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"
39 #include "nir_opcodes.h"
46 struct gl_shader_program
;
49 #define NIR_TRUE (~0u)
51 /** Defines a cast function
53 * This macro defines a cast function from in_type to out_type where
54 * out_type is some structure type that contains a field of type out_type.
56 * Note that you have to be a bit careful as the generated cast function
59 #define NIR_DEFINE_CAST(name, in_type, out_type, field) \
60 static inline out_type * \
61 name(const in_type *parent) \
63 return exec_node_data(out_type, parent, field); \
66 struct nir_function_overload
;
72 * Description of built-in state associated with a uniform
74 * \sa nir_variable::state_slots
91 * Data stored in an nir_constant
93 union nir_constant_data
{
100 typedef struct nir_constant
{
102 * Value of the constant.
104 * The field used to back the values supplied by the constant is determined
105 * by the type associated with the \c nir_variable. Constants may be
106 * scalars, vectors, or matrices.
108 union nir_constant_data value
;
110 /* Array elements / Structure Fields */
111 struct nir_constant
**elements
;
115 * \brief Layout qualifiers for gl_FragDepth.
117 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
118 * with a layout qualifier.
121 nir_depth_layout_none
, /**< No depth layout is specified. */
122 nir_depth_layout_any
,
123 nir_depth_layout_greater
,
124 nir_depth_layout_less
,
125 nir_depth_layout_unchanged
129 * Either a uniform, global variable, shader input, or shader output. Based on
130 * ir_variable - it should be easy to translate between the two.
134 struct exec_node node
;
137 * Declared type of the variable
139 const struct glsl_type
*type
;
142 * Declared name of the variable
147 * For variables which satisfy the is_interface_instance() predicate, this
148 * points to an array of integers such that if the ith member of the
149 * interface block is an array, max_ifc_array_access[i] is the maximum
150 * array element of that member that has been accessed. If the ith member
151 * of the interface block is not an array, max_ifc_array_access[i] is
154 * For variables whose type is not an interface block, this pointer is
157 unsigned *max_ifc_array_access
;
159 struct nir_variable_data
{
162 * Is the variable read-only?
164 * This is set for variables declared as \c const, shader inputs,
167 unsigned read_only
:1;
170 unsigned invariant
:1;
173 * Storage class of the variable.
175 * \sa nir_variable_mode
177 nir_variable_mode mode
:4;
180 * Interpolation mode for shader inputs / outputs
182 * \sa glsl_interp_qualifier
184 unsigned interpolation
:2;
187 * \name ARB_fragment_coord_conventions
190 unsigned origin_upper_left
:1;
191 unsigned pixel_center_integer
:1;
195 * Was the location explicitly set in the shader?
197 * If the location is explicitly set in the shader, it \b cannot be changed
198 * by the linker or by the API (e.g., calls to \c glBindAttribLocation have
201 unsigned explicit_location
:1;
202 unsigned explicit_index
:1;
205 * Was an initial binding explicitly set in the shader?
207 * If so, constant_initializer contains an integer nir_constant
208 * representing the initial binding point.
210 unsigned explicit_binding
:1;
213 * Does this variable have an initializer?
215 * This is used by the linker to cross-validiate initializers of global
218 unsigned has_initializer
:1;
221 * Is this variable a generic output or input that has not yet been matched
222 * up to a variable in another stage of the pipeline?
224 * This is used by the linker as scratch storage while assigning locations
225 * to generic inputs and outputs.
227 unsigned is_unmatched_generic_inout
:1;
230 * If non-zero, then this variable may be packed along with other variables
231 * into a single varying slot, so this offset should be applied when
232 * accessing components. For example, an offset of 1 means that the x
233 * component of this variable is actually stored in component y of the
234 * location specified by \c location.
236 unsigned location_frac
:2;
239 * Non-zero if this variable was created by lowering a named interface
240 * block which was not an array.
242 * Note that this variable and \c from_named_ifc_block_array will never
245 unsigned from_named_ifc_block_nonarray
:1;
248 * Non-zero if this variable was created by lowering a named interface
249 * block which was an array.
251 * Note that this variable and \c from_named_ifc_block_nonarray will never
254 unsigned from_named_ifc_block_array
:1;
257 * \brief Layout qualifier for gl_FragDepth.
259 * This is not equal to \c ir_depth_layout_none if and only if this
260 * variable is \c gl_FragDepth and a layout qualifier is specified.
262 nir_depth_layout depth_layout
;
265 * Storage location of the base of this variable
267 * The precise meaning of this field depends on the nature of the variable.
269 * - Vertex shader input: one of the values from \c gl_vert_attrib.
270 * - Vertex shader output: one of the values from \c gl_varying_slot.
