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/list.h"
34 #include "util/ralloc.h"
36 #include "util/bitset.h"
37 #include "nir_types.h"
38 #include "glsl/shader_enums.h"
41 #include "nir_opcodes.h"
48 struct gl_shader_program
;
51 #define NIR_TRUE (~0u)
53 /** Defines a cast function
55 * This macro defines a cast function from in_type to out_type where
56 * out_type is some structure type that contains a field of type out_type.
58 * Note that you have to be a bit careful as the generated cast function
61 #define NIR_DEFINE_CAST(name, in_type, out_type, field) \
62 static inline out_type * \
63 name(const in_type *parent) \
65 return exec_node_data(out_type, parent, field); \
68 struct nir_function_overload
;
75 * Description of built-in state associated with a uniform
77 * \sa nir_variable::state_slots
90 nir_var_shader_storage
,
95 * Data stored in an nir_constant
97 union nir_constant_data
{
104 typedef struct nir_constant
{
106 * Value of the constant.
108 * The field used to back the values supplied by the constant is determined
109 * by the type associated with the \c nir_variable. Constants may be
110 * scalars, vectors, or matrices.
112 union nir_constant_data value
;
114 /* Array elements / Structure Fields */
115 struct nir_constant
**elements
;
119 * \brief Layout qualifiers for gl_FragDepth.
121 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
122 * with a layout qualifier.
125 nir_depth_layout_none
, /**< No depth layout is specified. */
126 nir_depth_layout_any
,
127 nir_depth_layout_greater
,
128 nir_depth_layout_less
,
129 nir_depth_layout_unchanged
133 * Either a uniform, global variable, shader input, or shader output. Based on
134 * ir_variable - it should be easy to translate between the two.
138 struct exec_node node
;
141 * Declared type of the variable
143 const struct glsl_type
*type
;
146 * Declared name of the variable
151 * For variables which satisfy the is_interface_instance() predicate, this
152 * points to an array of integers such that if the ith member of the
153 * interface block is an array, max_ifc_array_access[i] is the maximum
154 * array element of that member that has been accessed. If the ith member
155 * of the interface block is not an array, max_ifc_array_access[i] is
158 * For variables whose type is not an interface block, this pointer is
161 unsigned *max_ifc_array_access
;
163 struct nir_variable_data
{
166 * Is the variable read-only?
168 * This is set for variables declared as \c const, shader inputs,
171 unsigned read_only
:1;
174 unsigned invariant
:1;
177 * Storage class of the variable.
179 * \sa nir_variable_mode
181 nir_variable_mode mode
:4;
184 * Interpolation mode for shader inputs / outputs
186 * \sa glsl_interp_qualifier
188 unsigned interpolation
:2;
191 * \name ARB_fragment_coord_conventions
194 unsigned origin_upper_left
:1;
195 unsigned pixel_center_integer
:1;
199 * Was the location explicitly set in the shader?
201 * If the location is explicitly set in the shader, it \b cannot be changed
202 * by the linker or by the API (e.g., calls to \c glBindAttribLocation have
205 unsigned explicit_location
:1;
206 unsigned explicit_index
:1;
209 * Was an initial binding explicitly set in the shader?
211 * If so, constant_initializer contains an integer nir_constant
212 * representing the initial binding point.
214 unsigned explicit_binding
:1;
217 * Does this variable have an initializer?
219 * This is used by the linker to cross-validiate initializers of global
222 unsigned has_initializer
:1;
225 * Is this variable a generic output or input that has not yet been matched
226 * up to a variable in another stage of the pipeline?
228 * This is used by the linker as scratch storage while assigning locations
229 * to generic inputs and outputs.
231 unsigned is_unmatched_generic_inout
:1;
234 * If non-zero, then this variable may be packed along with other variables
235 * into a single varying slot, so this offset should be applied when
236 * accessing components. For example, an offset of 1 means that the x
237 * component of this variable is actually stored in component y of the
238 * location specified by \c location.
240 unsigned location_frac
:2;
243 * Non-zero if this variable was created by lowering a named interface
244 * block which was not an array.
246 * Note that this variable and \c from_named_ifc_block_array will never
249 unsigned from_named_ifc_block_nonarray
:1;
252 * Non-zero if this variable was created by lowering a named interface
253 * block which was an array.
255 * Note that this variable and \c from_named_ifc_block_nonarray will never
258 unsigned from_named_ifc_block_array
:1;
261 * \brief Layout qualifier for gl_FragDepth.
263 * This is not equal to \c ir_depth_layout_none if and only if this
264 * variable is \c gl_FragDepth and a layout qualifier is specified.
266 nir_depth_layout depth_layout
;
269 * Storage location of the base of this variable
271 * The precise meaning of this field depends on the nature of the variable.
273 * - Vertex shader input: one of the values from \c gl_vert_attrib.
274 * - Vertex shader output: one of the values from \c gl_varying_slot.
275 * - Geometry shader input: one of the values from \c gl_varying_slot.
276 * - Geometry shader output: one of the values from \c gl_varying_slot.
