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_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
.from_named_ifc_block
= ir
->data
.from_named_ifc_block
;
468 var
->data
.compact
= false;
470 switch(ir
->data
.mode
) {
472 case ir_var_temporary
:
474 var
->data
.mode
= nir_var_shader_temp
;
476 var
->data
.mode
= nir_var_function_temp
;
479 case ir_var_function_in
:
480 case ir_var_const_in
:
481 var
->data
.mode
= nir_var_function_temp
;
484 case ir_var_shader_in
:
485 if (shader
->info
.stage
== MESA_SHADER_GEOMETRY
&&
486 ir
->data
.location
== VARYING_SLOT_PRIMITIVE_ID
) {
487 /* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */
488 var
->data
.location
= SYSTEM_VALUE_PRIMITIVE_ID
;
489 var
->data
.mode
= nir_var_system_value
;
491 var
->data
.mode
= nir_var_shader_in
;
493 if (shader
->info
.stage
== MESA_SHADER_TESS_EVAL
&&
494 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
495 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
496 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
499 if (shader
->info
.stage
> MESA_SHADER_VERTEX
&&
500 ir
->data
.location
>= VARYING_SLOT_CLIP_DIST0
&&
501 ir
->data
.location
<= VARYING_SLOT_CULL_DIST1
) {
502 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
507 case ir_var_shader_out
:
508 var
->data
.mode
= nir_var_shader_out
;
509 if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
&&
510 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
511 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
512 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
515 if (shader
->info
.stage
<= MESA_SHADER_GEOMETRY
&&
516 ir
->data
.location
>= VARYING_SLOT_CLIP_DIST0
&&
517 ir
->data
.location
<= VARYING_SLOT_CULL_DIST1
) {
518 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
523 if (ir
->get_interface_type())
524 var
->data
.mode
= nir_var_mem_ubo
;
526 var
->data
.mode
= nir_var_uniform
;
529 case ir_var_shader_storage
:
530 var
->data
.mode
= nir_var_mem_ssbo
;
533 case ir_var_system_value
:
534 var
->data
.mode
= nir_var_system_value
;
538 unreachable("not reached");
541 unsigned mem_access
= 0;
542 if (ir
->data
.memory_read_only
)
543 mem_access
|= ACCESS_NON_WRITEABLE
;
544 if (ir
->data
.memory_write_only
)
545 mem_access
|= ACCESS_NON_READABLE
;
546 if (ir
->data
.memory_coherent
)
547 mem_access
|= ACCESS_COHERENT
;
548 if (ir
->data
.memory_volatile
)
549 mem_access
|= ACCESS_VOLATILE
;
550 if (ir
->data
.memory_restrict
)
551 mem_access
|= ACCESS_RESTRICT
;
553 var
->interface_type
= ir
->get_interface_type();
555 /* For UBO and SSBO variables, we need explicit types */
556 if (var
->data
.mode
& (nir_var_mem_ubo
| nir_var_mem_ssbo
)) {
557 const glsl_type
*explicit_ifc_type
=
558 ir
->get_interface_type()->get_explicit_interface_type(supports_std430
);
560 var
->interface_type
= explicit_ifc_type
;
562 if (ir
->type
->without_array()->is_interface()) {
563 /* If the type contains the interface, wrap the explicit type in the
564 * right number of arrays.
566 var
->type
= wrap_type_in_array(explicit_ifc_type
, ir
->type
);
568 /* Otherwise, this variable is one entry in the interface */
569 UNUSED
bool found
= false;
570 for (unsigned i
= 0; i
< explicit_ifc_type
->length
; i
++) {
571 const glsl_struct_field
*field
=
572 &explicit_ifc_type
->fields
.structure
[i
];
573 if (strcmp(ir
->name
, field
->name
) != 0)
576 var
->type
= field
->type
;
577 if (field
->memory_read_only
)
578 mem_access
|= ACCESS_NON_WRITEABLE
;
579 if (field
->memory_write_only
)
580 mem_access
|= ACCESS_NON_READABLE
;
581 if (field
->memory_coherent
)
582 mem_access
|= ACCESS_COHERENT
;
583 if (field
->memory_volatile
)
584 mem_access
|= ACCESS_VOLATILE
;
585 if (field
->memory_restrict
)
586 mem_access
|= ACCESS_RESTRICT
;
595 var
->data
.interpolation
= ir
->data
.interpolation
;
596 var
->data
.location_frac
= ir
->data
.location_frac
;
598 switch (ir
->data
.depth_layout
) {
599 case ir_depth_layout_none
:
600 var
->data
.depth_layout
= nir_depth_layout_none
;
602 case ir_depth_layout_any
:
603 var
->data
.depth_layout
= nir_depth_layout_any
;
605 case ir_depth_layout_greater
:
606 var
->data
.depth_layout
= nir_depth_layout_greater
;
608 case ir_depth_layout_less
:
609 var
->data
.depth_layout
= nir_depth_layout_less
;
611 case ir_depth_layout_unchanged
:
612 var
->data
.depth_layout
= nir_depth_layout_unchanged
;
615 unreachable("not reached");
618 var
->data
.index
= ir
->data
.index
;
619 var
->data
.descriptor_set
= 0;
620 var
->data
.binding
= ir
->data
.binding
;
621 var
->data
.explicit_binding
= ir
->data
.explicit_binding
;
622 var
->data
.bindless
= ir
->data
.bindless
;
623 var
->data
.offset
= ir
->data
.offset
;
624 var
->data
.access
= (gl_access_qualifier
)mem_access
;
626 if (var
->type
->without_array()->is_image()) {
627 var
->data
.image
.format
= ir
->data
.image_format
;
628 } else if (var
->data
.mode
== nir_var_shader_out
) {
629 var
->data
.xfb
.buffer
= ir
->data
.xfb_buffer
;
630 var
->data
.xfb
.stride
= ir
->data
.xfb_stride
;
633 var
->data
.fb_fetch_output
= ir
->data
.fb_fetch_output
;
634 var
->data
.explicit_xfb_buffer
= ir
->data
.explicit_xfb_buffer
;
635 var
->data
.explicit_xfb_stride
= ir
->data
.explicit_xfb_stride
;
637 var
->num_state_slots
= ir
->get_num_state_slots();
638 if (var
->num_state_slots
> 0) {
639 var
->state_slots
= rzalloc_array(var
, nir_state_slot
,
640 var
->num_state_slots
);
642 ir_state_slot
*state_slots
= ir
->get_state_slots();
643 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
644 for (unsigned j
= 0; j
< 5; j
++)
645 var
->state_slots
[i
].tokens
[j
] = state_slots
[i
].tokens
[j
];
646 var
->state_slots
[i
].swizzle
= state_slots
[i
].swizzle
;
649 var
->state_slots
= NULL
;
652 var
->constant_initializer
= constant_copy(ir
->constant_initializer
, var
);
654 if (var
->data
.mode
== nir_var_function_temp
)
655 nir_function_impl_add_variable(impl
, var
);
657 nir_shader_add_variable(shader
, var
);
659 _mesa_hash_table_insert(var_table
, ir
, var
);
663 nir_function_visitor::visit_enter(ir_function
*ir
)
665 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
) {
666 visitor
->create_function(sig
);
668 return visit_continue_with_parent
;
672 nir_visitor::create_function(ir_function_signature
*ir
)
674 if (ir
->is_intrinsic())
677 nir_function
*func
= nir_function_create(shader
, ir
->function_name());
678 if (strcmp(ir
->function_name(), "main") == 0)
679 func
->is_entrypoint
= true;
681 func
->num_params
= ir
->parameters
.length() +
682 (ir
->return_type
!= glsl_type::void_type
);
683 func
->params
= ralloc_array(shader
, nir_parameter
, func
->num_params
);
687 if (ir
->return_type
!= glsl_type::void_type
) {
688 /* The return value is a variable deref (basically an out parameter) */
689 func
->params
[np
].num_components
= 1;
690 func
->params
[np
].bit_size
= 32;
694 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
695 /* FINISHME: pass arrays, structs, etc by reference? */
696 assert(param
->type
->is_vector() || param
->type
->is_scalar());
698 if (param
->data
.mode
== ir_var_function_in
) {
699 func
->params
[np
].num_components
= param
->type
->vector_elements
;
700 func
->params
[np
].bit_size
= glsl_get_bit_size(param
->type
);
702 func
->params
[np
].num_components
= 1;
703 func
->params
[np
].bit_size
= 32;
707 assert(np
== func
->num_params
);
709 _mesa_hash_table_insert(this->overload_table
, ir
, func
);
713 nir_visitor::visit(ir_function
*ir
)
715 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
)
720 nir_visitor::visit(ir_function_signature
*ir
)
722 if (ir
->is_intrinsic())
727 struct hash_entry
*entry
=
728 _mesa_hash_table_search(this->overload_table
, ir
);
731 nir_function
*func
= (nir_function
*) entry
->data
;
733 if (ir
->is_defined
) {
734 nir_function_impl
*impl
= nir_function_impl_create(func
);
737 this->is_global
= false;
739 nir_builder_init(&b
, impl
);
740 b
.cursor
= nir_after_cf_list(&impl
->body
);
742 unsigned i
= (ir
->return_type
!= glsl_type::void_type
) ? 1 : 0;
744 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
746 nir_local_variable_create(impl
, param
->type
, param
->name
);
748 if (param
->data
.mode
== ir_var_function_in
) {
749 nir_store_var(&b
, var
, nir_load_param(&b
, i
), ~0);
752 _mesa_hash_table_insert(var_table
, param
, var
);
756 visit_exec_list(&ir
->body
, this);
758 this->is_global
= true;
765 nir_visitor::visit(ir_loop
*ir
)
768 visit_exec_list(&ir
->body_instructions
, this);
769 nir_pop_loop(&b
, NULL
);
773 nir_visitor::visit(ir_if
*ir
)
775 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
776 visit_exec_list(&ir
->then_instructions
, this);
777 nir_push_else(&b
, NULL
);
778 visit_exec_list(&ir
->else_instructions
, this);
779 nir_pop_if(&b
, NULL
);
783 nir_visitor::visit(ir_discard
*ir
)
786 * discards aren't treated as control flow, because before we lower them
787 * they can appear anywhere in the shader and the stuff after them may still
788 * be executed (yay, crazy GLSL rules!). However, after lowering, all the
789 * discards will be immediately followed by a return.
