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
;
299 /* create start & end blocks */
300 nir_block
*start_block
= nir_block_create(shader
);
301 nir_block
*end_block
= nir_block_create(shader
);
302 start_block
->cf_node
.parent
= &impl
->cf_node
;
303 end_block
->cf_node
.parent
= &impl
->cf_node
;
304 impl
->end_block
= end_block
;
306 exec_list_push_tail(&impl
->body
, &start_block
->cf_node
.node
);
308 start_block
->successors
[0] = end_block
;
309 _mesa_set_add(end_block
->predecessors
, start_block
);
314 nir_function_impl_create(nir_function
*function
)
316 assert(function
->impl
== NULL
);
318 nir_function_impl
*impl
= nir_function_impl_create_bare(function
->shader
);
320 function
->impl
= impl
;
321 impl
->function
= function
;
327 nir_block_create(nir_shader
*shader
)
329 nir_block
*block
= rzalloc(shader
, nir_block
);
331 cf_init(&block
->cf_node
, nir_cf_node_block
);
333 block
->successors
[0] = block
->successors
[1] = NULL
;
334 block
->predecessors
= _mesa_pointer_set_create(block
);
335 block
->imm_dom
= NULL
;
336 /* XXX maybe it would be worth it to defer allocation? This
337 * way it doesn't get allocated for shader refs that never run
338 * nir_calc_dominance? For example, state-tracker creates an
339 * initial IR, clones that, runs appropriate lowering pass, passes
340 * to driver which does common lowering/opt, and then stores ref
341 * which is later used to do state specific lowering and futher
342 * opt. Do any of the references not need dominance metadata?
344 block
->dom_frontier
= _mesa_pointer_set_create(block
);
346 exec_list_make_empty(&block
->instr_list
);
352 src_init(nir_src
*src
)
356 src
->reg
.indirect
= NULL
;
357 src
->reg
.base_offset
= 0;
361 nir_if_create(nir_shader
*shader
)
363 nir_if
*if_stmt
= ralloc(shader
, nir_if
);
365 if_stmt
->control
= nir_selection_control_none
;
367 cf_init(&if_stmt
->cf_node
, nir_cf_node_if
);
368 src_init(&if_stmt
->condition
);
370 nir_block
*then
= nir_block_create(shader
);
371 exec_list_make_empty(&if_stmt
->then_list
);
372 exec_list_push_tail(&if_stmt
->then_list
, &then
->cf_node
.node
);
373 then
->cf_node
.parent
= &if_stmt
->cf_node
;
375 nir_block
*else_stmt
= nir_block_create(shader
);
376 exec_list_make_empty(&if_stmt
->else_list
);
377 exec_list_push_tail(&if_stmt
->else_list
, &else_stmt
->cf_node
.node
);
378 else_stmt
->cf_node
.parent
= &if_stmt
->cf_node
;
384 nir_loop_create(nir_shader
*shader
)
386 nir_loop
*loop
= rzalloc(shader
, nir_loop
);
388 cf_init(&loop
->cf_node
, nir_cf_node_loop
);
390 nir_block
*body
= nir_block_create(shader
);
391 exec_list_make_empty(&loop
->body
);
392 exec_list_push_tail(&loop
->body
, &body
->cf_node
.node
);
393 body
->cf_node
.parent
= &loop
->cf_node
;
395 body
->successors
[0] = body
;
396 _mesa_set_add(body
->predecessors
, body
);
402 instr_init(nir_instr
*instr
, nir_instr_type type
)
406 exec_node_init(&instr
->node
);
410 dest_init(nir_dest
*dest
)
412 dest
->is_ssa
= false;
413 dest
->reg
.reg
= NULL
;
414 dest
->reg
.indirect
= NULL
;
415 dest
->reg
.base_offset
= 0;
419 alu_dest_init(nir_alu_dest
*dest
)
421 dest_init(&dest
->dest
);
422 dest
->saturate
= false;
423 dest
->write_mask
= 0xf;
427 alu_src_init(nir_alu_src
*src
)
430 src
->abs
= src
->negate
= false;
431 for (int i
= 0; i
< NIR_MAX_VEC_COMPONENTS
; ++i
)
436 nir_alu_instr_create(nir_shader
*shader
, nir_op op
)
438 unsigned num_srcs
= nir_op_infos
[op
].num_inputs
;
439 /* TODO: don't use rzalloc */
440 nir_alu_instr
*instr
=
442 sizeof(nir_alu_instr
) + num_srcs
* sizeof(nir_alu_src
));
444 instr_init(&instr
->instr
, nir_instr_type_alu
);
446 alu_dest_init(&instr
->dest
);
447 for (unsigned i
= 0; i
< num_srcs
; i
++)
448 alu_src_init(&instr
->src
[i
]);
454 nir_deref_instr_create(nir_shader
*shader
, nir_deref_type deref_type
)
456 nir_deref_instr
*instr
=
457 rzalloc_size(shader
, sizeof(nir_deref_instr
));
459 instr_init(&instr
->instr
, nir_instr_type_deref
);
461 instr
->deref_type
= deref_type
;
462 if (deref_type
!= nir_deref_type_var
)
463 src_init(&instr
->parent
);
465 if (deref_type
== nir_deref_type_array
||
466 deref_type
== nir_deref_type_ptr_as_array
)
467 src_init(&instr
->arr
.index
);
469 dest_init(&instr
->dest
);
475 nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
)
477 nir_jump_instr
*instr
= ralloc(shader
, nir_jump_instr
);
478 instr_init(&instr
->instr
, nir_instr_type_jump
);
483 nir_load_const_instr
*
484 nir_load_const_instr_create(nir_shader
*shader
, unsigned num_components
,
487 nir_load_const_instr
*instr
=
488 rzalloc_size(shader
, sizeof(*instr
) + num_components
* sizeof(*instr
->value
));
489 instr_init(&instr
->instr
, nir_instr_type_load_const
);
491 nir_ssa_def_init(&instr
->instr
, &instr
->def
, num_components
, bit_size
, NULL
);
496 nir_intrinsic_instr
*
497 nir_intrinsic_instr_create(nir_shader
*shader
, nir_intrinsic_op op
)
499 unsigned num_srcs
= nir_intrinsic_infos
[op
].num_srcs
;
500 /* TODO: don't use rzalloc */
501 nir_intrinsic_instr
*instr
=
503 sizeof(nir_intrinsic_instr
) + num_srcs
* sizeof(nir_src
));
505 instr_init(&instr
->instr
, nir_instr_type_intrinsic
);
506 instr
->intrinsic
= op
;
508 if (nir_intrinsic_infos
[op
].has_dest
)
509 dest_init(&instr
->dest
);
511 for (unsigned i
= 0; i
< num_srcs
; i
++)
512 src_init(&instr
->src
[i
]);
518 nir_call_instr_create(nir_shader
*shader
, nir_function
*callee
)
520 const unsigned num_params
= callee
->num_params
;
521 nir_call_instr
*instr
=
522 rzalloc_size(shader
, sizeof(*instr
) +
523 num_params
* sizeof(instr
->params
[0]));
525 instr_init(&instr
->instr
, nir_instr_type_call
);
526 instr
->callee
= callee
;
527 instr
->num_params
= num_params
;
528 for (unsigned i
= 0; i
< num_params
; i
++)
529 src_init(&instr
->params
[i
]);
534 static int8_t default_tg4_offsets
[4][2] =
543 nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
)
545 nir_tex_instr
*instr
= rzalloc(shader
, nir_tex_instr
);
546 instr_init(&instr
->instr
, nir_instr_type_tex
);
548 dest_init(&instr
->dest
);
550 instr
->num_srcs
= num_srcs
;
551 instr
