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.
308 * Location an atomic counter is stored at.
311 unsigned buffer_index
;
316 * ARB_shader_image_load_store qualifiers.
319 bool read_only
; /**< "readonly" qualifier. */
320 bool write_only
; /**< "writeonly" qualifier. */
325 /** Image internal format if specified explicitly, otherwise GL_NONE. */
330 * Highest element accessed with a constant expression array index
332 * Not used for non-array variables.
334 unsigned max_array_access
;
339 * Built-in state that backs this uniform
341 * Once set at variable creation, \c state_slots must remain invariant.
342 * This is because, ideally, this array would be shared by all clones of
343 * this variable in the IR tree. In other words, we'd really like for it
344 * to be a fly-weight.
346 * If the variable is not a uniform, \c num_state_slots will be zero and
347 * \c state_slots will be \c NULL.
350 unsigned num_state_slots
; /**< Number of state slots used */
351 nir_state_slot
*state_slots
; /**< State descriptors. */
355 * Constant expression assigned in the initializer of the variable
357 nir_constant
*constant_initializer
;
360 * For variables that are in an interface block or are an instance of an
361 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
363 * \sa ir_variable::location
365 const struct glsl_type
*interface_type
;
369 struct exec_node node
;
371 unsigned num_components
; /** < number of vector components */
372 unsigned num_array_elems
; /** < size of array (0 for no array) */
374 /** generic register index. */
377 /** only for debug purposes, can be NULL */
380 /** whether this register is local (per-function) or global (per-shader) */
384 * If this flag is set to true, then accessing channels >= num_components
385 * is well-defined, and simply spills over to the next array element. This
386 * is useful for backends that can do per-component accessing, in
387 * particular scalar backends. By setting this flag and making
388 * num_components equal to 1, structures can be packed tightly into
389 * registers and then registers can be accessed per-component to get to
390 * each structure member, even if it crosses vec4 boundaries.
394 /** set of nir_instr's where this register is used (read from) */
395 struct list_head uses
;
397 /** set of nir_instr's where this register is defined (written to) */
398 struct list_head defs
;
400 /** set of nir_if's where this register is used as a condition */
401 struct list_head if_uses
;
408 nir_instr_type_intrinsic
,
409 nir_instr_type_load_const
,
411 nir_instr_type_ssa_undef
,
413 nir_instr_type_parallel_copy
,
416 typedef struct nir_instr
{
417 struct exec_node node
;
419 struct nir_block
*block
;
421 /* A temporary for optimization and analysis passes to use for storing
422 * flags. For instance, DCE uses this to store the "dead/live" info.
427 static inline nir_instr
*
428 nir_instr_next(nir_instr
*instr
)
430 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
431 if (exec_node_is_tail_sentinel(next
))
434 return exec_node_data(nir_instr
, next
, node
);
437 static inline nir_instr
*
438 nir_instr_prev(nir_instr
*instr
)
440 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
441 if (exec_node_is_head_sentinel(prev
))
444 return exec_node_data(nir_instr
, prev
, node
);
448 nir_instr_is_first(nir_instr
*instr
)
450 return exec_node_is_head_sentinel(exec_node_get_prev(&instr
->node
));
454 nir_instr_is_last(nir_instr
*instr
)
456 return exec_node_is_tail_sentinel(exec_node_get_next(&instr
->node
));
460 /** for debugging only, can be NULL */
463 /** generic SSA definition index. */
466 /** Index into the live_in and live_out bitfields */
469 nir_instr
*parent_instr
;
471 /** set of nir_instr's where this register is used (read from) */
472 struct list_head uses
;
474 /** set of nir_if's where this register is used as a condition */
475 struct list_head if_uses
;
477 uint8_t num_components
;
484 struct nir_src
*indirect
; /** < NULL for no indirect offset */
485 unsigned base_offset
;
487 /* TODO use-def chain goes here */
491 nir_instr
*parent_instr
;