792 nir_intrinsic_instr
*discard
;
794 discard
= nir_intrinsic_instr_create(this->shader
,
795 nir_intrinsic_discard_if
);
797 nir_src_for_ssa(evaluate_rvalue(ir
->condition
));
799 discard
= nir_intrinsic_instr_create(this->shader
, nir_intrinsic_discard
);
802 nir_builder_instr_insert(&b
, &discard
->instr
);
806 nir_visitor::visit(ir_demote
*ir
)
808 nir_intrinsic_instr
*demote
=
809 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_demote
);
811 nir_builder_instr_insert(&b
, &demote
->instr
);
815 nir_visitor::visit(ir_emit_vertex
*ir
)
817 nir_intrinsic_instr
*instr
=
818 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_emit_vertex
);
819 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
820 nir_builder_instr_insert(&b
, &instr
->instr
);
824 nir_visitor::visit(ir_end_primitive
*ir
)
826 nir_intrinsic_instr
*instr
=
827 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_end_primitive
);
828 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
829 nir_builder_instr_insert(&b
, &instr
->instr
);
833 nir_visitor::visit(ir_loop_jump
*ir
)
837 case ir_loop_jump::jump_break
:
838 type
= nir_jump_break
;
840 case ir_loop_jump::jump_continue
:
841 type
= nir_jump_continue
;
844 unreachable("not reached");
847 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, type
);
848 nir_builder_instr_insert(&b
, &instr
->instr
);
852 nir_visitor::visit(ir_return
*ir
)
854 if (ir
->value
!= NULL
) {
855 nir_deref_instr
*ret_deref
=
856 nir_build_deref_cast(&b
, nir_load_param(&b
, 0),
857 nir_var_function_temp
, ir
->value
->type
, 0);
859 nir_ssa_def
*val
= evaluate_rvalue(ir
->value
);
860 nir_store_deref(&b
, ret_deref
, val
, ~0);
863 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, nir_jump_return
);
864 nir_builder_instr_insert(&b
, &instr
->instr
);
868 intrinsic_set_std430_align(nir_intrinsic_instr
*intrin
, const glsl_type
*type
)
870 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
871 unsigned pow2_components
= util_next_power_of_two(type
->vector_elements
);
872 nir_intrinsic_set_align(intrin
, (bit_size
/ 8) * pow2_components
, 0);
875 /* Accumulate any qualifiers along the deref chain to get the actual
876 * load/store qualifier.
879 static enum gl_access_qualifier
880 deref_get_qualifier(nir_deref_instr
*deref
)
883 nir_deref_path_init(&path
, deref
, NULL
);
885 unsigned qualifiers
= path
.path
[0]->var
->data
.access
;
887 const glsl_type
*parent_type
= path
.path
[0]->type
;
888 for (nir_deref_instr
**cur_ptr
= &path
.path
[1]; *cur_ptr
; cur_ptr
++) {
889 nir_deref_instr
*cur
= *cur_ptr
;
891 if (parent_type
->is_interface()) {
892 const struct glsl_struct_field
*field
=
893 &parent_type
->fields
.structure
[cur
->strct
.index
];
894 if (field
->memory_read_only
)
895 qualifiers
|= ACCESS_NON_WRITEABLE
;
896 if (field
->memory_write_only
)
897 qualifiers
|= ACCESS_NON_READABLE
;
898 if (field
->memory_coherent
)
899 qualifiers
|= ACCESS_COHERENT
;
900 if (field
->memory_volatile
)
901 qualifiers
|= ACCESS_VOLATILE
;
902 if (field
->memory_restrict
)
903 qualifiers
|= ACCESS_RESTRICT
;
906 parent_type
= cur
->type
;
909 nir_deref_path_finish(&path
);
911 return (gl_access_qualifier
) qualifiers
;
915 nir_visitor::visit(ir_call
*ir
)
917 if (ir
->callee
->is_intrinsic()) {
920 switch (ir
->callee
->intrinsic_id
) {
921 case ir_intrinsic_generic_atomic_add
:
922 op
= ir
->return_deref
->type
->is_integer_32_64()
923 ? nir_intrinsic_deref_atomic_add
: nir_intrinsic_deref_atomic_fadd
;
925 case ir_intrinsic_generic_atomic_and
:
926 op
= nir_intrinsic_deref_atomic_and
;
928 case ir_intrinsic_generic_atomic_or
:
929 op
= nir_intrinsic_deref_atomic_or
;
931 case ir_intrinsic_generic_atomic_xor
:
932 op
= nir_intrinsic_deref_atomic_xor
;
934 case ir_intrinsic_generic_atomic_min
:
935 assert(ir
->return_deref
);
936 if (ir
->return_deref
->type
== glsl_type::int_type
)
937 op
= nir_intrinsic_deref_atomic_imin
;
938 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
939 op
= nir_intrinsic_deref_atomic_umin
;
940 else if (ir
->return_deref
->type
== glsl_type::float_type
)
941 op
= nir_intrinsic_deref_atomic_fmin
;
943 unreachable("Invalid type");
945 case ir_intrinsic_generic_atomic_max
:
946 assert(ir
->return_deref
);
947 if (ir
->return_deref
->type
== glsl_type::int_type
)
948 op
= nir_intrinsic_deref_atomic_imax
;
949 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
950 op
= nir_intrinsic_deref_atomic_umax
;
951 else if (ir
->return_deref
->type
== glsl_type::float_type
)
952 op
= nir_intrinsic_deref_atomic_fmax
;
954 unreachable("Invalid type");
956 case ir_intrinsic_generic_atomic_exchange
:
957 op
= nir_intrinsic_deref_atomic_exchange
;
959 case ir_intrinsic_generic_atomic_comp_swap
:
960 op
= ir
->return_deref
->type
->is_integer_32_64()
961 ? nir_intrinsic_deref_atomic_comp_swap
962 : nir_intrinsic_deref_atomic_fcomp_swap
;
964 case ir_intrinsic_atomic_counter_read
:
965 op
= nir_intrinsic_atomic_counter_read_deref
;
967 case ir_intrinsic_atomic_counter_increment
:
968 op
= nir_intrinsic_atomic_counter_inc_deref
;
970 case ir_intrinsic_atomic_counter_predecrement
:
971 op
= nir_intrinsic_atomic_counter_pre_dec_deref
;
973 case ir_intrinsic_atomic_counter_add
:
974 op
= nir_intrinsic_atomic_counter_add_deref
;
976 case ir_intrinsic_atomic_counter_and
:
977 op
= nir_intrinsic_atomic_counter_and_deref
;
979 case ir_intrinsic_atomic_counter_or
:
980 op
= nir_intrinsic_atomic_counter_or_deref
;
982 case ir_intrinsic_atomic_counter_xor
:
983 op
= nir_intrinsic_atomic_counter_xor_deref
;
985 case ir_intrinsic_atomic_counter_min
:
986 op
= nir_intrinsic_atomic_counter_min_deref
;
988 case ir_intrinsic_atomic_counter_max
:
989 op
= nir_intrinsic_atomic_counter_max_deref
;
991 case ir_intrinsic_atomic_counter_exchange
:
992 op
= nir_intrinsic_atomic_counter_exchange_deref
;
994 case ir_intrinsic_atomic_counter_comp_swap
:
995 op
= nir_intrinsic_atomic_counter_comp_swap_deref
;
997 case ir_intrinsic_image_load
:
998 op
= nir_intrinsic_image_deref_load
;
1000 case ir_intrinsic_image_store
:
1001 op
= nir_intrinsic_image_deref_store
;
1003 case ir_intrinsic_image_atomic_add
:
1004 op
= ir
->return_deref
->type
->is_integer_32_64()
1005 ? nir_intrinsic_image_deref_atomic_add
1006 : nir_intrinsic_image_deref_atomic_fadd
;
1008 case ir_intrinsic_image_atomic_min
:
1009 if (ir
->return_deref
->type
== glsl_type::int_type
)
1010 op
= nir_intrinsic_image_deref_atomic_imin
;
1011 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1012 op
= nir_intrinsic_image_deref_atomic_umin
;
1014 unreachable("Invalid type");
1016 case ir_intrinsic_image_atomic_max
:
1017 if (ir
->return_deref
->type
== glsl_type::int_type
)
1018 op
= nir_intrinsic_image_deref_atomic_imax
;
1019 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1020 op
= nir_intrinsic_image_deref_atomic_umax
;
1022 unreachable("Invalid type");
1024 case ir_intrinsic_image_atomic_and
:
1025 op
= nir_intrinsic_image_deref_atomic_and
;
1027 case ir_intrinsic_image_atomic_or
:
1028 op
= nir_intrinsic_image_deref_atomic_or
;
1030 case ir_intrinsic_image_atomic_xor
:
1031 op
= nir_intrinsic_image_deref_atomic_xor
;
1033 case ir_intrinsic_image_atomic_exchange
:
1034 op
= nir_intrinsic_image_deref_atomic_exchange
;
1036 case ir_intrinsic_image_atomic_comp_swap
:
1037 op
= nir_intrinsic_image_deref_atomic_comp_swap
;
1039 case ir_intrinsic_image_atomic_inc_wrap
:
1040 op
= nir_intrinsic_image_deref_atomic_inc_wrap
;
1042 case ir_intrinsic_image_atomic_dec_wrap
:
1043 op
= nir_intrinsic_image_deref_atomic_dec_wrap
;
1045 case ir_intrinsic_memory_barrier
:
1046 op
= nir_intrinsic_memory_barrier
;
1048 case ir_intrinsic_image_size
:
1049 op
= nir_intrinsic_image_deref_size
;
1051 case ir_intrinsic_image_samples
:
1052 op
= nir_intrinsic_image_deref_samples
;
1054 case ir_intrinsic_ssbo_store
:
1055 op
= nir_intrinsic_store_ssbo
;
1057 case ir_intrinsic_ssbo_load
:
1058 op
= nir_intrinsic_load_ssbo
;
1060 case ir_intrinsic_ssbo_atomic_add
:
1061 op
= ir
->return_deref
->type
->is_integer_32_64()
1062 ? nir_intrinsic_ssbo_atomic_add
: nir_intrinsic_ssbo_atomic_fadd
;
1064 case ir_intrinsic_ssbo_atomic_and
:
1065 op
= nir_intrinsic_ssbo_atomic_and
;
1067 case ir_intrinsic_ssbo_atomic_or
:
1068 op
= nir_intrinsic_ssbo_atomic_or
;
1070 case ir_intrinsic_ssbo_atomic_xor
:
1071 op
= nir_intrinsic_ssbo_atomic_xor
;
1073 case ir_intrinsic_ssbo_atomic_min
:
1074 assert(ir
->return_deref
);
1075 if (ir
->return_deref
->type
== glsl_type::int_type
)
1076 op
= nir_intrinsic_ssbo_atomic_imin
;
1077 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1078 op
= nir_intrinsic_ssbo_atomic_umin
;
1079 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1080 op
= nir_intrinsic_ssbo_atomic_fmin
;
1082 unreachable("Invalid type");
1084 case ir_intrinsic_ssbo_atomic_max
:
1085 assert(ir
->return_deref
);
1086 if (ir
->return_deref
->type
== glsl_type::int_type
)
1087 op
= nir_intrinsic_ssbo_atomic_imax
;
1088 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1089 op
= nir_intrinsic_ssbo_atomic_umax
;
1090 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1091 op
= nir_intrinsic_ssbo_atomic_fmax
;
1093 unreachable("Invalid type");
1095 case ir_intrinsic_ssbo_atomic_exchange