->src
= ralloc_array(instr
, nir_tex_src
, num_srcs
);
552 for (unsigned i
= 0; i
< num_srcs
; i
++)
553 src_init(&instr
->src
[i
].src
);
555 instr
->texture_index
= 0;
556 instr
->sampler_index
= 0;
557 memcpy(instr
->tg4_offsets
, default_tg4_offsets
, sizeof(instr
->tg4_offsets
));
563 nir_tex_instr_add_src(nir_tex_instr
*tex
,
564 nir_tex_src_type src_type
,
567 nir_tex_src
*new_srcs
= rzalloc_array(tex
, nir_tex_src
,
570 for (unsigned i
= 0; i
< tex
->num_srcs
; i
++) {
571 new_srcs
[i
].src_type
= tex
->src
[i
].src_type
;
572 nir_instr_move_src(&tex
->instr
, &new_srcs
[i
].src
,
576 ralloc_free(tex
->src
);
579 tex
->src
[tex
->num_srcs
].src_type
= src_type
;
580 nir_instr_rewrite_src(&tex
->instr
, &tex
->src
[tex
->num_srcs
].src
, src
);
585 nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
)
587 assert(src_idx
< tex
->num_srcs
);
589 /* First rewrite the source to NIR_SRC_INIT */
590 nir_instr_rewrite_src(&tex
->instr
, &tex
->src
[src_idx
].src
, NIR_SRC_INIT
);
592 /* Now, move all of the other sources down */
593 for (unsigned i
= src_idx
+ 1; i
< tex
->num_srcs
; i
++) {
594 tex
->src
[i
-1].src_type
= tex
->src
[i
].src_type
;
595 nir_instr_move_src(&tex
->instr
, &tex
->src
[i
-1].src
, &tex
->src
[i
].src
);
601 nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
)
603 if (tex
->op
!= nir_texop_tg4
)
605 return memcmp(tex
->tg4_offsets
, default_tg4_offsets
,
606 sizeof(tex
->tg4_offsets
)) != 0;
610 nir_phi_instr_create(nir_shader
*shader
)
612 nir_phi_instr
*instr
= ralloc(shader
, nir_phi_instr
);
613 instr_init(&instr
->instr
, nir_instr_type_phi
);
615 dest_init(&instr
->dest
);
616 exec_list_make_empty(&instr
->srcs
);
620 nir_parallel_copy_instr
*
621 nir_parallel_copy_instr_create(nir_shader
*shader
)
623 nir_parallel_copy_instr
*instr
= ralloc(shader
, nir_parallel_copy_instr
);
624 instr_init(&instr
->instr
, nir_instr_type_parallel_copy
);
626 exec_list_make_empty(&instr
->entries
);
631 nir_ssa_undef_instr
*
632 nir_ssa_undef_instr_create(nir_shader
*shader
,
633 unsigned num_components
,
636 nir_ssa_undef_instr
*instr
= ralloc(shader
, nir_ssa_undef_instr
);
637 instr_init(&instr
->instr
, nir_instr_type_ssa_undef
);
639 nir_ssa_def_init(&instr
->instr
, &instr
->def
, num_components
, bit_size
, NULL
);
644 static nir_const_value
645 const_value_float(double d
, unsigned bit_size
)
648 memset(&v
, 0, sizeof(v
));
650 case 16: v
.u16
= _mesa_float_to_half(d
); break;
651 case 32: v
.f32
= d
; break;
652 case 64: v
.f64
= d
; break;
654 unreachable("Invalid bit size");
659 static nir_const_value
660 const_value_int(int64_t i
, unsigned bit_size
)
663 memset(&v
, 0, sizeof(v
));
665 case 1: v
.b
= i
& 1; break;
666 case 8: v
.i8
= i
; break;
667 case 16: v
.i16
= i
; break;
668 case 32: v
.i32
= i
; break;
669 case 64: v
.i64
= i
; break;
671 unreachable("Invalid bit size");
677 nir_alu_binop_identity(nir_op binop
, unsigned bit_size
)
679 const int64_t max_int
= (1ull << (bit_size
- 1)) - 1;
680 const int64_t min_int
= -max_int
- 1;
683 return const_value_int(0, bit_size
);
685 return const_value_float(0, bit_size
);
687 return const_value_int(1, bit_size
);
689 return const_value_float(1, bit_size
);
691 return const_value_int(max_int
, bit_size
);
693 return const_value_int(~0ull, bit_size
);
695 return const_value_float(INFINITY
, bit_size
);
697 return const_value_int(min_int
, bit_size
);
699 return const_value_int(0, bit_size
);
701 return const_value_float(-INFINITY
, bit_size
);
703 return const_value_int(~0ull, bit_size
);
705 return const_value_int(0, bit_size
);
707 return const_value_int(0, bit_size
);
709 unreachable("Invalid reduction operation");
714 nir_cf_node_get_function(nir_cf_node
*node
)
716 while (node
->type
!= nir_cf_node_function
) {
720 return nir_cf_node_as_function(node
);
723 /* Reduces a cursor by trying to convert everything to after and trying to
724 * go up to block granularity when possible.
727 reduce_cursor(nir_cursor cursor
)
729 switch (cursor
.option
) {
730 case nir_cursor_before_block
:
731 assert(nir_cf_node_prev(&cursor
.block
->cf_node
) == NULL
||
732 nir_cf_node_prev(&cursor
.block
->cf_node
)->type
!= nir_cf_node_block
);
733 if (exec_list_is_empty(&cursor
.block
->instr_list
)) {
734 /* Empty block. After is as good as before. */
735 cursor
.option
= nir_cursor_after_block
;
739 case nir_cursor_after_block
:
742 case nir_cursor_before_instr
: {
743 nir_instr
*prev_instr
= nir_instr_prev(cursor
.instr
);
745 /* Before this instruction is after the previous */
746 cursor
.instr
= prev_instr
;
747 cursor
.option
= nir_cursor_after_instr
;
749 /* No previous instruction. Switch to before block */
750 cursor
.block
= cursor
.instr
->block
;
751 cursor
.option
= nir_cursor_before_block
;
753 return reduce_cursor(cursor
);
756 case nir_cursor_after_instr
:
757 if (nir_instr_next(cursor
.instr
) == NULL
) {
758 /* This is the last instruction, switch to after block */
759 cursor
.option
= nir_cursor_after_block
;
760 cursor
.block
= cursor
.instr
->block
;
765 unreachable("Inavlid cursor option");
770 nir_cursors_equal(nir_cursor a
, nir_cursor b
)
772 /* Reduced cursors should be unique */
773 a
= reduce_cursor(a
);
774 b
= reduce_cursor(b
);
776 return a
.block
== b
.block
&& a
.option
== b
.option
;
780 add_use_cb(nir_src
*src
, void *state
)
782 nir_instr
*instr
= state
;
784 src
->parent_instr
= instr
;
785 list_addtail(&src
->use_link
,
786 src
->is_ssa
? &src
->ssa
->uses
: &src
->reg
.reg
->uses
);
792 add_ssa_def_cb(nir_ssa_def
*def
, void *state
)
794 nir_instr
*instr
= state
;
796 if (instr
->block
&& def
->index
== UINT_MAX
) {
797 nir_function_impl
*impl
=
798 nir_cf_node_get_function(&instr
->block
->cf_node
);
800 def
->index
= impl
->ssa_alloc
++;
807 add_reg_def_cb(nir_dest
*dest
, void *state
)
809 nir_instr
*instr
= state
;
812 dest
->reg
.parent_instr
= instr
;
813 list_addtail(&dest
->reg
.def_link
, &dest
->reg
.reg
->defs
);
820 add_defs_uses(nir_instr
*instr
)
822 nir_foreach_src(instr
, add_use_cb
, instr
);
823 nir_foreach_dest(instr
, add_reg_def_cb