271 * - Geometry shader input: one of the values from \c gl_varying_slot.
272 * - Geometry shader output: one of the values from \c gl_varying_slot.
273 * - Fragment shader input: one of the values from \c gl_varying_slot.
274 * - Fragment shader output: one of the values from \c gl_frag_result.
275 * - Uniforms: Per-stage uniform slot number for default uniform block.
276 * - Uniforms: Index within the uniform block definition for UBO members.
277 * - Other: This field is not currently used.
279 * If the variable is a uniform, shader input, or shader output, and the
280 * slot has not been assigned, the value will be -1.
285 * The actual location of the variable in the IR. Only valid for inputs
288 unsigned int driver_location
;
291 * output index for dual source blending.
296 * Initial binding point for a sampler or UBO.
298 * For array types, this represents the binding point for the first element.
303 * Location an atomic counter is stored at.
306 unsigned buffer_index
;
311 * ARB_shader_image_load_store qualifiers.
314 bool read_only
; /**< "readonly" qualifier. */
315 bool write_only
; /**< "writeonly" qualifier. */
320 /** Image internal format if specified explicitly, otherwise GL_NONE. */
325 * Highest element accessed with a constant expression array index
327 * Not used for non-array variables.
329 unsigned max_array_access
;
334 * Built-in state that backs this uniform
336 * Once set at variable creation, \c state_slots must remain invariant.
337 * This is because, ideally, this array would be shared by all clones of
338 * this variable in the IR tree. In other words, we'd really like for it
339 * to be a fly-weight.
341 * If the variable is not a uniform, \c num_state_slots will be zero and
342 * \c state_slots will be \c NULL.
345 unsigned num_state_slots
; /**< Number of state slots used */
346 nir_state_slot
*state_slots
; /**< State descriptors. */
350 * Constant expression assigned in the initializer of the variable
352 nir_constant
*constant_initializer
;
355 * For variables that are in an interface block or are an instance of an
356 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
358 * \sa ir_variable::location
360 const struct glsl_type
*interface_type
;
364 struct exec_node node
;
366 unsigned num_components
; /** < number of vector components */
367 unsigned num_array_elems
; /** < size of array (0 for no array) */
369 /** generic register index. */
372 /** only for debug purposes, can be NULL */
375 /** whether this register is local (per-function) or global (per-shader) */
379 * If this flag is set to true, then accessing channels >= num_components
380 * is well-defined, and simply spills over to the next array element. This
381 * is useful for backends that can do per-component accessing, in
382 * particular scalar backends. By setting this flag and making
383 * num_components equal to 1, structures can be packed tightly into
384 * registers and then registers can be accessed per-component to get to
385 * each structure member, even if it crosses vec4 boundaries.
389 /** set of nir_instr's where this register is used (read from) */
392 /** set of nir_instr's where this register is defined (written to) */
395 /** set of nir_if's where this register is used as a condition */
403 nir_instr_type_intrinsic
,
404 nir_instr_type_load_const
,
406 nir_instr_type_ssa_undef
,
408 nir_instr_type_parallel_copy
,
412 struct exec_node node
;
414 struct nir_block
*block
;
416 /* A temporary for optimization and analysis passes to use for storing
417 * flags. For instance, DCE uses this to store the "dead/live" info.
422 static inline nir_instr
*
423 nir_instr_next(nir_instr
*instr
)
425 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
426 if (exec_node_is_tail_sentinel(next
))
429 return exec_node_data(nir_instr
, next
, node
);
432 static inline nir_instr
*
433 nir_instr_prev(nir_instr
*instr
)
435 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
436 if (exec_node_is_head_sentinel(prev
))
439 return exec_node_data(nir_instr
, prev
, node
);
443 /** for debugging only, can be NULL */
446 /** generic SSA definition index. */
449 /** Index into the live_in and live_out bitfields */
452 nir_instr
*parent_instr
;
454 /** set of nir_instr's where this register is used (read from) */
457 /** set of nir_if's where this register is used as a condition */
460 uint8_t num_components
;
467 struct nir_src
*indirect
; /** < NULL for no indirect offset */
468 unsigned base_offset
;
470 /* TODO use-def chain goes here */
475 struct nir_src
*indirect
; /** < NULL for no indirect offset */
476 unsigned base_offset
;
478 /* TODO def-use chain goes here */
481 typedef struct nir_src
{
499 static inline nir_src
500 nir_src_for_ssa(nir_ssa_def
*def
)
510 static inline nir_src
511 nir_src_for_reg(nir_register
*reg
)
517 src
.reg
.indirect
= NULL
;
518 src
.reg
.base_offset
= 0;
523 static inline nir_dest
524 nir_dest_for_reg(nir_register
*reg
)
530 dest
.reg
.indirect
= NULL
;
531 dest
.reg
.base_offset
= 0;
536 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *mem_ctx
);
537 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, void *mem_ctx
);
543 * \name input modifiers
547 * For inputs interpreted as floating point, flips the sign bit. For
548 * inputs interpreted as integers, performs the two's complement negation.