277 * - Fragment shader input: one of the values from \c gl_varying_slot.
278 * - Fragment shader output: one of the values from \c gl_frag_result.
279 * - Uniforms: Per-stage uniform slot number for default uniform block.
280 * - Uniforms: Index within the uniform block definition for UBO members.
281 * - Other: This field is not currently used.
283 * If the variable is a uniform, shader input, or shader output, and the
284 * slot has not been assigned, the value will be -1.
289 * The actual location of the variable in the IR. Only valid for inputs
292 unsigned int driver_location
;
295 * output index for dual source blending.
300 * Initial binding point for a sampler or UBO.
302 * For array types, this represents the binding point for the first element.
307 * Location an atomic counter is stored at.
310 unsigned buffer_index
;
315 * ARB_shader_image_load_store qualifiers.
318 bool read_only
; /**< "readonly" qualifier. */
319 bool write_only
; /**< "writeonly" qualifier. */
324 /** Image internal format if specified explicitly, otherwise GL_NONE. */
329 * Highest element accessed with a constant expression array index
331 * Not used for non-array variables.
333 unsigned max_array_access
;
338 * Built-in state that backs this uniform
340 * Once set at variable creation, \c state_slots must remain invariant.
341 * This is because, ideally, this array would be shared by all clones of
342 * this variable in the IR tree. In other words, we'd really like for it
343 * to be a fly-weight.
345 * If the variable is not a uniform, \c num_state_slots will be zero and
346 * \c state_slots will be \c NULL.
349 unsigned num_state_slots
; /**< Number of state slots used */
350 nir_state_slot
*state_slots
; /**< State descriptors. */
354 * Constant expression assigned in the initializer of the variable
356 nir_constant
*constant_initializer
;
359 * For variables that are in an interface block or are an instance of an
360 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
362 * \sa ir_variable::location
364 const struct glsl_type
*interface_type
;
368 struct exec_node node
;
370 unsigned num_components
; /** < number of vector components */
371 unsigned num_array_elems
; /** < size of array (0 for no array) */
373 /** generic register index. */
376 /** only for debug purposes, can be NULL */
379 /** whether this register is local (per-function) or global (per-shader) */
383 * If this flag is set to true, then accessing channels >= num_components
384 * is well-defined, and simply spills over to the next array element. This
385 * is useful for backends that can do per-component accessing, in
386 * particular scalar backends. By setting this flag and making
387 * num_components equal to 1, structures can be packed tightly into
388 * registers and then registers can be accessed per-component to get to
389 * each structure member, even if it crosses vec4 boundaries.
393 /** set of nir_instr's where this register is used (read from) */
394 struct list_head uses
;
396 /** set of nir_instr's where this register is defined (written to) */
397 struct list_head defs
;
399 /** set of nir_if's where this register is used as a condition */
400 struct list_head if_uses
;
407 nir_instr_type_intrinsic
,
408 nir_instr_type_load_const
,
410 nir_instr_type_ssa_undef
,
412 nir_instr_type_parallel_copy
,
415 typedef struct nir_instr
{
416 struct exec_node node
;
418 struct nir_block
*block
;
420 /* A temporary for optimization and analysis passes to use for storing
421 * flags. For instance, DCE uses this to store the "dead/live" info.
426 static inline nir_instr
*
427 nir_instr_next(nir_instr
*instr
)
429 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
430 if (exec_node_is_tail_sentinel(next
))
433 return exec_node_data(nir_instr
, next
, node
);
436 static inline nir_instr
*
437 nir_instr_prev(nir_instr
*instr
)
439 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
440 if (exec_node_is_head_sentinel(prev
))
443 return exec_node_data(nir_instr
, prev
, node
);
447 nir_instr_is_first(nir_instr
*instr
)
449 return exec_node_is_head_sentinel(exec_node_get_prev(&instr
->node
));
453 nir_instr_is_last(nir_instr
*instr
)
455 return exec_node_is_tail_sentinel(exec_node_get_next(&instr
->node
));
459 /** for debugging only, can be NULL */
462 /** generic SSA definition index. */
465 /** Index into the live_in and live_out bitfields */
468 nir_instr
*parent_instr
;
470 /** set of nir_instr's where this register is used (read from) */
471 struct list_head uses
;
473 /** set of nir_if's where this register is used as a condition */
474 struct list_head if_uses
;
476 uint8_t num_components
;
483 struct nir_src
*indirect
; /** < NULL for no indirect offset */
484 unsigned base_offset
;
486 /* TODO use-def chain goes here */
490 nir_instr
*parent_instr
;
491 struct list_head def_link
;
494 struct nir_src
*indirect
; /** < NULL for no indirect offset */
495 unsigned base_offset
;
497 /* TODO def-use chain goes here */
502 typedef struct nir_src
{
504 nir_instr
*parent_instr
;
505 struct nir_if
*parent_if
;
508 struct list_head use_link
;
518 #define NIR_SRC_INIT (nir_src) { { NULL } }
520 #define nir_foreach_use(reg_or_ssa_def, src) \
521 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
523 #define nir_foreach_use_safe(reg_or_ssa_def, src) \
524 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
526 #define nir_foreach_if_use(reg_or_ssa_def, src) \
527 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
529 #define nir_foreach_if_use_safe(reg_or_ssa_def, src) \
530 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
541 #define NIR_DEST_INIT (nir_dest) { { { NULL } } }
543 #define nir_foreach_def(reg, dest) \
544 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
546 #define nir_foreach_def_safe(reg, dest) \
547 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
549 static inline nir_src
550 nir_src_for_ssa(nir_ssa_def
*def
)
552 nir_src src
= NIR_SRC_INIT
;
560 static inline nir_src
561 nir_src_for_reg(nir_register
*reg
)
563 nir_src src
= NIR_SRC_INIT
;
567 src
.reg
.indirect
= NULL
;
568 src
.reg
.base_offset
= 0;
573 static inline nir_dest
574 nir_dest_for_reg(nir_register
*reg
)
576 nir_dest dest
= NIR_DEST_INIT
;
583 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *mem_ctx
);
584 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, void *mem_ctx
);
590 * \name input modifiers
594 * For inputs interpreted as floating point, flips the sign bit. For
595 * inputs interpreted as integers, performs the two's complement negation.
