2 * Copyright © 2014 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Connor Abbott (cwabbott0@gmail.com)
28 #include "float64_glsl.h"
29 #include "glsl_to_nir.h"
30 #include "ir_visitor.h"
31 #include "ir_hierarchical_visitor.h"
33 #include "ir_optimization.h"
35 #include "compiler/nir/nir_control_flow.h"
36 #include "compiler/nir/nir_builder.h"
37 #include "compiler/nir/nir_builtin_builder.h"
38 #include "compiler/nir/nir_deref.h"
39 #include "main/errors.h"
40 #include "main/imports.h"
41 #include "main/mtypes.h"
42 #include "main/shaderobj.h"
43 #include "util/u_math.h"
46 * pass to lower GLSL IR to NIR
48 * This will lower variable dereferences to loads/stores of corresponding
49 * variables in NIR - the variables will be converted to registers in a later
55 class nir_visitor
: public ir_visitor
58 nir_visitor(gl_context
*ctx
, nir_shader
*shader
);
61 virtual void visit(ir_variable
*);
62 virtual void visit(ir_function
*);
63 virtual void visit(ir_function_signature
*);
64 virtual void visit(ir_loop
*);
65 virtual void visit(ir_if
*);
66 virtual void visit(ir_discard
*);
67 virtual void visit(ir_demote
*);
68 virtual void visit(ir_loop_jump
*);
69 virtual void visit(ir_return
*);
70 virtual void visit(ir_call
*);
71 virtual void visit(ir_assignment
*);
72 virtual void visit(ir_emit_vertex
*);
73 virtual void visit(ir_end_primitive
*);
74 virtual void visit(ir_expression
*);
75 virtual void visit(ir_swizzle
*);
76 virtual void visit(ir_texture
*);
77 virtual void visit(ir_constant
*);
78 virtual void visit(ir_dereference_variable
*);
79 virtual void visit(ir_dereference_record
*);
80 virtual void visit(ir_dereference_array
*);
81 virtual void visit(ir_barrier
*);
83 void create_function(ir_function_signature
*ir
);
86 void add_instr(nir_instr
*instr
, unsigned num_components
, unsigned bit_size
);
87 nir_ssa_def
*evaluate_rvalue(ir_rvalue
*ir
);
89 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
**srcs
);
90 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
);
91 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
,
93 nir_alu_instr
*emit(nir_op op
, unsigned dest_size
, nir_ssa_def
*src1
,
94 nir_ssa_def
*src2
, nir_ssa_def
*src3
);
99 nir_function_impl
*impl
;
101 nir_ssa_def
*result
; /* result of the expression tree last visited */
103 nir_deref_instr
*evaluate_deref(ir_instruction
*ir
);
105 nir_constant
*constant_copy(ir_constant
*ir
, void *mem_ctx
);
107 /* most recent deref instruction created */
108 nir_deref_instr
*deref
;
110 /* whether the IR we're operating on is per-function or global */
113 ir_function_signature
*sig
;
115 /* map of ir_variable -> nir_variable */
116 struct hash_table
*var_table
;
118 /* map of ir_function_signature -> nir_function_overload */
119 struct hash_table
*overload_table
;
123 * This visitor runs before the main visitor, calling create_function() for
124 * each function so that the main visitor can resolve forward references in
128 class nir_function_visitor
: public ir_hierarchical_visitor
131 nir_function_visitor(nir_visitor
*v
) : visitor(v
)
134 virtual ir_visitor_status
visit_enter(ir_function
*);
137 nir_visitor
*visitor
;
140 /* glsl_to_nir can only handle converting certain function paramaters
141 * to NIR. This visitor checks for parameters it can't currently handle.
143 class ir_function_param_visitor
: public ir_hierarchical_visitor
146 ir_function_param_visitor()
151 virtual ir_visitor_status
visit_enter(ir_function_signature
*ir
)
154 if (ir
->is_intrinsic())
155 return visit_continue
;
157 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
158 if (!param
->type
->is_vector() || !param
->type
->is_scalar()) {
163 if (param
->data
.mode
== ir_var_function_inout
) {
169 return visit_continue
;
175 } /* end of anonymous namespace */
179 has_unsupported_function_param(exec_list
*ir
)
181 ir_function_param_visitor visitor
;
182 visit_list_elements(&visitor
, ir
);
183 return visitor
.unsupported
;
187 glsl_to_nir(struct gl_context
*ctx
,
188 const struct gl_shader_program
*shader_prog
,
189 gl_shader_stage stage
,
190 const nir_shader_compiler_options
*options
)
192 struct gl_linked_shader
*sh
= shader_prog
->_LinkedShaders
[stage
];
194 const struct gl_shader_compiler_options
*gl_options
=
195 &ctx
->Const
.ShaderCompilerOptions
[stage
];
197 /* glsl_to_nir can only handle converting certain function paramaters
198 * to NIR. If we find something we can't handle then we get the GLSL IR
199 * opts to remove it before we continue on.
201 * TODO: add missing glsl ir to nir support and remove this loop.
203 while (has_unsupported_function_param(sh
->ir
)) {
204 do_common_optimization(sh
->ir
, true, true, gl_options
,
205 ctx
->Const
.NativeIntegers
);
208 nir_shader
*shader
= nir_shader_create(NULL
, stage
, options
,
211 nir_visitor
v1(ctx
, shader
);
212 nir_function_visitor
v2(&v1
);
214 visit_exec_list(sh
->ir
, &v1
);
216 nir_validate_shader(shader
, "after glsl to nir, before function inline");
218 /* We have to lower away local constant initializers right before we
219 * inline functions. That way they get properly initialized at the top
220 * of the function and not at the top of its caller.
222 nir_lower_constant_initializers(shader
, (nir_variable_mode
)~0);
223 nir_lower_returns(shader
);
224 nir_inline_functions(shader
);
225 nir_opt_deref(shader
);
227 nir_validate_shader(shader
, "after function inlining and return lowering");
229 /* Now that we have inlined everything remove all of the functions except
232 foreach_list_typed_safe(nir_function
, function
, node
, &(shader
)->functions
){
233 if (strcmp("main", function
->name
) != 0) {
234 exec_node_remove(&function
->node
);
238 shader
->info
.name
= ralloc_asprintf(shader
, "GLSL%d", shader_prog
->Name
);
239 if (shader_prog
->Label
)
240 shader
->info
.label
= ralloc_strdup(shader
, shader_prog
->Label
);
242 /* Check for transform feedback varyings specified via the API */
243 shader
->info
.has_transform_feedback_varyings
=
244 shader_prog
->TransformFeedback
.NumVarying
> 0;
246 /* Check for transform feedback varyings specified in the Shader */
247 if (shader_prog
->last_vert_prog
)
248 shader
->info
.has_transform_feedback_varyings
|=
249 shader_prog
->last_vert_prog
->sh
.LinkedTransformFeedback
->NumVarying
> 0;
251 if (shader
->info
.stage
== MESA_SHADER_FRAGMENT
) {
252 shader
->info
.fs
.pixel_center_integer
= sh
->Program
->info
.fs
.pixel_center_integer
;
253 shader
->info
.fs
.origin_upper_left
= sh
->Program
->info
.fs
.origin_upper_left
;
259 nir_visitor::nir_visitor(gl_context
*ctx
, nir_shader
*shader
)
261 this->supports_std430
= ctx
->Const
.UseSTD430AsDefaultPacking
;
262 this->shader
= shader
;
263 this->is_global
= true;
264 this->var_table
= _mesa_pointer_hash_table_create(NULL
);
265 this->overload_table
= _mesa_pointer_hash_table_create(NULL
);
270 memset(&this->b
, 0, sizeof(this->b
));
273 nir_visitor::~nir_visitor()
275 _mesa_hash_table_destroy(this->var_table
, NULL
);
276 _mesa_hash_table_destroy(this->overload_table
, NULL
);
280 nir_visitor::evaluate_deref(ir_instruction
*ir
)
287 nir_visitor::constant_copy(ir_constant
*ir
, void *mem_ctx
)
292 nir_constant
*ret
= rzalloc(mem_ctx
, nir_constant
);
294 const unsigned rows
= ir
->type
->vector_elements
;
295 const unsigned cols
= ir
->type
->matrix_columns
;
298 ret
->num_elements
= 0;
299 switch (ir
->type
->base_type
) {
301 /* Only float base types can be matrices. */
304 for (unsigned r
= 0; r
< rows
; r
++)
305 ret
->values
[r
].u32
= ir
->value
.u
[r
];
310 /* Only float base types can be matrices. */
313 for (unsigned r
= 0; r
< rows
; r
++)
314 ret
->values
[r
].i32
= ir
->value
.i
[r
];
318 case GLSL_TYPE_FLOAT
:
319 case GLSL_TYPE_DOUBLE
:
321 ret
->elements
= ralloc_array(mem_ctx
, nir_constant
*, cols
);
322 ret
->num_elements
= cols
;
323 for (unsigned c
= 0; c
< cols
; c
++) {
324 nir_constant
*col_const
= rzalloc(mem_ctx
, nir_constant
);
325 col_const
->num_elements
= 0;
326 switch (ir
->type
->base_type
) {
327 case GLSL_TYPE_FLOAT
:
328 for (unsigned r
= 0; r
< rows
; r
++)
329 col_const
->values
[r
].f32
= ir
->value
.f
[c
* rows
+ r
];
332 case GLSL_TYPE_DOUBLE
:
333 for (unsigned r
= 0; r
< rows
; r
++)
334 col_const
->values
[r
].f64
= ir
->value
.d
[c
* rows
+ r
];
338 unreachable("Cannot get here from the first level switch");
340 ret
->elements
[c
] = col_const
;
343 switch (ir
->type
->base_type
) {
344 case GLSL_TYPE_FLOAT
:
345 for (unsigned r
= 0; r
< rows
; r
++)
346 ret
->values
[r
].f32
= ir
->value
.f
[r
];
349 case GLSL_TYPE_DOUBLE
:
350 for (unsigned r
= 0; r
< rows
; r
++)
351 ret
->values
[r
].f64
= ir
->value
.d
[r
];
355 unreachable("Cannot get here from the first level switch");
360 case GLSL_TYPE_UINT64
:
361 /* Only float base types can be matrices. */
364 for (unsigned r
= 0; r
< rows
; r
++)
365 ret
->values
[r
].u64
= ir
->value
.u64
[r
];
368 case GLSL_TYPE_INT64
:
369 /* Only float base types can be matrices. */
372 for (unsigned r
= 0; r
< rows
; r
++)
373 ret
->values
[r
].i64
= ir
->value
.i64
[r
];
377 /* Only float base types can be matrices. */
380 for (unsigned r
= 0; r
< rows
; r
++)
381 ret
->values
[r
].b
= ir
->value
.b
[r
];
385 case GLSL_TYPE_STRUCT
:
386 case GLSL_TYPE_ARRAY
:
387 ret
->elements
= ralloc_array(mem_ctx
, nir_constant
*,
389 ret
->num_elements
= ir
->type
->length
;
391 for (i
= 0; i
< ir
->type
->length
; i
++)
392 ret
->elements
[i
] = constant_copy(ir
->const_elements
[i
], mem_ctx
);
396 unreachable("not reached");
402 static const glsl_type
*
403 wrap_type_in_array(const glsl_type
*elem_type
, const glsl_type
*array_type
)
405 if (!array_type
->is_array())
408 elem_type
= wrap_type_in_array(elem_type
, array_type
->fields
.array
);
410 return glsl_type::get_array_instance(elem_type
, array_type
->length
);
414 get_nir_how_declared(unsigned how_declared
)
416 if (how_declared
== ir_var_hidden
)
417 return nir_var_hidden
;
419 return nir_var_declared_normally
;
423 nir_visitor::visit(ir_variable
*ir
)
425 /* TODO: In future we should switch to using the NIR lowering pass but for
426 * now just ignore these variables as GLSL IR should have lowered them.
427 * Anything remaining are just dead vars that weren't cleaned up.
