2 * Copyright © 2014 Connor Abbott
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)
30 #include "util/hash_table.h"
32 #include "GL/gl.h" /* GLenum */
33 #include "util/ralloc.h"
35 #include "util/bitset.h"
36 #include "nir_types.h"
37 #include "glsl/shader_enums.h"
40 #include "nir_opcodes.h"
47 struct gl_shader_program
;
50 #define NIR_TRUE (~0u)
52 /** Defines a cast function
54 * This macro defines a cast function from in_type to out_type where
55 * out_type is some structure type that contains a field of type out_type.
57 * Note that you have to be a bit careful as the generated cast function
60 #define NIR_DEFINE_CAST(name, in_type, out_type, field) \
61 static inline out_type * \
62 name(const in_type *parent) \
64 return exec_node_data(out_type, parent, field); \
67 struct nir_function_overload
;
74 * Description of built-in state associated with a uniform
76 * \sa nir_variable::state_slots
93 * Data stored in an nir_constant
95 union nir_constant_data
{
102 typedef struct nir_constant
{
104 * Value of the constant.
106 * The field used to back the values supplied by the constant is determined
107 * by the type associated with the \c nir_variable. Constants may be
108 * scalars, vectors, or matrices.
110 union nir_constant_data value
;
112 /* Array elements / Structure Fields */
113 struct nir_constant
**elements
;
117 * \brief Layout qualifiers for gl_FragDepth.
119 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
120 * with a layout qualifier.
123 nir_depth_layout_none
, /**< No depth layout is specified. */
124 nir_depth_layout_any
,
125 nir_depth_layout_greater
,
126 nir_depth_layout_less
,
127 nir_depth_layout_unchanged
131 * Either a uniform, global variable, shader input, or shader output. Based on
132 * ir_variable - it should be easy to translate between the two.
136 struct exec_node node
;
139 * Declared type of the variable
141 const struct glsl_type
*type
;
144 * Declared name of the variable
149 * For variables which satisfy the is_interface_instance() predicate, this
150 * points to an array of integers such that if the ith member of the
151 * interface block is an array, max_ifc_array_access[i] is the maximum
152 * array element of that member that has been accessed. If the ith member
153 * of the interface block is not an array, max_ifc_array_access[i] is
156 * For variables whose type is not an interface block, this pointer is
159 unsigned *max_ifc_array_access
;
161 struct nir_variable_data
{
164 * Is the variable read-only?
166 * This is set for variables declared as \c const, shader inputs,
169 unsigned read_only
:1;
172 unsigned invariant
:1;
175 * Storage class of the variable.
177 * \sa nir_variable_mode
179 nir_variable_mode mode
:4;
182 * Interpolation mode for shader inputs / outputs
184 * \sa glsl_interp_qualifier
186 unsigned interpolation
:2;
189 * \name ARB_fragment_coord_conventions
192 unsigned origin_upper_left
:1;
193 unsigned pixel_center_integer
:1;
197 * Was the location explicitly set in the shader?
199 * If the location is explicitly set in the shader, it \b cannot be changed
200 * by the linker or by the API (e.g., calls to \c glBindAttribLocation have
203 unsigned explicit_location
:1;
204 unsigned explicit_index
:1;
207 * Was an initial binding explicitly set in the shader?
209 * If so, constant_initializer contains an integer nir_constant
210 * representing the initial binding point.
212 unsigned explicit_binding
:1;
215 * Does this variable have an initializer?
217 * This is used by the linker to cross-validiate initializers of global
220 unsigned has_initializer
:1;
223 * Is this variable a generic output or input that has not yet been matched
224 * up to a variable in another stage of the pipeline?
226 * This is used by the linker as scratch storage while assigning locations
227 * to generic inputs and outputs.
229 unsigned is_unmatched_generic_inout
:1;
232 * If non-zero, then this variable may be packed along with other variables
233 * into a single varying slot, so this offset should be applied when
234 * accessing components. For example, an offset of 1 means that the x
235 * component of this variable is actually stored in component y of the
236 * location specified by \c location.
238 unsigned location_frac
:2;
241 * Non-zero if this variable was created by lowering a named interface
242 * block which was not an array.
244 * Note that this variable and \c from_named_ifc_block_array will never
247 unsigned from_named_ifc_block_nonarray
:1;
250 * Non-zero if this variable was created by lowering a named interface
251 * block which was an array.
253 * Note that this variable and \c from_named_ifc_block_nonarray will never
256 unsigned from_named_ifc_block_array
:1;
259 * \brief Layout qualifier for gl_FragDepth.
261 * This is not equal to \c ir_depth_layout_none if and only if this
262 * variable is \c gl_FragDepth and a layout qualifier is specified.
264 nir_depth_layout depth_layout
;
267 * Storage location of the base of this variable
269 * The precise meaning of this field depends on the nature of the variable.
