2 * Copyright © 2014 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Connor Abbott (cwabbott0@gmail.com)
28 #include "glsl_to_nir.h"
29 #include "ir_visitor.h"
30 #include "ir_hierarchical_visitor.h"
32 #include "compiler/nir/nir_control_flow.h"
33 #include "compiler/nir/nir_builder.h"
34 #include "main/imports.h"
35 #include "main/mtypes.h"
38 * pass to lower GLSL IR to NIR
40 * This will lower variable dereferences to loads/stores of corresponding
41 * variables in NIR - the variables will be converted to registers in a later
47 class nir_visitor
: public ir_visitor
50 nir_visitor(nir_shader
*shader
);
53 virtual void visit(ir_variable
*);
54 virtual void visit(ir_function
*);
55 virtual void visit(ir_function_signature
*);
56 virtual void visit(ir_loop
*);
57 virtual void visit(ir_if
*);
58 virtual void visit(ir_discard
*);
59 virtual void visit(ir_loop_jump
*);
60 virtual void visit(ir_return
*);
61 virtual void visit(ir_call
*);
62 virtual void visit(ir_assignment
*);
63 virtual void visit(ir_emit_vertex
*);
64 virtual void visit(ir_end_primitive
*);
65 virtual void visit(ir_expression
*);
66 virtual void visit(ir_swizzle
*);
67 virtual void visit(ir_texture
*);
68 virtual void visit(ir_constant
*);
69 virtual void visit(ir_dereference_variable
*);
70 virtual void visit(ir_dereference_record
*);
71 virtual void visit(ir_dereference_array
*);
72 virtual void visit(ir_barrier
*);
74 void create_function(ir_function_signature
*ir
);
77 void add_instr(nir_instr
*instr
, unsigned num_components
, unsigned bit_size
);
78 nir_ssa_def
*evaluate_rvalue(ir_rvalue
*ir
);
80 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
**srcs
);
81 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
);
82 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
,
84 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
,
85 nir_ssa_def
*src2
, nir_ssa_def
*src3
);
90 nir_function_impl
*impl
;
92 nir_ssa_def
*result
; /* result of the expression tree last visited */
94 nir_deref_var
*evaluate_deref(nir_instr
*mem_ctx
, ir_instruction
*ir
);
96 /* the head of the dereference chain we're creating */
97 nir_deref_var
*deref_head
;
98 /* the tail of the dereference chain we're creating */
99 nir_deref
*deref_tail
;
101 nir_variable
*var
; /* variable created by ir_variable visitor */
103 /* whether the IR we're operating on is per-function or global */
106 /* map of ir_variable -> nir_variable */
107 struct hash_table
*var_table
;
109 /* map of ir_function_signature -> nir_function_overload */
110 struct hash_table
*overload_table
;
114 * This visitor runs before the main visitor, calling create_function() for
115 * each function so that the main visitor can resolve forward references in
119 class nir_function_visitor
: public ir_hierarchical_visitor
122 nir_function_visitor(nir_visitor
*v
) : visitor(v
)
125 virtual ir_visitor_status
visit_enter(ir_function
*);
128 nir_visitor
*visitor
;
131 } /* end of anonymous namespace */
134 nir_remap_attributes(nir_shader
*shader
,
135 const nir_shader_compiler_options
*options
)
137 if (options
->vs_inputs_dual_locations
) {
138 nir_foreach_variable(var
, &shader
->inputs
) {
139 var
->data
.location
+=
140 _mesa_bitcount_64(shader
->info
.vs
.double_inputs
&
141 BITFIELD64_MASK(var
->data
.location
));
145 /* Once the remap is done, reset double_inputs_read, so later it will have
146 * which location/slots are doubles */
147 shader
->info
.vs
.double_inputs
= 0;
151 glsl_to_nir(const struct gl_shader_program
*shader_prog
,
152 gl_shader_stage stage
,
153 const nir_shader_compiler_options
*options
)
155 struct gl_linked_shader
*sh
= shader_prog
->_LinkedShaders
[stage
];
157 nir_shader
*shader
= nir_shader_create(NULL
, stage
, options
,
160 nir_visitor
v1(shader
);
161 nir_function_visitor
v2(&v1
);
163 visit_exec_list(sh
->ir
, &v1
);
165 nir_lower_constant_initializers(shader
, (nir_variable_mode
)~0);
167 /* Remap the locations to slots so those requiring two slots will occupy
168 * two locations. For instance, if we have in the IR code a dvec3 attr0 in
169 * location 0 and vec4 attr1 in location 1, in NIR attr0 will use
170 * locations/slots 0 and 1, and attr1 will use location/slot 2 */
171 if (shader
->info
.stage
== MESA_SHADER_VERTEX
)
172 nir_remap_attributes(shader
, options
);
174 shader
->info
.name
= ralloc_asprintf(shader
, "GLSL%d", shader_prog
->Name
);
175 if (shader_prog
->Label
)
176 shader
->info
.label
= ralloc_strdup(shader
, shader_prog
->Label
);
178 /* Check for transform feedback varyings specified via the API */
179 shader
->info
.has_transform_feedback_varyings
=
180 shader_prog
->TransformFeedback
.NumVarying
> 0;
182 /* Check for transform feedback varyings specified in the Shader */
183 if (shader_prog
->last_vert_prog
)
184 shader
->info
.has_transform_feedback_varyings
|=
185 shader_prog
->last_vert_prog
->sh
.LinkedTransformFeedback
->NumVarying
> 0;
190 nir_visitor::nir_visitor(nir_shader
*shader
)
192 this->supports_ints
= shader
->options
->native_integers
;
193 this->shader
= shader
;
194 this->is_global
= true;
195 this->var_table
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
196 _mesa_key_pointer_equal
);
197 this->overload_table
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
198 _mesa_key_pointer_equal
);
202 this->deref_head
= NULL
;
203 this->deref_tail
= NULL
;
204 memset(&this->b
, 0, sizeof(this->b
));
207 nir_visitor::~nir_visitor()
209 _mesa_hash_table_destroy(this->var_table
, NULL
);
210 _mesa_hash_table_destroy(this->overload_table
, NULL
);
214 nir_visitor::evaluate_deref(nir_instr
*mem_ctx
, ir_instruction
*ir
)
217 ralloc_steal(mem_ctx
, this->deref_head
);
218 return this->deref_head
;
221 static nir_constant
*
222 constant_copy(ir_constant
*ir
, void *mem_ctx
)
227 nir_constant
*ret
= rzalloc(mem_ctx
, nir_constant
);
229 const unsigned rows
= ir
->type
->vector_elements
;
230 const unsigned cols
= ir
->type
->matrix_columns
;
233 ret
->num_elements
= 0;
234 switch (ir
->type
->base_type
) {
236 /* Only float base types can be matrices. */
239 for (unsigned r
= 0; r
< rows
; r
++)
240 ret
->values
[0].u32
[r
] = ir
->value
.u
[r
];
245 /* Only float base types can be matrices. */
248 for (unsigned r
= 0; r
< rows
; r
++)
249 ret
->values
[0].i32
[r
] = ir
->value
.i
[r
];
253 case GLSL_TYPE_FLOAT
:
254 for (unsigned c
= 0; c
< cols
; c
++) {
255 for (unsigned r
= 0; r
< rows
; r
++)
256 ret
->values
[c
].f32
[r
] = ir
->value
.f
[c
* rows
+ r
];
260 case GLSL_TYPE_DOUBLE
:
261 for (unsigned c
= 0; c
< cols
; c
++) {
262 for (unsigned r
= 0; r
< rows
; r
++)
263 ret
->values
[c
].f64
[r
] = ir
->value
.d
[c
* rows
+ r
];
267 case GLSL_TYPE_UINT64
:
268 /* Only float base types can be matrices. */
271 for (unsigned r
= 0; r
< rows
; r
++)
272 ret
->values
[0].u64
[r
] = ir
->value
.u64
[r
];
275 case GLSL_TYPE_INT64
:
276 /* Only float base types can be matrices. */
279 for (unsigned r
= 0; r
< rows
; r
++)
280 ret
->values
[0].i64
[r
] = ir
->value
.i64
[r
];
284 /* Only float base types can be matrices. */
287 for (unsigned r
= 0; r
< rows
; r
++)
288 ret
->values
[0].u32
[r
] = ir
->value
.b
[r
] ? NIR_TRUE
: NIR_FALSE
;
292 case GLSL_TYPE_STRUCT
:
293 case GLSL_TYPE_ARRAY
:
294 ret
->elements
= ralloc_array(mem_ctx
, nir_constant
*,
296 ret
->num_elements
= ir
->type
->length
;
298 for (i
= 0; i
< ir
->type
->length
; i
++)
299 ret
->elements
[i
] = constant_copy(ir
->const_elements
[i
], mem_ctx
);
303 unreachable("not reached");
310 nir_visitor::visit(ir_variable
*ir
)
312 /* TODO: In future we should switch to using the NIR lowering pass but for
313 * now just ignore these variables as GLSL IR should have lowered them.
314 * Anything remaining are just dead vars that weren't cleaned up.
