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)
31 #include "util/hash_table.h"
32 #include "compiler/glsl/list.h"
33 #include "GL/gl.h" /* GLenum */
34 #include "util/list.h"
35 #include "util/ralloc.h"
37 #include "util/bitscan.h"
38 #include "util/bitset.h"
39 #include "util/macros.h"
40 #include "compiler/nir_types.h"
41 #include "compiler/shader_enums.h"
42 #include "compiler/shader_info.h"
46 #include "util/debug.h"
49 #include "nir_opcodes.h"
51 #if defined(_WIN32) && !defined(snprintf)
52 #define snprintf _snprintf
60 #define NIR_TRUE (~0u)
61 #define NIR_MAX_VEC_COMPONENTS 4
62 #define NIR_MAX_MATRIX_COLUMNS 4
63 typedef uint8_t nir_component_mask_t
;
65 /** Defines a cast function
67 * This macro defines a cast function from in_type to out_type where
68 * out_type is some structure type that contains a field of type out_type.
70 * Note that you have to be a bit careful as the generated cast function
73 #define NIR_DEFINE_CAST(name, in_type, out_type, field, \
74 type_field, type_value) \
75 static inline out_type * \
76 name(const in_type *parent) \
78 assert(parent && parent->type_field == type_value); \
79 return exec_node_data(out_type, parent, field); \
89 * Description of built-in state associated with a uniform
91 * \sa nir_variable::state_slots
94 gl_state_index16 tokens
[STATE_LENGTH
];
99 nir_var_shader_in
= (1 << 0),
100 nir_var_shader_out
= (1 << 1),
101 nir_var_shader_temp
= (1 << 2),
102 nir_var_function_temp
= (1 << 3),
103 nir_var_uniform
= (1 << 4),
104 nir_var_mem_ubo
= (1 << 5),
105 nir_var_system_value
= (1 << 6),
106 nir_var_mem_ssbo
= (1 << 7),
107 nir_var_mem_shared
= (1 << 8),
108 nir_var_mem_global
= (1 << 9),
116 nir_rounding_mode_undef
= 0,
117 nir_rounding_mode_rtne
= 1, /* round to nearest even */
118 nir_rounding_mode_ru
= 2, /* round up */
119 nir_rounding_mode_rd
= 3, /* round down */
120 nir_rounding_mode_rtz
= 4, /* round towards zero */
137 #define nir_const_value_to_array(arr, c, components, m) \
139 for (unsigned i = 0; i < components; ++i) \
143 typedef struct nir_constant
{
145 * Value of the constant.
147 * The field used to back the values supplied by the constant is determined
148 * by the type associated with the \c nir_variable. Constants may be
149 * scalars, vectors, or matrices.
151 nir_const_value values
[NIR_MAX_VEC_COMPONENTS
];
153 /* we could get this from the var->type but makes clone *much* easier to
154 * not have to care about the type.
156 unsigned num_elements
;
158 /* Array elements / Structure Fields */
159 struct nir_constant
**elements
;
163 * \brief Layout qualifiers for gl_FragDepth.
165 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
166 * with a layout qualifier.
169 nir_depth_layout_none
, /**< No depth layout is specified. */
170 nir_depth_layout_any
,
171 nir_depth_layout_greater
,
172 nir_depth_layout_less
,
173 nir_depth_layout_unchanged
177 * Enum keeping track of how a variable was declared.
181 * Normal declaration.
183 nir_var_declared_normally
= 0,
186 * Variable is implicitly generated by the compiler and should not be
187 * visible via the API.
190 } nir_var_declaration_type
;
193 * Either a uniform, global variable, shader input, or shader output. Based on
194 * ir_variable - it should be easy to translate between the two.
197 typedef struct nir_variable
{
198 struct exec_node node
;
201 * Declared type of the variable
203 const struct glsl_type
*type
;
206 * Declared name of the variable
210 struct nir_variable_data
{
212 * Storage class of the variable.
214 * \sa nir_variable_mode
216 nir_variable_mode mode
;
219 * Is the variable read-only?
221 * This is set for variables declared as \c const, shader inputs,
224 unsigned read_only
:1;
228 unsigned invariant
:1;
231 * When separate shader programs are enabled, only input/outputs between
232 * the stages of a multi-stage separate program can be safely removed
233 * from the shader interface. Other input/outputs must remains active.
235 * This is also used to make sure xfb varyings that are unused by the
236 * fragment shader are not removed.
238 unsigned always_active_io
:1;
241 * Interpolation mode for shader inputs / outputs
243 * \sa glsl_interp_mode
245 unsigned interpolation
:2;
248 * If non-zero, then this variable may be packed along with other variables
249 * into a single varying slot, so this offset should be applied when
250 * accessing components. For example, an offset of 1 means that the x
251 * component of this variable is actually stored in component y of the
252 * location specified by \c location.
254 unsigned location_frac
:2;
257 * If true, this variable represents an array of scalars that should
258 * be tightly packed. In other words, consecutive array elements
259 * should be stored one component apart, rather than one slot apart.
264 * Whether this is a fragment shader output implicitly initialized with
265 * the previous contents of the specified render target at the
266 * framebuffer location corresponding to this shader invocation.
268 unsigned fb_fetch_output
:1;
271 * Non-zero if this variable is considered bindless as defined by
272 * ARB_bindless_texture.
277 * Was an explicit binding set in the shader?
279 unsigned explicit_binding
:1;
282 * Was a transfer feedback buffer set in the shader?
284 unsigned explicit_xfb_buffer
:1;
287 * Was a transfer feedback stride set in the shader?
289 unsigned explicit_xfb_stride
:1;
292 * Was an explicit offset set in the shader?
294 unsigned explicit_offset
:1;
297 * \brief Layout qualifier for gl_FragDepth.
299 * This is not equal to \c ir_depth_layout_none if and only if this
300 * variable is \c gl_FragDepth and a layout qualifier is specified.
302 nir_depth_layout depth_layout
;
305 * Storage location of the base of this variable
307 * The precise meaning of this field depends on the nature of the variable.
309 * - Vertex shader input: one of the values from \c gl_vert_attrib.
310 * - Vertex shader output: one of the values from \c gl_varying_slot.
311 * - Geometry shader input: one of the values from \c gl_varying_slot.
312 * - Geometry shader output: one of the values from \c gl_varying_slot.
313 * - Fragment shader input: one of the values from \c gl_varying_slot.
314 * - Fragment shader output: one of the values from \c gl_frag_result.
315 * - Uniforms: Per-stage uniform slot number for default uniform block.
316 * - Uniforms: Index within the uniform block definition for UBO members.
317 * - Non-UBO Uniforms: uniform slot number.
318 * - Other: This field is not currently used.
320 * If the variable is a uniform, shader input, or shader output, and the
321 * slot has not been assigned, the value will be -1.
326 * The actual location of the variable in the IR. Only valid for inputs
329 unsigned int driver_location
;
332 * Vertex stream output identifier.
334 * For packed outputs, bit 31 is set and bits [2*i+1,2*i] indicate the
335 * stream of the i-th component.
340 * output index for dual source blending.
345 * Descriptor set binding for sampler or UBO.
350 * Initial binding point for a sampler or UBO.
352 * For array types, this represents the binding point for the first element.
357 * Location an atomic counter or transform feedback is stored at.
362 * Transform feedback buffer.
367 * Transform feedback stride.
372 * How the variable was declared. See nir_var_declaration_type.
374 * This is used to detect variables generated by the compiler, so should
375 * not be visible via the API.
377 unsigned how_declared
:2;
380 * ARB_shader_image_load_store qualifiers.
383 enum gl_access_qualifier access
;
385 /** Image internal format if specified explicitly, otherwise GL_NONE. */
391 * Built-in state that backs this uniform
393 * Once set at variable creation, \c state_slots must remain invariant.
394 * This is because, ideally, this array would be shared by all clones of
395 * this variable in the IR tree. In other words, we'd really like for it
396 * to be a fly-weight.
398 * If the variable is not a uniform, \c num_state_slots will be zero and
399 * \c state_slots will be \c NULL.
402 unsigned num_state_slots
; /**< Number of state slots used */
403 nir_state_slot
*state_slots
; /**< State descriptors. */
407 * Constant expression assigned in the initializer of the variable
409 * This field should only be used temporarily by creators of NIR shaders
410 * and then lower_constant_initializers can be used to get rid of them.
411 * Most of the rest of NIR ignores this field or asserts that it's NULL.
413 nir_constant
*constant_initializer
;
416 * For variables that are in an interface block or are an instance of an
417 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
419 * \sa ir_variable::location
421 const struct glsl_type
*interface_type
;
424 * Description of per-member data for per-member struct variables
426 * This is used for variables which are actually an amalgamation of
427 * multiple entities such as a struct of built-in values or a struct of
428 * inputs each with their own layout specifier. This is only allowed on
429 * variables with a struct or array of array of struct type.
431 unsigned num_members
;
432 struct nir_variable_data
*members
;
435 #define nir_foreach_variable(var, var_list) \
436 foreach_list_typed(nir_variable, var, node, var_list)
438 #define nir_foreach_variable_safe(var, var_list) \
439 foreach_list_typed_safe(nir_variable, var, node, var_list)
442 nir_variable_is_global(const nir_variable
*var
)
444 return var
->data
.mode
!= nir_var_function_temp
;
447 typedef struct nir_register
{
448 struct exec_node node
;
450 unsigned num_components
; /** < number of vector components */
451 unsigned num_array_elems
; /** < size of array (0 for no array) */
453 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
456 /** generic register index. */
459 /** only for debug purposes, can be NULL */
462 /** set of nir_srcs where this register is used (read from) */
463 struct list_head uses
;
465 /** set of nir_dests where this register is defined (written to) */
466 struct list_head defs
;
468 /** set of nir_ifs where this register is used as a condition */
469 struct list_head if_uses
;
472 #define nir_foreach_register(reg, reg_list) \
473 foreach_list_typed(nir_register, reg, node, reg_list)
474 #define nir_foreach_register_safe(reg, reg_list) \
475 foreach_list_typed_safe(nir_register, reg, node, reg_list)
477 typedef enum PACKED
{
479 nir_instr_type_deref
,
482 nir_instr_type_intrinsic
,
483 nir_instr_type_load_const
,
485 nir_instr_type_ssa_undef
,
487 nir_instr_type_parallel_copy
,
490 typedef struct nir_instr
{
491 struct exec_node node
;
492 struct nir_block
*block
;
495 /* A temporary for optimization and analysis passes to use for storing
496 * flags. For instance, DCE uses this to store the "dead/live" info.
