2 * Copyright © 2014 Connor Abbott
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Connor Abbott (cwabbott0@gmail.com)
31 #include "util/hash_table.h"
32 #include "compiler/glsl/list.h"
33 #include "GL/gl.h" /* GLenum */
34 #include "util/list.h"
35 #include "util/ralloc.h"
37 #include "util/bitscan.h"
38 #include "util/bitset.h"
39 #include "util/macros.h"
40 #include "compiler/nir_types.h"
41 #include "compiler/shader_enums.h"
42 #include "compiler/shader_info.h"
46 #include "util/debug.h"
49 #include "nir_opcodes.h"
51 #if defined(_WIN32) && !defined(snprintf)
52 #define snprintf _snprintf
60 #define NIR_TRUE (~0u)
61 #define NIR_MAX_VEC_COMPONENTS 4
62 #define NIR_MAX_MATRIX_COLUMNS 4
63 typedef uint8_t nir_component_mask_t
;
65 /** Defines a cast function
67 * This macro defines a cast function from in_type to out_type where
68 * out_type is some structure type that contains a field of type out_type.
70 * Note that you have to be a bit careful as the generated cast function
73 #define NIR_DEFINE_CAST(name, in_type, out_type, field, \
74 type_field, type_value) \
75 static inline out_type * \
76 name(const in_type *parent) \
78 assert(parent && parent->type_field == type_value); \
79 return exec_node_data(out_type, parent, field); \
89 * Description of built-in state associated with a uniform
91 * \sa nir_variable::state_slots
94 gl_state_index16 tokens
[STATE_LENGTH
];
99 nir_var_shader_in
= (1 << 0),
100 nir_var_shader_out
= (1 << 1),
101 nir_var_shader_temp
= (1 << 2),
102 nir_var_function_temp
= (1 << 3),
103 nir_var_uniform
= (1 << 4),
104 nir_var_mem_ubo
= (1 << 5),
105 nir_var_system_value
= (1 << 6),
106 nir_var_mem_ssbo
= (1 << 7),
107 nir_var_mem_shared
= (1 << 8),
108 nir_var_mem_global
= (1 << 9),
116 nir_rounding_mode_undef
= 0,
117 nir_rounding_mode_rtne
= 1, /* round to nearest even */
118 nir_rounding_mode_ru
= 2, /* round up */
119 nir_rounding_mode_rd
= 3, /* round down */
120 nir_rounding_mode_rtz
= 4, /* round towards zero */
137 #define nir_const_value_to_array(arr, c, components, m) \
139 for (unsigned i = 0; i < components; ++i) \
143 static inline nir_const_value
144 nir_const_value_for_raw_uint(uint64_t x
, unsigned bit_size
)
147 memset(&v
, 0, sizeof(v
));
150 case 1: v
.b
= x
; break;
151 case 8: v
.u8
= x
; break;
152 case 16: v
.u16
= x
; break;
153 case 32: v
.u32
= x
; break;
154 case 64: v
.u64
= x
; break;
156 unreachable("Invalid bit size");
162 static inline nir_const_value
163 nir_const_value_for_int(int64_t i
, unsigned bit_size
)
166 memset(&v
, 0, sizeof(v
));
168 assert(bit_size
<= 64);
170 assert(i
>= (-(1ll << (bit_size
- 1))));
171 assert(i
< (1ll << (bit_size
- 1)));
174 return nir_const_value_for_raw_uint(i
, bit_size
);
177 static inline nir_const_value
178 nir_const_value_for_uint(uint64_t u
, unsigned bit_size
)
181 memset(&v
, 0, sizeof(v
));
183 assert(bit_size
<= 64);
185 assert(u
< (1ull << bit_size
));
187 return nir_const_value_for_raw_uint(u
, bit_size
);
190 static inline nir_const_value
191 nir_const_value_for_bool(bool b
, unsigned bit_size
)
193 /* Booleans use a 0/-1 convention */
194 return nir_const_value_for_int(-(int)b
, bit_size
);
197 /* This one isn't inline because it requires half-float conversion */
198 nir_const_value
nir_const_value_for_float(double b
, unsigned bit_size
);
200 static inline int64_t
201 nir_const_value_as_int(nir_const_value value
, unsigned bit_size
)
204 /* int1_t uses 0/-1 convention */
205 case 1: return -(int)value
.b
;
206 case 8: return value
.i8
;
207 case 16: return value
.i16
;
208 case 32: return value
.i32
;
209 case 64: return value
.i64
;
211 unreachable("Invalid bit size");
215 static inline uint64_t
216 nir_const_value_as_uint(nir_const_value value
, unsigned bit_size
)
219 case 1: return value
.b
;
220 case 8: return value
.u8
;
221 case 16: return value
.u16
;
222 case 32: return value
.u32
;
223 case 64: return value
.u64
;
225 unreachable("Invalid bit size");
230 nir_const_value_as_bool(nir_const_value value
, unsigned bit_size
)
232 int64_t i
= nir_const_value_as_int(value
, bit_size
);
234 /* Booleans of any size use 0/-1 convention */
235 assert(i
== 0 || i
== -1);
240 /* This one isn't inline because it requires half-float conversion */
241 double nir_const_value_as_float(nir_const_value value
, unsigned bit_size
);
243 typedef struct nir_constant
{
245 * Value of the constant.
247 * The field used to back the values supplied by the constant is determined
248 * by the type associated with the \c nir_variable. Constants may be
249 * scalars, vectors, or matrices.
251 nir_const_value values
[NIR_MAX_VEC_COMPONENTS
];
253 /* we could get this from the var->type but makes clone *much* easier to
254 * not have to care about the type.
256 unsigned num_elements
;
258 /* Array elements / Structure Fields */
259 struct nir_constant
**elements
;
263 * \brief Layout qualifiers for gl_FragDepth.
265 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
266 * with a layout qualifier.
269 nir_depth_layout_none
, /**< No depth layout is specified. */
270 nir_depth_layout_any
,
271 nir_depth_layout_greater
,
272 nir_depth_layout_less
,
273 nir_depth_layout_unchanged
277 * Enum keeping track of how a variable was declared.
281 * Normal declaration.
283 nir_var_declared_normally
= 0,
286 * Variable is implicitly generated by the compiler and should not be
287 * visible via the API.
290 } nir_var_declaration_type
;
293 * Either a uniform, global variable, shader input, or shader output. Based on
294 * ir_variable - it should be easy to translate between the two.
297 typedef struct nir_variable
{
298 struct exec_node node
;
301 * Declared type of the variable
303 const struct glsl_type
*type
;
306 * Declared name of the variable
310 struct nir_variable_data
{
312 * Storage class of the variable.
314 * \sa nir_variable_mode
316 nir_variable_mode mode
;
319 * Is the variable read-only?
321 * This is set for variables declared as \c const, shader inputs,
324 unsigned read_only
:1;
328 unsigned invariant
:1;
331 * Can this variable be coalesced with another?
333 * This is set by nir_lower_io_to_temporaries to say that any
334 * copies involving this variable should stay put. Propagating it can
335 * duplicate the resulting load/store, which is not wanted, and may
336 * result in a load/store of the variable with an indirect offset which
337 * the backend may not be able to handle.
339 unsigned cannot_coalesce
:1;
342 * When separate shader programs are enabled, only input/outputs between
343 * the stages of a multi-stage separate program can be safely removed
344 * from the shader interface. Other input/outputs must remains active.
346 * This is also used to make sure xfb varyings that are unused by the
347 * fragment shader are not removed.
349 unsigned always_active_io
:1;
352 * Interpolation mode for shader inputs / outputs
354 * \sa glsl_interp_mode
356 unsigned interpolation
:2;
359 * If non-zero, then this variable may be packed along with other variables
360 * into a single varying slot, so this offset should be applied when
361 * accessing components. For example, an offset of 1 means that the x
362 * component of this variable is actually stored in component y of the
363 * location specified by \c location.
365 unsigned location_frac
:2;
368 * If true, this variable represents an array of scalars that should
369 * be tightly packed. In other words, consecutive array elements
370 * should be stored one component apart, rather than one slot apart.
375 * Whether this is a fragment shader output implicitly initialized with
376 * the previous contents of the specified render target at the
377 * framebuffer location corresponding to this shader invocation.
379 unsigned fb_fetch_output
:1;
382 * Non-zero if this variable is considered bindless as defined by
383 * ARB_bindless_texture.
388 * Was an explicit binding set in the shader?
390 unsigned explicit_binding
:1;
393 * Was a transfer feedback buffer set in the shader?
395 unsigned explicit_xfb_buffer
:1;
398 * Was a transfer feedback stride set in the shader?
400 unsigned explicit_xfb_stride
:1;
403 * Was an explicit offset set in the shader?
405 unsigned explicit_offset
:1;
408 * How the variable was declared. See nir_var_declaration_type.
410 * This is used to detect variables generated by the compiler, so should
411 * not be visible via the API.
413 unsigned how_declared
:2;
416 * \brief Layout qualifier for gl_FragDepth.
418 * This is not equal to \c ir_depth_layout_none if and only if this
419 * variable is \c gl_FragDepth and a layout qualifier is specified.
421 nir_depth_layout depth_layout
;
424 * Storage location of the base of this variable
426 * The precise meaning of this field depends on the nature of the variable.
428 * - Vertex shader input: one of the values from \c gl_vert_attrib.
429 * - Vertex shader output: one of the values from \c gl_varying_slot.
430 * - Geometry shader input: one of the values from \c gl_varying_slot.
431 * - Geometry shader output: one of the values from \c gl_varying_slot.
432 * - Fragment shader input: one of the values from \c gl_varying_slot.
433 * - Fragment shader output: one of the values from \c gl_frag_result.
434 * - Uniforms: Per-stage uniform slot number for default uniform block.
435 * - Uniforms: Index within the uniform block definition for UBO members.
436 * - Non-UBO Uniforms: uniform slot number.
437 * - Other: This field is not currently used.
439 * If the variable is a uniform, shader input, or shader output, and the
440 * slot has not been assigned, the value will be -1.
445 * The actual location of the variable in the IR. Only valid for inputs
448 unsigned int driver_location
;
451 * Vertex stream output identifier.
453 * For packed outputs, bit 31 is set and bits [2*i+1,2*i] indicate the
454 * stream of the i-th component.
459 * output index for dual source blending.
464 * Descriptor set binding for sampler or UBO.
469 * Initial binding point for a sampler or UBO.
471 * For array types, this represents the binding point for the first element.
476 * Location an atomic counter or transform feedback is stored at.
481 * Transform feedback buffer.
486 * Transform feedback stride.
491 * ARB_shader_image_load_store qualifiers.
494 enum gl_access_qualifier access
;
496 /** Image internal format if specified explicitly, otherwise GL_NONE. */
501 /* Number of nir_variable_data members */
502 unsigned num_members
;
505 * Built-in state that backs this uniform
507 * Once set at variable creation, \c state_slots must remain invariant.
508 * This is because, ideally, this array would be shared by all clones of
509 * this variable in the IR tree. In other words, we'd really like for it
510 * to be a fly-weight.
512 * If the variable is not a uniform, \c num_state_slots will be zero and
513 * \c state_slots will be \c NULL.
516 unsigned num_state_slots
; /**< Number of state slots used */
517 nir_state_slot
*state_slots
; /**< State descriptors. */
521 * Constant expression assigned in the initializer of the variable
523 * This field should only be used temporarily by creators of NIR shaders
524 * and then lower_constant_initializers can be used to get rid of them.
525 * Most of the rest of NIR ignores this field or asserts that it's NULL.
527 nir_constant
*constant_initializer
;
530 * For variables that are in an interface block or are an instance of an
531 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
533 * \sa ir_variable::location
535 const struct glsl_type
*interface_type
;
538 * Description of per-member data for per-member struct variables
540 * This is used for variables which are actually an amalgamation of
541 * multiple entities such as a struct of built-in values or a struct of
542 * inputs each with their own layout specifier. This is only allowed on
543 * variables with a struct or array of array of struct type.
545 struct nir_variable_data
*members
;
548 #define nir_foreach_variable(var, var_list) \
549 foreach_list_typed(nir_variable, var, node, var_list)
551 #define nir_foreach_variable_safe(var, var_list) \
552 foreach_list_typed_safe(nir_variable, var, node, var_list)
555 nir_variable_is_global(const nir_variable
*var
)
557 return var
->data
.mode
!= nir_var_function_temp
;
560 typedef struct nir_register
{
561 struct exec_node node
;
563 unsigned num_components
; /** < number of vector components */
564 unsigned num_array_elems
; /** < size of array (0 for no array) */
566 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
569 /** generic register index. */
572 /** only for debug purposes, can be NULL */
575 /** set of nir_srcs where this register is used (read from) */
576 struct list_head uses
;
578 /** set of nir_dests where this register is defined (written to) */
579 struct list_head defs
;
581 /** set of nir_ifs where this register is used as a condition */
582 struct list_head if_uses
;
585 #define nir_foreach_register(reg, reg_list) \
586 foreach_list_typed(nir_register, reg, node, reg_list)
587 #define nir_foreach_register_safe(reg, reg_list) \
588 foreach_list_typed_safe(nir_register, reg, node, reg_list)
590 typedef enum PACKED
{
592 nir_instr_type_deref
,
595 nir_instr_type_intrinsic
,
596 nir_instr_type_load_const
,
598 nir_instr_type_ssa_undef
,
600 nir_instr_type_parallel_copy
,
603 typedef struct nir_instr
{
604 struct exec_node node
;
605 struct nir_block
*block
;
608 /* A temporary for optimization and analysis passes to use for storing
609 * flags. For instance, DCE uses this to store the "dead/live" info.
