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/bitset.h"
38 #include "util/macros.h"
39 #include "compiler/nir_types.h"
40 #include "compiler/shader_enums.h"
41 #include "compiler/shader_info.h"
45 #include "util/debug.h"
48 #include "nir_opcodes.h"
50 #if defined(_WIN32) && !defined(snprintf)
51 #define snprintf _snprintf
59 #define NIR_TRUE (~0u)
60 #define NIR_MAX_VEC_COMPONENTS 4
61 typedef uint8_t nir_component_mask_t
;
63 /** Defines a cast function
65 * This macro defines a cast function from in_type to out_type where
66 * out_type is some structure type that contains a field of type out_type.
68 * Note that you have to be a bit careful as the generated cast function
71 #define NIR_DEFINE_CAST(name, in_type, out_type, field, \
72 type_field, type_value) \
73 static inline out_type * \
74 name(const in_type *parent) \
76 assert(parent && parent->type_field == type_value); \
77 return exec_node_data(out_type, parent, field); \
87 * Description of built-in state associated with a uniform
89 * \sa nir_variable::state_slots
92 gl_state_index16 tokens
[STATE_LENGTH
];
97 nir_var_shader_in
= (1 << 0),
98 nir_var_shader_out
= (1 << 1),
99 nir_var_global
= (1 << 2),
100 nir_var_local
= (1 << 3),
101 nir_var_uniform
= (1 << 4),
102 nir_var_shader_storage
= (1 << 5),
103 nir_var_system_value
= (1 << 6),
104 nir_var_shared
= (1 << 8),
112 nir_rounding_mode_undef
= 0,
113 nir_rounding_mode_rtne
= 1, /* round to nearest even */
114 nir_rounding_mode_ru
= 2, /* round up */
115 nir_rounding_mode_rd
= 3, /* round down */
116 nir_rounding_mode_rtz
= 4, /* round towards zero */
120 float f32
[NIR_MAX_VEC_COMPONENTS
];
121 double f64
[NIR_MAX_VEC_COMPONENTS
];
122 int8_t i8
[NIR_MAX_VEC_COMPONENTS
];
123 uint8_t u8
[NIR_MAX_VEC_COMPONENTS
];
124 int16_t i16
[NIR_MAX_VEC_COMPONENTS
];
125 uint16_t u16
[NIR_MAX_VEC_COMPONENTS
];
126 int32_t i32
[NIR_MAX_VEC_COMPONENTS
];
127 uint32_t u32
[NIR_MAX_VEC_COMPONENTS
];
128 int64_t i64
[NIR_MAX_VEC_COMPONENTS
];
129 uint64_t u64
[NIR_MAX_VEC_COMPONENTS
];
132 typedef struct nir_constant
{
134 * Value of the constant.
136 * The field used to back the values supplied by the constant is determined
137 * by the type associated with the \c nir_variable. Constants may be
138 * scalars, vectors, or matrices.
140 nir_const_value values
[NIR_MAX_VEC_COMPONENTS
];
142 /* we could get this from the var->type but makes clone *much* easier to
143 * not have to care about the type.
145 unsigned num_elements
;
147 /* Array elements / Structure Fields */
148 struct nir_constant
**elements
;
152 * \brief Layout qualifiers for gl_FragDepth.
154 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
155 * with a layout qualifier.
158 nir_depth_layout_none
, /**< No depth layout is specified. */
159 nir_depth_layout_any
,
160 nir_depth_layout_greater
,
161 nir_depth_layout_less
,
162 nir_depth_layout_unchanged
166 * Either a uniform, global variable, shader input, or shader output. Based on
167 * ir_variable - it should be easy to translate between the two.
170 typedef struct nir_variable
{
171 struct exec_node node
;
174 * Declared type of the variable
176 const struct glsl_type
*type
;
179 * Declared name of the variable
183 struct nir_variable_data
{
185 * Storage class of the variable.
187 * \sa nir_variable_mode
189 nir_variable_mode mode
;
192 * Is the variable read-only?
194 * This is set for variables declared as \c const, shader inputs,
197 unsigned read_only
:1;
201 unsigned invariant
:1;
204 * When separate shader programs are enabled, only input/outputs between
205 * the stages of a multi-stage separate program can be safely removed
206 * from the shader interface. Other input/outputs must remains active.
208 * This is also used to make sure xfb varyings that are unused by the
209 * fragment shader are not removed.
211 unsigned always_active_io
:1;
214 * Interpolation mode for shader inputs / outputs
216 * \sa glsl_interp_mode
218 unsigned interpolation
:2;
221 * \name ARB_fragment_coord_conventions
224 unsigned origin_upper_left
:1;
225 unsigned pixel_center_integer
:1;
229 * If non-zero, then this variable may be packed along with other variables
230 * into a single varying slot, so this offset should be applied when
231 * accessing components. For example, an offset of 1 means that the x
232 * component of this variable is actually stored in component y of the
233 * location specified by \c location.
235 unsigned location_frac
:2;
238 * If true, this variable represents an array of scalars that should
239 * be tightly packed. In other words, consecutive array elements
240 * should be stored one component apart, rather than one slot apart.
245 * Whether this is a fragment shader output implicitly initialized with
246 * the previous contents of the specified render target at the
247 * framebuffer location corresponding to this shader invocation.
249 unsigned fb_fetch_output
:1;
252 * Non-zero if this variable is considered bindless as defined by
253 * ARB_bindless_texture.
258 * Was an explicit binding set in the shader?
260 unsigned explicit_binding
:1;
263 * Was a transfer feedback buffer set in the shader?
265 unsigned explicit_xfb_buffer
:1;
268 * Was a transfer feedback stride set in the shader?
270 unsigned explicit_xfb_stride
:1;
273 * Was an explicit offset set in the shader?
275 unsigned explicit_offset
:1;
278 * \brief Layout qualifier for gl_FragDepth.
280 * This is not equal to \c ir_depth_layout_none if and only if this
281 * variable is \c gl_FragDepth and a layout qualifier is specified.
283 nir_depth_layout depth_layout
;
286 * Storage location of the base of this variable
288 * The precise meaning of this field depends on the nature of the variable.
290 * - Vertex shader input: one of the values from \c gl_vert_attrib.
291 * - Vertex shader output: one of the values from \c gl_varying_slot.
292 * - Geometry shader input: one of the values from \c gl_varying_slot.
293 * - Geometry shader output: one of the values from \c gl_varying_slot.
294 * - Fragment shader input: one of the values from \c gl_varying_slot.
295 * - Fragment shader output: one of the values from \c gl_frag_result.
296 * - Uniforms: Per-stage uniform slot number for default uniform block.
297 * - Uniforms: Index within the uniform block definition for UBO members.
298 * - Non-UBO Uniforms: uniform slot number.
299 * - Other: This field is not currently used.
301 * If the variable is a uniform, shader input, or shader output, and the
302 * slot has not been assigned, the value will be -1.
307 * The actual location of the variable in the IR. Only valid for inputs
310 unsigned int driver_location
;
313 * Vertex stream output identifier.
315 * For packed outputs, bit 31 is set and bits [2*i+1,2*i] indicate the
316 * stream of the i-th component.
321 * output index for dual source blending.
326 * Descriptor set binding for sampler or UBO.
331 * Initial binding point for a sampler or UBO.
333 * For array types, this represents the binding point for the first element.
338 * Location an atomic counter or transform feedback is stored at.
343 * Transform feedback buffer.
348 * Transform feedback stride.
353 * ARB_shader_image_load_store qualifiers.
356 bool read_only
; /**< "readonly" qualifier. */
357 bool write_only
; /**< "writeonly" qualifier. */
362 /** Image internal format if specified explicitly, otherwise GL_NONE. */
368 * Built-in state that backs this uniform
370 * Once set at variable creation, \c state_slots must remain invariant.
371 * This is because, ideally, this array would be shared by all clones of
372 * this variable in the IR tree. In other words, we'd really like for it
373 * to be a fly-weight.
375 * If the variable is not a uniform, \c num_state_slots will be zero and
376 * \c state_slots will be \c NULL.
379 unsigned num_state_slots
; /**< Number of state slots used */
380 nir_state_slot
*state_slots
; /**< State descriptors. */
384 * Constant expression assigned in the initializer of the variable
386 * This field should only be used temporarily by creators of NIR shaders
387 * and then lower_constant_initializers can be used to get rid of them.
388 * Most of the rest of NIR ignores this field or asserts that it's NULL.