316 if (ir
->data
.mode
== ir_var_shader_shared
)
319 nir_variable
*var
= rzalloc(shader
, nir_variable
);
320 var
->type
= ir
->type
;
321 var
->name
= ralloc_strdup(var
, ir
->name
);
323 var
->data
.always_active_io
= ir
->data
.always_active_io
;
324 var
->data
.read_only
= ir
->data
.read_only
;
325 var
->data
.centroid
= ir
->data
.centroid
;
326 var
->data
.sample
= ir
->data
.sample
;
327 var
->data
.patch
= ir
->data
.patch
;
328 var
->data
.invariant
= ir
->data
.invariant
;
329 var
->data
.location
= ir
->data
.location
;
330 var
->data
.stream
= ir
->data
.stream
;
331 var
->data
.compact
= false;
333 switch(ir
->data
.mode
) {
335 case ir_var_temporary
:
337 var
->data
.mode
= nir_var_global
;
339 var
->data
.mode
= nir_var_local
;
342 case ir_var_function_in
:
343 case ir_var_function_out
:
344 case ir_var_function_inout
:
345 case ir_var_const_in
:
346 var
->data
.mode
= nir_var_local
;
349 case ir_var_shader_in
:
350 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
351 ir
->data
.location
== VARYING_SLOT_FACE
) {
352 /* For whatever reason, GLSL IR makes gl_FrontFacing an input */
353 var
->data
.location
= SYSTEM_VALUE_FRONT_FACE
;
354 var
->data
.mode
= nir_var_system_value
;
355 } else if (shader
->info
.stage
== MESA_SHADER_GEOMETRY
&&
356 ir
->data
.location
== VARYING_SLOT_PRIMITIVE_ID
) {
357 /* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */
358 var
->data
.location
= SYSTEM_VALUE_PRIMITIVE_ID
;
359 var
->data
.mode
= nir_var_system_value
;
361 var
->data
.mode
= nir_var_shader_in
;
363 if (shader
->info
.stage
== MESA_SHADER_TESS_EVAL
&&
364 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
365 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
366 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
370 /* Mark all the locations that require two slots */
371 if (shader
->info
.stage
== MESA_SHADER_VERTEX
&&
372 glsl_type_is_dual_slot(glsl_without_array(var
->type
))) {
373 for (unsigned i
= 0; i
< glsl_count_attribute_slots(var
->type
, true); i
++) {
374 uint64_t bitfield
= BITFIELD64_BIT(var
->data
.location
+ i
);
375 shader
->info
.vs
.double_inputs
|= bitfield
;
380 case ir_var_shader_out
:
381 var
->data
.mode
= nir_var_shader_out
;
382 if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
&&
383 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
384 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
385 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
390 var
->data
.mode
= nir_var_uniform
;
393 case ir_var_shader_storage
:
394 var
->data
.mode
= nir_var_shader_storage
;
397 case ir_var_system_value
:
398 var
->data
.mode
= nir_var_system_value
;
402 unreachable("not reached");
405 var
->data
.interpolation
= ir
->data
.interpolation
;
406 var
->data
.origin_upper_left
= ir
->data
.origin_upper_left
;
407 var
->data
.pixel_center_integer
= ir
->data
.pixel_center_integer
;
408 var
->data
.location_frac
= ir
->data
.location_frac
;
410 if (var
->data
.pixel_center_integer
) {
411 assert(shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
412 shader
->info
.fs
.pixel_center_integer
= true;
415 switch (ir
->data
.depth_layout
) {
416 case ir_depth_layout_none
:
417 var
->data
.depth_layout
= nir_depth_layout_none
;
419 case ir_depth_layout_any
:
420 var
->data
.depth_layout
= nir_depth_layout_any
;
422 case ir_depth_layout_greater
:
423 var
->data
.depth_layout
= nir_depth_layout_greater
;
425 case ir_depth_layout_less
:
426 var
->data
.depth_layout
= nir_depth_layout_less
;
428 case ir_depth_layout_unchanged
:
429 var
->data
.depth_layout
= nir_depth_layout_unchanged
;
432 unreachable("not reached");
435 var
->data
.index
= ir
->data
.index
;
436 var
->data
.descriptor_set
= 0;
437 var
->data
.binding
= ir
->data
.binding
;
438 var
->data
.explicit_binding
= ir
->data
.explicit_binding
;
439 var
->data
.bindless
= ir
->data
.bindless
;
440 var
->data
.offset
= ir
->data
.offset
;
441 var
->data
.image
.read_only
= ir
->data
.memory_read_only
;
442 var
->data
.image
.write_only
= ir
->data
.memory_write_only
;
443 var
->data
.image
.coherent
= ir
->data
.memory_coherent
;
444 var
->data
.image
._volatile
= ir
->data
.memory_volatile
;
445 var
->data
.image
.restrict_flag
= ir
->data
.memory_restrict
;
446 var
->data
.image
.format
= ir
->data
.image_format
;
447 var
->data
.fb_fetch_output
= ir
->data
.fb_fetch_output
;
449 var
->num_state_slots
= ir
->get_num_state_slots();
450 if (var
->num_state_slots
> 0) {
451 var
->state_slots
= rzalloc_array(var
, nir_state_slot
,
452 var
->num_state_slots
);
454 ir_state_slot
*state_slots
= ir
->get_state_slots();
455 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
456 for (unsigned j
= 0; j
< 5; j
++)
457 var
->state_slots
[i
].tokens
[j
] = state_slots
[i
].tokens
[j
];
458 var
->state_slots
[i
].swizzle
= state_slots
[i
].swizzle
;
461 var
->state_slots
= NULL
;
464 var
->constant_initializer
= constant_copy(ir
->constant_initializer
, var
);
466 var
->interface_type
= ir
->get_interface_type();
468 if (var
->data
.mode
== nir_var_local
)
469 nir_function_impl_add_variable(impl
, var
);
471 nir_shader_add_variable(shader
, var
);
473 _mesa_hash_table_insert(var_table
, ir
, var
);
478 nir_function_visitor::visit_enter(ir_function
*ir
)
480 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
) {
481 visitor
->create_function(sig
);
483 return visit_continue_with_parent
;
487 nir_visitor::create_function(ir_function_signature
*ir
)
489 if (ir
->is_intrinsic())
492 nir_function
*func
= nir_function_create(shader
, ir
->function_name());
494 assert(ir
->parameters
.is_empty());
495 assert(ir
->return_type
== glsl_type::void_type
);
497 _mesa_hash_table_insert(this->overload_table
, ir
, func
);
501 nir_visitor::visit(ir_function
*ir
)
503 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
)
508 nir_visitor::visit(ir_function_signature
*ir
)
510 if (ir
->is_intrinsic())
513 struct hash_entry
*entry
=
514 _mesa_hash_table_search(this->overload_table
, ir
);
517 nir_function
*func
= (nir_function
*) entry
->data
;
519 if (ir
->is_defined
) {
520 nir_function_impl
*impl
= nir_function_impl_create(func
);
523 assert(strcmp(func
->name
, "main") == 0);
524 assert(ir
->parameters
.is_empty());
525 assert(func
->return_type
== glsl_type::void_type
);
527 this->is_global
= false;
529 nir_builder_init(&b
, impl
);
530 b
.cursor
= nir_after_cf_list(&impl
->body
);
531 visit_exec_list(&ir
->body
, this);
533 this->is_global
= true;
540 nir_visitor::visit(ir_loop
*ir
)
543 visit_exec_list(&ir
->body_instructions
, this);
544 nir_pop_loop(&b
, NULL
);
548 nir_visitor::visit(ir_if
*ir
)
550 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
551 visit_exec_list(&ir
->then_instructions
, this);
552 nir_push_else(&b
, NULL
);
553 visit_exec_list(&ir
->else_instructions
, this);
554 nir_pop_if(&b
, NULL
);
558 nir_visitor::visit(ir_discard
*ir
)
561 * discards aren't treated as control flow, because before we lower them
562 * they can appear anywhere in the shader and the stuff after them may still
563 * be executed (yay, crazy GLSL rules!). However, after lowering, all the
564 * discards will be immediately followed by a return.
