2 * Copyright © 2014 Connor Abbott
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Connor Abbott (cwabbott0@gmail.com)
31 #include "util/hash_table.h"
32 #include "compiler/glsl/list.h"
33 #include "GL/gl.h" /* GLenum */
34 #include "util/list.h"
35 #include "util/ralloc.h"
37 #include "util/bitscan.h"
38 #include "util/bitset.h"
39 #include "util/macros.h"
40 #include "compiler/nir_types.h"
41 #include "compiler/shader_enums.h"
42 #include "compiler/shader_info.h"
46 #include "util/debug.h"
49 #include "nir_opcodes.h"
51 #if defined(_WIN32) && !defined(snprintf)
52 #define snprintf _snprintf
60 #define NIR_TRUE (~0u)
61 #define NIR_MAX_VEC_COMPONENTS 4
62 #define NIR_MAX_MATRIX_COLUMNS 4
63 #define NIR_STREAM_PACKED (1 << 8)
64 typedef uint8_t nir_component_mask_t
;
66 /** Defines a cast function
68 * This macro defines a cast function from in_type to out_type where
69 * out_type is some structure type that contains a field of type out_type.
71 * Note that you have to be a bit careful as the generated cast function
74 #define NIR_DEFINE_CAST(name, in_type, out_type, field, \
75 type_field, type_value) \
76 static inline out_type * \
77 name(const in_type *parent) \
79 assert(parent && parent->type_field == type_value); \
80 return exec_node_data(out_type, parent, field); \
90 * Description of built-in state associated with a uniform
92 * \sa nir_variable::state_slots
95 gl_state_index16 tokens
[STATE_LENGTH
];
100 nir_var_shader_in
= (1 << 0),
101 nir_var_shader_out
= (1 << 1),
102 nir_var_shader_temp
= (1 << 2),
103 nir_var_function_temp
= (1 << 3),
104 nir_var_uniform
= (1 << 4),
105 nir_var_mem_ubo
= (1 << 5),
106 nir_var_system_value
= (1 << 6),
107 nir_var_mem_ssbo
= (1 << 7),
108 nir_var_mem_shared
= (1 << 8),
109 nir_var_mem_global
= (1 << 9),
110 nir_var_mem_push_const
= (1 << 10), /* not actually used for variables */
111 nir_num_variable_modes
= 11,
112 nir_var_all
= (1 << nir_num_variable_modes
) - 1,
119 nir_rounding_mode_undef
= 0,
120 nir_rounding_mode_rtne
= 1, /* round to nearest even */
121 nir_rounding_mode_ru
= 2, /* round up */
122 nir_rounding_mode_rd
= 3, /* round down */
123 nir_rounding_mode_rtz
= 4, /* round towards zero */
140 #define nir_const_value_to_array(arr, c, components, m) \
142 for (unsigned i = 0; i < components; ++i) \
146 static inline nir_const_value
147 nir_const_value_for_raw_uint(uint64_t x
, unsigned bit_size
)
150 memset(&v
, 0, sizeof(v
));
153 case 1: v
.b
= x
; break;
154 case 8: v
.u8
= x
; break;
155 case 16: v
.u16
= x
; break;
156 case 32: v
.u32
= x
; break;
157 case 64: v
.u64
= x
; break;
159 unreachable("Invalid bit size");
165 static inline nir_const_value
166 nir_const_value_for_int(int64_t i
, unsigned bit_size
)
169 memset(&v
, 0, sizeof(v
));
171 assert(bit_size
<= 64);
173 assert(i
>= (-(1ll << (bit_size
- 1))));
174 assert(i
< (1ll << (bit_size
- 1)));
177 return nir_const_value_for_raw_uint(i
, bit_size
);
180 static inline nir_const_value
181 nir_const_value_for_uint(uint64_t u
, unsigned bit_size
)
184 memset(&v
, 0, sizeof(v
));
186 assert(bit_size
<= 64);
188 assert(u
< (1ull << bit_size
));
190 return nir_const_value_for_raw_uint(u
, bit_size
);
193 static inline nir_const_value
194 nir_const_value_for_bool(bool b
, unsigned bit_size
)
196 /* Booleans use a 0/-1 convention */
197 return nir_const_value_for_int(-(int)b
, bit_size
);
200 /* This one isn't inline because it requires half-float conversion */
201 nir_const_value
nir_const_value_for_float(double b
, unsigned bit_size
);
203 static inline int64_t
204 nir_const_value_as_int(nir_const_value value
, unsigned bit_size
)
207 /* int1_t uses 0/-1 convention */
208 case 1: return -(int)value
.b
;
209 case 8: return value
.i8
;
210 case 16: return value
.i16
;
211 case 32: return value
.i32
;
212 case 64: return value
.i64
;
214 unreachable("Invalid bit size");
218 static inline uint64_t
219 nir_const_value_as_uint(nir_const_value value
, unsigned bit_size
)
222 case 1: return value
.b
;
223 case 8: return value
.u8
;
224 case 16: return value
.u16
;
225 case 32: return value
.u32
;
226 case 64: return value
.u64
;
228 unreachable("Invalid bit size");
233 nir_const_value_as_bool(nir_const_value value
, unsigned bit_size
)
235 int64_t i
= nir_const_value_as_int(value
, bit_size
);
237 /* Booleans of any size use 0/-1 convention */
238 assert(i
== 0 || i
== -1);
243 /* This one isn't inline because it requires half-float conversion */
244 double nir_const_value_as_float(nir_const_value value
, unsigned bit_size
);
246 typedef struct nir_constant
{
248 * Value of the constant.