271 * - Vertex shader input: one of the values from \c gl_vert_attrib.
272 * - Vertex shader output: one of the values from \c gl_varying_slot.
273 * - Geometry shader input: one of the values from \c gl_varying_slot.
274 * - Geometry shader output: one of the values from \c gl_varying_slot.
275 * - Fragment shader input: one of the values from \c gl_varying_slot.
276 * - Fragment shader output: one of the values from \c gl_frag_result.
277 * - Uniforms: Per-stage uniform slot number for default uniform block.
278 * - Uniforms: Index within the uniform block definition for UBO members.
279 * - Other: This field is not currently used.
281 * If the variable is a uniform, shader input, or shader output, and the
282 * slot has not been assigned, the value will be -1.
287 * The actual location of the variable in the IR. Only valid for inputs
290 unsigned int driver_location
;
293 * output index for dual source blending.
298 * Initial binding point for a sampler or UBO.
300 * For array types, this represents the binding point for the first element.
305 * Location an atomic counter is stored at.
308 unsigned buffer_index
;
313 * ARB_shader_image_load_store qualifiers.
316 bool read_only
; /**< "readonly" qualifier. */
317 bool write_only
; /**< "writeonly" qualifier. */
322 /** Image internal format if specified explicitly, otherwise GL_NONE. */
327 * Highest element accessed with a constant expression array index
329 * Not used for non-array variables.
331 unsigned max_array_access
;
336 * Built-in state that backs this uniform
338 * Once set at variable creation, \c state_slots must remain invariant.
339 * This is because, ideally, this array would be shared by all clones of
340 * this variable in the IR tree. In other words, we'd really like for it
341 * to be a fly-weight.
343 * If the variable is not a uniform, \c num_state_slots will be zero and
344 * \c state_slots will be \c NULL.
347 unsigned num_state_slots
; /**< Number of state slots used */
348 nir_state_slot
*state_slots
; /**< State descriptors. */
352 * Constant expression assigned in the initializer of the variable
354 nir_constant
*constant_initializer
;
357 * For variables that are in an interface block or are an instance of an
358 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
360 * \sa ir_variable::location
362 const struct glsl_type
*interface_type
;
366 struct exec_node node
;
368 unsigned num_components
; /** < number of vector components */
369 unsigned num_array_elems
; /** < size of array (0 for no array) */
371 /** generic register index. */
374 /** only for debug purposes, can be NULL */
377 /** whether this register is local (per-function) or global (per-shader) */
381 * If this flag is set to true, then accessing channels >= num_components
382 * is well-defined, and simply spills over to the next array element. This
383 * is useful for backends that can do per-component accessing, in
384 * particular scalar backends. By setting this flag and making
385 * num_components equal to 1, structures can be packed tightly into
386 * registers and then registers can be accessed per-component to get to
387 * each structure member, even if it crosses vec4 boundaries.
392 * If this pointer is non-NULL then this register has exactly one
393 * definition and that definition dominates all of its uses. This is
394 * set by the out-of-SSA pass so that backends can get SSA-like
395 * information even once they have gone out of SSA.
397 struct nir_instr
*parent_instr
;
399 /** set of nir_instr's where this register is used (read from) */
402 /** set of nir_instr's where this register is defined (written to) */
405 /** set of nir_if's where this register is used as a condition */
413 nir_instr_type_intrinsic
,
414 nir_instr_type_load_const
,
416 nir_instr_type_ssa_undef
,
418 nir_instr_type_parallel_copy
,
421 typedef struct nir_instr
{
422 struct exec_node node
;
424 struct nir_block
*block
;
426 /* A temporary for optimization and analysis passes to use for storing
427 * flags. For instance, DCE uses this to store the "dead/live" info.
432 static inline nir_instr
*
433 nir_instr_next(nir_instr
*instr
)
435 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
436 if (exec_node_is_tail_sentinel(next
))
439 return exec_node_data(nir_instr
, next
, node
);
442 static inline nir_instr
*
443 nir_instr_prev(nir_instr
*instr
)
445 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
446 if (exec_node_is_head_sentinel(prev
))
449 return exec_node_data(nir_instr
, prev
, node
);
453 /** for debugging only, can be NULL */
456 /** generic SSA definition index. */
459 /** Index into the live_in and live_out bitfields */
462 nir_instr
*parent_instr
;
464 /** set of nir_instr's where this register is used (read from) */
467 /** set of nir_if's where this register is used as a condition */
470 uint8_t num_components
;
477 struct nir_src
*indirect
; /** < NULL for no indirect offset */
478 unsigned base_offset
;
480 /* TODO use-def chain goes here */
485 struct nir_src
*indirect
; /** < NULL for no indirect offset */
486 unsigned base_offset
;
488 /* TODO def-use chain goes here */
491 typedef struct nir_src
{
509 static inline nir_src
510 nir_src_for_ssa(nir_ssa_def
*def
)
520 static inline nir_src
521 nir_src_for_reg(nir_register
*reg
)
527 src
.reg
.indirect
= NULL
;
528 src
.reg
.base_offset
= 0;
533 static inline nir_instr
*
534 nir_src_get_parent_instr(const nir_src
*src
)
537 return src
->ssa
->parent_instr
;
539 return src
->reg
.reg
->parent_instr
;
543 static inline nir_dest
544 nir_dest_for_reg(nir_register
*reg
)
550 dest
.reg
.indirect
= NULL
;
551 dest
.reg
.base_offset
= 0;
556 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *mem_ctx
);
557 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, void *mem_ctx
);
563 * \name input modifiers
567 * For inputs interpreted as floating point, flips the sign bit. For
568 * inputs interpreted as integers, performs the two's complement negation.