600 * Clears the sign bit for floating point values, and computes the integer
601 * absolute value for integers. Note that the negate modifier acts after
602 * the absolute value modifier, therefore if both are set then all inputs
603 * will become negative.
609 * For each input component, says which component of the register it is
610 * chosen from. Note that which elements of the swizzle are used and which
611 * are ignored are based on the write mask for most opcodes - for example,
612 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
613 * a swizzle of {2, x, 1, 0} where x means "don't care."
622 * \name saturate output modifier
624 * Only valid for opcodes that output floating-point numbers. Clamps the
625 * output to between 0.0 and 1.0 inclusive.
630 unsigned write_mask
: 4; /* ignored if dest.is_ssa is true */
633 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
, void *mem_ctx
);
634 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
638 nir_type_invalid
= 0, /* Not a valid type */
646 NIR_OP_IS_COMMUTATIVE
= (1 << 0),
647 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
648 } nir_op_algebraic_property
;
656 * The number of components in the output
658 * If non-zero, this is the size of the output and input sizes are
659 * explicitly given; swizzle and writemask are still in effect, but if
660 * the output component is masked out, then the input component may
663 * If zero, the opcode acts in the standard, per-component manner; the
664 * operation is performed on each component (except the ones that are
665 * masked out) with the input being taken from the input swizzle for
668 * The size of some of the inputs may be given (i.e. non-zero) even
669 * though output_size is zero; in that case, the inputs with a zero
670 * size act per-component, while the inputs with non-zero size don't.
672 unsigned output_size
;
675 * The type of vector that the instruction outputs. Note that the
676 * staurate modifier is only allowed on outputs with the float type.
679 nir_alu_type output_type
;
682 * The number of components in each input
684 unsigned input_sizes
[4];
687 * The type of vector that each input takes. Note that negate and
688 * absolute value are only allowed on inputs with int or float type and
689 * behave differently on the two.
691 nir_alu_type input_types
[4];
693 nir_op_algebraic_property algebraic_properties
;
696 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
698 typedef struct nir_alu_instr
{
705 /* is this source channel used? */
707 nir_alu_instr_channel_used(nir_alu_instr
*instr
, unsigned src
, unsigned channel
)
709 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
710 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
712 return (instr
->dest
.write_mask
>> channel
) & 1;
716 * For instructions whose destinations are SSA, get the number of channels
719 static inline unsigned
720 nir_ssa_alu_instr_src_components(nir_alu_instr
*instr
, unsigned src
)
722 assert(instr
->dest
.dest
.is_ssa
);
724 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
725 return nir_op_infos
[instr
->op
].input_sizes
[src
];
727 return instr
->dest
.dest
.ssa
.num_components
;
732 nir_deref_type_array
,
733 nir_deref_type_struct
736 typedef struct nir_deref
{
737 nir_deref_type deref_type
;
738 struct nir_deref
*child
;
739 const struct glsl_type
*type
;
748 /* This enum describes how the array is referenced. If the deref is
749 * direct then the base_offset is used. If the deref is indirect then then
750 * offset is given by base_offset + indirect. If the deref is a wildcard
751 * then the deref refers to all of the elements of the array at the same
752 * time. Wildcard dereferences are only ever allowed in copy_var
753 * intrinsics and the source and destination derefs must have matching
757 nir_deref_array_type_direct
,
758 nir_deref_array_type_indirect
,
759 nir_deref_array_type_wildcard
,
760 } nir_deref_array_type
;
765 nir_deref_array_type deref_array_type
;
766 unsigned base_offset
;
776 NIR_DEFINE_CAST(nir_deref_as_var
, nir_deref
, nir_deref_var
, deref
)
777 NIR_DEFINE_CAST(nir_deref_as_array
, nir_deref
, nir_deref_array
, deref
)
778 NIR_DEFINE_CAST(nir_deref_as_struct
, nir_deref
, nir_deref_struct
, deref
)
784 nir_deref_var
**params
;
785 nir_deref_var
*return_deref
;
787 struct nir_function_overload
*callee
;
790 #define INTRINSIC(name, num_srcs, src_components, has_dest, dest_components, \
791 num_variables, num_indices, flags) \
792 nir_intrinsic_##name,
794 #define LAST_INTRINSIC(name) nir_last_intrinsic = nir_intrinsic_##name,
797 #include "nir_intrinsics.h"
798 nir_num_intrinsics
= nir_last_intrinsic
+ 1
802 #undef LAST_INTRINSIC
804 /** Represents an intrinsic
806 * An intrinsic is an instruction type for handling things that are
807 * more-or-less regular operations but don't just consume and produce SSA
808 * values like ALU operations do. Intrinsics are not for things that have
809 * special semantic meaning such as phi nodes and parallel copies.