390 nir_constant
*constant_initializer
;
393 * For variables that are in an interface block or are an instance of an
394 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
396 * \sa ir_variable::location
398 const struct glsl_type
*interface_type
;
401 * Description of per-member data for per-member struct variables
403 * This is used for variables which are actually an amalgamation of
404 * multiple entities such as a struct of built-in values or a struct of
405 * inputs each with their own layout specifier. This is only allowed on
406 * variables with a struct or array of array of struct type.
408 unsigned num_members
;
409 struct nir_variable_data
*members
;
412 #define nir_foreach_variable(var, var_list) \
413 foreach_list_typed(nir_variable, var, node, var_list)
415 #define nir_foreach_variable_safe(var, var_list) \
416 foreach_list_typed_safe(nir_variable, var, node, var_list)
419 nir_variable_is_global(const nir_variable
*var
)
421 return var
->data
.mode
!= nir_var_local
;
424 typedef struct nir_register
{
425 struct exec_node node
;
427 unsigned num_components
; /** < number of vector components */
428 unsigned num_array_elems
; /** < size of array (0 for no array) */
430 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
433 /** generic register index. */
436 /** only for debug purposes, can be NULL */
439 /** whether this register is local (per-function) or global (per-shader) */
443 * If this flag is set to true, then accessing channels >= num_components
444 * is well-defined, and simply spills over to the next array element. This
445 * is useful for backends that can do per-component accessing, in
446 * particular scalar backends. By setting this flag and making
447 * num_components equal to 1, structures can be packed tightly into
448 * registers and then registers can be accessed per-component to get to
449 * each structure member, even if it crosses vec4 boundaries.
453 /** set of nir_srcs where this register is used (read from) */
454 struct list_head uses
;
456 /** set of nir_dests where this register is defined (written to) */
457 struct list_head defs
;
459 /** set of nir_ifs where this register is used as a condition */
460 struct list_head if_uses
;
463 #define nir_foreach_register(reg, reg_list) \
464 foreach_list_typed(nir_register, reg, node, reg_list)
465 #define nir_foreach_register_safe(reg, reg_list) \
466 foreach_list_typed_safe(nir_register, reg, node, reg_list)
470 nir_instr_type_deref
,
473 nir_instr_type_intrinsic
,
474 nir_instr_type_load_const
,
476 nir_instr_type_ssa_undef
,
478 nir_instr_type_parallel_copy
,
481 typedef struct nir_instr
{
482 struct exec_node node
;
484 struct nir_block
*block
;
486 /** generic instruction index. */
489 /* A temporary for optimization and analysis passes to use for storing
490 * flags. For instance, DCE uses this to store the "dead/live" info.
495 static inline nir_instr
*
496 nir_instr_next(nir_instr
*instr
)
498 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
499 if (exec_node_is_tail_sentinel(next
))
502 return exec_node_data(nir_instr
, next
, node
);
505 static inline nir_instr
*
506 nir_instr_prev(nir_instr
*instr
)
508 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
509 if (exec_node_is_head_sentinel(prev
))
512 return exec_node_data(nir_instr
, prev
, node
);
516 nir_instr_is_first(const nir_instr
*instr
)
518 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr
->node
));
522 nir_instr_is_last(const nir_instr
*instr
)
524 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr
->node
));
527 typedef struct nir_ssa_def
{
528 /** for debugging only, can be NULL */
531 /** generic SSA definition index. */
534 /** Index into the live_in and live_out bitfields */
537 /** Instruction which produces this SSA value. */
538 nir_instr
*parent_instr
;
540 /** set of nir_instrs where this register is used (read from) */
541 struct list_head uses
;
543 /** set of nir_ifs where this register is used as a condition */
544 struct list_head if_uses
;
546 uint8_t num_components
;
548 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
556 struct nir_src
*indirect
; /** < NULL for no indirect offset */
557 unsigned base_offset
;
559 /* TODO use-def chain goes here */
563 nir_instr
*parent_instr
;
564 struct list_head def_link
;
567 struct nir_src
*indirect
; /** < NULL for no indirect offset */
568 unsigned base_offset
;
570 /* TODO def-use chain goes here */
575 typedef struct nir_src
{
577 /** Instruction that consumes this value as a source. */
578 nir_instr
*parent_instr
;
579 struct nir_if
*parent_if
;
582 struct list_head use_link
;
592 static inline nir_src
595 nir_src src
= { { NULL
} };
599 #define NIR_SRC_INIT nir_src_init()
601 #define nir_foreach_use(src, reg_or_ssa_def) \
602 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
604 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
605 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
607 #define nir_foreach_if_use(src, reg_or_ssa_def) \
608 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
610 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
611 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
622 static inline nir_dest
625 nir_dest dest
= { { { NULL
} } };
629 #define NIR_DEST_INIT nir_dest_init()
631 #define nir_foreach_def(dest, reg) \
632 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
634 #define nir_foreach_def_safe(dest, reg) \
635 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
637 static inline nir_src
638 nir_src_for_ssa(nir_ssa_def
*def
)
640 nir_src src
= NIR_SRC_INIT
;
648 static inline nir_src
649 nir_src_for_reg(nir_register
*reg
)
651 nir_src src
= NIR_SRC_INIT
;
655 src
.reg
.indirect
= NULL
;
656 src
.reg
.base_offset
= 0;
661 static inline nir_dest
662 nir_dest_for_reg(nir_register
*reg
)
664 nir_dest dest
= NIR_DEST_INIT
;
671 static inline unsigned
672 nir_src_bit_size(nir_src src
)
674 return src
.is_ssa
? src
.ssa
->bit_size
: src
.reg
.reg
->bit_size
;
677 static inline unsigned
678 nir_src_num_components(nir_src src
)
680 return src
.is_ssa
? src
.ssa
->num_components
: src
.reg
.reg
->num_components
;
683 static inline unsigned
684 nir_dest_bit_size(nir_dest dest
)
686 return dest
.is_ssa
? dest
.ssa
.bit_size
: dest
.reg
.reg
->bit_size
;
689 static inline unsigned
690 nir_dest_num_components(nir_dest dest
)
692 return dest
.is_ssa
? dest
.ssa
.num_components
: dest
.reg
.reg
->num_components
;
695 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
696 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
702 * \name input modifiers
706 * For inputs interpreted as floating point, flips the sign bit. For
707 * inputs interpreted as integers, performs the two's complement negation.
712 * Clears the sign bit for floating point values, and computes the integer
713 * absolute value for integers. Note that the negate modifier acts after
714 * the absolute value modifier, therefore if both are set then all inputs
715 * will become negative.
721 * For each input component, says which component of the register it is
722 * chosen from. Note that which elements of the swizzle are used and which
723 * are ignored are based on the write mask for most opcodes - for example,
724 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
725 * a swizzle of {2, x, 1, 0} where x means "don't care."
727 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
734 * \name saturate output modifier
736 * Only valid for opcodes that output floating-point numbers. Clamps the
737 * output to between 0.0 and 1.0 inclusive.
742 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
746 nir_type_invalid
= 0, /* Not a valid type */
751 nir_type_bool32
= 32 | nir_type_bool
,
752 nir_type_int8
= 8 | nir_type_int
,
753 nir_type_int16
= 16 | nir_type_int
,
754 nir_type_int32
= 32 | nir_type_int
,
755 nir_type_int64
= 64 | nir_type_int
,
756 nir_type_uint8
= 8 | nir_type_uint
,
757 nir_type_uint16
= 16 | nir_type_uint
,
758 nir_type_uint32
= 32 | nir_type_uint
,
759 nir_type_uint64
= 64 | nir_type_uint
,
760 nir_type_float16
= 16 | nir_type_float
,
761 nir_type_float32
= 32 | nir_type_float
,
762 nir_type_float64
= 64 | nir_type_float
,
765 #define NIR_ALU_TYPE_SIZE_MASK 0xfffffff8
766 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x00000007
768 static inline unsigned
769 nir_alu_type_get_type_size(nir_alu_type type
)
771 return type
& NIR_ALU_TYPE_SIZE_MASK
;
774 static inline unsigned
775 nir_alu_type_get_base_type(nir_alu_type type
)
777 return type
& NIR_ALU_TYPE_BASE_TYPE_MASK
;
780 static inline nir_alu_type
781 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
785 return nir_type_bool32
;
788 return nir_type_uint32
;
791 return nir_type_int32
;
793 case GLSL_TYPE_UINT16
:
794 return nir_type_uint16
;
796 case GLSL_TYPE_INT16
:
797 return nir_type_int16
;
799 case GLSL_TYPE_UINT8
:
800 return nir_type_uint8
;
802 return nir_type_int8
;
803 case GLSL_TYPE_UINT64
:
804 return nir_type_uint64
;
806 case GLSL_TYPE_INT64
:
807 return nir_type_int64
;
809 case GLSL_TYPE_FLOAT
:
810 return nir_type_float32
;
812 case GLSL_TYPE_FLOAT16
:
813 return nir_type_float16
;
815 case GLSL_TYPE_DOUBLE
:
816 return nir_type_float64
;
819 unreachable("unknown type");
823 static inline nir_alu_type
824 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
826 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
829 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
830 nir_rounding_mode rnd
);
833 NIR_OP_IS_COMMUTATIVE
= (1 << 0),
834 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
835 } nir_op_algebraic_property
;
843 * The number of components in the output
845 * If non-zero, this is the size of the output and input sizes are
846 * explicitly given; swizzle and writemask are still in effect, but if
847 * the output component is masked out, then the input component may
850 * If zero, the opcode acts in the standard, per-component manner; the
851 * operation is performed on each component (except the ones that are
852 * masked out) with the input being taken from the input swizzle for
855 * The size of some of the inputs may be given (i.e. non-zero) even
856 * though output_size is zero; in that case, the inputs with a zero
857 * size act per-component, while the inputs with non-zero size don't.
