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_instr
*evaluate_deref(ir_instruction
*ir
);
96 /* most recent deref instruction created */
97 nir_deref_instr
*deref
;
99 nir_variable
*var
; /* variable created by ir_variable visitor */
101 /* whether the IR we're operating on is per-function or global */
104 /* map of ir_variable -> nir_variable */
105 struct hash_table
*var_table
;
107 /* map of ir_function_signature -> nir_function_overload */
108 struct hash_table
*overload_table
;
112 * This visitor runs before the main visitor, calling create_function() for
113 * each function so that the main visitor can resolve forward references in
117 class nir_function_visitor
: public ir_hierarchical_visitor
120 nir_function_visitor(nir_visitor
*v
) : visitor(v
)
123 virtual ir_visitor_status
visit_enter(ir_function
*);
126 nir_visitor
*visitor
;
129 } /* end of anonymous namespace */
132 nir_remap_attributes(nir_shader
*shader
,
133 const nir_shader_compiler_options
*options
)
135 if (options
->vs_inputs_dual_locations
) {
136 nir_foreach_variable(var
, &shader
->inputs
) {
137 var
->data
.location
+=
138 _mesa_bitcount_64(shader
->info
.vs
.double_inputs
&
139 BITFIELD64_MASK(var
->data
.location
));
143 /* Once the remap is done, reset double_inputs_read, so later it will have
144 * which location/slots are doubles */
145 shader
->info
.vs
.double_inputs
= 0;
149 glsl_to_nir(const struct gl_shader_program
*shader_prog
,
150 gl_shader_stage stage
,
151 const nir_shader_compiler_options
*options
)
153 struct gl_linked_shader
*sh
= shader_prog
->_LinkedShaders
[stage
];
155 nir_shader
*shader
= nir_shader_create(NULL
, stage
, options
,
158 nir_visitor
v1(shader
);
159 nir_function_visitor
v2(&v1
);
161 visit_exec_list(sh
->ir
, &v1
);
163 nir_lower_constant_initializers(shader
, (nir_variable_mode
)~0);
165 /* Remap the locations to slots so those requiring two slots will occupy
166 * two locations. For instance, if we have in the IR code a dvec3 attr0 in
167 * location 0 and vec4 attr1 in location 1, in NIR attr0 will use
168 * locations/slots 0 and 1, and attr1 will use location/slot 2 */
169 if (shader
->info
.stage
== MESA_SHADER_VERTEX
)
170 nir_remap_attributes(shader
, options
);
172 shader
->info
.name
= ralloc_asprintf(shader
, "GLSL%d", shader_prog
->Name
);
173 if (shader_prog
->Label
)
174 shader
->info
.label
= ralloc_strdup(shader
, shader_prog
->Label
);
176 /* Check for transform feedback varyings specified via the API */
177 shader
->info
.has_transform_feedback_varyings
=
178 shader_prog
->TransformFeedback
.NumVarying
> 0;
180 /* Check for transform feedback varyings specified in the Shader */
181 if (shader_prog
->last_vert_prog
)
182 shader
->info
.has_transform_feedback_varyings
|=
183 shader_prog
->last_vert_prog
->sh
.LinkedTransformFeedback
->NumVarying
> 0;
188 nir_visitor::nir_visitor(nir_shader
*shader
)
190 this->supports_ints
= shader
->options
->native_integers
;
191 this->shader
= shader
;
192 this->is_global
= true;
193 this->var_table
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
194 _mesa_key_pointer_equal
);
195 this->overload_table
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
196 _mesa_key_pointer_equal
);
200 memset(&this->b
, 0, sizeof(this->b
));
203 nir_visitor::~nir_visitor()
205 _mesa_hash_table_destroy(this->var_table
, NULL
);
206 _mesa_hash_table_destroy(this->overload_table
, NULL
);
210 nir_visitor::evaluate_deref(ir_instruction
*ir
)
216 static nir_constant
*
217 constant_copy(ir_constant
*ir
, void *mem_ctx
)
222 nir_constant
*ret
= rzalloc(mem_ctx
, nir_constant
);
224 const unsigned rows
= ir
->type
->vector_elements
;
225 const unsigned cols
= ir
->type
->matrix_columns
;
228 ret
->num_elements
= 0;
229 switch (ir
->type
->base_type
) {
231 /* Only float base types can be matrices. */
234 for (unsigned r
= 0; r
< rows
; r
++)
235 ret
->values
[0].u32
[r
] = ir
->value
.u
[r
];
240 /* Only float base types can be matrices. */
243 for (unsigned r
= 0; r
< rows
; r
++)
244 ret
->values
[0].i32
[r
] = ir
->value
.i
[r
];
248 case GLSL_TYPE_FLOAT
:
249 for (unsigned c
= 0; c
< cols
; c
++) {
250 for (unsigned r
= 0; r
< rows
; r
++)
251 ret
->values
[c
].f32
[r
] = ir
->value
.f
[c
* rows
+ r
];
255 case GLSL_TYPE_DOUBLE
:
256 for (unsigned c
= 0; c
< cols
; c
++) {
257 for (unsigned r
= 0; r
< rows
; r
++)
258 ret
->values
[c
].f64
[r
] = ir
->value
.d
[c
* rows
+ r
];
262 case GLSL_TYPE_UINT64
:
263 /* Only float base types can be matrices. */
266 for (unsigned r
= 0; r
< rows
; r
++)
267 ret
->values
[0].u64
[r
] = ir
->value
.u64
[r
];
270 case GLSL_TYPE_INT64
:
271 /* Only float base types can be matrices. */
274 for (unsigned r
= 0; r
< rows
; r
++)
275 ret
->values
[0].i64
[r
] = ir
->value
.i64
[r
];
279 /* Only float base types can be matrices. */
282 for (unsigned r
= 0; r
< rows
; r
++)
283 ret
->values
[0].u32
[r
] = ir
->value
.b
[r
] ? NIR_TRUE
: NIR_FALSE
;
287 case GLSL_TYPE_STRUCT
:
288 case GLSL_TYPE_ARRAY
:
289 ret
->elements
= ralloc_array(mem_ctx
, nir_constant
*,
291 ret
->num_elements
= ir
->type
->length
;
293 for (i
= 0; i
< ir
->type
->length
; i
++)
294 ret
->elements
[i
] = constant_copy(ir
->const_elements
[i
], mem_ctx
);
298 unreachable("not reached");
305 nir_visitor::visit(ir_variable
*ir
)
307 /* TODO: In future we should switch to using the NIR lowering pass but for
308 * now just ignore these variables as GLSL IR should have lowered them.
309 * Anything remaining are just dead vars that weren't cleaned up.