573 * Clears the sign bit for floating point values, and computes the integer
574 * absolute value for integers. Note that the negate modifier acts after
575 * the absolute value modifier, therefore if both are set then all inputs
576 * will become negative.
582 * For each input component, says which component of the register it is
583 * chosen from. Note that which elements of the swizzle are used and which
584 * are ignored are based on the write mask for most opcodes - for example,
585 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
586 * a swizzle of {2, x, 1, 0} where x means "don't care."
595 * \name saturate output modifier
597 * Only valid for opcodes that output floating-point numbers. Clamps the
598 * output to between 0.0 and 1.0 inclusive.
603 unsigned write_mask
: 4; /* ignored if dest.is_ssa is true */
606 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
, void *mem_ctx
);
607 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
611 nir_type_invalid
= 0, /* Not a valid type */
619 NIR_OP_IS_COMMUTATIVE
= (1 << 0),
620 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
621 } nir_op_algebraic_property
;
629 * The number of components in the output
631 * If non-zero, this is the size of the output and input sizes are
632 * explicitly given; swizzle and writemask are still in effect, but if
633 * the output component is masked out, then the input component may
636 * If zero, the opcode acts in the standard, per-component manner; the
637 * operation is performed on each component (except the ones that are
638 * masked out) with the input being taken from the input swizzle for
641 * The size of some of the inputs may be given (i.e. non-zero) even
642 * though output_size is zero; in that case, the inputs with a zero
643 * size act per-component, while the inputs with non-zero size don't.
645 unsigned output_size
;
648 * The type of vector that the instruction outputs. Note that the
649 * staurate modifier is only allowed on outputs with the float type.
652 nir_alu_type output_type
;
655 * The number of components in each input
657 unsigned input_sizes
[4];
660 * The type of vector that each input takes. Note that negate and
661 * absolute value are only allowed on inputs with int or float type and
662 * behave differently on the two.
664 nir_alu_type input_types
[4];
666 nir_op_algebraic_property algebraic_properties
;
669 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
671 typedef struct nir_alu_instr
{
678 /* is this source channel used? */
680 nir_alu_instr_channel_used(nir_alu_instr
*instr
, unsigned src
, unsigned channel
)
682 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
683 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
685 return (instr
->dest
.write_mask
>> channel
) & 1;
689 * For instructions whose destinations are SSA, get the number of channels
692 static inline unsigned
693 nir_ssa_alu_instr_src_components(nir_alu_instr
*instr
, unsigned src
)
695 assert(instr
->dest
.dest
.is_ssa
);
697 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
698 return nir_op_infos
[instr
->op
].input_sizes
[src
];
700 return instr
->dest
.dest
.ssa
.num_components
;
705 nir_deref_type_array
,
706 nir_deref_type_struct
709 typedef struct nir_deref
{
710 nir_deref_type deref_type
;
711 struct nir_deref
*child
;
712 const struct glsl_type
*type
;
721 /* This enum describes how the array is referenced. If the deref is
722 * direct then the base_offset is used. If the deref is indirect then then
723 * offset is given by base_offset + indirect. If the deref is a wildcard
724 * then the deref refers to all of the elements of the array at the same
725 * time. Wildcard dereferences are only ever allowed in copy_var
726 * intrinsics and the source and destination derefs must have matching
730 nir_deref_array_type_direct
,
731 nir_deref_array_type_indirect
,
732 nir_deref_array_type_wildcard
,
733 } nir_deref_array_type
;
738 nir_deref_array_type deref_array_type
;
739 unsigned base_offset
;
749 NIR_DEFINE_CAST(nir_deref_as_var
, nir_deref
, nir_deref_var
, deref
)
750 NIR_DEFINE_CAST(nir_deref_as_array
, nir_deref
, nir_deref_array
, deref
)
751 NIR_DEFINE_CAST(nir_deref_as_struct
, nir_deref
, nir_deref_struct
, deref
)
757 nir_deref_var
**params
;
758 nir_deref_var
*return_deref
;
760 struct nir_function_overload
*callee
;
763 #define INTRINSIC(name, num_srcs, src_components, has_dest, dest_components, \
764 num_variables, num_indices, flags) \
765 nir_intrinsic_##name,
767 #define LAST_INTRINSIC(name) nir_last_intrinsic = nir_intrinsic_##name,
770 #include "nir_intrinsics.h"
771 nir_num_intrinsics
= nir_last_intrinsic
+ 1
775 #undef LAST_INTRINSIC
777 /** Represents an intrinsic
779 * An intrinsic is an instruction type for handling things that are
780 * more-or-less regular operations but don't just consume and produce SSA
781 * values like ALU operations do. Intrinsics are not for things that have
782 * special semantic meaning such as phi nodes and parallel copies.
