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"
37 * pass to lower GLSL IR to NIR
39 * This will lower variable dereferences to loads/stores of corresponding
40 * variables in NIR - the variables will be converted to registers in a later
46 class nir_visitor
: public ir_visitor
49 nir_visitor(nir_shader
*shader
);
52 virtual void visit(ir_variable
*);
53 virtual void visit(ir_function
*);
54 virtual void visit(ir_function_signature
*);
55 virtual void visit(ir_loop
*);
56 virtual void visit(ir_if
*);
57 virtual void visit(ir_discard
*);
58 virtual void visit(ir_loop_jump
*);
59 virtual void visit(ir_return
*);
60 virtual void visit(ir_call
*);
61 virtual void visit(ir_assignment
*);
62 virtual void visit(ir_emit_vertex
*);
63 virtual void visit(ir_end_primitive
*);
64 virtual void visit(ir_expression
*);
65 virtual void visit(ir_swizzle
*);
66 virtual void visit(ir_texture
*);
67 virtual void visit(ir_constant
*);
68 virtual void visit(ir_dereference_variable
*);
69 virtual void visit(ir_dereference_record
*);
70 virtual void visit(ir_dereference_array
*);
71 virtual void visit(ir_barrier
*);
73 void create_function(ir_function_signature
*ir
);
76 void add_instr(nir_instr
*instr
, unsigned num_components
, unsigned bit_size
);
77 nir_ssa_def
*evaluate_rvalue(ir_rvalue
*ir
);
79 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
**srcs
);
80 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
);
81 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
,
83 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
,
84 nir_ssa_def
*src2
, nir_ssa_def
*src3
);
89 nir_function_impl
*impl
;
91 nir_ssa_def
*result
; /* result of the expression tree last visited */
93 nir_deref_var
*evaluate_deref(nir_instr
*mem_ctx
, ir_instruction
*ir
);
95 /* the head of the dereference chain we're creating */
96 nir_deref_var
*deref_head
;
97 /* the tail of the dereference chain we're creating */
98 nir_deref
*deref_tail
;
100 nir_variable
*var
; /* variable created by ir_variable visitor */
102 /* whether the IR we're operating on is per-function or global */
105 /* map of ir_variable -> nir_variable */
106 struct hash_table
*var_table
;
108 /* map of ir_function_signature -> nir_function_overload */
109 struct hash_table
*overload_table
;
113 * This visitor runs before the main visitor, calling create_function() for
114 * each function so that the main visitor can resolve forward references in
118 class nir_function_visitor
: public ir_hierarchical_visitor
121 nir_function_visitor(nir_visitor
*v
) : visitor(v
)
124 virtual ir_visitor_status
visit_enter(ir_function
*);
127 nir_visitor
*visitor
;
130 } /* end of anonymous namespace */
133 nir_remap_attributes(nir_shader
*shader
)
135 nir_foreach_variable(var
, &shader
->inputs
) {
136 var
->data
.location
+= _mesa_bitcount_64(shader
->info
.double_inputs_read
&
137 BITFIELD64_MASK(var
->data
.location
));
140 /* Once the remap is done, reset double_inputs_read, so later it will have
141 * which location/slots are doubles */
142 shader
->info
.double_inputs_read
= 0;
146 glsl_to_nir(const struct gl_shader_program
*shader_prog
,
147 gl_shader_stage stage
,
148 const nir_shader_compiler_options
*options
)
150 struct gl_linked_shader
*sh
= shader_prog
->_LinkedShaders
[stage
];
152 nir_shader
*shader
= nir_shader_create(NULL
, stage
, options
,
155 nir_visitor
v1(shader
);
156 nir_function_visitor
v2(&v1
);
158 visit_exec_list(sh
->ir
, &v1
);
160 nir_lower_constant_initializers(shader
, (nir_variable_mode
)~0);
162 /* Remap the locations to slots so those requiring two slots will occupy
163 * two locations. For instance, if we have in the IR code a dvec3 attr0 in
164 * location 0 and vec4 attr1 in location 1, in NIR attr0 will use
165 * locations/slots 0 and 1, and attr1 will use location/slot 2 */
166 if (shader
->info
.stage
== MESA_SHADER_VERTEX
)
167 nir_remap_attributes(shader
);
169 shader
->info
.name
= ralloc_asprintf(shader
, "GLSL%d", shader_prog
->Name
);
170 if (shader_prog
->Label
)
171 shader
->info
.label
= ralloc_strdup(shader
, shader_prog
->Label
);
173 /* Check for transform feedback varyings specified via the API */
174 shader
->info
.has_transform_feedback_varyings
=
175 shader_prog
->TransformFeedback
.NumVarying
> 0;
177 /* Check for transform feedback varyings specified in the Shader */
178 if (shader_prog
->last_vert_prog
)
179 shader
->info
.has_transform_feedback_varyings
|=
180 shader_prog
->last_vert_prog
->sh
.LinkedTransformFeedback
->NumVarying
> 0;
185 nir_visitor::nir_visitor(nir_shader
*shader
)
187 this->supports_ints
= shader
->options
->native_integers
;
188 this->shader
= shader
;
189 this->is_global
= true;
190 this->var_table
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
191 _mesa_key_pointer_equal
);
192 this->overload_table
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
193 _mesa_key_pointer_equal
);
197 this->deref_head
= NULL
;
198 this->deref_tail
= NULL
;
199 memset(&this->b
, 0, sizeof(this->b
));
202 nir_visitor::~nir_visitor()
204 _mesa_hash_table_destroy(this->var_table
, NULL
);
205 _mesa_hash_table_destroy(this->overload_table
, NULL
);
209 nir_visitor::evaluate_deref(nir_instr
*mem_ctx
, ir_instruction
*ir
)
212 ralloc_steal(mem_ctx
, this->deref_head
);
213 return this->deref_head
;
216 static nir_constant
*
217 constant_copy(ir_constant
*ir
, void *mem_ctx
)
222 nir_constant
*ret
= ralloc(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
.compact
= false;
327 switch(ir
->data
.mode
) {
329 case ir_var_temporary
:
331 var
->data
.mode
= nir_var_global
;
333 var
->data
.mode
= nir_var_local
;
336 case ir_var_function_in
:
337 case ir_var_function_out
:
338 case ir_var_function_inout
:
339 case ir_var_const_in
:
340 var
->data
.mode
= nir_var_local
;
343 case ir_var_shader_in
:
344 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
345 ir
->data
.location
== VARYING_SLOT_FACE
) {
346 /* For whatever reason, GLSL IR makes gl_FrontFacing an input */
347 var
->data
.location
= SYSTEM_VALUE_FRONT_FACE
;
348 var
->data
.mode
= nir_var_system_value
;
349 } else if (shader
->info
.stage
== MESA_SHADER_GEOMETRY
&&
350 ir
->data
.location
== VARYING_SLOT_PRIMITIVE_ID
) {
351 /* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */
352 var
->data
.location
= SYSTEM_VALUE_PRIMITIVE_ID
;
353 var
->data
.mode
= nir_var_system_value
;
355 var
->data
.mode
= nir_var_shader_in
;
357 if (shader
->info
.stage
== MESA_SHADER_TESS_EVAL
&&
358 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
359 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
360 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
364 /* Mark all the locations that require two slots */
365 if (glsl_type_is_dual_slot(glsl_without_array(var
->type
))) {
366 for (uint i
= 0; i
< glsl_count_attribute_slots(var
->type
, true); i
++) {
367 uint64_t bitfield
= BITFIELD64_BIT(var
->data
.location
+ i
);
368 shader
->info
.double_inputs_read
|= bitfield
;
373 case ir_var_shader_out
:
374 var
->data
.mode
= nir_var_shader_out
;
375 if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
&&
376 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
377 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
378 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
383 var
->data
.mode
= nir_var_uniform
;
386 case ir_var_shader_storage
:
387 var
->data
.mode
= nir_var_shader_storage
;
390 case ir_var_system_value
:
391 var
->data
.mode
= nir_var_system_value
;
395 unreachable("not reached");
398 var
->data
.interpolation
= ir
->data
.interpolation
;
399 var
->data
.origin_upper_left
= ir
->data
.origin_upper_left
;
400 var
->data
.pixel_center_integer
= ir
->data
.pixel_center_integer
;
401 var
->data
.location_frac
= ir
->data
.location_frac
;
403 switch (ir
->data
.depth_layout
) {
404 case ir_depth_layout_none
:
405 var
->data
.depth_layout
= nir_depth_layout_none
;
407 case ir_depth_layout_any
:
408 var
->data
.depth_layout
= nir_depth_layout_any
;
410 case ir_depth_layout_greater
:
411 var
->data
.depth_layout
= nir_depth_layout_greater
;
413 case ir_depth_layout_less
:
414 var
->data
.depth_layout
= nir_depth_layout_less
;
416 case ir_depth_layout_unchanged
:
417 var
->data
.depth_layout
= nir_depth_layout_unchanged
;
420 unreachable("not reached");
423 var
->data
.index
= ir
->data
.index
;
424 var
->data
.descriptor_set
= 0;
425 var
->data
.binding
= ir
->data
.binding
;
426 var
->data
.offset
= ir
->data
.offset
;
427 var
->data
.image
.read_only
= ir
->data
