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"
36 #include "util/u_math.h"
39 * pass to lower GLSL IR to NIR
41 * This will lower variable dereferences to loads/stores of corresponding
42 * variables in NIR - the variables will be converted to registers in a later
48 class nir_visitor
: public ir_visitor
51 nir_visitor(nir_shader
*shader
);
54 virtual void visit(ir_variable
*);
55 virtual void visit(ir_function
*);
56 virtual void visit(ir_function_signature
*);
57 virtual void visit(ir_loop
*);
58 virtual void visit(ir_if
*);
59 virtual void visit(ir_discard
*);
60 virtual void visit(ir_loop_jump
*);
61 virtual void visit(ir_return
*);
62 virtual void visit(ir_call
*);
63 virtual void visit(ir_assignment
*);
64 virtual void visit(ir_emit_vertex
*);
65 virtual void visit(ir_end_primitive
*);
66 virtual void visit(ir_expression
*);
67 virtual void visit(ir_swizzle
*);
68 virtual void visit(ir_texture
*);
69 virtual void visit(ir_constant
*);
70 virtual void visit(ir_dereference_variable
*);
71 virtual void visit(ir_dereference_record
*);
72 virtual void visit(ir_dereference_array
*);
73 virtual void visit(ir_barrier
*);
75 void create_function(ir_function_signature
*ir
);
78 void add_instr(nir_instr
*instr
, unsigned num_components
, unsigned bit_size
);
79 nir_ssa_def
*evaluate_rvalue(ir_rvalue
*ir
);
81 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
**srcs
);
82 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
,
85 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
,
86 nir_ssa_def
*src2
, nir_ssa_def
*src3
);
91 nir_function_impl
*impl
;
93 nir_ssa_def
*result
; /* result of the expression tree last visited */
95 nir_deref_instr
*evaluate_deref(ir_instruction
*ir
);
97 /* most recent deref instruction created */
98 nir_deref_instr
*deref
;
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 glsl_to_nir(const struct gl_shader_program
*shader_prog
,
134 gl_shader_stage stage
,
135 const nir_shader_compiler_options
*options
)
137 struct gl_linked_shader
*sh
= shader_prog
->_LinkedShaders
[stage
];
139 nir_shader
*shader
= nir_shader_create(NULL
, stage
, options
,
142 nir_visitor
v1(shader
);
143 nir_function_visitor
v2(&v1
);
145 visit_exec_list(sh
->ir
, &v1
);
147 nir_lower_constant_initializers(shader
, (nir_variable_mode
)~0);
149 /* Remap the locations to slots so those requiring two slots will occupy
150 * two locations. For instance, if we have in the IR code a dvec3 attr0 in
151 * location 0 and vec4 attr1 in location 1, in NIR attr0 will use
152 * locations/slots 0 and 1, and attr1 will use location/slot 2 */
153 if (shader
->info
.stage
== MESA_SHADER_VERTEX
)
154 nir_remap_dual_slot_attributes(shader
, &sh
->Program
->DualSlotInputs
);
156 shader
->info
.name
= ralloc_asprintf(shader
, "GLSL%d", shader_prog
->Name
);
157 if (shader_prog
->Label
)
158 shader
->info
.label
= ralloc_strdup(shader
, shader_prog
->Label
);
160 /* Check for transform feedback varyings specified via the API */
161 shader
->info
.has_transform_feedback_varyings
=
162 shader_prog
->TransformFeedback
.NumVarying
> 0;
164 /* Check for transform feedback varyings specified in the Shader */
165 if (shader_prog
->last_vert_prog
)
166 shader
->info
.has_transform_feedback_varyings
|=
167 shader_prog
->last_vert_prog
->sh
.LinkedTransformFeedback
->NumVarying
> 0;
172 nir_visitor::nir_visitor(nir_shader
*shader
)
174 this->supports_ints
= shader
->options
->native_integers
;
175 this->shader
= shader
;
176 this->is_global
= true;
177 this->var_table
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
178 _mesa_key_pointer_equal
);
179 this->overload_table
= _mesa_hash_table_create(NULL
, _mesa_hash_pointer
,
180 _mesa_key_pointer_equal
);
184 memset(&this->b
, 0, sizeof(this->b
));
187 nir_visitor::~nir_visitor()
189 _mesa_hash_table_destroy(this->var_table
, NULL
);
190 _mesa_hash_table_destroy(this->overload_table
, NULL
);
194 nir_visitor::evaluate_deref(ir_instruction
*ir
)
200 static nir_constant
*
201 constant_copy(ir_constant
*ir
, void *mem_ctx
)
206 nir_constant
*ret
= rzalloc(mem_ctx
, nir_constant
);
208 const unsigned rows
= ir
->type
->vector_elements
;
209 const unsigned cols
= ir
->type
->matrix_columns
;
212 ret
->num_elements
= 0;
213 switch (ir
->type
->base_type
) {
215 /* Only float base types can be matrices. */
218 for (unsigned r
= 0; r
< rows
; r
++)
219 ret
->values
[0].u32
[r
] = ir
->value
.u
[r
];
224 /* Only float base types can be matrices. */
227 for (unsigned r
= 0; r
< rows
; r
++)
228 ret
->values
[0].i32
[r
] = ir
->value
.i
[r
];
232 case GLSL_TYPE_FLOAT
:
233 for (unsigned c
= 0; c
< cols
; c
++) {
234 for (unsigned r
= 0; r
< rows
; r
++)
235 ret
->values
[c
].f32
[r
] = ir
->value
.f
[c
* rows
+ r
];
239 case GLSL_TYPE_DOUBLE
:
240 for (unsigned c
= 0; c
< cols
; c
++) {
241 for (unsigned r
= 0; r
< rows
; r
++)
242 ret
->values
[c
].f64
[r
] = ir
->value
.d
[c
* rows
+ r
];
246 case GLSL_TYPE_UINT64
:
247 /* Only float base types can be matrices. */
250 for (unsigned r
= 0; r
< rows
; r
++)
251 ret
->values
[0].u64
[r
] = ir
->value
.u64
[r
];
254 case GLSL_TYPE_INT64
:
255 /* Only float base types can be matrices. */
258 for (unsigned r
= 0; r
< rows
; r
++)
259 ret
->values
[0].i64
[r
] = ir
->value
.i64
[r
];
263 /* Only float base types can be matrices. */
266 for (unsigned r
= 0; r
< rows
; r
++)
267 ret
->values
[0].u32
[r
] = ir
->value
.b
[r
] ? NIR_TRUE
: NIR_FALSE
;
271 case GLSL_TYPE_STRUCT
:
272 case GLSL_TYPE_ARRAY
:
273 ret
->elements
= ralloc_array(mem_ctx
, nir_constant
*,
275 ret
->num_elements
= ir
->type
->length
;
277 for (i
= 0; i
< ir
->type
->length
; i
++)
278 ret
->elements
[i
] = constant_copy(ir
->const_elements
[i
], mem_ctx
);
282 unreachable("not reached");
289 nir_visitor::visit(ir_variable
*ir
)
291 /* TODO: In future we should switch to using the NIR lowering pass but for
292 * now just ignore these variables as GLSL IR should have lowered them.
293 * Anything remaining are just dead vars that weren't cleaned up.