:
1096 op
= nir_intrinsic_ssbo_atomic_exchange
;
1098 case ir_intrinsic_ssbo_atomic_comp_swap
:
1099 op
= ir
->return_deref
->type
->is_integer_32_64()
1100 ? nir_intrinsic_ssbo_atomic_comp_swap
1101 : nir_intrinsic_ssbo_atomic_fcomp_swap
;
1103 case ir_intrinsic_shader_clock
:
1104 op
= nir_intrinsic_shader_clock
;
1106 case ir_intrinsic_begin_invocation_interlock
:
1107 op
= nir_intrinsic_begin_invocation_interlock
;
1109 case ir_intrinsic_end_invocation_interlock
:
1110 op
= nir_intrinsic_end_invocation_interlock
;
1112 case ir_intrinsic_group_memory_barrier
:
1113 op
= nir_intrinsic_group_memory_barrier
;
1115 case ir_intrinsic_memory_barrier_atomic_counter
:
1116 op
= nir_intrinsic_memory_barrier_atomic_counter
;
1118 case ir_intrinsic_memory_barrier_buffer
:
1119 op
= nir_intrinsic_memory_barrier_buffer
;
1121 case ir_intrinsic_memory_barrier_image
:
1122 op
= nir_intrinsic_memory_barrier_image
;
1124 case ir_intrinsic_memory_barrier_shared
:
1125 op
= nir_intrinsic_memory_barrier_shared
;
1127 case ir_intrinsic_shared_load
:
1128 op
= nir_intrinsic_load_shared
;
1130 case ir_intrinsic_shared_store
:
1131 op
= nir_intrinsic_store_shared
;
1133 case ir_intrinsic_shared_atomic_add
:
1134 op
= ir
->return_deref
->type
->is_integer_32_64()
1135 ? nir_intrinsic_shared_atomic_add
1136 : nir_intrinsic_shared_atomic_fadd
;
1138 case ir_intrinsic_shared_atomic_and
:
1139 op
= nir_intrinsic_shared_atomic_and
;
1141 case ir_intrinsic_shared_atomic_or
:
1142 op
= nir_intrinsic_shared_atomic_or
;
1144 case ir_intrinsic_shared_atomic_xor
:
1145 op
= nir_intrinsic_shared_atomic_xor
;
1147 case ir_intrinsic_shared_atomic_min
:
1148 assert(ir
->return_deref
);
1149 if (ir
->return_deref
->type
== glsl_type::int_type
)
1150 op
= nir_intrinsic_shared_atomic_imin
;
1151 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1152 op
= nir_intrinsic_shared_atomic_umin
;
1153 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1154 op
= nir_intrinsic_shared_atomic_fmin
;
1156 unreachable("Invalid type");
1158 case ir_intrinsic_shared_atomic_max
:
1159 assert(ir
->return_deref
);
1160 if (ir
->return_deref
->type
== glsl_type::int_type
)
1161 op
= nir_intrinsic_shared_atomic_imax
;
1162 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1163 op
= nir_intrinsic_shared_atomic_umax
;
1164 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1165 op
= nir_intrinsic_shared_atomic_fmax
;
1167 unreachable("Invalid type");
1169 case ir_intrinsic_shared_atomic_exchange
:
1170 op
= nir_intrinsic_shared_atomic_exchange
;
1172 case ir_intrinsic_shared_atomic_comp_swap
:
1173 op
= ir
->return_deref
->type
->is_integer_32_64()
1174 ? nir_intrinsic_shared_atomic_comp_swap
1175 : nir_intrinsic_shared_atomic_fcomp_swap
;
1177 case ir_intrinsic_vote_any
:
1178 op
= nir_intrinsic_vote_any
;
1180 case ir_intrinsic_vote_all
:
1181 op
= nir_intrinsic_vote_all
;
1183 case ir_intrinsic_vote_eq
:
1184 op
= nir_intrinsic_vote_ieq
;
1186 case ir_intrinsic_ballot
:
1187 op
= nir_intrinsic_ballot
;
1189 case ir_intrinsic_read_invocation
:
1190 op
= nir_intrinsic_read_invocation
;
1192 case ir_intrinsic_read_first_invocation
:
1193 op
= nir_intrinsic_read_first_invocation
;
1195 case ir_intrinsic_helper_invocation
:
1196 op
= nir_intrinsic_is_helper_invocation
;
1199 unreachable("not reached");
1202 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(shader
, op
);
1203 nir_ssa_def
*ret
= &instr
->dest
.ssa
;
1206 case nir_intrinsic_deref_atomic_add
:
1207 case nir_intrinsic_deref_atomic_imin
:
1208 case nir_intrinsic_deref_atomic_umin
:
1209 case nir_intrinsic_deref_atomic_imax
:
1210 case nir_intrinsic_deref_atomic_umax
:
1211 case nir_intrinsic_deref_atomic_and
:
1212 case nir_intrinsic_deref_atomic_or
:
1213 case nir_intrinsic_deref_atomic_xor
:
1214 case nir_intrinsic_deref_atomic_exchange
:
1215 case nir_intrinsic_deref_atomic_comp_swap
:
1216 case nir_intrinsic_deref_atomic_fadd
:
1217 case nir_intrinsic_deref_atomic_fmin
:
1218 case nir_intrinsic_deref_atomic_fmax
:
1219 case nir_intrinsic_deref_atomic_fcomp_swap
: {
1220 int param_count
= ir
->actual_parameters
.length();
1221 assert(param_count
== 2 || param_count
== 3);
1224 exec_node
*param
= ir
->actual_parameters
.get_head();
1225 ir_rvalue
*rvalue
= (ir_rvalue
*) param
;
1226 ir_dereference
*deref
= rvalue
->as_dereference();
1227 ir_swizzle
*swizzle
= NULL
;
1229 /* We may have a swizzle to pick off a single vec4 component */
1230 swizzle
= rvalue
->as_swizzle();
1231 assert(swizzle
&& swizzle
->type
->vector_elements
== 1);
1232 deref
= swizzle
->val
->as_dereference();
1235 nir_deref_instr
*nir_deref
= evaluate_deref(deref
);
1237 nir_deref
= nir_build_deref_array_imm(&b
, nir_deref
,
1240 instr
->src
[0] = nir_src_for_ssa(&nir_deref
->dest
.ssa
);
1242 nir_intrinsic_set_access(instr
, deref_get_qualifier(nir_deref
));
1244 /* data1 parameter (this is always present) */
1245 param
= param
->get_next();
1246 ir_instruction
*inst
= (ir_instruction
*) param
;
1247 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1249 /* data2 parameter (only with atomic_comp_swap) */
1250 if (param_count
== 3) {
1251 assert(op
== nir_intrinsic_deref_atomic_comp_swap
||
1252 op
== nir_intrinsic_deref_atomic_fcomp_swap
);
1253 param
= param
->get_next();
1254 inst
= (ir_instruction
*) param
;
1255 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1259 assert(ir
->return_deref
);
1260 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1261 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1262 nir_builder_instr_insert(&b
, &instr
->instr
);
1265 case nir_intrinsic_atomic_counter_read_deref
:
1266 case nir_intrinsic_atomic_counter_inc_deref
:
1267 case nir_intrinsic_atomic_counter_pre_dec_deref
:
1268 case nir_intrinsic_atomic_counter_add_deref
:
1269 case nir_intrinsic_atomic_counter_min_deref
:
1270 case nir_intrinsic_atomic_counter_max_deref
:
1271 case nir_intrinsic_atomic_counter_and_deref
:
1272 case nir_intrinsic_atomic_counter_or_deref
:
1273 case nir_intrinsic_atomic_counter_xor_deref
:
1274 case nir_intrinsic_atomic_counter_exchange_deref
:
1275 case nir_intrinsic_atomic_counter_comp_swap_deref
: {
1276 /* Set the counter variable dereference. */
1277 exec_node
*param
= ir
->actual_parameters
.get_head();
1278 ir_dereference
*counter
= (ir_dereference
*)param
;
1280 instr
->src
[0] = nir_src_for_ssa(&evaluate_deref(counter
)->dest
.ssa
);
1281 param
= param
->get_next();
1283 /* Set the intrinsic destination. */
1284 if (ir
->return_deref
) {
1285 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1288 /* Set the intrinsic parameters. */
1289 if (!param
->is_tail_sentinel()) {
1291 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1292 param
= param
->get_next();
1295 if (!param
->is_tail_sentinel()) {
1297 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1298 param
= param
->get_next();
1301 nir_builder_instr_insert(&b
, &instr
->instr
);
1304 case nir_intrinsic_image_deref_load
:
1305 case nir_intrinsic_image_deref_store
:
1306 case nir_intrinsic_image_deref_atomic_add
:
1307 case nir_intrinsic_image_deref_atomic_imin
:
1308 case nir_intrinsic_image_deref_atomic_umin
:
1309 case nir_intrinsic_image_deref_atomic_imax
:
1310 case nir_intrinsic_image_deref_atomic_umax
:
1311 case nir_intrinsic_image_deref_atomic_and
:
1312 case nir_intrinsic_image_deref_atomic_or
:
1313 case nir_intrinsic_image_deref_atomic_xor
:
1314 case nir_intrinsic_image_deref_atomic_exchange
:
1315 case nir_intrinsic_image_deref_atomic_comp_swap
:
1316 case nir_intrinsic_image_deref_atomic_fadd
:
1317 case nir_intrinsic_image_deref_samples
:
1318 case nir_intrinsic_image_deref_size
:
1319 case nir_intrinsic_image_deref_atomic_inc_wrap
:
1320 case nir_intrinsic_image_deref_atomic_dec_wrap
: {
1321 nir_ssa_undef_instr
*instr_undef
=
1322 nir_ssa_undef_instr_create(shader
, 1, 32);
1323 nir_builder_instr_insert(&b
, &instr_undef
->instr
);
1325 /* Set the image variable dereference. */
1326 exec_node
*param
= ir
->actual_parameters
.get_head();
1327 ir_dereference
*image
= (ir_dereference
*)param
;
1328 nir_deref_instr
*deref
= evaluate_deref(image
);
1329 const glsl_type
*type
= deref
->type
;
1331 nir_intrinsic_set_access(instr
, deref_get_qualifier(deref
));
1333 instr
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
1334 param
= param
->get_next();
1336 /* Set the intrinsic destination. */
1337 if (ir
->return_deref
) {
1338 unsigned num_components
= ir
->return_deref
->type
->vector_elements
;
1339 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1340 num_components
, 32, NULL
);
1343 if (op
== nir_intrinsic_image_deref_size
) {
1344 instr
->num_components
= instr
->dest
.ssa
.num_components
;
1345 } else if (op
== nir_intrinsic_image_deref_load
||
1346 op
== nir_intrinsic_image_deref_store
) {
1347 instr
->num_components
= 4;
1350 if (op
== nir_intrinsic_image_deref_size
||
1351 op
== nir_intrinsic_image_deref_samples
) {
1352 nir_builder_instr_insert(&b
, &instr
->instr
);
1356 /* Set the address argument, extending the coordinate vector to four
1359 nir_ssa_def
*src_addr
=
1360 evaluate_rvalue((ir_dereference
*)param
);
1361 nir_ssa_def
*srcs
[4];
1363 for (int i
= 0; i
< 4; i
++) {