810 * Examples of intrinsics include variable load/store operations, system
811 * value loads, and the like. Even though texturing more-or-less falls
812 * under this category, texturing is its own instruction type because
813 * trying to represent texturing with intrinsics would lead to a
814 * combinatorial explosion of intrinsic opcodes.
816 * By having a single instruction type for handling a lot of different
817 * cases, optimization passes can look for intrinsics and, for the most
818 * part, completely ignore them. Each intrinsic type also has a few
819 * possible flags that govern whether or not they can be reordered or
820 * eliminated. That way passes like dead code elimination can still work
821 * on intrisics without understanding the meaning of each.
823 * Each intrinsic has some number of constant indices, some number of
824 * variables, and some number of sources. What these sources, variables,
825 * and indices mean depends on the intrinsic and is documented with the
826 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
827 * instructions are the only types of instruction that can operate on
833 nir_intrinsic_op intrinsic
;
837 /** number of components if this is a vectorized intrinsic
839 * Similarly to ALU operations, some intrinsics are vectorized.
840 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
841 * For vectorized intrinsics, the num_components field specifies the
842 * number of destination components and the number of source components
843 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
845 uint8_t num_components
;
849 nir_deref_var
*variables
[2];
852 } nir_intrinsic_instr
;
855 * \name NIR intrinsics semantic flags
857 * information about what the compiler can do with the intrinsics.
859 * \sa nir_intrinsic_info::flags
863 * whether the intrinsic can be safely eliminated if none of its output
864 * value is not being used.
866 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
869 * Whether the intrinsic can be reordered with respect to any other
870 * intrinsic, i.e. whether the only reordering dependencies of the
871 * intrinsic are due to the register reads/writes.
873 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
874 } nir_intrinsic_semantic_flag
;
876 #define NIR_INTRINSIC_MAX_INPUTS 4
881 unsigned num_srcs
; /** < number of register/SSA inputs */
883 /** number of components of each input register
885 * If this value is 0, the number of components is given by the
886 * num_components field of nir_intrinsic_instr.
888 unsigned src_components
[NIR_INTRINSIC_MAX_INPUTS
];
892 /** number of components of the output register
894 * If this value is 0, the number of components is given by the
895 * num_components field of nir_intrinsic_instr.
897 unsigned dest_components
;
899 /** the number of inputs/outputs that are variables */
900 unsigned num_variables
;
902 /** the number of constant indices used by the intrinsic */
903 unsigned num_indices
;
905 /** semantic flags for calls to this intrinsic */
906 nir_intrinsic_semantic_flag flags
;
907 } nir_intrinsic_info
;
909 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
912 * \group texture information
914 * This gives semantic information about textures which is useful to the
915 * frontend, the backend, and lowering passes, but not the optimizer.
920 nir_tex_src_projector
,
921 nir_tex_src_comparitor
, /* shadow comparitor */
925 nir_tex_src_ms_index
, /* MSAA sample index */
928 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
929 nir_num_tex_src_types
934 nir_tex_src_type src_type
;
938 nir_texop_tex
, /**< Regular texture look-up */
939 nir_texop_txb
, /**< Texture look-up with LOD bias */
940 nir_texop_txl
, /**< Texture look-up with explicit LOD */
941 nir_texop_txd
, /**< Texture look-up with partial derivatvies */
942 nir_texop_txf
, /**< Texel fetch with explicit LOD */
943 nir_texop_txf_ms
, /**< Multisample texture fetch */
944 nir_texop_txs
, /**< Texture size */
945 nir_texop_lod
, /**< Texture lod query */
946 nir_texop_tg4
, /**< Texture gather */
947 nir_texop_query_levels
/**< Texture levels query */
953 enum glsl_sampler_dim sampler_dim
;
954 nir_alu_type dest_type
;
959 unsigned num_srcs
, coord_components
;
960 bool is_array
, is_shadow
;
963 * If is_shadow is true, whether this is the old-style shadow that outputs 4
964 * components or the new-style shadow that outputs 1 component.