859 unsigned output_size
;
862 * The type of vector that the instruction outputs. Note that the
863 * staurate modifier is only allowed on outputs with the float type.
866 nir_alu_type output_type
;
869 * The number of components in each input
871 unsigned input_sizes
[NIR_MAX_VEC_COMPONENTS
];
874 * The type of vector that each input takes. Note that negate and
875 * absolute value are only allowed on inputs with int or float type and
876 * behave differently on the two.
878 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
880 nir_op_algebraic_property algebraic_properties
;
883 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
885 typedef struct nir_alu_instr
{
889 /** Indicates that this ALU instruction generates an exact value
891 * This is kind of a mixture of GLSL "precise" and "invariant" and not
892 * really equivalent to either. This indicates that the value generated by
893 * this operation is high-precision and any code transformations that touch
894 * it must ensure that the resulting value is bit-for-bit identical to the
903 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
904 nir_alu_instr
*instr
);
905 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
906 nir_alu_instr
*instr
);
908 /* is this source channel used? */
910 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
913 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
914 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
916 return (instr
->dest
.write_mask
>> channel
) & 1;
920 * For instructions whose destinations are SSA, get the number of channels
923 static inline unsigned
924 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
926 assert(instr
->dest
.dest
.is_ssa
);
928 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
929 return nir_op_infos
[instr
->op
].input_sizes
[src
];
931 return instr
->dest
.dest
.ssa
.num_components
;
934 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
935 unsigned src1
, unsigned src2
);
939 nir_deref_type_array
,
940 nir_deref_type_array_wildcard
,
941 nir_deref_type_struct
,
948 /** The type of this deref instruction */
949 nir_deref_type deref_type
;
951 /** The mode of the underlying variable */
952 nir_variable_mode mode
;
954 /** The dereferenced type of the resulting pointer value */
955 const struct glsl_type
*type
;
958 /** Variable being dereferenced if deref_type is a deref_var */
961 /** Parent deref if deref_type is not deref_var */
965 /** Additional deref parameters */
976 /** Destination to store the resulting "pointer" */
980 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
981 type
, nir_instr_type_deref
)
983 static inline nir_deref_instr
*
984 nir_src_as_deref(nir_src src
)
989 if (src
.ssa
->parent_instr
->type
!= nir_instr_type_deref
)
992 return nir_instr_as_deref(src
.ssa
->parent_instr
);
995 static inline nir_deref_instr
*
996 nir_deref_instr_parent(const nir_deref_instr
*instr
)
998 if (instr
->deref_type
== nir_deref_type_var
)
1001 return nir_src_as_deref(instr
->parent
);
1004 static inline nir_variable
*
1005 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1007 while (instr
->deref_type
!= nir_deref_type_var
) {
1008 if (instr
->deref_type
== nir_deref_type_cast
)
1011 instr
= nir_deref_instr_parent(instr
);
1017 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1019 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1024 struct nir_function
*callee
;
1026 unsigned num_params
;
1030 #include "nir_intrinsics.h"
1032 #define NIR_INTRINSIC_MAX_CONST_INDEX 3
1034 /** Represents an intrinsic
1036 * An intrinsic is an instruction type for handling things that are
1037 * more-or-less regular operations but don't just consume and produce SSA
1038 * values like ALU operations do. Intrinsics are not for things that have
1039 * special semantic meaning such as phi nodes and parallel copies.
1040 * Examples of intrinsics include variable load/store operations, system
1041 * value loads, and the like. Even though texturing more-or-less falls
1042 * under this category, texturing is its own instruction type because
1043 * trying to represent texturing with intrinsics would lead to a
1044 * combinatorial explosion of intrinsic opcodes.
1046 * By having a single instruction type for handling a lot of different
1047 * cases, optimization passes can look for intrinsics and, for the most
1048 * part, completely ignore them. Each intrinsic type also has a few
1049 * possible flags that govern whether or not they can be reordered or
1050 * eliminated. That way passes like dead code elimination can still work
1051 * on intrisics without understanding the meaning of each.
1053 * Each intrinsic has some number of constant indices, some number of
1054 * variables, and some number of sources. What these sources, variables,
1055 * and indices mean depends on the intrinsic and is documented with the
1056 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1057 * instructions are the only types of instruction that can operate on
1063 nir_intrinsic_op intrinsic
;
1067 /** number of components if this is a vectorized intrinsic
1069 * Similarly to ALU operations, some intrinsics are vectorized.
1070 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1071 * For vectorized intrinsics, the num_components field specifies the
1072 * number of destination components and the number of source components
1073 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1075 uint8_t num_components
;
1077 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1080 } nir_intrinsic_instr
;
1082 static inline nir_variable
*
1083 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1085 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1089 * \name NIR intrinsics semantic flags
1091 * information about what the compiler can do with the intrinsics.
1093 * \sa nir_intrinsic_info::flags
1097 * whether the intrinsic can be safely eliminated if none of its output
1098 * value is not being used.
1100 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1103 * Whether the intrinsic can be reordered with respect to any other
1104 * intrinsic, i.e. whether the only reordering dependencies of the
1105 * intrinsic are due to the register reads/writes.
1107 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1108 } nir_intrinsic_semantic_flag
;
1111 * \name NIR intrinsics const-index flag
1113 * Indicates the usage of a const_index slot.
1115 * \sa nir_intrinsic_info::index_map
1119 * Generally instructions that take a offset src argument, can encode
1120 * a constant 'base' value which is added to the offset.
1122 NIR_INTRINSIC_BASE
= 1,
1125 * For store instructions, a writemask for the store.
1127 NIR_INTRINSIC_WRMASK
= 2,
1130 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1132 NIR_INTRINSIC_STREAM_ID
= 3,
1135 * The clip-plane id for load_user_clip_plane intrinsic.
1137 NIR_INTRINSIC_UCP_ID
= 4,
1140 * The amount of data, starting from BASE, that this instruction may
1141 * access. This is used to provide bounds if the offset is not constant.
1143 NIR_INTRINSIC_RANGE
= 5,
1146 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1148 NIR_INTRINSIC_DESC_SET
= 6,
1151 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1153 NIR_INTRINSIC_BINDING
= 7,
1158 NIR_INTRINSIC_COMPONENT
= 8,
1161 * Interpolation mode (only meaningful for FS inputs).
1163 NIR_INTRINSIC_INTERP_MODE
= 9,
1166 * A binary nir_op to use when performing a reduction or scan operation
1168 NIR_INTRINSIC_REDUCTION_OP
= 10,
1171 * Cluster size for reduction operations
1173 NIR_INTRINSIC_CLUSTER_SIZE
= 11,
1176 * Parameter index for a load_param intrinsic
1178 NIR_INTRINSIC_PARAM_IDX
= 12,
1180 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1182 } nir_intrinsic_index_flag
;
1184 #define NIR_INTRINSIC_MAX_INPUTS 5
1189 unsigned num_srcs
; /** < number of register/SSA inputs */
1191 /** number of components of each input register
1193 * If this value is 0, the number of components is given by the
1194 * num_components field of nir_intrinsic_instr.
1196 unsigned src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1200 /** number of components of the output register
1202 * If this value is 0, the number of components is given by the
1203 * num_components field of nir_intrinsic_instr.
