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;
171 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
) {
172 shader
->info
.fs
.pixel_center_integer
= sh
->Program
->info
.fs
.pixel_center_integer
;
173 shader
->info
.fs
.origin_upper_left
= sh
->Program
->info
.fs
.origin_upper_left
;
179 nir_visitor::nir_visitor(nir_shader
*shader
)
181 this->supports_ints
= shader
->options
->native_integers
;
182 this->shader
= shader
;
183 this->is_global
= true;
184 this->var_table
= _mesa_pointer_hash_table_create(NULL
);
185 this->overload_table
= _mesa_pointer_hash_table_create(NULL
);
189 memset(&this->b
, 0, sizeof(this->b
));
192 nir_visitor::~nir_visitor()
194 _mesa_hash_table_destroy(this->var_table
, NULL
);
195 _mesa_hash_table_destroy(this->overload_table
, NULL
);
199 nir_visitor::evaluate_deref(ir_instruction
*ir
)
206 nir_visitor::constant_copy(ir_constant
*ir
, void *mem_ctx
)
211 nir_constant
*ret
= rzalloc(mem_ctx
, nir_constant
);
213 const unsigned rows
= ir
->type
->vector_elements
;
214 const unsigned cols
= ir
->type
->matrix_columns
;
217 ret
->num_elements
= 0;
218 switch (ir
->type
->base_type
) {
220 /* Only float base types can be matrices. */
223 for (unsigned r
= 0; r
< rows
; r
++)
225 ret
->values
[0].u32
[r
] = ir
->value
.u
[r
];
227 ret
->values
[0].f32
[r
] = ir
->value
.u
[r
];
232 /* Only float base types can be matrices. */
235 for (unsigned r
= 0; r
< rows
; r
++)
237 ret
->values
[0].i32
[r
] = ir
->value
.i
[r
];
239 ret
->values
[0].f32
[r
] = ir
->value
.i
[r
];
243 case GLSL_TYPE_FLOAT
:
244 for (unsigned c
= 0; c
< cols
; c
++) {
245 for (unsigned r
= 0; r
< rows
; r
++)
246 ret
->values
[c
].f32
[r
] = ir
->value
.f
[c
* rows
+ r
];
250 case GLSL_TYPE_DOUBLE
:
251 for (unsigned c
= 0; c
< cols
; c
++) {
252 for (unsigned r
= 0; r
< rows
; r
++)
253 ret
->values
[c
].f64
[r
] = ir
->value
.d
[c
* rows
+ r
];
257 case GLSL_TYPE_UINT64
:
258 /* Only float base types can be matrices. */
261 for (unsigned r
= 0; r
< rows
; r
++)
262 ret
->values
[0].u64
[r
] = ir
->value
.u64
[r
];
265 case GLSL_TYPE_INT64
:
266 /* Only float base types can be matrices. */
269 for (unsigned r
= 0; r
< rows
; r
++)
270 ret
->values
[0].i64
[r
] = ir
->value
.i64
[r
];
274 /* Only float base types can be matrices. */
277 for (unsigned r
= 0; r
< rows
; r
++)
278 ret
->values
[0].b
[r
] = ir
->value
.b
[r
];
282 case GLSL_TYPE_STRUCT
:
283 case GLSL_TYPE_ARRAY
:
284 ret
->elements
= ralloc_array(mem_ctx
, nir_constant
*,
286 ret
->num_elements
= ir
->type
->length
;
288 for (i
= 0; i
< ir
->type
->length
; i
++)
289 ret
->elements
[i
] = constant_copy(ir
->const_elements
[i
], mem_ctx
);
293 unreachable("not reached");
300 nir_visitor::visit(ir_variable
*ir
)
302 /* TODO: In future we should switch to using the NIR lowering pass but for
303 * now just ignore these variables as GLSL IR should have lowered them.
304 * Anything remaining are just dead vars that weren't cleaned up.
306 if (ir
->data
.mode
== ir_var_shader_shared
)
309 /* FINISHME: inout parameters */
310 assert(ir
->data
.mode
!= ir_var_function_inout
);
312 if (ir
->data
.mode
== ir_var_function_out
)
315 nir_variable
*var
= rzalloc(shader
, nir_variable
);
316 var
->type
= ir
->type
;
317 var
->name
= ralloc_strdup(var
, ir
->name
);
319 var
->data
.always_active_io
= ir
->data
.always_active_io
;
320 var
->data
.read_only
= ir
->data
.read_only
;
321 var
->data
.centroid
= ir
->data
.centroid
;
322 var
->data
.sample
= ir
->data
.sample
;
323 var
->data
.patch
= ir
->data
.patch
;
324 var
->data
.invariant
= ir
->data
.invariant
;
325 var
->data
.location
= ir
->data
.location
;
326 var
->data
.stream
= ir
->data
.stream
;
327 var
->data
.compact
= false;
329 switch(ir
->data
.mode
) {
331 case ir_var_temporary
:
333 var
->data
.mode
= nir_var_shader_temp
;
335 var
->data
.mode
= nir_var_function_temp
;
338 case ir_var_function_in
:
339 case ir_var_const_in
:
340 var
->data
.mode
= nir_var_function_temp
;
343 case ir_var_shader_in
:
344 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
345 ir
->data
.location
== VARYING_SLOT_FACE
) {
346 /* For whatever reason, GLSL IR makes gl_FrontFacing an input */
347 var
->data
.location
= SYSTEM_VALUE_FRONT_FACE
;
348 var
->data
.mode
= nir_var_system_value
;
349 } else if (shader
->info
.stage
== MESA_SHADER_GEOMETRY
&&
350 ir
->data
.location
== VARYING_SLOT_PRIMITIVE_ID
) {
351 /* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */
352 var
->data
.location
= SYSTEM_VALUE_PRIMITIVE_ID
;
353 var
->data
.mode
= nir_var_system_value
;
355 var
->data
.mode
= nir_var_shader_in
;
357 if (shader
->info
.stage
== MESA_SHADER_TESS_EVAL
&&
358 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
359 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
360 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
363 if (shader
->info
.stage
> MESA_SHADER_VERTEX
&&
364 ir
->data
.location
>= VARYING_SLOT_CLIP_DIST0
&&
365 ir
->data
.location
<= VARYING_SLOT_CULL_DIST1
) {
366 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
371 case ir_var_shader_out
:
372 var
->data
.mode
= nir_var_shader_out
;
373 if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
&&
374 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
375 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
376 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
379 if (shader
->info
.stage
<= MESA_SHADER_GEOMETRY
&&
380 ir
->data
.location
>= VARYING_SLOT_CLIP_DIST0
&&
381 ir
->data
.location
<= VARYING_SLOT_CULL_DIST1
) {
382 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
387 if (ir
->get_interface_type())
388 var
->data
.mode
= nir_var_mem_ubo
;
390 var
->data
.mode
= nir_var_uniform
;
393 case ir_var_shader_storage
:
394 var
->data
.mode
= nir_var_mem_ssbo
;
397 case ir_var_system_value
:
398 var
->data
.mode
= nir_var_system_value
;
402 unreachable("not reached");
405 var
->data
.interpolation
= ir
->data
.interpolation
;
406 var
->data
.location_frac
= ir
->data
.location_frac
;
408 switch (ir
->data
.depth_layout
) {
409 case ir_depth_layout_none
:
410 var
->data
.depth_layout
= nir_depth_layout_none
;
412 case ir_depth_layout_any
:
413 var
->data
.depth_layout
= nir_depth_layout_any
;
415 case ir_depth_layout_greater
:
416 var
->data
.depth_layout
= nir_depth_layout_greater
;
418 case ir_depth_layout_less
:
419 var
->data
.depth_layout
= nir_depth_layout_less
;
421 case ir_depth_layout_unchanged
:
422 var
->data
.depth_layout
= nir_depth_layout_unchanged
;
425 unreachable("not reached");
428 var
->data
.index
= ir
->data
.index
;
429 var
->data
.descriptor_set
= 0;
430 var
->data
.binding
= ir
->data
.binding
;
431 var
->data
.explicit_binding
= ir
->data
.explicit_binding
;
432 var
->data
.bindless
= ir
->data
.bindless
;
433 var
->data
.offset
= ir
->data
.offset
;
435 unsigned image_access
= 0;
436 if (ir
->data
.memory_read_only
)
437 image_access
|= ACCESS_NON_WRITEABLE
;
438 if (ir
->data
.memory_write_only
)
439 image_access
|= ACCESS_NON_READABLE
;
440 if (ir
->data
.memory_coherent
)
441 image_access
|= ACCESS_COHERENT
;
442 if (ir
->data
.memory_volatile
)
443 image_access
|= ACCESS_VOLATILE
;
444 if (ir
->data
.memory_restrict
)
445 image_access
|= ACCESS_RESTRICT
;
446 var
->data
.image
.access
= (gl_access_qualifier
)image_access
;
447 var
->data
.image
.format
= ir
->data
.image_format
;
449 var
->data
.fb_fetch_output
= ir
->data
.fb_fetch_output
;
450 var
->data
.explicit_xfb_buffer
= ir
->data
.explicit_xfb_buffer
;
451 var
->data
.explicit_xfb_stride
= ir
->data
.explicit_xfb_stride
;
452 var
->data
.xfb_buffer
= ir
->data
.xfb_buffer
;
453 var
->data
.xfb_stride
= ir
->data
.xfb_stride
;
455 var
->num_state_slots
= ir
->get_num_state_slots();
456 if (var
->num_state_slots
> 0) {
457 var
->state_slots
= rzalloc_array(var
, nir_state_slot
,
458 var
->num_state_slots
);
460 ir_state_slot
*state_slots
= ir
->get_state_slots();
461 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
462 for (unsigned j
= 0; j
< 5; j
++)
463 var
->state_slots
[i
].tokens
[j
] = state_slots
[i
].tokens
[j
];
464 var
->state_slots
[i
].swizzle
= state_slots
[i
].swizzle
;
467 var
->state_slots
= NULL
;
470 var
->constant_initializer
= constant_copy(ir
->constant_initializer
, var
);
472 var
->interface_type
= ir
->get_interface_type();
474 if (var
->data
.mode
== nir_var_function_temp
)
