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 "float64_glsl.h"
29 #include "glsl_to_nir.h"
30 #include "ir_visitor.h"
31 #include "ir_hierarchical_visitor.h"
33 #include "ir_optimization.h"
35 #include "compiler/nir/nir_control_flow.h"
36 #include "compiler/nir/nir_builder.h"
37 #include "main/errors.h"
38 #include "main/imports.h"
39 #include "main/mtypes.h"
40 #include "main/shaderobj.h"
41 #include "util/u_math.h"
44 * pass to lower GLSL IR to NIR
46 * This will lower variable dereferences to loads/stores of corresponding
47 * variables in NIR - the variables will be converted to registers in a later
53 class nir_visitor
: public ir_visitor
56 nir_visitor(gl_context
*ctx
, nir_shader
*shader
);
59 virtual void visit(ir_variable
*);
60 virtual void visit(ir_function
*);
61 virtual void visit(ir_function_signature
*);
62 virtual void visit(ir_loop
*);
63 virtual void visit(ir_if
*);
64 virtual void visit(ir_discard
*);
65 virtual void visit(ir_loop_jump
*);
66 virtual void visit(ir_return
*);
67 virtual void visit(ir_call
*);
68 virtual void visit(ir_assignment
*);
69 virtual void visit(ir_emit_vertex
*);
70 virtual void visit(ir_end_primitive
*);
71 virtual void visit(ir_expression
*);
72 virtual void visit(ir_swizzle
*);
73 virtual void visit(ir_texture
*);
74 virtual void visit(ir_constant
*);
75 virtual void visit(ir_dereference_variable
*);
76 virtual void visit(ir_dereference_record
*);
77 virtual void visit(ir_dereference_array
*);
78 virtual void visit(ir_barrier
*);
80 void create_function(ir_function_signature
*ir
);
83 void add_instr(nir_instr
*instr
, unsigned num_components
, unsigned bit_size
);
84 nir_ssa_def
*evaluate_rvalue(ir_rvalue
*ir
);
86 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
**srcs
);
87 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
);
88 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
,
90 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
,
91 nir_ssa_def
*src2
, nir_ssa_def
*src3
);
96 nir_function_impl
*impl
;
98 nir_ssa_def
*result
; /* result of the expression tree last visited */
100 nir_deref_instr
*evaluate_deref(ir_instruction
*ir
);
102 nir_constant
*constant_copy(ir_constant
*ir
, void *mem_ctx
);
104 /* most recent deref instruction created */
105 nir_deref_instr
*deref
;
107 /* whether the IR we're operating on is per-function or global */
110 ir_function_signature
*sig
;
112 /* map of ir_variable -> nir_variable */
113 struct hash_table
*var_table
;
115 /* map of ir_function_signature -> nir_function_overload */
116 struct hash_table
*overload_table
;
120 * This visitor runs before the main visitor, calling create_function() for
121 * each function so that the main visitor can resolve forward references in
125 class nir_function_visitor
: public ir_hierarchical_visitor
128 nir_function_visitor(nir_visitor
*v
) : visitor(v
)
131 virtual ir_visitor_status
visit_enter(ir_function
*);
134 nir_visitor
*visitor
;
137 /* glsl_to_nir can only handle converting certain function paramaters
138 * to NIR. This visitor checks for parameters it can't currently handle.
140 class ir_function_param_visitor
: public ir_hierarchical_visitor
143 ir_function_param_visitor()
148 virtual ir_visitor_status
visit_enter(ir_function_signature
*ir
)
151 if (ir
->is_intrinsic())
152 return visit_continue
;
154 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
155 if (!param
->type
->is_vector() || !param
->type
->is_scalar()) {
160 if (param
->data
.mode
== ir_var_function_inout
) {
166 return visit_continue
;
172 } /* end of anonymous namespace */
176 has_unsupported_function_param(exec_list
*ir
)
178 ir_function_param_visitor visitor
;
179 visit_list_elements(&visitor
, ir
);
180 return visitor
.unsupported
;
184 glsl_to_nir(struct gl_context
*ctx
,
185 const struct gl_shader_program
*shader_prog
,
186 gl_shader_stage stage
,
187 const nir_shader_compiler_options
*options
)
189 struct gl_linked_shader
*sh
= shader_prog
->_LinkedShaders
[stage
];
191 const struct gl_shader_compiler_options
*gl_options
=
192 &ctx
->Const
.ShaderCompilerOptions
[stage
];
194 /* glsl_to_nir can only handle converting certain function paramaters
195 * to NIR. If we find something we can't handle then we get the GLSL IR
196 * opts to remove it before we continue on.
198 * TODO: add missing glsl ir to nir support and remove this loop.
200 while (has_unsupported_function_param(sh
->ir
)) {
201 do_common_optimization(sh
->ir
, true, true, gl_options
,
202 ctx
->Const
.NativeIntegers
);
205 nir_shader
*shader
= nir_shader_create(NULL
, stage
, options
,
208 nir_visitor
v1(ctx
, shader
);
209 nir_function_visitor
v2(&v1
);
211 visit_exec_list(sh
->ir
, &v1
);
213 nir_validate_shader(shader
, "after glsl to nir, before function inline");
215 /* We have to lower away local constant initializers right before we
216 * inline functions. That way they get properly initialized at the top
217 * of the function and not at the top of its caller.
219 nir_lower_constant_initializers(shader
, (nir_variable_mode
)~0);
220 nir_lower_returns(shader
);
221 nir_inline_functions(shader
);
222 nir_opt_deref(shader
);
224 nir_validate_shader(shader
, "after function inlining and return lowering");
226 /* Now that we have inlined everything remove all of the functions except
229 foreach_list_typed_safe(nir_function
, function
, node
, &(shader
)->functions
){
230 if (strcmp("main", function
->name
) != 0) {
231 exec_node_remove(&function
->node
);
235 /* Remap the locations to slots so those requiring two slots will occupy
236 * two locations. For instance, if we have in the IR code a dvec3 attr0 in
237 * location 0 and vec4 attr1 in location 1, in NIR attr0 will use
238 * locations/slots 0 and 1, and attr1 will use location/slot 2 */
239 if (shader
->info
.stage
== MESA_SHADER_VERTEX
)
240 nir_remap_dual_slot_attributes(shader
, &sh
->Program
->DualSlotInputs
);
242 shader
->info
.name
= ralloc_asprintf(shader
, "GLSL%d", shader_prog
->Name
);
243 if (shader_prog
->Label
)
244 shader
->info
.label
= ralloc_strdup(shader
, shader_prog
->Label
);
246 /* Check for transform feedback varyings specified via the API */
247 shader
->info
.has_transform_feedback_varyings
=
248 shader_prog
->TransformFeedback
.NumVarying
> 0;
250 /* Check for transform feedback varyings specified in the Shader */
251 if (shader_prog
->last_vert_prog
)
252 shader
->info
.has_transform_feedback_varyings
|=
253 shader_prog
->last_vert_prog
->sh
.LinkedTransformFeedback
->NumVarying
> 0;
255 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
) {
256 shader
->info
.fs
.pixel_center_integer
= sh
->Program
->info
.fs
.pixel_center_integer
;
257 shader
->info
.fs
.origin_upper_left
= sh
->Program
->info
.fs
.origin_upper_left
;
263 nir_visitor::nir_visitor(gl_context
*ctx
, nir_shader
*shader
)
265 this->supports_std430
= ctx
->Const
.UseSTD430AsDefaultPacking
;
266 this->shader
= shader
;
267 this->is_global
= true;
268 this->var_table
= _mesa_pointer_hash_table_create(NULL
);
269 this->overload_table
= _mesa_pointer_hash_table_create(NULL
);
274 memset(&this->b
, 0, sizeof(this->b
));
277 nir_visitor::~nir_visitor()
279 _mesa_hash_table_destroy(this->var_table
, NULL
);
280 _mesa_hash_table_destroy(this->overload_table
, NULL
);
284 nir_visitor::evaluate_deref(ir_instruction
*ir
)
291 nir_visitor::constant_copy(ir_constant
*ir
, void *mem_ctx
)
296 nir_constant
*ret
= rzalloc(mem_ctx
, nir_constant
);
298 const unsigned rows
= ir
->type
->vector_elements
;
299 const unsigned cols
= ir
->type
->matrix_columns
;
302 ret
->num_elements
= 0;
303 switch (ir
->type
->base_type
) {
305 /* Only float base types can be matrices. */
308 for (unsigned r
= 0; r
< rows
; r
++)
309 ret
->values
[0][r
].u32
= ir
->value
.u
[r
];
314 /* Only float base types can be matrices. */
317 for (unsigned r
= 0; r
< rows
; r
++)
318 ret
->values
[0][r
].i32
= ir
->value
.i
[r
];
322 case GLSL_TYPE_FLOAT
:
323 for (unsigned c
= 0; c
< cols
; c
++) {
324 for (unsigned r
= 0; r
< rows
; r
++)
325 ret
->values
[c
][r
].f32
= ir
->value
.f
[c
* rows
+ r
];
329 case GLSL_TYPE_DOUBLE
:
330 for (unsigned c
= 0; c
< cols
; c
++) {
331 for (unsigned r
= 0; r
< rows
; r
++)
332 ret
->values
[c
][r
].f64
= ir
->value
.d
[c
* rows
+ r
];
336 case GLSL_TYPE_UINT64
:
337 /* Only float base types can be matrices. */
340 for (unsigned r
= 0; r
< rows
; r
++)
341 ret
->values
[0][r
].u64
= ir
->value
.u64
[r
];
344 case GLSL_TYPE_INT64
:
345 /* Only float base types can be matrices. */
348 for (unsigned r
= 0; r
< rows
; r
++)
349 ret
->values
[0][r
].i64
= ir
->value
.i64
[r
];
353 /* Only float base types can be matrices. */
356 for (unsigned r
= 0; r
< rows
; r
++)
357 ret
->values
[0][r
].b
= ir
->value
.b
[r
];
361 case GLSL_TYPE_STRUCT
:
362 case GLSL_TYPE_ARRAY
:
363 ret
->elements
= ralloc_array(mem_ctx
, nir_constant
*,
365 ret
->num_elements
= ir
->type
->length
;
367 for (i
= 0; i
< ir
->type
->length
; i
++)
368 ret
->elements
[i
] = constant_copy(ir
->const_elements
[i
], mem_ctx
);
372 unreachable("not reached");
378 static const glsl_type
*
379 wrap_type_in_array(const glsl_type
*elem_type
, const glsl_type
*array_type
)
381 if (!array_type
->is_array())
384 elem_type
= wrap_type_in_array(elem_type
, array_type
->fields
.array
);
386 return glsl_type::get_array_instance(elem_type
, array_type
->length
);
390 nir_visitor::visit(ir_variable
*ir
)
392 /* TODO: In future we should switch to using the NIR lowering pass but for
393 * now just ignore these variables as GLSL IR should have lowered them.
394 * Anything remaining are just dead vars that weren't cleaned up.