295 if (ir
->data
.mode
== ir_var_shader_shared
)
298 nir_variable
*var
= rzalloc(shader
, nir_variable
);
299 var
->type
= ir
->type
;
300 var
->name
= ralloc_strdup(var
, ir
->name
);
302 var
->data
.always_active_io
= ir
->data
.always_active_io
;
303 var
->data
.read_only
= ir
->data
.read_only
;
304 var
->data
.centroid
= ir
->data
.centroid
;
305 var
->data
.sample
= ir
->data
.sample
;
306 var
->data
.patch
= ir
->data
.patch
;
307 var
->data
.invariant
= ir
->data
.invariant
;
308 var
->data
.location
= ir
->data
.location
;
309 var
->data
.stream
= ir
->data
.stream
;
310 var
->data
.compact
= false;
312 switch(ir
->data
.mode
) {
314 case ir_var_temporary
:
316 var
->data
.mode
= nir_var_global
;
318 var
->data
.mode
= nir_var_local
;
321 case ir_var_function_in
:
322 case ir_var_function_out
:
323 case ir_var_function_inout
:
324 case ir_var_const_in
:
325 var
->data
.mode
= nir_var_local
;
328 case ir_var_shader_in
:
329 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
330 ir
->data
.location
== VARYING_SLOT_FACE
) {
331 /* For whatever reason, GLSL IR makes gl_FrontFacing an input */
332 var
->data
.location
= SYSTEM_VALUE_FRONT_FACE
;
333 var
->data
.mode
= nir_var_system_value
;
334 } else if (shader
->info
.stage
== MESA_SHADER_GEOMETRY
&&
335 ir
->data
.location
== VARYING_SLOT_PRIMITIVE_ID
) {
336 /* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */
337 var
->data
.location
= SYSTEM_VALUE_PRIMITIVE_ID
;
338 var
->data
.mode
= nir_var_system_value
;
340 var
->data
.mode
= nir_var_shader_in
;
342 if (shader
->info
.stage
== MESA_SHADER_TESS_EVAL
&&
343 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
344 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
345 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
350 case ir_var_shader_out
:
351 var
->data
.mode
= nir_var_shader_out
;
352 if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
&&
353 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
354 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
355 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
360 var
->data
.mode
= nir_var_uniform
;
363 case ir_var_shader_storage
:
364 var
->data
.mode
= nir_var_shader_storage
;
367 case ir_var_system_value
:
368 var
->data
.mode
= nir_var_system_value
;
372 unreachable("not reached");
375 var
->data
.interpolation
= ir
->data
.interpolation
;
376 var
->data
.origin_upper_left
= ir
->data
.origin_upper_left
;
377 var
->data
.pixel_center_integer
= ir
->data
.pixel_center_integer
;
378 var
->data
.location_frac
= ir
->data
.location_frac
;
380 if (var
->data
.pixel_center_integer
) {
381 assert(shader
->info
.stage
== MESA_SHADER_FRAGMENT
);
382 shader
->info
.fs
.pixel_center_integer
= true;
385 switch (ir
->data
.depth_layout
) {
386 case ir_depth_layout_none
:
387 var
->data
.depth_layout
= nir_depth_layout_none
;
389 case ir_depth_layout_any
:
390 var
->data
.depth_layout
= nir_depth_layout_any
;
392 case ir_depth_layout_greater
:
393 var
->data
.depth_layout
= nir_depth_layout_greater
;
395 case ir_depth_layout_less
:
396 var
->data
.depth_layout
= nir_depth_layout_less
;
398 case ir_depth_layout_unchanged
:
399 var
->data
.depth_layout
= nir_depth_layout_unchanged
;
402 unreachable("not reached");
405 var
->data
.index
= ir
->data
.index
;
406 var
->data
.descriptor_set
= 0;
407 var
->data
.binding
= ir
->data
.binding
;
408 var
->data
.explicit_binding
= ir
->data
.explicit_binding
;
409 var
->data
.bindless
= ir
->data
.bindless
;
410 var
->data
.offset
= ir
->data
.offset
;
412 unsigned image_access
= 0;
413 if (ir
->data
.memory_read_only
)
414 image_access
|= ACCESS_NON_WRITEABLE
;
415 if (ir
->data
.memory_write_only
)
416 image_access
|= ACCESS_NON_READABLE
;
417 if (ir
->data
.memory_coherent
)
418 image_access
|= ACCESS_COHERENT
;
419 if (ir
->data
.memory_volatile
)
420 image_access
|= ACCESS_VOLATILE
;
421 if (ir
->data
.memory_restrict
)
422 image_access
|= ACCESS_RESTRICT
;
423 var
->data
.image
.access
= (gl_access_qualifier
)image_access
;
424 var
->data
.image
.format
= ir
->data
.image_format
;
426 var
->data
.fb_fetch_output
= ir
->data
.fb_fetch_output
;
427 var
->data
.explicit_xfb_buffer
= ir
->data
.explicit_xfb_buffer
;
428 var
->data
.explicit_xfb_stride
= ir
->data
.explicit_xfb_stride
;
429 var
->data
.xfb_buffer
= ir
->data
.xfb_buffer
;
430 var
->data
.xfb_stride
= ir
->data
.xfb_stride
;
432 var
->num_state_slots
= ir
->get_num_state_slots();
433 if (var
->num_state_slots
> 0) {
434 var
->state_slots
= rzalloc_array(var
, nir_state_slot
,
435 var
->num_state_slots
);
437 ir_state_slot
*state_slots
= ir
->get_state_slots();
438 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
439 for (unsigned j
= 0; j
< 5; j
++)
440 var
->state_slots
[i
].tokens
[j
] = state_slots
[i
].tokens
[j
];
441 var
->state_slots
[i
].swizzle
= state_slots
[i
].swizzle
;
444 var
->state_slots
= NULL
;
447 var
->constant_initializer
= constant_copy(ir
->constant_initializer
, var
);
449 var
->interface_type
= ir
->get_interface_type();
451 if (var
->data
.mode
== nir_var_local
)
452 nir_function_impl_add_variable(impl
, var
);
454 nir_shader_add_variable(shader
, var
);
456 _mesa_hash_table_insert(var_table
, ir
, var
);
461 nir_function_visitor::visit_enter(ir_function
*ir
)
463 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
) {
464 visitor
->create_function(sig
);
466 return visit_continue_with_parent
;
470 nir_visitor::create_function(ir_function_signature
*ir
)
472 if (ir
->is_intrinsic())
475 nir_function
*func
= nir_function_create(shader
, ir
->function_name());
477 assert(ir
->parameters
.is_empty());
478 assert(ir
->return_type
== glsl_type::void_type
);
480 _mesa_hash_table_insert(this->overload_table
, ir
, func
);
484 nir_visitor::visit(ir_function
*ir
)
486 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
)
491 nir_visitor::visit(ir_function_signature
*ir
)
493 if (ir
->is_intrinsic())
496 struct hash_entry
*entry
=
497 _mesa_hash_table_search(this->overload_table
, ir
);
500 nir_function
*func
= (nir_function
*) entry
->data
;
502 if (ir
->is_defined
) {
503 nir_function_impl
*impl
= nir_function_impl_create(func
);
506 assert(strcmp(func
->name
, "main") == 0);
507 assert(ir
->parameters
.is_empty());
509 this->is_global
= false;
511 nir_builder_init(&b
, impl
);
512 b
.cursor
= nir_after_cf_list(&impl
->body
);
513 visit_exec_list(&ir
->body
, this);
515 this->is_global
= true;
522 nir_visitor::visit(ir_loop
*ir
)
525 visit_exec_list(&ir
->body_instructions
, this);
526 nir_pop_loop(&b
, NULL
);
530 nir_visitor::visit(ir_if
*ir
)
532 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
533 visit_exec_list(&ir
->then_instructions
, this);
534 nir_push_else(&b
, NULL
);
535 visit_exec_list(&ir
->else_instructions
, this);
536 nir_pop_if(&b
, NULL
);
540 nir_visitor::visit(ir_discard
*ir
)
543 * discards aren't treated as control flow, because before we lower them
544 * they can appear anywhere in the shader and the stuff after them may still
545 * be executed (yay, crazy GLSL rules!). However, after lowering, all the
546 * discards will be immediately followed by a return.
549 nir_intrinsic_instr
*discard
;
551 discard
= nir_intrinsic_instr_create(this->shader
,
552 nir_intrinsic_discard_if
);
554 nir_src_for_ssa(evaluate_rvalue(ir
->condition
));
556 discard
= nir_intrinsic_instr_create(this->shader
, nir_intrinsic_discard
);
559 nir_builder_instr_insert(&b
, &discard
->instr
);
563 nir_visitor::visit(ir_emit_vertex
*ir
)
565 nir_intrinsic_instr
*instr
=
566 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_emit_vertex
);
567 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
568 nir_builder_instr_insert(&b
, &instr
->instr
);
572 nir_visitor::visit(ir_end_primitive
*ir
)
574 nir_intrinsic_instr
*instr
=
575 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_end_primitive
);
576 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
577 nir_builder_instr_insert(&b
, &instr
->instr
);
581 nir_visitor::visit(ir_loop_jump
*ir
)
585 case ir_loop_jump::jump_break
:
586 type
= nir_jump_break
;
588 case ir_loop_jump::jump_continue
:
589 type
= nir_jump_continue
;
592 unreachable("not reached");
595 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, type
);
596 nir_builder_instr_insert(&b
, &instr
->instr
);
600 nir_visitor::visit(ir_return
*ir
)
602 assert(ir
->value
== NULL
);
603 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, nir_jump_return
);
604 nir_builder_instr_insert(&b
, &instr
->instr
);
608 intrinsic_set_std430_align(nir_intrinsic_instr
*intrin
, const glsl_type
*type
)
610 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
611 unsigned pow2_components
= util_next_power_of_two(type
->vector_elements
);