1205 unsigned dest_components
;
1207 /** the number of constant indices used by the intrinsic */
1208 unsigned num_indices
;
1210 /** indicates the usage of intr->const_index[n] */
1211 unsigned index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1213 /** semantic flags for calls to this intrinsic */
1214 nir_intrinsic_semantic_flag flags
;
1215 } nir_intrinsic_info
;
1217 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1219 static inline unsigned
1220 nir_intrinsic_src_components(nir_intrinsic_instr
*intr
, unsigned srcn
)
1222 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1223 assert(srcn
< info
->num_srcs
);
1224 if (info
->src_components
[srcn
])
1225 return info
->src_components
[srcn
];
1227 return intr
->num_components
;
1230 static inline unsigned
1231 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1233 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1234 if (!info
->has_dest
)
1236 else if (info
->dest_components
)
1237 return info
->dest_components
;
1239 return intr
->num_components
;
1242 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1243 static inline type \
1244 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1246 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1247 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1248 return instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1250 static inline void \
1251 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1253 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1254 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1255 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1258 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1259 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1260 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1261 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1262 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1263 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1264 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1265 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1266 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1267 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1268 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1269 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1272 * \group texture information
1274 * This gives semantic information about textures which is useful to the
1275 * frontend, the backend, and lowering passes, but not the optimizer.
1280 nir_tex_src_projector
,
1281 nir_tex_src_comparator
, /* shadow comparator */
1285 nir_tex_src_ms_index
, /* MSAA sample index */
1286 nir_tex_src_ms_mcs
, /* MSAA compression value */
1289 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
1290 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
1291 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
1292 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
1293 nir_tex_src_plane
, /* < selects plane for planar textures */
1294 nir_num_tex_src_types
1299 nir_tex_src_type src_type
;
1303 nir_texop_tex
, /**< Regular texture look-up */
1304 nir_texop_txb
, /**< Texture look-up with LOD bias */
1305 nir_texop_txl
, /**< Texture look-up with explicit LOD */
1306 nir_texop_txd
, /**< Texture look-up with partial derivatives */
1307 nir_texop_txf
, /**< Texel fetch with explicit LOD */
1308 nir_texop_txf_ms
, /**< Multisample texture fetch */
1309 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
1310 nir_texop_txs
, /**< Texture size */
1311 nir_texop_lod
, /**< Texture lod query */
1312 nir_texop_tg4
, /**< Texture gather */
1313 nir_texop_query_levels
, /**< Texture levels query */
1314 nir_texop_texture_samples
, /**< Texture samples query */
1315 nir_texop_samples_identical
, /**< Query whether all samples are definitely
1323 enum glsl_sampler_dim sampler_dim
;
1324 nir_alu_type dest_type
;
1329 unsigned num_srcs
, coord_components
;
1330 bool is_array
, is_shadow
;
1333 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1334 * components or the new-style shadow that outputs 1 component.
1336 bool is_new_style_shadow
;
1338 /* gather component selector */
1339 unsigned component
: 2;
1341 /** The texture index
1343 * If this texture instruction has a nir_tex_src_texture_offset source,
1344 * then the texture index is given by texture_index + texture_offset.
1346 unsigned texture_index
;
1348 /** The size of the texture array or 0 if it's not an array */
1349 unsigned texture_array_size
;
1351 /** The sampler index
1353 * The following operations do not require a sampler and, as such, this
1354 * field should be ignored:
1356 * - nir_texop_txf_ms
1359 * - nir_texop_query_levels
1360 * - nir_texop_texture_samples
1361 * - nir_texop_samples_identical
1363 * If this texture instruction has a nir_tex_src_sampler_offset source,
1364 * then the sampler index is given by sampler_index + sampler_offset.
1366 unsigned sampler_index
;
1369 static inline unsigned
1370 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
1372 switch (instr
->op
) {
1373 case nir_texop_txs
: {
1375 switch (instr
->sampler_dim
) {
1376 case GLSL_SAMPLER_DIM_1D
:
1377 case GLSL_SAMPLER_DIM_BUF
:
1380 case GLSL_SAMPLER_DIM_2D
:
1381 case GLSL_SAMPLER_DIM_CUBE
:
1382 case GLSL_SAMPLER_DIM_MS
:
1383 case GLSL_SAMPLER_DIM_RECT
:
1384 case GLSL_SAMPLER_DIM_EXTERNAL
:
1385 case GLSL_SAMPLER_DIM_SUBPASS
:
1388 case GLSL_SAMPLER_DIM_3D
:
1392 unreachable("not reached");
1394 if (instr
->is_array
)
1402 case nir_texop_texture_samples
:
1403 case nir_texop_query_levels
:
1404 case nir_texop_samples_identical
:
1408 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
1415 /* Returns true if this texture operation queries something about the texture
1416 * rather than actually sampling it.
1419 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
1421 switch (instr
->op
) {
1424 case nir_texop_texture_samples
:
1425 case nir_texop_query_levels
:
1426 case nir_texop_txf_ms_mcs
:
1433 case nir_texop_txf_ms
:
1437 unreachable("Invalid texture opcode");
1442 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1444 switch (instr
->op
) {
1465 static inline nir_alu_type
1466 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
1468 switch (instr
->src
[src
].src_type
) {
1469 case nir_tex_src_coord
:
1470 switch (instr
->op
) {
1472 case nir_texop_txf_ms
:
1473 case nir_texop_txf_ms_mcs
:
1474 case nir_texop_samples_identical
:
1475 return nir_type_int
;
1478 return nir_type_float
;
1481 case nir_tex_src_lod
:
1482 switch (instr
->op
) {
1485 return nir_type_int
;
1488 return nir_type_float
;
1491 case nir_tex_src_projector
:
1492 case nir_tex_src_comparator
:
1493 case nir_tex_src_bias
:
1494 case nir_tex_src_ddx
:
1495 case nir_tex_src_ddy
:
1496 return nir_type_float
;
1498 case nir_tex_src_offset
:
1499 case nir_tex_src_ms_index
:
1500 case nir_tex_src_texture_offset
:
1501 case nir_tex_src_sampler_offset
:
1502 return nir_type_int
;
1505 unreachable("Invalid texture source type");
1509 static inline unsigned
1510 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
1512 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
1513 return instr
->coord_components
;
1515 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
1516 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
1519 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
1520 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
1521 if (instr
->is_array
)
1522 return instr
->coord_components
- 1;
1524 return instr
->coord_components
;
1527 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
1528 * the offset, since a cube maps to a single face.
1530 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
1531 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
1533 else if (instr
->is_array
)
1534 return instr
->coord_components
- 1;
1536 return instr
->coord_components
;
1543 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
1545 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
1546 if (instr
->src
[i
].src_type
== type
)
1552 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
1553 nir_tex_src_type src_type
,
1556 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
1561 nir_const_value value
;
1564 } nir_load_const_instr
;
1577 /* creates a new SSA variable in an undefined state */
1582 } nir_ssa_undef_instr
;
1585 struct exec_node node
;
1587 /* The predecessor block corresponding to this source */
1588 struct nir_block
*pred
;
1593 #define nir_foreach_phi_src(phi_src, phi) \
1594 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
1595 #define nir_foreach_phi_src_safe(phi_src, phi) \
1596 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
1601 struct exec_list srcs
; /** < list of nir_phi_src */
1607 struct exec_node node
;
1610 } nir_parallel_copy_entry
;
1612 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
1613 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
1618 /* A list of nir_parallel_copy_entrys. The sources of all of the
1619 * entries are copied to the corresponding destinations "in parallel".
1620 * In other words, if we have two entries: a -> b and b -> a, the values
1623 struct exec_list entries
;
1624 } nir_parallel_copy_instr
;
1626 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
1627 type
, nir_instr_type_alu
)
1628 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
1629 type
, nir_instr_type_call
)
1630 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
1631 type
, nir_instr_type_jump
)
1632 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
1633 type
, nir_instr_type_tex
)
1634 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
1635 type
, nir_instr_type_intrinsic
)
1636 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
1637 type
, nir_instr_type_load_const
)
1638 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
1639 type
, nir_instr_type_ssa_undef
)
1640 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
1641 type
, nir_instr_type_phi
)
1642 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
1643 nir_parallel_copy_instr
, instr
,
1644 type
, nir_instr_type_parallel_copy
)
1649 * Control flow consists of a tree of control flow nodes, which include
1650 * if-statements and loops. The leaves of the tree are basic blocks, lists of
1651 * instructions that always run start-to-finish. Each basic block also keeps
1652 * track of its successors (blocks which may run immediately after the current
1653 * block) and predecessors (blocks which could have run immediately before the
1654 * current block). Each function also has a start block and an end block which
1655 * all return statements point to (which is always empty). Together, all the
1656 * blocks with their predecessors and successors make up the control flow
1657 * graph (CFG) of the function. There are helpers that modify the tree of
1658 * control flow nodes while modifying the CFG appropriately; these should be
1659 * used instead of modifying the tree directly.