396 if (ir
->data
.mode
== ir_var_shader_shared
)
399 /* FINISHME: inout parameters */
400 assert(ir
->data
.mode
!= ir_var_function_inout
);
402 if (ir
->data
.mode
== ir_var_function_out
)
405 nir_variable
*var
= rzalloc(shader
, nir_variable
);
406 var
->type
= ir
->type
;
407 var
->name
= ralloc_strdup(var
, ir
->name
);
409 var
->data
.always_active_io
= ir
->data
.always_active_io
;
410 var
->data
.read_only
= ir
->data
.read_only
;
411 var
->data
.centroid
= ir
->data
.centroid
;
412 var
->data
.sample
= ir
->data
.sample
;
413 var
->data
.patch
= ir
->data
.patch
;
414 var
->data
.invariant
= ir
->data
.invariant
;
415 var
->data
.location
= ir
->data
.location
;
416 var
->data
.stream
= ir
->data
.stream
;
417 var
->data
.compact
= false;
419 switch(ir
->data
.mode
) {
421 case ir_var_temporary
:
423 var
->data
.mode
= nir_var_shader_temp
;
425 var
->data
.mode
= nir_var_function_temp
;
428 case ir_var_function_in
:
429 case ir_var_const_in
:
430 var
->data
.mode
= nir_var_function_temp
;
433 case ir_var_shader_in
:
434 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
&&
435 ir
->data
.location
== VARYING_SLOT_FACE
) {
436 /* For whatever reason, GLSL IR makes gl_FrontFacing an input */
437 var
->data
.location
= SYSTEM_VALUE_FRONT_FACE
;
438 var
->data
.mode
= nir_var_system_value
;
439 } else if (shader
->info
.stage
== MESA_SHADER_GEOMETRY
&&
440 ir
->data
.location
== VARYING_SLOT_PRIMITIVE_ID
) {
441 /* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */
442 var
->data
.location
= SYSTEM_VALUE_PRIMITIVE_ID
;
443 var
->data
.mode
= nir_var_system_value
;
445 var
->data
.mode
= nir_var_shader_in
;
447 if (shader
->info
.stage
== MESA_SHADER_TESS_EVAL
&&
448 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
449 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
450 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
453 if (shader
->info
.stage
> MESA_SHADER_VERTEX
&&
454 ir
->data
.location
>= VARYING_SLOT_CLIP_DIST0
&&
455 ir
->data
.location
<= VARYING_SLOT_CULL_DIST1
) {
456 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
461 case ir_var_shader_out
:
462 var
->data
.mode
= nir_var_shader_out
;
463 if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
&&
464 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
465 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
466 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
469 if (shader
->info
.stage
<= MESA_SHADER_GEOMETRY
&&
470 ir
->data
.location
>= VARYING_SLOT_CLIP_DIST0
&&
471 ir
->data
.location
<= VARYING_SLOT_CULL_DIST1
) {
472 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
477 if (ir
->get_interface_type())
478 var
->data
.mode
= nir_var_mem_ubo
;
480 var
->data
.mode
= nir_var_uniform
;
483 case ir_var_shader_storage
:
484 var
->data
.mode
= nir_var_mem_ssbo
;
487 case ir_var_system_value
:
488 var
->data
.mode
= nir_var_system_value
;
492 unreachable("not reached");
495 unsigned image_access
= 0;
496 if (ir
->data
.memory_read_only
)
497 image_access
|= ACCESS_NON_WRITEABLE
;
498 if (ir
->data
.memory_write_only
)
499 image_access
|= ACCESS_NON_READABLE
;
500 if (ir
->data
.memory_coherent
)
501 image_access
|= ACCESS_COHERENT
;
502 if (ir
->data
.memory_volatile
)
503 image_access
|= ACCESS_VOLATILE
;
504 if (ir
->data
.memory_restrict
)
505 image_access
|= ACCESS_RESTRICT
;
507 /* For UBO and SSBO variables, we need explicit types */
508 if (var
->data
.mode
& (nir_var_mem_ubo
| nir_var_mem_ssbo
)) {
509 const glsl_type
*explicit_ifc_type
=
510 ir
->get_interface_type()->get_explicit_interface_type(supports_std430
);
512 if (ir
->type
->without_array()->is_interface()) {
513 /* If the type contains the interface, wrap the explicit type in the
514 * right number of arrays.
516 var
->type
= wrap_type_in_array(explicit_ifc_type
, ir
->type
);
518 /* Otherwise, this variable is one entry in the interface */
519 UNUSED
bool found
= false;
520 for (unsigned i
= 0; i
< explicit_ifc_type
->length
; i
++) {
521 const glsl_struct_field
*field
=
522 &explicit_ifc_type
->fields
.structure
[i
];
523 if (strcmp(ir
->name
, field
->name
) != 0)
526 var
->type
= field
->type
;
527 if (field
->memory_read_only
)
528 image_access
|= ACCESS_NON_WRITEABLE
;
529 if (field
->memory_write_only
)
530 image_access
|= ACCESS_NON_READABLE
;
531 if (field
->memory_coherent
)
532 image_access
|= ACCESS_COHERENT
;
533 if (field
->memory_volatile
)
534 image_access
|= ACCESS_VOLATILE
;
535 if (field
->memory_restrict
)
536 image_access
|= ACCESS_RESTRICT
;
545 var
->data
.interpolation
= ir
->data
.interpolation
;
546 var
->data
.location_frac
= ir
->data
.location_frac
;
548 switch (ir
->data
.depth_layout
) {
549 case ir_depth_layout_none
:
550 var
->data
.depth_layout
= nir_depth_layout_none
;
552 case ir_depth_layout_any
:
553 var
->data
.depth_layout
= nir_depth_layout_any
;
555 case ir_depth_layout_greater
:
556 var
->data
.depth_layout
= nir_depth_layout_greater
;
558 case ir_depth_layout_less
:
559 var
->data
.depth_layout
= nir_depth_layout_less
;
561 case ir_depth_layout_unchanged
:
562 var
->data
.depth_layout
= nir_depth_layout_unchanged
;
565 unreachable("not reached");
568 var
->data
.index
= ir
->data
.index
;
569 var
->data
.descriptor_set
= 0;
570 var
->data
.binding
= ir
->data
.binding
;
571 var
->data
.explicit_binding
= ir
->data
.explicit_binding
;
572 var
->data
.bindless
= ir
->data
.bindless
;
573 var
->data
.offset
= ir
->data
.offset
;
575 var
->data
.image
.access
= (gl_access_qualifier
)image_access
;
576 var
->data
.image
.format
= ir
->data
.image_format
;
578 var
->data
.fb_fetch_output
= ir
->data
.fb_fetch_output
;
579 var
->data
.explicit_xfb_buffer
= ir
->data
.explicit_xfb_buffer
;
580 var
->data
.explicit_xfb_stride
= ir
->data
.explicit_xfb_stride
;
581 var
->data
.xfb_buffer
= ir
->data
.xfb_buffer
;
582 var
->data
.xfb_stride
= ir
->data
.xfb_stride
;
584 var
->num_state_slots
= ir
->get_num_state_slots();
585 if (var
->num_state_slots
> 0) {
586 var
->state_slots
= rzalloc_array(var
, nir_state_slot
,
587 var
->num_state_slots
);
589 ir_state_slot
*state_slots
= ir
->get_state_slots();
590 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
591 for (unsigned j
= 0; j
< 5; j
++)
592 var
->state_slots
[i
].tokens
[j
] = state_slots
[i
].tokens
[j
];
593 var
->state_slots
[i
].swizzle
= state_slots
[i
].swizzle
;
596 var
->state_slots
= NULL
;
599 var
->constant_initializer
= constant_copy(ir
->constant_initializer
, var
);
601 var
->interface_type
= ir
->get_interface_type();
603 if (var
->data
.mode
== nir_var_function_temp
)
604 nir_function_impl_add_variable(impl
, var
);
606 nir_shader_add_variable(shader
, var
);
608 _mesa_hash_table_insert(var_table
, ir
, var
);
612 nir_function_visitor::visit_enter(ir_function
*ir
)
614 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
) {
615 visitor
->create_function(sig
);
617 return visit_continue_with_parent
;
621 nir_visitor::create_function(ir_function_signature
*ir
)
623 if (ir
->is_intrinsic())
626 nir_function
*func
= nir_function_create(shader
, ir
->function_name());
627 if (strcmp(ir
->function_name(), "main") == 0)
628 func
->is_entrypoint
= true;
630 func
->num_params
= ir
->parameters
.length() +
631 (ir
->return_type
!= glsl_type::void_type
);
632 func
->params
= ralloc_array(shader
, nir_parameter
, func
->num_params
);
636 if (ir
->return_type
!= glsl_type::void_type
) {
637 /* The return value is a variable deref (basically an out parameter) */
638 func
->params
[np
].num_components
= 1;
639 func
->params
[np
].bit_size
= 32;
643 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
644 /* FINISHME: pass arrays, structs, etc by reference? */
645 assert(param
->type
->is_vector() || param
->type
->is_scalar());
647 if (param
->data
.mode
== ir_var_function_in
) {
648 func
->params
[np
].num_components
= param
->type
->vector_elements
;
649 func
->params
[np
].bit_size
= glsl_get_bit_size(param
->type
);
651 func
->params
[np
].num_components
= 1;
652 func
->params
[np
].bit_size
= 32;
656 assert(np
== func
->num_params
);
658 _mesa_hash_table_insert(this->overload_table
, ir
, func
);
662 nir_visitor::visit(ir_function
*ir
)
664 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
)
669 nir_visitor::visit(ir_function_signature
*ir
)
671 if (ir
->is_intrinsic())
676 struct hash_entry
*entry
=
677 _mesa_hash_table_search(this->overload_table
, ir
);
680 nir_function
*func
= (nir_function
*) entry
->data
;
682 if (ir
->is_defined
) {
683 nir_function_impl
*impl
= nir_function_impl_create(func
);
686 this->is_global
= false;
688 nir_builder_init(&b
, impl
);
689 b
.cursor
= nir_after_cf_list(&impl
->body
);
691 unsigned i
= (ir
->return_type
!= glsl_type::void_type
) ? 1 : 0;
693 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
695 nir_local_variable_create(impl
, param
->type
, param
->name
);
697 if (param
->data
.mode
== ir_var_function_in
) {
698 nir_store_var(&b
, var
, nir_load_param(&b
, i
), ~0);
701 _mesa_hash_table_insert(var_table
, param
, var
);
705 visit_exec_list(&ir
->body
, this);
707 this->is_global
= true;
714 nir_visitor::visit(ir_loop
*ir
)
717 visit_exec_list(&ir
->body_instructions
, this);
718 nir_pop_loop(&b
, NULL
);
722 nir_visitor::visit(ir_if
*ir
)
724 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
725 visit_exec_list(&ir
->then_instructions
, this);
726 nir_push_else(&b
, NULL
);
727 visit_exec_list(&ir
->else_instructions
, this);
728 nir_pop_if(&b
, NULL
);
732 nir_visitor::visit(ir_discard
*ir
)
735 * discards aren't treated as control flow, because before we lower them
736 * they can appear anywhere in the shader and the stuff after them may still
737 * be executed (yay, crazy GLSL rules!). However, after lowering, all the
738 * discards will be immediately followed by a return.