311 if (ir
->data
.mode
== ir_var_shader_shared
)
314 nir_variable
*var
= rzalloc(shader
, nir_variable
);
315 var
->type
= ir
->type
;
316 var
->name
= ralloc_strdup(var
, ir
->name
);
318 var
->data
.always_active_io
= ir
->data
.always_active_io
;
319 var
->data
.read_only
= ir
->data
.read_only
;
320 var
->data
.centroid
= ir
->data
.centroid
;
321 var
->data
.sample
= ir
->data
.sample
;
322 var
->data
.patch
= ir
->data
.patch
;
323 var
->data
.invariant
= ir
->data
.invariant
;
324 var
->data
.location
= ir
->data
.location
;
325 var
->data
.stream
= ir
->data
.stream
;
326 var
->data
.compact
= false;
328 switch(ir
->data
.mode
) {
330 case ir_var_temporary
:
332 var
->data
.mode
= nir_var_global
;
334 var
->data
.mode
= nir_var_local
;
337 case ir_var_function_in
:
338 case ir_var_function_out
:
339 case ir_var_function_inout
:
340 case ir_var_const_in
:
341 var
->data
.mode
= nir_var_local
;
344 case ir_var_shader_in
:
345 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
346 ir
->data
.location
== VARYING_SLOT_FACE
) {
347 /* For whatever reason, GLSL IR makes gl_FrontFacing an input */
348 var
->data
.location
= SYSTEM_VALUE_FRONT_FACE
;
349 var
->data
.mode
= nir_var_system_value
;
350 } else if (shader
->info
.stage
== MESA_SHADER_GEOMETRY
&&
351 ir
->data
.location
== VARYING_SLOT_PRIMITIVE_ID
) {
352 /* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */
353 var
->data
.location
= SYSTEM_VALUE_PRIMITIVE_ID
;
354 var
->data
.mode
= nir_var_system_value
;
356 var
->data
.mode
= nir_var_shader_in
;
358 if (shader
->info
.stage
== MESA_SHADER_TESS_EVAL
&&
359 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
360 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
361 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
365 /* Mark all the locations that require two slots */
366 if (shader
->info
.stage
== MESA_SHADER_VERTEX
&&
367 glsl_type_is_dual_slot(glsl_without_array(var
->type
))) {
368 for (unsigned i
= 0; i
< glsl_count_attribute_slots(var
->type
, true); i
++) {
369 uint64_t bitfield
= BITFIELD64_BIT(var
->data
.location
+ i
);
370 shader
->info
.vs
.double_inputs
|= bitfield
;
375 case ir_var_shader_out
:
376 var
->data
.mode
= nir_var_shader_out
;
377 if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
&&
378 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
379 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
380 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
385 var
->data
.mode
= nir_var_uniform
;
388 case ir_var_shader_storage
:
389 var
->data
.mode
= nir_var_shader_storage
;
392 case ir_var_system_value
:
393 var
->data
.mode
= nir_var_system_value
;
397 unreachable("not reached");
400 var
->data
.interpolation
= ir
->data
.interpolation
;
401 var
->data
.origin_upper_left
= ir
->data
.origin_upper_left
;
402 var
->data
.pixel_center_integer
= ir
->data
.pixel_center_integer
;
403 var
->data
.location_frac
= ir
->data
.location_frac
;
405 if (var
->data
.pixel_center_integer
) {
406 assert(shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
407 shader
->info
.fs
.pixel_center_integer
= true;
410 switch (ir
->data
.depth_layout
) {
411 case ir_depth_layout_none
:
412 var
->data
.depth_layout
= nir_depth_layout_none
;
414 case ir_depth_layout_any
:
415 var
->data
.depth_layout
= nir_depth_layout_any
;
417 case ir_depth_layout_greater
:
418 var
->data
.depth_layout
= nir_depth_layout_greater
;
420 case ir_depth_layout_less
:
421 var
->data
.depth_layout
= nir_depth_layout_less
;
423 case ir_depth_layout_unchanged
:
424 var
->data
.depth_layout
= nir_depth_layout_unchanged
;
427 unreachable("not reached");
430 var
->data
.index
= ir
->data
.index
;
431 var
->data
.descriptor_set
= 0;
432 var
->data
.binding
= ir
->data
.binding
;
433 var
->data
.explicit_binding
= ir
->data
.explicit_binding
;
434 var
->data
.bindless
= ir
->data
.bindless
;
435 var
->data
.offset
= ir
->data
.offset
;
436 var
->data
.image
.read_only
= ir
->data
.memory_read_only
;
437 var
->data
.image
.write_only
= ir
->data
.memory_write_only
;
438 var
->data
.image
.coherent
= ir
->data
.memory_coherent
;
439 var
->data
.image
._volatile
= ir
->data
.memory_volatile
;
440 var
->data
.image
.restrict_flag
= ir
->data
.memory_restrict
;
441 var
->data
.image
.format
= ir
->data
.image_format
;
442 var
->data
.fb_fetch_output
= ir
->data
.fb_fetch_output
;
444 var
->num_state_slots
= ir
->get_num_state_slots();
445 if (var
->num_state_slots
> 0) {
446 var
->state_slots
= rzalloc_array(var
, nir_state_slot
,
447 var
->num_state_slots
);
449 ir_state_slot
*state_slots
= ir
->get_state_slots();
450 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
451 for (unsigned j
= 0; j
< 5; j
++)
452 var
->state_slots
[i
].tokens
[j
] = state_slots
[i
].tokens
[j
];
453 var
->state_slots
[i
].swizzle
= state_slots
[i
].swizzle
;
456 var
->state_slots
= NULL
;
459 var
->constant_initializer
= constant_copy(ir
->constant_initializer
, var
);
461 var
->interface_type
= ir
->get_interface_type();
463 if (var
->data
.mode
== nir_var_local
)
464 nir_function_impl_add_variable(impl
, var
);
466 nir_shader_add_variable(shader
, var
);
468 _mesa_hash_table_insert(var_table
, ir
, var
);
473 nir_function_visitor::visit_enter(ir_function
*ir
)
475 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
) {
476 visitor
->create_function(sig
);
478 return visit_continue_with_parent
;
482 nir_visitor::create_function(ir_function_signature
*ir
)
484 if (ir
->is_intrinsic())
487 nir_function
*func
= nir_function_create(shader
, ir
->function_name());
489 assert(ir
->parameters
.is_empty());
490 assert(ir
->return_type
== glsl_type::void_type
);
492 _mesa_hash_table_insert(this->overload_table
, ir
, func
);
496 nir_visitor::visit(ir_function
*ir
)
498 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
)
503 nir_visitor::visit(ir_function_signature
*ir
)
505 if (ir
->is_intrinsic())
508 struct hash_entry
*entry
=
509 _mesa_hash_table_search(this->overload_table
, ir
);
512 nir_function
*func
= (nir_function
*) entry
->data
;
514 if (ir
->is_defined
) {
515 nir_function_impl
*impl
= nir_function_impl_create(func
);
518 assert(strcmp(func
->name
, "main") == 0);
519 assert(ir
->parameters
.is_empty());
521 this->is_global
= false;
523 nir_builder_init(&b
, impl
);
524 b
.cursor
= nir_after_cf_list(&impl
->body
);
525 visit_exec_list(&ir
->body
, this);
527 this->is_global
= true;
534 nir_visitor::visit(ir_loop
*ir
)
537 visit_exec_list(&ir
->body_instructions
, this);
538 nir_pop_loop(&b
, NULL
);
542 nir_visitor::visit(ir_if
*ir
)
544 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
545 visit_exec_list(&ir
->then_instructions
, this);
546 nir_push_else(&b
, NULL
);
547 visit_exec_list(&ir
->else_instructions
, this);
548 nir_pop_if(&b
, NULL
);
552 nir_visitor::visit(ir_discard
*ir
)
555 * discards aren't treated as control flow, because before we lower them
556 * they can appear anywhere in the shader and the stuff after them may still
557 * be executed (yay, crazy GLSL rules!). However, after lowering, all the
558 * discards will be immediately followed by a return.
561 nir_intrinsic_instr
*discard
;
563 discard
= nir_intrinsic_instr_create(this->shader
,
564 nir_intrinsic_discard_if
);
566 nir_src_for_ssa(evaluate_rvalue(ir
->condition
));
568 discard
= nir_intrinsic_instr_create(this->shader
, nir_intrinsic_discard
);
571 nir_builder_instr_insert(&b
, &discard
->instr
);
575 nir_visitor::visit(ir_emit_vertex
*ir
)
577 nir_intrinsic_instr
*instr
=
578 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_emit_vertex
);
579 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
580 nir_builder_instr_insert(&b
, &instr
->instr
);
584 nir_visitor::visit(ir_end_primitive
*ir
)
586 nir_intrinsic_instr
*instr
=
587 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_end_primitive
);
588 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
589 nir_builder_instr_insert(&b
, &instr
->instr
);
593 nir_visitor::visit(ir_loop_jump
*ir
)
597 case ir_loop_jump::jump_break
:
598 type
= nir_jump_break
;
600 case ir_loop_jump::jump_continue
:
601 type
= nir_jump_continue
;
604 unreachable("not reached");
607 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, type
);
608 nir_builder_instr_insert(&b
, &instr
->instr
);
612 nir_visitor::visit(ir_return
*ir
)
614 assert(ir
->value
== NULL
);