475 nir_function_impl_add_variable(impl
, var
);
477 nir_shader_add_variable(shader
, var
);
479 _mesa_hash_table_insert(var_table
, ir
, var
);
483 nir_function_visitor::visit_enter(ir_function
*ir
)
485 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
) {
486 visitor
->create_function(sig
);
488 return visit_continue_with_parent
;
492 nir_visitor::create_function(ir_function_signature
*ir
)
494 if (ir
->is_intrinsic())
497 nir_function
*func
= nir_function_create(shader
, ir
->function_name());
498 if (strcmp(ir
->function_name(), "main") == 0)
499 func
->is_entrypoint
= true;
501 func
->num_params
= ir
->parameters
.length() +
502 (ir
->return_type
!= glsl_type::void_type
);
503 func
->params
= ralloc_array(shader
, nir_parameter
, func
->num_params
);
507 if (ir
->return_type
!= glsl_type::void_type
) {
508 /* The return value is a variable deref (basically an out parameter) */
509 func
->params
[np
].num_components
= 1;
510 func
->params
[np
].bit_size
= 32;
514 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
515 /* FINISHME: pass arrays, structs, etc by reference? */
516 assert(param
->type
->is_vector() || param
->type
->is_scalar());
518 if (param
->data
.mode
== ir_var_function_in
) {
519 func
->params
[np
].num_components
= param
->type
->vector_elements
;
520 func
->params
[np
].bit_size
= glsl_get_bit_size(param
->type
);
522 func
->params
[np
].num_components
= 1;
523 func
->params
[np
].bit_size
= 32;
527 assert(np
== func
->num_params
);
529 _mesa_hash_table_insert(this->overload_table
, ir
, func
);
533 nir_visitor::visit(ir_function
*ir
)
535 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
)
540 nir_visitor::visit(ir_function_signature
*ir
)
542 if (ir
->is_intrinsic())
547 struct hash_entry
*entry
=
548 _mesa_hash_table_search(this->overload_table
, ir
);
551 nir_function
*func
= (nir_function
*) entry
->data
;
553 if (ir
->is_defined
) {
554 nir_function_impl
*impl
= nir_function_impl_create(func
);
557 this->is_global
= false;
559 nir_builder_init(&b
, impl
);
560 b
.cursor
= nir_after_cf_list(&impl
->body
);
562 unsigned i
= (ir
->return_type
!= glsl_type::void_type
) ? 1 : 0;
564 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
566 nir_local_variable_create(impl
, param
->type
, param
->name
);
568 if (param
->data
.mode
== ir_var_function_in
) {
569 nir_store_var(&b
, var
, nir_load_param(&b
, i
), ~0);
572 _mesa_hash_table_insert(var_table
, param
, var
);
576 visit_exec_list(&ir
->body
, this);
578 this->is_global
= true;
585 nir_visitor::visit(ir_loop
*ir
)
588 visit_exec_list(&ir
->body_instructions
, this);
589 nir_pop_loop(&b
, NULL
);
593 nir_visitor::visit(ir_if
*ir
)
595 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
596 visit_exec_list(&ir
->then_instructions
, this);
597 nir_push_else(&b
, NULL
);
598 visit_exec_list(&ir
->else_instructions
, this);
599 nir_pop_if(&b
, NULL
);
603 nir_visitor::visit(ir_discard
*ir
)
606 * discards aren't treated as control flow, because before we lower them
607 * they can appear anywhere in the shader and the stuff after them may still
608 * be executed (yay, crazy GLSL rules!). However, after lowering, all the
609 * discards will be immediately followed by a return.
612 nir_intrinsic_instr
*discard
;
614 discard
= nir_intrinsic_instr_create(this->shader
,
615 nir_intrinsic_discard_if
);
617 nir_src_for_ssa(evaluate_rvalue(ir
->condition
));
619 discard
= nir_intrinsic_instr_create(this->shader
, nir_intrinsic_discard
);
622 nir_builder_instr_insert(&b
, &discard
->instr
);
626 nir_visitor::visit(ir_emit_vertex
*ir
)
628 nir_intrinsic_instr
*instr
=
629 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_emit_vertex
);
630 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
631 nir_builder_instr_insert(&b
, &instr
->instr
);
635 nir_visitor::visit(ir_end_primitive
*ir
)
637 nir_intrinsic_instr
*instr
=
638 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_end_primitive
);
639 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
640 nir_builder_instr_insert(&b
, &instr
->instr
);
644 nir_visitor::visit(ir_loop_jump
*ir
)
648 case ir_loop_jump::jump_break
:
649 type
= nir_jump_break
;
651 case ir_loop_jump::jump_continue
:
652 type
= nir_jump_continue
;
655 unreachable("not reached");
658 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, type
);
659 nir_builder_instr_insert(&b
, &instr
->instr
);
663 nir_visitor::visit(ir_return
*ir
)
665 if (ir
->value
!= NULL
) {
666 nir_deref_instr
*ret_deref
=
667 nir_build_deref_cast(&b
, nir_load_param(&b
, 0),
668 nir_var_function_temp
, ir
->value
->type
, 0);
670 nir_ssa_def
*val
= evaluate_rvalue(ir
->value
);
671 nir_store_deref(&b
, ret_deref
, val
, ~0);
674 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, nir_jump_return
);
675 nir_builder_instr_insert(&b
, &instr
->instr
);
679 intrinsic_set_std430_align(nir_intrinsic_instr
*intrin
, const glsl_type
*type
)
681 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
682 unsigned pow2_components
= util_next_power_of_two(type
->vector_elements
);
683 nir_intrinsic_set_align(intrin
, (bit_size
/ 8) * pow2_components
, 0);
687 nir_visitor::visit(ir_call
*ir
)
689 if (ir
->callee
->is_intrinsic()) {
692 switch (ir
->callee
->intrinsic_id
) {
693 case ir_intrinsic_atomic_counter_read
:
694 op
= nir_intrinsic_atomic_counter_read_deref
;
696 case ir_intrinsic_atomic_counter_increment
:
697 op
= nir_intrinsic_atomic_counter_inc_deref
;
699 case ir_intrinsic_atomic_counter_predecrement
:
700 op
= nir_intrinsic_atomic_counter_pre_dec_deref
;
702 case ir_intrinsic_atomic_counter_add
:
703 op
= nir_intrinsic_atomic_counter_add_deref
;
705 case ir_intrinsic_atomic_counter_and
:
706 op
= nir_intrinsic_atomic_counter_and_deref
;
708 case ir_intrinsic_atomic_counter_or
:
709 op
= nir_intrinsic_atomic_counter_or_deref
;
711 case ir_intrinsic_atomic_counter_xor
:
712 op
= nir_intrinsic_atomic_counter_xor_deref
;
714 case ir_intrinsic_atomic_counter_min
:
715 op
= nir_intrinsic_atomic_counter_min_deref
;
717 case ir_intrinsic_atomic_counter_max
:
718 op
= nir_intrinsic_atomic_counter_max_deref
;
720 case ir_intrinsic_atomic_counter_exchange
:
721 op
= nir_intrinsic_atomic_counter_exchange_deref
;
723 case ir_intrinsic_atomic_counter_comp_swap
:
724 op
= nir_intrinsic_atomic_counter_comp_swap_deref
;
726 case ir_intrinsic_image_load
:
727 op
= nir_intrinsic_image_deref_load
;
729 case ir_intrinsic_image_store
:
730 op
= nir_intrinsic_image_deref_store
;
732 case ir_intrinsic_image_atomic_add
:
733 op
= ir
->return_deref
->type
->is_integer_32_64()
734 ? nir_intrinsic_image_deref_atomic_add
735 : nir_intrinsic_image_deref_atomic_fadd
;
737 case ir_intrinsic_image_atomic_min
:
738 op
= nir_intrinsic_image_deref_atomic_min
;
740 case ir_intrinsic_image_atomic_max
:
741 op
= nir_intrinsic_image_deref_atomic_max
;
743 case ir_intrinsic_image_atomic_and
:
744 op
= nir_intrinsic_image_deref_atomic_and
;
746 case ir_intrinsic_image_atomic_or
:
747 op
= nir_intrinsic_image_deref_atomic_or
;
749 case ir_intrinsic_image_atomic_xor
:
750 op
= nir_intrinsic_image_deref_atomic_xor
;
752 case ir_intrinsic_image_atomic_exchange
:
753 op
= nir_intrinsic_image_deref_atomic_exchange
;
755 case ir_intrinsic_image_atomic_comp_swap
:
756 op
= nir_intrinsic_image_deref_atomic_comp_swap
;
758 case ir_intrinsic_memory_barrier
:
759 op
= nir_intrinsic_memory_barrier
;
761 case ir_intrinsic_image_size
:
762 op
= nir_intrinsic_image_deref_size
;
764 case ir_intrinsic_image_samples
:
765 op
= nir_intrinsic_image_deref_samples
;
767 case ir_intrinsic_ssbo_store
:
768 op
= nir_intrinsic_store_ssbo
;
770 case ir_intrinsic_ssbo_load
:
771 op
= nir_intrinsic_load_ssbo
;
773 case ir_intrinsic_ssbo_atomic_add
:
774 op
= ir
->return_deref
->type
->is_integer_32_64()
775 ? nir_intrinsic_ssbo_atomic_add
: nir_intrinsic_ssbo_atomic_fadd
;
777 case ir_intrinsic_ssbo_atomic_and
:
778 op
= nir_intrinsic_ssbo_atomic_and
;
780 case ir_intrinsic_ssbo_atomic_or
:
781 op
= nir_intrinsic_ssbo_atomic_or
;
783 case ir_intrinsic_ssbo_atomic_xor
:
784 op
= nir_intrinsic_ssbo_atomic_xor
;
786 case ir_intrinsic_ssbo_atomic_min
:
787 assert(ir
->return_deref
);
788 if (ir
->return_deref
->type
== glsl_type::int_type
)
789 op
= nir_intrinsic_ssbo_atomic_imin
;
790 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
791 op
= nir_intrinsic_ssbo_atomic_umin
;