429 if (ir
->data
.mode
== ir_var_shader_shared
)
432 /* FINISHME: inout parameters */
433 assert(ir
->data
.mode
!= ir_var_function_inout
);
435 if (ir
->data
.mode
== ir_var_function_out
)
438 nir_variable
*var
= rzalloc(shader
, nir_variable
);
439 var
->type
= ir
->type
;
440 var
->name
= ralloc_strdup(var
, ir
->name
);
442 var
->data
.always_active_io
= ir
->data
.always_active_io
;
443 var
->data
.read_only
= ir
->data
.read_only
;
444 var
->data
.centroid
= ir
->data
.centroid
;
445 var
->data
.sample
= ir
->data
.sample
;
446 var
->data
.patch
= ir
->data
.patch
;
447 var
->data
.how_declared
= get_nir_how_declared(ir
->data
.how_declared
);
448 var
->data
.invariant
= ir
->data
.invariant
;
449 var
->data
.location
= ir
->data
.location
;
450 var
->data
.stream
= ir
->data
.stream
;
451 if (ir
->data
.stream
& (1u << 31))
452 var
->data
.stream
|= NIR_STREAM_PACKED
;
454 var
->data
.precision
= ir
->data
.precision
;
455 var
->data
.explicit_location
= ir
->data
.explicit_location
;
456 var
->data
.from_named_ifc_block
= ir
->data
.from_named_ifc_block
;
457 var
->data
.compact
= false;
459 switch(ir
->data
.mode
) {
461 case ir_var_temporary
:
463 var
->data
.mode
= nir_var_shader_temp
;
465 var
->data
.mode
= nir_var_function_temp
;
468 case ir_var_function_in
:
469 case ir_var_const_in
:
470 var
->data
.mode
= nir_var_function_temp
;
473 case ir_var_shader_in
:
474 if (shader
->info
.stage
== MESA_SHADER_GEOMETRY
&&
475 ir
->data
.location
== VARYING_SLOT_PRIMITIVE_ID
) {
476 /* For whatever reason, GLSL IR makes gl_PrimitiveIDIn an input */
477 var
->data
.location
= SYSTEM_VALUE_PRIMITIVE_ID
;
478 var
->data
.mode
= nir_var_system_value
;
480 var
->data
.mode
= nir_var_shader_in
;
482 if (shader
->info
.stage
== MESA_SHADER_TESS_EVAL
&&
483 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
484 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
485 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
488 if (shader
->info
.stage
> MESA_SHADER_VERTEX
&&
489 ir
->data
.location
>= VARYING_SLOT_CLIP_DIST0
&&
490 ir
->data
.location
<= VARYING_SLOT_CULL_DIST1
) {
491 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
496 case ir_var_shader_out
:
497 var
->data
.mode
= nir_var_shader_out
;
498 if (shader
->info
.stage
== MESA_SHADER_TESS_CTRL
&&
499 (ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_INNER
||
500 ir
->data
.location
== VARYING_SLOT_TESS_LEVEL_OUTER
)) {
501 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
504 if (shader
->info
.stage
<= MESA_SHADER_GEOMETRY
&&
505 ir
->data
.location
>= VARYING_SLOT_CLIP_DIST0
&&
506 ir
->data
.location
<= VARYING_SLOT_CULL_DIST1
) {
507 var
->data
.compact
= ir
->type
->without_array()->is_scalar();
512 if (ir
->get_interface_type())
513 var
->data
.mode
= nir_var_mem_ubo
;
515 var
->data
.mode
= nir_var_uniform
;
518 case ir_var_shader_storage
:
519 var
->data
.mode
= nir_var_mem_ssbo
;
522 case ir_var_system_value
:
523 var
->data
.mode
= nir_var_system_value
;
527 unreachable("not reached");
530 unsigned mem_access
= 0;
531 if (ir
->data
.memory_read_only
)
532 mem_access
|= ACCESS_NON_WRITEABLE
;
533 if (ir
->data
.memory_write_only
)
534 mem_access
|= ACCESS_NON_READABLE
;
535 if (ir
->data
.memory_coherent
)
536 mem_access
|= ACCESS_COHERENT
;
537 if (ir
->data
.memory_volatile
)
538 mem_access
|= ACCESS_VOLATILE
;
539 if (ir
->data
.memory_restrict
)
540 mem_access
|= ACCESS_RESTRICT
;
542 /* For UBO and SSBO variables, we need explicit types */
543 if (var
->data
.mode
& (nir_var_mem_ubo
| nir_var_mem_ssbo
)) {
544 const glsl_type
*explicit_ifc_type
=
545 ir
->get_interface_type()->get_explicit_interface_type(supports_std430
);
547 if (ir
->type
->without_array()->is_interface()) {
548 /* If the type contains the interface, wrap the explicit type in the
549 * right number of arrays.
551 var
->type
= wrap_type_in_array(explicit_ifc_type
, ir
->type
);
553 /* Otherwise, this variable is one entry in the interface */
554 UNUSED
bool found
= false;
555 for (unsigned i
= 0; i
< explicit_ifc_type
->length
; i
++) {
556 const glsl_struct_field
*field
=
557 &explicit_ifc_type
->fields
.structure
[i
];
558 if (strcmp(ir
->name
, field
->name
) != 0)
561 var
->type
= field
->type
;
562 if (field
->memory_read_only
)
563 mem_access
|= ACCESS_NON_WRITEABLE
;
564 if (field
->memory_write_only
)
565 mem_access
|= ACCESS_NON_READABLE
;
566 if (field
->memory_coherent
)
567 mem_access
|= ACCESS_COHERENT
;
568 if (field
->memory_volatile
)
569 mem_access
|= ACCESS_VOLATILE
;
570 if (field
->memory_restrict
)
571 mem_access
|= ACCESS_RESTRICT
;
580 var
->data
.interpolation
= ir
->data
.interpolation
;
581 var
->data
.location_frac
= ir
->data
.location_frac
;
583 switch (ir
->data
.depth_layout
) {
584 case ir_depth_layout_none
:
585 var
->data
.depth_layout
= nir_depth_layout_none
;
587 case ir_depth_layout_any
:
588 var
->data
.depth_layout
= nir_depth_layout_any
;
590 case ir_depth_layout_greater
:
591 var
->data
.depth_layout
= nir_depth_layout_greater
;
593 case ir_depth_layout_less
:
594 var
->data
.depth_layout
= nir_depth_layout_less
;
596 case ir_depth_layout_unchanged
:
597 var
->data
.depth_layout
= nir_depth_layout_unchanged
;
600 unreachable("not reached");
603 var
->data
.index
= ir
->data
.index
;
604 var
->data
.descriptor_set
= 0;
605 var
->data
.binding
= ir
->data
.binding
;
606 var
->data
.explicit_binding
= ir
->data
.explicit_binding
;
607 var
->data
.bindless
= ir
->data
.bindless
;
608 var
->data
.offset
= ir
->data
.offset
;
609 var
->data
.access
= (gl_access_qualifier
)mem_access
;
611 if (var
->type
->without_array()->is_image()) {
612 var
->data
.image
.format
= ir
->data
.image_format
;
613 } else if (var
->data
.mode
== nir_var_shader_out
) {
614 var
->data
.xfb
.buffer
= ir
->data
.xfb_buffer
;
615 var
->data
.xfb
.stride
= ir
->data
.xfb_stride
;
618 var
->data
.fb_fetch_output
= ir
->data
.fb_fetch_output
;
619 var
->data
.explicit_xfb_buffer
= ir
->data
.explicit_xfb_buffer
;
620 var
->data
.explicit_xfb_stride
= ir
->data
.explicit_xfb_stride
;
622 var
->num_state_slots
= ir
->get_num_state_slots();
623 if (var
->num_state_slots
> 0) {
624 var
->state_slots
= rzalloc_array(var
, nir_state_slot
,
625 var
->num_state_slots
);
627 ir_state_slot
*state_slots
= ir
->get_state_slots();
628 for (unsigned i
= 0; i
< var
->num_state_slots
; i
++) {
629 for (unsigned j
= 0; j
< 5; j
++)
630 var
->state_slots
[i
].tokens
[j
] = state_slots
[i
].tokens
[j
];
631 var
->state_slots
[i
].swizzle
= state_slots
[i
].swizzle
;
634 var
->state_slots
= NULL
;
637 var
->constant_initializer
= constant_copy(ir
->constant_initializer
, var
);
639 var
->interface_type
= ir
->get_interface_type();
641 if (var
->data
.mode
== nir_var_function_temp
)
642 nir_function_impl_add_variable(impl
, var
);
644 nir_shader_add_variable(shader
, var
);
646 _mesa_hash_table_insert(var_table
, ir
, var
);
650 nir_function_visitor::visit_enter(ir_function
*ir
)
652 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
) {
653 visitor
->create_function(sig
);
655 return visit_continue_with_parent
;
659 nir_visitor::create_function(ir_function_signature
*ir
)
661 if (ir
->is_intrinsic())
664 nir_function
*func
= nir_function_create(shader
, ir
->function_name());
665 if (strcmp(ir
->function_name(), "main") == 0)
666 func
->is_entrypoint
= true;
668 func
->num_params
= ir
->parameters
.length() +
669 (ir
->return_type
!= glsl_type::void_type
);
670 func
->params
= ralloc_array(shader
, nir_parameter
, func
->num_params
);
674 if (ir
->return_type
!= glsl_type::void_type
) {
675 /* The return value is a variable deref (basically an out parameter) */
676 func
->params
[np
].num_components
= 1;
677 func
->params
[np
].bit_size
= 32;
681 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
682 /* FINISHME: pass arrays, structs, etc by reference? */
683 assert(param
->type
->is_vector() || param
->type
->is_scalar());
685 if (param
->data
.mode
== ir_var_function_in
) {
686 func
->params
[np
].num_components
= param
->type
->vector_elements
;
687 func
->params
[np
].bit_size
= glsl_get_bit_size(param
->type
);
689 func
->params
[np
].num_components
= 1;
690 func
->params
[np
].bit_size
= 32;
694 assert(np
== func
->num_params
);
696 _mesa_hash_table_insert(this->overload_table
, ir
, func
);
700 nir_visitor::visit(ir_function
*ir
)
702 foreach_in_list(ir_function_signature
, sig
, &ir
->signatures
)
707 nir_visitor::visit(ir_function_signature
*ir
)
709 if (ir
->is_intrinsic())
714 struct hash_entry
*entry
=
715 _mesa_hash_table_search(this->overload_table
, ir
);
718 nir_function
*func
= (nir_function
*) entry
->data
;
720 if (ir
->is_defined
) {
721 nir_function_impl
*impl
= nir_function_impl_create(func
);
724 this->is_global
= false;
726 nir_builder_init(&b
, impl
);
727 b
.cursor
= nir_after_cf_list(&impl
->body
);
729 unsigned i
= (ir
->return_type
!= glsl_type::void_type
) ? 1 : 0;
731 foreach_in_list(ir_variable
, param
, &ir
->parameters
) {
733 nir_local_variable_create(impl
, param
->type
, param
->name
);
735 if (param
->data
.mode
== ir_var_function_in
) {
736 nir_store_var(&b
, var
, nir_load_param(&b
, i
), ~0);
739 _mesa_hash_table_insert(var_table
, param
, var
);
743 visit_exec_list(&ir
->body
, this);
745 this->is_global
= true;
752 nir_visitor::visit(ir_loop
*ir
)
755 visit_exec_list(&ir
->body_instructions
, this);
756 nir_pop_loop(&b
, NULL
);
760 nir_visitor::visit(ir_if
*ir
)
762 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
763 visit_exec_list(&ir
->then_instructions
, this);
764 nir_push_else(&b
, NULL
);
765 visit_exec_list(&ir
->else_instructions
, this);
766 nir_pop_if(&b
, NULL
);
770 nir_visitor::visit(ir_discard
*ir
)
773 * discards aren't treated as control flow, because before we lower them
774 * they can appear anywhere in the shader and the stuff after them may still
775 * be executed (yay, crazy GLSL rules!). However, after lowering, all the
776 * discards will be immediately followed by a return.