615 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, nir_jump_return
);
616 nir_builder_instr_insert(&b
, &instr
->instr
);
620 nir_visitor::visit(ir_call
*ir
)
622 if (ir
->callee
->is_intrinsic()) {
625 switch (ir
->callee
->intrinsic_id
) {
626 case ir_intrinsic_atomic_counter_read
:
627 op
= nir_intrinsic_atomic_counter_read_deref
;
629 case ir_intrinsic_atomic_counter_increment
:
630 op
= nir_intrinsic_atomic_counter_inc_deref
;
632 case ir_intrinsic_atomic_counter_predecrement
:
633 op
= nir_intrinsic_atomic_counter_dec_deref
;
635 case ir_intrinsic_atomic_counter_add
:
636 op
= nir_intrinsic_atomic_counter_add_deref
;
638 case ir_intrinsic_atomic_counter_and
:
639 op
= nir_intrinsic_atomic_counter_and_deref
;
641 case ir_intrinsic_atomic_counter_or
:
642 op
= nir_intrinsic_atomic_counter_or_deref
;
644 case ir_intrinsic_atomic_counter_xor
:
645 op
= nir_intrinsic_atomic_counter_xor_deref
;
647 case ir_intrinsic_atomic_counter_min
:
648 op
= nir_intrinsic_atomic_counter_min_deref
;
650 case ir_intrinsic_atomic_counter_max
:
651 op
= nir_intrinsic_atomic_counter_max_deref
;
653 case ir_intrinsic_atomic_counter_exchange
:
654 op
= nir_intrinsic_atomic_counter_exchange_deref
;
656 case ir_intrinsic_atomic_counter_comp_swap
:
657 op
= nir_intrinsic_atomic_counter_comp_swap_deref
;
659 case ir_intrinsic_image_load
:
660 op
= nir_intrinsic_image_deref_load
;
662 case ir_intrinsic_image_store
:
663 op
= nir_intrinsic_image_deref_store
;
665 case ir_intrinsic_image_atomic_add
:
666 op
= nir_intrinsic_image_deref_atomic_add
;
668 case ir_intrinsic_image_atomic_min
:
669 op
= nir_intrinsic_image_deref_atomic_min
;
671 case ir_intrinsic_image_atomic_max
:
672 op
= nir_intrinsic_image_deref_atomic_max
;
674 case ir_intrinsic_image_atomic_and
:
675 op
= nir_intrinsic_image_deref_atomic_and
;
677 case ir_intrinsic_image_atomic_or
:
678 op
= nir_intrinsic_image_deref_atomic_or
;
680 case ir_intrinsic_image_atomic_xor
:
681 op
= nir_intrinsic_image_deref_atomic_xor
;
683 case ir_intrinsic_image_atomic_exchange
:
684 op
= nir_intrinsic_image_deref_atomic_exchange
;
686 case ir_intrinsic_image_atomic_comp_swap
:
687 op
= nir_intrinsic_image_deref_atomic_comp_swap
;
689 case ir_intrinsic_memory_barrier
:
690 op
= nir_intrinsic_memory_barrier
;
692 case ir_intrinsic_image_size
:
693 op
= nir_intrinsic_image_deref_size
;
695 case ir_intrinsic_image_samples
:
696 op
= nir_intrinsic_image_deref_samples
;
698 case ir_intrinsic_ssbo_store
:
699 op
= nir_intrinsic_store_ssbo
;
701 case ir_intrinsic_ssbo_load
:
702 op
= nir_intrinsic_load_ssbo
;
704 case ir_intrinsic_ssbo_atomic_add
:
705 op
= nir_intrinsic_ssbo_atomic_add
;
707 case ir_intrinsic_ssbo_atomic_and
:
708 op
= nir_intrinsic_ssbo_atomic_and
;
710 case ir_intrinsic_ssbo_atomic_or
:
711 op
= nir_intrinsic_ssbo_atomic_or
;
713 case ir_intrinsic_ssbo_atomic_xor
:
714 op
= nir_intrinsic_ssbo_atomic_xor
;
716 case ir_intrinsic_ssbo_atomic_min
:
717 assert(ir
->return_deref
);
718 if (ir
->return_deref
->type
== glsl_type::int_type
)
719 op
= nir_intrinsic_ssbo_atomic_imin
;
720 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
721 op
= nir_intrinsic_ssbo_atomic_umin
;
723 unreachable("Invalid type");
725 case ir_intrinsic_ssbo_atomic_max
:
726 assert(ir
->return_deref
);
727 if (ir
->return_deref
->type
== glsl_type::int_type
)
728 op
= nir_intrinsic_ssbo_atomic_imax
;
729 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
730 op
= nir_intrinsic_ssbo_atomic_umax
;
732 unreachable("Invalid type");
734 case ir_intrinsic_ssbo_atomic_exchange
:
735 op
= nir_intrinsic_ssbo_atomic_exchange
;
737 case ir_intrinsic_ssbo_atomic_comp_swap
:
738 op
= nir_intrinsic_ssbo_atomic_comp_swap
;
740 case ir_intrinsic_shader_clock
:
741 op
= nir_intrinsic_shader_clock
;
743 case ir_intrinsic_begin_invocation_interlock
:
744 op
= nir_intrinsic_begin_invocation_interlock
;
746 case ir_intrinsic_end_invocation_interlock
:
747 op
= nir_intrinsic_end_invocation_interlock
;
749 case ir_intrinsic_group_memory_barrier
:
750 op
= nir_intrinsic_group_memory_barrier
;
752 case ir_intrinsic_memory_barrier_atomic_counter
:
753 op
= nir_intrinsic_memory_barrier_atomic_counter
;
755 case ir_intrinsic_memory_barrier_buffer
:
756 op
= nir_intrinsic_memory_barrier_buffer
;
758 case ir_intrinsic_memory_barrier_image
:
759 op
= nir_intrinsic_memory_barrier_image
;
761 case ir_intrinsic_memory_barrier_shared
:
762 op
= nir_intrinsic_memory_barrier_shared
;
764 case ir_intrinsic_shared_load
:
765 op
= nir_intrinsic_load_shared
;
767 case ir_intrinsic_shared_store
:
768 op
= nir_intrinsic_store_shared
;
770 case ir_intrinsic_shared_atomic_add
:
771 op
= nir_intrinsic_shared_atomic_add
;
773 case ir_intrinsic_shared_atomic_and
:
774 op
= nir_intrinsic_shared_atomic_and
;
776 case ir_intrinsic_shared_atomic_or
:
777 op
= nir_intrinsic_shared_atomic_or
;
779 case ir_intrinsic_shared_atomic_xor
:
780 op
= nir_intrinsic_shared_atomic_xor
;
782 case ir_intrinsic_shared_atomic_min
:
783 assert(ir
->return_deref
);
784 if (ir
->return_deref
->type
== glsl_type::int_type
)
785 op
= nir_intrinsic_shared_atomic_imin
;
786 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
787 op
= nir_intrinsic_shared_atomic_umin
;
789 unreachable("Invalid type");
791 case ir_intrinsic_shared_atomic_max
:
792 assert(ir
->return_deref
);
793 if (ir
->return_deref
->type
== glsl_type::int_type
)
794 op
= nir_intrinsic_shared_atomic_imax
;
795 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
796 op
= nir_intrinsic_shared_atomic_umax
;
798 unreachable("Invalid type");
800 case ir_intrinsic_shared_atomic_exchange
:
801 op
= nir_intrinsic_shared_atomic_exchange
;
803 case ir_intrinsic_shared_atomic_comp_swap
:
804 op
= nir_intrinsic_shared_atomic_comp_swap
;
806 case ir_intrinsic_vote_any
:
807 op
= nir_intrinsic_vote_any
;
809 case ir_intrinsic_vote_all
:
810 op
= nir_intrinsic_vote_all
;
812 case ir_intrinsic_vote_eq
:
813 op
= nir_intrinsic_vote_ieq
;
815 case ir_intrinsic_ballot
:
816 op
= nir_intrinsic_ballot
;
818 case ir_intrinsic_read_invocation
:
819 op
= nir_intrinsic_read_invocation
;
821 case ir_intrinsic_read_first_invocation
:
822 op
= nir_intrinsic_read_first_invocation
;
825 unreachable("not reached");
828 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(shader
, op
);
829 nir_dest
*dest
= &instr
->dest
;
832 case nir_intrinsic_atomic_counter_read_deref
:
833 case nir_intrinsic_atomic_counter_inc_deref
:
834 case nir_intrinsic_atomic_counter_dec_deref
:
835 case nir_intrinsic_atomic_counter_add_deref
:
836 case nir_intrinsic_atomic_counter_min_deref
:
837 case nir_intrinsic_atomic_counter_max_deref
:
838 case nir_intrinsic_atomic_counter_and_deref
:
839 case nir_intrinsic_atomic_counter_or_deref
:
840 case nir_intrinsic_atomic_counter_xor_deref
:
841 case nir_intrinsic_atomic_counter_exchange_deref
:
842 case nir_intrinsic_atomic_counter_comp_swap_deref
: {
843 /* Set the counter variable dereference. */
844 exec_node
*param
= ir
->actual_parameters
.get_head();
845 ir_dereference
*counter
= (ir_dereference
*)param
;
847 instr
->src
[0] = nir_src_for_ssa(&evaluate_deref(counter
)->dest
.ssa
);
848 param
= param
->get_next();
850 /* Set the intrinsic destination. */
851 if (ir
->return_deref
) {
852 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
855 /* Set the intrinsic parameters. */
856 if (!param
->is_tail_sentinel()) {
858 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
859 param
= param
->get_next();
862 if (!param
->is_tail_sentinel()) {
864 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
865 param
= param
->get_next();
868 nir_builder_instr_insert(&b
, &instr
->instr
);
871 case nir_intrinsic_image_deref_load
:
872 case nir_intrinsic_image_deref_store
:
873 case nir_intrinsic_image_deref_atomic_add
:
874 case nir_intrinsic_image_deref_atomic_min
:
875 case nir_intrinsic_image_deref_atomic_max
:
876 case nir_intrinsic_image_deref_atomic_and
:
877 case nir_intrinsic_image_deref_atomic_or
:
878 case nir_intrinsic_image_deref_atomic_xor
:
879 case nir_intrinsic_image_deref_atomic_exchange
:
880 case nir_intrinsic_image_deref_atomic_comp_swap
:
881 case nir_intrinsic_image_deref_samples
:
882 case nir_intrinsic_image_deref_size
: {
883 nir_ssa_undef_instr
*instr_undef
=
884 nir_ssa_undef_instr_create(shader
, 1, 32);
885 nir_builder_instr_insert(&b
, &instr_undef
->instr
);
887 /* Set the image variable dereference. */
