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 "compiler/nir/nir_builtin_builder.h"
38 #include "compiler/nir/nir_deref.h"
39 #include "main/errors.h"
40 #include "main/imports.h"
41 #include "main/mtypes.h"
42 #include "main/shaderobj.h"
43 #include "util/u_math.h"
46 * pass to lower GLSL IR to NIR
48 * This will lower variable dereferences to loads/stores of corresponding
49 * variables in NIR - the variables will be converted to registers in a later
55 class nir_visitor
: public ir_visitor
58 nir_visitor(gl_context
*ctx
, nir_shader
*shader
);
61 virtual void visit(ir_variable
*);
62 virtual void visit(ir_function
*);
63 virtual void visit(ir_function_signature
*);
64 virtual void visit(ir_loop
*);
65 virtual void visit(ir_if
*);
66 virtual void visit(ir_discard
*);
67 virtual void visit(ir_demote
*);
68 virtual void visit(ir_loop_jump
*);
69 virtual void visit(ir_return
*);
70 virtual void visit(ir_call
*);
71 virtual void visit(ir_assignment
*);
72 virtual void visit(ir_emit_vertex
*);
73 virtual void visit(ir_end_primitive
*);
74 virtual void visit(ir_expression
*);
75 virtual void visit(ir_swizzle
*);
76 virtual void visit(ir_texture
*);
77 virtual void visit(ir_constant
*);
78 virtual void visit(ir_dereference_variable
*);
79 virtual void visit(ir_dereference_record
*);
80 virtual void visit(ir_dereference_array
*);
81 virtual void visit(ir_barrier
*);
83 void create_function(ir_function_signature
*ir
);
86 void add_instr(nir_instr
*instr
, unsigned num_components
, unsigned bit_size
);
87 nir_ssa_def
*evaluate_rvalue(ir_rvalue
*ir
);
89 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
**srcs
);
90 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
);
91 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
,
93 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
,
94 nir_ssa_def
*src2
, nir_ssa_def
*src3
);
99 nir_function_impl
*impl
;
101 nir_ssa_def
*result
; /* result of the expression tree last visited */
103 nir_deref_instr
*evaluate_deref(ir_instruction
*ir
);
105 nir_constant
*constant_copy(ir_constant
*ir
, void *mem_ctx
);
107 /* most recent deref instruction created */
108 nir_deref_instr
*deref
;
110 /* whether the IR we're operating on is per-function or global */
113 ir_function_signature
*sig
;
115 /* map of ir_variable -> nir_variable */
116 struct hash_table
*var_table
;
118 /* map of ir_function_signature -> nir_function_overload */
119 struct hash_table
*overload_table
;
123 * This visitor runs before the main visitor, calling create_function() for
124 * each function so that the main visitor can resolve forward references in
128 class nir_function_visitor
: public ir_hierarchical_visitor
131 nir_function_visitor(nir_visitor
*v
) : visitor(v
)
134 virtual ir_visitor_status
visit_enter(ir_function
*);
137 nir_visitor
*visitor
;
140 /* glsl_to_nir can only handle converting certain function paramaters
141 * to NIR. This visitor checks for parameters it can't currently handle.
143 class ir_function_param_visitor
: public ir_hierarchical_visitor
146 ir_function_param_visitor()
151 virtual ir_visitor_status
visit_enter(ir_function_signature
*ir
)
154 if (ir
->is_intrinsic())
155 return visit_continue
;
157 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
158 if (!param
->type
->is_vector() || !param
->type
->is_scalar()) {
163 if (param
->data
.mode
== ir_var_function_inout
) {
169 return visit_continue
;
175 } /* end of anonymous namespace */
179 has_unsupported_function_param(exec_list
*ir
)
181 ir_function_param_visitor visitor
;
182 visit_list_elements(&visitor
, ir
);
183 return visitor
.unsupported
;
187 glsl_to_nir(struct gl_context
*ctx
,
188 const struct gl_shader_program
*shader_prog
,
189 gl_shader_stage stage
,
190 const nir_shader_compiler_options
*options
)
192 struct gl_linked_shader
*sh
= shader_prog
->_LinkedShaders
[stage
];
194 const struct gl_shader_compiler_options
*gl_options
=
195 &ctx
->Const
.ShaderCompilerOptions
[stage
];
197 /* glsl_to_nir can only handle converting certain function paramaters
198 * to NIR. If we find something we can't handle then we get the GLSL IR
199 * opts to remove it before we continue on.
201 * TODO: add missing glsl ir to nir support and remove this loop.
203 while (has_unsupported_function_param(sh
->ir
)) {
204 do_common_optimization(sh
->ir
, true, true, gl_options
,
205 ctx
->Const
.NativeIntegers
);
208 nir_shader
*shader
= nir_shader_create(NULL
, stage
, options
,
211 nir_visitor
v1(ctx
, shader
);
212 nir_function_visitor
v2(&v1
);
214 visit_exec_list(sh
->ir
, &v1
);
216 nir_validate_shader(shader
, "after glsl to nir, before function inline");
218 /* We have to lower away local constant initializers right before we
219 * inline functions. That way they get properly initialized at the top
220 * of the function and not at the top of its caller.
222 nir_lower_constant_initializers(shader
, (nir_variable_mode
)~0);
223 nir_lower_returns(shader
);
224 nir_inline_functions(shader
);
225 nir_opt_deref(shader
);
227 nir_validate_shader(shader
, "after function inlining and return lowering");
229 /* Now that we have inlined everything remove all of the functions except
232 foreach_list_typed_safe(nir_function
, function
, node
, &(shader
)->functions
){
233 if (strcmp("main", function
->name
) != 0) {
234 exec_node_remove(&function
->node
);
238 shader
->info
.name
= ralloc_asprintf(shader
, "GLSL%d", shader_prog
->Name
);
239 if (shader_prog
->Label
)
240 shader
->info
.label
= ralloc_strdup(shader
, shader_prog
->Label
);
242 /* Check for transform feedback varyings specified via the API */
243 shader
->info
.has_transform_feedback_varyings
=
244 shader_prog
->TransformFeedback
.NumVarying
> 0;
246 /* Check for transform feedback varyings specified in the Shader */
247 if (shader_prog
->last_vert_prog
)
248 shader
->info
.has_transform_feedback_varyings
|=
249 shader_prog
->last_vert_prog
->sh
.LinkedTransformFeedback
->NumVarying
> 0;
251 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
) {
252 shader
->info
.fs
.pixel_center_integer
= sh
->Program
->info
.fs
.pixel_center_integer
;
253 shader
->info
.fs
.origin_upper_left
= sh
->Program
->info
.fs
.origin_upper_left
;
259 nir_visitor::nir_visitor(gl_context
*ctx
, nir_shader
*shader
)
261 this->supports_std430
= ctx
->Const
.UseSTD430AsDefaultPacking
;
262 this->shader
= shader
;
263 this->is_global
= true;
264 this->var_table
= _mesa_pointer_hash_table_create(NULL
);
265 this->overload_table
= _mesa_pointer_hash_table_create(NULL
);
270 memset(&this->b
, 0, sizeof(this->b
));
273 nir_visitor::~nir_visitor()
275 _mesa_hash_table_destroy(this->var_table
, NULL
);
276 _mesa_hash_table_destroy(this->overload_table
, NULL
);
280 nir_visitor::evaluate_deref(ir_instruction
*ir
)
287 nir_visitor::constant_copy(ir_constant
*ir
, void *mem_ctx
)
292 nir_constant
*ret
= rzalloc(mem_ctx
, nir_constant
);
294 const unsigned rows
= ir
->type
->vector_elements
;
295 const unsigned cols
= ir
->type
->matrix_columns
;
298 ret
->num_elements
= 0;
299 switch (ir
->type
->base_type
) {
301 /* Only float base types can be matrices. */
304 for (unsigned r
= 0; r
< rows
; r
++)
305 ret
->values
[r
].u32
= ir
->value
.u
[r
];
310 /* Only float base types can be matrices. */
313 for (unsigned r
= 0; r
< rows
; r
++)
314 ret
->values
[r
].i32
= ir
->value
.i
[r
];
318 case GLSL_TYPE_FLOAT
:
319 case GLSL_TYPE_DOUBLE
:
321 ret
->elements
= ralloc_array(mem_ctx
, nir_constant
*, cols
);
322 ret
->num_elements
= cols
;
323 for (unsigned c
= 0; c
< cols
; c
++) {
324 nir_constant
*col_const
= rzalloc(mem_ctx
, nir_constant
);
325 col_const
->num_elements
= 0;
326 switch (ir
->type
->base_type
) {
327 case GLSL_TYPE_FLOAT
:
328 for (unsigned r
= 0; r
< rows
; r
++)
329 col_const
->values
[r
].f32
= ir
->value
.f
[c
* rows
+ r
];
332 case GLSL_TYPE_DOUBLE
:
333 for (unsigned r
= 0; r
< rows
; r
++)
334 col_const
->values
[r
].f64
= ir
->value
.d
[c
* rows
+ r
];
338 unreachable("Cannot get here from the first level switch");
340 ret
->elements
[c
] = col_const
;
343 switch (ir
->type
->base_type
) {
344 case GLSL_TYPE_FLOAT
:
345 for (unsigned r
= 0; r
< rows
; r
++)
346 ret
->values
[r
].f32
= ir
->value
.f
[r
];
349 case GLSL_TYPE_DOUBLE
:
350 for (unsigned r
= 0; r
< rows
; r
++)
351 ret
->values
[r
].f64
= ir
->value
.d
[r
];
355 unreachable("Cannot get here from the first level switch");
360 case GLSL_TYPE_UINT64
:
361 /* Only float base types can be matrices. */
364 for (unsigned r
= 0; r
< rows
; r
++)
365 ret
->values
[r
].u64
= ir
->value
.u64
[r
];
368 case GLSL_TYPE_INT64
:
369 /* Only float base types can be matrices. */
372 for (unsigned r
= 0; r
< rows
; r
++)
373 ret
->values
[r
].i64
= ir
->value
.i64
[r
];
377 /* Only float base types can be matrices. */
380 for (unsigned r
= 0; r
< rows
; r
++)
381 ret
->values
[r
].b
= ir
->value
.b
[r
];
385 case GLSL_TYPE_STRUCT
:
386 case GLSL_TYPE_ARRAY
:
387 ret
->elements
= ralloc_array(mem_ctx
, nir_constant
*,
389 ret
->num_elements
= ir
->type
->length
;
391 for (i
= 0; i
< ir
->type
->length
; i
++)
392 ret
->elements
[i
] = constant_copy(ir
->const_elements
[i
], mem_ctx
);
396 unreachable("not reached");
402 static const glsl_type
*
403 wrap_type_in_array(const glsl_type
*elem_type
, const glsl_type
*array_type
)
405 if (!array_type
->is_array())
408 elem_type
= wrap_type_in_array(elem_type
, array_type
->fields
.array
);
410 return glsl_type::get_array_instance(elem_type
, array_type
->length
);
414 get_nir_how_declared(unsigned how_declared
)
416 if (how_declared
== ir_var_hidden
)
417 return nir_var_hidden
;
419 return nir_var_declared_normally
;
423 nir_visitor::visit(ir_variable
*ir
)
425 /* TODO: In future we should switch to using the NIR lowering pass but for
426 * now just ignore these variables as GLSL IR should have lowered them.
427 * Anything remaining are just dead vars that weren't cleaned up.