783 * Examples of intrinsics include variable load/store operations, system
784 * value loads, and the like. Even though texturing more-or-less falls
785 * under this category, texturing is its own instruction type because
786 * trying to represent texturing with intrinsics would lead to a
787 * combinatorial explosion of intrinsic opcodes.
789 * By having a single instruction type for handling a lot of different
790 * cases, optimization passes can look for intrinsics and, for the most
791 * part, completely ignore them. Each intrinsic type also has a few
792 * possible flags that govern whether or not they can be reordered or
793 * eliminated. That way passes like dead code elimination can still work
794 * on intrisics without understanding the meaning of each.
796 * Each intrinsic has some number of constant indices, some number of
797 * variables, and some number of sources. What these sources, variables,
798 * and indices mean depends on the intrinsic and is documented with the
799 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
800 * instructions are the only types of instruction that can operate on
806 nir_intrinsic_op intrinsic
;
810 /** number of components if this is a vectorized intrinsic
812 * Similarly to ALU operations, some intrinsics are vectorized.
813 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
814 * For vectorized intrinsics, the num_components field specifies the
815 * number of destination components and the number of source components
816 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
818 uint8_t num_components
;
822 nir_deref_var
*variables
[2];
825 } nir_intrinsic_instr
;
828 * \name NIR intrinsics semantic flags
830 * information about what the compiler can do with the intrinsics.
832 * \sa nir_intrinsic_info::flags
836 * whether the intrinsic can be safely eliminated if none of its output
837 * value is not being used.
839 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
842 * Whether the intrinsic can be reordered with respect to any other
843 * intrinsic, i.e. whether the only reordering dependencies of the
844 * intrinsic are due to the register reads/writes.
846 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
847 } nir_intrinsic_semantic_flag
;
849 #define NIR_INTRINSIC_MAX_INPUTS 4
854 unsigned num_srcs
; /** < number of register/SSA inputs */
856 /** number of components of each input register
858 * If this value is 0, the number of components is given by the
859 * num_components field of nir_intrinsic_instr.
861 unsigned src_components
[NIR_INTRINSIC_MAX_INPUTS
];
865 /** number of components of the output register
867 * If this value is 0, the number of components is given by the
868 * num_components field of nir_intrinsic_instr.
870 unsigned dest_components
;
872 /** the number of inputs/outputs that are variables */
873 unsigned num_variables
;
875 /** the number of constant indices used by the intrinsic */
876 unsigned num_indices
;
878 /** semantic flags for calls to this intrinsic */
879 nir_intrinsic_semantic_flag flags
;
880 } nir_intrinsic_info
;
882 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
885 * \group texture information
887 * This gives semantic information about textures which is useful to the
888 * frontend, the backend, and lowering passes, but not the optimizer.
893 nir_tex_src_projector
,
894 nir_tex_src_comparitor
, /* shadow comparitor */
898 nir_tex_src_ms_index
, /* MSAA sample index */
901 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
902 nir_num_tex_src_types
907 nir_tex_src_type src_type
;
911 nir_texop_tex
, /**< Regular texture look-up */
912 nir_texop_txb
, /**< Texture look-up with LOD bias */
913 nir_texop_txl
, /**< Texture look-up with explicit LOD */
914 nir_texop_txd
, /**< Texture look-up with partial derivatvies */
915 nir_texop_txf
, /**< Texel fetch with explicit LOD */
916 nir_texop_txf_ms
, /**< Multisample texture fetch */
917 nir_texop_txs
, /**< Texture size */
918 nir_texop_lod
, /**< Texture lod query */
919 nir_texop_tg4
, /**< Texture gather */
920 nir_texop_query_levels
/**< Texture levels query */
926 enum glsl_sampler_dim sampler_dim
;
927 nir_alu_type dest_type
;
932 unsigned num_srcs
, coord_components
;
933 bool is_array
, is_shadow
;
936 * If is_shadow is true, whether this is the old-style shadow that outputs 4
937 * components or the new-style shadow that outputs 1 component.