1666 nir_cf_node_function
1669 typedef struct nir_cf_node
{
1670 struct exec_node node
;
1671 nir_cf_node_type type
;
1672 struct nir_cf_node
*parent
;
1675 typedef struct nir_block
{
1676 nir_cf_node cf_node
;
1678 struct exec_list instr_list
; /** < list of nir_instr */
1680 /** generic block index; generated by nir_index_blocks */
1684 * Each block can only have up to 2 successors, so we put them in a simple
1685 * array - no need for anything more complicated.
1687 struct nir_block
*successors
[2];
1689 /* Set of nir_block predecessors in the CFG */
1690 struct set
*predecessors
;
1693 * this node's immediate dominator in the dominance tree - set to NULL for
1696 struct nir_block
*imm_dom
;
1698 /* This node's children in the dominance tree */
1699 unsigned num_dom_children
;
1700 struct nir_block
**dom_children
;
1702 /* Set of nir_blocks on the dominance frontier of this block */
1703 struct set
*dom_frontier
;
1706 * These two indices have the property that dom_{pre,post}_index for each
1707 * child of this block in the dominance tree will always be between
1708 * dom_pre_index and dom_post_index for this block, which makes testing if
1709 * a given block is dominated by another block an O(1) operation.
1711 unsigned dom_pre_index
, dom_post_index
;
1713 /* live in and out for this block; used for liveness analysis */
1714 BITSET_WORD
*live_in
;
1715 BITSET_WORD
*live_out
;
1718 static inline nir_instr
*
1719 nir_block_first_instr(nir_block
*block
)
1721 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
1722 return exec_node_data(nir_instr
, head
, node
);
1725 static inline nir_instr
*
1726 nir_block_last_instr(nir_block
*block
)
1728 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
1729 return exec_node_data(nir_instr
, tail
, node
);
1732 #define nir_foreach_instr(instr, block) \
1733 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
1734 #define nir_foreach_instr_reverse(instr, block) \
1735 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
1736 #define nir_foreach_instr_safe(instr, block) \
1737 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
1738 #define nir_foreach_instr_reverse_safe(instr, block) \
1739 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
1741 typedef struct nir_if
{
1742 nir_cf_node cf_node
;
1745 struct exec_list then_list
; /** < list of nir_cf_node */
1746 struct exec_list else_list
; /** < list of nir_cf_node */
1752 nir_instr
*conditional_instr
;
1754 nir_block
*break_block
;
1755 nir_block
*continue_from_block
;
1757 bool continue_from_then
;
1759 struct list_head loop_terminator_link
;
1760 } nir_loop_terminator
;
1763 /* Number of instructions in the loop */
1764 unsigned num_instructions
;
1766 /* How many times the loop is run (if known) */
1767 unsigned trip_count
;
1768 bool is_trip_count_known
;
1770 /* Unroll the loop regardless of its size */
1773 nir_loop_terminator
*limiting_terminator
;
1775 /* A list of loop_terminators terminating this loop. */
1776 struct list_head loop_terminator_list
;
1780 nir_cf_node cf_node
;
1782 struct exec_list body
; /** < list of nir_cf_node */
1784 nir_loop_info
*info
;
1788 * Various bits of metadata that can may be created or required by
1789 * optimization and analysis passes
1792 nir_metadata_none
= 0x0,
1793 nir_metadata_block_index
= 0x1,
1794 nir_metadata_dominance
= 0x2,
1795 nir_metadata_live_ssa_defs
= 0x4,
1796 nir_metadata_not_properly_reset
= 0x8,
1797 nir_metadata_loop_analysis
= 0x10,
1801 nir_cf_node cf_node
;
1803 /** pointer to the function of which this is an implementation */
1804 struct nir_function
*function
;
1806 struct exec_list body
; /** < list of nir_cf_node */
1808 nir_block
*end_block
;
1810 /** list for all local variables in the function */
1811 struct exec_list locals
;
1813 /** list of local registers in the function */
1814 struct exec_list registers
;
1816 /** next available local register index */
1819 /** next available SSA value index */
1822 /* total number of basic blocks, only valid when block_index_dirty = false */
1823 unsigned num_blocks
;
1825 nir_metadata valid_metadata
;
1826 } nir_function_impl
;
1828 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
1829 nir_start_block(nir_function_impl
*impl
)
1831 return (nir_block
*) impl
->body
.head_sentinel
.next
;
1834 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
1835 nir_impl_last_block(nir_function_impl
*impl
)
1837 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
1840 static inline nir_cf_node
*
1841 nir_cf_node_next(nir_cf_node
*node
)
1843 struct exec_node
*next
= exec_node_get_next(&node
->node
);
1844 if (exec_node_is_tail_sentinel(next
))
1847 return exec_node_data(nir_cf_node
, next
, node
);
1850 static inline nir_cf_node
*
1851 nir_cf_node_prev(nir_cf_node
*node
)
1853 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
1854 if (exec_node_is_head_sentinel(prev
))
1857 return exec_node_data(nir_cf_node
, prev
, node
);
1861 nir_cf_node_is_first(const nir_cf_node
*node
)
1863 return exec_node_is_head_sentinel(node
->node
.prev
);
1867 nir_cf_node_is_last(const nir_cf_node
*node
)
1869 return exec_node_is_tail_sentinel(node
->node
.next
);
1872 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
1873 type
, nir_cf_node_block
)
1874 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
1875 type
, nir_cf_node_if
)
1876 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
1877 type
, nir_cf_node_loop
)
1878 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
1879 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
1881 static inline nir_block
*
1882 nir_if_first_then_block(nir_if
*if_stmt
)
1884 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
1885 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
1888 static inline nir_block
*
1889 nir_if_last_then_block(nir_if
*if_stmt
)
1891 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
1892 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
1895 static inline nir_block
*
1896 nir_if_first_else_block(nir_if
*if_stmt
)
1898 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
1899 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
1902 static inline nir_block
*
1903 nir_if_last_else_block(nir_if
*if_stmt
)
1905 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
1906 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
1909 static inline nir_block
*
1910 nir_loop_first_block(nir_loop
*loop
)
1912 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
1913 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
1916 static inline nir_block
*
1917 nir_loop_last_block(nir_loop
*loop
)
1919 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
1920 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
1924 uint8_t num_components
;
1928 typedef struct nir_function
{
1929 struct exec_node node
;
1932 struct nir_shader
*shader
;
1934 unsigned num_params
;
1935 nir_parameter
*params
;
1937 /** The implementation of this function.
1939 * If the function is only declared and not implemented, this is NULL.
1941 nir_function_impl
*impl
;
1944 typedef struct nir_shader_compiler_options
{
1949 /** Lowers flrp when it does not support doubles */
1956 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
1957 bool lower_bitfield_extract
;
1958 /** Lowers ibitfield_extract/ubitfield_extract to bfm, compares, shifts. */
1959 bool lower_bitfield_extract_to_shifts
;
1960 /** Lowers bitfield_insert to bfi/bfm */
1961 bool lower_bitfield_insert
;
1962 /** Lowers bitfield_insert to bfm, compares, and shifts. */
1963 bool lower_bitfield_insert_to_shifts
;
1964 /** Lowers bitfield_reverse to shifts. */
1965 bool lower_bitfield_reverse
;
1966 /** Lowers bit_count to shifts. */
1967 bool lower_bit_count
;
1968 /** Lowers bfm to shifts and subtracts. */
1970 /** Lowers ifind_msb to compare and ufind_msb */
1971 bool lower_ifind_msb
;
1972 /** Lowers find_lsb to ufind_msb and logic ops */
1973 bool lower_find_lsb
;
1974 bool lower_uadd_carry
;
1975 bool lower_usub_borrow
;
1976 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
1977 bool lower_mul_high
;
1978 /** lowers fneg and ineg to fsub and isub. */
1980 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
1983 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
1986 /** enables rules to lower idiv by power-of-two: */
1989 /* lower b2f to iand */
1992 /* Does the native fdot instruction replicate its result for four
1993 * components? If so, then opt_algebraic_late will turn all fdotN
1994 * instructions into fdot_replicatedN instructions.
