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)
28 #include "glsl_to_nir.h"
29 #include "ir_visitor.h"
30 #include "ir_hierarchical_visitor.h"
32 #include "compiler/nir/nir_control_flow.h"
33 #include "compiler/nir/nir_builder.h"
34 #include "main/imports.h"
35 #include "main/mtypes.h"
38 * pass to lower GLSL IR to NIR
40 * This will lower variable dereferences to loads/stores of corresponding
41 * variables in NIR - the variables will be converted to registers in a later
47 class nir_visitor
: public ir_visitor
50 nir_visitor(nir_shader
*shader
);
53 virtual void visit(ir_variable
*);
54 virtual void visit(ir_function
*);
55 virtual void visit(ir_function_signature
*);
56 virtual void visit(ir_loop
*);
57 virtual void visit(ir_if
*);
58 virtual void visit(ir_discard
*);
59 virtual void visit(ir_loop_jump
*);
60 virtual void visit(ir_return
*);
61 virtual void visit(ir_call
*);
62 virtual void visit(ir_assignment
*);
63 virtual void visit(ir_emit_vertex
*);
64 virtual void visit(ir_end_primitive
*);
65 virtual void visit(ir_expression
*);
66 virtual void visit(ir_swizzle
*);
67 virtual void visit(ir_texture
*);
68 virtual void visit(ir_constant
*);
69 virtual void visit(ir_dereference_variable
*);
70 virtual void visit(ir_dereference_record
*);
71 virtual void visit(ir_dereference_array
*);
72 virtual void visit(ir_barrier
*);
74 void create_function(ir_function_signature
*ir
);
77 void add_instr(nir_instr
*instr
, unsigned num_components
, unsigned bit_size
);
78 nir_ssa_def
*evaluate_rvalue(ir_rvalue
*ir
);
80 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
**srcs
);
81 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
);
82 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
,
84 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
,
85 nir_ssa_def
*src2
, nir_ssa_def
*src3
);
90 nir_function_impl
*impl
;
92 nir_ssa_def
*result
; /* result of the expression tree last visited */
94 nir_deref_var
*evaluate_deref(nir_instr
*mem_ctx
, ir_instruction
*ir
);
96 /* the head of the dereference chain we're creating */
97 nir_deref_var
*deref_head
;
98 /* the tail of the dereference chain we're creating */
99 nir_deref
*deref_tail
;
101 nir_variable
*var
; /* variable created by ir_variable visitor */
103 /* whether the IR we're operating on is per-function or global */
106 /* map of ir_variable -> nir_variable */
107 struct hash_table
*var_table
;
109 /* map of ir_function_signature -> nir_function_overload */
110 struct hash_table
*overload_table
;
114 * This visitor runs before the main visitor, calling create_function() for
115 * each function so that the main visitor can resolve forward references in
119 class nir_function_visitor
: public ir_hierarchical_visitor
122 nir_function_visitor(nir_visitor
*v
) : visitor(v
)
125 virtual ir_visitor_status
visit_enter(ir_function
*);
128 nir_visitor
*visitor
;
131 } /* end of anonymous namespace */
134 nir_remap_attributes(nir_shader
*shader
,
135 const nir_shader_compiler_options
*options
)
137 if (options
->vs_inputs_dual_locations
) {
138 nir_foreach_variable(var
, &shader
->inputs
) {
139 var
->data
.location
+=
140 _mesa_bitcount_64(shader
->info
.vs
.double_inputs
&
141 BITFIELD64_MASK(var
->data
.location
));
145 /* Once the remap is done, reset double_inputs_read, so later it will have
146 * which location/slots are doubles */
147 shader
->info
.vs
.double_inputs
= 0;
151 glsl_to_nir(const struct gl_shader_program
*shader_prog
,
152 gl_shader_stage stage
,
153 const nir_shader_compiler_options
*options
)
155 struct gl_linked_shader
*sh
= shader_prog
->_LinkedShaders
[stage
];
157 nir_shader
*shader
= nir_shader_create(NULL
, stage
, options
,
160 nir_visitor
v1(shader
);
161 nir_function_visitor
v2(&v1
);
163 visit_exec_list(sh
->ir
, &v1
);
165 nir_lower_constant_initializers(shader
, (nir_variable_mode
)~0);
167 /* Remap the locations to slots so those requiring two slots will occupy
168 * two locations. For instance, if we have in the IR code a dvec3 attr0 in
169 * location 0 and vec4 attr1 in location 1, in NIR attr0 will use
170 * locations/slots 0 and 1, and attr1 will use location/slot 2 */
171 if (shader
->info
.stage
== MESA_SHADER_VERTEX
)
172 nir_remap_attributes(shader
, options
);
174 shader
->info
.name
= ralloc_asprintf(shader
, "GLSL%d", shader_prog
->Name
);
175 if (shader_prog
->Label
)
176 shader
->info
.label
= ralloc_strdup(shader
, shader_prog
->Label
);
178 /* Check for transform feedback varyings specified via the API */
179 shader
->info
.has_transform_feedback_varyings
=
180 shader_prog
->TransformFeedback
.NumVarying
> 0;
182 /* Check for transform feedback varyings specified in the Shader */
183 if (shader_prog
->last_vert_prog
)
184 shader
->info
.has_transform_feedback_varyings
|=
185 shader_prog
->last_vert_prog
->sh
.LinkedTransformFeedback
->NumVarying
> 0;
190 nir_visitor::nir_visitor(nir_shader
*shader
)
192 this->supports_ints
= shader
->options
->native_integers
;
193 this->shader
= shader
;
194 this->is_global
= true;
195 this->var_table
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
196 _mesa_key_pointer_equal
);
197 this->overload_table
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
198 _mesa_key_pointer_equal
);
202 this->deref_head
= NULL
;
203 this->deref_tail
= NULL
;
204 memset(&this->b
, 0, sizeof(this->b
));
207 nir_visitor::~nir_visitor()
209 _mesa_hash_table_destroy(this->var_table
, NULL
);
210 _mesa_hash_table_destroy(this->overload_table
, NULL
);
214 nir_visitor::evaluate_deref(nir_instr
*mem_ctx
, ir_instruction
*ir
)
217 ralloc_steal(mem_ctx
, this->deref_head
);
218 return this->deref_head
;
221 static nir_constant
*
222 constant_copy(ir_constant
*ir
, void *mem_ctx
)
227 nir_constant
*ret
= rzalloc(mem_ctx
, nir_constant
);
229 const unsigned rows
= ir
->type
->vector_elements
;
230 const unsigned cols
= ir
->type
->matrix_columns
;
233 ret
->num_elements
= 0;
234 switch (ir
->type
->base_type
) {
236 /* Only float base types can be matrices. */
239 for (unsigned r
= 0; r
< rows
; r
++)
240 ret
->values
[0].u32
[r
] = ir
->value
.u
[r
];
245 /* Only float base types can be matrices. */
248 for (unsigned r
= 0; r
< rows
; r
++)
249 ret
->values
[0].i32
[r
] = ir
->value
.i
[r
];
253 case GLSL_TYPE_FLOAT
:
254 for (unsigned c
= 0; c
< cols
; c
++) {
255 for (unsigned r
= 0; r
< rows
; r
++)
256 ret
->values
[c
].f32
[r
] = ir
->value
.f
[c
* rows
+ r
];
260 case GLSL_TYPE_DOUBLE
:
261 for (unsigned c
= 0; c
< cols
; c
++) {
262 for (unsigned r
= 0; r
< rows
; r
++)
263 ret
->values
[c
].f64
[r
] = ir
->value
.d
[c
* rows
+ r
];
267 case GLSL_TYPE_UINT64
:
268 /* Only float base types can be matrices. */
271 for (unsigned r
= 0; r
< rows
; r
++)
272 ret
->values
[0].u64
[r
] = ir
->value
.u64
[r
];
275 case GLSL_TYPE_INT64
:
276 /* Only float base types can be matrices. */
279 for (unsigned r
= 0; r
< rows
; r
++)
280 ret
->values
[0].i64
[r
] = ir
->value
.i64
[r
];
284 /* Only float base types can be matrices. */
287 for (unsigned r
= 0; r
< rows
; r
++)
288 ret
->values
[0].u32
[r
] = ir
->value
.b
[r
] ? NIR_TRUE
: NIR_FALSE
;
292 case GLSL_TYPE_STRUCT
:
293 case GLSL_TYPE_ARRAY
:
294 ret
->elements
= ralloc_array(mem_ctx
, nir_constant
*,
296 ret
->num_elements
= ir
->type
->length
;
298 for (i
= 0; i
< ir
->type
->length
; i
++)
299 ret
->elements
[i
] = constant_copy(ir
->const_elements
[i
], mem_ctx
);
303 unreachable("not reached");
310 nir_visitor::visit(ir_variable
*ir
)
312 /* TODO: In future we should switch to using the NIR lowering pass but for
313 * now just ignore these variables as GLSL IR should have lowered them.
314 * Anything remaining are just dead vars that weren't cleaned up.