1364 if (i
< type
->coordinate_components())
1365 srcs
[i
] = nir_channel(&b
, src_addr
, i
);
1367 srcs
[i
] = &instr_undef
->def
;
1370 instr
->src
[1] = nir_src_for_ssa(nir_vec(&b
, srcs
, 4));
1371 param
= param
->get_next();
1373 /* Set the sample argument, which is undefined for single-sample
1376 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_MS
) {
1378 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1379 param
= param
->get_next();
1381 instr
->src
[2] = nir_src_for_ssa(&instr_undef
->def
);
1384 /* Set the intrinsic parameters. */
1385 if (!param
->is_tail_sentinel()) {
1387 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1388 param
= param
->get_next();
1389 } else if (op
== nir_intrinsic_image_deref_load
) {
1390 instr
->src
[3] = nir_src_for_ssa(nir_imm_int(&b
, 0)); /* LOD */
1393 if (!param
->is_tail_sentinel()) {
1395 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1396 param
= param
->get_next();
1397 } else if (op
== nir_intrinsic_image_deref_store
) {
1398 instr
->src
[4] = nir_src_for_ssa(nir_imm_int(&b
, 0)); /* LOD */
1401 nir_builder_instr_insert(&b
, &instr
->instr
);
1404 case nir_intrinsic_memory_barrier
:
1405 case nir_intrinsic_group_memory_barrier
:
1406 case nir_intrinsic_memory_barrier_atomic_counter
:
1407 case nir_intrinsic_memory_barrier_buffer
:
1408 case nir_intrinsic_memory_barrier_image
:
1409 case nir_intrinsic_memory_barrier_shared
:
1410 nir_builder_instr_insert(&b
, &instr
->instr
);
1412 case nir_intrinsic_shader_clock
:
1413 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 2, 32, NULL
);
1414 instr
->num_components
= 2;
1415 nir_builder_instr_insert(&b
, &instr
->instr
);
1417 case nir_intrinsic_begin_invocation_interlock
:
1418 nir_builder_instr_insert(&b
, &instr
->instr
);
1420 case nir_intrinsic_end_invocation_interlock
:
1421 nir_builder_instr_insert(&b
, &instr
->instr
);
1423 case nir_intrinsic_store_ssbo
: {
1424 exec_node
*param
= ir
->actual_parameters
.get_head();
1425 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1427 param
= param
->get_next();
1428 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1430 param
= param
->get_next();
1431 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1433 param
= param
->get_next();
1434 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1437 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1438 if (val
->type
->is_boolean())
1439 nir_val
= nir_b2i32(&b
, nir_val
);
1441 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1442 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(block
));
1443 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(offset
));
1444 intrinsic_set_std430_align(instr
, val
->type
);
1445 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1446 instr
->num_components
= val
->type
->vector_elements
;
1448 nir_builder_instr_insert(&b
, &instr
->instr
);
1451 case nir_intrinsic_load_ssbo
: {
1452 exec_node
*param
= ir
->actual_parameters
.get_head();
1453 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1455 param
= param
->get_next();
1456 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1458 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(block
));
1459 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1461 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1462 instr
->num_components
= type
->vector_elements
;
1463 intrinsic_set_std430_align(instr
, type
);
1465 /* Setup destination register */
1466 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1467 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1468 type
->vector_elements
, bit_size
, NULL
);
1470 /* Insert the created nir instruction now since in the case of boolean
1471 * result we will need to emit another instruction after it
1473 nir_builder_instr_insert(&b
, &instr
->instr
);
1476 * In SSBO/UBO's, a true boolean value is any non-zero value, but we
1477 * consider a true boolean to be ~0. Fix this up with a != 0
1480 if (type
->is_boolean())
1481 ret
= nir_i2b(&b
, &instr
->dest
.ssa
);
1484 case nir_intrinsic_ssbo_atomic_add
:
1485 case nir_intrinsic_ssbo_atomic_imin
:
1486 case nir_intrinsic_ssbo_atomic_umin
:
1487 case nir_intrinsic_ssbo_atomic_imax
:
1488 case nir_intrinsic_ssbo_atomic_umax
:
1489 case nir_intrinsic_ssbo_atomic_and
:
1490 case nir_intrinsic_ssbo_atomic_or
:
1491 case nir_intrinsic_ssbo_atomic_xor
:
1492 case nir_intrinsic_ssbo_atomic_exchange
:
1493 case nir_intrinsic_ssbo_atomic_comp_swap
:
1494 case nir_intrinsic_ssbo_atomic_fadd
:
1495 case nir_intrinsic_ssbo_atomic_fmin
:
1496 case nir_intrinsic_ssbo_atomic_fmax
:
1497 case nir_intrinsic_ssbo_atomic_fcomp_swap
: {
1498 int param_count
= ir
->actual_parameters
.length();
1499 assert(param_count
== 3 || param_count
== 4);
1502 exec_node
*param
= ir
->actual_parameters
.get_head();
1503 ir_instruction
*inst
= (ir_instruction
*) param
;
1504 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1507 param
= param
->get_next();
1508 inst
= (ir_instruction
*) param
;
1509 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1511 /* data1 parameter (this is always present) */
1512 param
= param
->get_next();
1513 inst
= (ir_instruction
*) param
;
1514 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1516 /* data2 parameter (only with atomic_comp_swap) */
1517 if (param_count
== 4) {
1518 assert(op
== nir_intrinsic_ssbo_atomic_comp_swap
||
1519 op
== nir_intrinsic_ssbo_atomic_fcomp_swap
);
1520 param
= param
->get_next();
1521 inst
= (ir_instruction
*) param
;
1522 instr
->src
[3] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1526 assert(ir
->return_deref
);
1527 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1528 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1529 nir_builder_instr_insert(&b
, &instr
->instr
);
1532 case nir_intrinsic_load_shared
: {
1533 exec_node
*param
= ir
->actual_parameters
.get_head();
1534 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1536 nir_intrinsic_set_base(instr
, 0);
1537 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(offset
));
1539 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1540 instr
->num_components
= type
->vector_elements
;
1541 intrinsic_set_std430_align(instr
, type
);
1543 /* Setup destination register */
1544 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1545 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1546 type
->vector_elements
, bit_size
, NULL
);
1548 nir_builder_instr_insert(&b
, &instr
->instr
);
1550 /* The value in shared memory is a 32-bit value */
1551 if (type
->is_boolean())
1552 ret
= nir_i2b(&b
, &instr
->dest
.ssa
);
1555 case nir_intrinsic_store_shared
: {
1556 exec_node
*param
= ir
->actual_parameters
.get_head();
1557 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1559 param
= param
->get_next();
1560 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1562 param
= param
->get_next();
1563 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1566 nir_intrinsic_set_base(instr
, 0);
1567 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1569 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1571 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1572 /* The value in shared memory is a 32-bit value */
1573 if (val
->type
->is_boolean())
1574 nir_val
= nir_b2i32(&b
, nir_val
);
1576 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1577 instr
->num_components
= val
->type
->vector_elements
;
1578 intrinsic_set_std430_align(instr
, val
->type
);
1580 nir_builder_instr_insert(&b
, &instr
->instr
);
1583 case nir_intrinsic_shared_atomic_add
:
1584 case nir_intrinsic_shared_atomic_imin
:
1585 case nir_intrinsic_shared_atomic_umin
:
1586 case nir_intrinsic_shared_atomic_imax
:
1587 case nir_intrinsic_shared_atomic_umax
:
1588 case nir_intrinsic_shared_atomic_and
:
1589 case nir_intrinsic_shared_atomic_or
:
1590 case nir_intrinsic_shared_atomic_xor
:
1591 case nir_intrinsic_shared_atomic_exchange
:
1592 case nir_intrinsic_shared_atomic_comp_swap
:
1593 case nir_intrinsic_shared_atomic_fadd
:
1594 case nir_intrinsic_shared_atomic_fmin
:
1595 case nir_intrinsic_shared_atomic_fmax
:
1596 case nir_intrinsic_shared_atomic_fcomp_swap
: {
1597 int param_count
= ir
->actual_parameters
.length();
1598 assert(param_count
== 2 || param_count
== 3);
1601 exec_node
*param
= ir
->actual_parameters
.get_head();
1602 ir_instruction
*inst
= (ir_instruction
*) param
;
1603 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1605 /* data1 parameter (this is always present) */
1606 param
= param
->get_next();
1607 inst
= (ir_instruction
*) param
;
1608 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1610 /* data2 parameter (only with atomic_comp_swap) */
1611 if (param_count
== 3) {
1612 assert(op
== nir_intrinsic_shared_atomic_comp_swap
||
1613 op
== nir_intrinsic_shared_atomic_fcomp_swap
);
1614 param
= param
->get_next();
1615 inst
= (ir_instruction
*) param
;
1617 nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1621 assert(ir
->return_deref
);
1622 unsigned bit_size
= glsl_get_bit_size(ir
->return_deref
->type
);
1623 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1624 ir
->return_deref
->type
->vector_elements
,
1626 nir_builder_instr_insert(&b
, &instr
->instr