1996 bool fdot_replicates
;
1998 /** lowers ffract to fsub+ffloor: */
2003 bool lower_pack_half_2x16
;
2004 bool lower_pack_unorm_2x16
;
2005 bool lower_pack_snorm_2x16
;
2006 bool lower_pack_unorm_4x8
;
2007 bool lower_pack_snorm_4x8
;
2008 bool lower_unpack_half_2x16
;
2009 bool lower_unpack_unorm_2x16
;
2010 bool lower_unpack_snorm_2x16
;
2011 bool lower_unpack_unorm_4x8
;
2012 bool lower_unpack_snorm_4x8
;
2014 bool lower_extract_byte
;
2015 bool lower_extract_word
;
2017 bool lower_all_io_to_temps
;
2020 * Does the driver support real 32-bit integers? (Otherwise, integers
2021 * are simulated by floats.)
2023 bool native_integers
;
2025 /* Indicates that the driver only has zero-based vertex id */
2026 bool vertex_id_zero_based
;
2029 * If enabled, gl_BaseVertex will be lowered as:
2030 * is_indexed_draw (~0/0) & firstvertex
2032 bool lower_base_vertex
;
2035 * If enabled, gl_HelperInvocation will be lowered as:
2037 * !((1 << sample_id) & sample_mask_in))
2039 * This depends on some possibly hw implementation details, which may
2040 * not be true for all hw. In particular that the FS is only executed
2041 * for covered samples or for helper invocations. So, do not blindly
2042 * enable this option.
2044 * Note: See also issue #22 in ARB_shader_image_load_store
2046 bool lower_helper_invocation
;
2048 bool lower_cs_local_index_from_id
;
2050 bool lower_device_index_to_zero
;
2052 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
2053 bool lower_wpos_pntc
;
2056 * Should nir_lower_io() create load_interpolated_input intrinsics?
2058 * If not, it generates regular load_input intrinsics and interpolation
2059 * information must be inferred from the list of input nir_variables.
2061 bool use_interpolated_input_intrinsics
;
2064 * Do vertex shader double inputs use two locations? The Vulkan spec
2065 * requires two locations to be used, OpenGL allows a single location.
2067 bool vs_inputs_dual_locations
;
2069 unsigned max_unroll_iterations
;
2070 } nir_shader_compiler_options
;
2072 typedef struct nir_shader
{
2073 /** list of uniforms (nir_variable) */
2074 struct exec_list uniforms
;
2076 /** list of inputs (nir_variable) */
2077 struct exec_list inputs
;
2079 /** list of outputs (nir_variable) */
2080 struct exec_list outputs
;
2082 /** list of shared compute variables (nir_variable) */
2083 struct exec_list shared
;
2085 /** Set of driver-specific options for the shader.
2087 * The memory for the options is expected to be kept in a single static
2088 * copy by the driver.
2090 const struct nir_shader_compiler_options
*options
;
2092 /** Various bits of compile-time information about a given shader */
2093 struct shader_info info
;
2095 /** list of global variables in the shader (nir_variable) */
2096 struct exec_list globals
;
2098 /** list of system value variables in the shader (nir_variable) */
2099 struct exec_list system_values
;
2101 struct exec_list functions
; /** < list of nir_function */
2103 /** list of global register in the shader */
2104 struct exec_list registers
;
2106 /** next available global register index */
2110 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
2113 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
2115 /** Constant data associated with this shader.
2117 * Constant data is loaded through load_constant intrinsics. See also
2118 * nir_opt_large_constants.
2120 void *constant_data
;
2121 unsigned constant_data_size
;
2124 static inline nir_function_impl
*
2125 nir_shader_get_entrypoint(nir_shader
*shader
)
2127 assert(exec_list_length(&shader
->functions
) == 1);
2128 struct exec_node
*func_node
= exec_list_get_head(&shader
->functions
);
2129 nir_function
*func
= exec_node_data(nir_function
, func_node
, node
);
2130 assert(func
->num_params
== 0);
2135 #define nir_foreach_function(func, shader) \
2136 foreach_list_typed(nir_function, func, node, &(shader)->functions)
2138 nir_shader
*nir_shader_create(void *mem_ctx
,
2139 gl_shader_stage stage
,
2140 const nir_shader_compiler_options
*options
,
2143 /** creates a register, including assigning it an index and adding it to the list */
2144 nir_register
*nir_global_reg_create(nir_shader
*shader
);
2146 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
2148 void nir_reg_remove(nir_register
*reg
);
2150 /** Adds a variable to the appropriate list in nir_shader */
2151 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
2154 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
2156 assert(var
->data
.mode
== nir_var_local
);
2157 exec_list_push_tail(&impl
->locals
, &var
->node
);
2160 /** creates a variable, sets a few defaults, and adds it to the list */
2161 nir_variable
*nir_variable_create(nir_shader
*shader
,
2162 nir_variable_mode mode
,
2163 const struct glsl_type
*type
,
2165 /** creates a local variable and adds it to the list */
2166 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
2167 const struct glsl_type
*type
,
2170 /** creates a function and adds it to the shader's list of functions */
2171 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
2173 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
2174 /** creates a function_impl that isn't tied to any particular function */
2175 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
2177 nir_block
*nir_block_create(nir_shader
*shader
);
2178 nir_if
*nir_if_create(nir_shader
*shader
);
2179 nir_loop
*nir_loop_create(nir_shader
*shader
);
2181 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
2183 /** requests that the given pieces of metadata be generated */
2184 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
2185 /** dirties all but the preserved metadata */
2186 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
2188 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
2189 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
2191 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
2192 nir_deref_type deref_type
);
2194 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
2196 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
2197 unsigned num_components
,
2200 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
2201 nir_intrinsic_op op
);
2203 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
2204 nir_function
*callee
);
2206 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
2208 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
2210 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
2212 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
2213 unsigned num_components
,
2216 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
2219 * NIR Cursors and Instruction Insertion API
2222 * A tiny struct representing a point to insert/extract instructions or
2223 * control flow nodes. Helps reduce the combinatorial explosion of possible
2224 * points to insert/extract.
2226 * \sa nir_control_flow.h
2229 nir_cursor_before_block
,
2230 nir_cursor_after_block
,
2231 nir_cursor_before_instr
,
2232 nir_cursor_after_instr
,
2233 } nir_cursor_option
;
2236 nir_cursor_option option
;
2243 static inline nir_block
*
2244 nir_cursor_current_block(nir_cursor cursor
)
2246 if (cursor
.option
== nir_cursor_before_instr
||
2247 cursor
.option
== nir_cursor_after_instr
) {
2248 return cursor
.instr
->block
;
2250 return cursor
.block
;
2254 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
2256 static inline nir_cursor
2257 nir_before_block(nir_block
*block
)
2260 cursor
.option
= nir_cursor_before_block
;
2261 cursor
.block
= block
;
2265 static inline nir_cursor
2266 nir_after_block(nir_block
*block
)
2269 cursor
.option
= nir_cursor_after_block
;
2270 cursor
.block
= block
;
2274 static inline nir_cursor
2275 nir_before_instr(nir_instr
*instr
)
2278 cursor
.option
= nir_cursor_before_instr
;
2279 cursor
.instr
= instr
;
2283 static inline nir_cursor
2284 nir_after_instr(nir_instr
*instr
)
2287 cursor
.option
= nir_cursor_after_instr
;
2288 cursor
.instr
= instr
;
2292 static inline nir_cursor
2293 nir_after_block_before_jump(nir_block
*block
)
2295 nir_instr
*last_instr
= nir_block_last_instr(block
);
2296 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
2297 return nir_before_instr(last_instr
);
2299 return nir_after_block(block
);
2303 static inline nir_cursor
2304 nir_before_cf_node(nir_cf_node
*node
)
2306 if (node
->type
== nir_cf_node_block
)
2307 return nir_before_block(nir_cf_node_as_block(node
));
2309 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
2312 static inline nir_cursor
2313 nir_after_cf_node(nir_cf_node
*node
)
2315 if (node
->type
== nir_cf_node_block
)
2316 return nir_after_block(nir_cf_node_as_block(node
));
2318 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
2321 static inline nir_cursor
2322 nir_after_phis(nir_block
*block
)
2324 nir_foreach_instr(instr
, block
) {
2325 if (instr
->type
!= nir_instr_type_phi
)
2326 return nir_before_instr(instr
);
2328 return nir_after_block(block
);
2331 static inline nir_cursor
2332 nir_after_cf_node_and_phis(nir_cf_node
*node
)
2334 if (node
->type
== nir_cf_node_block
)
2335 return nir_after_block(nir_cf_node_as_block(node
));
2337 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
2339 return nir_after_phis(block
);
2342 static inline nir_cursor
2343 nir_before_cf_list(struct exec_list
*cf_list
)
2345 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
2346 exec_list_get_head(cf_list
), node
);
2347 return nir_before_cf_node(first_node
);
2350 static inline nir_cursor
2351 nir_after_cf_list(struct exec_list
*cf_list
)
2353 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
2354 exec_list_get_tail(cf_list
), node
);
2355 return nir_after_cf_node(last_node
);
2359 * Insert a NIR instruction at the given cursor.