316 if (ir
->data
.mode
== ir_var_shader_shared
)
319 nir_variable
*var
= rzalloc(shader
, nir_variable
);
320 var
->type
= ir
->type
;
321 var
->name
= ralloc_strdup(var
, ir
->name
);
323 var
->data
.always_active_io
= ir
->data
.always_active_io
;
324 var
->data
.read_only
= ir
->data
.read_only
;
325 var
->data
.centroid
= ir
->data
.centroid
;
326 var
->data
.sample
= ir
->data
.sample
;
327 var
->data
.patch
= ir
->data
.patch
;
328 var
->data
.invariant
= ir
->data
.invariant
;
329 var
->data
.location
= ir
->data
.location
;
330 var
->data
.stream
= ir
->data
.stream
;
331 var
->data
.compact
= false;
333 switch(ir
->data
.mode
) {
335 case ir_var_temporary
:
337 var
->data
.mode
= nir_var_global
;
339 var
->data
.mode
= nir_var_local
;
342 case ir_var_function_in
:
343 case ir_var_function_out
:
344 case ir_var_function_inout
:
345 case ir_var_const_in
:
346 var
->data
.mode
= nir_var_local
;
349 case ir_var_shader_in
:
350 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
351 ir
->data
.location
== VARYING_SLOT_FACE
) {
352 /* For whatever reason, GLSL IR makes gl_FrontFacing an input */
353 var
->data
.location
= SYSTEM_VALUE_FRONT_FACE
;
354 var
->data
.mode
= nir_var_system_value
;
355 } else if (shader
->info
.stage
== MESA_SHADER_GEOMETRY
&&
356 ir
->data
.location
== VARYING_SLOT_PRIMITIVE_ID
) {
357 /* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */
358 var
->data
.location
= SYSTEM_VALUE_PRIMITIVE_ID
;
359 var
->data
.mode
= nir_var_system_value
;
361 var
->data
.mode
= nir_var_shader_in
;
363 if (shader
->info
.stage
== MESA_SHADER_TESS_EVAL
&&
364 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
365 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
366 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
370 /* Mark all the locations that require two slots */
371 if (shader
->info
.stage
== MESA_SHADER_VERTEX
&&
372 glsl_type_is_dual_slot(glsl_without_array(var
->type
))) {
373 for (unsigned i
= 0; i
< glsl_count_attribute_slots(var
->type
, true); i
++) {
374 uint64_t bitfield
= BITFIELD64_BIT(var
->data
.location
+ i
);
375 shader
->info
.vs
.double_inputs
|= bitfield
;
380 case ir_var_shader_out
:
381 var
->data
.mode
= nir_var_shader_out
;
382 if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
&&
383 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
384 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
385 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
390 var
->data
.mode
= nir_var_uniform
;
393 case ir_var_shader_storage
:
394 var
->data
.mode
= nir_var_shader_storage
;
397 case ir_var_system_value
:
398 var
->data
.mode
= nir_var_system_value
;
402 unreachable("not reached");
405 var
->data
.interpolation
= ir
->data
.interpolation
;
406 var
->data
.origin_upper_left
= ir
->data
.origin_upper_left
;
407 var
->data
.pixel_center_integer
= ir
->data
.pixel_center_integer
;
408 var
->data
.location_frac
= ir
->data
.location_frac
;
410 if (var
->data
.pixel_center_integer
) {
411 assert(shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
412 shader
->info
.fs
.pixel_center_integer
= true;
415 switch (ir
->data
.depth_layout
) {
416 case ir_depth_layout_none
:
417 var
->data
.depth_layout
= nir_depth_layout_none
;
419 case ir_depth_layout_any
:
420 var
->data
.depth_layout
= nir_depth_layout_any
;
422 case ir_depth_layout_greater
:
423 var
->data
.depth_layout
= nir_depth_layout_greater
;
425 case ir_depth_layout_less
:
426 var
->data
.depth_layout
= nir_depth_layout_less
;
428 case ir_depth_layout_unchanged
:
429 var
->data
.depth_layout
= nir_depth_layout_unchanged
;
432 unreachable("not reached");
435 var
->data
.index
= ir
->data
.index
;
436 var
->data
.descriptor_set
= 0;
437 var
->data
.binding
= ir
->data
.binding
;
438 var
->data
.explicit_binding
= ir
->data
.explicit_binding
;
439 var
->data
.bindless
= ir
->data
.bindless
;
440 var
->data
.offset
= ir
->data
.offset
;
441 var
->data
.image
.read_only
= ir
->data
.memory_read_only
;
442 var
->data
.image
.write_only
= ir
->data
.memory_write_only
;
443 var
->data
.image
.coherent
= ir
->data
.memory_coherent
;
444 var
->data
.image
._volatile
= ir
->data
.memory_volatile
;
445 var
->data
.image
.restrict_flag
= ir
->data
.memory_restrict
;
446 var
->data
.image
.format
= ir
->data
.image_format
;
447 var
->data
.fb_fetch_output
= ir
->data
.fb_fetch_output
;
449 var
->num_state_slots
= ir
->get_num_state_slots();
450 if (var
->num_state_slots
> 0) {
451 var
->state_slots
= rzalloc_array(var
, nir_state_slot
,
452 var
->num_state_slots
);
454 ir_state_slot
*state_slots
= ir
->get_state_slots();
455 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
456 for (unsigned j
= 0; j
< 5; j
++)
457 var
->state_slots
[i
].tokens
[j
] = state_slots
[i
].tokens
[j
];
458 var
->state_slots
[i
].swizzle
= state_slots
[i
].swizzle
;
461 var
->state_slots
= NULL
;
464 var
->constant_initializer
= constant_copy(ir
->constant_initializer
, var
);
466 var
->interface_type
= ir
->get_interface_type();
468 if (var
->data
.mode
== nir_var_local
)
469 nir_function_impl_add_variable(impl
, var
);
471 nir_shader_add_variable(shader
, var
);
473 _mesa_hash_table_insert(var_table
, ir
, var
);
478 nir_function_visitor::visit_enter(ir_function
*ir
)
480 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
) {
481 visitor
->create_function(sig
);
483 return visit_continue_with_parent
;
487 nir_visitor::create_function(ir_function_signature
*ir
)
489 if (ir
->is_intrinsic())
492 nir_function
*func
= nir_function_create(shader
, ir
->function_name());
494 assert(ir
->parameters
.is_empty());
495 assert(ir
->return_type
== glsl_type::void_type
);
497 _mesa_hash_table_insert(this->overload_table
, ir
, func
);
501 nir_visitor::visit(ir_function
*ir
)
503 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
)
508 nir_visitor::visit(ir_function_signature
*ir
)
510 if (ir
->is_intrinsic())
513 struct hash_entry
*entry
=
514 _mesa_hash_table_search(this->overload_table
, ir
);
517 nir_function
*func
= (nir_function
*) entry
->data
;
519 if (ir
->is_defined
) {
520 nir_function_impl
*impl
= nir_function_impl_create(func
);
523 assert(strcmp(func
->name
, "main") == 0);
524 assert(ir
->parameters
.is_empty());
525 assert(func
->return_type
== glsl_type::void_type
);
527 this->is_global
= false;
529 nir_builder_init(&b
, impl
);
530 b
.cursor
= nir_after_cf_list(&impl
->body
);
531 visit_exec_list(&ir
->body
, this);
533 this->is_global
= true;
540 nir_visitor::visit(ir_loop
*ir
)
543 visit_exec_list(&ir
->body_instructions
, this);
544 nir_pop_loop(&b
, NULL
);
548 nir_visitor::visit(ir_if
*ir
)
550 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
551 visit_exec_list(&ir
->then_instructions
, this);
552 nir_push_else(&b
, NULL
);
553 visit_exec_list(&ir
->else_instructions
, this);
554 nir_pop_if(&b
, NULL
);
558 nir_visitor::visit(ir_discard
*ir
)
561 * discards aren't treated as control flow, because before we lower them
562 * they can appear anywhere in the shader and the stuff after them may still
563 * be executed (yay, crazy GLSL rules!). However, after lowering, all the
564 * discards will be immediately followed by a return.