492 struct list_head def_link
;
495 struct nir_src
*indirect
; /** < NULL for no indirect offset */
496 unsigned base_offset
;
498 /* TODO def-use chain goes here */
503 typedef struct nir_src
{
505 nir_instr
*parent_instr
;
506 struct nir_if
*parent_if
;
509 struct list_head use_link
;
520 # define NIR_SRC_INIT nir_src()
522 # define NIR_SRC_INIT (nir_src) { { NULL } }
525 #define nir_foreach_use(reg_or_ssa_def, src) \
526 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
528 #define nir_foreach_use_safe(reg_or_ssa_def, src) \
529 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
531 #define nir_foreach_if_use(reg_or_ssa_def, src) \
532 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
534 #define nir_foreach_if_use_safe(reg_or_ssa_def, src) \
535 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
547 # define NIR_DEST_INIT nir_dest()
549 # define NIR_DEST_INIT (nir_dest) { { { NULL } } }
552 #define nir_foreach_def(reg, dest) \
553 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
555 #define nir_foreach_def_safe(reg, dest) \
556 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
558 static inline nir_src
559 nir_src_for_ssa(nir_ssa_def
*def
)
561 nir_src src
= NIR_SRC_INIT
;
569 static inline nir_src
570 nir_src_for_reg(nir_register
*reg
)
572 nir_src src
= NIR_SRC_INIT
;
576 src
.reg
.indirect
= NULL
;
577 src
.reg
.base_offset
= 0;
582 static inline nir_dest
583 nir_dest_for_reg(nir_register
*reg
)
585 nir_dest dest
= NIR_DEST_INIT
;
592 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *mem_ctx
);
593 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, void *mem_ctx
);
599 * \name input modifiers
603 * For inputs interpreted as floating point, flips the sign bit. For
604 * inputs interpreted as integers, performs the two's complement negation.
609 * Clears the sign bit for floating point values, and computes the integer
610 * absolute value for integers. Note that the negate modifier acts after
611 * the absolute value modifier, therefore if both are set then all inputs
612 * will become negative.
618 * For each input component, says which component of the register it is
619 * chosen from. Note that which elements of the swizzle are used and which
620 * are ignored are based on the write mask for most opcodes - for example,
621 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
622 * a swizzle of {2, x, 1, 0} where x means "don't care."
631 * \name saturate output modifier
633 * Only valid for opcodes that output floating-point numbers. Clamps the
634 * output to between 0.0 and 1.0 inclusive.
639 unsigned write_mask
: 4; /* ignored if dest.is_ssa is true */
642 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
, void *mem_ctx
);
643 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
647 nir_type_invalid
= 0, /* Not a valid type */
655 NIR_OP_IS_COMMUTATIVE
= (1 << 0),
656 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
657 } nir_op_algebraic_property
;
665 * The number of components in the output
667 * If non-zero, this is the size of the output and input sizes are
668 * explicitly given; swizzle and writemask are still in effect, but if
669 * the output component is masked out, then the input component may
672 * If zero, the opcode acts in the standard, per-component manner; the
673 * operation is performed on each component (except the ones that are
674 * masked out) with the input being taken from the input swizzle for
677 * The size of some of the inputs may be given (i.e. non-zero) even
678 * though output_size is zero; in that case, the inputs with a zero
679 * size act per-component, while the inputs with non-zero size don't.
681 unsigned output_size
;
684 * The type of vector that the instruction outputs. Note that the
685 * staurate modifier is only allowed on outputs with the float type.
688 nir_alu_type output_type
;
691 * The number of components in each input
693 unsigned input_sizes
[4];
696 * The type of vector that each input takes. Note that negate and
697 * absolute value are only allowed on inputs with int or float type and
698 * behave differently on the two.
