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 nir_constant
*constant_copy(ir_constant
*ir
, void *mem_ctx
);
99 /* most recent deref instruction created */
100 nir_deref_instr
*deref
;
102 /* whether the IR we're operating on is per-function or global */
105 ir_function_signature
*sig
;
107 /* map of ir_variable -> nir_variable */
108 struct hash_table
*var_table
;
110 /* map of ir_function_signature -> nir_function_overload */
111 struct hash_table
*overload_table
;
115 * This visitor runs before the main visitor, calling create_function() for
116 * each function so that the main visitor can resolve forward references in
120 class nir_function_visitor
: public ir_hierarchical_visitor
123 nir_function_visitor(nir_visitor
*v
) : visitor(v
)
126 virtual ir_visitor_status
visit_enter(ir_function
*);
129 nir_visitor
*visitor
;
132 } /* end of anonymous namespace */
135 glsl_to_nir(const struct gl_shader_program
*shader_prog
,
136 gl_shader_stage stage
,
137 const nir_shader_compiler_options
*options
)
139 struct gl_linked_shader
*sh
= shader_prog
->_LinkedShaders
[stage
];
141 nir_shader
*shader
= nir_shader_create(NULL
, stage
, options
,
144 nir_visitor
v1(shader
);
145 nir_function_visitor
v2(&v1
);
147 visit_exec_list(sh
->ir
, &v1
);
149 nir_lower_constant_initializers(shader
, (nir_variable_mode
)~0);
151 /* Remap the locations to slots so those requiring two slots will occupy
152 * two locations. For instance, if we have in the IR code a dvec3 attr0 in
153 * location 0 and vec4 attr1 in location 1, in NIR attr0 will use
154 * locations/slots 0 and 1, and attr1 will use location/slot 2 */
155 if (shader
->info
.stage
== MESA_SHADER_VERTEX
)
156 nir_remap_dual_slot_attributes(shader
, &sh
->Program
->DualSlotInputs
);
158 shader
->info
.name
= ralloc_asprintf(shader
, "GLSL%d", shader_prog
->Name
);
159 if (shader_prog
->Label
)
160 shader
->info
.label
= ralloc_strdup(shader
, shader_prog
->Label
);
162 /* Check for transform feedback varyings specified via the API */
163 shader
->info
.has_transform_feedback_varyings
=
164 shader_prog
->TransformFeedback
.NumVarying
> 0;
166 /* Check for transform feedback varyings specified in the Shader */
167 if (shader_prog
->last_vert_prog
)
168 shader
->info
.has_transform_feedback_varyings
|=
169 shader_prog
->last_vert_prog
->sh
.LinkedTransformFeedback
->NumVarying
> 0;
174 nir_visitor::nir_visitor(nir_shader
*shader
)
176 this->supports_ints
= shader
->options
->native_integers
;
177 this->shader
= shader
;
178 this->is_global
= true;
179 this->var_table
= _mesa_pointer_hash_table_create(NULL
);
180 this->overload_table
= _mesa_pointer_hash_table_create(NULL
);
183 memset(&this->b
, 0, sizeof(this->b
));
186 nir_visitor::~nir_visitor()
188 _mesa_hash_table_destroy(this->var_table
, NULL
);
189 _mesa_hash_table_destroy(this->overload_table
, NULL
);
193 nir_visitor::evaluate_deref(ir_instruction
*ir
)
200 nir_visitor::constant_copy(ir_constant
*ir
, void *mem_ctx
)
205 nir_constant
*ret
= rzalloc(mem_ctx
, nir_constant
);
207 const unsigned rows
= ir
->type
->vector_elements
;
208 const unsigned cols
= ir
->type
->matrix_columns
;
211 ret
->num_elements
= 0;
212 switch (ir
->type
->base_type
) {
214 /* Only float base types can be matrices. */
217 for (unsigned r
= 0; r
< rows
; r
++)
219 ret
->values
[0].u32
[r
] = ir
->value
.u
[r
];
221 ret
->values
[0].f32
[r
] = ir
->value
.u
[r
];
226 /* Only float base types can be matrices. */
229 for (unsigned r
= 0; r
< rows
; r
++)
231 ret
->values
[0].i32
[r
] = ir
->value
.i
[r
];
233 ret
->values
[0].f32
[r
] = ir
->value
.i
[r
];
237 case GLSL_TYPE_FLOAT
:
238 for (unsigned c
= 0; c
< cols
; c
++) {
239 for (unsigned r
= 0; r
< rows
; r
++)
240 ret
->values
[c
].f32
[r
] = ir
->value
.f
[c
* rows
+ r
];
244 case GLSL_TYPE_DOUBLE
:
245 for (unsigned c
= 0; c
< cols
; c
++) {
246 for (unsigned r
= 0; r
< rows
; r
++)
247 ret
->values
[c
].f64
[r
] = ir
->value
.d
[c
* rows
+ r
];
251 case GLSL_TYPE_UINT64
:
252 /* Only float base types can be matrices. */
255 for (unsigned r
= 0; r
< rows
; r
++)
256 ret
->values
[0].u64
[r
] = ir
->value
.u64
[r
];
259 case GLSL_TYPE_INT64
:
260 /* Only float base types can be matrices. */
263 for (unsigned r
= 0; r
< rows
; r
++)
264 ret
->values
[0].i64
[r
] = ir
->value
.i64
[r
];
268 /* Only float base types can be matrices. */
271 for (unsigned r
= 0; r
< rows
; r
++)
272 ret
->values
[0].b
[r
] = ir
->value
.b
[r
];
276 case GLSL_TYPE_STRUCT
:
277 case GLSL_TYPE_ARRAY
:
278 ret
->elements
= ralloc_array(mem_ctx
, nir_constant
*,
280 ret
->num_elements
= ir
->type
->length
;
282 for (i
= 0; i
< ir
->type
->length
; i
++)
283 ret
->elements
[i
] = constant_copy(ir
->const_elements
[i
], mem_ctx
);
287 unreachable("not reached");
294 nir_visitor::visit(ir_variable
*ir
)
296 /* TODO: In future we should switch to using the NIR lowering pass but for
297 * now just ignore these variables as GLSL IR should have lowered them.
298 * Anything remaining are just dead vars that weren't cleaned up.
300 if (ir
->data
.mode
== ir_var_shader_shared
)
303 /* FINISHME: inout parameters */
304 assert(ir
->data
.mode
!= ir_var_function_inout
);
306 if (ir
->data
.mode
== ir_var_function_out
)
309 nir_variable
*var
= rzalloc(shader
, nir_variable
);
310 var
->type
= ir
->type
;
311 var
->name
= ralloc_strdup(var
, ir
->name
);
313 var
->data
.always_active_io
= ir
->data
.always_active_io
;
314 var
->data
.read_only
= ir
->data
.read_only
;
315 var
->data
.centroid
= ir
->data
.centroid
;
316 var
->data
.sample
= ir
->data
.sample
;
317 var
->data
.patch
= ir
->data
.patch
;
318 var
->data
.invariant
= ir
->data
.invariant
;
319 var
->data
.location
= ir
->data
.location
;
320 var
->data
.stream
= ir
->data
.stream
;
321 var
->data
.compact
= false;
323 switch(ir
->data
.mode
) {
325 case ir_var_temporary
:
327 var
->data
.mode
= nir_var_shader_temp
;
329 var
->data
.mode
= nir_var_function_temp
;
332 case ir_var_function_in
:
333 case ir_var_const_in
:
334 var
->data
.mode
= nir_var_function_temp
;
337 case ir_var_shader_in
:
338 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
339 ir
->data
.location
== VARYING_SLOT_FACE
) {
340 /* For whatever reason, GLSL IR makes gl_FrontFacing an input */
341 var
->data
.location
= SYSTEM_VALUE_FRONT_FACE
;
342 var
->data
.mode
= nir_var_system_value
;
343 } else if (shader
->info
.stage
== MESA_SHADER_GEOMETRY
&&
344 ir
->data
.location
== VARYING_SLOT_PRIMITIVE_ID
) {
345 /* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */
346 var
->data
.location
= SYSTEM_VALUE_PRIMITIVE_ID
;
347 var
->data
.mode
= nir_var_system_value
;
349 var
->data
.mode
= nir_var_shader_in
;
351 if (shader
->info
.stage
== MESA_SHADER_TESS_EVAL
&&
352 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
353 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
354 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
359 case ir_var_shader_out
:
360 var
->data
.mode
= nir_var_shader_out
;
361 if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
&&
362 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
363 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
364 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
369 if (ir
->get_interface_type())
370 var
->data
.mode
= nir_var_mem_ubo
;
372 var
->data
.mode
= nir_var_uniform
;
375 case ir_var_shader_storage
:
376 var
->data
.mode
= nir_var_mem_ssbo
;
379 case ir_var_system_value
:
380 var
->data
.mode
= nir_var_system_value
;
384 unreachable("not reached");
387 var
->data
.interpolation
= ir
->data
.interpolation
;
388 var
->data
.origin_upper_left
= ir
->data
.origin_upper_left
;
389 var
->data
.pixel_center_integer
= ir
->data
.pixel_center_integer
;
390 var
->data
.location_frac
= ir
->data
.location_frac
;
392 switch (ir
->data
.depth_layout
) {
393 case ir_depth_layout_none
:
394 var
->data
.depth_layout
= nir_depth_layout_none
;
396 case ir_depth_layout_any
:
397 var
->data
.depth_layout
= nir_depth_layout_any
;
399 case ir_depth_layout_greater
:
400 var
->data
.depth_layout
= nir_depth_layout_greater
;
402 case ir_depth_layout_less
:
403 var
->data
.depth_layout
= nir_depth_layout_less
;
405 case ir_depth_layout_unchanged
:
406 var
->data
.depth_layout
= nir_depth_layout_unchanged
;
409 unreachable("not reached");
412 var
->data
.index
= ir
->data
.index
;
413 var
->data
.descriptor_set
= 0;
414 var
->data
.binding
= ir
->data
.binding
;
415 var
->data
.explicit_binding