888 exec_node
*param
= ir
->actual_parameters
.get_head();
889 ir_dereference
*image
= (ir_dereference
*)param
;
890 const glsl_type
*type
=
891 image
->variable_referenced()->type
->without_array();
893 instr
->src
[0] = nir_src_for_ssa(&evaluate_deref(image
)->dest
.ssa
);
894 param
= param
->get_next();
896 /* Set the intrinsic destination. */
897 if (ir
->return_deref
) {
898 unsigned num_components
= ir
->return_deref
->type
->vector_elements
;
899 if (instr
->intrinsic
== nir_intrinsic_image_deref_size
)
900 instr
->num_components
= num_components
;
901 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
902 num_components
, 32, NULL
);
905 if (op
== nir_intrinsic_image_deref_size
||
906 op
== nir_intrinsic_image_deref_samples
) {
907 nir_builder_instr_insert(&b
, &instr
->instr
);
911 /* Set the address argument, extending the coordinate vector to four
914 nir_ssa_def
*src_addr
=
915 evaluate_rvalue((ir_dereference
*)param
);
916 nir_ssa_def
*srcs
[4];
918 for (int i
= 0; i
< 4; i
++) {
919 if (i
< type
->coordinate_components())
920 srcs
[i
] = nir_channel(&b
, src_addr
, i
);
922 srcs
[i
] = &instr_undef
->def
;
925 instr
->src
[1] = nir_src_for_ssa(nir_vec(&b
, srcs
, 4));
926 param
= param
->get_next();
928 /* Set the sample argument, which is undefined for single-sample
931 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_MS
) {
933 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
934 param
= param
->get_next();
936 instr
->src
[2] = nir_src_for_ssa(&instr_undef
->def
);
939 /* Set the intrinsic parameters. */
940 if (!param
->is_tail_sentinel()) {
942 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
943 param
= param
->get_next();
946 if (!param
->is_tail_sentinel()) {
948 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
949 param
= param
->get_next();
951 nir_builder_instr_insert(&b
, &instr
->instr
);
954 case nir_intrinsic_memory_barrier
:
955 case nir_intrinsic_group_memory_barrier
:
956 case nir_intrinsic_memory_barrier_atomic_counter
:
957 case nir_intrinsic_memory_barrier_buffer
:
958 case nir_intrinsic_memory_barrier_image
:
959 case nir_intrinsic_memory_barrier_shared
:
960 nir_builder_instr_insert(&b
, &instr
->instr
);
962 case nir_intrinsic_shader_clock
:
963 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 2, 32, NULL
);
964 instr
->num_components
= 2;
965 nir_builder_instr_insert(&b
, &instr
->instr
);
967 case nir_intrinsic_begin_invocation_interlock
:
968 nir_builder_instr_insert(&b
, &instr
->instr
);
970 case nir_intrinsic_end_invocation_interlock
:
971 nir_builder_instr_insert(&b
, &instr
->instr
);
973 case nir_intrinsic_store_ssbo
: {
974 exec_node
*param
= ir
->actual_parameters
.get_head();
975 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
977 param
= param
->get_next();
978 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
980 param
= param
->get_next();
981 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
983 param
= param
->get_next();
984 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
987 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(val
));
988 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(block
));
989 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(offset
));
990 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
991 instr
->num_components
= val
->type
->vector_elements
;
993 nir_builder_instr_insert(&b
, &instr
->instr
);
996 case nir_intrinsic_load_ssbo
: {
997 exec_node
*param
= ir
->actual_parameters
.get_head();
998 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1000 param
= param
->get_next();
1001 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1003 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(block
));
1004 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1006 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1007 instr
->num_components
= type
->vector_elements
;
1009 /* Setup destination register */
1010 unsigned bit_size
= glsl_get_bit_size(type
);
1011 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1012 type
->vector_elements
, bit_size
, NULL
);
1014 /* Insert the created nir instruction now since in the case of boolean
1015 * result we will need to emit another instruction after it
1017 nir_builder_instr_insert(&b
, &instr
->instr
);
1020 * In SSBO/UBO's, a true boolean value is any non-zero value, but we
1021 * consider a true boolean to be ~0. Fix this up with a != 0
1024 if (type
->is_boolean()) {
1025 nir_alu_instr
*load_ssbo_compare
=
1026 nir_alu_instr_create(shader
, nir_op_ine
);
1027 load_ssbo_compare
->src
[0].src
.is_ssa
= true;
1028 load_ssbo_compare
->src
[0].src
.ssa
= &instr
->dest
.ssa
;
1029 load_ssbo_compare
->src
[1].src
=
1030 nir_src_for_ssa(nir_imm_int(&b
, 0));
1031 for (unsigned i
= 0; i
< type
->vector_elements
; i
++)
1032 load_ssbo_compare
->src
[1].swizzle
[i
] = 0;
1033 nir_ssa_dest_init(&load_ssbo_compare
->instr
,
1034 &load_ssbo_compare
->dest
.dest
,
1035 type
->vector_elements
, bit_size
, NULL
);
1036 load_ssbo_compare
->dest
.write_mask
= (1 << type
->vector_elements
) - 1;
1037 nir_builder_instr_insert(&b
, &load_ssbo_compare
->instr
);
1038 dest
= &load_ssbo_compare
->dest
.dest
;
1042 case nir_intrinsic_ssbo_atomic_add
:
1043 case nir_intrinsic_ssbo_atomic_imin
:
1044 case nir_intrinsic_ssbo_atomic_umin
:
1045 case nir_intrinsic_ssbo_atomic_imax
:
1046 case nir_intrinsic_ssbo_atomic_umax
:
1047 case nir_intrinsic_ssbo_atomic_and
:
1048 case nir_intrinsic_ssbo_atomic_or
:
1049 case nir_intrinsic_ssbo_atomic_xor
:
1050 case nir_intrinsic_ssbo_atomic_exchange
:
1051 case nir_intrinsic_ssbo_atomic_comp_swap
: {
1052 int param_count
= ir
->actual_parameters
.length();
1053 assert(param_count
== 3 || param_count
== 4);
1056 exec_node
*param
= ir
->actual_parameters
.get_head();
1057 ir_instruction
*inst
= (ir_instruction
*) param
;
1058 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1061 param
= param
->get_next();
1062 inst
= (ir_instruction
*) param
;
1063 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1065 /* data1 parameter (this is always present) */
1066 param
= param
->get_next();
1067 inst
= (ir_instruction
*) param
;
1068 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1070 /* data2 parameter (only with atomic_comp_swap) */
1071 if (param_count
== 4) {
1072 assert(op
== nir_intrinsic_ssbo_atomic_comp_swap
);
1073 param
= param
->get_next();
1074 inst
= (ir_instruction
*) param
;
1075 instr
->src
[3] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1079 assert(ir
->return_deref
);
1080 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1081 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1082 nir_builder_instr_insert(&b
, &instr
->instr
);
1085 case nir_intrinsic_load_shared
: {
1086 exec_node
*param
= ir
->actual_parameters
.get_head();
1087 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1089 nir_intrinsic_set_base(instr
, 0);
1090 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(offset
));
1092 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1093 instr
->num_components
= type
->vector_elements
;
1095 /* Setup destination register */
1096 unsigned bit_size
= glsl_get_bit_size(type
);
1097 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1098 type
->vector_elements
, bit_size
, NULL
);
1100 nir_builder_instr_insert(&b
, &instr
->instr
);
1103 case nir_intrinsic_store_shared
: {
1104 exec_node
*param
= ir
->actual_parameters
.get_head();
1105 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1107 param
= param
->get_next();
1108 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1110 param
= param
->get_next();
1111 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1114 nir_intrinsic_set_base(instr
, 0);
1115 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1117 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1119 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(val
));
1120 instr
->num_components
= val
->type
->vector_elements
;
1122 nir_builder_instr_insert(&b
, &instr
->instr
);
1125 case nir_intrinsic_shared_atomic_add
:
1126 case nir_intrinsic_shared_atomic_imin
:
1127 case nir_intrinsic_shared_atomic_umin
:
1128 case nir_intrinsic_shared_atomic_imax
:
1129 case nir_intrinsic_shared_atomic_umax
:
1130 case nir_intrinsic_shared_atomic_and
:
1131 case nir_intrinsic_shared_atomic_or
:
1132 case nir_intrinsic_shared_atomic_xor
