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 "util/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_variable_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_FLOAT16
:
320 case GLSL_TYPE_DOUBLE
:
322 ret
->elements
= ralloc_array(mem_ctx
, nir_constant
*, cols
);
323 ret
->num_elements
= cols
;
324 for (unsigned c
= 0; c
< cols
; c
++) {
325 nir_constant
*col_const
= rzalloc(mem_ctx
, nir_constant
);
326 col_const
->num_elements
= 0;
327 switch (ir
->type
->base_type
) {
328 case GLSL_TYPE_FLOAT
:
329 for (unsigned r
= 0; r
< rows
; r
++)
330 col_const
->values
[r
].f32
= ir
->value
.f
[c
* rows
+ r
];
333 case GLSL_TYPE_FLOAT16
:
334 for (unsigned r
= 0; r
< rows
; r
++)
335 col_const
->values
[r
].u16
= ir
->value
.f16
[c
* rows
+ r
];
338 case GLSL_TYPE_DOUBLE
:
339 for (unsigned r
= 0; r
< rows
; r
++)
340 col_const
->values
[r
].f64
= ir
->value
.d
[c
* rows
+ r
];
344 unreachable("Cannot get here from the first level switch");
346 ret
->elements
[c
] = col_const
;
349 switch (ir
->type
->base_type
) {
350 case GLSL_TYPE_FLOAT
:
351 for (unsigned r
= 0; r
< rows
; r
++)
352 ret
->values
[r
].f32
= ir
->value
.f
[r
];
355 case GLSL_TYPE_FLOAT16
:
356 for (unsigned r
= 0; r
< rows
; r
++)
357 ret
->values
[r
].u16
= ir
->value
.f16
[r
];
360 case GLSL_TYPE_DOUBLE
:
361 for (unsigned r
= 0; r
< rows
; r
++)
362 ret
->values
[r
].f64
= ir
->value
.d
[r
];
366 unreachable("Cannot get here from the first level switch");
371 case GLSL_TYPE_UINT64
:
372 /* Only float base types can be matrices. */
375 for (unsigned r
= 0; r
< rows
; r
++)
376 ret
->values
[r
].u64
= ir
->value
.u64
[r
];
379 case GLSL_TYPE_INT64
:
380 /* Only float base types can be matrices. */
383 for (unsigned r
= 0; r
< rows
; r
++)
384 ret
->values
[r
].i64
= ir
->value
.i64
[r
];
388 /* Only float base types can be matrices. */
391 for (unsigned r
= 0; r
< rows
; r
++)
392 ret
->values
[r
].b
= ir
->value
.b
[r
];
396 case GLSL_TYPE_STRUCT
:
397 case GLSL_TYPE_ARRAY
:
398 ret
->elements
= ralloc_array(mem_ctx
, nir_constant
*,
400 ret
->num_elements
= ir
->type
->length
;
402 for (i
= 0; i
< ir
->type
->length
; i
++)
403 ret
->elements
[i
] = constant_copy(ir
->const_elements
[i
], mem_ctx
);
407 unreachable("not reached");
413 static const glsl_type
*
414 wrap_type_in_array(const glsl_type
*elem_type
, const glsl_type
*array_type
)
416 if (!array_type
->is_array())
419 elem_type
= wrap_type_in_array(elem_type
, array_type
->fields
.array
);
421 return glsl_type::get_array_instance(elem_type
, array_type
->length
);
425 get_nir_how_declared(unsigned how_declared
)
427 if (how_declared
== ir_var_hidden
)
428 return nir_var_hidden
;
430 return nir_var_declared_normally
;
434 nir_visitor::visit(ir_variable
*ir
)
436 /* TODO: In future we should switch to using the NIR lowering pass but for
437 * now just ignore these variables as GLSL IR should have lowered them.
438 * Anything remaining are just dead vars that weren't cleaned up.
440 if (ir
->data
.mode
== ir_var_shader_shared
)
443 /* FINISHME: inout parameters */
444 assert(ir
->data
.mode
!= ir_var_function_inout
);
446 if (ir
->data
.mode
== ir_var_function_out
)
449 nir_variable
*var
= rzalloc(shader
, nir_variable
);
450 var
->type
= ir
->type
;
451 var
->name
= ralloc_strdup(var
, ir
->name
);
453 var
->data
.always_active_io
= ir
->data
.always_active_io
;
454 var
->data
.read_only
= ir
->data
.read_only
;
455 var
->data
.centroid
= ir
->data
.centroid
;
456 var
->data
.sample
= ir
->data
.sample
;
457 var
->data
.patch
= ir
->data
.patch
;
458 var
->data
.how_declared
= get_nir_how_declared(ir
->data
.how_declared
);
459 var
->data
.invariant
= ir
->data
.invariant
;
460 var
->data
.location
= ir
->data
.location
;
461 var
->data
.stream
= ir
->data
.stream
;
462 if (ir
->data
.stream
& (1u << 31))
463 var
->data
.stream
|= NIR_STREAM_PACKED
;
465 var
->data
.precision
= ir
->data
.precision
;
466 var
->data
.explicit_location
= ir
->data
.explicit_location
;
467 var
->data
.matrix_layout
= ir
->data
.matrix_layout
;
468 var
->data
.from_named_ifc_block
= ir
->data
.from_named_ifc_block
;
469 var
->data
.compact
= false;
471 switch(ir
->data
.mode
) {
473 case ir_var_temporary
:
475 var
->data
.mode
= nir_var_shader_temp
;
477 var
->data
.mode
= nir_var_function_temp
;
480 case ir_var_function_in
:
481 case ir_var_const_in
:
482 var
->data
.mode
= nir_var_function_temp
;
485 case ir_var_shader_in
:
486 if (shader
->info
.stage
== MESA_SHADER_GEOMETRY
&&
487 ir
->data
.location
== VARYING_SLOT_PRIMITIVE_ID
) {
488 /* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */
489 var
->data
.location
= SYSTEM_VALUE_PRIMITIVE_ID
;
490 var
->data
.mode
= nir_var_system_value
;
492 var
->data
.mode
= nir_var_shader_in
;
494 if (shader
->info
.stage
== MESA_SHADER_TESS_EVAL
&&
495 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
496 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
497 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
500 if (shader
->info
.stage
> MESA_SHADER_VERTEX
&&
501 ir
->data
.location
>= VARYING_SLOT_CLIP_DIST0
&&
502 ir
->data
.location
<= VARYING_SLOT_CULL_DIST1
) {
503 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
508 case ir_var_shader_out
:
509 var
->data
.mode
= nir_var_shader_out
;
510 if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
&&
511 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
512 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
513 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
516 if (shader
->info
.stage
<= MESA_SHADER_GEOMETRY
&&
517 ir
->data
.location
>= VARYING_SLOT_CLIP_DIST0
&&
518 ir
->data
.location
<= VARYING_SLOT_CULL_DIST1
) {
519 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
524 if (ir
->get_interface_type())
525 var
->data
.mode
= nir_var_mem_ubo
;
527 var
->data
.mode
= nir_var_uniform
;
530 case ir_var_shader_storage
:
531 var
->data
.mode
= nir_var_mem_ssbo
;
534 case ir_var_system_value
:
535 var
->data
.mode
= nir_var_system_value
;
539 unreachable("not reached");
542 unsigned mem_access
= 0;
543 if (ir
->data
.memory_read_only
)
544 mem_access
|= ACCESS_NON_WRITEABLE
;
545 if (ir
->data
.memory_write_only
)
546 mem_access
|= ACCESS_NON_READABLE
;
547 if (ir
->data
.memory_coherent
)
548 mem_access
|= ACCESS_COHERENT
;
549 if (ir
->data
.memory_volatile
)
550 mem_access
|= ACCESS_VOLATILE
;
551 if (ir
->data
.memory_restrict
)
552 mem_access
|= ACCESS_RESTRICT
;
554 var
->interface_type
= ir
->get_interface_type();
556 /* For UBO and SSBO variables, we need explicit types */
557 if (var
->data
.mode
& (nir_var_mem_ubo
| nir_var_mem_ssbo
)) {
558 const glsl_type
*explicit_ifc_type
=
559 ir
->get_interface_type()->get_explicit_interface_type(supports_std430
);
561 var
->interface_type
= explicit_ifc_type
;
563 if (ir
->type
->without_array()->is_interface()) {
564 /* If the type contains the interface, wrap the explicit type in the
565 * right number of arrays.
