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 "main/mtypes.h"
36 #include "main/bitset.h"
37 #include "nir_types.h"
40 #include "nir_opcodes.h"
47 #define NIR_TRUE (~0u)
49 /** Defines a cast function
51 * This macro defines a cast function from in_type to out_type where
52 * out_type is some structure type that contains a field of type out_type.
54 * Note that you have to be a bit careful as the generated cast function
57 #define NIR_DEFINE_CAST(name, in_type, out_type, field) \
58 static inline out_type * \
59 name(const in_type *parent) \
61 return exec_node_data(out_type, parent, field); \
64 struct nir_function_overload
;
70 * Description of built-in state associated with a uniform
72 * \sa nir_variable::state_slots
89 * Data stored in an nir_constant
91 union nir_constant_data
{
98 typedef struct nir_constant
{
100 * Value of the constant.
102 * The field used to back the values supplied by the constant is determined
103 * by the type associated with the \c nir_variable. Constants may be
104 * scalars, vectors, or matrices.
106 union nir_constant_data value
;
108 /* Array elements / Structure Fields */
109 struct nir_constant
**elements
;
113 * \brief Layout qualifiers for gl_FragDepth.
115 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
116 * with a layout qualifier.
119 nir_depth_layout_none
, /**< No depth layout is specified. */
120 nir_depth_layout_any
,
121 nir_depth_layout_greater
,
122 nir_depth_layout_less
,
123 nir_depth_layout_unchanged
127 * Either a uniform, global variable, shader input, or shader output. Based on
128 * ir_variable - it should be easy to translate between the two.
132 struct exec_node node
;
135 * Declared type of the variable
137 const struct glsl_type
*type
;
140 * Declared name of the variable
145 * For variables which satisfy the is_interface_instance() predicate, this
146 * points to an array of integers such that if the ith member of the
147 * interface block is an array, max_ifc_array_access[i] is the maximum
148 * array element of that member that has been accessed. If the ith member
149 * of the interface block is not an array, max_ifc_array_access[i] is
152 * For variables whose type is not an interface block, this pointer is
155 unsigned *max_ifc_array_access
;
157 struct nir_variable_data
{
160 * Is the variable read-only?
162 * This is set for variables declared as \c const, shader inputs,
165 unsigned read_only
:1;
168 unsigned invariant
:1;
171 * Storage class of the variable.
173 * \sa nir_variable_mode
175 nir_variable_mode mode
:4;
178 * Interpolation mode for shader inputs / outputs
180 * \sa glsl_interp_qualifier
182 unsigned interpolation
:2;
185 * \name ARB_fragment_coord_conventions
188 unsigned origin_upper_left
:1;
189 unsigned pixel_center_integer
:1;
193 * Was the location explicitly set in the shader?
195 * If the location is explicitly set in the shader, it \b cannot be changed
196 * by the linker or by the API (e.g., calls to \c glBindAttribLocation have
199 unsigned explicit_location
:1;
200 unsigned explicit_index
:1;
203 * Was an initial binding explicitly set in the shader?
205 * If so, constant_initializer contains an integer nir_constant
206 * representing the initial binding point.
208 unsigned explicit_binding
:1;
211 * Does this variable have an initializer?
213 * This is used by the linker to cross-validiate initializers of global
216 unsigned has_initializer
:1;
219 * Is this variable a generic output or input that has not yet been matched
220 * up to a variable in another stage of the pipeline?
222 * This is used by the linker as scratch storage while assigning locations
223 * to generic inputs and outputs.
225 unsigned is_unmatched_generic_inout
:1;
228 * If non-zero, then this variable may be packed along with other variables
229 * into a single varying slot, so this offset should be applied when
230 * accessing components. For example, an offset of 1 means that the x
231 * component of this variable is actually stored in component y of the
232 * location specified by \c location.
234 unsigned location_frac
:2;
237 * Non-zero if this variable was created by lowering a named interface
238 * block which was not an array.
240 * Note that this variable and \c from_named_ifc_block_array will never
243 unsigned from_named_ifc_block_nonarray
:1;
246 * Non-zero if this variable was created by lowering a named interface
247 * block which was an array.
249 * Note that this variable and \c from_named_ifc_block_nonarray will never
252 unsigned from_named_ifc_block_array
:1;
255 * \brief Layout qualifier for gl_FragDepth.
257 * This is not equal to \c ir_depth_layout_none if and only if this
258 * variable is \c gl_FragDepth and a layout qualifier is specified.
260 nir_depth_layout depth_layout
;
263 * Storage location of the base of this variable
265 * The precise meaning of this field depends on the nature of the variable.