792 else if (ir
->return_deref
->type
== glsl_type::float_type
)
793 op
= nir_intrinsic_ssbo_atomic_fmin
;
795 unreachable("Invalid type");
797 case ir_intrinsic_ssbo_atomic_max
:
798 assert(ir
->return_deref
);
799 if (ir
->return_deref
->type
== glsl_type::int_type
)
800 op
= nir_intrinsic_ssbo_atomic_imax
;
801 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
802 op
= nir_intrinsic_ssbo_atomic_umax
;
803 else if (ir
->return_deref
->type
== glsl_type::float_type
)
804 op
= nir_intrinsic_ssbo_atomic_fmax
;
806 unreachable("Invalid type");
808 case ir_intrinsic_ssbo_atomic_exchange
:
809 op
= nir_intrinsic_ssbo_atomic_exchange
;
811 case ir_intrinsic_ssbo_atomic_comp_swap
:
812 op
= ir
->return_deref
->type
->is_integer_32_64()
813 ? nir_intrinsic_ssbo_atomic_comp_swap
814 : nir_intrinsic_ssbo_atomic_fcomp_swap
;
816 case ir_intrinsic_shader_clock
:
817 op
= nir_intrinsic_shader_clock
;
819 case ir_intrinsic_begin_invocation_interlock
:
820 op
= nir_intrinsic_begin_invocation_interlock
;
822 case ir_intrinsic_end_invocation_interlock
:
823 op
= nir_intrinsic_end_invocation_interlock
;
825 case ir_intrinsic_group_memory_barrier
:
826 op
= nir_intrinsic_group_memory_barrier
;
828 case ir_intrinsic_memory_barrier_atomic_counter
:
829 op
= nir_intrinsic_memory_barrier_atomic_counter
;
831 case ir_intrinsic_memory_barrier_buffer
:
832 op
= nir_intrinsic_memory_barrier_buffer
;
834 case ir_intrinsic_memory_barrier_image
:
835 op
= nir_intrinsic_memory_barrier_image
;
837 case ir_intrinsic_memory_barrier_shared
:
838 op
= nir_intrinsic_memory_barrier_shared
;
840 case ir_intrinsic_shared_load
:
841 op
= nir_intrinsic_load_shared
;
843 case ir_intrinsic_shared_store
:
844 op
= nir_intrinsic_store_shared
;
846 case ir_intrinsic_shared_atomic_add
:
847 op
= ir
->return_deref
->type
->is_integer_32_64()
848 ? nir_intrinsic_shared_atomic_add
849 : nir_intrinsic_shared_atomic_fadd
;
851 case ir_intrinsic_shared_atomic_and
:
852 op
= nir_intrinsic_shared_atomic_and
;
854 case ir_intrinsic_shared_atomic_or
:
855 op
= nir_intrinsic_shared_atomic_or
;
857 case ir_intrinsic_shared_atomic_xor
:
858 op
= nir_intrinsic_shared_atomic_xor
;
860 case ir_intrinsic_shared_atomic_min
:
861 assert(ir
->return_deref
);
862 if (ir
->return_deref
->type
== glsl_type::int_type
)
863 op
= nir_intrinsic_shared_atomic_imin
;
864 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
865 op
= nir_intrinsic_shared_atomic_umin
;
866 else if (ir
->return_deref
->type
== glsl_type::float_type
)
867 op
= nir_intrinsic_shared_atomic_fmin
;
869 unreachable("Invalid type");
871 case ir_intrinsic_shared_atomic_max
:
872 assert(ir
->return_deref
);
873 if (ir
->return_deref
->type
== glsl_type::int_type
)
874 op
= nir_intrinsic_shared_atomic_imax
;
875 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
876 op
= nir_intrinsic_shared_atomic_umax
;
877 else if (ir
->return_deref
->type
== glsl_type::float_type
)
878 op
= nir_intrinsic_shared_atomic_fmax
;
880 unreachable("Invalid type");
882 case ir_intrinsic_shared_atomic_exchange
:
883 op
= nir_intrinsic_shared_atomic_exchange
;
885 case ir_intrinsic_shared_atomic_comp_swap
:
886 op
= ir
->return_deref
->type
->is_integer_32_64()
887 ? nir_intrinsic_shared_atomic_comp_swap
888 : nir_intrinsic_shared_atomic_fcomp_swap
;
890 case ir_intrinsic_vote_any
:
891 op
= nir_intrinsic_vote_any
;
893 case ir_intrinsic_vote_all
:
894 op
= nir_intrinsic_vote_all
;
896 case ir_intrinsic_vote_eq
:
897 op
= nir_intrinsic_vote_ieq
;
899 case ir_intrinsic_ballot
:
900 op
= nir_intrinsic_ballot
;
902 case ir_intrinsic_read_invocation
:
903 op
= nir_intrinsic_read_invocation
;
905 case ir_intrinsic_read_first_invocation
:
906 op
= nir_intrinsic_read_first_invocation
;
909 unreachable("not reached");
912 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(shader
, op
);
913 nir_ssa_def
*ret
= &instr
->dest
.ssa
;
916 case nir_intrinsic_atomic_counter_read_deref
:
917 case nir_intrinsic_atomic_counter_inc_deref
:
918 case nir_intrinsic_atomic_counter_pre_dec_deref
:
919 case nir_intrinsic_atomic_counter_add_deref
:
920 case nir_intrinsic_atomic_counter_min_deref
:
921 case nir_intrinsic_atomic_counter_max_deref
:
922 case nir_intrinsic_atomic_counter_and_deref
:
923 case nir_intrinsic_atomic_counter_or_deref
:
924 case nir_intrinsic_atomic_counter_xor_deref
:
925 case nir_intrinsic_atomic_counter_exchange_deref
:
926 case nir_intrinsic_atomic_counter_comp_swap_deref
: {
927 /* Set the counter variable dereference. */
928 exec_node
*param
= ir
->actual_parameters
.get_head();
929 ir_dereference
*counter
= (ir_dereference
*)param
;
931 instr
->src
[0] = nir_src_for_ssa(&evaluate_deref(counter
)->dest
.ssa
);
932 param
= param
->get_next();
934 /* Set the intrinsic destination. */
935 if (ir
->return_deref
) {
936 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
939 /* Set the intrinsic parameters. */
940 if (!param
->is_tail_sentinel()) {
942 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
943 param
= param
->get_next();
946 if (!param
->is_tail_sentinel()) {
948 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
949 param
= param
->get_next();
952 nir_builder_instr_insert(&b
, &instr
->instr
);
955 case nir_intrinsic_image_deref_load
:
956 case nir_intrinsic_image_deref_store
:
957 case nir_intrinsic_image_deref_atomic_add
:
958 case nir_intrinsic_image_deref_atomic_min
:
959 case nir_intrinsic_image_deref_atomic_max
:
960 case nir_intrinsic_image_deref_atomic_and
:
961 case nir_intrinsic_image_deref_atomic_or
:
962 case nir_intrinsic_image_deref_atomic_xor
:
963 case nir_intrinsic_image_deref_atomic_exchange
:
964 case nir_intrinsic_image_deref_atomic_comp_swap
:
965 case nir_intrinsic_image_deref_atomic_fadd
:
966 case nir_intrinsic_image_deref_samples
:
967 case nir_intrinsic_image_deref_size
: {
968 nir_ssa_undef_instr
*instr_undef
=
969 nir_ssa_undef_instr_create(shader
, 1, 32);
970 nir_builder_instr_insert(&b
, &instr_undef
->instr
);
972 /* Set the image variable dereference. */
973 exec_node
*param
= ir
->actual_parameters
.get_head();
974 ir_dereference
*image
= (ir_dereference
*)param
;
975 const glsl_type
*type
=
976 image
->variable_referenced()->type
->without_array();
978 instr
->src
[0] = nir_src_for_ssa(&evaluate_deref(image
)->dest
.ssa
);
979 param
= param
->get_next();
981 /* Set the intrinsic destination. */
982 if (ir
->return_deref
) {
983 unsigned num_components
= ir
->return_deref
->type
->vector_elements
;
984 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
985 num_components
, 32, NULL
);
988 if (op
== nir_intrinsic_image_deref_size
) {
989 instr
->num_components
= instr
->dest
.ssa
.num_components
;
990 } else if (op
== nir_intrinsic_image_deref_load
||
991 op
== nir_intrinsic_image_deref_store
) {
992 instr
->num_components
= 4;
995 if (op
== nir_intrinsic_image_deref_size
||
996 op
== nir_intrinsic_image_deref_samples
) {
997 nir_builder_instr_insert(&b
, &instr
->instr
);
1001 /* Set the address argument, extending the coordinate vector to four
1004 nir_ssa_def
*src_addr
=
1005 evaluate_rvalue((ir_dereference
*)param
);
1006 nir_ssa_def
*srcs
[4];
1008 for (int i
= 0; i
< 4; i
++) {
1009 if (i
< type
->coordinate_components())
1010 srcs
[i
] = nir_channel(&b
, src_addr
, i
);
1012 srcs
[i
] = &instr_undef
->def
;
1015 instr
->src
[1] = nir_src_for_ssa(nir_vec(&b
, srcs
, 4));
1016 param
= param
->get_next();
1018 /* Set the sample argument, which is undefined for single-sample
1021 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_MS
) {
1023 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1024 param
= param
->get_next();
1026 instr
->src
[2] = nir_src_for_ssa(&instr_undef
->def
);
1029 /* Set the intrinsic parameters. */
1030 if (!param
->is_tail_sentinel()) {
1032 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1033 param
= param
->get_next();
1036 if (!param
->is_tail_sentinel()) {
1038 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1039 param
= param
->get_next();
1041 nir_builder_instr_insert(&b
, &instr
->instr
);
1044 case nir_intrinsic_memory_barrier
:
1045 case nir_intrinsic_group_memory_barrier
:
1046 case nir_intrinsic_memory_barrier_atomic_counter
:
1047 case nir_intrinsic_memory_barrier_buffer
:
1048 case nir_intrinsic_memory_barrier_image
:
1049 case nir_intrinsic_memory_barrier_shared
:
1050 nir_builder_instr_insert(&b
, &instr