613 /** generic instruction index. */
617 static inline nir_instr
*
618 nir_instr_next(nir_instr
*instr
)
620 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
621 if (exec_node_is_tail_sentinel(next
))
624 return exec_node_data(nir_instr
, next
, node
);
627 static inline nir_instr
*
628 nir_instr_prev(nir_instr
*instr
)
630 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
631 if (exec_node_is_head_sentinel(prev
))
634 return exec_node_data(nir_instr
, prev
, node
);
638 nir_instr_is_first(const nir_instr
*instr
)
640 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr
->node
));
644 nir_instr_is_last(const nir_instr
*instr
)
646 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr
->node
));
649 typedef struct nir_ssa_def
{
650 /** for debugging only, can be NULL */
653 /** generic SSA definition index. */
656 /** Index into the live_in and live_out bitfields */
659 /** Instruction which produces this SSA value. */
660 nir_instr
*parent_instr
;
662 /** set of nir_instrs where this register is used (read from) */
663 struct list_head uses
;
665 /** set of nir_ifs where this register is used as a condition */
666 struct list_head if_uses
;
668 uint8_t num_components
;
670 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
678 struct nir_src
*indirect
; /** < NULL for no indirect offset */
679 unsigned base_offset
;
681 /* TODO use-def chain goes here */
685 nir_instr
*parent_instr
;
686 struct list_head def_link
;
689 struct nir_src
*indirect
; /** < NULL for no indirect offset */
690 unsigned base_offset
;
692 /* TODO def-use chain goes here */
697 typedef struct nir_src
{
699 /** Instruction that consumes this value as a source. */
700 nir_instr
*parent_instr
;
701 struct nir_if
*parent_if
;
704 struct list_head use_link
;
714 static inline nir_src
717 nir_src src
= { { NULL
} };
721 #define NIR_SRC_INIT nir_src_init()
723 #define nir_foreach_use(src, reg_or_ssa_def) \
724 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
726 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
727 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
729 #define nir_foreach_if_use(src, reg_or_ssa_def) \
730 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
732 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
733 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
744 static inline nir_dest
747 nir_dest dest
= { { { NULL
} } };
751 #define NIR_DEST_INIT nir_dest_init()
753 #define nir_foreach_def(dest, reg) \
754 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
756 #define nir_foreach_def_safe(dest, reg) \
757 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
759 static inline nir_src
760 nir_src_for_ssa(nir_ssa_def
*def
)
762 nir_src src
= NIR_SRC_INIT
;
770 static inline nir_src
771 nir_src_for_reg(nir_register
*reg
)
773 nir_src src
= NIR_SRC_INIT
;
777 src
.reg
.indirect
= NULL
;
778 src
.reg
.base_offset
= 0;
783 static inline nir_dest
784 nir_dest_for_reg(nir_register
*reg
)
786 nir_dest dest
= NIR_DEST_INIT
;
793 static inline unsigned
794 nir_src_bit_size(nir_src src
)
796 return src
.is_ssa
? src
.ssa
->bit_size
: src
.reg
.reg
->bit_size
;
799 static inline unsigned
800 nir_src_num_components(nir_src src
)
802 return src
.is_ssa
? src
.ssa
->num_components
: src
.reg
.reg
->num_components
;
806 nir_src_is_const(nir_src src
)
809 src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
;
812 static inline unsigned
813 nir_dest_bit_size(nir_dest dest
)
815 return dest
.is_ssa
? dest
.ssa
.bit_size
: dest
.reg
.reg
->bit_size
;
818 static inline unsigned
819 nir_dest_num_components(nir_dest dest
)
821 return dest
.is_ssa
? dest
.ssa
.num_components
: dest
.reg
.reg
->num_components
;
824 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
825 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
831 * \name input modifiers
835 * For inputs interpreted as floating point, flips the sign bit. For
836 * inputs interpreted as integers, performs the two's complement negation.
841 * Clears the sign bit for floating point values, and computes the integer
842 * absolute value for integers. Note that the negate modifier acts after
843 * the absolute value modifier, therefore if both are set then all inputs
844 * will become negative.
850 * For each input component, says which component of the register it is
851 * chosen from. Note that which elements of the swizzle are used and which
852 * are ignored are based on the write mask for most opcodes - for example,
853 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
854 * a swizzle of {2, x, 1, 0} where x means "don't care."
856 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
863 * \name saturate output modifier
865 * Only valid for opcodes that output floating-point numbers. Clamps the
866 * output to between 0.0 and 1.0 inclusive.
871 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
874 /** NIR sized and unsized types
876 * The values in this enum are carefully chosen so that the sized type is
877 * just the unsized type OR the number of bits.
880 nir_type_invalid
= 0, /* Not a valid type */
884 nir_type_float
= 128,
885 nir_type_bool1
= 1 | nir_type_bool
,
886 nir_type_bool32
= 32 | nir_type_bool
,
887 nir_type_int1
= 1 | nir_type_int
,
888 nir_type_int8
= 8 | nir_type_int
,
889 nir_type_int16
= 16 | nir_type_int
,
890 nir_type_int32
= 32 | nir_type_int
,
891 nir_type_int64
= 64 | nir_type_int
,
892 nir_type_uint1
= 1 | nir_type_uint
,
893 nir_type_uint8
= 8 | nir_type_uint
,
894 nir_type_uint16
= 16 | nir_type_uint
,
895 nir_type_uint32
= 32 | nir_type_uint
,
896 nir_type_uint64
= 64 | nir_type_uint
,
897 nir_type_float16
= 16 | nir_type_float
,
898 nir_type_float32
= 32 | nir_type_float
,
899 nir_type_float64
= 64 | nir_type_float
,
902 #define NIR_ALU_TYPE_SIZE_MASK 0x79
903 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
905 static inline unsigned
906 nir_alu_type_get_type_size(nir_alu_type type
)
908 return type
& NIR_ALU_TYPE_SIZE_MASK
;
911 static inline unsigned
912 nir_alu_type_get_base_type(nir_alu_type type
)
914 return type
& NIR_ALU_TYPE_BASE_TYPE_MASK
;
917 static inline nir_alu_type
918 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
922 return nir_type_bool1
;
925 return nir_type_uint32
;
928 return nir_type_int32
;
930 case GLSL_TYPE_UINT16
:
931 return nir_type_uint16
;
933 case GLSL_TYPE_INT16
:
934 return nir_type_int16
;
936 case GLSL_TYPE_UINT8
:
937 return nir_type_uint8
;
939 return nir_type_int8
;
940 case GLSL_TYPE_UINT64
:
941 return nir_type_uint64
;
943 case GLSL_TYPE_INT64
:
944 return nir_type_int64
;
946 case GLSL_TYPE_FLOAT
:
947 return nir_type_float32
;
949 case GLSL_TYPE_FLOAT16
:
950 return nir_type_float16
;
952 case GLSL_TYPE_DOUBLE
:
953 return nir_type_float64
;
956 case GLSL_TYPE_SAMPLER
:
957 case GLSL_TYPE_IMAGE
:
958 case GLSL_TYPE_ATOMIC_UINT
:
959 case GLSL_TYPE_STRUCT
:
960 case GLSL_TYPE_INTERFACE
:
961 case GLSL_TYPE_ARRAY
:
963 case GLSL_TYPE_SUBROUTINE
:
964 case GLSL_TYPE_FUNCTION
:
965 case GLSL_TYPE_ERROR
:
966 return nir_type_invalid
;
969 unreachable("unknown type");
972 static inline nir_alu_type
973 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
975 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
978 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
979 nir_rounding_mode rnd
);
982 nir_op_vec(unsigned components
)
984 switch (components
) {
985 case 1: return nir_op_mov
;
986 case 2: return nir_op_vec2
;
987 case 3: return nir_op_vec3
;
988 case 4: return nir_op_vec4
;
989 default: unreachable("bad component count");
994 nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode
, unsigned bit_size
)
996 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
) ||
997 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
) ||
998 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
);
1002 nir_is_denorm_flush_to_zero(unsigned execution_mode
, unsigned bit_size
)
1004 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
) ||
1005 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
) ||
1006 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
);
1010 nir_is_denorm_preserve(unsigned execution_mode
, unsigned bit_size
)
1012 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP16
) ||
1013 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP32
) ||
1014 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP64
);
1018 nir_is_rounding_mode_rtne(unsigned execution_mode
, unsigned bit_size
)
1020 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1021 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1022 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1026 nir_is_rounding_mode_rtz(unsigned execution_mode
, unsigned bit_size
)
1028 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1029 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1030 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1034 nir_has_any_rounding_mode_rtz(unsigned execution_mode
)
1036 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1037 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1038 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1042 nir_has_any_rounding_mode_rtne(unsigned execution_mode
)
1044 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1045 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1046 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1049 static inline nir_rounding_mode
1050 nir_get_rounding_mode_from_float_controls(unsigned execution_mode
,
1053 if (nir_alu_type_get_base_type(type
) != nir_type_float
)
1054 return nir_rounding_mode_undef
;
1056 unsigned bit_size
= nir_alu_type_get_type_size(type
);
1058 if (nir_is_rounding_mode_rtz(execution_mode
, bit_size
))
1059 return nir_rounding_mode_rtz
;
1060 if (nir_is_rounding_mode_rtne(execution_mode
, bit_size
))
1061 return nir_rounding_mode_rtne
;
1062 return nir_rounding_mode_undef
;
1066 nir_has_any_rounding_mode_enabled(unsigned execution_mode
)
1069 nir_has_any_rounding_mode_rtne(execution_mode
) ||
1070 nir_has_any_rounding_mode_rtz(execution_mode
);
1076 * Operation where the first two sources are commutative.
1078 * For 2-source operations, this just mathematical commutativity. Some
1079 * 3-source operations, like ffma, are only commutative in the first two
1082 NIR_OP_IS_2SRC_COMMUTATIVE
= (1 << 0),
1083 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
1084 } nir_op_algebraic_property
;
1089 unsigned num_inputs
;
1092 * The number of components in the output
1094 * If non-zero, this is the size of the output and input sizes are
1095 * explicitly given; swizzle and writemask are still in effect, but if
1096 * the output component is masked out, then the input component may
1099 * If zero, the opcode acts in the standard, per-component manner; the
1100 * operation is performed on each component (except the ones that are
1101 * masked out) with the input being taken from the input swizzle for
1104 * The size of some of the inputs may be given (i.e. non-zero) even
1105 * though output_size is zero; in that case, the inputs with a zero
1106 * size act per-component, while the inputs with non-zero size don't.
1108 unsigned output_size
;
1111 * The type of vector that the instruction outputs. Note that the
1112 * staurate modifier is only allowed on outputs with the float type.
1115 nir_alu_type output_type
;
1118 * The number of components in each input
1120 unsigned input_sizes
[NIR_MAX_VEC_COMPONENTS
];
1123 * The type of vector that each input takes. Note that negate and
1124 * absolute value are only allowed on inputs with int or float type and
1125 * behave differently on the two.
1127 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
1129 nir_op_algebraic_property algebraic_properties
;
1131 /* Whether this represents a numeric conversion opcode */
1135 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
1137 typedef struct nir_alu_instr
{
1141 /** Indicates that this ALU instruction generates an exact value
1143 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1144 * really equivalent to either. This indicates that the value generated by
1145 * this operation is high-precision and any code transformations that touch
1146 * it must ensure that the resulting value is bit-for-bit identical to the
1152 * Indicates that this instruction do not cause wrapping to occur, in the
1153 * form of overflow or underflow.
1155 bool no_signed_wrap
:1;
1156 bool no_unsigned_wrap
:1;
1162 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
1163 nir_alu_instr
*instr
);
1164 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
1165 nir_alu_instr
*instr
);
1167 /* is this source channel used? */
1169 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
1172 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1173 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
1175 return (instr
->dest
.write_mask
>> channel
) & 1;
1178 static inline nir_component_mask_t
1179 nir_alu_instr_src_read_mask(const nir_alu_instr
*instr
, unsigned src
)
1181 nir_component_mask_t read_mask
= 0;
1182 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; c
++) {
1183 if (!nir_alu_instr_channel_used(instr
, src
, c
))
1186 read_mask
|= (1 << instr
->src
[src
].swizzle
[c
]);
1192 * Get the number of channels used for a source
1194 static inline unsigned
1195 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
1197 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1198 return nir_op_infos
[instr
->op
].input_sizes
[src
];
1200 return nir_dest_num_components(instr
->dest
.dest
);
1204 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1206 switch (instr
->op
) {
1226 bool nir_const_value_negative_equal(nir_const_value c1
, nir_const_value c2
,
1227 nir_alu_type full_type
);
1229 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
1230 unsigned src1
, unsigned src2
);
1232 bool nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
1233 const nir_alu_instr
*alu2
,
1234 unsigned src1
, unsigned src2
);
1238 nir_deref_type_array
,
1239 nir_deref_type_array_wildcard
,
1240 nir_deref_type_ptr_as_array
,
1241 nir_deref_type_struct
,
1242 nir_deref_type_cast
,
1248 /** The type of this deref instruction */
1249 nir_deref_type deref_type
;
1251 /** The mode of the underlying variable */
1252 nir_variable_mode mode
;
1254 /** The dereferenced type of the resulting pointer value */
1255 const struct glsl_type
*type
;
1258 /** Variable being dereferenced if deref_type is a deref_var */
1261 /** Parent deref if deref_type is not deref_var */
1265 /** Additional deref parameters */
1276 unsigned ptr_stride
;
1280 /** Destination to store the resulting "pointer" */
1284 static inline nir_deref_instr
*nir_src_as_deref(nir_src src
);
1286 static inline nir_deref_instr
*
1287 nir_deref_instr_parent(const nir_deref_instr
*instr
)
1289 if (instr
->deref_type
== nir_deref_type_var
)
1292 return nir_src_as_deref(instr
->parent
);
1295 static inline nir_variable
*
1296 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1298 while (instr
->deref_type
!= nir_deref_type_var
) {
1299 if (instr
->deref_type
== nir_deref_type_cast
)
1302 instr
= nir_deref_instr_parent(instr
);
1308 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1309 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr
*instr
);
1310 bool nir_deref_instr_has_complex_use(nir_deref_instr
*instr
);
1312 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1314 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*instr
);
1319 struct nir_function
*callee
;
1321 unsigned num_params
;
1325 #include "nir_intrinsics.h"
1327 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1329 /** Represents an intrinsic
1331 * An intrinsic is an instruction type for handling things that are
1332 * more-or-less regular operations but don't just consume and produce SSA
1333 * values like ALU operations do. Intrinsics are not for things that have
1334 * special semantic meaning such as phi nodes and parallel copies.