567 nir_intrinsic_instr
*discard
;
569 discard
= nir_intrinsic_instr_create(this->shader
,
570 nir_intrinsic_discard_if
);
572 nir_src_for_ssa(evaluate_rvalue(ir
->condition
));
574 discard
= nir_intrinsic_instr_create(this->shader
, nir_intrinsic_discard
);
577 nir_builder_instr_insert(&b
, &discard
->instr
);
581 nir_visitor::visit(ir_emit_vertex
*ir
)
583 nir_intrinsic_instr
*instr
=
584 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_emit_vertex
);
585 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
586 nir_builder_instr_insert(&b
, &instr
->instr
);
590 nir_visitor::visit(ir_end_primitive
*ir
)
592 nir_intrinsic_instr
*instr
=
593 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_end_primitive
);
594 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
595 nir_builder_instr_insert(&b
, &instr
->instr
);
599 nir_visitor::visit(ir_loop_jump
*ir
)
603 case ir_loop_jump::jump_break
:
604 type
= nir_jump_break
;
606 case ir_loop_jump::jump_continue
:
607 type
= nir_jump_continue
;
610 unreachable("not reached");
613 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, type
);
614 nir_builder_instr_insert(&b
, &instr
->instr
);
618 nir_visitor::visit(ir_return
*ir
)
620 if (ir
->value
!= NULL
) {
621 nir_intrinsic_instr
*copy
=
622 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_copy_var
);
624 copy
->variables
[0] = nir_deref_var_create(copy
, this->impl
->return_var
);
625 copy
->variables
[1] = evaluate_deref(©
->instr
, ir
->value
);
628 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, nir_jump_return
);
629 nir_builder_instr_insert(&b
, &instr
->instr
);
633 nir_visitor::visit(ir_call
*ir
)
635 if (ir
->callee
->is_intrinsic()) {
638 switch (ir
->callee
->intrinsic_id
) {
639 case ir_intrinsic_atomic_counter_read
:
640 op
= nir_intrinsic_atomic_counter_read_var
;
642 case ir_intrinsic_atomic_counter_increment
:
643 op
= nir_intrinsic_atomic_counter_inc_var
;
645 case ir_intrinsic_atomic_counter_predecrement
:
646 op
= nir_intrinsic_atomic_counter_dec_var
;
648 case ir_intrinsic_atomic_counter_add
:
649 op
= nir_intrinsic_atomic_counter_add_var
;
651 case ir_intrinsic_atomic_counter_and
:
652 op
= nir_intrinsic_atomic_counter_and_var
;
654 case ir_intrinsic_atomic_counter_or
:
655 op
= nir_intrinsic_atomic_counter_or_var
;
657 case ir_intrinsic_atomic_counter_xor
:
658 op
= nir_intrinsic_atomic_counter_xor_var
;
660 case ir_intrinsic_atomic_counter_min
:
661 op
= nir_intrinsic_atomic_counter_min_var
;
663 case ir_intrinsic_atomic_counter_max
:
664 op
= nir_intrinsic_atomic_counter_max_var
;
666 case ir_intrinsic_atomic_counter_exchange
:
667 op
= nir_intrinsic_atomic_counter_exchange_var
;
669 case ir_intrinsic_atomic_counter_comp_swap
:
670 op
= nir_intrinsic_atomic_counter_comp_swap_var
;
672 case ir_intrinsic_image_load
:
673 op
= nir_intrinsic_image_var_load
;
675 case ir_intrinsic_image_store
:
676 op
= nir_intrinsic_image_var_store
;
678 case ir_intrinsic_image_atomic_add
:
679 op
= nir_intrinsic_image_var_atomic_add
;
681 case ir_intrinsic_image_atomic_min
:
682 op
= nir_intrinsic_image_var_atomic_min
;
684 case ir_intrinsic_image_atomic_max
:
685 op
= nir_intrinsic_image_var_atomic_max
;
687 case ir_intrinsic_image_atomic_and
:
688 op
= nir_intrinsic_image_var_atomic_and
;
690 case ir_intrinsic_image_atomic_or
:
691 op
= nir_intrinsic_image_var_atomic_or
;
693 case ir_intrinsic_image_atomic_xor
:
694 op
= nir_intrinsic_image_var_atomic_xor
;
696 case ir_intrinsic_image_atomic_exchange
:
697 op
= nir_intrinsic_image_var_atomic_exchange
;
699 case ir_intrinsic_image_atomic_comp_swap
:
700 op
= nir_intrinsic_image_var_atomic_comp_swap
;
702 case ir_intrinsic_memory_barrier
:
703 op
= nir_intrinsic_memory_barrier
;
705 case ir_intrinsic_image_size
:
706 op
= nir_intrinsic_image_var_size
;
708 case ir_intrinsic_image_samples
:
709 op
= nir_intrinsic_image_var_samples
;
711 case ir_intrinsic_ssbo_store
:
712 op
= nir_intrinsic_store_ssbo
;
714 case ir_intrinsic_ssbo_load
:
715 op
= nir_intrinsic_load_ssbo
;
717 case ir_intrinsic_ssbo_atomic_add
:
718 op
= nir_intrinsic_ssbo_atomic_add
;
720 case ir_intrinsic_ssbo_atomic_and
:
721 op
= nir_intrinsic_ssbo_atomic_and
;
723 case ir_intrinsic_ssbo_atomic_or
:
724 op
= nir_intrinsic_ssbo_atomic_or
;
726 case ir_intrinsic_ssbo_atomic_xor
:
727 op
= nir_intrinsic_ssbo_atomic_xor
;
729 case ir_intrinsic_ssbo_atomic_min
:
730 assert(ir
->return_deref
);
731 if (ir
->return_deref
->type
== glsl_type::int_type
)
732 op
= nir_intrinsic_ssbo_atomic_imin
;
733 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
734 op
= nir_intrinsic_ssbo_atomic_umin
;
736 unreachable("Invalid type");
738 case ir_intrinsic_ssbo_atomic_max
:
739 assert(ir
->return_deref
);
740 if (ir
->return_deref
->type
== glsl_type::int_type
)
741 op
= nir_intrinsic_ssbo_atomic_imax
;
742 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
743 op
= nir_intrinsic_ssbo_atomic_umax
;
745 unreachable("Invalid type");
747 case ir_intrinsic_ssbo_atomic_exchange
:
748 op
= nir_intrinsic_ssbo_atomic_exchange
;
750 case ir_intrinsic_ssbo_atomic_comp_swap
:
751 op
= nir_intrinsic_ssbo_atomic_comp_swap
;
753 case ir_intrinsic_shader_clock
:
754 op
= nir_intrinsic_shader_clock
;
756 case ir_intrinsic_begin_invocation_interlock
:
757 op
= nir_intrinsic_begin_invocation_interlock
;
759 case ir_intrinsic_end_invocation_interlock
:
760 op
= nir_intrinsic_end_invocation_interlock
;
762 case ir_intrinsic_group_memory_barrier
:
763 op
= nir_intrinsic_group_memory_barrier
;
765 case ir_intrinsic_memory_barrier_atomic_counter
:
766 op
= nir_intrinsic_memory_barrier_atomic_counter
;
768 case ir_intrinsic_memory_barrier_buffer
:
769 op
= nir_intrinsic_memory_barrier_buffer
;
771 case ir_intrinsic_memory_barrier_image
:
772 op
= nir_intrinsic_memory_barrier_image
;
774 case ir_intrinsic_memory_barrier_shared
:
775 op
= nir_intrinsic_memory_barrier_shared
;
777 case ir_intrinsic_shared_load
:
778 op
= nir_intrinsic_load_shared
;
780 case ir_intrinsic_shared_store
:
781 op
= nir_intrinsic_store_shared
;
783 case ir_intrinsic_shared_atomic_add
:
784 op
= nir_intrinsic_shared_atomic_add
;
786 case ir_intrinsic_shared_atomic_and
:
787 op
= nir_intrinsic_shared_atomic_and
;
789 case ir_intrinsic_shared_atomic_or
:
790 op
= nir_intrinsic_shared_atomic_or
;
792 case ir_intrinsic_shared_atomic_xor
:
793 op
= nir_intrinsic_shared_atomic_xor
;
795 case ir_intrinsic_shared_atomic_min
:
796 assert(ir
->return_deref
);
797 if (ir
->return_deref
->type
== glsl_type::int_type
)
798 op
= nir_intrinsic_shared_atomic_imin
;
799 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
800 op
= nir_intrinsic_shared_atomic_umin
;
802 unreachable("Invalid type");
804 case ir_intrinsic_shared_atomic_max
:
805 assert(ir
->return_deref
);
806 if (ir
->return_deref
->type
== glsl_type::int_type
)
807 op
= nir_intrinsic_shared_atomic_imax
;
808 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
809 op
= nir_intrinsic_shared_atomic_umax
;
811 unreachable("Invalid type");
813 case ir_intrinsic_shared_atomic_exchange
:
814 op
= nir_intrinsic_shared_atomic_exchange
;
816 case ir_intrinsic_shared_atomic_comp_swap
:
817 op
= nir_intrinsic_shared_atomic_comp_swap
;
819 case ir_intrinsic_vote_any
:
820 op
= nir_intrinsic_vote_any
;
822 case ir_intrinsic_vote_all
:
823 op
= nir_intrinsic_vote_all
;
825 case ir_intrinsic_vote_eq
:
826 op
= nir_intrinsic_vote_ieq
;
828 case ir_intrinsic_ballot
:
829 op
= nir_intrinsic_ballot
;
831 case ir_intrinsic_read_invocation
:
832 op
= nir_intrinsic_read_invocation
;
834 case ir_intrinsic_read_first_invocation
:
835 op
= nir_intrinsic_read_first_invocation
;
838 unreachable("not reached");
841 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(shader
, op
);
842 nir_dest
*dest
= &instr
->dest
;
845 case nir_intrinsic_atomic_counter_read_var
:
846 case nir_intrinsic_atomic_counter_inc_var
:
847 case nir_intrinsic_atomic_counter_dec_var
:
848 case nir_intrinsic_atomic_counter_add_var
:
849 case nir_intrinsic_atomic_counter_min_var
:
850 case nir_intrinsic_atomic_counter_max_var
:
851 case nir_intrinsic_atomic_counter_and_var
:
852 case nir_intrinsic_atomic_counter_or_var
:
853 case nir_intrinsic_atomic_counter_xor_var
:
854 case nir_intrinsic_atomic_counter_exchange_var
:
855 case nir_intrinsic_atomic_counter_comp_swap_var
: {
856 /* Set the counter variable dereference. */
857 exec_node
*param
= ir
->actual_parameters
.get_head();
858 ir_dereference
*counter
= (ir_dereference
*)param
;
860 instr
->variables
[0] = evaluate_deref(&instr
->instr
, counter
);
861 param
= param
->get_next();
863 /* Set the intrinsic destination. */
864 if (ir
->return_deref
) {
865 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
868 /* Set the intrinsic parameters. */
869 if (!param
->is_tail_sentinel()) {
871 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
872 param
= param
->get_next();
875 if (!param
->is_tail_sentinel()) {
877 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
878 param
= param
->get_next();
881 nir_builder_instr_insert(&b
, &instr
->instr
);
884 case nir_intrinsic_image_var_load
:
885 case nir_intrinsic_image_var_store
:
886 case nir_intrinsic_image_var_atomic_add
:
887 case nir_intrinsic_image_var_atomic_min
:
888 case nir_intrinsic_image_var_atomic_max
:
889 case nir_intrinsic_image_var_atomic_and
:
890 case nir_intrinsic_image_var_atomic_or
:
891 case nir_intrinsic_image_var_atomic_xor
:
892 case nir_intrinsic_image_var_atomic_exchange
:
893 case nir_intrinsic_image_var_atomic_comp_swap
:
894 case nir_intrinsic_image_var_samples
:
895 case nir_intrinsic_image_var_size
: {
896 nir_ssa_undef_instr
*instr_undef
=
897 nir_ssa_undef_instr_create(shader
, 1, 32);
898 nir_builder_instr_insert(&b
, &instr_undef
->instr
);
900 /* Set the image variable dereference. */
901 exec_node
*param
= ir
->actual_parameters
.get_head();
902 ir_dereference
*image
= (ir_dereference
*)param
;
903 const glsl_type
*type
=
904 image
->variable_referenced()->type
->without_array();
906 instr
->variables
[0] = evaluate_deref(&instr
->instr
, image
);
907 param
= param
->get_next();
909 /* Set the intrinsic destination. */
910 if (ir
->return_deref
) {
911 unsigned num_components
= ir
->return_deref
->type
->vector_elements
;
912 if (instr
->intrinsic
== nir_intrinsic_image_var_size
)
913 instr
->num_components
= num_components
;
914 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
915 num_components
, 32, NULL
);
918 if (op
== nir_intrinsic_image_var_size
||
919 op
== nir_intrinsic_image_var_samples
) {
920 nir_builder_instr_insert(&b
, &instr
->instr
);
924 /* Set the address argument, extending the coordinate vector to four
927 nir_ssa_def
*src_addr
=
928 evaluate_rvalue((ir_dereference
*)param
);
929 nir_ssa_def
*srcs
[4];
931 for (int i
= 0; i
< 4; i
++) {
932 if (i
< type
->coordinate_components())
933 srcs
[i
] = nir_channel(&b
, src_addr
, i
);
935 srcs
[i
] = &instr_undef
->def
;
938 instr
->src
[0] = nir_src_for_ssa(nir_vec(&b
, srcs
, 4));
939 param
= param
->get_next();
941 /* Set the sample argument, which is undefined for single-sample
944 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_MS
) {
946 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
947 param
= param
->get_next();
949 instr
->src
[1] = nir_src_for_ssa(&instr_undef
->def
);
952 /* Set the intrinsic parameters. */
953 if (!param
->is_tail_sentinel()) {
955 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
956 param
= param
->get_next();
959 if (!param
->is_tail_sentinel()) {
961 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
962 param
= param
->get_next();
964 nir_builder_instr_insert(&b
, &instr
->instr
);
967 case nir_intrinsic_memory_barrier
:
968 case nir_intrinsic_group_memory_barrier
:
969 case nir_intrinsic_memory_barrier_atomic_counter
:
970 case nir_intrinsic_memory_barrier_buffer
:
971 case nir_intrinsic_memory_barrier_image
:
972 case nir_intrinsic_memory_barrier_shared
:
973 nir_builder_instr_insert(&b
, &instr
->instr
);
975 case nir_intrinsic_shader_clock
:
976 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 2, 32, NULL
);
977 instr
->num_components
= 2;
978 nir_builder_instr_insert(&b
, &instr
->instr
);
980 case nir_intrinsic_begin_invocation_interlock
:
981 nir_builder_instr_insert(&b
, &instr
->instr
);
983 case nir_intrinsic_end_invocation_interlock
:
984 nir_builder_instr_insert(&b
, &instr
->instr
);
986 case nir_intrinsic_store_ssbo
: {
987 exec_node
*param
= ir
->actual_parameters
.get_head();
988 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
990 param
= param
->get_next();
991 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
993 param
= param
->get_next();
994 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
996 param
= param
->get_next();
997 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1000 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(val
));
1001 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(block
));
1002 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(offset
));
1003 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1004 instr
->num_components
= val
->type
->vector_elements
;
1006 nir_builder_instr_insert(&b
, &instr
->instr
);
1009 case nir_intrinsic_load_ssbo
: {
1010 exec_node
*param
= ir
->actual_parameters
.get_head();
1011 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1013 param
= param
->get_next();
1014 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1016 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(block
));
1017 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1019 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1020 instr
->num_components
= type
->vector_elements
;
1022 /* Setup destination register */
1023 unsigned bit_size
= glsl_get_bit_size(type
);
1024 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1025 type
->vector_elements
, bit_size
, NULL
);
1027 /* Insert the created nir instruction now since in the case of boolean
1028 * result we will need to emit another instruction after it
1030 nir_builder_instr_insert(&b
, &instr
->instr
);
1033 * In SSBO/UBO's, a true boolean value is any non-zero value, but we
1034 * consider a true boolean to be ~0. Fix this up with a != 0
1037 if (type
->is_boolean()) {
1038 nir_alu_instr
*load_ssbo_compare
=
1039 nir_alu_instr_create(shader
, nir_op_ine
);
1040 load_ssbo_compare
->src
[0].src
.is_ssa
= true;
1041 load_ssbo_compare
->src
[0].src
.ssa
= &instr
->dest
.ssa
;
1042 load_ssbo_compare
->src
[1].src
=
1043 nir_src_for_ssa(nir_imm_int(&b
, 0));
1044 for (unsigned i
= 0; i
< type
->vector_elements
; i
++)
1045 load_ssbo_compare
->src
[1].swizzle
[i
] = 0;
1046 nir_ssa_dest_init(&load_ssbo_compare
->instr
,
1047 &load_ssbo_compare
->dest
.dest
,
1048 type
->vector_elements
, bit_size
, NULL
);
1049 load_ssbo_compare
->dest
.write_mask
= (1 << type
->vector_elements
) - 1;
1050 nir_builder_instr_insert(&b
, &load_ssbo_compare
->instr
);
1051 dest
= &load_ssbo_compare
->dest
.dest
;
1055 case nir_intrinsic_ssbo_atomic_add
:
1056 case nir_intrinsic_ssbo_atomic_imin
:
1057 case nir_intrinsic_ssbo_atomic_umin
:
1058 case nir_intrinsic_ssbo_atomic_imax
:
1059 case nir_intrinsic_ssbo_atomic_umax
:
1060 case nir_intrinsic_ssbo_atomic_and
:
1061 case nir_intrinsic_ssbo_atomic_or
:
1062 case nir_intrinsic_ssbo_atomic_xor
:
1063 case nir_intrinsic_ssbo_atomic_exchange
:
1064 case nir_intrinsic_ssbo_atomic_comp_swap
: {
1065 int param_count
= ir
->actual_parameters
.length();
1066 assert(param_count
== 3 || param_count
== 4);
1069 exec_node
*param
= ir
->actual_parameters
.get_head();
1070 ir_instruction
*inst
= (ir_instruction
*) param
;
1071 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1074 param
= param
->get_next();
1075 inst
= (ir_instruction
*) param
;
1076 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1078 /* data1 parameter (this is always present) */
1079 param
= param
->get_next();
1080 inst
= (ir_instruction
*) param
;
1081 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1083 /* data2 parameter (only with atomic_comp_swap) */
1084 if (param_count
== 4) {
1085 assert(op
== nir_intrinsic_ssbo_atomic_comp_swap
);
1086 param
= param
->get_next();
1087 inst
= (ir_instruction
*) param
;
1088 instr
->src
[3] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1092 assert(ir
->return_deref
);
1093 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1094 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1095 nir_builder_instr_insert(&b
, &instr
->instr
);
1098 case nir_intrinsic_load_shared
: {
1099 exec_node
*param
= ir
->actual_parameters
.get_head();
1100 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1102 nir_intrinsic_set_base(instr
, 0);
1103 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(offset
));
1105 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1106 instr
->num_components
= type
->vector_elements
;
1108 /* Setup destination register */
1109 unsigned bit_size
= glsl_get_bit_size(type
);
1110 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1111 type
->vector_elements
, bit_size
, NULL
);
1113 nir_builder_instr_insert(&b
, &instr
->instr
);
1116 case nir_intrinsic_store_shared
: {
1117 exec_node
*param
= ir
->actual_parameters
.get_head();
1118 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1120 param
= param
->get_next();
1121 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1123 param
= param
->get_next();
1124 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1127 nir_intrinsic_set_base(instr
, 0);
1128 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1130 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1132 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(val
));
1133 instr
->num_components
= val
->type
->vector_elements
;
1135 nir_builder_instr_insert(&b
, &instr
->instr
);
1138 case nir_intrinsic_shared_atomic_add
:
1139 case nir_intrinsic_shared_atomic_imin
:
1140 case nir_intrinsic_shared_atomic_umin
:
1141 case nir_intrinsic_shared_atomic_imax
:
1142 case nir_intrinsic_shared_atomic_umax
:
1143 case nir_intrinsic_shared_atomic_and
:
1144 case nir_intrinsic_shared_atomic_or
:
1145 case nir_intrinsic_shared_atomic_xor
:
1146 case nir_intrinsic_shared_atomic_exchange
:
1147 case nir_intrinsic_shared_atomic_comp_swap
: {
1148 int param_count
= ir
->actual_parameters
.length();
1149 assert(param_count
== 2 || param_count
== 3);
1152 exec_node
*param
= ir
->actual_parameters
.get_head();
1153 ir_instruction
*inst
= (ir_instruction
*) param
;
1154 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1156 /* data1 parameter (this is always present) */
1157 param
= param
->get_next();
1158 inst
= (ir_instruction
*) param
;
1159 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1161 /* data2 parameter (only with atomic_comp_swap) */
1162 if (param_count
== 3) {
1163 assert(op
== nir_intrinsic_shared_atomic_comp_swap
);
1164 param