= ir
->data
.explicit_binding
;
416 var
->data
.bindless
= ir
->data
.bindless
;
417 var
->data
.offset
= ir
->data
.offset
;
419 unsigned image_access
= 0;
420 if (ir
->data
.memory_read_only
)
421 image_access
|= ACCESS_NON_WRITEABLE
;
422 if (ir
->data
.memory_write_only
)
423 image_access
|= ACCESS_NON_READABLE
;
424 if (ir
->data
.memory_coherent
)
425 image_access
|= ACCESS_COHERENT
;
426 if (ir
->data
.memory_volatile
)
427 image_access
|= ACCESS_VOLATILE
;
428 if (ir
->data
.memory_restrict
)
429 image_access
|= ACCESS_RESTRICT
;
430 var
->data
.image
.access
= (gl_access_qualifier
)image_access
;
431 var
->data
.image
.format
= ir
->data
.image_format
;
433 var
->data
.fb_fetch_output
= ir
->data
.fb_fetch_output
;
434 var
->data
.explicit_xfb_buffer
= ir
->data
.explicit_xfb_buffer
;
435 var
->data
.explicit_xfb_stride
= ir
->data
.explicit_xfb_stride
;
436 var
->data
.xfb_buffer
= ir
->data
.xfb_buffer
;
437 var
->data
.xfb_stride
= ir
->data
.xfb_stride
;
439 var
->num_state_slots
= ir
->get_num_state_slots();
440 if (var
->num_state_slots
> 0) {
441 var
->state_slots
= rzalloc_array(var
, nir_state_slot
,
442 var
->num_state_slots
);
444 ir_state_slot
*state_slots
= ir
->get_state_slots();
445 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
446 for (unsigned j
= 0; j
< 5; j
++)
447 var
->state_slots
[i
].tokens
[j
] = state_slots
[i
].tokens
[j
];
448 var
->state_slots
[i
].swizzle
= state_slots
[i
].swizzle
;
451 var
->state_slots
= NULL
;
454 var
->constant_initializer
= constant_copy(ir
->constant_initializer
, var
);
456 var
->interface_type
= ir
->get_interface_type();
458 if (var
->data
.mode
== nir_var_function_temp
)
459 nir_function_impl_add_variable(impl
, var
);
461 nir_shader_add_variable(shader
, var
);
463 _mesa_hash_table_insert(var_table
, ir
, var
);
467 nir_function_visitor::visit_enter(ir_function
*ir
)
469 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
) {
470 visitor
->create_function(sig
);
472 return visit_continue_with_parent
;
476 nir_visitor::create_function(ir_function_signature
*ir
)
478 if (ir
->is_intrinsic())
481 nir_function
*func
= nir_function_create(shader
, ir
->function_name());
482 if (strcmp(ir
->function_name(), "main") == 0)
483 func
->is_entrypoint
= true;
485 func
->num_params
= ir
->parameters
.length() +
486 (ir
->return_type
!= glsl_type::void_type
);
487 func
->params
= ralloc_array(shader
, nir_parameter
, func
->num_params
);
491 if (ir
->return_type
!= glsl_type::void_type
) {
492 /* The return value is a variable deref (basically an out parameter) */
493 func
->params
[np
].num_components
= 1;
494 func
->params
[np
].bit_size
= 32;
498 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
499 /* FINISHME: pass arrays, structs, etc by reference? */
500 assert(param
->type
->is_vector() || param
->type
->is_scalar());
502 if (param
->data
.mode
== ir_var_function_in
) {
503 func
->params
[np
].num_components
= param
->type
->vector_elements
;
504 func
->params
[np
].bit_size
= glsl_get_bit_size(param
->type
);
506 func
->params
[np
].num_components
= 1;
507 func
->params
[np
].bit_size
= 32;
511 assert(np
== func
->num_params
);
513 _mesa_hash_table_insert(this->overload_table
, ir
, func
);
517 nir_visitor::visit(ir_function
*ir
)
519 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
)
524 nir_visitor::visit(ir_function_signature
*ir
)
526 if (ir
->is_intrinsic())
531 struct hash_entry
*entry
=
532 _mesa_hash_table_search(this->overload_table
, ir
);
535 nir_function
*func
= (nir_function
*) entry
->data
;
537 if (ir
->is_defined
) {
538 nir_function_impl
*impl
= nir_function_impl_create(func
);
541 this->is_global
= false;
543 nir_builder_init(&b
, impl
);
544 b
.cursor
= nir_after_cf_list(&impl
->body
);
546 unsigned i
= (ir
->return_type
!= glsl_type::void_type
) ? 1 : 0;
548 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
550 nir_local_variable_create(impl
, param
->type
, param
->name
);
552 if (param
->data
.mode
== ir_var_function_in
) {
553 nir_store_var(&b
, var
, nir_load_param(&b
, i
), ~0);
556 _mesa_hash_table_insert(var_table
, param
, var
);
560 visit_exec_list(&ir
->body
, this);
562 this->is_global
= true;
569 nir_visitor::visit(ir_loop
*ir
)
572 visit_exec_list(&ir
->body_instructions
, this);
573 nir_pop_loop(&b
, NULL
);
577 nir_visitor::visit(ir_if
*ir
)
579 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
580 visit_exec_list(&ir
->then_instructions
, this);
581 nir_push_else(&b
, NULL
);
582 visit_exec_list(&ir
->else_instructions
, this);
583 nir_pop_if(&b
, NULL
);
587 nir_visitor::visit(ir_discard
*ir
)
590 * discards aren't treated as control flow, because before we lower them
591 * they can appear anywhere in the shader and the stuff after them may still
592 * be executed (yay, crazy GLSL rules!). However, after lowering, all the
593 * discards will be immediately followed by a return.
596 nir_intrinsic_instr
*discard
;
598 discard
= nir_intrinsic_instr_create(this->shader
,
599 nir_intrinsic_discard_if
);
601 nir_src_for_ssa(evaluate_rvalue(ir
->condition
));
603 discard
= nir_intrinsic_instr_create(this->shader
, nir_intrinsic_discard
);
606 nir_builder_instr_insert(&b
, &discard
->instr
);
610 nir_visitor::visit(ir_emit_vertex
*ir
)
612 nir_intrinsic_instr
*instr
=
613 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_emit_vertex
);
614 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
615 nir_builder_instr_insert(&b
, &instr
->instr
);
619 nir_visitor::visit(ir_end_primitive
*ir
)
621 nir_intrinsic_instr
*instr
=
622 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_end_primitive
);
623 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
624 nir_builder_instr_insert(&b
, &instr
->instr
);
628 nir_visitor::visit(ir_loop_jump
*ir
)
632 case ir_loop_jump::jump_break
:
633 type
= nir_jump_break
;
635 case ir_loop_jump::jump_continue
:
636 type
= nir_jump_continue
;
639 unreachable("not reached");
642 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, type
);
643 nir_builder_instr_insert(&b
, &instr
->instr
);
647 nir_visitor::visit(ir_return
*ir
)
649 if (ir
->value
!= NULL
) {
650 nir_deref_instr
*ret_deref
=
651 nir_build_deref_cast(&b
, nir_load_param(&b
, 0),
652 nir_var_function_temp
, ir
->value
->type
, 0);
654 nir_ssa_def
*val
= evaluate_rvalue(ir
->value
);
655 nir_store_deref(&b
, ret_deref
, val
, ~0);
658 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, nir_jump_return
);
659 nir_builder_instr_insert(&b
, &instr
->instr
);
663 intrinsic_set_std430_align(nir_intrinsic_instr
*intrin
, const glsl_type
*type
)
665 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
666 unsigned pow2_components
= util_next_power_of_two(type
->vector_elements
);
667 nir_intrinsic_set_align(intrin
, (bit_size
/ 8) * pow2_components
, 0);
671 nir_visitor::visit(ir_call
*ir
)
673 if (ir
->callee
->is_intrinsic()) {
676 switch (ir
->callee
->intrinsic_id
) {
677 case ir_intrinsic_atomic_counter_read
:
678 op
= nir_intrinsic_atomic_counter_read_deref
;
680 case ir_intrinsic_atomic_counter_increment
:
681 op
= nir_intrinsic_atomic_counter_inc_deref
;
683 case ir_intrinsic_atomic_counter_predecrement
:
684 op
= nir_intrinsic_atomic_counter_pre_dec_deref
;
686 case ir_intrinsic_atomic_counter_add
:
687 op
= nir_intrinsic_atomic_counter_add_deref
;
689 case ir_intrinsic_atomic_counter_and
:
690 op
= nir_intrinsic_atomic_counter_and_deref
;
692 case ir_intrinsic_atomic_counter_or
:
693 op
= nir_intrinsic_atomic_counter_or_deref
;
695 case ir_intrinsic_atomic_counter_xor
:
696 op
= nir_intrinsic_atomic_counter_xor_deref
;
698 case ir_intrinsic_atomic_counter_min
:
699 op
= nir_intrinsic_atomic_counter_min_deref
;
701 case ir_intrinsic_atomic_counter_max
:
702 op
= nir_intrinsic_atomic_counter_max_deref
;
704 case ir_intrinsic_atomic_counter_exchange
:
705 op
= nir_intrinsic_atomic_counter_exchange_deref
;
707 case ir_intrinsic_atomic_counter_comp_swap
:
708 op
= nir_intrinsic_atomic_counter_comp_swap_deref
;
710 case ir_intrinsic_image_load
:
711 op
= nir_intrinsic_image_deref_load
;
713 case ir_intrinsic_image_store
:
714 op
= nir_intrinsic_image_deref_store
;
716 case ir_intrinsic_image_atomic_add
:
717 op
= ir
->return_deref
->type
->is_integer_32_64()
718 ? nir_intrinsic_image_deref_atomic_add
719 : nir_intrinsic_image_deref_atomic_fadd
;
721 case ir_intrinsic_image_atomic_min
:
722 op
= nir_intrinsic_image_deref_atomic_min
;
724 case ir_intrinsic_image_atomic_max