.memory_read_only
;
428 var
->data
.image
.write_only
= ir
->data
.memory_write_only
;
429 var
->data
.image
.coherent
= ir
->data
.memory_coherent
;
430 var
->data
.image
._volatile
= ir
->data
.memory_volatile
;
431 var
->data
.image
.restrict_flag
= ir
->data
.memory_restrict
;
432 var
->data
.image
.format
= ir
->data
.image_format
;
433 var
->data
.fb_fetch_output
= ir
->data
.fb_fetch_output
;
435 var
->num_state_slots
= ir
->get_num_state_slots();
436 if (var
->num_state_slots
> 0) {
437 var
->state_slots
= ralloc_array(var
, nir_state_slot
,
438 var
->num_state_slots
);
440 ir_state_slot
*state_slots
= ir
->get_state_slots();
441 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
442 for (unsigned j
= 0; j
< 5; j
++)
443 var
->state_slots
[i
].tokens
[j
] = state_slots
[i
].tokens
[j
];
444 var
->state_slots
[i
].swizzle
= state_slots
[i
].swizzle
;
447 var
->state_slots
= NULL
;
450 var
->constant_initializer
= constant_copy(ir
->constant_initializer
, var
);
452 var
->interface_type
= ir
->get_interface_type();
454 if (var
->data
.mode
== nir_var_local
)
455 nir_function_impl_add_variable(impl
, var
);
457 nir_shader_add_variable(shader
, var
);
459 _mesa_hash_table_insert(var_table
, ir
, var
);
464 nir_function_visitor::visit_enter(ir_function
*ir
)
466 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
) {
467 visitor
->create_function(sig
);
469 return visit_continue_with_parent
;
473 nir_visitor::create_function(ir_function_signature
*ir
)
475 if (ir
->is_intrinsic())
478 nir_function
*func
= nir_function_create(shader
, ir
->function_name());
480 assert(ir
->parameters
.is_empty());
481 assert(ir
->return_type
== glsl_type::void_type
);
483 _mesa_hash_table_insert(this->overload_table
, ir
, func
);
487 nir_visitor::visit(ir_function
*ir
)
489 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
)
494 nir_visitor::visit(ir_function_signature
*ir
)
496 if (ir
->is_intrinsic())
499 struct hash_entry
*entry
=
500 _mesa_hash_table_search(this->overload_table
, ir
);
503 nir_function
*func
= (nir_function
*) entry
->data
;
505 if (ir
->is_defined
) {
506 nir_function_impl
*impl
= nir_function_impl_create(func
);
509 assert(strcmp(func
->name
, "main") == 0);
510 assert(ir
->parameters
.is_empty());
511 assert(func
->return_type
== glsl_type::void_type
);
513 this->is_global
= false;
515 nir_builder_init(&b
, impl
);
516 b
.cursor
= nir_after_cf_list(&impl
->body
);
517 visit_exec_list(&ir
->body
, this);
519 this->is_global
= true;
526 nir_visitor::visit(ir_loop
*ir
)
529 visit_exec_list(&ir
->body_instructions
, this);
530 nir_pop_loop(&b
, NULL
);
534 nir_visitor::visit(ir_if
*ir
)
536 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
537 visit_exec_list(&ir
->then_instructions
, this);
538 nir_push_else(&b
, NULL
);
539 visit_exec_list(&ir
->else_instructions
, this);
540 nir_pop_if(&b
, NULL
);
544 nir_visitor::visit(ir_discard
*ir
)
547 * discards aren't treated as control flow, because before we lower them
548 * they can appear anywhere in the shader and the stuff after them may still
549 * be executed (yay, crazy GLSL rules!). However, after lowering, all the
550 * discards will be immediately followed by a return.
553 nir_intrinsic_instr
*discard
;
555 discard
= nir_intrinsic_instr_create(this->shader
,
556 nir_intrinsic_discard_if
);
558 nir_src_for_ssa(evaluate_rvalue(ir
->condition
));
560 discard
= nir_intrinsic_instr_create(this->shader
, nir_intrinsic_discard
);
563 nir_builder_instr_insert(&b
, &discard
->instr
);
567 nir_visitor::visit(ir_emit_vertex
*ir
)
569 nir_intrinsic_instr
*instr
=
570 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_emit_vertex
);
571 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
572 nir_builder_instr_insert(&b
, &instr
->instr
);
576 nir_visitor::visit(ir_end_primitive
*ir
)
578 nir_intrinsic_instr
*instr
=
579 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_end_primitive
);
580 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
581 nir_builder_instr_insert(&b
, &instr
->instr
);
585 nir_visitor::visit(ir_loop_jump
*ir
)
589 case ir_loop_jump::jump_break
:
590 type
= nir_jump_break
;
592 case ir_loop_jump::jump_continue
:
593 type
= nir_jump_continue
;
596 unreachable("not reached");
599 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, type
);
600 nir_builder_instr_insert(&b
, &instr
->instr
);
604 nir_visitor::visit(ir_return
*ir
)
606 if (ir
->value
!= NULL
) {
607 nir_intrinsic_instr
*copy
=
608 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_copy_var
);
610 copy
->variables
[0] = nir_deref_var_create(copy
, this->impl
->return_var
);
611 copy
->variables
[1] = evaluate_deref(©
->instr
, ir
->value
);
614 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, nir_jump_return
);
615 nir_builder_instr_insert(&b
, &instr
->instr
);
619 nir_visitor::visit(ir_call
*ir
)
621 if (ir
->callee
->is_intrinsic()) {
624 switch (ir
->callee
->intrinsic_id
) {
625 case ir_intrinsic_atomic_counter_read
:
626 op
= nir_intrinsic_atomic_counter_read_var
;
628 case ir_intrinsic_atomic_counter_increment
:
629 op
= nir_intrinsic_atomic_counter_inc_var
;
631 case ir_intrinsic_atomic_counter_predecrement
:
632 op
= nir_intrinsic_atomic_counter_dec_var
;
634 case ir_intrinsic_atomic_counter_add
:
635 op
= nir_intrinsic_atomic_counter_add_var
;
637 case ir_intrinsic_atomic_counter_and
:
638 op
= nir_intrinsic_atomic_counter_and_var
;
640 case ir_intrinsic_atomic_counter_or
:
641 op
= nir_intrinsic_atomic_counter_or_var
;
643 case ir_intrinsic_atomic_counter_xor
:
644 op
= nir_intrinsic_atomic_counter_xor_var
;
646 case ir_intrinsic_atomic_counter_min
:
647 op
= nir_intrinsic_atomic_counter_min_var
;
649 case ir_intrinsic_atomic_counter_max
:
650 op
= nir_intrinsic_atomic_counter_max_var
;
652 case ir_intrinsic_atomic_counter_exchange
:
653 op
= nir_intrinsic_atomic_counter_exchange_var
;
655 case ir_intrinsic_atomic_counter_comp_swap
:
656 op
= nir_intrinsic_atomic_counter_comp_swap_var
;
658 case ir_intrinsic_image_load
:
659 op
= nir_intrinsic_image_load
;
661 case ir_intrinsic_image_store
:
662 op
= nir_intrinsic_image_store
;
664 case ir_intrinsic_image_atomic_add
:
665 op
= nir_intrinsic_image_atomic_add
;
667 case ir_intrinsic_image_atomic_min
:
668 op
= nir_intrinsic_image_atomic_min
;
670 case ir_intrinsic_image_atomic_max
:
671 op
= nir_intrinsic_image_atomic_max
;
673 case ir_intrinsic_image_atomic_and
:
674 op
= nir_intrinsic_image_atomic_and
;
676 case ir_intrinsic_image_atomic_or
:
677 op
= nir_intrinsic_image_atomic_or
;
679 case ir_intrinsic_image_atomic_xor
:
680 op
= nir_intrinsic_image_atomic_xor
;
682 case ir_intrinsic_image_atomic_exchange
:
683 op
= nir_intrinsic_image_atomic_exchange
;
685 case ir_intrinsic_image_atomic_comp_swap
:
686 op
= nir_intrinsic_image_atomic_comp_swap
;
688 case ir_intrinsic_memory_barrier
:
689 op
= nir_intrinsic_memory_barrier
;
691 case ir_intrinsic_image_size
:
692 op
= nir_intrinsic_image_size
;
694 case ir_intrinsic_image_samples
:
695 op
= nir_intrinsic_image_samples
;
697 case ir_intrinsic_ssbo_store
:
698 op
= nir_intrinsic_store_ssbo
;
700 case ir_intrinsic_ssbo_load
:
701 op
= nir_intrinsic_load_ssbo
;
703 case ir_intrinsic_ssbo_atomic_add
:
704 op
= nir_intrinsic_ssbo_atomic_add
;
706 case ir_intrinsic_ssbo_atomic_and
:
707 op
= nir_intrinsic_ssbo_atomic_and
;
709 case ir_intrinsic_ssbo_atomic_or
:
710 op
= nir_intrinsic_ssbo_atomic_or
;
712 case ir_intrinsic_ssbo_atomic_xor
:
713 op
= nir_intrinsic_ssbo_atomic_xor
;
715 case ir_intrinsic_ssbo_atomic_min
:
716 assert(ir
->return_deref
);
717 if (ir
->return_deref
->type
== glsl_type::int_type
)
718 op
= nir_intrinsic_ssbo_atomic_imin
;
719 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
720 op
= nir_intrinsic_ssbo_atomic_umin
;
722 unreachable("Invalid type");
724 case ir_intrinsic_ssbo_atomic_max
:
725 assert(ir
->return_deref
);
726 if (ir
->return_deref
->type
== glsl_type::int_type
)
727 op
= nir_intrinsic_ssbo_atomic_imax
;
728 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
729 op
= nir_intrinsic_ssbo_atomic_umax
;
731 unreachable("Invalid type");
733 case ir_intrinsic_ssbo_atomic_exchange
:
734 op
= nir_intrinsic_ssbo_atomic_exchange
;
736 case ir_intrinsic_ssbo_atomic_comp_swap
:
737 op
= nir_intrinsic_ssbo_atomic_comp_swap
;
739 case ir_intrinsic_shader_clock
:
740 op
= nir_intrinsic_shader_clock
;
742 case ir_intrinsic_group_memory_barrier
:
743 op
= nir_intrinsic_group_memory_barrier
;
745 case ir_intrinsic_memory_barrier_atomic_counter