429 if (ir
->data
.mode
== ir_var_shader_shared
)
432 /* FINISHME: inout parameters */
433 assert(ir
->data
.mode
!= ir_var_function_inout
);
435 if (ir
->data
.mode
== ir_var_function_out
)
438 nir_variable
*var
= rzalloc(shader
, nir_variable
);
439 var
->type
= ir
->type
;
440 var
->name
= ralloc_strdup(var
, ir
->name
);
442 var
->data
.always_active_io
= ir
->data
.always_active_io
;
443 var
->data
.read_only
= ir
->data
.read_only
;
444 var
->data
.centroid
= ir
->data
.centroid
;
445 var
->data
.sample
= ir
->data
.sample
;
446 var
->data
.patch
= ir
->data
.patch
;
447 var
->data
.how_declared
= get_nir_how_declared(ir
->data
.how_declared
);
448 var
->data
.invariant
= ir
->data
.invariant
;
449 var
->data
.location
= ir
->data
.location
;
450 var
->data
.stream
= ir
->data
.stream
;
451 if (ir
->data
.stream
& (1u << 31))
452 var
->data
.stream
|= NIR_STREAM_PACKED
;
454 var
->data
.precision
= ir
->data
.precision
;
455 var
->data
.explicit_location
= ir
->data
.explicit_location
;
456 var
->data
.from_named_ifc_block
= ir
->data
.from_named_ifc_block
;
457 var
->data
.compact
= false;
459 switch(ir
->data
.mode
) {
461 case ir_var_temporary
:
463 var
->data
.mode
= nir_var_shader_temp
;
465 var
->data
.mode
= nir_var_function_temp
;
468 case ir_var_function_in
:
469 case ir_var_const_in
:
470 var
->data
.mode
= nir_var_function_temp
;
473 case ir_var_shader_in
:
474 if (shader
->info
.stage
== MESA_SHADER_GEOMETRY
&&
475 ir
->data
.location
== VARYING_SLOT_PRIMITIVE_ID
) {
476 /* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */
477 var
->data
.location
= SYSTEM_VALUE_PRIMITIVE_ID
;
478 var
->data
.mode
= nir_var_system_value
;
480 var
->data
.mode
= nir_var_shader_in
;
482 if (shader
->info
.stage
== MESA_SHADER_TESS_EVAL
&&
483 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
484 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
485 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
488 if (shader
->info
.stage
> MESA_SHADER_VERTEX
&&
489 ir
->data
.location
>= VARYING_SLOT_CLIP_DIST0
&&
490 ir
->data
.location
<= VARYING_SLOT_CULL_DIST1
) {
491 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
496 case ir_var_shader_out
:
497 var
->data
.mode
= nir_var_shader_out
;
498 if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
&&
499 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
500 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
501 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
504 if (shader
->info
.stage
<= MESA_SHADER_GEOMETRY
&&
505 ir
->data
.location
>= VARYING_SLOT_CLIP_DIST0
&&
506 ir
->data
.location
<= VARYING_SLOT_CULL_DIST1
) {
507 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
512 if (ir
->get_interface_type())
513 var
->data
.mode
= nir_var_mem_ubo
;
515 var
->data
.mode
= nir_var_uniform
;
518 case ir_var_shader_storage
:
519 var
->data
.mode
= nir_var_mem_ssbo
;
522 case ir_var_system_value
:
523 var
->data
.mode
= nir_var_system_value
;
527 unreachable("not reached");
530 unsigned mem_access
= 0;
531 if (ir
->data
.memory_read_only
)
532 mem_access
|= ACCESS_NON_WRITEABLE
;
533 if (ir
->data
.memory_write_only
)
534 mem_access
|= ACCESS_NON_READABLE
;
535 if (ir
->data
.memory_coherent
)
536 mem_access
|= ACCESS_COHERENT
;
537 if (ir
->data
.memory_volatile
)
538 mem_access
|= ACCESS_VOLATILE
;
539 if (ir
->data
.memory_restrict
)
540 mem_access
|= ACCESS_RESTRICT
;
542 var
->interface_type
= ir
->get_interface_type();
544 /* For UBO and SSBO variables, we need explicit types */
545 if (var
->data
.mode
& (nir_var_mem_ubo
| nir_var_mem_ssbo
)) {
546 const glsl_type
*explicit_ifc_type
=
547 ir
->get_interface_type()->get_explicit_interface_type(supports_std430
);
549 var
->interface_type
= explicit_ifc_type
;
551 if (ir
->type
->without_array()->is_interface()) {
552 /* If the type contains the interface, wrap the explicit type in the
553 * right number of arrays.
555 var
->type
= wrap_type_in_array(explicit_ifc_type
, ir
->type
);
557 /* Otherwise, this variable is one entry in the interface */
558 UNUSED
bool found
= false;
559 for (unsigned i
= 0; i
< explicit_ifc_type
->length
; i
++) {
560 const glsl_struct_field
*field
=
561 &explicit_ifc_type
->fields
.structure
[i
];
562 if (strcmp(ir
->name
, field
->name
) != 0)
565 var
->type
= field
->type
;
566 if (field
->memory_read_only
)
567 mem_access
|= ACCESS_NON_WRITEABLE
;
568 if (field
->memory_write_only
)
569 mem_access
|= ACCESS_NON_READABLE
;
570 if (field
->memory_coherent
)
571 mem_access
|= ACCESS_COHERENT
;
572 if (field
->memory_volatile
)
573 mem_access
|= ACCESS_VOLATILE
;
574 if (field
->memory_restrict
)
575 mem_access
|= ACCESS_RESTRICT
;
584 var
->data
.interpolation
= ir
->data
.interpolation
;
585 var
->data
.location_frac
= ir
->data
.location_frac
;
587 switch (ir
->data
.depth_layout
) {
588 case ir_depth_layout_none
:
589 var
->data
.depth_layout
= nir_depth_layout_none
;
591 case ir_depth_layout_any
:
592 var
->data
.depth_layout
= nir_depth_layout_any
;
594 case ir_depth_layout_greater
:
595 var
->data
.depth_layout
= nir_depth_layout_greater
;
597 case ir_depth_layout_less
:
598 var
->data
.depth_layout
= nir_depth_layout_less
;
600 case ir_depth_layout_unchanged
:
601 var
->data
.depth_layout
= nir_depth_layout_unchanged
;
604 unreachable("not reached");
607 var
->data
.index
= ir
->data
.index
;
608 var
->data
.descriptor_set
= 0;
609 var
->data
.binding
= ir
->data
.binding
;
610 var
->data
.explicit_binding
= ir
->data
.explicit_binding
;
611 var
->data
.bindless
= ir
->data
.bindless
;
612 var
->data
.offset
= ir
->data
.offset
;
613 var
->data
.access
= (gl_access_qualifier
)mem_access
;
615 if (var
->type
->without_array()->is_image()) {
616 var
->data
.image
.format
= ir
->data
.image_format
;
617 } else if (var
->data
.mode
== nir_var_shader_out
) {
618 var
->data
.xfb
.buffer
= ir
->data
.xfb_buffer
;
619 var
->data
.xfb
.stride
= ir
->data
.xfb_stride
;
622 var
->data
.fb_fetch_output
= ir
->data
.fb_fetch_output
;
623 var
->data
.explicit_xfb_buffer
= ir
->data
.explicit_xfb_buffer
;
624 var
->data
.explicit_xfb_stride
= ir
->data
.explicit_xfb_stride
;
626 var
->num_state_slots
= ir
->get_num_state_slots();
627 if (var
->num_state_slots
> 0) {
628 var
->state_slots
= rzalloc_array(var
, nir_state_slot
,
629 var
->num_state_slots
);
631 ir_state_slot
*state_slots
= ir
->get_state_slots();
632 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
633 for (unsigned j
= 0; j
< 5; j
++)
634 var
->state_slots
[i
].tokens
[j
] = state_slots
[i
].tokens
[j
];
635 var
->state_slots
[i
].swizzle
= state_slots
[i
].swizzle
;
638 var
->state_slots
= NULL
;
641 var
->constant_initializer
= constant_copy(ir
->constant_initializer
, var
);
643 if (var
->data
.mode
== nir_var_function_temp
)
644 nir_function_impl_add_variable(impl
, var
);
646 nir_shader_add_variable(shader
, var
);
648 _mesa_hash_table_insert(var_table
, ir
, var
);
652 nir_function_visitor::visit_enter(ir_function
*ir
)
654 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
) {
655 visitor
->create_function(sig
);
657 return visit_continue_with_parent
;
661 nir_visitor::create_function(ir_function_signature
*ir
)
663 if (ir
->is_intrinsic())
666 nir_function
*func
= nir_function_create(shader
, ir
->function_name());
667 if (strcmp(ir
->function_name(), "main") == 0)
668 func
->is_entrypoint
= true;
670 func
->num_params
= ir
->parameters
.length() +
671 (ir
->return_type
!= glsl_type::void_type
);
672 func
->params
= ralloc_array(shader
, nir_parameter
, func
->num_params
);
676 if (ir
->return_type
!= glsl_type::void_type
) {
677 /* The return value is a variable deref (basically an out parameter) */
678 func
->params
[np
].num_components
= 1;
679 func
->params
[np
].bit_size
= 32;
683 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
684 /* FINISHME: pass arrays, structs, etc by reference? */
685 assert(param
->type
->is_vector() || param
->type
->is_scalar());
687 if (param
->data
.mode
== ir_var_function_in
) {
688 func
->params
[np
].num_components
= param
->type
->vector_elements
;
689 func
->params
[np
].bit_size
= glsl_get_bit_size(param
->type
);
691 func
->params
[np
].num_components
= 1;
692 func
->params
[np
].bit_size
= 32;
696 assert(np
== func
->num_params
);
698 _mesa_hash_table_insert(this->overload_table
, ir
, func
);
702 nir_visitor::visit(ir_function
*ir
)
704 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
)
709 nir_visitor::visit(ir_function_signature
*ir
)
711 if (ir
->is_intrinsic())
716 struct hash_entry
*entry
=
717 _mesa_hash_table_search(this->overload_table
, ir
);
720 nir_function
*func
= (nir_function
*) entry
->data
;
722 if (ir
->is_defined
) {
723 nir_function_impl
*impl
= nir_function_impl_create(func
);
726 this->is_global
= false;
728 nir_builder_init(&b
, impl
);
729 b
.cursor
= nir_after_cf_list(&impl
->body
);
731 unsigned i
= (ir
->return_type
!= glsl_type::void_type
) ? 1 : 0;
733 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
735 nir_local_variable_create(impl
, param
->type
, param
->name
);
737 if (param
->data
.mode
== ir_var_function_in
) {
738 nir_store_var(&b
, var
, nir_load_param(&b
, i
), ~0);
741 _mesa_hash_table_insert(var_table
, param
, var
);
745 visit_exec_list(&ir
->body
, this);
747 this->is_global
= true;
754 nir_visitor::visit(ir_loop
*ir
)
757 visit_exec_list(&ir
->body_instructions
, this);
758 nir_pop_loop(&b
, NULL
);
762 nir_visitor::visit(ir_if
*ir
)
764 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
765 visit_exec_list(&ir
->then_instructions
, this);
766 nir_push_else(&b
, NULL
);
767 visit_exec_list(&ir
->else_instructions
, this);
768 nir_pop_if(&b
, NULL
);
772 nir_visitor::visit(ir_discard
*ir
)
775 * discards aren't treated as control flow, because before we lower them
776 * they can appear anywhere in the shader and the stuff after them may still
777 * be executed (yay, crazy GLSL rules!). However, after lowering, all the
778 * discards will be immediately followed by a return.
781 nir_intrinsic_instr
*discard
;
783 discard
= nir_intrinsic_instr_create(this->shader
,
784 nir_intrinsic_discard_if
);
786 nir_src_for_ssa(evaluate_rvalue(ir
->condition
));
788 discard
= nir_intrinsic_instr_create(this->shader
, nir_intrinsic_discard
);
791 nir_builder_instr_insert(&b
, &discard
->instr
);
795 nir_visitor::visit(ir_demote
*ir
)
797 nir_intrinsic_instr
*demote
=
798 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_demote
);
800 nir_builder_instr_insert(&b
, &demote
->instr
);
804 nir_visitor::visit(ir_emit_vertex
*ir
)
806 nir_intrinsic_instr
*instr
=
807 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_emit_vertex
);
808 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
809 nir_builder_instr_insert(&b
, &instr
->instr
);
813 nir_visitor::visit(ir_end_primitive
*ir
)
815 nir_intrinsic_instr
*instr
=
816 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_end_primitive
);
817 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
818 nir_builder_instr_insert(&b
, &instr
->instr
);
822 nir_visitor::visit(ir_loop_jump
*ir
)
826 case ir_loop_jump::jump_break
:
827 type
= nir_jump_break
;
829 case ir_loop_jump::jump_continue
:
830 type
= nir_jump_continue
;
833 unreachable("not reached");
836 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, type
);
837 nir_builder_instr_insert(&b
, &instr
->instr
);
841 nir_visitor::visit(ir_return
*ir
)
843 if (ir
->value
!= NULL
) {
844 nir_deref_instr
*ret_deref
=
845 nir_build_deref_cast(&b
, nir_load_param(&b
, 0),
846 nir_var_function_temp
, ir
->value
->type
, 0);
848 nir_ssa_def
*val
= evaluate_rvalue(ir
->value
);
849 nir_store_deref(&b
, ret_deref
, val
, ~0);
852 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, nir_jump_return
);
853 nir_builder_instr_insert(&b
, &instr
->instr
);
857 intrinsic_set_std430_align(nir_intrinsic_instr
*intrin
, const glsl_type
*type
)
859 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
860 unsigned pow2_components
= util_next_power_of_two(type
->vector_elements
);
861 nir_intrinsic_set_align(intrin
, (bit_size
/ 8) * pow2_components
, 0);
864 /* Accumulate any qualifiers along the deref chain to get the actual
865 * load/store qualifier.