567 var
->type
= wrap_type_in_array(explicit_ifc_type
, ir
->type
);
569 /* Otherwise, this variable is one entry in the interface */
570 UNUSED
bool found
= false;
571 for (unsigned i
= 0; i
< explicit_ifc_type
->length
; i
++) {
572 const glsl_struct_field
*field
=
573 &explicit_ifc_type
->fields
.structure
[i
];
574 if (strcmp(ir
->name
, field
->name
) != 0)
577 var
->type
= field
->type
;
578 if (field
->memory_read_only
)
579 mem_access
|= ACCESS_NON_WRITEABLE
;
580 if (field
->memory_write_only
)
581 mem_access
|= ACCESS_NON_READABLE
;
582 if (field
->memory_coherent
)
583 mem_access
|= ACCESS_COHERENT
;
584 if (field
->memory_volatile
)
585 mem_access
|= ACCESS_VOLATILE
;
586 if (field
->memory_restrict
)
587 mem_access
|= ACCESS_RESTRICT
;
596 var
->data
.interpolation
= ir
->data
.interpolation
;
597 var
->data
.location_frac
= ir
->data
.location_frac
;
599 switch (ir
->data
.depth_layout
) {
600 case ir_depth_layout_none
:
601 var
->data
.depth_layout
= nir_depth_layout_none
;
603 case ir_depth_layout_any
:
604 var
->data
.depth_layout
= nir_depth_layout_any
;
606 case ir_depth_layout_greater
:
607 var
->data
.depth_layout
= nir_depth_layout_greater
;
609 case ir_depth_layout_less
:
610 var
->data
.depth_layout
= nir_depth_layout_less
;
612 case ir_depth_layout_unchanged
:
613 var
->data
.depth_layout
= nir_depth_layout_unchanged
;
616 unreachable("not reached");
619 var
->data
.index
= ir
->data
.index
;
620 var
->data
.descriptor_set
= 0;
621 var
->data
.binding
= ir
->data
.binding
;
622 var
->data
.explicit_binding
= ir
->data
.explicit_binding
;
623 var
->data
.bindless
= ir
->data
.bindless
;
624 var
->data
.offset
= ir
->data
.offset
;
625 var
->data
.access
= (gl_access_qualifier
)mem_access
;
627 if (var
->type
->without_array()->is_image()) {
628 var
->data
.image
.format
= ir
->data
.image_format
;
629 } else if (var
->data
.mode
== nir_var_shader_out
) {
630 var
->data
.xfb
.buffer
= ir
->data
.xfb_buffer
;
631 var
->data
.xfb
.stride
= ir
->data
.xfb_stride
;
634 var
->data
.fb_fetch_output
= ir
->data
.fb_fetch_output
;
635 var
->data
.explicit_xfb_buffer
= ir
->data
.explicit_xfb_buffer
;
636 var
->data
.explicit_xfb_stride
= ir
->data
.explicit_xfb_stride
;
638 var
->num_state_slots
= ir
->get_num_state_slots();
639 if (var
->num_state_slots
> 0) {
640 var
->state_slots
= rzalloc_array(var
, nir_state_slot
,
641 var
->num_state_slots
);
643 ir_state_slot
*state_slots
= ir
->get_state_slots();
644 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
645 for (unsigned j
= 0; j
< 5; j
++)
646 var
->state_slots
[i
].tokens
[j
] = state_slots
[i
].tokens
[j
];
647 var
->state_slots
[i
].swizzle
= state_slots
[i
].swizzle
;
650 var
->state_slots
= NULL
;
653 var
->constant_initializer
= constant_copy(ir
->constant_initializer
, var
);
655 if (var
->data
.mode
== nir_var_function_temp
)
656 nir_function_impl_add_variable(impl
, var
);
658 nir_shader_add_variable(shader
, var
);
660 _mesa_hash_table_insert(var_table
, ir
, var
);
664 nir_function_visitor::visit_enter(ir_function
*ir
)
666 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
) {
667 visitor
->create_function(sig
);
669 return visit_continue_with_parent
;
673 nir_visitor::create_function(ir_function_signature
*ir
)
675 if (ir
->is_intrinsic())
678 nir_function
*func
= nir_function_create(shader
, ir
->function_name());
679 if (strcmp(ir
->function_name(), "main") == 0)
680 func
->is_entrypoint
= true;
682 func
->num_params
= ir
->parameters
.length() +
683 (ir
->return_type
!= glsl_type::void_type
);
684 func
->params
= ralloc_array(shader
, nir_parameter
, func
->num_params
);
688 if (ir
->return_type
!= glsl_type::void_type
) {
689 /* The return value is a variable deref (basically an out parameter) */
690 func
->params
[np
].num_components
= 1;
691 func
->params
[np
].bit_size
= 32;
695 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
696 /* FINISHME: pass arrays, structs, etc by reference? */
697 assert(param
->type
->is_vector() || param
->type
->is_scalar());
699 if (param
->data
.mode
== ir_var_function_in
) {
700 func
->params
[np
].num_components
= param
->type
->vector_elements
;
701 func
->params
[np
].bit_size
= glsl_get_bit_size(param
->type
);
703 func
->params
[np
].num_components
= 1;
704 func
->params
[np
].bit_size
= 32;
708 assert(np
== func
->num_params
);
710 _mesa_hash_table_insert(this->overload_table
, ir
, func
);
714 nir_visitor::visit(ir_function
*ir
)
716 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
)
721 nir_visitor::visit(ir_function_signature
*ir
)
723 if (ir
->is_intrinsic())
728 struct hash_entry
*entry
=
729 _mesa_hash_table_search(this->overload_table
, ir
);
732 nir_function
*func
= (nir_function
*) entry
->data
;
734 if (ir
->is_defined
) {
735 nir_function_impl
*impl
= nir_function_impl_create(func
);
738 this->is_global
= false;
740 nir_builder_init(&b
, impl
);
741 b
.cursor
= nir_after_cf_list(&impl
->body
);
743 unsigned i
= (ir
->return_type
!= glsl_type::void_type
) ? 1 : 0;
745 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
747 nir_local_variable_create(impl
, param
->type
, param
->name
);
749 if (param
->data
.mode
== ir_var_function_in
) {
750 nir_store_var(&b
, var
, nir_load_param(&b
, i
), ~0);
753 _mesa_hash_table_insert(var_table
, param
, var
);
757 visit_exec_list(&ir
->body
, this);
759 this->is_global
= true;
766 nir_visitor::visit(ir_loop
*ir
)
769 visit_exec_list(&ir
->body_instructions
, this);
770 nir_pop_loop(&b
, NULL
);
774 nir_visitor::visit(ir_if
*ir
)
776 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
777 visit_exec_list(&ir
->then_instructions
, this);
778 nir_push_else(&b
, NULL
);
779 visit_exec_list(&ir
->else_instructions
, this);
780 nir_pop_if(&b
, NULL
);
784 nir_visitor::visit(ir_discard
*ir
)
787 * discards aren't treated as control flow, because before we lower them
788 * they can appear anywhere in the shader and the stuff after them may still
789 * be executed (yay, crazy GLSL rules!). However, after lowering, all the
790 * discards will be immediately followed by a return.
793 nir_intrinsic_instr
*discard
;
795 discard
= nir_intrinsic_instr_create(this->shader
,
796 nir_intrinsic_discard_if
);
798 nir_src_for_ssa(evaluate_rvalue(ir
->condition
));
800 discard
= nir_intrinsic_instr_create(this->shader
, nir_intrinsic_discard
);
803 nir_builder_instr_insert(&b
, &discard
->instr
);
807 nir_visitor::visit(ir_demote
*ir
)
809 nir_intrinsic_instr
*demote
=
810 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_demote
);
812 nir_builder_instr_insert(&b
, &demote
->instr
);
816 nir_visitor::visit(ir_emit_vertex
*ir
)
818 nir_intrinsic_instr
*instr
=
819 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_emit_vertex
);
820 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
821 nir_builder_instr_insert(&b
, &instr
->instr
);
825 nir_visitor::visit(ir_end_primitive
*ir
)
827 nir_intrinsic_instr
*instr
=
828 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_end_primitive
);
829 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
830 nir_builder_instr_insert(&b
, &instr
->instr
);
834 nir_visitor::visit(ir_loop_jump
*ir
)
838 case ir_loop_jump::jump_break
:
839 type
= nir_jump_break
;
841 case ir_loop_jump::jump_continue
:
842 type
= nir_jump_continue
;
845 unreachable("not reached");
848 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, type
);
849 nir_builder_instr_insert(&b
, &instr
->instr
);
853 nir_visitor::visit(ir_return
*ir
)
855 if (ir
->value
!= NULL
) {
856 nir_deref_instr
*ret_deref
=
857 nir_build_deref_cast(&b
, nir_load_param(&b
, 0),
858 nir_var_function_temp
, ir
->value
->type
, 0);
860 nir_ssa_def
*val
= evaluate_rvalue(ir
->value
);
861 nir_store_deref(&b
, ret_deref
, val
, ~0);
864 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, nir_jump_return
);
865 nir_builder_instr_insert(&b
, &instr
->instr
);
869 intrinsic_set_std430_align(nir_intrinsic_instr
*intrin
, const glsl_type
*type
)
871 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
872 unsigned pow2_components
= util_next_power_of_two(type
->vector_elements
);
873 nir_intrinsic_set_align(intrin
, (bit_size
/ 8) * pow2_components
, 0);
876 /* Accumulate any qualifiers along the deref chain to get the actual
877 * load/store qualifier.
880 static enum gl_access_qualifier
881 deref_get_qualifier(nir_deref_instr
*deref
)
884 nir_deref_path_init(&path
, deref
, NULL
);
886 unsigned qualifiers
= path
.path
[0]->var
->data
.access
;
888 const glsl_type
*parent_type
= path
.path
[0]->type
;
889 for (nir_deref_instr
**cur_ptr
= &path
.path
[1]; *cur_ptr
; cur_ptr
++) {
890 nir_deref_instr
*cur
= *cur_ptr
;
892 if (parent_type
->is_interface()) {
893 const struct glsl_struct_field
*field
=
894 &parent_type
->fields
.structure
[cur
->strct
.index
];
895 if (field
->memory_read_only
)
896 qualifiers
|= ACCESS_NON_WRITEABLE
;
897 if (field
->memory_write_only
)
898 qualifiers
|= ACCESS_NON_READABLE
;
899 if (field
->memory_coherent
)
900 qualifiers
|= ACCESS_COHERENT
;
901 if (field
->memory_volatile
)
902 qualifiers
|= ACCESS_VOLATILE
;
903 if (field
->memory_restrict
)
904 qualifiers
|= ACCESS_RESTRICT
;
907 parent_type
= cur
->type
;
910 nir_deref_path_finish(&path
);
912 return (gl_access_qualifier
) qualifiers
;
916 nir_visitor::visit(ir_call
*ir
)
918 if (ir
->callee
->is_intrinsic()) {
921 switch (ir
->callee
->intrinsic_id
) {
922 case ir_intrinsic_generic_atomic_add
:
923 op
= ir
->return_deref
->type
->is_integer_32_64()
924 ? nir_intrinsic_deref_atomic_add
: nir_intrinsic_deref_atomic_fadd
;
926 case ir_intrinsic_generic_atomic_and
:
927 op
= nir_intrinsic_deref_atomic_and
;
929 case ir_intrinsic_generic_atomic_or
:
930 op
= nir_intrinsic_deref_atomic_or
;
932 case ir_intrinsic_generic_atomic_xor
:
933 op
= nir_intrinsic_deref_atomic_xor
;
935 case ir_intrinsic_generic_atomic_min
:
936 assert(ir
->return_deref
);
937 if (ir
->return_deref