, instr
);
824 nir_foreach_ssa_def(instr
, add_ssa_def_cb
, instr
);
828 nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
)
830 switch (cursor
.option
) {
831 case nir_cursor_before_block
:
832 /* Only allow inserting jumps into empty blocks. */
833 if (instr
->type
== nir_instr_type_jump
)
834 assert(exec_list_is_empty(&cursor
.block
->instr_list
));
836 instr
->block
= cursor
.block
;
837 add_defs_uses(instr
);
838 exec_list_push_head(&cursor
.block
->instr_list
, &instr
->node
);
840 case nir_cursor_after_block
: {
841 /* Inserting instructions after a jump is illegal. */
842 nir_instr
*last
= nir_block_last_instr(cursor
.block
);
843 assert(last
== NULL
|| last
->type
!= nir_instr_type_jump
);
846 instr
->block
= cursor
.block
;
847 add_defs_uses(instr
);
848 exec_list_push_tail(&cursor
.block
->instr_list
, &instr
->node
);
851 case nir_cursor_before_instr
:
852 assert(instr
->type
!= nir_instr_type_jump
);
853 instr
->block
= cursor
.instr
->block
;
854 add_defs_uses(instr
);
855 exec_node_insert_node_before(&cursor
.instr
->node
, &instr
->node
);
857 case nir_cursor_after_instr
:
858 /* Inserting instructions after a jump is illegal. */
859 assert(cursor
.instr
->type
!= nir_instr_type_jump
);
861 /* Only allow inserting jumps at the end of the block. */
862 if (instr
->type
== nir_instr_type_jump
)
863 assert(cursor
.instr
== nir_block_last_instr(cursor
.instr
->block
));
865 instr
->block
= cursor
.instr
->block
;
866 add_defs_uses(instr
);
867 exec_node_insert_after(&cursor
.instr
->node
, &instr
->node
);
871 if (instr
->type
== nir_instr_type_jump
)
872 nir_handle_add_jump(instr
->block
);
876 src_is_valid(const nir_src
*src
)
878 return src
->is_ssa
? (src
->ssa
!= NULL
) : (src
->reg
.reg
!= NULL
);
882 remove_use_cb(nir_src
*src
, void *state
)
886 if (src_is_valid(src
))
887 list_del(&src
->use_link
);
893 remove_def_cb(nir_dest
*dest
, void *state
)
898 list_del(&dest
->reg
.def_link
);
904 remove_defs_uses(nir_instr
*instr
)
906 nir_foreach_dest(instr
, remove_def_cb
, instr
);
907 nir_foreach_src(instr
, remove_use_cb
, instr
);
910 void nir_instr_remove_v(nir_instr
*instr
)
912 remove_defs_uses(instr
);
913 exec_node_remove(&instr
->node
);
915 if (instr
->type
== nir_instr_type_jump
) {
916 nir_jump_instr
*jump_instr
= nir_instr_as_jump(instr
);
917 nir_handle_remove_jump(instr
->block
, jump_instr
->type
);
924 nir_index_local_regs(nir_function_impl
*impl
)
927 foreach_list_typed(nir_register
, reg
, node
, &impl
->registers
) {
928 reg
->index
= index
++;
930 impl
->reg_alloc
= index
;
934 visit_alu_dest(nir_alu_instr
*instr
, nir_foreach_dest_cb cb
, void *state
)
936 return cb(&instr
->dest
.dest
, state
);
940 visit_deref_dest(nir_deref_instr
*instr
, nir_foreach_dest_cb cb
, void *state
)
942 return cb(&instr
->dest
, state
);
946 visit_intrinsic_dest(nir_intrinsic_instr
*instr
, nir_foreach_dest_cb cb
,
949 if (nir_intrinsic_infos
[instr
->intrinsic
].has_dest
)
950 return cb(&instr
->dest
, state
);
956 visit_texture_dest(nir_tex_instr
*instr
, nir_foreach_dest_cb cb
,
959 return cb(&instr
->dest
, state
);
963 visit_phi_dest(nir_phi_instr
*instr
, nir_foreach_dest_cb cb
, void *state
)
965 return cb(&instr
->dest
, state
);
969 visit_parallel_copy_dest(nir_parallel_copy_instr
*instr
,
970 nir_foreach_dest_cb cb
, void *state
)
972 nir_foreach_parallel_copy_entry(entry
, instr
) {
973 if (!cb(&entry
->dest
, state
))
981 nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
)
983 switch (instr
->type
) {
984 case nir_instr_type_alu
:
985 return visit_alu_dest(nir_instr_as_alu(instr
), cb
, state
);
986 case nir_instr_type_deref
:
987 return visit_deref_dest(nir_instr_as_deref(instr
), cb
, state
);
988 case nir_instr_type_intrinsic
:
989 return visit_intrinsic_dest(nir_instr_as_intrinsic(instr
), cb
, state
);
990 case nir_instr_type_tex
:
991 return visit_texture_dest(nir_instr_as_tex(instr
), cb
, state
);
992 case nir_instr_type_phi
:
993 return visit_phi_dest(nir_instr_as_phi(instr
), cb
, state
);
994 case nir_instr_type_parallel_copy
:
995 return visit_parallel_copy_dest(nir_instr_as_parallel_copy(instr
),
998 case nir_instr_type_load_const
:
999 case nir_instr_type_ssa_undef
:
1000 case nir_instr_type_call
:
1001 case nir_instr_type_jump
:
1005 unreachable("Invalid instruction type");
1012 struct foreach_ssa_def_state
{
1013 nir_foreach_ssa_def_cb cb
;
1018 nir_ssa_def_visitor(nir_dest
*dest
, void *void_state
)
1020 struct foreach_ssa_def_state
*state
= void_state
;
1023 return state
->cb(&dest
->ssa
, state
->client_state
);
1029 nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
, void *state
)
1031 switch (instr
->type
) {
1032 case nir_instr_type_alu
:
1033 case nir_instr_type_deref
:
1034 case nir_instr_type_tex
:
1035 case nir_instr_type_intrinsic
:
1036 case nir_instr_type_phi
:
1037 case nir_instr_type_parallel_copy
: {
1038 struct foreach_ssa_def_state foreach_state
= {cb
, state
};
1039 return nir_foreach_dest(instr
, nir_ssa_def_visitor
, &foreach_state
);
1042 case nir_instr_type_load_const
:
1043 return cb(&nir_instr_as_load_const(instr
)->def
, state
);
1044 case nir_instr_type_ssa_undef
:
1045 return cb(&nir_instr_as_ssa_undef(instr
)->def
, state
);
1046 case nir_instr_type_call
:
1047 case nir_instr_type_jump
:
1050 unreachable("Invalid instruction type");
1055 nir_instr_ssa_def(nir_instr
*instr
)
1057 switch (instr
->type
) {
1058 case nir_instr_type_alu
:
1059 assert(nir_instr_as_alu(instr
)->dest
.dest
.is_ssa
);
1060 return &nir_instr_as_alu(instr
)->dest
.dest
.ssa
;
1062 case nir_instr_type_deref
:
1063 assert(nir_instr_as_deref(instr
)->dest
.is_ssa
);
1064 return &nir_instr_as_deref(instr
)->dest
.ssa
;
1066 case nir_instr_type_tex
:
1067 assert(nir_instr_as_tex(instr
)->dest
.is_ssa
);
1068 return &nir_instr_as_tex(instr
)->dest
.ssa
;
1070 case nir_instr_type_intrinsic
: {
1071 nir_intrinsic_instr
*intrin
= nir_instr_as_intrinsic(instr
);
1072 if (nir_intrinsic_infos
[intrin
->intrinsic
].has_dest
) {
1073 assert(intrin
->dest
.is_ssa
);
1074 return &intrin
->dest
.ssa
;