:
725 op
= nir_intrinsic_image_deref_atomic_max
;
727 case ir_intrinsic_image_atomic_and
:
728 op
= nir_intrinsic_image_deref_atomic_and
;
730 case ir_intrinsic_image_atomic_or
:
731 op
= nir_intrinsic_image_deref_atomic_or
;
733 case ir_intrinsic_image_atomic_xor
:
734 op
= nir_intrinsic_image_deref_atomic_xor
;
736 case ir_intrinsic_image_atomic_exchange
:
737 op
= nir_intrinsic_image_deref_atomic_exchange
;
739 case ir_intrinsic_image_atomic_comp_swap
:
740 op
= nir_intrinsic_image_deref_atomic_comp_swap
;
742 case ir_intrinsic_memory_barrier
:
743 op
= nir_intrinsic_memory_barrier
;
745 case ir_intrinsic_image_size
:
746 op
= nir_intrinsic_image_deref_size
;
748 case ir_intrinsic_image_samples
:
749 op
= nir_intrinsic_image_deref_samples
;
751 case ir_intrinsic_ssbo_store
:
752 op
= nir_intrinsic_store_ssbo
;
754 case ir_intrinsic_ssbo_load
:
755 op
= nir_intrinsic_load_ssbo
;
757 case ir_intrinsic_ssbo_atomic_add
:
758 op
= ir
->return_deref
->type
->is_integer_32_64()
759 ? nir_intrinsic_ssbo_atomic_add
: nir_intrinsic_ssbo_atomic_fadd
;
761 case ir_intrinsic_ssbo_atomic_and
:
762 op
= nir_intrinsic_ssbo_atomic_and
;
764 case ir_intrinsic_ssbo_atomic_or
:
765 op
= nir_intrinsic_ssbo_atomic_or
;
767 case ir_intrinsic_ssbo_atomic_xor
:
768 op
= nir_intrinsic_ssbo_atomic_xor
;
770 case ir_intrinsic_ssbo_atomic_min
:
771 assert(ir
->return_deref
);
772 if (ir
->return_deref
->type
== glsl_type::int_type
)
773 op
= nir_intrinsic_ssbo_atomic_imin
;
774 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
775 op
= nir_intrinsic_ssbo_atomic_umin
;
776 else if (ir
->return_deref
->type
== glsl_type::float_type
)
777 op
= nir_intrinsic_ssbo_atomic_fmin
;
779 unreachable("Invalid type");
781 case ir_intrinsic_ssbo_atomic_max
:
782 assert(ir
->return_deref
);
783 if (ir
->return_deref
->type
== glsl_type::int_type
)
784 op
= nir_intrinsic_ssbo_atomic_imax
;
785 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
786 op
= nir_intrinsic_ssbo_atomic_umax
;
787 else if (ir
->return_deref
->type
== glsl_type::float_type
)
788 op
= nir_intrinsic_ssbo_atomic_fmax
;
790 unreachable("Invalid type");
792 case ir_intrinsic_ssbo_atomic_exchange
:
793 op
= nir_intrinsic_ssbo_atomic_exchange
;
795 case ir_intrinsic_ssbo_atomic_comp_swap
:
796 op
= ir
->return_deref
->type
->is_integer_32_64()
797 ? nir_intrinsic_ssbo_atomic_comp_swap
798 : nir_intrinsic_ssbo_atomic_fcomp_swap
;
800 case ir_intrinsic_shader_clock
:
801 op
= nir_intrinsic_shader_clock
;
803 case ir_intrinsic_begin_invocation_interlock
:
804 op
= nir_intrinsic_begin_invocation_interlock
;
806 case ir_intrinsic_end_invocation_interlock
:
807 op
= nir_intrinsic_end_invocation_interlock
;
809 case ir_intrinsic_group_memory_barrier
:
810 op
= nir_intrinsic_group_memory_barrier
;
812 case ir_intrinsic_memory_barrier_atomic_counter
:
813 op
= nir_intrinsic_memory_barrier_atomic_counter
;
815 case ir_intrinsic_memory_barrier_buffer
:
816 op
= nir_intrinsic_memory_barrier_buffer
;
818 case ir_intrinsic_memory_barrier_image
:
819 op
= nir_intrinsic_memory_barrier_image
;
821 case ir_intrinsic_memory_barrier_shared
:
822 op
= nir_intrinsic_memory_barrier_shared
;
824 case ir_intrinsic_shared_load
:
825 op
= nir_intrinsic_load_shared
;
827 case ir_intrinsic_shared_store
:
828 op
= nir_intrinsic_store_shared
;
830 case ir_intrinsic_shared_atomic_add
:
831 op
= ir
->return_deref
->type
->is_integer_32_64()
832 ? nir_intrinsic_shared_atomic_add
833 : nir_intrinsic_shared_atomic_fadd
;
835 case ir_intrinsic_shared_atomic_and
:
836 op
= nir_intrinsic_shared_atomic_and
;
838 case ir_intrinsic_shared_atomic_or
:
839 op
= nir_intrinsic_shared_atomic_or
;
841 case ir_intrinsic_shared_atomic_xor
:
842 op
= nir_intrinsic_shared_atomic_xor
;
844 case ir_intrinsic_shared_atomic_min
:
845 assert(ir
->return_deref
);
846 if (ir
->return_deref
->type
== glsl_type::int_type
)
847 op
= nir_intrinsic_shared_atomic_imin
;
848 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
849 op
= nir_intrinsic_shared_atomic_umin
;
850 else if (ir
->return_deref
->type
== glsl_type::float_type
)
851 op
= nir_intrinsic_shared_atomic_fmin
;
853 unreachable("Invalid type");
855 case ir_intrinsic_shared_atomic_max
:
856 assert(ir
->return_deref
);
857 if (ir
->return_deref
->type
== glsl_type::int_type
)
858 op
= nir_intrinsic_shared_atomic_imax
;
859 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
860 op
= nir_intrinsic_shared_atomic_umax
;
861 else if (ir
->return_deref
->type
== glsl_type::float_type
)
862 op
= nir_intrinsic_shared_atomic_fmax
;
864 unreachable("Invalid type");
866 case ir_intrinsic_shared_atomic_exchange
:
867 op
= nir_intrinsic_shared_atomic_exchange
;
869 case ir_intrinsic_shared_atomic_comp_swap
:
870 op
= ir
->return_deref
->type
->is_integer_32_64()
871 ? nir_intrinsic_shared_atomic_comp_swap
872 : nir_intrinsic_shared_atomic_fcomp_swap
;
874 case ir_intrinsic_vote_any
:
875 op
= nir_intrinsic_vote_any
;
877 case ir_intrinsic_vote_all
:
878 op
= nir_intrinsic_vote_all
;
880 case ir_intrinsic_vote_eq
:
881 op
= nir_intrinsic_vote_ieq
;
883 case ir_intrinsic_ballot
:
884 op
= nir_intrinsic_ballot
;
886 case ir_intrinsic_read_invocation
:
887 op
= nir_intrinsic_read_invocation
;
889 case ir_intrinsic_read_first_invocation
:
890 op
= nir_intrinsic_read_first_invocation
;
893 unreachable("not reached");
896 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(shader
, op
);
897 nir_ssa_def
*ret
= &instr
->dest
.ssa
;
900 case nir_intrinsic_atomic_counter_read_deref
:
901 case nir_intrinsic_atomic_counter_inc_deref
:
902 case nir_intrinsic_atomic_counter_pre_dec_deref
:
903 case nir_intrinsic_atomic_counter_add_deref
:
904 case nir_intrinsic_atomic_counter_min_deref
:
905 case nir_intrinsic_atomic_counter_max_deref
:
906 case nir_intrinsic_atomic_counter_and_deref
:
907 case nir_intrinsic_atomic_counter_or_deref
:
908 case nir_intrinsic_atomic_counter_xor_deref
:
909 case nir_intrinsic_atomic_counter_exchange_deref
:
910 case nir_intrinsic_atomic_counter_comp_swap_deref
: {
911 /* Set the counter variable dereference. */
912 exec_node
*param
= ir
->actual_parameters
.get_head();
913 ir_dereference
*counter
= (ir_dereference
*)param
;
915 instr
->src
[0] = nir_src_for_ssa(&evaluate_deref(counter
)->dest
.ssa
);
916 param
= param
->get_next();
918 /* Set the intrinsic destination. */
919 if (ir
->return_deref
) {
920 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
923 /* Set the intrinsic parameters. */
924 if (!param
->is_tail_sentinel()) {
926 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
927 param
= param
->get_next();
930 if (!param
->is_tail_sentinel()) {
932 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
933 param
= param
->get_next();
936 nir_builder_instr_insert(&b
, &instr
->instr
);
939 case nir_intrinsic_image_deref_load
:
940 case nir_intrinsic_image_deref_store
:
941 case nir_intrinsic_image_deref_atomic_add
:
942 case nir_intrinsic_image_deref_atomic_min
:
943 case nir_intrinsic_image_deref_atomic_max
:
944 case nir_intrinsic_image_deref_atomic_and
:
945 case nir_intrinsic_image_deref_atomic_or
:
946 case nir_intrinsic_image_deref_atomic_xor
:
947 case nir_intrinsic_image_deref_atomic_exchange
:
948 case nir_intrinsic_image_deref_atomic_comp_swap
:
949 case nir_intrinsic_image_deref_atomic_fadd
:
950 case nir_intrinsic_image_deref_samples
:
951 case nir_intrinsic_image_deref_size
: {
952 nir_ssa_undef_instr
*instr_undef
=
953 nir_ssa_undef_instr_create(shader
, 1, 32);
954 nir_builder_instr_insert(&b
, &instr_undef
->instr
);
956 /* Set the image variable dereference. */
957 exec_node
*param
= ir
->actual_parameters
.get_head();
958 ir_dereference
*image
= (ir_dereference
*)param
;
959 const glsl_type
*type
=
960 image
->variable_referenced()->type
->without_array();
962 instr
->src
[0] = nir_src_for_ssa(&evaluate_deref(image
)->dest
.ssa
);
963 param
= param
->get_next();
965 /* Set the intrinsic destination. */
966 if (ir
->return_deref
) {
967 unsigned num_components
= ir
->return_deref
->type
->vector_elements
;
968 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
969 num_components
, 32, NULL
);
972 if (op
== nir_intrinsic_image_deref_size
) {
973 instr
->num_components
= instr
->dest
.ssa
.num_components
;
974 } else if (op
== nir_intrinsic_image_deref_load
||
975 op
== nir_intrinsic_image_deref_store
) {
976 instr
->num_components
= 4;
979 if (op
== nir_intrinsic_image_deref_size
||
980 op