);
1629 case nir_intrinsic_vote_any
:
1630 case nir_intrinsic_vote_all
:
1631 case nir_intrinsic_vote_ieq
: {
1632 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 1, NULL
);
1633 instr
->num_components
= 1;
1635 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1636 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1638 nir_builder_instr_insert(&b
, &instr
->instr
);
1642 case nir_intrinsic_ballot
: {
1643 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1644 ir
->return_deref
->type
->vector_elements
, 64, 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 nir_builder_instr_insert(&b
, &instr
->instr
);
1653 case nir_intrinsic_read_invocation
: {
1654 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1655 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1656 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1658 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1659 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1661 ir_rvalue
*invocation
= (ir_rvalue
*) ir
->actual_parameters
.get_head()->next
;
1662 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(invocation
));
1664 nir_builder_instr_insert(&b
, &instr
->instr
);
1667 case nir_intrinsic_read_first_invocation
: {
1668 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1669 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1670 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1672 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1673 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1675 nir_builder_instr_insert(&b
, &instr
->instr
);
1678 case nir_intrinsic_is_helper_invocation
: {
1679 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 1, NULL
);
1680 instr
->num_components
= 1;
1681 nir_builder_instr_insert(&b
, &instr
->instr
);
1685 unreachable("not reached");
1688 if (ir
->return_deref
)
1689 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), ret
, ~0);
1694 struct hash_entry
*entry
=
1695 _mesa_hash_table_search(this->overload_table
, ir
->callee
);
1697 nir_function
*callee
= (nir_function
*) entry
->data
;
1699 nir_call_instr
*call
= nir_call_instr_create(this->shader
, callee
);
1702 nir_deref_instr
*ret_deref
= NULL
;
1703 if (ir
->return_deref
) {
1704 nir_variable
*ret_tmp
=
1705 nir_local_variable_create(this->impl
, ir
->return_deref
->type
,
1707 ret_deref
= nir_build_deref_var(&b
, ret_tmp
);
1708 call
->params
[i
++] = nir_src_for_ssa(&ret_deref
->dest
.ssa
);
1711 foreach_two_lists(formal_node
, &ir
->callee
->parameters
,
1712 actual_node
, &ir
->actual_parameters
) {
1713 ir_rvalue
*param_rvalue
= (ir_rvalue
*) actual_node
;
1714 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
1716 if (sig_param
->data
.mode
== ir_var_function_out
) {
1717 nir_deref_instr
*out_deref
= evaluate_deref(param_rvalue
);
1718 call
->params
[i
] = nir_src_for_ssa(&out_deref
->dest
.ssa
);
1719 } else if (sig_param
->data
.mode
== ir_var_function_in
) {
1720 nir_ssa_def
*val
= evaluate_rvalue(param_rvalue
);
1721 nir_src src
= nir_src_for_ssa(val
);
1723 nir_src_copy(&call
->params
[i
], &src
, call
);
1724 } else if (sig_param
->data
.mode
== ir_var_function_inout
) {
1725 unreachable("unimplemented: inout parameters");
1731 nir_builder_instr_insert(&b
, &call
->instr
);
1733 if (ir
->return_deref
)
1734 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), nir_load_deref(&b
, ret_deref
), ~0);
1738 nir_visitor::visit(ir_assignment
*ir
)
1740 unsigned num_components
= ir
->lhs
->type
->vector_elements
;
1742 b
.exact
= ir
->lhs
->variable_referenced()->data
.invariant
||
1743 ir
->lhs
->variable_referenced()->data
.precise
;
1745 if ((ir
->rhs
->as_dereference() || ir
->rhs
->as_constant()) &&
1746 (ir
->write_mask
== (1 << num_components
) - 1 || ir
->write_mask
== 0)) {
1747 nir_deref_instr
*lhs
= evaluate_deref(ir
->lhs
);
1748 nir_deref_instr
*rhs
= evaluate_deref(ir
->rhs
);
1749 enum gl_access_qualifier lhs_qualifiers
= deref_get_qualifier(lhs
);
1750 enum gl_access_qualifier rhs_qualifiers
= deref_get_qualifier(rhs
);
1751 if (ir
->condition
) {
1752 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1753 nir_copy_deref_with_access(&b
, lhs
, rhs
, lhs_qualifiers
,
1755 nir_pop_if(&b
, NULL
);
1757 nir_copy_deref_with_access(&b
, lhs
, rhs
, lhs_qualifiers
,
1763 assert(ir
->rhs
->type
->is_scalar() || ir
->rhs
->type
->is_vector());
1765 ir
->lhs
->accept(this);
1766 nir_deref_instr
*lhs_deref
= this->deref
;
1767 nir_ssa_def
*src
= evaluate_rvalue(ir
->rhs
);
1769 if (ir
->write_mask
!= (1 << num_components
) - 1 && ir
->write_mask
!= 0) {
1770 /* GLSL IR will give us the input to the write-masked assignment in a
1771 * single packed vector. So, for example, if the writemask is xzw, then
1772 * we have to swizzle x -> x, y -> z, and z -> w and get the y component
1776 unsigned component
= 0;
1777 for (unsigned i
= 0; i
< 4; i
++) {
1778 swiz
[i
] = ir
->write_mask
& (1 << i
) ? component
++ : 0;
1780 src
= nir_swizzle(&b
, src
, swiz
, num_components
);
1783 enum gl_access_qualifier qualifiers
= deref_get_qualifier(lhs_deref
);
1784 if (ir
->condition
) {
1785 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1786 nir_store_deref_with_access(&b
, lhs_deref
, src
, ir
->write_mask
,
1788 nir_pop_if(&b
, NULL
);
1790 nir_store_deref_with_access(&b
, lhs_deref
, src
, ir
->write_mask
,
1796 * Given an instruction, returns a pointer to its destination or NULL if there
1797 * is no destination.
1799 * Note that this only handles instructions we generate at this level.
1802 get_instr_dest(nir_instr
*instr
)
1804 nir_alu_instr
*alu_instr
;
1805 nir_intrinsic_instr
*intrinsic_instr
;
1806 nir_tex_instr
*tex_instr
;
1808 switch (instr
->type
) {
1809 case nir_instr_type_alu
:
1810 alu_instr
= nir_instr_as_alu(instr
);
1811 return &alu_instr
->dest
.dest
;
1813 case nir_instr_type_intrinsic
:
1814 intrinsic_instr
= nir_instr_as_intrinsic(instr
);
1815 if (nir_intrinsic_infos
[intrinsic_instr
->intrinsic
].has_dest
)
1816 return &intrinsic_instr
->dest
;
1820 case nir_instr_type_tex
:
1821 tex_instr
= nir_instr_as_tex(instr
);
1822 return &tex_instr
->dest
;
1825 unreachable("not reached");
1832 nir_visitor::add_instr(nir_instr
*instr
, unsigned num_components
,
1835 nir_dest
*dest
= get_instr_dest(instr
);
1838 nir_ssa_dest_init(instr
, dest
, num_components
, bit_size
, NULL
);
1840 nir_builder_instr_insert(&b
, instr
);
1843 assert(dest
->is_ssa
);
1844 this->result
= &dest
->ssa
;
1849 nir_visitor::evaluate_rvalue(ir_rvalue
* ir
)
1852 if (ir
->as_dereference() || ir
->as_constant()) {
1854 * A dereference is being used on the right hand side, which means we
1855 * must emit a variable load.
1858 enum gl_access_qualifier access
= deref_get_qualifier(this->deref
);
1859 this->result
= nir_load_deref_with_access(&b
, this->deref
, access
);
1862 return this->result
;
1866 type_is_float(glsl_base_type type
)
1868 return type
== GLSL_TYPE_FLOAT
|| type
== GLSL_TYPE_DOUBLE
||
1869 type
== GLSL_TYPE_FLOAT16
;
1873 type_is_signed(glsl_base_type type
)
1875 return type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_INT64
||
1876 type
== GLSL_TYPE_INT16
;
1880 nir_visitor::visit(ir_expression
*ir
)
1882 /* Some special cases */
1883 switch (ir
->operation
) {
1884 case ir_binop_ubo_load
: {
1885 nir_intrinsic_instr
*load
=
1886 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_ubo
);
1887 unsigned bit_size
= ir
->type
->is_boolean() ? 32 :
1888 glsl_get_bit_size(ir
->type
);
1889 load
->num_components
= ir
->type
->vector_elements
;
1890 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1891 load
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1892 intrinsic_set_std430_align(load
, ir
->type
);
1893 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1896 * In UBO's, a true boolean value is any non-zero value, but we consider
1897 * a true boolean to be ~0. Fix this up with a != 0 comparison.
1900 if (ir
->type
->is_boolean())
1901 this->result
= nir_i2b(&b
, &load
->dest
.ssa
);
1906 case ir_unop_interpolate_at_centroid
:
1907 case ir_binop_interpolate_at_offset
:
1908 case ir_binop_interpolate_at_sample
: {
1909 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1910 ir_swizzle
*swizzle
= NULL
;
1912 /* the api does not allow a swizzle here, but the varying packing code
1913 * may have pushed one into here.
1915 swizzle
= ir
->operands
[0]->as_swizzle();
1917 deref
= swizzle
->val
->as_dereference();
1921 deref
->accept(this);
1923 nir_intrinsic_op op
;
1924 if (this->deref
->mode
== nir_var_shader_in
) {
1925 switch (ir
->operation
) {
1926 case ir_unop_interpolate_at_centroid
:
1927 op
= nir_intrinsic_interp_deref_at_centroid
;
1929 case ir_binop_interpolate_at_offset
:
1930 op
= nir_intrinsic_interp_deref_at_offset
;
1932 case ir_binop_interpolate_at_sample
:
1933 op
= nir_intrinsic_interp_deref_at_sample
;
1936 unreachable("Invalid interpolation intrinsic");
1939 /* This case can happen if the vertex shader does not write the
1940 * given varying. In this case, the linker will lower it to a
1941 * global variable. Since interpolating a variable makes no
1942 * sense, we'll just turn it into a load which will probably
1943 * eventually end up as an SSA definition.