868 static enum gl_access_qualifier
869 deref_get_qualifier(nir_deref_instr
*deref
)
872 nir_deref_path_init(&path
, deref
, NULL
);
874 unsigned qualifiers
= path
.path
[0]->var
->data
.access
;
876 const glsl_type
*parent_type
= path
.path
[0]->type
;
877 for (nir_deref_instr
**cur_ptr
= &path
.path
[1]; *cur_ptr
; cur_ptr
++) {
878 nir_deref_instr
*cur
= *cur_ptr
;
880 if (parent_type
->is_interface()) {
881 const struct glsl_struct_field
*field
=
882 &parent_type
->fields
.structure
[cur
->strct
.index
];
883 if (field
->memory_read_only
)
884 qualifiers
|= ACCESS_NON_WRITEABLE
;
885 if (field
->memory_write_only
)
886 qualifiers
|= ACCESS_NON_READABLE
;
887 if (field
->memory_coherent
)
888 qualifiers
|= ACCESS_COHERENT
;
889 if (field
->memory_volatile
)
890 qualifiers
|= ACCESS_VOLATILE
;
891 if (field
->memory_restrict
)
892 qualifiers
|= ACCESS_RESTRICT
;
895 parent_type
= cur
->type
;
898 nir_deref_path_finish(&path
);
900 return (gl_access_qualifier
) qualifiers
;
904 nir_visitor::visit(ir_call
*ir
)
906 if (ir
->callee
->is_intrinsic()) {
909 switch (ir
->callee
->intrinsic_id
) {
910 case ir_intrinsic_generic_atomic_add
:
911 op
= ir
->return_deref
->type
->is_integer_32_64()
912 ? nir_intrinsic_deref_atomic_add
: nir_intrinsic_deref_atomic_fadd
;
914 case ir_intrinsic_generic_atomic_and
:
915 op
= nir_intrinsic_deref_atomic_and
;
917 case ir_intrinsic_generic_atomic_or
:
918 op
= nir_intrinsic_deref_atomic_or
;
920 case ir_intrinsic_generic_atomic_xor
:
921 op
= nir_intrinsic_deref_atomic_xor
;
923 case ir_intrinsic_generic_atomic_min
:
924 assert(ir
->return_deref
);
925 if (ir
->return_deref
->type
== glsl_type::int_type
)
926 op
= nir_intrinsic_deref_atomic_imin
;
927 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
928 op
= nir_intrinsic_deref_atomic_umin
;
929 else if (ir
->return_deref
->type
== glsl_type::float_type
)
930 op
= nir_intrinsic_deref_atomic_fmin
;
932 unreachable("Invalid type");
934 case ir_intrinsic_generic_atomic_max
:
935 assert(ir
->return_deref
);
936 if (ir
->return_deref
->type
== glsl_type::int_type
)
937 op
= nir_intrinsic_deref_atomic_imax
;
938 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
939 op
= nir_intrinsic_deref_atomic_umax
;
940 else if (ir
->return_deref
->type
== glsl_type::float_type
)
941 op
= nir_intrinsic_deref_atomic_fmax
;
943 unreachable("Invalid type");
945 case ir_intrinsic_generic_atomic_exchange
:
946 op
= nir_intrinsic_deref_atomic_exchange
;
948 case ir_intrinsic_generic_atomic_comp_swap
:
949 op
= ir
->return_deref
->type
->is_integer_32_64()
950 ? nir_intrinsic_deref_atomic_comp_swap
951 : nir_intrinsic_deref_atomic_fcomp_swap
;
953 case ir_intrinsic_atomic_counter_read
:
954 op
= nir_intrinsic_atomic_counter_read_deref
;
956 case ir_intrinsic_atomic_counter_increment
:
957 op
= nir_intrinsic_atomic_counter_inc_deref
;
959 case ir_intrinsic_atomic_counter_predecrement
:
960 op
= nir_intrinsic_atomic_counter_pre_dec_deref
;
962 case ir_intrinsic_atomic_counter_add
:
963 op
= nir_intrinsic_atomic_counter_add_deref
;
965 case ir_intrinsic_atomic_counter_and
:
966 op
= nir_intrinsic_atomic_counter_and_deref
;
968 case ir_intrinsic_atomic_counter_or
:
969 op
= nir_intrinsic_atomic_counter_or_deref
;
971 case ir_intrinsic_atomic_counter_xor
:
972 op
= nir_intrinsic_atomic_counter_xor_deref
;
974 case ir_intrinsic_atomic_counter_min
:
975 op
= nir_intrinsic_atomic_counter_min_deref
;
977 case ir_intrinsic_atomic_counter_max
:
978 op
= nir_intrinsic_atomic_counter_max_deref
;
980 case ir_intrinsic_atomic_counter_exchange
:
981 op
= nir_intrinsic_atomic_counter_exchange_deref
;
983 case ir_intrinsic_atomic_counter_comp_swap
:
984 op
= nir_intrinsic_atomic_counter_comp_swap_deref
;
986 case ir_intrinsic_image_load
:
987 op
= nir_intrinsic_image_deref_load
;
989 case ir_intrinsic_image_store
:
990 op
= nir_intrinsic_image_deref_store
;
992 case ir_intrinsic_image_atomic_add
:
993 op
= ir
->return_deref
->type
->is_integer_32_64()
994 ? nir_intrinsic_image_deref_atomic_add
995 : nir_intrinsic_image_deref_atomic_fadd
;
997 case ir_intrinsic_image_atomic_min
:
998 if (ir
->return_deref
->type
== glsl_type::int_type
)
999 op
= nir_intrinsic_image_deref_atomic_imin
;
1000 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1001 op
= nir_intrinsic_image_deref_atomic_umin
;
1003 unreachable("Invalid type");
1005 case ir_intrinsic_image_atomic_max
:
1006 if (ir
->return_deref
->type
== glsl_type::int_type
)
1007 op
= nir_intrinsic_image_deref_atomic_imax
;
1008 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1009 op
= nir_intrinsic_image_deref_atomic_umax
;
1011 unreachable("Invalid type");
1013 case ir_intrinsic_image_atomic_and
:
1014 op
= nir_intrinsic_image_deref_atomic_and
;
1016 case ir_intrinsic_image_atomic_or
:
1017 op
= nir_intrinsic_image_deref_atomic_or
;
1019 case ir_intrinsic_image_atomic_xor
:
1020 op
= nir_intrinsic_image_deref_atomic_xor
;
1022 case ir_intrinsic_image_atomic_exchange
:
1023 op
= nir_intrinsic_image_deref_atomic_exchange
;
1025 case ir_intrinsic_image_atomic_comp_swap
:
1026 op
= nir_intrinsic_image_deref_atomic_comp_swap
;
1028 case ir_intrinsic_image_atomic_inc_wrap
:
1029 op
= nir_intrinsic_image_deref_atomic_inc_wrap
;
1031 case ir_intrinsic_image_atomic_dec_wrap
:
1032 op
= nir_intrinsic_image_deref_atomic_dec_wrap
;
1034 case ir_intrinsic_memory_barrier
:
1035 op
= nir_intrinsic_memory_barrier
;
1037 case ir_intrinsic_image_size
:
1038 op
= nir_intrinsic_image_deref_size
;
1040 case ir_intrinsic_image_samples
:
1041 op
= nir_intrinsic_image_deref_samples
;
1043 case ir_intrinsic_ssbo_store
:
1044 op
= nir_intrinsic_store_ssbo
;
1046 case ir_intrinsic_ssbo_load
:
1047 op
= nir_intrinsic_load_ssbo
;
1049 case ir_intrinsic_ssbo_atomic_add
:
1050 op
= ir
->return_deref
->type
->is_integer_32_64()
1051 ? nir_intrinsic_ssbo_atomic_add
: nir_intrinsic_ssbo_atomic_fadd
;
1053 case ir_intrinsic_ssbo_atomic_and
:
1054 op
= nir_intrinsic_ssbo_atomic_and
;
1056 case ir_intrinsic_ssbo_atomic_or
:
1057 op
= nir_intrinsic_ssbo_atomic_or
;
1059 case ir_intrinsic_ssbo_atomic_xor
:
1060 op
= nir_intrinsic_ssbo_atomic_xor
;
1062 case ir_intrinsic_ssbo_atomic_min
:
1063 assert(ir
->return_deref
);
1064 if (ir
->return_deref
->type
== glsl_type::int_type
)
1065 op
= nir_intrinsic_ssbo_atomic_imin
;
1066 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1067 op
= nir_intrinsic_ssbo_atomic_umin
;
1068 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1069 op
= nir_intrinsic_ssbo_atomic_fmin
;
1071 unreachable("Invalid type");
1073 case ir_intrinsic_ssbo_atomic_max
:
1074 assert(ir
->return_deref
);
1075 if (ir
->return_deref
->type
== glsl_type::int_type
)
1076 op
= nir_intrinsic_ssbo_atomic_imax
;
1077 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1078 op
= nir_intrinsic_ssbo_atomic_umax
;
1079 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1080 op
= nir_intrinsic_ssbo_atomic_fmax
;
1082 unreachable("Invalid type");
1084 case ir_intrinsic_ssbo_atomic_exchange
:
1085 op
= nir_intrinsic_ssbo_atomic_exchange
;
1087 case ir_intrinsic_ssbo_atomic_comp_swap
:
1088 op
= ir
->return_deref
->type
->is_integer_32_64()
1089 ? nir_intrinsic_ssbo_atomic_comp_swap
1090 : nir_intrinsic_ssbo_atomic_fcomp_swap
;
1092 case ir_intrinsic_shader_clock
:
1093 op
= nir_intrinsic_shader_clock
;
1095 case ir_intrinsic_begin_invocation_interlock
:
1096 op
= nir_intrinsic_begin_invocation_interlock
;
1098 case ir_intrinsic_end_invocation_interlock
:
1099 op
= nir_intrinsic_end_invocation_interlock
;
1101 case ir_intrinsic_group_memory_barrier
:
1102 op
= nir_intrinsic_group_memory_barrier
;
1104 case ir_intrinsic_memory_barrier_atomic_counter
:
1105 op
= nir_intrinsic_memory_barrier_atomic_counter
;
1107 case ir_intrinsic_memory_barrier_buffer
:
1108 op
= nir_intrinsic_memory_barrier_buffer
;
1110 case ir_intrinsic_memory_barrier_image
:
1111 op
= nir_intrinsic_memory_barrier_image
;
1113 case ir_intrinsic_memory_barrier_shared
:
1114 op
= nir_intrinsic_memory_barrier_shared
;
1116 case ir_intrinsic_shared_load
:
1117 op
= nir_intrinsic_load_shared
;
1119 case ir_intrinsic_shared_store
:
1120 op
= nir_intrinsic_store_shared
;
1122 case ir_intrinsic_shared_atomic_add
:
1123 op
= ir
->return_deref
->type
->is_integer_32_64()
1124 ? nir_intrinsic_shared_atomic_add
1125 : nir_intrinsic_shared_atomic_fadd
;
1127 case ir_intrinsic_shared_atomic_and
:
1128 op
= nir_intrinsic_shared_atomic_and
;
1130 case ir_intrinsic_shared_atomic_or
:
1131 op
= nir_intrinsic_shared_atomic_or
;
1133 case ir_intrinsic_shared_atomic_xor
:
1134 op
= nir_intrinsic_shared_atomic_xor
;
1136 case ir_intrinsic_shared_atomic_min
:
1137 assert(ir
->return_deref
);
1138 if (ir
->return_deref
->type
== glsl_type::int_type
)
1139 op
= nir_intrinsic_shared_atomic_imin
;
1140 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1141 op
= nir_intrinsic_shared_atomic_umin
;
1142 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1143 op
= nir_intrinsic_shared_atomic_fmin
;
1145 unreachable("Invalid type");
1147 case ir_intrinsic_shared_atomic_max
:
1148 assert(ir
->return_deref
);
1149 if (ir
->return_deref
->type
== glsl_type::int_type
)
1150 op
= nir_intrinsic_shared_atomic_imax
;
1151 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1152 op
= nir_intrinsic_shared_atomic_umax
;
1153 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1154 op
= nir_intrinsic_shared_atomic_fmax
;
1156 unreachable("Invalid type");
1158 case ir_intrinsic_shared_atomic_exchange
:
1159 op
= nir_intrinsic_shared_atomic_exchange
;
1161 case ir_intrinsic_shared_atomic_comp_swap
:
1162 op
= ir
->return_deref
->type
->is_integer_32_64()
1163 ? nir_intrinsic_shared_atomic_comp_swap
1164 : nir_intrinsic_shared_atomic_fcomp_swap
;
1166 case ir_intrinsic_vote_any
:
1167 op
= nir_intrinsic_vote_any
;
1169 case ir_intrinsic_vote_all
:
1170 op
= nir_intrinsic_vote_all
;
1172 case ir_intrinsic_vote_eq
:
1173 op
= nir_intrinsic_vote_ieq
;
1175 case ir_intrinsic_ballot
:
1176 op
= nir_intrinsic_ballot
;
1178 case ir_intrinsic_read_invocation
:
1179 op
= nir_intrinsic_read_invocation
;
1181 case ir_intrinsic_read_first_invocation
:
1182 op
= nir_intrinsic_read_first_invocation
;
1184 case ir_intrinsic_helper_invocation
:
1185 op
= nir_intrinsic_is_helper_invocation
;
1188 unreachable("not reached");