== nir_intrinsic_image_deref_samples
) {
981 nir_builder_instr_insert(&b
, &instr
->instr
);
985 /* Set the address argument, extending the coordinate vector to four
988 nir_ssa_def
*src_addr
=
989 evaluate_rvalue((ir_dereference
*)param
);
990 nir_ssa_def
*srcs
[4];
992 for (int i
= 0; i
< 4; i
++) {
993 if (i
< type
->coordinate_components())
994 srcs
[i
] = nir_channel(&b
, src_addr
, i
);
996 srcs
[i
] = &instr_undef
->def
;
999 instr
->src
[1] = nir_src_for_ssa(nir_vec(&b
, srcs
, 4));
1000 param
= param
->get_next();
1002 /* Set the sample argument, which is undefined for single-sample
1005 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_MS
) {
1007 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1008 param
= param
->get_next();
1010 instr
->src
[2] = nir_src_for_ssa(&instr_undef
->def
);
1013 /* Set the intrinsic parameters. */
1014 if (!param
->is_tail_sentinel()) {
1016 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1017 param
= param
->get_next();
1020 if (!param
->is_tail_sentinel()) {
1022 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1023 param
= param
->get_next();
1025 nir_builder_instr_insert(&b
, &instr
->instr
);
1028 case nir_intrinsic_memory_barrier
:
1029 case nir_intrinsic_group_memory_barrier
:
1030 case nir_intrinsic_memory_barrier_atomic_counter
:
1031 case nir_intrinsic_memory_barrier_buffer
:
1032 case nir_intrinsic_memory_barrier_image
:
1033 case nir_intrinsic_memory_barrier_shared
:
1034 nir_builder_instr_insert(&b
, &instr
->instr
);
1036 case nir_intrinsic_shader_clock
:
1037 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 2, 32, NULL
);
1038 instr
->num_components
= 2;
1039 nir_builder_instr_insert(&b
, &instr
->instr
);
1041 case nir_intrinsic_begin_invocation_interlock
:
1042 nir_builder_instr_insert(&b
, &instr
->instr
);
1044 case nir_intrinsic_end_invocation_interlock
:
1045 nir_builder_instr_insert(&b
, &instr
->instr
);
1047 case nir_intrinsic_store_ssbo
: {
1048 exec_node
*param
= ir
->actual_parameters
.get_head();
1049 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1051 param
= param
->get_next();
1052 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1054 param
= param
->get_next();
1055 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1057 param
= param
->get_next();
1058 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1061 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1062 if (val
->type
->is_boolean())
1063 nir_val
= nir_b2i32(&b
, nir_val
);
1065 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1066 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(block
));
1067 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(offset
));
1068 intrinsic_set_std430_align(instr
, val
->type
);
1069 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1070 instr
->num_components
= val
->type
->vector_elements
;
1072 nir_builder_instr_insert(&b
, &instr
->instr
);
1075 case nir_intrinsic_load_ssbo
: {
1076 exec_node
*param
= ir
->actual_parameters
.get_head();
1077 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1079 param
= param
->get_next();
1080 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1082 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(block
));
1083 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1085 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1086 instr
->num_components
= type
->vector_elements
;
1087 intrinsic_set_std430_align(instr
, type
);
1089 /* Setup destination register */
1090 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1091 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1092 type
->vector_elements
, bit_size
, NULL
);
1094 /* Insert the created nir instruction now since in the case of boolean
1095 * result we will need to emit another instruction after it
1097 nir_builder_instr_insert(&b
, &instr
->instr
);
1100 * In SSBO/UBO's, a true boolean value is any non-zero value, but we
1101 * consider a true boolean to be ~0. Fix this up with a != 0
1104 if (type
->is_boolean())
1105 ret
= nir_i2b(&b
, &instr
->dest
.ssa
);
1108 case nir_intrinsic_ssbo_atomic_add
:
1109 case nir_intrinsic_ssbo_atomic_imin
:
1110 case nir_intrinsic_ssbo_atomic_umin
:
1111 case nir_intrinsic_ssbo_atomic_imax
:
1112 case nir_intrinsic_ssbo_atomic_umax
:
1113 case nir_intrinsic_ssbo_atomic_and
:
1114 case nir_intrinsic_ssbo_atomic_or
:
1115 case nir_intrinsic_ssbo_atomic_xor
:
1116 case nir_intrinsic_ssbo_atomic_exchange
:
1117 case nir_intrinsic_ssbo_atomic_comp_swap
:
1118 case nir_intrinsic_ssbo_atomic_fadd
:
1119 case nir_intrinsic_ssbo_atomic_fmin
:
1120 case nir_intrinsic_ssbo_atomic_fmax
:
1121 case nir_intrinsic_ssbo_atomic_fcomp_swap
: {
1122 int param_count
= ir
->actual_parameters
.length();
1123 assert(param_count
== 3 || param_count
== 4);
1126 exec_node
*param
= ir
->actual_parameters
.get_head();
1127 ir_instruction
*inst
= (ir_instruction
*) param
;
1128 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1131 param
= param
->get_next();
1132 inst
= (ir_instruction
*) param
;
1133 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1135 /* data1 parameter (this is always present) */
1136 param
= param
->get_next();
1137 inst
= (ir_instruction
*) param
;
1138 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1140 /* data2 parameter (only with atomic_comp_swap) */
1141 if (param_count
== 4) {
1142 assert(op
== nir_intrinsic_ssbo_atomic_comp_swap
||
1143 op
== nir_intrinsic_ssbo_atomic_fcomp_swap
);
1144 param
= param
->get_next();
1145 inst
= (ir_instruction
*) param
;
1146 instr
->src
[3] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1150 assert(ir
->return_deref
);
1151 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1152 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1153 nir_builder_instr_insert(&b
, &instr
->instr
);
1156 case nir_intrinsic_load_shared
: {
1157 exec_node
*param
= ir
->actual_parameters
.get_head();
1158 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1160 nir_intrinsic_set_base(instr
, 0);
1161 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(offset
));
1163 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1164 instr
->num_components
= type
->vector_elements
;
1165 intrinsic_set_std430_align(instr
, type
);
1167 /* Setup destination register */
1168 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1169 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1170 type
->vector_elements
, bit_size
, NULL
);
1172 nir_builder_instr_insert(&b
, &instr
->instr
);
1174 /* The value in shared memory is a 32-bit value */
1175 if (type
->is_boolean())
1176 ret
= nir_i2b(&b
, &instr
->dest
.ssa
);
1179 case nir_intrinsic_store_shared
: {
1180 exec_node
*param
= ir
->actual_parameters
.get_head();
1181 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1183 param
= param
->get_next();
1184 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1186 param
= param
->get_next();
1187 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1190 nir_intrinsic_set_base(instr
, 0);
1191 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1193 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1195 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1196 /* The value in shared memory is a 32-bit value */
1197 if (val
->type
->is_boolean())
1198 nir_val
= nir_b2i32(&b
, nir_val
);
1200 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1201 instr
->num_components
= val
->type
->vector_elements
;
1202 intrinsic_set_std430_align(instr
, val
->type
);
1204 nir_builder_instr_insert(&b
, &instr
->instr
);
1207 case nir_intrinsic_shared_atomic_add
:
1208 case nir_intrinsic_shared_atomic_imin
:
1209 case nir_intrinsic_shared_atomic_umin
:
1210 case nir_intrinsic_shared_atomic_imax
:
1211 case nir_intrinsic_shared_atomic_umax
:
1212 case nir_intrinsic_shared_atomic_and
:
1213 case nir_intrinsic_shared_atomic_or
:
1214 case nir_intrinsic_shared_atomic_xor
:
1215 case nir_intrinsic_shared_atomic_exchange
:
1216 case nir_intrinsic_shared_atomic_comp_swap
:
1217 case nir_intrinsic_shared_atomic_fadd
:
1218 case nir_intrinsic_shared_atomic_fmin
:
1219 case nir_intrinsic_shared_atomic_fmax
:
1220 case nir_intrinsic_shared_atomic_fcomp_swap
: {
1221 int param_count
= ir
->actual_parameters
.length();
1222 assert(param_count
== 2 || param_count
== 3);
1225 exec_node
*param
= ir
->actual_parameters
.get_head();
1226 ir_instruction
*inst
= (ir_instruction
*) param
;
1227 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1229 /* data1 parameter (this is always present) */