:
746 op
= nir_intrinsic_memory_barrier_atomic_counter
;
748 case ir_intrinsic_memory_barrier_buffer
:
749 op
= nir_intrinsic_memory_barrier_buffer
;
751 case ir_intrinsic_memory_barrier_image
:
752 op
= nir_intrinsic_memory_barrier_image
;
754 case ir_intrinsic_memory_barrier_shared
:
755 op
= nir_intrinsic_memory_barrier_shared
;
757 case ir_intrinsic_shared_load
:
758 op
= nir_intrinsic_load_shared
;
760 case ir_intrinsic_shared_store
:
761 op
= nir_intrinsic_store_shared
;
763 case ir_intrinsic_shared_atomic_add
:
764 op
= nir_intrinsic_shared_atomic_add
;
766 case ir_intrinsic_shared_atomic_and
:
767 op
= nir_intrinsic_shared_atomic_and
;
769 case ir_intrinsic_shared_atomic_or
:
770 op
= nir_intrinsic_shared_atomic_or
;
772 case ir_intrinsic_shared_atomic_xor
:
773 op
= nir_intrinsic_shared_atomic_xor
;
775 case ir_intrinsic_shared_atomic_min
:
776 assert(ir
->return_deref
);
777 if (ir
->return_deref
->type
== glsl_type::int_type
)
778 op
= nir_intrinsic_shared_atomic_imin
;
779 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
780 op
= nir_intrinsic_shared_atomic_umin
;
782 unreachable("Invalid type");
784 case ir_intrinsic_shared_atomic_max
:
785 assert(ir
->return_deref
);
786 if (ir
->return_deref
->type
== glsl_type::int_type
)
787 op
= nir_intrinsic_shared_atomic_imax
;
788 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
789 op
= nir_intrinsic_shared_atomic_umax
;
791 unreachable("Invalid type");
793 case ir_intrinsic_shared_atomic_exchange
:
794 op
= nir_intrinsic_shared_atomic_exchange
;
796 case ir_intrinsic_shared_atomic_comp_swap
:
797 op
= nir_intrinsic_shared_atomic_comp_swap
;
799 case ir_intrinsic_vote_any
:
800 op
= nir_intrinsic_vote_any
;
802 case ir_intrinsic_vote_all
:
803 op
= nir_intrinsic_vote_all
;
805 case ir_intrinsic_vote_eq
:
806 op
= nir_intrinsic_vote_eq
;
808 case ir_intrinsic_ballot
:
809 op
= nir_intrinsic_ballot
;
811 case ir_intrinsic_read_invocation
:
812 op
= nir_intrinsic_read_invocation
;
814 case ir_intrinsic_read_first_invocation
:
815 op
= nir_intrinsic_read_first_invocation
;
818 unreachable("not reached");
821 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(shader
, op
);
822 nir_dest
*dest
= &instr
->dest
;
825 case nir_intrinsic_atomic_counter_read_var
:
826 case nir_intrinsic_atomic_counter_inc_var
:
827 case nir_intrinsic_atomic_counter_dec_var
:
828 case nir_intrinsic_atomic_counter_add_var
:
829 case nir_intrinsic_atomic_counter_min_var
:
830 case nir_intrinsic_atomic_counter_max_var
:
831 case nir_intrinsic_atomic_counter_and_var
:
832 case nir_intrinsic_atomic_counter_or_var
:
833 case nir_intrinsic_atomic_counter_xor_var
:
834 case nir_intrinsic_atomic_counter_exchange_var
:
835 case nir_intrinsic_atomic_counter_comp_swap_var
: {
836 /* Set the counter variable dereference. */
837 exec_node
*param
= ir
->actual_parameters
.get_head();
838 ir_dereference
*counter
= (ir_dereference
*)param
;
840 instr
->variables
[0] = evaluate_deref(&instr
->instr
, counter
);
841 param
= param
->get_next();
843 /* Set the intrinsic destination. */
844 if (ir
->return_deref
) {
845 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
848 /* Set the intrinsic parameters. */
849 if (!param
->is_tail_sentinel()) {
851 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
852 param
= param
->get_next();
855 if (!param
->is_tail_sentinel()) {
857 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
858 param
= param
->get_next();
861 nir_builder_instr_insert(&b
, &instr
->instr
);
864 case nir_intrinsic_image_load
:
865 case nir_intrinsic_image_store
:
866 case nir_intrinsic_image_atomic_add
:
867 case nir_intrinsic_image_atomic_min
:
868 case nir_intrinsic_image_atomic_max
:
869 case nir_intrinsic_image_atomic_and
:
870 case nir_intrinsic_image_atomic_or
:
871 case nir_intrinsic_image_atomic_xor
:
872 case nir_intrinsic_image_atomic_exchange
:
873 case nir_intrinsic_image_atomic_comp_swap
:
874 case nir_intrinsic_image_samples
:
875 case nir_intrinsic_image_size
: {
876 nir_ssa_undef_instr
*instr_undef
=
877 nir_ssa_undef_instr_create(shader
, 1, 32);
878 nir_builder_instr_insert(&b
, &instr_undef
->instr
);
880 /* Set the image variable dereference. */
881 exec_node
*param
= ir
->actual_parameters
.get_head();
882 ir_dereference
*image
= (ir_dereference
*)param
;
883 const glsl_type
*type
=
884 image
->variable_referenced()->type
->without_array();
886 instr
->variables
[0] = evaluate_deref(&instr
->instr
, image
);
887 param
= param
->get_next();
889 /* Set the intrinsic destination. */
890 if (ir
->return_deref
) {
891 unsigned num_components
= ir
->return_deref
->type
->vector_elements
;
892 if (instr
->intrinsic
== nir_intrinsic_image_size
)
893 instr
->num_components
= num_components
;
894 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
895 num_components
, 32, NULL
);
898 if (op
== nir_intrinsic_image_size
||
899 op
== nir_intrinsic_image_samples
) {
900 nir_builder_instr_insert(&b
, &instr
->instr
);
904 /* Set the address argument, extending the coordinate vector to four
907 nir_ssa_def
*src_addr
=
908 evaluate_rvalue((ir_dereference
*)param
);
909 nir_ssa_def
*srcs
[4];
911 for (int i
= 0; i
< 4; i
++) {
912 if (i
< type
->coordinate_components())
913 srcs
[i
] = nir_channel(&b
, src_addr
, i
);
915 srcs
[i
] = &instr_undef
->def
;
918 instr
->src
[0] = nir_src_for_ssa(nir_vec(&b
, srcs
, 4));
919 param
= param
->get_next();
921 /* Set the sample argument, which is undefined for single-sample
924 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_MS
) {
926 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
927 param
= param
->get_next();
929 instr
->src
[1] = nir_src_for_ssa(&instr_undef
->def
);
932 /* Set the intrinsic parameters. */
933 if (!param
->is_tail_sentinel()) {
935 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
936 param
= param
->get_next();
939 if (!param
->is_tail_sentinel()) {
941 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
942 param
= param
->get_next();
944 nir_builder_instr_insert(&b
, &instr
->instr
);
947 case nir_intrinsic_memory_barrier
:
948 case nir_intrinsic_group_memory_barrier
:
949 case nir_intrinsic_memory_barrier_atomic_counter
:
950 case nir_intrinsic_memory_barrier_buffer
:
951 case nir_intrinsic_memory_barrier_image
:
952 case nir_intrinsic_memory_barrier_shared
:
953 nir_builder_instr_insert(&b
, &instr
->instr
);
955 case nir_intrinsic_shader_clock
:
956 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 2, 32, NULL
);
957 instr
->num_components
= 2;
958 nir_builder_instr_insert(&b
, &instr
->instr
);
960 case nir_intrinsic_store_ssbo
: {
961 exec_node
*param
= ir
->actual_parameters
.get_head();
962 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
964 param
= param
->get_next();
965 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
967 param
= param
->get_next();
968 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
970 param
= param
->get_next();
971 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
974 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(val
));
975 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(block
));
976 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(offset
));
977 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
978 instr
->num_components
= val
->type
->vector_elements
;
980 nir_builder_instr_insert(&b
, &instr
->instr
);
983 case nir_intrinsic_load_ssbo
: {
984 exec_node
*param
= ir
->actual_parameters
.get_head();
985 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
987 param
= param
->get_next();
988 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
990 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(block
));
991 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
993 const glsl_type
*type
= ir
->return_deref
->var
->type
;
994 instr
->num_components
= type
->vector_elements
;
996 /* Setup destination register */
997 unsigned bit_size
= glsl_get_bit_size(type
);
998 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
999 type
->vector_elements
, bit_size
, NULL
);
1001 /* Insert the created nir instruction now since in the case of boolean
1002 * result we will need to emit another instruction after it
1004 nir_builder_instr_insert(&b
, &instr
->instr
);
1007 * In SSBO/UBO's, a true boolean value is any non-zero value, but we
1008 * consider a true boolean to be ~0. Fix this up with a != 0
1011 if (type
->is_boolean()) {
1012 nir_alu_instr
*load_ssbo_compare
=
1013 nir_alu_instr_create(shader