->type
== glsl_type::int_type
)
938 op
= nir_intrinsic_deref_atomic_imin
;
939 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
940 op
= nir_intrinsic_deref_atomic_umin
;
941 else if (ir
->return_deref
->type
== glsl_type::float_type
)
942 op
= nir_intrinsic_deref_atomic_fmin
;
944 unreachable("Invalid type");
946 case ir_intrinsic_generic_atomic_max
:
947 assert(ir
->return_deref
);
948 if (ir
->return_deref
->type
== glsl_type::int_type
)
949 op
= nir_intrinsic_deref_atomic_imax
;
950 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
951 op
= nir_intrinsic_deref_atomic_umax
;
952 else if (ir
->return_deref
->type
== glsl_type::float_type
)
953 op
= nir_intrinsic_deref_atomic_fmax
;
955 unreachable("Invalid type");
957 case ir_intrinsic_generic_atomic_exchange
:
958 op
= nir_intrinsic_deref_atomic_exchange
;
960 case ir_intrinsic_generic_atomic_comp_swap
:
961 op
= ir
->return_deref
->type
->is_integer_32_64()
962 ? nir_intrinsic_deref_atomic_comp_swap
963 : nir_intrinsic_deref_atomic_fcomp_swap
;
965 case ir_intrinsic_atomic_counter_read
:
966 op
= nir_intrinsic_atomic_counter_read_deref
;
968 case ir_intrinsic_atomic_counter_increment
:
969 op
= nir_intrinsic_atomic_counter_inc_deref
;
971 case ir_intrinsic_atomic_counter_predecrement
:
972 op
= nir_intrinsic_atomic_counter_pre_dec_deref
;
974 case ir_intrinsic_atomic_counter_add
:
975 op
= nir_intrinsic_atomic_counter_add_deref
;
977 case ir_intrinsic_atomic_counter_and
:
978 op
= nir_intrinsic_atomic_counter_and_deref
;
980 case ir_intrinsic_atomic_counter_or
:
981 op
= nir_intrinsic_atomic_counter_or_deref
;
983 case ir_intrinsic_atomic_counter_xor
:
984 op
= nir_intrinsic_atomic_counter_xor_deref
;
986 case ir_intrinsic_atomic_counter_min
:
987 op
= nir_intrinsic_atomic_counter_min_deref
;
989 case ir_intrinsic_atomic_counter_max
:
990 op
= nir_intrinsic_atomic_counter_max_deref
;
992 case ir_intrinsic_atomic_counter_exchange
:
993 op
= nir_intrinsic_atomic_counter_exchange_deref
;
995 case ir_intrinsic_atomic_counter_comp_swap
:
996 op
= nir_intrinsic_atomic_counter_comp_swap_deref
;
998 case ir_intrinsic_image_load
:
999 op
= nir_intrinsic_image_deref_load
;
1001 case ir_intrinsic_image_store
:
1002 op
= nir_intrinsic_image_deref_store
;
1004 case ir_intrinsic_image_atomic_add
:
1005 op
= ir
->return_deref
->type
->is_integer_32_64()
1006 ? nir_intrinsic_image_deref_atomic_add
1007 : nir_intrinsic_image_deref_atomic_fadd
;
1009 case ir_intrinsic_image_atomic_min
:
1010 if (ir
->return_deref
->type
== glsl_type::int_type
)
1011 op
= nir_intrinsic_image_deref_atomic_imin
;
1012 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1013 op
= nir_intrinsic_image_deref_atomic_umin
;
1015 unreachable("Invalid type");
1017 case ir_intrinsic_image_atomic_max
:
1018 if (ir
->return_deref
->type
== glsl_type::int_type
)
1019 op
= nir_intrinsic_image_deref_atomic_imax
;
1020 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1021 op
= nir_intrinsic_image_deref_atomic_umax
;
1023 unreachable("Invalid type");
1025 case ir_intrinsic_image_atomic_and
:
1026 op
= nir_intrinsic_image_deref_atomic_and
;
1028 case ir_intrinsic_image_atomic_or
:
1029 op
= nir_intrinsic_image_deref_atomic_or
;
1031 case ir_intrinsic_image_atomic_xor
:
1032 op
= nir_intrinsic_image_deref_atomic_xor
;
1034 case ir_intrinsic_image_atomic_exchange
:
1035 op
= nir_intrinsic_image_deref_atomic_exchange
;
1037 case ir_intrinsic_image_atomic_comp_swap
:
1038 op
= nir_intrinsic_image_deref_atomic_comp_swap
;
1040 case ir_intrinsic_image_atomic_inc_wrap
:
1041 op
= nir_intrinsic_image_deref_atomic_inc_wrap
;
1043 case ir_intrinsic_image_atomic_dec_wrap
:
1044 op
= nir_intrinsic_image_deref_atomic_dec_wrap
;
1046 case ir_intrinsic_memory_barrier
:
1047 op
= nir_intrinsic_memory_barrier
;
1049 case ir_intrinsic_image_size
:
1050 op
= nir_intrinsic_image_deref_size
;
1052 case ir_intrinsic_image_samples
:
1053 op
= nir_intrinsic_image_deref_samples
;
1055 case ir_intrinsic_ssbo_store
:
1056 case ir_intrinsic_ssbo_load
:
1057 case ir_intrinsic_ssbo_atomic_add
:
1058 case ir_intrinsic_ssbo_atomic_and
:
1059 case ir_intrinsic_ssbo_atomic_or
:
1060 case ir_intrinsic_ssbo_atomic_xor
:
1061 case ir_intrinsic_ssbo_atomic_min
:
1062 case ir_intrinsic_ssbo_atomic_max
:
1063 case ir_intrinsic_ssbo_atomic_exchange
:
1064 case ir_intrinsic_ssbo_atomic_comp_swap
:
1065 /* SSBO store/loads should only have been lowered in GLSL IR for
1066 * non-nir drivers, NIR drivers make use of gl_nir_lower_buffers()
1069 unreachable("Invalid operation nir doesn't want lowered ssbo "
1071 case ir_intrinsic_shader_clock
:
1072 op
= nir_intrinsic_shader_clock
;
1074 case ir_intrinsic_begin_invocation_interlock
:
1075 op
= nir_intrinsic_begin_invocation_interlock
;
1077 case ir_intrinsic_end_invocation_interlock
:
1078 op
= nir_intrinsic_end_invocation_interlock
;
1080 case ir_intrinsic_group_memory_barrier
:
1081 op
= nir_intrinsic_group_memory_barrier
;
1083 case ir_intrinsic_memory_barrier_atomic_counter
:
1084 op
= nir_intrinsic_memory_barrier_atomic_counter
;
1086 case ir_intrinsic_memory_barrier_buffer
:
1087 op
= nir_intrinsic_memory_barrier_buffer
;
1089 case ir_intrinsic_memory_barrier_image
:
1090 op
= nir_intrinsic_memory_barrier_image
;
1092 case ir_intrinsic_memory_barrier_shared
:
1093 op
= nir_intrinsic_memory_barrier_shared
;
1095 case ir_intrinsic_shared_load
:
1096 op
= nir_intrinsic_load_shared
;
1098 case ir_intrinsic_shared_store
:
1099 op
= nir_intrinsic_store_shared
;
1101 case ir_intrinsic_shared_atomic_add
:
1102 op
= ir
->return_deref
->type
->is_integer_32_64()
1103 ? nir_intrinsic_shared_atomic_add
1104 : nir_intrinsic_shared_atomic_fadd
;
1106 case ir_intrinsic_shared_atomic_and
:
1107 op
= nir_intrinsic_shared_atomic_and
;
1109 case ir_intrinsic_shared_atomic_or
:
1110 op
= nir_intrinsic_shared_atomic_or
;
1112 case ir_intrinsic_shared_atomic_xor
:
1113 op
= nir_intrinsic_shared_atomic_xor
;
1115 case ir_intrinsic_shared_atomic_min
:
1116 assert(ir
->return_deref
);
1117 if (ir
->return_deref
->type
== glsl_type::int_type
)
1118 op
= nir_intrinsic_shared_atomic_imin
;
1119 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1120 op
= nir_intrinsic_shared_atomic_umin
;
1121 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1122 op
= nir_intrinsic_shared_atomic_fmin
;
1124 unreachable("Invalid type");
1126 case ir_intrinsic_shared_atomic_max
:
1127 assert(ir
->return_deref
);
1128 if (ir
->return_deref
->type
== glsl_type::int_type
)
1129 op
= nir_intrinsic_shared_atomic_imax
;
1130 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1131 op
= nir_intrinsic_shared_atomic_umax
;
1132 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1133 op
= nir_intrinsic_shared_atomic_fmax
;
1135 unreachable("Invalid type");
1137 case ir_intrinsic_shared_atomic_exchange
:
1138 op
= nir_intrinsic_shared_atomic_exchange
;
1140 case ir_intrinsic_shared_atomic_comp_swap
:
1141 op
= ir
->return_deref
->type
->is_integer_32_64()
1142 ? nir_intrinsic_shared_atomic_comp_swap
1143 : nir_intrinsic_shared_atomic_fcomp_swap
;
1145 case ir_intrinsic_vote_any
:
1146 op
= nir_intrinsic_vote_any
;
1148 case ir_intrinsic_vote_all
:
1149 op
= nir_intrinsic_vote_all
;
1151 case ir_intrinsic_vote_eq
:
1152 op
= nir_intrinsic_vote_ieq
;
1154 case ir_intrinsic_ballot
:
1155 op
= nir_intrinsic_ballot
;
1157 case ir_intrinsic_read_invocation
:
1158 op
= nir_intrinsic_read_invocation
;
1160 case ir_intrinsic_read_first_invocation
:
1161 op
= nir_intrinsic_read_first_invocation
;
1163 case ir_intrinsic_helper_invocation
:
1164 op
= nir_intrinsic_is_helper_invocation
;
1167 unreachable("not reached");
1170 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(shader
, op
);
1171 nir_ssa_def
*ret
= &instr
->dest
.ssa
;
1174 case nir_intrinsic_deref_atomic_add
:
1175 case nir_intrinsic_deref_atomic_imin
:
1176 case nir_intrinsic_deref_atomic_umin
:
1177 case nir_intrinsic_deref_atomic_imax
:
1178 case nir_intrinsic_deref_atomic_umax
:
1179 case nir_intrinsic_deref_atomic_and
:
1180 case nir_intrinsic_deref_atomic_or
:
1181 case nir_intrinsic_deref_atomic_xor
:
1182 case nir_intrinsic_deref_atomic_exchange
:
1183 case nir_intrinsic_deref_atomic_comp_swap
:
1184 case nir_intrinsic_deref_atomic_fadd
:
1185 case nir_intrinsic_deref_atomic_fmin
:
1186 case nir_intrinsic_deref_atomic_fmax
:
1187 case nir_intrinsic_deref_atomic_fcomp_swap
: {
1188 int param_count
= ir
->actual_parameters
.length();
1189 assert(param_count
== 2 || param_count
== 3);
1192 exec_node
*param
= ir
->actual_parameters
.get_head();
1193 ir_rvalue
*rvalue
= (ir_rvalue
*) param
;
1194 ir_dereference
*deref
= rvalue
->as_dereference();
1195 ir_swizzle
*swizzle
= NULL
;
1197 /* We may have a swizzle to pick off a single vec4 component */
1198 swizzle
= rvalue
->as_swizzle();
1199 assert(swizzle
&& swizzle
->type
->vector_elements
== 1);
1200 deref
= swizzle
->val
->as_dereference();
1203 nir_deref_instr
*nir_deref
= evaluate_deref(deref
);
1205 nir_deref
= nir_build_deref_array_imm(&b
, nir_deref
,
1208 instr
->src
[0] = nir_src_for_ssa(&nir_deref
->dest
.ssa
);
1210 nir_intrinsic_set_access(instr
, deref_get_qualifier(nir_deref
));
1212 /* data1 parameter (this is always present) */
1213 param
= param
->get_next();
1214 ir_instruction
*inst
= (ir_instruction
*) param
;
1215 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1217 /* data2 parameter (only with atomic_comp_swap) */
1218 if (param_count
== 3) {
1219 assert(op
== nir_intrinsic_deref_atomic_comp_swap
||
1220 op
== nir_intrinsic_deref_atomic_fcomp_swap
);
1221 param
= param
->get_next();
1222 inst
= (ir_instruction
*) param
;
1223 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1227 assert(ir
->return_deref
);
1228 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1229 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1230 nir_builder_instr_insert(&b
, &instr
->instr
);
1233 case nir_intrinsic_atomic_counter_read_deref