553 * Clears the sign bit for floating point values, and computes the integer
554 * absolute value for integers. Note that the negate modifier acts after
555 * the absolute value modifier, therefore if both are set then all inputs
556 * will become negative.
562 * For each input component, says which component of the register it is
563 * chosen from. Note that which elements of the swizzle are used and which
564 * are ignored are based on the write mask for most opcodes - for example,
565 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
566 * a swizzle of {2, x, 1, 0} where x means "don't care."
575 * \name saturate output modifier
577 * Only valid for opcodes that output floating-point numbers. Clamps the
578 * output to between 0.0 and 1.0 inclusive.
583 unsigned write_mask
: 4; /* ignored if dest.is_ssa is true */
586 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
, void *mem_ctx
);
587 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
591 nir_type_invalid
= 0, /* Not a valid type */
599 NIR_OP_IS_COMMUTATIVE
= (1 << 0),
600 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
601 } nir_op_algebraic_property
;
609 * The number of components in the output
611 * If non-zero, this is the size of the output and input sizes are
612 * explicitly given; swizzle and writemask are still in effect, but if
613 * the output component is masked out, then the input component may
616 * If zero, the opcode acts in the standard, per-component manner; the
617 * operation is performed on each component (except the ones that are
618 * masked out) with the input being taken from the input swizzle for
621 * The size of some of the inputs may be given (i.e. non-zero) even
622 * though output_size is zero; in that case, the inputs with a zero
623 * size act per-component, while the inputs with non-zero size don't.
625 unsigned output_size
;
628 * The type of vector that the instruction outputs. Note that the
629 * staurate modifier is only allowed on outputs with the float type.
632 nir_alu_type output_type
;
635 * The number of components in each input
637 unsigned input_sizes
[4];
640 * The type of vector that each input takes. Note that negate and
641 * absolute value are only allowed on inputs with int or float type and
642 * behave differently on the two.
644 nir_alu_type input_types
[4];
646 nir_op_algebraic_property algebraic_properties
;
649 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
651 typedef struct nir_alu_instr
{
658 /* is this source channel used? */
660 nir_alu_instr_channel_used(nir_alu_instr
*instr
, unsigned src
, unsigned channel
)
662 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
663 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
665 return (instr
->dest
.write_mask
>> channel
) & 1;
669 * For instructions whose destinations are SSA, get the number of channels
672 static inline unsigned
673 nir_ssa_alu_instr_src_components(nir_alu_instr
*instr
, unsigned src
)
675 assert(instr
->dest
.dest
.is_ssa
);
677 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
678 return nir_op_infos
[instr
->op
].input_sizes
[src
];
680 return instr
->dest
.dest
.ssa
.num_components
;
685 nir_deref_type_array
,
686 nir_deref_type_struct
689 typedef struct nir_deref
{
690 nir_deref_type deref_type
;
691 struct nir_deref
*child
;
692 const struct glsl_type
*type
;
701 /* This enum describes how the array is referenced. If the deref is
702 * direct then the base_offset is used. If the deref is indirect then then
703 * offset is given by base_offset + indirect. If the deref is a wildcard
704 * then the deref refers to all of the elements of the array at the same
705 * time. Wildcard dereferences are only ever allowed in copy_var
706 * intrinsics and the source and destination derefs must have matching
710 nir_deref_array_type_direct
,
711 nir_deref_array_type_indirect
,
712 nir_deref_array_type_wildcard
,
713 } nir_deref_array_type
;
718 nir_deref_array_type deref_array_type
;
719 unsigned base_offset
;
729 NIR_DEFINE_CAST(nir_deref_as_var
, nir_deref
, nir_deref_var
, deref
)
730 NIR_DEFINE_CAST(nir_deref_as_array
, nir_deref
, nir_deref_array
, deref
)
731 NIR_DEFINE_CAST(nir_deref_as_struct
, nir_deref
, nir_deref_struct
, deref
)
737 nir_deref_var
**params
;
738 nir_deref_var
*return_deref
;
740 struct nir_function_overload
*callee
;
743 #define INTRINSIC(name, num_srcs, src_components, has_dest, dest_components, \
744 num_variables, num_indices, flags) \
745 nir_intrinsic_##name,
747 #define LAST_INTRINSIC(name) nir_last_intrinsic = nir_intrinsic_##name,
750 #include "nir_intrinsics.h"
751 nir_num_intrinsics
= nir_last_intrinsic
+ 1
755 #undef LAST_INTRINSIC
757 /** Represents an intrinsic
759 * An intrinsic is an instruction type for handling things that are
760 * more-or-less regular operations but don't just consume and produce SSA
761 * values like ALU operations do. Intrinsics are not for things that have
762 * special semantic meaning such as phi nodes and parallel copies.