567 nir_intrinsic_instr
*discard
;
569 discard
= nir_intrinsic_instr_create(this->shader
,
570 nir_intrinsic_discard_if
);
572 nir_src_for_ssa(evaluate_rvalue(ir
->condition
));
574 discard
= nir_intrinsic_instr_create(this->shader
, nir_intrinsic_discard
);
577 nir_builder_instr_insert(&b
, &discard
->instr
);
581 nir_visitor::visit(ir_emit_vertex
*ir
)
583 nir_intrinsic_instr
*instr
=
584 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_emit_vertex
);
585 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
586 nir_builder_instr_insert(&b
, &instr
->instr
);
590 nir_visitor::visit(ir_end_primitive
*ir
)
592 nir_intrinsic_instr
*instr
=
593 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_end_primitive
);
594 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
595 nir_builder_instr_insert(&b
, &instr
->instr
);
599 nir_visitor::visit(ir_loop_jump
*ir
)
603 case ir_loop_jump::jump_break
:
604 type
= nir_jump_break
;
606 case ir_loop_jump::jump_continue
:
607 type
= nir_jump_continue
;
610 unreachable("not reached");
613 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, type
);
614 nir_builder_instr_insert(&b
, &instr
->instr
);
618 nir_visitor::visit(ir_return
*ir
)
620 assert(ir
->value
== NULL
);
621 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, nir_jump_return
);
622 nir_builder_instr_insert(&b
, &instr
->instr
);
626 nir_visitor::visit(ir_call
*ir
)
628 if (ir
->callee
->is_intrinsic()) {
631 switch (ir
->callee
->intrinsic_id
) {
632 case ir_intrinsic_atomic_counter_read
:
633 op
= nir_intrinsic_atomic_counter_read_var
;
635 case ir_intrinsic_atomic_counter_increment
:
636 op
= nir_intrinsic_atomic_counter_inc_var
;
638 case ir_intrinsic_atomic_counter_predecrement
:
639 op
= nir_intrinsic_atomic_counter_dec_var
;
641 case ir_intrinsic_atomic_counter_add
:
642 op
= nir_intrinsic_atomic_counter_add_var
;
644 case ir_intrinsic_atomic_counter_and
:
645 op
= nir_intrinsic_atomic_counter_and_var
;
647 case ir_intrinsic_atomic_counter_or
:
648 op
= nir_intrinsic_atomic_counter_or_var
;
650 case ir_intrinsic_atomic_counter_xor
:
651 op
= nir_intrinsic_atomic_counter_xor_var
;
653 case ir_intrinsic_atomic_counter_min
:
654 op
= nir_intrinsic_atomic_counter_min_var
;
656 case ir_intrinsic_atomic_counter_max
:
657 op
= nir_intrinsic_atomic_counter_max_var
;
659 case ir_intrinsic_atomic_counter_exchange
:
660 op
= nir_intrinsic_atomic_counter_exchange_var
;
662 case ir_intrinsic_atomic_counter_comp_swap
:
663 op
= nir_intrinsic_atomic_counter_comp_swap_var
;
665 case ir_intrinsic_image_load
:
666 op
= nir_intrinsic_image_var_load
;
668 case ir_intrinsic_image_store
:
669 op
= nir_intrinsic_image_var_store
;
671 case ir_intrinsic_image_atomic_add
:
672 op
= nir_intrinsic_image_var_atomic_add
;
674 case ir_intrinsic_image_atomic_min
:
675 op
= nir_intrinsic_image_var_atomic_min
;
677 case ir_intrinsic_image_atomic_max
:
678 op
= nir_intrinsic_image_var_atomic_max
;
680 case ir_intrinsic_image_atomic_and
:
681 op
= nir_intrinsic_image_var_atomic_and
;
683 case ir_intrinsic_image_atomic_or
:
684 op
= nir_intrinsic_image_var_atomic_or
;
686 case ir_intrinsic_image_atomic_xor
:
687 op
= nir_intrinsic_image_var_atomic_xor
;
689 case ir_intrinsic_image_atomic_exchange
:
690 op
= nir_intrinsic_image_var_atomic_exchange
;
692 case ir_intrinsic_image_atomic_comp_swap
:
693 op
= nir_intrinsic_image_var_atomic_comp_swap
;
695 case ir_intrinsic_memory_barrier
:
696 op
= nir_intrinsic_memory_barrier
;
698 case ir_intrinsic_image_size
:
699 op
= nir_intrinsic_image_var_size
;
701 case ir_intrinsic_image_samples
:
702 op
= nir_intrinsic_image_var_samples
;
704 case ir_intrinsic_ssbo_store
:
705 op
= nir_intrinsic_store_ssbo
;
707 case ir_intrinsic_ssbo_load
:
708 op
= nir_intrinsic_load_ssbo
;
710 case ir_intrinsic_ssbo_atomic_add
:
711 op
= nir_intrinsic_ssbo_atomic_add
;
713 case ir_intrinsic_ssbo_atomic_and
:
714 op
= nir_intrinsic_ssbo_atomic_and
;
716 case ir_intrinsic_ssbo_atomic_or
:
717 op
= nir_intrinsic_ssbo_atomic_or
;
719 case ir_intrinsic_ssbo_atomic_xor
:
720 op
= nir_intrinsic_ssbo_atomic_xor
;
722 case ir_intrinsic_ssbo_atomic_min
:
723 assert(ir
->return_deref
);
724 if (ir
->return_deref
->type
== glsl_type::int_type
)
725 op
= nir_intrinsic_ssbo_atomic_imin
;
726 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
727 op
= nir_intrinsic_ssbo_atomic_umin
;
729 unreachable("Invalid type");
731 case ir_intrinsic_ssbo_atomic_max
:
732 assert(ir
->return_deref
);
733 if (ir
->return_deref
->type
== glsl_type::int_type
)
734 op
= nir_intrinsic_ssbo_atomic_imax
;
735 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
736 op
= nir_intrinsic_ssbo_atomic_umax
;
738 unreachable("Invalid type");
740 case ir_intrinsic_ssbo_atomic_exchange
:
741 op
= nir_intrinsic_ssbo_atomic_exchange
;
743 case ir_intrinsic_ssbo_atomic_comp_swap
:
744 op
= nir_intrinsic_ssbo_atomic_comp_swap
;
746 case ir_intrinsic_shader_clock
:
747 op
= nir_intrinsic_shader_clock
;
749 case ir_intrinsic_begin_invocation_interlock
:
750 op
= nir_intrinsic_begin_invocation_interlock
;
752 case ir_intrinsic_end_invocation_interlock
:
753 op
= nir_intrinsic_end_invocation_interlock
;
755 case ir_intrinsic_group_memory_barrier
:
756 op
= nir_intrinsic_group_memory_barrier
;
758 case ir_intrinsic_memory_barrier_atomic_counter
:
759 op
= nir_intrinsic_memory_barrier_atomic_counter
;
761 case ir_intrinsic_memory_barrier_buffer
:
762 op
= nir_intrinsic_memory_barrier_buffer
;
764 case ir_intrinsic_memory_barrier_image
:
765 op
= nir_intrinsic_memory_barrier_image
;
767 case ir_intrinsic_memory_barrier_shared
:
768 op
= nir_intrinsic_memory_barrier_shared
;
770 case ir_intrinsic_shared_load
:
771 op
= nir_intrinsic_load_shared
;
773 case ir_intrinsic_shared_store
:
774 op
= nir_intrinsic_store_shared
;
776 case ir_intrinsic_shared_atomic_add
:
777 op
= nir_intrinsic_shared_atomic_add
;
779 case ir_intrinsic_shared_atomic_and
:
780 op
= nir_intrinsic_shared_atomic_and
;
782 case ir_intrinsic_shared_atomic_or
:
783 op
= nir_intrinsic_shared_atomic_or
;
785 case ir_intrinsic_shared_atomic_xor
:
786 op
= nir_intrinsic_shared_atomic_xor
;
788 case ir_intrinsic_shared_atomic_min
:
789 assert(ir
->return_deref
);
790 if (ir
->return_deref
->type
== glsl_type::int_type
)
791 op
= nir_intrinsic_shared_atomic_imin
;
792 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
793 op
= nir_intrinsic_shared_atomic_umin
;
795 unreachable("Invalid type");
797 case ir_intrinsic_shared_atomic_max
:
798 assert(ir
->return_deref
);
799 if (ir
->return_deref
->type
== glsl_type::int_type
)
800 op
= nir_intrinsic_shared_atomic_imax
;
801 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
802 op
= nir_intrinsic_shared_atomic_umax
;
804 unreachable("Invalid type");
806 case ir_intrinsic_shared_atomic_exchange
:
807 op
= nir_intrinsic_shared_atomic_exchange
;
809 case ir_intrinsic_shared_atomic_comp_swap
:
810 op
= nir_intrinsic_shared_atomic_comp_swap
;
812 case ir_intrinsic_vote_any
:
813 op
= nir_intrinsic_vote_any
;
815 case ir_intrinsic_vote_all
:
816 op
= nir_intrinsic_vote_all
;
818 case ir_intrinsic_vote_eq
:
819 op
= nir_intrinsic_vote_ieq
;
821 case ir_intrinsic_ballot
:
822 op
= nir_intrinsic_ballot
;
824 case ir_intrinsic_read_invocation
:
825 op
= nir_intrinsic_read_invocation
;
827 case ir_intrinsic_read_first_invocation
:
828 op
= nir_intrinsic_read_first_invocation
;
831 unreachable("not reached");
834 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(shader
, op
);
835 nir_dest
*dest
= &instr
->dest
;
838 case nir_intrinsic_atomic_counter_read_var
:
839 case nir_intrinsic_atomic_counter_inc_var
:
840 case nir_intrinsic_atomic_counter_dec_var
:
841 case nir_intrinsic_atomic_counter_add_var
:
842 case nir_intrinsic_atomic_counter_min_var
:
843 case nir_intrinsic_atomic_counter_max_var
:
844 case nir_intrinsic_atomic_counter_and_var
:
845 case nir_intrinsic_atomic_counter_or_var
:
846 case nir_intrinsic_atomic_counter_xor_var
:
847 case nir_intrinsic_atomic_counter_exchange_var
:
848 case nir_intrinsic_atomic_counter_comp_swap_var
: {
849 /* Set the counter variable dereference. */
850 exec_node
*param
= ir
->actual_parameters
.get_head();
851 ir_dereference
*counter
= (ir_dereference
*)param
;
853 instr
->variables
[0] = evaluate_deref(&instr
->instr
, counter
);
854 param
= param
->get_next();
856 /* Set the intrinsic destination. */
857 if (ir
->return_deref
) {
858 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
861 /* Set the intrinsic parameters. */
862 if (!param
->is_tail_sentinel()) {
864 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
865 param
= param
->get_next();
868 if (!param
->is_tail_sentinel()) {
870 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
871 param
= param
->get_next();
874 nir_builder_instr_insert(&b
, &instr
->instr
);
877 case nir_intrinsic_image_var_load
:
878 case nir_intrinsic_image_var_store
:
879 case nir_intrinsic_image_var_atomic_add
:
880 case nir_intrinsic_image_var_atomic_min
:
881 case nir_intrinsic_image_var_atomic_max
:
882 case nir_intrinsic_image_var_atomic_and
:
883 case nir_intrinsic_image_var_atomic_or
:
884 case nir_intrinsic_image_var_atomic_xor
:
885 case nir_intrinsic_image_var_atomic_exchange