= param
->get_next();
1165 inst
= (ir_instruction
*) param
;
1167 nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1171 assert(ir
->return_deref
);
1172 unsigned bit_size
= glsl_get_bit_size(ir
->return_deref
->type
);
1173 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1174 ir
->return_deref
->type
->vector_elements
,
1176 nir_builder_instr_insert(&b
, &instr
->instr
);
1179 case nir_intrinsic_vote_any
:
1180 case nir_intrinsic_vote_all
:
1181 case nir_intrinsic_vote_ieq
: {
1182 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1183 instr
->num_components
= 1;
1185 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1186 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1188 nir_builder_instr_insert(&b
, &instr
->instr
);
1192 case nir_intrinsic_ballot
: {
1193 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1194 ir
->return_deref
->type
->vector_elements
, 64, NULL
);
1195 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1197 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1198 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1200 nir_builder_instr_insert(&b
, &instr
->instr
);
1203 case nir_intrinsic_read_invocation
: {
1204 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1205 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1206 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1208 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1209 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1211 ir_rvalue
*invocation
= (ir_rvalue
*) ir
->actual_parameters
.get_head()->next
;
1212 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(invocation
));
1214 nir_builder_instr_insert(&b
, &instr
->instr
);
1217 case nir_intrinsic_read_first_invocation
: {
1218 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1219 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1220 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1222 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1223 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1225 nir_builder_instr_insert(&b
, &instr
->instr
);
1229 unreachable("not reached");
1232 if (ir
->return_deref
) {
1233 nir_intrinsic_instr
*store_instr
=
1234 nir_intrinsic_instr_create(shader
, nir_intrinsic_store_var
);
1235 store_instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1236 nir_intrinsic_set_write_mask(store_instr
,
1237 (1 << store_instr
->num_components
) - 1);
1239 store_instr
->variables
[0] =
1240 evaluate_deref(&store_instr
->instr
, ir
->return_deref
);
1241 store_instr
->src
[0] = nir_src_for_ssa(&dest
->ssa
);
1243 nir_builder_instr_insert(&b
, &store_instr
->instr
);
1249 struct hash_entry
*entry
=
1250 _mesa_hash_table_search(this->overload_table
, ir
->callee
);
1252 nir_function
*callee
= (nir_function
*) entry
->data
;
1254 nir_call_instr
*instr
= nir_call_instr_create(this->shader
, callee
);
1257 foreach_in_list(ir_dereference
, param
, &ir
->actual_parameters
) {
1258 instr
->params
[i
] = evaluate_deref(&instr
->instr
, param
);
1262 instr
->return_deref
= evaluate_deref(&instr
->instr
, ir
->return_deref
);
1263 nir_builder_instr_insert(&b
, &instr
->instr
);
1267 nir_visitor::visit(ir_assignment
*ir
)
1269 unsigned num_components
= ir
->lhs
->type
->vector_elements
;
1271 b
.exact
= ir
->lhs
->variable_referenced()->data
.invariant
||
1272 ir
->lhs
->variable_referenced()->data
.precise
;
1274 if ((ir
->rhs
->as_dereference() || ir
->rhs
->as_constant()) &&
1275 (ir
->write_mask
== (1 << num_components
) - 1 || ir
->write_mask
== 0)) {
1276 /* We're doing a plain-as-can-be copy, so emit a copy_var */
1277 nir_intrinsic_instr
*copy
=
1278 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_copy_var
);
1280 copy
->variables
[0] = evaluate_deref(©
->instr
, ir
->lhs
);
1281 copy
->variables
[1] = evaluate_deref(©
->instr
, ir
->rhs
);
1283 if (ir
->condition
) {
1284 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1285 nir_builder_instr_insert(&b
, ©
->instr
);
1286 nir_pop_if(&b
, NULL
);
1288 nir_builder_instr_insert(&b
, ©
->instr
);
1293 assert(ir
->rhs
->type
->is_scalar() || ir
->rhs
->type
->is_vector());
1295 ir
->lhs
->accept(this);
1296 nir_deref_var
*lhs_deref
= this->deref_head
;
1297 nir_ssa_def
*src
= evaluate_rvalue(ir
->rhs
);
1299 if (ir
->write_mask
!= (1 << num_components
) - 1 && ir
->write_mask
!= 0) {
1300 /* GLSL IR will give us the input to the write-masked assignment in a
1301 * single packed vector. So, for example, if the writemask is xzw, then
1302 * we have to swizzle x -> x, y -> z, and z -> w and get the y component
1306 unsigned component
= 0;
1307 for (unsigned i
= 0; i
< 4; i
++) {
1308 swiz
[i
] = ir
->write_mask
& (1 << i
) ? component
++ : 0;
1310 src
= nir_swizzle(&b
, src
, swiz
, num_components
, !supports_ints
);
1313 nir_intrinsic_instr
*store
=
1314 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_store_var
);
1315 store
->num_components
= ir
->lhs
->type
->vector_elements
;
1316 nir_intrinsic_set_write_mask(store
, ir
->write_mask
);
1317 store
->variables
[0] = nir_deref_var_clone(lhs_deref
, store
);
1318 store
->src
[0] = nir_src_for_ssa(src
);
1320 if (ir
->condition
) {
1321 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1322 nir_builder_instr_insert(&b
, &store
->instr
);
1323 nir_pop_if(&b
, NULL
);
1325 nir_builder_instr_insert(&b
, &store
->instr
);
1330 * Given an instruction, returns a pointer to its destination or NULL if there
1331 * is no destination.
1333 * Note that this only handles instructions we generate at this level.
1336 get_instr_dest(nir_instr
*instr
)
1338 nir_alu_instr
*alu_instr
;
1339 nir_intrinsic_instr
*intrinsic_instr
;
1340 nir_tex_instr
*tex_instr
;
1342 switch (instr
->type
) {
1343 case nir_instr_type_alu
:
1344 alu_instr
= nir_instr_as_alu(instr
);
1345 return &alu_instr
->dest
.dest
;
1347 case nir_instr_type_intrinsic
:
1348 intrinsic_instr
= nir_instr_as_intrinsic(instr
);
1349 if (nir_intrinsic_infos
[intrinsic_instr
->intrinsic
].has_dest
)
1350 return &intrinsic_instr
->dest
;
1354 case nir_instr_type_tex
:
1355 tex_instr
= nir_instr_as_tex(instr
);
1356 return &tex_instr
->dest
;
1359 unreachable("not reached");
1366 nir_visitor::add_instr(nir_instr
*instr
, unsigned num_components
,
1369 nir_dest
*dest
= get_instr_dest(instr
);
1372 nir_ssa_dest_init(instr
, dest
, num_components
, bit_size
, NULL
);
1374 nir_builder_instr_insert(&b
, instr
);
1377 assert(dest
->is_ssa
);
1378 this->result
= &dest
->ssa
;
1383 nir_visitor::evaluate_rvalue(ir_rvalue
* ir
)
1386 if (ir
->as_dereference() || ir
->as_constant()) {
1388 * A dereference is being used on the right hand side, which means we
1389 * must emit a variable load.
1392 nir_intrinsic_instr
*load_instr
=
1393 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_var
);
1394 load_instr
->num_components
= ir
->type
->vector_elements
;
1395 load_instr
->variables
[0] = this->deref_head
;
1396 ralloc_steal(load_instr
, load_instr
->variables
[0]);
1397 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1398 add_instr(&load_instr
->instr
, ir
->type
->vector_elements
, bit_size
);
1401 return this->result
;
1405 type_is_float(glsl_base_type type
)
1407 return type
== GLSL_TYPE_FLOAT
|| type
== GLSL_TYPE_DOUBLE
||
1408 type
== GLSL_TYPE_FLOAT16
;
1412 type_is_signed(glsl_base_type type
)
1414 return type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_INT64
||
1415 type
== GLSL_TYPE_INT16
;
1419 nir_visitor::visit(ir_expression
*ir
)
1421 /* Some special cases */
1422 switch (ir
->operation
) {
1423 case ir_binop_ubo_load
: {
1424 nir_intrinsic_instr
*load
=
1425 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_ubo
);
1426 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1427 load
->num_components
= ir
->type
->vector_elements
;
1428 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1429 load
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1430 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1433 * In UBO's, a true boolean value is any non-zero value, but we consider
1434 * a true boolean to be ~0. Fix this up with a != 0 comparison.
1437 if (ir
->type
->is_boolean())
1438 this->result
= nir_ine(&b
, &load
->dest
.ssa
, nir_imm_int(&b
, 0));
1443 case ir_unop_interpolate_at_centroid
:
1444 case ir_binop_interpolate_at_offset
:
1445 case ir_binop_interpolate_at_sample
: {
1446 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1447 ir_swizzle
*swizzle
= NULL
;
1449 /* the api does not allow a swizzle here, but the varying packing code
1450 * may have pushed one into here.
1452 swizzle
= ir
->operands
[0]->as_swizzle();
1454 deref
= swizzle
->val
->as_dereference();
1458 deref
->accept(this);
1460 nir_intrinsic_op op
;
1461 if (this->deref_head
->var
->data
.mode
== nir_var_shader_in
) {
1462 switch (ir
->operation
) {
1463 case ir_unop_interpolate_at_centroid
:
1464 op
= nir_intrinsic_interp_var_at_centroid
;
1466 case ir_binop_interpolate_at_offset
:
1467 op
= nir_intrinsic_interp_var_at_offset
;
1469 case ir_binop_interpolate_at_sample
:
1470 op
= nir_intrinsic_interp_var_at_sample
;
1473 unreachable("Invalid interpolation intrinsic");
1476 /* This case can happen if the vertex shader does not write the
1477 * given varying. In this case, the linker will lower it to a
1478 * global variable. Since interpolating a variable makes no
1479 * sense, we'll just turn it into a load which will probably
1480 * eventually end up as an SSA definition.