2361 * Note: This does not update the cursor.
2363 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
2366 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
2368 nir_instr_insert(nir_before_instr(instr
), before
);
2372 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
2374 nir_instr_insert(nir_after_instr(instr
), after
);
2378 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
2380 nir_instr_insert(nir_before_block(block
), before
);
2384 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
2386 nir_instr_insert(nir_after_block(block
), after
);
2390 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
2392 nir_instr_insert(nir_before_cf_node(node
), before
);
2396 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
2398 nir_instr_insert(nir_after_cf_node(node
), after
);
2402 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
2404 nir_instr_insert(nir_before_cf_list(list
), before
);
2408 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
2410 nir_instr_insert(nir_after_cf_list(list
), after
);
2413 void nir_instr_remove_v(nir_instr
*instr
);
2415 static inline nir_cursor
2416 nir_instr_remove(nir_instr
*instr
)
2419 nir_instr
*prev
= nir_instr_prev(instr
);
2421 cursor
= nir_after_instr(prev
);
2423 cursor
= nir_before_block(instr
->block
);
2425 nir_instr_remove_v(instr
);
2431 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
2432 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
2433 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
2434 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
2436 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
2437 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
2439 nir_const_value
*nir_src_as_const_value(nir_src src
);
2440 bool nir_src_is_dynamically_uniform(nir_src src
);
2441 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
2442 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
2443 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
2444 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
2445 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
2448 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
2449 unsigned num_components
, unsigned bit_size
,
2451 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
2452 unsigned num_components
, unsigned bit_size
,
2455 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
2456 const struct glsl_type
*type
,
2459 assert(glsl_type_is_vector_or_scalar(type
));
2460 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
2461 glsl_get_bit_size(type
), name
);
2463 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
2464 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
2465 nir_instr
*after_me
);
2467 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
2470 * finds the next basic block in source-code order, returns NULL if there is
2474 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
2476 /* Performs the opposite of nir_block_cf_tree_next() */
2478 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
2480 /* Gets the first block in a CF node in source-code order */
2482 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
2484 /* Gets the last block in a CF node in source-code order */
2486 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
2488 /* Gets the next block after a CF node in source-code order */
2490 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
2492 /* Macros for loops that visit blocks in source-code order */
2494 #define nir_foreach_block(block, impl) \
2495 for (nir_block *block = nir_start_block(impl); block != NULL; \
2496 block = nir_block_cf_tree_next(block))
2498 #define nir_foreach_block_safe(block, impl) \
2499 for (nir_block *block = nir_start_block(impl), \
2500 *next = nir_block_cf_tree_next(block); \
2502 block = next, next = nir_block_cf_tree_next(block))
2504 #define nir_foreach_block_reverse(block, impl) \
2505 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
2506 block = nir_block_cf_tree_prev(block))
2508 #define nir_foreach_block_reverse_safe(block, impl) \
2509 for (nir_block *block = nir_impl_last_block(impl), \
2510 *prev = nir_block_cf_tree_prev(block); \
2512 block = prev, prev = nir_block_cf_tree_prev(block))
2514 #define nir_foreach_block_in_cf_node(block, node) \
2515 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
2516 block != nir_cf_node_cf_tree_next(node); \
2517 block = nir_block_cf_tree_next(block))
2519 /* If the following CF node is an if, this function returns that if.
2520 * Otherwise, it returns NULL.
2522 nir_if
*nir_block_get_following_if(nir_block
*block
);
2524 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
2526 void nir_index_local_regs(nir_function_impl
*impl
);
2527 void nir_index_global_regs(nir_shader
*shader
);
2528 void nir_index_ssa_defs(nir_function_impl
*impl
);
2529 unsigned nir_index_instrs(nir_function_impl
*impl
);
2531 void nir_index_blocks(nir_function_impl
*impl
);
2533 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
2534 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
2535 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
2537 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
2538 nir_function_impl
*nir_function_impl_clone(const nir_function_impl
*fi
);
2539 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
2540 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
2542 nir_shader
*nir_shader_serialize_deserialize(void *mem_ctx
, nir_shader
*s
);
2545 void nir_validate_shader(nir_shader
*shader
);
2546 void nir_metadata_set_validation_flag(nir_shader
*shader
);
2547 void nir_metadata_check_validation_flag(nir_shader
*shader
);
2550 should_clone_nir(void)
2552 static int should_clone
= -1;
2553 if (should_clone
< 0)
2554 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
2556 return should_clone
;
2560 should_serialize_deserialize_nir(void)
2562 static int test_serialize
= -1;
2563 if (test_serialize
< 0)
2564 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
2566 return test_serialize
;
2570 should_print_nir(void)
2572 static int should_print
= -1;
2573 if (should_print
< 0)
2574 should_print
= env_var_as_boolean("NIR_PRINT", false);
2576 return should_print
;
2579 static inline void nir_validate_shader(nir_shader
*shader
) { (void) shader
; }
2580 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
2581 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
2582 static inline bool should_clone_nir(void) { return false; }
2583 static inline bool should_serialize_deserialize_nir(void) { return false; }
2584 static inline bool should_print_nir(void) { return false; }
2587 #define _PASS(nir, do_pass) do { \
2589 nir_validate_shader(nir); \
2590 if (should_clone_nir()) { \
2591 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
2595 if (should_serialize_deserialize_nir()) { \
2596 void *mem_ctx = ralloc_parent(nir); \
2597 nir = nir_shader_serialize_deserialize(mem_ctx, nir); \
2601 #define NIR_PASS(progress, nir, pass, ...) _PASS(nir, \
2602 nir_metadata_set_validation_flag(nir); \
2603 if (should_print_nir()) \
2604 printf("%s\n", #pass); \
2605 if (pass(nir, ##__VA_ARGS__)) { \
2607 if (should_print_nir()) \
2608 nir_print_shader(nir, stdout); \
2609 nir_metadata_check_validation_flag(nir); \
2613 #define NIR_PASS_V(nir, pass, ...) _PASS(nir, \
2614 if (should_print_nir()) \
2615 printf("%s\n", #pass); \
2616 pass(nir, ##__VA_ARGS__); \
2617 if (should_print_nir()) \
2618 nir_print_shader(nir, stdout); \
2621 void nir_calc_dominance_impl(nir_function_impl
*impl
);
2622 void nir_calc_dominance(nir_shader
*shader
);
2624 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
2625 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
2627 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
2628 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
2630 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
2631 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
2633 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
2634 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
2636 int nir_gs_count_vertices(const nir_shader
*shader
);
2638 bool nir_split_var_copies(nir_shader
*shader
);
2639 bool nir_split_per_member_structs(nir_shader
*shader
);
2641 bool nir_lower_returns_impl(nir_function_impl
*impl
);
2642 bool nir_lower_returns(nir_shader
*shader
);
2644 bool nir_inline_functions(nir_shader
*shader
);
2646 bool nir_propagate_invariant(nir_shader
*shader
);
2648 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
2649 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
2650 nir_intrinsic_instr
*copy
);
2651 bool nir_lower_var_copies(nir_shader
*shader
);
2653 void nir_fixup_deref_modes(nir_shader
*shader
);
2655 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
2657 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
2659 bool nir_lower_locals_to_regs(nir_shader
*shader
);
2661 void nir_lower_io_to_temporaries(nir_shader
*shader
,
2662 nir_function_impl
*entrypoint
,
2663 bool outputs
, bool inputs
);
2665 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
2667 void nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
2668 int (*type_size
)(const struct glsl_type
*));
2670 /* Some helpers to do very simple linking */
2671 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
2672 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
2673 bool default_to_smooth_interp
);
2676 /* If set, this forces all non-flat fragment shader inputs to be
2677 * interpolated as if with the "sample" qualifier. This requires
2678 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
2680 nir_lower_io_force_sample_interpolation
= (1 << 1),
2681 } nir_lower_io_options
;
2682 bool nir_lower_io(nir_shader
*shader
,
2683 nir_variable_mode modes
,
2684 int (*type_size
)(const struct glsl_type
*),
2685 nir_lower_io_options
);
2686 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
2687 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
2689 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
2691 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
2692 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
2693 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
2695 bool nir_remove_dead_derefs(nir_shader
*shader
);
2696 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
2697 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
);
2698 bool nir_lower_constant_initializers(nir_shader
*shader
,
2699 nir_variable_mode modes
);
2701 bool nir_move_load_const(nir_shader
*shader
);
2702 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
2703 bool nir_lower_vec_to_movs(nir_shader
*shader
);
2704 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
2706 bool nir_lower_alu(nir_shader
*shader
);
2707 bool nir_lower_alu_to_scalar(nir_shader
*shader
);
2708 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
2709 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
2710 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
2711 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
2712 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
2714 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
2715 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
2717 typedef struct nir_lower_subgroups_options
{
2718 uint8_t subgroup_size
;
2719 uint8_t ballot_bit_size
;
2720 bool lower_to_scalar
:1;
2721 bool lower_vote_trivial
:1;
2722 bool lower_vote_eq_to_ballot
:1;
2723 bool lower_subgroup_masks
:1;
2724 bool lower_shuffle
:1;
2725 bool lower_shuffle_to_32bit
:1;
2727 } nir_lower_subgroups_options
;
2729 bool nir_lower_subgroups(nir_shader
*shader
,
2730 const nir_lower_subgroups_options
*options
);
2732 bool nir_lower_system_values(nir_shader
*shader
);
2734 typedef struct nir_lower_tex_options
{
2736 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
2737 * sampler types a texture projector is lowered.