1080 case nir_instr_type_phi
:
1081 assert(nir_instr_as_phi(instr
)->dest
.is_ssa
);
1082 return &nir_instr_as_phi(instr
)->dest
.ssa
;
1084 case nir_instr_type_parallel_copy
:
1085 unreachable("Parallel copies are unsupported by this function");
1087 case nir_instr_type_load_const
:
1088 return &nir_instr_as_load_const(instr
)->def
;
1090 case nir_instr_type_ssa_undef
:
1091 return &nir_instr_as_ssa_undef(instr
)->def
;
1093 case nir_instr_type_call
:
1094 case nir_instr_type_jump
:
1098 unreachable("Invalid instruction type");
1102 visit_src(nir_src
*src
, nir_foreach_src_cb cb
, void *state
)
1104 if (!cb(src
, state
))
1106 if (!src
->is_ssa
&& src
->reg
.indirect
)
1107 return cb(src
->reg
.indirect
, state
);
1112 visit_alu_src(nir_alu_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1114 for (unsigned i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
1115 if (!visit_src(&instr
->src
[i
].src
, cb
, state
))
1122 visit_deref_instr_src(nir_deref_instr
*instr
,
1123 nir_foreach_src_cb cb
, void *state
)
1125 if (instr
->deref_type
!= nir_deref_type_var
) {
1126 if (!visit_src(&instr
->parent
, cb
, state
))
1130 if (instr
->deref_type
== nir_deref_type_array
||
1131 instr
->deref_type
== nir_deref_type_ptr_as_array
) {
1132 if (!visit_src(&instr
->arr
.index
, cb
, state
))
1140 visit_tex_src(nir_tex_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1142 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++) {
1143 if (!visit_src(&instr
->src
[i
].src
, cb
, state
))
1151 visit_intrinsic_src(nir_intrinsic_instr
*instr
, nir_foreach_src_cb cb
,
1154 unsigned num_srcs
= nir_intrinsic_infos
[instr
->intrinsic
].num_srcs
;
1155 for (unsigned i
= 0; i
< num_srcs
; i
++) {
1156 if (!visit_src(&instr
->src
[i
], cb
, state
))
1164 visit_call_src(nir_call_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1166 for (unsigned i
= 0; i
< instr
->num_params
; i
++) {
1167 if (!visit_src(&instr
->params
[i
], cb
, state
))
1175 visit_phi_src(nir_phi_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1177 nir_foreach_phi_src(src
, instr
) {
1178 if (!visit_src(&src
->src
, cb
, state
))
1186 visit_parallel_copy_src(nir_parallel_copy_instr
*instr
,
1187 nir_foreach_src_cb cb
, void *state
)
1189 nir_foreach_parallel_copy_entry(entry
, instr
) {
1190 if (!visit_src(&entry
->src
, cb
, state
))
1199 nir_foreach_src_cb cb
;
1200 } visit_dest_indirect_state
;
1203 visit_dest_indirect(nir_dest
*dest
, void *_state
)
1205 visit_dest_indirect_state
*state
= (visit_dest_indirect_state
*) _state
;
1207 if (!dest
->is_ssa
&& dest
->reg
.indirect
)
1208 return state
->cb(dest
->reg
.indirect
, state
->state
);
1214 nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
)
1216 switch (instr
->type
) {
1217 case nir_instr_type_alu
:
1218 if (!visit_alu_src(nir_instr_as_alu(instr
), cb
, state
))
1221 case nir_instr_type_deref
:
1222 if (!visit_deref_instr_src(nir_instr_as_deref(instr
), cb
, state
))
1225 case nir_instr_type_intrinsic
:
1226 if (!visit_intrinsic_src(nir_instr_as_intrinsic(instr
), cb
, state
))
1229 case nir_instr_type_tex
:
1230 if (!visit_tex_src(nir_instr_as_tex(instr
), cb
, state
))
1233 case nir_instr_type_call
:
1234 if (!visit_call_src(nir_instr_as_call(instr
), cb
, state
))
1237 case nir_instr_type_load_const
:
1238 /* Constant load instructions have no regular sources */
1240 case nir_instr_type_phi
:
1241 if (!visit_phi_src(nir_instr_as_phi(instr
), cb
, state
))
1244 case nir_instr_type_parallel_copy
:
1245 if (!visit_parallel_copy_src(nir_instr_as_parallel_copy(instr
),
1249 case nir_instr_type_jump
:
1250 case nir_instr_type_ssa_undef
:
1254 unreachable("Invalid instruction type");
1258 visit_dest_indirect_state dest_state
;
1259 dest_state
.state
= state
;
1261 return nir_foreach_dest(instr
, visit_dest_indirect
, &dest_state
);
1265 nir_foreach_phi_src_leaving_block(nir_block
*block
,
1266 nir_foreach_src_cb cb
,
1269 for (unsigned i
= 0; i
< ARRAY_SIZE(block
->successors
); i
++) {
1270 if (block
->successors
[i
] == NULL
)
1273 nir_foreach_instr(instr
, block
->successors
[i
]) {
1274 if (instr
->type
!= nir_instr_type_phi
)
1277 nir_phi_instr
*phi
= nir_instr_as_phi(instr
);
1278 nir_foreach_phi_src(phi_src
, phi
) {
1279 if (phi_src
->pred
== block
) {
1280 if (!cb(&phi_src
->src
, state
))
1291 nir_const_value_for_float(double f
, unsigned bit_size
)
1294 memset(&v
, 0, sizeof(v
));
1298 v
.u16
= _mesa_float_to_half(f
);
1307 unreachable("Invalid bit size");
1314 nir_const_value_as_float(nir_const_value value
, unsigned bit_size
)
1317 case 16: return _mesa_half_to_float(value
.u16
);
1318 case 32: return value
.f32
;
1319 case 64: return value
.f64
;
1321 unreachable("Invalid bit size");
1326 nir_src_as_const_value(nir_src src
)
1331 if (src
.ssa
->parent_instr
->type
!= nir_instr_type_load_const
)
1334 nir_load_const_instr
*load
= nir_instr_as_load_const(src
.ssa
->parent_instr
);
1340 * Returns true if the source is known to be dynamically uniform. Otherwise it
1341 * returns false which means it may or may not be dynamically uniform but it
1342 * can't be determined.
1345 nir_src_is_dynamically_uniform(nir_src src
)
1350 /* Constants are trivially dynamically uniform */
1351 if (src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
)
1354 /* As are uniform variables */
1355 if (src
.ssa
->parent_instr
->type
== nir_instr_type_intrinsic
) {
1356 nir_intrinsic_instr
*intr
= nir_instr_as_intrinsic(src
.ssa
->parent_instr
);
1357 if (intr
->intrinsic
== nir_intrinsic_load_uniform
&&
1358 nir_src_is_dynamically_uniform(intr
->src
[0]))
1362 /* Operating together dynamically uniform expressions produces a
1363 * dynamically uniform result
1365 if (src
.ssa
->parent_instr
->type
== nir_instr_type_alu
) {
1366 nir_alu_instr
*alu
= nir_instr_as_alu(src
.ssa
->parent_instr
);
1367 for (int i
= 0; i
< nir_op_infos
[alu
->op
].num_inputs
; i
++) {
1368 if (!nir_src_is_dynamically_uniform(alu
->src
[i
].src
))
1375 /* XXX: this could have many more tests, such as when a sampler function is
1376 * called with dynamically uniform arguments.