:
886 case nir_intrinsic_image_var_atomic_comp_swap
:
887 case nir_intrinsic_image_var_samples
:
888 case nir_intrinsic_image_var_size
: {
889 nir_ssa_undef_instr
*instr_undef
=
890 nir_ssa_undef_instr_create(shader
, 1, 32);
891 nir_builder_instr_insert(&b
, &instr_undef
->instr
);
893 /* Set the image variable dereference. */
894 exec_node
*param
= ir
->actual_parameters
.get_head();
895 ir_dereference
*image
= (ir_dereference
*)param
;
896 const glsl_type
*type
=
897 image
->variable_referenced()->type
->without_array();
899 instr
->variables
[0] = evaluate_deref(&instr
->instr
, image
);
900 param
= param
->get_next();
902 /* Set the intrinsic destination. */
903 if (ir
->return_deref
) {
904 unsigned num_components
= ir
->return_deref
->type
->vector_elements
;
905 if (instr
->intrinsic
== nir_intrinsic_image_var_size
)
906 instr
->num_components
= num_components
;
907 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
908 num_components
, 32, NULL
);
911 if (op
== nir_intrinsic_image_var_size
||
912 op
== nir_intrinsic_image_var_samples
) {
913 nir_builder_instr_insert(&b
, &instr
->instr
);
917 /* Set the address argument, extending the coordinate vector to four
920 nir_ssa_def
*src_addr
=
921 evaluate_rvalue((ir_dereference
*)param
);
922 nir_ssa_def
*srcs
[4];
924 for (int i
= 0; i
< 4; i
++) {
925 if (i
< type
->coordinate_components())
926 srcs
[i
] = nir_channel(&b
, src_addr
, i
);
928 srcs
[i
] = &instr_undef
->def
;
931 instr
->src
[0] = nir_src_for_ssa(nir_vec(&b
, srcs
, 4));
932 param
= param
->get_next();
934 /* Set the sample argument, which is undefined for single-sample
937 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_MS
) {
939 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
940 param
= param
->get_next();
942 instr
->src
[1] = nir_src_for_ssa(&instr_undef
->def
);
945 /* Set the intrinsic parameters. */
946 if (!param
->is_tail_sentinel()) {
948 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
949 param
= param
->get_next();
952 if (!param
->is_tail_sentinel()) {
954 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
955 param
= param
->get_next();
957 nir_builder_instr_insert(&b
, &instr
->instr
);
960 case nir_intrinsic_memory_barrier
:
961 case nir_intrinsic_group_memory_barrier
:
962 case nir_intrinsic_memory_barrier_atomic_counter
:
963 case nir_intrinsic_memory_barrier_buffer
:
964 case nir_intrinsic_memory_barrier_image
:
965 case nir_intrinsic_memory_barrier_shared
:
966 nir_builder_instr_insert(&b
, &instr
->instr
);
968 case nir_intrinsic_shader_clock
:
969 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 2, 32, NULL
);
970 instr
->num_components
= 2;
971 nir_builder_instr_insert(&b
, &instr
->instr
);
973 case nir_intrinsic_begin_invocation_interlock
:
974 nir_builder_instr_insert(&b
, &instr
->instr
);
976 case nir_intrinsic_end_invocation_interlock
:
977 nir_builder_instr_insert(&b
, &instr
->instr
);
979 case nir_intrinsic_store_ssbo
: {
980 exec_node
*param
= ir
->actual_parameters
.get_head();
981 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
983 param
= param
->get_next();
984 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
986 param
= param
->get_next();
987 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
989 param
= param
->get_next();
990 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
993 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(val
));
994 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(block
));
995 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(offset
));
996 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
997 instr
->num_components
= val
->type
->vector_elements
;
999 nir_builder_instr_insert(&b
, &instr
->instr
);
1002 case nir_intrinsic_load_ssbo
: {
1003 exec_node
*param
= ir
->actual_parameters
.get_head();
1004 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1006 param
= param
->get_next();
1007 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1009 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(block
));
1010 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1012 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1013 instr
->num_components
= type
->vector_elements
;
1015 /* Setup destination register */
1016 unsigned bit_size
= glsl_get_bit_size(type
);
1017 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1018 type
->vector_elements
, bit_size
, NULL
);
1020 /* Insert the created nir instruction now since in the case of boolean
1021 * result we will need to emit another instruction after it
1023 nir_builder_instr_insert(&b
, &instr
->instr
);
1026 * In SSBO/UBO's, a true boolean value is any non-zero value, but we
1027 * consider a true boolean to be ~0. Fix this up with a != 0
1030 if (type
->is_boolean()) {
1031 nir_alu_instr
*load_ssbo_compare
=
1032 nir_alu_instr_create(shader
, nir_op_ine
);
1033 load_ssbo_compare
->src
[0].src
.is_ssa
= true;
1034 load_ssbo_compare
->src
[0].src
.ssa
= &instr
->dest
.ssa
;
1035 load_ssbo_compare
->src
[1].src
=
1036 nir_src_for_ssa(nir_imm_int(&b
, 0));
1037 for (unsigned i
= 0; i
< type
->vector_elements
; i
++)
1038 load_ssbo_compare
->src
[1].swizzle
[i
] = 0;
1039 nir_ssa_dest_init(&load_ssbo_compare
->instr
,
1040 &load_ssbo_compare
->dest
.dest
,
1041 type
->vector_elements
, bit_size
, NULL
);
1042 load_ssbo_compare
->dest
.write_mask
= (1 << type
->vector_elements
) - 1;
1043 nir_builder_instr_insert(&b
, &load_ssbo_compare
->instr
);
1044 dest
= &load_ssbo_compare
->dest
.dest
;
1048 case nir_intrinsic_ssbo_atomic_add
:
1049 case nir_intrinsic_ssbo_atomic_imin
:
1050 case nir_intrinsic_ssbo_atomic_umin
:
1051 case nir_intrinsic_ssbo_atomic_imax
:
1052 case nir_intrinsic_ssbo_atomic_umax
:
1053 case nir_intrinsic_ssbo_atomic_and
:
1054 case nir_intrinsic_ssbo_atomic_or
:
1055 case nir_intrinsic_ssbo_atomic_xor
:
1056 case nir_intrinsic_ssbo_atomic_exchange
:
1057 case nir_intrinsic_ssbo_atomic_comp_swap
: {
1058 int param_count
= ir
->actual_parameters
.length();
1059 assert(param_count
== 3 || param_count
== 4);
1062 exec_node
*param
= ir
->actual_parameters
.get_head();
1063 ir_instruction
*inst
= (ir_instruction
*) param
;
1064 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1067 param
= param
->get_next();
1068 inst
= (ir_instruction
*) param
;
1069 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1071 /* data1 parameter (this is always present) */
1072 param
= param
->get_next();
1073 inst
= (ir_instruction
*) param
;
1074 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1076 /* data2 parameter (only with atomic_comp_swap) */
1077 if (param_count
== 4) {
1078 assert(op
== nir_intrinsic_ssbo_atomic_comp_swap
);
1079 param
= param
->get_next();
1080 inst
= (ir_instruction
*) param
;
1081 instr
->src
[3] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1085 assert(ir
->return_deref
);
1086 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1087 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1088 nir_builder_instr_insert(&b
, &instr
->instr
);
1091 case nir_intrinsic_load_shared
: {
1092 exec_node
*param
= ir
->actual_parameters
.get_head();
1093 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1095 nir_intrinsic_set_base(instr
, 0);
1096 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(offset
));
1098 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1099 instr
->num_components
= type
->vector_elements
;
1101 /* Setup destination register */
1102 unsigned bit_size
= glsl_get_bit_size(type
);
1103 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1104 type
->vector_elements
, bit_size
, NULL
);
1106 nir_builder_instr_insert(&b
, &instr
->instr
);
1109 case nir_intrinsic_store_shared
: {
1110 exec_node
*param
= ir
->actual_parameters
.get_head();
1111 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1113 param
= param
->get_next();
1114 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1116 param
= param
->get_next();
1117 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1120 nir_intrinsic_set_base(instr
, 0);
1121 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1123 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1125 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(val
));
1126 instr
->num_components
= val
->type
->vector_elements
;
1128 nir_builder_instr_insert(&b
, &instr
->instr
);
1131 case nir_intrinsic_shared_atomic_add
:
1132 case nir_intrinsic_shared_atomic_imin
:
1133 case nir_intrinsic_shared_atomic_umin
:
1134 case nir_intrinsic_shared_atomic_imax
:
1135 case nir_intrinsic_shared_atomic_umax
:
1136 case nir_intrinsic_shared_atomic_and
:
1137 case nir_intrinsic_shared_atomic_or
:
1138 case nir_intrinsic_shared_atomic_xor
:
1139 case nir_intrinsic_shared_atomic_exchange
:
1140 case nir_intrinsic_shared_atomic_comp_swap
: {
1141 int param_count
= ir
->actual_parameters
.length();
1142 assert(param_count
== 2 || param_count
== 3);
1145 exec_node
*param
= ir
->actual_parameters
.get_head();
1146 ir_instruction
*inst
= (ir_instruction
*) param
;
1147 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1149 /* data1 parameter (this is always present) */
1150 param
= param
->get_next();
1151 inst
= (ir_instruction
*) param
;