250 * The field used to back the values supplied by the constant is determined
251 * by the type associated with the \c nir_variable. Constants may be
252 * scalars, vectors, or matrices.
254 nir_const_value values
[NIR_MAX_VEC_COMPONENTS
];
256 /* we could get this from the var->type but makes clone *much* easier to
257 * not have to care about the type.
259 unsigned num_elements
;
261 /* Array elements / Structure Fields */
262 struct nir_constant
**elements
;
266 * \brief Layout qualifiers for gl_FragDepth.
268 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
269 * with a layout qualifier.
272 nir_depth_layout_none
, /**< No depth layout is specified. */
273 nir_depth_layout_any
,
274 nir_depth_layout_greater
,
275 nir_depth_layout_less
,
276 nir_depth_layout_unchanged
280 * Enum keeping track of how a variable was declared.
284 * Normal declaration.
286 nir_var_declared_normally
= 0,
289 * Variable is implicitly generated by the compiler and should not be
290 * visible via the API.
293 } nir_var_declaration_type
;
296 * Either a uniform, global variable, shader input, or shader output. Based on
297 * ir_variable - it should be easy to translate between the two.
300 typedef struct nir_variable
{
301 struct exec_node node
;
304 * Declared type of the variable
306 const struct glsl_type
*type
;
309 * Declared name of the variable
313 struct nir_variable_data
{
315 * Storage class of the variable.
317 * \sa nir_variable_mode
319 nir_variable_mode mode
:11;
322 * Is the variable read-only?
324 * This is set for variables declared as \c const, shader inputs,
327 unsigned read_only
:1;
331 unsigned invariant
:1;
334 * Can this variable be coalesced with another?
336 * This is set by nir_lower_io_to_temporaries to say that any
337 * copies involving this variable should stay put. Propagating it can
338 * duplicate the resulting load/store, which is not wanted, and may
339 * result in a load/store of the variable with an indirect offset which
340 * the backend may not be able to handle.
342 unsigned cannot_coalesce
:1;
345 * When separate shader programs are enabled, only input/outputs between
346 * the stages of a multi-stage separate program can be safely removed
347 * from the shader interface. Other input/outputs must remains active.
349 * This is also used to make sure xfb varyings that are unused by the
350 * fragment shader are not removed.
352 unsigned always_active_io
:1;
355 * Interpolation mode for shader inputs / outputs
357 * \sa glsl_interp_mode
359 unsigned interpolation
:2;
362 * If non-zero, then this variable may be packed along with other variables
363 * into a single varying slot, so this offset should be applied when
364 * accessing components. For example, an offset of 1 means that the x
365 * component of this variable is actually stored in component y of the
366 * location specified by \c location.
368 unsigned location_frac
:2;
371 * If true, this variable represents an array of scalars that should
372 * be tightly packed. In other words, consecutive array elements
373 * should be stored one component apart, rather than one slot apart.
378 * Whether this is a fragment shader output implicitly initialized with
379 * the previous contents of the specified render target at the
380 * framebuffer location corresponding to this shader invocation.
382 unsigned fb_fetch_output
:1;
385 * Non-zero if this variable is considered bindless as defined by
386 * ARB_bindless_texture.
391 * Was an explicit binding set in the shader?
393 unsigned explicit_binding
:1;
396 * Was a transfer feedback buffer set in the shader?