1382 src_remove_all_uses(nir_src
*src
)
1384 for (; src
; src
= src
->is_ssa
? NULL
: src
->reg
.indirect
) {
1385 if (!src_is_valid(src
))
1388 list_del(&src
->use_link
);
1393 src_add_all_uses(nir_src
*src
, nir_instr
*parent_instr
, nir_if
*parent_if
)
1395 for (; src
; src
= src
->is_ssa
? NULL
: src
->reg
.indirect
) {
1396 if (!src_is_valid(src
))
1400 src
->parent_instr
= parent_instr
;
1402 list_addtail(&src
->use_link
, &src
->ssa
->uses
);
1404 list_addtail(&src
->use_link
, &src
->reg
.reg
->uses
);
1407 src
->parent_if
= parent_if
;
1409 list_addtail(&src
->use_link
, &src
->ssa
->if_uses
);
1411 list_addtail(&src
->use_link
, &src
->reg
.reg
->if_uses
);
1417 nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
)
1419 assert(!src_is_valid(src
) || src
->parent_instr
== instr
);
1421 src_remove_all_uses(src
);
1423 src_add_all_uses(src
, instr
, NULL
);
1427 nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
)
1429 assert(!src_is_valid(dest
) || dest
->parent_instr
== dest_instr
);
1431 src_remove_all_uses(dest
);
1432 src_remove_all_uses(src
);
1434 *src
= NIR_SRC_INIT
;
1435 src_add_all_uses(dest
, dest_instr
, NULL
);
1439 nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
)
1441 nir_src
*src
= &if_stmt
->condition
;
1442 assert(!src_is_valid(src
) || src
->parent_if
== if_stmt
);
1444 src_remove_all_uses(src
);
1446 src_add_all_uses(src
, NULL
, if_stmt
);
1450 nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
, nir_dest new_dest
)
1453 /* We can only overwrite an SSA destination if it has no uses. */
1454 assert(list_is_empty(&dest
->ssa
.uses
) && list_is_empty(&dest
->ssa
.if_uses
));
1456 list_del(&dest
->reg
.def_link
);
1457 if (dest
->reg
.indirect
)
1458 src_remove_all_uses(dest
->reg
.indirect
);
1461 /* We can't re-write with an SSA def */
1462 assert(!new_dest
.is_ssa
);
1464 nir_dest_copy(dest
, &new_dest
, instr
);
1466 dest
->reg
.parent_instr
= instr
;
1467 list_addtail(&dest
->reg
.def_link
, &new_dest
.reg
.reg
->defs
);
1469 if (dest
->reg
.indirect
)
1470 src_add_all_uses(dest
->reg
.indirect
, instr
, NULL
);
1473 /* note: does *not* take ownership of 'name' */
1475 nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
1476 unsigned num_components
,
1477 unsigned bit_size
, const char *name
)
1479 def
->name
= ralloc_strdup(instr
, name
);
1480 def
->parent_instr
= instr
;
1481 list_inithead(&def
->uses
);
1482 list_inithead(&def
->if_uses
);
1483 def
->num_components
= num_components
;
1484 def
->bit_size
= bit_size
;
1485 def
->divergent
= true; /* This is the safer default */
1488 nir_function_impl
*impl
=
1489 nir_cf_node_get_function(&instr
->block
->cf_node
);
1491 def
->index
= impl
->ssa_alloc
++;
1493 def
->index
= UINT_MAX
;
1497 /* note: does *not* take ownership of 'name' */
1499 nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
1500 unsigned num_components
, unsigned bit_size
,
1503 dest
->is_ssa
= true;
1504 nir_ssa_def_init(instr
, &dest
->ssa
, num_components
, bit_size
, name
);
1508 nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
)
1510 assert(!new_src
.is_ssa
|| def
!= new_src
.ssa
);
1512 nir_foreach_use_safe(use_src
, def
)
1513 nir_instr_rewrite_src(use_src
->parent_instr
, use_src
, new_src
);
1515 nir_foreach_if_use_safe(use_src
, def
)
1516 nir_if_rewrite_condition(use_src
->parent_if
, new_src
);
1520 is_instr_between(nir_instr
*start
, nir_instr
*end
, nir_instr
*between
)
1522 assert(start
->block
== end
->block
);
1524 if (between
->block
!= start
->block
)
1527 /* Search backwards looking for "between" */
1528 while (start
!= end
) {
1532 end
= nir_instr_prev(end
);
1539 /* Replaces all uses of the given SSA def with the given source but only if
1540 * the use comes after the after_me instruction. This can be useful if you
1541 * are emitting code to fix up the result of some instruction: you can freely
1542 * use the result in that code and then call rewrite_uses_after and pass the
1543 * last fixup instruction as after_me and it will replace all of the uses you
1544 * want without touching the fixup code.
1546 * This function assumes that after_me is in the same block as
1547 * def->parent_instr and that after_me comes after def->parent_instr.
1550 nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
1551 nir_instr
*after_me
)
1553 if (new_src
.is_ssa
&& def
== new_src
.ssa
)
1556 nir_foreach_use_safe(use_src
, def
) {
1557 assert(use_src
->parent_instr
!= def
->parent_instr
);
1558 /* Since def already dominates all of its uses, the only way a use can
1559 * not be dominated by after_me is if it is between def and after_me in
1560 * the instruction list.
1562 if (!is_instr_between(def
->parent_instr
, after_me
, use_src
->parent_instr
))
1563 nir_instr_rewrite_src(use_src
->parent_instr
, use_src
, new_src
);
1566 nir_foreach_if_use_safe(use_src
, def
)
1567 nir_if_rewrite_condition(use_src
->parent_if
, new_src
);
1570 nir_component_mask_t
1571 nir_ssa_def_components_read(const nir_ssa_def
*def
)
1573 nir_component_mask_t read_mask
= 0;
1574 nir_foreach_use(use
, def
) {
1575 if (use
->parent_instr
->type
== nir_instr_type_alu
) {
1576 nir_alu_instr
*alu
= nir_instr_as_alu(use
->parent_instr
);
1577 nir_alu_src
*alu_src
= exec_node_data(nir_alu_src
, use
, src
);
1578 int src_idx
= alu_src
- &alu
->src
[0];
1579 assert(src_idx
>= 0 && src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
1580 read_mask
|= nir_alu_instr_src_read_mask(alu
, src_idx
);
1582 return (1 << def
->num_components
) - 1;
1586 if (!list_is_empty(&def
->if_uses
))
1593 nir_block_cf_tree_next(nir_block
*block
)
1595 if (block
== NULL
) {
1596 /* nir_foreach_block_safe() will call this function on a NULL block
1597 * after the last iteration, but it won't use the result so just return
1603 nir_cf_node
*cf_next
= nir_cf_node_next(&block
->cf_node
);
1605 return nir_cf_node_cf_tree_first(cf_next
);
1607 nir_cf_node
*parent
= block
->cf_node
.parent
;
1609 switch (parent
->type
) {
1610 case nir_cf_node_if
: {
1611 /* Are we at the end of the if? Go to the beginning of the else */
1612 nir_if
*if_stmt
= nir_cf_node_as_if(parent
);
1613 if (block
== nir_if_last_then_block(if_stmt
))
1614 return nir_if_first_else_block(if_stmt
);
1616 assert(block
== nir_if_last_else_block(if_stmt
));
1620 case nir_cf_node_loop
:
1621 return nir_cf_node_as_block(nir_cf_node_next(parent
));
1623 case nir_cf_node_function
:
1627 unreachable("unknown cf node type");
1632 nir_block_cf_tree_prev(nir_block
*block
)
1634 if (block
== NULL
) {