939 bool is_new_style_shadow
;
941 /* constant offset - must be 0 if the offset source is used */
944 /* gather component selector */
945 unsigned component
: 2;
947 /** The sampler index
949 * If this texture instruction has a nir_tex_src_sampler_offset source,
950 * then the sampler index is given by sampler_index + sampler_offset.
952 unsigned sampler_index
;
954 /** The size of the sampler array or 0 if it's not an array */
955 unsigned sampler_array_size
;
957 nir_deref_var
*sampler
; /* if this is NULL, use sampler_index instead */
960 static inline unsigned
961 nir_tex_instr_dest_size(nir_tex_instr
*instr
)
963 if (instr
->op
== nir_texop_txs
) {
965 switch (instr
->sampler_dim
) {
966 case GLSL_SAMPLER_DIM_1D
:
967 case GLSL_SAMPLER_DIM_BUF
:
970 case GLSL_SAMPLER_DIM_2D
:
971 case GLSL_SAMPLER_DIM_CUBE
:
972 case GLSL_SAMPLER_DIM_MS
:
973 case GLSL_SAMPLER_DIM_RECT
:
974 case GLSL_SAMPLER_DIM_EXTERNAL
:
977 case GLSL_SAMPLER_DIM_3D
:
981 unreachable("not reached");
988 if (instr
->op
== nir_texop_query_levels
)
991 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
997 static inline unsigned
998 nir_tex_instr_src_size(nir_tex_instr
*instr
, unsigned src
)
1000 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
1001 return instr
->coord_components
;
1004 if (instr
->src
[src
].src_type
== nir_tex_src_offset
||
1005 instr
->src
[src
].src_type
== nir_tex_src_ddx
||
1006 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
1007 if (instr
->is_array
)
1008 return instr
->coord_components
- 1;
1010 return instr
->coord_components
;
1017 nir_tex_instr_src_index(nir_tex_instr
*instr
, nir_tex_src_type type
)
1019 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
1020 if (instr
->src
[i
].src_type
== type
)
1037 nir_const_value value
;
1040 } nir_load_const_instr
;
1053 /* creates a new SSA variable in an undefined state */
1058 } nir_ssa_undef_instr
;
1061 struct exec_node node
;
1063 /* The predecessor block corresponding to this source */
1064 struct nir_block
*pred
;
1069 #define nir_foreach_phi_src(phi, entry) \
1070 foreach_list_typed(nir_phi_src, entry, node, &(phi)->srcs)
1075 struct exec_list srcs
; /** < list of nir_phi_src */
1081 struct exec_node node
;
1084 } nir_parallel_copy_entry
;
1086 #define nir_foreach_parallel_copy_entry(pcopy, entry) \
1087 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
1092 /* A list of nir_parallel_copy_entry's. The sources of all of the
1093 * entries are copied to the corresponding destinations "in parallel".
1094 * In other words, if we have two entries: a -> b and b -> a, the values
1097 struct exec_list entries
;
1098 } nir_parallel_copy_instr
;
1100 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
)
1101 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
)
1102 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
)
1103 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
)
1104 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
)
1105 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
)
1106 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
)
1107 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
)
1108 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
1109 nir_parallel_copy_instr
, instr
)
1114 * Control flow consists of a tree of control flow nodes, which include
1115 * if-statements and loops. The leaves of the tree are basic blocks, lists of
1116 * instructions that always run start-to-finish. Each basic block also keeps
1117 * track of its successors (blocks which may run immediately after the current
1118 * block) and predecessors (blocks which could have run immediately before the
1119 * current block). Each function also has a start block and an end block which
1120 * all return statements point to (which is always empty). Together, all the
1121 * blocks with their predecessors and successors make up the control flow
1122 * graph (CFG) of the function. There are helpers that modify the tree of
1123 * control flow nodes while modifying the CFG appropriately; these should be
1124 * used instead of modifying the tree directly.
1131 nir_cf_node_function
1134 typedef struct nir_cf_node
{
1135 struct exec_node node
;
1136 nir_cf_node_type type
;
1137 struct nir_cf_node
*parent
;
1140 typedef struct nir_block
{
1141 nir_cf_node cf_node
;
1143 struct exec_list instr_list
; /** < list of nir_instr */
1145 /** generic block index; generated by nir_index_blocks */
1149 * Each block can only have up to 2 successors, so we put them in a simple
1150 * array - no need for anything more complicated.