500 /** generic instruction index. */
504 static inline nir_instr
*
505 nir_instr_next(nir_instr
*instr
)
507 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
508 if (exec_node_is_tail_sentinel(next
))
511 return exec_node_data(nir_instr
, next
, node
);
514 static inline nir_instr
*
515 nir_instr_prev(nir_instr
*instr
)
517 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
518 if (exec_node_is_head_sentinel(prev
))
521 return exec_node_data(nir_instr
, prev
, node
);
525 nir_instr_is_first(const nir_instr
*instr
)
527 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr
->node
));
531 nir_instr_is_last(const nir_instr
*instr
)
533 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr
->node
));
536 typedef struct nir_ssa_def
{
537 /** for debugging only, can be NULL */
540 /** generic SSA definition index. */
543 /** Index into the live_in and live_out bitfields */
546 /** Instruction which produces this SSA value. */
547 nir_instr
*parent_instr
;
549 /** set of nir_instrs where this register is used (read from) */
550 struct list_head uses
;
552 /** set of nir_ifs where this register is used as a condition */
553 struct list_head if_uses
;
555 uint8_t num_components
;
557 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
565 struct nir_src
*indirect
; /** < NULL for no indirect offset */
566 unsigned base_offset
;
568 /* TODO use-def chain goes here */
572 nir_instr
*parent_instr
;
573 struct list_head def_link
;
576 struct nir_src
*indirect
; /** < NULL for no indirect offset */
577 unsigned base_offset
;
579 /* TODO def-use chain goes here */
584 typedef struct nir_src
{
586 /** Instruction that consumes this value as a source. */
587 nir_instr
*parent_instr
;
588 struct nir_if
*parent_if
;
591 struct list_head use_link
;
601 static inline nir_src
604 nir_src src
= { { NULL
} };
608 #define NIR_SRC_INIT nir_src_init()
610 #define nir_foreach_use(src, reg_or_ssa_def) \
611 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
613 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
614 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
616 #define nir_foreach_if_use(src, reg_or_ssa_def) \
617 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
619 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
620 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
631 static inline nir_dest
634 nir_dest dest
= { { { NULL
} } };
638 #define NIR_DEST_INIT nir_dest_init()
640 #define nir_foreach_def(dest, reg) \
641 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
643 #define nir_foreach_def_safe(dest, reg) \
644 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
646 static inline nir_src
647 nir_src_for_ssa(nir_ssa_def
*def
)
649 nir_src src
= NIR_SRC_INIT
;
657 static inline nir_src
658 nir_src_for_reg(nir_register
*reg
)
660 nir_src src
= NIR_SRC_INIT
;
664 src
.reg
.indirect
= NULL
;
665 src
.reg
.base_offset
= 0;
670 static inline nir_dest
671 nir_dest_for_reg(nir_register
*reg
)
673 nir_dest dest
= NIR_DEST_INIT
;
680 static inline unsigned
681 nir_src_bit_size(nir_src src
)
683 return src
.is_ssa
? src
.ssa
->bit_size
: src
.reg
.reg
->bit_size
;
686 static inline unsigned
687 nir_src_num_components(nir_src src
)
689 return src
.is_ssa
? src
.ssa
->num_components
: src
.reg
.reg
->num_components
;
693 nir_src_is_const(nir_src src
)
696 src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
;
699 int64_t nir_src_as_int(nir_src src
);
700 uint64_t nir_src_as_uint(nir_src src
);
701 bool nir_src_as_bool(nir_src src
);
702 double nir_src_as_float(nir_src src
);
703 int64_t nir_src_comp_as_int(nir_src src
, unsigned component
);
704 uint64_t nir_src_comp_as_uint(nir_src src
, unsigned component
);
705 bool nir_src_comp_as_bool(nir_src src
, unsigned component
);
706 double nir_src_comp_as_float(nir_src src
, unsigned component
);
708 static inline unsigned
709 nir_dest_bit_size(nir_dest dest
)
711 return dest
.is_ssa
? dest
.ssa
.bit_size
: dest
.reg
.reg
->bit_size
;
714 static inline unsigned
715 nir_dest_num_components(nir_dest dest
)
717 return dest
.is_ssa
? dest
.ssa
.num_components
: dest
.reg
.reg
->num_components
;
720 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
721 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
727 * \name input modifiers
731 * For inputs interpreted as floating point, flips the sign bit. For
732 * inputs interpreted as integers, performs the two's complement negation.
737 * Clears the sign bit for floating point values, and computes the integer
738 * absolute value for integers. Note that the negate modifier acts after
739 * the absolute value modifier, therefore if both are set then all inputs
740 * will become negative.
746 * For each input component, says which component of the register it is
747 * chosen from. Note that which elements of the swizzle are used and which
748 * are ignored are based on the write mask for most opcodes - for example,
749 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
750 * a swizzle of {2, x, 1, 0} where x means "don't care."
752 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
759 * \name saturate output modifier
761 * Only valid for opcodes that output floating-point numbers. Clamps the
762 * output to between 0.0 and 1.0 inclusive.
767 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
770 /** NIR sized and unsized types
772 * The values in this enum are carefully chosen so that the sized type is
773 * just the unsized type OR the number of bits.
776 nir_type_invalid
= 0, /* Not a valid type */
780 nir_type_float
= 128,
781 nir_type_bool1
= 1 | nir_type_bool
,
782 nir_type_bool32
= 32 | nir_type_bool
,
783 nir_type_int1
= 1 | nir_type_int
,
784 nir_type_int8
= 8 | nir_type_int
,
785 nir_type_int16
= 16 | nir_type_int
,
786 nir_type_int32
= 32 | nir_type_int
,
787 nir_type_int64
= 64 | nir_type_int
,
788 nir_type_uint1
= 1 | nir_type_uint
,
789 nir_type_uint8
= 8 | nir_type_uint
,
790 nir_type_uint16
= 16 | nir_type_uint
,
791 nir_type_uint32
= 32 | nir_type_uint
,
792 nir_type_uint64
= 64 | nir_type_uint
,
793 nir_type_float16
= 16 | nir_type_float
,
794 nir_type_float32
= 32 | nir_type_float
,
795 nir_type_float64
= 64 | nir_type_float
,
798 #define NIR_ALU_TYPE_SIZE_MASK 0x79
799 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
801 static inline unsigned
802 nir_alu_type_get_type_size(nir_alu_type type
)
804 return type
& NIR_ALU_TYPE_SIZE_MASK
;
807 static inline unsigned
808 nir_alu_type_get_base_type(nir_alu_type type
)
810 return type
& NIR_ALU_TYPE_BASE_TYPE_MASK
;
813 static inline nir_alu_type
814 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
818 return nir_type_bool1
;
821 return nir_type_uint32
;
824 return nir_type_int32
;
826 case GLSL_TYPE_UINT16
:
827 return nir_type_uint16
;
829 case GLSL_TYPE_INT16
:
830 return nir_type_int16
;
832 case GLSL_TYPE_UINT8
:
833 return nir_type_uint8
;
835 return nir_type_int8
;
836 case GLSL_TYPE_UINT64
:
837 return nir_type_uint64
;
839 case GLSL_TYPE_INT64
:
840 return nir_type_int64
;
842 case GLSL_TYPE_FLOAT
:
843 return nir_type_float32
;
845 case GLSL_TYPE_FLOAT16
:
846 return nir_type_float16
;
848 case GLSL_TYPE_DOUBLE
:
849 return nir_type_float64
;
852 case GLSL_TYPE_SAMPLER
:
853 case GLSL_TYPE_IMAGE
:
854 case GLSL_TYPE_ATOMIC_UINT
:
855 case GLSL_TYPE_STRUCT
:
856 case GLSL_TYPE_INTERFACE
:
857 case GLSL_TYPE_ARRAY
:
859 case GLSL_TYPE_SUBROUTINE
:
860 case GLSL_TYPE_FUNCTION
:
861 case GLSL_TYPE_ERROR
:
862 return nir_type_invalid
;
865 unreachable("unknown type");
868 static inline nir_alu_type
869 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
871 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
874 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
875 nir_rounding_mode rnd
);
878 nir_op_vec(unsigned components
)
880 switch (components
) {
881 case 1: return nir_op_mov
;
882 case 2: return nir_op_vec2
;
883 case 3: return nir_op_vec3
;
884 case 4: return nir_op_vec4
;
885 default: unreachable("bad component count");
891 * Operation where the first two sources are commutative.
893 * For 2-source operations, this just mathematical commutativity. Some
894 * 3-source operations, like ffma, are only commutative in the first two
897 NIR_OP_IS_2SRC_COMMUTATIVE
= (1 << 0),
898 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
899 } nir_op_algebraic_property
;
907 * The number of components in the output
909 * If non-zero, this is the size of the output and input sizes are
910 * explicitly given; swizzle and writemask are still in effect, but if
911 * the output component is masked out, then the input component may
914 * If zero, the opcode acts in the standard, per-component manner; the
915 * operation is performed on each component (except the ones that are
916 * masked out) with the input being taken from the input swizzle for
919 * The size of some of the inputs may be given (i.e. non-zero) even
920 * though output_size is zero; in that case, the inputs with a zero
921 * size act per-component, while the inputs with non-zero size don't.
923 unsigned output_size
;
926 * The type of vector that the instruction outputs. Note that the
927 * staurate modifier is only allowed on outputs with the float type.
930 nir_alu_type output_type
;
933 * The number of components in each input
935 unsigned input_sizes
[NIR_MAX_VEC_COMPONENTS
];
938 * The type of vector that each input takes. Note that negate and
939 * absolute value are only allowed on inputs with int or float type and
940 * behave differently on the two.
942 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
944 nir_op_algebraic_property algebraic_properties
;
946 /* Whether this represents a numeric conversion opcode */
950 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
952 typedef struct nir_alu_instr
{
956 /** Indicates that this ALU instruction generates an exact value
958 * This is kind of a mixture of GLSL "precise" and "invariant" and not
959 * really equivalent to either. This indicates that the value generated by
960 * this operation is high-precision and any code transformations that touch
961 * it must ensure that the resulting value is bit-for-bit identical to the
970 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
971 nir_alu_instr
*instr
);
972 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
973 nir_alu_instr
*instr
);
975 /* is this source channel used? */
977 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
980 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
981 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
983 return (instr
->dest
.write_mask
>> channel
) & 1;
986 static inline nir_component_mask_t
987 nir_alu_instr_src_read_mask(const nir_alu_instr
*instr
, unsigned src
)
989 nir_component_mask_t read_mask
= 0;
990 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; c
++) {
991 if (!nir_alu_instr_channel_used(instr
, src
, c
))
994 read_mask
|= (1 << instr
->src
[src
].swizzle
[c
]);
1000 * For instructions whose destinations are SSA, get the number of channels
1003 static inline unsigned
1004 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
1006 assert(instr
->dest
.dest
.is_ssa
);
1008 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1009 return nir_op_infos
[instr
->op
].input_sizes
[src
];
1011 return instr
->dest
.dest
.ssa
.num_components
;
1014 bool nir_const_value_negative_equal(const nir_const_value
*c1
,
1015 const nir_const_value
*c2
,
1016 unsigned components
,
1017 nir_alu_type base_type
,
1020 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
1021 unsigned src1
, unsigned src2
);
1023 bool nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
1024 const nir_alu_instr
*alu2
,
1025 unsigned src1
, unsigned src2
);
1029 nir_deref_type_array
,
1030 nir_deref_type_array_wildcard
,
1031 nir_deref_type_ptr_as_array
,
1032 nir_deref_type_struct
,
1033 nir_deref_type_cast
,
1039 /** The type of this deref instruction */
1040 nir_deref_type deref_type
;
1042 /** The mode of the underlying variable */
1043 nir_variable_mode mode
;
1045 /** The dereferenced type of the resulting pointer value */
1046 const struct glsl_type
*type
;
1049 /** Variable being dereferenced if deref_type is a deref_var */
1052 /** Parent deref if deref_type is not deref_var */
1056 /** Additional deref parameters */
1067 unsigned ptr_stride
;
1071 /** Destination to store the resulting "pointer" */
1075 static inline nir_deref_instr
*nir_src_as_deref(nir_src src
);
1077 static inline nir_deref_instr
*
1078 nir_deref_instr_parent(const nir_deref_instr
*instr
)
1080 if (instr
->deref_type
== nir_deref_type_var
)
1083 return nir_src_as_deref(instr
->parent
);
1086 static inline nir_variable
*
1087 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1089 while (instr
->deref_type
!= nir_deref_type_var
) {
1090 if (instr
->deref_type
== nir_deref_type_cast
)
1093 instr
= nir_deref_instr_parent(instr
);
1099 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1100 bool nir_deref_instr_has_complex_use(nir_deref_instr
*instr
);
1102 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1104 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*instr
);
1109 struct nir_function
*callee
;
1111 unsigned num_params
;
1115 #include "nir_intrinsics.h"
1117 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1119 /** Represents an intrinsic
1121 * An intrinsic is an instruction type for handling things that are
1122 * more-or-less regular operations but don't just consume and produce SSA
1123 * values like ALU operations do. Intrinsics are not for things that have
1124 * special semantic meaning such as phi nodes and parallel copies.