700 nir_alu_type input_types
[4];
702 nir_op_algebraic_property algebraic_properties
;
705 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
707 typedef struct nir_alu_instr
{
714 /* is this source channel used? */
716 nir_alu_instr_channel_used(nir_alu_instr
*instr
, unsigned src
, unsigned channel
)
718 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
719 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
721 return (instr
->dest
.write_mask
>> channel
) & 1;
725 * For instructions whose destinations are SSA, get the number of channels
728 static inline unsigned
729 nir_ssa_alu_instr_src_components(nir_alu_instr
*instr
, unsigned src
)
731 assert(instr
->dest
.dest
.is_ssa
);
733 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
734 return nir_op_infos
[instr
->op
].input_sizes
[src
];
736 return instr
->dest
.dest
.ssa
.num_components
;
741 nir_deref_type_array
,
742 nir_deref_type_struct
745 typedef struct nir_deref
{
746 nir_deref_type deref_type
;
747 struct nir_deref
*child
;
748 const struct glsl_type
*type
;
757 /* This enum describes how the array is referenced. If the deref is
758 * direct then the base_offset is used. If the deref is indirect then then
759 * offset is given by base_offset + indirect. If the deref is a wildcard
760 * then the deref refers to all of the elements of the array at the same
761 * time. Wildcard dereferences are only ever allowed in copy_var
762 * intrinsics and the source and destination derefs must have matching
766 nir_deref_array_type_direct
,
767 nir_deref_array_type_indirect
,
768 nir_deref_array_type_wildcard
,
769 } nir_deref_array_type
;
774 nir_deref_array_type deref_array_type
;
775 unsigned base_offset
;
785 NIR_DEFINE_CAST(nir_deref_as_var
, nir_deref
, nir_deref_var
, deref
)
786 NIR_DEFINE_CAST(nir_deref_as_array
, nir_deref
, nir_deref_array
, deref
)
787 NIR_DEFINE_CAST(nir_deref_as_struct
, nir_deref
, nir_deref_struct
, deref
)
789 /** Returns the tail of a deref chain */
790 static inline nir_deref
*
791 nir_deref_tail(nir_deref
*deref
)
794 deref
= deref
->child
;
802 nir_deref_var
**params
;
803 nir_deref_var
*return_deref
;
805 struct nir_function_overload
*callee
;
808 #define INTRINSIC(name, num_srcs, src_components, has_dest, dest_components, \
809 num_variables, num_indices, flags) \
810 nir_intrinsic_##name,
812 #define LAST_INTRINSIC(name) nir_last_intrinsic = nir_intrinsic_##name,
815 #include "nir_intrinsics.h"
816 nir_num_intrinsics
= nir_last_intrinsic
+ 1
820 #undef LAST_INTRINSIC
822 /** Represents an intrinsic
824 * An intrinsic is an instruction type for handling things that are
825 * more-or-less regular operations but don't just consume and produce SSA
826 * values like ALU operations do. Intrinsics are not for things that have
827 * special semantic meaning such as phi nodes and parallel copies.
828 * Examples of intrinsics include variable load/store operations, system
829 * value loads, and the like. Even though texturing more-or-less falls
830 * under this category, texturing is its own instruction type because
831 * trying to represent texturing with intrinsics would lead to a
832 * combinatorial explosion of intrinsic opcodes.
834 * By having a single instruction type for handling a lot of different
835 * cases, optimization passes can look for intrinsics and, for the most
836 * part, completely ignore them. Each intrinsic type also has a few
837 * possible flags that govern whether or not they can be reordered or
838 * eliminated. That way passes like dead code elimination can still work
839 * on intrisics without understanding the meaning of each.
841 * Each intrinsic has some number of constant indices, some number of
842 * variables, and some number of sources. What these sources, variables,
843 * and indices mean depends on the intrinsic and is documented with the
844 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
845 * instructions are the only types of instruction that can operate on
851 nir_intrinsic_op intrinsic
;
855 /** number of components if this is a vectorized intrinsic
857 * Similarly to ALU operations, some intrinsics are vectorized.
858 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
859 * For vectorized intrinsics, the num_components field specifies the
860 * number of destination components and the number of source components
861 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
863 uint8_t num_components
;
867 nir_deref_var
*variables
[2];
870 } nir_intrinsic_instr
;
873 * \name NIR intrinsics semantic flags
875 * information about what the compiler can do with the intrinsics.
877 * \sa nir_intrinsic_info::flags
881 * whether the intrinsic can be safely eliminated if none of its output
882 * value is not being used.
884 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
887 * Whether the intrinsic can be reordered with respect to any other
888 * intrinsic, i.e. whether the only reordering dependencies of the
889 * intrinsic are due to the register reads/writes.
891 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
892 } nir_intrinsic_semantic_flag
;
894 #define NIR_INTRINSIC_MAX_INPUTS 4
899 unsigned num_srcs
; /** < number of register/SSA inputs */
901 /** number of components of each input register
903 * If this value is 0, the number of components is given by the
904 * num_components field of nir_intrinsic_instr.
906 unsigned src_components
[NIR_INTRINSIC_MAX_INPUTS
];
910 /** number of components of the output register
912 * If this value is 0, the number of components is given by the
913 * num_components field of nir_intrinsic_instr.