763 * Examples of intrinsics include variable load/store operations, system
764 * value loads, and the like. Even though texturing more-or-less falls
765 * under this category, texturing is its own instruction type because
766 * trying to represent texturing with intrinsics would lead to a
767 * combinatorial explosion of intrinsic opcodes.
769 * By having a single instruction type for handling a lot of different
770 * cases, optimization passes can look for intrinsics and, for the most
771 * part, completely ignore them. Each intrinsic type also has a few
772 * possible flags that govern whether or not they can be reordered or
773 * eliminated. That way passes like dead code elimination can still work
774 * on intrisics without understanding the meaning of each.
776 * Each intrinsic has some number of constant indices, some number of
777 * variables, and some number of sources. What these sources, variables,
778 * and indices mean depends on the intrinsic and is documented with the
779 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
780 * instructions are the only types of instruction that can operate on
786 nir_intrinsic_op intrinsic
;
790 /** number of components if this is a vectorized intrinsic
792 * Similarly to ALU operations, some intrinsics are vectorized.
793 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
794 * For vectorized intrinsics, the num_components field specifies the
795 * number of destination components and the number of source components
796 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
798 uint8_t num_components
;
802 nir_deref_var
*variables
[2];
805 } nir_intrinsic_instr
;
808 * \name NIR intrinsics semantic flags
810 * information about what the compiler can do with the intrinsics.
812 * \sa nir_intrinsic_info::flags
816 * whether the intrinsic can be safely eliminated if none of its output
817 * value is not being used.
819 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
822 * Whether the intrinsic can be reordered with respect to any other
823 * intrinsic, i.e. whether the only reordering dependencies of the
824 * intrinsic are due to the register reads/writes.
826 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
827 } nir_intrinsic_semantic_flag
;
829 #define NIR_INTRINSIC_MAX_INPUTS 4
834 unsigned num_srcs
; /** < number of register/SSA inputs */
836 /** number of components of each input register
838 * If this value is 0, the number of components is given by the
839 * num_components field of nir_intrinsic_instr.
841 unsigned src_components
[NIR_INTRINSIC_MAX_INPUTS
];
845 /** number of components of the output register
847 * If this value is 0, the number of components is given by the
848 * num_components field of nir_intrinsic_instr.
850 unsigned dest_components
;
852 /** the number of inputs/outputs that are variables */
853 unsigned num_variables
;
855 /** the number of constant indices used by the intrinsic */
856 unsigned num_indices
;
858 /** semantic flags for calls to this intrinsic */
859 nir_intrinsic_semantic_flag flags
;
860 } nir_intrinsic_info
;
862 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
865 * \group texture information
867 * This gives semantic information about textures which is useful to the
868 * frontend, the backend, and lowering passes, but not the optimizer.
873 nir_tex_src_projector
,
874 nir_tex_src_comparitor
, /* shadow comparitor */
878 nir_tex_src_ms_index
, /* MSAA sample index */
881 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
882 nir_num_tex_src_types
887 nir_tex_src_type src_type
;
891 nir_texop_tex
, /**< Regular texture look-up */
892 nir_texop_txb
, /**< Texture look-up with LOD bias */
893 nir_texop_txl
, /**< Texture look-up with explicit LOD */
894 nir_texop_txd
, /**< Texture look-up with partial derivatvies */
895 nir_texop_txf
, /**< Texel fetch with explicit LOD */
896 nir_texop_txf_ms
, /**< Multisample texture fetch */
897 nir_texop_txs
, /**< Texture size */
898 nir_texop_lod
, /**< Texture lod query */
899 nir_texop_tg4
, /**< Texture gather */
900 nir_texop_query_levels
/**< Texture levels query */
906 enum glsl_sampler_dim sampler_dim
;
907 nir_alu_type dest_type
;
912 unsigned num_srcs
, coord_components
;
913 bool is_array
, is_shadow
;
916 * If is_shadow is true, whether this is the old-style shadow that outputs 4
917 * components or the new-style shadow that outputs 1 component.
919 bool is_new_style_shadow
;
921 /* constant offset - must be 0 if the offset source is used */
924 /* gather component selector */
925 unsigned component
: 2;
927 /** The sampler index
929 * If this texture instruction has a nir_tex_src_sampler_offset source,
930 * then the sampler index is given by sampler_index + sampler_offset.