267 * - Vertex shader input: one of the values from \c gl_vert_attrib.
268 * - Vertex shader output: one of the values from \c gl_varying_slot.
269 * - Geometry shader input: one of the values from \c gl_varying_slot.
270 * - Geometry shader output: one of the values from \c gl_varying_slot.
271 * - Fragment shader input: one of the values from \c gl_varying_slot.
272 * - Fragment shader output: one of the values from \c gl_frag_result.
273 * - Uniforms: Per-stage uniform slot number for default uniform block.
274 * - Uniforms: Index within the uniform block definition for UBO members.
275 * - Other: This field is not currently used.
277 * If the variable is a uniform, shader input, or shader output, and the
278 * slot has not been assigned, the value will be -1.
283 * The actual location of the variable in the IR. Only valid for inputs
286 unsigned int driver_location
;
289 * output index for dual source blending.
294 * Initial binding point for a sampler or UBO.
296 * For array types, this represents the binding point for the first element.
301 * Location an atomic counter is stored at.
304 unsigned buffer_index
;
309 * ARB_shader_image_load_store qualifiers.
312 bool read_only
; /**< "readonly" qualifier. */
313 bool write_only
; /**< "writeonly" qualifier. */
318 /** Image internal format if specified explicitly, otherwise GL_NONE. */
323 * Highest element accessed with a constant expression array index
325 * Not used for non-array variables.
327 unsigned max_array_access
;
332 * Built-in state that backs this uniform
334 * Once set at variable creation, \c state_slots must remain invariant.
335 * This is because, ideally, this array would be shared by all clones of
336 * this variable in the IR tree. In other words, we'd really like for it
337 * to be a fly-weight.
339 * If the variable is not a uniform, \c num_state_slots will be zero and
340 * \c state_slots will be \c NULL.
343 unsigned num_state_slots
; /**< Number of state slots used */
344 nir_state_slot
*state_slots
; /**< State descriptors. */
348 * Constant expression assigned in the initializer of the variable
350 nir_constant
*constant_initializer
;
353 * For variables that are in an interface block or are an instance of an
354 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
356 * \sa ir_variable::location
358 const struct glsl_type
*interface_type
;
362 struct exec_node node
;
364 unsigned num_components
; /** < number of vector components */
365 unsigned num_array_elems
; /** < size of array (0 for no array) */
367 /** generic register index. */
370 /** only for debug purposes, can be NULL */
373 /** whether this register is local (per-function) or global (per-shader) */
377 * If this flag is set to true, then accessing channels >= num_components
378 * is well-defined, and simply spills over to the next array element. This
379 * is useful for backends that can do per-component accessing, in
380 * particular scalar backends. By setting this flag and making
381 * num_components equal to 1, structures can be packed tightly into
382 * registers and then registers can be accessed per-component to get to
383 * each structure member, even if it crosses vec4 boundaries.
387 /** set of nir_instr's where this register is used (read from) */
390 /** set of nir_instr's where this register is defined (written to) */
393 /** set of nir_if's where this register is used as a condition */
401 nir_instr_type_intrinsic
,
402 nir_instr_type_load_const
,
404 nir_instr_type_ssa_undef
,
406 nir_instr_type_parallel_copy
,
410 struct exec_node node
;
412 struct nir_block
*block
;
414 /* flag for dead code elimination (see nir_opt_dce.c) */
418 static inline nir_instr
*
419 nir_instr_next(const nir_instr
*instr
)
421 return exec_node_data(nir_instr
, (instr
)->node
.next
, node
);
424 static inline nir_instr
*
425 nir_instr_prev(const nir_instr
*instr
)
427 return exec_node_data(nir_instr
, (instr
)->node
.prev
, node
);
431 /** for debugging only, can be NULL */
434 /** generic SSA definition index. */
437 /** Index into the live_in and live_out bitfields */
440 nir_instr
*parent_instr
;
442 /** set of nir_instr's where this register is used (read from) */
445 /** set of nir_if's where this register is used as a condition */
448 uint8_t num_components
;
455 struct nir_src
*indirect
; /** < NULL for no indirect offset */
456 unsigned base_offset
;
458 /* TODO use-def chain goes here */
463 struct nir_src
*indirect
; /** < NULL for no indirect offset */
464 unsigned base_offset
;
466 /* TODO def-use chain goes here */
469 typedef struct nir_src
{
487 static inline nir_src
488 nir_src_for_ssa(nir_ssa_def
*def
)
498 static inline nir_src
499 nir_src_for_reg(nir_register
*reg
)
505 src
.reg
.indirect
= NULL
;
506 src
.reg
.base_offset
= 0;
511 static inline nir_dest
512 nir_dest_for_reg(nir_register
*reg
)
518 dest
.reg
.indirect
= NULL
;
519 dest
.reg
.base_offset
= 0;
524 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *mem_ctx
);
525 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, void *mem_ctx
);
531 * \name input modifiers
535 * For inputs interpreted as floating point, flips the sign bit. For
536 * inputs interpreted as integers, performs the two's complement negation.