612 nir_intrinsic_set_align(intrin
, (bit_size
/ 8) * pow2_components
, 0);
616 nir_visitor::visit(ir_call
*ir
)
618 if (ir
->callee
->is_intrinsic()) {
621 switch (ir
->callee
->intrinsic_id
) {
622 case ir_intrinsic_atomic_counter_read
:
623 op
= nir_intrinsic_atomic_counter_read_deref
;
625 case ir_intrinsic_atomic_counter_increment
:
626 op
= nir_intrinsic_atomic_counter_inc_deref
;
628 case ir_intrinsic_atomic_counter_predecrement
:
629 op
= nir_intrinsic_atomic_counter_pre_dec_deref
;
631 case ir_intrinsic_atomic_counter_add
:
632 op
= nir_intrinsic_atomic_counter_add_deref
;
634 case ir_intrinsic_atomic_counter_and
:
635 op
= nir_intrinsic_atomic_counter_and_deref
;
637 case ir_intrinsic_atomic_counter_or
:
638 op
= nir_intrinsic_atomic_counter_or_deref
;
640 case ir_intrinsic_atomic_counter_xor
:
641 op
= nir_intrinsic_atomic_counter_xor_deref
;
643 case ir_intrinsic_atomic_counter_min
:
644 op
= nir_intrinsic_atomic_counter_min_deref
;
646 case ir_intrinsic_atomic_counter_max
:
647 op
= nir_intrinsic_atomic_counter_max_deref
;
649 case ir_intrinsic_atomic_counter_exchange
:
650 op
= nir_intrinsic_atomic_counter_exchange_deref
;
652 case ir_intrinsic_atomic_counter_comp_swap
:
653 op
= nir_intrinsic_atomic_counter_comp_swap_deref
;
655 case ir_intrinsic_image_load
:
656 op
= nir_intrinsic_image_deref_load
;
658 case ir_intrinsic_image_store
:
659 op
= nir_intrinsic_image_deref_store
;
661 case ir_intrinsic_image_atomic_add
:
662 op
= ir
->return_deref
->type
->is_integer_32_64()
663 ? nir_intrinsic_image_deref_atomic_add
664 : nir_intrinsic_image_deref_atomic_fadd
;
666 case ir_intrinsic_image_atomic_min
:
667 op
= nir_intrinsic_image_deref_atomic_min
;
669 case ir_intrinsic_image_atomic_max
:
670 op
= nir_intrinsic_image_deref_atomic_max
;
672 case ir_intrinsic_image_atomic_and
:
673 op
= nir_intrinsic_image_deref_atomic_and
;
675 case ir_intrinsic_image_atomic_or
:
676 op
= nir_intrinsic_image_deref_atomic_or
;
678 case ir_intrinsic_image_atomic_xor
:
679 op
= nir_intrinsic_image_deref_atomic_xor
;
681 case ir_intrinsic_image_atomic_exchange
:
682 op
= nir_intrinsic_image_deref_atomic_exchange
;
684 case ir_intrinsic_image_atomic_comp_swap
:
685 op
= nir_intrinsic_image_deref_atomic_comp_swap
;
687 case ir_intrinsic_memory_barrier
:
688 op
= nir_intrinsic_memory_barrier
;
690 case ir_intrinsic_image_size
:
691 op
= nir_intrinsic_image_deref_size
;
693 case ir_intrinsic_image_samples
:
694 op
= nir_intrinsic_image_deref_samples
;
696 case ir_intrinsic_ssbo_store
:
697 op
= nir_intrinsic_store_ssbo
;
699 case ir_intrinsic_ssbo_load
:
700 op
= nir_intrinsic_load_ssbo
;
702 case ir_intrinsic_ssbo_atomic_add
:
703 op
= ir
->return_deref
->type
->is_integer_32_64()
704 ? nir_intrinsic_ssbo_atomic_add
: nir_intrinsic_ssbo_atomic_fadd
;
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
;
721 else if (ir
->return_deref
->type
== glsl_type::float_type
)
722 op
= nir_intrinsic_ssbo_atomic_fmin
;
724 unreachable("Invalid type");
726 case ir_intrinsic_ssbo_atomic_max
:
727 assert(ir
->return_deref
);
728 if (ir
->return_deref
->type
== glsl_type::int_type
)
729 op
= nir_intrinsic_ssbo_atomic_imax
;
730 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
731 op
= nir_intrinsic_ssbo_atomic_umax
;
732 else if (ir
->return_deref
->type
== glsl_type::float_type
)
733 op
= nir_intrinsic_ssbo_atomic_fmax
;
735 unreachable("Invalid type");
737 case ir_intrinsic_ssbo_atomic_exchange
:
738 op
= nir_intrinsic_ssbo_atomic_exchange
;
740 case ir_intrinsic_ssbo_atomic_comp_swap
:
741 op
= ir
->return_deref
->type
->is_integer_32_64()
742 ? nir_intrinsic_ssbo_atomic_comp_swap
743 : nir_intrinsic_ssbo_atomic_fcomp_swap
;
745 case ir_intrinsic_shader_clock
:
746 op
= nir_intrinsic_shader_clock
;
748 case ir_intrinsic_begin_invocation_interlock
:
749 op
= nir_intrinsic_begin_invocation_interlock
;
751 case ir_intrinsic_end_invocation_interlock
:
752 op
= nir_intrinsic_end_invocation_interlock
;
754 case ir_intrinsic_begin_fragment_shader_ordering
:
755 op
= nir_intrinsic_begin_fragment_shader_ordering
;
757 case ir_intrinsic_group_memory_barrier
:
758 op
= nir_intrinsic_group_memory_barrier
;
760 case ir_intrinsic_memory_barrier_atomic_counter
:
761 op
= nir_intrinsic_memory_barrier_atomic_counter
;
763 case ir_intrinsic_memory_barrier_buffer
:
764 op
= nir_intrinsic_memory_barrier_buffer
;
766 case ir_intrinsic_memory_barrier_image
:
767 op
= nir_intrinsic_memory_barrier_image
;
769 case ir_intrinsic_memory_barrier_shared
:
770 op
= nir_intrinsic_memory_barrier_shared
;
772 case ir_intrinsic_shared_load
:
773 op
= nir_intrinsic_load_shared
;
775 case ir_intrinsic_shared_store
:
776 op
= nir_intrinsic_store_shared
;
778 case ir_intrinsic_shared_atomic_add
:
779 op
= ir
->return_deref
->type
->is_integer_32_64()
780 ? nir_intrinsic_shared_atomic_add
781 : nir_intrinsic_shared_atomic_fadd
;
783 case ir_intrinsic_shared_atomic_and
:
784 op
= nir_intrinsic_shared_atomic_and
;
786 case ir_intrinsic_shared_atomic_or
:
787 op
= nir_intrinsic_shared_atomic_or
;
789 case ir_intrinsic_shared_atomic_xor
:
790 op
= nir_intrinsic_shared_atomic_xor
;
792 case ir_intrinsic_shared_atomic_min
:
793 assert(ir
->return_deref
);
794 if (ir
->return_deref
->type
== glsl_type::int_type
)
795 op
= nir_intrinsic_shared_atomic_imin
;
796 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
797 op
= nir_intrinsic_shared_atomic_umin
;
798 else if (ir
->return_deref
->type
== glsl_type::float_type
)
799 op
= nir_intrinsic_shared_atomic_fmin
;
801 unreachable("Invalid type");
803 case ir_intrinsic_shared_atomic_max
:
804 assert(ir
->return_deref
);
805 if (ir
->return_deref
->type
== glsl_type::int_type
)
806 op
= nir_intrinsic_shared_atomic_imax
;
807 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
808 op
= nir_intrinsic_shared_atomic_umax
;
809 else if (ir
->return_deref
->type
== glsl_type::float_type
)
810 op
= nir_intrinsic_shared_atomic_fmax
;
812 unreachable("Invalid type");
814 case ir_intrinsic_shared_atomic_exchange
:
815 op
= nir_intrinsic_shared_atomic_exchange
;
817 case ir_intrinsic_shared_atomic_comp_swap
:
818 op
= ir
->return_deref
->type
->is_integer_32_64()
819 ? nir_intrinsic_shared_atomic_comp_swap
820 : nir_intrinsic_shared_atomic_fcomp_swap
;
822 case ir_intrinsic_vote_any
:
823 op
= nir_intrinsic_vote_any
;
825 case ir_intrinsic_vote_all
:
826 op
= nir_intrinsic_vote_all
;
828 case ir_intrinsic_vote_eq
:
829 op
= nir_intrinsic_vote_ieq
;
831 case ir_intrinsic_ballot
:
832 op
= nir_intrinsic_ballot
;
834 case ir_intrinsic_read_invocation
:
835 op
= nir_intrinsic_read_invocation
;
837 case ir_intrinsic_read_first_invocation
:
838 op
= nir_intrinsic_read_first_invocation
;
841 unreachable("not reached");
844 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(shader
, op
);
845 nir_ssa_def
*ret
= &instr
->dest
.ssa
;
848 case nir_intrinsic_atomic_counter_read_deref
:
849 case nir_intrinsic_atomic_counter_inc_deref
:
850 case nir_intrinsic_atomic_counter_pre_dec_deref
:
851 case nir_intrinsic_atomic_counter_add_deref
:
852 case nir_intrinsic_atomic_counter_min_deref
:
853 case nir_intrinsic_atomic_counter_max_deref
:
854 case nir_intrinsic_atomic_counter_and_deref
:
855 case nir_intrinsic_atomic_counter_or_deref
:
856 case nir_intrinsic_atomic_counter_xor_deref
:
857 case nir_intrinsic_atomic_counter_exchange_deref
:
858 case nir_intrinsic_atomic_counter_comp_swap_deref
: {
859 /* Set the counter variable dereference. */
860 exec_node
*param
= ir
->actual_parameters
.get_head();
861 ir_dereference
*counter
= (ir_dereference
*)param
;
863 instr
->src
[0] = nir_src_for_ssa(&evaluate_deref(counter
)->dest
.ssa
);
864 param
= param
->get_next();
866 /* Set the intrinsic destination. */
867 if (ir
->return_deref
) {
868 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
871 /* Set the intrinsic parameters. */
872 if (!param
->is_tail_sentinel()) {
874 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
875 param
= param
->get_next();
878 if (!param
->is_tail_sentinel()) {
880 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
881 param
= param
->get_next();
884 nir_builder_instr_insert(&b
, &instr
->instr
);
887 case nir_intrinsic_image_deref_load
:
888 case nir_intrinsic_image_deref_store
:
889 case nir_intrinsic_image_deref_atomic_add
:
890 case nir_intrinsic_image_deref_atomic_min
:
891 case nir_intrinsic_image_deref_atomic_max
:
892 case nir_intrinsic_image_deref_atomic_and
:
893 case nir_intrinsic_image_deref_atomic_or
:
894 case nir_intrinsic_image_deref_atomic_xor
:
895 case nir_intrinsic_image_deref_atomic_exchange
:
896 case nir_intrinsic_image_deref_atomic_comp_swap
:
897 case nir_intrinsic_image_deref_atomic_fadd
:
898 case nir_intrinsic_image_deref_samples
:
899 case nir_intrinsic_image_deref_size
: {
900 nir_ssa_undef_instr
*instr_undef
=
901 nir_ssa_undef_instr_create(shader
, 1, 32);
902 nir_builder_instr_insert(&b
, &instr_undef
->instr
);
904 /* Set the image variable dereference. */
905 exec_node
*param
= ir
->actual_parameters
.get_head();
906 ir_dereference
*image
= (ir_dereference
*)param
;
907 const glsl_type
*type
=
908 image
->variable_referenced()->type
->without_array();
910 instr
->src
[0] = nir_src_for_ssa(&evaluate_deref(image
)->dest
.ssa
);
911 param
= param
->get_next();
913 /* Set the intrinsic destination. */
914 if (ir
->return_deref
) {
915 unsigned num_components
= ir
->return_deref
->type
->vector_elements
;
916 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
917 num_components
, 32, NULL
);
920 if (op
== nir_intrinsic_image_deref_size
) {
921 instr
->num_components
= instr
->dest
.ssa
.num_components
;
922 } else if (op
== nir_intrinsic_image_deref_load
||
923 op
== nir_intrinsic_image_deref_store
) {
924 instr
->num_components
= 4;
927 if (op
== nir_intrinsic_image_deref_size
||
928 op
== nir_intrinsic_image_deref_samples
) {
929 nir_builder_instr_insert(&b
, &instr
->instr
);
933 /* Set the address argument, extending the coordinate vector to four
936 nir_ssa_def
*src_addr
=
937 evaluate_rvalue((ir_dereference
*)param
);
938 nir_ssa_def
*srcs
[4];
940 for (int i
= 0; i
< 4; i
++) {
941 if (i
< type
->coordinate_components())
942 srcs
[i
] = nir_channel(&b
, src_addr
, i
);
944 srcs
[i
] = &instr_undef
->def
;
947 instr
->src
[1] = nir_src_for_ssa(nir_vec(&b
, srcs
, 4));
948 param
= param
->get_next();
950 /* Set the sample argument, which is undefined for single-sample
953 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_MS
) {
955 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
956 param
= param
->get_next();
958 instr
->src
[2] = nir_src_for_ssa(&instr_undef
->def
);
961 /* Set the intrinsic parameters. */
962 if (!param
->is_tail_sentinel()) {
964 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
965 param
= param
->get_next();
968 if (!param
->is_tail_sentinel()) {
970 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
971 param
= param
->get_next();
973 nir_builder_instr_insert(&b
, &instr
->instr
);
976 case nir_intrinsic_memory_barrier
:
977 case nir_intrinsic_group_memory_barrier
:
978 case nir_intrinsic_memory_barrier_atomic_counter
:
979 case nir_intrinsic_memory_barrier_buffer
:
980 case nir_intrinsic_memory_barrier_image
:
981 case nir_intrinsic_memory_barrier_shared
:
982 nir_builder_instr_insert(&b
, &instr
->instr
);
984 case nir_intrinsic_shader_clock
:
985 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 2, 32, NULL
);
986 instr
->num_components
= 2;
987 nir_builder_instr_insert(&b
, &instr
->instr
);
989 case nir_intrinsic_begin_invocation_interlock
:
990 nir_builder_instr_insert(&b
, &instr
->instr
);
992 case nir_intrinsic_end_invocation_interlock
:
993 nir_builder_instr_insert(&b
, &instr
->instr
);
995 case nir_intrinsic_begin_fragment_shader_ordering
:
996 nir_builder_instr_insert(&b
, &instr
->instr
);
998 case nir_intrinsic_store_ssbo
: {
999 exec_node
*param
= ir
->actual_parameters
.get_head();
1000 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1002 param
= param
->get_next();
1003 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1005 param
= param
->get_next();
1006 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1008 param
= param
->get_next();
1009 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1012 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1013 assert(!val
->type
->is_boolean() || nir_val
->bit_size
== 32);
1015 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1016 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(block
));
1017 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(offset
));
1018 intrinsic_set_std430_align(instr
, val
->type
);
1019 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1020 instr
->num_components
= val
->type
->vector_elements
;
1022 nir_builder_instr_insert(&b
, &instr
->instr
);
1025 case nir_intrinsic_load_ssbo
: {
1026 exec_node
*param
= ir
->actual_parameters
.get_head();
1027 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1029 param
= param
->get_next();
1030 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1032 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(block
));
1033 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1035 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1036 instr
->num_components
= type
->vector_elements
;
1037 intrinsic_set_std430_align(instr
, type
);
1039 /* Setup destination register */
1040 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1041 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1042 type
->vector_elements
, bit_size
, NULL
);
1044 /* Insert the created nir instruction now since in the case of boolean
1045 * result we will need to emit another instruction after it
1047 nir_builder_instr_insert(&b
, &instr
->instr
);
1050 * In SSBO/UBO's, a true boolean value is any non-zero value, but we
1051 * consider a true boolean to be ~0. Fix this up with a != 0
1054 if (type
->is_boolean())
1055 ret
= nir_i2b(&b
, &instr
->dest
.ssa
);
1058 case nir_intrinsic_ssbo_atomic_add
:
1059 case nir_intrinsic_ssbo_atomic_imin
:
1060 case nir_intrinsic_ssbo_atomic_umin
:
1061 case nir_intrinsic_ssbo_atomic_imax
:
1062 case nir_intrinsic_ssbo_atomic_umax
:
1063 case nir_intrinsic_ssbo_atomic_and
:
1064 case nir_intrinsic_ssbo_atomic_or
:
1065 case nir_intrinsic_ssbo_atomic_xor
:
1066 case nir_intrinsic_ssbo_atomic_exchange
:
1067 case nir_intrinsic_ssbo_atomic_comp_swap
:
1068 case nir_intrinsic_ssbo_atomic_fadd
:
1069 case nir_intrinsic_ssbo_atomic_fmin
:
1070 case nir_intrinsic_ssbo_atomic_fmax
:
1071 case nir_intrinsic_ssbo_atomic_fcomp_swap
: {
1072 int param_count
= ir
->actual_parameters
.length();
1073 assert(param_count
== 3 || param_count
== 4);
1076 exec_node
*param
= ir
->actual_parameters
.get_head();
1077 ir_instruction
*inst
= (ir_instruction
*) param
;
1078 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1081 param
= param
->get_next();
1082 inst
= (ir_instruction
*) param
;
1083 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1085 /* data1 parameter (this is always present) */
1086 param
= param
->get_next();
1087 inst
= (ir_instruction
*) param
;
1088 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1090 /* data2 parameter (only with atomic_comp_swap) */
1091 if (param_count
== 4) {
1092 assert(op
== nir_intrinsic_ssbo_atomic_comp_swap
||
1093 op
== nir_intrinsic_ssbo_atomic_fcomp_swap
);
1094 param
= param
->get_next();
1095 inst
= (ir_instruction
*) param
;
1096 instr
->src
[3] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1100 assert(ir
->return_deref
);
1101 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1102 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1103 nir_builder_instr_insert(&b
, &instr
->instr
);
1106 case nir_intrinsic_load_shared
: {
1107 exec_node
*param
= ir
->actual_parameters
.get_head();
1108 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1110 nir_intrinsic_set_base(instr
, 0);
1111 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(offset
));
1113 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1114 instr
->num_components
= type
->vector_elements
;
1115 intrinsic_set_std430_align(instr
, type
);
1117 /* Setup destination register */
1118 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1119 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1120 type
->vector_elements
, bit_size
, NULL
);
1122 nir_builder_instr_insert(&b
, &instr
->instr
);
1125 case nir_intrinsic_store_shared
: {
1126 exec_node
*param
= ir
->actual_parameters
.get_head();
1127 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1129 param
= param
->get_next();
1130 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1132 param
= param
->get_next();
1133 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1136 nir_intrinsic_set_base(instr
, 0);
1137 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1139 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1141 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1142 assert(!val
->type
->is_boolean() || nir_val
->bit_size
== 32);
1144 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1145 instr
->num_components
= val
->type
->vector_elements
;
1146 intrinsic_set_std430_align(instr
, val
->type
);
1148 nir_builder_instr_insert(&b
, &instr
->instr
);
1151 case nir_intrinsic_shared_atomic_add
:
1152 case nir_intrinsic_shared_atomic_imin
:
1153 case nir_intrinsic_shared_atomic_umin
:
1154 case nir_intrinsic_shared_atomic_imax
:
1155 case nir_intrinsic_shared_atomic_umax
:
1156 case nir_intrinsic_shared_atomic_and