398 unsigned explicit_xfb_buffer
:1;
401 * Was a transfer feedback stride set in the shader?
403 unsigned explicit_xfb_stride
:1;
406 * Was an explicit offset set in the shader?
408 unsigned explicit_offset
:1;
411 * How the variable was declared. See nir_var_declaration_type.
413 * This is used to detect variables generated by the compiler, so should
414 * not be visible via the API.
416 unsigned how_declared
:2;
419 * \brief Layout qualifier for gl_FragDepth.
421 * This is not equal to \c ir_depth_layout_none if and only if this
422 * variable is \c gl_FragDepth and a layout qualifier is specified.
424 nir_depth_layout depth_layout
:3;
427 * Vertex stream output identifier.
429 * For packed outputs, NIR_STREAM_PACKED is set and bits [2*i+1,2*i]
430 * indicate the stream of the i-th component.
435 * Access flags for memory variables (SSBO/global), image uniforms, and
436 * bindless images in uniforms/inputs/outputs.
438 enum gl_access_qualifier access
:8;
441 * Descriptor set binding for sampler or UBO.
443 unsigned descriptor_set
:5;
446 * output index for dual source blending.
451 * Initial binding point for a sampler or UBO.
453 * For array types, this represents the binding point for the first element.
458 * Storage location of the base of this variable
460 * The precise meaning of this field depends on the nature of the variable.
462 * - Vertex shader input: one of the values from \c gl_vert_attrib.
463 * - Vertex shader output: one of the values from \c gl_varying_slot.
464 * - Geometry shader input: one of the values from \c gl_varying_slot.
465 * - Geometry shader output: one of the values from \c gl_varying_slot.
466 * - Fragment shader input: one of the values from \c gl_varying_slot.
467 * - Fragment shader output: one of the values from \c gl_frag_result.
468 * - Uniforms: Per-stage uniform slot number for default uniform block.
469 * - Uniforms: Index within the uniform block definition for UBO members.
470 * - Non-UBO Uniforms: uniform slot number.
471 * - Other: This field is not currently used.
473 * If the variable is a uniform, shader input, or shader output, and the
474 * slot has not been assigned, the value will be -1.
479 * The actual location of the variable in the IR. Only valid for inputs,
480 * outputs, and uniforms (including samplers and images).
482 unsigned driver_location
;
485 * Location an atomic counter or transform feedback is stored at.
491 /** Image internal format if specified explicitly, otherwise GL_NONE. */
492 uint16_t format
; /* GLenum */
497 * Transform feedback buffer.
502 * Transform feedback stride.
510 * Identifier for this variable generated by nir_index_vars() that is unique
511 * among other variables in the same exec_list.
515 /* Number of nir_variable_data members */
516 uint16_t num_members
;
519 * Built-in state that backs this uniform
521 * Once set at variable creation, \c state_slots must remain invariant.
522 * This is because, ideally, this array would be shared by all clones of
523 * this variable in the IR tree. In other words, we'd really like for it
524 * to be a fly-weight.
526 * If the variable is not a uniform, \c num_state_slots will be zero and
527 * \c state_slots will be \c NULL.
530 uint16_t num_state_slots
; /**< Number of state slots used */
531 nir_state_slot
*state_slots
; /**< State descriptors. */
535 * Constant expression assigned in the initializer of the variable
537 * This field should only be used temporarily by creators of NIR shaders
538 * and then lower_constant_initializers can be used to get rid of them.
539 * Most of the rest of NIR ignores this field or asserts that it's NULL.
541 nir_constant
*constant_initializer
;
544 * For variables that are in an interface block or are an instance of an
545 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
547 * \sa ir_variable::location
549 const struct glsl_type
*interface_type
;
552 * Description of per-member data for per-member struct variables
554 * This is used for variables which are actually an amalgamation of
555 * multiple entities such as a struct of built-in values or a struct of
556 * inputs each with their own layout specifier. This is only allowed on
557 * variables with a struct or array of array of struct type.