966 bool is_new_style_shadow
;
968 /* constant offset - must be 0 if the offset source is used */
971 /* gather component selector */
972 unsigned component
: 2;
974 /** The sampler index
976 * If this texture instruction has a nir_tex_src_sampler_offset source,
977 * then the sampler index is given by sampler_index + sampler_offset.
979 unsigned sampler_index
;
981 /** The size of the sampler array or 0 if it's not an array */
982 unsigned sampler_array_size
;
984 nir_deref_var
*sampler
; /* if this is NULL, use sampler_index instead */
987 static inline unsigned
988 nir_tex_instr_dest_size(nir_tex_instr
*instr
)
991 case nir_texop_txs
: {
993 switch (instr
->sampler_dim
) {
994 case GLSL_SAMPLER_DIM_1D
:
995 case GLSL_SAMPLER_DIM_BUF
:
998 case GLSL_SAMPLER_DIM_2D
:
999 case GLSL_SAMPLER_DIM_CUBE
:
1000 case GLSL_SAMPLER_DIM_MS
:
1001 case GLSL_SAMPLER_DIM_RECT
:
1002 case GLSL_SAMPLER_DIM_EXTERNAL
:
1005 case GLSL_SAMPLER_DIM_3D
:
1009 unreachable("not reached");
1011 if (instr
->is_array
)
1019 case nir_texop_query_levels
:
1023 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
1030 static inline unsigned
1031 nir_tex_instr_src_size(nir_tex_instr
*instr
, unsigned src
)
1033 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
1034 return instr
->coord_components
;
1037 if (instr
->src
[src
].src_type
== nir_tex_src_offset
||
1038 instr
->src
[src
].src_type
== nir_tex_src_ddx
||
1039 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
1040 if (instr
->is_array
)
1041 return instr
->coord_components
- 1;
1043 return instr
->coord_components
;
1050 nir_tex_instr_src_index(nir_tex_instr
*instr
, nir_tex_src_type type
)
1052 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
1053 if (instr
->src
[i
].src_type
== type
)
1070 nir_const_value value
;
1073 } nir_load_const_instr
;
1086 /* creates a new SSA variable in an undefined state */
1091 } nir_ssa_undef_instr
;
1094 struct exec_node node
;
1096 /* The predecessor block corresponding to this source */
1097 struct nir_block
*pred
;
1102 #define nir_foreach_phi_src(phi, entry) \
1103 foreach_list_typed(nir_phi_src, entry, node, &(phi)->srcs)
1108 struct exec_list srcs
; /** < list of nir_phi_src */
1114 struct exec_node node
;
1117 } nir_parallel_copy_entry
;
1119 #define nir_foreach_parallel_copy_entry(pcopy, entry) \
1120 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
1125 /* A list of nir_parallel_copy_entry's. The sources of all of the
1126 * entries are copied to the corresponding destinations "in parallel".
1127 * In other words, if we have two entries: a -> b and b -> a, the values
1130 struct exec_list entries
;
1131 } nir_parallel_copy_instr
;
1133 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
)
1134 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
)
1135 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
)
1136 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
)
1137 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
)
1138 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
)
1139 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
)
1140 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
)
1141 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
1142 nir_parallel_copy_instr
, instr
)
1147 * Control flow consists of a tree of control flow nodes, which include
1148 * if-statements and loops. The leaves of the tree are basic blocks, lists of
1149 * instructions that always run start-to-finish. Each basic block also keeps
1150 * track of its successors (blocks which may run immediately after the current
1151 * block) and predecessors (blocks which could have run immediately before the
1152 * current block). Each function also has a start block and an end block which
1153 * all return statements point to (which is always empty). Together, all the
1154 * blocks with their predecessors and successors make up the control flow
1155 * graph (CFG) of the function. There are helpers that modify the tree of
1156 * control flow nodes while modifying the CFG appropriately; these should be
1157 * used instead of modifying the tree directly.
1164 nir_cf_node_function
1167 typedef struct nir_cf_node
{
1168 struct exec_node node
;
1169 nir_cf_node_type type
;
1170 struct nir_cf_node
*parent
;
1173 typedef struct nir_block
{
1174 nir_cf_node cf_node
;
1176 struct exec_list instr_list
; /** < list of nir_instr */
1178 /** generic block index; generated by nir_index_blocks */
1182 * Each block can only have up to 2 successors, so we put them in a simple
1183 * array - no need for anything more complicated.
1185 struct nir_block
*successors
[2];
1187 /* Set of nir_block predecessors in the CFG */
1188 struct set
*predecessors
;
1191 * this node's immediate dominator in the dominance tree - set to NULL for
1194 struct nir_block
*imm_dom
;
1196 /* This node's children in the dominance tree */
1197 unsigned num_dom_children
;
1198 struct nir_block
**dom_children
;
1200 /* Set of nir_block's on the dominance frontier of this block */
1201 struct set
*dom_frontier
;
1204 * These two indices have the property that dom_{pre,post}_index for each
1205 * child of this block in the dominance tree will always be between
1206 * dom_pre_index and dom_post_index for this block, which makes testing if
1207 * a given block is dominated by another block an O(1) operation.