741 nir_intrinsic_instr
*discard
;
743 discard
= nir_intrinsic_instr_create(this->shader
,
744 nir_intrinsic_discard_if
);
746 nir_src_for_ssa(evaluate_rvalue(ir
->condition
));
748 discard
= nir_intrinsic_instr_create(this->shader
, nir_intrinsic_discard
);
751 nir_builder_instr_insert(&b
, &discard
->instr
);
755 nir_visitor::visit(ir_emit_vertex
*ir
)
757 nir_intrinsic_instr
*instr
=
758 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_emit_vertex
);
759 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
760 nir_builder_instr_insert(&b
, &instr
->instr
);
764 nir_visitor::visit(ir_end_primitive
*ir
)
766 nir_intrinsic_instr
*instr
=
767 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_end_primitive
);
768 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
769 nir_builder_instr_insert(&b
, &instr
->instr
);
773 nir_visitor::visit(ir_loop_jump
*ir
)
777 case ir_loop_jump::jump_break
:
778 type
= nir_jump_break
;
780 case ir_loop_jump::jump_continue
:
781 type
= nir_jump_continue
;
784 unreachable("not reached");
787 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, type
);
788 nir_builder_instr_insert(&b
, &instr
->instr
);
792 nir_visitor::visit(ir_return
*ir
)
794 if (ir
->value
!= NULL
) {
795 nir_deref_instr
*ret_deref
=
796 nir_build_deref_cast(&b
, nir_load_param(&b
, 0),
797 nir_var_function_temp
, ir
->value
->type
, 0);
799 nir_ssa_def
*val
= evaluate_rvalue(ir
->value
);
800 nir_store_deref(&b
, ret_deref
, val
, ~0);
803 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, nir_jump_return
);
804 nir_builder_instr_insert(&b
, &instr
->instr
);
808 intrinsic_set_std430_align(nir_intrinsic_instr
*intrin
, const glsl_type
*type
)
810 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
811 unsigned pow2_components
= util_next_power_of_two(type
->vector_elements
);
812 nir_intrinsic_set_align(intrin
, (bit_size
/ 8) * pow2_components
, 0);
816 nir_visitor::visit(ir_call
*ir
)
818 if (ir
->callee
->is_intrinsic()) {
821 switch (ir
->callee
->intrinsic_id
) {
822 case ir_intrinsic_generic_atomic_add
:
823 op
= ir
->return_deref
->type
->is_integer_32_64()
824 ? nir_intrinsic_deref_atomic_add
: nir_intrinsic_deref_atomic_fadd
;
826 case ir_intrinsic_generic_atomic_and
:
827 op
= nir_intrinsic_deref_atomic_and
;
829 case ir_intrinsic_generic_atomic_or
:
830 op
= nir_intrinsic_deref_atomic_or
;
832 case ir_intrinsic_generic_atomic_xor
:
833 op
= nir_intrinsic_deref_atomic_xor
;
835 case ir_intrinsic_generic_atomic_min
:
836 assert(ir
->return_deref
);
837 if (ir
->return_deref
->type
== glsl_type::int_type
)
838 op
= nir_intrinsic_deref_atomic_imin
;
839 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
840 op
= nir_intrinsic_deref_atomic_umin
;
841 else if (ir
->return_deref
->type
== glsl_type::float_type
)
842 op
= nir_intrinsic_deref_atomic_fmin
;
844 unreachable("Invalid type");
846 case ir_intrinsic_generic_atomic_max
:
847 assert(ir
->return_deref
);
848 if (ir
->return_deref
->type
== glsl_type::int_type
)
849 op
= nir_intrinsic_deref_atomic_imax
;
850 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
851 op
= nir_intrinsic_deref_atomic_umax
;
852 else if (ir
->return_deref
->type
== glsl_type::float_type
)
853 op
= nir_intrinsic_deref_atomic_fmax
;
855 unreachable("Invalid type");
857 case ir_intrinsic_generic_atomic_exchange
:
858 op
= nir_intrinsic_deref_atomic_exchange
;
860 case ir_intrinsic_generic_atomic_comp_swap
:
861 op
= ir
->return_deref
->type
->is_integer_32_64()
862 ? nir_intrinsic_deref_atomic_comp_swap
863 : nir_intrinsic_deref_atomic_fcomp_swap
;
865 case ir_intrinsic_atomic_counter_read
:
866 op
= nir_intrinsic_atomic_counter_read_deref
;
868 case ir_intrinsic_atomic_counter_increment
:
869 op
= nir_intrinsic_atomic_counter_inc_deref
;
871 case ir_intrinsic_atomic_counter_predecrement
:
872 op
= nir_intrinsic_atomic_counter_pre_dec_deref
;
874 case ir_intrinsic_atomic_counter_add
:
875 op
= nir_intrinsic_atomic_counter_add_deref
;
877 case ir_intrinsic_atomic_counter_and
:
878 op
= nir_intrinsic_atomic_counter_and_deref
;
880 case ir_intrinsic_atomic_counter_or
:
881 op
= nir_intrinsic_atomic_counter_or_deref
;
883 case ir_intrinsic_atomic_counter_xor
:
884 op
= nir_intrinsic_atomic_counter_xor_deref
;
886 case ir_intrinsic_atomic_counter_min
:
887 op
= nir_intrinsic_atomic_counter_min_deref
;
889 case ir_intrinsic_atomic_counter_max
:
890 op
= nir_intrinsic_atomic_counter_max_deref
;
892 case ir_intrinsic_atomic_counter_exchange
:
893 op
= nir_intrinsic_atomic_counter_exchange_deref
;
895 case ir_intrinsic_atomic_counter_comp_swap
:
896 op
= nir_intrinsic_atomic_counter_comp_swap_deref
;
898 case ir_intrinsic_image_load
:
899 op
= nir_intrinsic_image_deref_load
;
901 case ir_intrinsic_image_store
:
902 op
= nir_intrinsic_image_deref_store
;
904 case ir_intrinsic_image_atomic_add
:
905 op
= ir
->return_deref
->type
->is_integer_32_64()
906 ? nir_intrinsic_image_deref_atomic_add
907 : nir_intrinsic_image_deref_atomic_fadd
;
909 case ir_intrinsic_image_atomic_min
:
910 op
= nir_intrinsic_image_deref_atomic_min
;
912 case ir_intrinsic_image_atomic_max
:
913 op
= nir_intrinsic_image_deref_atomic_max
;
915 case ir_intrinsic_image_atomic_and
:
916 op
= nir_intrinsic_image_deref_atomic_and
;
918 case ir_intrinsic_image_atomic_or
:
919 op
= nir_intrinsic_image_deref_atomic_or
;
921 case ir_intrinsic_image_atomic_xor
:
922 op
= nir_intrinsic_image_deref_atomic_xor
;
924 case ir_intrinsic_image_atomic_exchange
:
925 op
= nir_intrinsic_image_deref_atomic_exchange
;
927 case ir_intrinsic_image_atomic_comp_swap
:
928 op
= nir_intrinsic_image_deref_atomic_comp_swap
;
930 case ir_intrinsic_memory_barrier
:
931 op
= nir_intrinsic_memory_barrier
;
933 case ir_intrinsic_image_size
:
934 op
= nir_intrinsic_image_deref_size
;
936 case ir_intrinsic_image_samples
:
937 op
= nir_intrinsic_image_deref_samples
;
939 case ir_intrinsic_ssbo_store
:
940 op
= nir_intrinsic_store_ssbo
;
942 case ir_intrinsic_ssbo_load
:
943 op
= nir_intrinsic_load_ssbo
;
945 case ir_intrinsic_ssbo_atomic_add
:
946 op
= ir
->return_deref
->type
->is_integer_32_64()
947 ? nir_intrinsic_ssbo_atomic_add
: nir_intrinsic_ssbo_atomic_fadd
;
949 case ir_intrinsic_ssbo_atomic_and
:
950 op
= nir_intrinsic_ssbo_atomic_and
;
952 case ir_intrinsic_ssbo_atomic_or
:
953 op
= nir_intrinsic_ssbo_atomic_or
;
955 case ir_intrinsic_ssbo_atomic_xor
:
956 op
= nir_intrinsic_ssbo_atomic_xor
;
958 case ir_intrinsic_ssbo_atomic_min
:
959 assert(ir
->return_deref
);
960 if (ir
->return_deref
->type
== glsl_type::int_type
)
961 op
= nir_intrinsic_ssbo_atomic_imin
;
962 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
963 op
= nir_intrinsic_ssbo_atomic_umin
;
964 else if (ir
->return_deref
->type
== glsl_type::float_type
)
965 op
= nir_intrinsic_ssbo_atomic_fmin
;
967 unreachable("Invalid type");
969 case ir_intrinsic_ssbo_atomic_max
:
970 assert(ir
->return_deref
);
971 if (ir
->return_deref
->type
== glsl_type::int_type
)
972 op
= nir_intrinsic_ssbo_atomic_imax
;
973 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
974 op
= nir_intrinsic_ssbo_atomic_umax
;
975 else if (ir
->return_deref
->type
== glsl_type::float_type
)
976 op
= nir_intrinsic_ssbo_atomic_fmax
;
978 unreachable("Invalid type");
980 case ir_intrinsic_ssbo_atomic_exchange
:
981 op
= nir_intrinsic_ssbo_atomic_exchange
;
983 case ir_intrinsic_ssbo_atomic_comp_swap
:
984 op
= ir
->return_deref
->type
->is_integer_32_64()
985 ? nir_intrinsic_ssbo_atomic_comp_swap
986 : nir_intrinsic_ssbo_atomic_fcomp_swap
;
988 case ir_intrinsic_shader_clock
:
989 op
= nir_intrinsic_shader_clock
;
991 case ir_intrinsic_begin_invocation_interlock
:
992 op
= nir_intrinsic_begin_invocation_interlock
;
994 case ir_intrinsic_end_invocation_interlock
:
995 op
= nir_intrinsic_end_invocation_interlock
;
997 case ir_intrinsic_group_memory_barrier
:
998 op
= nir_intrinsic_group_memory_barrier
;
1000 case ir_intrinsic_memory_barrier_atomic_counter
:
1001 op
= nir_intrinsic_memory_barrier_atomic_counter
;
1003 case ir_intrinsic_memory_barrier_buffer
:
1004 op
= nir_intrinsic_memory_barrier_buffer
;
1006 case ir_intrinsic_memory_barrier_image
:
1007 op
= nir_intrinsic_memory_barrier_image
;
1009 case ir_intrinsic_memory_barrier_shared
:
1010 op
= nir_intrinsic_memory_barrier_shared
;
1012 case ir_intrinsic_shared_load
:
1013 op
= nir_intrinsic_load_shared
;
1015 case ir_intrinsic_shared_store
:
1016 op
= nir_intrinsic_store_shared
;
1018 case ir_intrinsic_shared_atomic_add
:
1019 op
= ir
->return_deref
->type
->is_integer_32_64()
1020 ? nir_intrinsic_shared_atomic_add
1021 : nir_intrinsic_shared_atomic_fadd
;
1023 case ir_intrinsic_shared_atomic_and
:
1024 op
= nir_intrinsic_shared_atomic_and
;
1026 case ir_intrinsic_shared_atomic_or
:
1027 op
= nir_intrinsic_shared_atomic_or
;
1029 case ir_intrinsic_shared_atomic_xor
:
1030 op
= nir_intrinsic_shared_atomic_xor
;
1032 case ir_intrinsic_shared_atomic_min
:
1033 assert(ir
->return_deref
);
1034 if (ir
->return_deref
->type
== glsl_type::int_type
)
1035 op
= nir_intrinsic_shared_atomic_imin
;
1036 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1037 op
= nir_intrinsic_shared_atomic_umin
;
1038 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1039 op
= nir_intrinsic_shared_atomic_fmin
;
1041 unreachable("Invalid type");
1043 case ir_intrinsic_shared_atomic_max
:
1044 assert(ir
->return_deref
);
1045 if (ir
->return_deref
->type
== glsl_type::int_type
)
1046 op
= nir_intrinsic_shared_atomic_imax
;
1047 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1048 op
= nir_intrinsic_shared_atomic_umax
;
1049 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1050 op
= nir_intrinsic_shared_atomic_fmax
;
1052 unreachable("Invalid type");
1054 case ir_intrinsic_shared_atomic_exchange
:
1055 op
= nir_intrinsic_shared_atomic_exchange
;
1057 case ir_intrinsic_shared_atomic_comp_swap
:
1058 op
= ir
->return_deref
->type
->is_integer_32_64()
1059 ? nir_intrinsic_shared_atomic_comp_swap
1060 : nir_intrinsic_shared_atomic_fcomp_swap
;