->instr
);
1052 case nir_intrinsic_shader_clock
:
1053 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 2, 32, NULL
);
1054 instr
->num_components
= 2;
1055 nir_builder_instr_insert(&b
, &instr
->instr
);
1057 case nir_intrinsic_begin_invocation_interlock
:
1058 nir_builder_instr_insert(&b
, &instr
->instr
);
1060 case nir_intrinsic_end_invocation_interlock
:
1061 nir_builder_instr_insert(&b
, &instr
->instr
);
1063 case nir_intrinsic_store_ssbo
: {
1064 exec_node
*param
= ir
->actual_parameters
.get_head();
1065 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1067 param
= param
->get_next();
1068 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1070 param
= param
->get_next();
1071 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1073 param
= param
->get_next();
1074 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1077 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1078 if (val
->type
->is_boolean())
1079 nir_val
= nir_b2i32(&b
, nir_val
);
1081 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1082 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(block
));
1083 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(offset
));
1084 intrinsic_set_std430_align(instr
, val
->type
);
1085 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1086 instr
->num_components
= val
->type
->vector_elements
;
1088 nir_builder_instr_insert(&b
, &instr
->instr
);
1091 case nir_intrinsic_load_ssbo
: {
1092 exec_node
*param
= ir
->actual_parameters
.get_head();
1093 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1095 param
= param
->get_next();
1096 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1098 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(block
));
1099 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1101 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1102 instr
->num_components
= type
->vector_elements
;
1103 intrinsic_set_std430_align(instr
, type
);
1105 /* Setup destination register */
1106 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1107 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1108 type
->vector_elements
, bit_size
, NULL
);
1110 /* Insert the created nir instruction now since in the case of boolean
1111 * result we will need to emit another instruction after it
1113 nir_builder_instr_insert(&b
, &instr
->instr
);
1116 * In SSBO/UBO's, a true boolean value is any non-zero value, but we
1117 * consider a true boolean to be ~0. Fix this up with a != 0
1120 if (type
->is_boolean())
1121 ret
= nir_i2b(&b
, &instr
->dest
.ssa
);
1124 case nir_intrinsic_ssbo_atomic_add
:
1125 case nir_intrinsic_ssbo_atomic_imin
:
1126 case nir_intrinsic_ssbo_atomic_umin
:
1127 case nir_intrinsic_ssbo_atomic_imax
:
1128 case nir_intrinsic_ssbo_atomic_umax
:
1129 case nir_intrinsic_ssbo_atomic_and
:
1130 case nir_intrinsic_ssbo_atomic_or
:
1131 case nir_intrinsic_ssbo_atomic_xor
:
1132 case nir_intrinsic_ssbo_atomic_exchange
:
1133 case nir_intrinsic_ssbo_atomic_comp_swap
:
1134 case nir_intrinsic_ssbo_atomic_fadd
:
1135 case nir_intrinsic_ssbo_atomic_fmin
:
1136 case nir_intrinsic_ssbo_atomic_fmax
:
1137 case nir_intrinsic_ssbo_atomic_fcomp_swap
: {
1138 int param_count
= ir
->actual_parameters
.length();
1139 assert(param_count
== 3 || param_count
== 4);
1142 exec_node
*param
= ir
->actual_parameters
.get_head();
1143 ir_instruction
*inst
= (ir_instruction
*) param
;
1144 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1147 param
= param
->get_next();
1148 inst
= (ir_instruction
*) param
;
1149 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1151 /* data1 parameter (this is always present) */
1152 param
= param
->get_next();
1153 inst
= (ir_instruction
*) param
;
1154 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1156 /* data2 parameter (only with atomic_comp_swap) */
1157 if (param_count
== 4) {
1158 assert(op
== nir_intrinsic_ssbo_atomic_comp_swap
||
1159 op
== nir_intrinsic_ssbo_atomic_fcomp_swap
);
1160 param
= param
->get_next();
1161 inst
= (ir_instruction
*) param
;
1162 instr
->src
[3] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1166 assert(ir
->return_deref
);
1167 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1168 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1169 nir_builder_instr_insert(&b
, &instr
->instr
);
1172 case nir_intrinsic_load_shared
: {
1173 exec_node
*param
= ir
->actual_parameters
.get_head();
1174 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1176 nir_intrinsic_set_base(instr
, 0);
1177 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(offset
));
1179 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1180 instr
->num_components
= type
->vector_elements
;
1181 intrinsic_set_std430_align(instr
, type
);
1183 /* Setup destination register */
1184 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1185 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1186 type
->vector_elements
, bit_size
, NULL
);
1188 nir_builder_instr_insert(&b
, &instr
->instr
);
1190 /* The value in shared memory is a 32-bit value */
1191 if (type
->is_boolean())
1192 ret
= nir_i2b(&b
, &instr
->dest
.ssa
);
1195 case nir_intrinsic_store_shared
: {
1196 exec_node
*param
= ir
->actual_parameters
.get_head();
1197 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1199 param
= param
->get_next();
1200 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1202 param
= param
->get_next();
1203 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1206 nir_intrinsic_set_base(instr
, 0);
1207 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1209 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1211 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1212 /* The value in shared memory is a 32-bit value */
1213 if (val
->type
->is_boolean())
1214 nir_val
= nir_b2i32(&b
, nir_val
);
1216 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1217 instr
->num_components
= val
->type
->vector_elements
;
1218 intrinsic_set_std430_align(instr
, val
->type
);
1220 nir_builder_instr_insert(&b
, &instr
->instr
);
1223 case nir_intrinsic_shared_atomic_add
:
1224 case nir_intrinsic_shared_atomic_imin
:
1225 case nir_intrinsic_shared_atomic_umin
:
1226 case nir_intrinsic_shared_atomic_imax
:
1227 case nir_intrinsic_shared_atomic_umax
:
1228 case nir_intrinsic_shared_atomic_and
:
1229 case nir_intrinsic_shared_atomic_or
:
1230 case nir_intrinsic_shared_atomic_xor
:
1231 case nir_intrinsic_shared_atomic_exchange
:
1232 case nir_intrinsic_shared_atomic_comp_swap
:
1233 case nir_intrinsic_shared_atomic_fadd
:
1234 case nir_intrinsic_shared_atomic_fmin
:
1235 case nir_intrinsic_shared_atomic_fmax
:
1236 case nir_intrinsic_shared_atomic_fcomp_swap
: {
1237 int param_count
= ir
->actual_parameters
.length();
1238 assert(param_count
== 2 || param_count
== 3);
1241 exec_node
*param
= ir
->actual_parameters
.get_head();
1242 ir_instruction
*inst
= (ir_instruction
*) param
;
1243 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1245 /* data1 parameter (this is always present) */
1246 param
= param
->get_next();
1247 inst
= (ir_instruction
*) param
;
1248 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1250 /* data2 parameter (only with atomic_comp_swap) */
1251 if (param_count
== 3) {
1252 assert(op
== nir_intrinsic_shared_atomic_comp_swap
||
1253 op
== nir_intrinsic_shared_atomic_fcomp_swap
);
1254 param
= param
->get_next();
1255 inst
= (ir_instruction
*) param
;
1257 nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1261 assert(ir
->return_deref
);
1262 unsigned bit_size
= glsl_get_bit_size(ir
->return_deref
->type
);
1263 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1264 ir
->return_deref
->type
->vector_elements
,
1266 nir_builder_instr_insert(&b
, &instr
->instr
);
1269 case nir_intrinsic_vote_any
:
1270 case nir_intrinsic_vote_all
:
1271 case nir_intrinsic_vote_ieq
: {
1272 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 1, NULL
);
1273 instr
->num_components
= 1;
1275 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1276 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1278 nir_builder_instr_insert(&b
, &instr
->instr
);
1282 case nir_intrinsic_ballot
: {
1283 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1284 ir
->return_deref
->type
->vector_elements
, 64, NULL
);