541 * Clears the sign bit for floating point values, and computes the integer
542 * absolute value for integers. Note that the negate modifier acts after
543 * the absolute value modifier, therefore if both are set then all inputs
544 * will become negative.
550 * For each input component, says which component of the register it is
551 * chosen from. Note that which elements of the swizzle are used and which
552 * are ignored are based on the write mask for most opcodes - for example,
553 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
554 * a swizzle of {2, x, 1, 0} where x means "don't care."
563 * \name saturate output modifier
565 * Only valid for opcodes that output floating-point numbers. Clamps the
566 * output to between 0.0 and 1.0 inclusive.
571 unsigned write_mask
: 4; /* ignored if dest.is_ssa is true */
574 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
, void *mem_ctx
);
575 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
579 nir_type_invalid
= 0, /* Not a valid type */
587 NIR_OP_IS_COMMUTATIVE
= (1 << 0),
588 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
589 } nir_op_algebraic_property
;
597 * The number of components in the output
599 * If non-zero, this is the size of the output and input sizes are
600 * explicitly given; swizzle and writemask are still in effect, but if
601 * the output component is masked out, then the input component may
604 * If zero, the opcode acts in the standard, per-component manner; the
605 * operation is performed on each component (except the ones that are
606 * masked out) with the input being taken from the input swizzle for
609 * The size of some of the inputs may be given (i.e. non-zero) even
610 * though output_size is zero; in that case, the inputs with a zero
611 * size act per-component, while the inputs with non-zero size don't.
613 unsigned output_size
;
616 * The type of vector that the instruction outputs. Note that the
617 * staurate modifier is only allowed on outputs with the float type.
620 nir_alu_type output_type
;
623 * The number of components in each input
625 unsigned input_sizes
[4];
628 * The type of vector that each input takes. Note that negate and
629 * absolute value are only allowed on inputs with int or float type and
630 * behave differently on the two.
632 nir_alu_type input_types
[4];
634 nir_op_algebraic_property algebraic_properties
;
637 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
639 typedef struct nir_alu_instr
{
646 /* is this source channel used? */
648 nir_alu_instr_channel_used(nir_alu_instr
*instr
, unsigned src
, unsigned channel
)
650 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
651 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
653 return (instr
->dest
.write_mask
>> channel
) & 1;
657 * For instructions whose destinations are SSA, get the number of channels
660 static inline unsigned
661 nir_ssa_alu_instr_src_components(nir_alu_instr
*instr
, unsigned src
)
663 assert(instr
->dest
.dest
.is_ssa
);
665 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
666 return nir_op_infos
[instr
->op
].input_sizes
[src
];
668 return instr
->dest
.dest
.ssa
.num_components
;
673 nir_deref_type_array
,
674 nir_deref_type_struct
677 typedef struct nir_deref
{
678 nir_deref_type deref_type
;
679 struct nir_deref
*child
;
680 const struct glsl_type
*type
;
689 /* This enum describes how the array is referenced. If the deref is
690 * direct then the base_offset is used. If the deref is indirect then then
691 * offset is given by base_offset + indirect. If the deref is a wildcard
692 * then the deref refers to all of the elements of the array at the same
693 * time. Wildcard dereferences are only ever allowed in copy_var
694 * intrinsics and the source and destination derefs must have matching
698 nir_deref_array_type_direct
,
699 nir_deref_array_type_indirect
,
700 nir_deref_array_type_wildcard
,
701 } nir_deref_array_type
;
706 nir_deref_array_type deref_array_type
;
707 unsigned base_offset
;
717 NIR_DEFINE_CAST(nir_deref_as_var
, nir_deref
, nir_deref_var
, deref
)
718 NIR_DEFINE_CAST(nir_deref_as_array
, nir_deref
, nir_deref_array
, deref
)
719 NIR_DEFINE_CAST(nir_deref_as_struct
, nir_deref
, nir_deref_struct
, deref
)
725 nir_deref_var
**params
;
726 nir_deref_var
*return_deref
;
728 struct nir_function_overload
*callee
;
731 #define INTRINSIC(name, num_srcs, src_components, has_dest, dest_components, \
732 num_variables, num_indices, flags) \
733 nir_intrinsic_##name,
735 #define LAST_INTRINSIC(name) nir_last_intrinsic = nir_intrinsic_##name,
738 #include "nir_intrinsics.h"
739 nir_num_intrinsics
= nir_last_intrinsic
+ 1
743 #undef LAST_INTRINSIC
745 /** Represents an intrinsic
747 * An intrinsic is an instruction type for handling things that are
748 * more-or-less regular operations but don't just consume and produce SSA
749 * values like ALU operations do. Intrinsics are not for things that have
750 * special semantic meaning such as phi nodes and parallel copies.