:
1133 case nir_intrinsic_shared_atomic_exchange
:
1134 case nir_intrinsic_shared_atomic_comp_swap
: {
1135 int param_count
= ir
->actual_parameters
.length();
1136 assert(param_count
== 2 || param_count
== 3);
1139 exec_node
*param
= ir
->actual_parameters
.get_head();
1140 ir_instruction
*inst
= (ir_instruction
*) param
;
1141 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1143 /* data1 parameter (this is always present) */
1144 param
= param
->get_next();
1145 inst
= (ir_instruction
*) param
;
1146 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1148 /* data2 parameter (only with atomic_comp_swap) */
1149 if (param_count
== 3) {
1150 assert(op
== nir_intrinsic_shared_atomic_comp_swap
);
1151 param
= param
->get_next();
1152 inst
= (ir_instruction
*) param
;
1154 nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1158 assert(ir
->return_deref
);
1159 unsigned bit_size
= glsl_get_bit_size(ir
->return_deref
->type
);
1160 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1161 ir
->return_deref
->type
->vector_elements
,
1163 nir_builder_instr_insert(&b
, &instr
->instr
);
1166 case nir_intrinsic_vote_any
:
1167 case nir_intrinsic_vote_all
:
1168 case nir_intrinsic_vote_ieq
: {
1169 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1170 instr
->num_components
= 1;
1172 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1173 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1175 nir_builder_instr_insert(&b
, &instr
->instr
);
1179 case nir_intrinsic_ballot
: {
1180 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1181 ir
->return_deref
->type
->vector_elements
, 64, NULL
);
1182 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1184 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1185 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1187 nir_builder_instr_insert(&b
, &instr
->instr
);
1190 case nir_intrinsic_read_invocation
: {
1191 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1192 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1193 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1195 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1196 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1198 ir_rvalue
*invocation
= (ir_rvalue
*) ir
->actual_parameters
.get_head()->next
;
1199 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(invocation
));
1201 nir_builder_instr_insert(&b
, &instr
->instr
);
1204 case nir_intrinsic_read_first_invocation
: {
1205 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1206 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1207 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1209 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1210 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1212 nir_builder_instr_insert(&b
, &instr
->instr
);
1216 unreachable("not reached");
1219 if (ir
->return_deref
)
1220 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), &dest
->ssa
, ~0);
1225 unreachable("glsl_to_nir only handles function calls to intrinsics");
1229 nir_visitor::visit(ir_assignment
*ir
)
1231 unsigned num_components
= ir
->lhs
->type
->vector_elements
;
1233 b
.exact
= ir
->lhs
->variable_referenced()->data
.invariant
||
1234 ir
->lhs
->variable_referenced()->data
.precise
;
1236 if ((ir
->rhs
->as_dereference() || ir
->rhs
->as_constant()) &&
1237 (ir
->write_mask
== (1 << num_components
) - 1 || ir
->write_mask
== 0)) {
1238 if (ir
->condition
) {
1239 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1240 nir_copy_deref(&b
, evaluate_deref(ir
->lhs
), evaluate_deref(ir
->rhs
));
1241 nir_pop_if(&b
, NULL
);
1243 nir_copy_deref(&b
, evaluate_deref(ir
->lhs
), evaluate_deref(ir
->rhs
));
1248 assert(ir
->rhs
->type
->is_scalar() || ir
->rhs
->type
->is_vector());
1250 ir
->lhs
->accept(this);
1251 nir_deref_instr
*lhs_deref
= this->deref
;
1252 nir_ssa_def
*src
= evaluate_rvalue(ir
->rhs
);
1254 if (ir
->write_mask
!= (1 << num_components
) - 1 && ir
->write_mask
!= 0) {
1255 /* GLSL IR will give us the input to the write-masked assignment in a
1256 * single packed vector. So, for example, if the writemask is xzw, then
1257 * we have to swizzle x -> x, y -> z, and z -> w and get the y component
1261 unsigned component
= 0;
1262 for (unsigned i
= 0; i
< 4; i
++) {
1263 swiz
[i
] = ir
->write_mask
& (1 << i
) ? component
++ : 0;
1265 src
= nir_swizzle(&b
, src
, swiz
, num_components
, !supports_ints
);
1268 if (ir
->condition
) {
1269 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1270 nir_store_deref(&b
, lhs_deref
, src
, ir
->write_mask
);
1271 nir_pop_if(&b
, NULL
);
1273 nir_store_deref(&b
, lhs_deref
, src
, ir
->write_mask
);
1278 * Given an instruction, returns a pointer to its destination or NULL if there
1279 * is no destination.
1281 * Note that this only handles instructions we generate at this level.
1284 get_instr_dest(nir_instr
*instr
)
1286 nir_alu_instr
*alu_instr
;
1287 nir_intrinsic_instr
*intrinsic_instr
;
1288 nir_tex_instr
*tex_instr
;
1290 switch (instr
->type
) {
1291 case nir_instr_type_alu
:
1292 alu_instr
= nir_instr_as_alu(instr
);
1293 return &alu_instr
->dest
.dest
;
1295 case nir_instr_type_intrinsic
:
1296 intrinsic_instr
= nir_instr_as_intrinsic(instr
);
1297 if (nir_intrinsic_infos
[intrinsic_instr
->intrinsic
].has_dest
)
1298 return &intrinsic_instr
->dest
;
1302 case nir_instr_type_tex
:
1303 tex_instr
= nir_instr_as_tex(instr
);
1304 return &tex_instr
->dest
;
1307 unreachable("not reached");
1314 nir_visitor::add_instr(nir_instr
*instr
, unsigned num_components
,
1317 nir_dest
*dest
= get_instr_dest(instr
);
1320 nir_ssa_dest_init(instr
, dest
, num_components
, bit_size
, NULL
);
1322 nir_builder_instr_insert(&b
, instr
);
1325 assert(dest
->is_ssa
);
1326 this->result
= &dest
->ssa
;
1331 nir_visitor::evaluate_rvalue(ir_rvalue
* ir
)
1334 if (ir
->as_dereference() || ir
->as_constant()) {
1336 * A dereference is being used on the right hand side, which means we
1337 * must emit a variable load.
1340 this->result
= nir_load_deref(&b
, this->deref
);
1343 return this->result
;
1347 type_is_float(glsl_base_type type
)
1349 return type
== GLSL_TYPE_FLOAT
|| type
== GLSL_TYPE_DOUBLE
||
1350 type
== GLSL_TYPE_FLOAT16
;
1354 type_is_signed(glsl_base_type type
)
1356 return type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_INT64
||
1357 type
== GLSL_TYPE_INT16
;
1361 nir_visitor::visit(ir_expression
*ir
)
1363 /* Some special cases */
1364 switch (ir
->operation
) {
1365 case ir_binop_ubo_load
: {
1366 nir_intrinsic_instr
*load
=
1367 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_ubo
);
1368 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1369 load
->num_components
= ir
->type
->vector_elements
;
1370 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1371 load
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1372 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1375 * In UBO's, a true boolean value is any non-zero value, but we consider
1376 * a true boolean to be ~0. Fix this up with a != 0 comparison.
1379 if (ir
->type
->is_boolean())
1380 this->result
= nir_ine(&b
, &load
->dest
.ssa
, nir_imm_int(&b
, 0));
1385 case ir_unop_interpolate_at_centroid
:
1386 case ir_binop_interpolate_at_offset
:
1387 case ir_binop_interpolate_at_sample
: {
1388 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1389 ir_swizzle
*swizzle
= NULL
;
1391 /* the api does not allow a swizzle here, but the varying packing code
1392 * may have pushed one into here.
1394 swizzle
= ir
->operands
[0]->as_swizzle();
1396 deref
= swizzle
->val
->as_dereference();
1400 deref
->accept(this);
1402 nir_intrinsic_op op
;
1403 if (this->deref
->mode
== nir_var_shader_in
) {
1404 switch (ir
->operation
) {
1405 case ir_unop_interpolate_at_centroid
:
1406 op
= nir_intrinsic_interp_deref_at_centroid
;
1408 case ir_binop_interpolate_at_offset
:
1409 op
= nir_intrinsic_interp_deref_at_offset
;
1411 case ir_binop_interpolate_at_sample
:
1412 op
= nir_intrinsic_interp_deref_at_sample
;
1415 unreachable("Invalid interpolation intrinsic");
1418 /* This case can happen if the vertex shader does not write the
1419 * given varying. In this case, the linker will lower it to a
1420 * global variable. Since interpolating a variable makes no
1421 * sense, we'll just turn it into a load which will probably
1422 * eventually end up as an SSA definition.