779 nir_intrinsic_instr
*discard
;
781 discard
= nir_intrinsic_instr_create(this->shader
,
782 nir_intrinsic_discard_if
);
784 nir_src_for_ssa(evaluate_rvalue(ir
->condition
));
786 discard
= nir_intrinsic_instr_create(this->shader
, nir_intrinsic_discard
);
789 nir_builder_instr_insert(&b
, &discard
->instr
);
793 nir_visitor::visit(ir_demote
*ir
)
795 nir_intrinsic_instr
*demote
=
796 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_demote
);
798 nir_builder_instr_insert(&b
, &demote
->instr
);
802 nir_visitor::visit(ir_emit_vertex
*ir
)
804 nir_intrinsic_instr
*instr
=
805 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_emit_vertex
);
806 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
807 nir_builder_instr_insert(&b
, &instr
->instr
);
811 nir_visitor::visit(ir_end_primitive
*ir
)
813 nir_intrinsic_instr
*instr
=
814 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_end_primitive
);
815 nir_intrinsic_set_stream_id(instr
, ir
->stream_id());
816 nir_builder_instr_insert(&b
, &instr
->instr
);
820 nir_visitor::visit(ir_loop_jump
*ir
)
824 case ir_loop_jump::jump_break
:
825 type
= nir_jump_break
;
827 case ir_loop_jump::jump_continue
:
828 type
= nir_jump_continue
;
831 unreachable("not reached");
834 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, type
);
835 nir_builder_instr_insert(&b
, &instr
->instr
);
839 nir_visitor::visit(ir_return
*ir
)
841 if (ir
->value
!= NULL
) {
842 nir_deref_instr
*ret_deref
=
843 nir_build_deref_cast(&b
, nir_load_param(&b
, 0),
844 nir_var_function_temp
, ir
->value
->type
, 0);
846 nir_ssa_def
*val
= evaluate_rvalue(ir
->value
);
847 nir_store_deref(&b
, ret_deref
, val
, ~0);
850 nir_jump_instr
*instr
= nir_jump_instr_create(this->shader
, nir_jump_return
);
851 nir_builder_instr_insert(&b
, &instr
->instr
);
855 intrinsic_set_std430_align(nir_intrinsic_instr
*intrin
, const glsl_type
*type
)
857 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
858 unsigned pow2_components
= util_next_power_of_two(type
->vector_elements
);
859 nir_intrinsic_set_align(intrin
, (bit_size
/ 8) * pow2_components
, 0);
862 /* Accumulate any qualifiers along the deref chain to get the actual
863 * load/store qualifier.
866 static enum gl_access_qualifier
867 deref_get_qualifier(nir_deref_instr
*deref
)
870 nir_deref_path_init(&path
, deref
, NULL
);
872 unsigned qualifiers
= path
.path
[0]->var
->data
.access
;
874 const glsl_type
*parent_type
= path
.path
[0]->type
;
875 for (nir_deref_instr
**cur_ptr
= &path
.path
[1]; *cur_ptr
; cur_ptr
++) {
876 nir_deref_instr
*cur
= *cur_ptr
;
878 if (parent_type
->is_interface()) {
879 const struct glsl_struct_field
*field
=
880 &parent_type
->fields
.structure
[cur
->strct
.index
];
881 if (field
->memory_read_only
)
882 qualifiers
|= ACCESS_NON_WRITEABLE
;
883 if (field
->memory_write_only
)
884 qualifiers
|= ACCESS_NON_READABLE
;
885 if (field
->memory_coherent
)
886 qualifiers
|= ACCESS_COHERENT
;
887 if (field
->memory_volatile
)
888 qualifiers
|= ACCESS_VOLATILE
;
889 if (field
->memory_restrict
)
890 qualifiers
|= ACCESS_RESTRICT
;
893 parent_type
= cur
->type
;
896 nir_deref_path_finish(&path
);
898 return (gl_access_qualifier
) qualifiers
;
902 nir_visitor::visit(ir_call
*ir
)
904 if (ir
->callee
->is_intrinsic()) {
907 switch (ir
->callee
->intrinsic_id
) {
908 case ir_intrinsic_generic_atomic_add
:
909 op
= ir
->return_deref
->type
->is_integer_32_64()
910 ? nir_intrinsic_deref_atomic_add
: nir_intrinsic_deref_atomic_fadd
;
912 case ir_intrinsic_generic_atomic_and
:
913 op
= nir_intrinsic_deref_atomic_and
;
915 case ir_intrinsic_generic_atomic_or
:
916 op
= nir_intrinsic_deref_atomic_or
;
918 case ir_intrinsic_generic_atomic_xor
:
919 op
= nir_intrinsic_deref_atomic_xor
;
921 case ir_intrinsic_generic_atomic_min
:
922 assert(ir
->return_deref
);
923 if (ir
->return_deref
->type
== glsl_type::int_type
)
924 op
= nir_intrinsic_deref_atomic_imin
;
925 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
926 op
= nir_intrinsic_deref_atomic_umin
;
927 else if (ir
->return_deref
->type
== glsl_type::float_type
)
928 op
= nir_intrinsic_deref_atomic_fmin
;
930 unreachable("Invalid type");
932 case ir_intrinsic_generic_atomic_max
:
933 assert(ir
->return_deref
);
934 if (ir
->return_deref
->type
== glsl_type::int_type
)
935 op
= nir_intrinsic_deref_atomic_imax
;
936 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
937 op
= nir_intrinsic_deref_atomic_umax
;
938 else if (ir
->return_deref
->type
== glsl_type::float_type
)
939 op
= nir_intrinsic_deref_atomic_fmax
;
941 unreachable("Invalid type");
943 case ir_intrinsic_generic_atomic_exchange
:
944 op
= nir_intrinsic_deref_atomic_exchange
;
946 case ir_intrinsic_generic_atomic_comp_swap
:
947 op
= ir
->return_deref
->type
->is_integer_32_64()
948 ? nir_intrinsic_deref_atomic_comp_swap
949 : nir_intrinsic_deref_atomic_fcomp_swap
;
951 case ir_intrinsic_atomic_counter_read
:
952 op
= nir_intrinsic_atomic_counter_read_deref
;
954 case ir_intrinsic_atomic_counter_increment
:
955 op
= nir_intrinsic_atomic_counter_inc_deref
;
957 case ir_intrinsic_atomic_counter_predecrement
:
958 op
= nir_intrinsic_atomic_counter_pre_dec_deref
;
960 case ir_intrinsic_atomic_counter_add
:
961 op
= nir_intrinsic_atomic_counter_add_deref
;
963 case ir_intrinsic_atomic_counter_and
:
964 op
= nir_intrinsic_atomic_counter_and_deref
;
966 case ir_intrinsic_atomic_counter_or
:
967 op
= nir_intrinsic_atomic_counter_or_deref
;
969 case ir_intrinsic_atomic_counter_xor
:
970 op
= nir_intrinsic_atomic_counter_xor_deref
;
972 case ir_intrinsic_atomic_counter_min
:
973 op
= nir_intrinsic_atomic_counter_min_deref
;
975 case ir_intrinsic_atomic_counter_max
:
976 op
= nir_intrinsic_atomic_counter_max_deref
;
978 case ir_intrinsic_atomic_counter_exchange
:
979 op
= nir_intrinsic_atomic_counter_exchange_deref
;
981 case ir_intrinsic_atomic_counter_comp_swap
:
982 op
= nir_intrinsic_atomic_counter_comp_swap_deref
;
984 case ir_intrinsic_image_load
:
985 op
= nir_intrinsic_image_deref_load
;
987 case ir_intrinsic_image_store
:
988 op
= nir_intrinsic_image_deref_store
;
990 case ir_intrinsic_image_atomic_add
:
991 op
= ir
->return_deref
->type
->is_integer_32_64()
992 ? nir_intrinsic_image_deref_atomic_add
993 : nir_intrinsic_image_deref_atomic_fadd
;
995 case ir_intrinsic_image_atomic_min
:
996 if (ir
->return_deref
->type
== glsl_type::int_type
)
997 op
= nir_intrinsic_image_deref_atomic_imin
;
998 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
999 op
= nir_intrinsic_image_deref_atomic_umin
;
1001 unreachable("Invalid type");
1003 case ir_intrinsic_image_atomic_max
:
1004 if (ir
->return_deref
->type
== glsl_type::int_type
)
1005 op
= nir_intrinsic_image_deref_atomic_imax
;
1006 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1007 op
= nir_intrinsic_image_deref_atomic_umax
;
1009 unreachable("Invalid type");
1011 case ir_intrinsic_image_atomic_and
:
1012 op
= nir_intrinsic_image_deref_atomic_and
;
1014 case ir_intrinsic_image_atomic_or
:
1015 op
= nir_intrinsic_image_deref_atomic_or
;
1017 case ir_intrinsic_image_atomic_xor
:
1018 op
= nir_intrinsic_image_deref_atomic_xor
;
1020 case ir_intrinsic_image_atomic_exchange
:
1021 op
= nir_intrinsic_image_deref_atomic_exchange
;
1023 case ir_intrinsic_image_atomic_comp_swap
:
1024 op
= nir_intrinsic_image_deref_atomic_comp_swap
;
1026 case ir_intrinsic_image_atomic_inc_wrap
:
1027 op
= nir_intrinsic_image_deref_atomic_inc_wrap
;
1029 case ir_intrinsic_image_atomic_dec_wrap
:
1030 op
= nir_intrinsic_image_deref_atomic_dec_wrap
;
1032 case ir_intrinsic_memory_barrier
:
1033 op
= nir_intrinsic_memory_barrier
;
1035 case ir_intrinsic_image_size
:
1036 op
= nir_intrinsic_image_deref_size
;
1038 case ir_intrinsic_image_samples
:
1039 op
= nir_intrinsic_image_deref_samples
;
1041 case ir_intrinsic_ssbo_store
:
1042 op
= nir_intrinsic_store_ssbo
;
1044 case ir_intrinsic_ssbo_load
:
1045 op
= nir_intrinsic_load_ssbo
;
1047 case ir_intrinsic_ssbo_atomic_add
:
1048 op
= ir
->return_deref
->type
->is_integer_32_64()
1049 ? nir_intrinsic_ssbo_atomic_add
: nir_intrinsic_ssbo_atomic_fadd
;
1051 case ir_intrinsic_ssbo_atomic_and
:
1052 op
= nir_intrinsic_ssbo_atomic_and
;
1054 case ir_intrinsic_ssbo_atomic_or
:
1055 op
= nir_intrinsic_ssbo_atomic_or
;
1057 case ir_intrinsic_ssbo_atomic_xor
:
1058 op
= nir_intrinsic_ssbo_atomic_xor
;
1060 case ir_intrinsic_ssbo_atomic_min
:
1061 assert(ir
->return_deref
);
1062 if (ir
->return_deref
->type
== glsl_type::int_type
)
1063 op
= nir_intrinsic_ssbo_atomic_imin
;
1064 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1065 op
= nir_intrinsic_ssbo_atomic_umin
;
1066 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1067 op
= nir_intrinsic_ssbo_atomic_fmin
;
1069 unreachable("Invalid type");
1071 case ir_intrinsic_ssbo_atomic_max
:
1072 assert(ir
->return_deref
);
1073 if (ir
->return_deref
->type
== glsl_type::int_type
)
1074 op
= nir_intrinsic_ssbo_atomic_imax
;
1075 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1076 op
= nir_intrinsic_ssbo_atomic_umax
;
1077 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1078 op
= nir_intrinsic_ssbo_atomic_fmax
;
1080 unreachable("Invalid type");
1082 case ir_intrinsic_ssbo_atomic_exchange
:
1083 op
= nir_intrinsic_ssbo_atomic_exchange
;
1085 case ir_intrinsic_ssbo_atomic_comp_swap
:
1086 op
= ir
->return_deref
->type
->is_integer_32_64()
1087 ? nir_intrinsic_ssbo_atomic_comp_swap
1088 : nir_intrinsic_ssbo_atomic_fcomp_swap
;
1090 case ir_intrinsic_shader_clock
:
1091 op
= nir_intrinsic_shader_clock
;
1093 case ir_intrinsic_begin_invocation_interlock
:
1094 op
= nir_intrinsic_begin_invocation_interlock
;
1096 case ir_intrinsic_end_invocation_interlock
:
1097 op
= nir_intrinsic_end_invocation_interlock
;
1099 case ir_intrinsic_group_memory_barrier
:
1100 op
= nir_intrinsic_group_memory_barrier
;
1102 case ir_intrinsic_memory_barrier_atomic_counter
:
1103 op
= nir_intrinsic_memory_barrier_atomic_counter
;
1105 case ir_intrinsic_memory_barrier_buffer
:
1106 op
= nir_intrinsic_memory_barrier_buffer
;
1108 case ir_intrinsic_memory_barrier_image
:
1109 op
= nir_intrinsic_memory_barrier_image
;
1111 case ir_intrinsic_memory_barrier_shared
:
1112 op
= nir_intrinsic_memory_barrier_shared
;
1114 case ir_intrinsic_shared_load
:
1115 op
= nir_intrinsic_load_shared
;
1117 case ir_intrinsic_shared_store
:
1118 op
= nir_intrinsic_store_shared
;
1120 case ir_intrinsic_shared_atomic_add
:
1121 op
= ir
->return_deref
->type
->is_integer_32_64()
1122 ? nir_intrinsic_shared_atomic_add
1123 : nir_intrinsic_shared_atomic_fadd
;
1125 case ir_intrinsic_shared_atomic_and
:
1126 op
= nir_intrinsic_shared_atomic_and
;
1128 case ir_intrinsic_shared_atomic_or
:
1129 op
= nir_intrinsic_shared_atomic_or
;
1131 case ir_intrinsic_shared_atomic_xor
:
1132 op
= nir_intrinsic_shared_atomic_xor
;
1134 case ir_intrinsic_shared_atomic_min
:
1135 assert(ir
->return_deref
);
1136 if (ir
->return_deref
->type
== glsl_type::int_type
)
1137 op
= nir_intrinsic_shared_atomic_imin
;