751 * Examples of intrinsics include variable load/store operations, system
752 * value loads, and the like. Even though texturing more-or-less falls
753 * under this category, texturing is its own instruction type because
754 * trying to represent texturing with intrinsics would lead to a
755 * combinatorial explosion of intrinsic opcodes.
757 * By having a single instruction type for handling a lot of different
758 * cases, optimization passes can look for intrinsics and, for the most
759 * part, completely ignore them. Each intrinsic type also has a few
760 * possible flags that govern whether or not they can be reordered or
761 * eliminated. That way passes like dead code elimination can still work
762 * on intrisics without understanding the meaning of each.
764 * Each intrinsic has some number of constant indices, some number of
765 * variables, and some number of sources. What these sources, variables,
766 * and indices mean depends on the intrinsic and is documented with the
767 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
768 * instructions are the only types of instruction that can operate on
774 nir_intrinsic_op intrinsic
;
778 /** number of components if this is a vectorized intrinsic
780 * Similarly to ALU operations, some intrinsics are vectorized.
781 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
782 * For vectorized intrinsics, the num_components field specifies the
783 * number of destination components and the number of source components
784 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
786 uint8_t num_components
;
790 nir_deref_var
*variables
[2];
793 } nir_intrinsic_instr
;
796 * \name NIR intrinsics semantic flags
798 * information about what the compiler can do with the intrinsics.
800 * \sa nir_intrinsic_info::flags
804 * whether the intrinsic can be safely eliminated if none of its output
805 * value is not being used.
807 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
810 * Whether the intrinsic can be reordered with respect to any other
811 * intrinsic, i.e. whether the only reordering dependencies of the
812 * intrinsic are due to the register reads/writes.
814 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
815 } nir_intrinsic_semantic_flag
;
817 #define NIR_INTRINSIC_MAX_INPUTS 4
822 unsigned num_srcs
; /** < number of register/SSA inputs */
824 /** number of components of each input register
826 * If this value is 0, the number of components is given by the
827 * num_components field of nir_intrinsic_instr.
829 unsigned src_components
[NIR_INTRINSIC_MAX_INPUTS
];
833 /** number of components of the output register
835 * If this value is 0, the number of components is given by the
836 * num_components field of nir_intrinsic_instr.
838 unsigned dest_components
;
840 /** the number of inputs/outputs that are variables */
841 unsigned num_variables
;
843 /** the number of constant indices used by the intrinsic */
844 unsigned num_indices
;
846 /** semantic flags for calls to this intrinsic */
847 nir_intrinsic_semantic_flag flags
;
848 } nir_intrinsic_info
;
850 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
853 * \group texture information
855 * This gives semantic information about textures which is useful to the
856 * frontend, the backend, and lowering passes, but not the optimizer.
861 nir_tex_src_projector
,
862 nir_tex_src_comparitor
, /* shadow comparitor */
866 nir_tex_src_ms_index
, /* MSAA sample index */
869 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
870 nir_num_tex_src_types
875 nir_tex_src_type src_type
;
879 nir_texop_tex
, /**< Regular texture look-up */
880 nir_texop_txb
, /**< Texture look-up with LOD bias */
881 nir_texop_txl
, /**< Texture look-up with explicit LOD */
882 nir_texop_txd
, /**< Texture look-up with partial derivatvies */
883 nir_texop_txf
, /**< Texel fetch with explicit LOD */
884 nir_texop_txf_ms
, /**< Multisample texture fetch */
885 nir_texop_txs
, /**< Texture size */
886 nir_texop_lod
, /**< Texture lod query */
887 nir_texop_tg4
, /**< Texture gather */
888 nir_texop_query_levels
/**< Texture levels query */
894 enum glsl_sampler_dim sampler_dim
;
895 nir_alu_type dest_type
;
900 unsigned num_srcs
, coord_components
;
901 bool is_array
, is_shadow
;
904 * If is_shadow is true, whether this is the old-style shadow that outputs 4
905 * components or the new-style shadow that outputs 1 component.