915 unsigned dest_components
;
917 /** the number of inputs/outputs that are variables */
918 unsigned num_variables
;
920 /** the number of constant indices used by the intrinsic */
921 unsigned num_indices
;
923 /** semantic flags for calls to this intrinsic */
924 nir_intrinsic_semantic_flag flags
;
925 } nir_intrinsic_info
;
927 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
930 * \group texture information
932 * This gives semantic information about textures which is useful to the
933 * frontend, the backend, and lowering passes, but not the optimizer.
938 nir_tex_src_projector
,
939 nir_tex_src_comparitor
, /* shadow comparitor */
943 nir_tex_src_ms_index
, /* MSAA sample index */
946 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
947 nir_num_tex_src_types
952 nir_tex_src_type src_type
;
956 nir_texop_tex
, /**< Regular texture look-up */
957 nir_texop_txb
, /**< Texture look-up with LOD bias */
958 nir_texop_txl
, /**< Texture look-up with explicit LOD */
959 nir_texop_txd
, /**< Texture look-up with partial derivatvies */
960 nir_texop_txf
, /**< Texel fetch with explicit LOD */
961 nir_texop_txf_ms
, /**< Multisample texture fetch */
962 nir_texop_txs
, /**< Texture size */
963 nir_texop_lod
, /**< Texture lod query */
964 nir_texop_tg4
, /**< Texture gather */
965 nir_texop_query_levels
/**< Texture levels query */
971 enum glsl_sampler_dim sampler_dim
;
972 nir_alu_type dest_type
;
977 unsigned num_srcs
, coord_components
;
978 bool is_array
, is_shadow
;
981 * If is_shadow is true, whether this is the old-style shadow that outputs 4
982 * components or the new-style shadow that outputs 1 component.
984 bool is_new_style_shadow
;
986 /* constant offset - must be 0 if the offset source is used */
989 /* gather component selector */
990 unsigned component
: 2;
992 /* The descriptor set containing this texture */
993 unsigned sampler_set
;
995 /** The sampler index
997 * If this texture instruction has a nir_tex_src_sampler_offset source,
998 * then the sampler index is given by sampler_index + sampler_offset.
1000 unsigned sampler_index
;
1002 /** The size of the sampler array or 0 if it's not an array */
1003 unsigned sampler_array_size
;
1005 nir_deref_var
*sampler
; /* if this is NULL, use sampler_index instead */
1008 static inline unsigned
1009 nir_tex_instr_dest_size(nir_tex_instr
*instr
)
1011 switch (instr
->op
) {
1012 case nir_texop_txs
: {
1014 switch (instr
->sampler_dim
) {
1015 case GLSL_SAMPLER_DIM_1D
:
1016 case GLSL_SAMPLER_DIM_BUF
:
1019 case GLSL_SAMPLER_DIM_2D
:
1020 case GLSL_SAMPLER_DIM_CUBE
:
1021 case GLSL_SAMPLER_DIM_MS
:
1022 case GLSL_SAMPLER_DIM_RECT
:
1023 case GLSL_SAMPLER_DIM_EXTERNAL
:
1026 case GLSL_SAMPLER_DIM_3D
:
1030 unreachable("not reached");
1032 if (instr
->is_array
)
1040 case nir_texop_query_levels
:
1044 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
1051 static inline unsigned
1052 nir_tex_instr_src_size(nir_tex_instr
*instr
, unsigned src
)
1054 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
1055 return instr
->coord_components
;
1058 if (instr
->src
[src
].src_type
== nir_tex_src_offset
||
1059 instr
->src
[src
].src_type
== nir_tex_src_ddx
||
1060 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
1061 if (instr
->is_array
)
1062 return instr
->coord_components
- 1;
1064 return instr
->coord_components
;
1071 nir_tex_instr_src_index(nir_tex_instr
*instr
, nir_tex_src_type type
)
1073 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
1074 if (instr
->src
[i
].src_type
== type
)
1091 nir_const_value value
;
1094 } nir_load_const_instr
;
1107 /* creates a new SSA variable in an undefined state */
1112 } nir_ssa_undef_instr
;
1115 struct exec_node node
;
1117 /* The predecessor block corresponding to this source */
1118 struct nir_block
*pred
;
1123 #define nir_foreach_phi_src(phi, entry) \
1124 foreach_list_typed(nir_phi_src, entry, node, &(phi)->srcs)
1129 struct exec_list srcs
; /** < list of nir_phi_src */
1135 struct exec_node node
;
1138 } nir_parallel_copy_entry
;
1140 #define nir_foreach_parallel_copy_entry(pcopy, entry) \
1141 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
1146 /* A list of nir_parallel_copy_entry's. The sources of all of the
1147 * entries are copied to the corresponding destinations "in parallel".