:
1234 case nir_intrinsic_atomic_counter_inc_deref
:
1235 case nir_intrinsic_atomic_counter_pre_dec_deref
:
1236 case nir_intrinsic_atomic_counter_add_deref
:
1237 case nir_intrinsic_atomic_counter_min_deref
:
1238 case nir_intrinsic_atomic_counter_max_deref
:
1239 case nir_intrinsic_atomic_counter_and_deref
:
1240 case nir_intrinsic_atomic_counter_or_deref
:
1241 case nir_intrinsic_atomic_counter_xor_deref
:
1242 case nir_intrinsic_atomic_counter_exchange_deref
:
1243 case nir_intrinsic_atomic_counter_comp_swap_deref
: {
1244 /* Set the counter variable dereference. */
1245 exec_node
*param
= ir
->actual_parameters
.get_head();
1246 ir_dereference
*counter
= (ir_dereference
*)param
;
1248 instr
->src
[0] = nir_src_for_ssa(&evaluate_deref(counter
)->dest
.ssa
);
1249 param
= param
->get_next();
1251 /* Set the intrinsic destination. */
1252 if (ir
->return_deref
) {
1253 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1256 /* Set the intrinsic parameters. */
1257 if (!param
->is_tail_sentinel()) {
1259 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1260 param
= param
->get_next();
1263 if (!param
->is_tail_sentinel()) {
1265 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1266 param
= param
->get_next();
1269 nir_builder_instr_insert(&b
, &instr
->instr
);
1272 case nir_intrinsic_image_deref_load
:
1273 case nir_intrinsic_image_deref_store
:
1274 case nir_intrinsic_image_deref_atomic_add
:
1275 case nir_intrinsic_image_deref_atomic_imin
:
1276 case nir_intrinsic_image_deref_atomic_umin
:
1277 case nir_intrinsic_image_deref_atomic_imax
:
1278 case nir_intrinsic_image_deref_atomic_umax
:
1279 case nir_intrinsic_image_deref_atomic_and
:
1280 case nir_intrinsic_image_deref_atomic_or
:
1281 case nir_intrinsic_image_deref_atomic_xor
:
1282 case nir_intrinsic_image_deref_atomic_exchange
:
1283 case nir_intrinsic_image_deref_atomic_comp_swap
:
1284 case nir_intrinsic_image_deref_atomic_fadd
:
1285 case nir_intrinsic_image_deref_samples
:
1286 case nir_intrinsic_image_deref_size
:
1287 case nir_intrinsic_image_deref_atomic_inc_wrap
:
1288 case nir_intrinsic_image_deref_atomic_dec_wrap
: {
1289 nir_ssa_undef_instr
*instr_undef
=
1290 nir_ssa_undef_instr_create(shader
, 1, 32);
1291 nir_builder_instr_insert(&b
, &instr_undef
->instr
);
1293 /* Set the image variable dereference. */
1294 exec_node
*param
= ir
->actual_parameters
.get_head();
1295 ir_dereference
*image
= (ir_dereference
*)param
;
1296 nir_deref_instr
*deref
= evaluate_deref(image
);
1297 const glsl_type
*type
= deref
->type
;
1299 nir_intrinsic_set_access(instr
, deref_get_qualifier(deref
));
1301 instr
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
1302 param
= param
->get_next();
1304 /* Set the intrinsic destination. */
1305 if (ir
->return_deref
) {
1306 unsigned num_components
= ir
->return_deref
->type
->vector_elements
;
1307 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1308 num_components
, 32, NULL
);
1311 if (op
== nir_intrinsic_image_deref_size
) {
1312 instr
->num_components
= instr
->dest
.ssa
.num_components
;
1313 } else if (op
== nir_intrinsic_image_deref_load
||
1314 op
== nir_intrinsic_image_deref_store
) {
1315 instr
->num_components
= 4;
1318 if (op
== nir_intrinsic_image_deref_size
||
1319 op
== nir_intrinsic_image_deref_samples
) {
1320 nir_builder_instr_insert(&b
, &instr
->instr
);
1324 /* Set the address argument, extending the coordinate vector to four
1327 nir_ssa_def
*src_addr
=
1328 evaluate_rvalue((ir_dereference
*)param
);
1329 nir_ssa_def
*srcs
[4];
1331 for (int i
= 0; i
< 4; i
++) {
1332 if (i
< type
->coordinate_components())
1333 srcs
[i
] = nir_channel(&b
, src_addr
, i
);
1335 srcs
[i
] = &instr_undef
->def
;
1338 instr
->src
[1] = nir_src_for_ssa(nir_vec(&b
, srcs
, 4));
1339 param
= param
->get_next();
1341 /* Set the sample argument, which is undefined for single-sample
1344 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_MS
) {
1346 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1347 param
= param
->get_next();
1349 instr
->src
[2] = nir_src_for_ssa(&instr_undef
->def
);
1352 /* Set the intrinsic parameters. */
1353 if (!param
->is_tail_sentinel()) {
1355 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1356 param
= param
->get_next();
1357 } else if (op
== nir_intrinsic_image_deref_load
) {
1358 instr
->src
[3] = nir_src_for_ssa(nir_imm_int(&b
, 0)); /* LOD */
1361 if (!param
->is_tail_sentinel()) {
1363 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1364 param
= param
->get_next();
1365 } else if (op
== nir_intrinsic_image_deref_store
) {
1366 instr
->src
[4] = nir_src_for_ssa(nir_imm_int(&b
, 0)); /* LOD */
1369 nir_builder_instr_insert(&b
, &instr
->instr
);
1372 case nir_intrinsic_memory_barrier
:
1373 case nir_intrinsic_group_memory_barrier
:
1374 case nir_intrinsic_memory_barrier_atomic_counter
:
1375 case nir_intrinsic_memory_barrier_buffer
:
1376 case nir_intrinsic_memory_barrier_image
:
1377 case nir_intrinsic_memory_barrier_shared
:
1378 nir_builder_instr_insert(&b
, &instr
->instr
);
1380 case nir_intrinsic_shader_clock
:
1381 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 2, 32, NULL
);
1382 instr
->num_components
= 2;
1383 nir_builder_instr_insert(&b
, &instr
->instr
);
1385 case nir_intrinsic_begin_invocation_interlock
:
1386 nir_builder_instr_insert(&b
, &instr
->instr
);
1388 case nir_intrinsic_end_invocation_interlock
:
1389 nir_builder_instr_insert(&b
, &instr
->instr
);
1391 case nir_intrinsic_store_ssbo
: {
1392 exec_node
*param
= ir
->actual_parameters
.get_head();
1393 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1395 param
= param
->get_next();
1396 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1398 param
= param
->get_next();
1399 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1401 param
= param
->get_next();
1402 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1405 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1406 if (val
->type
->is_boolean())
1407 nir_val
= nir_b2i32(&b
, nir_val
);
1409 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1410 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(block
));
1411 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(offset
));
1412 intrinsic_set_std430_align(instr
, val
->type
);
1413 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1414 instr
->num_components
= val
->type
->vector_elements
;
1416 nir_builder_instr_insert(&b
, &instr
->instr
);
1419 case nir_intrinsic_load_shared
: {
1420 exec_node
*param
= ir
->actual_parameters
.get_head();
1421 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1423 nir_intrinsic_set_base(instr
, 0);
1424 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(offset
));
1426 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1427 instr
->num_components
= type
->vector_elements
;
1428 intrinsic_set_std430_align(instr
, type
);
1430 /* Setup destination register */
1431 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1432 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1433 type
->vector_elements
, bit_size
, NULL
);
1435 nir_builder_instr_insert(&b
, &instr
->instr
);
1437 /* The value in shared memory is a 32-bit value */
1438 if (type
->is_boolean())
1439 ret
= nir_b2b1(&b
, &instr
->dest
.ssa
);
1442 case nir_intrinsic_store_shared
: {
1443 exec_node
*param
= ir
->actual_parameters
.get_head();
1444 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1446 param
= param
->get_next();
1447 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1449 param
= param
->get_next();
1450 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1453 nir_intrinsic_set_base(instr
, 0);
1454 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1456 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1458 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1459 /* The value in shared memory is a 32-bit value */
1460 if (val
->type
->is_boolean())
1461 nir_val
= nir_b2b32(&b
, nir_val
);
1463 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1464 instr
->num_components
= val
->type
->vector_elements
;
1465 intrinsic_set_std430_align(instr
, val
->type
);
1467 nir_builder_instr_insert(&b
, &instr
->instr
);
1470 case nir_intrinsic_shared_atomic_add
:
1471 case nir_intrinsic_shared_atomic_imin
:
1472 case nir_intrinsic_shared_atomic_umin
:
1473 case nir_intrinsic_shared_atomic_imax
:
1474 case nir_intrinsic_shared_atomic_umax
:
1475 case nir_intrinsic_shared_atomic_and
:
1476 case nir_intrinsic_shared_atomic_or
:
1477 case nir_intrinsic_shared_atomic_xor
:
1478 case nir_intrinsic_shared_atomic_exchange
:
1479 case nir_intrinsic_shared_atomic_comp_swap
:
1480 case nir_intrinsic_shared_atomic_fadd
:
1481 case nir_intrinsic_shared_atomic_fmin
:
1482 case nir_intrinsic_shared_atomic_fmax
:
1483 case nir_intrinsic_shared_atomic_fcomp_swap
: {
1484 int param_count
= ir
->actual_parameters
.length();
1485 assert(param_count
== 2 || param_count
== 3);
1488 exec_node
*param
= ir
->actual_parameters
.get_head();
1489 ir_instruction
*inst
= (ir_instruction
*) param
;
1490 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1492 /* data1 parameter (this is always present) */
1493 param
= param
->get_next();
1494 inst
= (ir_instruction
*) param
;
1495 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1497 /* data2 parameter (only with atomic_comp_swap) */
1498 if (param_count
== 3) {
1499 assert(op
== nir_intrinsic_shared_atomic_comp_swap
||
1500 op
== nir_intrinsic_shared_atomic_fcomp_swap
);
1501 param
= param
->get_next();
1502 inst
= (ir_instruction
*) param
;
1504 nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1508 assert(ir
->return_deref
);
1509 unsigned bit_size
= glsl_get_bit_size(ir
->return_deref
->type
);
1510 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1511 ir
->return_deref
->type
->vector_elements
,
1513 nir_builder_instr_insert(&b
, &instr
->instr
);
1516 case nir_intrinsic_vote_any