907 bool is_new_style_shadow
;
909 /* constant offset - must be 0 if the offset source is used */
912 /* gather component selector */
913 unsigned component
: 2;
915 /** The sampler index
917 * If this texture instruction has a nir_tex_src_sampler_offset source,
918 * then the sampler index is given by sampler_index + sampler_offset.
920 unsigned sampler_index
;
922 /** The size of the sampler array or 0 if it's not an array */
923 unsigned sampler_array_size
;
925 nir_deref_var
*sampler
; /* if this is NULL, use sampler_index instead */
928 static inline unsigned
929 nir_tex_instr_dest_size(nir_tex_instr
*instr
)
931 if (instr
->op
== nir_texop_txs
) {
933 switch (instr
->sampler_dim
) {
934 case GLSL_SAMPLER_DIM_1D
:
935 case GLSL_SAMPLER_DIM_BUF
:
938 case GLSL_SAMPLER_DIM_2D
:
939 case GLSL_SAMPLER_DIM_CUBE
:
940 case GLSL_SAMPLER_DIM_MS
:
941 case GLSL_SAMPLER_DIM_RECT
:
942 case GLSL_SAMPLER_DIM_EXTERNAL
:
945 case GLSL_SAMPLER_DIM_3D
:
949 unreachable("not reached");
956 if (instr
->op
== nir_texop_query_levels
)
959 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
965 static inline unsigned
966 nir_tex_instr_src_size(nir_tex_instr
*instr
, unsigned src
)
968 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
969 return instr
->coord_components
;
972 if (instr
->src
[src
].src_type
== nir_tex_src_offset
||
973 instr
->src
[src
].src_type
== nir_tex_src_ddx
||
974 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
976 return instr
->coord_components
- 1;
978 return instr
->coord_components
;
985 nir_tex_instr_src_index(nir_tex_instr
*instr
, nir_tex_src_type type
)
987 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
988 if (instr
->src
[i
].src_type
== type
)
1005 nir_const_value value
;
1008 } nir_load_const_instr
;
1021 /* creates a new SSA variable in an undefined state */
1026 } nir_ssa_undef_instr
;
1029 struct exec_node node
;
1031 /* The predecessor block corresponding to this source */
1032 struct nir_block
*pred
;
1037 #define nir_foreach_phi_src(phi, entry) \
1038 foreach_list_typed(nir_phi_src, entry, node, &(phi)->srcs)
1043 struct exec_list srcs
; /** < list of nir_phi_src */
1049 struct exec_node node
;
1052 } nir_parallel_copy_entry
;
1054 #define nir_foreach_parallel_copy_entry(pcopy, entry) \
1055 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
1060 /* A list of nir_parallel_copy_entry's. The sources of all of the
1061 * entries are copied to the corresponding destinations "in parallel".
1062 * In other words, if we have two entries: a -> b and b -> a, the values
1065 struct exec_list entries
;
1066 } nir_parallel_copy_instr
;
1068 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
)
1069 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
)
1070 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
)
1071 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
)
1072 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
)
1073 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
)
1074 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
)
1075 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
)
1076 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
1077 nir_parallel_copy_instr
, instr
)
1082 * Control flow consists of a tree of control flow nodes, which include
1083 * if-statements and loops. The leaves of the tree are basic blocks, lists of
1084 * instructions that always run start-to-finish. Each basic block also keeps
1085 * track of its successors (blocks which may run immediately after the current
1086 * block) and predecessors (blocks which could have run immediately before the
1087 * current block). Each function also has a start block and an end block which
1088 * all return statements point to (which is always empty). Together, all the
1089 * blocks with their predecessors and successors make up the control flow
1090 * graph (CFG) of the function. There are helpers that modify the tree of
1091 * control flow nodes while modifying the CFG appropriately; these should be
1092 * used instead of modifying the tree directly.
1099 nir_cf_node_function
1102 typedef struct nir_cf_node
{
1103 struct exec_node node
;
1104 nir_cf_node_type type
;
1105 struct nir_cf_node
*parent
;
1108 typedef struct nir_block
{
1109 nir_cf_node cf_node
;
1111 struct exec_list instr_list
; /** < list of nir_instr */
1113 /** generic block index; generated by nir_index_blocks */
1117 * Each block can only have up to 2 successors, so we put them in a simple
1118 * array - no need for anything more complicated.
