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"
33 #include "GL/gl.h" /* GLenum */
34 #include "util/ralloc.h"
35 #include "main/mtypes.h"
36 #include "main/bitset.h"
37 #include "nir_types.h"
45 #define NIR_TRUE (~0u)
47 /** Defines a cast function
49 * This macro defines a cast function from in_type to out_type where
50 * out_type is some structure type that contains a field of type out_type.
52 * Note that you have to be a bit careful as the generated cast function
55 #define NIR_DEFINE_CAST(name, in_type, out_type, field) \
56 static inline out_type * \
57 name(const in_type *parent) \
59 return exec_node_data(out_type, parent, field); \
62 struct nir_function_overload
;
68 * Description of built-in state associated with a uniform
70 * \sa nir_variable::state_slots
87 * Data stored in an nir_constant
89 union nir_constant_data
{
96 typedef struct nir_constant
{
98 * Value of the constant.
100 * The field used to back the values supplied by the constant is determined
101 * by the type associated with the \c nir_variable. Constants may be
102 * scalars, vectors, or matrices.
104 union nir_constant_data value
;
106 /* Array elements / Structure Fields */
107 struct nir_constant
**elements
;
111 * \brief Layout qualifiers for gl_FragDepth.
113 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
114 * with a layout qualifier.
117 nir_depth_layout_none
, /**< No depth layout is specified. */
118 nir_depth_layout_any
,
119 nir_depth_layout_greater
,
120 nir_depth_layout_less
,
121 nir_depth_layout_unchanged
125 * Either a uniform, global variable, shader input, or shader output. Based on
126 * ir_variable - it should be easy to translate between the two.
130 struct exec_node node
;
133 * Declared type of the variable
135 const struct glsl_type
*type
;
138 * Declared name of the variable
143 * For variables which satisfy the is_interface_instance() predicate, this
144 * points to an array of integers such that if the ith member of the
145 * interface block is an array, max_ifc_array_access[i] is the maximum
146 * array element of that member that has been accessed. If the ith member
147 * of the interface block is not an array, max_ifc_array_access[i] is
150 * For variables whose type is not an interface block, this pointer is
153 unsigned *max_ifc_array_access
;
155 struct nir_variable_data
{
158 * Is the variable read-only?
160 * This is set for variables declared as \c const, shader inputs,
163 unsigned read_only
:1;
166 unsigned invariant
:1;
169 * Storage class of the variable.
171 * \sa nir_variable_mode
173 nir_variable_mode mode
:4;
176 * Interpolation mode for shader inputs / outputs
178 * \sa glsl_interp_qualifier
180 unsigned interpolation
:2;
183 * \name ARB_fragment_coord_conventions
186 unsigned origin_upper_left
:1;
187 unsigned pixel_center_integer
:1;
191 * Was the location explicitly set in the shader?
193 * If the location is explicitly set in the shader, it \b cannot be changed
194 * by the linker or by the API (e.g., calls to \c glBindAttribLocation have
197 unsigned explicit_location
:1;
198 unsigned explicit_index
:1;
201 * Was an initial binding explicitly set in the shader?
203 * If so, constant_initializer contains an integer nir_constant
204 * representing the initial binding point.
206 unsigned explicit_binding
:1;
209 * Does this variable have an initializer?
211 * This is used by the linker to cross-validiate initializers of global
214 unsigned has_initializer
:1;
217 * Is this variable a generic output or input that has not yet been matched
218 * up to a variable in another stage of the pipeline?
220 * This is used by the linker as scratch storage while assigning locations
221 * to generic inputs and outputs.
223 unsigned is_unmatched_generic_inout
:1;
226 * If non-zero, then this variable may be packed along with other variables
227 * into a single varying slot, so this offset should be applied when
228 * accessing components. For example, an offset of 1 means that the x
229 * component of this variable is actually stored in component y of the
230 * location specified by \c location.
232 unsigned location_frac
:2;
235 * Non-zero if this variable was created by lowering a named interface
236 * block which was not an array.
238 * Note that this variable and \c from_named_ifc_block_array will never
241 unsigned from_named_ifc_block_nonarray
:1;
244 * Non-zero if this variable was created by lowering a named interface
245 * block which was an array.
247 * Note that this variable and \c from_named_ifc_block_nonarray will never
250 unsigned from_named_ifc_block_array
:1;
253 * \brief Layout qualifier for gl_FragDepth.