1945 assert(this->deref
->mode
== nir_var_shader_temp
);
1946 op
= nir_intrinsic_load_deref
;
1949 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(shader
, op
);
1950 intrin
->num_components
= deref
->type
->vector_elements
;
1951 intrin
->src
[0] = nir_src_for_ssa(&this->deref
->dest
.ssa
);
1953 if (intrin
->intrinsic
== nir_intrinsic_interp_deref_at_offset
||
1954 intrin
->intrinsic
== nir_intrinsic_interp_deref_at_sample
)
1955 intrin
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1957 unsigned bit_size
= glsl_get_bit_size(deref
->type
);
1958 add_instr(&intrin
->instr
, deref
->type
->vector_elements
, bit_size
);
1961 unsigned swiz
[4] = {
1962 swizzle
->mask
.x
, swizzle
->mask
.y
, swizzle
->mask
.z
, swizzle
->mask
.w
1965 result
= nir_swizzle(&b
, result
, swiz
,
1966 swizzle
->type
->vector_elements
);
1972 case ir_unop_ssbo_unsized_array_length
: {
1973 nir_intrinsic_instr
*intrin
=
1974 nir_intrinsic_instr_create(b
.shader
,
1975 nir_intrinsic_deref_buffer_array_length
);
1977 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1978 intrin
->src
[0] = nir_src_for_ssa(&evaluate_deref(deref
)->dest
.ssa
);
1980 add_instr(&intrin
->instr
, 1, 32);
1988 nir_ssa_def
*srcs
[4];
1989 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1990 srcs
[i
] = evaluate_rvalue(ir
->operands
[i
]);
1992 glsl_base_type types
[4];
1993 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1994 types
[i
] = ir
->operands
[i
]->type
->base_type
;
1996 glsl_base_type out_type
= ir
->type
->base_type
;
1998 switch (ir
->operation
) {
1999 case ir_unop_bit_not
: result
= nir_inot(&b
, srcs
[0]); break;
2000 case ir_unop_logic_not
:
2001 result
= nir_inot(&b
, srcs
[0]);
2004 result
= type_is_float(types
[0]) ? nir_fneg(&b
, srcs
[0])
2005 : nir_ineg(&b
, srcs
[0]);
2008 result
= type_is_float(types
[0]) ? nir_fabs(&b
, srcs
[0])
2009 : nir_iabs(&b
, srcs
[0]);
2012 result
= nir_uclz(&b
, srcs
[0]);
2014 case ir_unop_saturate
:
2015 assert(type_is_float(types
[0]));
2016 result
= nir_fsat(&b
, srcs
[0]);
2019 result
= type_is_float(types
[0]) ? nir_fsign(&b
, srcs
[0])
2020 : nir_isign(&b
, srcs
[0]);
2022 case ir_unop_rcp
: result
= nir_frcp(&b
, srcs
[0]); break;
2023 case ir_unop_rsq
: result
= nir_frsq(&b
, srcs
[0]); break;
2024 case ir_unop_sqrt
: result
= nir_fsqrt(&b
, srcs
[0]); break;
2025 case ir_unop_exp
: unreachable("ir_unop_exp should have been lowered");
2026 case ir_unop_log
: unreachable("ir_unop_log should have been lowered");
2027 case ir_unop_exp2
: result
= nir_fexp2(&b
, srcs
[0]); break;
2028 case ir_unop_log2
: result
= nir_flog2(&b
, srcs
[0]); break;
2068 case ir_unop_i642u64
:
2069 case ir_unop_u642i64
: {
2070 nir_alu_type src_type
= nir_get_nir_type_for_glsl_base_type(types
[0]);
2071 nir_alu_type dst_type
= nir_get_nir_type_for_glsl_base_type(out_type
);
2072 result
= nir_build_alu(&b
, nir_type_conversion_op(src_type
, dst_type
,
2073 nir_rounding_mode_undef
),
2074 srcs
[0], NULL
, NULL
, NULL
);
2075 /* b2i and b2f don't have fixed bit-size versions so the builder will
2076 * just assume 32 and we have to fix it up here.
2078 result
->bit_size
= nir_alu_type_get_type_size(dst_type
);
2082 case ir_unop_f2fmp
: {
2083 result
= nir_build_alu(&b
, nir_op_f2fmp
, srcs
[0], NULL
, NULL
, NULL
);
2087 case ir_unop_bitcast_i2f
:
2088 case ir_unop_bitcast_f2i
:
2089 case ir_unop_bitcast_u2f
:
2090 case ir_unop_bitcast_f2u
:
2091 case ir_unop_bitcast_i642d
:
2092 case ir_unop_bitcast_d2i64
:
2093 case ir_unop_bitcast_u642d
:
2094 case ir_unop_bitcast_d2u64
:
2095 case ir_unop_subroutine_to_int
:
2097 result
= nir_mov(&b
, srcs
[0]);
2099 case ir_unop_trunc
: result
= nir_ftrunc(&b
, srcs
[0]); break;
2100 case ir_unop_ceil
: result
= nir_fceil(&b
, srcs
[0]); break;
2101 case ir_unop_floor
: result
= nir_ffloor(&b
, srcs
[0]); break;
2102 case ir_unop_fract
: result
= nir_ffract(&b
, srcs
[0]); break;
2103 case ir_unop_frexp_exp
: result
= nir_frexp_exp(&b
, srcs
[0]); break;
2104 case ir_unop_frexp_sig
: result
= nir_frexp_sig(&b
, srcs
[0]); break;
2105 case ir_unop_round_even
: result
= nir_fround_even(&b
, srcs
[0]); break;
2106 case ir_unop_sin
: result
= nir_fsin(&b
, srcs
[0]); break;
2107 case ir_unop_cos
: result
= nir_fcos(&b
, srcs
[0]); break;
2108 case ir_unop_dFdx
: result
= nir_fddx(&b
, srcs
[0]); break;
2109 case ir_unop_dFdy
: result
= nir_fddy(&b
, srcs
[0]); break;
2110 case ir_unop_dFdx_fine
: result
= nir_fddx_fine(&b
, srcs
[0]); break;
2111 case ir_unop_dFdy_fine
: result
= nir_fddy_fine(&b
, srcs
[0]); break;
2112 case ir_unop_dFdx_coarse
: result
= nir_fddx_coarse(&b
, srcs
[0]); break;
2113 case ir_unop_dFdy_coarse
: result
= nir_fddy_coarse(&b
, srcs
[0]); break;
2114 case ir_unop_pack_snorm_2x16
:
2115 result
= nir_pack_snorm_2x16(&b
, srcs
[0]);
2117 case ir_unop_pack_snorm_4x8
:
2118 result
= nir_pack_snorm_4x8(&b
, srcs
[0]);
2120 case ir_unop_pack_unorm_2x16
:
2121 result
= nir_pack_unorm_2x16(&b
, srcs
[0]);
2123 case ir_unop_pack_unorm_4x8
:
2124 result
= nir_pack_unorm_4x8(&b
, srcs
[0]);
2126 case ir_unop_pack_half_2x16
:
2127 result
= nir_pack_half_2x16(&b
, srcs
[0]);
2129 case ir_unop_unpack_snorm_2x16
:
2130 result
= nir_unpack_snorm_2x16(&b
, srcs
[0]);
2132 case ir_unop_unpack_snorm_4x8
:
2133 result
= nir_unpack_snorm_4x8(&b
, srcs
[0]);
2135 case ir_unop_unpack_unorm_2x16
:
2136 result
= nir_unpack_unorm_2x16(&b
, srcs
[0]);
2138 case ir_unop_unpack_unorm_4x8
:
2139 result
= nir_unpack_unorm_4x8(&b
, srcs
[0]);
2141 case ir_unop_unpack_half_2x16
:
2142 result
= nir_unpack_half_2x16(&b
, srcs
[0]);
2144 case ir_unop_pack_sampler_2x32
:
2145 case ir_unop_pack_image_2x32
:
2146 case ir_unop_pack_double_2x32
:
2147 case ir_unop_pack_int_2x32
:
2148 case ir_unop_pack_uint_2x32
:
2149 result
= nir_pack_64_2x32(&b
, srcs
[0]);
2151 case ir_unop_unpack_sampler_2x32
:
2152 case ir_unop_unpack_image_2x32
:
2153 case ir_unop_unpack_double_2x32
:
2154 case ir_unop_unpack_int_2x32
:
2155 case ir_unop_unpack_uint_2x32
:
2156 result
= nir_unpack_64_2x32(&b
, srcs
[0]);
2158 case ir_unop_bitfield_reverse
:
2159 result
= nir_bitfield_reverse(&b
, srcs
[0]);
2161 case ir_unop_bit_count
:
2162 result
= nir_bit_count(&b
, srcs
[0]);
2164 case ir_unop_find_msb
:
2166 case GLSL_TYPE_UINT
:
2167 result
= nir_ufind_msb(&b
, srcs
[0]);
2170 result
= nir_ifind_msb(&b
, srcs
[0]);
2173 unreachable("Invalid type for findMSB()");
2176 case ir_unop_find_lsb
:
2177 result
= nir_find_lsb(&b
, srcs
[0]);
2181 switch (ir
->type
->vector_elements
) {
2183 switch (ir
->operands
[0]->type
->vector_elements
) {
2184 case 1: result
= nir_fnoise1_1(&b
, srcs
[0]); break;
2185 case 2: result
= nir_fnoise1_2(&b
, srcs
[0]); break;
2186 case 3: result
= nir_fnoise1_3(&b
, srcs
[0]); break;
2187 case 4: result
= nir_fnoise1_4(&b
, srcs
[0]); break;
2188 default: unreachable("not reached");
2192 switch (ir
->operands
[0]->type
->vector_elements
) {
2193 case 1: result
= nir_fnoise2_1(&b
, srcs
[0]); break;
2194 case 2: result
= nir_fnoise2_2(&b
, srcs
[0]); break;
2195 case 3: result
= nir_fnoise2_3(&b
, srcs
[0]); break;
2196 case 4: result
= nir_fnoise2_4(&b
, srcs
[0]); break;
2197 default: unreachable("not reached");
2201 switch (ir
->operands
[0]->type
->vector_elements
) {
2202 case 1: result
= nir_fnoise3_1(&b
, srcs
[0]); break;
2203 case 2: result
= nir_fnoise3_2(&b
, srcs
[0]); break;
2204 case 3: result
= nir_fnoise3_3(&b
, srcs
[0]); break;
2205 case 4: result
= nir_fnoise3_4(&b
, srcs
[0]); break;
2206 default: unreachable("not reached");
2210 switch (ir
->operands
[0]->type
->vector_elements
) {
2211 case 1: result
= nir_fnoise4_1(&b
, srcs
[0]); break;
2212 case 2: result
= nir_fnoise4_2(&b
, srcs
[0]); break;
2213 case 3: result
= nir_fnoise4_3(&b
, srcs
[0]); break;
2214 case 4: result
= nir_fnoise4_4(&b
, srcs
[0]); break;
2215 default: unreachable("not reached");
2219 unreachable("not reached");
2222 case ir_unop_get_buffer_size
: {
2223 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(
2225 nir_intrinsic_get_buffer_size
);
2226 load
->num_components
= ir
->type
->vector_elements
;
2227 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
2228 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2229 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
2234 result
= nir_atan(&b
, srcs
[0]);
2238 result
= type_is_float(out_type
) ? nir_fadd(&b
, srcs
[0], srcs
[1])
2239 : nir_iadd(&b
, srcs
[0], srcs
[1]);
2241 case ir_binop_add_sat
:
2242 result
= type_is_signed(out_type
) ? nir_iadd_sat(&b
, srcs
[0], srcs
[1])
2243 : nir_uadd_sat(&b
, srcs
[0], srcs
[1]);
2246 result
= type_is_float(out_type
) ? nir_fsub(&b
, srcs
[0], srcs
[1])
2247 : nir_isub(&b
, srcs
[0], srcs
[1]);
2249 case ir_binop_sub_sat
:
2250 result
= type_is_signed(out_type
) ? nir_isub_sat(&b
, srcs
[0], srcs
[1])
2251 : nir_usub_sat(&b
, srcs
[0], srcs
[1]);
2253 case ir_binop_abs_sub
:
2254 /* out_type is always unsigned for ir_binop_abs_sub, so we have to key
2255 * on the type of the sources.