1230 param
= param
->get_next();
1231 inst
= (ir_instruction
*) param
;
1232 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1234 /* data2 parameter (only with atomic_comp_swap) */
1235 if (param_count
== 3) {
1236 assert(op
== nir_intrinsic_shared_atomic_comp_swap
||
1237 op
== nir_intrinsic_shared_atomic_fcomp_swap
);
1238 param
= param
->get_next();
1239 inst
= (ir_instruction
*) param
;
1241 nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1245 assert(ir
->return_deref
);
1246 unsigned bit_size
= glsl_get_bit_size(ir
->return_deref
->type
);
1247 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1248 ir
->return_deref
->type
->vector_elements
,
1250 nir_builder_instr_insert(&b
, &instr
->instr
);
1253 case nir_intrinsic_vote_any
:
1254 case nir_intrinsic_vote_all
:
1255 case nir_intrinsic_vote_ieq
: {
1256 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 1, NULL
);
1257 instr
->num_components
= 1;
1259 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1260 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1262 nir_builder_instr_insert(&b
, &instr
->instr
);
1266 case nir_intrinsic_ballot
: {
1267 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1268 ir
->return_deref
->type
->vector_elements
, 64, NULL
);
1269 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1271 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1272 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1274 nir_builder_instr_insert(&b
, &instr
->instr
);
1277 case nir_intrinsic_read_invocation
: {
1278 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1279 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1280 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1282 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1283 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1285 ir_rvalue
*invocation
= (ir_rvalue
*) ir
->actual_parameters
.get_head()->next
;
1286 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(invocation
));
1288 nir_builder_instr_insert(&b
, &instr
->instr
);
1291 case nir_intrinsic_read_first_invocation
: {
1292 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1293 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1294 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1296 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1297 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1299 nir_builder_instr_insert(&b
, &instr
->instr
);
1303 unreachable("not reached");
1306 if (ir
->return_deref
)
1307 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), ret
, ~0);
1312 struct hash_entry
*entry
=
1313 _mesa_hash_table_search(this->overload_table
, ir
->callee
);
1315 nir_function
*callee
= (nir_function
*) entry
->data
;
1317 nir_call_instr
*call
= nir_call_instr_create(this->shader
, callee
);
1320 nir_deref_instr
*ret_deref
= NULL
;
1321 if (ir
->return_deref
) {
1322 nir_variable
*ret_tmp
=
1323 nir_local_variable_create(this->impl
, ir
->return_deref
->type
,
1325 ret_deref
= nir_build_deref_var(&b
, ret_tmp
);
1326 call
->params
[i
++] = nir_src_for_ssa(&ret_deref
->dest
.ssa
);
1329 foreach_two_lists(formal_node
, &ir
->callee
->parameters
,
1330 actual_node
, &ir
->actual_parameters
) {
1331 ir_rvalue
*param_rvalue
= (ir_rvalue
*) actual_node
;
1332 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
1334 if (sig_param
->data
.mode
== ir_var_function_out
) {
1335 nir_deref_instr
*out_deref
= evaluate_deref(param_rvalue
);
1336 call
->params
[i
] = nir_src_for_ssa(&out_deref
->dest
.ssa
);
1337 } else if (sig_param
->data
.mode
== ir_var_function_in
) {
1338 nir_ssa_def
*val
= evaluate_rvalue(param_rvalue
);
1339 nir_src src
= nir_src_for_ssa(val
);
1341 nir_src_copy(&call
->params
[i
], &src
, call
);
1342 } else if (sig_param
->data
.mode
== ir_var_function_inout
) {
1343 unreachable("unimplemented: inout parameters");
1349 nir_builder_instr_insert(&b
, &call
->instr
);
1351 if (ir
->return_deref
)
1352 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), nir_load_deref(&b
, ret_deref
), ~0);
1356 nir_visitor::visit(ir_assignment
*ir
)
1358 unsigned num_components
= ir
->lhs
->type
->vector_elements
;
1360 b
.exact
= ir
->lhs
->variable_referenced()->data
.invariant
||
1361 ir
->lhs
->variable_referenced()->data
.precise
;
1363 if ((ir
->rhs
->as_dereference() || ir
->rhs
->as_constant()) &&
1364 (ir
->write_mask
== (1 << num_components
) - 1 || ir
->write_mask
== 0)) {
1365 if (ir
->condition
) {
1366 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1367 nir_copy_deref(&b
, evaluate_deref(ir
->lhs
), evaluate_deref(ir
->rhs
));
1368 nir_pop_if(&b
, NULL
);
1370 nir_copy_deref(&b
, evaluate_deref(ir
->lhs
), evaluate_deref(ir
->rhs
));
1375 assert(ir
->rhs
->type
->is_scalar() || ir
->rhs
->type
->is_vector());
1377 ir
->lhs
->accept(this);
1378 nir_deref_instr
*lhs_deref
= this->deref
;
1379 nir_ssa_def
*src
= evaluate_rvalue(ir
->rhs
);
1381 if (ir
->write_mask
!= (1 << num_components
) - 1 && ir
->write_mask
!= 0) {
1382 /* GLSL IR will give us the input to the write-masked assignment in a
1383 * single packed vector. So, for example, if the writemask is xzw, then
1384 * we have to swizzle x -> x, y -> z, and z -> w and get the y component
1388 unsigned component
= 0;
1389 for (unsigned i
= 0; i
< 4; i
++) {
1390 swiz
[i
] = ir
->write_mask
& (1 << i
) ? component
++ : 0;
1392 src
= nir_swizzle(&b
, src
, swiz
, num_components
, false);
1395 if (ir
->condition
) {
1396 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1397 nir_store_deref(&b
, lhs_deref
, src
, ir
->write_mask
);
1398 nir_pop_if(&b
, NULL
);
1400 nir_store_deref(&b
, lhs_deref
, src
, ir
->write_mask
);
1405 * Given an instruction, returns a pointer to its destination or NULL if there
1406 * is no destination.
1408 * Note that this only handles instructions we generate at this level.
1411 get_instr_dest(nir_instr
*instr
)
1413 nir_alu_instr
*alu_instr
;
1414 nir_intrinsic_instr
*intrinsic_instr
;
1415 nir_tex_instr
*tex_instr
;
1417 switch (instr
->type
) {
1418 case nir_instr_type_alu
:
1419 alu_instr
= nir_instr_as_alu(instr
);
1420 return &alu_instr
->dest
.dest
;
1422 case nir_instr_type_intrinsic
:
1423 intrinsic_instr
= nir_instr_as_intrinsic(instr
);
1424 if (nir_intrinsic_infos
[intrinsic_instr
->intrinsic
].has_dest
)
1425 return &intrinsic_instr
->dest
;
1429 case nir_instr_type_tex
:
1430 tex_instr
= nir_instr_as_tex(instr
);
1431 return &tex_instr
->dest
;
1434 unreachable("not reached");
1441 nir_visitor::add_instr(nir_instr
*instr
, unsigned num_components
,
1444 nir_dest
*dest
= get_instr_dest(instr
);
1447 nir_ssa_dest_init(instr
, dest
, num_components
, bit_size
, NULL
);
1449 nir_builder_instr_insert(&b
, instr
);
1452 assert(dest
->is_ssa
);
1453 this->result
= &dest
->ssa
;
1458 nir_visitor::evaluate_rvalue(ir_rvalue
* ir
)
1461 if (ir
->as_dereference() || ir
->as_constant()) {
1463 * A dereference is being used on the right hand side, which means we
1464 * must emit a variable load.
1467 this->result
= nir_load_deref(&b
, this->deref
);
1470 return this->result
;
1474 type_is_float(glsl_base_type type
)
1476 return type
== GLSL_TYPE_FLOAT
|| type
== GLSL_TYPE_DOUBLE
||
1477 type
== GLSL_TYPE_FLOAT16
;
1481 type_is_signed(glsl_base_type type
)
1483 return type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_INT64
||
1484 type
== GLSL_TYPE_INT16
;
1488 type_is_int(glsl_base_type type
)
1490 return type
== GLSL_TYPE_UINT
|| type
== GLSL_TYPE_INT
||
1491 type
== GLSL_TYPE_UINT8
|| type
== GLSL_TYPE_INT8
||
1492 type
== GLSL_TYPE_UINT16
|| type
== GLSL_TYPE_INT16
||
1493 type
== GLSL_TYPE_UINT64
|| type
== GLSL_TYPE_INT64
;
1497 nir_visitor::visit(ir_expression
*ir
)
1499 /* Some special cases */
1500 switch (ir
->operation
) {
1501 case ir_binop_ubo_load
: {
1502 nir_intrinsic_instr
*load
=
1503 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_ubo
);
1504 unsigned bit_size
= ir
->type
->is_boolean() ? 32 :
1505 glsl_get_bit_size(ir
->type
);
1506 load
->num_components
= ir
->type
->vector_elements
;
1507 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1508 load
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1509 intrinsic_set_std430_align(load
, ir
->type
);
1510 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1513 * In UBO's, a true boolean value is any non-zero value, but we consider
1514 * a true boolean to be ~0. Fix this up with a != 0 comparison.