1152 struct nir_block
*successors
[2];
1154 /* Set of nir_block predecessors in the CFG */
1155 struct set
*predecessors
;
1158 * this node's immediate dominator in the dominance tree - set to NULL for
1161 struct nir_block
*imm_dom
;
1163 /* This node's children in the dominance tree */
1164 unsigned num_dom_children
;
1165 struct nir_block
**dom_children
;
1167 /* Set of nir_block's on the dominance frontier of this block */
1168 struct set
*dom_frontier
;
1171 * These two indices have the property that dom_{pre,post}_index for each
1172 * child of this block in the dominance tree will always be between
1173 * dom_pre_index and dom_post_index for this block, which makes testing if
1174 * a given block is dominated by another block an O(1) operation.
1176 unsigned dom_pre_index
, dom_post_index
;
1178 /* live in and out for this block; used for liveness analysis */
1179 BITSET_WORD
*live_in
;
1180 BITSET_WORD
*live_out
;
1183 static inline nir_instr
*
1184 nir_block_first_instr(nir_block
*block
)
1186 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
1187 return exec_node_data(nir_instr
, head
, node
);
1190 static inline nir_instr
*
1191 nir_block_last_instr(nir_block
*block
)
1193 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
1194 return exec_node_data(nir_instr
, tail
, node
);
1197 #define nir_foreach_instr(block, instr) \
1198 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
1199 #define nir_foreach_instr_reverse(block, instr) \
1200 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
1201 #define nir_foreach_instr_safe(block, instr) \
1202 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
1205 nir_cf_node cf_node
;
1208 struct exec_list then_list
; /** < list of nir_cf_node */
1209 struct exec_list else_list
; /** < list of nir_cf_node */
1212 static inline nir_cf_node
*
1213 nir_if_first_then_node(nir_if
*if_stmt
)
1215 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
1216 return exec_node_data(nir_cf_node
, head
, node
);
1219 static inline nir_cf_node
*
1220 nir_if_last_then_node(nir_if
*if_stmt
)
1222 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
1223 return exec_node_data(nir_cf_node
, tail
, node
);
1226 static inline nir_cf_node
*
1227 nir_if_first_else_node(nir_if
*if_stmt
)
1229 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
1230 return exec_node_data(nir_cf_node
, head
, node
);
1233 static inline nir_cf_node
*
1234 nir_if_last_else_node(nir_if
*if_stmt
)
1236 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
1237 return exec_node_data(nir_cf_node
, tail
, node
);
1241 nir_cf_node cf_node
;
1243 struct exec_list body
; /** < list of nir_cf_node */
1246 static inline nir_cf_node
*
1247 nir_loop_first_cf_node(nir_loop
*loop
)
1249 return exec_node_data(nir_cf_node
, exec_list_get_head(&loop
->body
), node
);
1252 static inline nir_cf_node
*
1253 nir_loop_last_cf_node(nir_loop
*loop
)
1255 return exec_node_data(nir_cf_node
, exec_list_get_tail(&loop
->body
), node
);
1259 * Various bits of metadata that can may be created or required by
1260 * optimization and analysis passes
1263 nir_metadata_none
= 0x0,
1264 nir_metadata_block_index
= 0x1,
1265 nir_metadata_dominance
= 0x2,
1266 nir_metadata_live_variables
= 0x4,
1270 nir_cf_node cf_node
;
1272 /** pointer to the overload of which this is an implementation */
1273 struct nir_function_overload
*overload
;
1275 struct exec_list body
; /** < list of nir_cf_node */
1277 nir_block
*start_block
, *end_block
;
1279 /** list for all local variables in the function */
1280 struct exec_list locals
;
1282 /** array of variables used as parameters */
1283 unsigned num_params
;
1284 nir_variable
**params
;
1286 /** variable used to hold the result of the function */
1287 nir_variable
*return_var
;
1289 /** list of local registers in the function */
1290 struct exec_list registers
;
1292 /** next available local register index */
1295 /** next available SSA value index */
1298 /* total number of basic blocks, only valid when block_index_dirty = false */
1299 unsigned num_blocks
;
1301 nir_metadata valid_metadata
;
1302 } nir_function_impl
;
1304 static inline nir_cf_node
*
1305 nir_cf_node_next(nir_cf_node
*node
)
1307 struct exec_node
*next
= exec_node_get_next(&node
->node
);
1308 if (exec_node_is_tail_sentinel(next
))
1311 return exec_node_data(nir_cf_node
, next
, node
);
1314 static inline nir_cf_node
*
1315 nir_cf_node_prev(nir_cf_node
*node
)
1317 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
1318 if (exec_node_is_head_sentinel(prev
))
1321 return exec_node_data(nir_cf_node
, prev
, node
);
1325 nir_cf_node_is_first(const nir_cf_node
*node
)
1327 return exec_node_is_head_sentinel(node
->node
.prev
);
1331 nir_cf_node_is_last(const nir_cf_node
*node
)
1333 return exec_node_is_tail_sentinel(node
->node
.next
);
1336 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
)
1337 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
)
1338 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
)
1339 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
, nir_function_impl
, cf_node
)
1344 nir_parameter_inout
,
1345 } nir_parameter_type
;
1348 nir_parameter_type param_type
;
1349 const struct glsl_type
*type
;
1352 typedef struct nir_function_overload
{
1353 struct exec_node node
;
1355 unsigned num_params
;
1356 nir_parameter
*params
;
1357 const struct glsl_type
*return_type
;
1359 nir_function_impl
*impl
; /** < NULL if the overload is only declared yet */
1361 /** pointer to the function of which this is an overload */
1362 struct nir_function
*function
;
1363 } nir_function_overload
;
1365 typedef struct nir_function
{
1366 struct exec_node node
;
1368 struct exec_list overload_list
; /** < list of nir_function_overload */
1370 struct nir_shader
*shader
;
1373 #define nir_function_first_overload(func) \
1374 exec_node_data(nir_function_overload, \
1375 exec_list_get_head(&(func)->overload_list), node)
1377 typedef struct nir_shader_compiler_options
{
1383 /** lowers fneg and ineg to fsub and isub. */
1386 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
1390 * Does the driver support real 32-bit integers? (Otherwise, integers
1391 * are simulated by floats.)