1335 * Examples of intrinsics include variable load/store operations, system
1336 * value loads, and the like. Even though texturing more-or-less falls
1337 * under this category, texturing is its own instruction type because
1338 * trying to represent texturing with intrinsics would lead to a
1339 * combinatorial explosion of intrinsic opcodes.
1341 * By having a single instruction type for handling a lot of different
1342 * cases, optimization passes can look for intrinsics and, for the most
1343 * part, completely ignore them. Each intrinsic type also has a few
1344 * possible flags that govern whether or not they can be reordered or
1345 * eliminated. That way passes like dead code elimination can still work
1346 * on intrisics without understanding the meaning of each.
1348 * Each intrinsic has some number of constant indices, some number of
1349 * variables, and some number of sources. What these sources, variables,
1350 * and indices mean depends on the intrinsic and is documented with the
1351 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1352 * instructions are the only types of instruction that can operate on
1358 nir_intrinsic_op intrinsic
;
1362 /** number of components if this is a vectorized intrinsic
1364 * Similarly to ALU operations, some intrinsics are vectorized.
1365 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1366 * For vectorized intrinsics, the num_components field specifies the
1367 * number of destination components and the number of source components
1368 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1370 uint8_t num_components
;
1372 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1375 } nir_intrinsic_instr
;
1377 static inline nir_variable
*
1378 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1380 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1384 /* Memory ordering. */
1385 NIR_MEMORY_ACQUIRE
= 1 << 0,
1386 NIR_MEMORY_RELEASE
= 1 << 1,
1388 /* Memory visibility operations. */
1389 NIR_MEMORY_MAKE_AVAILABLE
= 1 << 3,
1390 NIR_MEMORY_MAKE_VISIBLE
= 1 << 4,
1391 } nir_memory_semantics
;
1395 NIR_SCOPE_QUEUE_FAMILY
,
1396 NIR_SCOPE_WORKGROUP
,
1398 NIR_SCOPE_INVOCATION
,
1402 * \name NIR intrinsics semantic flags
1404 * information about what the compiler can do with the intrinsics.
1406 * \sa nir_intrinsic_info::flags
1410 * whether the intrinsic can be safely eliminated if none of its output
1411 * value is not being used.
1413 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1416 * Whether the intrinsic can be reordered with respect to any other
1417 * intrinsic, i.e. whether the only reordering dependencies of the
1418 * intrinsic are due to the register reads/writes.
1420 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1421 } nir_intrinsic_semantic_flag
;
1424 * \name NIR intrinsics const-index flag
1426 * Indicates the usage of a const_index slot.
1428 * \sa nir_intrinsic_info::index_map
1432 * Generally instructions that take a offset src argument, can encode
1433 * a constant 'base' value which is added to the offset.
1435 NIR_INTRINSIC_BASE
= 1,
1438 * For store instructions, a writemask for the store.
1440 NIR_INTRINSIC_WRMASK
,
1443 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1445 NIR_INTRINSIC_STREAM_ID
,
1448 * The clip-plane id for load_user_clip_plane intrinsic.
1450 NIR_INTRINSIC_UCP_ID
,
1453 * The amount of data, starting from BASE, that this instruction may
1454 * access. This is used to provide bounds if the offset is not constant.
1456 NIR_INTRINSIC_RANGE
,
1459 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1461 NIR_INTRINSIC_DESC_SET
,
1464 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1466 NIR_INTRINSIC_BINDING
,
1471 NIR_INTRINSIC_COMPONENT
,
1474 * Interpolation mode (only meaningful for FS inputs).
1476 NIR_INTRINSIC_INTERP_MODE
,
1479 * A binary nir_op to use when performing a reduction or scan operation
1481 NIR_INTRINSIC_REDUCTION_OP
,
1484 * Cluster size for reduction operations
1486 NIR_INTRINSIC_CLUSTER_SIZE
,
1489 * Parameter index for a load_param intrinsic
1491 NIR_INTRINSIC_PARAM_IDX
,
1494 * Image dimensionality for image intrinsics
1496 * One of GLSL_SAMPLER_DIM_*
1498 NIR_INTRINSIC_IMAGE_DIM
,
1501 * Non-zero if we are accessing an array image
1503 NIR_INTRINSIC_IMAGE_ARRAY
,
1506 * Image format for image intrinsics
1508 NIR_INTRINSIC_FORMAT
,
1511 * Access qualifiers for image and memory access intrinsics
1513 NIR_INTRINSIC_ACCESS
,
1516 * Alignment for offsets and addresses
1518 * These two parameters, specify an alignment in terms of a multiplier and
1519 * an offset. The offset or address parameter X of the intrinsic is
1520 * guaranteed to satisfy the following:
1522 * (X - align_offset) % align_mul == 0
1524 NIR_INTRINSIC_ALIGN_MUL
,
1525 NIR_INTRINSIC_ALIGN_OFFSET
,
1528 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1530 NIR_INTRINSIC_DESC_TYPE
,
1533 * The nir_alu_type of a uniform/input/output
1538 * The swizzle mask for the instructions
1539 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1541 NIR_INTRINSIC_SWIZZLE_MASK
,
1543 /* Separate source/dest access flags for copies */
1544 NIR_INTRINSIC_SRC_ACCESS
,
1545 NIR_INTRINSIC_DST_ACCESS
,
1547 /* Driver location for nir_load_patch_location_ir3 */
1548 NIR_INTRINSIC_DRIVER_LOCATION
,
1551 * Mask of nir_memory_semantics, includes ordering and visibility.
1553 NIR_INTRINSIC_MEMORY_SEMANTICS
,
1556 * Mask of nir_variable_modes affected by the memory operation.
1558 NIR_INTRINSIC_MEMORY_MODES
,
1561 * Value of nir_scope.
1563 NIR_INTRINSIC_MEMORY_SCOPE
,
1565 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1567 } nir_intrinsic_index_flag
;
1569 #define NIR_INTRINSIC_MAX_INPUTS 5
1574 unsigned num_srcs
; /** < number of register/SSA inputs */
1576 /** number of components of each input register
1578 * If this value is 0, the number of components is given by the
1579 * num_components field of nir_intrinsic_instr. If this value is -1, the
1580 * intrinsic consumes however many components are provided and it is not
1583 int src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1587 /** number of components of the output register
1589 * If this value is 0, the number of components is given by the
1590 * num_components field of nir_intrinsic_instr.
1592 unsigned dest_components
;
1594 /** bitfield of legal bit sizes */
1595 unsigned dest_bit_sizes
;
1597 /** the number of constant indices used by the intrinsic */
1598 unsigned num_indices
;
1600 /** indicates the usage of intr->const_index[n] */
1601 unsigned index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1603 /** semantic flags for calls to this intrinsic */
1604 nir_intrinsic_semantic_flag flags
;
1605 } nir_intrinsic_info
;
1607 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1609 static inline unsigned
1610 nir_intrinsic_src_components(nir_intrinsic_instr
*intr
, unsigned srcn
)
1612 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1613 assert(srcn
< info
->num_srcs
);
1614 if (info
->src_components
[srcn
] > 0)
1615 return info
->src_components
[srcn
];
1616 else if (info
->src_components
[srcn
] == 0)
1617 return intr
->num_components
;
1619 return nir_src_num_components(intr
->src
[srcn
]);
1622 static inline unsigned
1623 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1625 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1626 if (!info
->has_dest
)
1628 else if (info
->dest_components
)
1629 return info
->dest_components
;
1631 return intr
->num_components
;
1634 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1635 static inline type \
1636 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1638 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1639 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1640 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1642 static inline void \
1643 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1645 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1646 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1647 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1650 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1651 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1652 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1653 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1654 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1655 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1656 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1657 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1658 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1659 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1660 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1661 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1662 INTRINSIC_IDX_ACCESSORS(image_dim
, IMAGE_DIM
, enum glsl_sampler_dim
)
1663 INTRINSIC_IDX_ACCESSORS(image_array
, IMAGE_ARRAY
, bool)
1664 INTRINSIC_IDX_ACCESSORS(access
, ACCESS
, enum gl_access_qualifier
)
1665 INTRINSIC_IDX_ACCESSORS(src_access
, SRC_ACCESS
, enum gl_access_qualifier
)
1666 INTRINSIC_IDX_ACCESSORS(dst_access
, DST_ACCESS
, enum gl_access_qualifier
)
1667 INTRINSIC_IDX_ACCESSORS(format
, FORMAT
, unsigned)
1668 INTRINSIC_IDX_ACCESSORS(align_mul
, ALIGN_MUL
, unsigned)
1669 INTRINSIC_IDX_ACCESSORS(align_offset
, ALIGN_OFFSET
, unsigned)
1670 INTRINSIC_IDX_ACCESSORS(desc_type
, DESC_TYPE
, unsigned)
1671 INTRINSIC_IDX_ACCESSORS(type
, TYPE
, nir_alu_type
)
1672 INTRINSIC_IDX_ACCESSORS(swizzle_mask
, SWIZZLE_MASK
, unsigned)
1673 INTRINSIC_IDX_ACCESSORS(driver_location
, DRIVER_LOCATION
, unsigned)
1674 INTRINSIC_IDX_ACCESSORS(memory_semantics
, MEMORY_SEMANTICS
, nir_memory_semantics
)
1675 INTRINSIC_IDX_ACCESSORS(memory_modes
, MEMORY_MODES
, nir_variable_mode
)
1676 INTRINSIC_IDX_ACCESSORS(memory_scope
, MEMORY_SCOPE
, nir_scope
)
1679 nir_intrinsic_set_align(nir_intrinsic_instr
*intrin
,
1680 unsigned align_mul
, unsigned align_offset
)
1682 assert(util_is_power_of_two_nonzero(align_mul
));
1683 assert(align_offset
< align_mul
);
1684 nir_intrinsic_set_align_mul(intrin
, align_mul
);
1685 nir_intrinsic_set_align_offset(intrin
, align_offset
);
1688 /** Returns a simple alignment for a load/store intrinsic offset
1690 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1691 * and ALIGN_OFFSET parameters, this helper takes both into account and
1692 * provides a single simple alignment parameter. The offset X is guaranteed
1693 * to satisfy X % align == 0.
1695 static inline unsigned
1696 nir_intrinsic_align(const nir_intrinsic_instr
*intrin
)
1698 const unsigned align_mul
= nir_intrinsic_align_mul(intrin
);
1699 const unsigned align_offset
= nir_intrinsic_align_offset(intrin
);
1700 assert(align_offset
< align_mul
);
1701 return align_offset
? 1 << (ffs(align_offset
) - 1) : align_mul
;
1704 /* Converts a image_deref_* intrinsic into a image_* one */
1705 void nir_rewrite_image_intrinsic(nir_intrinsic_instr
*instr
,
1706 nir_ssa_def
*handle
, bool bindless
);
1708 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1710 nir_intrinsic_can_reorder(nir_intrinsic_instr
*instr
)
1712 if (instr
->intrinsic
== nir_intrinsic_load_deref
||
1713 instr
->intrinsic
== nir_intrinsic_load_ssbo
||
1714 instr
->intrinsic
== nir_intrinsic_bindless_image_load
||
1715 instr
->intrinsic
== nir_intrinsic_image_deref_load
||
1716 instr
->intrinsic
== nir_intrinsic_image_load
) {
1717 return nir_intrinsic_access(instr
) & ACCESS_CAN_REORDER
;
1719 const nir_intrinsic_info
*info
=
1720 &nir_intrinsic_infos
[instr
->intrinsic
];
1721 return (info
->flags
& NIR_INTRINSIC_CAN_ELIMINATE
) &&
1722 (info
->flags
& NIR_INTRINSIC_CAN_REORDER
);
1727 * \group texture information
1729 * This gives semantic information about textures which is useful to the
1730 * frontend, the backend, and lowering passes, but not the optimizer.
