEM_VALUE_FRONT_FACE
;
2126 case nir_intrinsic_load_sample_id
:
2127 return SYSTEM_VALUE_SAMPLE_ID
;
2128 case nir_intrinsic_load_sample_pos
:
2129 return SYSTEM_VALUE_SAMPLE_POS
;
2130 case nir_intrinsic_load_sample_mask_in
:
2131 return SYSTEM_VALUE_SAMPLE_MASK_IN
;
2132 case nir_intrinsic_load_local_invocation_id
:
2133 return SYSTEM_VALUE_LOCAL_INVOCATION_ID
;
2134 case nir_intrinsic_load_local_invocation_index
:
2135 return SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
;
2136 case nir_intrinsic_load_num_work_groups
:
2137 return SYSTEM_VALUE_NUM_WORK_GROUPS
;
2138 case nir_intrinsic_load_work_group_id
:
2139 return SYSTEM_VALUE_WORK_GROUP_ID
;
2140 case nir_intrinsic_load_primitive_id
:
2141 return SYSTEM_VALUE_PRIMITIVE_ID
;
2142 case nir_intrinsic_load_tess_coord
:
2143 return SYSTEM_VALUE_TESS_COORD
;
2144 case nir_intrinsic_load_tess_level_outer
:
2145 return SYSTEM_VALUE_TESS_LEVEL_OUTER
;
2146 case nir_intrinsic_load_tess_level_inner
:
2147 return SYSTEM_VALUE_TESS_LEVEL_INNER
;
2148 case nir_intrinsic_load_tess_level_outer_default
:
2149 return SYSTEM_VALUE_TESS_LEVEL_OUTER_DEFAULT
;
2150 case nir_intrinsic_load_tess_level_inner_default
:
2151 return SYSTEM_VALUE_TESS_LEVEL_INNER_DEFAULT
;
2152 case nir_intrinsic_load_patch_vertices_in
:
2153 return SYSTEM_VALUE_VERTICES_IN
;
2154 case nir_intrinsic_load_helper_invocation
:
2155 return SYSTEM_VALUE_HELPER_INVOCATION
;
2156 case nir_intrinsic_load_color0
:
2157 return SYSTEM_VALUE_COLOR0
;
2158 case nir_intrinsic_load_color1
:
2159 return SYSTEM_VALUE_COLOR1
;
2160 case nir_intrinsic_load_view_index
:
2161 return SYSTEM_VALUE_VIEW_INDEX
;
2162 case nir_intrinsic_load_subgroup_size
:
2163 return SYSTEM_VALUE_SUBGROUP_SIZE
;
2164 case nir_intrinsic_load_subgroup_invocation
:
2165 return SYSTEM_VALUE_SUBGROUP_INVOCATION
;
2166 case nir_intrinsic_load_subgroup_eq_mask
:
2167 return SYSTEM_VALUE_SUBGROUP_EQ_MASK
;
2168 case nir_intrinsic_load_subgroup_ge_mask
:
2169 return SYSTEM_VALUE_SUBGROUP_GE_MASK
;
2170 case nir_intrinsic_load_subgroup_gt_mask
:
2171 return SYSTEM_VALUE_SUBGROUP_GT_MASK
;
2172 case nir_intrinsic_load_subgroup_le_mask
:
2173 return SYSTEM_VALUE_SUBGROUP_LE_MASK
;
2174 case nir_intrinsic_load_subgroup_lt_mask
:
2175 return SYSTEM_VALUE_SUBGROUP_LT_MASK
;
2176 case nir_intrinsic_load_num_subgroups
:
2177 return SYSTEM_VALUE_NUM_SUBGROUPS
;
2178 case nir_intrinsic_load_subgroup_id
:
2179 return SYSTEM_VALUE_SUBGROUP_ID
;
2180 case nir_intrinsic_load_local_group_size
:
2181 return SYSTEM_VALUE_LOCAL_GROUP_SIZE
;
2182 case nir_intrinsic_load_global_invocation_id
:
2183 return SYSTEM_VALUE_GLOBAL_INVOCATION_ID
;
2184 case nir_intrinsic_load_user_data_amd
:
2185 return SYSTEM_VALUE_USER_DATA_AMD
;
2187 unreachable("intrinsic doesn't produce a system value");
2191 /* OpenGL utility method that remaps the location attributes if they are
2192 * doubles. Not needed for vulkan due the differences on the input location
2193 * count for doubles on vulkan vs OpenGL
2195 * The bitfield returned in dual_slot is one bit for each double input slot in
2196 * the original OpenGL single-slot input numbering. The mapping from old
2197 * locations to new locations is as follows:
2199 * new_loc = loc + util_bitcount(dual_slot & BITFIELD64_MASK(loc))
2202 nir_remap_dual_slot_attributes(nir_shader
*shader
, uint64_t *dual_slot
)
2204 assert(shader
->info
.stage
== MESA_SHADER_VERTEX
);
2207 nir_foreach_shader_in_variable(var
, shader
) {
2208 if (glsl_type_is_dual_slot(glsl_without_array(var
->type
))) {
2209 unsigned slots
= glsl_count_attribute_slots(var
->type
, true);
2210 *dual_slot
|= BITFIELD64_MASK(slots
) << var
->data
.location
;
2214 nir_foreach_shader_in_variable(var
, shader
) {
2215 var
->data
.location
+=
2216 util_bitcount64(*dual_slot
& BITFIELD64_MASK(var
->data
.location
));
2220 /* Returns an attribute mask that has been re-compacted using the given
2224 nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
)
2227 unsigned loc
= u_bit_scan64(&dual_slot
);
2228 /* mask of all bits up to and including loc */
2229 uint64_t mask
= BITFIELD64_MASK(loc
+ 1);
2230 attribs
= (attribs
& mask
) | ((attribs
& ~mask
) >> 1);
2236 nir_rewrite_image_intrinsic(nir_intrinsic_instr
*intrin
, nir_ssa_def
*src
,
2239 enum gl_access_qualifier access
= nir_intrinsic_access(intrin
);
2241 switch (intrin
->intrinsic
) {
2243 case nir_intrinsic_image_deref_##op: \
2244 intrin->intrinsic = bindless ? nir_intrinsic_bindless_image_##op \
2245 : nir_intrinsic_image_##op; \
2257 CASE(atomic_exchange
)
2258 CASE(atomic_comp_swap
)
2260 CASE(atomic_inc_wrap
)
2261 CASE(atomic_dec_wrap
)
2264 CASE(load_raw_intel
)
2265 CASE(store_raw_intel
)
2268 unreachable("Unhanded image intrinsic");
2271 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
2272 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
2274 nir_intrinsic_set_image_dim(intrin
, glsl_get_sampler_dim(deref
->type
));
2275 nir_intrinsic_set_image_array(intrin
, glsl_sampler_type_is_array(deref
->type
));
2276 nir_intrinsic_set_access(intrin
, access
| var
->data
.access
);
2277 nir_intrinsic_set_format(intrin
, var
->data
.image
.format
);
2279 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[0],
2280 nir_src_for_ssa(src
));
2284 nir_image_intrinsic_coord_components(const nir_intrinsic_instr
*instr
)
2286 enum glsl_sampler_dim dim
= nir_intrinsic_image_dim(instr
);
2287 int coords
= glsl_get_sampler_dim_coordinate_components(dim
);
2288 if (dim
== GLSL_SAMPLER_DIM_CUBE
)
2291 return coords
+ nir_intrinsic_image_array(instr
);