255 * This is not equal to \c ir_depth_layout_none if and only if this
256 * variable is \c gl_FragDepth and a layout qualifier is specified.
258 nir_depth_layout depth_layout
;
261 * Storage location of the base of this variable
263 * The precise meaning of this field depends on the nature of the variable.
265 * - Vertex shader input: one of the values from \c gl_vert_attrib.
266 * - Vertex shader output: one of the values from \c gl_varying_slot.
267 * - Geometry shader input: one of the values from \c gl_varying_slot.
268 * - Geometry shader output: one of the values from \c gl_varying_slot.
269 * - Fragment shader input: one of the values from \c gl_varying_slot.
270 * - Fragment shader output: one of the values from \c gl_frag_result.
271 * - Uniforms: Per-stage uniform slot number for default uniform block.
272 * - Uniforms: Index within the uniform block definition for UBO members.
273 * - Other: This field is not currently used.
275 * If the variable is a uniform, shader input, or shader output, and the
276 * slot has not been assigned, the value will be -1.
281 * The actual location of the variable in the IR. Only valid for inputs
284 unsigned int driver_location
;
287 * output index for dual source blending.
292 * Initial binding point for a sampler or UBO.
294 * For array types, this represents the binding point for the first element.
299 * Location an atomic counter is stored at.
302 unsigned buffer_index
;
307 * ARB_shader_image_load_store qualifiers.
310 bool read_only
; /**< "readonly" qualifier. */
311 bool write_only
; /**< "writeonly" qualifier. */
316 /** Image internal format if specified explicitly, otherwise GL_NONE. */
321 * Highest element accessed with a constant expression array index
323 * Not used for non-array variables.
325 unsigned max_array_access
;
330 * Built-in state that backs this uniform
332 * Once set at variable creation, \c state_slots must remain invariant.
333 * This is because, ideally, this array would be shared by all clones of
334 * this variable in the IR tree. In other words, we'd really like for it
335 * to be a fly-weight.
337 * If the variable is not a uniform, \c num_state_slots will be zero and
338 * \c state_slots will be \c NULL.
341 unsigned num_state_slots
; /**< Number of state slots used */
342 nir_state_slot
*state_slots
; /**< State descriptors. */
346 * Constant expression assigned in the initializer of the variable
348 nir_constant
*constant_initializer
;
351 * For variables that are in an interface block or are an instance of an
352 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
354 * \sa ir_variable::location
356 const struct glsl_type
*interface_type
;
360 struct exec_node node
;
362 unsigned num_components
; /** < number of vector components */
363 unsigned num_array_elems
; /** < size of array (0 for no array) */
365 /** generic register index. */
368 /** only for debug purposes, can be NULL */
371 /** whether this register is local (per-function) or global (per-shader) */
375 * If this flag is set to true, then accessing channels >= num_components
376 * is well-defined, and simply spills over to the next array element. This
377 * is useful for backends that can do per-component accessing, in
378 * particular scalar backends. By setting this flag and making
379 * num_components equal to 1, structures can be packed tightly into
380 * registers and then registers can be accessed per-component to get to
381 * each structure member, even if it crosses vec4 boundaries.
385 /** set of nir_instr's where this register is used (read from) */
388 /** set of nir_instr's where this register is defined (written to) */
391 /** set of nir_if's where this register is used as a condition */
399 nir_instr_type_intrinsic
,
400 nir_instr_type_load_const
,
402 nir_instr_type_ssa_undef
,
404 nir_instr_type_parallel_copy
,
408 struct exec_node node
;
410 struct nir_block
*block
;
412 /* flag for dead code elimination (see nir_opt_dce.c) */
416 static inline nir_instr
*
417 nir_instr_next(const nir_instr
*instr
)
419 return exec_node_data(nir_instr
, (instr
)->node
.next
, node
);
422 static inline nir_instr
*
423 nir_instr_prev(const nir_instr
*instr
)
425 return exec_node_data(nir_instr
, (instr
)->node
.prev
, node
);
429 /** for debugging only, can be NULL */
432 /** generic SSA definition index. */
435 /** Index into the live_in and live_out bitfields */
438 nir_instr
*parent_instr
;
440 /** set of nir_instr's where this register is used (read from) */
443 /** set of nir_if's where this register is used as a condition */
446 uint8_t num_components
;
453 struct nir_src
*indirect
; /** < NULL for no indirect offset */
454 unsigned base_offset
;
456 /* TODO use-def chain goes here */
461 struct nir_src
*indirect
; /** < NULL for no indirect offset */
462 unsigned base_offset
;
464 /* TODO def-use chain goes here */
467 typedef struct nir_src
{
485 nir_src
nir_src_copy(nir_src src
, void *mem_ctx
);
486 nir_dest
nir_dest_copy(nir_dest dest
, void *mem_ctx
);
492 * \name input modifiers
496 * For inputs interpreted as floating point, flips the sign bit. For
497 * inputs interpreted as integers, performs the two's complement negation.