1120 struct nir_block
*successors
[2];
1122 /* Set of nir_block predecessors in the CFG */
1123 struct set
*predecessors
;
1126 * this node's immediate dominator in the dominance tree - set to NULL for
1129 struct nir_block
*imm_dom
;
1131 /* This node's children in the dominance tree */
1132 unsigned num_dom_children
;
1133 struct nir_block
**dom_children
;
1135 /* Set of nir_block's on the dominance frontier of this block */
1136 struct set
*dom_frontier
;
1138 /* live in and out for this block; used for liveness analysis */
1139 BITSET_WORD
*live_in
;
1140 BITSET_WORD
*live_out
;
1143 static inline nir_instr
*
1144 nir_block_first_instr(nir_block
*block
)
1146 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
1147 return exec_node_data(nir_instr
, head
, node
);
1150 static inline nir_instr
*
1151 nir_block_last_instr(nir_block
*block
)
1153 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
1154 return exec_node_data(nir_instr
, tail
, node
);
1157 #define nir_foreach_instr(block, instr) \
1158 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
1159 #define nir_foreach_instr_reverse(block, instr) \
1160 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
1161 #define nir_foreach_instr_safe(block, instr) \
1162 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
1165 nir_cf_node cf_node
;
1168 struct exec_list then_list
; /** < list of nir_cf_node */
1169 struct exec_list else_list
; /** < list of nir_cf_node */
1172 static inline nir_cf_node
*
1173 nir_if_first_then_node(nir_if
*if_stmt
)
1175 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
1176 return exec_node_data(nir_cf_node
, head
, node
);
1179 static inline nir_cf_node
*
1180 nir_if_last_then_node(nir_if
*if_stmt
)
1182 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
1183 return exec_node_data(nir_cf_node
, tail
, node
);
1186 static inline nir_cf_node
*
1187 nir_if_first_else_node(nir_if
*if_stmt
)
1189 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
1190 return exec_node_data(nir_cf_node
, head
, node
);
1193 static inline nir_cf_node
*
1194 nir_if_last_else_node(nir_if
*if_stmt
)
1196 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
1197 return exec_node_data(nir_cf_node
, tail
, node
);
1201 nir_cf_node cf_node
;
1203 struct exec_list body
; /** < list of nir_cf_node */
1206 static inline nir_cf_node
*
1207 nir_loop_first_cf_node(nir_loop
*loop
)
1209 return exec_node_data(nir_cf_node
, exec_list_get_head(&loop
->body
), node
);
1212 static inline nir_cf_node
*
1213 nir_loop_last_cf_node(nir_loop
*loop
)
1215 return exec_node_data(nir_cf_node
, exec_list_get_tail(&loop
->body
), node
);
1219 * Various bits of metadata that can may be created or required by
1220 * optimization and analysis passes
1223 nir_metadata_none
= 0x0,
1224 nir_metadata_block_index
= 0x1,
1225 nir_metadata_dominance
= 0x2,
1226 nir_metadata_live_variables
= 0x4,
1230 nir_cf_node cf_node
;
1232 /** pointer to the overload of which this is an implementation */
1233 struct nir_function_overload
*overload
;
1235 struct exec_list body
; /** < list of nir_cf_node */
1237 nir_block
*start_block
, *end_block
;
1239 /** list for all local variables in the function */
1240 struct exec_list locals
;
1242 /** array of variables used as parameters */
1243 unsigned num_params
;
1244 nir_variable
**params
;
1246 /** variable used to hold the result of the function */
1247 nir_variable
*return_var
;
1249 /** list of local registers in the function */
1250 struct exec_list registers
;
1252 /** next available local register index */
1255 /** next available SSA value index */
1258 /* total number of basic blocks, only valid when block_index_dirty = false */
1259 unsigned num_blocks
;
1261 nir_metadata valid_metadata
;
1262 } nir_function_impl
;
1264 static inline nir_cf_node
*
1265 nir_cf_node_next(nir_cf_node
*node
)
1267 return exec_node_data(nir_cf_node
, exec_node_get_next(&node
->node
), node
);
1270 static inline nir_cf_node
*
1271 nir_cf_node_prev(nir_cf_node
*node
)
1273 return exec_node_data(nir_cf_node
, exec_node_get_prev(&node
->node
), node
);
1277 nir_cf_node_is_first(const nir_cf_node
*node
)
1279 return exec_node_is_head_sentinel(node
->node
.prev
);
1283 nir_cf_node_is_last(const nir_cf_node
*node
)
1285 return exec_node_is_tail_sentinel(node
->node
.next
);
1288 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
)
1289 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
)
1290 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
)
1291 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
, nir_function_impl
, cf_node
)
1296 nir_parameter_inout
,
1297 } nir_parameter_type
;
1300 nir_parameter_type param_type
;
1301 const struct glsl_type