1285 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1287 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1288 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1290 nir_builder_instr_insert(&b
, &instr
->instr
);
1293 case nir_intrinsic_read_invocation
: {
1294 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1295 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1296 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1298 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1299 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1301 ir_rvalue
*invocation
= (ir_rvalue
*) ir
->actual_parameters
.get_head()->next
;
1302 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(invocation
));
1304 nir_builder_instr_insert(&b
, &instr
->instr
);
1307 case nir_intrinsic_read_first_invocation
: {
1308 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1309 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1310 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1312 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1313 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1315 nir_builder_instr_insert(&b
, &instr
->instr
);
1319 unreachable("not reached");
1322 if (ir
->return_deref
)
1323 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), ret
, ~0);
1328 struct hash_entry
*entry
=
1329 _mesa_hash_table_search(this->overload_table
, ir
->callee
);
1331 nir_function
*callee
= (nir_function
*) entry
->data
;
1333 nir_call_instr
*call
= nir_call_instr_create(this->shader
, callee
);
1336 nir_deref_instr
*ret_deref
= NULL
;
1337 if (ir
->return_deref
) {
1338 nir_variable
*ret_tmp
=
1339 nir_local_variable_create(this->impl
, ir
->return_deref
->type
,
1341 ret_deref
= nir_build_deref_var(&b
, ret_tmp
);
1342 call
->params
[i
++] = nir_src_for_ssa(&ret_deref
->dest
.ssa
);
1345 foreach_two_lists(formal_node
, &ir
->callee
->parameters
,
1346 actual_node
, &ir
->actual_parameters
) {
1347 ir_rvalue
*param_rvalue
= (ir_rvalue
*) actual_node
;
1348 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
1350 if (sig_param
->data
.mode
== ir_var_function_out
) {
1351 nir_deref_instr
*out_deref
= evaluate_deref(param_rvalue
);
1352 call
->params
[i
] = nir_src_for_ssa(&out_deref
->dest
.ssa
);
1353 } else if (sig_param
->data
.mode
== ir_var_function_in
) {
1354 nir_ssa_def
*val
= evaluate_rvalue(param_rvalue
);
1355 nir_src src
= nir_src_for_ssa(val
);
1357 nir_src_copy(&call
->params
[i
], &src
, call
);
1358 } else if (sig_param
->data
.mode
== ir_var_function_inout
) {
1359 unreachable("unimplemented: inout parameters");
1365 nir_builder_instr_insert(&b
, &call
->instr
);
1367 if (ir
->return_deref
)
1368 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), nir_load_deref(&b
, ret_deref
), ~0);
1372 nir_visitor::visit(ir_assignment
*ir
)
1374 unsigned num_components
= ir
->lhs
->type
->vector_elements
;
1376 b
.exact
= ir
->lhs
->variable_referenced()->data
.invariant
||
1377 ir
->lhs
->variable_referenced()->data
.precise
;
1379 if ((ir
->rhs
->as_dereference() || ir
->rhs
->as_constant()) &&
1380 (ir
->write_mask
== (1 << num_components
) - 1 || ir
->write_mask
== 0)) {
1381 if (ir
->condition
) {
1382 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1383 nir_copy_deref(&b
, evaluate_deref(ir
->lhs
), evaluate_deref(ir
->rhs
));
1384 nir_pop_if(&b
, NULL
);
1386 nir_copy_deref(&b
, evaluate_deref(ir
->lhs
), evaluate_deref(ir
->rhs
));
1391 assert(ir
->rhs
->type
->is_scalar() || ir
->rhs
->type
->is_vector());
1393 ir
->lhs
->accept(this);
1394 nir_deref_instr
*lhs_deref
= this->deref
;
1395 nir_ssa_def
*src
= evaluate_rvalue(ir
->rhs
);
1397 if (ir
->write_mask
!= (1 << num_components
) - 1 && ir
->write_mask
!= 0) {
1398 /* GLSL IR will give us the input to the write-masked assignment in a
1399 * single packed vector. So, for example, if the writemask is xzw, then
1400 * we have to swizzle x -> x, y -> z, and z -> w and get the y component
1404 unsigned component
= 0;
1405 for (unsigned i
= 0; i
< 4; i
++) {
1406 swiz
[i
] = ir
->write_mask
& (1 << i
) ? component
++ : 0;
1408 src
= nir_swizzle(&b
, src
, swiz
, num_components
, false);
1411 if (ir
->condition
) {
1412 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1413 nir_store_deref(&b
, lhs_deref
, src
, ir
->write_mask
);
1414 nir_pop_if(&b
, NULL
);
1416 nir_store_deref(&b
, lhs_deref
, src
, ir
->write_mask
);
1421 * Given an instruction, returns a pointer to its destination or NULL if there
1422 * is no destination.
1424 * Note that this only handles instructions we generate at this level.
1427 get_instr_dest(nir_instr
*instr
)
1429 nir_alu_instr
*alu_instr
;
1430 nir_intrinsic_instr
*intrinsic_instr
;
1431 nir_tex_instr
*tex_instr
;
1433 switch (instr
->type
) {
1434 case nir_instr_type_alu
:
1435 alu_instr
= nir_instr_as_alu(instr
);
1436 return &alu_instr
->dest
.dest
;
1438 case nir_instr_type_intrinsic
:
1439 intrinsic_instr
= nir_instr_as_intrinsic(instr
);
1440 if (nir_intrinsic_infos
[intrinsic_instr
->intrinsic
].has_dest
)
1441 return &intrinsic_instr
->dest
;
1445 case nir_instr_type_tex
:
1446 tex_instr
= nir_instr_as_tex(instr
);
1447 return &tex_instr
->dest
;
1450 unreachable("not reached");
1457 nir_visitor::add_instr(nir_instr
*instr
, unsigned num_components
,
1460 nir_dest
*dest
= get_instr_dest(instr
);
1463 nir_ssa_dest_init(instr
, dest
, num_components
, bit_size
, NULL
);
1465 nir_builder_instr_insert(&b
, instr
);
1468 assert(dest
->is_ssa
);
1469 this->result
= &dest
->ssa
;
1474 nir_visitor::evaluate_rvalue(ir_rvalue
* ir
)
1477 if (ir
->as_dereference() || ir
->as_constant()) {
1479 * A dereference is being used on the right hand side, which means we
1480 * must emit a variable load.
1483 this->result
= nir_load_deref(&b
, this->deref
);
1486 return this->result
;
1490 type_is_float(glsl_base_type type
)
1492 return type
== GLSL_TYPE_FLOAT
|| type
== GLSL_TYPE_DOUBLE
||
1493 type
== GLSL_TYPE_FLOAT16
;
1497 type_is_signed(glsl_base_type type
)
1499 return type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_INT64
||
1500 type
== GLSL_TYPE_INT16
;
1504 type_is_int(glsl_base_type type
)
1506 return type
== GLSL_TYPE_UINT
|| type
== GLSL_TYPE_INT
||
1507 type
== GLSL_TYPE_UINT8
|| type
== GLSL_TYPE_INT8
||
1508 type
== GLSL_TYPE_UINT16
|| type
== GLSL_TYPE_INT16
||
1509 type
== GLSL_TYPE_UINT64
|| type
== GLSL_TYPE_INT64
;
1513 nir_visitor::visit(ir_expression
*ir
)
1515 /* Some special cases */
1516 switch (ir
->operation
) {
1517 case ir_binop_ubo_load
: {
1518 nir_intrinsic_instr
*load
=
1519 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_ubo
);
1520 unsigned bit_size
= ir
->type
->is_boolean() ? 32 :
1521 glsl_get_bit_size(ir
->type
);
1522 load
->num_components
= ir
->type
->vector_elements
;
1523 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1524 load
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1525 intrinsic_set_std430_align(load
, ir
->type
);
1526 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1529 * In UBO's, a true boolean value is any non-zero value, but we consider
1530 * a true boolean to be ~0. Fix this up with a != 0 comparison.
1533 if (ir
->type
->is_boolean())
1534 this->result
= nir_i2b(&b
, &load
->dest
.ssa
);
1539 case ir_unop_interpolate_at_centroid
:
1540 case ir_binop_interpolate_at_offset
:
1541 case ir_binop_interpolate_at_sample
: {
1542 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1543 ir_swizzle
*swizzle
= NULL
;
1545 /* the api does not allow a swizzle here, but the varying packing code
1546 * may have pushed one into here.
1548 swizzle
= ir
->operands
[0]->as_swizzle();
1550 deref
= swizzle
->val
->as_dereference();
1554 deref
->accept(this);
1556 nir_intrinsic_op op
;
1557 if (this->deref
->mode
== nir_var_shader_in
) {
1558 switch (ir
->operation
) {
1559 case ir_unop_interpolate_at_centroid
:
1560 op
= nir_intrinsic_interp_deref_at_centroid
;
1562 case ir_binop_interpolate_at_offset
:
1563 op
= nir_intrinsic_interp_deref_at_offset
;
1565 case ir_binop_interpolate_at_sample
:
1566 op
= nir_intrinsic_interp_deref_at_sample
;
1569 unreachable("Invalid interpolation intrinsic");
1572 /* This case can happen if the vertex shader does not write the
1573 * given varying. In this case, the linker will lower it to a
1574 * global variable. Since interpolating a variable makes no
1575 * sense, we'll just turn it into a load which will probably
1576 * eventually end up as an SSA definition.