:
1157 case nir_intrinsic_shared_atomic_or
:
1158 case nir_intrinsic_shared_atomic_xor
:
1159 case nir_intrinsic_shared_atomic_exchange
:
1160 case nir_intrinsic_shared_atomic_comp_swap
:
1161 case nir_intrinsic_shared_atomic_fadd
:
1162 case nir_intrinsic_shared_atomic_fmin
:
1163 case nir_intrinsic_shared_atomic_fmax
:
1164 case nir_intrinsic_shared_atomic_fcomp_swap
: {
1165 int param_count
= ir
->actual_parameters
.length();
1166 assert(param_count
== 2 || param_count
== 3);
1169 exec_node
*param
= ir
->actual_parameters
.get_head();
1170 ir_instruction
*inst
= (ir_instruction
*) param
;
1171 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1173 /* data1 parameter (this is always present) */
1174 param
= param
->get_next();
1175 inst
= (ir_instruction
*) param
;
1176 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1178 /* data2 parameter (only with atomic_comp_swap) */
1179 if (param_count
== 3) {
1180 assert(op
== nir_intrinsic_shared_atomic_comp_swap
||
1181 op
== nir_intrinsic_shared_atomic_fcomp_swap
);
1182 param
= param
->get_next();
1183 inst
= (ir_instruction
*) param
;
1185 nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1189 assert(ir
->return_deref
);
1190 unsigned bit_size
= glsl_get_bit_size(ir
->return_deref
->type
);
1191 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1192 ir
->return_deref
->type
->vector_elements
,
1194 nir_builder_instr_insert(&b
, &instr
->instr
);
1197 case nir_intrinsic_vote_any
:
1198 case nir_intrinsic_vote_all
:
1199 case nir_intrinsic_vote_ieq
: {
1200 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1201 instr
->num_components
= 1;
1203 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1204 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1206 nir_builder_instr_insert(&b
, &instr
->instr
);
1210 case nir_intrinsic_ballot
: {
1211 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1212 ir
->return_deref
->type
->vector_elements
, 64, NULL
);
1213 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1215 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1216 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1218 nir_builder_instr_insert(&b
, &instr
->instr
);
1221 case nir_intrinsic_read_invocation
: {
1222 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1223 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1224 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1226 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1227 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1229 ir_rvalue
*invocation
= (ir_rvalue
*) ir
->actual_parameters
.get_head()->next
;
1230 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(invocation
));
1232 nir_builder_instr_insert(&b
, &instr
->instr
);
1235 case nir_intrinsic_read_first_invocation
: {
1236 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1237 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1238 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1240 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1241 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1243 nir_builder_instr_insert(&b
, &instr
->instr
);
1247 unreachable("not reached");
1250 if (ir
->return_deref
)
1251 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), ret
, ~0);
1256 unreachable("glsl_to_nir only handles function calls to intrinsics");
1260 nir_visitor::visit(ir_assignment
*ir
)
1262 unsigned num_components
= ir
->lhs
->type
->vector_elements
;
1264 b
.exact
= ir
->lhs
->variable_referenced()->data
.invariant
||
1265 ir
->lhs
->variable_referenced()->data
.precise
;
1267 if ((ir
->rhs
->as_dereference() || ir
->rhs
->as_constant()) &&
1268 (ir
->write_mask
== (1 << num_components
) - 1 || ir
->write_mask
== 0)) {
1269 if (ir
->condition
) {
1270 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1271 nir_copy_deref(&b
, evaluate_deref(ir
->lhs
), evaluate_deref(ir
->rhs
));
1272 nir_pop_if(&b
, NULL
);
1274 nir_copy_deref(&b
, evaluate_deref(ir
->lhs
), evaluate_deref(ir
->rhs
));
1279 assert(ir
->rhs
->type
->is_scalar() || ir
->rhs
->type
->is_vector());
1281 ir
->lhs
->accept(this);
1282 nir_deref_instr
*lhs_deref
= this->deref
;
1283 nir_ssa_def
*src
= evaluate_rvalue(ir
->rhs
);
1285 if (ir
->write_mask
!= (1 << num_components
) - 1 && ir
->write_mask
!= 0) {
1286 /* GLSL IR will give us the input to the write-masked assignment in a
1287 * single packed vector. So, for example, if the writemask is xzw, then
1288 * we have to swizzle x -> x, y -> z, and z -> w and get the y component
1292 unsigned component
= 0;
1293 for (unsigned i
= 0; i
< 4; i
++) {
1294 swiz
[i
] = ir
->write_mask
& (1 << i
) ? component
++ : 0;
1296 src
= nir_swizzle(&b
, src
, swiz
, num_components
, !supports_ints
);
1299 if (ir
->condition
) {
1300 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1301 nir_store_deref(&b
, lhs_deref
, src
, ir
->write_mask
);
1302 nir_pop_if(&b
, NULL
);
1304 nir_store_deref(&b
, lhs_deref
, src
, ir
->write_mask
);
1309 * Given an instruction, returns a pointer to its destination or NULL if there
1310 * is no destination.
1312 * Note that this only handles instructions we generate at this level.
1315 get_instr_dest(nir_instr
*instr
)
1317 nir_alu_instr
*alu_instr
;
1318 nir_intrinsic_instr
*intrinsic_instr
;
1319 nir_tex_instr
*tex_instr
;
1321 switch (instr
->type
) {
1322 case nir_instr_type_alu
:
1323 alu_instr
= nir_instr_as_alu(instr
);
1324 return &alu_instr
->dest
.dest
;
1326 case nir_instr_type_intrinsic
:
1327 intrinsic_instr
= nir_instr_as_intrinsic(instr
);
1328 if (nir_intrinsic_infos
[intrinsic_instr
->intrinsic
].has_dest
)
1329 return &intrinsic_instr
->dest
;
1333 case nir_instr_type_tex
:
1334 tex_instr
= nir_instr_as_tex(instr
);
1335 return &tex_instr
->dest
;
1338 unreachable("not reached");
1345 nir_visitor::add_instr(nir_instr
*instr
, unsigned num_components
,
1348 nir_dest
*dest
= get_instr_dest(instr
);
1351 nir_ssa_dest_init(instr
, dest
, num_components
, bit_size
, NULL
);
1353 nir_builder_instr_insert(&b
, instr
);
1356 assert(dest
->is_ssa
);
1357 this->result
= &dest
->ssa
;
1362 nir_visitor::evaluate_rvalue(ir_rvalue
* ir
)
1365 if (ir
->as_dereference() || ir
->as_constant()) {
1367 * A dereference is being used on the right hand side, which means we
1368 * must emit a variable load.
1371 this->result
= nir_load_deref(&b
, this->deref
);
1374 return this->result
;
1378 type_is_float(glsl_base_type type
)
1380 return type
== GLSL_TYPE_FLOAT
|| type
== GLSL_TYPE_DOUBLE
||
1381 type
== GLSL_TYPE_FLOAT16
;
1385 type_is_signed(glsl_base_type type
)
1387 return type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_INT64
||
1388 type
== GLSL_TYPE_INT16
;
1392 nir_visitor::visit(ir_expression
*ir
)
1394 /* Some special cases */
1395 switch (ir
->operation
) {
1396 case ir_binop_ubo_load
: {
1397 nir_intrinsic_instr
*load
=
1398 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_ubo
);
1399 unsigned bit_size
= ir
->type
->is_boolean() ? 32 :
1400 glsl_get_bit_size(ir
->type
);
1401 load
->num_components
= ir
->type
->vector_elements
;
1402 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1403 load
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1404 intrinsic_set_std430_align(load
, ir
->type
);
1405 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1408 * In UBO's, a true boolean value is any non-zero value, but we consider
1409 * a true boolean to be ~0. Fix this up with a != 0 comparison.
1412 if (ir
->type
->is_boolean())
1413 this->result
= nir_i2b(&b
, &load
->dest
.ssa
);
1418 case ir_unop_interpolate_at_centroid
:
1419 case ir_binop_interpolate_at_offset
:
1420 case ir_binop_interpolate_at_sample
: {
1421 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1422 ir_swizzle
*swizzle
= NULL
;
1424 /* the api does not allow a swizzle here, but the varying packing code
1425 * may have pushed one into here.
1427 swizzle
= ir
->operands
[0]->as_swizzle();
1429 deref
= swizzle
->val
->as_dereference();
1433 deref
->accept(this);
1435 nir_intrinsic_op op
;
1436 if (this->deref
->mode
== nir_var_shader_in
) {
1437 switch (ir
->operation
) {
1438 case ir_unop_interpolate_at_centroid
:
1439 op
= nir_intrinsic_interp_deref_at_centroid
;
1441 case ir_binop_interpolate_at_offset
:
1442 op
= nir_intrinsic_interp_deref_at_offset
;
1444 case ir_binop_interpolate_at_sample
:
1445 op
= nir_intrinsic_interp_deref_at_sample
;
1448 unreachable("Invalid interpolation intrinsic");
1451 /* This case can happen if the vertex shader does not write the
1452 * given varying. In this case, the linker will lower it to a
1453 * global variable. Since interpolating a variable makes no
1454 * sense, we'll just turn it into a load which will probably
1455 * eventually end up as an SSA definition.