502 * Clears the sign bit for floating point values, and computes the integer
503 * absolute value for integers. Note that the negate modifier acts after
504 * the absolute value modifier, therefore if both are set then all inputs
505 * will become negative.
511 * For each input component, says which component of the register it is
512 * chosen from. Note that which elements of the swizzle are used and which
513 * are ignored are based on the write mask for most opcodes - for example,
514 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
515 * a swizzle of {2, x, 1, 0} where x means "don't care."
524 * \name saturate output modifier
526 * Only valid for opcodes that output floating-point numbers. Clamps the
527 * output to between 0.0 and 1.0 inclusive.
532 unsigned write_mask
: 4; /* ignored if dest.is_ssa is true */
535 #define OPCODE(name, num_inputs, output_size, output_type, \
536 input_sizes, input_types, algebraic_props) \
539 #define LAST_OPCODE(name) nir_last_opcode = nir_op_##name,
542 #include "nir_opcodes.h"
543 nir_num_opcodes
= nir_last_opcode
+ 1
557 NIR_OP_IS_COMMUTATIVE
= (1 << 0),
558 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
559 } nir_op_algebraic_property
;
567 * The number of components in the output
569 * If non-zero, this is the size of the output and input sizes are
570 * explicitly given; swizzle and writemask are still in effect, but if
571 * the output component is masked out, then the input component may
574 * If zero, the opcode acts in the standard, per-component manner; the
575 * operation is performed on each component (except the ones that are
576 * masked out) with the input being taken from the input swizzle for
579 * The size of some of the inputs may be given (i.e. non-zero) even
580 * though output_size is zero; in that case, the inputs with a zero
581 * size act per-component, while the inputs with non-zero size don't.
583 unsigned output_size
;
586 * The type of vector that the instruction outputs. Note that the
587 * staurate modifier is only allowed on outputs with the float type.
590 nir_alu_type output_type
;
593 * The number of components in each input
595 unsigned input_sizes
[4];
598 * The type of vector that each input takes. Note that negate and
599 * absolute value are only allowed on inputs with int or float type and
600 * behave differently on the two.
602 nir_alu_type input_types
[4];
604 nir_op_algebraic_property algebraic_properties
;
607 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
609 typedef struct nir_alu_instr
{
616 /* is this source channel used? */
618 nir_alu_instr_channel_used(nir_alu_instr
*instr
, unsigned src
, unsigned channel
)
620 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
621 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
623 return (instr
->dest
.write_mask
>> channel
) & 1;
628 nir_deref_type_array
,
629 nir_deref_type_struct
632 typedef struct nir_deref
{
633 nir_deref_type deref_type
;
634 struct nir_deref
*child
;
635 const struct glsl_type
*type
;
644 /* This enum describes how the array is referenced. If the deref is
645 * direct then the base_offset is used. If the deref is indirect then then
646 * offset is given by base_offset + indirect. If the deref is a wildcard
647 * then the deref refers to all of the elements of the array at the same
648 * time. Wildcard dereferences are only ever allowed in copy_var
649 * intrinsics and the source and destination derefs must have matching
653 nir_deref_array_type_direct
,
654 nir_deref_array_type_indirect
,
655 nir_deref_array_type_wildcard
,
656 } nir_deref_array_type
;
661 nir_deref_array_type deref_array_type
;
662 unsigned base_offset
;
672 NIR_DEFINE_CAST(nir_deref_as_var
, nir_deref
, nir_deref_var
, deref
)
673 NIR_DEFINE_CAST(nir_deref_as_array
, nir_deref
, nir_deref_array
, deref
)
674 NIR_DEFINE_CAST(nir_deref_as_struct
, nir_deref
, nir_deref_struct
, deref
)
680 nir_deref_var
**params
;
681 nir_deref_var
*return_deref
;
683 struct nir_function_overload
*callee
;
686 #define INTRINSIC(name, num_srcs, src_components, has_dest, dest_components, \
687 num_variables, num_indices, flags) \
688 nir_intrinsic_##name,
690 #define LAST_INTRINSIC(name) nir_last_intrinsic = nir_intrinsic_##name,
693 #include "nir_intrinsics.h"
694 nir_num_intrinsics
= nir_last_intrinsic
+ 1
698 #undef LAST_INTRINSIC
700 /** Represents an intrinsic
702 * An intrinsic is an instruction type for handling things that are
703 * more-or-less regular operations but don't just consume and produce SSA
704 * values like ALU operations do. Intrinsics are not for things that have
705 * special semantic meaning such as phi nodes and parallel copies.