1125 * Examples of intrinsics include variable load/store operations, system
1126 * value loads, and the like. Even though texturing more-or-less falls
1127 * under this category, texturing is its own instruction type because
1128 * trying to represent texturing with intrinsics would lead to a
1129 * combinatorial explosion of intrinsic opcodes.
1131 * By having a single instruction type for handling a lot of different
1132 * cases, optimization passes can look for intrinsics and, for the most
1133 * part, completely ignore them. Each intrinsic type also has a few
1134 * possible flags that govern whether or not they can be reordered or
1135 * eliminated. That way passes like dead code elimination can still work
1136 * on intrisics without understanding the meaning of each.
1138 * Each intrinsic has some number of constant indices, some number of
1139 * variables, and some number of sources. What these sources, variables,
1140 * and indices mean depends on the intrinsic and is documented with the
1141 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1142 * instructions are the only types of instruction that can operate on
1148 nir_intrinsic_op intrinsic
;
1152 /** number of components if this is a vectorized intrinsic
1154 * Similarly to ALU operations, some intrinsics are vectorized.
1155 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1156 * For vectorized intrinsics, the num_components field specifies the
1157 * number of destination components and the number of source components
1158 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1160 uint8_t num_components
;
1162 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1165 } nir_intrinsic_instr
;
1167 static inline nir_variable
*
1168 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1170 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1174 * \name NIR intrinsics semantic flags
1176 * information about what the compiler can do with the intrinsics.
1178 * \sa nir_intrinsic_info::flags
1182 * whether the intrinsic can be safely eliminated if none of its output
1183 * value is not being used.
1185 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1188 * Whether the intrinsic can be reordered with respect to any other
1189 * intrinsic, i.e. whether the only reordering dependencies of the
1190 * intrinsic are due to the register reads/writes.
1192 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1193 } nir_intrinsic_semantic_flag
;
1196 * \name NIR intrinsics const-index flag
1198 * Indicates the usage of a const_index slot.
1200 * \sa nir_intrinsic_info::index_map
1204 * Generally instructions that take a offset src argument, can encode
1205 * a constant 'base' value which is added to the offset.
1207 NIR_INTRINSIC_BASE
= 1,
1210 * For store instructions, a writemask for the store.
1212 NIR_INTRINSIC_WRMASK
,
1215 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1217 NIR_INTRINSIC_STREAM_ID
,
1220 * The clip-plane id for load_user_clip_plane intrinsic.
1222 NIR_INTRINSIC_UCP_ID
,
1225 * The amount of data, starting from BASE, that this instruction may
1226 * access. This is used to provide bounds if the offset is not constant.
1228 NIR_INTRINSIC_RANGE
,
1231 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1233 NIR_INTRINSIC_DESC_SET
,
1236 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1238 NIR_INTRINSIC_BINDING
,
1243 NIR_INTRINSIC_COMPONENT
,
1246 * Interpolation mode (only meaningful for FS inputs).
1248 NIR_INTRINSIC_INTERP_MODE
,
1251 * A binary nir_op to use when performing a reduction or scan operation
1253 NIR_INTRINSIC_REDUCTION_OP
,
1256 * Cluster size for reduction operations
1258 NIR_INTRINSIC_CLUSTER_SIZE
,
1261 * Parameter index for a load_param intrinsic
1263 NIR_INTRINSIC_PARAM_IDX
,
1266 * Image dimensionality for image intrinsics
1268 * One of GLSL_SAMPLER_DIM_*
1270 NIR_INTRINSIC_IMAGE_DIM
,
1273 * Non-zero if we are accessing an array image
1275 NIR_INTRINSIC_IMAGE_ARRAY
,
1278 * Image format for image intrinsics
1280 NIR_INTRINSIC_FORMAT
,
1283 * Access qualifiers for image and memory access intrinsics
1285 NIR_INTRINSIC_ACCESS
,
1288 * Alignment for offsets and addresses
1290 * These two parameters, specify an alignment in terms of a multiplier and
1291 * an offset. The offset or address parameter X of the intrinsic is
1292 * guaranteed to satisfy the following:
1294 * (X - align_offset) % align_mul == 0
1296 NIR_INTRINSIC_ALIGN_MUL
,
1297 NIR_INTRINSIC_ALIGN_OFFSET
,
1300 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1302 NIR_INTRINSIC_DESC_TYPE
,
1305 * The nir_alu_type of a uniform/input/output
1310 * The swizzle mask for the instructions
1311 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1313 NIR_INTRINSIC_SWIZZLE_MASK
,
1315 /* Separate source/dest access flags for copies */
1316 NIR_INTRINSIC_SRC_ACCESS
= 21,
1317 NIR_INTRINSIC_DST_ACCESS
= 22,
1319 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1321 } nir_intrinsic_index_flag
;
1323 #define NIR_INTRINSIC_MAX_INPUTS 5
1328 unsigned num_srcs
; /** < number of register/SSA inputs */
1330 /** number of components of each input register
1332 * If this value is 0, the number of components is given by the
1333 * num_components field of nir_intrinsic_instr. If this value is -1, the
1334 * intrinsic consumes however many components are provided and it is not
1337 int src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1341 /** number of components of the output register
1343 * If this value is 0, the number of components is given by the
1344 * num_components field of nir_intrinsic_instr.
1346 unsigned dest_components
;
1348 /** bitfield of legal bit sizes */
1349 unsigned dest_bit_sizes
;
1351 /** the number of constant indices used by the intrinsic */
1352 unsigned num_indices
;
1354 /** indicates the usage of intr->const_index[n] */
1355 unsigned index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1357 /** semantic flags for calls to this intrinsic */
1358 nir_intrinsic_semantic_flag flags
;
1359 } nir_intrinsic_info
;
1361 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1363 static inline unsigned
1364 nir_intrinsic_src_components(nir_intrinsic_instr
*intr
, unsigned srcn
)
1366 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1367 assert(srcn
< info
->num_srcs
);
1368 if (info
->src_components
[srcn
] > 0)
1369 return info
->src_components
[srcn
];
1370 else if (info
->src_components
[srcn
] == 0)
1371 return intr
->num_components
;
1373 return nir_src_num_components(intr
->src
[srcn
]);
1376 static inline unsigned
1377 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1379 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1380 if (!info
->has_dest
)
1382 else if (info
->dest_components
)
1383 return info
->dest_components
;
1385 return intr
->num_components
;
1388 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1389 static inline type \
1390 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1392 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1393 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1394 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1396 static inline void \
1397 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1399 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1400 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1401 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1404 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1405 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1406 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1407 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1408 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1409 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1410 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1411 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1412 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1413 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1414 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1415 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1416 INTRINSIC_IDX_ACCESSORS(image_dim
, IMAGE_DIM
, enum glsl_sampler_dim
)
1417 INTRINSIC_IDX_ACCESSORS(image_array
, IMAGE_ARRAY
, bool)
1418 INTRINSIC_IDX_ACCESSORS(access
, ACCESS
, enum gl_access_qualifier
)
1419 INTRINSIC_IDX_ACCESSORS(src_access
, SRC_ACCESS
, enum gl_access_qualifier
)
1420 INTRINSIC_IDX_ACCESSORS(dst_access
, DST_ACCESS
, enum gl_access_qualifier
)
1421 INTRINSIC_IDX_ACCESSORS(format
, FORMAT
, unsigned)
1422 INTRINSIC_IDX_ACCESSORS(align_mul
, ALIGN_MUL
, unsigned)
1423 INTRINSIC_IDX_ACCESSORS(align_offset
, ALIGN_OFFSET
, unsigned)
1424 INTRINSIC_IDX_ACCESSORS(desc_type
, DESC_TYPE
, unsigned)
1425 INTRINSIC_IDX_ACCESSORS(type
, TYPE
, nir_alu_type
)
1426 INTRINSIC_IDX_ACCESSORS(swizzle_mask
, SWIZZLE_MASK
, unsigned)
1429 nir_intrinsic_set_align(nir_intrinsic_instr
*intrin
,
1430 unsigned align_mul
, unsigned align_offset
)
1432 assert(util_is_power_of_two_nonzero(align_mul
));
1433 assert(align_offset
< align_mul
);
1434 nir_intrinsic_set_align_mul(intrin
, align_mul
);
1435 nir_intrinsic_set_align_offset(intrin
, align_offset
);
1438 /** Returns a simple alignment for a load/store intrinsic offset
1440 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1441 * and ALIGN_OFFSET parameters, this helper takes both into account and
1442 * provides a single simple alignment parameter. The offset X is guaranteed
1443 * to satisfy X % align == 0.
1445 static inline unsigned
1446 nir_intrinsic_align(const nir_intrinsic_instr
*intrin
)
1448 const unsigned align_mul
= nir_intrinsic_align_mul(intrin
);
1449 const unsigned align_offset
= nir_intrinsic_align_offset(intrin
);
1450 assert(align_offset
< align_mul
);
1451 return align_offset
? 1 << (ffs(align_offset
) - 1) : align_mul
;
1454 /* Converts a image_deref_* intrinsic into a image_* one */
1455 void nir_rewrite_image_intrinsic(nir_intrinsic_instr
*instr
,
1456 nir_ssa_def
*handle
, bool bindless
);
1458 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1460 nir_intrinsic_can_reorder(nir_intrinsic_instr
*instr
)
1462 if (instr
->intrinsic
== nir_intrinsic_load_deref
||
1463 instr
->intrinsic
== nir_intrinsic_load_ssbo
||
1464 instr
->intrinsic
== nir_intrinsic_bindless_image_load
||
1465 instr
->intrinsic
== nir_intrinsic_image_deref_load
||
1466 instr
->intrinsic
== nir_intrinsic_image_load
) {
1467 return nir_intrinsic_access(instr
) & ACCESS_CAN_REORDER
;
1469 const nir_intrinsic_info
*info
=
1470 &nir_intrinsic_infos
[instr
->intrinsic
];
1471 return (info
->flags
& NIR_INTRINSIC_CAN_ELIMINATE
) &&
1472 (info
->flags
& NIR_INTRINSIC_CAN_REORDER
);
1477 * \group texture information
1479 * This gives semantic information about textures which is useful to the
1480 * frontend, the backend, and lowering passes, but not the optimizer.