1457 assert(this->deref
->mode
== nir_var_global
);
1458 op
= nir_intrinsic_load_deref
;
1461 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(shader
, op
);
1462 intrin
->num_components
= deref
->type
->vector_elements
;
1463 intrin
->src
[0] = nir_src_for_ssa(&this->deref
->dest
.ssa
);
1465 if (intrin
->intrinsic
== nir_intrinsic_interp_deref_at_offset
||
1466 intrin
->intrinsic
== nir_intrinsic_interp_deref_at_sample
)
1467 intrin
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1469 unsigned bit_size
= glsl_get_bit_size(deref
->type
);
1470 add_instr(&intrin
->instr
, deref
->type
->vector_elements
, bit_size
);
1473 unsigned swiz
[4] = {
1474 swizzle
->mask
.x
, swizzle
->mask
.y
, swizzle
->mask
.z
, swizzle
->mask
.w
1477 result
= nir_swizzle(&b
, result
, swiz
,
1478 swizzle
->type
->vector_elements
, false);
1488 nir_ssa_def
*srcs
[4];
1489 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1490 srcs
[i
] = evaluate_rvalue(ir
->operands
[i
]);
1492 glsl_base_type types
[4];
1493 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1495 types
[i
] = ir
->operands
[i
]->type
->base_type
;
1497 types
[i
] = GLSL_TYPE_FLOAT
;
1499 glsl_base_type out_type
;
1501 out_type
= ir
->type
->base_type
;
1503 out_type
= GLSL_TYPE_FLOAT
;
1505 switch (ir
->operation
) {
1506 case ir_unop_bit_not
: result
= nir_inot(&b
, srcs
[0]); break;
1507 case ir_unop_logic_not
:
1508 result
= supports_ints
? nir_inot(&b
, srcs
[0]) : nir_fnot(&b
, srcs
[0]);
1511 result
= type_is_float(types
[0]) ? nir_fneg(&b
, srcs
[0])
1512 : nir_ineg(&b
, srcs
[0]);
1515 result
= type_is_float(types
[0]) ? nir_fabs(&b
, srcs
[0])
1516 : nir_iabs(&b
, srcs
[0]);
1518 case ir_unop_saturate
:
1519 assert(type_is_float(types
[0]));
1520 result
= nir_fsat(&b
, srcs
[0]);
1523 result
= type_is_float(types
[0]) ? nir_fsign(&b
, srcs
[0])
1524 : nir_isign(&b
, srcs
[0]);
1526 case ir_unop_rcp
: result
= nir_frcp(&b
, srcs
[0]); break;
1527 case ir_unop_rsq
: result
= nir_frsq(&b
, srcs
[0]); break;
1528 case ir_unop_sqrt
: result
= nir_fsqrt(&b
, srcs
[0]); break;
1529 case ir_unop_exp
: unreachable("ir_unop_exp should have been lowered");
1530 case ir_unop_log
: unreachable("ir_unop_log should have been lowered");
1531 case ir_unop_exp2
: result
= nir_fexp2(&b
, srcs
[0]); break;
1532 case ir_unop_log2
: result
= nir_flog2(&b
, srcs
[0]); break;
1534 result
= supports_ints
? nir_i2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1537 result
= supports_ints
? nir_u2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1540 result
= supports_ints
? nir_b2f(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1574 case ir_unop_i642u64
:
1575 case ir_unop_u642i64
: {
1576 nir_alu_type src_type
= nir_get_nir_type_for_glsl_base_type(types
[0]);
1577 nir_alu_type dst_type
= nir_get_nir_type_for_glsl_base_type(out_type
);
1578 result
= nir_build_alu(&b
, nir_type_conversion_op(src_type
, dst_type
,
1579 nir_rounding_mode_undef
),
1580 srcs
[0], NULL
, NULL
, NULL
);
1581 /* b2i and b2f don't have fixed bit-size versions so the builder will
1582 * just assume 32 and we have to fix it up here.
1584 result
->bit_size
= nir_alu_type_get_type_size(dst_type
);
1588 case ir_unop_bitcast_i2f
:
1589 case ir_unop_bitcast_f2i
:
1590 case ir_unop_bitcast_u2f
:
1591 case ir_unop_bitcast_f2u
:
1592 case ir_unop_bitcast_i642d
:
1593 case ir_unop_bitcast_d2i64
:
1594 case ir_unop_bitcast_u642d
:
1595 case ir_unop_bitcast_d2u64
:
1596 case ir_unop_subroutine_to_int
:
1598 result
= nir_imov(&b
, srcs
[0]);
1600 case ir_unop_trunc
: result
= nir_ftrunc(&b
, srcs
[0]); break;
1601 case ir_unop_ceil
: result
= nir_fceil(&b
, srcs
[0]); break;
1602 case ir_unop_floor
: result
= nir_ffloor(&b
, srcs
[0]); break;
1603 case ir_unop_fract
: result
= nir_ffract(&b
, srcs
[0]); break;
1604 case ir_unop_frexp_exp
: result
= nir_frexp_exp(&b
, srcs
[0]); break;
1605 case ir_unop_frexp_sig
: result
= nir_frexp_sig(&b
, srcs
[0]); break;
1606 case ir_unop_round_even
: result
= nir_fround_even(&b
, srcs
[0]); break;
1607 case ir_unop_sin
: result
= nir_fsin(&b
, srcs
[0]); break;
1608 case ir_unop_cos
: result
= nir_fcos(&b
, srcs
[0]); break;
1609 case ir_unop_dFdx
: result
= nir_fddx(&b
, srcs
[0]); break;
1610 case ir_unop_dFdy
: result
= nir_fddy(&b
, srcs
[0]); break;
1611 case ir_unop_dFdx_fine
: result
= nir_fddx_fine(&b
, srcs
[0]); break;
1612 case ir_unop_dFdy_fine
: result
= nir_fddy_fine(&b
, srcs
[0]); break;
1613 case ir_unop_dFdx_coarse
: result
= nir_fddx_coarse(&b
, srcs
[0]); break;
1614 case ir_unop_dFdy_coarse
: result
= nir_fddy_coarse(&b
, srcs
[0]); break;
1615 case ir_unop_pack_snorm_2x16
:
1616 result
= nir_pack_snorm_2x16(&b
, srcs
[0]);
1618 case ir_unop_pack_snorm_4x8
:
1619 result
= nir_pack_snorm_4x8(&b
, srcs
[0]);
1621 case ir_unop_pack_unorm_2x16
:
1622 result
= nir_pack_unorm_2x16(&b
, srcs
[0]);
1624 case ir_unop_pack_unorm_4x8
:
1625 result
= nir_pack_unorm_4x8(&b
, srcs
[0]);
1627 case ir_unop_pack_half_2x16
:
1628 result
= nir_pack_half_2x16(&b
, srcs
[0]);
1630 case ir_unop_unpack_snorm_2x16
:
1631 result
= nir_unpack_snorm_2x16(&b
, srcs
[0]);
1633 case ir_unop_unpack_snorm_4x8
:
1634 result
= nir_unpack_snorm_4x8(&b
, srcs
[0]);
1636 case ir_unop_unpack_unorm_2x16
:
1637 result
= nir_unpack_unorm_2x16(&b
, srcs
[0]);
1639 case ir_unop_unpack_unorm_4x8
:
1640 result
= nir_unpack_unorm_4x8(&b
, srcs
[0]);
1642 case ir_unop_unpack_half_2x16
:
1643 result
= nir_unpack_half_2x16(&b
, srcs
[0]);
1645 case ir_unop_pack_sampler_2x32
:
1646 case ir_unop_pack_image_2x32
:
1647 case ir_unop_pack_double_2x32
:
1648 case ir_unop_pack_int_2x32
:
1649 case ir_unop_pack_uint_2x32
:
1650 result
= nir_pack_64_2x32(&b
, srcs
[0]);
1652 case ir_unop_unpack_sampler_2x32
:
1653 case ir_unop_unpack_image_2x32
:
1654 case ir_unop_unpack_double_2x32
:
1655 case ir_unop_unpack_int_2x32
:
1656 case ir_unop_unpack_uint_2x32
:
1657 result
= nir_unpack_64_2x32(&b
, srcs
[0]);
1659 case ir_unop_bitfield_reverse
:
1660 result
= nir_bitfield_reverse(&b
, srcs
[0]);
1662 case ir_unop_bit_count
:
1663 result
= nir_bit_count(&b
, srcs
[0]);
1665 case ir_unop_find_msb
:
1667 case GLSL_TYPE_UINT
:
1668 result
= nir_ufind_msb(&b
, srcs
[0]);
1671 result
= nir_ifind_msb(&b
, srcs
[0]);
1674 unreachable("Invalid type for findMSB()");
1677 case ir_unop_find_lsb
:
1678 result
= nir_find_lsb(&b
, srcs
[0]);
1682 switch (ir
->type
->vector_elements
) {
1684 switch (ir
->operands
[0]->type
->vector_elements
) {
1685 case 1: result
= nir_fnoise1_1(&b
, srcs
[0]); break;
1686 case 2: result
= nir_fnoise1_2(&b
, srcs
[0]); break;
1687 case 3: result
= nir_fnoise1_3(&b
, srcs
[0]); break;
1688 case 4: result
= nir_fnoise1_4(&b
, srcs
[0]); break;
1689 default: unreachable("not reached");
1693 switch (ir
->operands
[0]->type
->vector_elements
) {
1694 case 1: result
= nir_fnoise2_1(&b
, srcs
[0]); break;
1695 case 2: result
= nir_fnoise2_2(&b
, srcs
[0]); break;
1696 case 3: result
= nir_fnoise2_3(&b
, srcs
[0]); break;
1697 case 4: result
= nir_fnoise2_4(&b
, srcs
[0]); break;
1698 default: unreachable("not reached");
1702 switch (ir
->operands
[0]->type
->vector_elements
) {
1703 case 1: result
= nir_fnoise3_1(&b
, srcs
[0]); break;
1704 case 2: result
= nir_fnoise3_2(&b
, srcs
[0]); break;
1705 case 3: result
= nir_fnoise3_3(&b
, srcs
[0]); break;
1706 case 4: result
= nir_fnoise3_4(&b
, srcs
[0]); break;
1707 default: unreachable("not reached");
1711 switch (ir
->operands
[0]->type
->vector_elements
) {
1712 case 1: result
= nir_fnoise4_1(&b
, srcs
[0]); break;
1713 case 2: result