, nir_op_ine
);
1014 load_ssbo_compare
->src
[0].src
.is_ssa
= true;
1015 load_ssbo_compare
->src
[0].src
.ssa
= &instr
->dest
.ssa
;
1016 load_ssbo_compare
->src
[1].src
=
1017 nir_src_for_ssa(nir_imm_int(&b
, 0));
1018 for (unsigned i
= 0; i
< type
->vector_elements
; i
++)
1019 load_ssbo_compare
->src
[1].swizzle
[i
] = 0;
1020 nir_ssa_dest_init(&load_ssbo_compare
->instr
,
1021 &load_ssbo_compare
->dest
.dest
,
1022 type
->vector_elements
, bit_size
, NULL
);
1023 load_ssbo_compare
->dest
.write_mask
= (1 << type
->vector_elements
) - 1;
1024 nir_builder_instr_insert(&b
, &load_ssbo_compare
->instr
);
1025 dest
= &load_ssbo_compare
->dest
.dest
;
1029 case nir_intrinsic_ssbo_atomic_add
:
1030 case nir_intrinsic_ssbo_atomic_imin
:
1031 case nir_intrinsic_ssbo_atomic_umin
:
1032 case nir_intrinsic_ssbo_atomic_imax
:
1033 case nir_intrinsic_ssbo_atomic_umax
:
1034 case nir_intrinsic_ssbo_atomic_and
:
1035 case nir_intrinsic_ssbo_atomic_or
:
1036 case nir_intrinsic_ssbo_atomic_xor
:
1037 case nir_intrinsic_ssbo_atomic_exchange
:
1038 case nir_intrinsic_ssbo_atomic_comp_swap
: {
1039 int param_count
= ir
->actual_parameters
.length();
1040 assert(param_count
== 3 || param_count
== 4);
1043 exec_node
*param
= ir
->actual_parameters
.get_head();
1044 ir_instruction
*inst
= (ir_instruction
*) param
;
1045 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1048 param
= param
->get_next();
1049 inst
= (ir_instruction
*) param
;
1050 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1052 /* data1 parameter (this is always present) */
1053 param
= param
->get_next();
1054 inst
= (ir_instruction
*) param
;
1055 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1057 /* data2 parameter (only with atomic_comp_swap) */
1058 if (param_count
== 4) {
1059 assert(op
== nir_intrinsic_ssbo_atomic_comp_swap
);
1060 param
= param
->get_next();
1061 inst
= (ir_instruction
*) param
;
1062 instr
->src
[3] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1066 assert(ir
->return_deref
);
1067 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1068 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1069 nir_builder_instr_insert(&b
, &instr
->instr
);
1072 case nir_intrinsic_load_shared
: {
1073 exec_node
*param
= ir
->actual_parameters
.get_head();
1074 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1076 nir_intrinsic_set_base(instr
, 0);
1077 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(offset
));
1079 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1080 instr
->num_components
= type
->vector_elements
;
1082 /* Setup destination register */
1083 unsigned bit_size
= glsl_get_bit_size(type
);
1084 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1085 type
->vector_elements
, bit_size
, NULL
);
1087 nir_builder_instr_insert(&b
, &instr
->instr
);
1090 case nir_intrinsic_store_shared
: {
1091 exec_node
*param
= ir
->actual_parameters
.get_head();
1092 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1094 param
= param
->get_next();
1095 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1097 param
= param
->get_next();
1098 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1101 nir_intrinsic_set_base(instr
, 0);
1102 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1104 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1106 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(val
));
1107 instr
->num_components
= val
->type
->vector_elements
;
1109 nir_builder_instr_insert(&b
, &instr
->instr
);
1112 case nir_intrinsic_shared_atomic_add
:
1113 case nir_intrinsic_shared_atomic_imin
:
1114 case nir_intrinsic_shared_atomic_umin
:
1115 case nir_intrinsic_shared_atomic_imax
:
1116 case nir_intrinsic_shared_atomic_umax
:
1117 case nir_intrinsic_shared_atomic_and
:
1118 case nir_intrinsic_shared_atomic_or
:
1119 case nir_intrinsic_shared_atomic_xor
:
1120 case nir_intrinsic_shared_atomic_exchange
:
1121 case nir_intrinsic_shared_atomic_comp_swap
: {
1122 int param_count
= ir
->actual_parameters
.length();
1123 assert(param_count
== 2 || param_count
== 3);
1126 exec_node
*param
= ir
->actual_parameters
.get_head();
1127 ir_instruction
*inst
= (ir_instruction
*) param
;
1128 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1130 /* data1 parameter (this is always present) */
1131 param
= param
->get_next();
1132 inst
= (ir_instruction
*) param
;
1133 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1135 /* data2 parameter (only with atomic_comp_swap) */
1136 if (param_count
== 3) {
1137 assert(op
== nir_intrinsic_shared_atomic_comp_swap
);
1138 param
= param
->get_next();
1139 inst
= (ir_instruction
*) param
;
1141 nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1145 assert(ir
->return_deref
);
1146 unsigned bit_size
= glsl_get_bit_size(ir
->return_deref
->type
);
1147 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1148 ir
->return_deref
->type
->vector_elements
,
1150 nir_builder_instr_insert(&b
, &instr
->instr
);
1153 case nir_intrinsic_vote_any
:
1154 case nir_intrinsic_vote_all
:
1155 case nir_intrinsic_vote_eq
: {
1156 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1158 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1159 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1161 nir_builder_instr_insert(&b
, &instr
->instr
);
1165 case nir_intrinsic_ballot
: {
1166 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1167 ir
->return_deref
->type
->vector_elements
, 64, NULL
);
1169 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1170 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1172 nir_builder_instr_insert(&b
, &instr
->instr
);
1175 case nir_intrinsic_read_invocation
: {
1176 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1177 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1178 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1180 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1181 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1183 ir_rvalue
*invocation
= (ir_rvalue
*) ir
->actual_parameters
.get_head()->next
;
1184 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(invocation
));
1186 nir_builder_instr_insert(&b
, &instr
->instr
);
1189 case nir_intrinsic_read_first_invocation
: {
1190 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1191 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1192 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1194 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1195 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1197 nir_builder_instr_insert(&b
, &instr
->instr
);
1201 unreachable("not reached");
1204 if (ir
->return_deref
) {
1205 nir_intrinsic_instr
*store_instr
=
1206 nir_intrinsic_instr_create(shader
, nir_intrinsic_store_var
);
1207 store_instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1208 nir_intrinsic_set_write_mask(store_instr
,
1209 (1 << store_instr
->num_components
) - 1);
1211 store_instr
->variables
[0] =
1212 evaluate_deref(&store_instr
->instr
, ir
->return_deref
);
1213 store_instr
->src
[0] = nir_src_for_ssa(&dest
->ssa
);
1215 nir_builder_instr_insert(&b
, &store_instr
->instr
);
1221 struct hash_entry
*entry
=
1222 _mesa_hash_table_search(this->overload_table
, ir
->callee
);
1224 nir_function
*callee
= (nir_function
*) entry
->data
;
1226 nir_call_instr
*instr
= nir_call_instr_create(this->shader
, callee
);
1229 foreach_in_list(ir_dereference
, param
, &ir
->actual_parameters
) {
1230 instr
->params
[i
] = evaluate_deref(&instr
->instr
, param
);
1234 instr
->return_deref
= evaluate_deref(&instr
->instr
, ir
->return_deref
);
1235 nir_builder_instr_insert(&b
, &instr
->instr
);
1239 nir_visitor::visit(ir_assignment
*ir
)
1241 unsigned num_components
= ir
->lhs
->type
->vector_elements
;
1243 b
.exact
= ir
->lhs
->variable_referenced()->data
.invariant
||
1244 ir
->lhs
->variable_referenced()->data
.precise
;
1246 if ((ir
->rhs
->as_dereference() || ir
->rhs
->as_constant()) &&
1247 (ir
->write_mask
== (1 << num_components
) - 1 || ir
->write_mask
== 0)) {
1248 /* We're doing a plain-as-can-be copy, so emit a copy_var */
1249 nir_intrinsic_instr
*copy
=
1250 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_copy_var
);
1252 copy
->variables
[0] = evaluate_deref(©
->instr
, ir
->lhs
);
1253 copy
->variables
[1] = evaluate_deref(©
->instr
, ir
->rhs
);
1255 if (ir
->condition
) {
1256 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1257 nir_builder_instr_insert(&b
, ©
->instr
);
1258 nir_pop_if(&b
, NULL
);