1485 nir_tex_src_projector
,
1486 nir_tex_src_comparator
, /* shadow comparator */
1490 nir_tex_src_min_lod
,
1491 nir_tex_src_ms_index
, /* MSAA sample index */
1492 nir_tex_src_ms_mcs
, /* MSAA compression value */
1495 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
1496 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
1497 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
1498 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
1499 nir_tex_src_texture_handle
, /* < bindless texture handle */
1500 nir_tex_src_sampler_handle
, /* < bindless sampler handle */
1501 nir_tex_src_plane
, /* < selects plane for planar textures */
1502 nir_num_tex_src_types
1507 nir_tex_src_type src_type
;
1511 nir_texop_tex
, /**< Regular texture look-up */
1512 nir_texop_txb
, /**< Texture look-up with LOD bias */
1513 nir_texop_txl
, /**< Texture look-up with explicit LOD */
1514 nir_texop_txd
, /**< Texture look-up with partial derivatives */
1515 nir_texop_txf
, /**< Texel fetch with explicit LOD */
1516 nir_texop_txf_ms
, /**< Multisample texture fetch */
1517 nir_texop_txf_ms_fb
, /**< Multisample texture fetch from framebuffer */
1518 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
1519 nir_texop_txs
, /**< Texture size */
1520 nir_texop_lod
, /**< Texture lod query */
1521 nir_texop_tg4
, /**< Texture gather */
1522 nir_texop_query_levels
, /**< Texture levels query */
1523 nir_texop_texture_samples
, /**< Texture samples query */
1524 nir_texop_samples_identical
, /**< Query whether all samples are definitely
1532 enum glsl_sampler_dim sampler_dim
;
1533 nir_alu_type dest_type
;
1538 unsigned num_srcs
, coord_components
;
1539 bool is_array
, is_shadow
;
1542 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1543 * components or the new-style shadow that outputs 1 component.
1545 bool is_new_style_shadow
;
1547 /* gather component selector */
1548 unsigned component
: 2;
1550 /* gather offsets */
1551 int8_t tg4_offsets
[4][2];
1553 /* True if the texture index or handle is not dynamically uniform */
1554 bool texture_non_uniform
;
1556 /* True if the sampler index or handle is not dynamically uniform */
1557 bool sampler_non_uniform
;
1559 /** The texture index
1561 * If this texture instruction has a nir_tex_src_texture_offset source,
1562 * then the texture index is given by texture_index + texture_offset.
1564 unsigned texture_index
;
1566 /** The size of the texture array or 0 if it's not an array */
1567 unsigned texture_array_size
;
1569 /** The sampler index
1571 * The following operations do not require a sampler and, as such, this
1572 * field should be ignored:
1574 * - nir_texop_txf_ms
1577 * - nir_texop_query_levels
1578 * - nir_texop_texture_samples
1579 * - nir_texop_samples_identical
1581 * If this texture instruction has a nir_tex_src_sampler_offset source,
1582 * then the sampler index is given by sampler_index + sampler_offset.
1584 unsigned sampler_index
;
1587 static inline unsigned
1588 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
1590 switch (instr
->op
) {
1591 case nir_texop_txs
: {
1593 switch (instr
->sampler_dim
) {
1594 case GLSL_SAMPLER_DIM_1D
:
1595 case GLSL_SAMPLER_DIM_BUF
:
1598 case GLSL_SAMPLER_DIM_2D
:
1599 case GLSL_SAMPLER_DIM_CUBE
:
1600 case GLSL_SAMPLER_DIM_MS
:
1601 case GLSL_SAMPLER_DIM_RECT
:
1602 case GLSL_SAMPLER_DIM_EXTERNAL
:
1603 case GLSL_SAMPLER_DIM_SUBPASS
:
1606 case GLSL_SAMPLER_DIM_3D
:
1610 unreachable("not reached");
1612 if (instr
->is_array
)
1620 case nir_texop_texture_samples
:
1621 case nir_texop_query_levels
:
1622 case nir_texop_samples_identical
:
1626 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
1633 /* Returns true if this texture operation queries something about the texture
1634 * rather than actually sampling it.
1637 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
1639 switch (instr
->op
) {
1642 case nir_texop_texture_samples
:
1643 case nir_texop_query_levels
:
1644 case nir_texop_txf_ms_mcs
:
1651 case nir_texop_txf_ms
:
1652 case nir_texop_txf_ms_fb
:
1656 unreachable("Invalid texture opcode");
1661 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1663 switch (instr
->op
) {
1683 static inline nir_alu_type
1684 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
1686 switch (instr
->src
[src
].src_type
) {
1687 case nir_tex_src_coord
:
1688 switch (instr
->op
) {
1690 case nir_texop_txf_ms
:
1691 case nir_texop_txf_ms_fb
:
1692 case nir_texop_txf_ms_mcs
:
1693 case nir_texop_samples_identical
:
1694 return nir_type_int
;
1697 return nir_type_float
;
1700 case nir_tex_src_lod
:
1701 switch (instr
->op
) {
1704 return nir_type_int
;
1707 return nir_type_float
;
1710 case nir_tex_src_projector
:
1711 case nir_tex_src_comparator
:
1712 case nir_tex_src_bias
:
1713 case nir_tex_src_ddx
:
1714 case nir_tex_src_ddy
:
1715 return nir_type_float
;
1717 case nir_tex_src_offset
:
1718 case nir_tex_src_ms_index
:
1719 case nir_tex_src_texture_offset
:
1720 case nir_tex_src_sampler_offset
:
1721 return nir_type_int
;
1724 unreachable("Invalid texture source type");
1728 static inline unsigned
1729 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
1731 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
1732 return instr
->coord_components
;
1734 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
1735 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
1738 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
1739 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
1740 if (instr
->is_array
)
1741 return instr
->coord_components
- 1;
1743 return instr
->coord_components
;
1746 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
1747 * the offset, since a cube maps to a single face.
1749 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
1750 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
1752 else if (instr
->is_array
)
1753 return instr
->coord_components
- 1;
1755 return instr
->coord_components
;
1762 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
1764 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
1765 if (instr
->src
[i
].src_type
== type
)
1771 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
1772 nir_tex_src_type src_type
,
1775 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
1777 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
);
1784 nir_const_value value
[];
1785 } nir_load_const_instr
;
1787 #define nir_const_load_to_arr(arr, l, m) \
1789 nir_const_value_to_array(arr, l->value, l->def.num_components, m); \
1803 /* creates a new SSA variable in an undefined state */
1808 } nir_ssa_undef_instr
;
1811 struct exec_node node
;
1813 /* The predecessor block corresponding to this source */
1814 struct nir_block
*pred
;
1819 #define nir_foreach_phi_src(phi_src, phi) \
1820 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
1821 #define nir_foreach_phi_src_safe(phi_src, phi) \
1822 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
1827 struct exec_list srcs
; /** < list of nir_phi_src */
1833 struct exec_node node
;
1836 } nir_parallel_copy_entry
;
1838 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
1839 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
1844 /* A list of nir_parallel_copy_entrys. The sources of all of the
1845 * entries are copied to the corresponding destinations "in parallel".
1846 * In other words, if we have two entries: a -> b and b -> a, the values
1849 struct exec_list entries
;
1850 } nir_parallel_copy_instr
;
1852 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
1853 type
, nir_instr_type_alu
)
1854 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
1855 type
, nir_instr_type_deref
)
1856 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
1857 type
, nir_instr_type_call
)
1858 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
1859 type
, nir_instr_type_jump
)
1860 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
1861 type
, nir_instr_type_tex
)
1862 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
1863 type
, nir_instr_type_intrinsic
)
1864 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
1865 type
, nir_instr_type_load_const
)
1866 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
1867 type
, nir_instr_type_ssa_undef
)
1868 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
1869 type
, nir_instr_type_phi
)
1870 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
1871 nir_parallel_copy_instr
, instr
,
1872 type
, nir_instr_type_parallel_copy
)
1877 * Control flow consists of a tree of control flow nodes, which include
1878 * if-statements and loops. The leaves of the tree are basic blocks, lists of
1879 * instructions that always run start-to-finish. Each basic block also keeps
1880 * track of its successors (blocks which may run immediately after the current
1881 * block) and predecessors (blocks which could have run immediately before the
1882 * current block). Each function also has a start block and an end block which
1883 * all return statements point to (which is always empty). Together, all the
1884 * blocks with their predecessors and successors make up the control flow
1885 * graph (CFG) of the function. There are helpers that modify the tree of
1886 * control flow nodes while modifying the CFG appropriately; these should be
1887 * used instead of modifying the tree directly.
1894 nir_cf_node_function
1897 typedef struct nir_cf_node
{
1898 struct exec_node node
;
1899 nir_cf_node_type type
;
1900 struct nir_cf_node
*parent
;
1903 typedef struct nir_block
{
1904 nir_cf_node cf_node
;
1906 struct exec_list instr_list
; /** < list of nir_instr */
1908 /** generic block index; generated by nir_index_blocks */
1912 * Each block can only have up to 2 successors, so we put them in a simple
1913 * array - no need for anything more complicated.
1915 struct nir_block
*successors
[2];
1917 /* Set of nir_block predecessors in the CFG */
1918 struct set
*predecessors
;
1921 * this node's immediate dominator in the dominance tree - set to NULL for
1924 struct nir_block
*imm_dom
;
1926 /* This node's children in the dominance tree */
1927 unsigned num_dom_children
;
1928 struct nir_block
**dom_children
;
1930 /* Set of nir_blocks on the dominance frontier of this block */
1931 struct set
*dom_frontier
;
1934 * These two indices have the property that dom_{pre,post}_index for each
1935 * child of this block in the dominance tree will always be between
1936 * dom_pre_index and dom_post_index for this block, which makes testing if
1937 * a given block is dominated by another block an O(1) operation.
1939 unsigned dom_pre_index
, dom_post_index
;
1941 /* live in and out for this block; used for liveness analysis */
1942 BITSET_WORD
*live_in
;
1943 BITSET_WORD
*live_out
;
1946 static inline nir_instr
*
1947 nir_block_first_instr(nir_block
*block
)
1949 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
1950 return exec_node_data(nir_instr
, head
, node
);
1953 static inline nir_instr
*
1954 nir_block_last_instr(nir_block
*block
)
1956 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
1957 return exec_node_data(nir_instr
, tail
, node
);
1961 nir_block_ends_in_jump(nir_block
*block
)
1963 return !exec_list_is_empty(&block
->instr_list
) &&
1964 nir_block_last_instr(block
)->type
== nir_instr_type_jump
;
1967 #define nir_foreach_instr(instr, block) \
1968 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
1969 #define nir_foreach_instr_reverse(instr, block) \
1970 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
1971 #define nir_foreach_instr_safe(instr, block) \
1972 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
1973 #define nir_foreach_instr_reverse_safe(instr, block) \
1974 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
1977 nir_selection_control_none
= 0x0,
1978 nir_selection_control_flatten
= 0x1,
1979 nir_selection_control_dont_flatten
= 0x2,
1980 } nir_selection_control
;
1982 typedef struct nir_if
{
1983 nir_cf_node cf_node
;
1985 nir_selection_control control
;
1987 struct exec_list then_list
; /** < list of nir_cf_node */
1988 struct exec_list else_list
; /** < list of nir_cf_node */
1994 /** Instruction that generates nif::condition. */
1995 nir_instr
*conditional_instr
;
1997 /** Block within ::nif that has the break instruction. */
1998 nir_block
*break_block
;
2000 /** Last block for the then- or else-path that does not contain the break. */
2001 nir_block
*continue_from_block
;
2003 /** True when ::break_block is in the else-path of ::nif. */
2004 bool continue_from_then
;
2007 /* This is true if the terminators exact trip count is unknown. For
2010 * for (int i = 0; i < imin(x, 4); i++)
2013 * Here loop analysis would have set a max_trip_count of 4 however we dont
2014 * know for sure that this is the exact trip count.