1062 case ir_intrinsic_vote_any
:
1063 op
= nir_intrinsic_vote_any
;
1065 case ir_intrinsic_vote_all
:
1066 op
= nir_intrinsic_vote_all
;
1068 case ir_intrinsic_vote_eq
:
1069 op
= nir_intrinsic_vote_ieq
;
1071 case ir_intrinsic_ballot
:
1072 op
= nir_intrinsic_ballot
;
1074 case ir_intrinsic_read_invocation
:
1075 op
= nir_intrinsic_read_invocation
;
1077 case ir_intrinsic_read_first_invocation
:
1078 op
= nir_intrinsic_read_first_invocation
;
1081 unreachable("not reached");
1084 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(shader
, op
);
1085 nir_ssa_def
*ret
= &instr
->dest
.ssa
;
1088 case nir_intrinsic_deref_atomic_add
:
1089 case nir_intrinsic_deref_atomic_imin
:
1090 case nir_intrinsic_deref_atomic_umin
:
1091 case nir_intrinsic_deref_atomic_imax
:
1092 case nir_intrinsic_deref_atomic_umax
:
1093 case nir_intrinsic_deref_atomic_and
:
1094 case nir_intrinsic_deref_atomic_or
:
1095 case nir_intrinsic_deref_atomic_xor
:
1096 case nir_intrinsic_deref_atomic_exchange
:
1097 case nir_intrinsic_deref_atomic_comp_swap
:
1098 case nir_intrinsic_deref_atomic_fadd
:
1099 case nir_intrinsic_deref_atomic_fmin
:
1100 case nir_intrinsic_deref_atomic_fmax
:
1101 case nir_intrinsic_deref_atomic_fcomp_swap
: {
1102 int param_count
= ir
->actual_parameters
.length();
1103 assert(param_count
== 2 || param_count
== 3);
1106 exec_node
*param
= ir
->actual_parameters
.get_head();
1107 ir_rvalue
*rvalue
= (ir_rvalue
*) param
;
1108 ir_dereference
*deref
= rvalue
->as_dereference();
1109 ir_swizzle
*swizzle
= NULL
;
1111 /* We may have a swizzle to pick off a single vec4 component */
1112 swizzle
= rvalue
->as_swizzle();
1113 assert(swizzle
&& swizzle
->type
->vector_elements
== 1);
1114 deref
= swizzle
->val
->as_dereference();
1117 nir_deref_instr
*nir_deref
= evaluate_deref(deref
);
1119 nir_deref
= nir_build_deref_array_imm(&b
, nir_deref
,
1122 instr
->src
[0] = nir_src_for_ssa(&nir_deref
->dest
.ssa
);
1124 /* data1 parameter (this is always present) */
1125 param
= param
->get_next();
1126 ir_instruction
*inst
= (ir_instruction
*) param
;
1127 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1129 /* data2 parameter (only with atomic_comp_swap) */
1130 if (param_count
== 3) {
1131 assert(op
== nir_intrinsic_deref_atomic_comp_swap
||
1132 op
== nir_intrinsic_deref_atomic_fcomp_swap
);
1133 param
= param
->get_next();
1134 inst
= (ir_instruction
*) param
;
1135 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1139 assert(ir
->return_deref
);
1140 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1141 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1142 nir_builder_instr_insert(&b
, &instr
->instr
);
1145 case nir_intrinsic_atomic_counter_read_deref
:
1146 case nir_intrinsic_atomic_counter_inc_deref
:
1147 case nir_intrinsic_atomic_counter_pre_dec_deref
:
1148 case nir_intrinsic_atomic_counter_add_deref
:
1149 case nir_intrinsic_atomic_counter_min_deref
:
1150 case nir_intrinsic_atomic_counter_max_deref
:
1151 case nir_intrinsic_atomic_counter_and_deref
:
1152 case nir_intrinsic_atomic_counter_or_deref
:
1153 case nir_intrinsic_atomic_counter_xor_deref
:
1154 case nir_intrinsic_atomic_counter_exchange_deref
:
1155 case nir_intrinsic_atomic_counter_comp_swap_deref
: {
1156 /* Set the counter variable dereference. */
1157 exec_node
*param
= ir
->actual_parameters
.get_head();
1158 ir_dereference
*counter
= (ir_dereference
*)param
;
1160 instr
->src
[0] = nir_src_for_ssa(&evaluate_deref(counter
)->dest
.ssa
);
1161 param
= param
->get_next();
1163 /* Set the intrinsic destination. */
1164 if (ir
->return_deref
) {
1165 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1168 /* Set the intrinsic parameters. */
1169 if (!param
->is_tail_sentinel()) {
1171 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1172 param
= param
->get_next();
1175 if (!param
->is_tail_sentinel()) {
1177 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1178 param
= param
->get_next();
1181 nir_builder_instr_insert(&b
, &instr
->instr
);
1184 case nir_intrinsic_image_deref_load
:
1185 case nir_intrinsic_image_deref_store
:
1186 case nir_intrinsic_image_deref_atomic_add
:
1187 case nir_intrinsic_image_deref_atomic_min
:
1188 case nir_intrinsic_image_deref_atomic_max
:
1189 case nir_intrinsic_image_deref_atomic_and
:
1190 case nir_intrinsic_image_deref_atomic_or
:
1191 case nir_intrinsic_image_deref_atomic_xor
:
1192 case nir_intrinsic_image_deref_atomic_exchange
:
1193 case nir_intrinsic_image_deref_atomic_comp_swap
:
1194 case nir_intrinsic_image_deref_atomic_fadd
:
1195 case nir_intrinsic_image_deref_samples
:
1196 case nir_intrinsic_image_deref_size
: {
1197 nir_ssa_undef_instr
*instr_undef
=
1198 nir_ssa_undef_instr_create(shader
, 1, 32);
1199 nir_builder_instr_insert(&b
, &instr_undef
->instr
);
1201 /* Set the image variable dereference. */
1202 exec_node
*param
= ir
->actual_parameters
.get_head();
1203 ir_dereference
*image
= (ir_dereference
*)param
;
1204 nir_deref_instr
*deref
= evaluate_deref(image
);
1205 const glsl_type
*type
= deref
->type
;
1207 instr
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
1208 param
= param
->get_next();
1210 /* Set the intrinsic destination. */
1211 if (ir
->return_deref
) {
1212 unsigned num_components
= ir
->return_deref
->type
->vector_elements
;
1213 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1214 num_components
, 32, NULL
);
1217 if (op
== nir_intrinsic_image_deref_size
) {
1218 instr
->num_components
= instr
->dest
.ssa
.num_components
;
1219 } else if (op
== nir_intrinsic_image_deref_load
||
1220 op
== nir_intrinsic_image_deref_store
) {
1221 instr
->num_components
= 4;
1224 if (op
== nir_intrinsic_image_deref_size
||
1225 op
== nir_intrinsic_image_deref_samples
) {
1226 nir_builder_instr_insert(&b
, &instr
->instr
);
1230 /* Set the address argument, extending the coordinate vector to four
1233 nir_ssa_def
*src_addr
=
1234 evaluate_rvalue((ir_dereference
*)param
);
1235 nir_ssa_def
*srcs
[4];
1237 for (int i
= 0; i
< 4; i
++) {
1238 if (i
< type
->coordinate_components())
1239 srcs
[i
] = nir_channel(&b
, src_addr
, i
);
1241 srcs
[i
] = &instr_undef
->def
;
1244 instr
->src
[1] = nir_src_for_ssa(nir_vec(&b
, srcs
, 4));
1245 param
= param
->get_next();
1247 /* Set the sample argument, which is undefined for single-sample
1250 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_MS
) {
1252 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1253 param
= param
->get_next();
1255 instr
->src
[2] = nir_src_for_ssa(&instr_undef
->def
);
1258 /* Set the intrinsic parameters. */
1259 if (!param
->is_tail_sentinel()) {
1261 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1262 param
= param
->get_next();
1265 if (!param
->is_tail_sentinel()) {
1267 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1268 param
= param
->get_next();
1270 nir_builder_instr_insert(&b
, &instr
->instr
);
1273 case nir_intrinsic_memory_barrier
:
1274 case nir_intrinsic_group_memory_barrier
:
1275 case nir_intrinsic_memory_barrier_atomic_counter
:
1276 case nir_intrinsic_memory_barrier_buffer
:
1277 case nir_intrinsic_memory_barrier_image
:
1278 case nir_intrinsic_memory_barrier_shared
:
1279 nir_builder_instr_insert(&b
, &instr
->instr
);
1281 case nir_intrinsic_shader_clock
:
1282 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 2, 32, NULL
);
1283 instr
->num_components
= 2;
1284 nir_builder_instr_insert(&b
, &instr
->instr
);
1286 case nir_intrinsic_begin_invocation_interlock
:
1287 nir_builder_instr_insert(&b
, &instr
->instr
);
1289 case nir_intrinsic_end_invocation_interlock
:
1290 nir_builder_instr_insert(&b
, &instr
->instr
);
1292 case nir_intrinsic_store_ssbo
: {
1293 exec_node
*param
= ir
->actual_parameters
.get_head();
1294 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1296 param
= param
->get_next();
1297 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1299 param
= param
->get_next();
1300 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1302 param
= param
->get_next();
1303 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1306 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1307 if (val
->type
->is_boolean())
1308 nir_val
= nir_b2i32(&b
, nir_val
);
1310 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1311 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(block
));
1312 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(offset
));
1313 intrinsic_set_std430_align(instr
, val
->type
);
1314 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1315 instr
->num_components
= val
->type
->vector_elements
;
1317 nir_builder_instr_insert(&b
, &instr
->instr
);
1320 case nir_intrinsic_load_ssbo
: {
1321 exec_node
*param
= ir
->actual_parameters
.get_head();
1322 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1324 param
= param
->get_next();
1325 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1327 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(block
));
1328 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1330 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1331 instr
->num_components
= type
->vector_elements
;
1332 intrinsic_set_std430_align(instr
, type
);
1334 /* Setup destination register */
1335 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1336 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1337 type
->vector_elements
, bit_size
, NULL
);
1339 /* Insert the created nir instruction now since in the case of boolean
1340 * result we will need to emit another instruction after it
1342 nir_builder_instr_insert(&b
, &instr
->instr
);
1345 * In SSBO/UBO's, a true boolean value is any non-zero value, but we
1346 * consider a true boolean to be ~0. Fix this up with a != 0
1349 if (type
->is_boolean())
1350 ret
= nir_i2b(&b
, &instr
->dest
.ssa
);
1353 case nir_intrinsic_ssbo_atomic_add
:
1354 case nir_intrinsic_ssbo_atomic_imin
:
1355 case nir_intrinsic_ssbo_atomic_umin
:
1356 case nir_intrinsic_ssbo_atomic_imax
:
1357 case nir_intrinsic_ssbo_atomic_umax
:
1358 case nir_intrinsic_ssbo_atomic_and
:
1359 case nir_intrinsic_ssbo_atomic_or
:
1360 case nir_intrinsic_ssbo_atomic_xor
:
1361 case nir_intrinsic_ssbo_atomic_exchange
:
1362 case nir_intrinsic_ssbo_atomic_comp_swap
:
1363 case nir_intrinsic_ssbo_atomic_fadd
:
1364 case nir_intrinsic_ssbo_atomic_fmin
:
1365 case nir_intrinsic_ssbo_atomic_fmax
:
1366 case nir_intrinsic_ssbo_atomic_fcomp_swap
: {
1367 int param_count
= ir
->actual_parameters
.length();
1368 assert(param_count
== 3 || param_count
== 4);
1371 exec_node
*param
= ir
->actual_parameters
.get_head();
1372 ir_instruction
*inst
= (ir_instruction
*) param
;
1373 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1376 param
= param
->get_next();
1377 inst
= (ir_instruction
*) param
;
1378 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1380 /* data1 parameter (this is always present) */
1381 param
= param
->get_next();
1382 inst
= (ir_instruction
*) param
;
1383 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1385 /* data2 parameter (only with atomic_comp_swap) */
1386 if (param_count
== 4) {
1387 assert(op
== nir_intrinsic_ssbo_atomic_comp_swap
||
1388 op
== nir_intrinsic_ssbo_atomic_fcomp_swap
);
1389 param
= param
->get_next();
1390 inst
= (ir_instruction
*) param
;
1391 instr
->src
[3] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1395 assert(ir
->return_deref
);
1396 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1397 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1398 nir_builder_instr_insert(&b
, &instr
->instr
);
1401 case nir_intrinsic_load_shared
: {
1402 exec_node
*param
= ir
->actual_parameters
.get_head();
1403 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1405 nir_intrinsic_set_base(instr
, 0);
1406 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(offset
));
1408 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1409 instr
->num_components
= type
->vector_elements
;
1410 intrinsic_set_std430_align(instr
, type
);
1412 /* Setup destination register */
1413 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1414 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1415 type
->vector_elements
, bit_size
, NULL
);
1417 nir_builder_instr_insert(&b
, &instr
->instr
);
1419 /* The value in shared memory is a 32-bit value */
1420 if (type
->is_boolean())
1421 ret
= nir_i2b(&b
, &instr
->dest
.ssa
);
1424 case nir_intrinsic_store_shared
: {
1425 exec_node
*param
= ir
->actual_parameters
.get_head();
1426 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1428 param
= param
->get_next();
1429 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1431 param
= param
->get_next();
1432 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1435 nir_intrinsic_set_base(instr
, 0);
1436 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1438 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1440 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1441 /* The value in shared memory is a 32-bit value */
1442 if (val
->type
->is_boolean())
1443 nir_val
= nir_b2i32(&b
, nir_val
);
1445 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1446 instr
->num_components
= val
->type
->vector_elements
;
1447 intrinsic_set_std430_align(instr
, val
->type
);
1449 nir_builder_instr_insert(&b
, &instr
->instr
);
1452 case nir_intrinsic_shared_atomic_add
:
1453 case nir_intrinsic_shared_atomic_imin
:
1454 case nir_intrinsic_shared_atomic_umin
:
1455 case nir_intrinsic_shared_atomic_imax
:
1456 case nir_intrinsic_shared_atomic_umax
:
1457 case nir_intrinsic_shared_atomic_and
:
1458 case nir_intrinsic_shared_atomic_or
:
1459 case nir_intrinsic_shared_atomic_xor
:
1460 case nir_intrinsic_shared_atomic_exchange
:
1461 case nir_intrinsic_shared_atomic_comp_swap
:
1462 case nir_intrinsic_shared_atomic_fadd
:
1463 case nir_intrinsic_shared_atomic_fmin
:
1464 case nir_intrinsic_shared_atomic_fmax
:
1465 case nir_intrinsic_shared_atomic_fcomp_swap
: {
1466 int param_count
= ir
->actual_parameters
.length();
1467 assert(param_count
== 2 || param_count
== 3);
1470 exec_node
*param
= ir
->actual_parameters
.get_head();
1471 ir_instruction
*inst
= (ir_instruction
*) param
;
1472 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1474 /* data1 parameter (this is always present) */
1475 param
= param
->get_next();
1476 inst
= (ir_instruction
*) param
;
1477 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1479 /* data2 parameter (only with atomic_comp_swap) */
1480 if (param_count
== 3) {
1481 assert(op
== nir_intrinsic_shared_atomic_comp_swap
||
1482 op
== nir_intrinsic_shared_atomic_fcomp_swap
);
1483 param
= param
->get_next();
1484 inst
= (ir_instruction
*) param
;
1486 nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1490 assert(ir
->return_deref
);
1491 unsigned bit_size
= glsl_get_bit_size(ir
->return_deref
->type
);
1492 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1493 ir
->return_deref
->type
->vector_elements
,
1495 nir_builder_instr_insert(&b
, &instr
->instr
);
1498 case nir_intrinsic_vote_any
:
1499 case nir_intrinsic_vote_all
:
1500 case nir_intrinsic_vote_ieq
: {
1501 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 1, NULL
);
1502 instr
->num_components
= 1;
1504 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1505 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1507 nir_builder_instr_insert(&b
, &instr
->instr
);
1511 case nir_intrinsic_ballot
: {
1512 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1513 ir
->return_deref
->type
->vector_elements
, 64, NULL
);
1514 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1516 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1517 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1519 nir_builder_instr_insert(&b
, &instr
->instr
);
1522 case nir_intrinsic_read_invocation
: {
1523 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1524 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1525 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1527 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1528 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1530 ir_rvalue
*invocation
= (ir_rvalue
*) ir
->actual_parameters
.get_head()->next
;
1531 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(invocation
));
1533 nir_builder_instr_insert(&b
, &instr
->instr
);
1536 case nir_intrinsic_read_first_invocation
: {
1537 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1538 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1539 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1541 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1542 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1544 nir_builder_instr_insert(&b
, &instr
->instr
);
1548 unreachable("not reached");
1551 if (ir
->return_deref
)
1552 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), ret
, ~0);
1557 struct hash_entry
*entry
=
1558 _mesa_hash_table_search(this->overload_table
, ir
->callee
);
1560 nir_function
*callee
= (nir_function
*) entry
->data
;
1562 nir_call_instr
*call
= nir_call_instr_create(this->shader
, callee
);
1565 nir_deref_instr
*ret_deref
= NULL
;
1566 if (ir
->return_deref
) {
1567 nir_variable
*ret_tmp
=
1568 nir_local_variable_create(this->impl
, ir
->return_deref
->type
,
1570 ret_deref
= nir_build_deref_var(&b
, ret_tmp
);
1571 call
->params
[i
++] = nir_src_for_ssa(&ret_deref
->dest
.ssa
);
1574 foreach_two_lists(formal_node
, &ir
->callee
->parameters
,
1575 actual_node
, &ir
->actual_parameters
) {
1576 ir_rvalue
*param_rvalue
= (ir_rvalue
*) actual_node
;
1577 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
1579 if (sig_param
->data
.mode
== ir_var_function_out
) {
1580 nir_deref_instr
*out_deref
= evaluate_deref(param_rvalue
);
1581 call
->params
[i
] = nir_src_for_ssa(&out_deref
->dest
.ssa
);
1582 } else if (sig_param
->data
.mode
== ir_var_function_in
) {
1583 nir_ssa_def
*val
= evaluate_rvalue(param_rvalue
);
1584 nir_src src
= nir_src_for_ssa(val
);
1586 nir_src_copy(&call
->params
[i
], &src
, call
);
1587 } else if (sig_param
->data
.mode
== ir_var_function_inout
) {
1588 unreachable("unimplemented: inout parameters");
1594 nir_builder_instr_insert(&b
, &call
->instr
);
1596 if (ir
->return_deref
)
1597 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), nir_load_deref(&b
, ret_deref
), ~0);
1601 nir_visitor::visit(ir_assignment
*ir
)
1603 unsigned num_components
= ir
->lhs
->type
->vector_elements
;
1605 b
.exact
= ir
->lhs
->variable_referenced()->data
.invariant
||
1606 ir
->lhs
->variable_referenced()->data
.precise
;
1608 if ((ir
->rhs
->as_dereference() || ir
->rhs
->as_constant()) &&
1609 (ir
->write_mask
== (1 << num_components
) - 1 || ir
->write_mask
== 0)) {
1610 if (ir
->condition
) {
1611 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1612 nir_copy_deref(&b
, evaluate_deref(ir
->lhs
), evaluate_deref(ir
->rhs
));
1613 nir_pop_if(&b
, NULL
);
1615 nir_copy_deref(&b
, evaluate_deref(ir
->lhs
), evaluate_deref(ir
->rhs
));
1620 assert(ir
->rhs
->type
->is_scalar() || ir
->rhs
->type
->is_vector());
1622 ir
->lhs
->accept(this);
1623 nir_deref_instr
*lhs_deref
= this->deref
;
1624 nir_ssa_def
*src
= evaluate_rvalue(ir
->rhs
);
1626 if (ir
->write_mask
!= (1 << num_components
) - 1 && ir
->write_mask
!= 0) {
1627 /* GLSL IR will give us the input to the write-masked assignment in a
1628 * single packed vector. So, for example, if the writemask is xzw, then
1629 * we have to swizzle x -> x, y -> z, and z -> w and get the y component
1633 unsigned component
= 0;
1634 for (unsigned i
= 0; i
< 4; i
++) {
1635 swiz
[i
] = ir
->write_mask
& (1 << i
) ? component
++ : 0;
1637 src
= nir_swizzle(&b
, src
, swiz
, num_components
, false);
1640 if (ir
->condition
) {
1641 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1642 nir_store_deref(&b
, lhs_deref
, src
, ir
->write_mask
);
1643 nir_pop_if(&b
, NULL
);
1645 nir_store_deref(&b
, lhs_deref
, src
, ir
->write_mask
);
1650 * Given an instruction, returns a pointer to its destination or NULL if there
1651 * is no destination.
1653 * Note that this only handles instructions we generate at this level.