= nir_fnoise4_2(&b
, srcs
[0]); break;
1714 case 3: result
= nir_fnoise4_3(&b
, srcs
[0]); break;
1715 case 4: result
= nir_fnoise4_4(&b
, srcs
[0]); break;
1716 default: unreachable("not reached");
1720 unreachable("not reached");
1723 case ir_unop_get_buffer_size
: {
1724 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(
1726 nir_intrinsic_get_buffer_size
);
1727 load
->num_components
= ir
->type
->vector_elements
;
1728 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1729 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1730 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1735 result
= type_is_float(out_type
) ? nir_fadd(&b
, srcs
[0], srcs
[1])
1736 : nir_iadd(&b
, srcs
[0], srcs
[1]);
1739 result
= type_is_float(out_type
) ? nir_fsub(&b
, srcs
[0], srcs
[1])
1740 : nir_isub(&b
, srcs
[0], srcs
[1]);
1743 result
= type_is_float(out_type
) ? nir_fmul(&b
, srcs
[0], srcs
[1])
1744 : nir_imul(&b
, srcs
[0], srcs
[1]);
1747 if (type_is_float(out_type
))
1748 result
= nir_fdiv(&b
, srcs
[0], srcs
[1]);
1749 else if (type_is_signed(out_type
))
1750 result
= nir_idiv(&b
, srcs
[0], srcs
[1]);
1752 result
= nir_udiv(&b
, srcs
[0], srcs
[1]);
1755 result
= type_is_float(out_type
) ? nir_fmod(&b
, srcs
[0], srcs
[1])
1756 : nir_umod(&b
, srcs
[0], srcs
[1]);
1759 if (type_is_float(out_type
))
1760 result
= nir_fmin(&b
, srcs
[0], srcs
[1]);
1761 else if (type_is_signed(out_type
))
1762 result
= nir_imin(&b
, srcs
[0], srcs
[1]);
1764 result
= nir_umin(&b
, srcs
[0], srcs
[1]);
1767 if (type_is_float(out_type
))
1768 result
= nir_fmax(&b
, srcs
[0], srcs
[1]);
1769 else if (type_is_signed(out_type
))
1770 result
= nir_imax(&b
, srcs
[0], srcs
[1]);
1772 result
= nir_umax(&b
, srcs
[0], srcs
[1]);
1774 case ir_binop_pow
: result
= nir_fpow(&b
, srcs
[0], srcs
[1]); break;
1775 case ir_binop_bit_and
: result
= nir_iand(&b
, srcs
[0], srcs
[1]); break;
1776 case ir_binop_bit_or
: result
= nir_ior(&b
, srcs
[0], srcs
[1]); break;
1777 case ir_binop_bit_xor
: result
= nir_ixor(&b
, srcs
[0], srcs
[1]); break;
1778 case ir_binop_logic_and
:
1779 result
= supports_ints
? nir_iand(&b
, srcs
[0], srcs
[1])
1780 : nir_fand(&b
, srcs
[0], srcs
[1]);
1782 case ir_binop_logic_or
:
1783 result
= supports_ints
? nir_ior(&b
, srcs
[0], srcs
[1])
1784 : nir_for(&b
, srcs
[0], srcs
[1]);
1786 case ir_binop_logic_xor
:
1787 result
= supports_ints
? nir_ixor(&b
, srcs
[0], srcs
[1])
1788 : nir_fxor(&b
, srcs
[0], srcs
[1]);
1790 case ir_binop_lshift
: result
= nir_ishl(&b
, srcs
[0], srcs
[1]); break;
1791 case ir_binop_rshift
:
1792 result
= (type_is_signed(out_type
)) ? nir_ishr(&b
, srcs
[0], srcs
[1])
1793 : nir_ushr(&b
, srcs
[0], srcs
[1]);
1795 case ir_binop_imul_high
:
1796 result
= (out_type
== GLSL_TYPE_INT
) ? nir_imul_high(&b
, srcs
[0], srcs
[1])
1797 : nir_umul_high(&b
, srcs
[0], srcs
[1]);
1799 case ir_binop_carry
: result
= nir_uadd_carry(&b
, srcs
[0], srcs
[1]); break;
1800 case ir_binop_borrow
: result
= nir_usub_borrow(&b
, srcs
[0], srcs
[1]); break;
1802 if (supports_ints
) {
1803 if (type_is_float(types
[0]))
1804 result
= nir_flt(&b
, srcs
[0], srcs
[1]);
1805 else if (type_is_signed(types
[0]))
1806 result
= nir_ilt(&b
, srcs
[0], srcs
[1]);
1808 result
= nir_ult(&b
, srcs
[0], srcs
[1]);
1810 result
= nir_slt(&b
, srcs
[0], srcs
[1]);
1813 case ir_binop_gequal
:
1814 if (supports_ints
) {
1815 if (type_is_float(types
[0]))
1816 result
= nir_fge(&b
, srcs
[0], srcs
[1]);
1817 else if (type_is_signed(types
[0]))
1818 result
= nir_ige(&b
, srcs
[0], srcs
[1]);
1820 result
= nir_uge(&b
, srcs
[0], srcs
[1]);
1822 result
= nir_sge(&b
, srcs
[0], srcs
[1]);
1825 case ir_binop_equal
:
1826 if (supports_ints
) {
1827 if (type_is_float(types
[0]))
1828 result
= nir_feq(&b
, srcs
[0], srcs
[1]);
1830 result
= nir_ieq(&b
, srcs
[0], srcs
[1]);
1832 result
= nir_seq(&b
, srcs
[0], srcs
[1]);
1835 case ir_binop_nequal
:
1836 if (supports_ints
) {
1837 if (type_is_float(types
[0]))
1838 result
= nir_fne(&b
, srcs
[0], srcs
[1]);
1840 result
= nir_ine(&b
, srcs
[0], srcs
[1]);
1842 result
= nir_sne(&b
, srcs
[0], srcs
[1]);
1845 case ir_binop_all_equal
:
1846 if (supports_ints
) {
1847 if (type_is_float(types
[0])) {
1848 switch (ir
->operands
[0]->type
->vector_elements
) {
1849 case 1: result
= nir_feq(&b
, srcs
[0], srcs
[1]); break;
1850 case 2: result
= nir_ball_fequal2(&b
, srcs
[0], srcs
[1]); break;
1851 case 3: result
= nir_ball_fequal3(&b
, srcs
[0], srcs
[1]); break;
1852 case 4: result
= nir_ball_fequal4(&b
, srcs
[0], srcs
[1]); break;
1854 unreachable("not reached");
1857 switch (ir
->operands
[0]->type
->vector_elements
) {
1858 case 1: result
= nir_ieq(&b
, srcs
[0], srcs
[1]); break;
1859 case 2: result
= nir_ball_iequal2(&b
, srcs
[0], srcs
[1]); break;
1860 case 3: result
= nir_ball_iequal3(&b
, srcs
[0], srcs
[1]); break;
1861 case 4: result
= nir_ball_iequal4(&b
, srcs
[0], srcs
[1]); break;
1863 unreachable("not reached");
1867 switch (ir
->operands
[0]->type
->vector_elements
) {
1868 case 1: result
= nir_seq(&b
, srcs
[0], srcs
[1]); break;
1869 case 2: result
= nir_fall_equal2(&b
, srcs
[0], srcs
[1]); break;
1870 case 3: result
= nir_fall_equal3(&b
, srcs
[0], srcs
[1]); break;
1871 case 4: result
= nir_fall_equal4(&b
, srcs
[0], srcs
[1]); break;
1873 unreachable("not reached");
1877 case ir_binop_any_nequal
:
1878 if (supports_ints
) {
1879 if (type_is_float(types
[0])) {
1880 switch (ir
->operands
[0]->type
->vector_elements
) {
1881 case 1: result
= nir_fne(&b
, srcs
[0], srcs
[1]); break;
1882 case 2: result
= nir_bany_fnequal2(&b
, srcs
[0], srcs
[1]); break;
1883 case 3: result
= nir_bany_fnequal3(&b
, srcs
[0], srcs
[1]); break;
1884 case 4: result
= nir_bany_fnequal4(&b
, srcs
[0], srcs
[1]); break;
1886 unreachable("not reached");
1889 switch (ir
->operands
[0]->type
->vector_elements
) {
1890 case 1: result
= nir_ine(&b
, srcs
[0], srcs
[1]); break;
1891 case 2: result
= nir_bany_inequal2(&b
, srcs
[0], srcs
[1]); break;
1892 case 3: result
= nir_bany_inequal3(&b
, srcs
[0], srcs
[1]); break;
1893 case 4: result
= nir_bany_inequal4(&b
, srcs
[0], srcs
[1]); break;
1895 unreachable("not reached");
1899 switch (ir
->operands
[0]->type
->vector_elements
) {
1900 case 1: result
= nir_sne(&b
, srcs
[0], srcs
[1]); break;
1901 case 2: result
= nir_fany_nequal2(&b
, srcs
[0], srcs
[1]); break;
1902 case 3: result
= nir_fany_nequal3(&b
, srcs
[0], srcs
[1]); break;
1903 case 4: result
= nir_fany_nequal4(&b
, srcs
[0], srcs
[1]); break;
1905 unreachable("not reached");
1910 switch (ir
->operands
[0]->type
->vector_elements
) {
1911 case 2: result
= nir_fdot2(&b
, srcs
[0], srcs
[1]); break;
1912 case 3: result
= nir_fdot3(&b
, srcs
[0], srcs
[1]); break;
1913 case 4: result
= nir_fdot4(&b
, srcs
[0], srcs
[1]); break;
1915 unreachable("not reached");
1918 case ir_binop_vector_extract
: {
1919 result
= nir_channel(&b
, srcs
[0], 0);
1920 for (unsigned i
= 1; i
< ir
->operands
[0]->type
->vector_elements
; i
++) {
1921 nir_ssa_def
*swizzled
= nir_channel(&b
, srcs
[0], i
);
1922 result
= nir_bcsel(&b
, nir_ieq(&b
, srcs
[1], nir_imm_int(&b
, i
)),
1928 case ir_binop_ldexp
: result
= nir_ldexp(&b
, srcs
[0], srcs
[1]); break;
1930 result
= nir_ffma(&b
, srcs
[0], srcs
[1], srcs
[2]);
1933 result
= nir_flrp(&b
, srcs
[0], srcs
[1], srcs
[2]);
1937 result
= nir_bcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
1939 result
= nir_fcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
1941 case ir_triop_bitfield_extract
:
1942 result
= (out_type
== GLSL_TYPE_INT
) ?