2016 bool exact_trip_count_unknown
;
2018 struct list_head loop_terminator_link
;
2019 } nir_loop_terminator
;
2022 /* Estimated cost (in number of instructions) of the loop */
2023 unsigned instr_cost
;
2025 /* Guessed trip count based on array indexing */
2026 unsigned guessed_trip_count
;
2028 /* Maximum number of times the loop is run (if known) */
2029 unsigned max_trip_count
;
2031 /* Do we know the exact number of times the loop will be run */
2032 bool exact_trip_count_known
;
2034 /* Unroll the loop regardless of its size */
2037 /* Does the loop contain complex loop terminators, continues or other
2038 * complex behaviours? If this is true we can't rely on
2039 * loop_terminator_list to be complete or accurate.
2043 nir_loop_terminator
*limiting_terminator
;
2045 /* A list of loop_terminators terminating this loop. */
2046 struct list_head loop_terminator_list
;
2050 nir_loop_control_none
= 0x0,
2051 nir_loop_control_unroll
= 0x1,
2052 nir_loop_control_dont_unroll
= 0x2,
2056 nir_cf_node cf_node
;
2058 struct exec_list body
; /** < list of nir_cf_node */
2060 nir_loop_info
*info
;
2061 nir_loop_control control
;
2062 bool partially_unrolled
;
2066 * Various bits of metadata that can may be created or required by
2067 * optimization and analysis passes
2070 nir_metadata_none
= 0x0,
2071 nir_metadata_block_index
= 0x1,
2072 nir_metadata_dominance
= 0x2,
2073 nir_metadata_live_ssa_defs
= 0x4,
2074 nir_metadata_not_properly_reset
= 0x8,
2075 nir_metadata_loop_analysis
= 0x10,
2079 nir_cf_node cf_node
;
2081 /** pointer to the function of which this is an implementation */
2082 struct nir_function
*function
;
2084 struct exec_list body
; /** < list of nir_cf_node */
2086 nir_block
*end_block
;
2088 /** list for all local variables in the function */
2089 struct exec_list locals
;
2091 /** list of local registers in the function */
2092 struct exec_list registers
;
2094 /** next available local register index */
2097 /** next available SSA value index */
2100 /* total number of basic blocks, only valid when block_index_dirty = false */
2101 unsigned num_blocks
;
2103 nir_metadata valid_metadata
;
2104 } nir_function_impl
;
2106 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2107 nir_start_block(nir_function_impl
*impl
)
2109 return (nir_block
*) impl
->body
.head_sentinel
.next
;
2112 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2113 nir_impl_last_block(nir_function_impl
*impl
)
2115 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
2118 static inline nir_cf_node
*
2119 nir_cf_node_next(nir_cf_node
*node
)
2121 struct exec_node
*next
= exec_node_get_next(&node
->node
);
2122 if (exec_node_is_tail_sentinel(next
))
2125 return exec_node_data(nir_cf_node
, next
, node
);
2128 static inline nir_cf_node
*
2129 nir_cf_node_prev(nir_cf_node
*node
)
2131 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
2132 if (exec_node_is_head_sentinel(prev
))
2135 return exec_node_data(nir_cf_node
, prev
, node
);
2139 nir_cf_node_is_first(const nir_cf_node
*node
)
2141 return exec_node_is_head_sentinel(node
->node
.prev
);
2145 nir_cf_node_is_last(const nir_cf_node
*node
)
2147 return exec_node_is_tail_sentinel(node
->node
.next
);
2150 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
2151 type
, nir_cf_node_block
)
2152 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
2153 type
, nir_cf_node_if
)
2154 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
2155 type
, nir_cf_node_loop
)
2156 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
2157 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
2159 static inline nir_block
*
2160 nir_if_first_then_block(nir_if
*if_stmt
)
2162 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
2163 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2166 static inline nir_block
*
2167 nir_if_last_then_block(nir_if
*if_stmt
)
2169 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
2170 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2173 static inline nir_block
*
2174 nir_if_first_else_block(nir_if
*if_stmt
)
2176 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
2177 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2180 static inline nir_block
*
2181 nir_if_last_else_block(nir_if
*if_stmt
)
2183 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
2184 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2187 static inline nir_block
*
2188 nir_loop_first_block(nir_loop
*loop
)
2190 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
2191 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2194 static inline nir_block
*
2195 nir_loop_last_block(nir_loop
*loop
)
2197 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
2198 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2202 * Return true if this list of cf_nodes contains a single empty block.
2205 nir_cf_list_is_empty_block(struct exec_list
*cf_list
)
2207 if (exec_list_is_singular(cf_list
)) {
2208 struct exec_node
*head
= exec_list_get_head(cf_list
);
2210 nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2211 return exec_list_is_empty(&block
->instr_list
);
2217 uint8_t num_components
;
2221 typedef struct nir_function
{
2222 struct exec_node node
;
2225 struct nir_shader
*shader
;
2227 unsigned num_params
;
2228 nir_parameter
*params
;
2230 /** The implementation of this function.
2232 * If the function is only declared and not implemented, this is NULL.
2234 nir_function_impl
*impl
;
2240 nir_lower_imul64
= (1 << 0),
2241 nir_lower_isign64
= (1 << 1),
2242 /** Lower all int64 modulus and division opcodes */
2243 nir_lower_divmod64
= (1 << 2),
2244 /** Lower all 64-bit umul_high and imul_high opcodes */
2245 nir_lower_imul_high64
= (1 << 3),
2246 nir_lower_mov64
= (1 << 4),
2247 nir_lower_icmp64
= (1 << 5),
2248 nir_lower_iadd64
= (1 << 6),
2249 nir_lower_iabs64
= (1 << 7),
2250 nir_lower_ineg64
= (1 << 8),
2251 nir_lower_logic64
= (1 << 9),
2252 nir_lower_minmax64
= (1 << 10),
2253 nir_lower_shift64
= (1 << 11),
2254 nir_lower_imul_2x32_64
= (1 << 12),
2255 } nir_lower_int64_options
;
2258 nir_lower_drcp
= (1 << 0),
2259 nir_lower_dsqrt
= (1 << 1),
2260 nir_lower_drsq
= (1 << 2),
2261 nir_lower_dtrunc
= (1 << 3),
2262 nir_lower_dfloor
= (1 << 4),
2263 nir_lower_dceil
= (1 << 5),
2264 nir_lower_dfract
= (1 << 6),
2265 nir_lower_dround_even
= (1 << 7),
2266 nir_lower_dmod
= (1 << 8),
2267 nir_lower_fp64_full_software
= (1 << 9),
2268 } nir_lower_doubles_options
;
2270 typedef struct nir_shader_compiler_options
{
2276 /** Lowers flrp when it does not support doubles */
2282 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2283 bool lower_bitfield_extract
;
2284 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
2285 bool lower_bitfield_extract_to_shifts
;
2286 /** Lowers bitfield_insert to bfi/bfm */
2287 bool lower_bitfield_insert
;
2288 /** Lowers bitfield_insert to compares, and shifts. */
2289 bool lower_bitfield_insert_to_shifts
;
2290 /** Lowers bitfield_insert to bfm/bitfield_select. */
2291 bool lower_bitfield_insert_to_bitfield_select
;
2292 /** Lowers bitfield_reverse to shifts. */
2293 bool lower_bitfield_reverse
;
2294 /** Lowers bit_count to shifts. */
2295 bool lower_bit_count
;
2296 /** Lowers ifind_msb to compare and ufind_msb */
2297 bool lower_ifind_msb
;
2298 /** Lowers find_lsb to ufind_msb and logic ops */
2299 bool lower_find_lsb
;
2300 bool lower_uadd_carry
;
2301 bool lower_usub_borrow
;
2302 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2303 bool lower_mul_high
;
2304 /** lowers fneg and ineg to fsub and isub. */
2306 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2309 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2312 /** enables rules to lower idiv by power-of-two: */
2315 /** enable rules to avoid bit shifts */
2316 bool lower_bitshift
;
2318 /** enables rules to lower isign to imin+imax */
2321 /** enables rules to lower fsign to fsub and flt */
2324 /* Does the native fdot instruction replicate its result for four
2325 * components? If so, then opt_algebraic_late will turn all fdotN
2326 * instructions into fdot_replicatedN instructions.
2328 bool fdot_replicates
;
2330 /** lowers ffloor to fsub+ffract: */
2333 /** lowers ffract to fsub+ffloor: */
2336 /** lowers fceil to fneg+ffloor+fneg: */
2343 bool lower_pack_half_2x16
;
2344 bool lower_pack_unorm_2x16
;
2345 bool lower_pack_snorm_2x16
;
2346 bool lower_pack_unorm_4x8
;
2347 bool lower_pack_snorm_4x8
;
2348 bool lower_unpack_half_2x16
;
2349 bool lower_unpack_unorm_2x16
;
2350 bool lower_unpack_snorm_2x16
;
2351 bool lower_unpack_unorm_4x8
;
2352 bool lower_unpack_snorm_4x8
;
2354 bool lower_extract_byte
;
2355 bool lower_extract_word
;
2357 bool lower_all_io_to_temps
;
2358 bool lower_all_io_to_elements
;
2360 /* Indicates that the driver only has zero-based vertex id */
2361 bool vertex_id_zero_based
;
2364 * If enabled, gl_BaseVertex will be lowered as:
2365 * is_indexed_draw (~0/0) & firstvertex
2367 bool lower_base_vertex
;
2370 * If enabled, gl_HelperInvocation will be lowered as:
2372 * !((1 << sample_id) & sample_mask_in))
2374 * This depends on some possibly hw implementation details, which may
2375 * not be true for all hw. In particular that the FS is only executed
2376 * for covered samples or for helper invocations. So, do not blindly
2377 * enable this option.
2379 * Note: See also issue #22 in ARB_shader_image_load_store
2381 bool lower_helper_invocation
;
2384 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
2386 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
2387 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
2389 bool optimize_sample_mask_in
;
2391 bool lower_cs_local_index_from_id
;
2392 bool lower_cs_local_id_from_index
;
2394 bool lower_device_index_to_zero
;
2396 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
2397 bool lower_wpos_pntc
;
2403 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
2404 * for IO purposes and would prefer loads/stores be vectorized.
2409 * Should nir_lower_io() create load_interpolated_input intrinsics?
2411 * If not, it generates regular load_input intrinsics and interpolation
2412 * information must be inferred from the list of input nir_variables.