2742 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
2744 bool lower_txf_offset
;
2747 * If true, lower away nir_tex_src_offset for all rect textures.
2749 bool lower_rect_offset
;
2752 * If true, lower rect textures to 2D, using txs to fetch the
2753 * texture dimensions and dividing the texture coords by the
2754 * texture dims to normalize.
2759 * If true, convert yuv to rgb.
2761 unsigned lower_y_uv_external
;
2762 unsigned lower_y_u_v_external
;
2763 unsigned lower_yx_xuxv_external
;
2764 unsigned lower_xy_uxvx_external
;
2767 * To emulate certain texture wrap modes, this can be used
2768 * to saturate the specified tex coord to [0.0, 1.0]. The
2769 * bits are according to sampler #, ie. if, for example:
2771 * (conf->saturate_s & (1 << n))
2773 * is true, then the s coord for sampler n is saturated.
2775 * Note that clamping must happen *after* projector lowering
2776 * so any projected texture sample instruction with a clamped
2777 * coordinate gets automatically lowered, regardless of the
2778 * 'lower_txp' setting.
2780 unsigned saturate_s
;
2781 unsigned saturate_t
;
2782 unsigned saturate_r
;
2784 /* Bitmask of textures that need swizzling.
2786 * If (swizzle_result & (1 << texture_index)), then the swizzle in
2787 * swizzles[texture_index] is applied to the result of the texturing
2790 unsigned swizzle_result
;
2792 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
2793 * while 4 and 5 represent 0 and 1 respectively.
2795 uint8_t swizzles
[32][4];
2798 * Bitmap of textures that need srgb to linear conversion. If
2799 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
2800 * of the texture are lowered to linear.
2802 unsigned lower_srgb
;
2805 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
2807 bool lower_txd_cube_map
;
2810 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
2811 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
2812 * with lower_txd_cube_map.
2814 bool lower_txd_shadow
;
2817 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
2818 * Implies lower_txd_cube_map and lower_txd_shadow.
2821 } nir_lower_tex_options
;
2823 bool nir_lower_tex(nir_shader
*shader
,
2824 const nir_lower_tex_options
*options
);
2826 bool nir_lower_idiv(nir_shader
*shader
);
2828 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
);
2829 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
);
2830 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
2832 void nir_lower_two_sided_color(nir_shader
*shader
);
2834 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
2836 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
2837 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
2838 const gl_state_index16
*uniform_state_tokens
);
2840 typedef struct nir_lower_wpos_ytransform_options
{
2841 gl_state_index16 state_tokens
[STATE_LENGTH
];
2842 bool fs_coord_origin_upper_left
:1;
2843 bool fs_coord_origin_lower_left
:1;
2844 bool fs_coord_pixel_center_integer
:1;
2845 bool fs_coord_pixel_center_half_integer
:1;
2846 } nir_lower_wpos_ytransform_options
;
2848 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
2849 const nir_lower_wpos_ytransform_options
*options
);
2850 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
2852 typedef struct nir_lower_drawpixels_options
{
2853 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
2854 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
2855 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
2856 unsigned drawpix_sampler
;
2857 unsigned pixelmap_sampler
;
2859 bool scale_and_bias
:1;
2860 } nir_lower_drawpixels_options
;
2862 void nir_lower_drawpixels(nir_shader
*shader
,
2863 const nir_lower_drawpixels_options
*options
);
2865 typedef struct nir_lower_bitmap_options
{
2868 } nir_lower_bitmap_options
;
2870 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
2872 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
, unsigned ssbo_offset
);
2873 bool nir_lower_to_source_mods(nir_shader
*shader
);
2875 bool nir_lower_gs_intrinsics(nir_shader
*shader
);
2877 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
2879 bool nir_lower_bit_size(nir_shader
*shader
,
2880 nir_lower_bit_size_callback callback
,
2881 void *callback_data
);
2884 nir_lower_imul64
= (1 << 0),
2885 nir_lower_isign64
= (1 << 1),
2886 /** Lower all int64 modulus and division opcodes */
2887 nir_lower_divmod64
= (1 << 2),
2888 } nir_lower_int64_options
;
2890 bool nir_lower_int64(nir_shader
*shader
, nir_lower_int64_options options
);
2893 nir_lower_drcp
= (1 << 0),
2894 nir_lower_dsqrt
= (1 << 1),
2895 nir_lower_drsq
= (1 << 2),
2896 nir_lower_dtrunc
= (1 << 3),
2897 nir_lower_dfloor
= (1 << 4),
2898 nir_lower_dceil
= (1 << 5),
2899 nir_lower_dfract
= (1 << 6),
2900 nir_lower_dround_even
= (1 << 7),
2901 nir_lower_dmod
= (1 << 8)
2902 } nir_lower_doubles_options
;
2904 bool nir_lower_doubles(nir_shader
*shader
, nir_lower_doubles_options options
);
2905 bool nir_lower_pack(nir_shader
*shader
);
2907 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
2909 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
2911 void nir_loop_analyze_impl(nir_function_impl
*impl
,
2912 nir_variable_mode indirect_mask
);
2914 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
2916 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
2917 bool nir_repair_ssa(nir_shader
*shader
);
2919 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
2921 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
2922 * registers. If false, convert all values (even those not involved in a phi
2923 * node) to registers.
2925 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
2927 bool nir_lower_phis_to_regs_block(nir_block
*block
);
2928 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
2930 bool nir_opt_algebraic(nir_shader
*shader
);
2931 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
2932 bool nir_opt_algebraic_late(nir_shader
*shader
);
2933 bool nir_opt_constant_folding(nir_shader
*shader
);
2935 bool nir_opt_global_to_local(nir_shader
*shader
);
2937 bool nir_copy_prop(nir_shader
*shader
);
2939 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
2941 bool nir_opt_cse(nir_shader
*shader
);
2943 bool nir_opt_dce(nir_shader
*shader
);
2945 bool nir_opt_dead_cf(nir_shader
*shader
);
2947 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
2949 bool nir_opt_if(nir_shader
*shader
);
2951 bool nir_opt_intrinsics(nir_shader
*shader
);
2953 bool nir_opt_large_constants(nir_shader
*shader
,
2954 glsl_type_size_align_func size_align
,
2955 unsigned threshold
);
2957 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
2959 bool nir_opt_move_comparisons(nir_shader
*shader
);
2961 bool nir_opt_move_load_ubo(nir_shader
*shader
);
2963 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
);
2965 bool nir_opt_remove_phis_impl(nir_function_impl
*impl
);
2966 bool nir_opt_remove_phis(nir_shader
*shader
);
2968 bool nir_opt_shrink_load(nir_shader
*shader
);
2970 bool nir_opt_trivial_continues(nir_shader
*shader
);
2972 bool nir_opt_undef(nir_shader
*shader
);
2974 bool nir_opt_conditional_discard(nir_shader
*shader
);
2976 void nir_sweep(nir_shader
*shader
);
2978 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
2979 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);