1635 /* do this for consistency with nir_block_cf_tree_next() */
1639 nir_cf_node
*cf_prev
= nir_cf_node_prev(&block
->cf_node
);
1641 return nir_cf_node_cf_tree_last(cf_prev
);
1643 nir_cf_node
*parent
= block
->cf_node
.parent
;
1645 switch (parent
->type
) {
1646 case nir_cf_node_if
: {
1647 /* Are we at the beginning of the else? Go to the end of the if */
1648 nir_if
*if_stmt
= nir_cf_node_as_if(parent
);
1649 if (block
== nir_if_first_else_block(if_stmt
))
1650 return nir_if_last_then_block(if_stmt
);
1652 assert(block
== nir_if_first_then_block(if_stmt
));
1656 case nir_cf_node_loop
:
1657 return nir_cf_node_as_block(nir_cf_node_prev(parent
));
1659 case nir_cf_node_function
:
1663 unreachable("unknown cf node type");
1667 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
)
1669 switch (node
->type
) {
1670 case nir_cf_node_function
: {
1671 nir_function_impl
*impl
= nir_cf_node_as_function(node
);
1672 return nir_start_block(impl
);
1675 case nir_cf_node_if
: {
1676 nir_if
*if_stmt
= nir_cf_node_as_if(node
);
1677 return nir_if_first_then_block(if_stmt
);
1680 case nir_cf_node_loop
: {
1681 nir_loop
*loop
= nir_cf_node_as_loop(node
);
1682 return nir_loop_first_block(loop
);
1685 case nir_cf_node_block
: {
1686 return nir_cf_node_as_block(node
);
1690 unreachable("unknown node type");
1694 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
)
1696 switch (node
->type
) {
1697 case nir_cf_node_function
: {
1698 nir_function_impl
*impl
= nir_cf_node_as_function(node
);
1699 return nir_impl_last_block(impl
);
1702 case nir_cf_node_if
: {
1703 nir_if
*if_stmt
= nir_cf_node_as_if(node
);
1704 return nir_if_last_else_block(if_stmt
);
1707 case nir_cf_node_loop
: {
1708 nir_loop
*loop
= nir_cf_node_as_loop(node
);
1709 return nir_loop_last_block(loop
);
1712 case nir_cf_node_block
: {
1713 return nir_cf_node_as_block(node
);
1717 unreachable("unknown node type");
1721 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
)
1723 if (node
->type
== nir_cf_node_block
)
1724 return nir_block_cf_tree_next(nir_cf_node_as_block(node
));
1725 else if (node
->type
== nir_cf_node_function
)
1728 return nir_cf_node_as_block(nir_cf_node_next(node
));
1732 nir_block_get_following_if(nir_block
*block
)
1734 if (exec_node_is_tail_sentinel(&block
->cf_node
.node
))
1737 if (nir_cf_node_is_last(&block
->cf_node
))
1740 nir_cf_node
*next_node
= nir_cf_node_next(&block
->cf_node
);
1742 if (next_node
->type
!= nir_cf_node_if
)
1745 return nir_cf_node_as_if(next_node
);
1749 nir_block_get_following_loop(nir_block
*block
)
1751 if (exec_node_is_tail_sentinel(&block
->cf_node
.node
))
1754 if (nir_cf_node_is_last(&block
->cf_node
))
1757 nir_cf_node
*next_node
= nir_cf_node_next(&block
->cf_node
);
1759 if (next_node
->type
!= nir_cf_node_loop
)
1762 return nir_cf_node_as_loop(next_node
);
1766 nir_index_blocks(nir_function_impl
*impl
)
1770 if (impl
->valid_metadata
& nir_metadata_block_index
)
1773 nir_foreach_block(block
, impl
) {
1774 block
->index
= index
++;
1777 /* The end_block isn't really part of the program, which is why its index
1780 impl
->num_blocks
= impl
->end_block
->index
= index
;
1784 index_ssa_def_cb(nir_ssa_def
*def
, void *state
)
1786 unsigned *index
= (unsigned *) state
;
1787 def
->index
= (*index
)++;
1793 * The indices are applied top-to-bottom which has the very nice property
1794 * that, if A dominates B, then A->index <= B->index.
1797 nir_index_ssa_defs(nir_function_impl
*impl
)
1801 nir_foreach_block(block
, impl
) {
1802 nir_foreach_instr(instr
, block
)
1803 nir_foreach_ssa_def(instr
, index_ssa_def_cb
, &index
);
1806 impl
->ssa_alloc
= index
;
1810 * The indices are applied top-to-bottom which has the very nice property
1811 * that, if A dominates B, then A->index <= B->index.
1814 nir_index_instrs(nir_function_impl
*impl
)
1818 nir_foreach_block(block
, impl
) {
1819 nir_foreach_instr(instr
, block
)
1820 instr
->index
= index
++;
1827 nir_shader_index_vars(nir_shader
*shader
, nir_variable_mode modes
)
1830 nir_foreach_variable_with_modes(var
, shader
, modes
)
1831 var
->index
= count
++;
1836 nir_function_impl_index_vars(nir_function_impl
*impl
)
1839 nir_foreach_function_temp_variable(var
, impl
)
1840 var
->index
= count
++;
1845 cursor_next_instr(nir_cursor cursor
)
1847 switch (cursor
.option
) {
1848 case nir_cursor_before_block
:
1849 for (nir_block
*block
= cursor
.block
; block
;
1850 block
= nir_block_cf_tree_next(block
)) {
1851 nir_instr
*instr
= nir_block_first_instr(block
);
1857 case nir_cursor_after_block
:
1858 cursor
.block
= nir_block_cf_tree_next(cursor
.block
);
1859 if (cursor
.block
== NULL
)
1862 cursor
.option
= nir_cursor_before_block
;
1863 return cursor_next_instr(cursor
);
1865 case nir_cursor_before_instr
:
1866 return cursor
.instr
;
1868 case nir_cursor_after_instr
:
1869 if (nir_instr_next(cursor
.instr
))
1870 return nir_instr_next(cursor
.instr
);
1872 cursor
.option
= nir_cursor_after_block
;
1873 cursor
.block
= cursor
.instr
->block
;
1874 return cursor_next_instr(cursor
);
1877 unreachable("Inavlid cursor option");
1880 ASSERTED
static bool
1881 dest_is_ssa(nir_dest
*dest
, void *_state
)
1884 return dest
->is_ssa
;
1888 nir_function_impl_lower_instructions(nir_function_impl
*impl
,
1889 nir_instr_filter_cb filter
,
1890 nir_lower_instr_cb lower
,
1894 nir_builder_init(&b
, impl
);
1896 nir_metadata preserved
= nir_metadata_block_index
|
1897 nir_metadata_dominance
;
1899 bool progress
= false;
1900 nir_cursor iter
= nir_before_cf_list(&impl
->body
);
1902 while ((instr
= cursor_next_instr(iter
)) != NULL
) {
1903 if (filter
&& !filter(instr
, cb_data
)) {
1904 iter
= nir_after_instr(instr
);
1908 assert(nir_foreach_dest(instr
, dest_is_ssa
, NULL
));
1909 nir_ssa_def
*old_def
= nir_instr_ssa_def(instr
);
1910 if (old_def
== NULL
) {
1911 iter
= nir_after_instr(instr
);
1915 /* We're about to ask the callback to generate a replacement for instr.
1916 * Save off the uses from instr's SSA def so we know what uses to
1917 * rewrite later. If we use nir_ssa_def_rewrite_uses, it fails in the
1918 * case where the generated replacement code uses the result of instr
1919 * itself. If we use nir_ssa_def_rewrite_uses_after (which is the
1920 * normal solution to this problem), it doesn't work well if control-
1921 * flow is inserted as part of the replacement, doesn't handle cases
1922 * where the replacement is something consumed by instr, and suffers
1923 * from performance issues. This is the only way to 100% guarantee
1924 * that we rewrite the correct set efficiently.