1209 unsigned dom_pre_index
, dom_post_index
;
1211 /* live in and out for this block; used for liveness analysis */
1212 BITSET_WORD
*live_in
;
1213 BITSET_WORD
*live_out
;
1216 static inline nir_instr
*
1217 nir_block_first_instr(nir_block
*block
)
1219 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
1220 return exec_node_data(nir_instr
, head
, node
);
1223 static inline nir_instr
*
1224 nir_block_last_instr(nir_block
*block
)
1226 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
1227 return exec_node_data(nir_instr
, tail
, node
);
1230 #define nir_foreach_instr(block, instr) \
1231 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
1232 #define nir_foreach_instr_reverse(block, instr) \
1233 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
1234 #define nir_foreach_instr_safe(block, instr) \
1235 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
1236 #define nir_foreach_instr_safe_reverse(block, instr) \
1237 foreach_list_typed_safe_reverse(nir_instr, instr, node, &(block)->instr_list)
1239 typedef struct nir_if
{
1240 nir_cf_node cf_node
;
1243 struct exec_list then_list
; /** < list of nir_cf_node */
1244 struct exec_list else_list
; /** < list of nir_cf_node */
1247 static inline nir_cf_node
*
1248 nir_if_first_then_node(nir_if
*if_stmt
)
1250 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
1251 return exec_node_data(nir_cf_node
, head
, node
);
1254 static inline nir_cf_node
*
1255 nir_if_last_then_node(nir_if
*if_stmt
)
1257 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
1258 return exec_node_data(nir_cf_node
, tail
, node
);
1261 static inline nir_cf_node
*
1262 nir_if_first_else_node(nir_if
*if_stmt
)
1264 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
1265 return exec_node_data(nir_cf_node
, head
, node
);
1268 static inline nir_cf_node
*
1269 nir_if_last_else_node(nir_if
*if_stmt
)
1271 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
1272 return exec_node_data(nir_cf_node
, tail
, node
);
1276 nir_cf_node cf_node
;
1278 struct exec_list body
; /** < list of nir_cf_node */
1281 static inline nir_cf_node
*
1282 nir_loop_first_cf_node(nir_loop
*loop
)
1284 return exec_node_data(nir_cf_node
, exec_list_get_head(&loop
->body
), node
);
1287 static inline nir_cf_node
*
1288 nir_loop_last_cf_node(nir_loop
*loop
)
1290 return exec_node_data(nir_cf_node
, exec_list_get_tail(&loop
->body
), node
);
1294 * Various bits of metadata that can may be created or required by
1295 * optimization and analysis passes
1298 nir_metadata_none
= 0x0,
1299 nir_metadata_block_index
= 0x1,
1300 nir_metadata_dominance
= 0x2,
1301 nir_metadata_live_variables
= 0x4,
1305 nir_cf_node cf_node
;
1307 /** pointer to the overload of which this is an implementation */
1308 struct nir_function_overload
*overload
;
1310 struct exec_list body
; /** < list of nir_cf_node */
1312 nir_block
*end_block
;
1314 /** list for all local variables in the function */
1315 struct exec_list locals
;
1317 /** array of variables used as parameters */
1318 unsigned num_params
;
1319 nir_variable
**params
;
1321 /** variable used to hold the result of the function */
1322 nir_variable
*return_var
;
1324 /** list of local registers in the function */
1325 struct exec_list registers
;
1327 /** next available local register index */
1330 /** next available SSA value index */
1333 /* total number of basic blocks, only valid when block_index_dirty = false */
1334 unsigned num_blocks
;
1336 nir_metadata valid_metadata
;
1337 } nir_function_impl
;
1339 static inline nir_block
*
1340 nir_start_block(nir_function_impl
*impl
)
1342 return (nir_block
*) exec_list_get_head(&impl
->body
);
1345 static inline nir_cf_node
*
1346 nir_cf_node_next(nir_cf_node
*node
)
1348 struct exec_node
*next
= exec_node_get_next(&node
->node
);
1349 if (exec_node_is_tail_sentinel(next
))
1352 return exec_node_data(nir_cf_node
, next
, node
);
1355 static inline nir_cf_node
*
1356 nir_cf_node_prev(nir_cf_node
*node
)
1358 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
1359 if (exec_node_is_head_sentinel(prev
))
1362 return exec_node_data(nir_cf_node
, prev
, node
);
1366 nir_cf_node_is_first(const nir_cf_node
*node
)
1368 return exec_node_is_head_sentinel(node
->node
.prev
);
1372 nir_cf_node_is_last(const nir_cf_node
*node
)
1374 return exec_node_is_tail_sentinel(node
->node
.next
);
1377 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
)
1378 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
)
1379 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
)
1380 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
, nir_function_impl
, cf_node
)
1385 nir_parameter_inout
,
1386 } nir_parameter_type
;
1389 nir_parameter_type param_type
;
1390 const struct glsl_type
*type
;
1393 typedef struct nir_function_overload
{
1394 struct exec_node node
;
1396 unsigned num_params
;
1397 nir_parameter
*params
;
1398 const struct glsl_type
*return_type
;
1400 nir_function_impl
*impl
; /** < NULL if the overload is only declared yet */
1402 /** pointer to the function of which this is an overload */
1403 struct nir_function
*function
;
1404 } nir_function_overload
;
1406 typedef struct nir_function
{
1407 struct exec_node node
;
1409 struct exec_list overload_list
; /** < list of nir_function_overload */
1411 struct nir_shader
*shader
;
1414 #define nir_function_first_overload(func) \
1415 exec_node_data(nir_function_overload, \
1416 exec_list_get_head(&(func)->overload_list), node)
1418 typedef struct nir_shader_compiler_options
{
1424 /** lowers fneg and ineg to fsub and isub. */
1426 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
1429 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
1433 * Does the driver support real 32-bit integers? (Otherwise, integers
1434 * are simulated by floats.)