1191 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(shader
, op
);
1192 nir_ssa_def
*ret
= &instr
->dest
.ssa
;
1195 case nir_intrinsic_deref_atomic_add
:
1196 case nir_intrinsic_deref_atomic_imin
:
1197 case nir_intrinsic_deref_atomic_umin
:
1198 case nir_intrinsic_deref_atomic_imax
:
1199 case nir_intrinsic_deref_atomic_umax
:
1200 case nir_intrinsic_deref_atomic_and
:
1201 case nir_intrinsic_deref_atomic_or
:
1202 case nir_intrinsic_deref_atomic_xor
:
1203 case nir_intrinsic_deref_atomic_exchange
:
1204 case nir_intrinsic_deref_atomic_comp_swap
:
1205 case nir_intrinsic_deref_atomic_fadd
:
1206 case nir_intrinsic_deref_atomic_fmin
:
1207 case nir_intrinsic_deref_atomic_fmax
:
1208 case nir_intrinsic_deref_atomic_fcomp_swap
: {
1209 int param_count
= ir
->actual_parameters
.length();
1210 assert(param_count
== 2 || param_count
== 3);
1213 exec_node
*param
= ir
->actual_parameters
.get_head();
1214 ir_rvalue
*rvalue
= (ir_rvalue
*) param
;
1215 ir_dereference
*deref
= rvalue
->as_dereference();
1216 ir_swizzle
*swizzle
= NULL
;
1218 /* We may have a swizzle to pick off a single vec4 component */
1219 swizzle
= rvalue
->as_swizzle();
1220 assert(swizzle
&& swizzle
->type
->vector_elements
== 1);
1221 deref
= swizzle
->val
->as_dereference();
1224 nir_deref_instr
*nir_deref
= evaluate_deref(deref
);
1226 nir_deref
= nir_build_deref_array_imm(&b
, nir_deref
,
1229 instr
->src
[0] = nir_src_for_ssa(&nir_deref
->dest
.ssa
);
1231 nir_intrinsic_set_access(instr
, deref_get_qualifier(nir_deref
));
1233 /* data1 parameter (this is always present) */
1234 param
= param
->get_next();
1235 ir_instruction
*inst
= (ir_instruction
*) param
;
1236 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1238 /* data2 parameter (only with atomic_comp_swap) */
1239 if (param_count
== 3) {
1240 assert(op
== nir_intrinsic_deref_atomic_comp_swap
||
1241 op
== nir_intrinsic_deref_atomic_fcomp_swap
);
1242 param
= param
->get_next();
1243 inst
= (ir_instruction
*) param
;
1244 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1248 assert(ir
->return_deref
);
1249 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1250 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1251 nir_builder_instr_insert(&b
, &instr
->instr
);
1254 case nir_intrinsic_atomic_counter_read_deref
:
1255 case nir_intrinsic_atomic_counter_inc_deref
:
1256 case nir_intrinsic_atomic_counter_pre_dec_deref
:
1257 case nir_intrinsic_atomic_counter_add_deref
:
1258 case nir_intrinsic_atomic_counter_min_deref
:
1259 case nir_intrinsic_atomic_counter_max_deref
:
1260 case nir_intrinsic_atomic_counter_and_deref
:
1261 case nir_intrinsic_atomic_counter_or_deref
:
1262 case nir_intrinsic_atomic_counter_xor_deref
:
1263 case nir_intrinsic_atomic_counter_exchange_deref
:
1264 case nir_intrinsic_atomic_counter_comp_swap_deref
: {
1265 /* Set the counter variable dereference. */
1266 exec_node
*param
= ir
->actual_parameters
.get_head();
1267 ir_dereference
*counter
= (ir_dereference
*)param
;
1269 instr
->src
[0] = nir_src_for_ssa(&evaluate_deref(counter
)->dest
.ssa
);
1270 param
= param
->get_next();
1272 /* Set the intrinsic destination. */
1273 if (ir
->return_deref
) {
1274 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1277 /* Set the intrinsic parameters. */
1278 if (!param
->is_tail_sentinel()) {
1280 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1281 param
= param
->get_next();
1284 if (!param
->is_tail_sentinel()) {
1286 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1287 param
= param
->get_next();
1290 nir_builder_instr_insert(&b
, &instr
->instr
);
1293 case nir_intrinsic_image_deref_load
:
1294 case nir_intrinsic_image_deref_store
:
1295 case nir_intrinsic_image_deref_atomic_add
:
1296 case nir_intrinsic_image_deref_atomic_imin
:
1297 case nir_intrinsic_image_deref_atomic_umin
:
1298 case nir_intrinsic_image_deref_atomic_imax
:
1299 case nir_intrinsic_image_deref_atomic_umax
:
1300 case nir_intrinsic_image_deref_atomic_and
:
1301 case nir_intrinsic_image_deref_atomic_or
:
1302 case nir_intrinsic_image_deref_atomic_xor
:
1303 case nir_intrinsic_image_deref_atomic_exchange
:
1304 case nir_intrinsic_image_deref_atomic_comp_swap
:
1305 case nir_intrinsic_image_deref_atomic_fadd
:
1306 case nir_intrinsic_image_deref_samples
:
1307 case nir_intrinsic_image_deref_size
:
1308 case nir_intrinsic_image_deref_atomic_inc_wrap
:
1309 case nir_intrinsic_image_deref_atomic_dec_wrap
: {
1310 nir_ssa_undef_instr
*instr_undef
=
1311 nir_ssa_undef_instr_create(shader
, 1, 32);
1312 nir_builder_instr_insert(&b
, &instr_undef
->instr
);
1314 /* Set the image variable dereference. */
1315 exec_node
*param
= ir
->actual_parameters
.get_head();
1316 ir_dereference
*image
= (ir_dereference
*)param
;
1317 nir_deref_instr
*deref
= evaluate_deref(image
);
1318 const glsl_type
*type
= deref
->type
;
1320 nir_intrinsic_set_access(instr
, deref_get_qualifier(deref
));
1322 instr
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
1323 param
= param
->get_next();
1325 /* Set the intrinsic destination. */
1326 if (ir
->return_deref
) {
1327 unsigned num_components
= ir
->return_deref
->type
->vector_elements
;
1328 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1329 num_components
, 32, NULL
);
1332 if (op
== nir_intrinsic_image_deref_size
) {
1333 instr
->num_components
= instr
->dest
.ssa
.num_components
;
1334 } else if (op
== nir_intrinsic_image_deref_load
||
1335 op
== nir_intrinsic_image_deref_store
) {
1336 instr
->num_components
= 4;
1339 if (op
== nir_intrinsic_image_deref_size
||
1340 op
== nir_intrinsic_image_deref_samples
) {
1341 nir_builder_instr_insert(&b
, &instr
->instr
);
1345 /* Set the address argument, extending the coordinate vector to four
1348 nir_ssa_def
*src_addr
=
1349 evaluate_rvalue((ir_dereference
*)param
);
1350 nir_ssa_def
*srcs
[4];
1352 for (int i
= 0; i
< 4; i
++) {
1353 if (i
< type
->coordinate_components())
1354 srcs
[i
] = nir_channel(&b
, src_addr
, i
);
1356 srcs
[i
] = &instr_undef
->def
;
1359 instr
->src
[1] = nir_src_for_ssa(nir_vec(&b
, srcs
, 4));
1360 param
= param
->get_next();
1362 /* Set the sample argument, which is undefined for single-sample
1365 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_MS
) {
1367 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1368 param
= param
->get_next();
1370 instr
->src
[2] = nir_src_for_ssa(&instr_undef
->def
);
1373 /* Set the intrinsic parameters. */
1374 if (!param
->is_tail_sentinel()) {
1376 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1377 param
= param
->get_next();
1378 } else if (op
== nir_intrinsic_image_deref_load
) {
1379 instr
->src
[3] = nir_src_for_ssa(nir_imm_int(&b
, 0)); /* LOD */
1382 if (!param
->is_tail_sentinel()) {
1384 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1385 param
= param
->get_next();
1386 } else if (op
== nir_intrinsic_image_deref_store
) {
1387 instr
->src
[4] = nir_src_for_ssa(nir_imm_int(&b
, 0)); /* LOD */
1390 nir_builder_instr_insert(&b
, &instr
->instr
);
1393 case nir_intrinsic_memory_barrier
:
1394 case nir_intrinsic_group_memory_barrier
:
1395 case nir_intrinsic_memory_barrier_atomic_counter
:
1396 case nir_intrinsic_memory_barrier_buffer
:
1397 case nir_intrinsic_memory_barrier_image
:
1398 case nir_intrinsic_memory_barrier_shared
:
1399 nir_builder_instr_insert(&b
, &instr
->instr
);
1401 case nir_intrinsic_shader_clock
:
1402 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 2, 32, NULL
);
1403 instr
->num_components
= 2;
1404 nir_builder_instr_insert(&b
, &instr
->instr
);
1406 case nir_intrinsic_begin_invocation_interlock
:
1407 nir_builder_instr_insert(&b
, &instr
->instr
);
1409 case nir_intrinsic_end_invocation_interlock
:
1410 nir_builder_instr_insert(&b
, &instr
->instr
);
1412 case nir_intrinsic_store_ssbo
: {
1413 exec_node
*param
= ir
->actual_parameters
.get_head();
1414 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1416 param
= param
->get_next();
1417 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1419 param
= param
->get_next();
1420 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1422 param
= param
->get_next();
1423 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1426 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1427 if (val
->type
->is_boolean())
1428 nir_val
= nir_b2i32(&b
, nir_val
);
1430 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1431 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(block
));
1432 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(offset
));
1433 intrinsic_set_std430_align(instr
, val
->type
);
1434 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1435 instr
->num_components
= val
->type
->vector_elements
;
1437 nir_builder_instr_insert(&b
, &instr
->instr
);
1440 case nir_intrinsic_load_ssbo
: {
1441 exec_node
*param
= ir
->actual_parameters
.get_head();
1442 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1444 param
= param
->get_next();
1445 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1447 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(block
));
1448 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1450 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1451 instr
->num_components
= type
->vector_elements
;
1452 intrinsic_set_std430_align(instr
, type
);
1454 /* Setup destination register */
1455 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1456 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1457 type
->vector_elements
, bit_size
, NULL
);
1459 /* Insert the created nir instruction now since in the case of boolean
1460 * result we will need to emit another instruction after it
1462 nir_builder_instr_insert(&b
, &instr
->instr
);
1465 * In SSBO/UBO's, a true boolean value is any non-zero value, but we
1466 * consider a true boolean to be ~0. Fix this up with a != 0
1469 if (type
->is_boolean())
1470 ret
= nir_i2b(&b
, &instr
->dest
.ssa
);
1473 case nir_intrinsic_ssbo_atomic_add
:
1474 case nir_intrinsic_ssbo_atomic_imin
:
1475 case nir_intrinsic_ssbo_atomic_umin
:
1476 case nir_intrinsic_ssbo_atomic_imax
:
1477 case nir_intrinsic_ssbo_atomic_umax
:
1478 case nir_intrinsic_ssbo_atomic_and
:
1479 case nir_intrinsic_ssbo_atomic_or
:
1480 case nir_intrinsic_ssbo_atomic_xor
:
1481 case nir_intrinsic_ssbo_atomic_exchange
:
1482 case nir_intrinsic_ssbo_atomic_comp_swap
:
1483 case nir_intrinsic_ssbo_atomic_fadd
:
1484 case nir_intrinsic_ssbo_atomic_fmin
:
1485 case nir_intrinsic_ssbo_atomic_fmax
:
1486 case nir_intrinsic_ssbo_atomic_fcomp_swap
: {
1487 int param_count
= ir
->actual_parameters
.length();
1488 assert(param_count
== 3 || param_count
== 4);
1491 exec_node
*param
= ir
->actual_parameters
.get_head();
1492 ir_instruction
*inst
= (ir_instruction
*) param
;
1493 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1496 param
= param
->get_next();
1497 inst
= (ir_instruction
*) param
;
1498 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1500 /* data1 parameter (this is always present) */
1501 param
= param
->get_next();
1502 inst
= (ir_instruction
*) param
;
1503 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1505 /* data2 parameter (only with atomic_comp_swap) */
1506 if (param_count
== 4) {
1507 assert(op
== nir_intrinsic_ssbo_atomic_comp_swap
||
1508 op
== nir_intrinsic_ssbo_atomic_fcomp_swap
);
1509 param
= param
->get_next();
1510 inst
= (ir_instruction
*) param
;
1511 instr
->src
[3] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1515 assert(ir
->return_deref
);
1516 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1517 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1518 nir_builder_instr_insert(&b
, &instr
->instr
);
1521 case nir_intrinsic_load_shared
: {
1522 exec_node
*param
= ir
->actual_parameters
.get_head();
1523 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1525 nir_intrinsic_set_base(instr
, 0);
1526 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(offset
));
1528 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1529 instr
->num_components
= type
->vector_elements
;
1530 intrinsic_set_std430_align(instr
, type
);
1532 /* Setup destination register */
1533 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1534 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1535 type
->vector_elements
, bit_size
, NULL
);
1537 nir_builder_instr_insert(&b
, &instr
->instr
);
1539 /* The value in shared memory is a 32-bit value */
1540 if (type
->is_boolean())
1541 ret
= nir_i2b(&b
, &instr
->dest
.ssa
);
1544 case nir_intrinsic_store_shared
: {
1545 exec_node
*param
= ir
->actual_parameters
.get_head();
1546 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1548 param
= param
->get_next();
1549 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1551 param
= param
->get_next();
1552 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1555 nir_intrinsic_set_base(instr
, 0);
1556 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1558 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1560 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1561 /* The value in shared memory is a 32-bit value */
1562 if (val
->type
->is_boolean())
1563 nir_val
= nir_b2i32(&b
, nir_val
);
1565 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1566 instr
->num_components
= val
->type
->vector_elements
;
1567 intrinsic_set_std430_align(instr
, val
->type
);
1569 nir_builder_instr_insert(&b
, &instr
->instr
);
1572 case nir_intrinsic_shared_atomic_add
:
1573 case nir_intrinsic_shared_atomic_imin
:
1574 case nir_intrinsic_shared_atomic_umin
:
1575 case nir_intrinsic_shared_atomic_imax
:
1576 case nir_intrinsic_shared_atomic_umax
:
1577 case nir_intrinsic_shared_atomic_and
:
1578 case nir_intrinsic_shared_atomic_or
:
1579 case nir_intrinsic_shared_atomic_xor
:
1580 case nir_intrinsic_shared_atomic_exchange
:
1581 case nir_intrinsic_shared_atomic_comp_swap
:
1582 case nir_intrinsic_shared_atomic_fadd
:
1583 case nir_intrinsic_shared_atomic_fmin
:
1584 case nir_intrinsic_shared_atomic_fmax
:
1585 case nir_intrinsic_shared_atomic_fcomp_swap
: {
1586 int param_count
= ir
->actual_parameters
.length();
1587 assert(param_count
== 2 || param_count
== 3);
1590 exec_node
*param
= ir
->actual_parameters
.get_head();
1591 ir_instruction
*inst
= (ir_instruction
*) param
;
1592 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1594 /* data1 parameter (this is always present) */
1595 param
= param
->get_next();
1596 inst
= (ir_instruction
*) param
;
1597 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1599 /* data2 parameter (only with atomic_comp_swap) */
1600 if (param_count
== 3) {
1601 assert(op
== nir_intrinsic_shared_atomic_comp_swap
||
1602 op
== nir_intrinsic_shared_atomic_fcomp_swap
);
1603 param
= param
->get_next();
1604 inst
= (ir_instruction
*) param
;
1606 nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1610 assert(ir
->return_deref
);
1611 unsigned bit_size
= glsl_get_bit_size(ir
->return_deref
->type
);
1612 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1613 ir
->return_deref
->type
->vector_elements
,
1615 nir_builder_instr_insert(&b
, &instr
->instr
);
1618 case nir_intrinsic_vote_any
:
1619 case nir_intrinsic_vote_all
:
1620 case nir_intrinsic_vote_ieq
: {
1621 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 1, NULL
);
1622 instr
->num_components
= 1;
1624 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1625 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1627 nir_builder_instr_insert(&b
, &instr
->instr
);
1631 case nir_intrinsic_ballot
: {
1632 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1633 ir
->return_deref
->type
->vector_elements
, 64, NULL
);
1634 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1636 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1637 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1639 nir_builder_instr_insert(&b
, &instr
->instr
);
1642 case nir_intrinsic_read_invocation
: {
1643 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1644 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1645 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1647 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1648 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1650 ir_rvalue
*invocation
= (ir_rvalue
*) ir
->actual_parameters
.get_head()->next
;
1651 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(invocation
));
1653 nir_builder_instr_insert(&b
, &instr
->instr
);
1656 case nir_intrinsic_read_first_invocation
: {
1657 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1658 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1659 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1661 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1662 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1664 nir_builder_instr_insert(&b
, &instr
->instr
);
1667 case nir_intrinsic_is_helper_invocation
: {
1668 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 1, NULL
);
1669 instr
->num_components
= 1;
1670 nir_builder_instr_insert(&b
, &instr
->instr
);
1674 unreachable("not reached");
1677 if (ir
->return_deref
)
1678 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), ret
, ~0);
1683 struct hash_entry
*entry
=
1684 _mesa_hash_table_search(this->overload_table
, ir
->callee
);
1686 nir_function
*callee
= (nir_function
*) entry
->data
;
1688 nir_call_instr
*call
= nir_call_instr_create(this->shader
, callee
);
1691 nir_deref_instr
*ret_deref
= NULL
;
1692 if (ir
->return_deref
) {
1693 nir_variable
*ret_tmp
=
1694 nir_local_variable_create(this->impl
, ir
->return_deref
->type
,
1696 ret_deref
= nir_build_deref_var(&b
, ret_tmp
);
1697 call
->params
[i
++] = nir_src_for_ssa(&ret_deref
->dest
.ssa
);
1700 foreach_two_lists(formal_node
, &ir
->callee
->parameters
,
1701 actual_node
, &ir
->actual_parameters
) {
1702 ir_rvalue
*param_rvalue
= (ir_rvalue
*) actual_node
;
1703 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
1705 if (sig_param
->data
.mode
== ir_var_function_out
) {
1706 nir_deref_instr
*out_deref
= evaluate_deref(param_rvalue
);
1707 call
->params
[i
] = nir_src_for_ssa(&out_deref
->dest
.ssa
);
1708 } else if (sig_param
->data
.mode
== ir_var_function_in
) {
1709 nir_ssa_def
*val
= evaluate_rvalue(param_rvalue
);
1710 nir_src src
= nir_src_for_ssa(val
);
1712 nir_src_copy(&call
->params
[i
], &src
, call
);
1713 } else if (sig_param
->data
.mode
== ir_var_function_inout
) {
1714 unreachable("unimplemented: inout parameters");
1720 nir_builder_instr_insert(&b
, &call
->instr
);
1722 if (ir
->return_deref
)
1723 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), nir_load_deref(&b
, ret_deref
), ~0);
1727 nir_visitor::visit(ir_assignment
*ir
)
1729 unsigned num_components
= ir
->lhs
->type
->vector_elements
;
1731 b
.exact
= ir
->lhs
->variable_referenced()->data
.invariant
||
1732 ir
->lhs
->variable_referenced()->data
.precise
;
1734 if ((ir
->rhs
->as_dereference() || ir
->rhs
->as_constant()) &&
1735 (ir
->write_mask
== (1 << num_components
) - 1 || ir
->write_mask
== 0)) {
1736 nir_deref_instr
*lhs
= evaluate_deref(ir
->lhs
);
1737 nir_deref_instr
*rhs
= evaluate_deref(ir
->rhs
);
1738 enum gl_access_qualifier lhs_qualifiers
= deref_get_qualifier(lhs
);
1739 enum gl_access_qualifier rhs_qualifiers
= deref_get_qualifier(rhs
);
1740 if (ir
->condition
) {
1741 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1742 nir_copy_deref_with_access(&b
, lhs
, rhs
, lhs_qualifiers
,
1744 nir_pop_if(&b
, NULL
);
1746 nir_copy_deref_with_access(&b
, lhs
, rhs
, lhs_qualifiers
,
1752 assert(ir
->rhs
->type
->is_scalar() || ir
->rhs
->type
->is_vector());
1754 ir
->lhs
->accept(this);
1755 nir_deref_instr
*lhs_deref
= this->deref
;
1756 nir_ssa_def
*src
= evaluate_rvalue(ir
->rhs
);
1758 if (ir
->write_mask
!= (1 << num_components
) - 1 && ir
->write_mask
!= 0) {
1759 /* GLSL IR will give us the input to the write-masked assignment in a
1760 * single packed vector. So, for example, if the writemask is xzw, then
1761 * we have to swizzle x -> x, y -> z, and z -> w and get the y component
1765 unsigned component
= 0;
1766 for (unsigned i
= 0; i
< 4; i
++) {
1767 swiz
[i
] = ir
->write_mask
& (1 << i
) ? component
++ : 0;
1769 src
= nir_swizzle(&b
, src
, swiz
, num_components
);
1772 enum gl_access_qualifier qualifiers
= deref_get_qualifier(lhs_deref
);
1773 if (ir
->condition
) {
1774 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1775 nir_store_deref_with_access(&b
, lhs_deref
, src
, ir
->write_mask
,
1777 nir_pop_if(&b
, NULL
);
1779 nir_store_deref_with_access(&b
, lhs_deref
, src
, ir
->write_mask
,
1785 * Given an instruction, returns a pointer to its destination or NULL if there
1786 * is no destination.
1788 * Note that this only handles instructions we generate at this level.