1482 assert(this->deref_head
->var
->data
.mode
== nir_var_global
);
1483 op
= nir_intrinsic_load_var
;
1486 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(shader
, op
);
1487 intrin
->num_components
= deref
->type
->vector_elements
;
1488 intrin
->variables
[0] = this->deref_head
;
1489 ralloc_steal(intrin
, intrin
->variables
[0]);
1491 if (intrin
->intrinsic
== nir_intrinsic_interp_var_at_offset
||
1492 intrin
->intrinsic
== nir_intrinsic_interp_var_at_sample
)
1493 intrin
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1495 unsigned bit_size
= glsl_get_bit_size(deref
->type
);
1496 add_instr(&intrin
->instr
, deref
->type
->vector_elements
, bit_size
);
1499 unsigned swiz
[4] = {
1500 swizzle
->mask
.x
, swizzle
->mask
.y
, swizzle
->mask
.z
, swizzle
->mask
.w
1503 result
= nir_swizzle(&b
, result
, swiz
,
1504 swizzle
->type
->vector_elements
, false);
1514 nir_ssa_def
*srcs
[4];
1515 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1516 srcs
[i
] = evaluate_rvalue(ir
->operands
[i
]);
1518 glsl_base_type types
[4];
1519 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1521 types
[i
] = ir
->operands
[i
]->type
->base_type
;
1523 types
[i
] = GLSL_TYPE_FLOAT
;
1525 glsl_base_type out_type
;
1527 out_type
= ir
->type
->base_type
;
1529 out_type
= GLSL_TYPE_FLOAT
;
1531 switch (ir
->operation
) {
1532 case ir_unop_bit_not
: result
= nir_inot(&b
, srcs
[0]); break;
1533 case ir_unop_logic_not
:
1534 result
= supports_ints
? nir_inot(&b
, srcs
[0]) : nir_fnot(&b
, srcs
[0]);
1537 result
= type_is_float(types
[0]) ? nir_fneg(&b
, srcs
[0])
1538 : nir_ineg(&b
, srcs
[0]);
1541 result
= type_is_float(types
[0]) ? nir_fabs(&b
, srcs
[0])
1542 : nir_iabs(&b
, srcs
[0]);
1544 case ir_unop_saturate
:
1545 assert(type_is_float(types
[0]));
1546 result
= nir_fsat(&b
, srcs
[0]);
1549 result
= type_is_float(types
[0]) ? nir_fsign(&b
, srcs
[0])
1550 : nir_isign(&b
, srcs
[0]);
1552 case ir_unop_rcp
: result
= nir_frcp(&b
, srcs
[0]); break;
1553 case ir_unop_rsq
: result
= nir_frsq(&b
, srcs
[0]); break;
1554 case ir_unop_sqrt
: result
= nir_fsqrt(&b
, srcs
[0]); break;
1555 case ir_unop_exp
: unreachable("ir_unop_exp should have been lowered");
1556 case ir_unop_log
: unreachable("ir_unop_log should have been lowered");
1557 case ir_unop_exp2
: result
= nir_fexp2(&b
, srcs
[0]); break;
1558 case ir_unop_log2
: result
= nir_flog2(&b
, srcs
[0]); break;
1560 result
= supports_ints
? nir_i2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1563 result
= supports_ints
? nir_u2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1566 result
= supports_ints
? nir_b2f(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1600 case ir_unop_i642u64
:
1601 case ir_unop_u642i64
: {
1602 nir_alu_type src_type
= nir_get_nir_type_for_glsl_base_type(types
[0]);
1603 nir_alu_type dst_type
= nir_get_nir_type_for_glsl_base_type(out_type
);
1604 result
= nir_build_alu(&b
, nir_type_conversion_op(src_type
, dst_type
,
1605 nir_rounding_mode_undef
),
1606 srcs
[0], NULL
, NULL
, NULL
);
1607 /* b2i and b2f don't have fixed bit-size versions so the builder will
1608 * just assume 32 and we have to fix it up here.
1610 result
->bit_size
= nir_alu_type_get_type_size(dst_type
);
1614 case ir_unop_bitcast_i2f
:
1615 case ir_unop_bitcast_f2i
:
1616 case ir_unop_bitcast_u2f
:
1617 case ir_unop_bitcast_f2u
:
1618 case ir_unop_bitcast_i642d
:
1619 case ir_unop_bitcast_d2i64
:
1620 case ir_unop_bitcast_u642d
:
1621 case ir_unop_bitcast_d2u64
:
1622 case ir_unop_subroutine_to_int
:
1624 result
= nir_imov(&b
, srcs
[0]);
1626 case ir_unop_trunc
: result
= nir_ftrunc(&b
, srcs
[0]); break;
1627 case ir_unop_ceil
: result
= nir_fceil(&b
, srcs
[0]); break;
1628 case ir_unop_floor
: result
= nir_ffloor(&b
, srcs
[0]); break;
1629 case ir_unop_fract
: result
= nir_ffract(&b
, srcs
[0]); break;
1630 case ir_unop_frexp_exp
: result
= nir_frexp_exp(&b
, srcs
[0]); break;
1631 case ir_unop_frexp_sig
: result
= nir_frexp_sig(&b
, srcs
[0]); break;
1632 case ir_unop_round_even
: result
= nir_fround_even(&b
, srcs
[0]); break;
1633 case ir_unop_sin
: result
= nir_fsin(&b
, srcs
[0]); break;
1634 case ir_unop_cos
: result
= nir_fcos(&b
, srcs
[0]); break;
1635 case ir_unop_dFdx
: result
= nir_fddx(&b
, srcs
[0]); break;
1636 case ir_unop_dFdy
: result
= nir_fddy(&b
, srcs
[0]); break;
1637 case ir_unop_dFdx_fine
: result
= nir_fddx_fine(&b
, srcs
[0]); break;
1638 case ir_unop_dFdy_fine
: result
= nir_fddy_fine(&b
, srcs
[0]); break;
1639 case ir_unop_dFdx_coarse
: result
= nir_fddx_coarse(&b
, srcs
[0]); break;
1640 case ir_unop_dFdy_coarse
: result
= nir_fddy_coarse(&b
, srcs
[0]); break;
1641 case ir_unop_pack_snorm_2x16
:
1642 result
= nir_pack_snorm_2x16(&b
, srcs
[0]);
1644 case ir_unop_pack_snorm_4x8
:
1645 result
= nir_pack_snorm_4x8(&b
, srcs
[0]);
1647 case ir_unop_pack_unorm_2x16
:
1648 result
= nir_pack_unorm_2x16(&b
, srcs
[0]);
1650 case ir_unop_pack_unorm_4x8
:
1651 result
= nir_pack_unorm_4x8(&b
, srcs
[0]);
1653 case ir_unop_pack_half_2x16
:
1654 result
= nir_pack_half_2x16(&b
, srcs
[0]);
1656 case ir_unop_unpack_snorm_2x16
:
1657 result
= nir_unpack_snorm_2x16(&b
, srcs
[0]);
1659 case ir_unop_unpack_snorm_4x8
:
1660 result
= nir_unpack_snorm_4x8(&b
, srcs
[0]);
1662 case ir_unop_unpack_unorm_2x16
:
1663 result
= nir_unpack_unorm_2x16(&b
, srcs
[0]);
1665 case ir_unop_unpack_unorm_4x8
:
1666 result
= nir_unpack_unorm_4x8(&b
, srcs
[0]);
1668 case ir_unop_unpack_half_2x16
:
1669 result
= nir_unpack_half_2x16(&b
, srcs
[0]);
1671 case ir_unop_pack_sampler_2x32
:
1672 case ir_unop_pack_image_2x32
:
1673 case ir_unop_pack_double_2x32
:
1674 case ir_unop_pack_int_2x32
:
1675 case ir_unop_pack_uint_2x32
:
1676 result
= nir_pack_64_2x32(&b
, srcs
[0]);
1678 case ir_unop_unpack_sampler_2x32
:
1679 case ir_unop_unpack_image_2x32
:
1680 case ir_unop_unpack_double_2x32
:
1681 case ir_unop_unpack_int_2x32
:
1682 case ir_unop_unpack_uint_2x32
:
1683 result
= nir_unpack_64_2x32(&b
, srcs
[0]);
1685 case ir_unop_bitfield_reverse
:
1686 result
= nir_bitfield_reverse(&b
, srcs
[0]);
1688 case ir_unop_bit_count
:
1689 result
= nir_bit_count(&b
, srcs
[0]);
1691 case ir_unop_find_msb
:
1693 case GLSL_TYPE_UINT
:
1694 result
= nir_ufind_msb(&b
, srcs
[0]);
1697 result
= nir_ifind_msb(&b
, srcs
[0]);
1700 unreachable("Invalid type for findMSB()");
1703 case ir_unop_find_lsb
:
1704 result
= nir_find_lsb(&b
, srcs
[0]);
1708 switch (ir
->type
->vector_elements
) {
1710 switch (ir
->operands
[0]->type
->vector_elements
) {
1711 case 1: result
= nir_fnoise1_1(&b
, srcs
[0]); break;
1712 case 2: result
= nir_fnoise1_2(&b
, srcs
[0]); break;
1713 case 3: result
= nir_fnoise1_3(&b
, srcs
[0]); break;
1714 case 4: result
= nir_fnoise1_4(&b
, srcs
[0]); break;
1715 default: unreachable("not reached");
1719 switch (ir
->operands
[0]->type
->vector_elements
) {
1720 case 1: result
= nir_fnoise2_1(&b
, srcs
[0]); break;
1721 case 2: result
= nir_fnoise2_2(&b
, srcs
[0]); break;
1722 case 3: result
= nir_fnoise2_3(&b
, srcs
[0]); break;
1723 case 4: result
= nir_fnoise2_4(&b
, srcs
[0]); break;
1724 default: unreachable("not reached");
1728 switch (ir
->operands
[0]->type
->vector_elements
) {
1729 case 1: result
= nir_fnoise3_1(&b
, srcs
[0]); break;
1730 case 2: result
= nir_fnoise3_2(&b
, srcs
[0]); break;
1731 case 3: result
= nir_fnoise3_3(&b
, srcs
[0]); break;
1732 case 4: result
= nir_fnoise3_4(&b
, srcs
[0]); break;
1733 default: unreachable("not reached");
1737 switch (ir
->operands
[0]->type
->vector_elements
) {
1738 case 1: result
= nir_fnoise4_1(&b
, srcs