1735 nir_tex_src_projector
,
1736 nir_tex_src_comparator
, /* shadow comparator */
1740 nir_tex_src_min_lod
,
1741 nir_tex_src_ms_index
, /* MSAA sample index */
1742 nir_tex_src_ms_mcs
, /* MSAA compression value */
1745 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
1746 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
1747 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
1748 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
1749 nir_tex_src_texture_handle
, /* < bindless texture handle */
1750 nir_tex_src_sampler_handle
, /* < bindless sampler handle */
1751 nir_tex_src_plane
, /* < selects plane for planar textures */
1752 nir_num_tex_src_types
1757 nir_tex_src_type src_type
;
1761 nir_texop_tex
, /**< Regular texture look-up */
1762 nir_texop_txb
, /**< Texture look-up with LOD bias */
1763 nir_texop_txl
, /**< Texture look-up with explicit LOD */
1764 nir_texop_txd
, /**< Texture look-up with partial derivatives */
1765 nir_texop_txf
, /**< Texel fetch with explicit LOD */
1766 nir_texop_txf_ms
, /**< Multisample texture fetch */
1767 nir_texop_txf_ms_fb
, /**< Multisample texture fetch from framebuffer */
1768 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
1769 nir_texop_txs
, /**< Texture size */
1770 nir_texop_lod
, /**< Texture lod query */
1771 nir_texop_tg4
, /**< Texture gather */
1772 nir_texop_query_levels
, /**< Texture levels query */
1773 nir_texop_texture_samples
, /**< Texture samples query */
1774 nir_texop_samples_identical
, /**< Query whether all samples are definitely
1777 nir_texop_tex_prefetch
, /**< Regular texture look-up, eligible for pre-dispatch */
1783 enum glsl_sampler_dim sampler_dim
;
1784 nir_alu_type dest_type
;
1789 unsigned num_srcs
, coord_components
;
1790 bool is_array
, is_shadow
;
1793 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1794 * components or the new-style shadow that outputs 1 component.
1796 bool is_new_style_shadow
;
1798 /* gather component selector */
1799 unsigned component
: 2;
1801 /* gather offsets */
1802 int8_t tg4_offsets
[4][2];
1804 /* True if the texture index or handle is not dynamically uniform */
1805 bool texture_non_uniform
;
1807 /* True if the sampler index or handle is not dynamically uniform */
1808 bool sampler_non_uniform
;
1810 /** The texture index
1812 * If this texture instruction has a nir_tex_src_texture_offset source,
1813 * then the texture index is given by texture_index + texture_offset.
1815 unsigned texture_index
;
1817 /** The size of the texture array or 0 if it's not an array */
1818 unsigned texture_array_size
;
1820 /** The sampler index
1822 * The following operations do not require a sampler and, as such, this
1823 * field should be ignored:
1825 * - nir_texop_txf_ms
1828 * - nir_texop_query_levels
1829 * - nir_texop_texture_samples
1830 * - nir_texop_samples_identical
1832 * If this texture instruction has a nir_tex_src_sampler_offset source,
1833 * then the sampler index is given by sampler_index + sampler_offset.
1835 unsigned sampler_index
;
1838 static inline unsigned
1839 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
1841 switch (instr
->op
) {
1842 case nir_texop_txs
: {
1844 switch (instr
->sampler_dim
) {
1845 case GLSL_SAMPLER_DIM_1D
:
1846 case GLSL_SAMPLER_DIM_BUF
:
1849 case GLSL_SAMPLER_DIM_2D
:
1850 case GLSL_SAMPLER_DIM_CUBE
:
1851 case GLSL_SAMPLER_DIM_MS
:
1852 case GLSL_SAMPLER_DIM_RECT
:
1853 case GLSL_SAMPLER_DIM_EXTERNAL
:
1854 case GLSL_SAMPLER_DIM_SUBPASS
:
1857 case GLSL_SAMPLER_DIM_3D
:
1861 unreachable("not reached");
1863 if (instr
->is_array
)
1871 case nir_texop_texture_samples
:
1872 case nir_texop_query_levels
:
1873 case nir_texop_samples_identical
:
1877 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
1884 /* Returns true if this texture operation queries something about the texture
1885 * rather than actually sampling it.
1888 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
1890 switch (instr
->op
) {
1893 case nir_texop_texture_samples
:
1894 case nir_texop_query_levels
:
1895 case nir_texop_txf_ms_mcs
:
1902 case nir_texop_txf_ms
:
1903 case nir_texop_txf_ms_fb
:
1907 unreachable("Invalid texture opcode");
1912 nir_tex_instr_has_implicit_derivative(const nir_tex_instr
*instr
)
1914 switch (instr
->op
) {
1924 static inline nir_alu_type
1925 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
1927 switch (instr
->src
[src
].src_type
) {
1928 case nir_tex_src_coord
:
1929 switch (instr
->op
) {
1931 case nir_texop_txf_ms
:
1932 case nir_texop_txf_ms_fb
:
1933 case nir_texop_txf_ms_mcs
:
1934 case nir_texop_samples_identical
:
1935 return nir_type_int
;
1938 return nir_type_float
;
1941 case nir_tex_src_lod
:
1942 switch (instr
->op
) {
1945 return nir_type_int
;
1948 return nir_type_float
;
1951 case nir_tex_src_projector
:
1952 case nir_tex_src_comparator
:
1953 case nir_tex_src_bias
:
1954 case nir_tex_src_min_lod
:
1955 case nir_tex_src_ddx
:
1956 case nir_tex_src_ddy
:
1957 return nir_type_float
;
1959 case nir_tex_src_offset
:
1960 case nir_tex_src_ms_index
:
1961 case nir_tex_src_plane
:
1962 return nir_type_int
;
1964 case nir_tex_src_ms_mcs
:
1965 case nir_tex_src_texture_deref
:
1966 case nir_tex_src_sampler_deref
:
1967 case nir_tex_src_texture_offset
:
1968 case nir_tex_src_sampler_offset
:
1969 case nir_tex_src_texture_handle
:
1970 case nir_tex_src_sampler_handle
:
1971 return nir_type_uint
;
1973 case nir_num_tex_src_types
:
1974 unreachable("nir_num_tex_src_types is not a valid source type");
1977 unreachable("Invalid texture source type");
1980 static inline unsigned
1981 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
1983 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
1984 return instr
->coord_components
;
1986 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
1987 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
1990 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
1991 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
1992 if (instr
->is_array
)
1993 return instr
->coord_components
- 1;
1995 return instr
->coord_components
;
1998 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
1999 * the offset, since a cube maps to a single face.
2001 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
2002 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2004 else if (instr
->is_array
)
2005 return instr
->coord_components
- 1;
2007 return instr
->coord_components
;
2014 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
2016 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
2017 if (instr
->src
[i
].src_type
== type
)
2023 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
2024 nir_tex_src_type src_type
,
2027 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
2029 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
);
2036 nir_const_value value
[];
2037 } nir_load_const_instr
;
2050 /* creates a new SSA variable in an undefined state */
2055 } nir_ssa_undef_instr
;
2058 struct exec_node node
;
2060 /* The predecessor block corresponding to this source */
2061 struct nir_block
*pred
;
2066 #define nir_foreach_phi_src(phi_src, phi) \
2067 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
2068 #define nir_foreach_phi_src_safe(phi_src, phi) \
2069 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
2074 struct exec_list srcs
; /** < list of nir_phi_src */
2080 struct exec_node node
;
2083 } nir_parallel_copy_entry
;
2085 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
2086 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
2091 /* A list of nir_parallel_copy_entrys. The sources of all of the
2092 * entries are copied to the corresponding destinations "in parallel".
2093 * In other words, if we have two entries: a -> b and b -> a, the values
2096 struct exec_list entries
;
2097 } nir_parallel_copy_instr
;
2099 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
2100 type
, nir_instr_type_alu
)
2101 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
2102 type
, nir_instr_type_deref
)
2103 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
2104 type
, nir_instr_type_call
)
2105 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
2106 type
, nir_instr_type_jump
)
2107 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
2108 type
, nir_instr_type_tex
)
2109 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
2110 type
, nir_instr_type_intrinsic
)
2111 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
2112 type
, nir_instr_type_load_const
)
2113 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
2114 type
, nir_instr_type_ssa_undef
)
2115 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
2116 type
, nir_instr_type_phi
)
2117 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
2118 nir_parallel_copy_instr
, instr
,
2119 type
, nir_instr_type_parallel_copy
)
2122 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
2123 static inline type \
2124 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2126 assert(nir_src_is_const(src)); \
2127 nir_load_const_instr *load = \
2128 nir_instr_as_load_const(src.ssa->parent_instr); \
2129 assert(comp < load->def.num_components); \
2130 return nir_const_value_as_##suffix(load->value[comp], \
2131 load->def.bit_size); \
2134 static inline type \
2135 nir_src_as_##suffix(nir_src src) \
2137 assert(nir_src_num_components(src) == 1); \
2138 return nir_src_comp_as_##suffix(src, 0); \
2141 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2142 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint
)
2143 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2144 NIR_DEFINE_SRC_AS_CONST(double, float)
2146 #undef NIR_DEFINE_SRC_AS_CONST
2155 nir_ssa_scalar_is_const(nir_ssa_scalar s
)
2157 return s
.def
->parent_instr
->type
== nir_instr_type_load_const
;
2160 static inline nir_const_value
2161 nir_ssa_scalar_as_const_value(nir_ssa_scalar s
)
2163 assert(s
.comp
< s
.def
->num_components
);
2164 nir_load_const_instr
*load
= nir_instr_as_load_const(s
.def
->parent_instr
);
2165 return load
->value
[s
.comp
];
2168 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2169 static inline type \
2170 nir_ssa_scalar_as_##suffix(nir_ssa_scalar s) \
2172 return nir_const_value_as_##suffix( \
2173 nir_ssa_scalar_as_const_value(s), s.def->bit_size); \
2176 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2177 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint
)
2178 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2179 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2181 #undef NIR_DEFINE_SCALAR_AS_CONST
2184 nir_ssa_scalar_is_alu(nir_ssa_scalar s
)
2186 return s
.def
->parent_instr
->type
== nir_instr_type_alu
;
2189 static inline nir_op
2190 nir_ssa_scalar_alu_op(nir_ssa_scalar s
)
2192 return nir_instr_as_alu(s
.def
->parent_instr
)->op
;
2195 static inline nir_ssa_scalar
2196 nir_ssa_scalar_chase_alu_src(nir_ssa_scalar s
, unsigned alu_src_idx
)
2198 nir_ssa_scalar out
= { NULL
, 0 };
2200 nir_alu_instr
*alu
= nir_instr_as_alu(s
.def
->parent_instr
);
2201 assert(alu_src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
2203 /* Our component must be written */
2204 assert(s
.comp
< s
.def
->num_components
);
2205 assert(alu
->dest
.write_mask
& (1u << s
.comp
));
2207 assert(alu
->src
[alu_src_idx
].src
.is_ssa
);
2208 out
.def
= alu
->src
[alu_src_idx
].src
.ssa
;
2210 if (nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 0) {
2211 /* The ALU src is unsized so the source component follows the
2212 * destination component.
2214 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[s
.comp
];
2216 /* This is a sized source so all source components work together to
2217 * produce all the destination components. Since we need to return a
2218 * scalar, this only works if the source is a scalar.
2220 assert(nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 1);
2221 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[0];
2223 assert(out
.comp
< out
.def
->num_components
);
2232 * Control flow consists of a tree of control flow nodes, which include
2233 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2234 * instructions that always run start-to-finish. Each basic block also keeps
2235 * track of its successors (blocks which may run immediately after the current
2236 * block) and predecessors (blocks which could have run immediately before the
2237 * current block). Each function also has a start block and an end block which
2238 * all return statements point to (which is always empty). Together, all the
2239 * blocks with their predecessors and successors make up the control flow
2240 * graph (CFG) of the function. There are helpers that modify the tree of
2241 * control flow nodes while modifying the CFG appropriately; these should be
2242 * used instead of modifying the tree directly.
2249 nir_cf_node_function
2252 typedef struct nir_cf_node
{
2253 struct exec_node node
;
2254 nir_cf_node_type type
;
2255 struct nir_cf_node
*parent
;
2258 typedef struct nir_block
{
2259 nir_cf_node cf_node
;
2261 struct exec_list instr_list
; /** < list of nir_instr */
2263 /** generic block index; generated by nir_index_blocks */
2267 * Each block can only have up to 2 successors, so we put them in a simple
2268 * array - no need for anything more complicated.
2270 struct nir_block
*successors
[2];
2272 /* Set of nir_block predecessors in the CFG */
2273 struct set
*predecessors
;
2276 * this node's immediate dominator in the dominance tree - set to NULL for
2279 struct nir_block
*imm_dom
;
2281 /* This node's children in the dominance tree */
2282 unsigned num_dom_children
;
2283 struct nir_block
**dom_children
;
2285 /* Set of nir_blocks on the dominance frontier of this block */
2286 struct set
*dom_frontier
;
2289 * These two indices have the property that dom_{pre,post}_index for each
2290 * child of this block in the dominance tree will always be between
2291 * dom_pre_index and dom_post_index for this block, which makes testing if
2292 * a given block is dominated by another block an O(1) operation.