1260 nir_builder_instr_insert(&b
, ©
->instr
);
1265 assert(ir
->rhs
->type
->is_scalar() || ir
->rhs
->type
->is_vector());
1267 ir
->lhs
->accept(this);
1268 nir_deref_var
*lhs_deref
= this->deref_head
;
1269 nir_ssa_def
*src
= evaluate_rvalue(ir
->rhs
);
1271 if (ir
->write_mask
!= (1 << num_components
) - 1 && ir
->write_mask
!= 0) {
1272 /* GLSL IR will give us the input to the write-masked assignment in a
1273 * single packed vector. So, for example, if the writemask is xzw, then
1274 * we have to swizzle x -> x, y -> z, and z -> w and get the y component
1278 unsigned component
= 0;
1279 for (unsigned i
= 0; i
< 4; i
++) {
1280 swiz
[i
] = ir
->write_mask
& (1 << i
) ? component
++ : 0;
1282 src
= nir_swizzle(&b
, src
, swiz
, num_components
, !supports_ints
);
1285 nir_intrinsic_instr
*store
=
1286 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_store_var
);
1287 store
->num_components
= ir
->lhs
->type
->vector_elements
;
1288 nir_intrinsic_set_write_mask(store
, ir
->write_mask
);
1289 store
->variables
[0] = nir_deref_var_clone(lhs_deref
, store
);
1290 store
->src
[0] = nir_src_for_ssa(src
);
1292 if (ir
->condition
) {
1293 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1294 nir_builder_instr_insert(&b
, &store
->instr
);
1295 nir_pop_if(&b
, NULL
);
1297 nir_builder_instr_insert(&b
, &store
->instr
);
1302 * Given an instruction, returns a pointer to its destination or NULL if there
1303 * is no destination.
1305 * Note that this only handles instructions we generate at this level.
1308 get_instr_dest(nir_instr
*instr
)
1310 nir_alu_instr
*alu_instr
;
1311 nir_intrinsic_instr
*intrinsic_instr
;
1312 nir_tex_instr
*tex_instr
;
1314 switch (instr
->type
) {
1315 case nir_instr_type_alu
:
1316 alu_instr
= nir_instr_as_alu(instr
);
1317 return &alu_instr
->dest
.dest
;
1319 case nir_instr_type_intrinsic
:
1320 intrinsic_instr
= nir_instr_as_intrinsic(instr
);
1321 if (nir_intrinsic_infos
[intrinsic_instr
->intrinsic
].has_dest
)
1322 return &intrinsic_instr
->dest
;
1326 case nir_instr_type_tex
:
1327 tex_instr
= nir_instr_as_tex(instr
);
1328 return &tex_instr
->dest
;
1331 unreachable("not reached");
1338 nir_visitor::add_instr(nir_instr
*instr
, unsigned num_components
,
1341 nir_dest
*dest
= get_instr_dest(instr
);
1344 nir_ssa_dest_init(instr
, dest
, num_components
, bit_size
, NULL
);
1346 nir_builder_instr_insert(&b
, instr
);
1349 assert(dest
->is_ssa
);
1350 this->result
= &dest
->ssa
;
1355 nir_visitor::evaluate_rvalue(ir_rvalue
* ir
)
1358 if (ir
->as_dereference() || ir
->as_constant()) {
1360 * A dereference is being used on the right hand side, which means we
1361 * must emit a variable load.
1364 nir_intrinsic_instr
*load_instr
=
1365 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_var
);
1366 load_instr
->num_components
= ir
->type
->vector_elements
;
1367 load_instr
->variables
[0] = this->deref_head
;
1368 ralloc_steal(load_instr
, load_instr
->variables
[0]);
1369 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1370 add_instr(&load_instr
->instr
, ir
->type
->vector_elements
, bit_size
);
1373 return this->result
;
1377 type_is_float(glsl_base_type type
)
1379 return type
== GLSL_TYPE_FLOAT
|| type
== GLSL_TYPE_DOUBLE
;
1383 type_is_signed(glsl_base_type type
)
1385 return type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_INT64
;
1389 nir_visitor::visit(ir_expression
*ir
)
1391 /* Some special cases */
1392 switch (ir
->operation
) {
1393 case ir_binop_ubo_load
: {
1394 nir_intrinsic_instr
*load
=
1395 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_ubo
);
1396 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1397 load
->num_components
= ir
->type
->vector_elements
;
1398 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1399 load
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1400 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1403 * In UBO's, a true boolean value is any non-zero value, but we consider
1404 * a true boolean to be ~0. Fix this up with a != 0 comparison.
1407 if (ir
->type
->is_boolean())
1408 this->result
= nir_ine(&b
, &load
->dest
.ssa
, nir_imm_int(&b
, 0));
1413 case ir_unop_interpolate_at_centroid
:
1414 case ir_binop_interpolate_at_offset
:
1415 case ir_binop_interpolate_at_sample
: {
1416 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1417 ir_swizzle
*swizzle
= NULL
;
1419 /* the api does not allow a swizzle here, but the varying packing code
1420 * may have pushed one into here.
1422 swizzle
= ir
->operands
[0]->as_swizzle();
1424 deref
= swizzle
->val
->as_dereference();
1428 deref
->accept(this);
1430 nir_intrinsic_op op
;
1431 if (this->deref_head
->var
->data
.mode
== nir_var_shader_in
) {
1432 switch (ir
->operation
) {
1433 case ir_unop_interpolate_at_centroid
:
1434 op
= nir_intrinsic_interp_var_at_centroid
;
1436 case ir_binop_interpolate_at_offset
:
1437 op
= nir_intrinsic_interp_var_at_offset
;
1439 case ir_binop_interpolate_at_sample
:
1440 op
= nir_intrinsic_interp_var_at_sample
;
1443 unreachable("Invalid interpolation intrinsic");
1446 /* This case can happen if the vertex shader does not write the
1447 * given varying. In this case, the linker will lower it to a
1448 * global variable. Since interpolating a variable makes no
1449 * sense, we'll just turn it into a load which will probably
1450 * eventually end up as an SSA definition.
1452 assert(this->deref_head
->var
->data
.mode
== nir_var_global
);
1453 op
= nir_intrinsic_load_var
;
1456 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(shader
, op
);
1457 intrin
->num_components
= deref
->type
->vector_elements
;
1458 intrin
->variables
[0] = this->deref_head
;
1459 ralloc_steal(intrin
, intrin
->variables
[0]);
1461 if (intrin
->intrinsic
== nir_intrinsic_interp_var_at_offset
||
1462 intrin
->intrinsic
== nir_intrinsic_interp_var_at_sample
)
1463 intrin
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1465 unsigned bit_size
= glsl_get_bit_size(deref
->type
);
1466 add_instr(&intrin
->instr
, deref
->type
->vector_elements
, bit_size
);
1469 unsigned swiz
[4] = {
1470 swizzle
->mask
.x
, swizzle
->mask
.y
, swizzle
->mask
.z
, swizzle
->mask
.w
1473 result
= nir_swizzle(&b
, result
, swiz
,
1474 swizzle
->type
->vector_elements
, false);
1484 nir_ssa_def
*srcs
[4];
1485 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1486 srcs
[i
] = evaluate_rvalue(ir
->operands
[i
]);
1488 glsl_base_type types
[4];
1489 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1491 types
[i
] = ir
->operands
[i
]->type
->base_type
;
1493 types
[i
] = GLSL_TYPE_FLOAT
;
1495 glsl_base_type out_type
;
1497 out_type
= ir
->type
->base_type
;
1499 out_type
= GLSL_TYPE_FLOAT
;
1501 switch (ir
->operation
) {
1502 case ir_unop_bit_not
: result
= nir_inot(&b
, srcs
[0]); break;
1503 case ir_unop_logic_not
:
1504 result
= supports_ints
? nir_inot(&b
, srcs
[0]) : nir_fnot(&b
, srcs
[0]);
1507 result
= type_is_float(types
[0]) ? nir_fneg(&b
, srcs
[0])
1508 : nir_ineg(&b
, srcs
[0]);
1511 result
= type_is_float(types
[0]) ? nir_fabs(&b
, srcs
[0])
1512 : nir_iabs(&b
, srcs
[0]);
1514 case ir_unop_saturate
:
1515 assert(type_is_float(types
[0]));
1516 result
= nir_fsat(&b
, srcs
[0]);
1519 result
= type_is_float(types
[0]) ? nir_fsign(&b
, srcs
[0])
1520 : nir_isign(&b
, srcs
[0]);
1522 case ir_unop_rcp
: result
= nir_frcp(&b
, srcs
[0]); break;
1523 case ir_unop_rsq
: result
= nir_frsq(&b
, srcs
[0]); break;
1524 case ir_unop_sqrt
: result
= nir_fsqrt(&b
, srcs
[0]); break;
1525 case ir_unop_exp
: unreachable("ir_unop_exp should have been lowered");
1526 case ir_unop_log
: unreachable("ir_unop_log should have been lowered");
1527 case ir_unop_exp2
: result
= nir_fexp2(&b
, srcs
[0]); break;
1528 case ir_unop_log2
: result
= nir_flog2(&b
, srcs
[0]); break;
1530 result
= supports_ints
? nir_i2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1533 result
= supports_ints
? nir_u2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1536 result
= supports_ints
? nir_b2f(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1570 case ir_unop_i642u64
:
1571 case ir_unop_u642i64
: {
1572 nir_alu_type src_type
= nir_get_nir_type_for_glsl_base_type(types
[0]);
1573 nir_alu_type dst_type
= nir_get_nir_type_for_glsl_base_type(out_type
);
1574 result
= nir_build_alu(&b
, nir_type_conversion_op(src_type
, dst_type
),
1575 srcs
[0], NULL
, NULL
, NULL
);
1576 /* b2i and b2f don't have fixed bit-size versions so the builder will
1577 * just assume 32 and we have to fix it up here.