1926 struct list_head old_uses
, old_if_uses
;
1927 list_replace(&old_def
->uses
, &old_uses
);
1928 list_inithead(&old_def
->uses
);
1929 list_replace(&old_def
->if_uses
, &old_if_uses
);
1930 list_inithead(&old_def
->if_uses
);
1932 b
.cursor
= nir_after_instr(instr
);
1933 nir_ssa_def
*new_def
= lower(&b
, instr
, cb_data
);
1934 if (new_def
&& new_def
!= NIR_LOWER_INSTR_PROGRESS
) {
1935 assert(old_def
!= NULL
);
1936 if (new_def
->parent_instr
->block
!= instr
->block
)
1937 preserved
= nir_metadata_none
;
1939 nir_src new_src
= nir_src_for_ssa(new_def
);
1940 list_for_each_entry_safe(nir_src
, use_src
, &old_uses
, use_link
)
1941 nir_instr_rewrite_src(use_src
->parent_instr
, use_src
, new_src
);
1943 list_for_each_entry_safe(nir_src
, use_src
, &old_if_uses
, use_link
)
1944 nir_if_rewrite_condition(use_src
->parent_if
, new_src
);
1946 if (list_is_empty(&old_def
->uses
) && list_is_empty(&old_def
->if_uses
)) {
1947 iter
= nir_instr_remove(instr
);
1949 iter
= nir_after_instr(instr
);
1953 /* We didn't end up lowering after all. Put the uses back */
1955 list_replace(&old_uses
, &old_def
->uses
);
1956 list_replace(&old_if_uses
, &old_def
->if_uses
);
1958 iter
= nir_after_instr(instr
);
1960 if (new_def
== NIR_LOWER_INSTR_PROGRESS
)
1966 nir_metadata_preserve(impl
, preserved
);
1968 nir_metadata_preserve(impl
, nir_metadata_all
);
1975 nir_shader_lower_instructions(nir_shader
*shader
,
1976 nir_instr_filter_cb filter
,
1977 nir_lower_instr_cb lower
,
1980 bool progress
= false;
1982 nir_foreach_function(function
, shader
) {
1983 if (function
->impl
&&
1984 nir_function_impl_lower_instructions(function
->impl
,
1985 filter
, lower
, cb_data
))
1993 nir_intrinsic_from_system_value(gl_system_value val
)
1996 case SYSTEM_VALUE_VERTEX_ID
:
1997 return nir_intrinsic_load_vertex_id
;
1998 case SYSTEM_VALUE_INSTANCE_ID
:
1999 return nir_intrinsic_load_instance_id
;
2000 case SYSTEM_VALUE_DRAW_ID
:
2001 return nir_intrinsic_load_draw_id
;
2002 case SYSTEM_VALUE_BASE_INSTANCE
:
2003 return nir_intrinsic_load_base_instance
;
2004 case SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
:
2005 return nir_intrinsic_load_vertex_id_zero_base
;
2006 case SYSTEM_VALUE_IS_INDEXED_DRAW
:
2007 return nir_intrinsic_load_is_indexed_draw
;
2008 case SYSTEM_VALUE_FIRST_VERTEX
:
2009 return nir_intrinsic_load_first_vertex
;
2010 case SYSTEM_VALUE_BASE_VERTEX
:
2011 return nir_intrinsic_load_base_vertex
;
2012 case SYSTEM_VALUE_INVOCATION_ID
:
2013 return nir_intrinsic_load_invocation_id
;
2014 case SYSTEM_VALUE_FRAG_COORD
:
2015 return nir_intrinsic_load_frag_coord
;
2016 case SYSTEM_VALUE_POINT_COORD
:
2017 return nir_intrinsic_load_point_coord
;
2018 case SYSTEM_VALUE_LINE_COORD
:
2019 return nir_intrinsic_load_line_coord
;
2020 case SYSTEM_VALUE_FRONT_FACE
:
2021 return nir_intrinsic_load_front_face
;
2022 case SYSTEM_VALUE_SAMPLE_ID
:
2023 return nir_intrinsic_load_sample_id
;
2024 case SYSTEM_VALUE_SAMPLE_POS
:
2025 return nir_intrinsic_load_sample_pos
;
2026 case SYSTEM_VALUE_SAMPLE_MASK_IN
:
2027 return nir_intrinsic_load_sample_mask_in
;
2028 case SYSTEM_VALUE_LOCAL_INVOCATION_ID
:
2029 return nir_intrinsic_load_local_invocation_id
;
2030 case SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
:
2031 return nir_intrinsic_load_local_invocation_index
;
2032 case SYSTEM_VALUE_WORK_GROUP_ID
:
2033 return nir_intrinsic_load_work_group_id
;
2034 case SYSTEM_VALUE_NUM_WORK_GROUPS
:
2035 return nir_intrinsic_load_num_work_groups
;
2036 case SYSTEM_VALUE_PRIMITIVE_ID
:
2037 return nir_intrinsic_load_primitive_id
;
2038 case SYSTEM_VALUE_TESS_COORD
:
2039 return nir_intrinsic_load_tess_coord
;
2040 case SYSTEM_VALUE_TESS_LEVEL_OUTER
:
2041 return nir_intrinsic_load_tess_level_outer
;
2042 case SYSTEM_VALUE_TESS_LEVEL_INNER
:
2043 return nir_intrinsic_load_tess_level_inner
;
2044 case SYSTEM_VALUE_TESS_LEVEL_OUTER_DEFAULT
:
2045 return nir_intrinsic_load_tess_level_outer_default
;
2046 case SYSTEM_VALUE_TESS_LEVEL_INNER_DEFAULT
:
2047 return nir_intrinsic_load_tess_level_inner_default
;
2048 case SYSTEM_VALUE_VERTICES_IN
:
2049 return nir_intrinsic_load_patch_vertices_in
;
2050 case SYSTEM_VALUE_HELPER_INVOCATION
:
2051 return nir_intrinsic_load_helper_invocation
;
2052 case SYSTEM_VALUE_COLOR0
:
2053 return nir_intrinsic_load_color0
;
2054 case SYSTEM_VALUE_COLOR1
:
2055 return nir_intrinsic_load_color1
;
2056 case SYSTEM_VALUE_VIEW_INDEX
:
2057 return nir_intrinsic_load_view_index
;
2058 case SYSTEM_VALUE_SUBGROUP_SIZE
:
2059 return nir_intrinsic_load_subgroup_size
;
2060 case SYSTEM_VALUE_SUBGROUP_INVOCATION
:
2061 return nir_intrinsic_load_subgroup_invocation
;
2062 case SYSTEM_VALUE_SUBGROUP_EQ_MASK
:
2063 return nir_intrinsic_load_subgroup_eq_mask
;
2064 case SYSTEM_VALUE_SUBGROUP_GE_MASK
:
2065 return nir_intrinsic_load_subgroup_ge_mask
;
2066 case SYSTEM_VALUE_SUBGROUP_GT_MASK
:
2067 return nir_intrinsic_load_subgroup_gt_mask
;
2068 case SYSTEM_VALUE_SUBGROUP_LE_MASK
:
2069 return nir_intrinsic_load_subgroup_le_mask
;
2070 case SYSTEM_VALUE_SUBGROUP_LT_MASK
:
2071 return nir_intrinsic_load_subgroup_lt_mask
;
2072 case SYSTEM_VALUE_NUM_SUBGROUPS
:
2073 return nir_intrinsic_load_num_subgroups
;
2074 case SYSTEM_VALUE_SUBGROUP_ID
:
2075 return nir_intrinsic_load_subgroup_id
;
2076 case SYSTEM_VALUE_LOCAL_GROUP_SIZE
:
2077 return nir_intrinsic_load_local_group_size
;
2078 case SYSTEM_VALUE_GLOBAL_INVOCATION_ID
:
2079 return nir_intrinsic_load_global_invocation_id
;
2080 case SYSTEM_VALUE_GLOBAL_INVOCATION_INDEX
:
2081 return nir_intrinsic_load_global_invocation_index
;
2082 case SYSTEM_VALUE_WORK_DIM
:
2083 return nir_intrinsic_load_work_dim
;
2084 case SYSTEM_VALUE_USER_DATA_AMD
:
2085 return nir_intrinsic_load_user_data_amd
;
2087 unreachable("system value does not directly correspond to intrinsic");
2092 nir_system_value_from_intrinsic(nir_intrinsic_op intrin
)
2095 case nir_intrinsic_load_vertex_id
:
2096 return SYSTEM_VALUE_VERTEX_ID
;
2097 case nir_intrinsic_load_instance_id
:
2098 return SYSTEM_VALUE_INSTANCE_ID
;
2099 case nir_intrinsic_load_draw_id
:
2100 return SYSTEM_VALUE_DRAW_ID
;
2101 case nir_intrinsic_load_base_instance
:
2102 return SYSTEM_VALUE_BASE_INSTANCE
;
2103 case nir_intrinsic_load_vertex_id_zero_base