932 unsigned sampler_index
;
934 /** The size of the sampler array or 0 if it's not an array */
935 unsigned sampler_array_size
;
937 nir_deref_var
*sampler
; /* if this is NULL, use sampler_index instead */
940 static inline unsigned
941 nir_tex_instr_dest_size(nir_tex_instr
*instr
)
943 if (instr
->op
== nir_texop_txs
) {
945 switch (instr
->sampler_dim
) {
946 case GLSL_SAMPLER_DIM_1D
:
947 case GLSL_SAMPLER_DIM_BUF
:
950 case GLSL_SAMPLER_DIM_2D
:
951 case GLSL_SAMPLER_DIM_CUBE
:
952 case GLSL_SAMPLER_DIM_MS
:
953 case GLSL_SAMPLER_DIM_RECT
:
954 case GLSL_SAMPLER_DIM_EXTERNAL
:
957 case GLSL_SAMPLER_DIM_3D
:
961 unreachable("not reached");
968 if (instr
->op
== nir_texop_query_levels
)
971 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
977 static inline unsigned
978 nir_tex_instr_src_size(nir_tex_instr
*instr
, unsigned src
)
980 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
981 return instr
->coord_components
;
984 if (instr
->src
[src
].src_type
== nir_tex_src_offset
||
985 instr
->src
[src
].src_type
== nir_tex_src_ddx
||
986 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
988 return instr
->coord_components
- 1;
990 return instr
->coord_components
;
997 nir_tex_instr_src_index(nir_tex_instr
*instr
, nir_tex_src_type type
)
999 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
1000 if (instr
->src
[i
].src_type
== type
)
1017 nir_const_value value
;
1020 } nir_load_const_instr
;
1033 /* creates a new SSA variable in an undefined state */
1038 } nir_ssa_undef_instr
;
1041 struct exec_node node
;
1043 /* The predecessor block corresponding to this source */
1044 struct nir_block
*pred
;
1049 #define nir_foreach_phi_src(phi, entry) \
1050 foreach_list_typed(nir_phi_src, entry, node, &(phi)->srcs)
1055 struct exec_list srcs
; /** < list of nir_phi_src */
1061 struct exec_node node
;
1064 } nir_parallel_copy_entry
;
1066 #define nir_foreach_parallel_copy_entry(pcopy, entry) \
1067 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
1072 /* A list of nir_parallel_copy_entry's. The sources of all of the
1073 * entries are copied to the corresponding destinations "in parallel".
1074 * In other words, if we have two entries: a -> b and b -> a, the values
1077 struct exec_list entries
;
1078 } nir_parallel_copy_instr
;
1080 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
)
1081 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
)
1082 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
)
1083 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
)
1084 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
)
1085 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
)
1086 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
)
1087 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
)
1088 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
1089 nir_parallel_copy_instr
, instr
)
1094 * Control flow consists of a tree of control flow nodes, which include
1095 * if-statements and loops. The leaves of the tree are basic blocks, lists of
1096 * instructions that always run start-to-finish. Each basic block also keeps
1097 * track of its successors (blocks which may run immediately after the current
1098 * block) and predecessors (blocks which could have run immediately before the
1099 * current block). Each function also has a start block and an end block which
1100 * all return statements point to (which is always empty). Together, all the
1101 * blocks with their predecessors and successors make up the control flow
1102 * graph (CFG) of the function. There are helpers that modify the tree of
1103 * control flow nodes while modifying the CFG appropriately; these should be
1104 * used instead of modifying the tree directly.
1111 nir_cf_node_function
1114 typedef struct nir_cf_node
{
1115 struct exec_node node
;
1116 nir_cf_node_type type
;
1117 struct nir_cf_node
*parent
;
1120 typedef struct nir_block
{
1121 nir_cf_node cf_node
;
1123 struct exec_list instr_list
; /** < list of nir_instr */
1125 /** generic block index; generated by nir_index_blocks */
1129 * Each block can only have up to 2 successors, so we put them in a simple
1130 * array - no need for anything more complicated.
1132 struct nir_block
*successors
[2];
1134 /* Set of nir_block predecessors in the CFG */
1135 struct set
*predecessors
;
1138 * this node's immediate dominator in the dominance tree - set to NULL for
1141 struct nir_block
*imm_dom
;
1143 /* This node's children in the dominance tree */
1144 unsigned num_dom_children
;
1145 struct nir_block
**dom_children
;
1147 /* Set of nir_block's on the dominance frontier of this block */
1148 struct set
*dom_frontier
;
1151 * These two indices have the property that dom_{pre,post}_index for each
1152 * child of this block in the dominance tree will always be between
1153 * dom_pre_index and dom_post_index for this block, which makes testing if
1154 * a given block is dominated by another block an O(1) operation.