1436 bool native_integers
;
1437 } nir_shader_compiler_options
;
1439 typedef struct nir_shader
{
1440 /** hash table of name -> uniform nir_variable */
1441 struct exec_list uniforms
;
1443 /** hash table of name -> input nir_variable */
1444 struct exec_list inputs
;
1446 /** hash table of name -> output nir_variable */
1447 struct exec_list outputs
;
1449 /** Set of driver-specific options for the shader.
1451 * The memory for the options is expected to be kept in a single static
1452 * copy by the driver.
1454 const struct nir_shader_compiler_options
*options
;
1456 /** list of global variables in the shader */
1457 struct exec_list globals
;
1459 /** list of system value variables in the shader */
1460 struct exec_list system_values
;
1462 struct exec_list functions
; /** < list of nir_function */
1464 /** list of global register in the shader */
1465 struct exec_list registers
;
1467 /** next available global register index */
1471 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
1474 unsigned num_inputs
, num_uniforms
, num_outputs
;
1476 /** the number of uniforms that are only accessed directly */
1477 unsigned num_direct_uniforms
;
1480 #define nir_foreach_overload(shader, overload) \
1481 foreach_list_typed(nir_function, func, node, &(shader)->functions) \
1482 foreach_list_typed(nir_function_overload, overload, node, \
1483 &(func)->overload_list)
1485 nir_shader
*nir_shader_create(void *mem_ctx
,
1486 const nir_shader_compiler_options
*options
);
1488 /** creates a register, including assigning it an index and adding it to the list */
1489 nir_register
*nir_global_reg_create(nir_shader
*shader
);
1491 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
1493 void nir_reg_remove(nir_register
*reg
);
1495 /** creates a function and adds it to the shader's list of functions */
1496 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
1498 /** creates a null function returning null */
1499 nir_function_overload
*nir_function_overload_create(nir_function
*func
);
1501 nir_function_impl
*nir_function_impl_create(nir_function_overload
*func
);
1503 nir_block
*nir_block_create(void *mem_ctx
);
1504 nir_if
*nir_if_create(void *mem_ctx
);
1505 nir_loop
*nir_loop_create(void *mem_ctx
);
1507 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
1509 /** puts a control flow node immediately after another control flow node */
1510 void nir_cf_node_insert_after(nir_cf_node
*node
, nir_cf_node
*after
);
1512 /** puts a control flow node immediately before another control flow node */
1513 void nir_cf_node_insert_before(nir_cf_node
*node
, nir_cf_node
*before
);
1515 /** puts a control flow node at the beginning of a list from an if, loop, or function */
1516 void nir_cf_node_insert_begin(struct exec_list
*list
, nir_cf_node
*node
);
1518 /** puts a control flow node at the end of a list from an if, loop, or function */
1519 void nir_cf_node_insert_end(struct exec_list
*list
, nir_cf_node
*node
);
1521 /** removes a control flow node, doing any cleanup necessary */
1522 void nir_cf_node_remove(nir_cf_node
*node
);
1524 /** requests that the given pieces of metadata be generated */
1525 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
);
1526 /** dirties all but the preserved metadata */
1527 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
1529 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
1530 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
1532 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
1534 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
1535 unsigned num_components
);
1537 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
1538 nir_intrinsic_op op
);
1540 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
1541 nir_function_overload
*callee
);
1543 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
1545 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
1547 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
1549 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
1550 unsigned num_components
);
1552 nir_deref_var
*nir_deref_var_create(void *mem_ctx
, nir_variable
*var
);
1553 nir_deref_array
*nir_deref_array_create(void *mem_ctx
);
1554 nir_deref_struct
*nir_deref_struct_create(void *mem_ctx
, unsigned field_index
);
1556 nir_deref
*nir_copy_deref(void *mem_ctx
, nir_deref
*deref
);
1558 nir_load_const_instr
*
1559 nir_deref_get_const_initializer_load(nir_shader
*shader
, nir_deref_var
*deref
);
1561 void nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
);
1562 void nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
);
1564 void nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
);
1565 void nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
);
1567 void nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
);
1568 void nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
);
1570 void nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
);
1571 void nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
);
1573 void nir_instr_remove(nir_instr
*instr
);
1575 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
1576 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
1577 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
1578 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
1580 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
1581 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
1583 nir_const_value
*nir_src_as_const_value(nir_src src
);
1584 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
1585 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
1586 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
1587 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
1589 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
1590 unsigned num_components
, const char *name
);
1591 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
1592 unsigned num_components
, const char *name
);
1593 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
, void *mem_ctx
);
1595 /* visits basic blocks in source-code order */
1596 typedef bool (*nir_foreach_block_cb
)(nir_block
*block
, void *state
);
1597 bool nir_foreach_block(nir_function_impl
*impl
, nir_foreach_block_cb cb
,
1599 bool nir_foreach_block_reverse(nir_function_impl
*impl
, nir_foreach_block_cb cb
,
1602 /* If the following CF node is an if, this function returns that if.
