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 #define NIR_STREAM_PACKED (1 << 8)
64 typedef uint8_t nir_component_mask_t
;
65 typedef unsigned short GLenum16
;
67 /** Defines a cast function
69 * This macro defines a cast function from in_type to out_type where
70 * out_type is some structure type that contains a field of type out_type.
72 * Note that you have to be a bit careful as the generated cast function
75 #define NIR_DEFINE_CAST(name, in_type, out_type, field, \
76 type_field, type_value) \
77 static inline out_type * \
78 name(const in_type *parent) \
80 assert(parent && parent->type_field == type_value); \
81 return exec_node_data(out_type, parent, field); \
91 * Description of built-in state associated with a uniform
93 * \sa nir_variable::state_slots
96 gl_state_index16 tokens
[STATE_LENGTH
];
101 nir_var_shader_in
= (1 << 0),
102 nir_var_shader_out
= (1 << 1),
103 nir_var_shader_temp
= (1 << 2),
104 nir_var_function_temp
= (1 << 3),
105 nir_var_uniform
= (1 << 4),
106 nir_var_mem_ubo
= (1 << 5),
107 nir_var_system_value
= (1 << 6),
108 nir_var_mem_ssbo
= (1 << 7),
109 nir_var_mem_shared
= (1 << 8),
110 nir_var_mem_global
= (1 << 9),
111 nir_var_all
= (1 << 10) - 1,
118 nir_rounding_mode_undef
= 0,
119 nir_rounding_mode_rtne
= 1, /* round to nearest even */
120 nir_rounding_mode_ru
= 2, /* round up */
121 nir_rounding_mode_rd
= 3, /* round down */
122 nir_rounding_mode_rtz
= 4, /* round towards zero */
139 #define nir_const_value_to_array(arr, c, components, m) \
141 for (unsigned i = 0; i < components; ++i) \
145 static inline nir_const_value
146 nir_const_value_for_raw_uint(uint64_t x
, unsigned bit_size
)
149 memset(&v
, 0, sizeof(v
));
152 case 1: v
.b
= x
; break;
153 case 8: v
.u8
= x
; break;
154 case 16: v
.u16
= x
; break;
155 case 32: v
.u32
= x
; break;
156 case 64: v
.u64
= x
; break;
158 unreachable("Invalid bit size");
164 static inline nir_const_value
165 nir_const_value_for_int(int64_t i
, unsigned bit_size
)
168 memset(&v
, 0, sizeof(v
));
170 assert(bit_size
<= 64);
172 assert(i
>= (-(1ll << (bit_size
- 1))));
173 assert(i
< (1ll << (bit_size
- 1)));
176 return nir_const_value_for_raw_uint(i
, bit_size
);
179 static inline nir_const_value
180 nir_const_value_for_uint(uint64_t u
, unsigned bit_size
)
183 memset(&v
, 0, sizeof(v
));
185 assert(bit_size
<= 64);
187 assert(u
< (1ull << bit_size
));
189 return nir_const_value_for_raw_uint(u
, bit_size
);
192 static inline nir_const_value
193 nir_const_value_for_bool(bool b
, unsigned bit_size
)
195 /* Booleans use a 0/-1 convention */
196 return nir_const_value_for_int(-(int)b
, bit_size
);
199 /* This one isn't inline because it requires half-float conversion */
200 nir_const_value
nir_const_value_for_float(double b
, unsigned bit_size
);
202 static inline int64_t
203 nir_const_value_as_int(nir_const_value value
, unsigned bit_size
)
206 /* int1_t uses 0/-1 convention */
207 case 1: return -(int)value
.b
;
208 case 8: return value
.i8
;
209 case 16: return value
.i16
;
210 case 32: return value
.i32
;
211 case 64: return value
.i64
;
213 unreachable("Invalid bit size");
217 static inline uint64_t
218 nir_const_value_as_uint(nir_const_value value
, unsigned bit_size
)
221 case 1: return value
.b
;
222 case 8: return value
.u8
;
223 case 16: return value
.u16
;
224 case 32: return value
.u32
;
225 case 64: return value
.u64
;
227 unreachable("Invalid bit size");
232 nir_const_value_as_bool(nir_const_value value
, unsigned bit_size
)
234 int64_t i
= nir_const_value_as_int(value
, bit_size
);
236 /* Booleans of any size use 0/-1 convention */
237 assert(i
== 0 || i
== -1);
242 /* This one isn't inline because it requires half-float conversion */
243 double nir_const_value_as_float(nir_const_value value
, unsigned bit_size
);
245 typedef struct nir_constant
{
247 * Value of the constant.
249 * The field used to back the values supplied by the constant is determined
250 * by the type associated with the \c nir_variable. Constants may be
251 * scalars, vectors, or matrices.
253 nir_const_value values
[NIR_MAX_VEC_COMPONENTS
];
255 /* we could get this from the var->type but makes clone *much* easier to
256 * not have to care about the type.
258 unsigned num_elements
;
260 /* Array elements / Structure Fields */
261 struct nir_constant
**elements
;
265 * \brief Layout qualifiers for gl_FragDepth.
267 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
268 * with a layout qualifier.
271 nir_depth_layout_none
, /**< No depth layout is specified. */
272 nir_depth_layout_any
,
273 nir_depth_layout_greater
,
274 nir_depth_layout_less
,
275 nir_depth_layout_unchanged
279 * Enum keeping track of how a variable was declared.
283 * Normal declaration.
285 nir_var_declared_normally
= 0,
288 * Variable is implicitly generated by the compiler and should not be
289 * visible via the API.
292 } nir_var_declaration_type
;
295 * Either a uniform, global variable, shader input, or shader output. Based on
296 * ir_variable - it should be easy to translate between the two.
299 typedef struct nir_variable
{
300 struct exec_node node
;
303 * Declared type of the variable
305 const struct glsl_type
*type
;
308 * Declared name of the variable
312 struct nir_variable_data
{
314 * Storage class of the variable.
316 * \sa nir_variable_mode
318 nir_variable_mode mode
:10;
321 * Is the variable read-only?
323 * This is set for variables declared as \c const, shader inputs,
326 unsigned read_only
:1;
330 unsigned invariant
:1;
333 * Can this variable be coalesced with another?
335 * This is set by nir_lower_io_to_temporaries to say that any
336 * copies involving this variable should stay put. Propagating it can
337 * duplicate the resulting load/store, which is not wanted, and may
338 * result in a load/store of the variable with an indirect offset which
339 * the backend may not be able to handle.
341 unsigned cannot_coalesce
:1;
344 * When separate shader programs are enabled, only input/outputs between
345 * the stages of a multi-stage separate program can be safely removed
346 * from the shader interface. Other input/outputs must remains active.
348 * This is also used to make sure xfb varyings that are unused by the
349 * fragment shader are not removed.
351 unsigned always_active_io
:1;
354 * Interpolation mode for shader inputs / outputs
356 * \sa glsl_interp_mode
358 unsigned interpolation
:2;
361 * If non-zero, then this variable may be packed along with other variables
362 * into a single varying slot, so this offset should be applied when
363 * accessing components. For example, an offset of 1 means that the x
364 * component of this variable is actually stored in component y of the
365 * location specified by \c location.
367 unsigned location_frac
:2;
370 * If true, this variable represents an array of scalars that should
371 * be tightly packed. In other words, consecutive array elements
372 * should be stored one component apart, rather than one slot apart.
377 * Whether this is a fragment shader output implicitly initialized with
378 * the previous contents of the specified render target at the
379 * framebuffer location corresponding to this shader invocation.
381 unsigned fb_fetch_output
:1;
384 * Non-zero if this variable is considered bindless as defined by
385 * ARB_bindless_texture.
390 * Was an explicit binding set in the shader?
392 unsigned explicit_binding
:1;
395 * Was a transfer feedback buffer set in the shader?
397 unsigned explicit_xfb_buffer
:1;
400 * Was a transfer feedback stride set in the shader?
402 unsigned explicit_xfb_stride
:1;
405 * Was an explicit offset set in the shader?
407 unsigned explicit_offset
:1;
410 * How the variable was declared. See nir_var_declaration_type.
412 * This is used to detect variables generated by the compiler, so should
413 * not be visible via the API.
415 unsigned how_declared
:2;
418 * \brief Layout qualifier for gl_FragDepth.
420 * This is not equal to \c ir_depth_layout_none if and only if this
421 * variable is \c gl_FragDepth and a layout qualifier is specified.
423 nir_depth_layout depth_layout
:3;
426 * Vertex stream output identifier.
428 * For packed outputs, NIR_STREAM_PACKED is set and bits [2*i+1,2*i]
429 * indicate the stream of the i-th component.
434 * Access flags for memory variables (SSBO/global), image uniforms, and
435 * bindless images in uniforms/inputs/outputs.
437 enum gl_access_qualifier access
:8;
440 * Descriptor set binding for sampler or UBO.
442 unsigned descriptor_set
:5;
445 * output index for dual source blending.
450 * Initial binding point for a sampler or UBO.
452 * For array types, this represents the binding point for the first element.
457 * Storage location of the base of this variable
459 * The precise meaning of this field depends on the nature of the variable.
461 * - Vertex shader input: one of the values from \c gl_vert_attrib.
462 * - Vertex shader output: one of the values from \c gl_varying_slot.
463 * - Geometry shader input: one of the values from \c gl_varying_slot.
464 * - Geometry shader output: one of the values from \c gl_varying_slot.
465 * - Fragment shader input: one of the values from \c gl_varying_slot.
466 * - Fragment shader output: one of the values from \c gl_frag_result.
467 * - Uniforms: Per-stage uniform slot number for default uniform block.
468 * - Uniforms: Index within the uniform block definition for UBO members.
469 * - Non-UBO Uniforms: uniform slot number.
470 * - Other: This field is not currently used.
472 * If the variable is a uniform, shader input, or shader output, and the
473 * slot has not been assigned, the value will be -1.
478 * The actual location of the variable in the IR. Only valid for inputs,
479 * outputs, and uniforms (including samplers and images).
481 unsigned driver_location
;
484 * Location an atomic counter or transform feedback is stored at.
490 /** Image internal format if specified explicitly, otherwise GL_NONE. */
496 * Transform feedback buffer.
501 * Transform feedback stride.
508 /* Number of nir_variable_data members */
509 unsigned num_members
;
512 * Built-in state that backs this uniform
514 * Once set at variable creation, \c state_slots must remain invariant.
515 * This is because, ideally, this array would be shared by all clones of
516 * this variable in the IR tree. In other words, we'd really like for it
517 * to be a fly-weight.
519 * If the variable is not a uniform, \c num_state_slots will be zero and
520 * \c state_slots will be \c NULL.
523 unsigned num_state_slots
; /**< Number of state slots used */
524 nir_state_slot
*state_slots
; /**< State descriptors. */
528 * Constant expression assigned in the initializer of the variable
530 * This field should only be used temporarily by creators of NIR shaders
531 * and then lower_constant_initializers can be used to get rid of them.
532 * Most of the rest of NIR ignores this field or asserts that it's NULL.
534 nir_constant
*constant_initializer
;
537 * For variables that are in an interface block or are an instance of an
538 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
540 * \sa ir_variable::location
542 const struct glsl_type
*interface_type
;
545 * Description of per-member data for per-member struct variables
547 * This is used for variables which are actually an amalgamation of
548 * multiple entities such as a struct of built-in values or a struct of
549 * inputs each with their own layout specifier. This is only allowed on
550 * variables with a struct or array of array of struct type.
552 struct nir_variable_data
*members
;
555 #define nir_foreach_variable(var, var_list) \
556 foreach_list_typed(nir_variable, var, node, var_list)
558 #define nir_foreach_variable_safe(var, var_list) \
559 foreach_list_typed_safe(nir_variable, var, node, var_list)
562 nir_variable_is_global(const nir_variable
*var
)
564 return var
->data
.mode
!= nir_var_function_temp
;
567 typedef struct nir_register
{
568 struct exec_node node
;
570 unsigned num_components
; /** < number of vector components */
571 unsigned num_array_elems
; /** < size of array (0 for no array) */
573 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
576 /** generic register index. */
579 /** only for debug purposes, can be NULL */
582 /** set of nir_srcs where this register is used (read from) */
583 struct list_head uses
;
585 /** set of nir_dests where this register is defined (written to) */
586 struct list_head defs
;
588 /** set of nir_ifs where this register is used as a condition */
589 struct list_head if_uses
;
592 #define nir_foreach_register(reg, reg_list) \
593 foreach_list_typed(nir_register, reg, node, reg_list)
594 #define nir_foreach_register_safe(reg, reg_list) \
595 foreach_list_typed_safe(nir_register, reg, node, reg_list)
597 typedef enum PACKED
{
599 nir_instr_type_deref
,
602 nir_instr_type_intrinsic
,
603 nir_instr_type_load_const
,
605 nir_instr_type_ssa_undef
,
607 nir_instr_type_parallel_copy
,
610 typedef struct nir_instr
{
611 struct exec_node node
;
612 struct nir_block
*block
;
615 /* A temporary for optimization and analysis passes to use for storing
616 * flags. For instance, DCE uses this to store the "dead/live" info.