1156 unsigned dom_pre_index
, dom_post_index
;
1158 /* live in and out for this block; used for liveness analysis */
1159 BITSET_WORD
*live_in
;
1160 BITSET_WORD
*live_out
;
1163 static inline nir_instr
*
1164 nir_block_first_instr(nir_block
*block
)
1166 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
1167 return exec_node_data(nir_instr
, head
, node
);
1170 static inline nir_instr
*
1171 nir_block_last_instr(nir_block
*block
)
1173 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
1174 return exec_node_data(nir_instr
, tail
, node
);
1177 #define nir_foreach_instr(block, instr) \
1178 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
1179 #define nir_foreach_instr_reverse(block, instr) \
1180 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
1181 #define nir_foreach_instr_safe(block, instr) \
1182 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
1185 nir_cf_node cf_node
;
1188 struct exec_list then_list
; /** < list of nir_cf_node */
1189 struct exec_list else_list
; /** < list of nir_cf_node */
1192 static inline nir_cf_node
*
1193 nir_if_first_then_node(nir_if
*if_stmt
)
1195 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
1196 return exec_node_data(nir_cf_node
, head
, node
);
1199 static inline nir_cf_node
*
1200 nir_if_last_then_node(nir_if
*if_stmt
)
1202 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
1203 return exec_node_data(nir_cf_node
, tail
, node
);
1206 static inline nir_cf_node
*
1207 nir_if_first_else_node(nir_if
*if_stmt
)
1209 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
1210 return exec_node_data(nir_cf_node
, head
, node
);
1213 static inline nir_cf_node
*
1214 nir_if_last_else_node(nir_if
*if_stmt
)
1216 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
1217 return exec_node_data(nir_cf_node
, tail
, node
);
1221 nir_cf_node cf_node
;
1223 struct exec_list body
; /** < list of nir_cf_node */
1226 static inline nir_cf_node
*
1227 nir_loop_first_cf_node(nir_loop
*loop
)
1229 return exec_node_data(nir_cf_node
, exec_list_get_head(&loop
->body
), node
);
1232 static inline nir_cf_node
*
1233 nir_loop_last_cf_node(nir_loop
*loop
)
1235 return exec_node_data(nir_cf_node
, exec_list_get_tail(&loop
->body
), node
);
1239 * Various bits of metadata that can may be created or required by
1240 * optimization and analysis passes
1243 nir_metadata_none
= 0x0,
1244 nir_metadata_block_index
= 0x1,
1245 nir_metadata_dominance
= 0x2,
1246 nir_metadata_live_variables
= 0x4,
1250 nir_cf_node cf_node
;
1252 /** pointer to the overload of which this is an implementation */
1253 struct nir_function_overload
*overload
;
1255 struct exec_list body
; /** < list of nir_cf_node */
1257 nir_block
*start_block
, *end_block
;
1259 /** list for all local variables in the function */
1260 struct exec_list locals
;
1262 /** array of variables used as parameters */
1263 unsigned num_params
;
1264 nir_variable
**params
;
1266 /** variable used to hold the result of the function */
1267 nir_variable
*return_var
;
1269 /** list of local registers in the function */
1270 struct exec_list registers
;
1272 /** next available local register index */
1275 /** next available SSA value index */
1278 /* total number of basic blocks, only valid when block_index_dirty = false */
1279 unsigned num_blocks
;
1281 nir_metadata valid_metadata
;
1282 } nir_function_impl
;
1284 static inline nir_cf_node
*
1285 nir_cf_node_next(nir_cf_node
*node
)
1287 struct exec_node
*next
= exec_node_get_next(&node
->node
);
1288 if (exec_node_is_tail_sentinel(next
))
1291 return exec_node_data(nir_cf_node
, next
, node
);
1294 static inline nir_cf_node
*
1295 nir_cf_node_prev(nir_cf_node
*node
)
1297 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
1298 if (exec_node_is_head_sentinel(prev
))
1301 return exec_node_data(nir_cf_node
, prev
, node
);
1305 nir_cf_node_is_first(const nir_cf_node
*node
)
1307 return exec_node_is_head_sentinel(node
->node
.prev
);
1311 nir_cf_node_is_last(const nir_cf_node
*node
)
1313 return exec_node_is_tail_sentinel(node
->node
.next
);
1316 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
)
1317 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
)
1318 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
)
1319 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
, nir_function_impl
, cf_node
)
1324 nir_parameter_inout
,
1325 } nir_parameter_type
;
1328 nir_parameter_type param_type
;
1329 const struct glsl_type
*type
;
1332 typedef struct nir_function_overload
{
1333 struct exec_node node
;
1335 unsigned num_params
;
1336 nir_parameter
*params
;
1337 const struct glsl_type
*return_type
;
1339 nir_function_impl
*impl
; /** < NULL if the overload is only declared yet */
1341 /** pointer to the function of which this is an overload */
1342 struct nir_function
*function
;
1343 } nir_function_overload
;
1345 typedef struct nir_function
{
1346 struct exec_node node
;
1348 struct exec_list overload_list
; /** < list of nir_function_overload */
1350 struct nir_shader
*shader
;
1353 #define nir_function_first_overload(func) \
1354 exec_node_data(nir_function_overload, \
1355 exec_list_get_head(&(func)->overload_list), node)
1357 typedef struct nir_shader_compiler_options
{
1362 /** lowers fneg and ineg to fsub and isub. */
1364 } nir_shader_compiler_options
;
1366 typedef struct nir_shader
{
1367 /** hash table of name -> uniform nir_variable */
1368 struct hash_table
*uniforms
;
1370 /** hash table of name -> input nir_variable */
1371 struct hash_table
*inputs
;
1373 /** hash table of name -> output nir_variable */
1374 struct hash_table
*outputs
;
1376 /** Set of driver-specific options for the shader.