1603 * Otherwise, it returns NULL.
1605 nir_if
*nir_block_get_following_if(nir_block
*block
);
1607 void nir_index_local_regs(nir_function_impl
*impl
);
1608 void nir_index_global_regs(nir_shader
*shader
);
1609 void nir_index_ssa_defs(nir_function_impl
*impl
);
1611 void nir_index_blocks(nir_function_impl
*impl
);
1613 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
1614 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
1617 void nir_validate_shader(nir_shader
*shader
);
1619 static inline void nir_validate_shader(nir_shader
*shader
) { (void) shader
; }
1622 void nir_calc_dominance_impl(nir_function_impl
*impl
);
1623 void nir_calc_dominance(nir_shader
*shader
);
1625 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
1626 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
1628 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
1629 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
1631 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
1632 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
1634 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
1635 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
1637 void nir_split_var_copies(nir_shader
*shader
);
1639 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, void *mem_ctx
);
1640 void nir_lower_var_copies(nir_shader
*shader
);
1642 void nir_lower_global_vars_to_local(nir_shader
*shader
);
1644 void nir_lower_locals_to_regs(nir_shader
*shader
);
1646 void nir_assign_var_locations(struct exec_list
*var_list
,
1649 void nir_assign_var_locations_direct_first(nir_shader
*shader
,
1650 struct exec_list
*var_list
,
1651 unsigned *direct_size
,
1655 void nir_lower_io(nir_shader
*shader
, bool is_scalar
);
1657 void nir_lower_vars_to_ssa(nir_shader
*shader
);
1659 void nir_remove_dead_variables(nir_shader
*shader
);
1661 void nir_lower_vec_to_movs(nir_shader
*shader
);
1662 void nir_lower_alu_to_scalar(nir_shader
*shader
);
1663 void nir_lower_load_const_to_scalar(nir_shader
*shader
);
1665 void nir_lower_phis_to_scalar(nir_shader
*shader
);
1667 void nir_lower_samplers(nir_shader
*shader
,
1668 const struct gl_shader_program
*shader_program
,
1669 gl_shader_stage stage
);
1671 void nir_lower_system_values(nir_shader
*shader
);
1672 void nir_lower_tex_projector(nir_shader
*shader
);
1673 void nir_lower_idiv(nir_shader
*shader
);
1675 void nir_lower_atomics(nir_shader
*shader
);
1676 void nir_lower_to_source_mods(nir_shader
*shader
);
1678 void nir_normalize_cubemap_coords(nir_shader
*shader
);
1680 void nir_live_variables_impl(nir_function_impl
*impl
);
1681 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
1683 void nir_convert_to_ssa_impl(nir_function_impl
*impl
);
1684 void nir_convert_to_ssa(nir_shader
*shader
);
1686 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
1687 * registers. If false, convert all values (even those not involved in a phi
1688 * node) to registers.
1690 void nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
1692 bool nir_opt_algebraic(nir_shader
*shader
);
1693 bool nir_opt_algebraic_late(nir_shader
*shader
);
1694 bool nir_opt_constant_folding(nir_shader
*shader
);
1696 bool nir_opt_global_to_local(nir_shader
*shader
);
1698 bool nir_copy_prop_impl(nir_function_impl
*impl
);
1699 bool nir_copy_prop(nir_shader
*shader
);
1701 bool nir_opt_cse(nir_shader
*shader
);
1703 bool nir_opt_dce_impl(nir_function_impl
*impl
);
1704 bool nir_opt_dce(nir_shader
*shader
);
1706 void nir_opt_gcm(nir_shader
*shader
);
1708 bool nir_opt_peephole_select(nir_shader
*shader
);
1709 bool nir_opt_peephole_ffma(nir_shader
*shader
);
1711 bool nir_opt_remove_phis(nir_shader
*shader
);
1713 bool nir_opt_undef(nir_shader
*shader
);
1715 void nir_sweep(nir_shader
*shader
);