2294 unsigned dom_pre_index
, dom_post_index
;
2296 /* live in and out for this block; used for liveness analysis */
2297 BITSET_WORD
*live_in
;
2298 BITSET_WORD
*live_out
;
2301 static inline nir_instr
*
2302 nir_block_first_instr(nir_block
*block
)
2304 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
2305 return exec_node_data(nir_instr
, head
, node
);
2308 static inline nir_instr
*
2309 nir_block_last_instr(nir_block
*block
)
2311 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
2312 return exec_node_data(nir_instr
, tail
, node
);
2316 nir_block_ends_in_jump(nir_block
*block
)
2318 return !exec_list_is_empty(&block
->instr_list
) &&
2319 nir_block_last_instr(block
)->type
== nir_instr_type_jump
;
2322 #define nir_foreach_instr(instr, block) \
2323 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
2324 #define nir_foreach_instr_reverse(instr, block) \
2325 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
2326 #define nir_foreach_instr_safe(instr, block) \
2327 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
2328 #define nir_foreach_instr_reverse_safe(instr, block) \
2329 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
2332 nir_selection_control_none
= 0x0,
2333 nir_selection_control_flatten
= 0x1,
2334 nir_selection_control_dont_flatten
= 0x2,
2335 } nir_selection_control
;
2337 typedef struct nir_if
{
2338 nir_cf_node cf_node
;
2340 nir_selection_control control
;
2342 struct exec_list then_list
; /** < list of nir_cf_node */
2343 struct exec_list else_list
; /** < list of nir_cf_node */
2349 /** Instruction that generates nif::condition. */
2350 nir_instr
*conditional_instr
;
2352 /** Block within ::nif that has the break instruction. */
2353 nir_block
*break_block
;
2355 /** Last block for the then- or else-path that does not contain the break. */
2356 nir_block
*continue_from_block
;
2358 /** True when ::break_block is in the else-path of ::nif. */
2359 bool continue_from_then
;
2362 /* This is true if the terminators exact trip count is unknown. For
2365 * for (int i = 0; i < imin(x, 4); i++)
2368 * Here loop analysis would have set a max_trip_count of 4 however we dont
2369 * know for sure that this is the exact trip count.
2371 bool exact_trip_count_unknown
;
2373 struct list_head loop_terminator_link
;
2374 } nir_loop_terminator
;
2377 /* Estimated cost (in number of instructions) of the loop */
2378 unsigned instr_cost
;
2380 /* Guessed trip count based on array indexing */
2381 unsigned guessed_trip_count
;
2383 /* Maximum number of times the loop is run (if known) */
2384 unsigned max_trip_count
;
2386 /* Do we know the exact number of times the loop will be run */
2387 bool exact_trip_count_known
;
2389 /* Unroll the loop regardless of its size */
2392 /* Does the loop contain complex loop terminators, continues or other
2393 * complex behaviours? If this is true we can't rely on
2394 * loop_terminator_list to be complete or accurate.
2398 nir_loop_terminator
*limiting_terminator
;
2400 /* A list of loop_terminators terminating this loop. */
2401 struct list_head loop_terminator_list
;
2405 nir_loop_control_none
= 0x0,
2406 nir_loop_control_unroll
= 0x1,
2407 nir_loop_control_dont_unroll
= 0x2,
2411 nir_cf_node cf_node
;
2413 struct exec_list body
; /** < list of nir_cf_node */
2415 nir_loop_info
*info
;
2416 nir_loop_control control
;
2417 bool partially_unrolled
;
2421 * Various bits of metadata that can may be created or required by
2422 * optimization and analysis passes
2425 nir_metadata_none
= 0x0,
2426 nir_metadata_block_index
= 0x1,
2427 nir_metadata_dominance
= 0x2,
2428 nir_metadata_live_ssa_defs
= 0x4,
2429 nir_metadata_not_properly_reset
= 0x8,
2430 nir_metadata_loop_analysis
= 0x10,
2434 nir_cf_node cf_node
;
2436 /** pointer to the function of which this is an implementation */
2437 struct nir_function
*function
;
2439 struct exec_list body
; /** < list of nir_cf_node */
2441 nir_block
*end_block
;
2443 /** list for all local variables in the function */
2444 struct exec_list locals
;
2446 /** list of local registers in the function */
2447 struct exec_list registers
;
2449 /** next available local register index */
2452 /** next available SSA value index */
2455 /* total number of basic blocks, only valid when block_index_dirty = false */
2456 unsigned num_blocks
;
2458 nir_metadata valid_metadata
;
2459 } nir_function_impl
;
2461 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2462 nir_start_block(nir_function_impl
*impl
)
2464 return (nir_block
*) impl
->body
.head_sentinel
.next
;
2467 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2468 nir_impl_last_block(nir_function_impl
*impl
)
2470 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
2473 static inline nir_cf_node
*
2474 nir_cf_node_next(nir_cf_node
*node
)
2476 struct exec_node
*next
= exec_node_get_next(&node
->node
);
2477 if (exec_node_is_tail_sentinel(next
))
2480 return exec_node_data(nir_cf_node
, next
, node
);
2483 static inline nir_cf_node
*
2484 nir_cf_node_prev(nir_cf_node
*node
)
2486 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
2487 if (exec_node_is_head_sentinel(prev
))
2490 return exec_node_data(nir_cf_node
, prev
, node
);
2494 nir_cf_node_is_first(const nir_cf_node
*node
)
2496 return exec_node_is_head_sentinel(node
->node
.prev
);
2500 nir_cf_node_is_last(const nir_cf_node
*node
)
2502 return exec_node_is_tail_sentinel(node
->node
.next
);
2505 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
2506 type
, nir_cf_node_block
)
2507 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
2508 type
, nir_cf_node_if
)
2509 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
2510 type
, nir_cf_node_loop
)
2511 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
2512 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
2514 static inline nir_block
*
2515 nir_if_first_then_block(nir_if
*if_stmt
)
2517 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
2518 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2521 static inline nir_block
*
2522 nir_if_last_then_block(nir_if
*if_stmt
)
2524 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
2525 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2528 static inline nir_block
*
2529 nir_if_first_else_block(nir_if
*if_stmt
)
2531 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
2532 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2535 static inline nir_block
*
2536 nir_if_last_else_block(nir_if
*if_stmt
)
2538 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
2539 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2542 static inline nir_block
*
2543 nir_loop_first_block(nir_loop
*loop
)
2545 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
2546 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2549 static inline nir_block
*
2550 nir_loop_last_block(nir_loop
*loop
)
2552 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
2553 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2557 * Return true if this list of cf_nodes contains a single empty block.
2560 nir_cf_list_is_empty_block(struct exec_list
*cf_list
)
2562 if (exec_list_is_singular(cf_list
)) {
2563 struct exec_node
*head
= exec_list_get_head(cf_list
);
2565 nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2566 return exec_list_is_empty(&block
->instr_list
);
2572 uint8_t num_components
;
2576 typedef struct nir_function
{
2577 struct exec_node node
;
2580 struct nir_shader
*shader
;
2582 unsigned num_params
;
2583 nir_parameter
*params
;
2585 /** The implementation of this function.
2587 * If the function is only declared and not implemented, this is NULL.
2589 nir_function_impl
*impl
;
2595 nir_lower_imul64
= (1 << 0),
2596 nir_lower_isign64
= (1 << 1),
2597 /** Lower all int64 modulus and division opcodes */
2598 nir_lower_divmod64
= (1 << 2),
2599 /** Lower all 64-bit umul_high and imul_high opcodes */
2600 nir_lower_imul_high64
= (1 << 3),
2601 nir_lower_mov64
= (1 << 4),
2602 nir_lower_icmp64
= (1 << 5),
2603 nir_lower_iadd64
= (1 << 6),
2604 nir_lower_iabs64
= (1 << 7),
2605 nir_lower_ineg64
= (1 << 8),
2606 nir_lower_logic64
= (1 << 9),
2607 nir_lower_minmax64
= (1 << 10),
2608 nir_lower_shift64
= (1 << 11),
2609 nir_lower_imul_2x32_64
= (1 << 12),
2610 nir_lower_extract64
= (1 << 13),
2611 } nir_lower_int64_options
;
2614 nir_lower_drcp
= (1 << 0),
2615 nir_lower_dsqrt
= (1 << 1),
2616 nir_lower_drsq
= (1 << 2),
2617 nir_lower_dtrunc
= (1 << 3),
2618 nir_lower_dfloor
= (1 << 4),
2619 nir_lower_dceil
= (1 << 5),
2620 nir_lower_dfract
= (1 << 6),
2621 nir_lower_dround_even
= (1 << 7),
2622 nir_lower_dmod
= (1 << 8),
2623 nir_lower_dsub
= (1 << 9),
2624 nir_lower_ddiv
= (1 << 10),
2625 nir_lower_fp64_full_software
= (1 << 11),
2626 } nir_lower_doubles_options
;
2629 nir_divergence_single_prim_per_subgroup
= (1 << 0),
2630 nir_divergence_single_patch_per_tcs_subgroup
= (1 << 1),
2631 nir_divergence_single_patch_per_tes_subgroup
= (1 << 2),
2632 nir_divergence_view_index_uniform
= (1 << 3),
2633 } nir_divergence_options
;
2635 typedef struct nir_shader_compiler_options
{
2641 /** Lowers flrp when it does not support doubles */
2648 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2649 bool lower_bitfield_extract
;
2650 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
2651 bool lower_bitfield_extract_to_shifts
;
2652 /** Lowers bitfield_insert to bfi/bfm */
2653 bool lower_bitfield_insert
;
2654 /** Lowers bitfield_insert to compares, and shifts. */
2655 bool lower_bitfield_insert_to_shifts
;
2656 /** Lowers bitfield_insert to bfm/bitfield_select. */
2657 bool lower_bitfield_insert_to_bitfield_select
;
2658 /** Lowers bitfield_reverse to shifts. */
2659 bool lower_bitfield_reverse
;
2660 /** Lowers bit_count to shifts. */
2661 bool lower_bit_count
;
2662 /** Lowers ifind_msb to compare and ufind_msb */
2663 bool lower_ifind_msb
;
2664 /** Lowers find_lsb to ufind_msb and logic ops */
2665 bool lower_find_lsb
;
2666 bool lower_uadd_carry
;
2667 bool lower_usub_borrow
;
2668 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2669 bool lower_mul_high
;
2670 /** lowers fneg and ineg to fsub and isub. */
2672 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2675 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2678 /* lower fall_equalN/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
2679 bool lower_vector_cmp
;
2681 /** enables rules to lower idiv by power-of-two: */
2684 /** enable rules to avoid bit ops */
2687 /** enables rules to lower isign to imin+imax */
2690 /** enables rules to lower fsign to fsub and flt */
2693 /* lower fdph to fdot4 */
2696 /** lower fdot to fmul and fsum/fadd. */
2699 /* Does the native fdot instruction replicate its result for four
2700 * components? If so, then opt_algebraic_late will turn all fdotN
2701 * instructions into fdot_replicatedN instructions.
2703 bool fdot_replicates
;
2705 /** lowers ffloor to fsub+ffract: */
2708 /** lowers ffract to fsub+ffloor: */
2711 /** lowers fceil to fneg+ffloor+fneg: */
2718 bool lower_pack_half_2x16
;
2719 bool lower_pack_unorm_2x16
;
2720 bool lower_pack_snorm_2x16
;
2721 bool lower_pack_unorm_4x8
;
2722 bool lower_pack_snorm_4x8
;
2723 bool lower_unpack_half_2x16
;
2724 bool lower_unpack_unorm_2x16
;
2725 bool lower_unpack_snorm_2x16
;
2726 bool lower_unpack_unorm_4x8
;
2727 bool lower_unpack_snorm_4x8
;
2729 bool lower_extract_byte
;
2730 bool lower_extract_word
;
2732 bool lower_all_io_to_temps
;
2733 bool lower_all_io_to_elements
;
2735 /* Indicates that the driver only has zero-based vertex id */
2736 bool vertex_id_zero_based
;
2739 * If enabled, gl_BaseVertex will be lowered as:
2740 * is_indexed_draw (~0/0) & firstvertex
2742 bool lower_base_vertex
;
2745 * If enabled, gl_HelperInvocation will be lowered as:
2747 * !((1 << sample_id) & sample_mask_in))
2749 * This depends on some possibly hw implementation details, which may
2750 * not be true for all hw. In particular that the FS is only executed
2751 * for covered samples or for helper invocations. So, do not blindly
2752 * enable this option.
2754 * Note: See also issue #22 in ARB_shader_image_load_store
2756 bool lower_helper_invocation
;
2759 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
2761 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
2762 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
2764 bool optimize_sample_mask_in
;
2766 bool lower_cs_local_index_from_id
;
2767 bool lower_cs_local_id_from_index
;
2769 bool lower_device_index_to_zero
;
2771 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
2772 bool lower_wpos_pntc
;
2778 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
2779 * for IO purposes and would prefer loads/stores be vectorized.
2782 bool lower_to_scalar
;
2785 * Should nir_lower_io() create load_interpolated_input intrinsics?
2787 * If not, it generates regular load_input intrinsics and interpolation
2788 * information must be inferred from the list of input nir_variables.
2790 bool use_interpolated_input_intrinsics
;
2792 /* Lowers when 32x32->64 bit multiplication is not supported */
2793 bool lower_mul_2x32_64
;
2795 /* Lowers when rotate instruction is not supported */
2799 * Backend supports imul24, and would like to use it (when possible)
2800 * for address/offset calculation. If true, driver should call
2801 * nir_lower_amul(). (If not set, amul will automatically be lowered
2807 * Is this the Intel vec4 backend?