706 * Examples of intrinsics include variable load/store operations, system
707 * value loads, and the like. Even though texturing more-or-less falls
708 * under this category, texturing is its own instruction type because
709 * trying to represent texturing with intrinsics would lead to a
710 * combinatorial explosion of intrinsic opcodes.
712 * By having a single instruction type for handling a lot of different
713 * cases, optimization passes can look for intrinsics and, for the most
714 * part, completely ignore them. Each intrinsic type also has a few
715 * possible flags that govern whether or not they can be reordered or
716 * eliminated. That way passes like dead code elimination can still work
717 * on intrisics without understanding the meaning of each.
719 * Each intrinsic has some number of constant indices, some number of
720 * variables, and some number of sources. What these sources, variables,
721 * and indices mean depends on the intrinsic and is documented with the
722 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
723 * instructions are the only types of instruction that can operate on
729 nir_intrinsic_op intrinsic
;
733 /** number of components if this is a vectorized intrinsic
735 * Similarly to ALU operations, some intrinsics are vectorized.
736 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
737 * For vectorized intrinsics, the num_components field specifies the
738 * number of destination components and the number of source components
739 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
741 uint8_t num_components
;
745 nir_deref_var
*variables
[2];
748 } nir_intrinsic_instr
;
751 * \name NIR intrinsics semantic flags
753 * information about what the compiler can do with the intrinsics.
755 * \sa nir_intrinsic_info::flags
759 * whether the intrinsic can be safely eliminated if none of its output
760 * value is not being used.
762 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
765 * Whether the intrinsic can be reordered with respect to any other
766 * intrinsic, i.e. whether the only reordering dependencies of the
767 * intrinsic are due to the register reads/writes.
769 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
770 } nir_intrinsic_semantic_flag
;
772 #define NIR_INTRINSIC_MAX_INPUTS 4
777 unsigned num_srcs
; /** < number of register/SSA inputs */
779 /** number of components of each input register
781 * If this value is 0, the number of components is given by the
782 * num_components field of nir_intrinsic_instr.
784 unsigned src_components
[NIR_INTRINSIC_MAX_INPUTS
];
788 /** number of components of the output register
790 * If this value is 0, the number of components is given by the
791 * num_components field of nir_intrinsic_instr.
793 unsigned dest_components
;
795 /** the number of inputs/outputs that are variables */
796 unsigned num_variables
;
798 /** the number of constant indices used by the intrinsic */
799 unsigned num_indices
;
801 /** semantic flags for calls to this intrinsic */
802 nir_intrinsic_semantic_flag flags
;
803 } nir_intrinsic_info
;
805 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
808 * \group texture information
810 * This gives semantic information about textures which is useful to the
811 * frontend, the backend, and lowering passes, but not the optimizer.
816 nir_tex_src_projector
,
817 nir_tex_src_comparitor
, /* shadow comparitor */
821 nir_tex_src_ms_index
, /* MSAA sample index */
824 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
825 nir_num_texinput_types
829 nir_texop_tex
, /**< Regular texture look-up */
830 nir_texop_txb
, /**< Texture look-up with LOD bias */
831 nir_texop_txl
, /**< Texture look-up with explicit LOD */
832 nir_texop_txd
, /**< Texture look-up with partial derivatvies */
833 nir_texop_txf
, /**< Texel fetch with explicit LOD */
834 nir_texop_txf_ms
, /**< Multisample texture fetch */
835 nir_texop_txs
, /**< Texture size */
836 nir_texop_lod
, /**< Texture lod query */
837 nir_texop_tg4
, /**< Texture gather */
838 nir_texop_query_levels
/**< Texture levels query */
844 enum glsl_sampler_dim sampler_dim
;
845 nir_alu_type dest_type
;
850 nir_texinput_type src_type
[4];
851 unsigned num_srcs
, coord_components
;
852 bool is_array
, is_shadow
;
855 * If is_shadow is true, whether this is the old-style shadow that outputs 4
856 * components or the new-style shadow that outputs 1 component.