:
1517 case nir_intrinsic_vote_all
:
1518 case nir_intrinsic_vote_ieq
: {
1519 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 1, NULL
);
1520 instr
->num_components
= 1;
1522 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1523 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1525 nir_builder_instr_insert(&b
, &instr
->instr
);
1529 case nir_intrinsic_ballot
: {
1530 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1531 ir
->return_deref
->type
->vector_elements
, 64, NULL
);
1532 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1534 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1535 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1537 nir_builder_instr_insert(&b
, &instr
->instr
);
1540 case nir_intrinsic_read_invocation
: {
1541 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1542 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1543 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1545 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1546 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1548 ir_rvalue
*invocation
= (ir_rvalue
*) ir
->actual_parameters
.get_head()->next
;
1549 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(invocation
));
1551 nir_builder_instr_insert(&b
, &instr
->instr
);
1554 case nir_intrinsic_read_first_invocation
: {
1555 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1556 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1557 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1559 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1560 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1562 nir_builder_instr_insert(&b
, &instr
->instr
);
1565 case nir_intrinsic_is_helper_invocation
: {
1566 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 1, NULL
);
1567 instr
->num_components
= 1;
1568 nir_builder_instr_insert(&b
, &instr
->instr
);
1572 unreachable("not reached");
1575 if (ir
->return_deref
)
1576 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), ret
, ~0);
1581 struct hash_entry
*entry
=
1582 _mesa_hash_table_search(this->overload_table
, ir
->callee
);
1584 nir_function
*callee
= (nir_function
*) entry
->data
;
1586 nir_call_instr
*call
= nir_call_instr_create(this->shader
, callee
);
1589 nir_deref_instr
*ret_deref
= NULL
;
1590 if (ir
->return_deref
) {
1591 nir_variable
*ret_tmp
=
1592 nir_local_variable_create(this->impl
, ir
->return_deref
->type
,
1594 ret_deref
= nir_build_deref_var(&b
, ret_tmp
);
1595 call
->params
[i
++] = nir_src_for_ssa(&ret_deref
->dest
.ssa
);
1598 foreach_two_lists(formal_node
, &ir
->callee
->parameters
,
1599 actual_node
, &ir
->actual_parameters
) {
1600 ir_rvalue
*param_rvalue
= (ir_rvalue
*) actual_node
;
1601 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
1603 if (sig_param
->data
.mode
== ir_var_function_out
) {
1604 nir_deref_instr
*out_deref
= evaluate_deref(param_rvalue
);
1605 call
->params
[i
] = nir_src_for_ssa(&out_deref
->dest
.ssa
);
1606 } else if (sig_param
->data
.mode
== ir_var_function_in
) {
1607 nir_ssa_def
*val
= evaluate_rvalue(param_rvalue
);
1608 nir_src src
= nir_src_for_ssa(val
);
1610 nir_src_copy(&call
->params
[i
], &src
, call
);
1611 } else if (sig_param
->data
.mode
== ir_var_function_inout
) {
1612 unreachable("unimplemented: inout parameters");
1618 nir_builder_instr_insert(&b
, &call
->instr
);
1620 if (ir
->return_deref
)
1621 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), nir_load_deref(&b
, ret_deref
), ~0);
1625 nir_visitor::visit(ir_assignment
*ir
)
1627 unsigned num_components
= ir
->lhs
->type
->vector_elements
;
1629 b
.exact
= ir
->lhs
->variable_referenced()->data
.invariant
||
1630 ir
->lhs
->variable_referenced()->data
.precise
;
1632 if ((ir
->rhs
->as_dereference() || ir
->rhs
->as_constant()) &&
1633 (ir
->write_mask
== (1 << num_components
) - 1 || ir
->write_mask
== 0)) {
1634 nir_deref_instr
*lhs
= evaluate_deref(ir
->lhs
);
1635 nir_deref_instr
*rhs
= evaluate_deref(ir
->rhs
);
1636 enum gl_access_qualifier lhs_qualifiers
= deref_get_qualifier(lhs
);
1637 enum gl_access_qualifier rhs_qualifiers
= deref_get_qualifier(rhs
);
1638 if (ir
->condition
) {
1639 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1640 nir_copy_deref_with_access(&b
, lhs
, rhs
, lhs_qualifiers
,
1642 nir_pop_if(&b
, NULL
);
1644 nir_copy_deref_with_access(&b
, lhs
, rhs
, lhs_qualifiers
,
1650 assert(ir
->rhs
->type
->is_scalar() || ir
->rhs
->type
->is_vector());
1652 ir
->lhs
->accept(this);
1653 nir_deref_instr
*lhs_deref
= this->deref
;
1654 nir_ssa_def
*src
= evaluate_rvalue(ir
->rhs
);
1656 if (ir
->write_mask
!= (1 << num_components
) - 1 && ir
->write_mask
!= 0) {
1657 /* GLSL IR will give us the input to the write-masked assignment in a
1658 * single packed vector. So, for example, if the writemask is xzw, then
1659 * we have to swizzle x -> x, y -> z, and z -> w and get the y component
1663 unsigned component
= 0;
1664 for (unsigned i
= 0; i
< 4; i
++) {
1665 swiz
[i
] = ir
->write_mask
& (1 << i
) ? component
++ : 0;
1667 src
= nir_swizzle(&b
, src
, swiz
, num_components
);
1670 enum gl_access_qualifier qualifiers
= deref_get_qualifier(lhs_deref
);
1671 if (ir
->condition
) {
1672 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1673 nir_store_deref_with_access(&b
, lhs_deref
, src
, ir
->write_mask
,
1675 nir_pop_if(&b
, NULL
);
1677 nir_store_deref_with_access(&b
, lhs_deref
, src
, ir
->write_mask
,
1683 * Given an instruction, returns a pointer to its destination or NULL if there
1684 * is no destination.
1686 * Note that this only handles instructions we generate at this level.
1689 get_instr_dest(nir_instr
*instr
)
1691 nir_alu_instr
*alu_instr
;
1692 nir_intrinsic_instr
*intrinsic_instr
;
1693 nir_tex_instr
*tex_instr
;
1695 switch (instr
->type
) {
1696 case nir_instr_type_alu
:
1697 alu_instr
= nir_instr_as_alu(instr
);
1698 return &alu_instr
->dest
.dest
;
1700 case nir_instr_type_intrinsic
:
1701 intrinsic_instr
= nir_instr_as_intrinsic(instr
);
1702 if (nir_intrinsic_infos
[intrinsic_instr
->intrinsic
].has_dest
)
1703 return &intrinsic_instr
->dest
;
1707 case nir_instr_type_tex
:
1708 tex_instr
= nir_instr_as_tex(instr
);
1709 return &tex_instr
->dest
;
1712 unreachable("not reached");
1719 nir_visitor::add_instr(nir_instr
*instr
, unsigned num_components
,
1722 nir_dest
*dest
= get_instr_dest(instr
);
1725 nir_ssa_dest_init(instr
, dest
, num_components
, bit_size
, NULL
);
1727 nir_builder_instr_insert(&b
, instr
);
1730 assert(dest
->is_ssa
);
1731 this->result
= &dest
->ssa
;
1736 nir_visitor::evaluate_rvalue(ir_rvalue
* ir
)
1739 if (ir
->as_dereference() || ir
->as_constant()) {
1741 * A dereference is being used on the right hand side, which means we
1742 * must emit a variable load.
1745 enum gl_access_qualifier access
= deref_get_qualifier(this->deref
);
1746 this->result
= nir_load_deref_with_access(&b
, this->deref
, access
);
1749 return this->result
;
1753 type_is_float(glsl_base_type type
)
1755 return type
== GLSL_TYPE_FLOAT
|| type
== GLSL_TYPE_DOUBLE
||
1756 type
== GLSL_TYPE_FLOAT16
;
1760 type_is_signed(glsl_base_type type
)
1762 return type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_INT64
||
1763 type
== GLSL_TYPE_INT16
;
1767 nir_visitor::visit(ir_expression
*ir
)
1769 /* Some special cases */
1770 switch (ir
->operation
) {
1771 case ir_unop_interpolate_at_centroid
:
1772 case ir_binop_interpolate_at_offset
:
1773 case ir_binop_interpolate_at_sample
: {
1774 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1775 ir_swizzle
*swizzle
= NULL
;
1777 /* the api does not allow a swizzle here, but the varying packing code
1778 * may have pushed one into here.
1780 swizzle
= ir
->operands
[0]->as_swizzle();
1782 deref
= swizzle
->val
->as_dereference();
1786 deref
->accept(this);
1788 nir_intrinsic_op op
;
1789 if (this->deref
->mode
== nir_var_shader_in
) {
1790 switch (ir
->operation
) {
1791 case ir_unop_interpolate_at_centroid
:
1792 op
= nir_intrinsic_interp_deref_at_centroid
;
1794 case ir_binop_interpolate_at_offset
:
1795 op
= nir_intrinsic_interp_deref_at_offset
;
1797 case ir_binop_interpolate_at_sample
:
1798 op
= nir_intrinsic_interp_deref_at_sample
;
1801 unreachable("Invalid interpolation intrinsic");
1804 /* This case can happen if the vertex shader does not write the
1805 * given varying. In this case, the linker will lower it to a
1806 * global variable. Since interpolating a variable makes no
1807 * sense, we'll just turn it into a load which will probably
1808 * eventually end up as an SSA definition.
1810 assert(this->deref
->mode
== nir_var_shader_temp
);
1811 op
= nir_intrinsic_load_deref
;
1814 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(shader
, op
);
1815 intrin
->num_components
= deref
->type
->vector_elements
;
1816 intrin
->src
[0] = nir_src_for_ssa(&this->deref
->dest
.ssa
);
1818 if (intrin
->intrinsic
== nir_intrinsic_interp_deref_at_offset
||
1819 intrin
->intrinsic
== nir_intrinsic_interp_deref_at_sample
)
1820 intrin
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1822 unsigned bit_size
= glsl_get_bit_size(deref
->type
);
1823 add_instr(&intrin
->instr
, deref
->type
->vector_elements
, bit_size
);
1826 unsigned swiz
[4] = {
1827 swizzle
->mask
.x
, swizzle
->mask
.y
, swizzle
->mask
.z
, swizzle
->mask
.w
1830 result
= nir_swizzle(&b
, result
, swiz
,
1831 swizzle
->type
->vector_elements
);
1837 case ir_unop_ssbo_unsized_array_length
: {
1838 nir_intrinsic_instr
*intrin
=
1839 nir_intrinsic_instr_create(b
.shader
,
1840 nir_intrinsic_deref_buffer_array_length
);
1842 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1843 intrin
->src
[0] = nir_src_for_ssa(&evaluate_deref(deref
)->dest
.ssa
);
1845 add_instr(&intrin
->instr
, 1, 32);
1849 case ir_binop_ubo_load
:
1850 /* UBO loads should only have been lowered in GLSL IR for non-nir drivers,
1851 * NIR drivers make use of gl_nir_lower_buffers() instead.