1791 get_instr_dest(nir_instr
*instr
)
1793 nir_alu_instr
*alu_instr
;
1794 nir_intrinsic_instr
*intrinsic_instr
;
1795 nir_tex_instr
*tex_instr
;
1797 switch (instr
->type
) {
1798 case nir_instr_type_alu
:
1799 alu_instr
= nir_instr_as_alu(instr
);
1800 return &alu_instr
->dest
.dest
;
1802 case nir_instr_type_intrinsic
:
1803 intrinsic_instr
= nir_instr_as_intrinsic(instr
);
1804 if (nir_intrinsic_infos
[intrinsic_instr
->intrinsic
].has_dest
)
1805 return &intrinsic_instr
->dest
;
1809 case nir_instr_type_tex
:
1810 tex_instr
= nir_instr_as_tex(instr
);
1811 return &tex_instr
->dest
;
1814 unreachable("not reached");
1821 nir_visitor::add_instr(nir_instr
*instr
, unsigned num_components
,
1824 nir_dest
*dest
= get_instr_dest(instr
);
1827 nir_ssa_dest_init(instr
, dest
, num_components
, bit_size
, NULL
);
1829 nir_builder_instr_insert(&b
, instr
);
1832 assert(dest
->is_ssa
);
1833 this->result
= &dest
->ssa
;
1838 nir_visitor::evaluate_rvalue(ir_rvalue
* ir
)
1841 if (ir
->as_dereference() || ir
->as_constant()) {
1843 * A dereference is being used on the right hand side, which means we
1844 * must emit a variable load.
1847 enum gl_access_qualifier access
= deref_get_qualifier(this->deref
);
1848 this->result
= nir_load_deref_with_access(&b
, this->deref
, access
);
1851 return this->result
;
1855 type_is_float(glsl_base_type type
)
1857 return type
== GLSL_TYPE_FLOAT
|| type
== GLSL_TYPE_DOUBLE
||
1858 type
== GLSL_TYPE_FLOAT16
;
1862 type_is_signed(glsl_base_type type
)
1864 return type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_INT64
||
1865 type
== GLSL_TYPE_INT16
;
1869 nir_visitor::visit(ir_expression
*ir
)
1871 /* Some special cases */
1872 switch (ir
->operation
) {
1873 case ir_binop_ubo_load
: {
1874 nir_intrinsic_instr
*load
=
1875 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_ubo
);
1876 unsigned bit_size
= ir
->type
->is_boolean() ? 32 :
1877 glsl_get_bit_size(ir
->type
);
1878 load
->num_components
= ir
->type
->vector_elements
;
1879 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1880 load
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1881 intrinsic_set_std430_align(load
, ir
->type
);
1882 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1885 * In UBO's, a true boolean value is any non-zero value, but we consider
1886 * a true boolean to be ~0. Fix this up with a != 0 comparison.
1889 if (ir
->type
->is_boolean())
1890 this->result
= nir_i2b(&b
, &load
->dest
.ssa
);
1895 case ir_unop_interpolate_at_centroid
:
1896 case ir_binop_interpolate_at_offset
:
1897 case ir_binop_interpolate_at_sample
: {
1898 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1899 ir_swizzle
*swizzle
= NULL
;
1901 /* the api does not allow a swizzle here, but the varying packing code
1902 * may have pushed one into here.
1904 swizzle
= ir
->operands
[0]->as_swizzle();
1906 deref
= swizzle
->val
->as_dereference();
1910 deref
->accept(this);
1912 nir_intrinsic_op op
;
1913 if (this->deref
->mode
== nir_var_shader_in
) {
1914 switch (ir
->operation
) {
1915 case ir_unop_interpolate_at_centroid
:
1916 op
= nir_intrinsic_interp_deref_at_centroid
;
1918 case ir_binop_interpolate_at_offset
:
1919 op
= nir_intrinsic_interp_deref_at_offset
;
1921 case ir_binop_interpolate_at_sample
:
1922 op
= nir_intrinsic_interp_deref_at_sample
;
1925 unreachable("Invalid interpolation intrinsic");
1928 /* This case can happen if the vertex shader does not write the
1929 * given varying. In this case, the linker will lower it to a
1930 * global variable. Since interpolating a variable makes no
1931 * sense, we'll just turn it into a load which will probably
1932 * eventually end up as an SSA definition.
1934 assert(this->deref
->mode
== nir_var_shader_temp
);
1935 op
= nir_intrinsic_load_deref
;
1938 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(shader
, op
);
1939 intrin
->num_components
= deref
->type
->vector_elements
;
1940 intrin
->src
[0] = nir_src_for_ssa(&this->deref
->dest
.ssa
);
1942 if (intrin
->intrinsic
== nir_intrinsic_interp_deref_at_offset
||
1943 intrin
->intrinsic
== nir_intrinsic_interp_deref_at_sample
)
1944 intrin
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1946 unsigned bit_size
= glsl_get_bit_size(deref
->type
);
1947 add_instr(&intrin
->instr
, deref
->type
->vector_elements
, bit_size
);
1950 unsigned swiz
[4] = {
1951 swizzle
->mask
.x
, swizzle
->mask
.y
, swizzle
->mask
.z
, swizzle
->mask
.w
1954 result
= nir_swizzle(&b
, result
, swiz
,
1955 swizzle
->type
->vector_elements
);
1961 case ir_unop_ssbo_unsized_array_length
: {
1962 nir_intrinsic_instr
*intrin
=
1963 nir_intrinsic_instr_create(b
.shader
,
1964 nir_intrinsic_deref_buffer_array_length
);
1966 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1967 intrin
->src
[0] = nir_src_for_ssa(&evaluate_deref(deref
)->dest
.ssa
);
1969 add_instr(&intrin
->instr
, 1, 32);
1977 nir_ssa_def
*srcs
[4];
1978 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1979 srcs
[i
] = evaluate_rvalue(ir
->operands
[i
]);
1981 glsl_base_type types
[4];
1982 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1983 types
[i
] = ir
->operands
[i
]->type
->base_type
;
1985 glsl_base_type out_type
= ir
->type
->base_type
;
1987 switch (ir
->operation
) {
1988 case ir_unop_bit_not
: result
= nir_inot(&b
, srcs
[0]); break;
1989 case ir_unop_logic_not
:
1990 result
= nir_inot(&b
, srcs
[0]);
1993 result
= type_is_float(types
[0]) ? nir_fneg(&b
, srcs
[0])
1994 : nir_ineg(&b
, srcs
[0]);
1997 result
= type_is_float(types
[0]) ? nir_fabs(&b
, srcs
[0])
1998 : nir_iabs(&b
, srcs
[0]);
2001 result
= nir_uclz(&b
, srcs
[0]);
2003 case ir_unop_saturate
:
2004 assert(type_is_float(types
[0]));
2005 result
= nir_fsat(&b
, srcs
[0]);
2008 result
= type_is_float(types
[0]) ? nir_fsign(&b
, srcs
[0])
2009 : nir_isign(&b
, srcs
[0]);
2011 case ir_unop_rcp
: result
= nir_frcp(&b
, srcs
[0]); break;
2012 case ir_unop_rsq
: result
= nir_frsq(&b
, srcs
[0]); break;
2013 case ir_unop_sqrt
: result
= nir_fsqrt(&b
, srcs
[0]); break;
2014 case ir_unop_exp
: unreachable("ir_unop_exp should have been lowered");
2015 case ir_unop_log
: unreachable("ir_unop_log should have been lowered");
2016 case ir_unop_exp2
: result
= nir_fexp2(&b
, srcs
[0]); break;
2017 case ir_unop_log2
: result
= nir_flog2(&b
, srcs
[0]); break;
2053 case ir_unop_i642u64
:
2054 case ir_unop_u642i64
: {
2055 nir_alu_type src_type
= nir_get_nir_type_for_glsl_base_type(types
[0]);
2056 nir_alu_type dst_type
= nir_get_nir_type_for_glsl_base_type(out_type
);
2057 result
= nir_build_alu(&b
, nir_type_conversion_op(src_type
, dst_type
,
2058 nir_rounding_mode_undef
),
2059 srcs
[0], NULL
, NULL
, NULL
);
2060 /* b2i and b2f don't have fixed bit-size versions so the builder will
2061 * just assume 32 and we have to fix it up here.
2063 result
->bit_size
= nir_alu_type_get_type_size(dst_type
);
2067 case ir_unop_bitcast_i2f
:
2068 case ir_unop_bitcast_f2i
:
2069 case ir_unop_bitcast_u2f
:
2070 case ir_unop_bitcast_f2u
:
2071 case ir_unop_bitcast_i642d
:
2072 case ir_unop_bitcast_d2i64
:
2073 case ir_unop_bitcast_u642d
:
2074 case ir_unop_bitcast_d2u64
:
2075 case ir_unop_subroutine_to_int
:
2077 result
= nir_mov(&b
, srcs
[0]);
2079 case ir_unop_trunc
: result
= nir_ftrunc(&b
, srcs
[0]); break;
2080 case ir_unop_ceil
: result
= nir_fceil(&b
, srcs
[0]); break;
2081 case ir_unop_floor
: result
= nir_ffloor(&b
, srcs
[0]); break;
2082 case ir_unop_fract
: result
= nir_ffract(&b
, srcs
[0]); break;
2083 case ir_unop_frexp_exp
: result
= nir_frexp_exp(&b
, srcs
[0]); break;
2084 case ir_unop_frexp_sig
: result
= nir_frexp_sig(&b
, srcs
[0]); break;
2085 case ir_unop_round_even
: result
= nir_fround_even(&b
, srcs
[0]); break;
2086 case ir_unop_sin
: result
= nir_fsin(&b
, srcs
[0]); break;
2087 case ir_unop_cos
: result
= nir_fcos(&b
, srcs
[0]); break;
2088 case ir_unop_dFdx
: result
= nir_fddx(&b
, srcs
[0]); break;
2089 case ir_unop_dFdy
: result
= nir_fddy(&b
, srcs
[0]); break;
2090 case ir_unop_dFdx_fine
: result
= nir_fddx_fine(&b
, srcs
[0]); break;
2091 case ir_unop_dFdy_fine
: result
= nir_fddy_fine(&b
, srcs
[0]); break;
2092 case ir_unop_dFdx_coarse
: result
= nir_fddx_coarse(&b
, srcs
[0]); break;
2093 case ir_unop_dFdy_coarse
: result
= nir_fddy_coarse(&b
, srcs
[0]); break;
2094 case ir_unop_pack_snorm_2x16
:
2095 result
= nir_pack_snorm_2x16(&b
, srcs
[0]);
2097 case ir_unop_pack_snorm_4x8
:
2098 result
= nir_pack_snorm_4x8(&b
, srcs
[0]);
2100 case ir_unop_pack_unorm_2x16
:
2101 result
= nir_pack_unorm_2x16(&b
, srcs
[0]);
2103 case ir_unop_pack_unorm_4x8
:
2104 result
= nir_pack_unorm_4x8(&b
, srcs
[0]);
2106 case ir_unop_pack_half_2x16
:
2107 result
= nir_pack_half_2x16(&b
, srcs
[0]);
2109 case ir_unop_unpack_snorm_2x16
:
2110 result
= nir_unpack_snorm_2x16(&b
, srcs
[0]);
2112 case ir_unop_unpack_snorm_4x8
:
2113 result
= nir_unpack_snorm_4x8(&b
, srcs
[0]);
2115 case ir_unop_unpack_unorm_2x16
:
2116 result
= nir_unpack_unorm_2x16(&b
, srcs
[0]);
2118 case ir_unop_unpack_unorm_4x8
:
2119 result
= nir_unpack_unorm_4x8(&b
, srcs
[0]);
2121 case ir_unop_unpack_half_2x16
:
2122 result
= nir_unpack_half_2x16(&b
, srcs
[0]);
2124 case ir_unop_pack_sampler_2x32
:
2125 case ir_unop_pack_image_2x32
:
2126 case ir_unop_pack_double_2x32
:
2127 case ir_unop_pack_int_2x32
:
2128 case ir_unop_pack_uint_2x32
:
2129 result
= nir_pack_64_2x32(&b
, srcs
[0]);
2131 case ir_unop_unpack_sampler_2x32
:
2132 case ir_unop_unpack_image_2x32
:
2133 case ir_unop_unpack_double_2x32
:
2134 case ir_unop_unpack_int_2x32
:
2135 case ir_unop_unpack_uint_2x32
:
2136 result
= nir_unpack_64_2x32(&b
, srcs
[0]);
2138 case ir_unop_bitfield_reverse
:
2139 result
= nir_bitfield_reverse(&b
, srcs
[0]);
2141 case ir_unop_bit_count
:
2142 result
= nir_bit_count(&b
, srcs
[0]);
2144 case ir_unop_find_msb
:
2146 case GLSL_TYPE_UINT
:
2147 result
= nir_ufind_msb(&b
, srcs
[0]);
2150 result
= nir_ifind_msb(&b
, srcs
[0]);
2153 unreachable("Invalid type for findMSB()");
2156 case ir_unop_find_lsb
:
2157 result
= nir_find_lsb(&b
, srcs
[0]);
2161 switch (ir
->type
->vector_elements
) {
2163 switch (ir
->operands
[0]->type
->vector_elements
) {
2164 case 1: result
= nir_fnoise1_1(&b
, srcs
[0]); break;
2165 case 2: result
= nir_fnoise1_2(&b
, srcs
[0]); break;
2166 case 3: result
= nir_fnoise1_3(&b
, srcs
[0]); break;
2167 case 4: result
= nir_fnoise1_4(&b
, srcs
[0]); break;
2168 default: unreachable("not reached");
2172 switch (ir
->operands
[0]->type
->vector_elements
) {
2173 case 1: result
= nir_fnoise2_1(&b
, srcs
[0]); break;
2174 case 2: result
= nir_fnoise2_2(&b
, srcs
[0]); break;
2175 case 3: result
= nir_fnoise2_3(&b
, srcs
[0]); break;
2176 case 4: result
= nir_fnoise2_4(&b
, srcs
[0]); break;
2177 default: unreachable("not reached");
2181 switch (ir
->operands
[0]->type
->vector_elements
) {
2182 case 1: result
= nir_fnoise3_1(&b
, srcs
[0]); break;
2183 case 2: result
= nir_fnoise3_2(&b
, srcs
[0]); break;
2184 case 3: result
= nir_fnoise3_3(&b
, srcs
[0]); break;
2185 case 4: result
= nir_fnoise3_4(&b
, srcs
[0]); break;
2186 default: unreachable("not reached");
2190 switch (ir
->operands
[0]->type
->vector_elements
) {
2191 case 1: result
= nir_fnoise4_1(&b
, srcs
[0]); break;
2192 case 2: result
= nir_fnoise4_2(&b
, srcs
[0]); break;
2193 case 3: result
= nir_fnoise4_3(&b
, srcs
[0]); break;
2194 case 4: result
= nir_fnoise4_4(&b
, srcs
[0]); break;
2195 default: unreachable("not reached");
2199 unreachable("not reached");
2202 case ir_unop_get_buffer_size
: {
2203 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(
2205 nir_intrinsic_get_buffer_size
);
2206 load
->num_components
= ir
->type
->vector_elements
;
2207 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
2208 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2209 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
2214 result
= nir_atan(&b
, srcs
[0]);
2218 result
= type_is_float(out_type
) ? nir_fadd(&b
, srcs
[0], srcs
[1])
2219 : nir_iadd(&b
, srcs
[0], srcs
[1]);
2221 case ir_binop_add_sat
:
2222 result
= type_is_signed(out_type
) ? nir_iadd_sat(&b
, srcs
[0], srcs
[1])
2223 : nir_uadd_sat(&b
, srcs
[0], srcs
[1]);
2226 result
= type_is_float(out_type
) ? nir_fsub(&b
, srcs
[0], srcs
[1])
2227 : nir_isub(&b
, srcs
[0], srcs
[1]);
2229 case ir_binop_sub_sat
:
2230 result
= type_is_signed(out_type
) ? nir_isub_sat(&b
, srcs
[0], srcs
[1])
2231 : nir_usub_sat(&b
, srcs
[0], srcs
[1]);
2233 case ir_binop_abs_sub
:
2234 /* out_type is always unsigned for ir_binop_abs_sub, so we have to key
2235 * on the type of the sources.