858 bool is_new_style_shadow
;
860 /* constant offset - must be 0 if the offset source is used */
863 /* gather component selector */
864 unsigned component
: 2;
866 /** The sampler index
868 * If this texture instruction has a nir_tex_src_sampler_offset source,
869 * then the sampler index is given by sampler_index + sampler_offset.
871 unsigned sampler_index
;
873 /** The size of the sampler array or 0 if it's not an array */
874 unsigned sampler_array_size
;
876 nir_deref_var
*sampler
; /* if this is NULL, use sampler_index instead */
879 static inline unsigned
880 nir_tex_instr_dest_size(nir_tex_instr
*instr
)
882 if (instr
->op
== nir_texop_txs
) {
884 switch (instr
->sampler_dim
) {
885 case GLSL_SAMPLER_DIM_1D
:
886 case GLSL_SAMPLER_DIM_BUF
:
889 case GLSL_SAMPLER_DIM_2D
:
890 case GLSL_SAMPLER_DIM_CUBE
:
891 case GLSL_SAMPLER_DIM_MS
:
892 case GLSL_SAMPLER_DIM_RECT
:
893 case GLSL_SAMPLER_DIM_EXTERNAL
:
896 case GLSL_SAMPLER_DIM_3D
:
908 if (instr
->op
== nir_texop_query_levels
)
911 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
917 static inline unsigned
918 nir_tex_instr_src_size(nir_tex_instr
*instr
, unsigned src
)
920 if (instr
->src_type
[src
] == nir_tex_src_coord
)
921 return instr
->coord_components
;
924 if (instr
->src_type
[src
] == nir_tex_src_offset
||
925 instr
->src_type
[src
] == nir_tex_src_ddx
||
926 instr
->src_type
[src
] == nir_tex_src_ddy
) {
928 return instr
->coord_components
- 1;
930 return instr
->coord_components
;
937 nir_tex_instr_src_index(nir_tex_instr
*instr
, nir_texinput_type type
)
939 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
940 if (instr
->src_type
[i
] == type
)
957 nir_const_value value
;
960 } nir_load_const_instr
;
973 /* creates a new SSA variable in an undefined state */
978 } nir_ssa_undef_instr
;
981 struct exec_node node
;
983 /* The predecessor block corresponding to this source */
984 struct nir_block
*pred
;
992 struct exec_list srcs
; /** < list of nir_phi_src */
998 struct exec_node node
;
1001 } nir_parallel_copy_entry
;
1003 #define nir_foreach_parallel_copy_entry(pcopy, entry) \
1004 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
1009 /* A list of nir_parallel_copy_entry's. The sources of all of the
1010 * entries are copied to the corresponding destinations "in parallel".
1011 * In other words, if we have two entries: a -> b and b -> a, the values
1014 struct exec_list entries
;
1015 } nir_parallel_copy_instr
;
1017 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
)
1018 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
)
1019 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
)
1020 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
)
1021 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
)
1022 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
)
1023 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
)
1024 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
)
1025 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
1026 nir_parallel_copy_instr
, instr
)
1031 * Control flow consists of a tree of control flow nodes, which include
1032 * if-statements and loops. The leaves of the tree are basic blocks, lists of
1033 * instructions that always run start-to-finish. Each basic block also keeps
1034 * track of its successors (blocks which may run immediately after the current
1035 * block) and predecessors (blocks which could have run immediately before the
1036 * current block). Each function also has a start block and an end block which
1037 * all return statements point to (which is always empty). Together, all the
1038 * blocks with their predecessors and successors make up the control flow
1039 * graph (CFG) of the function. There are helpers that modify the tree of
1040 * control flow nodes while modifying the CFG appropriately; these should be
1041 * used instead of modifying the tree directly.
1048 nir_cf_node_function
1051 typedef struct nir_cf_node
{
1052 struct exec_node node
;
1053 nir_cf_node_type type
;
1054 struct nir_cf_node
*parent
;
1057 typedef struct nir_block
{
1058 nir_cf_node cf_node
;
1060 struct exec_list instr_list
; /** < list of nir_instr */
1062 /** generic block index; generated by nir_index_blocks */
1066 * Each block can only have up to 2 successors, so we put them in a simple
1067 * array - no need for anything more complicated.