1579 result
->bit_size
= nir_alu_type_get_type_size(dst_type
);
1583 case ir_unop_bitcast_i2f
:
1584 case ir_unop_bitcast_f2i
:
1585 case ir_unop_bitcast_u2f
:
1586 case ir_unop_bitcast_f2u
:
1587 case ir_unop_bitcast_i642d
:
1588 case ir_unop_bitcast_d2i64
:
1589 case ir_unop_bitcast_u642d
:
1590 case ir_unop_bitcast_d2u64
:
1591 case ir_unop_subroutine_to_int
:
1593 result
= nir_imov(&b
, srcs
[0]);
1595 case ir_unop_trunc
: result
= nir_ftrunc(&b
, srcs
[0]); break;
1596 case ir_unop_ceil
: result
= nir_fceil(&b
, srcs
[0]); break;
1597 case ir_unop_floor
: result
= nir_ffloor(&b
, srcs
[0]); break;
1598 case ir_unop_fract
: result
= nir_ffract(&b
, srcs
[0]); break;
1599 case ir_unop_round_even
: result
= nir_fround_even(&b
, srcs
[0]); break;
1600 case ir_unop_sin
: result
= nir_fsin(&b
, srcs
[0]); break;
1601 case ir_unop_cos
: result
= nir_fcos(&b
, srcs
[0]); break;
1602 case ir_unop_dFdx
: result
= nir_fddx(&b
, srcs
[0]); break;
1603 case ir_unop_dFdy
: result
= nir_fddy(&b
, srcs
[0]); break;
1604 case ir_unop_dFdx_fine
: result
= nir_fddx_fine(&b
, srcs
[0]); break;
1605 case ir_unop_dFdy_fine
: result
= nir_fddy_fine(&b
, srcs
[0]); break;
1606 case ir_unop_dFdx_coarse
: result
= nir_fddx_coarse(&b
, srcs
[0]); break;
1607 case ir_unop_dFdy_coarse
: result
= nir_fddy_coarse(&b
, srcs
[0]); break;
1608 case ir_unop_pack_snorm_2x16
:
1609 result
= nir_pack_snorm_2x16(&b
, srcs
[0]);
1611 case ir_unop_pack_snorm_4x8
:
1612 result
= nir_pack_snorm_4x8(&b
, srcs
[0]);
1614 case ir_unop_pack_unorm_2x16
:
1615 result
= nir_pack_unorm_2x16(&b
, srcs
[0]);
1617 case ir_unop_pack_unorm_4x8
:
1618 result
= nir_pack_unorm_4x8(&b
, srcs
[0]);
1620 case ir_unop_pack_half_2x16
:
1621 result
= nir_pack_half_2x16(&b
, srcs
[0]);
1623 case ir_unop_unpack_snorm_2x16
:
1624 result
= nir_unpack_snorm_2x16(&b
, srcs
[0]);
1626 case ir_unop_unpack_snorm_4x8
:
1627 result
= nir_unpack_snorm_4x8(&b
, srcs
[0]);
1629 case ir_unop_unpack_unorm_2x16
:
1630 result
= nir_unpack_unorm_2x16(&b
, srcs
[0]);
1632 case ir_unop_unpack_unorm_4x8
:
1633 result
= nir_unpack_unorm_4x8(&b
, srcs
[0]);
1635 case ir_unop_unpack_half_2x16
:
1636 result
= nir_unpack_half_2x16(&b
, srcs
[0]);
1638 case ir_unop_pack_double_2x32
:
1639 case ir_unop_pack_int_2x32
:
1640 case ir_unop_pack_uint_2x32
:
1641 result
= nir_pack_64_2x32(&b
, srcs
[0]);
1643 case ir_unop_unpack_double_2x32
:
1644 case ir_unop_unpack_int_2x32
:
1645 case ir_unop_unpack_uint_2x32
:
1646 result
= nir_unpack_64_2x32(&b
, srcs
[0]);
1648 case ir_unop_bitfield_reverse
:
1649 result
= nir_bitfield_reverse(&b
, srcs
[0]);
1651 case ir_unop_bit_count
:
1652 result
= nir_bit_count(&b
, srcs
[0]);
1654 case ir_unop_find_msb
:
1656 case GLSL_TYPE_UINT
:
1657 result
= nir_ufind_msb(&b
, srcs
[0]);
1660 result
= nir_ifind_msb(&b
, srcs
[0]);
1663 unreachable("Invalid type for findMSB()");
1666 case ir_unop_find_lsb
:
1667 result
= nir_find_lsb(&b
, srcs
[0]);
1671 switch (ir
->type
->vector_elements
) {
1673 switch (ir
->operands
[0]->type
->vector_elements
) {
1674 case 1: result
= nir_fnoise1_1(&b
, srcs
[0]); break;
1675 case 2: result
= nir_fnoise1_2(&b
, srcs
[0]); break;
1676 case 3: result
= nir_fnoise1_3(&b
, srcs
[0]); break;
1677 case 4: result
= nir_fnoise1_4(&b
, srcs
[0]); break;
1678 default: unreachable("not reached");
1682 switch (ir
->operands
[0]->type
->vector_elements
) {
1683 case 1: result
= nir_fnoise2_1(&b
, srcs
[0]); break;
1684 case 2: result
= nir_fnoise2_2(&b
, srcs
[0]); break;
1685 case 3: result
= nir_fnoise2_3(&b
, srcs
[0]); break;
1686 case 4: result
= nir_fnoise2_4(&b
, srcs
[0]); break;
1687 default: unreachable("not reached");
1691 switch (ir
->operands
[0]->type
->vector_elements
) {
1692 case 1: result
= nir_fnoise3_1(&b
, srcs
[0]); break;
1693 case 2: result
= nir_fnoise3_2(&b
, srcs
[0]); break;
1694 case 3: result
= nir_fnoise3_3(&b
, srcs
[0]); break;
1695 case 4: result
= nir_fnoise3_4(&b
, srcs
[0]); break;
1696 default: unreachable("not reached");
1700 switch (ir
->operands
[0]->type
->vector_elements
) {
1701 case 1: result
= nir_fnoise4_1(&b
, srcs
[0]); break;
1702 case 2: result
= nir_fnoise4_2(&b
, srcs
[0]); break;
1703 case 3: result
= nir_fnoise4_3(&b
, srcs
[0]); break;
1704 case 4: result
= nir_fnoise4_4(&b
, srcs
[0]); break;
1705 default: unreachable("not reached");
1709 unreachable("not reached");
1712 case ir_unop_get_buffer_size
: {
1713 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(
1715 nir_intrinsic_get_buffer_size
);
1716 load
->num_components
= ir
->type
->vector_elements
;
1717 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1718 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1719 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1724 result
= type_is_float(out_type
) ? nir_fadd(&b
, srcs
[0], srcs
[1])
1725 : nir_iadd(&b
, srcs
[0], srcs
[1]);
1728 result
= type_is_float(out_type
) ? nir_fsub(&b
, srcs
[0], srcs
[1])
1729 : nir_isub(&b
, srcs
[0], srcs
[1]);
1732 result
= type_is_float(out_type
) ? nir_fmul(&b
, srcs
[0], srcs
[1])
1733 : nir_imul(&b
, srcs
[0], srcs
[1]);
1736 if (type_is_float(out_type
))
1737 result
= nir_fdiv(&b
, srcs
[0], srcs
[1]);
1738 else if (type_is_signed(out_type
))
1739 result
= nir_idiv(&b
, srcs
[0], srcs
[1]);
1741 result
= nir_udiv(&b
, srcs
[0], srcs
[1]);
1744 result
= type_is_float(out_type
) ? nir_fmod(&b
, srcs
[0], srcs
[1])
1745 : nir_umod(&b
, srcs
[0], srcs
[1]);
1748 if (type_is_float(out_type
))
1749 result
= nir_fmin(&b
, srcs
[0], srcs
[1]);
1750 else if (type_is_signed(out_type
))
1751 result
= nir_imin(&b
, srcs
[0], srcs
[1]);
1753 result
= nir_umin(&b
, srcs
[0], srcs
[1]);
1756 if (type_is_float(out_type
))
1757 result
= nir_fmax(&b
, srcs
[0], srcs
[1]);
1758 else if (type_is_signed(out_type
))
1759 result
= nir_imax(&b
, srcs
[0], srcs
[1]);
1761 result
= nir_umax(&b
, srcs
[0], srcs
[1]);
1763 case ir_binop_pow
: result
= nir_fpow(&b
, srcs
[0], srcs
[1]); break;
1764 case ir_binop_bit_and
: result
= nir_iand(&b
, srcs