1943 nir_ibitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]) :
1944 nir_ubitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]);
1946 case ir_quadop_bitfield_insert
:
1947 result
= nir_bitfield_insert(&b
, srcs
[0], srcs
[1], srcs
[2], srcs
[3]);
1949 case ir_quadop_vector
:
1950 result
= nir_vec(&b
, srcs
, ir
->type
->vector_elements
);
1954 unreachable("not reached");
1959 nir_visitor::visit(ir_swizzle
*ir
)
1961 unsigned swizzle
[4] = { ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
};
1962 result
= nir_swizzle(&b
, evaluate_rvalue(ir
->val
), swizzle
,
1963 ir
->type
->vector_elements
, !supports_ints
);
1967 nir_visitor::visit(ir_texture
*ir
)
1974 num_srcs
= 1; /* coordinate */
1979 op
= (ir
->op
== ir_txb
) ? nir_texop_txb
: nir_texop_txl
;
1980 num_srcs
= 2; /* coordinate, bias/lod */
1984 op
= nir_texop_txd
; /* coordinate, dPdx, dPdy */
1990 if (ir
->lod_info
.lod
!= NULL
)
1991 num_srcs
= 2; /* coordinate, lod */
1993 num_srcs
= 1; /* coordinate */
1997 op
= nir_texop_txf_ms
;
1998 num_srcs
= 2; /* coordinate, sample_index */
2003 if (ir
->lod_info
.lod
!= NULL
)
2004 num_srcs
= 1; /* lod */
2011 num_srcs
= 1; /* coordinate */
2016 num_srcs
= 1; /* coordinate */
2019 case ir_query_levels
:
2020 op
= nir_texop_query_levels
;
2024 case ir_texture_samples
:
2025 op
= nir_texop_texture_samples
;
2029 case ir_samples_identical
:
2030 op
= nir_texop_samples_identical
;
2031 num_srcs
= 1; /* coordinate */
2035 unreachable("not reached");
2038 if (ir
->projector
!= NULL
)
2040 if (ir
->shadow_comparator
!= NULL
)
2042 if (ir
->offset
!= NULL
)
2045 /* Add one for the texture deref */
2048 nir_tex_instr
*instr
= nir_tex_instr_create(this->shader
, num_srcs
);
2051 instr
->sampler_dim
=
2052 (glsl_sampler_dim
) ir
->sampler
->type
->sampler_dimensionality
;
2053 instr
->is_array
= ir
->sampler
->type
->sampler_array
;
2054 instr
->is_shadow
= ir
->sampler
->type
->sampler_shadow
;
2055 if (instr
->is_shadow
)
2056 instr
->is_new_style_shadow
= (ir
->type
->vector_elements
== 1);
2057 switch (ir
->type
->base_type
) {
2058 case GLSL_TYPE_FLOAT
:
2059 instr
->dest_type
= nir_type_float
;
2062 instr
->dest_type
= nir_type_int
;
2064 case GLSL_TYPE_BOOL
:
2065 case GLSL_TYPE_UINT
:
2066 instr
->dest_type
= nir_type_uint
;
2069 unreachable("not reached");
2072 nir_deref_instr
*sampler_deref
= evaluate_deref(ir
->sampler
);
2073 instr
->src
[0].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2074 instr
->src
[0].src_type
= nir_tex_src_texture_deref
;
2075 instr
->src
[1].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2076 instr
->src
[1].src_type
= nir_tex_src_sampler_deref
;
2078 unsigned src_number
= 2;
2080 if (ir
->coordinate
!= NULL
) {
2081 instr
->coord_components
= ir
->coordinate
->type
->vector_elements
;
2082 instr
->src
[src_number
].src
=
2083 nir_src_for_ssa(evaluate_rvalue(ir
->coordinate
));
2084 instr
->src
[src_number
].src_type
= nir_tex_src_coord
;
2088 if (ir
->projector
!= NULL
) {
2089 instr
->src
[src_number
].src
=
2090 nir_src_for_ssa(evaluate_rvalue(ir
->projector
));
2091 instr
->src
[src_number
].src_type
= nir_tex_src_projector
;
2095 if (ir
->shadow_comparator
!= NULL
) {
2096 instr
->src
[src_number
].src
=
2097 nir_src_for_ssa(evaluate_rvalue(ir
->shadow_comparator
));
2098 instr
->src
[src_number
].src_type
= nir_tex_src_comparator
;
2102 if (ir
->offset
!= NULL
) {
2103 /* we don't support multiple offsets yet */
2104 assert(ir
->offset
->type
->is_vector() || ir
->offset
->type
->is_scalar());
2106 instr
->src
[src_number
].src
=
2107 nir_src_for_ssa(evaluate_rvalue(ir
->offset
));
2108 instr
->src
[src_number
].src_type
= nir_tex_src_offset
;
2114 instr
->src
[src_number
].src
=
2115 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.bias
));
2116 instr
->src
[src_number
].src_type
= nir_tex_src_bias
;
2123 if (ir
->lod_info
.lod
!= NULL
) {
2124 instr
->src
[src_number
].src
=
2125 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.lod
));
2126 instr
->src
[src_number
].src_type
= nir_tex_src_lod
;
2132 instr
->src
[src_number
].src
=
2133 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdx
));
2134 instr
->src
[src_number
].src_type
= nir_tex_src_ddx
;
2136 instr
->src
[src_number
].src
=
2137 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdy
));
2138 instr
->src
[src_number
].src_type
= nir_tex_src_ddy
;
2143 instr
->src
[src_number
].src
=
2144 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.sample_index
));
2145 instr
->src
[src_number
].src_type
= nir_tex_src_ms_index
;
2150 instr
->component
= ir
->lod_info
.component
->as_constant()->value
.u
[0];
2157 assert(src_number
== num_srcs
);
2159 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2160 add_instr(&instr
->instr
, nir_tex_instr_dest_size(instr
), bit_size
);
2164 nir_visitor::visit(ir_constant
*ir
)
2167 * We don't know if this variable is an array or struct that gets
2168 * dereferenced, so do the safe thing an make it a variable with a
2169 * constant initializer and return a dereference.
2173 nir_local_variable_create(this->impl
, ir
->type
, "const_temp");
2174 var
->data
.read_only
= true;
2175 var
->constant_initializer
= constant_copy(ir
, var
);
2177 this->deref
= nir_build_deref_var(&b
, var
);
2181 nir_visitor::visit(ir_dereference_variable
*ir
)
2183 struct hash_entry
*entry
=
2184 _mesa_hash_table_search(this->var_table
, ir
->var
);
2186 nir_variable
*var
= (nir_variable
*) entry
->data
;
2188 this->deref
= nir_build_deref_var(&b
, var
);
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 this->deref
= nir_build_deref_struct(&b
, this->deref
, field_index
);
2203 nir_visitor::visit(ir_dereference_array
*ir
)
2205 nir_ssa_def
*index
= evaluate_rvalue(ir
->array_index
);
2207 ir
->array
->accept(this);
2209 this->deref
= nir_build_deref_array(&b
, this->deref
, index
);
2213 nir_visitor::visit(ir_barrier
*)
2215 nir_intrinsic_instr
*instr
=
2216 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_barrier
);
2217 nir_builder_instr_insert(&b
, &instr
->instr
);