1578 assert(this->deref
->mode
== nir_var_shader_temp
);
1579 op
= nir_intrinsic_load_deref
;
1582 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(shader
, op
);
1583 intrin
->num_components
= deref
->type
->vector_elements
;
1584 intrin
->src
[0] = nir_src_for_ssa(&this->deref
->dest
.ssa
);
1586 if (intrin
->intrinsic
== nir_intrinsic_interp_deref_at_offset
||
1587 intrin
->intrinsic
== nir_intrinsic_interp_deref_at_sample
)
1588 intrin
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1590 unsigned bit_size
= glsl_get_bit_size(deref
->type
);
1591 add_instr(&intrin
->instr
, deref
->type
->vector_elements
, bit_size
);
1594 unsigned swiz
[4] = {
1595 swizzle
->mask
.x
, swizzle
->mask
.y
, swizzle
->mask
.z
, swizzle
->mask
.w
1598 result
= nir_swizzle(&b
, result
, swiz
,
1599 swizzle
->type
->vector_elements
, false);
1609 nir_ssa_def
*srcs
[4];
1610 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1611 srcs
[i
] = evaluate_rvalue(ir
->operands
[i
]);
1613 glsl_base_type types
[4];
1614 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1615 if (supports_ints
|| !type_is_int(ir
->operands
[i
]->type
->base_type
))
1616 types
[i
] = ir
->operands
[i
]->type
->base_type
;
1618 types
[i
] = GLSL_TYPE_FLOAT
;
1620 glsl_base_type out_type
;
1621 if (supports_ints
|| !type_is_int(ir
->type
->base_type
))
1622 out_type
= ir
->type
->base_type
;
1624 out_type
= GLSL_TYPE_FLOAT
;
1626 switch (ir
->operation
) {
1627 case ir_unop_bit_not
: result
= nir_inot(&b
, srcs
[0]); break;
1628 case ir_unop_logic_not
:
1629 result
= nir_inot(&b
, srcs
[0]);
1632 result
= type_is_float(types
[0]) ? nir_fneg(&b
, srcs
[0])
1633 : nir_ineg(&b
, srcs
[0]);
1636 result
= type_is_float(types
[0]) ? nir_fabs(&b
, srcs
[0])
1637 : nir_iabs(&b
, srcs
[0]);
1639 case ir_unop_saturate
:
1640 assert(type_is_float(types
[0]));
1641 result
= nir_fsat(&b
, srcs
[0]);
1644 result
= type_is_float(types
[0]) ? nir_fsign(&b
, srcs
[0])
1645 : nir_isign(&b
, srcs
[0]);
1647 case ir_unop_rcp
: result
= nir_frcp(&b
, srcs
[0]); break;
1648 case ir_unop_rsq
: result
= nir_frsq(&b
, srcs
[0]); break;
1649 case ir_unop_sqrt
: result
= nir_fsqrt(&b
, srcs
[0]); break;
1650 case ir_unop_exp
: unreachable("ir_unop_exp should have been lowered");
1651 case ir_unop_log
: unreachable("ir_unop_log should have been lowered");
1652 case ir_unop_exp2
: result
= nir_fexp2(&b
, srcs
[0]); break;
1653 case ir_unop_log2
: result
= nir_flog2(&b
, srcs
[0]); break;
1655 result
= supports_ints
? nir_i2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1658 result
= supports_ints
? nir_u2f32(&b
, srcs
[0]) : nir_fmov(&b
, srcs
[0]);
1661 result
= nir_b2f32(&b
, srcs
[0]);
1664 result
= supports_ints
? nir_f2i32(&b
, srcs
[0]) : nir_ftrunc(&b
, srcs
[0]);
1667 result
= supports_ints
? nir_f2u32(&b
, srcs
[0]) : nir_ftrunc(&b
, srcs
[0]);
1699 case ir_unop_i642u64
:
1700 case ir_unop_u642i64
: {
1701 nir_alu_type src_type
= nir_get_nir_type_for_glsl_base_type(types
[0]);
1702 nir_alu_type dst_type
= nir_get_nir_type_for_glsl_base_type(out_type
);
1703 result
= nir_build_alu(&b
, nir_type_conversion_op(src_type
, dst_type
,
1704 nir_rounding_mode_undef
),
1705 srcs
[0], NULL
, NULL
, NULL
);
1706 /* b2i and b2f don't have fixed bit-size versions so the builder will
1707 * just assume 32 and we have to fix it up here.
1709 result
->bit_size
= nir_alu_type_get_type_size(dst_type
);
1713 case ir_unop_bitcast_i2f
:
1714 case ir_unop_bitcast_f2i
:
1715 case ir_unop_bitcast_u2f
:
1716 case ir_unop_bitcast_f2u
:
1717 case ir_unop_bitcast_i642d
:
1718 case ir_unop_bitcast_d2i64
:
1719 case ir_unop_bitcast_u642d
:
1720 case ir_unop_bitcast_d2u64
:
1721 case ir_unop_subroutine_to_int
:
1723 result
= nir_imov(&b
, srcs
[0]);
1725 case ir_unop_trunc
: result
= nir_ftrunc(&b
, srcs
[0]); break;
1726 case ir_unop_ceil
: result
= nir_fceil(&b
, srcs
[0]); break;
1727 case ir_unop_floor
: result
= nir_ffloor(&b
, srcs
[0]); break;
1728 case ir_unop_fract
: result
= nir_ffract(&b
, srcs
[0]); break;
1729 case ir_unop_frexp_exp
: result
= nir_frexp_exp(&b
, srcs
[0]); break;
1730 case ir_unop_frexp_sig
: result
= nir_frexp_sig(&b
, srcs
[0]); break;
1731 case ir_unop_round_even
: result
= nir_fround_even(&b
, srcs
[0]); break;
1732 case ir_unop_sin
: result
= nir_fsin(&b
, srcs
[0]); break;
1733 case ir_unop_cos
: result
= nir_fcos(&b
, srcs
[0]); break;
1734 case ir_unop_dFdx
: result
= nir_fddx(&b
, srcs
[0]); break;
1735 case ir_unop_dFdy
: result
= nir_fddy(&b
, srcs
[0]); break;
1736 case ir_unop_dFdx_fine
: result
= nir_fddx_fine(&b
, srcs
[0]); break;
1737 case ir_unop_dFdy_fine
: result
= nir_fddy_fine(&b
, srcs
[0]); break;
1738 case ir_unop_dFdx_coarse
: result
= nir_fddx_coarse(&b
, srcs
[0]); break;
1739 case ir_unop_dFdy_coarse
: result
= nir_fddy_coarse(&b
, srcs
[0]); break;
1740 case ir_unop_pack_snorm_2x16
:
1741 result
= nir_pack_snorm_2x16(&b
, srcs
[0]);
1743 case ir_unop_pack_snorm_4x8
:
1744 result
= nir_pack_snorm_4x8(&b
, srcs
[0]);
1746 case ir_unop_pack_unorm_2x16
:
1747 result
= nir_pack_unorm_2x16(&b
, srcs
[0]);
1749 case ir_unop_pack_unorm_4x8
:
1750 result
= nir_pack_unorm_4x8(&b
, srcs
[0]);
1752 case ir_unop_pack_half_2x16
:
1753 result
= nir_pack_half_2x16(&b
, srcs
[0]);
1755 case ir_unop_unpack_snorm_2x16
:
1756 result
= nir_unpack_snorm_2x16(&b
, srcs
[0]);
1758 case ir_unop_unpack_snorm_4x8
:
1759 result
= nir_unpack_snorm_4x8(&b
, srcs
[0]);
1761 case ir_unop_unpack_unorm_2x16
:
1762 result
= nir_unpack_unorm_2x16(&b
, srcs
[0]);
1764 case ir_unop_unpack_unorm_4x8
:
1765 result
= nir_unpack_unorm_4x8(&b
, srcs
[0]);
1767 case ir_unop_unpack_half_2x16
:
1768 result
= nir_unpack_half_2x16(&b
, srcs
[0]);
1770 case ir_unop_pack_sampler_2x32
:
1771 case ir_unop_pack_image_2x32
:
1772 case ir_unop_pack_double_2x32
:
1773 case ir_unop_pack_int_2x32
:
1774 case ir_unop_pack_uint_2x32
:
1775 result
= nir_pack_64_2x32(&b
, srcs
[0]);
1777 case ir_unop_unpack_sampler_2x32
:
1778 case ir_unop_unpack_image_2x32
:
1779 case ir_unop_unpack_double_2x32
:
1780 case ir_unop_unpack_int_2x32
:
1781 case ir_unop_unpack_uint_2x32
:
1782 result
= nir_unpack_64_2x32(&b
, srcs
[0]);
1784 case ir_unop_bitfield_reverse
:
1785 result
= nir_bitfield_reverse(&b
, srcs
[0]);
1787 case ir_unop_bit_count
:
1788 result
= nir_bit_count(&b
, srcs
[0]);
1790 case ir_unop_find_msb
:
1792 case GLSL_TYPE_UINT
:
1793 result
= nir_ufind_msb(&b
, srcs
[0]);
1796 result
= nir_ifind_msb(&b
, srcs
[0]);
1799 unreachable("Invalid type for findMSB()");
1802 case ir_unop_find_lsb
:
1803 result
= nir_find_lsb(&b
, srcs
[0]);
1807 switch (ir
->type
->vector_elements
) {
1809 switch (ir
->operands
[0]->type
->vector_elements
) {
1810 case 1: result
= nir_fnoise1_1(&b
, srcs
[0]); break;
1811 case 2: result
= nir_fnoise1_2(&b
, srcs
[0]); break;
1812 case 3: result
= nir_fnoise1_3(&b
, srcs
[0]); break;
1813 case 4: result
= nir_fnoise1_4(&b
, srcs
[0]); break;
1814 default: unreachable("not reached");
1818 switch (ir
->operands
[0]->type
->vector_elements
) {
1819 case 1: result
= nir_fnoise2_1(&b
, srcs
[0]); break;
1820 case 2: result
= nir_fnoise2_2(&b
, srcs
[0]); break;
1821 case 3: result
= nir_fnoise2_3(&b
, srcs
[0]); break;
1822 case 4: result