1069 struct nir_block
*successors
[2];
1071 /* Set of nir_block predecessors in the CFG */
1072 struct set
*predecessors
;
1075 * this node's immediate dominator in the dominance tree - set to NULL for
1078 struct nir_block
*imm_dom
;
1080 /* This node's children in the dominance tree */
1081 unsigned num_dom_children
;
1082 struct nir_block
**dom_children
;
1084 /* Set of nir_block's on the dominance frontier of this block */
1085 struct set
*dom_frontier
;
1087 /* live in and out for this block; used for liveness analysis */
1088 BITSET_WORD
*live_in
;
1089 BITSET_WORD
*live_out
;
1092 static inline nir_instr
*
1093 nir_block_first_instr(nir_block
*block
)
1095 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
1096 return exec_node_data(nir_instr
, head
, node
);
1099 static inline nir_instr
*
1100 nir_block_last_instr(nir_block
*block
)
1102 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
1103 return exec_node_data(nir_instr
, tail
, node
);
1106 #define nir_foreach_instr(block, instr) \
1107 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
1108 #define nir_foreach_instr_reverse(block, instr) \
1109 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
1110 #define nir_foreach_instr_safe(block, instr) \
1111 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
1114 nir_cf_node cf_node
;
1117 struct exec_list then_list
; /** < list of nir_cf_node */
1118 struct exec_list else_list
; /** < list of nir_cf_node */
1121 static inline nir_cf_node
*
1122 nir_if_first_then_node(nir_if
*if_stmt
)
1124 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
1125 return exec_node_data(nir_cf_node
, head
, node
);
1128 static inline nir_cf_node
*
1129 nir_if_last_then_node(nir_if
*if_stmt
)
1131 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
1132 return exec_node_data(nir_cf_node
, tail
, node
);
1135 static inline nir_cf_node
*
1136 nir_if_first_else_node(nir_if
*if_stmt
)
1138 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
1139 return exec_node_data(nir_cf_node
, head
, node
);
1142 static inline nir_cf_node
*
1143 nir_if_last_else_node(nir_if
*if_stmt
)
1145 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
1146 return exec_node_data(nir_cf_node
, tail
, node
);
1150 nir_cf_node cf_node
;
1152 struct exec_list body
; /** < list of nir_cf_node */
1155 static inline nir_cf_node
*
1156 nir_loop_first_cf_node(nir_loop
*loop
)
1158 return exec_node_data(nir_cf_node
, exec_list_get_head(&loop
->body
), node
);
1161 static inline nir_cf_node
*
1162 nir_loop_last_cf_node(nir_loop
*loop
)
1164 return exec_node_data(nir_cf_node
, exec_list_get_tail(&loop
->body
), node
);
1168 * Various bits of metadata that can may be created or required by
1169 * optimization and analysis passes
1172 nir_metadata_none
= 0x0,
1173 nir_metadata_block_index
= 0x1,
1174 nir_metadata_dominance
= 0x2,
1175 nir_metadata_live_variables
= 0x4,
1179 nir_cf_node cf_node
;
1181 /** pointer to the overload of which this is an implementation */
1182 struct nir_function_overload
*overload
;
1184 struct exec_list body
; /** < list of nir_cf_node */
1186 nir_block
*start_block
, *end_block
;
1188 /** list for all local variables in the function */
1189 struct exec_list locals
;
1191 /** array of variables used as parameters */
1192 unsigned num_params
;
1193 nir_variable
**params
;
1195 /** variable used to hold the result of the function */
1196 nir_variable
*return_var
;
1198 /** list of local registers in the function */
1199 struct exec_list registers
;
1201 /** next available local register index */
1204 /** next available SSA value index */
1207 /* total number of basic blocks, only valid when block_index_dirty = false */
1208 unsigned num_blocks
;
1210 nir_metadata valid_metadata
;
1211 } nir_function_impl
;
1213 static inline nir_cf_node
*
1214 nir_cf_node_next(nir_cf_node
*node
)
1216 return exec_node_data(nir_cf_node
, exec_node_get_next(&node
->node
), node
);
1219 static inline nir_cf_node
*
1220 nir_cf_node_prev(nir_cf_node
*node
)
1222 return exec_node_data(nir_cf_node
, exec_node_get_prev(&node
->node
), node
);
1226 nir_cf_node_is_first(const nir_cf_node
*node
)
1228 return exec_node_is_head_sentinel(node
->node
.prev
);
1232 nir_cf_node_is_last(const nir_cf_node
*node
)
1234 return exec_node_is_tail_sentinel(node
->node
.next
);
1237 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
)
1238 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
)
1239 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
)
1240 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