1853 unreachable("Invalid operation nir doesn't want lowered ubo loads");
1858 nir_ssa_def
*srcs
[4];
1859 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1860 srcs
[i
] = evaluate_rvalue(ir
->operands
[i
]);
1862 glsl_base_type types
[4];
1863 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1864 types
[i
] = ir
->operands
[i
]->type
->base_type
;
1866 glsl_base_type out_type
= ir
->type
->base_type
;
1868 switch (ir
->operation
) {
1869 case ir_unop_bit_not
: result
= nir_inot(&b
, srcs
[0]); break;
1870 case ir_unop_logic_not
:
1871 result
= nir_inot(&b
, srcs
[0]);
1874 result
= type_is_float(types
[0]) ? nir_fneg(&b
, srcs
[0])
1875 : nir_ineg(&b
, srcs
[0]);
1878 result
= type_is_float(types
[0]) ? nir_fabs(&b
, srcs
[0])
1879 : nir_iabs(&b
, srcs
[0]);
1882 result
= nir_uclz(&b
, srcs
[0]);
1884 case ir_unop_saturate
:
1885 assert(type_is_float(types
[0]));
1886 result
= nir_fsat(&b
, srcs
[0]);
1889 result
= type_is_float(types
[0]) ? nir_fsign(&b
, srcs
[0])
1890 : nir_isign(&b
, srcs
[0]);
1892 case ir_unop_rcp
: result
= nir_frcp(&b
, srcs
[0]); break;
1893 case ir_unop_rsq
: result
= nir_frsq(&b
, srcs
[0]); break;
1894 case ir_unop_sqrt
: result
= nir_fsqrt(&b
, srcs
[0]); break;
1895 case ir_unop_exp
: unreachable("ir_unop_exp should have been lowered");
1896 case ir_unop_log
: unreachable("ir_unop_log should have been lowered");
1897 case ir_unop_exp2
: result
= nir_fexp2(&b
, srcs
[0]); break;
1898 case ir_unop_log2
: result
= nir_flog2(&b
, srcs
[0]); break;
1938 case ir_unop_i642u64
:
1939 case ir_unop_u642i64
: {
1940 nir_alu_type src_type
= nir_get_nir_type_for_glsl_base_type(types
[0]);
1941 nir_alu_type dst_type
= nir_get_nir_type_for_glsl_base_type(out_type
);
1942 result
= nir_build_alu(&b
, nir_type_conversion_op(src_type
, dst_type
,
1943 nir_rounding_mode_undef
),
1944 srcs
[0], NULL
, NULL
, NULL
);
1945 /* b2i and b2f don't have fixed bit-size versions so the builder will
1946 * just assume 32 and we have to fix it up here.
1948 result
->bit_size
= nir_alu_type_get_type_size(dst_type
);
1952 case ir_unop_f2fmp
: {
1953 result
= nir_build_alu(&b
, nir_op_f2fmp
, srcs
[0], NULL
, NULL
, NULL
);
1957 case ir_unop_bitcast_i2f
:
1958 case ir_unop_bitcast_f2i
:
1959 case ir_unop_bitcast_u2f
:
1960 case ir_unop_bitcast_f2u
:
1961 case ir_unop_bitcast_i642d
:
1962 case ir_unop_bitcast_d2i64
:
1963 case ir_unop_bitcast_u642d
:
1964 case ir_unop_bitcast_d2u64
:
1965 case ir_unop_subroutine_to_int
:
1967 result
= nir_mov(&b
, srcs
[0]);
1969 case ir_unop_trunc
: result
= nir_ftrunc(&b
, srcs
[0]); break;
1970 case ir_unop_ceil
: result
= nir_fceil(&b
, srcs
[0]); break;
1971 case ir_unop_floor
: result
= nir_ffloor(&b
, srcs
[0]); break;
1972 case ir_unop_fract
: result
= nir_ffract(&b
, srcs
[0]); break;
1973 case ir_unop_frexp_exp
: result
= nir_frexp_exp(&b
, srcs
[0]); break;
1974 case ir_unop_frexp_sig
: result
= nir_frexp_sig(&b
, srcs
[0]); break;
1975 case ir_unop_round_even
: result
= nir_fround_even(&b
, srcs
[0]); break;
1976 case ir_unop_sin
: result
= nir_fsin(&b
, srcs
[0]); break;
1977 case ir_unop_cos
: result
= nir_fcos(&b
, srcs
[0]); break;
1978 case ir_unop_dFdx
: result
= nir_fddx(&b
, srcs
[0]); break;
1979 case ir_unop_dFdy
: result
= nir_fddy(&b
, srcs
[0]); break;
1980 case ir_unop_dFdx_fine
: result
= nir_fddx_fine(&b
, srcs
[0]); break;
1981 case ir_unop_dFdy_fine
: result
= nir_fddy_fine(&b
, srcs
[0]); break;
1982 case ir_unop_dFdx_coarse
: result
= nir_fddx_coarse(&b
, srcs
[0]); break;
1983 case ir_unop_dFdy_coarse
: result
= nir_fddy_coarse(&b
, srcs
[0]); break;
1984 case ir_unop_pack_snorm_2x16
:
1985 result
= nir_pack_snorm_2x16(&b
, srcs
[0]);
1987 case ir_unop_pack_snorm_4x8
:
1988 result
= nir_pack_snorm_4x8(&b
, srcs
[0]);
1990 case ir_unop_pack_unorm_2x16
:
1991 result
= nir_pack_unorm_2x16(&b
, srcs
[0]);
1993 case ir_unop_pack_unorm_4x8
:
1994 result
= nir_pack_unorm_4x8(&b
, srcs
[0]);
1996 case ir_unop_pack_half_2x16
:
1997 result
= nir_pack_half_2x16(&b
, srcs
[0]);
1999 case ir_unop_unpack_snorm_2x16
:
2000 result
= nir_unpack_snorm_2x16(&b
, srcs
[0]);
2002 case ir_unop_unpack_snorm_4x8
:
2003 result
= nir_unpack_snorm_4x8(&b
, srcs
[0]);
2005 case ir_unop_unpack_unorm_2x16
:
2006 result
= nir_unpack_unorm_2x16(&b
, srcs
[0]);
2008 case ir_unop_unpack_unorm_4x8
:
2009 result
= nir_unpack_unorm_4x8(&b
, srcs
[0]);
2011 case ir_unop_unpack_half_2x16
:
2012 result
= nir_unpack_half_2x16(&b
, srcs
[0]);
2014 case ir_unop_pack_sampler_2x32
:
2015 case ir_unop_pack_image_2x32
:
2016 case ir_unop_pack_double_2x32
:
2017 case ir_unop_pack_int_2x32
:
2018 case ir_unop_pack_uint_2x32
:
2019 result
= nir_pack_64_2x32(&b
, srcs
[0]);
2021 case ir_unop_unpack_sampler_2x32
:
2022 case ir_unop_unpack_image_2x32
:
2023 case ir_unop_unpack_double_2x32
:
2024 case ir_unop_unpack_int_2x32
:
2025 case ir_unop_unpack_uint_2x32
:
2026 result
= nir_unpack_64_2x32(&b
, srcs
[0]);
2028 case ir_unop_bitfield_reverse
:
2029 result
= nir_bitfield_reverse(&b
, srcs
[0]);
2031 case ir_unop_bit_count
:
2032 result
= nir_bit_count(&b
, srcs
[0]);
2034 case ir_unop_find_msb
:
2036 case GLSL_TYPE_UINT
:
2037 result
= nir_ufind_msb(&b
, srcs
[0]);
2040 result
= nir_ifind_msb(&b
, srcs
[0]);
2043 unreachable("Invalid type for findMSB()");
2046 case ir_unop_find_lsb
:
2047 result
= nir_find_lsb(&b
, srcs
[0]);
2051 switch (ir
->type
->vector_elements
) {
2053 switch (ir
->operands
[0]->type
->vector_elements
) {
2054 case 1: result
= nir_fnoise1_1(&b
, srcs
[0]); break;
2055 case 2: result
= nir_fnoise1_2(&b
, srcs
[0]); break;
2056 case 3: result
= nir_fnoise1_3(&b
, srcs
[0]); break;
2057 case 4: result
= nir_fnoise1_4(&b
, srcs
[0]); break;
2058 default: unreachable("not reached");
2062 switch (ir
->operands
[0]->type
->vector_elements
) {
2063 case 1: result
= nir_fnoise2_1(&b
, srcs
[0]); break;
2064 case 2: result
= nir_fnoise2_2(&b
, srcs
[0]); break;
2065 case 3: result
= nir_fnoise2_3(&b
, srcs
[0]); break;
2066 case 4: result
= nir_fnoise2_4(&b
, srcs
[0]); break;
2067 default: unreachable("not reached");
2071 switch (ir
->operands
[0]->type
->vector_elements
) {
2072 case 1: result
= nir_fnoise3_1(&b
, srcs
[0]); break;
2073 case 2: result
= nir_fnoise3_2(&b
, srcs
[0]); break;
2074 case 3: result
= nir_fnoise3_3(&b
, srcs
[0]); break;
2075 case 4: result
= nir_fnoise3_4(&b
, srcs
[0]); break;
2076 default: unreachable("not reached");
2080 switch (ir
->operands
[0]->type
->vector_elements
) {
2081 case 1: result
= nir_fnoise4_1(&b
, srcs
[0]); break;
2082 case 2: result
= nir_fnoise4_2(&b
, srcs
[0]); break;
2083 case 3: result
= nir_fnoise4_3(&b
, srcs
[0]); break;
2084 case 4: result
= nir_fnoise4_4(&b
, srcs
[0]); break;
2085 default: unreachable("not reached");
2089 unreachable("not reached");
2092 case ir_unop_get_buffer_size
: {
2093 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(
2095 nir_intrinsic_get_buffer_size
);
2096 load
->num_components
= ir
->type
->vector_elements
;
2097 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
2098 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2099 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
2104 result
= nir_atan(&b
, srcs
[0]);
2108 result
= type_is_float(out_type
) ? nir_fadd(&b
, srcs
[0], srcs
[1])
2109 : nir_iadd(&b
, srcs
[0], srcs
[1]);
2111 case ir_binop_add_sat
:
2112 result
= type_is_signed(out_type
) ? nir_iadd_sat(&b
, srcs
[0], srcs
[1])
2113 : nir_uadd_sat(&b
, srcs
[0], srcs
[1]);
2116 result
= type_is_float(out_type
) ? nir_fsub(&b
, srcs
[0], srcs
[1])
2117 : nir_isub(&b
, srcs
[0], srcs
[1]);
2119 case ir_binop_sub_sat
:
2120 result
= type_is_signed(out_type
) ? nir_isub_sat(&b
, srcs
[0], srcs
[1])
2121 : nir_usub_sat(&b
, srcs
[0], srcs
[1]);
2123 case ir_binop_abs_sub
:
2124 /* out_type is always unsigned for ir_binop_abs_sub, so we have to key
2125 * on the type of the sources.