1517 if (ir
->type
->is_boolean())
1518 this->result
= nir_i2b(&b
, &load
->dest
.ssa
);
1523 case ir_unop_interpolate_at_centroid
:
1524 case ir_binop_interpolate_at_offset
:
1525 case ir_binop_interpolate_at_sample
: {
1526 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1527 ir_swizzle
*swizzle
= NULL
;
1529 /* the api does not allow a swizzle here, but the varying packing code
1530 * may have pushed one into here.
1532 swizzle
= ir
->operands
[0]->as_swizzle();
1534 deref
= swizzle
->val
->as_dereference();
1538 deref
->accept(this);
1540 nir_intrinsic_op op
;
1541 if (this->deref
->mode
== nir_var_shader_in
) {
1542 switch (ir
->operation
) {
1543 case ir_unop_interpolate_at_centroid
:
1544 op
= nir_intrinsic_interp_deref_at_centroid
;
1546 case ir_binop_interpolate_at_offset
:
1547 op
= nir_intrinsic_interp_deref_at_offset
;
1549 case ir_binop_interpolate_at_sample
:
1550 op
= nir_intrinsic_interp_deref_at_sample
;
1553 unreachable("Invalid interpolation intrinsic");
1556 /* This case can happen if the vertex shader does not write the
1557 * given varying. In this case, the linker will lower it to a
1558 * global variable. Since interpolating a variable makes no
1559 * sense, we'll just turn it into a load which will probably
1560 * eventually end up as an SSA definition.
1562 assert(this->deref
->mode
== nir_var_shader_temp
);
1563 op
= nir_intrinsic_load_deref
;
1566 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(shader
, op
);
1567 intrin
->num_components
= deref
->type
->vector_elements
;
1568 intrin
->src
[0] = nir_src_for_ssa(&this->deref
->dest
.ssa
);
1570 if (intrin
->intrinsic
== nir_intrinsic_interp_deref_at_offset
||
1571 intrin
->intrinsic
== nir_intrinsic_interp_deref_at_sample
)
1572 intrin
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1574 unsigned bit_size
= glsl_get_bit_size(deref
->type
);
1575 add_instr(&intrin
->instr
, deref
->type
->vector_elements
, bit_size
);
1578 unsigned swiz
[4] = {
1579 swizzle
->mask
.x
, swizzle
->mask
.y
, swizzle
->mask
.z
, swizzle
->mask
.w
1582 result
= nir_swizzle(&b
, result
, swiz
,
1583 swizzle
->type
->vector_elements
, false);
1593 nir_ssa_def
*srcs
[4];
1594 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1595 srcs
[i
] = evaluate_rvalue(ir
->operands
[i
]);
1597 glsl_base_type types
[4];
1598 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1599 if (supports_ints
|| !type_is_int(ir
->operands
[i
]->type
->base_type
))
1600 types
[i
] = ir
->operands
[i
]->type
->base_type
;
1602 types
[i
] = GLSL_TYPE_FLOAT
;
1604 glsl_base_type out_type
;
1605 if (supports_ints
|| !type_is_int(ir
->type
->base_type
))
1606 out_type
= ir
->type
->base_type
;
1608 out_type
= GLSL_TYPE_FLOAT
;
1610 switch (ir
->operation
) {
1611 case ir_unop_bit_not
: result
= nir_inot(&b
, srcs
[0]); break;
1612 case ir_unop_logic_not
:
1613 result
= nir_inot(&b
, srcs
[0]);
1616 result
= type_is_float(types
[0]) ? nir_fneg(&b
, srcs
[0])
1617 : nir_ineg(&b
, srcs
[0]);
1620 result
= type_is_float(types
[0]) ? nir_fabs(&b
, srcs
[0])
1621 : nir_iabs(&b
, srcs
[0]);
1623 case ir_unop_saturate
:
1624 assert(type_is_float(types
[0]));
1625 result
= nir_fsat(&b
, srcs
[0]);
1628 result
= type_is_float(types
[0]) ? nir_fsign(&b
, srcs
[0])
1629 : nir_isign(&b
, srcs
[0]);
1631 case ir_unop_rcp
: result
= nir_frcp(&b
, srcs
[0]); break;
1632 case ir_unop_rsq
: result
= nir_frsq(&b
, srcs
[0]); break;
1633 case ir_unop_sqrt
: result
= nir_fsqrt(&b
, srcs
[0]); break;
1634 case ir_unop_exp
: unreachable("ir_unop_exp should have been lowered");
1635 case ir_unop_log
: unreachable("ir_unop_log should have been lowered");
1636 case ir_unop_exp2
: result
= nir_fexp2(&b
, srcs
[0]); break;
1637 case ir_unop_log2
: result
= nir_flog2(&b
, srcs
[0]); break;
1639 result
= supports_ints
? nir_i2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1642 result
= supports_ints
? nir_u2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1645 result
= nir_b2f32(&b
, srcs
[0]);
1648 result
= supports_ints
? nir_f2i32(&b
, srcs
[0]) : nir_ftrunc(&b
, srcs
[0]);
1651 result
= supports_ints
? nir_f2u32(&b
, srcs
[0]) : nir_ftrunc(&b
, srcs
[0]);
1683 case ir_unop_i642u64
:
1684 case ir_unop_u642i64
: {
1685 nir_alu_type src_type
= nir_get_nir_type_for_glsl_base_type(types
[0]);
1686 nir_alu_type dst_type
= nir_get_nir_type_for_glsl_base_type(out_type
);
1687 result
= nir_build_alu(&b
, nir_type_conversion_op(src_type
, dst_type
,
1688 nir_rounding_mode_undef
),
1689 srcs
[0], NULL
, NULL
, NULL
);
1690 /* b2i and b2f don't have fixed bit-size versions so the builder will
1691 * just assume 32 and we have to fix it up here.
1693 result
->bit_size
= nir_alu_type_get_type_size(dst_type
);
1697 case ir_unop_bitcast_i2f
:
1698 case ir_unop_bitcast_f2i
:
1699 case ir_unop_bitcast_u2f
:
1700 case ir_unop_bitcast_f2u
:
1701 case ir_unop_bitcast_i642d
:
1702 case ir_unop_bitcast_d2i64
:
1703 case ir_unop_bitcast_u642d
:
1704 case ir_unop_bitcast_d2u64
:
1705 case ir_unop_subroutine_to_int
:
1707 result
= nir_imov(&b
, srcs
[0]);
1709 case ir_unop_trunc
: result
= nir_ftrunc(&b
, srcs
[0]); break;
1710 case ir_unop_ceil
: result
= nir_fceil(&b
, srcs
[0]); break;
1711 case ir_unop_floor
: result
= nir_ffloor(&b
, srcs
[0]); break;
1712 case ir_unop_fract
: result
= nir_ffract(&b
, srcs
[0]); break;
1713 case ir_unop_frexp_exp
: result
= nir_frexp_exp(&b
, srcs
[0]); break;
1714 case ir_unop_frexp_sig
: result
= nir_frexp_sig(&b
, srcs
[0]); break;
1715 case ir_unop_round_even
: result
= nir_fround_even(&b
, srcs
[0]); break;
1716 case ir_unop_sin
: result
= nir_fsin(&b
, srcs
[0]); break;
1717 case ir_unop_cos
: result
= nir_fcos(&b
, srcs
[0]); break;
1718 case ir_unop_dFdx
: result
= nir_fddx(&b
, srcs
[0]); break;
1719 case ir_unop_dFdy
: result
= nir_fddy(&b
, srcs
[0]); break;
1720 case ir_unop_dFdx_fine
: result
= nir_fddx_fine(&b
, srcs
[0]); break;
1721 case ir_unop_dFdy_fine
: result
= nir_fddy_fine(&b
, srcs
[0]); break;
1722 case ir_unop_dFdx_coarse
: result
= nir_fddx_coarse(&b
, srcs
[0]); break;
1723 case ir_unop_dFdy_coarse
: result
= nir_fddy_coarse(&b
, srcs
[0]); break;
1724 case ir_unop_pack_snorm_2x16
:
1725 result
= nir_pack_snorm_2x16(&b
, srcs
[0]);
1727 case ir_unop_pack_snorm_4x8
:
1728 result
= nir_pack_snorm_4x8(&b
, srcs
[0]);
1730 case ir_unop_pack_unorm_2x16
:
1731 result
= nir_pack_unorm_2x16(&b
, srcs
[0]);
1733 case ir_unop_pack_unorm_4x8
:
1734 result
= nir_pack_unorm_4x8(&b
, srcs
[0]);
1736 case ir_unop_pack_half_2x16
:
1737 result
= nir_pack_half_2x16(&b
, srcs
[0]);
1739 case ir_unop_unpack_snorm_2x16
:
1740 result
= nir_unpack_snorm_2x16(&b
, srcs
[0]);
1742 case ir_unop_unpack_snorm_4x8
:
1743 result
= nir_unpack_snorm_4x8(&b
, srcs
[0]);
1745 case ir_unop_unpack_unorm_2x16
:
1746 result
= nir_unpack_unorm_2x16(&b
, srcs
[0]);
1748 case ir_unop_unpack_unorm_4x8
:
1749 result
= nir_unpack_unorm_4x8(&b
, srcs
[0]);
1751 case ir_unop_unpack_half_2x16
:
1752 result
= nir_unpack_half_2x16(&b
, srcs
[0]);
1754 case ir_unop_pack_sampler_2x32
:
1755 case ir_unop_pack_image_2x32
:
1756 case ir_unop_pack_double_2x32
:
1757 case ir_unop_pack_int_2x32
:
1758 case ir_unop_pack_uint_2x32
:
1759 result
= nir_pack_64_2x32(&b
, srcs
[0]);
1761 case ir_unop_unpack_sampler_2x32
:
1762 case ir_unop_unpack_image_2x32
:
1763 case ir_unop_unpack_double_2x32
:
1764 case ir_unop_unpack_int_2x32
:
1765 case ir_unop_unpack_uint_2x32
:
1766 result
= nir_unpack_64_2x32(&b
, srcs
[0]);
1768 case ir_unop_bitfield_reverse
:
1769 result
= nir_bitfield_reverse(&b
, srcs
[0]);
1771 case ir_unop_bit_count
:
1772 result
= nir_bit_count(&b
, srcs
[0]);
1774 case ir_unop_find_msb