620 /** generic instruction index. */
624 static inline nir_instr
*
625 nir_instr_next(nir_instr
*instr
)
627 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
628 if (exec_node_is_tail_sentinel(next
))
631 return exec_node_data(nir_instr
, next
, node
);
634 static inline nir_instr
*
635 nir_instr_prev(nir_instr
*instr
)
637 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
638 if (exec_node_is_head_sentinel(prev
))
641 return exec_node_data(nir_instr
, prev
, node
);
645 nir_instr_is_first(const nir_instr
*instr
)
647 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr
->node
));
651 nir_instr_is_last(const nir_instr
*instr
)
653 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr
->node
));
656 typedef struct nir_ssa_def
{
657 /** for debugging only, can be NULL */
660 /** generic SSA definition index. */
663 /** Index into the live_in and live_out bitfields */
666 /** Instruction which produces this SSA value. */
667 nir_instr
*parent_instr
;
669 /** set of nir_instrs where this register is used (read from) */
670 struct list_head uses
;
672 /** set of nir_ifs where this register is used as a condition */
673 struct list_head if_uses
;
675 uint8_t num_components
;
677 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
685 struct nir_src
*indirect
; /** < NULL for no indirect offset */
686 unsigned base_offset
;
688 /* TODO use-def chain goes here */
692 nir_instr
*parent_instr
;
693 struct list_head def_link
;
696 struct nir_src
*indirect
; /** < NULL for no indirect offset */
697 unsigned base_offset
;
699 /* TODO def-use chain goes here */
704 typedef struct nir_src
{
706 /** Instruction that consumes this value as a source. */
707 nir_instr
*parent_instr
;
708 struct nir_if
*parent_if
;
711 struct list_head use_link
;
721 static inline nir_src
724 nir_src src
= { { NULL
} };
728 #define NIR_SRC_INIT nir_src_init()
730 #define nir_foreach_use(src, reg_or_ssa_def) \
731 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
733 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
734 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
736 #define nir_foreach_if_use(src, reg_or_ssa_def) \
737 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
739 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
740 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
751 static inline nir_dest
754 nir_dest dest
= { { { NULL
} } };
758 #define NIR_DEST_INIT nir_dest_init()
760 #define nir_foreach_def(dest, reg) \
761 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
763 #define nir_foreach_def_safe(dest, reg) \
764 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
766 static inline nir_src
767 nir_src_for_ssa(nir_ssa_def
*def
)
769 nir_src src
= NIR_SRC_INIT
;
777 static inline nir_src
778 nir_src_for_reg(nir_register
*reg
)
780 nir_src src
= NIR_SRC_INIT
;
784 src
.reg
.indirect
= NULL
;
785 src
.reg
.base_offset
= 0;
790 static inline nir_dest
791 nir_dest_for_reg(nir_register
*reg
)
793 nir_dest dest
= NIR_DEST_INIT
;
800 static inline unsigned
801 nir_src_bit_size(nir_src src
)
803 return src
.is_ssa
? src
.ssa
->bit_size
: src
.reg
.reg
->bit_size
;
806 static inline unsigned
807 nir_src_num_components(nir_src src
)
809 return src
.is_ssa
? src
.ssa
->num_components
: src
.reg
.reg
->num_components
;
813 nir_src_is_const(nir_src src
)
816 src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
;
819 static inline unsigned
820 nir_dest_bit_size(nir_dest dest
)
822 return dest
.is_ssa
? dest
.ssa
.bit_size
: dest
.reg
.reg
->bit_size
;
825 static inline unsigned
826 nir_dest_num_components(nir_dest dest
)
828 return dest
.is_ssa
? dest
.ssa
.num_components
: dest
.reg
.reg
->num_components
;
831 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
832 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
838 * \name input modifiers
842 * For inputs interpreted as floating point, flips the sign bit. For
843 * inputs interpreted as integers, performs the two's complement negation.
848 * Clears the sign bit for floating point values, and computes the integer
849 * absolute value for integers. Note that the negate modifier acts after
850 * the absolute value modifier, therefore if both are set then all inputs
851 * will become negative.
857 * For each input component, says which component of the register it is
858 * chosen from. Note that which elements of the swizzle are used and which
859 * are ignored are based on the write mask for most opcodes - for example,
860 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
861 * a swizzle of {2, x, 1, 0} where x means "don't care."
863 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
870 * \name saturate output modifier
872 * Only valid for opcodes that output floating-point numbers. Clamps the
873 * output to between 0.0 and 1.0 inclusive.
878 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
881 /** NIR sized and unsized types
883 * The values in this enum are carefully chosen so that the sized type is
884 * just the unsized type OR the number of bits.
887 nir_type_invalid
= 0, /* Not a valid type */
891 nir_type_float
= 128,
892 nir_type_bool1
= 1 | nir_type_bool
,
893 nir_type_bool32
= 32 | nir_type_bool
,
894 nir_type_int1
= 1 | nir_type_int
,
895 nir_type_int8
= 8 | nir_type_int
,
896 nir_type_int16
= 16 | nir_type_int
,
897 nir_type_int32
= 32 | nir_type_int
,
898 nir_type_int64
= 64 | nir_type_int
,
899 nir_type_uint1
= 1 | nir_type_uint
,
900 nir_type_uint8
= 8 | nir_type_uint
,
901 nir_type_uint16
= 16 | nir_type_uint
,
902 nir_type_uint32
= 32 | nir_type_uint
,
903 nir_type_uint64
= 64 | nir_type_uint
,
904 nir_type_float16
= 16 | nir_type_float
,
905 nir_type_float32
= 32 | nir_type_float
,
906 nir_type_float64
= 64 | nir_type_float
,
909 #define NIR_ALU_TYPE_SIZE_MASK 0x79
910 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
912 static inline unsigned
913 nir_alu_type_get_type_size(nir_alu_type type
)
915 return type
& NIR_ALU_TYPE_SIZE_MASK
;
918 static inline unsigned
919 nir_alu_type_get_base_type(nir_alu_type type
)
921 return type
& NIR_ALU_TYPE_BASE_TYPE_MASK
;
924 static inline nir_alu_type
925 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
929 return nir_type_bool1
;
932 return nir_type_uint32
;
935 return nir_type_int32
;
937 case GLSL_TYPE_UINT16
:
938 return nir_type_uint16
;
940 case GLSL_TYPE_INT16
:
941 return nir_type_int16
;
943 case GLSL_TYPE_UINT8
:
944 return nir_type_uint8
;
946 return nir_type_int8
;
947 case GLSL_TYPE_UINT64
:
948 return nir_type_uint64
;
950 case GLSL_TYPE_INT64
:
951 return nir_type_int64
;
953 case GLSL_TYPE_FLOAT
:
954 return nir_type_float32
;
956 case GLSL_TYPE_FLOAT16
:
957 return nir_type_float16
;
959 case GLSL_TYPE_DOUBLE
:
960 return nir_type_float64
;
963 case GLSL_TYPE_SAMPLER
:
964 case GLSL_TYPE_IMAGE
:
965 case GLSL_TYPE_ATOMIC_UINT
:
966 case GLSL_TYPE_STRUCT
:
967 case GLSL_TYPE_INTERFACE
:
968 case GLSL_TYPE_ARRAY
:
970 case GLSL_TYPE_SUBROUTINE
:
971 case GLSL_TYPE_FUNCTION
:
972 case GLSL_TYPE_ERROR
:
973 return nir_type_invalid
;
976 unreachable("unknown type");
979 static inline nir_alu_type
980 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
982 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
985 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
986 nir_rounding_mode rnd
);
989 nir_op_vec(unsigned components
)
991 switch (components
) {
992 case 1: return nir_op_mov
;
993 case 2: return nir_op_vec2
;
994 case 3: return nir_op_vec3
;
995 case 4: return nir_op_vec4
;
996 default: unreachable("bad component count");
1001 nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode
, unsigned bit_size
)
1003 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
) ||
1004 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
) ||
1005 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
);
1009 nir_is_denorm_flush_to_zero(unsigned execution_mode
, unsigned bit_size
)
1011 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
) ||
1012 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
) ||
1013 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
);
1017 nir_is_denorm_preserve(unsigned execution_mode
, unsigned bit_size
)
1019 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP16
) ||
1020 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP32
) ||
1021 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP64
);
1025 nir_is_rounding_mode_rtne(unsigned execution_mode
, unsigned bit_size
)
1027 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1028 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1029 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1033 nir_is_rounding_mode_rtz(unsigned execution_mode
, unsigned bit_size
)
1035 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1036 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1037 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1041 nir_has_any_rounding_mode_rtz(unsigned execution_mode
)
1043 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1044 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1045 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1049 nir_has_any_rounding_mode_rtne(unsigned execution_mode
)
1051 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1052 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1053 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1056 static inline nir_rounding_mode
1057 nir_get_rounding_mode_from_float_controls(unsigned execution_mode
,
1060 if (nir_alu_type_get_base_type(type
) != nir_type_float
)
1061 return nir_rounding_mode_undef
;
1063 unsigned bit_size
= nir_alu_type_get_type_size(type
);
1065 if (nir_is_rounding_mode_rtz(execution_mode
, bit_size
))
1066 return nir_rounding_mode_rtz
;
1067 if (nir_is_rounding_mode_rtne(execution_mode
, bit_size
))
1068 return nir_rounding_mode_rtne
;
1069 return nir_rounding_mode_undef
;
1073 nir_has_any_rounding_mode_enabled(unsigned execution_mode
)
1076 nir_has_any_rounding_mode_rtne(execution_mode
) ||
1077 nir_has_any_rounding_mode_rtz(execution_mode
);
1083 * Operation where the first two sources are commutative.
1085 * For 2-source operations, this just mathematical commutativity. Some
1086 * 3-source operations, like ffma, are only commutative in the first two
1089 NIR_OP_IS_2SRC_COMMUTATIVE
= (1 << 0),
1090 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
1091 } nir_op_algebraic_property
;
1096 unsigned num_inputs
;
1099 * The number of components in the output
1101 * If non-zero, this is the size of the output and input sizes are
1102 * explicitly given; swizzle and writemask are still in effect, but if
1103 * the output component is masked out, then the input component may
1106 * If zero, the opcode acts in the standard, per-component manner; the
1107 * operation is performed on each component (except the ones that are
1108 * masked out) with the input being taken from the input swizzle for
1111 * The size of some of the inputs may be given (i.e. non-zero) even
1112 * though output_size is zero; in that case, the inputs with a zero
1113 * size act per-component, while the inputs with non-zero size don't.
1115 unsigned output_size
;
1118 * The type of vector that the instruction outputs. Note that the
1119 * staurate modifier is only allowed on outputs with the float type.
1122 nir_alu_type output_type
;
1125 * The number of components in each input
1127 unsigned input_sizes
[NIR_MAX_VEC_COMPONENTS
];
1130 * The type of vector that each input takes. Note that negate and
1131 * absolute value are only allowed on inputs with int or float type and
1132 * behave differently on the two.
1134 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
1136 nir_op_algebraic_property algebraic_properties
;
1138 /* Whether this represents a numeric conversion opcode */
1142 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
1144 typedef struct nir_alu_instr
{
1148 /** Indicates that this ALU instruction generates an exact value
1150 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1151 * really equivalent to either. This indicates that the value generated by
1152 * this operation is high-precision and any code transformations that touch
1153 * it must ensure that the resulting value is bit-for-bit identical to the
1159 * Indicates that this instruction do not cause wrapping to occur, in the
1160 * form of overflow or underflow.
1162 bool no_signed_wrap
:1;
1163 bool no_unsigned_wrap
:1;
1169 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
1170 nir_alu_instr
*instr
);
1171 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
1172 nir_alu_instr
*instr
);
1174 /* is this source channel used? */
1176 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
1179 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1180 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
1182 return (instr
->dest
.write_mask
>> channel
) & 1;
1185 static inline nir_component_mask_t
1186 nir_alu_instr_src_read_mask(const nir_alu_instr
*instr
, unsigned src
)
1188 nir_component_mask_t read_mask
= 0;
1189 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; c
++) {
1190 if (!nir_alu_instr_channel_used(instr
, src
, c
))
1193 read_mask
|= (1 << instr
->src
[src
].swizzle
[c
]);
1199 * Get the number of channels used for a source
1201 static inline unsigned
1202 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
1204 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1205 return nir_op_infos
[instr
->op
].input_sizes
[src
];
1207 return nir_dest_num_components(instr
->dest
.dest
);
1211 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1213 switch (instr
->op
) {
1233 bool nir_const_value_negative_equal(nir_const_value c1
, nir_const_value c2
,
1234 nir_alu_type full_type
);
1236 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
1237 unsigned src1
, unsigned src2
);
1239 bool nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
1240 const nir_alu_instr
*alu2
,
1241 unsigned src1
, unsigned src2
);
1245 nir_deref_type_array
,
1246 nir_deref_type_array_wildcard
,
1247 nir_deref_type_ptr_as_array
,
1248 nir_deref_type_struct
,
1249 nir_deref_type_cast
,
1255 /** The type of this deref instruction */
1256 nir_deref_type deref_type
;
1258 /** The mode of the underlying variable */
1259 nir_variable_mode mode
;
1261 /** The dereferenced type of the resulting pointer value */
1262 const struct glsl_type
*type
;
1265 /** Variable being dereferenced if deref_type is a deref_var */
1268 /** Parent deref if deref_type is not deref_var */
1272 /** Additional deref parameters */
1283 unsigned ptr_stride
;
1287 /** Destination to store the resulting "pointer" */
1291 static inline nir_deref_instr
*nir_src_as_deref(nir_src src
);
1293 static inline nir_deref_instr
*
1294 nir_deref_instr_parent(const nir_deref_instr
*instr
)
1296 if (instr
->deref_type
== nir_deref_type_var
)
1299 return nir_src_as_deref(instr
->parent
);
1302 static inline nir_variable
*
1303 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1305 while (instr
->deref_type
!= nir_deref_type_var
) {
1306 if (instr
->deref_type
== nir_deref_type_cast
)
1309 instr
= nir_deref_instr_parent(instr
);
1315 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1316 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr
*instr
);
1317 bool nir_deref_instr_has_complex_use(nir_deref_instr
*instr
);
1319 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1321 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*instr
);
1326 struct nir_function
*callee
;
1328 unsigned num_params
;
1332 #include "nir_intrinsics.h"
1334 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1336 /** Represents an intrinsic
1338 * An intrinsic is an instruction type for handling things that are
1339 * more-or-less regular operations but don't just consume and produce SSA
1340 * values like ALU operations do. Intrinsics are not for things that have
1341 * special semantic meaning such as phi nodes and parallel copies.