2237 result
= type_is_signed(types
[0]) ? nir_uabs_isub(&b
, srcs
[0], srcs
[1])
2238 : nir_uabs_usub(&b
, srcs
[0], srcs
[1]);
2241 result
= type_is_signed(out_type
) ? nir_ihadd(&b
, srcs
[0], srcs
[1])
2242 : nir_uhadd(&b
, srcs
[0], srcs
[1]);
2244 case ir_binop_avg_round
:
2245 result
= type_is_signed(out_type
) ? nir_irhadd(&b
, srcs
[0], srcs
[1])
2246 : nir_urhadd(&b
, srcs
[0], srcs
[1]);
2248 case ir_binop_mul_32x16
:
2249 result
= type_is_signed(out_type
) ? nir_imul_32x16(&b
, srcs
[0], srcs
[1])
2250 : nir_umul_32x16(&b
, srcs
[0], srcs
[1]);
2253 if (type_is_float(out_type
))
2254 result
= nir_fmul(&b
, srcs
[0], srcs
[1]);
2255 else if (out_type
== GLSL_TYPE_INT64
&&
2256 (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
||
2257 ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
))
2258 result
= nir_imul_2x32_64(&b
, srcs
[0], srcs
[1]);
2259 else if (out_type
== GLSL_TYPE_UINT64
&&
2260 (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
||
2261 ir
->operands
[1]->type
->base_type
== GLSL_TYPE_UINT
))
2262 result
= nir_umul_2x32_64(&b
, srcs
[0], srcs
[1]);
2264 result
= nir_imul(&b
, srcs
[0], srcs
[1]);
2267 if (type_is_float(out_type
))
2268 result
= nir_fdiv(&b
, srcs
[0], srcs
[1]);
2269 else if (type_is_signed(out_type
))
2270 result
= nir_idiv(&b
, srcs
[0], srcs
[1]);
2272 result
= nir_udiv(&b
, srcs
[0], srcs
[1]);
2275 result
= type_is_float(out_type
) ? nir_fmod(&b
, srcs
[0], srcs
[1])
2276 : nir_umod(&b
, srcs
[0], srcs
[1]);
2279 if (type_is_float(out_type
))
2280 result
= nir_fmin(&b
, srcs
[0], srcs
[1]);
2281 else if (type_is_signed(out_type
))
2282 result
= nir_imin(&b
, srcs
[0], srcs
[1]);
2284 result
= nir_umin(&b
, srcs
[0], srcs
[1]);
2287 if (type_is_float(out_type
))
2288 result
= nir_fmax(&b
, srcs
[0], srcs
[1]);
2289 else if (type_is_signed(out_type
))
2290 result
= nir_imax(&b
, srcs
[0], srcs
[1]);
2292 result
= nir_umax(&b
, srcs
[0], srcs
[1]);
2294 case ir_binop_pow
: result
= nir_fpow(&b
, srcs
[0], srcs
[1]); break;
2295 case ir_binop_bit_and
: result
= nir_iand(&b
, srcs
[0], srcs
[1]); break;
2296 case ir_binop_bit_or
: result
= nir_ior(&b
, srcs
[0], srcs
[1]); break;
2297 case ir_binop_bit_xor
: result
= nir_ixor(&b
, srcs
[0], srcs
[1]); break;
2298 case ir_binop_logic_and
:
2299 result
= nir_iand(&b
, srcs
[0], srcs
[1]);
2301 case ir_binop_logic_or
:
2302 result
= nir_ior(&b
, srcs
[0], srcs
[1]);
2304 case ir_binop_logic_xor
:
2305 result
= nir_ixor(&b
, srcs
[0], srcs
[1]);
2307 case ir_binop_lshift
: result
= nir_ishl(&b
, srcs
[0], srcs
[1]); break;
2308 case ir_binop_rshift
:
2309 result
= (type_is_signed(out_type
)) ? nir_ishr(&b
, srcs
[0], srcs
[1])
2310 : nir_ushr(&b
, srcs
[0], srcs
[1]);
2312 case ir_binop_imul_high
:
2313 result
= (out_type
== GLSL_TYPE_INT
) ? nir_imul_high(&b
, srcs
[0], srcs
[1])
2314 : nir_umul_high(&b
, srcs
[0], srcs
[1]);
2316 case ir_binop_carry
: result
= nir_uadd_carry(&b
, srcs
[0], srcs
[1]); break;
2317 case ir_binop_borrow
: result
= nir_usub_borrow(&b
, srcs
[0], srcs
[1]); break;
2319 if (type_is_float(types
[0]))
2320 result
= nir_flt(&b
, srcs
[0], srcs
[1]);
2321 else if (type_is_signed(types
[0]))
2322 result
= nir_ilt(&b
, srcs
[0], srcs
[1]);
2324 result
= nir_ult(&b
, srcs
[0], srcs
[1]);
2326 case ir_binop_gequal
:
2327 if (type_is_float(types
[0]))
2328 result
= nir_fge(&b
, srcs
[0], srcs
[1]);
2329 else if (type_is_signed(types
[0]))
2330 result
= nir_ige(&b
, srcs
[0], srcs
[1]);
2332 result
= nir_uge(&b
, srcs
[0], srcs
[1]);
2334 case ir_binop_equal
:
2335 if (type_is_float(types
[0]))
2336 result
= nir_feq(&b
, srcs
[0], srcs
[1]);
2338 result
= nir_ieq(&b
, srcs
[0], srcs
[1]);
2340 case ir_binop_nequal
:
2341 if (type_is_float(types
[0]))
2342 result
= nir_fne(&b
, srcs
[0], srcs
[1]);
2344 result
= nir_ine(&b
, srcs
[0], srcs
[1]);
2346 case ir_binop_all_equal
:
2347 if (type_is_float(types
[0])) {
2348 switch (ir
->operands
[0]->type
->vector_elements
) {
2349 case 1: result
= nir_feq(&b
, srcs
[0], srcs
[1]); break;
2350 case 2: result
= nir_ball_fequal2(&b
, srcs
[0], srcs
[1]); break;
2351 case 3: result
= nir_ball_fequal3(&b
, srcs
[0], srcs
[1]); break;
2352 case 4: result
= nir_ball_fequal4(&b
, srcs
[0], srcs
[1]); break;
2354 unreachable("not reached");
2357 switch (ir
->operands
[0]->type
->vector_elements
) {
2358 case 1: result
= nir_ieq(&b
, srcs
[0], srcs
[1]); break;
2359 case 2: result
= nir_ball_iequal2(&b
, srcs
[0], srcs
[1]); break;
2360 case 3: result
= nir_ball_iequal3(&b
, srcs
[0], srcs
[1]); break;
2361 case 4: result
= nir_ball_iequal4(&b
, srcs
[0], srcs
[1]); break;
2363 unreachable("not reached");
2367 case ir_binop_any_nequal
:
2368 if (type_is_float(types
[0])) {
2369 switch (ir
->operands
[0]->type
->vector_elements
) {
2370 case 1: result
= nir_fne(&b
, srcs
[0], srcs
[1]); break;
2371 case 2: result
= nir_bany_fnequal2(&b
, srcs
[0], srcs
[1]); break;
2372 case 3: result
= nir_bany_fnequal3(&b
, srcs
[0], srcs
[1]); break;
2373 case 4: result
= nir_bany_fnequal4(&b
, srcs
[0], srcs
[1]); break;
2375 unreachable("not reached");
2378 switch (ir
->operands
[0]->type
->vector_elements
) {
2379 case 1: result
= nir_ine(&b
, srcs
[0], srcs
[1]); break;
2380 case 2: result
= nir_bany_inequal2(&b
, srcs
[0], srcs
[1]); break;
2381 case 3: result
= nir_bany_inequal3(&b
, srcs
[0], srcs
[1]); break;
2382 case 4: result
= nir_bany_inequal4(&b
, srcs
[0], srcs
[1]); break;
2384 unreachable("not reached");
2389 switch (ir
->operands
[0]->type
->vector_elements
) {
2390 case 2: result
= nir_fdot2(&b
, srcs
[0], srcs
[1]); break;
2391 case 3: result
= nir_fdot3(&b
, srcs
[0], srcs
[1]); break;
2392 case 4: result
= nir_fdot4(&b
, srcs
[0], srcs
[1]); break;
2394 unreachable("not reached");
2397 case ir_binop_vector_extract
: {
2398 result
= nir_channel(&b
, srcs
[0], 0);
2399 for (unsigned i
= 1; i
< ir
->operands
[0]->type
->vector_elements
; i
++) {
2400 nir_ssa_def
*swizzled
= nir_channel(&b
, srcs
[0], i
);
2401 result
= nir_bcsel(&b
, nir_ieq(&b
, srcs
[1], nir_imm_int(&b
, i
)),
2407 case ir_binop_atan2
:
2408 result
= nir_atan2(&b
, srcs
[0], srcs
[1]);
2411 case ir_binop_ldexp
: result
= nir_ldexp(&b
, srcs
[0], srcs
[1]); break;
2413 result
= nir_ffma(&b
, srcs
[0], srcs
[1], srcs
[2]);
2416 result
= nir_flrp(&b
, srcs
[0], srcs
[1], srcs
[2]);
2419 result
= nir_bcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
2421 case ir_triop_bitfield_extract
:
2422 result
= (out_type
== GLSL_TYPE_INT
) ?