2414 bool use_interpolated_input_intrinsics
;
2416 /* Lowers when 32x32->64 bit multiplication is not supported */
2417 bool lower_mul_2x32_64
;
2419 unsigned max_unroll_iterations
;
2421 nir_lower_int64_options lower_int64_options
;
2422 nir_lower_doubles_options lower_doubles_options
;
2423 } nir_shader_compiler_options
;
2425 typedef struct nir_shader
{
2426 /** list of uniforms (nir_variable) */
2427 struct exec_list uniforms
;
2429 /** list of inputs (nir_variable) */
2430 struct exec_list inputs
;
2432 /** list of outputs (nir_variable) */
2433 struct exec_list outputs
;
2435 /** list of shared compute variables (nir_variable) */
2436 struct exec_list shared
;
2438 /** Set of driver-specific options for the shader.
2440 * The memory for the options is expected to be kept in a single static
2441 * copy by the driver.
2443 const struct nir_shader_compiler_options
*options
;
2445 /** Various bits of compile-time information about a given shader */
2446 struct shader_info info
;
2448 /** list of global variables in the shader (nir_variable) */
2449 struct exec_list globals
;
2451 /** list of system value variables in the shader (nir_variable) */
2452 struct exec_list system_values
;
2454 struct exec_list functions
; /** < list of nir_function */
2457 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
2460 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
2462 /** Size in bytes of required scratch space */
2463 unsigned scratch_size
;
2465 /** Constant data associated with this shader.
2467 * Constant data is loaded through load_constant intrinsics. See also
2468 * nir_opt_large_constants.
2470 void *constant_data
;
2471 unsigned constant_data_size
;
2474 #define nir_foreach_function(func, shader) \
2475 foreach_list_typed(nir_function, func, node, &(shader)->functions)
2477 static inline nir_function_impl
*
2478 nir_shader_get_entrypoint(nir_shader
*shader
)
2480 nir_function
*func
= NULL
;
2482 nir_foreach_function(function
, shader
) {
2483 assert(func
== NULL
);
2484 if (function
->is_entrypoint
) {
2495 assert(func
->num_params
== 0);
2500 nir_shader
*nir_shader_create(void *mem_ctx
,
2501 gl_shader_stage stage
,
2502 const nir_shader_compiler_options
*options
,
2505 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
2507 void nir_reg_remove(nir_register
*reg
);
2509 /** Adds a variable to the appropriate list in nir_shader */
2510 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
2513 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
2515 assert(var
->data
.mode
== nir_var_function_temp
);
2516 exec_list_push_tail(&impl
->locals
, &var
->node
);
2519 /** creates a variable, sets a few defaults, and adds it to the list */
2520 nir_variable
*nir_variable_create(nir_shader
*shader
,
2521 nir_variable_mode mode
,
2522 const struct glsl_type
*type
,
2524 /** creates a local variable and adds it to the list */
2525 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
2526 const struct glsl_type
*type
,
2529 /** creates a function and adds it to the shader's list of functions */
2530 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
2532 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
2533 /** creates a function_impl that isn't tied to any particular function */
2534 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
2536 nir_block
*nir_block_create(nir_shader
*shader
);
2537 nir_if
*nir_if_create(nir_shader
*shader
);
2538 nir_loop
*nir_loop_create(nir_shader
*shader
);
2540 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
2542 /** requests that the given pieces of metadata be generated */
2543 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
2544 /** dirties all but the preserved metadata */
2545 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
2547 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
2548 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
2550 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
2551 nir_deref_type deref_type
);
2553 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
2555 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
2556 unsigned num_components
,
2559 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
2560 nir_intrinsic_op op
);
2562 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
2563 nir_function
*callee
);
2565 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
2567 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
2569 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
2571 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
2572 unsigned num_components
,
2575 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
2578 * NIR Cursors and Instruction Insertion API
2581 * A tiny struct representing a point to insert/extract instructions or
2582 * control flow nodes. Helps reduce the combinatorial explosion of possible
2583 * points to insert/extract.
2585 * \sa nir_control_flow.h
2588 nir_cursor_before_block
,
2589 nir_cursor_after_block
,
2590 nir_cursor_before_instr
,
2591 nir_cursor_after_instr
,
2592 } nir_cursor_option
;
2595 nir_cursor_option option
;
2602 static inline nir_block
*
2603 nir_cursor_current_block(nir_cursor cursor
)
2605 if (cursor
.option
== nir_cursor_before_instr
||
2606 cursor
.option
== nir_cursor_after_instr
) {
2607 return cursor
.instr
->block
;
2609 return cursor
.block
;
2613 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
2615 static inline nir_cursor
2616 nir_before_block(nir_block
*block
)
2619 cursor
.option
= nir_cursor_before_block
;
2620 cursor
.block
= block
;
2624 static inline nir_cursor
2625 nir_after_block(nir_block
*block
)
2628 cursor
.option
= nir_cursor_after_block
;
2629 cursor
.block
= block
;
2633 static inline nir_cursor
2634 nir_before_instr(nir_instr
*instr
)
2637 cursor
.option
= nir_cursor_before_instr
;
2638 cursor
.instr
= instr
;
2642 static inline nir_cursor
2643 nir_after_instr(nir_instr
*instr
)
2646 cursor
.option
= nir_cursor_after_instr
;
2647 cursor
.instr
= instr
;
2651 static inline nir_cursor
2652 nir_after_block_before_jump(nir_block
*block
)
2654 nir_instr
*last_instr
= nir_block_last_instr(block
);
2655 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
2656 return nir_before_instr(last_instr
);
2658 return nir_after_block(block
);
2662 static inline nir_cursor
2663 nir_before_src(nir_src
*src
, bool is_if_condition
)
2665 if (is_if_condition
) {
2666 nir_block
*prev_block
=
2667 nir_cf_node_as_block(nir_cf_node_prev(&src
->parent_if
->cf_node
));
2668 assert(!nir_block_ends_in_jump(prev_block
));
2669 return nir_after_block(prev_block
);
2670 } else if (src
->parent_instr
->type
== nir_instr_type_phi
) {
2672 nir_phi_instr
*cond_phi
= nir_instr_as_phi(src
->parent_instr
);
2674 nir_foreach_phi_src(phi_src
, cond_phi
) {
2675 if (phi_src
->src
.ssa
== src
->ssa
) {
2682 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
2683 * to have a more specific name.
2685 nir_phi_src
*phi_src
= LIST_ENTRY(nir_phi_src
, src
, src
);
2686 return nir_after_block_before_jump(phi_src
->pred
);
2688 return nir_before_instr(src
->parent_instr
);
2692 static inline nir_cursor
2693 nir_before_cf_node(nir_cf_node
*node
)
2695 if (node
->type
== nir_cf_node_block
)
2696 return nir_before_block(nir_cf_node_as_block(node
));
2698 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
2701 static inline nir_cursor
2702 nir_after_cf_node(nir_cf_node
*node
)
2704 if (node
->type
== nir_cf_node_block
)
2705 return nir_after_block(nir_cf_node_as_block(node
));
2707 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
2710 static inline nir_cursor
2711 nir_after_phis(nir_block
*block
)
2713 nir_foreach_instr(instr
, block
) {
2714 if (instr
->type
!= nir_instr_type_phi
)
2715 return nir_before_instr(instr
);
2717 return nir_after_block(block
);
2720 static inline nir_cursor
2721 nir_after_cf_node_and_phis(nir_cf_node
*node
)
2723 if (node
->type
== nir_cf_node_block
)
2724 return nir_after_block(nir_cf_node_as_block(node
));
2726 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
2728 return nir_after_phis(block
);
2731 static inline nir_cursor
2732 nir_before_cf_list(struct exec_list
*cf_list
)
2734 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
2735 exec_list_get_head(cf_list
), node
);
2736 return nir_before_cf_node(first_node
);
2739 static inline nir_cursor
2740 nir_after_cf_list(struct exec_list
*cf_list
)
2742 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
2743 exec_list_get_tail(cf_list
), node
);
2744 return nir_after_cf_node(last_node
);
2748 * Insert a NIR instruction at the given cursor.
2750 * Note: This does not update the cursor.
2752 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
2755 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
2757 nir_instr_insert(nir_before_instr(instr
), before
);
2761 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
2763 nir_instr_insert(nir_after_instr(instr
), after
);
2767 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
2769 nir_instr_insert(nir_before_block(block
), before
);
2773 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
2775 nir_instr_insert(nir_after_block(block
), after
);
2779 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
2781 nir_instr_insert(nir_before_cf_node(node
), before
);
2785 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
2787 nir_instr_insert(nir_after_cf_node(node
), after
);
2791 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
2793 nir_instr_insert(nir_before_cf_list(list
), before
);
2797 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
2799 nir_instr_insert(nir_after_cf_list(list
), after
);
2802 void nir_instr_remove_v(nir_instr
*instr
);
2804 static inline nir_cursor
2805 nir_instr_remove(nir_instr
*instr
)
2808 nir_instr
*prev
= nir_instr_prev(instr
);
2810 cursor
= nir_after_instr(prev
);
2812 cursor
= nir_before_block(instr
->block
);
2814 nir_instr_remove_v(instr
);
2820 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
2821 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
2822 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
2823 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
2825 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
2826 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
2828 nir_const_value
*nir_src_as_const_value(nir_src src
);
2830 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
2831 static inline c_type * \
2832 nir_src_as_ ## name (nir_src src) \
2834 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
2835 ? cast_macro(src.ssa->parent_instr) : NULL; \
2838 NIR_SRC_AS_(alu_instr
, nir_alu_instr
, nir_instr_type_alu
, nir_instr_as_alu
)
2839 NIR_SRC_AS_(intrinsic
, nir_intrinsic_instr
,
2840 nir_instr_type_intrinsic
, nir_instr_as_intrinsic
)
2841 NIR_SRC_AS_(deref
, nir_deref_instr
, nir_instr_type_deref
, nir_instr_as_deref
)
2843 bool nir_src_is_dynamically_uniform(nir_src src
);
2844 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
2845 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
2846 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
2847 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
2848 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
2851 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
2852 unsigned num_components
, unsigned bit_size
,
2854 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
2855 unsigned num_components
, unsigned bit_size
,
2858 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
2859 const struct glsl_type
*type
,
2862 assert(glsl_type_is_vector_or_scalar(type
));
2863 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
2864 glsl_get_bit_size(type
), name
);
2866 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
2867 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
2868 nir_instr
*after_me
);
2870 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
2873 * finds the next basic block in source-code order, returns NULL if there is
2877 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
2879 /* Performs the opposite of nir_block_cf_tree_next() */
2881 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
2883 /* Gets the first block in a CF node in source-code order */
2885 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
2887 /* Gets the last block in a CF node in source-code order */
2889 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
2891 /* Gets the next block after a CF node in source-code order */
2893 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
2895 /* Macros for loops that visit blocks in source-code order */
2897 #define nir_foreach_block(block, impl) \
2898 for (nir_block *block = nir_start_block(impl); block != NULL; \
2899 block = nir_block_cf_tree_next(block))
2901 #define nir_foreach_block_safe(block, impl) \
2902 for (nir_block *block = nir_start_block(impl), \
2903 *next = nir_block_cf_tree_next(block); \
2905 block = next, next = nir_block_cf_tree_next(block))
2907 #define nir_foreach_block_reverse(block, impl) \
2908 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
2909 block = nir_block_cf_tree_prev(block))
2911 #define nir_foreach_block_reverse_safe(block, impl) \
2912 for (nir_block *block = nir_impl_last_block(impl), \
2913 *prev = nir_block_cf_tree_prev(block); \
2915 block = prev, prev = nir_block_cf_tree_prev(block))
2917 #define nir_foreach_block_in_cf_node(block, node) \
2918 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
2919 block != nir_cf_node_cf_tree_next(node); \
2920 block = nir_block_cf_tree_next(block))
2922 /* If the following CF node is an if, this function returns that if.