*type
;
1304 typedef struct nir_function_overload
{
1305 struct exec_node node
;
1307 unsigned num_params
;
1308 nir_parameter
*params
;
1309 const struct glsl_type
*return_type
;
1311 nir_function_impl
*impl
; /** < NULL if the overload is only declared yet */
1313 /** pointer to the function of which this is an overload */
1314 struct nir_function
*function
;
1315 } nir_function_overload
;
1317 typedef struct nir_function
{
1318 struct exec_node node
;
1320 struct exec_list overload_list
; /** < list of nir_function_overload */
1322 struct nir_shader
*shader
;
1325 #define nir_function_first_overload(func) \
1326 exec_node_data(nir_function_overload, \
1327 exec_list_get_head(&(func)->overload_list), node)
1329 typedef struct nir_shader
{
1330 /** hash table of name -> uniform nir_variable */
1331 struct hash_table
*uniforms
;
1333 /** hash table of name -> input nir_variable */
1334 struct hash_table
*inputs
;
1336 /** hash table of name -> output nir_variable */
1337 struct hash_table
*outputs
;
1339 /** list of global variables in the shader */
1340 struct exec_list globals
;
1342 /** list of system value variables in the shader */
1343 struct exec_list system_values
;
1345 struct exec_list functions
; /** < list of nir_function */
1347 /** list of global register in the shader */
1348 struct exec_list registers
;
1350 /** structures used in this shader */
1351 unsigned num_user_structures
;
1352 struct glsl_type
**user_structures
;
1354 /** next available global register index */
1358 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
1361 unsigned num_inputs
, num_uniforms
, num_outputs
;
1364 #define nir_foreach_overload(shader, overload) \
1365 foreach_list_typed(nir_function, func, node, &(shader)->functions) \
1366 foreach_list_typed(nir_function_overload, overload, node, \
1367 &(func)->overload_list)
1369 nir_shader
*nir_shader_create(void *mem_ctx
);
1371 /** creates a register, including assigning it an index and adding it to the list */
1372 nir_register
*nir_global_reg_create(nir_shader
*shader
);
1374 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
1376 void nir_reg_remove(nir_register
*reg
);
1378 /** creates a function and adds it to the shader's list of functions */
1379 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
1381 /** creates a null function returning null */
1382 nir_function_overload
*nir_function_overload_create(nir_function
*func
);
1384 nir_function_impl
*nir_function_impl_create(nir_function_overload
*func
);
1386 nir_block
*nir_block_create(void *mem_ctx
);
1387 nir_if
*nir_if_create(void *mem_ctx
);
1388 nir_loop
*nir_loop_create(void *mem_ctx
);
1390 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
1392 /** puts a control flow node immediately after another control flow node */
1393 void nir_cf_node_insert_after(nir_cf_node
*node
, nir_cf_node
*after
);
1395 /** puts a control flow node immediately before another control flow node */
1396 void nir_cf_node_insert_before(nir_cf_node
*node
, nir_cf_node
*before
);
1398 /** puts a control flow node at the beginning of a list from an if, loop, or function */
1399 void nir_cf_node_insert_begin(struct exec_list
*list
, nir_cf_node
*node
);
1401 /** puts a control flow node at the end of a list from an if, loop, or function */
1402 void nir_cf_node_insert_end(struct exec_list
*list
, nir_cf_node
*node
);
1404 /** removes a control flow node, doing any cleanup necessary */
1405 void nir_cf_node_remove(nir_cf_node
*node
);
1407 /** requests that the given pieces of metadata be generated */
1408 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
);
1409 /** dirties all but the preserved metadata */
1410 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
1412 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
1413 nir_alu_instr
*nir_alu_instr_create(void *mem_ctx
, nir_op op
);
1415 nir_jump_instr
*nir_jump_instr_create(void *mem_ctx
, nir_jump_type type
);
1417 nir_load_const_instr
*nir_load_const_instr_create(void *mem_ctx
,
1418 unsigned num_components
);
1420 nir_intrinsic_instr
*nir_intrinsic_instr_create(void *mem_ctx
,
1421 nir_intrinsic_op op
);
1423 nir_call_instr
*nir_call_instr_create(void *mem_ctx
,
1424 nir_function_overload
*callee
);
1426 nir_tex_instr
*nir_tex_instr_create(void *mem_ctx
, unsigned num_srcs
);
1428 nir_phi_instr
*nir_phi_instr_create(void *mem_ctx
);
1430 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(void *mem_ctx
);
1432 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(void *mem_ctx
,
1433 unsigned num_components
);
1435 nir_deref_var
*nir_deref_var_create(void *mem_ctx
, nir_variable
*var
);
1436 nir_deref_array
*nir_deref_array_create(void *mem_ctx