:
1776 case GLSL_TYPE_UINT
:
1777 result
= nir_ufind_msb(&b
, srcs
[0]);
1780 result
= nir_ifind_msb(&b
, srcs
[0]);
1783 unreachable("Invalid type for findMSB()");
1786 case ir_unop_find_lsb
:
1787 result
= nir_find_lsb(&b
, srcs
[0]);
1791 switch (ir
->type
->vector_elements
) {
1793 switch (ir
->operands
[0]->type
->vector_elements
) {
1794 case 1: result
= nir_fnoise1_1(&b
, srcs
[0]); break;
1795 case 2: result
= nir_fnoise1_2(&b
, srcs
[0]); break;
1796 case 3: result
= nir_fnoise1_3(&b
, srcs
[0]); break;
1797 case 4: result
= nir_fnoise1_4(&b
, srcs
[0]); break;
1798 default: unreachable("not reached");
1802 switch (ir
->operands
[0]->type
->vector_elements
) {
1803 case 1: result
= nir_fnoise2_1(&b
, srcs
[0]); break;
1804 case 2: result
= nir_fnoise2_2(&b
, srcs
[0]); break;
1805 case 3: result
= nir_fnoise2_3(&b
, srcs
[0]); break;
1806 case 4: result
= nir_fnoise2_4(&b
, srcs
[0]); break;
1807 default: unreachable("not reached");
1811 switch (ir
->operands
[0]->type
->vector_elements
) {
1812 case 1: result
= nir_fnoise3_1(&b
, srcs
[0]); break;
1813 case 2: result
= nir_fnoise3_2(&b
, srcs
[0]); break;
1814 case 3: result
= nir_fnoise3_3(&b
, srcs
[0]); break;
1815 case 4: result
= nir_fnoise3_4(&b
, srcs
[0]); break;
1816 default: unreachable("not reached");
1820 switch (ir
->operands
[0]->type
->vector_elements
) {
1821 case 1: result
= nir_fnoise4_1(&b
, srcs
[0]); break;
1822 case 2: result
= nir_fnoise4_2(&b
, srcs
[0]); break;
1823 case 3: result
= nir_fnoise4_3(&b
, srcs
[0]); break;
1824 case 4: result
= nir_fnoise4_4(&b
, srcs
[0]); break;
1825 default: unreachable("not reached");
1829 unreachable("not reached");
1832 case ir_unop_get_buffer_size
: {
1833 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(
1835 nir_intrinsic_get_buffer_size
);
1836 load
->num_components
= ir
->type
->vector_elements
;
1837 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1838 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1839 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1844 result
= type_is_float(out_type
) ? nir_fadd(&b
, srcs
[0], srcs
[1])
1845 : nir_iadd(&b
, srcs
[0], srcs
[1]);
1848 result
= type_is_float(out_type
) ? nir_fsub(&b
, srcs
[0], srcs
[1])
1849 : nir_isub(&b
, srcs
[0], srcs
[1]);
1852 result
= type_is_float(out_type
) ? nir_fmul(&b
, srcs
[0], srcs
[1])
1853 : nir_imul(&b
, srcs
[0], srcs
[1]);
1856 if (type_is_float(out_type
))
1857 result
= nir_fdiv(&b
, srcs
[0], srcs
[1]);
1858 else if (type_is_signed(out_type
))
1859 result
= nir_idiv(&b
, srcs
[0], srcs
[1]);
1861 result
= nir_udiv(&b
, srcs
[0], srcs
[1]);
1864 result
= type_is_float(out_type
) ? nir_fmod(&b
, srcs
[0], srcs
[1])
1865 : nir_umod(&b
, srcs
[0], srcs
[1]);
1868 if (type_is_float(out_type
))
1869 result
= nir_fmin(&b
, srcs
[0], srcs
[1]);
1870 else if (type_is_signed(out_type
))
1871 result
= nir_imin(&b
, srcs
[0], srcs
[1]);
1873 result
= nir_umin(&b
, srcs
[0], srcs
[1]);
1876 if (type_is_float(out_type
))
1877 result
= nir_fmax(&b
, srcs
[0], srcs
[1]);
1878 else if (type_is_signed(out_type
))
1879 result
= nir_imax(&b
, srcs
[0], srcs
[1]);
1881 result
= nir_umax(&b
, srcs
[0], srcs
[1]);
1883 case ir_binop_pow
: result
= nir_fpow(&b
, srcs
[0], srcs
[1]); break;
1884 case ir_binop_bit_and
: result
= nir_iand(&b
, srcs
[0], srcs
[1]); break;
1885 case ir_binop_bit_or
: result
= nir_ior(&b
, srcs
[0], srcs
[1]); break;
1886 case ir_binop_bit_xor
: result
= nir_ixor(&b
, srcs
[0], srcs
[1]); break;
1887 case ir_binop_logic_and
:
1888 result
= nir_iand(&b
, srcs
[0], srcs
[1]);
1890 case ir_binop_logic_or
:
1891 result
= nir_ior(&b
, srcs
[0], srcs
[1]);
1893 case ir_binop_logic_xor
:
1894 result
= nir_ixor(&b
, srcs
[0], srcs
[1]);
1896 case ir_binop_lshift
: result
= nir_ishl(&b
, srcs
[0], srcs
[1]); break;
1897 case ir_binop_rshift
:
1898 result
= (type_is_signed(out_type
)) ? nir_ishr(&b
, srcs
[0], srcs
[1])
1899 : nir_ushr(&b
, srcs
[0], srcs
[1]);
1901 case ir_binop_imul_high
:
1902 result
= (out_type
== GLSL_TYPE_INT
) ? nir_imul_high(&b
, srcs
[0], srcs
[1])
1903 : nir_umul_high(&b
, srcs
[0], srcs
[1]);
1905 case ir_binop_carry
: result
= nir_uadd_carry(&b
, srcs
[0], srcs
[1]); break;
1906 case ir_binop_borrow
: result
= nir_usub_borrow(&b
, srcs
[0], srcs
[1]); break;
1908 if (type_is_float(types
[0]))
1909 result
= nir_flt(&b
, srcs
[0], srcs
[1]);
1910 else if (type_is_signed(types
[0]))
1911 result
= nir_ilt(&b
, srcs
[0], srcs
[1]);
1913 result
= nir_ult(&b
, srcs
[0], srcs
[1]);
1915 case ir_binop_gequal
:
1916 if (type_is_float(types
[0]))
1917 result
= nir_fge(&b
, srcs
[0], srcs
[1]);
1918 else if (type_is_signed(types
[0]))
1919 result
= nir_ige(&b
, srcs
[0], srcs
[1]);
1921 result
= nir_uge(&b
, srcs
[0], srcs
[1]);
1923 case ir_binop_equal
:
1924 if (type_is_float(types
[0]))
1925 result
= nir_feq(&b
, srcs
[0], srcs
[1]);
1927 result
= nir_ieq(&b
, srcs
[0], srcs
[1]);
1929 case ir_binop_nequal
:
1930 if (type_is_float(types
[0]))
1931 result
= nir_fne(&b
, srcs
[0], srcs
[1]);
1933 result
= nir_ine(&b
, srcs
[0], srcs
[1]);
1935 case ir_binop_all_equal
:
1936 if (type_is_float(types
[0])) {
1937 switch (ir
->operands
[0]->type
->vector_elements
) {
1938 case 1: result
= nir_feq(&b
, srcs
[0], srcs
[1]); break;
1939 case 2: result
= nir_ball_fequal2(&b
, srcs
[0], srcs
[1]); break;
1940 case 3: result
= nir_ball_fequal3(&b
, srcs
[0], srcs
[1]); break;
1941 case 4: result
= nir_ball_fequal4(&b
, srcs
[0], srcs
[1]); break;
1943 unreachable("not reached");
1946 switch (ir
->operands
[0]->type
->vector_elements
) {
1947 case 1: result
= nir_ieq(&b
, srcs
[0], srcs
[1]); break;
1948 case 2: result
= nir_ball_iequal2(&b
, srcs
[0], srcs
[1]); break;
1949 case 3: result
= nir_ball_iequal3(&b
, srcs
[0], srcs
[1]); break;
1950 case 4: result
= nir_ball_iequal4(&b
, srcs
[0], srcs
[1]); break;
1952 unreachable("not reached");
1956 case ir_binop_any_nequal
:
1957 if (type_is_float(types
[0])) {
1958 switch (ir
->operands
[0]->type
->vector_elements
) {
1959 case 1: result
= nir_fne(&b
, srcs
[0], srcs
[1]); break;
1960 case 2: result
= nir_bany_fnequal2(&b
, srcs
[0], srcs
[1]); break;
1961 case 3: result
= nir_bany_fnequal3(&b
, srcs
[0], srcs
[1]); break;
1962 case 4: result
= nir_bany_fnequal4(&b
, srcs
[0], srcs
[1]); break;
1964 unreachable("not reached");
1967 switch (ir
->operands
[0]->type
->vector_elements
) {
1968 case 1: result
= nir_ine(&b
, srcs
[0], srcs
[1]); break;
1969 case 2: result
= nir_bany_inequal2(&b
, srcs
[0], srcs
[1]); break;
1970 case 3: result
= nir_bany_inequal3(&b
, srcs
[0], srcs
[1]); break;
1971 case 4: result
= nir_bany_inequal4(&b
, srcs
[0], srcs
[1]); break;
1973 unreachable("not reached");
1978 switch (ir
->operands
[0]->type
->vector_elements
) {
1979 case 2: result
= nir_fdot2(&b
, srcs
[0], srcs
[1]); break;
1980 case 3: result
= nir_fdot3(&b
, srcs
[0], srcs
[1]); break;
1981 case 4: result
= nir_fdot4(&b
, srcs
[0], srcs
[1]); break;
1983 unreachable("not reached");
1986 case ir_binop_vector_extract
: {
1987 result
= nir_channel(&b
, srcs
[0], 0);
1988 for (unsigned i
= 1; i
< ir
->operands
[0]->type
->vector_elements
; i
++) {
1989 nir_ssa_def
*swizzled
= nir_channel(&b
, srcs
[0], i
);
1990 result
= nir_bcsel(&b
, nir_ieq(&b
, srcs
[1], nir_imm_int(&b
, i
)),
1996 case ir_binop_ldexp
: result
= nir_ldexp(&b
, srcs
[0], srcs
[1]); break;
1998 result
= nir_ffma(&b
, srcs
[0], srcs
[1], srcs
[2]);
2001 result
= nir_flrp(&b
, srcs
[0], srcs
[1], srcs
[2]);
2004 result
= nir_bcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
2006 case ir_triop_bitfield_extract
:
2007 result
= (out_type
== GLSL_TYPE_INT
) ?