1424 assert(this->deref
->mode
== nir_var_global
);
1425 op
= nir_intrinsic_load_deref
;
1428 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(shader
, op
);
1429 intrin
->num_components
= deref
->type
->vector_elements
;
1430 intrin
->src
[0] = nir_src_for_ssa(&this->deref
->dest
.ssa
);
1432 if (intrin
->intrinsic
== nir_intrinsic_interp_deref_at_offset
||
1433 intrin
->intrinsic
== nir_intrinsic_interp_deref_at_sample
)
1434 intrin
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1436 unsigned bit_size
= glsl_get_bit_size(deref
->type
);
1437 add_instr(&intrin
->instr
, deref
->type
->vector_elements
, bit_size
);
1440 unsigned swiz
[4] = {
1441 swizzle
->mask
.x
, swizzle
->mask
.y
, swizzle
->mask
.z
, swizzle
->mask
.w
1444 result
= nir_swizzle(&b
, result
, swiz
,
1445 swizzle
->type
->vector_elements
, false);
1455 nir_ssa_def
*srcs
[4];
1456 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1457 srcs
[i
] = evaluate_rvalue(ir
->operands
[i
]);
1459 glsl_base_type types
[4];
1460 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1462 types
[i
] = ir
->operands
[i
]->type
->base_type
;
1464 types
[i
] = GLSL_TYPE_FLOAT
;
1466 glsl_base_type out_type
;
1468 out_type
= ir
->type
->base_type
;
1470 out_type
= GLSL_TYPE_FLOAT
;
1472 switch (ir
->operation
) {
1473 case ir_unop_bit_not
: result
= nir_inot(&b
, srcs
[0]); break;
1474 case ir_unop_logic_not
:
1475 result
= supports_ints
? nir_inot(&b
, srcs
[0]) : nir_fnot(&b
, srcs
[0]);
1478 result
= type_is_float(types
[0]) ? nir_fneg(&b
, srcs
[0])
1479 : nir_ineg(&b
, srcs
[0]);
1482 result
= type_is_float(types
[0]) ? nir_fabs(&b
, srcs
[0])
1483 : nir_iabs(&b
, srcs
[0]);
1485 case ir_unop_saturate
:
1486 assert(type_is_float(types
[0]));
1487 result
= nir_fsat(&b
, srcs
[0]);
1490 result
= type_is_float(types
[0]) ? nir_fsign(&b
, srcs
[0])
1491 : nir_isign(&b
, srcs
[0]);
1493 case ir_unop_rcp
: result
= nir_frcp(&b
, srcs
[0]); break;
1494 case ir_unop_rsq
: result
= nir_frsq(&b
, srcs
[0]); break;
1495 case ir_unop_sqrt
: result
= nir_fsqrt(&b
, srcs
[0]); break;
1496 case ir_unop_exp
: unreachable("ir_unop_exp should have been lowered");
1497 case ir_unop_log
: unreachable("ir_unop_log should have been lowered");
1498 case ir_unop_exp2
: result
= nir_fexp2(&b
, srcs
[0]); break;
1499 case ir_unop_log2
: result
= nir_flog2(&b
, srcs
[0]); break;
1501 result
= supports_ints
? nir_i2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1504 result
= supports_ints
? nir_u2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1507 result
= supports_ints
? nir_b2f(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1541 case ir_unop_i642u64
:
1542 case ir_unop_u642i64
: {
1543 nir_alu_type src_type
= nir_get_nir_type_for_glsl_base_type(types
[0]);
1544 nir_alu_type dst_type
= nir_get_nir_type_for_glsl_base_type(out_type
);
1545 result
= nir_build_alu(&b
, nir_type_conversion_op(src_type
, dst_type
,
1546 nir_rounding_mode_undef
),
1547 srcs
[0], NULL
, NULL
, NULL
);
1548 /* b2i and b2f don't have fixed bit-size versions so the builder will
1549 * just assume 32 and we have to fix it up here.
1551 result
->bit_size
= nir_alu_type_get_type_size(dst_type
);
1555 case ir_unop_bitcast_i2f
:
1556 case ir_unop_bitcast_f2i
:
1557 case ir_unop_bitcast_u2f
:
1558 case ir_unop_bitcast_f2u
:
1559 case ir_unop_bitcast_i642d
:
1560 case ir_unop_bitcast_d2i64
:
1561 case ir_unop_bitcast_u642d
:
1562 case ir_unop_bitcast_d2u64
:
1563 case ir_unop_subroutine_to_int
:
1565 result
= nir_imov(&b
, srcs
[0]);
1567 case ir_unop_trunc
: result
= nir_ftrunc(&b
, srcs
[0]); break;
1568 case ir_unop_ceil
: result
= nir_fceil(&b
, srcs
[0]); break;
1569 case ir_unop_floor
: result
= nir_ffloor(&b
, srcs
[0]); break;
1570 case ir_unop_fract
: result
= nir_ffract(&b
, srcs
[0]); break;
1571 case ir_unop_frexp_exp
: result
= nir_frexp_exp(&b
, srcs
[0]); break;
1572 case ir_unop_frexp_sig
: result
= nir_frexp_sig(&b
, srcs
[0]); break;
1573 case ir_unop_round_even
: result
= nir_fround_even(&b
, srcs
[0]); break;
1574 case ir_unop_sin
: result
= nir_fsin(&b
, srcs
[0]); break;
1575 case ir_unop_cos
: result
= nir_fcos(&b
, srcs
[0]); break;
1576 case ir_unop_dFdx
: result
= nir_fddx(&b
, srcs
[0]); break;
1577 case ir_unop_dFdy
: result
= nir_fddy(&b
, srcs
[0]); break;
1578 case ir_unop_dFdx_fine
: result
= nir_fddx_fine(&b
, srcs
[0]); break;
1579 case ir_unop_dFdy_fine
: result
= nir_fddy_fine(&b
, srcs
[0]); break;
1580 case ir_unop_dFdx_coarse
: result
= nir_fddx_coarse(&b
, srcs
[0]); break;
1581 case ir_unop_dFdy_coarse
: result
= nir_fddy_coarse(&b
, srcs
[0]); break;
1582 case ir_unop_pack_snorm_2x16
:
1583 result
= nir_pack_snorm_2x16(&b
, srcs
[0]);
1585 case ir_unop_pack_snorm_4x8
:
1586 result
= nir_pack_snorm_4x8(&b
, srcs
[0]);
1588 case ir_unop_pack_unorm_2x16
:
1589 result
= nir_pack_unorm_2x16(&b
, srcs
[0]);
1591 case ir_unop_pack_unorm_4x8
:
1592 result
= nir_pack_unorm_4x8(&b
, srcs
[0]);
1594 case ir_unop_pack_half_2x16
:
1595 result
= nir_pack_half_2x16(&b
, srcs
[0]);
1597 case ir_unop_unpack_snorm_2x16
:
1598 result
= nir_unpack_snorm_2x16(&b
, srcs
[0]);
1600 case ir_unop_unpack_snorm_4x8
:
1601 result
= nir_unpack_snorm_4x8(&b
, srcs
[0]);
1603 case ir_unop_unpack_unorm_2x16
:
1604 result
= nir_unpack_unorm_2x16(&b
, srcs
[0]);
1606 case ir_unop_unpack_unorm_4x8
:
1607 result
= nir_unpack_unorm_4x8(&b
, srcs
[0]);
1609 case ir_unop_unpack_half_2x16
:
1610 result
= nir_unpack_half_2x16(&b
, srcs
[0]);
1612 case ir_unop_pack_sampler_2x32
:
1613 case ir_unop_pack_image_2x32
:
1614 case ir_unop_pack_double_2x32
:
1615 case ir_unop_pack_int_2x32
:
1616 case ir_unop_pack_uint_2x32
:
1617 result
= nir_pack_64_2x32(&b
, srcs
[0]);
1619 case ir_unop_unpack_sampler_2x32
:
1620 case ir_unop_unpack_image_2x32
:
1621 case ir_unop_unpack_double_2x32
:
1622 case ir_unop_unpack_int_2x32
:
1623 case ir_unop_unpack_uint_2x32
:
1624 result
= nir_unpack_64_2x32(&b
, srcs
[0]);
1626 case ir_unop_bitfield_reverse
:
1627 result
= nir_bitfield_reverse(&b
, srcs
[0]);
1629 case ir_unop_bit_count
:
1630 result
= nir_bit_count(&b
, srcs
[0]);
1632 case ir_unop_find_msb
:
1634 case GLSL_TYPE_UINT
:
1635 result
= nir_ufind_msb(&b
, srcs
[0]);
1638 result
= nir_ifind_msb(&b
, srcs
[0]);
1641 unreachable("Invalid type for findMSB()");
1644 case ir_unop_find_lsb
:
1645 result
= nir_find_lsb(&b
, srcs
[0]);
1649 switch (ir
->type
->vector_elements
) {
1651 switch (ir
->operands
[0]->type
->vector_elements
) {
1652 case 1: result
= nir_fnoise1_1(&b
, srcs
[0]); break;
1653 case 2: result
= nir_fnoise1_2(&b
, srcs
[0]); break;
1654 case 3: result
= nir_fnoise1_3(&b
, srcs
[0]); break;
1655 case 4: result
= nir_fnoise1_4(&b
, srcs
[0]); break;
1656 default: unreachable("not reached");
1660 switch (ir
->operands
[0]->type
->vector_elements
) {
1661 case 1: result
= nir_fnoise2_1(&b
, srcs
[0]); break;
1662 case 2: result
= nir_fnoise2_2(&b
, srcs
[0]); break;
1663 case 3: result
= nir_fnoise2_3(&b
, srcs
[0]); break;
1664 case 4: result
= nir_fnoise2_4(&b
, srcs
[0]); break;
1665 default: unreachable("not reached");
1669 switch (ir
->operands
[0]->type
->vector_elements
) {
1670 case 1: result
= nir_fnoise3_1(&b
, srcs
[0]); break;
1671 case 2: result
= nir_fnoise3_2(&b
, srcs
[0]); break;
1672 case 3: result
= nir_fnoise3_3(&b
, srcs
[0]); break;
1673 case 4: result
= nir_fnoise3_4(&b
, srcs
[0]); break;
1674 default: unreachable("not reached");
1678 switch (ir
->operands
[0]->type
->vector_elements
) {
1679 case 1: result
= nir_fnoise4_1(&b
, srcs
[0]); break;
1680 case 2: result