= nir_fnoise2_4(&b
, srcs
[0]); break;
1823 default: unreachable("not reached");
1827 switch (ir
->operands
[0]->type
->vector_elements
) {
1828 case 1: result
= nir_fnoise3_1(&b
, srcs
[0]); break;
1829 case 2: result
= nir_fnoise3_2(&b
, srcs
[0]); break;
1830 case 3: result
= nir_fnoise3_3(&b
, srcs
[0]); break;
1831 case 4: result
= nir_fnoise3_4(&b
, srcs
[0]); break;
1832 default: unreachable("not reached");
1836 switch (ir
->operands
[0]->type
->vector_elements
) {
1837 case 1: result
= nir_fnoise4_1(&b
, srcs
[0]); break;
1838 case 2: result
= nir_fnoise4_2(&b
, srcs
[0]); break;
1839 case 3: result
= nir_fnoise4_3(&b
, srcs
[0]); break;
1840 case 4: result
= nir_fnoise4_4(&b
, srcs
[0]); break;
1841 default: unreachable("not reached");
1845 unreachable("not reached");
1848 case ir_unop_get_buffer_size
: {
1849 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(
1851 nir_intrinsic_get_buffer_size
);
1852 load
->num_components
= ir
->type
->vector_elements
;
1853 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1854 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
1855 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1860 result
= type_is_float(out_type
) ? nir_fadd(&b
, srcs
[0], srcs
[1])
1861 : nir_iadd(&b
, srcs
[0], srcs
[1]);
1864 result
= type_is_float(out_type
) ? nir_fsub(&b
, srcs
[0], srcs
[1])
1865 : nir_isub(&b
, srcs
[0], srcs
[1]);
1868 if (type_is_float(out_type
))
1869 result
= nir_fmul(&b
, srcs
[0], srcs
[1]);
1870 else if (out_type
== GLSL_TYPE_INT64
&&
1871 (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
||
1872 ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
))
1873 result
= nir_imul_2x32_64(&b
, srcs
[0], srcs
[1]);
1874 else if (out_type
== GLSL_TYPE_UINT64
&&
1875 (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
||
1876 ir
->operands
[1]->type
->base_type
== GLSL_TYPE_UINT
))
1877 result
= nir_umul_2x32_64(&b
, srcs
[0], srcs
[1]);
1879 result
= nir_imul(&b
, srcs
[0], srcs
[1]);
1882 if (type_is_float(out_type
))
1883 result
= nir_fdiv(&b
, srcs
[0], srcs
[1]);
1884 else if (type_is_signed(out_type
))
1885 result
= nir_idiv(&b
, srcs
[0], srcs
[1]);
1887 result
= nir_udiv(&b
, srcs
[0], srcs
[1]);
1890 result
= type_is_float(out_type
) ? nir_fmod(&b
, srcs
[0], srcs
[1])
1891 : nir_umod(&b
, srcs
[0], srcs
[1]);
1894 if (type_is_float(out_type
))
1895 result
= nir_fmin(&b
, srcs
[0], srcs
[1]);
1896 else if (type_is_signed(out_type
))
1897 result
= nir_imin(&b
, srcs
[0], srcs
[1]);
1899 result
= nir_umin(&b
, srcs
[0], srcs
[1]);
1902 if (type_is_float(out_type
))
1903 result
= nir_fmax(&b
, srcs
[0], srcs
[1]);
1904 else if (type_is_signed(out_type
))
1905 result
= nir_imax(&b
, srcs
[0], srcs
[1]);
1907 result
= nir_umax(&b
, srcs
[0], srcs
[1]);
1909 case ir_binop_pow
: result
= nir_fpow(&b
, srcs
[0], srcs
[1]); break;
1910 case ir_binop_bit_and
: result
= nir_iand(&b
, srcs
[0], srcs
[1]); break;
1911 case ir_binop_bit_or
: result
= nir_ior(&b
, srcs
[0], srcs
[1]); break;
1912 case ir_binop_bit_xor
: result
= nir_ixor(&b
, srcs
[0], srcs
[1]); break;
1913 case ir_binop_logic_and
:
1914 result
= nir_iand(&b
, srcs
[0], srcs
[1]);
1916 case ir_binop_logic_or
:
1917 result
= nir_ior(&b
, srcs
[0], srcs
[1]);
1919 case ir_binop_logic_xor
:
1920 result
= nir_ixor(&b
, srcs
[0], srcs
[1]);
1922 case ir_binop_lshift
: result
= nir_ishl(&b
, srcs
[0], srcs
[1]); break;
1923 case ir_binop_rshift
:
1924 result
= (type_is_signed(out_type
)) ? nir_ishr(&b
, srcs
[0], srcs
[1])
1925 : nir_ushr(&b
, srcs
[0], srcs
[1]);
1927 case ir_binop_imul_high
:
1928 result
= (out_type
== GLSL_TYPE_INT
) ? nir_imul_high(&b
, srcs
[0], srcs
[1])
1929 : nir_umul_high(&b
, srcs
[0], srcs
[1]);
1931 case ir_binop_carry
: result
= nir_uadd_carry(&b
, srcs
[0], srcs
[1]); break;
1932 case ir_binop_borrow
: result
= nir_usub_borrow(&b
, srcs
[0], srcs
[1]); break;
1934 if (type_is_float(types
[0]))
1935 result
= nir_flt(&b
, srcs
[0], srcs
[1]);
1936 else if (type_is_signed(types
[0]))
1937 result
= nir_ilt(&b
, srcs
[0], srcs
[1]);
1939 result
= nir_ult(&b
, srcs
[0], srcs
[1]);
1941 case ir_binop_gequal
:
1942 if (type_is_float(types
[0]))
1943 result
= nir_fge(&b
, srcs
[0], srcs
[1]);
1944 else if (type_is_signed(types
[0]))
1945 result
= nir_ige(&b
, srcs
[0], srcs
[1]);
1947 result
= nir_uge(&b
, srcs
[0], srcs
[1]);
1949 case ir_binop_equal
:
1950 if (type_is_float(types
[0]))
1951 result
= nir_feq(&b
, srcs
[0], srcs
[1]);
1953 result
= nir_ieq(&b
, srcs
[0], srcs
[1]);
1955 case ir_binop_nequal
:
1956 if (type_is_float(types
[0]))
1957 result
= nir_fne(&b
, srcs
[0], srcs
[1]);
1959 result
= nir_ine(&b
, srcs
[0], srcs
[1]);
1961 case ir_binop_all_equal
:
1962 if (type_is_float(types
[0])) {
1963 switch (ir
->operands
[0]->type
->vector_elements
) {
1964 case 1: result
= nir_feq(&b
, srcs
[0], srcs
[1]); break;
1965 case 2: result
= nir_ball_fequal2(&b
, srcs
[0], srcs
[1]); break;
1966 case 3: result
= nir_ball_fequal3(&b
, srcs
[0], srcs
[1]); break;
1967 case 4: result
= nir_ball_fequal4(&b
, srcs
[0], srcs
[1]); break;
1969 unreachable("not reached");
1972 switch (ir
->operands
[0]->type
->vector_elements
) {
1973 case 1: result
= nir_ieq(&b
, srcs
[0], srcs
[1]); break;
1974 case 2: result
= nir_ball_iequal2(&b
, srcs
[0], srcs
[1]); break;
1975 case 3: result
= nir_ball_iequal3(&b
, srcs
[0], srcs
[1]); break;
1976 case 4: result
= nir_ball_iequal4(&b
, srcs
[0], srcs
[1]); break;
1978 unreachable("not reached");
1982 case ir_binop_any_nequal
:
1983 if (type_is_float(types
[0])) {
1984 switch (ir
->operands
[0]->type
->vector_elements
) {
1985 case 1: result
= nir_fne(&b
, srcs
[0], srcs
[1]); break;
1986 case 2: result
= nir_bany_fnequal2(&b
, srcs
[0], srcs
[1]); break;
1987 case 3: result
= nir_bany_fnequal3(&b
, srcs
[0], srcs
[1]); break;
1988 case 4: result
= nir_bany_fnequal4(&b
, srcs
[0], srcs
[1]); break;
1990 unreachable("not reached");
1993 switch (ir
->operands
[0]->type
->vector_elements
) {
1994 case 1: result
= nir_ine(&b
, srcs
[0], srcs
[1]); break;
1995 case 2: result
= nir_bany_inequal2(&b
, srcs
[0], srcs
[1]); break;
1996 case 3: result
= nir_bany_inequal3(&b
, srcs
[0], srcs
[1]); break;
1997 case 4: result
= nir_bany_inequal4(&b
, srcs
[0], srcs
[1]); break;
1999 unreachable("not reached");
2004 switch (ir
->operands
[0]->type
->vector_elements
) {
2005 case 2: result
= nir_fdot2(&b
, srcs
[0], srcs
[1]); break;
2006 case 3: result
= nir_fdot3(&b
, srcs
[0], srcs
[1]); break;
2007 case 4: result
= nir_fdot4(&b
, srcs
[0], srcs
[1]); break;
2009 unreachable("not reached");
2012 case ir_binop_vector_extract
: {
2013 result
= nir_channel(&b
, srcs
[0], 0);
2014 for (unsigned i
= 1; i
< ir
->operands
[0]->type
->vector_elements
; i
++) {
2015 nir_ssa_def
*swizzled
= nir_channel(&b
, srcs
[0], i
);
2016 result
= nir_bcsel(&b
, nir_ieq(&b
, srcs
[1], nir_imm_int(&b
, i
)),
2022 case ir_binop_ldexp
: result
= nir_ldexp(&b
, srcs
[0], srcs
[1]); break;
2024 result
= nir_ffma(&b
, srcs
[0], srcs
[1], srcs
[2]);
2027 result
= nir_flrp(&b
, srcs
[0], srcs
[1], srcs
[2]);
2030 result
= nir_bcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
2032 case ir_triop_bitfield_extract
:
2033 result
= (out_type
== GLSL_TYPE_INT
) ?