1378 * The memory for the options is expected to be kept in a single static
1379 * copy by the driver.
1381 const struct nir_shader_compiler_options
*options
;
1383 /** list of global variables in the shader */
1384 struct exec_list globals
;
1386 /** list of system value variables in the shader */
1387 struct exec_list system_values
;
1389 struct exec_list functions
; /** < list of nir_function */
1391 /** list of global register in the shader */
1392 struct exec_list registers
;
1394 /** structures used in this shader */
1395 unsigned num_user_structures
;
1396 struct glsl_type
**user_structures
;
1398 /** next available global register index */
1402 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
1405 unsigned num_inputs
, num_uniforms
, num_outputs
;
1408 #define nir_foreach_overload(shader, overload) \
1409 foreach_list_typed(nir_function, func, node, &(shader)->functions) \
1410 foreach_list_typed(nir_function_overload, overload, node, \
1411 &(func)->overload_list)
1413 nir_shader
*nir_shader_create(void *mem_ctx
,
1414 const nir_shader_compiler_options
*options
);
1416 /** creates a register, including assigning it an index and adding it to the list */
1417 nir_register
*nir_global_reg_create(nir_shader
*shader
);
1419 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
1421 void nir_reg_remove(nir_register
*reg
);
1423 /** creates a function and adds it to the shader's list of functions */
1424 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
1426 /** creates a null function returning null */
1427 nir_function_overload
*nir_function_overload_create(nir_function
*func
);
1429 nir_function_impl
*nir_function_impl_create(nir_function_overload
*func
);
1431 nir_block
*nir_block_create(void *mem_ctx
);
1432 nir_if
*nir_if_create(void *mem_ctx
);
1433 nir_loop
*nir_loop_create(void *mem_ctx
);
1435 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
1437 /** puts a control flow node immediately after another control flow node */
1438 void nir_cf_node_insert_after(nir_cf_node
*node
, nir_cf_node
*after
);
1440 /** puts a control flow node immediately before another control flow node */
1441 void nir_cf_node_insert_before(nir_cf_node
*node
, nir_cf_node
*before
);
1443 /** puts a control flow node at the beginning of a list from an if, loop, or function */
1444 void nir_cf_node_insert_begin(struct exec_list
*list
, nir_cf_node
*node
);
1446 /** puts a control flow node at the end of a list from an if, loop, or function */
1447 void nir_cf_node_insert_end(struct exec_list
*list
, nir_cf_node
*node
);
1449 /** removes a control flow node, doing any cleanup necessary */
1450 void nir_cf_node_remove(nir_cf_node
*node
);
1452 /** requests that the given pieces of metadata be generated */
1453 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
);
1454 /** dirties all but the preserved metadata */
1455 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
1457 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
1458 nir_alu_instr
*nir_alu_instr_create(void *mem_ctx
, nir_op op
);
1460 nir_jump_instr
*nir_jump_instr_create(void *mem_ctx
, nir_jump_type type
);
1462 nir_load_const_instr
*nir_load_const_instr_create(void *mem_ctx
,
1463 unsigned num_components
);
1465 nir_intrinsic_instr
*nir_intrinsic_instr_create(void *mem_ctx
,
1466 nir_intrinsic_op op
);
1468 nir_call_instr
*nir_call_instr_create(void *mem_ctx
,
1469 nir_function_overload
*callee
);
1471 nir_tex_instr
*nir_tex_instr_create(void *mem_ctx
, unsigned num_srcs
);
1473 nir_phi_instr
*nir_phi_instr_create(void *mem_ctx
);
1475 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(void *mem_ctx
);
1477 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(void *mem_ctx
,
1478 unsigned num_components
);
1480 nir_deref_var
*nir_deref_var_create(void *mem_ctx
, nir_variable
*var
);
1481 nir_deref_array
*nir_deref_array_create(void *mem_ctx
);
1482 nir_deref_struct
*nir_deref_struct_create(void *mem_ctx
, unsigned field_index
);
1484 nir_deref
*nir_copy_deref(void *mem_ctx
, nir_deref
*deref
);
1486 void nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
);
1487 void nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
);
1489 void nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
);
1490 void nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
);
1492 void nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
);