1152 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1154 /* data2 parameter (only with atomic_comp_swap) */
1155 if (param_count
== 3) {
1156 assert(op
== nir_intrinsic_shared_atomic_comp_swap
);
1157 param
= param
->get_next();
1158 inst
= (ir_instruction
*) param
;
1160 nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1164 assert(ir
->return_deref
);
1165 unsigned bit_size
= glsl_get_bit_size(ir
->return_deref
->type
);
1166 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1167 ir
->return_deref
->type
->vector_elements
,
1169 nir_builder_instr_insert(&b
, &instr
->instr
);
1172 case nir_intrinsic_vote_any
:
1173 case nir_intrinsic_vote_all
:
1174 case nir_intrinsic_vote_ieq
: {
1175 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1176 instr
->num_components
= 1;
1178 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1179 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1181 nir_builder_instr_insert(&b
, &instr
->instr
);
1185 case nir_intrinsic_ballot
: {
1186 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1187 ir
->return_deref
->type
->vector_elements
, 64, NULL
);
1188 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1190 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1191 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1193 nir_builder_instr_insert(&b
, &instr
->instr
);
1196 case nir_intrinsic_read_invocation
: {
1197 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1198 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1199 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1201 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1202 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1204 ir_rvalue
*invocation
= (ir_rvalue
*) ir
->actual_parameters
.get_head()->next
;
1205 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(invocation
));
1207 nir_builder_instr_insert(&b
, &instr
->instr
);
1210 case nir_intrinsic_read_first_invocation
: {
1211 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1212 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1213 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1215 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1216 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1218 nir_builder_instr_insert(&b
, &instr
->instr
);
1222 unreachable("not reached");
1225 if (ir
->return_deref
) {
1226 nir_intrinsic_instr
*store_instr
=
1227 nir_intrinsic_instr_create(shader
, nir_intrinsic_store_var
);
1228 store_instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1229 nir_intrinsic_set_write_mask(store_instr
,
1230 (1 << store_instr
->num_components
) - 1);
1232 store_instr
->variables
[0] =
1233 evaluate_deref(&store_instr
->instr
, ir
->return_deref
);
1234 store_instr
->src
[0] = nir_src_for_ssa(&dest
->ssa
);
1236 nir_builder_instr_insert(&b
, &store_instr
->instr
);
1242 unreachable("glsl_to_nir only handles function calls to intrinsics");
1246 nir_visitor::visit(ir_assignment
*ir
)
1248 unsigned num_components
= ir
->lhs
->type
->vector_elements
;
1250 b
.exact
= ir
->lhs
->variable_referenced()->data
.invariant
||
1251 ir
->lhs
->variable_referenced()->data
.precise
;
1253 if ((ir
->rhs
->as_dereference() || ir
->rhs
->as_constant()) &&
1254 (ir
->write_mask
== (1 << num_components
) - 1 || ir
->write_mask
== 0)) {
1255 /* We're doing a plain-as-can-be copy, so emit a copy_var */
1256 nir_intrinsic_instr
*copy
=
1257 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_copy_var
);
1259 copy
->variables
[0] = evaluate_deref(©
->instr
, ir
->lhs
);
1260 copy
->variables
[1] = evaluate_deref(©
->instr
, ir
->rhs
);
1262 if (ir
->condition
) {
1263 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1264 nir_builder_instr_insert(&b
, ©
->instr
);
1265 nir_pop_if(&b
, NULL
);
1267 nir_builder_instr_insert(&b
, ©
->instr
);
1272 assert(ir
->rhs
->type
->is_scalar() || ir
->rhs
->type
->is_vector());
1274 ir
->lhs
->accept(this);
1275 nir_deref_var
*lhs_deref
= this->deref_head
;
1276 nir_ssa_def
*src
= evaluate_rvalue(ir
->rhs
);
1278 if (ir
->write_mask
!= (1 << num_components
) - 1 && ir
->write_mask
!= 0) {
1279 /* GLSL IR will give us the input to the write-masked assignment in a
1280 * single packed vector. So, for example, if the writemask is xzw, then
1281 * we have to swizzle x -> x, y -> z, and z -> w and get the y component
1285 unsigned component
= 0;
1286 for (unsigned i
= 0; i
< 4; i
++) {
1287 swiz
[i
] = ir
->write_mask
& (1 << i
) ? component
++ : 0;
1289 src
= nir_swizzle(&b
, src
, swiz
, num_components
, !supports_ints
);
1292 nir_intrinsic_instr
*store
=
1293 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_store_var
);
1294 store
->num_components
= ir
->lhs
->type
->vector_elements
;
1295 nir_intrinsic_set_write_mask(store
, ir
->write_mask
);
1296 store
->variables
[0] = nir_deref_var_clone(lhs_deref
, store
);
1297 store
->src
[0] = nir_src_for_ssa(src
);
1299 if (ir
->condition
) {
1300 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1301 nir_builder_instr_insert(&b
, &store
->instr
);
1302 nir_pop_if(&b
, NULL
);
1304 nir_builder_instr_insert(&b
, &store
->instr
);
1309 * Given an instruction, returns a pointer to its destination or NULL if there
1310 * is no destination.
1312 * Note that this only handles instructions we generate at this level.
1315 get_instr_dest(nir_instr
*instr
)
1317 nir_alu_instr
*alu_instr
;
1318 nir_intrinsic_instr
*intrinsic_instr
;
1319 nir_tex_instr
*tex_instr
;
1321 switch (instr
->type
) {
1322 case nir_instr_type_alu
:
1323 alu_instr
= nir_instr_as_alu(instr
);
1324 return &alu_instr
->dest
.dest
;
1326 case nir_instr_type_intrinsic
:
1327 intrinsic_instr
= nir_instr_as_intrinsic(instr
);
1328 if (nir_intrinsic_infos
[intrinsic_instr
->intrinsic
].has_dest
)
1329 return &intrinsic_instr
->dest
;
1333 case nir_instr_type_tex
:
1334 tex_instr
= nir_instr_as_tex(instr
);
1335 return &tex_instr
->dest
;
1338 unreachable("not reached");
1345 nir_visitor::add_instr(nir_instr
*instr
, unsigned num_components
,
1348 nir_dest
*dest
= get_instr_dest(instr
);
1351 nir_ssa_dest_init(instr
, dest
, num_components
, bit_size
, NULL
);
1353 nir_builder_instr_insert(&b
, instr
);
1356 assert(dest
->is_ssa
);
1357 this->result
= &dest
->ssa
;
1362 nir_visitor::evaluate_rvalue(ir_rvalue
* ir
)
1365 if (ir
->as_dereference() || ir
->as_constant()) {
1367 * A dereference is being used on the right hand side, which means we
1368 * must emit a variable load.
1371 nir_intrinsic_instr
*load_instr
=
1372 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_var
);
1373 load_instr
->num_components
= ir
->type
->vector_elements
;
1374 load_instr
->variables
[0] = this->deref_head
;
1375 ralloc_steal(load_instr
, load_instr
->variables
[0]);
1376 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1377 add_instr(&load_instr
->instr
, ir
->type
->vector_elements
, bit_size
);
1380 return this->result
;
1384 type_is_float(glsl_base_type type
)
1386 return type
== GLSL_TYPE_FLOAT
|| type
== GLSL_TYPE_DOUBLE
||
1387 type
== GLSL_TYPE_FLOAT16
;
1391 type_is_signed(glsl_base_type type
)
1393 return type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_INT64
||
1394 type
== GLSL_TYPE_INT16
;
1398 nir_visitor::visit(ir_expression
*ir
)
1400 /* Some special cases */
1401 switch (ir
->operation
) {
1402 case ir_binop_ubo_load
: {
1403 nir_intrinsic_instr
*load
=
1404 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_ubo
);
1405 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1406 load
->num_components
= ir
->type
->vector_elements
;
1407 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1408 load
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1409 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1412 * In UBO's, a true boolean value is any non-zero value, but we consider
1413 * a true boolean to be ~0. Fix this up with a != 0 comparison.
1416 if (ir
->type
->is_boolean())
1417 this->result
= nir_ine(&b
, &load
->dest
.ssa
, nir_imm_int(&b
, 0));
1422 case ir_unop_interpolate_at_centroid
:
1423 case ir_binop_interpolate_at_offset
:
1424 case ir_binop_interpolate_at_sample
: {
1425 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1426 ir_swizzle
*swizzle
= NULL
;
1428 /* the api does not allow a swizzle here, but the varying packing code
1429 * may have pushed one into here.
1431 swizzle
= ir
->operands
[0]->as_swizzle();
1433 deref
= swizzle
->val
->as_dereference();
1437 deref
->accept(this);
1439 nir_intrinsic_op op
;
1440 if (this->deref_head
->var
->data
.mode
== nir_var_shader_in
) {
1441 switch (ir
->operation
) {
1442 case ir_unop_interpolate_at_centroid
:
1443 op
= nir_intrinsic_interp_var_at_centroid
;
1445 case ir_binop_interpolate_at_offset
:
1446 op
= nir_intrinsic_interp_var_at_offset
;
1448 case ir_binop_interpolate_at_sample
:
1449 op
= nir_intrinsic_interp_var_at_sample
;
1452 unreachable("Invalid interpolation intrinsic");
1455 /* This case can happen if the vertex shader does not write the
1456 * given varying. In this case, the linker will lower it to a
1457 * global variable. Since interpolating a variable makes no
1458 * sense, we'll just turn it into a load which will probably
1459 * eventually end up as an SSA definition.