2809 * Used to inhibit algebraic optimizations that are known to be harmful on
2810 * the Intel vec4 backend. This is generally applicable to any
2811 * optimization that might cause more immediate values to be used in
2812 * 3-source (e.g., ffma and flrp) instructions.
2816 unsigned max_unroll_iterations
;
2818 nir_lower_int64_options lower_int64_options
;
2819 nir_lower_doubles_options lower_doubles_options
;
2820 } nir_shader_compiler_options
;
2822 typedef struct nir_shader
{
2823 /** list of uniforms (nir_variable) */
2824 struct exec_list uniforms
;
2826 /** list of inputs (nir_variable) */
2827 struct exec_list inputs
;
2829 /** list of outputs (nir_variable) */
2830 struct exec_list outputs
;
2832 /** list of shared compute variables (nir_variable) */
2833 struct exec_list shared
;
2835 /** Set of driver-specific options for the shader.
2837 * The memory for the options is expected to be kept in a single static
2838 * copy by the driver.
2840 const struct nir_shader_compiler_options
*options
;
2842 /** Various bits of compile-time information about a given shader */
2843 struct shader_info info
;
2845 /** list of global variables in the shader (nir_variable) */
2846 struct exec_list globals
;
2848 /** list of system value variables in the shader (nir_variable) */
2849 struct exec_list system_values
;
2851 struct exec_list functions
; /** < list of nir_function */
2854 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
2857 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
2859 /** Size in bytes of required scratch space */
2860 unsigned scratch_size
;
2862 /** Constant data associated with this shader.
2864 * Constant data is loaded through load_constant intrinsics. See also
2865 * nir_opt_large_constants.
2867 void *constant_data
;
2868 unsigned constant_data_size
;
2871 #define nir_foreach_function(func, shader) \
2872 foreach_list_typed(nir_function, func, node, &(shader)->functions)
2874 static inline nir_function_impl
*
2875 nir_shader_get_entrypoint(nir_shader
*shader
)
2877 nir_function
*func
= NULL
;
2879 nir_foreach_function(function
, shader
) {
2880 assert(func
== NULL
);
2881 if (function
->is_entrypoint
) {
2892 assert(func
->num_params
== 0);
2897 nir_shader
*nir_shader_create(void *mem_ctx
,
2898 gl_shader_stage stage
,
2899 const nir_shader_compiler_options
*options
,
2902 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
2904 void nir_reg_remove(nir_register
*reg
);
2906 /** Adds a variable to the appropriate list in nir_shader */
2907 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
2910 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
2912 assert(var
->data
.mode
== nir_var_function_temp
);
2913 exec_list_push_tail(&impl
->locals
, &var
->node
);
2916 /** creates a variable, sets a few defaults, and adds it to the list */
2917 nir_variable
*nir_variable_create(nir_shader
*shader
,
2918 nir_variable_mode mode
,
2919 const struct glsl_type
*type
,
2921 /** creates a local variable and adds it to the list */
2922 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
2923 const struct glsl_type
*type
,
2926 /** creates a function and adds it to the shader's list of functions */
2927 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
2929 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
2930 /** creates a function_impl that isn't tied to any particular function */
2931 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
2933 nir_block
*nir_block_create(nir_shader
*shader
);
2934 nir_if
*nir_if_create(nir_shader
*shader
);
2935 nir_loop
*nir_loop_create(nir_shader
*shader
);
2937 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
2939 /** requests that the given pieces of metadata be generated */
2940 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
2941 /** dirties all but the preserved metadata */
2942 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
2944 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
2945 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
2947 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
2948 nir_deref_type deref_type
);
2950 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
2952 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
2953 unsigned num_components
,
2956 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
2957 nir_intrinsic_op op
);
2959 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
2960 nir_function
*callee
);
2962 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
2964 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
2966 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
2968 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
2969 unsigned num_components
,
2972 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
2975 * NIR Cursors and Instruction Insertion API
2978 * A tiny struct representing a point to insert/extract instructions or
2979 * control flow nodes. Helps reduce the combinatorial explosion of possible
2980 * points to insert/extract.
2982 * \sa nir_control_flow.h
2985 nir_cursor_before_block
,
2986 nir_cursor_after_block
,
2987 nir_cursor_before_instr
,
2988 nir_cursor_after_instr
,
2989 } nir_cursor_option
;
2992 nir_cursor_option option
;
2999 static inline nir_block
*
3000 nir_cursor_current_block(nir_cursor cursor
)
3002 if (cursor
.option
== nir_cursor_before_instr
||
3003 cursor
.option
== nir_cursor_after_instr
) {
3004 return cursor
.instr
->block
;
3006 return cursor
.block
;
3010 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
3012 static inline nir_cursor
3013 nir_before_block(nir_block
*block
)
3016 cursor
.option
= nir_cursor_before_block
;
3017 cursor
.block
= block
;
3021 static inline nir_cursor
3022 nir_after_block(nir_block
*block
)
3025 cursor
.option
= nir_cursor_after_block
;
3026 cursor
.block
= block
;
3030 static inline nir_cursor
3031 nir_before_instr(nir_instr
*instr
)
3034 cursor
.option
= nir_cursor_before_instr
;
3035 cursor
.instr
= instr
;
3039 static inline nir_cursor
3040 nir_after_instr(nir_instr
*instr
)
3043 cursor
.option
= nir_cursor_after_instr
;
3044 cursor
.instr
= instr
;
3048 static inline nir_cursor
3049 nir_after_block_before_jump(nir_block
*block
)
3051 nir_instr
*last_instr
= nir_block_last_instr(block
);
3052 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
3053 return nir_before_instr(last_instr
);
3055 return nir_after_block(block
);
3059 static inline nir_cursor
3060 nir_before_src(nir_src
*src
, bool is_if_condition
)
3062 if (is_if_condition
) {
3063 nir_block
*prev_block
=
3064 nir_cf_node_as_block(nir_cf_node_prev(&src
->parent_if
->cf_node
));
3065 assert(!nir_block_ends_in_jump(prev_block
));
3066 return nir_after_block(prev_block
);
3067 } else if (src
->parent_instr
->type
== nir_instr_type_phi
) {
3069 nir_phi_instr
*cond_phi
= nir_instr_as_phi(src
->parent_instr
);
3071 nir_foreach_phi_src(phi_src
, cond_phi
) {
3072 if (phi_src
->src
.ssa
== src
->ssa
) {
3079 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
3080 * to have a more specific name.
3082 nir_phi_src
*phi_src
= LIST_ENTRY(nir_phi_src
, src
, src
);
3083 return nir_after_block_before_jump(phi_src
->pred
);
3085 return nir_before_instr(src
->parent_instr
);
3089 static inline nir_cursor
3090 nir_before_cf_node(nir_cf_node
*node
)
3092 if (node
->type
== nir_cf_node_block
)
3093 return nir_before_block(nir_cf_node_as_block(node
));
3095 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
3098 static inline nir_cursor
3099 nir_after_cf_node(nir_cf_node
*node
)
3101 if (node
->type
== nir_cf_node_block
)
3102 return nir_after_block(nir_cf_node_as_block(node
));
3104 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
3107 static inline nir_cursor
3108 nir_after_phis(nir_block
*block
)
3110 nir_foreach_instr(instr
, block
) {
3111 if (instr
->type
!= nir_instr_type_phi
)
3112 return nir_before_instr(instr
);
3114 return nir_after_block(block
);
3117 static inline nir_cursor
3118 nir_after_cf_node_and_phis(nir_cf_node
*node
)
3120 if (node
->type
== nir_cf_node_block
)
3121 return nir_after_block(nir_cf_node_as_block(node
));
3123 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
3125 return nir_after_phis(block
);
3128 static inline nir_cursor
3129 nir_before_cf_list(struct exec_list
*cf_list
)
3131 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
3132 exec_list_get_head(cf_list
), node
);
3133 return nir_before_cf_node(first_node
);
3136 static inline nir_cursor
3137 nir_after_cf_list(struct exec_list
*cf_list
)
3139 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
3140 exec_list_get_tail(cf_list
), node
);
3141 return nir_after_cf_node(last_node
);
3145 * Insert a NIR instruction at the given cursor.
3147 * Note: This does not update the cursor.
3149 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
3152 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
3154 nir_instr_insert(nir_before_instr(instr
), before
);
3158 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
3160 nir_instr_insert(nir_after_instr(instr
), after
);
3164 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
3166 nir_instr_insert(nir_before_block(block
), before
);
3170 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
3172 nir_instr_insert(nir_after_block(block
), after
);
3176 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
3178 nir_instr_insert(nir_before_cf_node(node
), before
);
3182 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
3184 nir_instr_insert(nir_after_cf_node(node
), after
);
3188 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
3190 nir_instr_insert(nir_before_cf_list(list
), before
);
3194 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
3196 nir_instr_insert(nir_after_cf_list(list
), after
);
3199 void nir_instr_remove_v(nir_instr
*instr
);
3201 static inline nir_cursor
3202 nir_instr_remove(nir_instr
*instr
)
3205 nir_instr
*prev
= nir_instr_prev(instr
);
3207 cursor
= nir_after_instr(prev
);
3209 cursor
= nir_before_block(instr
->block
);
3211 nir_instr_remove_v(instr
);
3217 nir_ssa_def
*nir_instr_ssa_def(nir_instr
*instr
);
3219 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
3220 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
3221 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
3222 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
3224 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
3225 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
3227 nir_const_value
*nir_src_as_const_value(nir_src src
);
3229 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
3230 static inline c_type * \
3231 nir_src_as_ ## name (nir_src src) \
3233 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
3234 ? cast_macro(src.ssa->parent_instr) : NULL; \
3237 NIR_SRC_AS_(alu_instr
, nir_alu_instr
, nir_instr_type_alu
, nir_instr_as_alu
)
3238 NIR_SRC_AS_(intrinsic
, nir_intrinsic_instr
,
3239 nir_instr_type_intrinsic
, nir_instr_as_intrinsic
)
3240 NIR_SRC_AS_(deref
, nir_deref_instr
, nir_instr_type_deref
, nir_instr_as_deref
)
3242 bool nir_src_is_dynamically_uniform(nir_src src
);
3243 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
3244 bool nir_instrs_equal(const nir_instr
*instr1
, const nir_instr
*instr2
);
3245 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
3246 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
3247 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
3248 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
3251 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
3252 unsigned num_components
, unsigned bit_size
,
3254 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
3255 unsigned num_components
, unsigned bit_size
,
3258 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
3259 const struct glsl_type
*type
,
3262 assert(glsl_type_is_vector_or_scalar(type
));
3263 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
3264 glsl_get_bit_size(type
), name
);
3266 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
3267 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
3268 nir_instr
*after_me
);
3270 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
3273 * finds the next basic block in source-code order, returns NULL if there is
3277 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
3279 /* Performs the opposite of nir_block_cf_tree_next() */
3281 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
3283 /* Gets the first block in a CF node in source-code order */
3285 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
3287 /* Gets the last block in a CF node in source-code order */
3289 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
3291 /* Gets the next block after a CF node in source-code order */
3293 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
3295 /* Macros for loops that visit blocks in source-code order */
3297 #define nir_foreach_block(block, impl) \
3298 for (nir_block *block = nir_start_block(impl); block != NULL; \
3299 block = nir_block_cf_tree_next(block))
3301 #define nir_foreach_block_safe(block, impl) \
3302 for (nir_block *block = nir_start_block(impl), \
3303 *next = nir_block_cf_tree_next(block); \
3305 block = next, next = nir_block_cf_tree_next(block))
3307 #define nir_foreach_block_reverse(block, impl) \
3308 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
3309 block = nir_block_cf_tree_prev(block))
3311 #define nir_foreach_block_reverse_safe(block, impl) \
3312 for (nir_block *block = nir_impl_last_block(impl), \
3313 *prev = nir_block_cf_tree_prev(block); \
3315 block = prev, prev = nir_block_cf_tree_prev(block))
3317 #define nir_foreach_block_in_cf_node(block, node) \
3318 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
3319 block != nir_cf_node_cf_tree_next(node); \
3320 block = nir_block_cf_tree_next(block))
3322 /* If the following CF node is an if, this function returns that if.
3323 * Otherwise, it returns NULL.