1138 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1139 op
= nir_intrinsic_shared_atomic_umin
;
1140 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1141 op
= nir_intrinsic_shared_atomic_fmin
;
1143 unreachable("Invalid type");
1145 case ir_intrinsic_shared_atomic_max
:
1146 assert(ir
->return_deref
);
1147 if (ir
->return_deref
->type
== glsl_type::int_type
)
1148 op
= nir_intrinsic_shared_atomic_imax
;
1149 else if (ir
->return_deref
->type
== glsl_type::uint_type
)
1150 op
= nir_intrinsic_shared_atomic_umax
;
1151 else if (ir
->return_deref
->type
== glsl_type::float_type
)
1152 op
= nir_intrinsic_shared_atomic_fmax
;
1154 unreachable("Invalid type");
1156 case ir_intrinsic_shared_atomic_exchange
:
1157 op
= nir_intrinsic_shared_atomic_exchange
;
1159 case ir_intrinsic_shared_atomic_comp_swap
:
1160 op
= ir
->return_deref
->type
->is_integer_32_64()
1161 ? nir_intrinsic_shared_atomic_comp_swap
1162 : nir_intrinsic_shared_atomic_fcomp_swap
;
1164 case ir_intrinsic_vote_any
:
1165 op
= nir_intrinsic_vote_any
;
1167 case ir_intrinsic_vote_all
:
1168 op
= nir_intrinsic_vote_all
;
1170 case ir_intrinsic_vote_eq
:
1171 op
= nir_intrinsic_vote_ieq
;
1173 case ir_intrinsic_ballot
:
1174 op
= nir_intrinsic_ballot
;
1176 case ir_intrinsic_read_invocation
:
1177 op
= nir_intrinsic_read_invocation
;
1179 case ir_intrinsic_read_first_invocation
:
1180 op
= nir_intrinsic_read_first_invocation
;
1182 case ir_intrinsic_helper_invocation
:
1183 op
= nir_intrinsic_is_helper_invocation
;
1186 unreachable("not reached");
1189 nir_intrinsic_instr
*instr
= nir_intrinsic_instr_create(shader
, op
);
1190 nir_ssa_def
*ret
= &instr
->dest
.ssa
;
1193 case nir_intrinsic_deref_atomic_add
:
1194 case nir_intrinsic_deref_atomic_imin
:
1195 case nir_intrinsic_deref_atomic_umin
:
1196 case nir_intrinsic_deref_atomic_imax
:
1197 case nir_intrinsic_deref_atomic_umax
:
1198 case nir_intrinsic_deref_atomic_and
:
1199 case nir_intrinsic_deref_atomic_or
:
1200 case nir_intrinsic_deref_atomic_xor
:
1201 case nir_intrinsic_deref_atomic_exchange
:
1202 case nir_intrinsic_deref_atomic_comp_swap
:
1203 case nir_intrinsic_deref_atomic_fadd
:
1204 case nir_intrinsic_deref_atomic_fmin
:
1205 case nir_intrinsic_deref_atomic_fmax
:
1206 case nir_intrinsic_deref_atomic_fcomp_swap
: {
1207 int param_count
= ir
->actual_parameters
.length();
1208 assert(param_count
== 2 || param_count
== 3);
1211 exec_node
*param
= ir
->actual_parameters
.get_head();
1212 ir_rvalue
*rvalue
= (ir_rvalue
*) param
;
1213 ir_dereference
*deref
= rvalue
->as_dereference();
1214 ir_swizzle
*swizzle
= NULL
;
1216 /* We may have a swizzle to pick off a single vec4 component */
1217 swizzle
= rvalue
->as_swizzle();
1218 assert(swizzle
&& swizzle
->type
->vector_elements
== 1);
1219 deref
= swizzle
->val
->as_dereference();
1222 nir_deref_instr
*nir_deref
= evaluate_deref(deref
);
1224 nir_deref
= nir_build_deref_array_imm(&b
, nir_deref
,
1227 instr
->src
[0] = nir_src_for_ssa(&nir_deref
->dest
.ssa
);
1229 nir_intrinsic_set_access(instr
, deref_get_qualifier(nir_deref
));
1231 /* data1 parameter (this is always present) */
1232 param
= param
->get_next();
1233 ir_instruction
*inst
= (ir_instruction
*) param
;
1234 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1236 /* data2 parameter (only with atomic_comp_swap) */
1237 if (param_count
== 3) {
1238 assert(op
== nir_intrinsic_deref_atomic_comp_swap
||
1239 op
== nir_intrinsic_deref_atomic_fcomp_swap
);
1240 param
= param
->get_next();
1241 inst
= (ir_instruction
*) param
;
1242 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1246 assert(ir
->return_deref
);
1247 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1248 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1249 nir_builder_instr_insert(&b
, &instr
->instr
);
1252 case nir_intrinsic_atomic_counter_read_deref
:
1253 case nir_intrinsic_atomic_counter_inc_deref
:
1254 case nir_intrinsic_atomic_counter_pre_dec_deref
:
1255 case nir_intrinsic_atomic_counter_add_deref
:
1256 case nir_intrinsic_atomic_counter_min_deref
:
1257 case nir_intrinsic_atomic_counter_max_deref
:
1258 case nir_intrinsic_atomic_counter_and_deref
:
1259 case nir_intrinsic_atomic_counter_or_deref
:
1260 case nir_intrinsic_atomic_counter_xor_deref
:
1261 case nir_intrinsic_atomic_counter_exchange_deref
:
1262 case nir_intrinsic_atomic_counter_comp_swap_deref
: {
1263 /* Set the counter variable dereference. */
1264 exec_node
*param
= ir
->actual_parameters
.get_head();
1265 ir_dereference
*counter
= (ir_dereference
*)param
;
1267 instr
->src
[0] = nir_src_for_ssa(&evaluate_deref(counter
)->dest
.ssa
);
1268 param
= param
->get_next();
1270 /* Set the intrinsic destination. */
1271 if (ir
->return_deref
) {
1272 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 32, NULL
);
1275 /* Set the intrinsic parameters. */
1276 if (!param
->is_tail_sentinel()) {
1278 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1279 param
= param
->get_next();
1282 if (!param
->is_tail_sentinel()) {
1284 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1285 param
= param
->get_next();
1288 nir_builder_instr_insert(&b
, &instr
->instr
);
1291 case nir_intrinsic_image_deref_load
:
1292 case nir_intrinsic_image_deref_store
:
1293 case nir_intrinsic_image_deref_atomic_add
:
1294 case nir_intrinsic_image_deref_atomic_imin
:
1295 case nir_intrinsic_image_deref_atomic_umin
:
1296 case nir_intrinsic_image_deref_atomic_imax
:
1297 case nir_intrinsic_image_deref_atomic_umax
:
1298 case nir_intrinsic_image_deref_atomic_and
:
1299 case nir_intrinsic_image_deref_atomic_or
:
1300 case nir_intrinsic_image_deref_atomic_xor
:
1301 case nir_intrinsic_image_deref_atomic_exchange
:
1302 case nir_intrinsic_image_deref_atomic_comp_swap
:
1303 case nir_intrinsic_image_deref_atomic_fadd
:
1304 case nir_intrinsic_image_deref_samples
:
1305 case nir_intrinsic_image_deref_size
:
1306 case nir_intrinsic_image_deref_atomic_inc_wrap
:
1307 case nir_intrinsic_image_deref_atomic_dec_wrap
: {
1308 nir_ssa_undef_instr
*instr_undef
=
1309 nir_ssa_undef_instr_create(shader
, 1, 32);
1310 nir_builder_instr_insert(&b
, &instr_undef
->instr
);
1312 /* Set the image variable dereference. */
1313 exec_node
*param
= ir
->actual_parameters
.get_head();
1314 ir_dereference
*image
= (ir_dereference
*)param
;
1315 nir_deref_instr
*deref
= evaluate_deref(image
);
1316 const glsl_type
*type
= deref
->type
;
1318 nir_intrinsic_set_access(instr
, deref_get_qualifier(deref
));
1320 instr
->src
[0] = nir_src_for_ssa(&deref
->dest
.ssa
);
1321 param
= param
->get_next();
1323 /* Set the intrinsic destination. */
1324 if (ir
->return_deref
) {
1325 unsigned num_components
= ir
->return_deref
->type
->vector_elements
;
1326 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1327 num_components
, 32, NULL
);
1330 if (op
== nir_intrinsic_image_deref_size
) {
1331 instr
->num_components
= instr
->dest
.ssa
.num_components
;
1332 } else if (op
== nir_intrinsic_image_deref_load
||
1333 op
== nir_intrinsic_image_deref_store
) {
1334 instr
->num_components
= 4;
1337 if (op
== nir_intrinsic_image_deref_size
||
1338 op
== nir_intrinsic_image_deref_samples
) {
1339 nir_builder_instr_insert(&b
, &instr
->instr
);
1343 /* Set the address argument, extending the coordinate vector to four
1346 nir_ssa_def
*src_addr
=
1347 evaluate_rvalue((ir_dereference
*)param
);
1348 nir_ssa_def
*srcs
[4];
1350 for (int i
= 0; i
< 4; i
++) {
1351 if (i
< type
->coordinate_components())
1352 srcs
[i
] = nir_channel(&b
, src_addr
, i
);
1354 srcs
[i
] = &instr_undef
->def
;
1357 instr
->src
[1] = nir_src_for_ssa(nir_vec(&b
, srcs
, 4));
1358 param
= param
->get_next();
1360 /* Set the sample argument, which is undefined for single-sample
1363 if (type
->sampler_dimensionality
== GLSL_SAMPLER_DIM_MS
) {
1365 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1366 param
= param
->get_next();
1368 instr
->src
[2] = nir_src_for_ssa(&instr_undef
->def
);
1371 /* Set the intrinsic parameters. */
1372 if (!param
->is_tail_sentinel()) {
1374 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1375 param
= param
->get_next();
1378 if (!param
->is_tail_sentinel()) {
1380 nir_src_for_ssa(evaluate_rvalue((ir_dereference
*)param
));
1381 param
= param
->get_next();
1383 nir_builder_instr_insert(&b
, &instr
->instr
);
1386 case nir_intrinsic_memory_barrier
:
1387 case nir_intrinsic_group_memory_barrier
:
1388 case nir_intrinsic_memory_barrier_atomic_counter
:
1389 case nir_intrinsic_memory_barrier_buffer
:
1390 case nir_intrinsic_memory_barrier_image
:
1391 case nir_intrinsic_memory_barrier_shared
:
1392 nir_builder_instr_insert(&b
, &instr
->instr
);
1394 case nir_intrinsic_shader_clock
:
1395 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 2, 32, NULL
);
1396 instr
->num_components
= 2;
1397 nir_builder_instr_insert(&b
, &instr
->instr
);
1399 case nir_intrinsic_begin_invocation_interlock
:
1400 nir_builder_instr_insert(&b
, &instr
->instr
);
1402 case nir_intrinsic_end_invocation_interlock
:
1403 nir_builder_instr_insert(&b
, &instr
->instr
);
1405 case nir_intrinsic_store_ssbo
: {
1406 exec_node
*param
= ir
->actual_parameters
.get_head();
1407 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1409 param
= param
->get_next();
1410 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1412 param
= param
->get_next();
1413 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1415 param
= param
->get_next();
1416 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1419 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1420 if (val
->type
->is_boolean())
1421 nir_val
= nir_b2i32(&b
, nir_val
);
1423 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1424 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(block
));
1425 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(offset
));
1426 intrinsic_set_std430_align(instr
, val
->type
);
1427 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1428 instr
->num_components
= val
->type
->vector_elements
;
1430 nir_builder_instr_insert(&b
, &instr
->instr
);
1433 case nir_intrinsic_load_ssbo
: {
1434 exec_node
*param
= ir
->actual_parameters
.get_head();
1435 ir_rvalue
*block
= ((ir_instruction
*)param
)->as_rvalue();
1437 param
= param
->get_next();
1438 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1440 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(block
));
1441 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1443 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1444 instr
->num_components
= type
->vector_elements
;
1445 intrinsic_set_std430_align(instr
, type
);
1447 /* Setup destination register */
1448 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1449 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1450 type
->vector_elements
, bit_size
, NULL
);
1452 /* Insert the created nir instruction now since in the case of boolean
1453 * result we will need to emit another instruction after it
1455 nir_builder_instr_insert(&b
, &instr
->instr
);
1458 * In SSBO/UBO's, a true boolean value is any non-zero value, but we