1148 * In other words, if we have two entries: a -> b and b -> a, the values
1151 struct exec_list entries
;
1152 } nir_parallel_copy_instr
;
1154 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
)
1155 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
)
1156 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
)
1157 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
)
1158 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
)
1159 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
)
1160 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
)
1161 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
)
1162 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
1163 nir_parallel_copy_instr
, instr
)
1168 * Control flow consists of a tree of control flow nodes, which include
1169 * if-statements and loops. The leaves of the tree are basic blocks, lists of
1170 * instructions that always run start-to-finish. Each basic block also keeps
1171 * track of its successors (blocks which may run immediately after the current
1172 * block) and predecessors (blocks which could have run immediately before the
1173 * current block). Each function also has a start block and an end block which
1174 * all return statements point to (which is always empty). Together, all the
1175 * blocks with their predecessors and successors make up the control flow
1176 * graph (CFG) of the function. There are helpers that modify the tree of
1177 * control flow nodes while modifying the CFG appropriately; these should be
1178 * used instead of modifying the tree directly.
1185 nir_cf_node_function
1188 typedef struct nir_cf_node
{
1189 struct exec_node node
;
1190 nir_cf_node_type type
;
1191 struct nir_cf_node
*parent
;
1194 typedef struct nir_block
{
1195 nir_cf_node cf_node
;
1197 struct exec_list instr_list
; /** < list of nir_instr */
1199 /** generic block index; generated by nir_index_blocks */
1203 * Each block can only have up to 2 successors, so we put them in a simple
1204 * array - no need for anything more complicated.
1206 struct nir_block
*successors
[2];
1208 /* Set of nir_block predecessors in the CFG */
1209 struct set
*predecessors
;
1212 * this node's immediate dominator in the dominance tree - set to NULL for
1215 struct nir_block
*imm_dom
;
1217 /* This node's children in the dominance tree */
1218 unsigned num_dom_children
;
1219 struct nir_block
**dom_children
;
1221 /* Set of nir_block's on the dominance frontier of this block */
1222 struct set
*dom_frontier
;
1225 * These two indices have the property that dom_{pre,post}_index for each
1226 * child of this block in the dominance tree will always be between
1227 * dom_pre_index and dom_post_index for this block, which makes testing if
1228 * a given block is dominated by another block an O(1) operation.
1230 unsigned dom_pre_index
, dom_post_index
;
1232 /* live in and out for this block; used for liveness analysis */
1233 BITSET_WORD
*live_in
;
1234 BITSET_WORD
*live_out
;
1237 static inline nir_instr
*
1238 nir_block_first_instr(nir_block
*block
)
1240 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
1241 return exec_node_data(nir_instr
, head
, node
);
1244 static inline nir_instr
*
1245 nir_block_last_instr(nir_block
*block
)
1247 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
1248 return exec_node_data(nir_instr
, tail
, node
);
1251 #define nir_foreach_instr(block, instr) \
1252 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
1253 #define nir_foreach_instr_reverse(block, instr) \
1254 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
1255 #define nir_foreach_instr_safe(block, instr) \
1256 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
1257 #define nir_foreach_instr_safe_reverse(block, instr) \
1258 foreach_list_typed_safe_reverse(nir_instr, instr, node, &(block)->instr_list)
1260 typedef struct nir_if
{
1261 nir_cf_node cf_node
;
1264 struct exec_list then_list
; /** < list of nir_cf_node */
1265 struct exec_list else_list
; /** < list of nir_cf_node */
1268 static inline nir_cf_node
*
1269 nir_if_first_then_node(nir_if
*if_stmt
)
1271 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
1272 return exec_node_data(nir_cf_node
, head
, node
);
1275 static inline nir_cf_node
*
1276 nir_if_last_then_node(nir_if
*if_stmt
)
1278 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
1279 return exec_node_data(nir_cf_node
, tail
, node
);
1282 static inline nir_cf_node