2008 nir_ibitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]) :
2009 nir_ubitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]);
2011 case ir_quadop_bitfield_insert
:
2012 result
= nir_bitfield_insert(&b
, srcs
[0], srcs
[1], srcs
[2], srcs
[3]);
2014 case ir_quadop_vector
:
2015 result
= nir_vec(&b
, srcs
, ir
->type
->vector_elements
);
2019 unreachable("not reached");
2024 nir_visitor::visit(ir_swizzle
*ir
)
2026 unsigned swizzle
[4] = { ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
};
2027 result
= nir_swizzle(&b
, evaluate_rvalue(ir
->val
), swizzle
,
2028 ir
->type
->vector_elements
, false);
2032 nir_visitor::visit(ir_texture
*ir
)
2039 num_srcs
= 1; /* coordinate */
2044 op
= (ir
->op
== ir_txb
) ? nir_texop_txb
: nir_texop_txl
;
2045 num_srcs
= 2; /* coordinate, bias/lod */
2049 op
= nir_texop_txd
; /* coordinate, dPdx, dPdy */
2055 if (ir
->lod_info
.lod
!= NULL
)
2056 num_srcs
= 2; /* coordinate, lod */
2058 num_srcs
= 1; /* coordinate */
2062 op
= nir_texop_txf_ms
;
2063 num_srcs
= 2; /* coordinate, sample_index */
2068 if (ir
->lod_info
.lod
!= NULL
)
2069 num_srcs
= 1; /* lod */
2076 num_srcs
= 1; /* coordinate */
2081 num_srcs
= 1; /* coordinate */
2084 case ir_query_levels
:
2085 op
= nir_texop_query_levels
;
2089 case ir_texture_samples
:
2090 op
= nir_texop_texture_samples
;
2094 case ir_samples_identical
:
2095 op
= nir_texop_samples_identical
;
2096 num_srcs
= 1; /* coordinate */
2100 unreachable("not reached");
2103 if (ir
->projector
!= NULL
)
2105 if (ir
->shadow_comparator
!= NULL
)
2107 if (ir
->offset
!= NULL
)
2110 /* Add one for the texture deref */
2113 nir_tex_instr
*instr
= nir_tex_instr_create(this->shader
, num_srcs
);
2116 instr
->sampler_dim
=
2117 (glsl_sampler_dim
) ir
->sampler
->type
->sampler_dimensionality
;
2118 instr
->is_array
= ir
->sampler
->type
->sampler_array
;
2119 instr
->is_shadow
= ir
->sampler
->type
->sampler_shadow
;
2120 if (instr
->is_shadow
)
2121 instr
->is_new_style_shadow
= (ir
->type
->vector_elements
== 1);
2122 switch (ir
->type
->base_type
) {
2123 case GLSL_TYPE_FLOAT
:
2124 instr
->dest_type
= nir_type_float
;
2127 instr
->dest_type
= nir_type_int
;
2129 case GLSL_TYPE_BOOL
:
2130 case GLSL_TYPE_UINT
:
2131 instr
->dest_type
= nir_type_uint
;
2134 unreachable("not reached");
2137 nir_deref_instr
*sampler_deref
= evaluate_deref(ir
->sampler
);
2138 instr
->src
[0].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2139 instr
->src
[0].src_type
= nir_tex_src_texture_deref
;
2140 instr
->src
[1].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2141 instr
->src
[1].src_type
= nir_tex_src_sampler_deref
;
2143 unsigned src_number
= 2;
2145 if (ir
->coordinate
!= NULL
) {
2146 instr
->coord_components
= ir
->coordinate
->type
->vector_elements
;
2147 instr
->src
[src_number
].src
=
2148 nir_src_for_ssa(evaluate_rvalue(ir
->coordinate
));
2149 instr
->src
[src_number
].src_type
= nir_tex_src_coord
;
2153 if (ir
->projector
!= NULL
) {
2154 instr
->src
[src_number
].src
=
2155 nir_src_for_ssa(evaluate_rvalue(ir
->projector
));
2156 instr
->src
[src_number
].src_type
= nir_tex_src_projector
;
2160 if (ir
->shadow_comparator
!= NULL
) {
2161 instr
->src
[src_number
].src
=
2162 nir_src_for_ssa(evaluate_rvalue(ir
->shadow_comparator
));
2163 instr
->src
[src_number
].src_type
= nir_tex_src_comparator
;
2167 if (ir
->offset
!= NULL
) {
2168 /* we don't support multiple offsets yet */
2169 assert(ir
->offset
->type
->is_vector() || ir
->offset
->type
->is_scalar());
2171 instr
->src
[src_number
].src
=
2172 nir_src_for_ssa(evaluate_rvalue(ir
->offset
));
2173 instr
->src
[src_number
].src_type
= nir_tex_src_offset
;
2179 instr
->src
[src_number
].src
=
2180 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.bias
));
2181 instr
->src
[src_number
].src_type
= nir_tex_src_bias
;
2188 if (ir
->lod_info
.lod
!= NULL
) {
2189 instr
->src
[src_number
].src
=
2190 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.lod
));
2191 instr
->src
[src_number
].src_type
= nir_tex_src_lod
;
2197 instr
->src
[src_number
].src
=
2198 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdx
));
2199 instr
->src
[src_number
].src_type
= nir_tex_src_ddx
;
2201 instr
->src
[src_number
].src
=
2202 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdy
));
2203 instr
->src
[src_number
].src_type
= nir_tex_src_ddy
;
2208 instr
->src
[src_number
].src
=
2209 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.sample_index
));
2210 instr
->src
[src_number
].src_type
= nir_tex_src_ms_index
;
2215 instr
->component
= ir
->lod_info
.component
->as_constant()->value
.u
[0];
2222 assert(src_number
== num_srcs
);
2224 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2225 add_instr(&instr
->instr
, nir_tex_instr_dest_size(instr
), bit_size
);
2229 nir_visitor::visit(ir_constant
*ir
)
2232 * We don't know if this variable is an array or struct that gets
2233 * dereferenced, so do the safe thing an make it a variable with a
2234 * constant initializer and return a dereference.
2238 nir_local_variable_create(this->impl
, ir
->type
, "const_temp");
2239 var
->data
.read_only
= true;
2240 var
->constant_initializer
= constant_copy(ir
, var
);
2242 this->deref
= nir_build_deref_var(&b
, var
);
2246 nir_visitor::visit(ir_dereference_variable
*ir
)
2248 if (ir
->variable_referenced()->data
.mode
== ir_var_function_out
) {
2249 unsigned i
= (sig
->return_type
!= glsl_type::void_type
) ? 1 : 0;
2251 foreach_in_list(ir_variable
, param
, &sig
->parameters
) {
2252 if (param
== ir
->variable_referenced()) {
2258 this->deref
= nir_build_deref_cast(&b
, nir_load_param(&b
, i
),
2259 nir_var_function_temp
, ir
->type
, 0);
2263 assert(ir
->variable_referenced()->data
.mode
!= ir_var_function_inout
);
2265 struct hash_entry
*entry
=
2266 _mesa_hash_table_search(this->var_table
, ir
->var
);
2268 nir_variable
*var
= (nir_variable
*) entry
->data
;
2270 this->deref
= nir_build_deref_var(&b
, var
);
2274 nir_visitor::visit(ir_dereference_record
*ir
)
2276 ir
->record
->accept(this);
2278 int field_index
= ir
->field_idx
;
2279 assert(field_index
>= 0);
2281 this->deref
= nir_build_deref_struct(&b
, this->deref
, field_index
);
2285 nir_visitor::visit(ir_dereference_array
*ir
)
2287 nir_ssa_def
*index
= evaluate_rvalue(ir
->array_index
);
2289 ir
->array
->accept(this);
2291 this->deref
= nir_build_deref_array(&b
, this->deref
, index
);
2295 nir_visitor::visit(ir_barrier
*)
2297 nir_intrinsic_instr
*instr
=
2298 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_barrier
);
2299 nir_builder_instr_insert(&b
, &instr
->instr
);