1459 * consider a true boolean to be ~0. Fix this up with a != 0
1462 if (type
->is_boolean())
1463 ret
= nir_i2b(&b
, &instr
->dest
.ssa
);
1466 case nir_intrinsic_ssbo_atomic_add
:
1467 case nir_intrinsic_ssbo_atomic_imin
:
1468 case nir_intrinsic_ssbo_atomic_umin
:
1469 case nir_intrinsic_ssbo_atomic_imax
:
1470 case nir_intrinsic_ssbo_atomic_umax
:
1471 case nir_intrinsic_ssbo_atomic_and
:
1472 case nir_intrinsic_ssbo_atomic_or
:
1473 case nir_intrinsic_ssbo_atomic_xor
:
1474 case nir_intrinsic_ssbo_atomic_exchange
:
1475 case nir_intrinsic_ssbo_atomic_comp_swap
:
1476 case nir_intrinsic_ssbo_atomic_fadd
:
1477 case nir_intrinsic_ssbo_atomic_fmin
:
1478 case nir_intrinsic_ssbo_atomic_fmax
:
1479 case nir_intrinsic_ssbo_atomic_fcomp_swap
: {
1480 int param_count
= ir
->actual_parameters
.length();
1481 assert(param_count
== 3 || param_count
== 4);
1484 exec_node
*param
= ir
->actual_parameters
.get_head();
1485 ir_instruction
*inst
= (ir_instruction
*) param
;
1486 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1489 param
= param
->get_next();
1490 inst
= (ir_instruction
*) param
;
1491 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1493 /* data1 parameter (this is always present) */
1494 param
= param
->get_next();
1495 inst
= (ir_instruction
*) param
;
1496 instr
->src
[2] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1498 /* data2 parameter (only with atomic_comp_swap) */
1499 if (param_count
== 4) {
1500 assert(op
== nir_intrinsic_ssbo_atomic_comp_swap
||
1501 op
== nir_intrinsic_ssbo_atomic_fcomp_swap
);
1502 param
= param
->get_next();
1503 inst
= (ir_instruction
*) param
;
1504 instr
->src
[3] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1508 assert(ir
->return_deref
);
1509 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1510 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1511 nir_builder_instr_insert(&b
, &instr
->instr
);
1514 case nir_intrinsic_load_shared
: {
1515 exec_node
*param
= ir
->actual_parameters
.get_head();
1516 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1518 nir_intrinsic_set_base(instr
, 0);
1519 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(offset
));
1521 const glsl_type
*type
= ir
->return_deref
->var
->type
;
1522 instr
->num_components
= type
->vector_elements
;
1523 intrinsic_set_std430_align(instr
, type
);
1525 /* Setup destination register */
1526 unsigned bit_size
= type
->is_boolean() ? 32 : glsl_get_bit_size(type
);
1527 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1528 type
->vector_elements
, bit_size
, NULL
);
1530 nir_builder_instr_insert(&b
, &instr
->instr
);
1532 /* The value in shared memory is a 32-bit value */
1533 if (type
->is_boolean())
1534 ret
= nir_i2b(&b
, &instr
->dest
.ssa
);
1537 case nir_intrinsic_store_shared
: {
1538 exec_node
*param
= ir
->actual_parameters
.get_head();
1539 ir_rvalue
*offset
= ((ir_instruction
*)param
)->as_rvalue();
1541 param
= param
->get_next();
1542 ir_rvalue
*val
= ((ir_instruction
*)param
)->as_rvalue();
1544 param
= param
->get_next();
1545 ir_constant
*write_mask
= ((ir_instruction
*)param
)->as_constant();
1548 nir_intrinsic_set_base(instr
, 0);
1549 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(offset
));
1551 nir_intrinsic_set_write_mask(instr
, write_mask
->value
.u
[0]);
1553 nir_ssa_def
*nir_val
= evaluate_rvalue(val
);
1554 /* The value in shared memory is a 32-bit value */
1555 if (val
->type
->is_boolean())
1556 nir_val
= nir_b2i32(&b
, nir_val
);
1558 instr
->src
[0] = nir_src_for_ssa(nir_val
);
1559 instr
->num_components
= val
->type
->vector_elements
;
1560 intrinsic_set_std430_align(instr
, val
->type
);
1562 nir_builder_instr_insert(&b
, &instr
->instr
);
1565 case nir_intrinsic_shared_atomic_add
:
1566 case nir_intrinsic_shared_atomic_imin
:
1567 case nir_intrinsic_shared_atomic_umin
:
1568 case nir_intrinsic_shared_atomic_imax
:
1569 case nir_intrinsic_shared_atomic_umax
:
1570 case nir_intrinsic_shared_atomic_and
:
1571 case nir_intrinsic_shared_atomic_or
:
1572 case nir_intrinsic_shared_atomic_xor
:
1573 case nir_intrinsic_shared_atomic_exchange
:
1574 case nir_intrinsic_shared_atomic_comp_swap
:
1575 case nir_intrinsic_shared_atomic_fadd
:
1576 case nir_intrinsic_shared_atomic_fmin
:
1577 case nir_intrinsic_shared_atomic_fmax
:
1578 case nir_intrinsic_shared_atomic_fcomp_swap
: {
1579 int param_count
= ir
->actual_parameters
.length();
1580 assert(param_count
== 2 || param_count
== 3);
1583 exec_node
*param
= ir
->actual_parameters
.get_head();
1584 ir_instruction
*inst
= (ir_instruction
*) param
;
1585 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1587 /* data1 parameter (this is always present) */
1588 param
= param
->get_next();
1589 inst
= (ir_instruction
*) param
;
1590 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1592 /* data2 parameter (only with atomic_comp_swap) */
1593 if (param_count
== 3) {
1594 assert(op
== nir_intrinsic_shared_atomic_comp_swap
||
1595 op
== nir_intrinsic_shared_atomic_fcomp_swap
);
1596 param
= param
->get_next();
1597 inst
= (ir_instruction
*) param
;
1599 nir_src_for_ssa(evaluate_rvalue(inst
->as_rvalue()));
1603 assert(ir
->return_deref
);
1604 unsigned bit_size
= glsl_get_bit_size(ir
->return_deref
->type
);
1605 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1606 ir
->return_deref
->type
->vector_elements
,
1608 nir_builder_instr_insert(&b
, &instr
->instr
);
1611 case nir_intrinsic_vote_any
:
1612 case nir_intrinsic_vote_all
:
1613 case nir_intrinsic_vote_ieq
: {
1614 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 1, NULL
);
1615 instr
->num_components
= 1;
1617 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1618 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1620 nir_builder_instr_insert(&b
, &instr
->instr
);
1624 case nir_intrinsic_ballot
: {
1625 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1626 ir
->return_deref
->type
->vector_elements
, 64, NULL
);
1627 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1629 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1630 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1632 nir_builder_instr_insert(&b
, &instr
->instr
);
1635 case nir_intrinsic_read_invocation
: {
1636 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1637 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1638 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1640 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1641 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1643 ir_rvalue
*invocation
= (ir_rvalue
*) ir
->actual_parameters
.get_head()->next
;
1644 instr
->src
[1] = nir_src_for_ssa(evaluate_rvalue(invocation
));
1646 nir_builder_instr_insert(&b
, &instr
->instr
);
1649 case nir_intrinsic_read_first_invocation
: {
1650 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
,
1651 ir
->return_deref
->type
->vector_elements
, 32, NULL
);
1652 instr
->num_components
= ir
->return_deref
->type
->vector_elements
;
1654 ir_rvalue
*value
= (ir_rvalue
*) ir
->actual_parameters
.get_head();
1655 instr
->src
[0] = nir_src_for_ssa(evaluate_rvalue(value
));
1657 nir_builder_instr_insert(&b
, &instr
->instr
);
1660 case nir_intrinsic_is_helper_invocation
: {
1661 nir_ssa_dest_init(&instr
->instr
, &instr
->dest
, 1, 1, NULL
);
1662 instr
->num_components
= 1;
1663 nir_builder_instr_insert(&b
, &instr
->instr
);
1667 unreachable("not reached");
1670 if (ir
->return_deref
)
1671 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), ret
, ~0);
1676 struct hash_entry
*entry
=
1677 _mesa_hash_table_search(this->overload_table
, ir
->callee
);
1679 nir_function
*callee
= (nir_function
*) entry
->data
;
1681 nir_call_instr
*call
= nir_call_instr_create(this->shader
, callee
);
1684 nir_deref_instr
*ret_deref
= NULL
;
1685 if (ir
->return_deref
) {
1686 nir_variable
*ret_tmp
=
1687 nir_local_variable_create(this->impl
, ir
->return_deref
->type
,
1689 ret_deref
= nir_build_deref_var(&b
, ret_tmp
);
1690 call
->params
[i
++] = nir_src_for_ssa(&ret_deref
->dest
.ssa
);
1693 foreach_two_lists(formal_node
, &ir
->callee
->parameters
,
1694 actual_node
, &ir
->actual_parameters
) {
1695 ir_rvalue
*param_rvalue
= (ir_rvalue
*) actual_node
;
1696 ir_variable
*sig_param
= (ir_variable
*) formal_node
;
1698 if (sig_param
->data
.mode
== ir_var_function_out
) {
1699 nir_deref_instr
*out_deref
= evaluate_deref(param_rvalue
);
1700 call
->params
[i
] = nir_src_for_ssa(&out_deref
->dest
.ssa
);
1701 } else if (sig_param
->data
.mode
== ir_var_function_in
) {
1702 nir_ssa_def
*val
= evaluate_rvalue(param_rvalue
);
1703 nir_src src
= nir_src_for_ssa(val
);
1705 nir_src_copy(&call
->params
[i
], &src
, call
);
1706 } else if (sig_param
->data
.mode
== ir_var_function_inout
) {
1707 unreachable("unimplemented: inout parameters");
1713 nir_builder_instr_insert(&b
, &call
->instr
);
1715 if (ir
->return_deref
)
1716 nir_store_deref(&b
, evaluate_deref(ir
->return_deref
), nir_load_deref(&b
, ret_deref
), ~0);
1720 nir_visitor::visit(ir_assignment
*ir
)
1722 unsigned num_components
= ir
->lhs
->type
->vector_elements
;
1724 b
.exact
= ir
->lhs
->variable_referenced()->data
.invariant
||
1725 ir
->lhs
->variable_referenced()->data
.precise
;
1727 if ((ir
->rhs
->as_dereference() || ir
->rhs
->as_constant()) &&
1728 (ir
->write_mask
== (1 << num_components
) - 1 || ir
->write_mask
== 0)) {
1729 nir_deref_instr
*lhs
= evaluate_deref(ir
->lhs
);
1730 nir_deref_instr
*rhs
= evaluate_deref(ir
->rhs
);
1731 enum gl_access_qualifier lhs_qualifiers
= deref_get_qualifier(lhs
);
1732 enum gl_access_qualifier rhs_qualifiers
= deref_get_qualifier(rhs
);
1733 if (ir
->condition
) {
1734 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1735 nir_copy_deref_with_access(&b
, lhs
, rhs
, lhs_qualifiers
,
1737 nir_pop_if(&b
, NULL
);
1739 nir_copy_deref_with_access(&b
, lhs
, rhs
, lhs_qualifiers
,
1745 assert(ir
->rhs
->type
->is_scalar() || ir
->rhs
->type
->is_vector());
1747 ir
->lhs
->accept(this);
1748 nir_deref_instr
*lhs_deref
= this->deref
;
1749 nir_ssa_def
*src
= evaluate_rvalue(ir
->rhs
);
1751 if (ir
->write_mask
!= (1 << num_components
) - 1 && ir
->write_mask
!= 0) {
1752 /* GLSL IR will give us the input to the write-masked assignment in a
1753 * single packed vector. So, for example, if the writemask is xzw, then
1754 * we have to swizzle x -> x, y -> z, and z -> w and get the y component
1758 unsigned component
= 0;
1759 for (unsigned i
= 0; i
< 4; i
++) {
1760 swiz
[i
] = ir
->write_mask
& (1 << i
) ? component
++ : 0;
1762 src
= nir_swizzle(&b
, src
, swiz
, num_components
);
1765 enum gl_access_qualifier qualifiers
= deref_get_qualifier(lhs_deref
);
1766 if (ir
->condition
) {
1767 nir_push_if(&b
, evaluate_rvalue(ir
->condition
));
1768 nir_store_deref_with_access(&b
, lhs_deref
, src
, ir
->write_mask
,
1770 nir_pop_if(&b
, NULL
);
1772 nir_store_deref_with_access(&b
, lhs_deref
, src
, ir
->write_mask
,
1778 * Given an instruction, returns a pointer to its destination or NULL if there
1779 * is no destination.