1393 bool native_integers
;
1394 } nir_shader_compiler_options
;
1396 typedef struct nir_shader
{
1397 /** hash table of name -> uniform nir_variable */
1398 struct exec_list uniforms
;
1400 /** hash table of name -> input nir_variable */
1401 struct exec_list inputs
;
1403 /** hash table of name -> output nir_variable */
1404 struct exec_list outputs
;
1406 /** Set of driver-specific options for the shader.
1408 * The memory for the options is expected to be kept in a single static
1409 * copy by the driver.
1411 const struct nir_shader_compiler_options
*options
;
1413 /** list of global variables in the shader */
1414 struct exec_list globals
;
1416 /** list of system value variables in the shader */
1417 struct exec_list system_values
;
1419 struct exec_list functions
; /** < list of nir_function */
1421 /** list of global register in the shader */
1422 struct exec_list registers
;
1424 /** next available global register index */
1428 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
1431 unsigned num_inputs
, num_uniforms
, num_outputs
;
1434 #define nir_foreach_overload(shader, overload) \
1435 foreach_list_typed(nir_function, func, node, &(shader)->functions) \
1436 foreach_list_typed(nir_function_overload, overload, node, \
1437 &(func)->overload_list)
1439 nir_shader
*nir_shader_create(void *mem_ctx
,
1440 const nir_shader_compiler_options
*options
);
1442 /** creates a register, including assigning it an index and adding it to the list */
1443 nir_register
*nir_global_reg_create(nir_shader
*shader
);
1445 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
1447 void nir_reg_remove(nir_register
*reg
);
1449 /** creates a function and adds it to the shader's list of functions */
1450 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
1452 /** creates a null function returning null */
1453 nir_function_overload
*nir_function_overload_create(nir_function
*func
);
1455 nir_function_impl
*nir_function_impl_create(nir_function_overload
*func
);
1457 nir_block
*nir_block_create(void *mem_ctx
);
1458 nir_if
*nir_if_create(void *mem_ctx
);
1459 nir_loop
*nir_loop_create(void *mem_ctx
);
1461 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
1463 /** puts a control flow node immediately after another control flow node */
1464 void nir_cf_node_insert_after(nir_cf_node
*node
, nir_cf_node
*after
);
1466 /** puts a control flow node immediately before another control flow node */
1467 void nir_cf_node_insert_before(nir_cf_node
*node
, nir_cf_node
*before
);
1469 /** puts a control flow node at the beginning of a list from an if, loop, or function */
1470 void nir_cf_node_insert_begin(struct exec_list
*list
, nir_cf_node
*node
);
1472 /** puts a control flow node at the end of a list from an if, loop, or function */
1473 void nir_cf_node_insert_end(struct exec_list
*list
, nir_cf_node
*node
);
1475 /** removes a control flow node, doing any cleanup necessary */
1476 void nir_cf_node_remove(nir_cf_node
*node
);
1478 /** requests that the given pieces of metadata be generated */
1479 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
);
1480 /** dirties all but the preserved metadata */
1481 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
1483 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
1484 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
1486 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
1488 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
1489 unsigned num_components
);
1491 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
1492 nir_intrinsic_op op
);
1494 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
1495 nir_function_overload
*callee
);
1497 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
1499 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
1501 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
1503 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
1504 unsigned num_components
);
1506 nir_deref_var
*nir_deref_var_create(void *mem_ctx
, nir_variable
*var
);
1507 nir_deref_array
*nir_deref_array_create(void *mem_ctx
);
1508 nir_deref_struct
*nir_deref_struct_create(void *mem_ctx
, unsigned field_index
);
1510 nir_deref
*nir_copy_deref(void *mem_ctx
, nir_deref
*deref
);
1512 void nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
);
1513 void nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
);
1515 void nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
);
1516 void nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
);
1518 void nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
);
1519 void nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
);
1521 void nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
);
1522 void nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
);
1524 void nir_instr_remove(nir_instr
*instr
);
1526 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
1527 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
1528 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
1529 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
1531 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
1532 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
1534 nir_const_value
*nir_src_as_const_value(nir_src src
);
1535 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
1536 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
1538 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
1539 unsigned num_components
, const char *name
);
1540 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
1541 unsigned num_components
, const char *name
);
1542 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
, void *mem_ctx
);
1544 /* visits basic blocks in source-code order */
1545 typedef bool (*nir_foreach_block_cb
)(nir_block
*block
, void *state
);
1546 bool nir_foreach_block(nir_function_impl
*impl
, nir_foreach_block_cb cb
,
1548 bool nir_foreach_block_reverse(nir_function_impl
*impl
, nir_foreach_block_cb cb
,
1551 /* If the following CF node is an if, this function returns that if.