);
1437 nir_deref_struct
*nir_deref_struct_create(void *mem_ctx
, unsigned field_index
);
1439 nir_deref
*nir_copy_deref(void *mem_ctx
, nir_deref
*deref
);
1441 void nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
);
1442 void nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
);
1444 void nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
);
1445 void nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
);
1447 void nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
);
1448 void nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
);
1450 void nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
);
1451 void nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
);
1453 void nir_instr_remove(nir_instr
*instr
);
1455 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
1456 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
1457 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
1458 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
1460 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
1461 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
1463 nir_const_value
*nir_src_as_const_value(nir_src src
);
1464 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
1465 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
1467 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
1468 unsigned num_components
, const char *name
);
1469 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
1470 unsigned num_components
, const char *name
);
1471 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
, void *mem_ctx
);
1473 /* visits basic blocks in source-code order */
1474 typedef bool (*nir_foreach_block_cb
)(nir_block
*block
, void *state
);
1475 bool nir_foreach_block(nir_function_impl
*impl
, nir_foreach_block_cb cb
,
1477 bool nir_foreach_block_reverse(nir_function_impl
*impl
, nir_foreach_block_cb cb
,
1480 /* If the following CF node is an if, this function returns that if.
1481 * Otherwise, it returns NULL.
1483 nir_if
*nir_block_get_following_if(nir_block
*block
);
1485 void nir_index_local_regs(nir_function_impl
*impl
);
1486 void nir_index_global_regs(nir_shader
*shader
);
1487 void nir_index_ssa_defs(nir_function_impl
*impl
);
1489 void nir_index_blocks(nir_function_impl
*impl
);
1491 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
1492 void nir_print_instr(nir_instr
*instr
, FILE *fp
);
1495 void nir_validate_shader(nir_shader
*shader
);
1497 static inline void nir_validate_shader(nir_shader
*shader
) { }
1500 void nir_calc_dominance_impl(nir_function_impl
*impl
);
1501 void nir_calc_dominance(nir_shader
*shader
);
1503 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
1504 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
1506 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
1507 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
1509 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
1510 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
1512 void nir_split_var_copies(nir_shader
*shader
);
1514 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, void *mem_ctx
);
1515 void nir_lower_var_copies(nir_shader
*shader
);
1517 void nir_lower_global_vars_to_local(nir_shader
*shader
);
1519 void nir_lower_locals_to_regs(nir_shader
*shader
);
1521 void nir_lower_io(nir_shader
*shader
);
1523 void nir_lower_vars_to_ssa(nir_shader
*shader
);
1525 void nir_remove_dead_variables(nir_shader
*shader
);
1527 void nir_lower_vec_to_movs(nir_shader
*shader
);
1528 void nir_lower_alu_to_scalar(nir_shader
*shader
);
1530 void nir_lower_phis_to_scalar(nir_shader
*shader
);
1532 void nir_lower_samplers(nir_shader
*shader
,
1533 struct gl_shader_program
*shader_program
,
1534 struct gl_program
*prog
);
1536 void nir_lower_system_values(nir_shader
*shader
);
1538 void nir_lower_atomics(nir_shader
*shader
);
1539 void nir_lower_to_source_mods(nir_shader
*shader
);
1541 void nir_live_variables_impl(nir_function_impl
*impl
);
1542 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
1544 void nir_convert_to_ssa_impl(nir_function_impl
*impl
);
1545 void nir_convert_to_ssa(nir_shader
*shader
);
1546 void nir_convert_from_ssa(nir_shader
*shader
);
1548 bool nir_opt_algebraic(nir_shader
*shader
);
1549 bool nir_opt_constant_folding(nir_shader
*shader
);
1551 bool nir_opt_global_to_local(nir_shader
*shader
);
1553 bool nir_copy_prop_impl(nir_function_impl
*impl
);
1554 bool nir_copy_prop(nir_shader
*shader
);
1556 bool nir_opt_cse(nir_shader
*shader
);
1558 bool nir_opt_dce_impl(nir_function_impl
*impl
);
1559 bool nir_opt_dce(nir_shader
*shader
);
1561 bool nir_opt_peephole_select(nir_shader
*shader
);
1562 bool nir_opt_peephole_ffma(nir_shader
*shader
);