1781 * Note that this only handles instructions we generate at this level.
1784 get_instr_dest(nir_instr
*instr
)
1786 nir_alu_instr
*alu_instr
;
1787 nir_intrinsic_instr
*intrinsic_instr
;
1788 nir_tex_instr
*tex_instr
;
1790 switch (instr
->type
) {
1791 case nir_instr_type_alu
:
1792 alu_instr
= nir_instr_as_alu(instr
);
1793 return &alu_instr
->dest
.dest
;
1795 case nir_instr_type_intrinsic
:
1796 intrinsic_instr
= nir_instr_as_intrinsic(instr
);
1797 if (nir_intrinsic_infos
[intrinsic_instr
->intrinsic
].has_dest
)
1798 return &intrinsic_instr
->dest
;
1802 case nir_instr_type_tex
:
1803 tex_instr
= nir_instr_as_tex(instr
);
1804 return &tex_instr
->dest
;
1807 unreachable("not reached");
1814 nir_visitor::add_instr(nir_instr
*instr
, unsigned num_components
,
1817 nir_dest
*dest
= get_instr_dest(instr
);
1820 nir_ssa_dest_init(instr
, dest
, num_components
, bit_size
, NULL
);
1822 nir_builder_instr_insert(&b
, instr
);
1825 assert(dest
->is_ssa
);
1826 this->result
= &dest
->ssa
;
1831 nir_visitor::evaluate_rvalue(ir_rvalue
* ir
)
1834 if (ir
->as_dereference() || ir
->as_constant()) {
1836 * A dereference is being used on the right hand side, which means we
1837 * must emit a variable load.
1840 enum gl_access_qualifier access
= deref_get_qualifier(this->deref
);
1841 this->result
= nir_load_deref_with_access(&b
, this->deref
, access
);
1844 return this->result
;
1848 type_is_float(glsl_base_type type
)
1850 return type
== GLSL_TYPE_FLOAT
|| type
== GLSL_TYPE_DOUBLE
||
1851 type
== GLSL_TYPE_FLOAT16
;
1855 type_is_signed(glsl_base_type type
)
1857 return type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_INT64
||
1858 type
== GLSL_TYPE_INT16
;
1862 nir_visitor::visit(ir_expression
*ir
)
1864 /* Some special cases */
1865 switch (ir
->operation
) {
1866 case ir_binop_ubo_load
: {
1867 nir_intrinsic_instr
*load
=
1868 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_load_ubo
);
1869 unsigned bit_size
= ir
->type
->is_boolean() ? 32 :
1870 glsl_get_bit_size(ir
->type
);
1871 load
->num_components
= ir
->type
->vector_elements
;
1872 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
1873 load
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1874 intrinsic_set_std430_align(load
, ir
->type
);
1875 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
1878 * In UBO's, a true boolean value is any non-zero value, but we consider
1879 * a true boolean to be ~0. Fix this up with a != 0 comparison.
1882 if (ir
->type
->is_boolean())
1883 this->result
= nir_i2b(&b
, &load
->dest
.ssa
);
1888 case ir_unop_interpolate_at_centroid
:
1889 case ir_binop_interpolate_at_offset
:
1890 case ir_binop_interpolate_at_sample
: {
1891 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1892 ir_swizzle
*swizzle
= NULL
;
1894 /* the api does not allow a swizzle here, but the varying packing code
1895 * may have pushed one into here.
1897 swizzle
= ir
->operands
[0]->as_swizzle();
1899 deref
= swizzle
->val
->as_dereference();
1903 deref
->accept(this);
1905 nir_intrinsic_op op
;
1906 if (this->deref
->mode
== nir_var_shader_in
) {
1907 switch (ir
->operation
) {
1908 case ir_unop_interpolate_at_centroid
:
1909 op
= nir_intrinsic_interp_deref_at_centroid
;
1911 case ir_binop_interpolate_at_offset
:
1912 op
= nir_intrinsic_interp_deref_at_offset
;
1914 case ir_binop_interpolate_at_sample
:
1915 op
= nir_intrinsic_interp_deref_at_sample
;
1918 unreachable("Invalid interpolation intrinsic");
1921 /* This case can happen if the vertex shader does not write the
1922 * given varying. In this case, the linker will lower it to a
1923 * global variable. Since interpolating a variable makes no
1924 * sense, we'll just turn it into a load which will probably
1925 * eventually end up as an SSA definition.
1927 assert(this->deref
->mode
== nir_var_shader_temp
);
1928 op
= nir_intrinsic_load_deref
;
1931 nir_intrinsic_instr
*intrin
= nir_intrinsic_instr_create(shader
, op
);
1932 intrin
->num_components
= deref
->type
->vector_elements
;
1933 intrin
->src
[0] = nir_src_for_ssa(&this->deref
->dest
.ssa
);
1935 if (intrin
->intrinsic
== nir_intrinsic_interp_deref_at_offset
||
1936 intrin
->intrinsic
== nir_intrinsic_interp_deref_at_sample
)
1937 intrin
->src
[1] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[1]));
1939 unsigned bit_size
= glsl_get_bit_size(deref
->type
);
1940 add_instr(&intrin
->instr
, deref
->type
->vector_elements
, bit_size
);
1943 unsigned swiz
[4] = {
1944 swizzle
->mask
.x
, swizzle
->mask
.y
, swizzle
->mask
.z
, swizzle
->mask
.w
1947 result
= nir_swizzle(&b
, result
, swiz
,
1948 swizzle
->type
->vector_elements
);
1954 case ir_unop_ssbo_unsized_array_length
: {
1955 nir_intrinsic_instr
*intrin
=
1956 nir_intrinsic_instr_create(b
.shader
,
1957 nir_intrinsic_deref_buffer_array_length
);
1959 ir_dereference
*deref
= ir
->operands
[0]->as_dereference();
1960 intrin
->src
[0] = nir_src_for_ssa(&evaluate_deref(deref
)->dest
.ssa
);
1962 add_instr(&intrin
->instr
, 1, 32);
1970 nir_ssa_def
*srcs
[4];
1971 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1972 srcs
[i
] = evaluate_rvalue(ir
->operands
[i
]);
1974 glsl_base_type types
[4];
1975 for (unsigned i
= 0; i
< ir
->num_operands
; i
++)
1976 types
[i
] = ir
->operands
[i
]->type
->base_type
;
1978 glsl_base_type out_type
= ir
->type
->base_type
;
1980 switch (ir
->operation
) {
1981 case ir_unop_bit_not
: result
= nir_inot(&b
, srcs
[0]); break;
1982 case ir_unop_logic_not
:
1983 result
= nir_inot(&b
, srcs
[0]);
1986 result
= type_is_float(types
[0]) ? nir_fneg(&b
, srcs
[0])
1987 : nir_ineg(&b
, srcs
[0]);
1990 result
= type_is_float(types
[0]) ? nir_fabs(&b
, srcs
[0])
1991 : nir_iabs(&b
, srcs
[0]);
1993 case ir_unop_saturate
:
1994 assert(type_is_float(types
[0]));
1995 result
= nir_fsat(&b
, srcs
[0]);
1998 result
= type_is_float(types
[0]) ? nir_fsign(&b
, srcs
[0])
1999 : nir_isign(&b
, srcs
[0]);
2001 case ir_unop_rcp
: result
= nir_frcp(&b
, srcs
[0]); break;
2002 case ir_unop_rsq
: result
= nir_frsq(&b
, srcs
[0]); break;
2003 case ir_unop_sqrt
: result
= nir_fsqrt(&b
, srcs
[0]); break;
2004 case ir_unop_exp
: unreachable("ir_unop_exp should have been lowered");
2005 case ir_unop_log
: unreachable("ir_unop_log should have been lowered");
2006 case ir_unop_exp2
: result
= nir_fexp2(&b
, srcs
[0]); break;
2007 case ir_unop_log2
: result
= nir_flog2(&b
, srcs
[0]); break;
2043 case ir_unop_i642u64
:
2044 case ir_unop_u642i64
: {
2045 nir_alu_type src_type
= nir_get_nir_type_for_glsl_base_type(types
[0]);
2046 nir_alu_type dst_type
= nir_get_nir_type_for_glsl_base_type(out_type
);
2047 result
= nir_build_alu(&b
, nir_type_conversion_op(src_type
, dst_type
,
2048 nir_rounding_mode_undef
),
2049 srcs
[0], NULL
, NULL
, NULL
);
2050 /* b2i and b2f don't have fixed bit-size versions so the builder will
2051 * just assume 32 and we have to fix it up here.