1342 * Examples of intrinsics include variable load/store operations, system
1343 * value loads, and the like. Even though texturing more-or-less falls
1344 * under this category, texturing is its own instruction type because
1345 * trying to represent texturing with intrinsics would lead to a
1346 * combinatorial explosion of intrinsic opcodes.
1348 * By having a single instruction type for handling a lot of different
1349 * cases, optimization passes can look for intrinsics and, for the most
1350 * part, completely ignore them. Each intrinsic type also has a few
1351 * possible flags that govern whether or not they can be reordered or
1352 * eliminated. That way passes like dead code elimination can still work
1353 * on intrisics without understanding the meaning of each.
1355 * Each intrinsic has some number of constant indices, some number of
1356 * variables, and some number of sources. What these sources, variables,
1357 * and indices mean depends on the intrinsic and is documented with the
1358 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1359 * instructions are the only types of instruction that can operate on
1365 nir_intrinsic_op intrinsic
;
1369 /** number of components if this is a vectorized intrinsic
1371 * Similarly to ALU operations, some intrinsics are vectorized.
1372 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1373 * For vectorized intrinsics, the num_components field specifies the
1374 * number of destination components and the number of source components
1375 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1377 uint8_t num_components
;
1379 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1382 } nir_intrinsic_instr
;
1384 static inline nir_variable
*
1385 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1387 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1391 /* Memory ordering. */
1392 NIR_MEMORY_ACQUIRE
= 1 << 0,
1393 NIR_MEMORY_RELEASE
= 1 << 1,
1395 /* Memory visibility operations. */
1396 NIR_MEMORY_MAKE_AVAILABLE
= 1 << 3,
1397 NIR_MEMORY_MAKE_VISIBLE
= 1 << 4,
1398 } nir_memory_semantics
;
1402 NIR_SCOPE_QUEUE_FAMILY
,
1403 NIR_SCOPE_WORKGROUP
,
1405 NIR_SCOPE_INVOCATION
,
1409 * \name NIR intrinsics semantic flags
1411 * information about what the compiler can do with the intrinsics.
1413 * \sa nir_intrinsic_info::flags
1417 * whether the intrinsic can be safely eliminated if none of its output
1418 * value is not being used.
1420 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1423 * Whether the intrinsic can be reordered with respect to any other
1424 * intrinsic, i.e. whether the only reordering dependencies of the
1425 * intrinsic are due to the register reads/writes.
1427 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1428 } nir_intrinsic_semantic_flag
;
1431 * \name NIR intrinsics const-index flag
1433 * Indicates the usage of a const_index slot.
1435 * \sa nir_intrinsic_info::index_map
1439 * Generally instructions that take a offset src argument, can encode
1440 * a constant 'base' value which is added to the offset.
1442 NIR_INTRINSIC_BASE
= 1,
1445 * For store instructions, a writemask for the store.
1447 NIR_INTRINSIC_WRMASK
,
1450 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1452 NIR_INTRINSIC_STREAM_ID
,
1455 * The clip-plane id for load_user_clip_plane intrinsic.
1457 NIR_INTRINSIC_UCP_ID
,
1460 * The amount of data, starting from BASE, that this instruction may
1461 * access. This is used to provide bounds if the offset is not constant.
1463 NIR_INTRINSIC_RANGE
,
1466 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1468 NIR_INTRINSIC_DESC_SET
,
1471 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1473 NIR_INTRINSIC_BINDING
,
1478 NIR_INTRINSIC_COMPONENT
,
1481 * Interpolation mode (only meaningful for FS inputs).
1483 NIR_INTRINSIC_INTERP_MODE
,
1486 * A binary nir_op to use when performing a reduction or scan operation
1488 NIR_INTRINSIC_REDUCTION_OP
,
1491 * Cluster size for reduction operations
1493 NIR_INTRINSIC_CLUSTER_SIZE
,
1496 * Parameter index for a load_param intrinsic
1498 NIR_INTRINSIC_PARAM_IDX
,
1501 * Image dimensionality for image intrinsics
1503 * One of GLSL_SAMPLER_DIM_*
1505 NIR_INTRINSIC_IMAGE_DIM
,
1508 * Non-zero if we are accessing an array image
1510 NIR_INTRINSIC_IMAGE_ARRAY
,
1513 * Image format for image intrinsics
1515 NIR_INTRINSIC_FORMAT
,
1518 * Access qualifiers for image and memory access intrinsics
1520 NIR_INTRINSIC_ACCESS
,
1523 * Alignment for offsets and addresses
1525 * These two parameters, specify an alignment in terms of a multiplier and
1526 * an offset. The offset or address parameter X of the intrinsic is
1527 * guaranteed to satisfy the following:
1529 * (X - align_offset) % align_mul == 0
1531 NIR_INTRINSIC_ALIGN_MUL
,
1532 NIR_INTRINSIC_ALIGN_OFFSET
,
1535 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1537 NIR_INTRINSIC_DESC_TYPE
,
1540 * The nir_alu_type of a uniform/input/output
1545 * The swizzle mask for the instructions
1546 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1548 NIR_INTRINSIC_SWIZZLE_MASK
,
1550 /* Separate source/dest access flags for copies */
1551 NIR_INTRINSIC_SRC_ACCESS
,
1552 NIR_INTRINSIC_DST_ACCESS
,
1554 /* Driver location for nir_load_patch_location_ir3 */
1555 NIR_INTRINSIC_DRIVER_LOCATION
,
1558 * Mask of nir_memory_semantics, includes ordering and visibility.
1560 NIR_INTRINSIC_MEMORY_SEMANTICS
,
1563 * Mask of nir_variable_modes affected by the memory operation.
1565 NIR_INTRINSIC_MEMORY_MODES
,
1568 * Value of nir_scope.
1570 NIR_INTRINSIC_MEMORY_SCOPE
,
1572 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1574 } nir_intrinsic_index_flag
;
1576 #define NIR_INTRINSIC_MAX_INPUTS 5
1581 unsigned num_srcs
; /** < number of register/SSA inputs */
1583 /** number of components of each input register
1585 * If this value is 0, the number of components is given by the
1586 * num_components field of nir_intrinsic_instr. If this value is -1, the
1587 * intrinsic consumes however many components are provided and it is not
1590 int src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1594 /** number of components of the output register
1596 * If this value is 0, the number of components is given by the
1597 * num_components field of nir_intrinsic_instr.
1599 unsigned dest_components
;
1601 /** bitfield of legal bit sizes */
1602 unsigned dest_bit_sizes
;
1604 /** the number of constant indices used by the intrinsic */
1605 unsigned num_indices
;
1607 /** indicates the usage of intr->const_index[n] */
1608 unsigned index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1610 /** semantic flags for calls to this intrinsic */
1611 nir_intrinsic_semantic_flag flags
;
1612 } nir_intrinsic_info
;
1614 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1616 static inline unsigned
1617 nir_intrinsic_src_components(nir_intrinsic_instr
*intr
, unsigned srcn
)
1619 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1620 assert(srcn
< info
->num_srcs
);
1621 if (info
->src_components
[srcn
] > 0)
1622 return info
->src_components
[srcn
];
1623 else if (info
->src_components
[srcn
] == 0)
1624 return intr
->num_components
;
1626 return nir_src_num_components(intr
->src
[srcn
]);
1629 static inline unsigned
1630 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1632 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1633 if (!info
->has_dest
)
1635 else if (info
->dest_components
)
1636 return info
->dest_components
;
1638 return intr
->num_components
;
1641 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1642 static inline type \
1643 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1645 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1646 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1647 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1649 static inline void \
1650 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1652 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1653 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1654 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1657 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1658 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1659 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1660 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1661 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1662 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1663 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1664 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1665 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1666 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1667 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1668 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1669 INTRINSIC_IDX_ACCESSORS(image_dim
, IMAGE_DIM
, enum glsl_sampler_dim
)
1670 INTRINSIC_IDX_ACCESSORS(image_array
, IMAGE_ARRAY
, bool)
1671 INTRINSIC_IDX_ACCESSORS(access
, ACCESS
, enum gl_access_qualifier
)
1672 INTRINSIC_IDX_ACCESSORS(src_access
, SRC_ACCESS
, enum gl_access_qualifier
)
1673 INTRINSIC_IDX_ACCESSORS(dst_access
, DST_ACCESS
, enum gl_access_qualifier
)
1674 INTRINSIC_IDX_ACCESSORS(format
, FORMAT
, unsigned)
1675 INTRINSIC_IDX_ACCESSORS(align_mul
, ALIGN_MUL
, unsigned)
1676 INTRINSIC_IDX_ACCESSORS(align_offset
, ALIGN_OFFSET
, unsigned)
1677 INTRINSIC_IDX_ACCESSORS(desc_type
, DESC_TYPE
, unsigned)
1678 INTRINSIC_IDX_ACCESSORS(type
, TYPE
, nir_alu_type
)
1679 INTRINSIC_IDX_ACCESSORS(swizzle_mask
, SWIZZLE_MASK
, unsigned)
1680 INTRINSIC_IDX_ACCESSORS(driver_location
, DRIVER_LOCATION
, unsigned)
1681 INTRINSIC_IDX_ACCESSORS(memory_semantics
, MEMORY_SEMANTICS
, nir_memory_semantics
)
1682 INTRINSIC_IDX_ACCESSORS(memory_modes
, MEMORY_MODES
, nir_variable_mode
)
1683 INTRINSIC_IDX_ACCESSORS(memory_scope
, MEMORY_SCOPE
, nir_scope
)
1686 nir_intrinsic_set_align(nir_intrinsic_instr
*intrin
,
1687 unsigned align_mul
, unsigned align_offset
)
1689 assert(util_is_power_of_two_nonzero(align_mul
));
1690 assert(align_offset
< align_mul
);
1691 nir_intrinsic_set_align_mul(intrin
, align_mul
);
1692 nir_intrinsic_set_align_offset(intrin
, align_offset
);
1695 /** Returns a simple alignment for a load/store intrinsic offset
1697 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1698 * and ALIGN_OFFSET parameters, this helper takes both into account and
1699 * provides a single simple alignment parameter. The offset X is guaranteed
1700 * to satisfy X % align == 0.
1702 static inline unsigned
1703 nir_intrinsic_align(const nir_intrinsic_instr
*intrin
)
1705 const unsigned align_mul
= nir_intrinsic_align_mul(intrin
);
1706 const unsigned align_offset
= nir_intrinsic_align_offset(intrin
);
1707 assert(align_offset
< align_mul
);
1708 return align_offset
? 1 << (ffs(align_offset
) - 1) : align_mul
;
1711 /* Converts a image_deref_* intrinsic into a image_* one */
1712 void nir_rewrite_image_intrinsic(nir_intrinsic_instr
*instr
,
1713 nir_ssa_def
*handle
, bool bindless
);
1715 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1717 nir_intrinsic_can_reorder(nir_intrinsic_instr
*instr
)
1719 if (instr
->intrinsic
== nir_intrinsic_load_deref
||
1720 instr
->intrinsic
== nir_intrinsic_load_ssbo
||
1721 instr
->intrinsic
== nir_intrinsic_bindless_image_load
||
1722 instr
->intrinsic
== nir_intrinsic_image_deref_load
||
1723 instr
->intrinsic
== nir_intrinsic_image_load
) {
1724 return nir_intrinsic_access(instr
) & ACCESS_CAN_REORDER
;
1726 const nir_intrinsic_info
*info
=
1727 &nir_intrinsic_infos
[instr
->intrinsic
];
1728 return (info
->flags
& NIR_INTRINSIC_CAN_ELIMINATE
) &&
1729 (info
->flags
& NIR_INTRINSIC_CAN_REORDER
);
1734 * \group texture information
1736 * This gives semantic information about textures which is useful to the
1737 * frontend, the backend, and lowering passes, but not the optimizer.