2257 result
= type_is_signed(types
[0]) ? nir_uabs_isub(&b
, srcs
[0], srcs
[1])
2258 : nir_uabs_usub(&b
, srcs
[0], srcs
[1]);
2261 result
= type_is_signed(out_type
) ? nir_ihadd(&b
, srcs
[0], srcs
[1])
2262 : nir_uhadd(&b
, srcs
[0], srcs
[1]);
2264 case ir_binop_avg_round
:
2265 result
= type_is_signed(out_type
) ? nir_irhadd(&b
, srcs
[0], srcs
[1])
2266 : nir_urhadd(&b
, srcs
[0], srcs
[1]);
2268 case ir_binop_mul_32x16
:
2269 result
= type_is_signed(out_type
) ? nir_imul_32x16(&b
, srcs
[0], srcs
[1])
2270 : nir_umul_32x16(&b
, srcs
[0], srcs
[1]);
2273 if (type_is_float(out_type
))
2274 result
= nir_fmul(&b
, srcs
[0], srcs
[1]);
2275 else if (out_type
== GLSL_TYPE_INT64
&&
2276 (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
||
2277 ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
))
2278 result
= nir_imul_2x32_64(&b
, srcs
[0], srcs
[1]);
2279 else if (out_type
== GLSL_TYPE_UINT64
&&
2280 (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
||
2281 ir
->operands
[1]->type
->base_type
== GLSL_TYPE_UINT
))
2282 result
= nir_umul_2x32_64(&b
, srcs
[0], srcs
[1]);
2284 result
= nir_imul(&b
, srcs
[0], srcs
[1]);
2287 if (type_is_float(out_type
))
2288 result
= nir_fdiv(&b
, srcs
[0], srcs
[1]);
2289 else if (type_is_signed(out_type
))
2290 result
= nir_idiv(&b
, srcs
[0], srcs
[1]);
2292 result
= nir_udiv(&b
, srcs
[0], srcs
[1]);
2295 result
= type_is_float(out_type
) ? nir_fmod(&b
, srcs
[0], srcs
[1])
2296 : nir_umod(&b
, srcs
[0], srcs
[1]);
2299 if (type_is_float(out_type
))
2300 result
= nir_fmin(&b
, srcs
[0], srcs
[1]);
2301 else if (type_is_signed(out_type
))
2302 result
= nir_imin(&b
, srcs
[0], srcs
[1]);
2304 result
= nir_umin(&b
, srcs
[0], srcs
[1]);
2307 if (type_is_float(out_type
))
2308 result
= nir_fmax(&b
, srcs
[0], srcs
[1]);
2309 else if (type_is_signed(out_type
))
2310 result
= nir_imax(&b
, srcs
[0], srcs
[1]);
2312 result
= nir_umax(&b
, srcs
[0], srcs
[1]);
2314 case ir_binop_pow
: result
= nir_fpow(&b
, srcs
[0], srcs
[1]); break;
2315 case ir_binop_bit_and
: result
= nir_iand(&b
, srcs
[0], srcs
[1]); break;
2316 case ir_binop_bit_or
: result
= nir_ior(&b
, srcs
[0], srcs
[1]); break;
2317 case ir_binop_bit_xor
: result
= nir_ixor(&b
, srcs
[0], srcs
[1]); break;
2318 case ir_binop_logic_and
:
2319 result
= nir_iand(&b
, srcs
[0], srcs
[1]);
2321 case ir_binop_logic_or
:
2322 result
= nir_ior(&b
, srcs
[0], srcs
[1]);
2324 case ir_binop_logic_xor
:
2325 result
= nir_ixor(&b
, srcs
[0], srcs
[1]);
2327 case ir_binop_lshift
: result
= nir_ishl(&b
, srcs
[0], srcs
[1]); break;
2328 case ir_binop_rshift
:
2329 result
= (type_is_signed(out_type
)) ? nir_ishr(&b
, srcs
[0], srcs
[1])
2330 : nir_ushr(&b
, srcs
[0], srcs
[1]);
2332 case ir_binop_imul_high
:
2333 result
= (out_type
== GLSL_TYPE_INT
) ? nir_imul_high(&b
, srcs
[0], srcs
[1])
2334 : nir_umul_high(&b
, srcs
[0], srcs
[1]);
2336 case ir_binop_carry
: result
= nir_uadd_carry(&b
, srcs
[0], srcs
[1]); break;
2337 case ir_binop_borrow
: result
= nir_usub_borrow(&b
, srcs
[0], srcs
[1]); break;
2339 if (type_is_float(types
[0]))
2340 result
= nir_flt(&b
, srcs
[0], srcs
[1]);
2341 else if (type_is_signed(types
[0]))
2342 result
= nir_ilt(&b
, srcs
[0], srcs
[1]);
2344 result
= nir_ult(&b
, srcs
[0], srcs
[1]);
2346 case ir_binop_gequal
:
2347 if (type_is_float(types
[0]))
2348 result
= nir_fge(&b
, srcs
[0], srcs
[1]);
2349 else if (type_is_signed(types
[0]))
2350 result
= nir_ige(&b
, srcs
[0], srcs
[1]);
2352 result
= nir_uge(&b
, srcs
[0], srcs
[1]);
2354 case ir_binop_equal
:
2355 if (type_is_float(types
[0]))
2356 result
= nir_feq(&b
, srcs
[0], srcs
[1]);
2358 result
= nir_ieq(&b
, srcs
[0], srcs
[1]);
2360 case ir_binop_nequal
:
2361 if (type_is_float(types
[0]))
2362 result
= nir_fne(&b
, srcs
[0], srcs
[1]);
2364 result
= nir_ine(&b
, srcs
[0], srcs
[1]);
2366 case ir_binop_all_equal
:
2367 if (type_is_float(types
[0])) {
2368 switch (ir
->operands
[0]->type
->vector_elements
) {
2369 case 1: result
= nir_feq(&b
, srcs
[0], srcs
[1]); break;
2370 case 2: result
= nir_ball_fequal2(&b
, srcs
[0], srcs
[1]); break;
2371 case 3: result
= nir_ball_fequal3(&b
, srcs
[0], srcs
[1]); break;
2372 case 4: result
= nir_ball_fequal4(&b
, srcs
[0], srcs
[1]); break;
2374 unreachable("not reached");
2377 switch (ir
->operands
[0]->type
->vector_elements
) {
2378 case 1: result
= nir_ieq(&b
, srcs
[0], srcs
[1]); break;
2379 case 2: result
= nir_ball_iequal2(&b
, srcs
[0], srcs
[1]); break;
2380 case 3: result
= nir_ball_iequal3(&b
, srcs
[0], srcs
[1]); break;
2381 case 4: result
= nir_ball_iequal4(&b
, srcs
[0], srcs
[1]); break;
2383 unreachable("not reached");
2387 case ir_binop_any_nequal
:
2388 if (type_is_float(types
[0])) {
2389 switch (ir
->operands
[0]->type
->vector_elements
) {
2390 case 1: result
= nir_fne(&b
, srcs
[0], srcs
[1]); break;
2391 case 2: result
= nir_bany_fnequal2(&b
, srcs
[0], srcs
[1]); break;
2392 case 3: result
= nir_bany_fnequal3(&b
, srcs
[0], srcs
[1]); break;
2393 case 4: result
= nir_bany_fnequal4(&b
, srcs
[0], srcs
[1]); break;
2395 unreachable("not reached");
2398 switch (ir
->operands
[0]->type
->vector_elements
) {
2399 case 1: result
= nir_ine(&b
, srcs
[0], srcs
[1]); break;
2400 case 2: result
= nir_bany_inequal2(&b
, srcs
[0], srcs
[1]); break;
2401 case 3: result
= nir_bany_inequal3(&b
, srcs
[0], srcs
[1]); break;
2402 case 4: result
= nir_bany_inequal4(&b
, srcs
[0], srcs
[1]); break;
2404 unreachable("not reached");
2409 switch (ir
->operands
[0]->type
->vector_elements
) {
2410 case 2: result
= nir_fdot2(&b
, srcs
[0], srcs
[1]); break;
2411 case 3: result
= nir_fdot3(&b
, srcs
[0], srcs
[1]); break;
2412 case 4: result
= nir_fdot4(&b
, srcs
[0], srcs
[1]); break;
2414 unreachable("not reached");
2417 case ir_binop_vector_extract
: {
2418 result
= nir_channel(&b
, srcs
[0], 0);
2419 for (unsigned i
= 1; i
< ir
->operands
[0]->type
->vector_elements
; i
++) {
2420 nir_ssa_def
*swizzled
= nir_channel(&b
, srcs
[0], i
);
2421 result
= nir_bcsel(&b
, nir_ieq(&b
, srcs
[1], nir_imm_int(&b
, i
)),
2427 case ir_binop_atan2
:
2428 result
= nir_atan2(&b
, srcs
[0], srcs
[1]);
2431 case ir_binop_ldexp
: result
= nir_ldexp(&b
, srcs
[0], srcs
[1]); break;
2433 result
= nir_ffma(&b
, srcs
[0], srcs
[1], srcs
[2]);
2436 result
= nir_flrp(&b
, srcs
[0], srcs
[1], srcs
[2]);
2439 result
= nir_bcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
2441 case ir_triop_bitfield_extract
:
2442 result
= (out_type
== GLSL_TYPE_INT
) ?