2053 result
->bit_size
= nir_alu_type_get_type_size(dst_type
);
2057 case ir_unop_bitcast_i2f
:
2058 case ir_unop_bitcast_f2i
:
2059 case ir_unop_bitcast_u2f
:
2060 case ir_unop_bitcast_f2u
:
2061 case ir_unop_bitcast_i642d
:
2062 case ir_unop_bitcast_d2i64
:
2063 case ir_unop_bitcast_u642d
:
2064 case ir_unop_bitcast_d2u64
:
2065 case ir_unop_subroutine_to_int
:
2067 result
= nir_mov(&b
, srcs
[0]);
2069 case ir_unop_trunc
: result
= nir_ftrunc(&b
, srcs
[0]); break;
2070 case ir_unop_ceil
: result
= nir_fceil(&b
, srcs
[0]); break;
2071 case ir_unop_floor
: result
= nir_ffloor(&b
, srcs
[0]); break;
2072 case ir_unop_fract
: result
= nir_ffract(&b
, srcs
[0]); break;
2073 case ir_unop_frexp_exp
: result
= nir_frexp_exp(&b
, srcs
[0]); break;
2074 case ir_unop_frexp_sig
: result
= nir_frexp_sig(&b
, srcs
[0]); break;
2075 case ir_unop_round_even
: result
= nir_fround_even(&b
, srcs
[0]); break;
2076 case ir_unop_sin
: result
= nir_fsin(&b
, srcs
[0]); break;
2077 case ir_unop_cos
: result
= nir_fcos(&b
, srcs
[0]); break;
2078 case ir_unop_dFdx
: result
= nir_fddx(&b
, srcs
[0]); break;
2079 case ir_unop_dFdy
: result
= nir_fddy(&b
, srcs
[0]); break;
2080 case ir_unop_dFdx_fine
: result
= nir_fddx_fine(&b
, srcs
[0]); break;
2081 case ir_unop_dFdy_fine
: result
= nir_fddy_fine(&b
, srcs
[0]); break;
2082 case ir_unop_dFdx_coarse
: result
= nir_fddx_coarse(&b
, srcs
[0]); break;
2083 case ir_unop_dFdy_coarse
: result
= nir_fddy_coarse(&b
, srcs
[0]); break;
2084 case ir_unop_pack_snorm_2x16
:
2085 result
= nir_pack_snorm_2x16(&b
, srcs
[0]);
2087 case ir_unop_pack_snorm_4x8
:
2088 result
= nir_pack_snorm_4x8(&b
, srcs
[0]);
2090 case ir_unop_pack_unorm_2x16
:
2091 result
= nir_pack_unorm_2x16(&b
, srcs
[0]);
2093 case ir_unop_pack_unorm_4x8
:
2094 result
= nir_pack_unorm_4x8(&b
, srcs
[0]);
2096 case ir_unop_pack_half_2x16
:
2097 result
= nir_pack_half_2x16(&b
, srcs
[0]);
2099 case ir_unop_unpack_snorm_2x16
:
2100 result
= nir_unpack_snorm_2x16(&b
, srcs
[0]);
2102 case ir_unop_unpack_snorm_4x8
:
2103 result
= nir_unpack_snorm_4x8(&b
, srcs
[0]);
2105 case ir_unop_unpack_unorm_2x16
:
2106 result
= nir_unpack_unorm_2x16(&b
, srcs
[0]);
2108 case ir_unop_unpack_unorm_4x8
:
2109 result
= nir_unpack_unorm_4x8(&b
, srcs
[0]);
2111 case ir_unop_unpack_half_2x16
:
2112 result
= nir_unpack_half_2x16(&b
, srcs
[0]);
2114 case ir_unop_pack_sampler_2x32
:
2115 case ir_unop_pack_image_2x32
:
2116 case ir_unop_pack_double_2x32
:
2117 case ir_unop_pack_int_2x32
:
2118 case ir_unop_pack_uint_2x32
:
2119 result
= nir_pack_64_2x32(&b
, srcs
[0]);
2121 case ir_unop_unpack_sampler_2x32
:
2122 case ir_unop_unpack_image_2x32
:
2123 case ir_unop_unpack_double_2x32
:
2124 case ir_unop_unpack_int_2x32
:
2125 case ir_unop_unpack_uint_2x32
:
2126 result
= nir_unpack_64_2x32(&b
, srcs
[0]);
2128 case ir_unop_bitfield_reverse
:
2129 result
= nir_bitfield_reverse(&b
, srcs
[0]);
2131 case ir_unop_bit_count
:
2132 result
= nir_bit_count(&b
, srcs
[0]);
2134 case ir_unop_find_msb
:
2136 case GLSL_TYPE_UINT
:
2137 result
= nir_ufind_msb(&b
, srcs
[0]);
2140 result
= nir_ifind_msb(&b
, srcs
[0]);
2143 unreachable("Invalid type for findMSB()");
2146 case ir_unop_find_lsb
:
2147 result
= nir_find_lsb(&b
, srcs
[0]);
2151 switch (ir
->type
->vector_elements
) {
2153 switch (ir
->operands
[0]->type
->vector_elements
) {
2154 case 1: result
= nir_fnoise1_1(&b
, srcs
[0]); break;
2155 case 2: result
= nir_fnoise1_2(&b
, srcs
[0]); break;
2156 case 3: result
= nir_fnoise1_3(&b
, srcs
[0]); break;
2157 case 4: result
= nir_fnoise1_4(&b
, srcs
[0]); break;
2158 default: unreachable("not reached");
2162 switch (ir
->operands
[0]->type
->vector_elements
) {
2163 case 1: result
= nir_fnoise2_1(&b
, srcs
[0]); break;
2164 case 2: result
= nir_fnoise2_2(&b
, srcs
[0]); break;
2165 case 3: result
= nir_fnoise2_3(&b
, srcs
[0]); break;
2166 case 4: result
= nir_fnoise2_4(&b
, srcs
[0]); break;
2167 default: unreachable("not reached");
2171 switch (ir
->operands
[0]->type
->vector_elements
) {
2172 case 1: result
= nir_fnoise3_1(&b
, srcs
[0]); break;
2173 case 2: result
= nir_fnoise3_2(&b
, srcs
[0]); break;
2174 case 3: result
= nir_fnoise3_3(&b
, srcs
[0]); break;
2175 case 4: result
= nir_fnoise3_4(&b
, srcs
[0]); break;
2176 default: unreachable("not reached");
2180 switch (ir
->operands
[0]->type
->vector_elements
) {
2181 case 1: result
= nir_fnoise4_1(&b
, srcs
[0]); break;
2182 case 2: result
= nir_fnoise4_2(&b
, srcs
[0]); break;
2183 case 3: result
= nir_fnoise4_3(&b
, srcs
[0]); break;
2184 case 4: result
= nir_fnoise4_4(&b
, srcs
[0]); break;
2185 default: unreachable("not reached");
2189 unreachable("not reached");
2192 case ir_unop_get_buffer_size
: {
2193 nir_intrinsic_instr
*load
= nir_intrinsic_instr_create(
2195 nir_intrinsic_get_buffer_size
);
2196 load
->num_components
= ir
->type
->vector_elements
;
2197 load
->src
[0] = nir_src_for_ssa(evaluate_rvalue(ir
->operands
[0]));
2198 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2199 add_instr(&load
->instr
, ir
->type
->vector_elements
, bit_size
);
2204 result
= nir_atan(&b
, srcs
[0]);
2208 result
= type_is_float(out_type
) ? nir_fadd(&b
, srcs
[0], srcs
[1])
2209 : nir_iadd(&b
, srcs
[0], srcs
[1]);
2212 result
= type_is_float(out_type
) ? nir_fsub(&b
, srcs
[0], srcs
[1])
2213 : nir_isub(&b
, srcs
[0], srcs
[1]);
2216 if (type_is_float(out_type
))
2217 result
= nir_fmul(&b
, srcs
[0], srcs
[1]);
2218 else if (out_type
== GLSL_TYPE_INT64
&&
2219 (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_INT
||
2220 ir
->operands
[1]->type
->base_type
== GLSL_TYPE_INT
))
2221 result
= nir_imul_2x32_64(&b
, srcs
[0], srcs
[1]);
2222 else if (out_type
== GLSL_TYPE_UINT64
&&
2223 (ir
->operands
[0]->type
->base_type
== GLSL_TYPE_UINT
||
2224 ir
->operands
[1]->type
->base_type
== GLSL_TYPE_UINT
))
2225 result
= nir_umul_2x32_64(&b
, srcs
[0], srcs
[1]);
2227 result
= nir_imul(&b
, srcs
[0], srcs
[1]);
2230 if (type_is_float(out_type
))
2231 result
= nir_fdiv(&b
, srcs
[0], srcs
[1]);
2232 else if (type_is_signed(out_type
))
2233 result
= nir_idiv(&b
, srcs
[0], srcs
[1]);
2235 result
= nir_udiv(&b
, srcs
[0], srcs
[1]);
2238 result
= type_is_float(out_type
) ? nir_fmod(&b
, srcs
[0], srcs
[1])
2239 : nir_umod(&b
, srcs
[0], srcs
[1]);
2242 if (type_is_float(out_type
))
2243 result
= nir_fmin(&b
, srcs
[0], srcs
[1]);
2244 else if (type_is_signed(out_type
))
2245 result
= nir_imin(&b
, srcs
[0], srcs
[1]);
2247 result
= nir_umin(&b
, srcs
[0], srcs
[1]);
2250 if (type_is_float(out_type
))
2251 result
= nir_fmax(&b
, srcs
[0], srcs
[1]);
2252 else if (type_is_signed(out_type
))
2253 result
= nir_imax(&b
, srcs
[0], srcs
[1]);
2255 result
= nir_umax(&b
, srcs
[0], srcs
[1]);
2257 case ir_binop_pow
: result
= nir_fpow(&b
, srcs
[0], srcs
[1]); break;
2258 case ir_binop_bit_and
: result
= nir_iand(&b
, srcs
[0], srcs
[1]); break;
2259 case ir_binop_bit_or
: result
= nir_ior(&b
, srcs
[0], srcs
[1]); break;
2260 case ir_binop_bit_xor
: result
= nir_ixor(&b
, srcs
[0], srcs
[1]); break;
2261 case ir_binop_logic_and
:
2262 result
= nir_iand(&b
, srcs
[0], srcs
[1]);
2264 case ir_binop_logic_or
:
2265 result
= nir_ior(&b
, srcs
[0], srcs
[1]);
2267 case ir_binop_logic_xor
:
2268 result
= nir_ixor(&b
, srcs
[0], srcs
[1]);
2270 case ir_binop_lshift
: result
= nir_ishl(&b
, srcs
[0], srcs
[1]); break;
2271 case ir_binop_rshift
:
2272 result
= (type_is_signed(out_type
)) ? nir_ishr(&b
, srcs
[0], srcs
[1])
2273 : nir_ushr(&b
, srcs
[0], srcs
[1]);
2275 case ir_binop_imul_high
:
2276 result
= (out_type
== GLSL_TYPE_INT
) ? nir_imul_high(&b
, srcs
[0], srcs
[1])
2277 : nir_umul_high(&b
, srcs
[0], srcs
[1]);
2279 case ir_binop_carry
: result
= nir_uadd_carry(&b
, srcs
[0], srcs
[1]); break;
2280 case ir_binop_borrow
: result
= nir_usub_borrow(&b
, srcs
[0], srcs
[1]); break;
2282 if (type_is_float(types
[0]))
2283 result
= nir_flt(&b
, srcs
[0], srcs
[1]);
2284 else if (type_is_signed(types
[0]))
2285 result
= nir_ilt(&b
, srcs
[0], srcs
[1]);
2287 result
= nir_ult(&b
, srcs
[0], srcs
[1]);
2289 case ir_binop_gequal
:
2290 if (type_is_float(types
[0]))
2291 result
= nir_fge(&b
, srcs
[0], srcs
[1]);
2292 else if (type_is_signed(types
[0]))
2293 result
= nir_ige(&b
, srcs
[0], srcs
[1]);
2295 result
= nir_uge(&b
, srcs
[0], srcs
[1]);
2297 case ir_binop_equal
:
2298 if (type_is_float(types
[0]))
2299 result
= nir_feq(&b
, srcs
[0], srcs
[1]);
2301 result
= nir_ieq(&b
, srcs
[0], srcs
[1]);
2303 case ir_binop_nequal
:
2304 if (type_is_float(types
[0]))
2305 result
= nir_fne(&b
, srcs
[0], srcs
[1]);
2307 result
= nir_ine(&b
, srcs
[0], srcs
[1]);
2309 case ir_binop_all_equal
:
2310 if (type_is_float(types
[0])) {
2311 switch (ir
->operands
[0]->type
->vector_elements
) {
2312 case 1: result
= nir_feq(&b
, srcs
[0], srcs
[1]); break;
2313 case 2: result
= nir_ball_fequal2(&b
, srcs
[0], srcs
[1]); break;
2314 case 3: result
= nir_ball_fequal3(&b
, srcs
[0], srcs
[1]); break;
2315 case 4: result
= nir_ball_fequal4(&b
, srcs
[0], srcs
[1]); break;
2317 unreachable("not reached");
2320 switch (ir
->operands
[0]->type
->vector_elements
) {
2321 case 1: result
= nir_ieq(&b
, srcs
[0], srcs
[1]); break;
2322 case 2: result
= nir_ball_iequal2(&b
, srcs
[0], srcs
[1]); break;
2323 case 3: result
= nir_ball_iequal3(&b
, srcs
[0], srcs
[1]); break;
2324 case 4: result
= nir_ball_iequal4(&b
, srcs
[0], srcs
[1]); break;
2326 unreachable("not reached");
2330 case ir_binop_any_nequal
:
2331 if (type_is_float(types
[0])) {
2332 switch (ir
->operands
[0]->type
->vector_elements
) {
2333 case 1: result
= nir_fne(&b
, srcs
[0], srcs
[1]); break;
2334 case 2: result
= nir_bany_fnequal2(&b
, srcs
[0], srcs
[1]); break;
2335 case 3: result
= nir_bany_fnequal3(&b
, srcs
[0], srcs
[1]); break;
2336 case 4: result
= nir_bany_fnequal4(&b
, srcs
[0], srcs
[1]); break;
2338 unreachable("not reached");
2341 switch (ir
->operands
[0]->type
->vector_elements
) {
2342 case 1: result
= nir_ine(&b
, srcs
[0], srcs
[1]); break;
2343 case 2: result
= nir_bany_inequal2(&b
, srcs
[0], srcs
[1]); break;
2344 case 3: result
= nir_bany_inequal3(&b
, srcs
[0], srcs
[1]); break;
2345 case 4: result
= nir_bany_inequal4(&b
, srcs
[0], srcs
[1]); break;
2347 unreachable("not reached");
2352 switch (ir
->operands
[0]->type
->vector_elements
) {
2353 case 2: result
= nir_fdot2(&b
, srcs
[0], srcs
[1]); break;
2354 case 3: result
= nir_fdot3(&b
, srcs
[0], srcs
[1]); break;
2355 case 4: result
= nir_fdot4(&b
, srcs
[0], srcs
[1]); break;
2357 unreachable("not reached");
2360 case ir_binop_vector_extract
: {
2361 result
= nir_channel(&b
, srcs
[0], 0);
2362 for (unsigned i
= 1; i
< ir
->operands
[0]->type
->vector_elements
; i
++) {
2363 nir_ssa_def
*swizzled
= nir_channel(&b
, srcs
[0], i
);
2364 result
= nir_bcsel(&b
, nir_ieq(&b
, srcs
[1], nir_imm_int(&b
, i
)),
2370 case ir_binop_atan2
:
2371 result
= nir_atan2(&b
, srcs
[0], srcs
[1]);
2374 case ir_binop_ldexp
: result
= nir_ldexp(&b
, srcs
[0], srcs
[1]); break;
2376 result
= nir_ffma(&b
, srcs
[0], srcs
[1], srcs
[2]);
2379 result
= nir_flrp(&b
, srcs
[0], srcs
[1], srcs
[2]);
2382 result
= nir_bcsel(&b
, srcs
[0], srcs
[1], srcs
[2]);
2384 case ir_triop_bitfield_extract
:
2385 result
= (out_type
== GLSL_TYPE_INT
) ?