*
1283 nir_if_first_else_node(nir_if
*if_stmt
)
1285 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
1286 return exec_node_data(nir_cf_node
, head
, node
);
1289 static inline nir_cf_node
*
1290 nir_if_last_else_node(nir_if
*if_stmt
)
1292 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
1293 return exec_node_data(nir_cf_node
, tail
, node
);
1297 nir_cf_node cf_node
;
1299 struct exec_list body
; /** < list of nir_cf_node */
1302 static inline nir_cf_node
*
1303 nir_loop_first_cf_node(nir_loop
*loop
)
1305 return exec_node_data(nir_cf_node
, exec_list_get_head(&loop
->body
), node
);
1308 static inline nir_cf_node
*
1309 nir_loop_last_cf_node(nir_loop
*loop
)
1311 return exec_node_data(nir_cf_node
, exec_list_get_tail(&loop
->body
), node
);
1315 * Various bits of metadata that can may be created or required by
1316 * optimization and analysis passes
1319 nir_metadata_none
= 0x0,
1320 nir_metadata_block_index
= 0x1,
1321 nir_metadata_dominance
= 0x2,
1322 nir_metadata_live_variables
= 0x4,
1326 nir_cf_node cf_node
;
1328 /** pointer to the overload of which this is an implementation */
1329 struct nir_function_overload
*overload
;
1331 struct exec_list body
; /** < list of nir_cf_node */
1333 nir_block
*start_block
, *end_block
;
1335 /** list for all local variables in the function */
1336 struct exec_list locals
;
1338 /** array of variables used as parameters */
1339 unsigned num_params
;
1340 nir_variable
**params
;
1342 /** variable used to hold the result of the function */
1343 nir_variable
*return_var
;
1345 /** list of local registers in the function */
1346 struct exec_list registers
;
1348 /** next available local register index */
1351 /** next available SSA value index */
1354 /* total number of basic blocks, only valid when block_index_dirty = false */
1355 unsigned num_blocks
;
1357 nir_metadata valid_metadata
;
1358 } nir_function_impl
;
1360 static inline nir_cf_node
*
1361 nir_cf_node_next(nir_cf_node
*node
)
1363 struct exec_node
*next
= exec_node_get_next(&node
->node
);
1364 if (exec_node_is_tail_sentinel(next
))
1367 return exec_node_data(nir_cf_node
, next
, node
);
1370 static inline nir_cf_node
*
1371 nir_cf_node_prev(nir_cf_node
*node
)
1373 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
1374 if (exec_node_is_head_sentinel(prev
))
1377 return exec_node_data(nir_cf_node
, prev
, node
);
1381 nir_cf_node_is_first(const nir_cf_node
*node
)
1383 return exec_node_is_head_sentinel(node
->node
.prev
);
1387 nir_cf_node_is_last(const nir_cf_node
*node
)
1389 return exec_node_is_tail_sentinel(node
->node
.next
);
1392 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
)
1393 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
)
1394 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
)
1395 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
, nir_function_impl
, cf_node
)
1400 nir_parameter_inout
,
1401 } nir_parameter_type
;
1404 nir_parameter_type param_type
;
1405 const struct glsl_type
*type
;
1408 typedef struct nir_function_overload
{
1409 struct exec_node node
;
1411 unsigned num_params
;
1412 nir_parameter
*params
;
1413 const struct glsl_type
*return_type
;
1415 nir_function_impl
*impl
; /** < NULL if the overload is only declared yet */
1417 /** pointer to the function of which this is an overload */
1418 struct nir_function
*function
;
1419 } nir_function_overload
;
1421 typedef struct nir_function
{
1422 struct exec_node node
;
1424 struct exec_list overload_list
; /** < list of nir_function_overload */
1426 struct nir_shader
*shader
;
1429 #define nir_function_first_overload(func) \
1430 exec_node_data(nir_function_overload, \
1431 exec_list_get_head(&(func)->overload_list), node)
1433 typedef struct nir_shader_compiler_options
{
1439 /** lowers fneg and ineg to fsub and isub. */
1441 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
1444 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
1448 * Does the driver support real 32-bit integers? (Otherwise, integers
1449 * are simulated by floats.)
1451 bool native_integers
;
1452 } nir_shader_compiler_options
;
1454 typedef struct nir_shader
{
1455 /** hash table of name -> uniform nir_variable */
1456 struct exec_list uniforms
;
1458 /** hash table of name -> input nir_variable */
1459 struct exec_list inputs
;
1461 /** hash table of name -> output nir_variable */
1462 struct exec_list outputs
;
1464 /** Set of driver-specific options for the shader.