2923 * Otherwise, it returns NULL.
2925 nir_if
*nir_block_get_following_if(nir_block
*block
);
2927 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
2929 void nir_index_local_regs(nir_function_impl
*impl
);
2930 void nir_index_ssa_defs(nir_function_impl
*impl
);
2931 unsigned nir_index_instrs(nir_function_impl
*impl
);
2933 void nir_index_blocks(nir_function_impl
*impl
);
2935 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
2936 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
2937 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
2938 void nir_print_deref(const nir_deref_instr
*deref
, FILE *fp
);
2940 /** Shallow clone of a single ALU instruction. */
2941 nir_alu_instr
*nir_alu_instr_clone(nir_shader
*s
, const nir_alu_instr
*orig
);
2943 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
2944 nir_function_impl
*nir_function_impl_clone(nir_shader
*shader
,
2945 const nir_function_impl
*fi
);
2946 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
2947 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
2949 void nir_shader_replace(nir_shader
*dest
, nir_shader
*src
);
2951 void nir_shader_serialize_deserialize(nir_shader
*s
);
2954 void nir_validate_shader(nir_shader
*shader
, const char *when
);
2955 void nir_metadata_set_validation_flag(nir_shader
*shader
);
2956 void nir_metadata_check_validation_flag(nir_shader
*shader
);
2959 should_skip_nir(const char *name
)
2961 static const char *list
= NULL
;
2963 /* Comma separated list of names to skip. */
2964 list
= getenv("NIR_SKIP");
2972 return comma_separated_list_contains(list
, name
);
2976 should_clone_nir(void)
2978 static int should_clone
= -1;
2979 if (should_clone
< 0)
2980 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
2982 return should_clone
;
2986 should_serialize_deserialize_nir(void)
2988 static int test_serialize
= -1;
2989 if (test_serialize
< 0)
2990 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
2992 return test_serialize
;
2996 should_print_nir(void)
2998 static int should_print
= -1;
2999 if (should_print
< 0)
3000 should_print
= env_var_as_boolean("NIR_PRINT", false);
3002 return should_print
;
3005 static inline void nir_validate_shader(nir_shader
*shader
, const char *when
) { (void) shader
; (void)when
; }
3006 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
3007 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
3008 static inline bool should_skip_nir(UNUSED
const char *pass_name
) { return false; }
3009 static inline bool should_clone_nir(void) { return false; }
3010 static inline bool should_serialize_deserialize_nir(void) { return false; }
3011 static inline bool should_print_nir(void) { return false; }
3014 #define _PASS(pass, nir, do_pass) do { \
3015 if (should_skip_nir(#pass)) { \
3016 printf("skipping %s\n", #pass); \
3020 nir_validate_shader(nir, "after " #pass); \
3021 if (should_clone_nir()) { \
3022 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3023 nir_shader_replace(nir, clone); \
3025 if (should_serialize_deserialize_nir()) { \
3026 nir_shader_serialize_deserialize(nir); \
3030 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3031 nir_metadata_set_validation_flag(nir); \
3032 if (should_print_nir()) \
3033 printf("%s\n", #pass); \
3034 if (pass(nir, ##__VA_ARGS__)) { \
3036 if (should_print_nir()) \
3037 nir_print_shader(nir, stdout); \
3038 nir_metadata_check_validation_flag(nir); \
3042 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3043 if (should_print_nir()) \
3044 printf("%s\n", #pass); \
3045 pass(nir, ##__VA_ARGS__); \
3046 if (should_print_nir()) \
3047 nir_print_shader(nir, stdout); \
3050 #define NIR_SKIP(name) should_skip_nir(#name)
3052 void nir_calc_dominance_impl(nir_function_impl
*impl
);
3053 void nir_calc_dominance(nir_shader
*shader
);
3055 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
3056 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
3058 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
3059 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
3061 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
3062 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
3064 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
3065 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
3067 int nir_gs_count_vertices(const nir_shader
*shader
);
3069 bool nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3070 bool nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3071 bool nir_split_var_copies(nir_shader
*shader
);
3072 bool nir_split_per_member_structs(nir_shader
*shader
);
3073 bool nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
);
3075 bool nir_lower_returns_impl(nir_function_impl
*impl
);
3076 bool nir_lower_returns(nir_shader
*shader
);
3078 void nir_inline_function_impl(struct nir_builder
*b
,
3079 const nir_function_impl
*impl
,
3080 nir_ssa_def
**params
);
3081 bool nir_inline_functions(nir_shader
*shader
);
3083 bool nir_propagate_invariant(nir_shader
*shader
);
3085 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
3086 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
3087 nir_intrinsic_instr
*copy
);
3088 bool nir_lower_var_copies(nir_shader
*shader
);
3090 void nir_fixup_deref_modes(nir_shader
*shader
);
3092 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
3095 nir_lower_direct_array_deref_of_vec_load
= (1 << 0),
3096 nir_lower_indirect_array_deref_of_vec_load
= (1 << 1),
3097 nir_lower_direct_array_deref_of_vec_store
= (1 << 2),
3098 nir_lower_indirect_array_deref_of_vec_store
= (1 << 3),
3099 } nir_lower_array_deref_of_vec_options
;
3101 bool nir_lower_array_deref_of_vec(nir_shader
*shader
, nir_variable_mode modes
,
3102 nir_lower_array_deref_of_vec_options options
);
3104 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
3106 bool nir_lower_locals_to_regs(nir_shader
*shader
);
3108 void nir_lower_io_to_temporaries(nir_shader
*shader
,
3109 nir_function_impl
*entrypoint
,
3110 bool outputs
, bool inputs
);
3112 bool nir_lower_vars_to_scratch(nir_shader
*shader
,
3113 nir_variable_mode modes
,
3115 glsl_type_size_align_func size_align
);
3117 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
3119 void nir_gather_ssa_types(nir_function_impl
*impl
,
3120 BITSET_WORD
*float_types
,
3121 BITSET_WORD
*int_types
);
3123 void nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
3124 int (*type_size
)(const struct glsl_type
*, bool));
3126 /* Some helpers to do very simple linking */
3127 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3128 bool nir_remove_unused_io_vars(nir_shader
*shader
, struct exec_list
*var_list
,
3129 uint64_t *used_by_other_stage
,
3130 uint64_t *used_by_other_stage_patches
);
3131 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
3132 bool default_to_smooth_interp
);
3133 void nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3134 bool nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3137 /* If set, this forces all non-flat fragment shader inputs to be
3138 * interpolated as if with the "sample" qualifier. This requires
3139 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
3141 nir_lower_io_force_sample_interpolation
= (1 << 1),
3142 } nir_lower_io_options
;
3143 bool nir_lower_io(nir_shader
*shader
,
3144 nir_variable_mode modes
,
3145 int (*type_size
)(const struct glsl_type
*, bool),
3146 nir_lower_io_options
);
3150 * An address format which is a simple 32-bit global GPU address.
3152 nir_address_format_32bit_global
,
3155 * An address format which is a simple 64-bit global GPU address.
3157 nir_address_format_64bit_global
,
3160 * An address format which is a bounds-checked 64-bit global GPU address.
3162 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
3163 * address stored with the low bits in .x and high bits in .y, .z is a
3164 * size, and .w is an offset. When the final I/O operation is lowered, .w
3165 * is checked against .z and the operation is predicated on the result.
3167 nir_address_format_64bit_bounded_global
,
3170 * An address format which is comprised of a vec2 where the first
3171 * component is a buffer index and the second is an offset.
3173 nir_address_format_32bit_index_offset
,
3176 * An address format which is a simple 32-bit offset.
3178 nir_address_format_32bit_offset
,
3181 * An address format representing a purely logical addressing model. In
3182 * this model, all deref chains must be complete from the dereference
3183 * operation to the variable. Cast derefs are not allowed. These
3184 * addresses will be 32-bit scalars but the format is immaterial because
3185 * you can always chase the chain.
3187 nir_address_format_logical
,
3188 } nir_address_format
;
3190 static inline unsigned
3191 nir_address_format_bit_size(nir_address_format addr_format
)
3193 switch (addr_format
) {
3194 case nir_address_format_32bit_global
: return 32;
3195 case nir_address_format_64bit_global
: return 64;
3196 case nir_address_format_64bit_bounded_global
: return 32;
3197 case nir_address_format_32bit_index_offset
: return 32;
3198 case nir_address_format_32bit_offset
: return 32;
3199 case nir_address_format_logical
: return 32;
3201 unreachable("Invalid address format");
3204 static inline unsigned
3205 nir_address_format_num_components(nir_address_format addr_format
)
3207 switch (addr_format
) {
3208 case nir_address_format_32bit_global
: return 1;
3209 case nir_address_format_64bit_global
: return 1;
3210 case nir_address_format_64bit_bounded_global
: return 4;
3211 case nir_address_format_32bit_index_offset
: return 2;
3212 case nir_address_format_32bit_offset
: return 1;
3213 case nir_address_format_logical
: return 1;
3215 unreachable("Invalid address format");
3218 static inline const struct glsl_type
*
3219 nir_address_format_to_glsl_type(nir_address_format addr_format
)
3221 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
3222 assert(bit_size
== 32 || bit_size
== 64);
3223 return glsl_vector_type(bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
,
3224 nir_address_format_num_components(addr_format
));
3227 const nir_const_value
*nir_address_format_null_value(nir_address_format addr_format
);
3229 nir_ssa_def
*nir_build_addr_ieq(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3230 nir_address_format addr_format
);
3232 nir_ssa_def
*nir_build_addr_isub(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3233 nir_address_format addr_format
);
3235 nir_ssa_def
* nir_explicit_io_address_from_deref(struct nir_builder
*b
,
3236 nir_deref_instr
*deref
,
3237 nir_ssa_def
*base_addr
,
3238 nir_address_format addr_format
);
3239 void nir_lower_explicit_io_instr(struct nir_builder
*b
,
3240 nir_intrinsic_instr
*io_instr
,
3242 nir_address_format addr_format
);
3244 bool nir_lower_explicit_io(nir_shader
*shader
,
3245 nir_variable_mode modes
,
3246 nir_address_format
);
3248 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
3249 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
3251 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
3253 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
3254 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
3255 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
3257 bool nir_remove_dead_derefs(nir_shader
*shader
);
3258 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
3259 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
);
3260 bool nir_lower_constant_initializers(nir_shader
*shader
,
3261 nir_variable_mode modes
);
3263 bool nir_move_load_const(nir_shader
*shader
);
3264 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
3265 bool nir_lower_vec_to_movs(nir_shader
*shader
);
3266 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
3268 bool nir_lower_alu(nir_shader
*shader
);
3270 bool nir_lower_flrp(nir_shader
*shader
, unsigned lowering_mask
,
3271 bool always_precise
, bool have_ffma
);
3273 bool nir_lower_alu_to_scalar(nir_shader
*shader
, BITSET_WORD
*lower_set
);
3274 bool nir_lower_bool_to_float(nir_shader
*shader
);
3275 bool nir_lower_bool_to_int32(nir_shader
*shader
);
3276 bool nir_lower_int_to_float(nir_shader
*shader
);
3277 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
3278 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
3279 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
3280 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
3281 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
3283 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
3284 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
3285 bool nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode mask
);
3287 void nir_lower_fragcoord_wtrans(nir_shader
*shader
);
3288 void nir_lower_viewport_transform(nir_shader
*shader
);
3289 bool nir_lower_uniforms_to_ubo(nir_shader
*shader
, int multiplier
);
3291 typedef struct nir_lower_subgroups_options
{
3292 uint8_t subgroup_size
;
3293 uint8_t ballot_bit_size
;
3294 bool lower_to_scalar
:1;
3295 bool lower_vote_trivial
:1;
3296 bool lower_vote_eq_to_ballot
:1;
3297 bool lower_subgroup_masks
:1;
3298 bool lower_shuffle
:1;
3299 bool lower_shuffle_to_32bit
:1;
3301 } nir_lower_subgroups_options
;
3303 bool nir_lower_subgroups(nir_shader
*shader
,
3304 const nir_lower_subgroups_options
*options
);
3306 bool nir_lower_system_values(nir_shader
*shader
);
3308 enum PACKED nir_lower_tex_packing
{
3309 nir_lower_tex_packing_none
= 0,
3310 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
3311 * or unsigned ints based on the sampler type
3313 nir_lower_tex_packing_16
,
3314 /* The sampler returns 1 32-bit word of 4x8 unorm */
3315 nir_lower_tex_packing_8
,
3318 typedef struct nir_lower_tex_options
{
3320 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
3321 * sampler types a texture projector is lowered.