2386 nir_ibitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]) :
2387 nir_ubitfield_extract(&b
, srcs
[0], srcs
[1], srcs
[2]);
2389 case ir_quadop_bitfield_insert
:
2390 result
= nir_bitfield_insert(&b
, srcs
[0], srcs
[1], srcs
[2], srcs
[3]);
2392 case ir_quadop_vector
:
2393 result
= nir_vec(&b
, srcs
, ir
->type
->vector_elements
);
2397 unreachable("not reached");
2402 nir_visitor::visit(ir_swizzle
*ir
)
2404 unsigned swizzle
[4] = { ir
->mask
.x
, ir
->mask
.y
, ir
->mask
.z
, ir
->mask
.w
};
2405 result
= nir_swizzle(&b
, evaluate_rvalue(ir
->val
), swizzle
,
2406 ir
->type
->vector_elements
);
2410 nir_visitor::visit(ir_texture
*ir
)
2417 num_srcs
= 1; /* coordinate */
2422 op
= (ir
->op
== ir_txb
) ? nir_texop_txb
: nir_texop_txl
;
2423 num_srcs
= 2; /* coordinate, bias/lod */
2427 op
= nir_texop_txd
; /* coordinate, dPdx, dPdy */
2433 if (ir
->lod_info
.lod
!= NULL
)
2434 num_srcs
= 2; /* coordinate, lod */
2436 num_srcs
= 1; /* coordinate */
2440 op
= nir_texop_txf_ms
;
2441 num_srcs
= 2; /* coordinate, sample_index */
2446 if (ir
->lod_info
.lod
!= NULL
)
2447 num_srcs
= 1; /* lod */
2454 num_srcs
= 1; /* coordinate */
2459 num_srcs
= 1; /* coordinate */
2462 case ir_query_levels
:
2463 op
= nir_texop_query_levels
;
2467 case ir_texture_samples
:
2468 op
= nir_texop_texture_samples
;
2472 case ir_samples_identical
:
2473 op
= nir_texop_samples_identical
;
2474 num_srcs
= 1; /* coordinate */
2478 unreachable("not reached");
2481 if (ir
->projector
!= NULL
)
2483 if (ir
->shadow_comparator
!= NULL
)
2485 /* offsets are constants we store inside nir_tex_intrs.offsets */
2486 if (ir
->offset
!= NULL
&& !ir
->offset
->type
->is_array())
2489 /* Add one for the texture deref */
2492 nir_tex_instr
*instr
= nir_tex_instr_create(this->shader
, num_srcs
);
2495 instr
->sampler_dim
=
2496 (glsl_sampler_dim
) ir
->sampler
->type
->sampler_dimensionality
;
2497 instr
->is_array
= ir
->sampler
->type
->sampler_array
;
2498 instr
->is_shadow
= ir
->sampler
->type
->sampler_shadow
;
2499 if (instr
->is_shadow
)
2500 instr
->is_new_style_shadow
= (ir
->type
->vector_elements
== 1);
2501 switch (ir
->type
->base_type
) {
2502 case GLSL_TYPE_FLOAT
:
2503 instr
->dest_type
= nir_type_float
;
2506 instr
->dest_type
= nir_type_int
;
2508 case GLSL_TYPE_BOOL
:
2509 case GLSL_TYPE_UINT
:
2510 instr
->dest_type
= nir_type_uint
;
2513 unreachable("not reached");
2516 nir_deref_instr
*sampler_deref
= evaluate_deref(ir
->sampler
);
2518 /* check for bindless handles */
2519 if (sampler_deref
->mode
!= nir_var_uniform
||
2520 nir_deref_instr_get_variable(sampler_deref
)->data
.bindless
) {
2521 nir_ssa_def
*load
= nir_load_deref(&b
, sampler_deref
);
2522 instr
->src
[0].src
= nir_src_for_ssa(load
);
2523 instr
->src
[0].src_type
= nir_tex_src_texture_handle
;
2524 instr
->src
[1].src
= nir_src_for_ssa(load
);
2525 instr
->src
[1].src_type
= nir_tex_src_sampler_handle
;
2527 instr
->src
[0].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2528 instr
->src
[0].src_type
= nir_tex_src_texture_deref
;
2529 instr
->src
[1].src
= nir_src_for_ssa(&sampler_deref
->dest
.ssa
);
2530 instr
->src
[1].src_type
= nir_tex_src_sampler_deref
;
2533 unsigned src_number
= 2;
2535 if (ir
->coordinate
!= NULL
) {
2536 instr
->coord_components
= ir
->coordinate
->type
->vector_elements
;
2537 instr
->src
[src_number
].src
=
2538 nir_src_for_ssa(evaluate_rvalue(ir
->coordinate
));
2539 instr
->src
[src_number
].src_type
= nir_tex_src_coord
;
2543 if (ir
->projector
!= NULL
) {
2544 instr
->src
[src_number
].src
=
2545 nir_src_for_ssa(evaluate_rvalue(ir
->projector
));
2546 instr
->src
[src_number
].src_type
= nir_tex_src_projector
;
2550 if (ir
->shadow_comparator
!= NULL
) {
2551 instr
->src
[src_number
].src
=
2552 nir_src_for_ssa(evaluate_rvalue(ir
->shadow_comparator
));
2553 instr
->src
[src_number
].src_type
= nir_tex_src_comparator
;
2557 if (ir
->offset
!= NULL
) {
2558 if (ir
->offset
->type
->is_array()) {
2559 for (int i
= 0; i
< ir
->offset
->type
->array_size(); i
++) {
2560 const ir_constant
*c
=
2561 ir
->offset
->as_constant()->get_array_element(i
);
2563 for (unsigned j
= 0; j
< 2; ++j
) {
2564 int val
= c
->get_int_component(j
);
2565 assert(val
<= 31 && val
>= -32);
2566 instr
->tg4_offsets
[i
][j
] = val
;
2570 assert(ir
->offset
->type
->is_vector() || ir
->offset
->type
->is_scalar());
2572 instr
->src
[src_number
].src
=
2573 nir_src_for_ssa(evaluate_rvalue(ir
->offset
));
2574 instr
->src
[src_number
].src_type
= nir_tex_src_offset
;
2581 instr
->src
[src_number
].src
=
2582 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.bias
));
2583 instr
->src
[src_number
].src_type
= nir_tex_src_bias
;
2590 if (ir
->lod_info
.lod
!= NULL
) {
2591 instr
->src
[src_number
].src
=
2592 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.lod
));
2593 instr
->src
[src_number
].src_type
= nir_tex_src_lod
;
2599 instr
->src
[src_number
].src
=
2600 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdx
));
2601 instr
->src
[src_number
].src_type
= nir_tex_src_ddx
;
2603 instr
->src
[src_number
].src
=
2604 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.grad
.dPdy
));
2605 instr
->src
[src_number
].src_type
= nir_tex_src_ddy
;
2610 instr
->src
[src_number
].src
=
2611 nir_src_for_ssa(evaluate_rvalue(ir
->lod_info
.sample_index
));
2612 instr
->src
[src_number
].src_type
= nir_tex_src_ms_index
;
2617 instr
->component
= ir
->lod_info
.component
->as_constant()->value
.u
[0];
2624 assert(src_number
== num_srcs
);
2626 unsigned bit_size
= glsl_get_bit_size(ir
->type
);
2627 add_instr(&instr
->instr
, nir_tex_instr_dest_size(instr
), bit_size
);
2631 nir_visitor::visit(ir_constant
*ir
)
2634 * We don't know if this variable is an array or struct that gets
2635 * dereferenced, so do the safe thing an make it a variable with a
2636 * constant initializer and return a dereference.
2640 nir_local_variable_create(this->impl
, ir
->type
, "const_temp");
2641 var
->data
.read_only
= true;
2642 var
->constant_initializer
= constant_copy(ir
, var
);
2644 this->deref
= nir_build_deref_var(&b
, var
);
2648 nir_visitor::visit(ir_dereference_variable
*ir
)
2650 if (ir
->variable_referenced()->data
.mode
== ir_var_function_out
) {
2651 unsigned i
= (sig
->return_type
!= glsl_type::void_type
) ? 1 : 0;
2653 foreach_in_list(ir_variable
, param
, &sig
->parameters
) {
2654 if (param
== ir
->variable_referenced()) {
2660 this->deref
= nir_build_deref_cast(&b
, nir_load_param(&b
, i
),
2661 nir_var_function_temp
, ir
->type
, 0);
2665 assert(ir
->variable_referenced()->data
.mode
!= ir_var_function_inout
);
2667 struct hash_entry
*entry
=
2668 _mesa_hash_table_search(this->var_table
, ir
->var
);
2670 nir_variable
*var
= (nir_variable
*) entry
->data
;
2672 this->deref
= nir_build_deref_var(&b
, var
);
2676 nir_visitor::visit(ir_dereference_record
*ir
)
2678 ir
->record
->accept(this);
2680 int field_index
= ir
->field_idx
;
2681 assert(field_index
>= 0);
2683 this->deref
= nir_build_deref_struct(&b
, this->deref
, field_index
);
2687 nir_visitor::visit(ir_dereference_array
*ir
)
2689 nir_ssa_def
*index
= evaluate_rvalue(ir
->array_index
);
2691 ir
->array
->accept(this);
2693 this->deref
= nir_build_deref_array(&b
, this->deref
, index
);
2697 nir_visitor::visit(ir_barrier
*)
2699 nir_intrinsic_instr
*instr
=
2700 nir_intrinsic_instr_create(this->shader
, nir_intrinsic_barrier
);
2701 nir_builder_instr_insert(&b
, &instr
->instr
);
2705 glsl_float64_funcs_to_nir(struct gl_context
*ctx
,
2706 const nir_shader_compiler_options
*options
)
2708 /* We pretend it's a vertex shader. Ultimately, the stage shouldn't
2709 * matter because we're not optimizing anything here.
2711 struct gl_shader
*sh
= _mesa_new_shader(-1, MESA_SHADER_VERTEX
);
2712 sh
->Source
= float64_source
;
2713 sh
->CompileStatus
= COMPILE_FAILURE
;
2714 _mesa_glsl_compile_shader(ctx
, sh
, false, false, true);
2716 if (!sh
->CompileStatus
) {
2719 "fp64 software impl compile failed:\n%s\nsource:\n%s\n",
2720 sh
->InfoLog
, float64_source
);
2725 nir_shader
*nir
= nir_shader_create(NULL
, MESA_SHADER_VERTEX
, options
, NULL
);
2727 nir_visitor
v1(ctx
, nir
);
2728 nir_function_visitor
v2(&v1
);
2730 visit_exec_list(sh
->ir
, &v1
);
2732 /* _mesa_delete_shader will try to free sh->Source but it's static const */
2734 _mesa_delete_shader(ctx
, sh
);
2736 nir_validate_shader(nir
, "float64_funcs_to_nir");
2738 NIR_PASS_V(nir
, nir_lower_constant_initializers
, nir_var_function_temp
);
2739 NIR_PASS_V(nir
, nir_lower_returns
);
2740 NIR_PASS_V(nir
, nir_inline_functions
);
2741 NIR_PASS_V(nir
, nir_opt_deref
);
2743 /* Do some optimizations to clean up the shader now. By optimizing the
2744 * functions in the library, we avoid having to re-do that work every
2745 * time we inline a copy of a function. Reducing basic blocks also helps
2746 * with compile times.
2748 NIR_PASS_V(nir
, nir_lower_vars_to_ssa
);
2749 NIR_PASS_V(nir
, nir_copy_prop
);
2750 NIR_PASS_V(nir
, nir_opt_dce
);
2751 NIR_PASS_V(nir
, nir_opt_cse
);
2752 NIR_PASS_V(nir
, nir_opt_gcm
, true);
2753 NIR_PASS_V(nir
, nir_opt_peephole_select
, 1, false, false);
2754 NIR_PASS_V(nir
, nir_opt_dce
);