1461 assert(this->deref_head
->var
->data
.mode
== nir_var_global
);
1462 op
= nir_intrinsic_load_var
;
1465 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(shader
, op
);
1466 intrin
->num_components
= deref
->type
->vector_elements
;
1467 intrin
->variables
[0] = this->deref_head
;
1468 ralloc_steal(intrin
, intrin
->variables
[0]);
1470 if (intrin
->intrinsic
== nir_intrinsic_interp_var_at_offset
||
1471 intrin
->intrinsic
== nir_intrinsic_interp_var_at_sample
)
1472 intrin
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1474 unsigned bit_size
= glsl_get_bit_size(deref
->type
);
1475 add_instr(&intrin
->instr
, deref
->type
->vector_elements
, bit_size
);
1478 unsigned swiz
[4] = {
1479 swizzle
->mask
.x
, swizzle
->mask
.y
, swizzle
->mask
.z
, swizzle
->mask
.w
1482 result
= nir_swizzle(&b
, result
, swiz
,
1483 swizzle
->type
->vector_elements
, false);
1493 nir_ssa_def
*srcs
[4];
1494 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1495 srcs
[i
] = evaluate_rvalue(ir
->operands
[i
]);
1497 glsl_base_type types
[4];
1498 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1500 types
[i
] = ir
->operands
[i
]->type
->base_type
;
1502 types
[i
] = GLSL_TYPE_FLOAT
;
1504 glsl_base_type out_type
;
1506 out_type
= ir
->type
->base_type
;
1508 out_type
= GLSL_TYPE_FLOAT
;
1510 switch (ir
->operation
) {
1511 case ir_unop_bit_not
: result
= nir_inot(&b
, srcs
[0]); break;
1512 case ir_unop_logic_not
:
1513 result
= supports_ints
? nir_inot(&b
, srcs
[0]) : nir_fnot(&b
, srcs
[0]);
1516 result
= type_is_float(types
[0]) ? nir_fneg(&b
, srcs
[0])
1517 : nir_ineg(&b
, srcs
[0]);
1520 result
= type_is_float(types
[0]) ? nir_fabs(&b
, srcs
[0])
1521 : nir_iabs(&b
, srcs
[0]);
1523 case ir_unop_saturate
:
1524 assert(type_is_float(types
[0]));
1525 result
= nir_fsat(&b
, srcs
[0]);
1528 result
= type_is_float(types
[0]) ? nir_fsign(&b
, srcs
[0])
1529 : nir_isign(&b
, srcs
[0]);
1531 case ir_unop_rcp
: result
= nir_frcp(&b
, srcs
[0]); break;
1532 case ir_unop_rsq
: result
= nir_frsq(&b
, srcs
[0]); break;
1533 case ir_unop_sqrt
: result
= nir_fsqrt(&b
, srcs
[0]); break;
1534 case ir_unop_exp
: unreachable("ir_unop_exp should have been lowered");
1535 case ir_unop_log
: unreachable("ir_unop_log should have been lowered");
1536 case ir_unop_exp2
: result
= nir_fexp2(&b
, srcs
[0]); break;
1537 case ir_unop_log2
: result
= nir_flog2(&b
, srcs
[0]); break;
1539 result
= supports_ints
? nir_i2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1542 result
= supports_ints
? nir_u2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1545 result
= supports_ints
? nir_b2f(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1579 case ir_unop_i642u64
:
1580 case ir_unop_u642i64
: {
1581 nir_alu_type src_type
= nir_get_nir_type_for_glsl_base_type(types
[0]);
1582 nir_alu_type dst_type
= nir_get_nir_type_for_glsl_base_type(out_type
);
1583 result
= nir_build_alu(&b
, nir_type_conversion_op(src_type
, dst_type
,
1584 nir_rounding_mode_undef
),
1585 srcs
[0], NULL
, NULL
, NULL
);
1586 /* b2i and b2f don't have fixed bit-size versions so the builder will
1587 * just assume 32 and we have to fix it up here.
1589 result
->bit_size
= nir_alu_type_get_type_size(dst_type
);
1593 case ir_unop_bitcast_i2f
:
1594 case ir_unop_bitcast_f2i
:
1595 case ir_unop_bitcast_u2f
:
1596 case ir_unop_bitcast_f2u
:
1597 case ir_unop_bitcast_i642d
:
1598 case ir_unop_bitcast_d2i64
:
1599 case ir_unop_bitcast_u642d
:
1600 case ir_unop_bitcast_d2u64
:
1601 case ir_unop_subroutine_to_int
:
1603 result
= nir_imov(&b
, srcs
[0]);
1605 case ir_unop_trunc
: result
= nir_ftrunc(&b
, srcs
[0]); break;
1606 case ir_unop_ceil
: result
= nir_fceil(&b
, srcs
[0]); break;
1607 case ir_unop_floor
: result
= nir_ffloor(&b
, srcs
[0]); break;
1608 case ir_unop_fract
: result
= nir_ffract(&b
, srcs
[0]); break;
1609 case ir_unop_frexp_exp
: result
= nir_frexp_exp(&b
, srcs
[0]); break;
1610 case ir_unop_frexp_sig
: result
= nir_frexp_sig(&b
, srcs
[0]); break;
1611 case ir_unop_round_even
: result
= nir_fround_even(&b
, srcs
[0]); break;
1612 case ir_unop_sin
: result
= nir_fsin(&b
, srcs
[0]); break;
1613 case ir_unop_cos
: result
= nir_fcos(&b
, srcs
[0]); break;
1614 case ir_unop_dFdx
: result
= nir_fddx(&b
, srcs
[0]); break;
1615 case ir_unop_dFdy
: result
= nir_fddy(&b
, srcs
[0]); break;
1616 case ir_unop_dFdx_fine
: result
= nir_fddx_fine(&b
, srcs
[0]); break;
1617 case ir_unop_dFdy_fine
: result
= nir_fddy_fine(&b
, srcs
[0]); break;
1618 case ir_unop_dFdx_coarse
: result
= nir_fddx_coarse(&b
, srcs
[0]); break;
1619 case ir_unop_dFdy_coarse
: result
= nir_fddy_coarse(&b
, srcs
[0]); break;
1620 case ir_unop_pack_snorm_2x16
:
1621 result
= nir_pack_snorm_2x16(&b
, srcs
[0]);
1623 case ir_unop_pack_snorm_4x8
:
1624 result
= nir_pack_snorm_4x8(&b
, srcs
[0]);
1626 case ir_unop_pack_unorm_2x16
:
1627 result
= nir_pack_unorm_2x16(&b
, srcs
[0]);
1629 case ir_unop_pack_unorm_4x8
:
1630 result
= nir_pack_unorm_4x8(&b
, srcs
[0]);
1632 case ir_unop_pack_half_2x16
:
1633 result
= nir_pack_half_2x16(&b
, srcs
[0]);
1635 case ir_unop_unpack_snorm_2x16
:
1636 result
= nir_unpack_snorm_2x16(&b
, srcs
[0]);
1638 case ir_unop_unpack_snorm_4x8
:
1639 result
= nir_unpack_snorm_4x8(&b
, srcs
[0]);
1641 case ir_unop_unpack_unorm_2x16
:
1642 result
= nir_unpack_unorm_2x16(&b
, srcs
[0]);
1644 case ir_unop_unpack_unorm_4x8
:
1645 result
= nir_unpack_unorm_4x8(&b
, srcs
[0]);
1647 case ir_unop_unpack_half_2x16
:
1648 result
= nir_unpack_half_2x16(&b
, srcs
[0]);
1650 case ir_unop_pack_sampler_2x32
:
1651 case ir_unop_pack_image_2x32
:
1652 case ir_unop_pack_double_2x32
:
1653 case ir_unop_pack_int_2x32
:
1654 case ir_unop_pack_uint_2x32
:
1655 result
= nir_pack_64_2x32(&b
, srcs
[0]);
1657 case ir_unop_unpack_sampler_2x32
:
1658 case ir_unop_unpack_image_2x32
:
1659 case ir_unop_unpack_double_2x32
:
1660 case ir_unop_unpack_int_2x32
:
1661 case ir_unop_unpack_uint_2x32
:
1662 result
= nir_unpack_64_2x32(&b
, srcs
[0]);
1664 case ir_unop_bitfield_reverse
:
1665 result
= nir_bitfield_reverse(&b
, srcs
[0]);
1667 case ir_unop_bit_count
:
1668 result
= nir_bit_count(&b
, srcs
[0]);
1670 case ir_unop_find_msb
:
1672 case GLSL_TYPE_UINT
:
1673 result
= nir_ufind_msb(&b
, srcs
[0]);
1676 result
= nir_ifind_msb(&b
, srcs
[0]);
1679 unreachable("Invalid type for findMSB()");
1682 case ir_unop_find_lsb
:
1683 result
= nir_find_lsb(&b
, srcs
[0]);
1687 switch (ir
->type
->vector_elements
) {
1689 switch (ir
->operands
[0]->type
->vector_elements
) {
1690 case 1: result
= nir_fnoise1_1(&b
, srcs
[0]); break;
1691 case 2: result
= nir_fnoise1_2(&b
, srcs
[0]); break;
1692 case 3: result
= nir_fnoise1_3(&b
, srcs
[0]); break;
1693 case 4: result
= nir_fnoise1_4(&b
, srcs
[0]); break;
1694 default: unreachable("not reached");
1698 switch (ir
->operands
[0]->type
->vector_elements
) {
1699 case 1: result
= nir_fnoise2_1(&b
, srcs
[0]); break;
1700 case 2: result
= nir_fnoise2_2(&b
, srcs
[0]); break;
1701 case 3: result
= nir_fnoise2_3(&b
, srcs
[0]); break;
1702 case 4: result
= nir_fnoise2_4(&b
, srcs
[0]); break;
1703 default: unreachable("not reached");
1707 switch (ir
->operands
[0]->type
->vector_elements
) {
1708 case 1: result
= nir_fnoise3_1(&b
, srcs
[0]); break;
1709 case 2: result
= nir_fnoise3_2(&b
, srcs
[0]); break;
1710 case 3: result
= nir_fnoise3_3(&b
, srcs
[0]); break;
1711 case 4: result
= nir_fnoise3_4(&b
, srcs
[0]); break;
1712 default: unreachable("not reached");
1716 switch (ir
->operands
[0]->type
->vector_elements
) {
1717 case 1: result
= nir_fnoise4_1(&b
, srcs