1656 get_instr_dest(nir_instr
*instr
)
1658 nir_alu_instr
*alu_instr
;
1659 nir_intrinsic_instr
*intrinsic_instr
;
1660 nir_tex_instr
*tex_instr
;
1662 switch (instr
->type
) {
1663 case nir_instr_type_alu
:
1664 alu_instr
= nir_instr_as_alu(instr
);
1665 return &alu_instr
->dest
.dest
;
1667 case nir_instr_type_intrinsic
:
1668 intrinsic_instr
= nir_instr_as_intrinsic(instr
);
1669 if (nir_intrinsic_infos
[intrinsic_instr
->intrinsic
].has_dest
)
1670 return &intrinsic_instr
->dest
;
1674 case nir_instr_type_tex
:
1675 tex_instr
= nir_instr_as_tex(instr
);
1676 return &tex_instr
->dest
;
1679 unreachable("not reached");
1686 nir_visitor::add_instr(nir_instr
*instr
, unsigned num_components
,
1689 nir_dest
*dest
= get_instr_dest(instr
);
1692 nir_ssa_dest_init(instr
, dest
, num_components
, bit_size
, NULL
);
1694 nir_builder_instr_insert(&b
, instr
);
1697 assert(dest
->is_ssa
);
1698 this->result
= &dest
->ssa
;
1703 nir_visitor::evaluate_rvalue(ir_rvalue
* ir
)
1706 if (ir
->as_dereference() || ir
->as_constant()) {
1708 * A dereference is being used on the right hand side, which means we
1709 * must emit a variable load.
1712 this->result
= nir_load_deref(&b
, this->deref
);
1715 return this->result
;
1719 type_is_float(glsl_base_type type
)
1721 return type
== GLSL_TYPE_FLOAT
|| type
== GLSL_TYPE_DOUBLE
||
1722 type
== GLSL_TYPE_FLOAT16
;
1726 type_is_signed(glsl_base_type type
)
1728 return type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_INT64
||
1729 type
== GLSL_TYPE_INT16
;
1733 nir_visitor::visit(ir_expression
*ir
)
1735 /* Some special cases */
1736 switch (ir
->operation
) {
1737 case ir_binop_ubo_load
: {
1738 nir_intrinsic_instr
*load
=
1739 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_ubo
);
1740 unsigned bit_size
= ir
->type
->is_boolean() ? 32 :
1741 glsl_get_bit_size(ir
->type
);
1742 load
->num_components
= ir
->type
->vector_elements
;
1743 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1744 load
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1745 intrinsic_set_std430_align(load
, ir
->type
);
1746 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1749 * In UBO's, a true boolean value is any non-zero value, but we consider
1750 * a true boolean to be ~0. Fix this up with a != 0 comparison.
1753 if (ir
->type
->is_boolean())
1754 this->result
= nir_i2b(&b
, &load
->dest
.ssa
);
1759 case ir_unop_interpolate_at_centroid
:
1760 case ir_binop_interpolate_at_offset
:
1761 case ir_binop_interpolate_at_sample
: {
1762 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1763 ir_swizzle
*swizzle
= NULL
;
1765 /* the api does not allow a swizzle here, but the varying packing code
1766 * may have pushed one into here.
1768 swizzle
= ir
->operands
[0]->as_swizzle();
1770 deref
= swizzle
->val
->as_dereference();
1774 deref
->accept(this);
1776 nir_intrinsic_op op
;
1777 if (this->deref
->mode
== nir_var_shader_in
) {
1778 switch (ir
->operation
) {
1779 case ir_unop_interpolate_at_centroid
:
1780 op
= nir_intrinsic_interp_deref_at_centroid
;
1782 case ir_binop_interpolate_at_offset
:
1783 op
= nir_intrinsic_interp_deref_at_offset
;
1785 case ir_binop_interpolate_at_sample
:
1786 op
= nir_intrinsic_interp_deref_at_sample
;
1789 unreachable("Invalid interpolation intrinsic");
1792 /* This case can happen if the vertex shader does not write the
1793 * given varying. In this case, the linker will lower it to a
1794 * global variable. Since interpolating a variable makes no
1795 * sense, we'll just turn it into a load which will probably
1796 * eventually end up as an SSA definition.
1798 assert(this->deref
->mode
== nir_var_shader_temp
);
1799 op
= nir_intrinsic_load_deref
;
1802 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(shader
, op
);
1803 intrin
->num_components
= deref
->type
->vector_elements
;
1804 intrin
->src
[0] = nir_src_for_ssa(&this->deref
->dest
.ssa
);
1806 if (intrin
->intrinsic
== nir_intrinsic_interp_deref_at_offset
||
1807 intrin
->intrinsic
== nir_intrinsic_interp_deref_at_sample
)
1808 intrin
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1810 unsigned bit_size
= glsl_get_bit_size(deref
->type
);
1811 add_instr(&intrin
->instr
, deref
->type
->vector_elements
, bit_size
);
1814 unsigned swiz
[4] = {
1815 swizzle
->mask
.x
, swizzle
->mask
.y
, swizzle
->mask
.z
, swizzle
->mask
.w
1818 result
= nir_swizzle(&b
, result
, swiz
,
1819 swizzle
->type
->vector_elements
, false);
1825 case ir_unop_ssbo_unsized_array_length
: {
1826 nir_intrinsic_instr
*intrin
=
1827 nir_intrinsic_instr_create(b
.shader
,
1828 nir_intrinsic_deref_buffer_array_length
);
1830 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1831 intrin
->src
[0] = nir_src_for_ssa(&evaluate_deref(deref
)->dest
.ssa
);
1833 add_instr(&intrin
->instr
, 1, 32);
1841 nir_ssa_def
*srcs
[4];
1842 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1843 srcs
[i
] = evaluate_rvalue(ir
->operands
[i
]);
1845 glsl_base_type types
[4];
1846 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1847 types
[i
] = ir
->operands
[i
]->type
->base_type
;
1849 glsl_base_type out_type
= ir
->type
->base_type
;
1851 switch (ir
->operation
) {
1852 case ir_unop_bit_not
: result
= nir_inot(&b
, srcs
[0]); break;
1853 case ir_unop_logic_not
:
1854 result
= nir_inot(&b
, srcs
[0]);
1857 result
= type_is_float(types
[0]) ? nir_fneg(&b
, srcs
[0])
1858 : nir_ineg(&b
, srcs
[0]);
1861 result
= type_is_float(types
[0]) ? nir_fabs(&b
, srcs
[0])
1862 : nir_iabs(&b
, srcs
[0]);
1864 case ir_unop_saturate
:
1865 assert(type_is_float(types
[0]));
1866 result
= nir_fsat(&b
, srcs
[0]);
1869 result
= type_is_float(types
[0]) ? nir_fsign(&b
, srcs
[0])
1870 : nir_isign(&b
, srcs
[0]);
1872 case ir_unop_rcp
: result
= nir_frcp(&b
, srcs
[0]); break;
1873 case ir_unop_rsq
: result
= nir_frsq(&b
, srcs
[0]); break;
1874 case ir_unop_sqrt
: result
= nir_fsqrt(&b
, srcs
[0]); break;
1875 case ir_unop_exp
: unreachable("ir_unop_exp should have been lowered");
1876 case ir_unop_log
: unreachable("ir_unop_log should have been lowered");
1877 case ir_unop_exp2
: result
= nir_fexp2(&b
, srcs
[0]); break;
1878 case ir_unop_log2
: result
= nir_flog2(&b
, srcs
[0]); break;
1880 result
= nir_i2f32(&b
, srcs
[0]);
1883 result
= nir_u2f32(&b
, srcs
[0]);
1886 result
= nir_b2f32(&b
, srcs
[0]);
1889 result
= nir_f2i32(&b
, srcs
[0]);
1892 result
= nir_f2u32(&b
, srcs
[0]);
1924 case ir_unop_i642u64
:
1925 case ir_unop_u642i64
: {
1926 nir_alu_type src_type
= nir_get_nir_type_for_glsl_base_type(types
[0]);
1927 nir_alu_type dst_type
= nir_get_nir_type_for_glsl_base_type(out_type
);
1928 result
= nir_build_alu(&b
, nir_type_conversion_op(src_type
, dst_type
,
1929 nir_rounding_mode_undef
),
1930 srcs
[0], NULL
, NULL
, NULL
);
1931 /* b2i and b2f don't have fixed bit-size versions so the builder will
1932 * just assume 32 and we have to fix it up here.
1934 result
->bit_size
= nir_alu_type_get_type_size(dst_type
);
1938 case ir_unop_bitcast_i2f
:
1939 case ir_unop_bitcast_f2i
:
1940 case ir_unop_bitcast_u2f
:
1941 case ir_unop_bitcast_f2u
:
1942 case ir_unop_bitcast_i642d
:
1943 case ir_unop_bitcast_d2i64
:
1944 case ir_unop_bitcast_u642d
:
1945 case ir_unop_bitcast_d2u64
:
1946 case ir_unop_subroutine_to_int
:
1948 result
= nir_imov(&b
, srcs
[0]);
1950 case ir_unop_trunc
: result
= nir_ftrunc(&b
, srcs
[0]); break;
1951 case ir_unop_ceil
: result
= nir_fceil(&b
, srcs
[0]); break;
1952 case ir_unop_floor
: result
= nir_ffloor(&b
, srcs
[0]); break;
1953 case ir_unop_fract
: result
= nir_ffract(&b
, srcs
[0]); break;
1954 case ir_unop_frexp_exp
: result
= nir_frexp_exp(&b
, srcs
[0]); break;
1955 case ir_unop_frexp_sig
: result
= nir_frexp_sig(&b
, srcs
[0]); break;
1956 case ir_unop_round_even
: result
= nir_fround_even(&b
, srcs
[0]); break;
1957 case ir_unop_sin
: result
= nir_fsin(&b
, srcs
[0]); break;
1958 case ir_unop_cos
: result
= nir_fcos(&b
, srcs
[0]); break;
1959 case ir_unop_dFdx
: result
= nir_fddx(&b
, srcs
[0]); break;
1960 case ir_unop_dFdy
: result
= nir_fddy(&b
, srcs
[0]); break;
1961 case ir_unop_dFdx_fine
: result
= nir_fddx_fine(&b
, srcs
[0]); break;
1962 case ir_unop_dFdy_fine
: result
= nir_fddy_fine(&b
, srcs
[0]); break;
1963 case ir_unop_dFdx_coarse
: result
= nir_fddx_coarse(&b
, srcs
[0]); break;
1964 case ir_unop_dFdy_coarse
: result
= nir_fddy_coarse(&b
, srcs
[0]); break;
1965 case ir_unop_pack_snorm_2x16
:
1966 result
= nir_pack_snorm_2x16(&b
, srcs
[0]);
1968 case ir_unop_pack_snorm_4x8
:
1969 result
= nir_pack_snorm_4x8(&b
, srcs
[0]);
1971 case ir_unop_pack_unorm_2x16
:
1972 result
= nir_pack_unorm_2x16(&b
, srcs
[0]);
1974 case ir_unop_pack_unorm_4x8
:
1975 result
= nir_pack_unorm_4x8(&b
, srcs
[0]);
1977 case ir_unop_pack_half_2x16
:
1978 result
= nir_pack_half_2x16(&b
, srcs
[0]);
1980 case ir_unop_unpack_snorm_2x16
:
1981 result
= nir_unpack_snorm_2x16(&b
, srcs
[0]);
1983 case ir_unop_unpack_snorm_4x8
:
1984 result
= nir_unpack_snorm_4x8(&b
, srcs
[0]);
1986 case ir_unop_unpack_unorm_2x16
:
1987 result
= nir_unpack_unorm_2x16(&b
, srcs
[0]);
1989 case ir_unop_unpack_unorm_4x8
:
1990 result
= nir_unpack_unorm_4x8(&b
, srcs
[0]);
1992 case ir_unop_unpack_half_2x16
:
1993 result
= nir_unpack_half_2x16(&b
, srcs
[0]);
1995 case ir_unop_pack_sampler_2x32
:
1996 case ir_unop_pack_image_2x32
:
1997 case ir_unop_pack_double_2x32
:
1998 case ir_unop_pack_int_2x32
:
1999 case ir_unop_pack_uint_2x32
:
2000 result
= nir_pack_64_2x32(&b
, srcs
[0]);
2002 case ir_unop_unpack_sampler_2x32
:
2003 case ir_unop_unpack_image_2x32
:
2004 case ir_unop_unpack_double_2x32
:
2005 case ir_unop_unpack_int_2x32
:
2006 case ir_unop_unpack_uint_2x32
:
2007 result
= nir_unpack_64_2x32(&b
, srcs
[0]);
2009 case ir_unop_bitfield_reverse
:
2010 result
= nir_bitfield_reverse(&b
, srcs
[0]);
2012 case ir_unop_bit_count
:
2013 result
= nir_bit_count(&b
, srcs
[0]);
2015 case ir_unop_find_msb
:
2017 case GLSL_TYPE_UINT
:
2018 result
= nir_ufind_msb(&b
, srcs
[0]);
2021 result
= nir_ifind_msb(&b
, srcs
[0]);
2024 unreachable("Invalid type for findMSB()");
2027 case ir_unop_find_lsb
:
2028 result
= nir_find_lsb(&b
, srcs
[0]);
2032 switch (ir
->type
->vector_elements
) {
2034 switch (ir
->operands
[0]->type
->vector_elements
) {
2035 case 1: result
= nir_fnoise1_1(&b
, srcs
[0]); break;
2036 case 2: result
= nir_fnoise1_2(&b
, srcs
[0]); break;
2037 case 3: result
= nir_fnoise1_3(&b
, srcs
[0]); break;
2038 case 4: result
= nir_fnoise1_4(&b
, srcs
[0]); break;
2039 default: unreachable("not reached");
2043 switch (ir
->operands
[0]->type
->vector_elements
) {
2044 case 1: result
= nir_fnoise2_1(&b
, srcs
[0]); break;
2045 case 2: result
= nir_fnoise2_2(&b
, srcs
[0]); break;
2046 case 3: result
= nir_fnoise2_3(&b
, srcs
[0]); break;
2047 case 4: result
= nir_fnoise2_4(&b
, srcs
[0]); break;
2048 default: unreachable("not reached");
2052 switch (ir
->operands
[0]->type
->vector_elements
) {
2053 case 1: result
= nir_fnoise3_1(&b
, srcs
[0]); break;
2054 case 2: result
= nir_fnoise3_2(&b
, srcs
[0]); break;
2055 case 3: result
= nir_fnoise3_3(&b
, srcs
[0]); break;
2056 case 4: result
= nir_fnoise3_4(&b
, srcs
[0]); break;
2057 default: unreachable("not reached");
2061 switch (ir
->operands
[0]->type
->vector_elements
) {
2062 case 1: result
= nir_fnoise4_1(&b
, srcs
[0]); break;
2063 case 2: result
= nir_fnoise4_2(&b
, srcs
[0]); break;
2064 case 3: result
= nir_fnoise4_3(&b
, srcs
[0]); break;
2065 case 4: result
= nir_fnoise4_4(&b
, srcs
[0]); break;
2066 default: unreachable("not reached");
2070 unreachable("not reached");
2073 case ir_unop_get_buffer_size
: {
2074 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(
2076 nir_intrinsic_get_buffer_size
);
2077 load
->num_components
= ir
->type
->vector_elements
;
2078 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
2079 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2080 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
2085 result
= type_is_float(out_type
) ? nir_fadd(&b
, srcs
[0], srcs
[1])
2086 : nir_iadd(&b
, srcs
[0], srcs
[1]);
2089 result
= type_is_float(out_type
) ? nir_fsub(&b
, srcs
[0], srcs
[1])
2090 : nir_isub(&b
, srcs
[0], srcs
[1]);
2093 if (type_is_float(out_type
))
2094 result
= nir_fmul(&b
, srcs
[0], srcs
[1]);
2095 else if (out_type
== GLSL_TYPE_INT64
&&
2096 (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
||
2097 ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
))
2098 result
= nir_imul_2x32_64(&b
, srcs
[0], srcs
[1]);
2099 else if (out_type
== GLSL_TYPE_UINT64
&&
2100 (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
||
2101 ir
->operands
[1]->type
->base_type
== GLSL_TYPE_UINT
))
2102 result
= nir_umul_2x32_64(&b
, srcs
[0], srcs
[1]);
2104 result
= nir_imul(&b
, srcs
[0], srcs
[1]);
2107 if (type_is_float(out_type
))
2108 result
= nir_fdiv(&b
, srcs
[0], srcs
[1]);
2109 else if (type_is_signed(out_type
))
2110 result
= nir_idiv(&b
, srcs
[0], srcs
[1]);
2112 result
= nir_udiv(&b
, srcs
[0], srcs
[1]);
2115 result
= type_is_float(out_type
) ? nir_fmod(&b
, srcs
[0], srcs
[1])
2116 : nir_umod(&b
, srcs
[0], srcs
[1]);
2119 if (type_is_float(out_type
))
2120 result
= nir_fmin(&b
, srcs
[0], srcs
[1]);
2121 else if (type_is_signed(out_type
))
2122 result
= nir_imin(&b
, srcs
[0], srcs
[1]);
2124 result
= nir_umin(&b
, srcs
[0], srcs
[1]);
2127 if (type_is_float(out_type
))
2128 result
= nir_fmax(&b
, srcs
[0], srcs
[1]);
2129 else if (type_is_signed(out_type
))
2130 result
= nir_imax(&b
, srcs
[0], srcs
[1]);
2132 result
= nir_umax(&b
, srcs
[0], srcs
[1]);
2134 case ir_binop_pow
: result
= nir_fpow(&b
, srcs
[0], srcs
[1]); break;
2135 case ir_binop_bit_and
: result
= nir_iand(&b
, srcs
[0], srcs
[1]); break;
2136 case ir_binop_bit_or
: result
= nir_ior(&b
, srcs
[0], srcs
[1]); break;
2137 case ir_binop_bit_xor
: result
= nir_ixor(&b
, srcs
[0], srcs
[1]); break;
2138 case ir_binop_logic_and
:
2139 result
= nir_iand(&b
, srcs
[0], srcs
[1]);
2141 case ir_binop_logic_or
:
2142 result
= nir_ior(&b
, srcs
[0], srcs
[1]);
2144 case ir_binop_logic_xor
:
2145 result
= nir_ixor(&b
, srcs
[0], srcs
[1]);
2147 case ir_binop_lshift
: result
= nir_ishl(&b
, srcs
[0], srcs
[1]); break;
2148 case ir_binop_rshift
:
2149 result
= (type_is_signed(out_type
)) ? nir_ishr(&b
, srcs
[0], srcs
[1])
2150 : nir_ushr(&b
, srcs
[0], srcs
[1]);
2152 case ir_binop_imul_high
:
2153 result
= (out_type
== GLSL_TYPE_INT
) ? nir_imul_high(&b
, srcs
[0], srcs
[1])
2154 : nir_umul_high(&b
, srcs
[0], srcs
[1]);
2156 case ir_binop_carry
: result
= nir_uadd_carry(&b
, srcs
[0], srcs
[1]); break;
2157 case ir_binop_borrow
: result
= nir_usub_borrow(&b
, srcs
[0], srcs
[1]); break;
2159 if (type_is_float(types
[0]))
2160 result
= nir_flt(&b
, srcs
[0], srcs
[1]);
2161 else if (type_is_signed(types
[0]))
2162 result
= nir_ilt(&b
, srcs
[0], srcs
[1]);
2164 result
= nir_ult(&b
, srcs
[0], srcs
[1]);
2166 case ir_binop_gequal
:
2167 if (type_is_float(types
[0]))
2168 result
= nir_fge(&b
, srcs
[0], srcs
[1]);
2169 else if (type_is_signed(types
[0]))
2170 result
= nir_ige(&b
, srcs
[0], srcs
[1]);
2172 result
= nir_uge(&b
, srcs
[0], srcs
[1]);
2174 case ir_binop_equal
:
2175 if (type_is_float(types
[0]))
2176 result
= nir_feq(&b
, srcs
[0], srcs
[1]);
2178 result
= nir_ieq(&b
, srcs
[0], srcs
[1]);
2180 case ir_binop_nequal
:
2181 if (type_is_float(types
[0]))
2182 result
= nir_fne(&b
, srcs
[0], srcs
[1]);
2184 result
= nir_ine(&b
, srcs
[0], srcs
[1]);
2186 case ir_binop_all_equal
:
2187 if (type_is_float(types
[0])) {
2188 switch (ir
->operands
[0]->type
->vector_elements
) {
2189 case 1: result
= nir_feq(&b
, srcs
[0], srcs
[1]); break;
2190 case 2: result
= nir_ball_fequal2(&b
, srcs
[0], srcs
[1]); break;
2191 case 3: result
= nir_ball_fequal3(&b
, srcs
[0], srcs
[1]); break;
2192 case 4: result
= nir_ball_fequal4(&b
, srcs
[0], srcs
[1]); break;
2194 unreachable("not reached");
2197 switch (ir
->operands
[0]->type
->vector_elements
) {
2198 case 1: result
= nir_ieq(&b
, srcs
[0], srcs
[1]); break;
2199 case 2: result
= nir_ball_iequal2(&b
, srcs
[0], srcs
[1]); break;
2200 case 3: result
= nir_ball_iequal3(&b
, srcs
[0], srcs
[1]); break;
2201 case 4: result
= nir_ball_iequal4(&b
, srcs
[0], srcs
[1]); break;
2203 unreachable("not reached");
2207 case ir_binop_any_nequal
:
2208 if (type_is_float(types
[0])) {
2209 switch (ir
->operands
[0]->type
->vector_elements
) {
2210 case 1: result
= nir_fne(&b
, srcs
[0], srcs
[1]); break;
2211 case 2: result
= nir_bany_fnequal2(&b
, srcs
[0], srcs
[1]); break;
2212 case 3: result
= nir_bany_fnequal3(&b
, srcs
[0], srcs
[1]); break;
2213 case 4: result
= nir_bany_fnequal4(&b
, srcs
[0], srcs
[1]); break;
2215 unreachable("not reached");
2218 switch (ir
->operands
[0]->type
->vector_elements
) {
2219 case 1: result
= nir_ine(&b
, srcs
[0], srcs
[1]); break;
2220 case 2: result
= nir_bany_inequal2(&b
, srcs
[0], srcs
[1]); break;
2221 case 3: result
= nir_bany_inequal3(&b
, srcs
[0], srcs
[1]); break;
2222 case 4: result
= nir_bany_inequal4(&b
, srcs
[0], srcs
[1]); break;
2224 unreachable("not reached");
2229 switch (ir
->operands
[0]->type
->vector_elements
) {
2230 case 2: result
= nir_fdot2(&b
, srcs
[0], srcs
[1]); break;
2231 case 3: result
= nir_fdot3(&b
, srcs
[0], srcs
[1]); break;
2232 case 4: result
= nir_fdot4(&b
, srcs
[0], srcs
[1]); break;
2234 unreachable("not reached");
2237 case ir_binop_vector_extract
: {
2238 result
= nir_channel(&b
, srcs
[0], 0);
2239 for (unsigned i
= 1; i
< ir
->operands
[0]->type
->vector_elements
; i
++) {
2240 nir_ssa_def
*swizzled
= nir_channel(&b
, srcs
[0], i
);
2241 result
= nir_bcsel(&b
, nir_ieq(&b
, srcs
[1], nir_imm_int(&b
, i
)),
2247 case ir_binop_ldexp
: result
= nir_ldexp(&b
, srcs
[0], srcs
[1]); break;
2249 result
= nir_ffma(&b
, srcs
[0], srcs
[1], srcs
[2]);
2252 result
= nir_flrp(&b
, srcs
[0], srcs
[1], srcs
[2]);
2255 result
= nir_bcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
2257 case ir_triop_bitfield_extract
:
2258 result
= (out_type
== GLSL_TYPE_INT
) ?