3325 nir_if
*nir_block_get_following_if(nir_block
*block
);
3327 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
3329 void nir_index_local_regs(nir_function_impl
*impl
);
3330 void nir_index_ssa_defs(nir_function_impl
*impl
);
3331 unsigned nir_index_instrs(nir_function_impl
*impl
);
3333 void nir_index_blocks(nir_function_impl
*impl
);
3335 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
3336 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
3337 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
3338 void nir_print_deref(const nir_deref_instr
*deref
, FILE *fp
);
3340 /** Shallow clone of a single ALU instruction. */
3341 nir_alu_instr
*nir_alu_instr_clone(nir_shader
*s
, const nir_alu_instr
*orig
);
3343 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
3344 nir_function_impl
*nir_function_impl_clone(nir_shader
*shader
,
3345 const nir_function_impl
*fi
);
3346 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
3347 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
3349 void nir_shader_replace(nir_shader
*dest
, nir_shader
*src
);
3351 void nir_shader_serialize_deserialize(nir_shader
*s
);
3354 void nir_validate_shader(nir_shader
*shader
, const char *when
);
3355 void nir_metadata_set_validation_flag(nir_shader
*shader
);
3356 void nir_metadata_check_validation_flag(nir_shader
*shader
);
3359 should_skip_nir(const char *name
)
3361 static const char *list
= NULL
;
3363 /* Comma separated list of names to skip. */
3364 list
= getenv("NIR_SKIP");
3372 return comma_separated_list_contains(list
, name
);
3376 should_clone_nir(void)
3378 static int should_clone
= -1;
3379 if (should_clone
< 0)
3380 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
3382 return should_clone
;
3386 should_serialize_deserialize_nir(void)
3388 static int test_serialize
= -1;
3389 if (test_serialize
< 0)
3390 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
3392 return test_serialize
;
3396 should_print_nir(void)
3398 static int should_print
= -1;
3399 if (should_print
< 0)
3400 should_print
= env_var_as_boolean("NIR_PRINT", false);
3402 return should_print
;
3405 static inline void nir_validate_shader(nir_shader
*shader
, const char *when
) { (void) shader
; (void)when
; }
3406 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
3407 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
3408 static inline bool should_skip_nir(UNUSED
const char *pass_name
) { return false; }
3409 static inline bool should_clone_nir(void) { return false; }
3410 static inline bool should_serialize_deserialize_nir(void) { return false; }
3411 static inline bool should_print_nir(void) { return false; }
3414 #define _PASS(pass, nir, do_pass) do { \
3415 if (should_skip_nir(#pass)) { \
3416 printf("skipping %s\n", #pass); \
3420 nir_validate_shader(nir, "after " #pass); \
3421 if (should_clone_nir()) { \
3422 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3423 nir_shader_replace(nir, clone); \
3425 if (should_serialize_deserialize_nir()) { \
3426 nir_shader_serialize_deserialize(nir); \
3430 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3431 nir_metadata_set_validation_flag(nir); \
3432 if (should_print_nir()) \
3433 printf("%s\n", #pass); \
3434 if (pass(nir, ##__VA_ARGS__)) { \
3436 if (should_print_nir()) \
3437 nir_print_shader(nir, stdout); \
3438 nir_metadata_check_validation_flag(nir); \
3442 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3443 if (should_print_nir()) \
3444 printf("%s\n", #pass); \
3445 pass(nir, ##__VA_ARGS__); \
3446 if (should_print_nir()) \
3447 nir_print_shader(nir, stdout); \
3450 #define NIR_SKIP(name) should_skip_nir(#name)
3452 /** An instruction filtering callback
3454 * Returns true if the instruction should be processed and false otherwise.
3456 typedef bool (*nir_instr_filter_cb
)(const nir_instr
*, const void *);
3458 /** A simple instruction lowering callback
3460 * Many instruction lowering passes can be written as a simple function which
3461 * takes an instruction as its input and returns a sequence of instructions
3462 * that implement the consumed instruction. This function type represents
3463 * such a lowering function. When called, a function with this prototype
3464 * should either return NULL indicating that no lowering needs to be done or
3465 * emit a sequence of instructions using the provided builder (whose cursor
3466 * will already be placed after the instruction to be lowered) and return the
3467 * resulting nir_ssa_def.
3469 typedef nir_ssa_def
*(*nir_lower_instr_cb
)(struct nir_builder
*,
3470 nir_instr
*, void *);
3473 * Special return value for nir_lower_instr_cb when some progress occurred
3474 * (like changing an input to the instr) that didn't result in a replacement
3475 * SSA def being generated.
3477 #define NIR_LOWER_INSTR_PROGRESS ((nir_ssa_def *)(uintptr_t)1)
3479 /** Iterate over all the instructions in a nir_function_impl and lower them
3480 * using the provided callbacks
3482 * This function implements the guts of a standard lowering pass for you. It
3483 * iterates over all of the instructions in a nir_function_impl and calls the
3484 * filter callback on each one. If the filter callback returns true, it then
3485 * calls the lowering call back on the instruction. (Splitting it this way
3486 * allows us to avoid some save/restore work for instructions we know won't be
3487 * lowered.) If the instruction is dead after the lowering is complete, it
3488 * will be removed. If new instructions are added, the lowering callback will
3489 * also be called on them in case multiple lowerings are required.
3491 * The metadata for the nir_function_impl will also be updated. If any blocks
3492 * are added (they cannot be removed), dominance and block indices will be
3495 bool nir_function_impl_lower_instructions(nir_function_impl
*impl
,
3496 nir_instr_filter_cb filter
,
3497 nir_lower_instr_cb lower
,
3499 bool nir_shader_lower_instructions(nir_shader
*shader
,
3500 nir_instr_filter_cb filter
,
3501 nir_lower_instr_cb lower
,
3504 void nir_calc_dominance_impl(nir_function_impl
*impl
);
3505 void nir_calc_dominance(nir_shader
*shader
);
3507 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
3508 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
3509 bool nir_block_is_unreachable(nir_block
*block
);
3511 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
3512 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
3514 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
3515 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
3517 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
3518 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
3520 int nir_gs_count_vertices(const nir_shader
*shader
);
3522 bool nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3523 bool nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3524 bool nir_split_var_copies(nir_shader
*shader
);
3525 bool nir_split_per_member_structs(nir_shader
*shader
);
3526 bool nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
);
3528 bool nir_lower_returns_impl(nir_function_impl
*impl
);
3529 bool nir_lower_returns(nir_shader
*shader
);
3531 void nir_inline_function_impl(struct nir_builder
*b
,
3532 const nir_function_impl
*impl
,
3533 nir_ssa_def
**params
);
3534 bool nir_inline_functions(nir_shader
*shader
);
3536 bool nir_propagate_invariant(nir_shader
*shader
);
3538 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
3539 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
3540 nir_intrinsic_instr
*copy
);
3541 bool nir_lower_var_copies(nir_shader
*shader
);
3543 void nir_fixup_deref_modes(nir_shader
*shader
);
3545 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
3548 nir_lower_direct_array_deref_of_vec_load
= (1 << 0),
3549 nir_lower_indirect_array_deref_of_vec_load
= (1 << 1),
3550 nir_lower_direct_array_deref_of_vec_store
= (1 << 2),
3551 nir_lower_indirect_array_deref_of_vec_store
= (1 << 3),
3552 } nir_lower_array_deref_of_vec_options
;
3554 bool nir_lower_array_deref_of_vec(nir_shader
*shader
, nir_variable_mode modes
,
3555 nir_lower_array_deref_of_vec_options options
);
3557 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
3559 bool nir_lower_locals_to_regs(nir_shader
*shader
);
3561 void nir_lower_io_to_temporaries(nir_shader
*shader
,
3562 nir_function_impl
*entrypoint
,
3563 bool outputs
, bool inputs
);
3565 bool nir_lower_vars_to_scratch(nir_shader
*shader
,
3566 nir_variable_mode modes
,
3568 glsl_type_size_align_func size_align
);
3570 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
3572 void nir_gather_ssa_types(nir_function_impl
*impl
,
3573 BITSET_WORD
*float_types
,
3574 BITSET_WORD
*int_types
);
3576 void nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
3577 int (*type_size
)(const struct glsl_type
*, bool));
3579 /* Some helpers to do very simple linking */
3580 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3581 bool nir_remove_unused_io_vars(nir_shader
*shader
, struct exec_list
*var_list
,
3582 uint64_t *used_by_other_stage
,
3583 uint64_t *used_by_other_stage_patches
);
3584 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
3585 bool default_to_smooth_interp
);
3586 void nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3587 bool nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3589 bool nir_lower_amul(nir_shader
*shader
,
3590 int (*type_size
)(const struct glsl_type
*, bool));
3592 void nir_assign_io_var_locations(struct exec_list
*var_list
,
3594 gl_shader_stage stage
);
3597 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
3598 * to 32-bit operations. This is only valid for nir_var_shader_in/out
3601 nir_lower_io_lower_64bit_to_32
= (1 << 0),
3603 /* If set, this forces all non-flat fragment shader inputs to be
3604 * interpolated as if with the "sample" qualifier. This requires
3605 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
3607 nir_lower_io_force_sample_interpolation
= (1 << 1),
3608 } nir_lower_io_options
;
3609 bool nir_lower_io(nir_shader
*shader
,
3610 nir_variable_mode modes
,
3611 int (*type_size
)(const struct glsl_type
*, bool),
3612 nir_lower_io_options
);
3614 bool nir_io_add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
);
3617 nir_lower_vars_to_explicit_types(nir_shader
*shader
,
3618 nir_variable_mode modes
,
3619 glsl_type_size_align_func type_info
);
3623 * An address format which is a simple 32-bit global GPU address.
3625 nir_address_format_32bit_global
,
3628 * An address format which is a simple 64-bit global GPU address.
3630 nir_address_format_64bit_global
,
3633 * An address format which is a bounds-checked 64-bit global GPU address.
3635 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
3636 * address stored with the low bits in .x and high bits in .y, .z is a
3637 * size, and .w is an offset. When the final I/O operation is lowered, .w
3638 * is checked against .z and the operation is predicated on the result.
3640 nir_address_format_64bit_bounded_global
,
3643 * An address format which is comprised of a vec2 where the first
3644 * component is a buffer index and the second is an offset.
3646 nir_address_format_32bit_index_offset
,
3649 * An address format which is a simple 32-bit offset.
3651 nir_address_format_32bit_offset
,
3654 * An address format representing a purely logical addressing model. In
3655 * this model, all deref chains must be complete from the dereference
3656 * operation to the variable. Cast derefs are not allowed. These
3657 * addresses will be 32-bit scalars but the format is immaterial because
3658 * you can always chase the chain.
3660 nir_address_format_logical
,
3661 } nir_address_format
;
3663 static inline unsigned
3664 nir_address_format_bit_size(nir_address_format addr_format
)
3666 switch (addr_format
) {
3667 case nir_address_format_32bit_global
: return 32;
3668 case nir_address_format_64bit_global
: return 64;
3669 case nir_address_format_64bit_bounded_global
: return 32;
3670 case nir_address_format_32bit_index_offset
: return 32;
3671 case nir_address_format_32bit_offset
: return 32;
3672 case nir_address_format_logical
: return 32;
3674 unreachable("Invalid address format");
3677 static inline unsigned
3678 nir_address_format_num_components(nir_address_format addr_format
)
3680 switch (addr_format
) {
3681 case nir_address_format_32bit_global
: return 1;
3682 case nir_address_format_64bit_global
: return 1;
3683 case nir_address_format_64bit_bounded_global
: return 4;
3684 case nir_address_format_32bit_index_offset
: return 2;
3685 case nir_address_format_32bit_offset
: return 1;
3686 case nir_address_format_logical
: return 1;
3688 unreachable("Invalid address format");
3691 static inline const struct glsl_type
*
3692 nir_address_format_to_glsl_type(nir_address_format addr_format
)
3694 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
3695 assert(bit_size
== 32 || bit_size
== 64);
3696 return glsl_vector_type(bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
,
3697 nir_address_format_num_components(addr_format
));
3700 const nir_const_value
*nir_address_format_null_value(nir_address_format addr_format
);
3702 nir_ssa_def
*nir_build_addr_ieq(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3703 nir_address_format addr_format
);
3705 nir_ssa_def
*nir_build_addr_isub(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3706 nir_address_format addr_format
);
3708 nir_ssa_def
* nir_explicit_io_address_from_deref(struct nir_builder
*b
,
3709 nir_deref_instr
*deref
,
3710 nir_ssa_def
*base_addr
,
3711 nir_address_format addr_format
);
3712 void nir_lower_explicit_io_instr(struct nir_builder
*b
,
3713 nir_intrinsic_instr
*io_instr
,
3715 nir_address_format addr_format
);
3717 bool nir_lower_explicit_io(nir_shader
*shader
,
3718 nir_variable_mode modes
,
3719 nir_address_format
);
3721 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
3722 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
3724 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
3726 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
3727 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
3728 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
3730 bool nir_remove_dead_derefs(nir_shader
*shader
);
3731 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
3732 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
);
3733 bool nir_lower_constant_initializers(nir_shader
*shader
,
3734 nir_variable_mode modes
);
3736 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
3737 bool nir_lower_vec_to_movs(nir_shader
*shader
);
3738 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
3740 const gl_state_index16
*alpha_ref_state_tokens
);
3741 bool nir_lower_alu(nir_shader
*shader
);
3743 bool nir_lower_flrp(nir_shader
*shader
, unsigned lowering_mask
,
3744 bool always_precise
, bool have_ffma
);
3746 bool nir_lower_alu_to_scalar(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
3747 bool nir_lower_bool_to_float(nir_shader
*shader
);
3748 bool nir_lower_bool_to_int32(nir_shader
*shader
);
3749 bool nir_lower_int_to_float(nir_shader
*shader
);
3750 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
3751 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
3752 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
3753 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
3754 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
3756 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
3757 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
3758 bool nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode mask
);
3760 void nir_lower_fragcoord_wtrans(nir_shader
*shader
);
3761 void nir_lower_viewport_transform(nir_shader
*shader
);
3762 bool nir_lower_uniforms_to_ubo(nir_shader
*shader
, int multiplier
);
3764 typedef struct nir_lower_subgroups_options
{
3765 uint8_t subgroup_size
;
3766 uint8_t ballot_bit_size
;
3767 bool lower_to_scalar
:1;
3768 bool lower_vote_trivial
:1;
3769 bool lower_vote_eq_to_ballot
:1;
3770 bool lower_subgroup_masks
:1;
3771 bool lower_shuffle
:1;
3772 bool lower_shuffle_to_32bit
:1;
3774 } nir_lower_subgroups_options
;
3776 bool nir_lower_subgroups(nir_shader
*shader
,
3777 const nir_lower_subgroups_options
*options
);
3779 bool nir_lower_system_values(nir_shader
*shader
);
3781 enum PACKED nir_lower_tex_packing
{
3782 nir_lower_tex_packing_none
= 0,
3783 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
3784 * or unsigned ints based on the sampler type
3786 nir_lower_tex_packing_16
,
3787 /* The sampler returns 1 32-bit word of 4x8 unorm */
3788 nir_lower_tex_packing_8
,
3791 typedef struct nir_lower_tex_options
{
3793 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
3794 * sampler types a texture projector is lowered.