1466 * The memory for the options is expected to be kept in a single static
1467 * copy by the driver.
1469 const struct nir_shader_compiler_options
*options
;
1471 /** list of global variables in the shader */
1472 struct exec_list globals
;
1474 /** list of system value variables in the shader */
1475 struct exec_list system_values
;
1477 struct exec_list functions
; /** < list of nir_function */
1479 /** list of global register in the shader */
1480 struct exec_list registers
;
1482 /** next available global register index */
1486 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
1489 unsigned num_inputs
, num_uniforms
, num_outputs
;
1491 /** the number of uniforms that are only accessed directly */
1492 unsigned num_direct_uniforms
;
1495 #define nir_foreach_overload(shader, overload) \
1496 foreach_list_typed(nir_function, func, node, &(shader)->functions) \
1497 foreach_list_typed(nir_function_overload, overload, node, \
1498 &(func)->overload_list)
1500 nir_shader
*nir_shader_create(void *mem_ctx
,
1501 const nir_shader_compiler_options
*options
);
1503 /** creates a register, including assigning it an index and adding it to the list */
1504 nir_register
*nir_global_reg_create(nir_shader
*shader
);
1506 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
1508 void nir_reg_remove(nir_register
*reg
);
1510 /** creates a function and adds it to the shader's list of functions */
1511 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
1513 /** creates a null function returning null */
1514 nir_function_overload
*nir_function_overload_create(nir_function
*func
);
1516 nir_function_impl
*nir_function_impl_create(nir_function_overload
*func
);
1518 nir_block
*nir_block_create(void *mem_ctx
);
1519 nir_if
*nir_if_create(void *mem_ctx
);
1520 nir_loop
*nir_loop_create(void *mem_ctx
);
1522 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
1524 /** puts a control flow node immediately after another control flow node */
1525 void nir_cf_node_insert_after(nir_cf_node
*node
, nir_cf_node
*after
);
1527 /** puts a control flow node immediately before another control flow node */
1528 void nir_cf_node_insert_before(nir_cf_node
*node
, nir_cf_node
*before
);
1530 /** puts a control flow node at the beginning of a list from an if, loop, or function */
1531 void nir_cf_node_insert_begin(struct exec_list
*list
, nir_cf_node
*node
);
1533 /** puts a control flow node at the end of a list from an if, loop, or function */
1534 void nir_cf_node_insert_end(struct exec_list
*list
, nir_cf_node
*node
);
1536 /** removes a control flow node, doing any cleanup necessary */
1537 void nir_cf_node_remove(nir_cf_node
*node
);
1539 /** requests that the given pieces of metadata be generated */
1540 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
);
1541 /** dirties all but the preserved metadata */
1542 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
1544 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
1545 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
1547 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
1549 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
1550 unsigned num_components
);
1552 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
1553 nir_intrinsic_op op
);
1555 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
1556 nir_function_overload
*callee
);
1558 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
1560 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
1562 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
1564 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
1565 unsigned num_components
);
1567 nir_deref_var
*nir_deref_var_create(void *mem_ctx
, nir_variable
*var
);
1568 nir_deref_array
*nir_deref_array_create(void *mem_ctx
);
1569 nir_deref_struct
*nir_deref_struct_create(void *mem_ctx
, unsigned field_index
);
1571 nir_deref
*nir_copy_deref(void *mem_ctx
, nir_deref
*deref
);
1573 nir_load_const_instr
*
1574 nir_deref_get_const_initializer_load(nir_shader
*shader
, nir_deref_var
*deref
);
1576 void nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
);
1577 void nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
);
1579 void nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
);
1580 void nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
);
1582 void nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
);
1583 void nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
);
1585 void nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
);
1586 void nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
);
1588 void nir_instr_remove(nir_instr
*instr
);
1590 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
1591 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
1592 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
1593 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
1595 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
1596 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
1598 nir_const_value
*nir_src_as_const_value(nir_src src
);
1599 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
1600 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
1601 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
1602 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
1604 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
1605 unsigned num_components
, const char *name
);
1606 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
1607 unsigned num_components
, const char *name
);
1608 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
, void *mem_ctx
);
1610 /* visits basic blocks in source-code order */
1611 typedef bool (*nir_foreach_block_cb
)(nir_block
*block
, void *state
);
1612 bool nir_foreach_block(nir_function_impl
*impl
, nir_foreach_block_cb cb
,
1614 bool nir_foreach_block_reverse(nir_function_impl
*impl
, nir_foreach_block_cb cb
,
1617 /* If the following CF node is an if, this function returns that if.