3326 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
3328 bool lower_txf_offset
;
3331 * If true, lower away nir_tex_src_offset for all rect textures.
3333 bool lower_rect_offset
;
3336 * If true, lower rect textures to 2D, using txs to fetch the
3337 * texture dimensions and dividing the texture coords by the
3338 * texture dims to normalize.
3343 * If true, convert yuv to rgb.
3345 unsigned lower_y_uv_external
;
3346 unsigned lower_y_u_v_external
;
3347 unsigned lower_yx_xuxv_external
;
3348 unsigned lower_xy_uxvx_external
;
3349 unsigned lower_ayuv_external
;
3350 unsigned lower_xyuv_external
;
3353 * To emulate certain texture wrap modes, this can be used
3354 * to saturate the specified tex coord to [0.0, 1.0]. The
3355 * bits are according to sampler #, ie. if, for example:
3357 * (conf->saturate_s & (1 << n))
3359 * is true, then the s coord for sampler n is saturated.
3361 * Note that clamping must happen *after* projector lowering
3362 * so any projected texture sample instruction with a clamped
3363 * coordinate gets automatically lowered, regardless of the
3364 * 'lower_txp' setting.
3366 unsigned saturate_s
;
3367 unsigned saturate_t
;
3368 unsigned saturate_r
;
3370 /* Bitmask of textures that need swizzling.
3372 * If (swizzle_result & (1 << texture_index)), then the swizzle in
3373 * swizzles[texture_index] is applied to the result of the texturing
3376 unsigned swizzle_result
;
3378 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
3379 * while 4 and 5 represent 0 and 1 respectively.
3381 uint8_t swizzles
[32][4];
3383 /* Can be used to scale sampled values in range required by the format. */
3384 float scale_factors
[32];
3387 * Bitmap of textures that need srgb to linear conversion. If
3388 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
3389 * of the texture are lowered to linear.
3391 unsigned lower_srgb
;
3394 * If true, lower nir_texop_tex on shaders that doesn't support implicit
3395 * LODs to nir_texop_txl.
3397 bool lower_tex_without_implicit_lod
;
3400 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
3402 bool lower_txd_cube_map
;
3405 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
3410 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
3411 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
3412 * with lower_txd_cube_map.
3414 bool lower_txd_shadow
;
3417 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
3418 * Implies lower_txd_cube_map and lower_txd_shadow.
3423 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
3424 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
3426 bool lower_txb_shadow_clamp
;
3429 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
3430 * with nir_texop_txl. This includes cube maps.
3432 bool lower_txd_shadow_clamp
;
3435 * If true, lower nir_texop_txd on when it uses both offset and min_lod
3436 * with nir_texop_txl. This includes cube maps.
3438 bool lower_txd_offset_clamp
;
3441 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
3442 * sampler is bindless.
3444 bool lower_txd_clamp_bindless_sampler
;
3447 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
3448 * sampler index is not statically determinable to be less than 16.
3450 bool lower_txd_clamp_if_sampler_index_not_lt_16
;
3453 * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
3454 * 0-lod followed by a nir_ishr.
3459 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
3460 * mixed-up tg4 locations.
3462 bool lower_tg4_broadcom_swizzle
;
3465 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
3467 bool lower_tg4_offsets
;
3469 enum nir_lower_tex_packing lower_tex_packing
[32];
3470 } nir_lower_tex_options
;
3472 bool nir_lower_tex(nir_shader
*shader
,
3473 const nir_lower_tex_options
*options
);
3475 enum nir_lower_non_uniform_access_type
{
3476 nir_lower_non_uniform_ubo_access
= (1 << 0),
3477 nir_lower_non_uniform_ssbo_access
= (1 << 1),
3478 nir_lower_non_uniform_texture_access
= (1 << 2),
3479 nir_lower_non_uniform_image_access
= (1 << 3),
3482 bool nir_lower_non_uniform_access(nir_shader
*shader
,
3483 enum nir_lower_non_uniform_access_type
);
3485 bool nir_lower_idiv(nir_shader
*shader
);
3487 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
, bool use_vars
);
3488 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
);
3489 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
3491 bool nir_lower_frexp(nir_shader
*nir
);
3493 void nir_lower_two_sided_color(nir_shader
*shader
);
3495 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
3497 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
3498 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
3499 const gl_state_index16
*uniform_state_tokens
);
3501 typedef struct nir_lower_wpos_ytransform_options
{
3502 gl_state_index16 state_tokens
[STATE_LENGTH
];
3503 bool fs_coord_origin_upper_left
:1;
3504 bool fs_coord_origin_lower_left
:1;
3505 bool fs_coord_pixel_center_integer
:1;
3506 bool fs_coord_pixel_center_half_integer
:1;
3507 } nir_lower_wpos_ytransform_options
;
3509 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
3510 const nir_lower_wpos_ytransform_options
*options
);
3511 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
3513 bool nir_lower_fb_read(nir_shader
*shader
);
3515 typedef struct nir_lower_drawpixels_options
{
3516 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
3517 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
3518 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
3519 unsigned drawpix_sampler
;
3520 unsigned pixelmap_sampler
;
3522 bool scale_and_bias
:1;
3523 } nir_lower_drawpixels_options
;
3525 void nir_lower_drawpixels(nir_shader
*shader
,
3526 const nir_lower_drawpixels_options
*options
);
3528 typedef struct nir_lower_bitmap_options
{
3531 } nir_lower_bitmap_options
;
3533 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
3535 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
, unsigned ssbo_offset
);
3538 nir_lower_int_source_mods
= 1 << 0,
3539 nir_lower_float_source_mods
= 1 << 1,
3540 nir_lower_triop_abs
= 1 << 2,
3541 nir_lower_all_source_mods
= (1 << 3) - 1
3542 } nir_lower_to_source_mods_flags
;
3545 bool nir_lower_to_source_mods(nir_shader
*shader
, nir_lower_to_source_mods_flags options
);
3547 bool nir_lower_gs_intrinsics(nir_shader
*shader
);
3549 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
3551 bool nir_lower_bit_size(nir_shader
*shader
,
3552 nir_lower_bit_size_callback callback
,
3553 void *callback_data
);
3555 nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode
);
3556 bool nir_lower_int64(nir_shader
*shader
, nir_lower_int64_options options
);
3558 nir_lower_doubles_options
nir_lower_doubles_op_to_options_mask(nir_op opcode
);
3559 bool nir_lower_doubles(nir_shader
*shader
, const nir_shader
*softfp64
,
3560 nir_lower_doubles_options options
);
3561 bool nir_lower_pack(nir_shader
*shader
);
3563 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
3565 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
3567 void nir_loop_analyze_impl(nir_function_impl
*impl
,
3568 nir_variable_mode indirect_mask
);
3570 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
3572 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
3573 bool nir_repair_ssa(nir_shader
*shader
);
3575 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
3577 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
3578 * registers. If false, convert all values (even those not involved in a phi
3579 * node) to registers.
3581 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
3583 bool nir_lower_phis_to_regs_block(nir_block
*block
);
3584 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
3585 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
);
3587 bool nir_opt_comparison_pre(nir_shader
*shader
);
3589 bool nir_opt_algebraic(nir_shader
*shader
);
3590 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
3591 bool nir_opt_algebraic_late(nir_shader
*shader
);
3592 bool nir_opt_constant_folding(nir_shader
*shader
);
3594 bool nir_opt_combine_stores(nir_shader
*shader
, nir_variable_mode modes
);
3596 bool nir_copy_prop(nir_shader
*shader
);
3598 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
3600 bool nir_opt_cse(nir_shader
*shader
);
3602 bool nir_opt_dce(nir_shader
*shader
);
3604 bool nir_opt_dead_cf(nir_shader
*shader
);
3606 bool nir_opt_dead_write_vars(nir_shader
*shader
);
3608 bool nir_opt_deref_impl(nir_function_impl
*impl
);
3609 bool nir_opt_deref(nir_shader
*shader
);
3611 bool nir_opt_find_array_copies(nir_shader
*shader
);
3613 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
3615 bool nir_opt_idiv_const(nir_shader
*shader
, unsigned min_bit_size
);
3617 bool nir_opt_if(nir_shader
*shader
, bool aggressive_last_continue
);
3619 bool nir_opt_intrinsics(nir_shader
*shader
);
3621 bool nir_opt_large_constants(nir_shader
*shader
,
3622 glsl_type_size_align_func size_align
,
3623 unsigned threshold
);
3625 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
3627 bool nir_opt_move_comparisons(nir_shader
*shader
);
3629 bool nir_opt_move_load_ubo(nir_shader
*shader
);
3631 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
,
3632 bool indirect_load_ok
, bool expensive_alu_ok
);
3634 bool nir_opt_rematerialize_compares(nir_shader
*shader
);
3636 bool nir_opt_remove_phis(nir_shader
*shader
);
3638 bool nir_opt_shrink_load(nir_shader
*shader
);
3640 bool nir_opt_trivial_continues(nir_shader
*shader
);
3642 bool nir_opt_undef(nir_shader
*shader
);
3644 bool nir_opt_vectorize(nir_shader
*shader
);
3646 bool nir_opt_conditional_discard(nir_shader
*shader
);
3648 void nir_strip(nir_shader
*shader
);
3650 void nir_sweep(nir_shader
*shader
);
3652 void nir_remap_dual_slot_attributes(nir_shader
*shader
,
3653 uint64_t *dual_slot_inputs
);
3654 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
);
3656 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
3657 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);
3659 bool nir_lower_sincos(nir_shader
*shader
);