[0], srcs
[1]); break;
1765 case ir_binop_bit_or
: result
= nir_ior(&b
, srcs
[0], srcs
[1]); break;
1766 case ir_binop_bit_xor
: result
= nir_ixor(&b
, srcs
[0], srcs
[1]); break;
1767 case ir_binop_logic_and
:
1768 result
= supports_ints
? nir_iand(&b
, srcs
[0], srcs
[1])
1769 : nir_fand(&b
, srcs
[0], srcs
[1]);
1771 case ir_binop_logic_or
:
1772 result
= supports_ints
? nir_ior(&b
, srcs
[0], srcs
[1])
1773 : nir_for(&b
, srcs
[0], srcs
[1]);
1775 case ir_binop_logic_xor
:
1776 result
= supports_ints
? nir_ixor(&b
, srcs
[0], srcs
[1])
1777 : nir_fxor(&b
, srcs
[0], srcs
[1]);
1779 case ir_binop_lshift
: result
= nir_ishl(&b
, srcs
[0], srcs
[1]); break;
1780 case ir_binop_rshift
:
1781 result
= (type_is_signed(out_type
)) ? nir_ishr(&b
, srcs
[0], srcs
[1])
1782 : nir_ushr(&b
, srcs
[0], srcs
[1]);
1784 case ir_binop_imul_high
:
1785 result
= (out_type
== GLSL_TYPE_INT
) ? nir_imul_high(&b
, srcs
[0], srcs
[1])
1786 : nir_umul_high(&b
, srcs
[0], srcs
[1]);
1788 case ir_binop_carry
: result
= nir_uadd_carry(&b
, srcs
[0], srcs
[1]); break;
1789 case ir_binop_borrow
: result
= nir_usub_borrow(&b
, srcs
[0], srcs
[1]); break;
1791 if (supports_ints
) {
1792 if (type_is_float(types
[0]))
1793 result
= nir_flt(&b
, srcs
[0], srcs
[1]);
1794 else if (type_is_signed(types
[0]))
1795 result
= nir_ilt(&b
, srcs
[0], srcs
[1]);
1797 result
= nir_ult(&b
, srcs
[0], srcs
[1]);
1799 result
= nir_slt(&b
, srcs
[0], srcs
[1]);
1802 case ir_binop_greater
:
1803 if (supports_ints
) {
1804 if (type_is_float(types
[0]))
1805 result
= nir_flt(&b
, srcs
[1], srcs
[0]);
1806 else if (type_is_signed(types
[0]))
1807 result
= nir_ilt(&b
, srcs
[1], srcs
[0]);
1809 result
= nir_ult(&b
, srcs
[1], srcs
[0]);
1811 result
= nir_slt(&b
, srcs
[1], srcs
[0]);
1814 case ir_binop_lequal
:
1815 if (supports_ints
) {
1816 if (type_is_float(types
[0]))
1817 result
= nir_fge(&b
, srcs
[1], srcs
[0]);
1818 else if (type_is_signed(types
[0]))
1819 result
= nir_ige(&b
, srcs
[1], srcs
[0]);
1821 result
= nir_uge(&b
, srcs
[1], srcs
[0]);
1823 result
= nir_slt(&b
, srcs
[1], srcs
[0]);
1826 case ir_binop_gequal
:
1827 if (supports_ints
) {
1828 if (type_is_float(types
[0]))
1829 result
= nir_fge(&b
, srcs
[0], srcs
[1]);
1830 else if (type_is_signed(types
[0]))
1831 result
= nir_ige(&b
, srcs
[0], srcs
[1]);
1833 result
= nir_uge(&b
, srcs
[0], srcs
[1]);
1835 result
= nir_slt(&b
, srcs
[0], srcs
[1]);
1838 case ir_binop_equal
:
1839 if (supports_ints
) {
1840 if (type_is_float(types
[0]))
1841 result
= nir_feq(&b
, srcs
[0], srcs
[1]);
1843 result
= nir_ieq(&b
, srcs
[0], srcs
[1]);
1845 result
= nir_seq(&b
, srcs
[0], srcs
[1]);
1848 case ir_binop_nequal
:
1849 if (supports_ints
) {
1850 if (type_is_float(types
[0]))
1851 result
= nir_fne(&b
, srcs
[0], srcs
[1]);
1853 result
= nir_ine(&b
, srcs
[0], srcs
[1]);
1855 result
= nir_sne(&b
, srcs
[0], srcs
[1]);
1858 case ir_binop_all_equal
:
1859 if (supports_ints
) {
1860 if (type_is_float(types
[0])) {
1861 switch (ir
->operands
[0]->type
->vector_elements
) {
1862 case 1: result
= nir_feq(&b
, srcs
[0], srcs
[1]); break;
1863 case 2: result
= nir_ball_fequal2(&b
, srcs
[0], srcs
[1]); break;
1864 case 3: result
= nir_ball_fequal3(&b
, srcs
[0], srcs
[1]); break;
1865 case 4: result
= nir_ball_fequal4(&b
, srcs
[0], srcs
[1]); break;
1867 unreachable("not reached");
1870 switch (ir
->operands
[0]->type
->vector_elements
) {
1871 case 1: result
= nir_ieq(&b
, srcs
[0], srcs
[1]); break;
1872 case 2: result
= nir_ball_iequal2(&b
, srcs
[0], srcs
[1]); break;
1873 case 3: result
= nir_ball_iequal3(&b
, srcs
[0], srcs
[1]); break;
1874 case 4: result
= nir_ball_iequal4(&b
, srcs
[0], srcs
[1]); break;
1876 unreachable("not reached");
1880 switch (ir
->operands
[0]->type
->vector_elements
) {
1881 case 1: result
= nir_seq(&b
, srcs
[0], srcs
[1]); break;
1882 case 2: result
= nir_fall_equal2(&b
, srcs
[0], srcs
[1]); break;
1883 case 3: result
= nir_fall_equal3(&b
, srcs
[0], srcs
[1]); break;
1884 case 4: result
= nir_fall_equal4(&b
, srcs
[0], srcs
[1]); break;
1886 unreachable("not reached");
1890 case ir_binop_any_nequal
:
1891 if (supports_ints
) {
1892 if (type_is_float(types
[0])) {
1893 switch (ir
->operands
[0]->type
->vector_elements
) {
1894 case 1: result
= nir_fne(&b
, srcs
[0], srcs
[1]); break;
1895 case 2: result
= nir_bany_fnequal2(&b
, srcs
[0], srcs
[1]); break;
1896 case 3: result
= nir_bany_fnequal3(&b
, srcs
[0], srcs
[1]); break;
1897 case 4: result
= nir_bany_fnequal4(&b
, srcs
[0], srcs
[1]); break;
1899 unreachable("not reached");
1902 switch (ir
->operands
[0]->type
->vector_elements
) {
1903 case 1: result
= nir_ine(&b
, srcs
[0], srcs
[1]); break;
1904 case 2: result
= nir_bany_inequal2(&b
, srcs
[0], srcs
[1]); break;
1905 case 3: result
= nir_bany_inequal3(&b
, srcs
[0], srcs
[1]); break;
1906 case 4: result
= nir_bany_inequal4(&b
, srcs
[0], srcs
[1]); break;
1908 unreachable("not reached");
1912 switch (ir
->operands
[0]->type
->vector_elements
) {
1913 case 1: result
= nir_sne(&b
, srcs
[0], srcs
[1]); break;
1914 case 2: result
= nir_fany_nequal2(&b
, srcs
[0], srcs
[1]); break;
1915 case 3: result
= nir_fany_nequal3(&b
, srcs
[0], srcs
[1]); break;
1916 case 4: result
= nir_fany_nequal4(&b
, srcs
[0], srcs
[1]); break;
1918 unreachable("not reached");
1923 switch (ir
->operands
[0]->type
->vector_elements
) {
1924 case 2: result
= nir_fdot2(&b
, srcs
[0], srcs
[1]); break;
1925 case 3: result
= nir_fdot3(&b
, srcs
[0], srcs
[1]); break;
1926 case 4: result
= nir_fdot4(&b
, srcs
[0], srcs
[1]); break;
1928 unreachable("not reached");
1932 case ir_binop_ldexp
: result
= nir_ldexp(&b
, srcs
[0], srcs
[1]); break;
1934 result
= nir_ffma(&b
, srcs
[0], srcs
[1], srcs
[2]);
1937 result
= nir_flrp(&b
, srcs
[0], srcs
[1], srcs
[2]);
1941 result
= nir_bcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
1943 result
= nir_fcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
1945 case ir_triop_bitfield_extract
:
1946 result
= (out_type
== GLSL_TYPE_INT
) ?