1618 * Otherwise, it returns NULL.
1620 nir_if
*nir_block_get_following_if(nir_block
*block
);
1622 void nir_index_local_regs(nir_function_impl
*impl
);
1623 void nir_index_global_regs(nir_shader
*shader
);
1624 void nir_index_ssa_defs(nir_function_impl
*impl
);
1626 void nir_index_blocks(nir_function_impl
*impl
);
1628 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
1629 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
1632 void nir_validate_shader(nir_shader
*shader
);
1634 static inline void nir_validate_shader(nir_shader
*shader
) { (void) shader
; }
1637 void nir_calc_dominance_impl(nir_function_impl
*impl
);
1638 void nir_calc_dominance(nir_shader
*shader
);
1640 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
1641 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
1643 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
1644 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
1646 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
1647 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
1649 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
1650 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
1652 void nir_split_var_copies(nir_shader
*shader
);
1654 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, void *mem_ctx
);
1655 void nir_lower_var_copies(nir_shader
*shader
);
1657 void nir_lower_global_vars_to_local(nir_shader
*shader
);
1659 void nir_lower_locals_to_regs(nir_shader
*shader
);
1661 void nir_assign_var_locations(struct exec_list
*var_list
,
1664 void nir_assign_var_locations_direct_first(nir_shader
*shader
,
1665 struct exec_list
*var_list
,
1666 unsigned *direct_size
,
1670 void nir_lower_io(nir_shader
*shader
, bool is_scalar
);
1672 void nir_lower_vars_to_ssa(nir_shader
*shader
);
1674 void nir_remove_dead_variables(nir_shader
*shader
);
1676 void nir_lower_vec_to_movs(nir_shader
*shader
);
1677 void nir_lower_alu_to_scalar(nir_shader
*shader
);
1678 void nir_lower_load_const_to_scalar(nir_shader
*shader
);
1680 void nir_lower_phis_to_scalar(nir_shader
*shader
);
1682 void nir_lower_samplers(nir_shader
*shader
,
1683 const struct gl_shader_program
*shader_program
,
1684 gl_shader_stage stage
);
1685 void nir_lower_samplers_for_vk(nir_shader
*shader
);
1687 void nir_lower_system_values(nir_shader
*shader
);
1688 void nir_lower_tex_projector(nir_shader
*shader
);
1689 void nir_lower_idiv(nir_shader
*shader
);
1691 void nir_lower_atomics(nir_shader
*shader
);
1692 void nir_lower_to_source_mods(nir_shader
*shader
);
1694 void nir_normalize_cubemap_coords(nir_shader
*shader
);
1696 void nir_live_variables_impl(nir_function_impl
*impl
);
1697 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
1699 void nir_convert_to_ssa_impl(nir_function_impl
*impl
);
1700 void nir_convert_to_ssa(nir_shader
*shader
);
1702 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
1703 * registers. If false, convert all values (even those not involved in a phi
1704 * node) to registers.
1706 void nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
1708 bool nir_opt_algebraic(nir_shader
*shader
);
1709 bool nir_opt_algebraic_late(nir_shader
*shader
);
1710 bool nir_opt_constant_folding(nir_shader
*shader
);
1712 bool nir_opt_global_to_local(nir_shader
*shader
);
1714 bool nir_copy_prop_impl(nir_function_impl
*impl
);
1715 bool nir_copy_prop(nir_shader
*shader
);
1717 bool nir_opt_cse(nir_shader
*shader
);
1719 bool nir_opt_dce_impl(nir_function_impl
*impl
);
1720 bool nir_opt_dce(nir_shader
*shader
);
1722 void nir_opt_gcm(nir_shader
*shader
);
1724 bool nir_opt_peephole_select(nir_shader
*shader
);
1725 bool nir_opt_peephole_ffma(nir_shader
*shader
);
1727 bool nir_opt_remove_phis(nir_shader
*shader
);
1729 bool nir_opt_undef(nir_shader
*shader
);
1731 void nir_sweep(nir_shader
*shader
);