1742 nir_tex_src_projector
,
1743 nir_tex_src_comparator
, /* shadow comparator */
1747 nir_tex_src_min_lod
,
1748 nir_tex_src_ms_index
, /* MSAA sample index */
1749 nir_tex_src_ms_mcs
, /* MSAA compression value */
1752 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
1753 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
1754 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
1755 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
1756 nir_tex_src_texture_handle
, /* < bindless texture handle */
1757 nir_tex_src_sampler_handle
, /* < bindless sampler handle */
1758 nir_tex_src_plane
, /* < selects plane for planar textures */
1759 nir_num_tex_src_types
1764 nir_tex_src_type src_type
;
1768 nir_texop_tex
, /**< Regular texture look-up */
1769 nir_texop_txb
, /**< Texture look-up with LOD bias */
1770 nir_texop_txl
, /**< Texture look-up with explicit LOD */
1771 nir_texop_txd
, /**< Texture look-up with partial derivatives */
1772 nir_texop_txf
, /**< Texel fetch with explicit LOD */
1773 nir_texop_txf_ms
, /**< Multisample texture fetch */
1774 nir_texop_txf_ms_fb
, /**< Multisample texture fetch from framebuffer */
1775 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
1776 nir_texop_txs
, /**< Texture size */
1777 nir_texop_lod
, /**< Texture lod query */
1778 nir_texop_tg4
, /**< Texture gather */
1779 nir_texop_query_levels
, /**< Texture levels query */
1780 nir_texop_texture_samples
, /**< Texture samples query */
1781 nir_texop_samples_identical
, /**< Query whether all samples are definitely
1784 nir_texop_tex_prefetch
, /**< Regular texture look-up, eligible for pre-dispatch */
1790 enum glsl_sampler_dim sampler_dim
;
1791 nir_alu_type dest_type
;
1796 unsigned num_srcs
, coord_components
;
1797 bool is_array
, is_shadow
;
1800 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1801 * components or the new-style shadow that outputs 1 component.
1803 bool is_new_style_shadow
;
1805 /* gather component selector */
1806 unsigned component
: 2;
1808 /* gather offsets */
1809 int8_t tg4_offsets
[4][2];
1811 /* True if the texture index or handle is not dynamically uniform */
1812 bool texture_non_uniform
;
1814 /* True if the sampler index or handle is not dynamically uniform */
1815 bool sampler_non_uniform
;
1817 /** The texture index
1819 * If this texture instruction has a nir_tex_src_texture_offset source,
1820 * then the texture index is given by texture_index + texture_offset.
1822 unsigned texture_index
;
1824 /** The size of the texture array or 0 if it's not an array */
1825 unsigned texture_array_size
;
1827 /** The sampler index
1829 * The following operations do not require a sampler and, as such, this
1830 * field should be ignored:
1832 * - nir_texop_txf_ms
1835 * - nir_texop_query_levels
1836 * - nir_texop_texture_samples
1837 * - nir_texop_samples_identical
1839 * If this texture instruction has a nir_tex_src_sampler_offset source,
1840 * then the sampler index is given by sampler_index + sampler_offset.
1842 unsigned sampler_index
;
1845 static inline unsigned
1846 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
1848 switch (instr
->op
) {
1849 case nir_texop_txs
: {
1851 switch (instr
->sampler_dim
) {
1852 case GLSL_SAMPLER_DIM_1D
:
1853 case GLSL_SAMPLER_DIM_BUF
:
1856 case GLSL_SAMPLER_DIM_2D
:
1857 case GLSL_SAMPLER_DIM_CUBE
:
1858 case GLSL_SAMPLER_DIM_MS
:
1859 case GLSL_SAMPLER_DIM_RECT
:
1860 case GLSL_SAMPLER_DIM_EXTERNAL
:
1861 case GLSL_SAMPLER_DIM_SUBPASS
:
1864 case GLSL_SAMPLER_DIM_3D
:
1868 unreachable("not reached");
1870 if (instr
->is_array
)
1878 case nir_texop_texture_samples
:
1879 case nir_texop_query_levels
:
1880 case nir_texop_samples_identical
:
1884 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
1891 /* Returns true if this texture operation queries something about the texture
1892 * rather than actually sampling it.
1895 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
1897 switch (instr
->op
) {
1900 case nir_texop_texture_samples
:
1901 case nir_texop_query_levels
:
1902 case nir_texop_txf_ms_mcs
:
1909 case nir_texop_txf_ms
:
1910 case nir_texop_txf_ms_fb
:
1914 unreachable("Invalid texture opcode");
1919 nir_tex_instr_has_implicit_derivative(const nir_tex_instr
*instr
)
1921 switch (instr
->op
) {
1931 static inline nir_alu_type
1932 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
1934 switch (instr
->src
[src
].src_type
) {
1935 case nir_tex_src_coord
:
1936 switch (instr
->op
) {
1938 case nir_texop_txf_ms
:
1939 case nir_texop_txf_ms_fb
:
1940 case nir_texop_txf_ms_mcs
:
1941 case nir_texop_samples_identical
:
1942 return nir_type_int
;
1945 return nir_type_float
;
1948 case nir_tex_src_lod
:
1949 switch (instr
->op
) {
1952 return nir_type_int
;
1955 return nir_type_float
;
1958 case nir_tex_src_projector
:
1959 case nir_tex_src_comparator
:
1960 case nir_tex_src_bias
:
1961 case nir_tex_src_min_lod
:
1962 case nir_tex_src_ddx
:
1963 case nir_tex_src_ddy
:
1964 return nir_type_float
;
1966 case nir_tex_src_offset
:
1967 case nir_tex_src_ms_index
:
1968 case nir_tex_src_plane
:
1969 return nir_type_int
;
1971 case nir_tex_src_ms_mcs
:
1972 case nir_tex_src_texture_deref
:
1973 case nir_tex_src_sampler_deref
:
1974 case nir_tex_src_texture_offset
:
1975 case nir_tex_src_sampler_offset
:
1976 case nir_tex_src_texture_handle
:
1977 case nir_tex_src_sampler_handle
:
1978 return nir_type_uint
;
1980 case nir_num_tex_src_types
:
1981 unreachable("nir_num_tex_src_types is not a valid source type");
1984 unreachable("Invalid texture source type");
1987 static inline unsigned
1988 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
1990 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
1991 return instr
->coord_components
;
1993 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
1994 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
1997 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
1998 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
1999 if (instr
->is_array
)
2000 return instr
->coord_components
- 1;
2002 return instr
->coord_components
;
2005 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
2006 * the offset, since a cube maps to a single face.
2008 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
2009 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2011 else if (instr
->is_array
)
2012 return instr
->coord_components
- 1;
2014 return instr
->coord_components
;
2021 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
2023 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
2024 if (instr
->src
[i
].src_type
== type
)
2030 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
2031 nir_tex_src_type src_type
,
2034 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
2036 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
);
2043 nir_const_value value
[];
2044 } nir_load_const_instr
;
2057 /* creates a new SSA variable in an undefined state */
2062 } nir_ssa_undef_instr
;
2065 struct exec_node node
;
2067 /* The predecessor block corresponding to this source */
2068 struct nir_block
*pred
;
2073 #define nir_foreach_phi_src(phi_src, phi) \
2074 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
2075 #define nir_foreach_phi_src_safe(phi_src, phi) \
2076 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
2081 struct exec_list srcs
; /** < list of nir_phi_src */
2087 struct exec_node node
;
2090 } nir_parallel_copy_entry
;
2092 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
2093 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
2098 /* A list of nir_parallel_copy_entrys. The sources of all of the
2099 * entries are copied to the corresponding destinations "in parallel".
2100 * In other words, if we have two entries: a -> b and b -> a, the values
2103 struct exec_list entries
;
2104 } nir_parallel_copy_instr
;
2106 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
2107 type
, nir_instr_type_alu
)
2108 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
2109 type
, nir_instr_type_deref
)
2110 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
2111 type
, nir_instr_type_call
)
2112 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
2113 type
, nir_instr_type_jump
)
2114 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
2115 type
, nir_instr_type_tex
)
2116 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
2117 type
, nir_instr_type_intrinsic
)
2118 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
2119 type
, nir_instr_type_load_const
)
2120 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
2121 type
, nir_instr_type_ssa_undef
)
2122 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
2123 type
, nir_instr_type_phi
)
2124 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
2125 nir_parallel_copy_instr
, instr
,
2126 type
, nir_instr_type_parallel_copy
)
2129 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
2130 static inline type \
2131 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2133 assert(nir_src_is_const(src)); \
2134 nir_load_const_instr *load = \
2135 nir_instr_as_load_const(src.ssa->parent_instr); \
2136 assert(comp < load->def.num_components); \
2137 return nir_const_value_as_##suffix(load->value[comp], \
2138 load->def.bit_size); \
2141 static inline type \
2142 nir_src_as_##suffix(nir_src src) \
2144 assert(nir_src_num_components(src) == 1); \
2145 return nir_src_comp_as_##suffix(src, 0); \
2148 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2149 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint
)
2150 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2151 NIR_DEFINE_SRC_AS_CONST(double, float)
2153 #undef NIR_DEFINE_SRC_AS_CONST
2162 nir_ssa_scalar_is_const(nir_ssa_scalar s
)
2164 return s
.def
->parent_instr
->type
== nir_instr_type_load_const
;
2167 static inline nir_const_value
2168 nir_ssa_scalar_as_const_value(nir_ssa_scalar s
)
2170 assert(s
.comp
< s
.def
->num_components
);
2171 nir_load_const_instr
*load
= nir_instr_as_load_const(s
.def
->parent_instr
);
2172 return load
->value
[s
.comp
];
2175 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2176 static inline type \
2177 nir_ssa_scalar_as_##suffix(nir_ssa_scalar s) \
2179 return nir_const_value_as_##suffix( \
2180 nir_ssa_scalar_as_const_value(s), s.def->bit_size); \
2183 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2184 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint
)
2185 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2186 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2188 #undef NIR_DEFINE_SCALAR_AS_CONST
2191 nir_ssa_scalar_is_alu(nir_ssa_scalar s
)
2193 return s
.def
->parent_instr
->type
== nir_instr_type_alu
;
2196 static inline nir_op
2197 nir_ssa_scalar_alu_op(nir_ssa_scalar s
)
2199 return nir_instr_as_alu(s
.def
->parent_instr
)->op
;
2202 static inline nir_ssa_scalar
2203 nir_ssa_scalar_chase_alu_src(nir_ssa_scalar s
, unsigned alu_src_idx
)
2205 nir_ssa_scalar out
= { NULL
, 0 };
2207 nir_alu_instr
*alu
= nir_instr_as_alu(s
.def
->parent_instr
);
2208 assert(alu_src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
2210 /* Our component must be written */
2211 assert(s
.comp
< s
.def
->num_components
);
2212 assert(alu
->dest
.write_mask
& (1u << s
.comp
));
2214 assert(alu
->src
[alu_src_idx
].src
.is_ssa
);
2215 out
.def
= alu
->src
[alu_src_idx
].src
.ssa
;
2217 if (nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 0) {
2218 /* The ALU src is unsized so the source component follows the
2219 * destination component.
2221 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[s
.comp
];
2223 /* This is a sized source so all source components work together to
2224 * produce all the destination components. Since we need to return a
2225 * scalar, this only works if the source is a scalar.
2227 assert(nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 1);
2228 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[0];
2230 assert(out
.comp
< out
.def
->num_components
);
2239 * Control flow consists of a tree of control flow nodes, which include
2240 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2241 * instructions that always run start-to-finish. Each basic block also keeps
2242 * track of its successors (blocks which may run immediately after the current
2243 * block) and predecessors (blocks which could have run immediately before the
2244 * current block). Each function also has a start block and an end block which
2245 * all return statements point to (which is always empty). Together, all the
2246 * blocks with their predecessors and successors make up the control flow
2247 * graph (CFG) of the function. There are helpers that modify the tree of
2248 * control flow nodes while modifying the CFG appropriately; these should be
2249 * used instead of modifying the tree directly.
2256 nir_cf_node_function
2259 typedef struct nir_cf_node
{
2260 struct exec_node node
;
2261 nir_cf_node_type type
;
2262 struct nir_cf_node
*parent
;
2265 typedef struct nir_block
{
2266 nir_cf_node cf_node
;
2268 struct exec_list instr_list
; /** < list of nir_instr */
2270 /** generic block index; generated by nir_index_blocks */
2274 * Each block can only have up to 2 successors, so we put them in a simple
2275 * array - no need for anything more complicated.
2277 struct nir_block
*successors
[2];
2279 /* Set of nir_block predecessors in the CFG */
2280 struct set
*predecessors
;
2283 * this node's immediate dominator in the dominance tree - set to NULL for
2286 struct nir_block
*imm_dom
;
2288 /* This node's children in the dominance tree */
2289 unsigned num_dom_children
;
2290 struct nir_block
**dom_children
;
2292 /* Set of nir_blocks on the dominance frontier of this block */
2293 struct set
*dom_frontier
;
2296 * These two indices have the property that dom_{pre,post}_index for each
2297 * child of this block in the dominance tree will always be between
2298 * dom_pre_index and dom_post_index for this block, which makes testing if
2299 * a given block is dominated by another block an O(1) operation.