2259 nir_ibitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]) :
2260 nir_ubitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]);
2262 case ir_quadop_bitfield_insert
:
2263 result
= nir_bitfield_insert(&b
, srcs
[0], srcs
[1], srcs
[2], srcs
[3]);
2265 case ir_quadop_vector
:
2266 result
= nir_vec(&b
, srcs
, ir
->type
->vector_elements
);
2270 unreachable("not reached");
2275 nir_visitor::visit(ir_swizzle
*ir
)
2277 unsigned swizzle
[4] = { ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
};
2278 result
= nir_swizzle(&b
, evaluate_rvalue(ir
->val
), swizzle
,
2279 ir
->type
->vector_elements
, false);
2283 nir_visitor::visit(ir_texture
*ir
)
2290 num_srcs
= 1; /* coordinate */
2295 op
= (ir
->op
== ir_txb
) ? nir_texop_txb
: nir_texop_txl
;
2296 num_srcs
= 2; /* coordinate, bias/lod */
2300 op
= nir_texop_txd
; /* coordinate, dPdx, dPdy */
2306 if (ir
->lod_info
.lod
!= NULL
)
2307 num_srcs
= 2; /* coordinate, lod */
2309 num_srcs
= 1; /* coordinate */
2313 op
= nir_texop_txf_ms
;
2314 num_srcs
= 2; /* coordinate, sample_index */
2319 if (ir
->lod_info
.lod
!= NULL
)
2320 num_srcs
= 1; /* lod */
2327 num_srcs
= 1; /* coordinate */
2332 num_srcs
= 1; /* coordinate */
2335 case ir_query_levels
:
2336 op
= nir_texop_query_levels
;
2340 case ir_texture_samples
:
2341 op
= nir_texop_texture_samples
;
2345 case ir_samples_identical
:
2346 op
= nir_texop_samples_identical
;
2347 num_srcs
= 1; /* coordinate */
2351 unreachable("not reached");
2354 if (ir
->projector
!= NULL
)
2356 if (ir
->shadow_comparator
!= NULL
)
2358 /* offsets are constants we store inside nir_tex_intrs.offsets */
2359 if (ir
->offset
!= NULL
&& !ir
->offset
->type
->is_array())
2362 /* Add one for the texture deref */
2365 nir_tex_instr
*instr
= nir_tex_instr_create(this->shader
, num_srcs
);
2368 instr
->sampler_dim
=
2369 (glsl_sampler_dim
) ir
->sampler
->type
->sampler_dimensionality
;
2370 instr
->is_array
= ir
->sampler
->type
->sampler_array
;
2371 instr
->is_shadow
= ir
->sampler
->type
->sampler_shadow
;
2372 if (instr
->is_shadow
)
2373 instr
->is_new_style_shadow
= (ir
->type
->vector_elements
== 1);
2374 switch (ir
->type
->base_type
) {
2375 case GLSL_TYPE_FLOAT
:
2376 instr
->dest_type
= nir_type_float
;
2379 instr
->dest_type
= nir_type_int
;
2381 case GLSL_TYPE_BOOL
:
2382 case GLSL_TYPE_UINT
:
2383 instr
->dest_type
= nir_type_uint
;
2386 unreachable("not reached");
2389 nir_deref_instr
*sampler_deref
= evaluate_deref(ir
->sampler
);
2391 /* check for bindless handles */
2392 if (sampler_deref
->mode
!= nir_var_uniform
||
2393 nir_deref_instr_get_variable(sampler_deref
)->data
.bindless
) {
2394 nir_ssa_def
*load
= nir_load_deref(&b
, sampler_deref
);
2395 instr
->src
[0].src
= nir_src_for_ssa(load
);
2396 instr
->src
[0].src_type
= nir_tex_src_texture_handle
;
2397 instr
->src
[1].src
= nir_src_for_ssa(load
);
2398 instr
->src
[1].src_type
= nir_tex_src_sampler_handle
;
2400 instr
->src
[0].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2401 instr
->src
[0].src_type
= nir_tex_src_texture_deref
;
2402 instr
->src
[1].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2403 instr
->src
[1].src_type
= nir_tex_src_sampler_deref
;
2406 unsigned src_number
= 2;
2408 if (ir
->coordinate
!= NULL
) {
2409 instr
->coord_components
= ir
->coordinate
->type
->vector_elements
;
2410 instr
->src
[src_number
].src
=
2411 nir_src_for_ssa(evaluate_rvalue(ir
->coordinate
));
2412 instr
->src
[src_number
].src_type
= nir_tex_src_coord
;
2416 if (ir
->projector
!= NULL
) {
2417 instr
->src
[src_number
].src
=
2418 nir_src_for_ssa(evaluate_rvalue(ir
->projector
));
2419 instr
->src
[src_number
].src_type
= nir_tex_src_projector
;
2423 if (ir
->shadow_comparator
!= NULL
) {
2424 instr
->src
[src_number
].src
=
2425 nir_src_for_ssa(evaluate_rvalue(ir
->shadow_comparator
));
2426 instr
->src
[src_number
].src_type
= nir_tex_src_comparator
;
2430 if (ir
->offset
!= NULL
) {
2431 if (ir
->offset
->type
->is_array()) {
2432 for (int i
= 0; i
< ir
->offset
->type
->array_size(); i
++) {
2433 const ir_constant
*c
=
2434 ir
->offset
->as_constant()->get_array_element(i
);
2436 for (unsigned j
= 0; j
< 2; ++j
) {
2437 int val
= c
->get_int_component(j
);
2438 assert(val
<= 31 && val
>= -32);
2439 instr
->tg4_offsets
[i
][j
] = val
;
2443 assert(ir
->offset
->type
->is_vector() || ir
->offset
->type
->is_scalar());
2445 instr
->src
[src_number
].src
=
2446 nir_src_for_ssa(evaluate_rvalue(ir
->offset
));
2447 instr
->src
[src_number
].src_type
= nir_tex_src_offset
;
2454 instr
->src
[src_number
].src
=
2455 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.bias
));
2456 instr
->src
[src_number
].src_type
= nir_tex_src_bias
;
2463 if (ir
->lod_info
.lod
!= NULL
) {
2464 instr
->src
[src_number
].src
=
2465 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.lod
));
2466 instr
->src
[src_number
].src_type
= nir_tex_src_lod
;
2472 instr
->src
[src_number
].src
=
2473 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdx
));
2474 instr
->src
[src_number
].src_type
= nir_tex_src_ddx
;
2476 instr
->src
[src_number
].src
=
2477 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdy
));
2478 instr
->src
[src_number
].src_type
= nir_tex_src_ddy
;
2483 instr
->src
[src_number
].src
=
2484 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.sample_index
));
2485 instr
->src
[src_number
].src_type
= nir_tex_src_ms_index
;
2490 instr
->component
= ir
->lod_info
.component
->as_constant()->value
.u
[0];
2497 assert(src_number
== num_srcs
);
2499 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2500 add_instr(&instr
->instr
, nir_tex_instr_dest_size(instr
), bit_size
);
2504 nir_visitor::visit(ir_constant
*ir
)
2507 * We don't know if this variable is an array or struct that gets
2508 * dereferenced, so do the safe thing an make it a variable with a
2509 * constant initializer and return a dereference.
2513 nir_local_variable_create(this->impl
, ir
->type
, "const_temp");
2514 var
->data
.read_only
= true;
2515 var
->constant_initializer
= constant_copy(ir
, var
);
2517 this->deref
= nir_build_deref_var(&b
, var
);
2521 nir_visitor::visit(ir_dereference_variable
*ir
)
2523 if (ir
->variable_referenced()->data
.mode
== ir_var_function_out
) {
2524 unsigned i
= (sig
->return_type
!= glsl_type::void_type
) ? 1 : 0;
2526 foreach_in_list(ir_variable
, param
, &sig
->parameters
) {
2527 if (param
== ir
->variable_referenced()) {
2533 this->deref
= nir_build_deref_cast(&b
, nir_load_param(&b
, i
),
2534 nir_var_function_temp
, ir
->type
, 0);
2538 assert(ir
->variable_referenced()->data
.mode
!= ir_var_function_inout
);
2540 struct hash_entry
*entry
=
2541 _mesa_hash_table_search(this->var_table
, ir
->var
);
2543 nir_variable
*var
= (nir_variable
*) entry
->data
;
2545 this->deref
= nir_build_deref_var(&b
, var
);
2549 nir_visitor::visit(ir_dereference_record
*ir
)
2551 ir
->record
->accept(this);
2553 int field_index
= ir
->field_idx
;
2554 assert(field_index
>= 0);
2556 this->deref
= nir_build_deref_struct(&b
, this->deref
, field_index
);
2560 nir_visitor::visit(ir_dereference_array
*ir
)
2562 nir_ssa_def
*index
= evaluate_rvalue(ir
->array_index
);
2564 ir
->array
->accept(this);
2566 this->deref
= nir_build_deref_array(&b
, this->deref
, index
);
2570 nir_visitor::visit(ir_barrier
*)
2572 nir_intrinsic_instr
*instr
=
2573 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_barrier
);
2574 nir_builder_instr_insert(&b
, &instr
->instr
);
2578 glsl_float64_funcs_to_nir(struct gl_context
*ctx
,
2579 const nir_shader_compiler_options
*options
)
2581 /* We pretend it's a vertex shader. Ultimately, the stage shouldn't
2582 * matter because we're not optimizing anything here.
2584 struct gl_shader
*sh
= _mesa_new_shader(-1, MESA_SHADER_VERTEX
);
2585 sh
->Source
= float64_source
;
2586 sh
->CompileStatus
= COMPILE_FAILURE
;
2587 _mesa_glsl_compile_shader(ctx
, sh
, false, false, true);
2589 if (!sh
->CompileStatus
) {
2592 "fp64 software impl compile failed:\n%s\nsource:\n%s\n",
2593 sh
->InfoLog
, float64_source
);
2598 nir_shader
*nir
= nir_shader_create(NULL
, MESA_SHADER_VERTEX
, options
, NULL
);
2600 nir_visitor
v1(ctx
, nir
);
2601 nir_function_visitor
v2(&v1
);
2603 visit_exec_list(sh
->ir
, &v1
);
2605 /* _mesa_delete_shader will try to free sh->Source but it's static const */
2607 _mesa_delete_shader(ctx
, sh
);
2609 nir_validate_shader(nir
, "float64_funcs_to_nir");
2611 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_function_temp
);
2612 NIR_PASS_V(nir
, nir_lower_returns
);
2613 NIR_PASS_V(nir
, nir_inline_functions
);
2614 NIR_PASS_V(nir
, nir_opt_deref
);