, nir_function_impl
, cf_node
)
1245 nir_parameter_inout
,
1246 } nir_parameter_type
;
1249 nir_parameter_type param_type
;
1250 const struct glsl_type
*type
;
1253 typedef struct nir_function_overload
{
1254 struct exec_node node
;
1256 unsigned num_params
;
1257 nir_parameter
*params
;
1258 const struct glsl_type
*return_type
;
1260 nir_function_impl
*impl
; /** < NULL if the overload is only declared yet */
1262 /** pointer to the function of which this is an overload */
1263 struct nir_function
*function
;
1264 } nir_function_overload
;
1266 typedef struct nir_function
{
1267 struct exec_node node
;
1269 struct exec_list overload_list
; /** < list of nir_function_overload */
1271 struct nir_shader
*shader
;
1274 #define nir_function_first_overload(func) \
1275 exec_node_data(nir_function_overload, \
1276 exec_list_get_head(&(func)->overload_list), node)
1278 typedef struct nir_shader
{
1279 /** hash table of name -> uniform nir_variable */
1280 struct hash_table
*uniforms
;
1282 /** hash table of name -> input nir_variable */
1283 struct hash_table
*inputs
;
1285 /** hash table of name -> output nir_variable */
1286 struct hash_table
*outputs
;
1288 /** list of global variables in the shader */
1289 struct exec_list globals
;
1291 /** list of system value variables in the shader */
1292 struct exec_list system_values
;
1294 struct exec_list functions
; /** < list of nir_function */
1296 /** list of global register in the shader */
1297 struct exec_list registers
;
1299 /** structures used in this shader */
1300 unsigned num_user_structures
;
1301 struct glsl_type
**user_structures
;
1303 /** next available global register index */
1307 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
1310 unsigned num_inputs
, num_uniforms
, num_outputs
;
1313 #define nir_foreach_overload(shader, overload) \
1314 foreach_list_typed(nir_function, func, node, &(shader)->functions) \
1315 foreach_list_typed(nir_function_overload, overload, node, \
1316 &(func)->overload_list)
1318 nir_shader
*nir_shader_create(void *mem_ctx
);
1320 /** creates a register, including assigning it an index and adding it to the list */
1321 nir_register
*nir_global_reg_create(nir_shader
*shader
);
1323 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
1325 void nir_reg_remove(nir_register
*reg
);
1327 /** creates a function and adds it to the shader's list of functions */
1328 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
1330 /** creates a null function returning null */
1331 nir_function_overload
*nir_function_overload_create(nir_function
*func
);
1333 nir_function_impl
*nir_function_impl_create(nir_function_overload
*func
);
1335 nir_block
*nir_block_create(void *mem_ctx
);
1336 nir_if
*nir_if_create(void *mem_ctx
);
1337 nir_loop
*nir_loop_create(void *mem_ctx
);
1339 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
1341 /** puts a control flow node immediately after another control flow node */
1342 void nir_cf_node_insert_after(nir_cf_node
*node
, nir_cf_node
*after
);
1344 /** puts a control flow node immediately before another control flow node */
1345 void nir_cf_node_insert_before(nir_cf_node
*node
, nir_cf_node
*before
);
1347 /** puts a control flow node at the beginning of a list from an if, loop, or function */
1348 void nir_cf_node_insert_begin(struct exec_list
*list
, nir_cf_node
*node
);
1350 /** puts a control flow node at the end of a list from an if, loop, or function */
1351 void nir_cf_node_insert_end(struct exec_list
*list
, nir_cf_node
*node
);
1353 /** removes a control flow node, doing any cleanup necessary */
1354 void nir_cf_node_remove(nir_cf_node
*node
);
1356 /** requests that the given pieces of metadata be generated */
1357 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
);
1358 /** dirties all but the preserved metadata */
1359 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
1361 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
1362 nir_alu_instr
*nir_alu_instr_create(void *mem_ctx
, nir_op op
);
1364 nir_jump_instr
*nir_jump_instr_create(void *mem_ctx
, nir_jump_type type
);
1366 nir_load_const_instr
*nir_load_const_instr_create(void *mem_ctx
,
1367 unsigned num_components
);
1369 nir_intrinsic_instr
*nir_intrinsic_instr_create(void *mem_ctx
,
1370 nir_intrinsic_op op
);
1372 nir_call_instr
*nir_call_instr_create(void *mem_ctx
,
1373 nir_function_overload
*callee
);
1375 nir_tex_instr
*nir_tex_instr_create(void *mem_ctx
, unsigned num_srcs
);
1377 nir_phi_instr
*nir_phi_instr_create(void *mem_ctx