2301 unsigned dom_pre_index
, dom_post_index
;
2303 /* live in and out for this block; used for liveness analysis */
2304 BITSET_WORD
*live_in
;
2305 BITSET_WORD
*live_out
;
2308 static inline nir_instr
*
2309 nir_block_first_instr(nir_block
*block
)
2311 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
2312 return exec_node_data(nir_instr
, head
, node
);
2315 static inline nir_instr
*
2316 nir_block_last_instr(nir_block
*block
)
2318 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
2319 return exec_node_data(nir_instr
, tail
, node
);
2323 nir_block_ends_in_jump(nir_block
*block
)
2325 return !exec_list_is_empty(&block
->instr_list
) &&
2326 nir_block_last_instr(block
)->type
== nir_instr_type_jump
;
2329 #define nir_foreach_instr(instr, block) \
2330 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
2331 #define nir_foreach_instr_reverse(instr, block) \
2332 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
2333 #define nir_foreach_instr_safe(instr, block) \
2334 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
2335 #define nir_foreach_instr_reverse_safe(instr, block) \
2336 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
2339 nir_selection_control_none
= 0x0,
2340 nir_selection_control_flatten
= 0x1,
2341 nir_selection_control_dont_flatten
= 0x2,
2342 } nir_selection_control
;
2344 typedef struct nir_if
{
2345 nir_cf_node cf_node
;
2347 nir_selection_control control
;
2349 struct exec_list then_list
; /** < list of nir_cf_node */
2350 struct exec_list else_list
; /** < list of nir_cf_node */
2356 /** Instruction that generates nif::condition. */
2357 nir_instr
*conditional_instr
;
2359 /** Block within ::nif that has the break instruction. */
2360 nir_block
*break_block
;
2362 /** Last block for the then- or else-path that does not contain the break. */
2363 nir_block
*continue_from_block
;
2365 /** True when ::break_block is in the else-path of ::nif. */
2366 bool continue_from_then
;
2369 /* This is true if the terminators exact trip count is unknown. For
2372 * for (int i = 0; i < imin(x, 4); i++)
2375 * Here loop analysis would have set a max_trip_count of 4 however we dont
2376 * know for sure that this is the exact trip count.
2378 bool exact_trip_count_unknown
;
2380 struct list_head loop_terminator_link
;
2381 } nir_loop_terminator
;
2384 /* Estimated cost (in number of instructions) of the loop */
2385 unsigned instr_cost
;
2387 /* Guessed trip count based on array indexing */
2388 unsigned guessed_trip_count
;
2390 /* Maximum number of times the loop is run (if known) */
2391 unsigned max_trip_count
;
2393 /* Do we know the exact number of times the loop will be run */
2394 bool exact_trip_count_known
;
2396 /* Unroll the loop regardless of its size */
2399 /* Does the loop contain complex loop terminators, continues or other
2400 * complex behaviours? If this is true we can't rely on
2401 * loop_terminator_list to be complete or accurate.
2405 nir_loop_terminator
*limiting_terminator
;
2407 /* A list of loop_terminators terminating this loop. */
2408 struct list_head loop_terminator_list
;
2412 nir_loop_control_none
= 0x0,
2413 nir_loop_control_unroll
= 0x1,
2414 nir_loop_control_dont_unroll
= 0x2,
2418 nir_cf_node cf_node
;
2420 struct exec_list body
; /** < list of nir_cf_node */
2422 nir_loop_info
*info
;
2423 nir_loop_control control
;
2424 bool partially_unrolled
;
2428 * Various bits of metadata that can may be created or required by
2429 * optimization and analysis passes
2432 nir_metadata_none
= 0x0,
2433 nir_metadata_block_index
= 0x1,
2434 nir_metadata_dominance
= 0x2,
2435 nir_metadata_live_ssa_defs
= 0x4,
2436 nir_metadata_not_properly_reset
= 0x8,
2437 nir_metadata_loop_analysis
= 0x10,
2441 nir_cf_node cf_node
;
2443 /** pointer to the function of which this is an implementation */
2444 struct nir_function
*function
;
2446 struct exec_list body
; /** < list of nir_cf_node */
2448 nir_block
*end_block
;
2450 /** list for all local variables in the function */
2451 struct exec_list locals
;
2453 /** list of local registers in the function */
2454 struct exec_list registers
;
2456 /** next available local register index */
2459 /** next available SSA value index */
2462 /* total number of basic blocks, only valid when block_index_dirty = false */
2463 unsigned num_blocks
;
2465 nir_metadata valid_metadata
;
2466 } nir_function_impl
;
2468 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2469 nir_start_block(nir_function_impl
*impl
)
2471 return (nir_block
*) impl
->body
.head_sentinel
.next
;
2474 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2475 nir_impl_last_block(nir_function_impl
*impl
)
2477 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
2480 static inline nir_cf_node
*
2481 nir_cf_node_next(nir_cf_node
*node
)
2483 struct exec_node
*next
= exec_node_get_next(&node
->node
);
2484 if (exec_node_is_tail_sentinel(next
))
2487 return exec_node_data(nir_cf_node
, next
, node
);
2490 static inline nir_cf_node
*
2491 nir_cf_node_prev(nir_cf_node
*node
)
2493 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
2494 if (exec_node_is_head_sentinel(prev
))
2497 return exec_node_data(nir_cf_node
, prev
, node
);
2501 nir_cf_node_is_first(const nir_cf_node
*node
)
2503 return exec_node_is_head_sentinel(node
->node
.prev
);
2507 nir_cf_node_is_last(const nir_cf_node
*node
)
2509 return exec_node_is_tail_sentinel(node
->node
.next
);
2512 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
2513 type
, nir_cf_node_block
)
2514 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
2515 type
, nir_cf_node_if
)
2516 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
2517 type
, nir_cf_node_loop
)
2518 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
2519 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
2521 static inline nir_block
*
2522 nir_if_first_then_block(nir_if
*if_stmt
)
2524 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
2525 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2528 static inline nir_block
*
2529 nir_if_last_then_block(nir_if
*if_stmt
)
2531 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
2532 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2535 static inline nir_block
*
2536 nir_if_first_else_block(nir_if
*if_stmt
)
2538 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
2539 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2542 static inline nir_block
*
2543 nir_if_last_else_block(nir_if
*if_stmt
)
2545 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
2546 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2549 static inline nir_block
*
2550 nir_loop_first_block(nir_loop
*loop
)
2552 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
2553 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2556 static inline nir_block
*
2557 nir_loop_last_block(nir_loop
*loop
)
2559 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
2560 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2564 * Return true if this list of cf_nodes contains a single empty block.
2567 nir_cf_list_is_empty_block(struct exec_list
*cf_list
)
2569 if (exec_list_is_singular(cf_list
)) {
2570 struct exec_node
*head
= exec_list_get_head(cf_list
);
2572 nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2573 return exec_list_is_empty(&block
->instr_list
);
2579 uint8_t num_components
;
2583 typedef struct nir_function
{
2584 struct exec_node node
;
2587 struct nir_shader
*shader
;
2589 unsigned num_params
;
2590 nir_parameter
*params
;
2592 /** The implementation of this function.
2594 * If the function is only declared and not implemented, this is NULL.
2596 nir_function_impl
*impl
;
2602 nir_lower_imul64
= (1 << 0),
2603 nir_lower_isign64
= (1 << 1),
2604 /** Lower all int64 modulus and division opcodes */
2605 nir_lower_divmod64
= (1 << 2),
2606 /** Lower all 64-bit umul_high and imul_high opcodes */
2607 nir_lower_imul_high64
= (1 << 3),
2608 nir_lower_mov64
= (1 << 4),
2609 nir_lower_icmp64
= (1 << 5),
2610 nir_lower_iadd64
= (1 << 6),
2611 nir_lower_iabs64
= (1 << 7),
2612 nir_lower_ineg64
= (1 << 8),
2613 nir_lower_logic64
= (1 << 9),
2614 nir_lower_minmax64
= (1 << 10),
2615 nir_lower_shift64
= (1 << 11),
2616 nir_lower_imul_2x32_64
= (1 << 12),
2617 nir_lower_extract64
= (1 << 13),
2618 } nir_lower_int64_options
;
2621 nir_lower_drcp
= (1 << 0),
2622 nir_lower_dsqrt
= (1 << 1),
2623 nir_lower_drsq
= (1 << 2),
2624 nir_lower_dtrunc
= (1 << 3),
2625 nir_lower_dfloor
= (1 << 4),
2626 nir_lower_dceil
= (1 << 5),
2627 nir_lower_dfract
= (1 << 6),
2628 nir_lower_dround_even
= (1 << 7),
2629 nir_lower_dmod
= (1 << 8),
2630 nir_lower_dsub
= (1 << 9),
2631 nir_lower_ddiv
= (1 << 10),
2632 nir_lower_fp64_full_software
= (1 << 11),
2633 } nir_lower_doubles_options
;
2636 nir_divergence_single_prim_per_subgroup
= (1 << 0),
2637 nir_divergence_single_patch_per_tcs_subgroup
= (1 << 1),
2638 nir_divergence_single_patch_per_tes_subgroup
= (1 << 2),
2639 nir_divergence_view_index_uniform
= (1 << 3),
2640 } nir_divergence_options
;
2642 typedef struct nir_shader_compiler_options
{
2648 /** Lowers flrp when it does not support doubles */
2655 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2656 bool lower_bitfield_extract
;
2657 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
2658 bool lower_bitfield_extract_to_shifts
;
2659 /** Lowers bitfield_insert to bfi/bfm */
2660 bool lower_bitfield_insert
;
2661 /** Lowers bitfield_insert to compares, and shifts. */
2662 bool lower_bitfield_insert_to_shifts
;
2663 /** Lowers bitfield_insert to bfm/bitfield_select. */
2664 bool lower_bitfield_insert_to_bitfield_select
;
2665 /** Lowers bitfield_reverse to shifts. */
2666 bool lower_bitfield_reverse
;
2667 /** Lowers bit_count to shifts. */
2668 bool lower_bit_count
;
2669 /** Lowers ifind_msb to compare and ufind_msb */
2670 bool lower_ifind_msb
;
2671 /** Lowers find_lsb to ufind_msb and logic ops */
2672 bool lower_find_lsb
;
2673 bool lower_uadd_carry
;
2674 bool lower_usub_borrow
;
2675 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2676 bool lower_mul_high
;
2677 /** lowers fneg and ineg to fsub and isub. */
2679 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2682 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2685 /* lower fall_equalN/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
2686 bool lower_vector_cmp
;
2688 /** enables rules to lower idiv by power-of-two: */
2691 /** enable rules to avoid bit ops */
2694 /** enables rules to lower isign to imin+imax */
2697 /** enables rules to lower fsign to fsub and flt */
2700 /* lower fdph to fdot4 */
2703 /** lower fdot to fmul and fsum/fadd. */
2706 /* Does the native fdot instruction replicate its result for four
2707 * components? If so, then opt_algebraic_late will turn all fdotN
2708 * instructions into fdot_replicatedN instructions.
2710 bool fdot_replicates
;
2712 /** lowers ffloor to fsub+ffract: */
2715 /** lowers ffract to fsub+ffloor: */
2718 /** lowers fceil to fneg+ffloor+fneg: */
2725 bool lower_pack_half_2x16
;
2726 bool lower_pack_unorm_2x16
;
2727 bool lower_pack_snorm_2x16
;
2728 bool lower_pack_unorm_4x8
;
2729 bool lower_pack_snorm_4x8
;
2730 bool lower_unpack_half_2x16
;
2731 bool lower_unpack_unorm_2x16
;
2732 bool lower_unpack_snorm_2x16
;
2733 bool lower_unpack_unorm_4x8
;
2734 bool lower_unpack_snorm_4x8
;
2736 bool lower_extract_byte
;
2737 bool lower_extract_word
;
2739 bool lower_all_io_to_temps
;
2740 bool lower_all_io_to_elements
;
2742 /* Indicates that the driver only has zero-based vertex id */
2743 bool vertex_id_zero_based
;
2746 * If enabled, gl_BaseVertex will be lowered as:
2747 * is_indexed_draw (~0/0) & firstvertex
2749 bool lower_base_vertex
;
2752 * If enabled, gl_HelperInvocation will be lowered as:
2754 * !((1 << sample_id) & sample_mask_in))
2756 * This depends on some possibly hw implementation details, which may
2757 * not be true for all hw. In particular that the FS is only executed
2758 * for covered samples or for helper invocations. So, do not blindly
2759 * enable this option.
2761 * Note: See also issue #22 in ARB_shader_image_load_store
2763 bool lower_helper_invocation
;
2766 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
2768 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
2769 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
2771 bool optimize_sample_mask_in
;
2773 bool lower_cs_local_index_from_id
;
2774 bool lower_cs_local_id_from_index
;
2776 bool lower_device_index_to_zero
;
2778 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
2779 bool lower_wpos_pntc
;
2785 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
2786 * for IO purposes and would prefer loads/stores be vectorized.
2789 bool lower_to_scalar
;
2792 * Should nir_lower_io() create load_interpolated_input intrinsics?
2794 * If not, it generates regular load_input intrinsics and interpolation
2795 * information must be inferred from the list of input nir_variables.
2797 bool use_interpolated_input_intrinsics
;
2799 /* Lowers when 32x32->64 bit multiplication is not supported */
2800 bool lower_mul_2x32_64
;
2802 /* Lowers when rotate instruction is not supported */
2806 * Backend supports imul24, and would like to use it (when possible)
2807 * for address/offset calculation. If true, driver should call
2808 * nir_lower_amul(). (If not set, amul will automatically be lowered
2814 * Is this the Intel vec4 backend?
2816 * Used to inhibit algebraic optimizations that are known to be harmful on
2817 * the Intel vec4 backend. This is generally applicable to any
2818 * optimization that might cause more immediate values to be used in
2819 * 3-source (e.g., ffma and flrp) instructions.