2034 nir_ibitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]) :
2035 nir_ubitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]);
2037 case ir_quadop_bitfield_insert
:
2038 result
= nir_bitfield_insert(&b
, srcs
[0], srcs
[1], srcs
[2], srcs
[3]);
2040 case ir_quadop_vector
:
2041 result
= nir_vec(&b
, srcs
, ir
->type
->vector_elements
);
2045 unreachable("not reached");
2050 nir_visitor::visit(ir_swizzle
*ir
)
2052 unsigned swizzle
[4] = { ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
};
2053 result
= nir_swizzle(&b
, evaluate_rvalue(ir
->val
), swizzle
,
2054 ir
->type
->vector_elements
, false);
2058 nir_visitor::visit(ir_texture
*ir
)
2065 num_srcs
= 1; /* coordinate */
2070 op
= (ir
->op
== ir_txb
) ? nir_texop_txb
: nir_texop_txl
;
2071 num_srcs
= 2; /* coordinate, bias/lod */
2075 op
= nir_texop_txd
; /* coordinate, dPdx, dPdy */
2081 if (ir
->lod_info
.lod
!= NULL
)
2082 num_srcs
= 2; /* coordinate, lod */
2084 num_srcs
= 1; /* coordinate */
2088 op
= nir_texop_txf_ms
;
2089 num_srcs
= 2; /* coordinate, sample_index */
2094 if (ir
->lod_info
.lod
!= NULL
)
2095 num_srcs
= 1; /* lod */
2102 num_srcs
= 1; /* coordinate */
2107 num_srcs
= 1; /* coordinate */
2110 case ir_query_levels
:
2111 op
= nir_texop_query_levels
;
2115 case ir_texture_samples
:
2116 op
= nir_texop_texture_samples
;
2120 case ir_samples_identical
:
2121 op
= nir_texop_samples_identical
;
2122 num_srcs
= 1; /* coordinate */
2126 unreachable("not reached");
2129 if (ir
->projector
!= NULL
)
2131 if (ir
->shadow_comparator
!= NULL
)
2133 if (ir
->offset
!= NULL
)
2136 /* Add one for the texture deref */
2139 nir_tex_instr
*instr
= nir_tex_instr_create(this->shader
, num_srcs
);
2142 instr
->sampler_dim
=
2143 (glsl_sampler_dim
) ir
->sampler
->type
->sampler_dimensionality
;
2144 instr
->is_array
= ir
->sampler
->type
->sampler_array
;
2145 instr
->is_shadow
= ir
->sampler
->type
->sampler_shadow
;
2146 if (instr
->is_shadow
)
2147 instr
->is_new_style_shadow
= (ir
->type
->vector_elements
== 1);
2148 switch (ir
->type
->base_type
) {
2149 case GLSL_TYPE_FLOAT
:
2150 instr
->dest_type
= nir_type_float
;
2153 instr
->dest_type
= nir_type_int
;
2155 case GLSL_TYPE_BOOL
:
2156 case GLSL_TYPE_UINT
:
2157 instr
->dest_type
= nir_type_uint
;
2160 unreachable("not reached");
2163 nir_deref_instr
*sampler_deref
= evaluate_deref(ir
->sampler
);
2164 instr
->src
[0].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2165 instr
->src
[0].src_type
= nir_tex_src_texture_deref
;
2166 instr
->src
[1].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2167 instr
->src
[1].src_type
= nir_tex_src_sampler_deref
;
2169 unsigned src_number
= 2;
2171 if (ir
->coordinate
!= NULL
) {
2172 instr
->coord_components
= ir
->coordinate
->type
->vector_elements
;
2173 instr
->src
[src_number
].src
=
2174 nir_src_for_ssa(evaluate_rvalue(ir
->coordinate
));
2175 instr
->src
[src_number
].src_type
= nir_tex_src_coord
;
2179 if (ir
->projector
!= NULL
) {
2180 instr
->src
[src_number
].src
=
2181 nir_src_for_ssa(evaluate_rvalue(ir
->projector
));
2182 instr
->src
[src_number
].src_type
= nir_tex_src_projector
;
2186 if (ir
->shadow_comparator
!= NULL
) {
2187 instr
->src
[src_number
].src
=
2188 nir_src_for_ssa(evaluate_rvalue(ir
->shadow_comparator
));
2189 instr
->src
[src_number
].src_type
= nir_tex_src_comparator
;
2193 if (ir
->offset
!= NULL
) {
2194 /* we don't support multiple offsets yet */
2195 assert(ir
->offset
->type
->is_vector() || ir
->offset
->type
->is_scalar());
2197 instr
->src
[src_number
].src
=
2198 nir_src_for_ssa(evaluate_rvalue(ir
->offset
));
2199 instr
->src
[src_number
].src_type
= nir_tex_src_offset
;
2205 instr
->src
[src_number
].src
=
2206 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.bias
));
2207 instr
->src
[src_number
].src_type
= nir_tex_src_bias
;
2214 if (ir
->lod_info
.lod
!= NULL
) {
2215 instr
->src
[src_number
].src
=
2216 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.lod
));
2217 instr
->src
[src_number
].src_type
= nir_tex_src_lod
;
2223 instr
->src
[src_number
].src
=
2224 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdx
));
2225 instr
->src
[src_number
].src_type
= nir_tex_src_ddx
;
2227 instr
->src
[src_number
].src
=
2228 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdy
));
2229 instr
->src
[src_number
].src_type
= nir_tex_src_ddy
;
2234 instr
->src
[src_number
].src
=
2235 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.sample_index
));
2236 instr
->src
[src_number
].src_type
= nir_tex_src_ms_index
;
2241 instr
->component
= ir
->lod_info
.component
->as_constant()->value
.u
[0];
2248 assert(src_number
== num_srcs
);
2250 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2251 add_instr(&instr
->instr
, nir_tex_instr_dest_size(instr
), bit_size
);
2255 nir_visitor::visit(ir_constant
*ir
)
2258 * We don't know if this variable is an array or struct that gets
2259 * dereferenced, so do the safe thing an make it a variable with a
2260 * constant initializer and return a dereference.
2264 nir_local_variable_create(this->impl
, ir
->type
, "const_temp");
2265 var
->data
.read_only
= true;
2266 var
->constant_initializer
= constant_copy(ir
, var
);
2268 this->deref
= nir_build_deref_var(&b
, var
);
2272 nir_visitor::visit(ir_dereference_variable
*ir
)
2274 if (ir
->variable_referenced()->data
.mode
== ir_var_function_out
) {
2275 unsigned i
= (sig
->return_type
!= glsl_type::void_type
) ? 1 : 0;
2277 foreach_in_list(ir_variable
, param
, &sig
->parameters
) {
2278 if (param
== ir
->variable_referenced()) {
2284 this->deref
= nir_build_deref_cast(&b
, nir_load_param(&b
, i
),
2285 nir_var_function_temp
, ir
->type
, 0);
2289 assert(ir
->variable_referenced()->data
.mode
!= ir_var_function_inout
);
2291 struct hash_entry
*entry
=
2292 _mesa_hash_table_search(this->var_table
, ir
->var
);
2294 nir_variable
*var
= (nir_variable
*) entry
->data
;
2296 this->deref
= nir_build_deref_var(&b
, var
);
2300 nir_visitor::visit(ir_dereference_record
*ir
)
2302 ir
->record
->accept(this);
2304 int field_index
= ir
->field_idx
;
2305 assert(field_index
>= 0);
2307 this->deref
= nir_build_deref_struct(&b
, this->deref
, field_index
);
2311 nir_visitor::visit(ir_dereference_array
*ir
)
2313 nir_ssa_def
*index
= evaluate_rvalue(ir
->array_index
);
2315 ir
->array
->accept(this);
2317 this->deref
= nir_build_deref_array(&b
, this->deref
, index
);
2321 nir_visitor::visit(ir_barrier
*)
2323 nir_intrinsic_instr
*instr
=
2324 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_barrier
);
2325 nir_builder_instr_insert(&b
, &instr
->instr
);