= nir_fnoise4_2(&b
, srcs
[0]); break;
1681 case 3: result
= nir_fnoise4_3(&b
, srcs
[0]); break;
1682 case 4: result
= nir_fnoise4_4(&b
, srcs
[0]); break;
1683 default: unreachable("not reached");
1687 unreachable("not reached");
1690 case ir_unop_get_buffer_size
: {
1691 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(
1693 nir_intrinsic_get_buffer_size
);
1694 load
->num_components
= ir
->type
->vector_elements
;
1695 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1696 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1697 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1702 result
= type_is_float(out_type
) ? nir_fadd(&b
, srcs
[0], srcs
[1])
1703 : nir_iadd(&b
, srcs
[0], srcs
[1]);
1706 result
= type_is_float(out_type
) ? nir_fsub(&b
, srcs
[0], srcs
[1])
1707 : nir_isub(&b
, srcs
[0], srcs
[1]);
1710 result
= type_is_float(out_type
) ? nir_fmul(&b
, srcs
[0], srcs
[1])
1711 : nir_imul(&b
, srcs
[0], srcs
[1]);
1714 if (type_is_float(out_type
))
1715 result
= nir_fdiv(&b
, srcs
[0], srcs
[1]);
1716 else if (type_is_signed(out_type
))
1717 result
= nir_idiv(&b
, srcs
[0], srcs
[1]);
1719 result
= nir_udiv(&b
, srcs
[0], srcs
[1]);
1722 result
= type_is_float(out_type
) ? nir_fmod(&b
, srcs
[0], srcs
[1])
1723 : nir_umod(&b
, srcs
[0], srcs
[1]);
1726 if (type_is_float(out_type
))
1727 result
= nir_fmin(&b
, srcs
[0], srcs
[1]);
1728 else if (type_is_signed(out_type
))
1729 result
= nir_imin(&b
, srcs
[0], srcs
[1]);
1731 result
= nir_umin(&b
, srcs
[0], srcs
[1]);
1734 if (type_is_float(out_type
))
1735 result
= nir_fmax(&b
, srcs
[0], srcs
[1]);
1736 else if (type_is_signed(out_type
))
1737 result
= nir_imax(&b
, srcs
[0], srcs
[1]);
1739 result
= nir_umax(&b
, srcs
[0], srcs
[1]);
1741 case ir_binop_pow
: result
= nir_fpow(&b
, srcs
[0], srcs
[1]); break;
1742 case ir_binop_bit_and
: result
= nir_iand(&b
, srcs
[0], srcs
[1]); break;
1743 case ir_binop_bit_or
: result
= nir_ior(&b
, srcs
[0], srcs
[1]); break;
1744 case ir_binop_bit_xor
: result
= nir_ixor(&b
, srcs
[0], srcs
[1]); break;
1745 case ir_binop_logic_and
:
1746 result
= supports_ints
? nir_iand(&b
, srcs
[0], srcs
[1])
1747 : nir_fand(&b
, srcs
[0], srcs
[1]);
1749 case ir_binop_logic_or
:
1750 result
= supports_ints
? nir_ior(&b
, srcs
[0], srcs
[1])
1751 : nir_for(&b
, srcs
[0], srcs
[1]);
1753 case ir_binop_logic_xor
:
1754 result
= supports_ints
? nir_ixor(&b
, srcs
[0], srcs
[1])
1755 : nir_fxor(&b
, srcs
[0], srcs
[1]);
1757 case ir_binop_lshift
: result
= nir_ishl(&b
, srcs
[0], srcs
[1]); break;
1758 case ir_binop_rshift
:
1759 result
= (type_is_signed(out_type
)) ? nir_ishr(&b
, srcs
[0], srcs
[1])
1760 : nir_ushr(&b
, srcs
[0], srcs
[1]);
1762 case ir_binop_imul_high
:
1763 result
= (out_type
== GLSL_TYPE_INT
) ? nir_imul_high(&b
, srcs
[0], srcs
[1])
1764 : nir_umul_high(&b
, srcs
[0], srcs
[1]);
1766 case ir_binop_carry
: result
= nir_uadd_carry(&b
, srcs
[0], srcs
[1]); break;
1767 case ir_binop_borrow
: result
= nir_usub_borrow(&b
, srcs
[0], srcs
[1]); break;
1769 if (supports_ints
) {
1770 if (type_is_float(types
[0]))
1771 result
= nir_flt(&b
, srcs
[0], srcs
[1]);
1772 else if (type_is_signed(types
[0]))
1773 result
= nir_ilt(&b
, srcs
[0], srcs
[1]);
1775 result
= nir_ult(&b
, srcs
[0], srcs
[1]);
1777 result
= nir_slt(&b
, srcs
[0], srcs
[1]);
1780 case ir_binop_gequal
:
1781 if (supports_ints
) {
1782 if (type_is_float(types
[0]))
1783 result
= nir_fge(&b
, srcs
[0], srcs
[1]);
1784 else if (type_is_signed(types
[0]))
1785 result
= nir_ige(&b
, srcs
[0], srcs
[1]);
1787 result
= nir_uge(&b
, srcs
[0], srcs
[1]);
1789 result
= nir_sge(&b
, srcs
[0], srcs
[1]);
1792 case ir_binop_equal
:
1793 if (supports_ints
) {
1794 if (type_is_float(types
[0]))
1795 result
= nir_feq(&b
, srcs
[0], srcs
[1]);
1797 result
= nir_ieq(&b
, srcs
[0], srcs
[1]);
1799 result
= nir_seq(&b
, srcs
[0], srcs
[1]);
1802 case ir_binop_nequal
:
1803 if (supports_ints
) {
1804 if (type_is_float(types
[0]))
1805 result
= nir_fne(&b
, srcs
[0], srcs
[1]);
1807 result
= nir_ine(&b
, srcs
[0], srcs
[1]);
1809 result
= nir_sne(&b
, srcs
[0], srcs
[1]);
1812 case ir_binop_all_equal
:
1813 if (supports_ints
) {
1814 if (type_is_float(types
[0])) {
1815 switch (ir
->operands
[0]->type
->vector_elements
) {
1816 case 1: result
= nir_feq(&b
, srcs
[0], srcs
[1]); break;
1817 case 2: result
= nir_ball_fequal2(&b
, srcs
[0], srcs
[1]); break;
1818 case 3: result
= nir_ball_fequal3(&b
, srcs
[0], srcs
[1]); break;
1819 case 4: result
= nir_ball_fequal4(&b
, srcs
[0], srcs
[1]); break;
1821 unreachable("not reached");
1824 switch (ir
->operands
[0]->type
->vector_elements
) {
1825 case 1: result
= nir_ieq(&b
, srcs
[0], srcs
[1]); break;
1826 case 2: result
= nir_ball_iequal2(&b
, srcs
[0], srcs
[1]); break;
1827 case 3: result
= nir_ball_iequal3(&b
, srcs
[0], srcs
[1]); break;
1828 case 4: result
= nir_ball_iequal4(&b
, srcs
[0], srcs
[1]); break;
1830 unreachable("not reached");
1834 switch (ir
->operands
[0]->type
->vector_elements
) {
1835 case 1: result
= nir_seq(&b
, srcs
[0], srcs
[1]); break;
1836 case 2: result
= nir_fall_equal2(&b
, srcs
[0], srcs
[1]); break;
1837 case 3: result
= nir_fall_equal3(&b
, srcs
[0], srcs
[1]); break;
1838 case 4: result
= nir_fall_equal4(&b
, srcs
[0], srcs
[1]); break;
1840 unreachable("not reached");
1844 case ir_binop_any_nequal
:
1845 if (supports_ints
) {
1846 if (type_is_float(types
[0])) {
1847 switch (ir
->operands
[0]->type
->vector_elements
) {
1848 case 1: result
= nir_fne(&b
, srcs
[0], srcs
[1]); break;
1849 case 2: result
= nir_bany_fnequal2(&b
, srcs
[0], srcs
[1]); break;
1850 case 3: result
= nir_bany_fnequal3(&b
, srcs
[0], srcs
[1]); break;
1851 case 4: result
= nir_bany_fnequal4(&b
, srcs
[0], srcs
[1]); break;
1853 unreachable("not reached");
1856 switch (ir
->operands
[0]->type
->vector_elements
) {
1857 case 1: result
= nir_ine(&b
, srcs
[0], srcs
[1]); break;
1858 case 2: result
= nir_bany_inequal2(&b
, srcs
[0], srcs
[1]); break;
1859 case 3: result
= nir_bany_inequal3(&b
, srcs
[0], srcs
[1]); break;
1860 case 4: result
= nir_bany_inequal4(&b
, srcs
[0], srcs
[1]); break;
1862 unreachable("not reached");
1866 switch (ir
->operands
[0]->type
->vector_elements
) {
1867 case 1: result
= nir_sne(&b
, srcs
[0], srcs
[1]); break;
1868 case 2: result
= nir_fany_nequal2(&b
, srcs
[0], srcs
[1]); break;
1869 case 3: result
= nir_fany_nequal3(&b
, srcs
[0], srcs
[1]); break;
1870 case 4: result
= nir_fany_nequal4(&b
, srcs
[0], srcs
[1]); break;
1872 unreachable("not reached");
1877 switch (ir
->operands
[0]->type
->vector_elements
) {
1878 case 2: result
= nir_fdot2(&b
, srcs
[0], srcs
[1]); break;
1879 case 3: result
= nir_fdot3(&b
, srcs
[0], srcs
[1]); break;
1880 case 4: result
= nir_fdot4(&b
, srcs
[0], srcs
[1]); break;
1882 unreachable("not reached");
1885 case ir_binop_vector_extract
: {
1886 result
= nir_channel(&b
, srcs
[0], 0);
1887 for (unsigned i
= 1; i
< ir
->operands
[0]->type
->vector_elements
; i
++) {
1888 nir_ssa_def
*swizzled
= nir_channel(&b
, srcs
[0], i
);
1889 result
= nir_bcsel(&b
, nir_ieq(&b
, srcs
[1], nir_imm_int(&b
, i
)),
1895 case ir_binop_ldexp
: result
= nir_ldexp(&b
, srcs
[0], srcs
[1]); break;
1897 result
= nir_ffma(&b
, srcs
[0], srcs
[1], srcs
[2]);
1900 result
= nir_flrp(&b
, srcs
[0], srcs
[1], srcs
[2]);
1904 result
= nir_bcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
1906 result
= nir_fcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
1908 case ir_triop_bitfield_extract
:
1909 result
= (out_type
== GLSL_TYPE_INT
) ?