2823 unsigned max_unroll_iterations
;
2825 nir_lower_int64_options lower_int64_options
;
2826 nir_lower_doubles_options lower_doubles_options
;
2827 } nir_shader_compiler_options
;
2829 typedef struct nir_shader
{
2830 /** list of uniforms (nir_variable) */
2831 struct exec_list uniforms
;
2833 /** list of inputs (nir_variable) */
2834 struct exec_list inputs
;
2836 /** list of outputs (nir_variable) */
2837 struct exec_list outputs
;
2839 /** list of shared compute variables (nir_variable) */
2840 struct exec_list shared
;
2842 /** Set of driver-specific options for the shader.
2844 * The memory for the options is expected to be kept in a single static
2845 * copy by the driver.
2847 const struct nir_shader_compiler_options
*options
;
2849 /** Various bits of compile-time information about a given shader */
2850 struct shader_info info
;
2852 /** list of global variables in the shader (nir_variable) */
2853 struct exec_list globals
;
2855 /** list of system value variables in the shader (nir_variable) */
2856 struct exec_list system_values
;
2858 struct exec_list functions
; /** < list of nir_function */
2861 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
2864 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
2866 /** Size in bytes of required scratch space */
2867 unsigned scratch_size
;
2869 /** Constant data associated with this shader.
2871 * Constant data is loaded through load_constant intrinsics. See also
2872 * nir_opt_large_constants.
2874 void *constant_data
;
2875 unsigned constant_data_size
;
2878 #define nir_foreach_function(func, shader) \
2879 foreach_list_typed(nir_function, func, node, &(shader)->functions)
2881 static inline nir_function_impl
*
2882 nir_shader_get_entrypoint(nir_shader
*shader
)
2884 nir_function
*func
= NULL
;
2886 nir_foreach_function(function
, shader
) {
2887 assert(func
== NULL
);
2888 if (function
->is_entrypoint
) {
2899 assert(func
->num_params
== 0);
2904 nir_shader
*nir_shader_create(void *mem_ctx
,
2905 gl_shader_stage stage
,
2906 const nir_shader_compiler_options
*options
,
2909 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
2911 void nir_reg_remove(nir_register
*reg
);
2913 /** Adds a variable to the appropriate list in nir_shader */
2914 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
2917 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
2919 assert(var
->data
.mode
== nir_var_function_temp
);
2920 exec_list_push_tail(&impl
->locals
, &var
->node
);
2923 /** creates a variable, sets a few defaults, and adds it to the list */
2924 nir_variable
*nir_variable_create(nir_shader
*shader
,
2925 nir_variable_mode mode
,
2926 const struct glsl_type
*type
,
2928 /** creates a local variable and adds it to the list */
2929 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
2930 const struct glsl_type
*type
,
2933 /** creates a function and adds it to the shader's list of functions */
2934 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
2936 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
2937 /** creates a function_impl that isn't tied to any particular function */
2938 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
2940 nir_block
*nir_block_create(nir_shader
*shader
);
2941 nir_if
*nir_if_create(nir_shader
*shader
);
2942 nir_loop
*nir_loop_create(nir_shader
*shader
);
2944 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
2946 /** requests that the given pieces of metadata be generated */
2947 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
2948 /** dirties all but the preserved metadata */
2949 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
2951 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
2952 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
2954 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
2955 nir_deref_type deref_type
);
2957 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
2959 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
2960 unsigned num_components
,
2963 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
2964 nir_intrinsic_op op
);
2966 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
2967 nir_function
*callee
);
2969 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
2971 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
2973 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
2975 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
2976 unsigned num_components
,
2979 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
2982 * NIR Cursors and Instruction Insertion API
2985 * A tiny struct representing a point to insert/extract instructions or
2986 * control flow nodes. Helps reduce the combinatorial explosion of possible
2987 * points to insert/extract.
2989 * \sa nir_control_flow.h
2992 nir_cursor_before_block
,
2993 nir_cursor_after_block
,
2994 nir_cursor_before_instr
,
2995 nir_cursor_after_instr
,
2996 } nir_cursor_option
;
2999 nir_cursor_option option
;
3006 static inline nir_block
*
3007 nir_cursor_current_block(nir_cursor cursor
)
3009 if (cursor
.option
== nir_cursor_before_instr
||
3010 cursor
.option
== nir_cursor_after_instr
) {
3011 return cursor
.instr
->block
;
3013 return cursor
.block
;
3017 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
3019 static inline nir_cursor
3020 nir_before_block(nir_block
*block
)
3023 cursor
.option
= nir_cursor_before_block
;
3024 cursor
.block
= block
;
3028 static inline nir_cursor
3029 nir_after_block(nir_block
*block
)
3032 cursor
.option
= nir_cursor_after_block
;
3033 cursor
.block
= block
;
3037 static inline nir_cursor
3038 nir_before_instr(nir_instr
*instr
)
3041 cursor
.option
= nir_cursor_before_instr
;
3042 cursor
.instr
= instr
;
3046 static inline nir_cursor
3047 nir_after_instr(nir_instr
*instr
)
3050 cursor
.option
= nir_cursor_after_instr
;
3051 cursor
.instr
= instr
;
3055 static inline nir_cursor
3056 nir_after_block_before_jump(nir_block
*block
)
3058 nir_instr
*last_instr
= nir_block_last_instr(block
);
3059 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
3060 return nir_before_instr(last_instr
);
3062 return nir_after_block(block
);
3066 static inline nir_cursor
3067 nir_before_src(nir_src
*src
, bool is_if_condition
)
3069 if (is_if_condition
) {
3070 nir_block
*prev_block
=
3071 nir_cf_node_as_block(nir_cf_node_prev(&src
->parent_if
->cf_node
));
3072 assert(!nir_block_ends_in_jump(prev_block
));
3073 return nir_after_block(prev_block
);
3074 } else if (src
->parent_instr
->type
== nir_instr_type_phi
) {
3076 nir_phi_instr
*cond_phi
= nir_instr_as_phi(src
->parent_instr
);
3078 nir_foreach_phi_src(phi_src
, cond_phi
) {
3079 if (phi_src
->src
.ssa
== src
->ssa
) {
3086 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
3087 * to have a more specific name.
3089 nir_phi_src
*phi_src
= LIST_ENTRY(nir_phi_src
, src
, src
);
3090 return nir_after_block_before_jump(phi_src
->pred
);
3092 return nir_before_instr(src
->parent_instr
);
3096 static inline nir_cursor
3097 nir_before_cf_node(nir_cf_node
*node
)
3099 if (node
->type
== nir_cf_node_block
)
3100 return nir_before_block(nir_cf_node_as_block(node
));
3102 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
3105 static inline nir_cursor
3106 nir_after_cf_node(nir_cf_node
*node
)
3108 if (node
->type
== nir_cf_node_block
)
3109 return nir_after_block(nir_cf_node_as_block(node
));
3111 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
3114 static inline nir_cursor
3115 nir_after_phis(nir_block
*block
)
3117 nir_foreach_instr(instr
, block
) {
3118 if (instr
->type
!= nir_instr_type_phi
)
3119 return nir_before_instr(instr
);
3121 return nir_after_block(block
);
3124 static inline nir_cursor
3125 nir_after_cf_node_and_phis(nir_cf_node
*node
)
3127 if (node
->type
== nir_cf_node_block
)
3128 return nir_after_block(nir_cf_node_as_block(node
));
3130 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
3132 return nir_after_phis(block
);
3135 static inline nir_cursor
3136 nir_before_cf_list(struct exec_list
*cf_list
)
3138 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
3139 exec_list_get_head(cf_list
), node
);
3140 return nir_before_cf_node(first_node
);
3143 static inline nir_cursor
3144 nir_after_cf_list(struct exec_list
*cf_list
)
3146 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
3147 exec_list_get_tail(cf_list
), node
);
3148 return nir_after_cf_node(last_node
);
3152 * Insert a NIR instruction at the given cursor.
3154 * Note: This does not update the cursor.
3156 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
3159 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
3161 nir_instr_insert(nir_before_instr(instr
), before
);
3165 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
3167 nir_instr_insert(nir_after_instr(instr
), after
);
3171 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
3173 nir_instr_insert(nir_before_block(block
), before
);
3177 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
3179 nir_instr_insert(nir_after_block(block
), after
);
3183 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
3185 nir_instr_insert(nir_before_cf_node(node
), before
);
3189 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
3191 nir_instr_insert(nir_after_cf_node(node
), after
);
3195 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
3197 nir_instr_insert(nir_before_cf_list(list
), before
);
3201 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
3203 nir_instr_insert(nir_after_cf_list(list
), after
);
3206 void nir_instr_remove_v(nir_instr
*instr
);
3208 static inline nir_cursor
3209 nir_instr_remove(nir_instr
*instr
)
3212 nir_instr
*prev
= nir_instr_prev(instr
);
3214 cursor
= nir_after_instr(prev
);
3216 cursor
= nir_before_block(instr
->block
);
3218 nir_instr_remove_v(instr
);
3224 nir_ssa_def
*nir_instr_ssa_def(nir_instr
*instr
);
3226 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
3227 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
3228 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
3229 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
3231 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
3232 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
3234 nir_const_value
*nir_src_as_const_value(nir_src src
);
3236 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
3237 static inline c_type * \
3238 nir_src_as_ ## name (nir_src src) \
3240 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
3241 ? cast_macro(src.ssa->parent_instr) : NULL; \
3244 NIR_SRC_AS_(alu_instr
, nir_alu_instr
, nir_instr_type_alu
, nir_instr_as_alu
)
3245 NIR_SRC_AS_(intrinsic
, nir_intrinsic_instr
,
3246 nir_instr_type_intrinsic
, nir_instr_as_intrinsic
)
3247 NIR_SRC_AS_(deref
, nir_deref_instr
, nir_instr_type_deref
, nir_instr_as_deref
)
3249 bool nir_src_is_dynamically_uniform(nir_src src
);
3250 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
3251 bool nir_instrs_equal(const nir_instr
*instr1
, const nir_instr
*instr2
);
3252 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
3253 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
3254 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
3255 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
3258 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
3259 unsigned num_components
, unsigned bit_size
,
3261 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
3262 unsigned num_components
, unsigned bit_size
,
3265 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
3266 const struct glsl_type
*type
,
3269 assert(glsl_type_is_vector_or_scalar(type
));
3270 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
3271 glsl_get_bit_size(type
), name
);
3273 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
3274 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
3275 nir_instr
*after_me
);
3277 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
3280 * finds the next basic block in source-code order, returns NULL if there is
3284 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
3286 /* Performs the opposite of nir_block_cf_tree_next() */
3288 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
3290 /* Gets the first block in a CF node in source-code order */
3292 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
3294 /* Gets the last block in a CF node in source-code order */
3296 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
3298 /* Gets the next block after a CF node in source-code order */
3300 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
3302 /* Macros for loops that visit blocks in source-code order */
3304 #define nir_foreach_block(block, impl) \
3305 for (nir_block *block = nir_start_block(impl); block != NULL; \
3306 block = nir_block_cf_tree_next(block))
3308 #define nir_foreach_block_safe(block, impl) \
3309 for (nir_block *block = nir_start_block(impl), \
3310 *next = nir_block_cf_tree_next(block); \
3312 block = next, next = nir_block_cf_tree_next(block))
3314 #define nir_foreach_block_reverse(block, impl) \
3315 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
3316 block = nir_block_cf_tree_prev(block))
3318 #define nir_foreach_block_reverse_safe(block, impl) \
3319 for (nir_block *block = nir_impl_last_block(impl), \
3320 *prev = nir_block_cf_tree_prev(block); \
3322 block = prev, prev = nir_block_cf_tree_prev(block))
3324 #define nir_foreach_block_in_cf_node(block, node) \
3325 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
3326 block != nir_cf_node_cf_tree_next(node); \
3327 block = nir_block_cf_tree_next(block))
3329 /* If the following CF node is an if, this function returns that if.
3330 * Otherwise, it returns NULL.