);
1379 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(void *mem_ctx
);
1381 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(void *mem_ctx
,
1382 unsigned num_components
);
1384 nir_deref_var
*nir_deref_var_create(void *mem_ctx
, nir_variable
*var
);
1385 nir_deref_array
*nir_deref_array_create(void *mem_ctx
);
1386 nir_deref_struct
*nir_deref_struct_create(void *mem_ctx
, unsigned field_index
);
1388 nir_deref
*nir_copy_deref(void *mem_ctx
, nir_deref
*deref
);
1390 void nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
);
1391 void nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
);
1393 void nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
);
1394 void nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
);
1396 void nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
);
1397 void nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
);
1399 void nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
);
1400 void nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
);
1402 void nir_instr_remove(nir_instr
*instr
);
1404 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
1405 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
1406 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
1407 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
1409 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
1410 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
1412 nir_const_value
*nir_src_as_const_value(nir_src src
);
1413 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
1414 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
1416 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
1417 unsigned num_components
, const char *name
);
1418 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
, void *mem_ctx
);
1420 /* visits basic blocks in source-code order */
1421 typedef bool (*nir_foreach_block_cb
)(nir_block
*block
, void *state
);
1422 bool nir_foreach_block(nir_function_impl
*impl
, nir_foreach_block_cb cb
,
1424 bool nir_foreach_block_reverse(nir_function_impl
*impl
, nir_foreach_block_cb cb
,
1427 /* If the following CF node is an if, this function returns that if.
1428 * Otherwise, it returns NULL.
1430 nir_if
*nir_block_get_following_if(nir_block
*block
);
1432 void nir_index_local_regs(nir_function_impl
*impl
);
1433 void nir_index_global_regs(nir_shader
*shader
);
1434 void nir_index_ssa_defs(nir_function_impl
*impl
);
1436 void nir_index_blocks(nir_function_impl
*impl
);
1438 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
1440 void nir_validate_shader(nir_shader
*shader
);
1442 void nir_calc_dominance_impl(nir_function_impl
*impl
);
1443 void nir_calc_dominance(nir_shader
*shader
);
1445 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
1446 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
1448 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
1449 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
1451 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
1452 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
1454 void nir_split_var_copies(nir_shader
*shader
);
1456 void nir_lower_global_vars_to_local(nir_shader
*shader
);
1458 void nir_lower_locals_to_regs(nir_shader
*shader
);
1460 void nir_lower_io(nir_shader
*shader
);
1462 void nir_lower_variables(nir_shader
*shader
);
1464 void nir_remove_dead_variables(nir_shader
*shader
);
1466 void nir_lower_vec_to_movs(nir_shader
*shader
);
1468 void nir_lower_samplers(nir_shader
*shader
,
1469 struct gl_shader_program
*shader_program
,
1470 struct gl_program
*prog
);
1472 void nir_lower_system_values(nir_shader
*shader
);
1474 void nir_lower_atomics(nir_shader
*shader
);
1475 void nir_lower_to_source_mods(nir_shader
*shader
);
1477 void nir_live_variables_impl(nir_function_impl
*impl
);
1478 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
1480 void nir_convert_to_ssa_impl(nir_function_impl
*impl
);
1481 void nir_convert_to_ssa(nir_shader
*shader
);
1482 void nir_convert_from_ssa(nir_shader
*shader
);
1484 bool nir_opt_algebraic(nir_shader
*shader
);
1485 bool nir_opt_constant_folding(nir_shader
*shader
);
1487 bool nir_opt_global_to_local(nir_shader
*shader
);
1489 bool nir_copy_prop_impl(nir_function_impl
*impl
);
1490 bool nir_copy_prop(nir_shader
*shader
);
1492 bool nir_opt_cse(nir_shader
*shader
);
1494 bool nir_opt_dce_impl(nir_function_impl
*impl
);
1495 bool nir_opt_dce(nir_shader
*shader
);
1497 bool nir_opt_peephole_select(nir_shader
*shader
);
1498 bool nir_opt_peephole_ffma(nir_shader
*shader
);