2127 result
= type_is_signed(types
[0]) ? nir_uabs_isub(&b
, srcs
[0], srcs
[1])
2128 : nir_uabs_usub(&b
, srcs
[0], srcs
[1]);
2131 result
= type_is_signed(out_type
) ? nir_ihadd(&b
, srcs
[0], srcs
[1])
2132 : nir_uhadd(&b
, srcs
[0], srcs
[1]);
2134 case ir_binop_avg_round
:
2135 result
= type_is_signed(out_type
) ? nir_irhadd(&b
, srcs
[0], srcs
[1])
2136 : nir_urhadd(&b
, srcs
[0], srcs
[1]);
2138 case ir_binop_mul_32x16
:
2139 result
= type_is_signed(out_type
) ? nir_imul_32x16(&b
, srcs
[0], srcs
[1])
2140 : nir_umul_32x16(&b
, srcs
[0], srcs
[1]);
2143 if (type_is_float(out_type
))
2144 result
= nir_fmul(&b
, srcs
[0], srcs
[1]);
2145 else if (out_type
== GLSL_TYPE_INT64
&&
2146 (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
||
2147 ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
))
2148 result
= nir_imul_2x32_64(&b
, srcs
[0], srcs
[1]);
2149 else if (out_type
== GLSL_TYPE_UINT64
&&
2150 (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
||
2151 ir
->operands
[1]->type
->base_type
== GLSL_TYPE_UINT
))
2152 result
= nir_umul_2x32_64(&b
, srcs
[0], srcs
[1]);
2154 result
= nir_imul(&b
, srcs
[0], srcs
[1]);
2157 if (type_is_float(out_type
))
2158 result
= nir_fdiv(&b
, srcs
[0], srcs
[1]);
2159 else if (type_is_signed(out_type
))
2160 result
= nir_idiv(&b
, srcs
[0], srcs
[1]);
2162 result
= nir_udiv(&b
, srcs
[0], srcs
[1]);
2165 result
= type_is_float(out_type
) ? nir_fmod(&b
, srcs
[0], srcs
[1])
2166 : nir_umod(&b
, srcs
[0], srcs
[1]);
2169 if (type_is_float(out_type
))
2170 result
= nir_fmin(&b
, srcs
[0], srcs
[1]);
2171 else if (type_is_signed(out_type
))
2172 result
= nir_imin(&b
, srcs
[0], srcs
[1]);
2174 result
= nir_umin(&b
, srcs
[0], srcs
[1]);
2177 if (type_is_float(out_type
))
2178 result
= nir_fmax(&b
, srcs
[0], srcs
[1]);
2179 else if (type_is_signed(out_type
))
2180 result
= nir_imax(&b
, srcs
[0], srcs
[1]);
2182 result
= nir_umax(&b
, srcs
[0], srcs
[1]);
2184 case ir_binop_pow
: result
= nir_fpow(&b
, srcs
[0], srcs
[1]); break;
2185 case ir_binop_bit_and
: result
= nir_iand(&b
, srcs
[0], srcs
[1]); break;
2186 case ir_binop_bit_or
: result
= nir_ior(&b
, srcs
[0], srcs
[1]); break;
2187 case ir_binop_bit_xor
: result
= nir_ixor(&b
, srcs
[0], srcs
[1]); break;
2188 case ir_binop_logic_and
:
2189 result
= nir_iand(&b
, srcs
[0], srcs
[1]);
2191 case ir_binop_logic_or
:
2192 result
= nir_ior(&b
, srcs
[0], srcs
[1]);
2194 case ir_binop_logic_xor
:
2195 result
= nir_ixor(&b
, srcs
[0], srcs
[1]);
2197 case ir_binop_lshift
: result
= nir_ishl(&b
, srcs
[0], srcs
[1]); break;
2198 case ir_binop_rshift
:
2199 result
= (type_is_signed(out_type
)) ? nir_ishr(&b
, srcs
[0], srcs
[1])
2200 : nir_ushr(&b
, srcs
[0], srcs
[1]);
2202 case ir_binop_imul_high
:
2203 result
= (out_type
== GLSL_TYPE_INT
) ? nir_imul_high(&b
, srcs
[0], srcs
[1])
2204 : nir_umul_high(&b
, srcs
[0], srcs
[1]);
2206 case ir_binop_carry
: result
= nir_uadd_carry(&b
, srcs
[0], srcs
[1]); break;
2207 case ir_binop_borrow
: result
= nir_usub_borrow(&b
, srcs
[0], srcs
[1]); break;
2209 if (type_is_float(types
[0]))
2210 result
= nir_flt(&b
, srcs
[0], srcs
[1]);
2211 else if (type_is_signed(types
[0]))
2212 result
= nir_ilt(&b
, srcs
[0], srcs
[1]);
2214 result
= nir_ult(&b
, srcs
[0], srcs
[1]);
2216 case ir_binop_gequal
:
2217 if (type_is_float(types
[0]))
2218 result
= nir_fge(&b
, srcs
[0], srcs
[1]);
2219 else if (type_is_signed(types
[0]))
2220 result
= nir_ige(&b
, srcs
[0], srcs
[1]);
2222 result
= nir_uge(&b
, srcs
[0], srcs
[1]);
2224 case ir_binop_equal
:
2225 if (type_is_float(types
[0]))
2226 result
= nir_feq(&b
, srcs
[0], srcs
[1]);
2228 result
= nir_ieq(&b
, srcs
[0], srcs
[1]);
2230 case ir_binop_nequal
:
2231 if (type_is_float(types
[0]))
2232 result
= nir_fne(&b
, srcs
[0], srcs
[1]);
2234 result
= nir_ine(&b
, srcs
[0], srcs
[1]);
2236 case ir_binop_all_equal
:
2237 if (type_is_float(types
[0])) {
2238 switch (ir
->operands
[0]->type
->vector_elements
) {
2239 case 1: result
= nir_feq(&b
, srcs
[0], srcs
[1]); break;
2240 case 2: result
= nir_ball_fequal2(&b
, srcs
[0], srcs
[1]); break;
2241 case 3: result
= nir_ball_fequal3(&b
, srcs
[0], srcs
[1]); break;
2242 case 4: result
= nir_ball_fequal4(&b
, srcs
[0], srcs
[1]); break;
2244 unreachable("not reached");
2247 switch (ir
->operands
[0]->type
->vector_elements
) {
2248 case 1: result
= nir_ieq(&b
, srcs
[0], srcs
[1]); break;
2249 case 2: result
= nir_ball_iequal2(&b
, srcs
[0], srcs
[1]); break;
2250 case 3: result
= nir_ball_iequal3(&b
, srcs
[0], srcs
[1]); break;
2251 case 4: result
= nir_ball_iequal4(&b
, srcs
[0], srcs
[1]); break;
2253 unreachable("not reached");
2257 case ir_binop_any_nequal
:
2258 if (type_is_float(types
[0])) {
2259 switch (ir
->operands
[0]->type
->vector_elements
) {
2260 case 1: result
= nir_fne(&b
, srcs
[0], srcs
[1]); break;
2261 case 2: result
= nir_bany_fnequal2(&b
, srcs
[0], srcs
[1]); break;
2262 case 3: result
= nir_bany_fnequal3(&b
, srcs
[0], srcs
[1]); break;
2263 case 4: result
= nir_bany_fnequal4(&b
, srcs
[0], srcs
[1]); break;
2265 unreachable("not reached");
2268 switch (ir
->operands
[0]->type
->vector_elements
) {
2269 case 1: result
= nir_ine(&b
, srcs
[0], srcs
[1]); break;
2270 case 2: result
= nir_bany_inequal2(&b
, srcs
[0], srcs
[1]); break;
2271 case 3: result
= nir_bany_inequal3(&b
, srcs
[0], srcs
[1]); break;
2272 case 4: result
= nir_bany_inequal4(&b
, srcs
[0], srcs
[1]); break;
2274 unreachable("not reached");
2279 switch (ir
->operands
[0]->type
->vector_elements
) {
2280 case 2: result
= nir_fdot2(&b
, srcs
[0], srcs
[1]); break;
2281 case 3: result
= nir_fdot3(&b
, srcs
[0], srcs
[1]); break;
2282 case 4: result
= nir_fdot4(&b
, srcs
[0], srcs
[1]); break;
2284 unreachable("not reached");
2287 case ir_binop_vector_extract
: {
2288 result
= nir_channel(&b
, srcs
[0], 0);
2289 for (unsigned i
= 1; i
< ir
->operands
[0]->type
->vector_elements
; i
++) {
2290 nir_ssa_def
*swizzled
= nir_channel(&b
, srcs
[0], i
);
2291 result
= nir_bcsel(&b
, nir_ieq(&b
, srcs
[1], nir_imm_int(&b
, i
)),
2297 case ir_binop_atan2
:
2298 result
= nir_atan2(&b
, srcs
[0], srcs
[1]);
2301 case ir_binop_ldexp
: result
= nir_ldexp(&b
, srcs
[0], srcs
[1]); break;
2303 result
= nir_ffma(&b
, srcs
[0], srcs
[1], srcs
[2]);
2306 result
= nir_flrp(&b
, srcs
[0], srcs
[1], srcs
[2]);
2309 result
= nir_bcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
2311 case ir_triop_bitfield_extract
:
2312 result
= (out_type
== GLSL_TYPE_INT
) ?