1910 nir_ibitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]) :
1911 nir_ubitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]);
1913 case ir_quadop_bitfield_insert
:
1914 result
= nir_bitfield_insert(&b
, srcs
[0], srcs
[1], srcs
[2], srcs
[3]);
1916 case ir_quadop_vector
:
1917 result
= nir_vec(&b
, srcs
, ir
->type
->vector_elements
);
1921 unreachable("not reached");
1926 nir_visitor::visit(ir_swizzle
*ir
)
1928 unsigned swizzle
[4] = { ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
};
1929 result
= nir_swizzle(&b
, evaluate_rvalue(ir
->val
), swizzle
,
1930 ir
->type
->vector_elements
, !supports_ints
);
1934 nir_visitor::visit(ir_texture
*ir
)
1941 num_srcs
= 1; /* coordinate */
1946 op
= (ir
->op
== ir_txb
) ? nir_texop_txb
: nir_texop_txl
;
1947 num_srcs
= 2; /* coordinate, bias/lod */
1951 op
= nir_texop_txd
; /* coordinate, dPdx, dPdy */
1957 if (ir
->lod_info
.lod
!= NULL
)
1958 num_srcs
= 2; /* coordinate, lod */
1960 num_srcs
= 1; /* coordinate */
1964 op
= nir_texop_txf_ms
;
1965 num_srcs
= 2; /* coordinate, sample_index */
1970 if (ir
->lod_info
.lod
!= NULL
)
1971 num_srcs
= 1; /* lod */
1978 num_srcs
= 1; /* coordinate */
1983 num_srcs
= 1; /* coordinate */
1986 case ir_query_levels
:
1987 op
= nir_texop_query_levels
;
1991 case ir_texture_samples
:
1992 op
= nir_texop_texture_samples
;
1996 case ir_samples_identical
:
1997 op
= nir_texop_samples_identical
;
1998 num_srcs
= 1; /* coordinate */
2002 unreachable("not reached");
2005 if (ir
->projector
!= NULL
)
2007 if (ir
->shadow_comparator
!= NULL
)
2009 if (ir
->offset
!= NULL
)
2012 /* Add one for the texture deref */
2015 nir_tex_instr
*instr
= nir_tex_instr_create(this->shader
, num_srcs
);
2018 instr
->sampler_dim
=
2019 (glsl_sampler_dim
) ir
->sampler
->type
->sampler_dimensionality
;
2020 instr
->is_array
= ir
->sampler
->type
->sampler_array
;
2021 instr
->is_shadow
= ir
->sampler
->type
->sampler_shadow
;
2022 if (instr
->is_shadow
)
2023 instr
->is_new_style_shadow
= (ir
->type
->vector_elements
== 1);
2024 switch (ir
->type
->base_type
) {
2025 case GLSL_TYPE_FLOAT
:
2026 instr
->dest_type
= nir_type_float
;
2029 instr
->dest_type
= nir_type_int
;
2031 case GLSL_TYPE_BOOL
:
2032 case GLSL_TYPE_UINT
:
2033 instr
->dest_type
= nir_type_uint
;
2036 unreachable("not reached");
2039 nir_deref_instr
*sampler_deref
= evaluate_deref(ir
->sampler
);
2040 instr
->src
[0].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2041 instr
->src
[0].src_type
= nir_tex_src_texture_deref
;
2043 unsigned src_number
= 1;
2045 if (ir
->coordinate
!= NULL
) {
2046 instr
->coord_components
= ir
->coordinate
->type
->vector_elements
;
2047 instr
->src
[src_number
].src
=
2048 nir_src_for_ssa(evaluate_rvalue(ir
->coordinate
));
2049 instr
->src
[src_number
].src_type
= nir_tex_src_coord
;
2053 if (ir
->projector
!= NULL
) {
2054 instr
->src
[src_number
].src
=
2055 nir_src_for_ssa(evaluate_rvalue(ir
->projector
));
2056 instr
->src
[src_number
].src_type
= nir_tex_src_projector
;
2060 if (ir
->shadow_comparator
!= NULL
) {
2061 instr
->src
[src_number
].src
=
2062 nir_src_for_ssa(evaluate_rvalue(ir
->shadow_comparator
));
2063 instr
->src
[src_number
].src_type
= nir_tex_src_comparator
;
2067 if (ir
->offset
!= NULL
) {
2068 /* we don't support multiple offsets yet */
2069 assert(ir
->offset
->type
->is_vector() || ir
->offset
->type
->is_scalar());
2071 instr
->src
[src_number
].src
=
2072 nir_src_for_ssa(evaluate_rvalue(ir
->offset
));
2073 instr
->src
[src_number
].src_type
= nir_tex_src_offset
;
2079 instr
->src
[src_number
].src
=
2080 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.bias
));
2081 instr
->src
[src_number
].src_type
= nir_tex_src_bias
;
2088 if (ir
->lod_info
.lod
!= NULL
) {
2089 instr
->src
[src_number
].src
=
2090 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.lod
));
2091 instr
->src
[src_number
].src_type
= nir_tex_src_lod
;
2097 instr
->src
[src_number
].src
=
2098 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdx
));
2099 instr
->src
[src_number
].src_type
= nir_tex_src_ddx
;
2101 instr
->src
[src_number
].src
=
2102 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdy
));
2103 instr
->src
[src_number
].src_type
= nir_tex_src_ddy
;
2108 instr
->src
[src_number
].src
=
2109 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.sample_index
));
2110 instr
->src
[src_number
].src_type
= nir_tex_src_ms_index
;
2115 instr
->component
= ir
->lod_info
.component
->as_constant()->value
.u
[0];
2122 assert(src_number
== num_srcs
);
2124 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2125 add_instr(&instr
->instr
, nir_tex_instr_dest_size(instr
), bit_size
);
2129 nir_visitor::visit(ir_constant
*ir
)
2132 * We don't know if this variable is an array or struct that gets
2133 * dereferenced, so do the safe thing an make it a variable with a
2134 * constant initializer and return a dereference.
2138 nir_local_variable_create(this->impl
, ir
->type
, "const_temp");
2139 var
->data
.read_only
= true;
2140 var
->constant_initializer
= constant_copy(ir
, var
);
2142 this->deref
= nir_build_deref_var(&b
, var
);
2146 nir_visitor::visit(ir_dereference_variable
*ir
)
2148 struct hash_entry
*entry
=
2149 _mesa_hash_table_search(this->var_table
, ir
->var
);
2151 nir_variable
*var
= (nir_variable
*) entry
->data
;
2153 this->deref
= nir_build_deref_var(&b
, var
);
2157 nir_visitor::visit(ir_dereference_record
*ir
)
2159 ir
->record
->accept(this);
2161 int field_index
= ir
->field_idx
;
2162 assert(field_index
>= 0);
2164 this->deref
= nir_build_deref_struct(&b
, this->deref
, field_index
);
2168 nir_visitor::visit(ir_dereference_array
*ir
)
2170 nir_ssa_def
*index
= evaluate_rvalue(ir
->array_index
);
2172 ir
->array
->accept(this);
2174 this->deref
= nir_build_deref_array(&b
, this->deref
, index
);
2178 nir_visitor::visit(ir_barrier
*)
2180 nir_intrinsic_instr
*instr
=
2181 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_barrier
);
2182 nir_builder_instr_insert(&b
, &instr
->instr
);