2443 nir_ibitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]) :
2444 nir_ubitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]);
2446 case ir_quadop_bitfield_insert
:
2447 result
= nir_bitfield_insert(&b
, srcs
[0], srcs
[1], srcs
[2], srcs
[3]);
2449 case ir_quadop_vector
:
2450 result
= nir_vec(&b
, srcs
, ir
->type
->vector_elements
);
2454 unreachable("not reached");
2459 nir_visitor::visit(ir_swizzle
*ir
)
2461 unsigned swizzle
[4] = { ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
};
2462 result
= nir_swizzle(&b
, evaluate_rvalue(ir
->val
), swizzle
,
2463 ir
->type
->vector_elements
);
2467 nir_visitor::visit(ir_texture
*ir
)
2474 num_srcs
= 1; /* coordinate */
2479 op
= (ir
->op
== ir_txb
) ? nir_texop_txb
: nir_texop_txl
;
2480 num_srcs
= 2; /* coordinate, bias/lod */
2484 op
= nir_texop_txd
; /* coordinate, dPdx, dPdy */
2490 if (ir
->lod_info
.lod
!= NULL
)
2491 num_srcs
= 2; /* coordinate, lod */
2493 num_srcs
= 1; /* coordinate */
2497 op
= nir_texop_txf_ms
;
2498 num_srcs
= 2; /* coordinate, sample_index */
2503 if (ir
->lod_info
.lod
!= NULL
)
2504 num_srcs
= 1; /* lod */
2511 num_srcs
= 1; /* coordinate */
2516 num_srcs
= 1; /* coordinate */
2519 case ir_query_levels
:
2520 op
= nir_texop_query_levels
;
2524 case ir_texture_samples
:
2525 op
= nir_texop_texture_samples
;
2529 case ir_samples_identical
:
2530 op
= nir_texop_samples_identical
;
2531 num_srcs
= 1; /* coordinate */
2535 unreachable("not reached");
2538 if (ir
->projector
!= NULL
)
2540 if (ir
->shadow_comparator
!= NULL
)
2542 /* offsets are constants we store inside nir_tex_intrs.offsets */
2543 if (ir
->offset
!= NULL
&& !ir
->offset
->type
->is_array())
2546 /* Add one for the texture deref */
2549 nir_tex_instr
*instr
= nir_tex_instr_create(this->shader
, num_srcs
);
2552 instr
->sampler_dim
=
2553 (glsl_sampler_dim
) ir
->sampler
->type
->sampler_dimensionality
;
2554 instr
->is_array
= ir
->sampler
->type
->sampler_array
;
2555 instr
->is_shadow
= ir
->sampler
->type
->sampler_shadow
;
2556 if (instr
->is_shadow
)
2557 instr
->is_new_style_shadow
= (ir
->type
->vector_elements
== 1);
2558 switch (ir
->type
->base_type
) {
2559 case GLSL_TYPE_FLOAT
:
2560 instr
->dest_type
= nir_type_float
;
2562 case GLSL_TYPE_FLOAT16
:
2563 instr
->dest_type
= nir_type_float16
;
2566 instr
->dest_type
= nir_type_int
;
2568 case GLSL_TYPE_BOOL
:
2569 case GLSL_TYPE_UINT
:
2570 instr
->dest_type
= nir_type_uint
;
2573 unreachable("not reached");
2576 nir_deref_instr
*sampler_deref
= evaluate_deref(ir
->sampler
);
2578 /* check for bindless handles */
2579 if (sampler_deref
->mode
!= nir_var_uniform
||
2580 nir_deref_instr_get_variable(sampler_deref
)->data
.bindless
) {
2581 nir_ssa_def
*load
= nir_load_deref(&b
, sampler_deref
);
2582 instr
->src
[0].src
= nir_src_for_ssa(load
);
2583 instr
->src
[0].src_type
= nir_tex_src_texture_handle
;
2584 instr
->src
[1].src
= nir_src_for_ssa(load
);
2585 instr
->src
[1].src_type
= nir_tex_src_sampler_handle
;
2587 instr
->src
[0].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2588 instr
->src
[0].src_type
= nir_tex_src_texture_deref
;
2589 instr
->src
[1].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2590 instr
->src
[1].src_type
= nir_tex_src_sampler_deref
;
2593 unsigned src_number
= 2;
2595 if (ir
->coordinate
!= NULL
) {
2596 instr
->coord_components
= ir
->coordinate
->type
->vector_elements
;
2597 instr
->src
[src_number
].src
=
2598 nir_src_for_ssa(evaluate_rvalue(ir
->coordinate
));
2599 instr
->src
[src_number
].src_type
= nir_tex_src_coord
;
2603 if (ir
->projector
!= NULL
) {
2604 instr
->src
[src_number
].src
=
2605 nir_src_for_ssa(evaluate_rvalue(ir
->projector
));
2606 instr
->src
[src_number
].src_type
= nir_tex_src_projector
;
2610 if (ir
->shadow_comparator
!= NULL
) {
2611 instr
->src
[src_number
].src
=
2612 nir_src_for_ssa(evaluate_rvalue(ir
->shadow_comparator
));
2613 instr
->src
[src_number
].src_type
= nir_tex_src_comparator
;
2617 if (ir
->offset
!= NULL
) {
2618 if (ir
->offset
->type
->is_array()) {
2619 for (int i
= 0; i
< ir
->offset
->type
->array_size(); i
++) {
2620 const ir_constant
*c
=
2621 ir
->offset
->as_constant()->get_array_element(i
);
2623 for (unsigned j
= 0; j
< 2; ++j
) {
2624 int val
= c
->get_int_component(j
);
2625 assert(val
<= 31 && val
>= -32);
2626 instr
->tg4_offsets
[i
][j
] = val
;
2630 assert(ir
->offset
->type
->is_vector() || ir
->offset
->type
->is_scalar());
2632 instr
->src
[src_number
].src
=
2633 nir_src_for_ssa(evaluate_rvalue(ir
->offset
));
2634 instr
->src
[src_number
].src_type
= nir_tex_src_offset
;
2641 instr
->src
[src_number
].src
=
2642 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.bias
));
2643 instr
->src
[src_number
].src_type
= nir_tex_src_bias
;
2650 if (ir
->lod_info
.lod
!= NULL
) {
2651 instr
->src
[src_number
].src
=
2652 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.lod
));
2653 instr
->src
[src_number
].src_type
= nir_tex_src_lod
;
2659 instr
->src
[src_number
].src
=
2660 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdx
));
2661 instr
->src
[src_number
].src_type
= nir_tex_src_ddx
;
2663 instr
->src
[src_number
].src
=
2664 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdy
));
2665 instr
->src
[src_number
].src_type
= nir_tex_src_ddy
;
2670 instr
->src
[src_number
].src
=
2671 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.sample_index
));
2672 instr
->src
[src_number
].src_type
= nir_tex_src_ms_index
;
2677 instr
->component
= ir
->lod_info
.component
->as_constant()->value
.u
[0];
2684 assert(src_number
== num_srcs
);
2686 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2687 add_instr(&instr
->instr
, nir_tex_instr_dest_size(instr
), bit_size
);
2691 nir_visitor::visit(ir_constant
*ir
)
2694 * We don't know if this variable is an array or struct that gets
2695 * dereferenced, so do the safe thing an make it a variable with a
2696 * constant initializer and return a dereference.
2700 nir_local_variable_create(this->impl
, ir
->type
, "const_temp");
2701 var
->data
.read_only
= true;
2702 var
->constant_initializer
= constant_copy(ir
, var
);
2704 this->deref
= nir_build_deref_var(&b
, var
);
2708 nir_visitor::visit(ir_dereference_variable
*ir
)
2710 if (ir
->variable_referenced()->data
.mode
== ir_var_function_out
) {
2711 unsigned i
= (sig
->return_type
!= glsl_type::void_type
) ? 1 : 0;
2713 foreach_in_list(ir_variable
, param
, &sig
->parameters
) {
2714 if (param
== ir
->variable_referenced()) {
2720 this->deref
= nir_build_deref_cast(&b
, nir_load_param(&b
, i
),
2721 nir_var_function_temp
, ir
->type
, 0);
2725 assert(ir
->variable_referenced()->data
.mode
!= ir_var_function_inout
);
2727 struct hash_entry
*entry
=
2728 _mesa_hash_table_search(this->var_table
, ir
->var
);
2730 nir_variable
*var
= (nir_variable
*) entry
->data
;
2732 this->deref
= nir_build_deref_var(&b
, var
);
2736 nir_visitor::visit(ir_dereference_record
*ir
)
2738 ir
->record
->accept(this);
2740 int field_index
= ir
->field_idx
;
2741 assert(field_index
>= 0);
2743 this->deref
= nir_build_deref_struct(&b
, this->deref
, field_index
);
2747 nir_visitor::visit(ir_dereference_array
*ir
)
2749 nir_ssa_def
*index
= evaluate_rvalue(ir
->array_index
);
2751 ir
->array
->accept(this);
2753 this->deref
= nir_build_deref_array(&b
, this->deref
, index
);
2757 nir_visitor::visit(ir_barrier
*)
2759 if (shader
->info
.stage
== MESA_SHADER_COMPUTE
) {
2760 nir_intrinsic_instr
*shared_barrier
=
2761 nir_intrinsic_instr_create(this->shader
,
2762 nir_intrinsic_memory_barrier_shared
);
2763 nir_builder_instr_insert(&b
, &shared_barrier
->instr
);
2764 } else if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
) {
2765 nir_intrinsic_instr
*patch_barrier
=
2766 nir_intrinsic_instr_create(this->shader
,
2767 nir_intrinsic_memory_barrier_tcs_patch
);
2768 nir_builder_instr_insert(&b
, &patch_barrier
->instr
);
2771 nir_intrinsic_instr
*instr
=
2772 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_control_barrier
);
2773 nir_builder_instr_insert(&b
, &instr
->instr
);
2777 glsl_float64_funcs_to_nir(struct gl_context
*ctx
,
2778 const nir_shader_compiler_options
*options
)
2780 /* We pretend it's a vertex shader. Ultimately, the stage shouldn't
2781 * matter because we're not optimizing anything here.
2783 struct gl_shader
*sh
= _mesa_new_shader(-1, MESA_SHADER_VERTEX
);
2784 sh
->Source
= float64_source
;
2785 sh
->CompileStatus
= COMPILE_FAILURE
;
2786 _mesa_glsl_compile_shader(ctx
, sh
, false, false, true);
2788 if (!sh
->CompileStatus
) {
2791 "fp64 software impl compile failed:\n%s\nsource:\n%s\n",
2792 sh
->InfoLog
, float64_source
);
2797 nir_shader
*nir
= nir_shader_create(NULL
, MESA_SHADER_VERTEX
, options
, NULL
);
2799 nir_visitor
v1(ctx
, nir
);
2800 nir_function_visitor
v2(&v1
);
2802 visit_exec_list(sh
->ir
, &v1
);
2804 /* _mesa_delete_shader will try to free sh->Source but it's static const */
2806 _mesa_delete_shader(ctx
, sh
);
2808 nir_validate_shader(nir
, "float64_funcs_to_nir");
2810 NIR_PASS_V(nir
, nir_lower_variable_initializers
, nir_var_function_temp
);
2811 NIR_PASS_V(nir
, nir_lower_returns
);
2812 NIR_PASS_V(nir
, nir_inline_functions
);
2813 NIR_PASS_V(nir
, nir_opt_deref
);
2815 /* Do some optimizations to clean up the shader now. By optimizing the
2816 * functions in the library, we avoid having to re-do that work every
2817 * time we inline a copy of a function. Reducing basic blocks also helps
2818 * with compile times.
2820 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2821 NIR_PASS_V(nir
, nir_copy_prop
);
2822 NIR_PASS_V(nir
, nir_opt_dce
);
2823 NIR_PASS_V(nir
, nir_opt_cse
);
2824 NIR_PASS_V(nir
, nir_opt_gcm
, true);
2825 NIR_PASS_V(nir
, nir_opt_peephole_select
, 1, false, false);
2826 NIR_PASS_V(nir
, nir_opt_dce
);