2313 nir_ibitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]) :
2314 nir_ubitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]);
2316 case ir_quadop_bitfield_insert
:
2317 result
= nir_bitfield_insert(&b
, srcs
[0], srcs
[1], srcs
[2], srcs
[3]);
2319 case ir_quadop_vector
:
2320 result
= nir_vec(&b
, srcs
, ir
->type
->vector_elements
);
2324 unreachable("not reached");
2329 nir_visitor::visit(ir_swizzle
*ir
)
2331 unsigned swizzle
[4] = { ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
};
2332 result
= nir_swizzle(&b
, evaluate_rvalue(ir
->val
), swizzle
,
2333 ir
->type
->vector_elements
);
2337 nir_visitor::visit(ir_texture
*ir
)
2344 num_srcs
= 1; /* coordinate */
2349 op
= (ir
->op
== ir_txb
) ? nir_texop_txb
: nir_texop_txl
;
2350 num_srcs
= 2; /* coordinate, bias/lod */
2354 op
= nir_texop_txd
; /* coordinate, dPdx, dPdy */
2360 if (ir
->lod_info
.lod
!= NULL
)
2361 num_srcs
= 2; /* coordinate, lod */
2363 num_srcs
= 1; /* coordinate */
2367 op
= nir_texop_txf_ms
;
2368 num_srcs
= 2; /* coordinate, sample_index */
2373 if (ir
->lod_info
.lod
!= NULL
)
2374 num_srcs
= 1; /* lod */
2381 num_srcs
= 1; /* coordinate */
2386 num_srcs
= 1; /* coordinate */
2389 case ir_query_levels
:
2390 op
= nir_texop_query_levels
;
2394 case ir_texture_samples
:
2395 op
= nir_texop_texture_samples
;
2399 case ir_samples_identical
:
2400 op
= nir_texop_samples_identical
;
2401 num_srcs
= 1; /* coordinate */
2405 unreachable("not reached");
2408 if (ir
->projector
!= NULL
)
2410 if (ir
->shadow_comparator
!= NULL
)
2412 /* offsets are constants we store inside nir_tex_intrs.offsets */
2413 if (ir
->offset
!= NULL
&& !ir
->offset
->type
->is_array())
2416 /* Add one for the texture deref */
2419 nir_tex_instr
*instr
= nir_tex_instr_create(this->shader
, num_srcs
);
2422 instr
->sampler_dim
=
2423 (glsl_sampler_dim
) ir
->sampler
->type
->sampler_dimensionality
;
2424 instr
->is_array
= ir
->sampler
->type
->sampler_array
;
2425 instr
->is_shadow
= ir
->sampler
->type
->sampler_shadow
;
2426 if (instr
->is_shadow
)
2427 instr
->is_new_style_shadow
= (ir
->type
->vector_elements
== 1);
2428 switch (ir
->type
->base_type
) {
2429 case GLSL_TYPE_FLOAT
:
2430 instr
->dest_type
= nir_type_float
;
2432 case GLSL_TYPE_FLOAT16
:
2433 instr
->dest_type
= nir_type_float16
;
2436 instr
->dest_type
= nir_type_int
;
2438 case GLSL_TYPE_BOOL
:
2439 case GLSL_TYPE_UINT
:
2440 instr
->dest_type
= nir_type_uint
;
2443 unreachable("not reached");
2446 nir_deref_instr
*sampler_deref
= evaluate_deref(ir
->sampler
);
2448 /* check for bindless handles */
2449 if (sampler_deref
->mode
!= nir_var_uniform
||
2450 nir_deref_instr_get_variable(sampler_deref
)->data
.bindless
) {
2451 nir_ssa_def
*load
= nir_load_deref(&b
, sampler_deref
);
2452 instr
->src
[0].src
= nir_src_for_ssa(load
);
2453 instr
->src
[0].src_type
= nir_tex_src_texture_handle
;
2454 instr
->src
[1].src
= nir_src_for_ssa(load
);
2455 instr
->src
[1].src_type
= nir_tex_src_sampler_handle
;
2457 instr
->src
[0].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2458 instr
->src
[0].src_type
= nir_tex_src_texture_deref
;
2459 instr
->src
[1].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2460 instr
->src
[1].src_type
= nir_tex_src_sampler_deref
;
2463 unsigned src_number
= 2;
2465 if (ir
->coordinate
!= NULL
) {
2466 instr
->coord_components
= ir
->coordinate
->type
->vector_elements
;
2467 instr
->src
[src_number
].src
=
2468 nir_src_for_ssa(evaluate_rvalue(ir
->coordinate
));
2469 instr
->src
[src_number
].src_type
= nir_tex_src_coord
;
2473 if (ir
->projector
!= NULL
) {
2474 instr
->src
[src_number
].src
=
2475 nir_src_for_ssa(evaluate_rvalue(ir
->projector
));
2476 instr
->src
[src_number
].src_type
= nir_tex_src_projector
;
2480 if (ir
->shadow_comparator
!= NULL
) {
2481 instr
->src
[src_number
].src
=
2482 nir_src_for_ssa(evaluate_rvalue(ir
->shadow_comparator
));
2483 instr
->src
[src_number
].src_type
= nir_tex_src_comparator
;
2487 if (ir
->offset
!= NULL
) {
2488 if (ir
->offset
->type
->is_array()) {
2489 for (int i
= 0; i
< ir
->offset
->type
->array_size(); i
++) {
2490 const ir_constant
*c
=
2491 ir
->offset
->as_constant()->get_array_element(i
);
2493 for (unsigned j
= 0; j
< 2; ++j
) {
2494 int val
= c
->get_int_component(j
);
2495 assert(val
<= 31 && val
>= -32);
2496 instr
->tg4_offsets
[i
][j
] = val
;
2500 assert(ir
->offset
->type
->is_vector() || ir
->offset
->type
->is_scalar());
2502 instr
->src
[src_number
].src
=
2503 nir_src_for_ssa(evaluate_rvalue(ir
->offset
));
2504 instr
->src
[src_number
].src_type
= nir_tex_src_offset
;
2511 instr
->src
[src_number
].src
=
2512 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.bias
));
2513 instr
->src
[src_number
].src_type
= nir_tex_src_bias
;
2520 if (ir
->lod_info
.lod
!= NULL
) {
2521 instr
->src
[src_number
].src
=
2522 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.lod
));
2523 instr
->src
[src_number
].src_type
= nir_tex_src_lod
;
2529 instr
->src
[src_number
].src
=
2530 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdx
));
2531 instr
->src
[src_number
].src_type
= nir_tex_src_ddx
;
2533 instr
->src
[src_number
].src
=
2534 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdy
));
2535 instr
->src
[src_number
].src_type
= nir_tex_src_ddy
;
2540 instr
->src
[src_number
].src
=
2541 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.sample_index
));
2542 instr
->src
[src_number
].src_type
= nir_tex_src_ms_index
;
2547 instr
->component
= ir
->lod_info
.component
->as_constant()->value
.u
[0];
2554 assert(src_number
== num_srcs
);
2556 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2557 add_instr(&instr
->instr
, nir_tex_instr_dest_size(instr
), bit_size
);
2561 nir_visitor::visit(ir_constant
*ir
)
2564 * We don't know if this variable is an array or struct that gets
2565 * dereferenced, so do the safe thing an make it a variable with a
2566 * constant initializer and return a dereference.
2570 nir_local_variable_create(this->impl
, ir
->type
, "const_temp");
2571 var
->data
.read_only
= true;
2572 var
->constant_initializer
= constant_copy(ir
, var
);
2574 this->deref
= nir_build_deref_var(&b
, var
);
2578 nir_visitor::visit(ir_dereference_variable
*ir
)
2580 if (ir
->variable_referenced()->data
.mode
== ir_var_function_out
) {
2581 unsigned i
= (sig
->return_type
!= glsl_type::void_type
) ? 1 : 0;
2583 foreach_in_list(ir_variable
, param
, &sig
->parameters
) {
2584 if (param
== ir
->variable_referenced()) {
2590 this->deref
= nir_build_deref_cast(&b
, nir_load_param(&b
, i
),
2591 nir_var_function_temp
, ir
->type
, 0);
2595 assert(ir
->variable_referenced()->data
.mode
!= ir_var_function_inout
);
2597 struct hash_entry
*entry
=
2598 _mesa_hash_table_search(this->var_table
, ir
->var
);
2600 nir_variable
*var
= (nir_variable
*) entry
->data
;
2602 this->deref
= nir_build_deref_var(&b
, var
);
2606 nir_visitor::visit(ir_dereference_record
*ir
)
2608 ir
->record
->accept(this);
2610 int field_index
= ir
->field_idx
;
2611 assert(field_index
>= 0);
2613 this->deref
= nir_build_deref_struct(&b
, this->deref
, field_index
);
2617 nir_visitor::visit(ir_dereference_array
*ir
)
2619 nir_ssa_def
*index
= evaluate_rvalue(ir
->array_index
);
2621 ir
->array
->accept(this);
2623 this->deref
= nir_build_deref_array(&b
, this->deref
, index
);
2627 nir_visitor::visit(ir_barrier
*)
2629 if (shader
->info
.stage
== MESA_SHADER_COMPUTE
) {
2630 nir_intrinsic_instr
*shared_barrier
=
2631 nir_intrinsic_instr_create(this->shader
,
2632 nir_intrinsic_memory_barrier_shared
);
2633 nir_builder_instr_insert(&b
, &shared_barrier
->instr
);
2634 } else if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
2635 nir_intrinsic_instr
*patch_barrier
=
2636 nir_intrinsic_instr_create(this->shader
,
2637 nir_intrinsic_memory_barrier_tcs_patch
);
2638 nir_builder_instr_insert(&b
, &patch_barrier
->instr
);
2641 nir_intrinsic_instr
*instr
=
2642 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_control_barrier
);
2643 nir_builder_instr_insert(&b
, &instr
->instr
);
2647 glsl_float64_funcs_to_nir(struct gl_context
*ctx
,
2648 const nir_shader_compiler_options
*options
)
2650 /* We pretend it's a vertex shader. Ultimately, the stage shouldn't
2651 * matter because we're not optimizing anything here.
2653 struct gl_shader
*sh
= _mesa_new_shader(-1, MESA_SHADER_VERTEX
);
2654 sh
->Source
= float64_source
;
2655 sh
->CompileStatus
= COMPILE_FAILURE
;
2656 _mesa_glsl_compile_shader(ctx
, sh
, false, false, true);
2658 if (!sh
->CompileStatus
) {
2661 "fp64 software impl compile failed:\n%s\nsource:\n%s\n",
2662 sh
->InfoLog
, float64_source
);
2667 nir_shader
*nir
= nir_shader_create(NULL
, MESA_SHADER_VERTEX
, options
, NULL
);
2669 nir_visitor
v1(ctx
, nir
);
2670 nir_function_visitor
v2(&v1
);
2672 visit_exec_list(sh
->ir
, &v1
);
2674 /* _mesa_delete_shader will try to free sh->Source but it's static const */
2676 _mesa_delete_shader(ctx
, sh
);
2678 nir_validate_shader(nir
, "float64_funcs_to_nir");
2680 NIR_PASS_V(nir
, nir_lower_variable_initializers
, nir_var_function_temp
);
2681 NIR_PASS_V(nir
, nir_lower_returns
);
2682 NIR_PASS_V(nir
, nir_inline_functions
);
2683 NIR_PASS_V(nir
, nir_opt_deref
);
2685 /* Do some optimizations to clean up the shader now. By optimizing the
2686 * functions in the library, we avoid having to re-do that work every
2687 * time we inline a copy of a function. Reducing basic blocks also helps
2688 * with compile times.
2690 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2691 NIR_PASS_V(nir
, nir_copy_prop
);
2692 NIR_PASS_V(nir
, nir_opt_dce
);
2693 NIR_PASS_V(nir
, nir_opt_cse
);
2694 NIR_PASS_V(nir
, nir_opt_gcm
, true);
2695 NIR_PASS_V(nir
, nir_opt_peephole_select
, 1, false, false);
2696 NIR_PASS_V(nir
, nir_opt_dce
);