3799 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
3801 bool lower_txf_offset
;
3804 * If true, lower away nir_tex_src_offset for all rect textures.
3806 bool lower_rect_offset
;
3809 * If true, lower rect textures to 2D, using txs to fetch the
3810 * texture dimensions and dividing the texture coords by the
3811 * texture dims to normalize.
3816 * If true, convert yuv to rgb.
3818 unsigned lower_y_uv_external
;
3819 unsigned lower_y_u_v_external
;
3820 unsigned lower_yx_xuxv_external
;
3821 unsigned lower_xy_uxvx_external
;
3822 unsigned lower_ayuv_external
;
3823 unsigned lower_xyuv_external
;
3826 * To emulate certain texture wrap modes, this can be used
3827 * to saturate the specified tex coord to [0.0, 1.0]. The
3828 * bits are according to sampler #, ie. if, for example:
3830 * (conf->saturate_s & (1 << n))
3832 * is true, then the s coord for sampler n is saturated.
3834 * Note that clamping must happen *after* projector lowering
3835 * so any projected texture sample instruction with a clamped
3836 * coordinate gets automatically lowered, regardless of the
3837 * 'lower_txp' setting.
3839 unsigned saturate_s
;
3840 unsigned saturate_t
;
3841 unsigned saturate_r
;
3843 /* Bitmask of textures that need swizzling.
3845 * If (swizzle_result & (1 << texture_index)), then the swizzle in
3846 * swizzles[texture_index] is applied to the result of the texturing
3849 unsigned swizzle_result
;
3851 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
3852 * while 4 and 5 represent 0 and 1 respectively.
3854 uint8_t swizzles
[32][4];
3856 /* Can be used to scale sampled values in range required by the format. */
3857 float scale_factors
[32];
3860 * Bitmap of textures that need srgb to linear conversion. If
3861 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
3862 * of the texture are lowered to linear.
3864 unsigned lower_srgb
;
3867 * If true, lower nir_texop_tex on shaders that doesn't support implicit
3868 * LODs to nir_texop_txl.
3870 bool lower_tex_without_implicit_lod
;
3873 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
3875 bool lower_txd_cube_map
;
3878 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
3883 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
3884 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
3885 * with lower_txd_cube_map.
3887 bool lower_txd_shadow
;
3890 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
3891 * Implies lower_txd_cube_map and lower_txd_shadow.
3896 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
3897 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
3899 bool lower_txb_shadow_clamp
;
3902 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
3903 * with nir_texop_txl. This includes cube maps.
3905 bool lower_txd_shadow_clamp
;
3908 * If true, lower nir_texop_txd on when it uses both offset and min_lod
3909 * with nir_texop_txl. This includes cube maps.
3911 bool lower_txd_offset_clamp
;
3914 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
3915 * sampler is bindless.
3917 bool lower_txd_clamp_bindless_sampler
;
3920 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
3921 * sampler index is not statically determinable to be less than 16.
3923 bool lower_txd_clamp_if_sampler_index_not_lt_16
;
3926 * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
3927 * 0-lod followed by a nir_ishr.
3932 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
3933 * mixed-up tg4 locations.
3935 bool lower_tg4_broadcom_swizzle
;
3938 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
3940 bool lower_tg4_offsets
;
3942 enum nir_lower_tex_packing lower_tex_packing
[32];
3943 } nir_lower_tex_options
;
3945 bool nir_lower_tex(nir_shader
*shader
,
3946 const nir_lower_tex_options
*options
);
3948 enum nir_lower_non_uniform_access_type
{
3949 nir_lower_non_uniform_ubo_access
= (1 << 0),
3950 nir_lower_non_uniform_ssbo_access
= (1 << 1),
3951 nir_lower_non_uniform_texture_access
= (1 << 2),
3952 nir_lower_non_uniform_image_access
= (1 << 3),
3955 bool nir_lower_non_uniform_access(nir_shader
*shader
,
3956 enum nir_lower_non_uniform_access_type
);
3958 enum nir_lower_idiv_path
{
3959 /* This path is based on NV50LegalizeSSA::handleDIV(). It is the faster of
3960 * the two but it is not exact in some cases (for example, 1091317713u /
3961 * 1034u gives 5209173 instead of 1055432) */
3962 nir_lower_idiv_fast
,
3963 /* This path is based on AMDGPUTargetLowering::LowerUDIVREM() and
3964 * AMDGPUTargetLowering::LowerSDIVREM(). It requires more instructions than
3965 * the nv50 path and many of them are integer multiplications, so it is
3966 * probably slower. It should always return the correct result, though. */
3967 nir_lower_idiv_precise
,
3970 bool nir_lower_idiv(nir_shader
*shader
, enum nir_lower_idiv_path path
);
3972 bool nir_lower_input_attachments(nir_shader
*shader
, bool use_fragcoord_sysval
);
3974 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
,
3976 bool use_clipdist_array
,
3977 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
3978 bool nir_lower_clip_gs(nir_shader
*shader
, unsigned ucp_enables
,
3979 bool use_clipdist_array
,
3980 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
3981 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
,
3982 bool use_clipdist_array
);
3983 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
3985 void nir_lower_point_size_mov(nir_shader
*shader
,
3986 const gl_state_index16
*pointsize_state_tokens
);
3988 bool nir_lower_frexp(nir_shader
*nir
);
3990 void nir_lower_two_sided_color(nir_shader
*shader
);
3992 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
3994 bool nir_lower_flatshade(nir_shader
*shader
);
3996 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
3997 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
3998 const gl_state_index16
*uniform_state_tokens
);
4000 typedef struct nir_lower_wpos_ytransform_options
{
4001 gl_state_index16 state_tokens
[STATE_LENGTH
];
4002 bool fs_coord_origin_upper_left
:1;
4003 bool fs_coord_origin_lower_left
:1;
4004 bool fs_coord_pixel_center_integer
:1;
4005 bool fs_coord_pixel_center_half_integer
:1;
4006 } nir_lower_wpos_ytransform_options
;
4008 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
4009 const nir_lower_wpos_ytransform_options
*options
);
4010 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
4012 bool nir_lower_fb_read(nir_shader
*shader
);
4014 typedef struct nir_lower_drawpixels_options
{
4015 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
4016 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
4017 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
4018 unsigned drawpix_sampler
;
4019 unsigned pixelmap_sampler
;
4021 bool scale_and_bias
:1;
4022 } nir_lower_drawpixels_options
;
4024 void nir_lower_drawpixels(nir_shader
*shader
,
4025 const nir_lower_drawpixels_options
*options
);
4027 typedef struct nir_lower_bitmap_options
{
4030 } nir_lower_bitmap_options
;
4032 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
4034 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
, unsigned ssbo_offset
);
4037 nir_lower_int_source_mods
= 1 << 0,
4038 nir_lower_float_source_mods
= 1 << 1,
4039 nir_lower_triop_abs
= 1 << 2,
4040 nir_lower_all_source_mods
= (1 << 3) - 1
4041 } nir_lower_to_source_mods_flags
;
4044 bool nir_lower_to_source_mods(nir_shader
*shader
, nir_lower_to_source_mods_flags options
);
4046 bool nir_lower_gs_intrinsics(nir_shader
*shader
);
4048 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
4050 bool nir_lower_bit_size(nir_shader
*shader
,
4051 nir_lower_bit_size_callback callback
,
4052 void *callback_data
);
4054 nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode
);
4055 bool nir_lower_int64(nir_shader
*shader
, nir_lower_int64_options options
);
4057 nir_lower_doubles_options
nir_lower_doubles_op_to_options_mask(nir_op opcode
);
4058 bool nir_lower_doubles(nir_shader
*shader
, const nir_shader
*softfp64
,
4059 nir_lower_doubles_options options
);
4060 bool nir_lower_pack(nir_shader
*shader
);
4062 bool nir_lower_point_size(nir_shader
*shader
, float min
, float max
);
4065 nir_lower_interpolation_at_sample
= (1 << 1),
4066 nir_lower_interpolation_at_offset
= (1 << 2),
4067 nir_lower_interpolation_centroid
= (1 << 3),
4068 nir_lower_interpolation_pixel
= (1 << 4),
4069 nir_lower_interpolation_sample
= (1 << 5),
4070 } nir_lower_interpolation_options
;
4072 bool nir_lower_interpolation(nir_shader
*shader
,
4073 nir_lower_interpolation_options options
);
4075 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
4077 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
4079 void nir_loop_analyze_impl(nir_function_impl
*impl
,
4080 nir_variable_mode indirect_mask
);
4082 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
4084 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
4085 bool nir_repair_ssa(nir_shader
*shader
);
4087 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
4088 bool nir_convert_to_lcssa(nir_shader
*shader
, bool skip_invariants
, bool skip_bool_invariants
);
4089 bool* nir_divergence_analysis(nir_shader
*shader
, nir_divergence_options options
);
4091 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
4092 * registers. If false, convert all values (even those not involved in a phi
4093 * node) to registers.
4095 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
4097 bool nir_lower_phis_to_regs_block(nir_block
*block
);
4098 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
4099 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
);
4101 bool nir_lower_samplers(nir_shader
*shader
);
4103 /* This is here for unit tests. */
4104 bool nir_opt_comparison_pre_impl(nir_function_impl
*impl
);
4106 bool nir_opt_comparison_pre(nir_shader
*shader
);
4108 bool nir_opt_access(nir_shader
*shader
);
4109 bool nir_opt_algebraic(nir_shader
*shader
);
4110 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
4111 bool nir_opt_algebraic_late(nir_shader
*shader
);
4112 bool nir_opt_constant_folding(nir_shader
*shader
);
4114 bool nir_opt_combine_stores(nir_shader
*shader
, nir_variable_mode modes
);
4116 bool nir_copy_prop(nir_shader
*shader
);
4118 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
4120 bool nir_opt_cse(nir_shader
*shader
);
4122 bool nir_opt_dce(nir_shader
*shader
);
4124 bool nir_opt_dead_cf(nir_shader
*shader
);
4126 bool nir_opt_dead_write_vars(nir_shader
*shader
);
4128 bool nir_opt_deref_impl(nir_function_impl
*impl
);
4129 bool nir_opt_deref(nir_shader
*shader
);
4131 bool nir_opt_find_array_copies(nir_shader
*shader
);
4133 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
4135 bool nir_opt_idiv_const(nir_shader
*shader
, unsigned min_bit_size
);
4137 bool nir_opt_if(nir_shader
*shader
, bool aggressive_last_continue
);
4139 bool nir_opt_intrinsics(nir_shader
*shader
);
4141 bool nir_opt_large_constants(nir_shader
*shader
,
4142 glsl_type_size_align_func size_align
,
4143 unsigned threshold
);
4145 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
4148 nir_move_const_undef
= (1 << 0),
4149 nir_move_load_ubo
= (1 << 1),
4150 nir_move_load_input
= (1 << 2),
4151 nir_move_comparisons
= (1 << 3),
4154 bool nir_can_move_instr(nir_instr
*instr
, nir_move_options options
);
4156 bool nir_opt_sink(nir_shader
*shader
, nir_move_options options
);
4158 bool nir_opt_move(nir_shader
*shader
, nir_move_options options
);
4160 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
,
4161 bool indirect_load_ok
, bool expensive_alu_ok
);
4163 bool nir_opt_rematerialize_compares(nir_shader
*shader
);
4165 bool nir_opt_remove_phis(nir_shader
*shader
);
4166 bool nir_opt_remove_phis_block(nir_block
*block
);
4168 bool nir_opt_shrink_load(nir_shader
*shader
);
4170 bool nir_opt_trivial_continues(nir_shader
*shader
);
4172 bool nir_opt_undef(nir_shader
*shader
);
4174 bool nir_opt_vectorize(nir_shader
*shader
);
4176 bool nir_opt_conditional_discard(nir_shader
*shader
);
4178 void nir_strip(nir_shader
*shader
);
4180 void nir_sweep(nir_shader
*shader
);
4182 void nir_remap_dual_slot_attributes(nir_shader
*shader
,
4183 uint64_t *dual_slot_inputs
);
4184 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
);
4186 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
4187 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);
4190 nir_variable_is_in_ubo(const nir_variable
*var
)
4192 return (var
->data
.mode
== nir_var_mem_ubo
&&
4193 var
->interface_type
!= NULL
);
4197 nir_variable_is_in_ssbo(const nir_variable
*var
)
4199 return (var
->data
.mode
== nir_var_mem_ssbo
&&
4200 var
->interface_type
!= NULL
);
4204 nir_variable_is_in_block(const nir_variable
*var
)
4206 return nir_variable_is_in_ubo(var
) || nir_variable_is_in_ssbo(var
);