[0]); break;
1739 case 2: result
= nir_fnoise4_2(&b
, srcs
[0]); break;
1740 case 3: result
= nir_fnoise4_3(&b
, srcs
[0]); break;
1741 case 4: result
= nir_fnoise4_4(&b
, srcs
[0]); break;
1742 default: unreachable("not reached");
1746 unreachable("not reached");
1749 case ir_unop_get_buffer_size
: {
1750 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(
1752 nir_intrinsic_get_buffer_size
);
1753 load
->num_components
= ir
->type
->vector_elements
;
1754 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1755 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1756 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1761 result
= type_is_float(out_type
) ? nir_fadd(&b
, srcs
[0], srcs
[1])
1762 : nir_iadd(&b
, srcs
[0], srcs
[1]);
1765 result
= type_is_float(out_type
) ? nir_fsub(&b
, srcs
[0], srcs
[1])
1766 : nir_isub(&b
, srcs
[0], srcs
[1]);
1769 result
= type_is_float(out_type
) ? nir_fmul(&b
, srcs
[0], srcs
[1])
1770 : nir_imul(&b
, srcs
[0], srcs
[1]);
1773 if (type_is_float(out_type
))
1774 result
= nir_fdiv(&b
, srcs
[0], srcs
[1]);
1775 else if (type_is_signed(out_type
))
1776 result
= nir_idiv(&b
, srcs
[0], srcs
[1]);
1778 result
= nir_udiv(&b
, srcs
[0], srcs
[1]);
1781 result
= type_is_float(out_type
) ? nir_fmod(&b
, srcs
[0], srcs
[1])
1782 : nir_umod(&b
, srcs
[0], srcs
[1]);
1785 if (type_is_float(out_type
))
1786 result
= nir_fmin(&b
, srcs
[0], srcs
[1]);
1787 else if (type_is_signed(out_type
))
1788 result
= nir_imin(&b
, srcs
[0], srcs
[1]);
1790 result
= nir_umin(&b
, srcs
[0], srcs
[1]);
1793 if (type_is_float(out_type
))
1794 result
= nir_fmax(&b
, srcs
[0], srcs
[1]);
1795 else if (type_is_signed(out_type
))
1796 result
= nir_imax(&b
, srcs
[0], srcs
[1]);
1798 result
= nir_umax(&b
, srcs
[0], srcs
[1]);
1800 case ir_binop_pow
: result
= nir_fpow(&b
, srcs
[0], srcs
[1]); break;
1801 case ir_binop_bit_and
: result
= nir_iand(&b
, srcs
[0], srcs
[1]); break;
1802 case ir_binop_bit_or
: result
= nir_ior(&b
, srcs
[0], srcs
[1]); break;
1803 case ir_binop_bit_xor
: result
= nir_ixor(&b
, srcs
[0], srcs
[1]); break;
1804 case ir_binop_logic_and
:
1805 result
= supports_ints
? nir_iand(&b
, srcs
[0], srcs
[1])
1806 : nir_fand(&b
, srcs
[0], srcs
[1]);
1808 case ir_binop_logic_or
:
1809 result
= supports_ints
? nir_ior(&b
, srcs
[0], srcs
[1])
1810 : nir_for(&b
, srcs
[0], srcs
[1]);
1812 case ir_binop_logic_xor
:
1813 result
= supports_ints
? nir_ixor(&b
, srcs
[0], srcs
[1])
1814 : nir_fxor(&b
, srcs
[0], srcs
[1]);
1816 case ir_binop_lshift
: result
= nir_ishl(&b
, srcs
[0], srcs
[1]); break;
1817 case ir_binop_rshift
:
1818 result
= (type_is_signed(out_type
)) ? nir_ishr(&b
, srcs
[0], srcs
[1])
1819 : nir_ushr(&b
, srcs
[0], srcs
[1]);
1821 case ir_binop_imul_high
:
1822 result
= (out_type
== GLSL_TYPE_INT
) ? nir_imul_high(&b
, srcs
[0], srcs
[1])
1823 : nir_umul_high(&b
, srcs
[0], srcs
[1]);
1825 case ir_binop_carry
: result
= nir_uadd_carry(&b
, srcs
[0], srcs
[1]); break;
1826 case ir_binop_borrow
: result
= nir_usub_borrow(&b
, srcs
[0], srcs
[1]); break;
1828 if (supports_ints
) {
1829 if (type_is_float(types
[0]))
1830 result
= nir_flt(&b
, srcs
[0], srcs
[1]);
1831 else if (type_is_signed(types
[0]))
1832 result
= nir_ilt(&b
, srcs
[0], srcs
[1]);
1834 result
= nir_ult(&b
, srcs
[0], srcs
[1]);
1836 result
= nir_slt(&b
, srcs
[0], srcs
[1]);
1839 case ir_binop_gequal
:
1840 if (supports_ints
) {
1841 if (type_is_float(types
[0]))
1842 result
= nir_fge(&b
, srcs
[0], srcs
[1]);
1843 else if (type_is_signed(types
[0]))
1844 result
= nir_ige(&b
, srcs
[0], srcs
[1]);
1846 result
= nir_uge(&b
, srcs
[0], srcs
[1]);
1848 result
= nir_sge(&b
, srcs
[0], srcs
[1]);
1851 case ir_binop_equal
:
1852 if (supports_ints
) {
1853 if (type_is_float(types
[0]))
1854 result
= nir_feq(&b
, srcs
[0], srcs
[1]);
1856 result
= nir_ieq(&b
, srcs
[0], srcs
[1]);
1858 result
= nir_seq(&b
, srcs
[0], srcs
[1]);
1861 case ir_binop_nequal
:
1862 if (supports_ints
) {
1863 if (type_is_float(types
[0]))
1864 result
= nir_fne(&b
, srcs
[0], srcs
[1]);
1866 result
= nir_ine(&b
, srcs
[0], srcs
[1]);
1868 result
= nir_sne(&b
, srcs
[0], srcs
[1]);
1871 case ir_binop_all_equal
:
1872 if (supports_ints
) {
1873 if (type_is_float(types
[0])) {
1874 switch (ir
->operands
[0]->type
->vector_elements
) {
1875 case 1: result
= nir_feq(&b
, srcs
[0], srcs
[1]); break;
1876 case 2: result
= nir_ball_fequal2(&b
, srcs
[0], srcs
[1]); break;
1877 case 3: result
= nir_ball_fequal3(&b
, srcs
[0], srcs
[1]); break;
1878 case 4: result
= nir_ball_fequal4(&b
, srcs
[0], srcs
[1]); break;
1880 unreachable("not reached");
1883 switch (ir
->operands
[0]->type
->vector_elements
) {
1884 case 1: result
= nir_ieq(&b
, srcs
[0], srcs
[1]); break;
1885 case 2: result
= nir_ball_iequal2(&b
, srcs
[0], srcs
[1]); break;
1886 case 3: result
= nir_ball_iequal3(&b
, srcs
[0], srcs
[1]); break;
1887 case 4: result
= nir_ball_iequal4(&b
, srcs
[0], srcs
[1]); break;
1889 unreachable("not reached");
1893 switch (ir
->operands
[0]->type
->vector_elements
) {
1894 case 1: result
= nir_seq(&b
, srcs
[0], srcs
[1]); break;
1895 case 2: result
= nir_fall_equal2(&b
, srcs
[0], srcs
[1]); break;
1896 case 3: result
= nir_fall_equal3(&b
, srcs
[0], srcs
[1]); break;
1897 case 4: result
= nir_fall_equal4(&b
, srcs
[0], srcs
[1]); break;
1899 unreachable("not reached");
1903 case ir_binop_any_nequal
:
1904 if (supports_ints
) {
1905 if (type_is_float(types
[0])) {
1906 switch (ir
->operands
[0]->type
->vector_elements
) {
1907 case 1: result
= nir_fne(&b
, srcs
[0], srcs
[1]); break;
1908 case 2: result
= nir_bany_fnequal2(&b
, srcs
[0], srcs
[1]); break;
1909 case 3: result
= nir_bany_fnequal3(&b
, srcs
[0], srcs
[1]); break;
1910 case 4: result
= nir_bany_fnequal4(&b
, srcs
[0], srcs
[1]); break;
1912 unreachable("not reached");
1915 switch (ir
->operands
[0]->type
->vector_elements
) {
1916 case 1: result
= nir_ine(&b
, srcs
[0], srcs
[1]); break;
1917 case 2: result
= nir_bany_inequal2(&b
, srcs
[0], srcs
[1]); break;
1918 case 3: result
= nir_bany_inequal3(&b
, srcs
[0], srcs
[1]); break;
1919 case 4: result
= nir_bany_inequal4(&b
, srcs
[0], srcs
[1]); break;
1921 unreachable("not reached");
1925 switch (ir
->operands
[0]->type
->vector_elements
) {
1926 case 1: result
= nir_sne(&b
, srcs
[0], srcs
[1]); break;
1927 case 2: result
= nir_fany_nequal2(&b
, srcs
[0], srcs
[1]); break;
1928 case 3: result
= nir_fany_nequal3(&b
, srcs
[0], srcs
[1]); break;
1929 case 4: result
= nir_fany_nequal4(&b
, srcs
[0], srcs
[1]); break;
1931 unreachable("not reached");
1936 switch (ir
->operands
[0]->type
->vector_elements
) {
1937 case 2: result
= nir_fdot2(&b
, srcs
[0], srcs
[1]); break;
1938 case 3: result
= nir_fdot3(&b
, srcs
[0], srcs
[1]); break;
1939 case 4: result
= nir_fdot4(&b
, srcs
[0], srcs
[1]); break;
1941 unreachable("not reached");
1944 case ir_binop_vector_extract
: {
1945 result
= nir_channel(&b
, srcs
[0], 0);
1946 for (unsigned i
= 1; i
< ir
->operands
[0]->type
->vector_elements
; i
++) {
1947 nir_ssa_def
*swizzled
= nir_channel(&b
, srcs
[0], i
);
1948 result
= nir_bcsel(&b
, nir_ieq(&b
, srcs
[1], nir_imm_int(&b
, i
)),
1954 case ir_binop_ldexp
: result
= nir_ldexp(&b
, srcs
[0], srcs
[1]); break;
1956 result
= nir_ffma(&b
, srcs
[0], srcs
[1], srcs
[2]);
1959 result
= nir_flrp(&b
, srcs
[0], srcs
[1], srcs
[2]);
1963 result
= nir_bcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
1965 result
= nir_fcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
1967 case ir_triop_bitfield_extract
:
1968 result
= (out_type
== GLSL_TYPE_INT
) ?