[0]); break;
1718 case 2: result
= nir_fnoise4_2(&b
, srcs
[0]); break;
1719 case 3: result
= nir_fnoise4_3(&b
, srcs
[0]); break;
1720 case 4: result
= nir_fnoise4_4(&b
, srcs
[0]); break;
1721 default: unreachable("not reached");
1725 unreachable("not reached");
1728 case ir_unop_get_buffer_size
: {
1729 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(
1731 nir_intrinsic_get_buffer_size
);
1732 load
->num_components
= ir
->type
->vector_elements
;
1733 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1734 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1735 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1740 result
= type_is_float(out_type
) ? nir_fadd(&b
, srcs
[0], srcs
[1])
1741 : nir_iadd(&b
, srcs
[0], srcs
[1]);
1744 result
= type_is_float(out_type
) ? nir_fsub(&b
, srcs
[0], srcs
[1])
1745 : nir_isub(&b
, srcs
[0], srcs
[1]);
1748 result
= type_is_float(out_type
) ? nir_fmul(&b
, srcs
[0], srcs
[1])
1749 : nir_imul(&b
, srcs
[0], srcs
[1]);
1752 if (type_is_float(out_type
))
1753 result
= nir_fdiv(&b
, srcs
[0], srcs
[1]);
1754 else if (type_is_signed(out_type
))
1755 result
= nir_idiv(&b
, srcs
[0], srcs
[1]);
1757 result
= nir_udiv(&b
, srcs
[0], srcs
[1]);
1760 result
= type_is_float(out_type
) ? nir_fmod(&b
, srcs
[0], srcs
[1])
1761 : nir_umod(&b
, srcs
[0], srcs
[1]);
1764 if (type_is_float(out_type
))
1765 result
= nir_fmin(&b
, srcs
[0], srcs
[1]);
1766 else if (type_is_signed(out_type
))
1767 result
= nir_imin(&b
, srcs
[0], srcs
[1]);
1769 result
= nir_umin(&b
, srcs
[0], srcs
[1]);
1772 if (type_is_float(out_type
))
1773 result
= nir_fmax(&b
, srcs
[0], srcs
[1]);
1774 else if (type_is_signed(out_type
))
1775 result
= nir_imax(&b
, srcs
[0], srcs
[1]);
1777 result
= nir_umax(&b
, srcs
[0], srcs
[1]);
1779 case ir_binop_pow
: result
= nir_fpow(&b
, srcs
[0], srcs
[1]); break;
1780 case ir_binop_bit_and
: result
= nir_iand(&b
, srcs
[0], srcs
[1]); break;
1781 case ir_binop_bit_or
: result
= nir_ior(&b
, srcs
[0], srcs
[1]); break;
1782 case ir_binop_bit_xor
: result
= nir_ixor(&b
, srcs
[0], srcs
[1]); break;
1783 case ir_binop_logic_and
:
1784 result
= supports_ints
? nir_iand(&b
, srcs
[0], srcs
[1])
1785 : nir_fand(&b
, srcs
[0], srcs
[1]);
1787 case ir_binop_logic_or
:
1788 result
= supports_ints
? nir_ior(&b
, srcs
[0], srcs
[1])
1789 : nir_for(&b
, srcs
[0], srcs
[1]);
1791 case ir_binop_logic_xor
:
1792 result
= supports_ints
? nir_ixor(&b
, srcs
[0], srcs
[1])
1793 : nir_fxor(&b
, srcs
[0], srcs
[1]);
1795 case ir_binop_lshift
: result
= nir_ishl(&b
, srcs
[0], srcs
[1]); break;
1796 case ir_binop_rshift
:
1797 result
= (type_is_signed(out_type
)) ? nir_ishr(&b
, srcs
[0], srcs
[1])
1798 : nir_ushr(&b
, srcs
[0], srcs
[1]);
1800 case ir_binop_imul_high
:
1801 result
= (out_type
== GLSL_TYPE_INT
) ? nir_imul_high(&b
, srcs
[0], srcs
[1])
1802 : nir_umul_high(&b
, srcs
[0], srcs
[1]);
1804 case ir_binop_carry
: result
= nir_uadd_carry(&b
, srcs
[0], srcs
[1]); break;
1805 case ir_binop_borrow
: result
= nir_usub_borrow(&b
, srcs
[0], srcs
[1]); break;
1807 if (supports_ints
) {
1808 if (type_is_float(types
[0]))
1809 result
= nir_flt(&b
, srcs
[0], srcs
[1]);
1810 else if (type_is_signed(types
[0]))
1811 result
= nir_ilt(&b
, srcs
[0], srcs
[1]);
1813 result
= nir_ult(&b
, srcs
[0], srcs
[1]);
1815 result
= nir_slt(&b
, srcs
[0], srcs
[1]);
1818 case ir_binop_gequal
:
1819 if (supports_ints
) {
1820 if (type_is_float(types
[0]))
1821 result
= nir_fge(&b
, srcs
[0], srcs
[1]);
1822 else if (type_is_signed(types
[0]))
1823 result
= nir_ige(&b
, srcs
[0], srcs
[1]);
1825 result
= nir_uge(&b
, srcs
[0], srcs
[1]);
1827 result
= nir_sge(&b
, srcs
[0], srcs
[1]);
1830 case ir_binop_equal
:
1831 if (supports_ints
) {
1832 if (type_is_float(types
[0]))
1833 result
= nir_feq(&b
, srcs
[0], srcs
[1]);
1835 result
= nir_ieq(&b
, srcs
[0], srcs
[1]);
1837 result
= nir_seq(&b
, srcs
[0], srcs
[1]);
1840 case ir_binop_nequal
:
1841 if (supports_ints
) {
1842 if (type_is_float(types
[0]))
1843 result
= nir_fne(&b
, srcs
[0], srcs
[1]);
1845 result
= nir_ine(&b
, srcs
[0], srcs
[1]);
1847 result
= nir_sne(&b
, srcs
[0], srcs
[1]);
1850 case ir_binop_all_equal
:
1851 if (supports_ints
) {
1852 if (type_is_float(types
[0])) {
1853 switch (ir
->operands
[0]->type
->vector_elements
) {
1854 case 1: result
= nir_feq(&b
, srcs
[0], srcs
[1]); break;
1855 case 2: result
= nir_ball_fequal2(&b
, srcs
[0], srcs
[1]); break;
1856 case 3: result
= nir_ball_fequal3(&b
, srcs
[0], srcs
[1]); break;
1857 case 4: result
= nir_ball_fequal4(&b
, srcs
[0], srcs
[1]); break;
1859 unreachable("not reached");
1862 switch (ir
->operands
[0]->type
->vector_elements
) {
1863 case 1: result
= nir_ieq(&b
, srcs
[0], srcs
[1]); break;
1864 case 2: result
= nir_ball_iequal2(&b
, srcs
[0], srcs
[1]); break;
1865 case 3: result
= nir_ball_iequal3(&b
, srcs
[0], srcs
[1]); break;
1866 case 4: result
= nir_ball_iequal4(&b
, srcs
[0], srcs
[1]); break;
1868 unreachable("not reached");
1872 switch (ir
->operands
[0]->type
->vector_elements
) {
1873 case 1: result
= nir_seq(&b
, srcs
[0], srcs
[1]); break;
1874 case 2: result
= nir_fall_equal2(&b
, srcs
[0], srcs
[1]); break;
1875 case 3: result
= nir_fall_equal3(&b
, srcs
[0], srcs
[1]); break;
1876 case 4: result
= nir_fall_equal4(&b
, srcs
[0], srcs
[1]); break;
1878 unreachable("not reached");
1882 case ir_binop_any_nequal
:
1883 if (supports_ints
) {
1884 if (type_is_float(types
[0])) {
1885 switch (ir
->operands
[0]->type
->vector_elements
) {
1886 case 1: result
= nir_fne(&b
, srcs
[0], srcs
[1]); break;
1887 case 2: result
= nir_bany_fnequal2(&b
, srcs
[0], srcs
[1]); break;
1888 case 3: result
= nir_bany_fnequal3(&b
, srcs
[0], srcs
[1]); break;
1889 case 4: result
= nir_bany_fnequal4(&b
, srcs
[0], srcs
[1]); break;
1891 unreachable("not reached");
1894 switch (ir
->operands
[0]->type
->vector_elements
) {
1895 case 1: result
= nir_ine(&b
, srcs
[0], srcs
[1]); break;
1896 case 2: result
= nir_bany_inequal2(&b
, srcs
[0], srcs
[1]); break;
1897 case 3: result
= nir_bany_inequal3(&b
, srcs
[0], srcs
[1]); break;
1898 case 4: result
= nir_bany_inequal4(&b
, srcs
[0], srcs
[1]); break;
1900 unreachable("not reached");
1904 switch (ir
->operands
[0]->type
->vector_elements
) {
1905 case 1: result
= nir_sne(&b
, srcs
[0], srcs
[1]); break;
1906 case 2: result
= nir_fany_nequal2(&b
, srcs
[0], srcs
[1]); break;
1907 case 3: result
= nir_fany_nequal3(&b
, srcs
[0], srcs
[1]); break;
1908 case 4: result
= nir_fany_nequal4(&b
, srcs
[0], srcs
[1]); break;
1910 unreachable("not reached");
1915 switch (ir
->operands
[0]->type
->vector_elements
) {
1916 case 2: result
= nir_fdot2(&b
, srcs
[0], srcs
[1]); break;
1917 case 3: result
= nir_fdot3(&b
, srcs
[0], srcs
[1]); break;
1918 case 4: result
= nir_fdot4(&b
, srcs
[0], srcs
[1]); break;
1920 unreachable("not reached");
1923 case ir_binop_vector_extract
: {
1924 result
= nir_channel(&b
, srcs
[0], 0);
1925 for (unsigned i
= 1; i
< ir
->operands
[0]->type
->vector_elements
; i
++) {
1926 nir_ssa_def
*swizzled
= nir_channel(&b
, srcs
[0], i
);
1927 result
= nir_bcsel(&b
, nir_ieq(&b
, srcs
[1], nir_imm_int(&b
, i
)),
1933 case ir_binop_ldexp
: result
= nir_ldexp(&b
, srcs
[0], srcs
[1]); break;
1935 result
= nir_ffma(&b
, srcs
[0], srcs
[1], srcs
[2]);
1938 result
= nir_flrp(&b
, srcs
[0], srcs
[1], srcs
[2]);
1942 result
= nir_bcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
1944 result
= nir_fcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
1946 case ir_triop_bitfield_extract
:
1947 result
= (out_type
== GLSL_TYPE_INT
) ?