:
2104 return SYSTEM_VALUE_VERTEX_ID_ZERO_BASE
;
2105 case nir_intrinsic_load_first_vertex
:
2106 return SYSTEM_VALUE_FIRST_VERTEX
;
2107 case nir_intrinsic_load_is_indexed_draw
:
2108 return SYSTEM_VALUE_IS_INDEXED_DRAW
;
2109 case nir_intrinsic_load_base_vertex
:
2110 return SYSTEM_VALUE_BASE_VERTEX
;
2111 case nir_intrinsic_load_invocation_id
:
2112 return SYSTEM_VALUE_INVOCATION_ID
;
2113 case nir_intrinsic_load_frag_coord
:
2114 return SYSTEM_VALUE_FRAG_COORD
;
2115 case nir_intrinsic_load_point_coord
:
2116 return SYSTEM_VALUE_POINT_COORD
;
2117 case nir_intrinsic_load_line_coord
:
2118 return SYSTEM_VALUE_LINE_COORD
;
2119 case nir_intrinsic_load_front_face
:
2120 return SYSTEM_VALUE_FRONT_FACE
;
2121 case nir_intrinsic_load_sample_id
:
2122 return SYSTEM_VALUE_SAMPLE_ID
;
2123 case nir_intrinsic_load_sample_pos
:
2124 return SYSTEM_VALUE_SAMPLE_POS
;
2125 case nir_intrinsic_load_sample_mask_in
:
2126 return SYSTEM_VALUE_SAMPLE_MASK_IN
;
2127 case nir_intrinsic_load_local_invocation_id
:
2128 return SYSTEM_VALUE_LOCAL_INVOCATION_ID
;
2129 case nir_intrinsic_load_local_invocation_index
:
2130 return SYSTEM_VALUE_LOCAL_INVOCATION_INDEX
;
2131 case nir_intrinsic_load_num_work_groups
:
2132 return SYSTEM_VALUE_NUM_WORK_GROUPS
;
2133 case nir_intrinsic_load_work_group_id
:
2134 return SYSTEM_VALUE_WORK_GROUP_ID
;
2135 case nir_intrinsic_load_primitive_id
:
2136 return SYSTEM_VALUE_PRIMITIVE_ID
;
2137 case nir_intrinsic_load_tess_coord
:
2138 return SYSTEM_VALUE_TESS_COORD
;
2139 case nir_intrinsic_load_tess_level_outer
:
2140 return SYSTEM_VALUE_TESS_LEVEL_OUTER
;
2141 case nir_intrinsic_load_tess_level_inner
:
2142 return SYSTEM_VALUE_TESS_LEVEL_INNER
;
2143 case nir_intrinsic_load_tess_level_outer_default
:
2144 return SYSTEM_VALUE_TESS_LEVEL_OUTER_DEFAULT
;
2145 case nir_intrinsic_load_tess_level_inner_default
:
2146 return SYSTEM_VALUE_TESS_LEVEL_INNER_DEFAULT
;
2147 case nir_intrinsic_load_patch_vertices_in
:
2148 return SYSTEM_VALUE_VERTICES_IN
;
2149 case nir_intrinsic_load_helper_invocation
:
2150 return SYSTEM_VALUE_HELPER_INVOCATION
;
2151 case nir_intrinsic_load_color0
:
2152 return SYSTEM_VALUE_COLOR0
;
2153 case nir_intrinsic_load_color1
:
2154 return SYSTEM_VALUE_COLOR1
;
2155 case nir_intrinsic_load_view_index
:
2156 return SYSTEM_VALUE_VIEW_INDEX
;
2157 case nir_intrinsic_load_subgroup_size
:
2158 return SYSTEM_VALUE_SUBGROUP_SIZE
;
2159 case nir_intrinsic_load_subgroup_invocation
:
2160 return SYSTEM_VALUE_SUBGROUP_INVOCATION
;
2161 case nir_intrinsic_load_subgroup_eq_mask
:
2162 return SYSTEM_VALUE_SUBGROUP_EQ_MASK
;
2163 case nir_intrinsic_load_subgroup_ge_mask
:
2164 return SYSTEM_VALUE_SUBGROUP_GE_MASK
;
2165 case nir_intrinsic_load_subgroup_gt_mask
:
2166 return SYSTEM_VALUE_SUBGROUP_GT_MASK
;
2167 case nir_intrinsic_load_subgroup_le_mask
:
2168 return SYSTEM_VALUE_SUBGROUP_LE_MASK
;
2169 case nir_intrinsic_load_subgroup_lt_mask
:
2170 return SYSTEM_VALUE_SUBGROUP_LT_MASK
;
2171 case nir_intrinsic_load_num_subgroups
:
2172 return SYSTEM_VALUE_NUM_SUBGROUPS
;
2173 case nir_intrinsic_load_subgroup_id
:
2174 return SYSTEM_VALUE_SUBGROUP_ID
;
2175 case nir_intrinsic_load_local_group_size
:
2176 return SYSTEM_VALUE_LOCAL_GROUP_SIZE
;
2177 case nir_intrinsic_load_global_invocation_id
:
2178 return SYSTEM_VALUE_GLOBAL_INVOCATION_ID
;
2179 case nir_intrinsic_load_user_data_amd
:
2180 return SYSTEM_VALUE_USER_DATA_AMD
;
2182 unreachable("intrinsic doesn't produce a system value");
2186 /* OpenGL utility method that remaps the location attributes if they are
2187 * doubles. Not needed for vulkan due the differences on the input location
2188 * count for doubles on vulkan vs OpenGL
2190 * The bitfield returned in dual_slot is one bit for each double input slot in
2191 * the original OpenGL single-slot input numbering. The mapping from old
2192 * locations to new locations is as follows:
2194 * new_loc = loc + util_bitcount(dual_slot & BITFIELD64_MASK(loc))
2197 nir_remap_dual_slot_attributes(nir_shader
*shader
, uint64_t *dual_slot
)
2199 assert(shader
->info
.stage
== MESA_SHADER_VERTEX
);
2202 nir_foreach_shader_in_variable(var
, shader
) {
2203 if (glsl_type_is_dual_slot(glsl_without_array(var
->type
))) {
2204 unsigned slots
= glsl_count_attribute_slots(var
->type
, true);
2205 *dual_slot
|= BITFIELD64_MASK(slots
) << var
->data
.location
;
2209 nir_foreach_shader_in_variable(var
, shader
) {
2210 var
->data
.location
+=
2211 util_bitcount64(*dual_slot
& BITFIELD64_MASK(var
->data
.location
));
2215 /* Returns an attribute mask that has been re-compacted using the given
2219 nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
)
2222 unsigned loc
= u_bit_scan64(&dual_slot
);
2223 /* mask of all bits up to and including loc */
2224 uint64_t mask
= BITFIELD64_MASK(loc
+ 1);
2225 attribs
= (attribs
& mask
) | ((attribs
& ~mask
) >> 1);
2231 nir_rewrite_image_intrinsic(nir_intrinsic_instr
*intrin
, nir_ssa_def
*src
,
2234 enum gl_access_qualifier access
= nir_intrinsic_access(intrin
);
2236 switch (intrin
->intrinsic
) {
2238 case nir_intrinsic_image_deref_##op: \
2239 intrin->intrinsic = bindless ? nir_intrinsic_bindless_image_##op \
2240 : nir_intrinsic_image_##op; \
2252 CASE(atomic_exchange
)
2253 CASE(atomic_comp_swap
)
2255 CASE(atomic_inc_wrap
)
2256 CASE(atomic_dec_wrap
)
2259 CASE(load_raw_intel
)
2260 CASE(store_raw_intel
)
2263 unreachable("Unhanded image intrinsic");
2266 nir_deref_instr
*deref
= nir_src_as_deref(intrin
->src
[0]);
2267 nir_variable
*var
= nir_deref_instr_get_variable(deref
);
2269 nir_intrinsic_set_image_dim(intrin
, glsl_get_sampler_dim(deref
->type
));
2270 nir_intrinsic_set_image_array(intrin
, glsl_sampler_type_is_array(deref
->type
));
2271 nir_intrinsic_set_access(intrin
, access
| var
->data
.access
);
2272 nir_intrinsic_set_format(intrin
, var
->data
.image
.format
);
2274 nir_instr_rewrite_src(&intrin
->instr
, &intrin
->src
[0],
2275 nir_src_for_ssa(src
));
2279 nir_image_intrinsic_coord_components(const nir_intrinsic_instr
*instr
)
2281 enum glsl_sampler_dim dim
= nir_intrinsic_image_dim(instr
);
2282 int coords
= glsl_get_sampler_dim_coordinate_components(dim
);
2283 if (dim
== GLSL_SAMPLER_DIM_CUBE
)
2286 return coords
+ nir_intrinsic_image_array(instr
);