1969 nir_ibitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]) :
1970 nir_ubitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]);
1972 case ir_quadop_bitfield_insert
:
1973 result
= nir_bitfield_insert(&b
, srcs
[0], srcs
[1], srcs
[2], srcs
[3]);
1975 case ir_quadop_vector
:
1976 result
= nir_vec(&b
, srcs
, ir
->type
->vector_elements
);
1980 unreachable("not reached");
1985 nir_visitor::visit(ir_swizzle
*ir
)
1987 unsigned swizzle
[4] = { ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
};
1988 result
= nir_swizzle(&b
, evaluate_rvalue(ir
->val
), swizzle
,
1989 ir
->type
->vector_elements
, !supports_ints
);
1993 nir_visitor::visit(ir_texture
*ir
)
2000 num_srcs
= 1; /* coordinate */
2005 op
= (ir
->op
== ir_txb
) ? nir_texop_txb
: nir_texop_txl
;
2006 num_srcs
= 2; /* coordinate, bias/lod */
2010 op
= nir_texop_txd
; /* coordinate, dPdx, dPdy */
2016 if (ir
->lod_info
.lod
!= NULL
)
2017 num_srcs
= 2; /* coordinate, lod */
2019 num_srcs
= 1; /* coordinate */
2023 op
= nir_texop_txf_ms
;
2024 num_srcs
= 2; /* coordinate, sample_index */
2029 if (ir
->lod_info
.lod
!= NULL
)
2030 num_srcs
= 1; /* lod */
2037 num_srcs
= 1; /* coordinate */
2042 num_srcs
= 1; /* coordinate */
2045 case ir_query_levels
:
2046 op
= nir_texop_query_levels
;
2050 case ir_texture_samples
:
2051 op
= nir_texop_texture_samples
;
2055 case ir_samples_identical
:
2056 op
= nir_texop_samples_identical
;
2057 num_srcs
= 1; /* coordinate */
2061 unreachable("not reached");
2064 if (ir
->projector
!= NULL
)
2066 if (ir
->shadow_comparator
!= NULL
)
2068 if (ir
->offset
!= NULL
)
2071 nir_tex_instr
*instr
= nir_tex_instr_create(this->shader
, num_srcs
);
2074 instr
->sampler_dim
=
2075 (glsl_sampler_dim
) ir
->sampler
->type
->sampler_dimensionality
;
2076 instr
->is_array
= ir
->sampler
->type
->sampler_array
;
2077 instr
->is_shadow
= ir
->sampler
->type
->sampler_shadow
;
2078 if (instr
->is_shadow
)
2079 instr
->is_new_style_shadow
= (ir
->type
->vector_elements
== 1);
2080 switch (ir
->type
->base_type
) {
2081 case GLSL_TYPE_FLOAT
:
2082 instr
->dest_type
= nir_type_float
;
2085 instr
->dest_type
= nir_type_int
;
2087 case GLSL_TYPE_BOOL
:
2088 case GLSL_TYPE_UINT
:
2089 instr
->dest_type
= nir_type_uint
;
2092 unreachable("not reached");
2095 instr
->texture
= evaluate_deref(&instr
->instr
, ir
->sampler
);
2097 unsigned src_number
= 0;
2099 if (ir
->coordinate
!= NULL
) {
2100 instr
->coord_components
= ir
->coordinate
->type
->vector_elements
;
2101 instr
->src
[src_number
].src
=
2102 nir_src_for_ssa(evaluate_rvalue(ir
->coordinate
));
2103 instr
->src
[src_number
].src_type
= nir_tex_src_coord
;
2107 if (ir
->projector
!= NULL
) {
2108 instr
->src
[src_number
].src
=
2109 nir_src_for_ssa(evaluate_rvalue(ir
->projector
));
2110 instr
->src
[src_number
].src_type
= nir_tex_src_projector
;
2114 if (ir
->shadow_comparator
!= NULL
) {
2115 instr
->src
[src_number
].src
=
2116 nir_src_for_ssa(evaluate_rvalue(ir
->shadow_comparator
));
2117 instr
->src
[src_number
].src_type
= nir_tex_src_comparator
;
2121 if (ir
->offset
!= NULL
) {
2122 /* we don't support multiple offsets yet */
2123 assert(ir
->offset
->type
->is_vector() || ir
->offset
->type
->is_scalar());
2125 instr
->src
[src_number
].src
=
2126 nir_src_for_ssa(evaluate_rvalue(ir
->offset
));
2127 instr
->src
[src_number
].src_type
= nir_tex_src_offset
;
2133 instr
->src
[src_number
].src
=
2134 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.bias
));
2135 instr
->src
[src_number
].src_type
= nir_tex_src_bias
;
2142 if (ir
->lod_info
.lod
!= NULL
) {
2143 instr
->src
[src_number
].src
=
2144 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.lod
));
2145 instr
->src
[src_number
].src_type
= nir_tex_src_lod
;
2151 instr
->src
[src_number
].src
=
2152 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdx
));
2153 instr
->src
[src_number
].src_type
= nir_tex_src_ddx
;
2155 instr
->src
[src_number
].src
=
2156 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdy
));
2157 instr
->src
[src_number
].src_type
= nir_tex_src_ddy
;
2162 instr
->src
[src_number
].src
=
2163 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.sample_index
));
2164 instr
->src
[src_number
].src_type
= nir_tex_src_ms_index
;
2169 instr
->component
= ir
->lod_info
.component
->as_constant()->value
.u
[0];
2176 assert(src_number
== num_srcs
);
2178 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2179 add_instr(&instr
->instr
, nir_tex_instr_dest_size(instr
), bit_size
);
2183 nir_visitor::visit(ir_constant
*ir
)
2186 * We don't know if this variable is an array or struct that gets
2187 * dereferenced, so do the safe thing an make it a variable with a
2188 * constant initializer and return a dereference.
2192 nir_local_variable_create(this->impl
, ir
->type
, "const_temp");
2193 var
->data
.read_only
= true;
2194 var
->constant_initializer
= constant_copy(ir
, var
);
2196 this->deref_head
= nir_deref_var_create(this->shader
, var
);
2197 this->deref_tail
= &this->deref_head
->deref
;
2201 nir_visitor::visit(ir_dereference_variable
*ir
)
2203 struct hash_entry
*entry
=
2204 _mesa_hash_table_search(this->var_table
, ir
->var
);
2206 nir_variable
*var
= (nir_variable
*) entry
->data
;
2208 nir_deref_var
*deref
= nir_deref_var_create(this->shader
, var
);
2209 this->deref_head
= deref
;
2210 this->deref_tail
= &deref
->deref
;
2214 nir_visitor::visit(ir_dereference_record
*ir
)
2216 ir
->record
->accept(this);
2218 int field_index
= ir
->field_idx
;
2219 assert(field_index
>= 0);
2221 nir_deref_struct
*deref
= nir_deref_struct_create(this->deref_tail
, field_index
);
2222 deref
->deref
.type
= ir
->type
;
2223 this->deref_tail
->child
= &deref
->deref
;
2224 this->deref_tail
= &deref
->deref
;
2228 nir_visitor::visit(ir_dereference_array
*ir
)
2230 nir_deref_array
*deref
= nir_deref_array_create(this->shader
);
2231 deref
->deref
.type
= ir
->type
;
2233 ir_constant
*const_index
= ir
->array_index
->as_constant();
2234 if (const_index
!= NULL
) {
2235 deref
->deref_array_type
= nir_deref_array_type_direct
;
2236 deref
->base_offset
= const_index
->value
.u
[0];
2238 deref
->deref_array_type
= nir_deref_array_type_indirect
;
2240 nir_src_for_ssa(evaluate_rvalue(ir
->array_index
));
2243 ir
->array
->accept(this);
2245 this->deref_tail
->child
= &deref
->deref
;
2246 ralloc_steal(this->deref_tail
, deref
);
2247 this->deref_tail
= &deref
->deref
;
2251 nir_visitor::visit(ir_barrier
*)
2253 nir_intrinsic_instr
*instr
=
2254 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_barrier
);
2255 nir_builder_instr_insert(&b
, &instr
->instr
);