3332 nir_if
*nir_block_get_following_if(nir_block
*block
);
3334 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
3336 void nir_index_local_regs(nir_function_impl
*impl
);
3337 void nir_index_ssa_defs(nir_function_impl
*impl
);
3338 unsigned nir_index_instrs(nir_function_impl
*impl
);
3340 void nir_index_blocks(nir_function_impl
*impl
);
3342 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
3343 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
3344 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
3345 void nir_print_deref(const nir_deref_instr
*deref
, FILE *fp
);
3347 /** Shallow clone of a single ALU instruction. */
3348 nir_alu_instr
*nir_alu_instr_clone(nir_shader
*s
, const nir_alu_instr
*orig
);
3350 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
3351 nir_function_impl
*nir_function_impl_clone(nir_shader
*shader
,
3352 const nir_function_impl
*fi
);
3353 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
3354 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
3356 void nir_shader_replace(nir_shader
*dest
, nir_shader
*src
);
3358 void nir_shader_serialize_deserialize(nir_shader
*s
);
3361 void nir_validate_shader(nir_shader
*shader
, const char *when
);
3362 void nir_metadata_set_validation_flag(nir_shader
*shader
);
3363 void nir_metadata_check_validation_flag(nir_shader
*shader
);
3366 should_skip_nir(const char *name
)
3368 static const char *list
= NULL
;
3370 /* Comma separated list of names to skip. */
3371 list
= getenv("NIR_SKIP");
3379 return comma_separated_list_contains(list
, name
);
3383 should_clone_nir(void)
3385 static int should_clone
= -1;
3386 if (should_clone
< 0)
3387 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
3389 return should_clone
;
3393 should_serialize_deserialize_nir(void)
3395 static int test_serialize
= -1;
3396 if (test_serialize
< 0)
3397 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
3399 return test_serialize
;
3403 should_print_nir(void)
3405 static int should_print
= -1;
3406 if (should_print
< 0)
3407 should_print
= env_var_as_boolean("NIR_PRINT", false);
3409 return should_print
;
3412 static inline void nir_validate_shader(nir_shader
*shader
, const char *when
) { (void) shader
; (void)when
; }
3413 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
3414 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
3415 static inline bool should_skip_nir(UNUSED
const char *pass_name
) { return false; }
3416 static inline bool should_clone_nir(void) { return false; }
3417 static inline bool should_serialize_deserialize_nir(void) { return false; }
3418 static inline bool should_print_nir(void) { return false; }
3421 #define _PASS(pass, nir, do_pass) do { \
3422 if (should_skip_nir(#pass)) { \
3423 printf("skipping %s\n", #pass); \
3427 nir_validate_shader(nir, "after " #pass); \
3428 if (should_clone_nir()) { \
3429 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3430 nir_shader_replace(nir, clone); \
3432 if (should_serialize_deserialize_nir()) { \
3433 nir_shader_serialize_deserialize(nir); \
3437 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3438 nir_metadata_set_validation_flag(nir); \
3439 if (should_print_nir()) \
3440 printf("%s\n", #pass); \
3441 if (pass(nir, ##__VA_ARGS__)) { \
3443 if (should_print_nir()) \
3444 nir_print_shader(nir, stdout); \
3445 nir_metadata_check_validation_flag(nir); \
3449 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3450 if (should_print_nir()) \
3451 printf("%s\n", #pass); \
3452 pass(nir, ##__VA_ARGS__); \
3453 if (should_print_nir()) \
3454 nir_print_shader(nir, stdout); \
3457 #define NIR_SKIP(name) should_skip_nir(#name)
3459 /** An instruction filtering callback
3461 * Returns true if the instruction should be processed and false otherwise.
3463 typedef bool (*nir_instr_filter_cb
)(const nir_instr
*, const void *);
3465 /** A simple instruction lowering callback
3467 * Many instruction lowering passes can be written as a simple function which
3468 * takes an instruction as its input and returns a sequence of instructions
3469 * that implement the consumed instruction. This function type represents
3470 * such a lowering function. When called, a function with this prototype
3471 * should either return NULL indicating that no lowering needs to be done or
3472 * emit a sequence of instructions using the provided builder (whose cursor
3473 * will already be placed after the instruction to be lowered) and return the
3474 * resulting nir_ssa_def.
3476 typedef nir_ssa_def
*(*nir_lower_instr_cb
)(struct nir_builder
*,
3477 nir_instr
*, void *);
3480 * Special return value for nir_lower_instr_cb when some progress occurred
3481 * (like changing an input to the instr) that didn't result in a replacement
3482 * SSA def being generated.
3484 #define NIR_LOWER_INSTR_PROGRESS ((nir_ssa_def *)(uintptr_t)1)
3486 /** Iterate over all the instructions in a nir_function_impl and lower them
3487 * using the provided callbacks
3489 * This function implements the guts of a standard lowering pass for you. It
3490 * iterates over all of the instructions in a nir_function_impl and calls the
3491 * filter callback on each one. If the filter callback returns true, it then
3492 * calls the lowering call back on the instruction. (Splitting it this way
3493 * allows us to avoid some save/restore work for instructions we know won't be
3494 * lowered.) If the instruction is dead after the lowering is complete, it
3495 * will be removed. If new instructions are added, the lowering callback will
3496 * also be called on them in case multiple lowerings are required.
3498 * The metadata for the nir_function_impl will also be updated. If any blocks
3499 * are added (they cannot be removed), dominance and block indices will be
3502 bool nir_function_impl_lower_instructions(nir_function_impl
*impl
,
3503 nir_instr_filter_cb filter
,
3504 nir_lower_instr_cb lower
,
3506 bool nir_shader_lower_instructions(nir_shader
*shader
,
3507 nir_instr_filter_cb filter
,
3508 nir_lower_instr_cb lower
,
3511 void nir_calc_dominance_impl(nir_function_impl
*impl
);
3512 void nir_calc_dominance(nir_shader
*shader
);
3514 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
3515 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
3516 bool nir_block_is_unreachable(nir_block
*block
);
3518 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
3519 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
3521 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
3522 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
3524 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
3525 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
3527 int nir_gs_count_vertices(const nir_shader
*shader
);
3529 bool nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3530 bool nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3531 bool nir_split_var_copies(nir_shader
*shader
);
3532 bool nir_split_per_member_structs(nir_shader
*shader
);
3533 bool nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
);
3535 bool nir_lower_returns_impl(nir_function_impl
*impl
);
3536 bool nir_lower_returns(nir_shader
*shader
);
3538 void nir_inline_function_impl(struct nir_builder
*b
,
3539 const nir_function_impl
*impl
,
3540 nir_ssa_def
**params
);
3541 bool nir_inline_functions(nir_shader
*shader
);
3543 bool nir_propagate_invariant(nir_shader
*shader
);
3545 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
3546 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
3547 nir_intrinsic_instr
*copy
);
3548 bool nir_lower_var_copies(nir_shader
*shader
);
3550 void nir_fixup_deref_modes(nir_shader
*shader
);
3552 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
3555 nir_lower_direct_array_deref_of_vec_load
= (1 << 0),
3556 nir_lower_indirect_array_deref_of_vec_load
= (1 << 1),
3557 nir_lower_direct_array_deref_of_vec_store
= (1 << 2),
3558 nir_lower_indirect_array_deref_of_vec_store
= (1 << 3),
3559 } nir_lower_array_deref_of_vec_options
;
3561 bool nir_lower_array_deref_of_vec(nir_shader
*shader
, nir_variable_mode modes
,
3562 nir_lower_array_deref_of_vec_options options
);
3564 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
3566 bool nir_lower_locals_to_regs(nir_shader
*shader
);
3568 void nir_lower_io_to_temporaries(nir_shader
*shader
,
3569 nir_function_impl
*entrypoint
,
3570 bool outputs
, bool inputs
);
3572 bool nir_lower_vars_to_scratch(nir_shader
*shader
,
3573 nir_variable_mode modes
,
3575 glsl_type_size_align_func size_align
);
3577 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
3579 void nir_gather_ssa_types(nir_function_impl
*impl
,
3580 BITSET_WORD
*float_types
,
3581 BITSET_WORD
*int_types
);
3583 void nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
3584 int (*type_size
)(const struct glsl_type
*, bool));
3586 /* Some helpers to do very simple linking */
3587 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3588 bool nir_remove_unused_io_vars(nir_shader
*shader
, struct exec_list
*var_list
,
3589 uint64_t *used_by_other_stage
,
3590 uint64_t *used_by_other_stage_patches
);
3591 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
3592 bool default_to_smooth_interp
);
3593 void nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3594 bool nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3596 bool nir_lower_amul(nir_shader
*shader
,
3597 int (*type_size
)(const struct glsl_type
*, bool));
3599 void nir_assign_io_var_locations(struct exec_list
*var_list
,
3601 gl_shader_stage stage
);
3604 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
3605 * to 32-bit operations. This is only valid for nir_var_shader_in/out
3608 nir_lower_io_lower_64bit_to_32
= (1 << 0),
3610 /* If set, this forces all non-flat fragment shader inputs to be
3611 * interpolated as if with the "sample" qualifier. This requires
3612 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
3614 nir_lower_io_force_sample_interpolation
= (1 << 1),
3615 } nir_lower_io_options
;
3616 bool nir_lower_io(nir_shader
*shader
,
3617 nir_variable_mode modes
,
3618 int (*type_size
)(const struct glsl_type
*, bool),
3619 nir_lower_io_options
);
3621 bool nir_io_add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
);
3624 nir_lower_vars_to_explicit_types(nir_shader
*shader
,
3625 nir_variable_mode modes
,
3626 glsl_type_size_align_func type_info
);
3630 * An address format which is a simple 32-bit global GPU address.
3632 nir_address_format_32bit_global
,
3635 * An address format which is a simple 64-bit global GPU address.
3637 nir_address_format_64bit_global
,
3640 * An address format which is a bounds-checked 64-bit global GPU address.
3642 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
3643 * address stored with the low bits in .x and high bits in .y, .z is a
3644 * size, and .w is an offset. When the final I/O operation is lowered, .w
3645 * is checked against .z and the operation is predicated on the result.
3647 nir_address_format_64bit_bounded_global
,
3650 * An address format which is comprised of a vec2 where the first
3651 * component is a buffer index and the second is an offset.
3653 nir_address_format_32bit_index_offset
,
3656 * An address format which is a simple 32-bit offset.
3658 nir_address_format_32bit_offset
,
3661 * An address format representing a purely logical addressing model. In
3662 * this model, all deref chains must be complete from the dereference
3663 * operation to the variable. Cast derefs are not allowed. These
3664 * addresses will be 32-bit scalars but the format is immaterial because
3665 * you can always chase the chain.
3667 nir_address_format_logical
,
3668 } nir_address_format
;
3670 static inline unsigned
3671 nir_address_format_bit_size(nir_address_format addr_format
)
3673 switch (addr_format
) {
3674 case nir_address_format_32bit_global
: return 32;
3675 case nir_address_format_64bit_global
: return 64;
3676 case nir_address_format_64bit_bounded_global
: return 32;
3677 case nir_address_format_32bit_index_offset
: return 32;
3678 case nir_address_format_32bit_offset
: return 32;
3679 case nir_address_format_logical
: return 32;
3681 unreachable("Invalid address format");
3684 static inline unsigned
3685 nir_address_format_num_components(nir_address_format addr_format
)
3687 switch (addr_format
) {
3688 case nir_address_format_32bit_global
: return 1;
3689 case nir_address_format_64bit_global
: return 1;
3690 case nir_address_format_64bit_bounded_global
: return 4;
3691 case nir_address_format_32bit_index_offset
: return 2;
3692 case nir_address_format_32bit_offset
: return 1;
3693 case nir_address_format_logical
: return 1;
3695 unreachable("Invalid address format");
3698 static inline const struct glsl_type
*
3699 nir_address_format_to_glsl_type(nir_address_format addr_format
)
3701 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
3702 assert(bit_size
== 32 || bit_size
== 64);
3703 return glsl_vector_type(bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
,
3704 nir_address_format_num_components(addr_format
));
3707 const nir_const_value
*nir_address_format_null_value(nir_address_format addr_format
);
3709 nir_ssa_def
*nir_build_addr_ieq(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3710 nir_address_format addr_format
);
3712 nir_ssa_def
*nir_build_addr_isub(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3713 nir_address_format addr_format
);
3715 nir_ssa_def
* nir_explicit_io_address_from_deref(struct nir_builder
*b
,
3716 nir_deref_instr
*deref
,
3717 nir_ssa_def
*base_addr
,
3718 nir_address_format addr_format
);
3719 void nir_lower_explicit_io_instr(struct nir_builder
*b
,
3720 nir_intrinsic_instr
*io_instr
,
3722 nir_address_format addr_format
);
3724 bool nir_lower_explicit_io(nir_shader
*shader
,
3725 nir_variable_mode modes
,
3726 nir_address_format
);
3728 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
3729 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
3731 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
3733 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
3734 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
3735 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
3737 bool nir_remove_dead_derefs(nir_shader
*shader
);
3738 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
3739 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
);
3740 bool nir_lower_constant_initializers(nir_shader
*shader
,
3741 nir_variable_mode modes
);
3743 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
3744 bool nir_lower_vec_to_movs(nir_shader
*shader
);
3745 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
3747 const gl_state_index16
*alpha_ref_state_tokens
);
3748 bool nir_lower_alu(nir_shader
*shader
);
3750 bool nir_lower_flrp(nir_shader
*shader
, unsigned lowering_mask
,
3751 bool always_precise
, bool have_ffma
);
3753 bool nir_lower_alu_to_scalar(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
3754 bool nir_lower_bool_to_float(nir_shader
*shader
);
3755 bool nir_lower_bool_to_int32(nir_shader
*shader
);
3756 bool nir_lower_int_to_float(nir_shader
*shader
);
3757 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
3758 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
3759 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
3760 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
3761 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
3763 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
3764 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
3765 bool nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode mask
);
3767 void nir_lower_fragcoord_wtrans(nir_shader
*shader
);
3768 void nir_lower_viewport_transform(nir_shader
*shader
);
3769 bool nir_lower_uniforms_to_ubo(nir_shader
*shader
, int multiplier
);
3771 typedef struct nir_lower_subgroups_options
{
3772 uint8_t subgroup_size
;
3773 uint8_t ballot_bit_size
;
3774 bool lower_to_scalar
:1;
3775 bool lower_vote_trivial
:1;
3776 bool lower_vote_eq_to_ballot
:1;
3777 bool lower_subgroup_masks
:1;
3778 bool lower_shuffle
:1;
3779 bool lower_shuffle_to_32bit
:1;
3781 } nir_lower_subgroups_options
;
3783 bool nir_lower_subgroups(nir_shader
*shader
,
3784 const nir_lower_subgroups_options
*options
);
3786 bool nir_lower_system_values(nir_shader
*shader
);
3788 enum PACKED nir_lower_tex_packing
{
3789 nir_lower_tex_packing_none
= 0,
3790 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
3791 * or unsigned ints based on the sampler type
3793 nir_lower_tex_packing_16
,
3794 /* The sampler returns 1 32-bit word of 4x8 unorm */
3795 nir_lower_tex_packing_8
,
3798 typedef struct nir_lower_tex_options
{
3800 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
3801 * sampler types a texture projector is lowered.