1948 nir_ibitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]) :
1949 nir_ubitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]);
1951 case ir_quadop_bitfield_insert
:
1952 result
= nir_bitfield_insert(&b
, srcs
[0], srcs
[1], srcs
[2], srcs
[3]);
1954 case ir_quadop_vector
:
1955 result
= nir_vec(&b
, srcs
, ir
->type
->vector_elements
);
1959 unreachable("not reached");
1964 nir_visitor::visit(ir_swizzle
*ir
)
1966 unsigned swizzle
[4] = { ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
};
1967 result
= nir_swizzle(&b
, evaluate_rvalue(ir
->val
), swizzle
,
1968 ir
->type
->vector_elements
, !supports_ints
);
1972 nir_visitor::visit(ir_texture
*ir
)
1979 num_srcs
= 1; /* coordinate */
1984 op
= (ir
->op
== ir_txb
) ? nir_texop_txb
: nir_texop_txl
;
1985 num_srcs
= 2; /* coordinate, bias/lod */
1989 op
= nir_texop_txd
; /* coordinate, dPdx, dPdy */
1995 if (ir
->lod_info
.lod
!= NULL
)
1996 num_srcs
= 2; /* coordinate, lod */
1998 num_srcs
= 1; /* coordinate */
2002 op
= nir_texop_txf_ms
;
2003 num_srcs
= 2; /* coordinate, sample_index */
2008 if (ir
->lod_info
.lod
!= NULL
)
2009 num_srcs
= 1; /* lod */
2016 num_srcs
= 1; /* coordinate */
2021 num_srcs
= 1; /* coordinate */
2024 case ir_query_levels
:
2025 op
= nir_texop_query_levels
;
2029 case ir_texture_samples
:
2030 op
= nir_texop_texture_samples
;
2034 case ir_samples_identical
:
2035 op
= nir_texop_samples_identical
;
2036 num_srcs
= 1; /* coordinate */
2040 unreachable("not reached");
2043 if (ir
->projector
!= NULL
)
2045 if (ir
->shadow_comparator
!= NULL
)
2047 if (ir
->offset
!= NULL
)
2050 nir_tex_instr
*instr
= nir_tex_instr_create(this->shader
, num_srcs
);
2053 instr
->sampler_dim
=
2054 (glsl_sampler_dim
) ir
->sampler
->type
->sampler_dimensionality
;
2055 instr
->is_array
= ir
->sampler
->type
->sampler_array
;
2056 instr
->is_shadow
= ir
->sampler
->type
->sampler_shadow
;
2057 if (instr
->is_shadow
)
2058 instr
->is_new_style_shadow
= (ir
->type
->vector_elements
== 1);
2059 switch (ir
->type
->base_type
) {
2060 case GLSL_TYPE_FLOAT
:
2061 instr
->dest_type
= nir_type_float
;
2064 instr
->dest_type
= nir_type_int
;
2066 case GLSL_TYPE_BOOL
:
2067 case GLSL_TYPE_UINT
:
2068 instr
->dest_type
= nir_type_uint
;
2071 unreachable("not reached");
2074 instr
->texture
= evaluate_deref(&instr
->instr
, ir
->sampler
);
2076 unsigned src_number
= 0;
2078 if (ir
->coordinate
!= NULL
) {
2079 instr
->coord_components
= ir
->coordinate
->type
->vector_elements
;
2080 instr
->src
[src_number
].src
=
2081 nir_src_for_ssa(evaluate_rvalue(ir
->coordinate
));
2082 instr
->src
[src_number
].src_type
= nir_tex_src_coord
;
2086 if (ir
->projector
!= NULL
) {
2087 instr
->src
[src_number
].src
=
2088 nir_src_for_ssa(evaluate_rvalue(ir
->projector
));
2089 instr
->src
[src_number
].src_type
= nir_tex_src_projector
;
2093 if (ir
->shadow_comparator
!= NULL
) {
2094 instr
->src
[src_number
].src
=
2095 nir_src_for_ssa(evaluate_rvalue(ir
->shadow_comparator
));
2096 instr
->src
[src_number
].src_type
= nir_tex_src_comparator
;
2100 if (ir
->offset
!= NULL
) {
2101 /* we don't support multiple offsets yet */
2102 assert(ir
->offset
->type
->is_vector() || ir
->offset
->type
->is_scalar());
2104 instr
->src
[src_number
].src
=
2105 nir_src_for_ssa(evaluate_rvalue(ir
->offset
));
2106 instr
->src
[src_number
].src_type
= nir_tex_src_offset
;
2112 instr
->src
[src_number
].src
=
2113 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.bias
));
2114 instr
->src
[src_number
].src_type
= nir_tex_src_bias
;
2121 if (ir
->lod_info
.lod
!= NULL
) {
2122 instr
->src
[src_number
].src
=
2123 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.lod
));
2124 instr
->src
[src_number
].src_type
= nir_tex_src_lod
;
2130 instr
->src
[src_number
].src
=
2131 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdx
));
2132 instr
->src
[src_number
].src_type
= nir_tex_src_ddx
;
2134 instr
->src
[src_number
].src
=
2135 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdy
));
2136 instr
->src
[src_number
].src_type
= nir_tex_src_ddy
;
2141 instr
->src
[src_number
].src
=
2142 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.sample_index
));
2143 instr
->src
[src_number
].src_type
= nir_tex_src_ms_index
;
2148 instr
->component
= ir
->lod_info
.component
->as_constant()->value
.u
[0];
2155 assert(src_number
== num_srcs
);
2157 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2158 add_instr(&instr
->instr
, nir_tex_instr_dest_size(instr
), bit_size
);
2162 nir_visitor::visit(ir_constant
*ir
)
2165 * We don't know if this variable is an array or struct that gets
2166 * dereferenced, so do the safe thing an make it a variable with a
2167 * constant initializer and return a dereference.
2171 nir_local_variable_create(this->impl
, ir
->type
, "const_temp");
2172 var
->data
.read_only
= true;
2173 var
->constant_initializer
= constant_copy(ir
, var
);
2175 this->deref_head
= nir_deref_var_create(this->shader
, var
);
2176 this->deref_tail
= &this->deref_head
->deref
;
2180 nir_visitor::visit(ir_dereference_variable
*ir
)
2182 struct hash_entry
*entry
=
2183 _mesa_hash_table_search(this->var_table
, ir
->var
);
2185 nir_variable
*var
= (nir_variable
*) entry
->data
;
2187 nir_deref_var
*deref
= nir_deref_var_create(this->shader
, var
);
2188 this->deref_head
= deref
;
2189 this->deref_tail
= &deref
->deref
;
2193 nir_visitor::visit(ir_dereference_record
*ir
)
2195 ir
->record
->accept(this);
2197 int field_index
= ir
->field_idx
;
2198 assert(field_index
>= 0);
2200 nir_deref_struct
*deref
= nir_deref_struct_create(this->deref_tail
, field_index
);
2201 deref
->deref
.type
= ir
->type
;
2202 this->deref_tail
->child
= &deref
->deref
;
2203 this->deref_tail
= &deref
->deref
;
2207 nir_visitor::visit(ir_dereference_array
*ir
)
2209 nir_deref_array
*deref
= nir_deref_array_create(this->shader
);
2210 deref
->deref
.type
= ir
->type
;
2212 ir_constant
*const_index
= ir
->array_index
->as_constant();
2213 if (const_index
!= NULL
) {
2214 deref
->deref_array_type
= nir_deref_array_type_direct
;
2215 deref
->base_offset
= const_index
->value
.u
[0];
2217 deref
->deref_array_type
= nir_deref_array_type_indirect
;
2219 nir_src_for_ssa(evaluate_rvalue(ir
->array_index
));
2222 ir
->array
->accept(this);
2224 this->deref_tail
->child
= &deref
->deref
;
2225 ralloc_steal(this->deref_tail
, deref
);
2226 this->deref_tail
= &deref
->deref
;
2230 nir_visitor::visit(ir_barrier
*)
2232 nir_intrinsic_instr
*instr
=
2233 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_barrier
);
2234 nir_builder_instr_insert(&b
, &instr
->instr
);