1493 void nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
);
1495 void nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
);
1496 void nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
);
1498 void nir_instr_remove(nir_instr
*instr
);
1500 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
1501 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
1502 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
1503 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
1505 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
1506 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
1508 nir_const_value
*nir_src_as_const_value(nir_src src
);
1509 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
1510 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
1512 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
1513 unsigned num_components
, const char *name
);
1514 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
1515 unsigned num_components
, const char *name
);
1516 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
, void *mem_ctx
);
1518 /* visits basic blocks in source-code order */
1519 typedef bool (*nir_foreach_block_cb
)(nir_block
*block
, void *state
);
1520 bool nir_foreach_block(nir_function_impl
*impl
, nir_foreach_block_cb cb
,
1522 bool nir_foreach_block_reverse(nir_function_impl
*impl
, nir_foreach_block_cb cb
,
1525 /* If the following CF node is an if, this function returns that if.
1526 * Otherwise, it returns NULL.
1528 nir_if
*nir_block_get_following_if(nir_block
*block
);
1530 void nir_index_local_regs(nir_function_impl
*impl
);
1531 void nir_index_global_regs(nir_shader
*shader
);
1532 void nir_index_ssa_defs(nir_function_impl
*impl
);
1534 void nir_index_blocks(nir_function_impl
*impl
);
1536 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
1537 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
1540 void nir_validate_shader(nir_shader
*shader
);
1542 static inline void nir_validate_shader(nir_shader
*shader
) { }
1545 void nir_calc_dominance_impl(nir_function_impl
*impl
);
1546 void nir_calc_dominance(nir_shader
*shader
);
1548 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
1549 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
1551 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
1552 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
1554 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
1555 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
1557 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
1558 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
1560 void nir_split_var_copies(nir_shader
*shader
);
1562 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, void *mem_ctx
);
1563 void nir_lower_var_copies(nir_shader
*shader
);
1565 void nir_lower_global_vars_to_local(nir_shader
*shader
);
1567 void nir_lower_locals_to_regs(nir_shader
*shader
);
1569 void nir_lower_io(nir_shader
*shader
);
1571 void nir_lower_vars_to_ssa(nir_shader
*shader
);
1573 void nir_remove_dead_variables(nir_shader
*shader
);
1575 void nir_lower_vec_to_movs(nir_shader
*shader
);
1576 void nir_lower_alu_to_scalar(nir_shader
*shader
);
1578 void nir_lower_phis_to_scalar(nir_shader
*shader
);
1580 void nir_lower_samplers(nir_shader
*shader
,
1581 struct gl_shader_program
*shader_program
,
1582 struct gl_program
*prog
);
1584 void nir_lower_system_values(nir_shader
*shader
);
1586 void nir_lower_atomics(nir_shader
*shader
);
1587 void nir_lower_to_source_mods(nir_shader
*shader
);
1589 void nir_live_variables_impl(nir_function_impl
*impl
);
1590 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
1592 void nir_convert_to_ssa_impl(nir_function_impl
*impl
);
1593 void nir_convert_to_ssa(nir_shader
*shader
);
1594 void nir_convert_from_ssa(nir_shader
*shader
);
1596 bool nir_opt_algebraic(nir_shader
*shader
);
1597 bool nir_opt_constant_folding(nir_shader
*shader
);
1599 bool nir_opt_global_to_local(nir_shader
*shader
);
1601 bool nir_copy_prop_impl(nir_function_impl
*impl
);
1602 bool nir_copy_prop(nir_shader
*shader
);
1604 bool nir_opt_cse(nir_shader
*shader
);
1606 bool nir_opt_dce_impl(nir_function_impl
*impl
);
1607 bool nir_opt_dce(nir_shader
*shader
);
1609 void nir_opt_gcm(nir_shader
*shader
);
1611 bool nir_opt_peephole_select(nir_shader
*shader
);
1612 bool nir_opt_peephole_ffma(nir_shader
*shader
);
1614 bool nir_opt_remove_phis(nir_shader
*shader
);