2423 nir_ibitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]) :
2424 nir_ubitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]);
2426 case ir_quadop_bitfield_insert
:
2427 result
= nir_bitfield_insert(&b
, srcs
[0], srcs
[1], srcs
[2], srcs
[3]);
2429 case ir_quadop_vector
:
2430 result
= nir_vec(&b
, srcs
, ir
->type
->vector_elements
);
2434 unreachable("not reached");
2439 nir_visitor::visit(ir_swizzle
*ir
)
2441 unsigned swizzle
[4] = { ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
};
2442 result
= nir_swizzle(&b
, evaluate_rvalue(ir
->val
), swizzle
,
2443 ir
->type
->vector_elements
);
2447 nir_visitor::visit(ir_texture
*ir
)
2454 num_srcs
= 1; /* coordinate */
2459 op
= (ir
->op
== ir_txb
) ? nir_texop_txb
: nir_texop_txl
;
2460 num_srcs
= 2; /* coordinate, bias/lod */
2464 op
= nir_texop_txd
; /* coordinate, dPdx, dPdy */
2470 if (ir
->lod_info
.lod
!= NULL
)
2471 num_srcs
= 2; /* coordinate, lod */
2473 num_srcs
= 1; /* coordinate */
2477 op
= nir_texop_txf_ms
;
2478 num_srcs
= 2; /* coordinate, sample_index */
2483 if (ir
->lod_info
.lod
!= NULL
)
2484 num_srcs
= 1; /* lod */
2491 num_srcs
= 1; /* coordinate */
2496 num_srcs
= 1; /* coordinate */
2499 case ir_query_levels
:
2500 op
= nir_texop_query_levels
;
2504 case ir_texture_samples
:
2505 op
= nir_texop_texture_samples
;
2509 case ir_samples_identical
:
2510 op
= nir_texop_samples_identical
;
2511 num_srcs
= 1; /* coordinate */
2515 unreachable("not reached");
2518 if (ir
->projector
!= NULL
)
2520 if (ir
->shadow_comparator
!= NULL
)
2522 /* offsets are constants we store inside nir_tex_intrs.offsets */
2523 if (ir
->offset
!= NULL
&& !ir
->offset
->type
->is_array())
2526 /* Add one for the texture deref */
2529 nir_tex_instr
*instr
= nir_tex_instr_create(this->shader
, num_srcs
);
2532 instr
->sampler_dim
=
2533 (glsl_sampler_dim
) ir
->sampler
->type
->sampler_dimensionality
;
2534 instr
->is_array
= ir
->sampler
->type
->sampler_array
;
2535 instr
->is_shadow
= ir
->sampler
->type
->sampler_shadow
;
2536 if (instr
->is_shadow
)
2537 instr
->is_new_style_shadow
= (ir
->type
->vector_elements
== 1);
2538 switch (ir
->type
->base_type
) {
2539 case GLSL_TYPE_FLOAT
:
2540 instr
->dest_type
= nir_type_float
;
2543 instr
->dest_type
= nir_type_int
;
2545 case GLSL_TYPE_BOOL
:
2546 case GLSL_TYPE_UINT
:
2547 instr
->dest_type
= nir_type_uint
;
2550 unreachable("not reached");
2553 nir_deref_instr
*sampler_deref
= evaluate_deref(ir
->sampler
);
2555 /* check for bindless handles */
2556 if (sampler_deref
->mode
!= nir_var_uniform
||
2557 nir_deref_instr_get_variable(sampler_deref
)->data
.bindless
) {
2558 nir_ssa_def
*load
= nir_load_deref(&b
, sampler_deref
);
2559 instr
->src
[0].src
= nir_src_for_ssa(load
);
2560 instr
->src
[0].src_type
= nir_tex_src_texture_handle
;
2561 instr
->src
[1].src
= nir_src_for_ssa(load
);
2562 instr
->src
[1].src_type
= nir_tex_src_sampler_handle
;
2564 instr
->src
[0].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2565 instr
->src
[0].src_type
= nir_tex_src_texture_deref
;
2566 instr
->src
[1].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2567 instr
->src
[1].src_type
= nir_tex_src_sampler_deref
;
2570 unsigned src_number
= 2;
2572 if (ir
->coordinate
!= NULL
) {
2573 instr
->coord_components
= ir
->coordinate
->type
->vector_elements
;
2574 instr
->src
[src_number
].src
=
2575 nir_src_for_ssa(evaluate_rvalue(ir
->coordinate
));
2576 instr
->src
[src_number
].src_type
= nir_tex_src_coord
;
2580 if (ir
->projector
!= NULL
) {
2581 instr
->src
[src_number
].src
=
2582 nir_src_for_ssa(evaluate_rvalue(ir
->projector
));
2583 instr
->src
[src_number
].src_type
= nir_tex_src_projector
;
2587 if (ir
->shadow_comparator
!= NULL
) {
2588 instr
->src
[src_number
].src
=
2589 nir_src_for_ssa(evaluate_rvalue(ir
->shadow_comparator
));
2590 instr
->src
[src_number
].src_type
= nir_tex_src_comparator
;
2594 if (ir
->offset
!= NULL
) {
2595 if (ir
->offset
->type
->is_array()) {
2596 for (int i
= 0; i
< ir
->offset
->type
->array_size(); i
++) {
2597 const ir_constant
*c
=
2598 ir
->offset
->as_constant()->get_array_element(i
);
2600 for (unsigned j
= 0; j
< 2; ++j
) {
2601 int val
= c
->get_int_component(j
);
2602 assert(val
<= 31 && val
>= -32);
2603 instr
->tg4_offsets
[i
][j
] = val
;
2607 assert(ir
->offset
->type
->is_vector() || ir
->offset
->type
->is_scalar());
2609 instr
->src
[src_number
].src
=
2610 nir_src_for_ssa(evaluate_rvalue(ir
->offset
));
2611 instr
->src
[src_number
].src_type
= nir_tex_src_offset
;
2618 instr
->src
[src_number
].src
=
2619 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.bias
));
2620 instr
->src
[src_number
].src_type
= nir_tex_src_bias
;
2627 if (ir
->lod_info
.lod
!= NULL
) {
2628 instr
->src
[src_number
].src
=
2629 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.lod
));
2630 instr
->src
[src_number
].src_type
= nir_tex_src_lod
;
2636 instr
->src
[src_number
].src
=
2637 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdx
));
2638 instr
->src
[src_number
].src_type
= nir_tex_src_ddx
;
2640 instr
->src
[src_number
].src
=
2641 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdy
));
2642 instr
->src
[src_number
].src_type
= nir_tex_src_ddy
;
2647 instr
->src
[src_number
].src
=
2648 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.sample_index
));
2649 instr
->src
[src_number
].src_type
= nir_tex_src_ms_index
;
2654 instr
->component
= ir
->lod_info
.component
->as_constant()->value
.u
[0];
2661 assert(src_number
== num_srcs
);
2663 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2664 add_instr(&instr
->instr
, nir_tex_instr_dest_size(instr
), bit_size
);
2668 nir_visitor::visit(ir_constant
*ir
)
2671 * We don't know if this variable is an array or struct that gets
2672 * dereferenced, so do the safe thing an make it a variable with a
2673 * constant initializer and return a dereference.
2677 nir_local_variable_create(this->impl
, ir
->type
, "const_temp");
2678 var
->data
.read_only
= true;
2679 var
->constant_initializer
= constant_copy(ir
, var
);
2681 this->deref
= nir_build_deref_var(&b
, var
);
2685 nir_visitor::visit(ir_dereference_variable
*ir
)
2687 if (ir
->variable_referenced()->data
.mode
== ir_var_function_out
) {
2688 unsigned i
= (sig
->return_type
!= glsl_type::void_type
) ? 1 : 0;
2690 foreach_in_list(ir_variable
, param
, &sig
->parameters
) {
2691 if (param
== ir
->variable_referenced()) {
2697 this->deref
= nir_build_deref_cast(&b
, nir_load_param(&b
, i
),
2698 nir_var_function_temp
, ir
->type
, 0);
2702 assert(ir
->variable_referenced()->data
.mode
!= ir_var_function_inout
);
2704 struct hash_entry
*entry
=
2705 _mesa_hash_table_search(this->var_table
, ir
->var
);
2707 nir_variable
*var
= (nir_variable
*) entry
->data
;
2709 this->deref
= nir_build_deref_var(&b
, var
);
2713 nir_visitor::visit(ir_dereference_record
*ir
)
2715 ir
->record
->accept(this);
2717 int field_index
= ir
->field_idx
;
2718 assert(field_index
>= 0);
2720 this->deref
= nir_build_deref_struct(&b
, this->deref
, field_index
);
2724 nir_visitor::visit(ir_dereference_array
*ir
)
2726 nir_ssa_def
*index
= evaluate_rvalue(ir
->array_index
);
2728 ir
->array
->accept(this);
2730 this->deref
= nir_build_deref_array(&b
, this->deref
, index
);
2734 nir_visitor::visit(ir_barrier
*)
2736 if (shader
->info
.stage
== MESA_SHADER_COMPUTE
) {
2737 nir_intrinsic_instr
*shared_barrier
=
2738 nir_intrinsic_instr_create(this->shader
,
2739 nir_intrinsic_memory_barrier_shared
);
2740 nir_builder_instr_insert(&b
, &shared_barrier
->instr
);
2741 } else if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
2742 nir_intrinsic_instr
*patch_barrier
=
2743 nir_intrinsic_instr_create(this->shader
,
2744 nir_intrinsic_memory_barrier_tcs_patch
);
2745 nir_builder_instr_insert(&b
, &patch_barrier
->instr
);
2748 nir_intrinsic_instr
*instr
=
2749 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_control_barrier
);
2750 nir_builder_instr_insert(&b
, &instr
->instr
);
2754 glsl_float64_funcs_to_nir(struct gl_context
*ctx
,
2755 const nir_shader_compiler_options
*options
)
2757 /* We pretend it's a vertex shader. Ultimately, the stage shouldn't
2758 * matter because we're not optimizing anything here.
2760 struct gl_shader
*sh
= _mesa_new_shader(-1, MESA_SHADER_VERTEX
);
2761 sh
->Source
= float64_source
;
2762 sh
->CompileStatus
= COMPILE_FAILURE
;
2763 _mesa_glsl_compile_shader(ctx
, sh
, false, false, true);
2765 if (!sh
->CompileStatus
) {
2768 "fp64 software impl compile failed:\n%s\nsource:\n%s\n",
2769 sh
->InfoLog
, float64_source
);
2774 nir_shader
*nir
= nir_shader_create(NULL
, MESA_SHADER_VERTEX
, options
, NULL
);
2776 nir_visitor
v1(ctx
, nir
);
2777 nir_function_visitor
v2(&v1
);
2779 visit_exec_list(sh
->ir
, &v1
);
2781 /* _mesa_delete_shader will try to free sh->Source but it's static const */
2783 _mesa_delete_shader(ctx
, sh
);
2785 nir_validate_shader(nir
, "float64_funcs_to_nir");
2787 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_function_temp
);
2788 NIR_PASS_V(nir
, nir_lower_returns
);
2789 NIR_PASS_V(nir
, nir_inline_functions
);
2790 NIR_PASS_V(nir
, nir_opt_deref
);
2792 /* Do some optimizations to clean up the shader now. By optimizing the
2793 * functions in the library, we avoid having to re-do that work every
2794 * time we inline a copy of a function. Reducing basic blocks also helps
2795 * with compile times.
2797 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2798 NIR_PASS_V(nir
, nir_copy_prop
);
2799 NIR_PASS_V(nir
, nir_opt_dce
);
2800 NIR_PASS_V(nir
, nir_opt_cse
);
2801 NIR_PASS_V(nir
, nir_opt_gcm
, true);
2802 NIR_PASS_V(nir
, nir_opt_peephole_select
, 1, false, false);
2803 NIR_PASS_V(nir
, nir_opt_dce
);