1552 * Otherwise, it returns NULL.
1554 nir_if
*nir_block_get_following_if(nir_block
*block
);
1556 void nir_index_local_regs(nir_function_impl
*impl
);
1557 void nir_index_global_regs(nir_shader
*shader
);
1558 void nir_index_ssa_defs(nir_function_impl
*impl
);
1560 void nir_index_blocks(nir_function_impl
*impl
);
1562 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
1563 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
1566 void nir_validate_shader(nir_shader
*shader
);
1568 static inline void nir_validate_shader(nir_shader
*shader
) { }
1571 void nir_calc_dominance_impl(nir_function_impl
*impl
);
1572 void nir_calc_dominance(nir_shader
*shader
);
1574 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
1575 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
1577 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
1578 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
1580 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
1581 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
1583 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
1584 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
1586 void nir_split_var_copies(nir_shader
*shader
);
1588 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, void *mem_ctx
);
1589 void nir_lower_var_copies(nir_shader
*shader
);
1591 void nir_lower_global_vars_to_local(nir_shader
*shader
);
1593 void nir_lower_locals_to_regs(nir_shader
*shader
);
1595 void nir_assign_var_locations_scalar(struct exec_list
*var_list
,
1597 void nir_assign_var_locations_scalar_direct_first(nir_shader
*shader
,
1598 struct exec_list
*var_list
,
1599 unsigned *direct_size
,
1602 void nir_lower_io(nir_shader
*shader
);
1604 void nir_lower_vars_to_ssa(nir_shader
*shader
);
1606 void nir_remove_dead_variables(nir_shader
*shader
);
1608 void nir_lower_vec_to_movs(nir_shader
*shader
);
1609 void nir_lower_alu_to_scalar(nir_shader
*shader
);
1611 void nir_lower_phis_to_scalar(nir_shader
*shader
);
1613 void nir_lower_samplers(nir_shader
*shader
,
1614 const struct gl_shader_program
*shader_program
,
1615 gl_shader_stage stage
);
1617 void nir_lower_system_values(nir_shader
*shader
);
1618 void nir_lower_tex_projector(nir_shader
*shader
);
1619 void nir_lower_idiv(nir_shader
*shader
);
1621 void nir_lower_atomics(nir_shader
*shader
);
1622 void nir_lower_to_source_mods(nir_shader
*shader
);
1624 void nir_normalize_cubemap_coords(nir_shader
*shader
);
1626 void nir_live_variables_impl(nir_function_impl
*impl
);
1627 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
1629 void nir_convert_to_ssa_impl(nir_function_impl
*impl
);
1630 void nir_convert_to_ssa(nir_shader
*shader
);
1631 void nir_convert_from_ssa(nir_shader
*shader
);
1633 bool nir_opt_algebraic(nir_shader
*shader
);
1634 bool nir_opt_algebraic_late(nir_shader
*shader
);
1635 bool nir_opt_constant_folding(nir_shader
*shader
);
1637 bool nir_opt_global_to_local(nir_shader
*shader
);
1639 bool nir_copy_prop_impl(nir_function_impl
*impl
);
1640 bool nir_copy_prop(nir_shader
*shader
);
1642 bool nir_opt_cse(nir_shader
*shader
);
1644 bool nir_opt_dce_impl(nir_function_impl
*impl
);
1645 bool nir_opt_dce(nir_shader
*shader
);
1647 void nir_opt_gcm(nir_shader
*shader
);
1649 bool nir_opt_peephole_select(nir_shader
*shader
);
1650 bool nir_opt_peephole_ffma(nir_shader
*shader
);
1652 bool nir_opt_remove_phis(nir_shader
*shader
);
1654 void nir_sweep(nir_shader
*shader
);