1947 nir_ibitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]) :
1948 nir_ubitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]);
1950 case ir_quadop_bitfield_insert
:
1951 result
= nir_bitfield_insert(&b
, srcs
[0], srcs
[1], srcs
[2], srcs
[3]);
1953 case ir_quadop_vector
:
1954 result
= nir_vec(&b
, srcs
, ir
->type
->vector_elements
);
1958 unreachable("not reached");
1963 nir_visitor::visit(ir_swizzle
*ir
)
1965 unsigned swizzle
[4] = { ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
};
1966 result
= nir_swizzle(&b
, evaluate_rvalue(ir
->val
), swizzle
,
1967 ir
->type
->vector_elements
, !supports_ints
);
1971 nir_visitor::visit(ir_texture
*ir
)
1978 num_srcs
= 1; /* coordinate */
1983 op
= (ir
->op
== ir_txb
) ? nir_texop_txb
: nir_texop_txl
;
1984 num_srcs
= 2; /* coordinate, bias/lod */
1988 op
= nir_texop_txd
; /* coordinate, dPdx, dPdy */
1994 if (ir
->lod_info
.lod
!= NULL
)
1995 num_srcs
= 2; /* coordinate, lod */
1997 num_srcs
= 1; /* coordinate */
2001 op
= nir_texop_txf_ms
;
2002 num_srcs
= 2; /* coordinate, sample_index */
2007 if (ir
->lod_info
.lod
!= NULL
)
2008 num_srcs
= 1; /* lod */
2015 num_srcs
= 1; /* coordinate */
2020 num_srcs
= 1; /* coordinate */
2023 case ir_query_levels
:
2024 op
= nir_texop_query_levels
;
2028 case ir_texture_samples
:
2029 op
= nir_texop_texture_samples
;
2033 case ir_samples_identical
:
2034 op
= nir_texop_samples_identical
;
2035 num_srcs
= 1; /* coordinate */
2039 unreachable("not reached");
2042 if (ir
->projector
!= NULL
)
2044 if (ir
->shadow_comparator
!= NULL
)
2046 if (ir
->offset
!= NULL
)
2049 nir_tex_instr
*instr
= nir_tex_instr_create(this->shader
, num_srcs
);
2052 instr
->sampler_dim
=
2053 (glsl_sampler_dim
) ir
->sampler
->type
->sampler_dimensionality
;
2054 instr
->is_array
= ir
->sampler
->type
->sampler_array
;
2055 instr
->is_shadow
= ir
->sampler
->type
->sampler_shadow
;
2056 if (instr
->is_shadow
)
2057 instr
->is_new_style_shadow
= (ir
->type
->vector_elements
== 1);
2058 switch (ir
->type
->base_type
) {
2059 case GLSL_TYPE_FLOAT
:
2060 instr
->dest_type
= nir_type_float
;
2063 instr
->dest_type
= nir_type_int
;
2065 case GLSL_TYPE_BOOL
:
2066 case GLSL_TYPE_UINT
:
2067 instr
->dest_type
= nir_type_uint
;
2070 unreachable("not reached");
2073 instr
->texture
= evaluate_deref(&instr
->instr
, ir
->sampler
);
2075 unsigned src_number
= 0;
2077 if (ir
->coordinate
!= NULL
) {
2078 instr
->coord_components
= ir
->coordinate
->type
->vector_elements
;
2079 instr
->src
[src_number
].src
=
2080 nir_src_for_ssa(evaluate_rvalue(ir
->coordinate
));
2081 instr
->src
[src_number
].src_type
= nir_tex_src_coord
;
2085 if (ir
->projector
!= NULL
) {
2086 instr
->src
[src_number
].src
=
2087 nir_src_for_ssa(evaluate_rvalue(ir
->projector
));
2088 instr
->src
[src_number
].src_type
= nir_tex_src_projector
;
2092 if (ir
->shadow_comparator
!= NULL
) {
2093 instr
->src
[src_number
].src
=
2094 nir_src_for_ssa(evaluate_rvalue(ir
->shadow_comparator
));
2095 instr
->src
[src_number
].src_type
= nir_tex_src_comparator
;
2099 if (ir
->offset
!= NULL
) {
2100 /* we don't support multiple offsets yet */
2101 assert(ir
->offset
->type
->is_vector() || ir
->offset
->type
->is_scalar());
2103 instr
->src
[src_number
].src
=
2104 nir_src_for_ssa(evaluate_rvalue(ir
->offset
));
2105 instr
->src
[src_number
].src_type
= nir_tex_src_offset
;
2111 instr
->src
[src_number
].src
=
2112 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.bias
));
2113 instr
->src
[src_number
].src_type
= nir_tex_src_bias
;
2120 if (ir
->lod_info
.lod
!= NULL
) {
2121 instr
->src
[src_number
].src
=
2122 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.lod
));
2123 instr
->src
[src_number
].src_type
= nir_tex_src_lod
;
2129 instr
->src
[src_number
].src
=
2130 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdx
));
2131 instr
->src
[src_number
].src_type
= nir_tex_src_ddx
;
2133 instr
->src
[src_number
].src
=
2134 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdy
));
2135 instr
->src
[src_number
].src_type
= nir_tex_src_ddy
;
2140 instr
->src
[src_number
].src
=
2141 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.sample_index
));
2142 instr
->src
[src_number
].src_type
= nir_tex_src_ms_index
;
2147 instr
->component
= ir
->lod_info
.component
->as_constant()->value
.u
[0];
2154 assert(src_number
== num_srcs
);
2156 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2157 add_instr(&instr
->instr
, nir_tex_instr_dest_size(instr
), bit_size
);
2161 nir_visitor::visit(ir_constant
*ir
)
2164 * We don't know if this variable is an array or struct that gets
2165 * dereferenced, so do the safe thing an make it a variable with a
2166 * constant initializer and return a dereference.
2170 nir_local_variable_create(this->impl
, ir
->type
, "const_temp");
2171 var
->data
.read_only
= true;
2172 var
->constant_initializer
= constant_copy(ir
, var
);
2174 this->deref_head
= nir_deref_var_create(this->shader
, var
);
2175 this->deref_tail
= &this->deref_head
->deref
;
2179 nir_visitor::visit(ir_dereference_variable
*ir
)
2181 struct hash_entry
*entry
=
2182 _mesa_hash_table_search(this->var_table
, ir
->var
);
2184 nir_variable
*var
= (nir_variable
*) entry
->data
;
2186 nir_deref_var
*deref
= nir_deref_var_create(this->shader
, var
);
2187 this->deref_head
= deref
;
2188 this->deref_tail
= &deref
->deref
;
2192 nir_visitor::visit(ir_dereference_record
*ir
)
2194 ir
->record
->accept(this);
2196 int field_index
= ir
->field_idx
;
2197 assert(field_index
>= 0);
2199 nir_deref_struct
*deref
= nir_deref_struct_create(this->deref_tail
, field_index
);
2200 deref
->deref
.type
= ir
->type
;
2201 this->deref_tail
->child
= &deref
->deref
;
2202 this->deref_tail
= &deref
->deref
;
2206 nir_visitor::visit(ir_dereference_array
*ir
)
2208 nir_deref_array
*deref
= nir_deref_array_create(this->shader
);
2209 deref
->deref
.type
= ir
->type
;
2211 ir_constant
*const_index
= ir
->array_index
->as_constant();
2212 if (const_index
!= NULL
) {
2213 deref
->deref_array_type
= nir_deref_array_type_direct
;
2214 deref
->base_offset
= const_index
->value
.u
[0];
2216 deref
->deref_array_type
= nir_deref_array_type_indirect
;
2218 nir_src_for_ssa(evaluate_rvalue(ir
->array_index
));
2221 ir
->array
->accept(this);
2223 this->deref_tail
->child
= &deref
->deref
;
2224 ralloc_steal(this->deref_tail
, deref
);
2225 this->deref_tail
= &deref
->deref
;
2229 nir_visitor::visit(ir_barrier
*)
2231 nir_intrinsic_instr
*instr
=
2232 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_barrier
);
2233 nir_builder_instr_insert(&b
, &instr
->instr
);