3806 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
3808 bool lower_txf_offset
;
3811 * If true, lower away nir_tex_src_offset for all rect textures.
3813 bool lower_rect_offset
;
3816 * If true, lower rect textures to 2D, using txs to fetch the
3817 * texture dimensions and dividing the texture coords by the
3818 * texture dims to normalize.
3823 * If true, convert yuv to rgb.
3825 unsigned lower_y_uv_external
;
3826 unsigned lower_y_u_v_external
;
3827 unsigned lower_yx_xuxv_external
;
3828 unsigned lower_xy_uxvx_external
;
3829 unsigned lower_ayuv_external
;
3830 unsigned lower_xyuv_external
;
3833 * To emulate certain texture wrap modes, this can be used
3834 * to saturate the specified tex coord to [0.0, 1.0]. The
3835 * bits are according to sampler #, ie. if, for example:
3837 * (conf->saturate_s & (1 << n))
3839 * is true, then the s coord for sampler n is saturated.
3841 * Note that clamping must happen *after* projector lowering
3842 * so any projected texture sample instruction with a clamped
3843 * coordinate gets automatically lowered, regardless of the
3844 * 'lower_txp' setting.
3846 unsigned saturate_s
;
3847 unsigned saturate_t
;
3848 unsigned saturate_r
;
3850 /* Bitmask of textures that need swizzling.
3852 * If (swizzle_result & (1 << texture_index)), then the swizzle in
3853 * swizzles[texture_index] is applied to the result of the texturing
3856 unsigned swizzle_result
;
3858 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
3859 * while 4 and 5 represent 0 and 1 respectively.
3861 uint8_t swizzles
[32][4];
3863 /* Can be used to scale sampled values in range required by the format. */
3864 float scale_factors
[32];
3867 * Bitmap of textures that need srgb to linear conversion. If
3868 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
3869 * of the texture are lowered to linear.
3871 unsigned lower_srgb
;
3874 * If true, lower nir_texop_tex on shaders that doesn't support implicit
3875 * LODs to nir_texop_txl.
3877 bool lower_tex_without_implicit_lod
;
3880 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
3882 bool lower_txd_cube_map
;
3885 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
3890 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
3891 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
3892 * with lower_txd_cube_map.
3894 bool lower_txd_shadow
;
3897 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
3898 * Implies lower_txd_cube_map and lower_txd_shadow.
3903 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
3904 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
3906 bool lower_txb_shadow_clamp
;
3909 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
3910 * with nir_texop_txl. This includes cube maps.
3912 bool lower_txd_shadow_clamp
;
3915 * If true, lower nir_texop_txd on when it uses both offset and min_lod
3916 * with nir_texop_txl. This includes cube maps.
3918 bool lower_txd_offset_clamp
;
3921 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
3922 * sampler is bindless.
3924 bool lower_txd_clamp_bindless_sampler
;
3927 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
3928 * sampler index is not statically determinable to be less than 16.
3930 bool lower_txd_clamp_if_sampler_index_not_lt_16
;
3933 * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
3934 * 0-lod followed by a nir_ishr.
3939 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
3940 * mixed-up tg4 locations.
3942 bool lower_tg4_broadcom_swizzle
;
3945 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
3947 bool lower_tg4_offsets
;
3949 enum nir_lower_tex_packing lower_tex_packing
[32];
3950 } nir_lower_tex_options
;
3952 bool nir_lower_tex(nir_shader
*shader
,
3953 const nir_lower_tex_options
*options
);
3955 enum nir_lower_non_uniform_access_type
{
3956 nir_lower_non_uniform_ubo_access
= (1 << 0),
3957 nir_lower_non_uniform_ssbo_access
= (1 << 1),
3958 nir_lower_non_uniform_texture_access
= (1 << 2),
3959 nir_lower_non_uniform_image_access
= (1 << 3),
3962 bool nir_lower_non_uniform_access(nir_shader
*shader
,
3963 enum nir_lower_non_uniform_access_type
);
3965 enum nir_lower_idiv_path
{
3966 /* This path is based on NV50LegalizeSSA::handleDIV(). It is the faster of
3967 * the two but it is not exact in some cases (for example, 1091317713u /
3968 * 1034u gives 5209173 instead of 1055432) */
3969 nir_lower_idiv_fast
,
3970 /* This path is based on AMDGPUTargetLowering::LowerUDIVREM() and
3971 * AMDGPUTargetLowering::LowerSDIVREM(). It requires more instructions than
3972 * the nv50 path and many of them are integer multiplications, so it is
3973 * probably slower. It should always return the correct result, though. */
3974 nir_lower_idiv_precise
,
3977 bool nir_lower_idiv(nir_shader
*shader
, enum nir_lower_idiv_path path
);
3979 bool nir_lower_input_attachments(nir_shader
*shader
, bool use_fragcoord_sysval
);
3981 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
,
3983 bool use_clipdist_array
,
3984 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
3985 bool nir_lower_clip_gs(nir_shader
*shader
, unsigned ucp_enables
,
3986 bool use_clipdist_array
,
3987 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
3988 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
,
3989 bool use_clipdist_array
);
3990 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
3992 void nir_lower_point_size_mov(nir_shader
*shader
,
3993 const gl_state_index16
*pointsize_state_tokens
);
3995 bool nir_lower_frexp(nir_shader
*nir
);
3997 void nir_lower_two_sided_color(nir_shader
*shader
);
3999 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
4001 bool nir_lower_flatshade(nir_shader
*shader
);
4003 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
4004 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
4005 const gl_state_index16
*uniform_state_tokens
);
4007 typedef struct nir_lower_wpos_ytransform_options
{
4008 gl_state_index16 state_tokens
[STATE_LENGTH
];
4009 bool fs_coord_origin_upper_left
:1;
4010 bool fs_coord_origin_lower_left
:1;
4011 bool fs_coord_pixel_center_integer
:1;
4012 bool fs_coord_pixel_center_half_integer
:1;
4013 } nir_lower_wpos_ytransform_options
;
4015 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
4016 const nir_lower_wpos_ytransform_options
*options
);
4017 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
4019 bool nir_lower_fb_read(nir_shader
*shader
);
4021 typedef struct nir_lower_drawpixels_options
{
4022 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
4023 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
4024 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
4025 unsigned drawpix_sampler
;
4026 unsigned pixelmap_sampler
;
4028 bool scale_and_bias
:1;
4029 } nir_lower_drawpixels_options
;
4031 void nir_lower_drawpixels(nir_shader
*shader
,
4032 const nir_lower_drawpixels_options
*options
);
4034 typedef struct nir_lower_bitmap_options
{
4037 } nir_lower_bitmap_options
;
4039 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
4041 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
, unsigned ssbo_offset
);
4044 nir_lower_int_source_mods
= 1 << 0,
4045 nir_lower_float_source_mods
= 1 << 1,
4046 nir_lower_triop_abs
= 1 << 2,
4047 nir_lower_all_source_mods
= (1 << 3) - 1
4048 } nir_lower_to_source_mods_flags
;
4051 bool nir_lower_to_source_mods(nir_shader
*shader
, nir_lower_to_source_mods_flags options
);
4053 bool nir_lower_gs_intrinsics(nir_shader
*shader
);
4055 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
4057 bool nir_lower_bit_size(nir_shader
*shader
,
4058 nir_lower_bit_size_callback callback
,
4059 void *callback_data
);
4061 nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode
);
4062 bool nir_lower_int64(nir_shader
*shader
, nir_lower_int64_options options
);
4064 nir_lower_doubles_options
nir_lower_doubles_op_to_options_mask(nir_op opcode
);
4065 bool nir_lower_doubles(nir_shader
*shader
, const nir_shader
*softfp64
,
4066 nir_lower_doubles_options options
);
4067 bool nir_lower_pack(nir_shader
*shader
);
4069 bool nir_lower_point_size(nir_shader
*shader
, float min
, float max
);
4072 nir_lower_interpolation_at_sample
= (1 << 1),
4073 nir_lower_interpolation_at_offset
= (1 << 2),
4074 nir_lower_interpolation_centroid
= (1 << 3),
4075 nir_lower_interpolation_pixel
= (1 << 4),
4076 nir_lower_interpolation_sample
= (1 << 5),
4077 } nir_lower_interpolation_options
;
4079 bool nir_lower_interpolation(nir_shader
*shader
,
4080 nir_lower_interpolation_options options
);
4082 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
4084 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
4086 void nir_loop_analyze_impl(nir_function_impl
*impl
,
4087 nir_variable_mode indirect_mask
);
4089 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
4091 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
4092 bool nir_repair_ssa(nir_shader
*shader
);
4094 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
4095 bool nir_convert_to_lcssa(nir_shader
*shader
, bool skip_invariants
, bool skip_bool_invariants
);
4096 bool* nir_divergence_analysis(nir_shader
*shader
, nir_divergence_options options
);
4098 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
4099 * registers. If false, convert all values (even those not involved in a phi
4100 * node) to registers.
4102 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
4104 bool nir_lower_phis_to_regs_block(nir_block
*block
);
4105 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
4106 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
);
4108 bool nir_lower_samplers(nir_shader
*shader
);
4110 /* This is here for unit tests. */
4111 bool nir_opt_comparison_pre_impl(nir_function_impl
*impl
);
4113 bool nir_opt_comparison_pre(nir_shader
*shader
);
4115 bool nir_opt_access(nir_shader
*shader
);
4116 bool nir_opt_algebraic(nir_shader
*shader
);
4117 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
4118 bool nir_opt_algebraic_late(nir_shader
*shader
);
4119 bool nir_opt_constant_folding(nir_shader
*shader
);
4121 bool nir_opt_combine_stores(nir_shader
*shader
, nir_variable_mode modes
);
4123 bool nir_copy_prop(nir_shader
*shader
);
4125 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
4127 bool nir_opt_cse(nir_shader
*shader
);
4129 bool nir_opt_dce(nir_shader
*shader
);
4131 bool nir_opt_dead_cf(nir_shader
*shader
);
4133 bool nir_opt_dead_write_vars(nir_shader
*shader
);
4135 bool nir_opt_deref_impl(nir_function_impl
*impl
);
4136 bool nir_opt_deref(nir_shader
*shader
);
4138 bool nir_opt_find_array_copies(nir_shader
*shader
);
4140 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
4142 bool nir_opt_idiv_const(nir_shader
*shader
, unsigned min_bit_size
);
4144 bool nir_opt_if(nir_shader
*shader
, bool aggressive_last_continue
);
4146 bool nir_opt_intrinsics(nir_shader
*shader
);
4148 bool nir_opt_large_constants(nir_shader
*shader
,
4149 glsl_type_size_align_func size_align
,
4150 unsigned threshold
);
4152 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
4155 nir_move_const_undef
= (1 << 0),
4156 nir_move_load_ubo
= (1 << 1),
4157 nir_move_load_input
= (1 << 2),
4158 nir_move_comparisons
= (1 << 3),
4161 bool nir_can_move_instr(nir_instr
*instr
, nir_move_options options
);
4163 bool nir_opt_sink(nir_shader
*shader
, nir_move_options options
);
4165 bool nir_opt_move(nir_shader
*shader
, nir_move_options options
);
4167 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
,
4168 bool indirect_load_ok
, bool expensive_alu_ok
);
4170 bool nir_opt_rematerialize_compares(nir_shader
*shader
);
4172 bool nir_opt_remove_phis(nir_shader
*shader
);
4173 bool nir_opt_remove_phis_block(nir_block
*block
);
4175 bool nir_opt_shrink_load(nir_shader
*shader
);
4177 bool nir_opt_trivial_continues(nir_shader
*shader
);
4179 bool nir_opt_undef(nir_shader
*shader
);
4181 bool nir_opt_vectorize(nir_shader
*shader
);
4183 bool nir_opt_conditional_discard(nir_shader
*shader
);
4185 void nir_strip(nir_shader
*shader
);
4187 void nir_sweep(nir_shader
*shader
);
4189 void nir_remap_dual_slot_attributes(nir_shader
*shader
,
4190 uint64_t *dual_slot_inputs
);
4191 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
);
4193 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
4194 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);
4197 nir_variable_is_in_ubo(const nir_variable
*var
)
4199 return (var
->data
.mode
== nir_var_mem_ubo
&&
4200 var
->interface_type
!= NULL
);
4204 nir_variable_is_in_ssbo(const nir_variable
*var
)
4206 return (var
->data
.mode
== nir_var_mem_ssbo
&&
4207 var
->interface_type
!= NULL
);
4211 nir_variable_is_in_block(const nir_variable
*var
)
4213 return nir_variable_is_in_ubo(var
) || nir_variable_is_in_ssbo(var
);