559 struct nir_variable_data
*members
;
562 #define nir_foreach_variable(var, var_list) \
563 foreach_list_typed(nir_variable, var, node, var_list)
565 #define nir_foreach_variable_safe(var, var_list) \
566 foreach_list_typed_safe(nir_variable, var, node, var_list)
569 nir_variable_is_global(const nir_variable
*var
)
571 return var
->data
.mode
!= nir_var_function_temp
;
574 typedef struct nir_register
{
575 struct exec_node node
;
577 unsigned num_components
; /** < number of vector components */
578 unsigned num_array_elems
; /** < size of array (0 for no array) */
580 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
583 /** generic register index. */
586 /** only for debug purposes, can be NULL */
589 /** set of nir_srcs where this register is used (read from) */
590 struct list_head uses
;
592 /** set of nir_dests where this register is defined (written to) */
593 struct list_head defs
;
595 /** set of nir_ifs where this register is used as a condition */
596 struct list_head if_uses
;
599 #define nir_foreach_register(reg, reg_list) \
600 foreach_list_typed(nir_register, reg, node, reg_list)
601 #define nir_foreach_register_safe(reg, reg_list) \
602 foreach_list_typed_safe(nir_register, reg, node, reg_list)
604 typedef enum PACKED
{
606 nir_instr_type_deref
,
609 nir_instr_type_intrinsic
,
610 nir_instr_type_load_const
,
612 nir_instr_type_ssa_undef
,
614 nir_instr_type_parallel_copy
,
617 typedef struct nir_instr
{
618 struct exec_node node
;
619 struct nir_block
*block
;
622 /* A temporary for optimization and analysis passes to use for storing
623 * flags. For instance, DCE uses this to store the "dead/live" info.
627 /** generic instruction index. */
631 static inline nir_instr
*
632 nir_instr_next(nir_instr
*instr
)
634 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
635 if (exec_node_is_tail_sentinel(next
))
638 return exec_node_data(nir_instr
, next
, node
);
641 static inline nir_instr
*
642 nir_instr_prev(nir_instr
*instr
)
644 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
645 if (exec_node_is_head_sentinel(prev
))
648 return exec_node_data(nir_instr
, prev
, node
);
652 nir_instr_is_first(const nir_instr
*instr
)
654 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr
->node
));
658 nir_instr_is_last(const nir_instr
*instr
)
660 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr
->node
));
663 typedef struct nir_ssa_def
{
664 /** for debugging only, can be NULL */
667 /** generic SSA definition index. */
670 /** Index into the live_in and live_out bitfields */
673 /** Instruction which produces this SSA value. */
674 nir_instr
*parent_instr
;
676 /** set of nir_instrs where this register is used (read from) */
677 struct list_head uses
;
679 /** set of nir_ifs where this register is used as a condition */
680 struct list_head if_uses
;
682 uint8_t num_components
;
684 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
692 struct nir_src
*indirect
; /** < NULL for no indirect offset */
693 unsigned base_offset
;
695 /* TODO use-def chain goes here */
699 nir_instr
*parent_instr
;
700 struct list_head def_link
;
703 struct nir_src
*indirect
; /** < NULL for no indirect offset */
704 unsigned base_offset
;
706 /* TODO def-use chain goes here */
711 typedef struct nir_src
{
713 /** Instruction that consumes this value as a source. */
714 nir_instr
*parent_instr
;
715 struct nir_if
*parent_if
;
718 struct list_head use_link
;
728 static inline nir_src
731 nir_src src
= { { NULL
} };
735 #define NIR_SRC_INIT nir_src_init()
737 #define nir_foreach_use(src, reg_or_ssa_def) \
738 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
740 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
741 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
743 #define nir_foreach_if_use(src, reg_or_ssa_def) \
744 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
746 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
747 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
758 static inline nir_dest
761 nir_dest dest
= { { { NULL
} } };
765 #define NIR_DEST_INIT nir_dest_init()
767 #define nir_foreach_def(dest, reg) \
768 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
770 #define nir_foreach_def_safe(dest, reg) \
771 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
773 static inline nir_src
774 nir_src_for_ssa(nir_ssa_def
*def
)
776 nir_src src
= NIR_SRC_INIT
;
784 static inline nir_src
785 nir_src_for_reg(nir_register
*reg
)
787 nir_src src
= NIR_SRC_INIT
;
791 src
.reg
.indirect
= NULL
;
792 src
.reg
.base_offset
= 0;
797 static inline nir_dest
798 nir_dest_for_reg(nir_register
*reg
)
800 nir_dest dest
= NIR_DEST_INIT
;
807 static inline unsigned
808 nir_src_bit_size(nir_src src
)
810 return src
.is_ssa
? src
.ssa
->bit_size
: src
.reg
.reg
->bit_size
;
813 static inline unsigned
814 nir_src_num_components(nir_src src
)
816 return src
.is_ssa
? src
.ssa
->num_components
: src
.reg
.reg
->num_components
;
820 nir_src_is_const(nir_src src
)
823 src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
;
826 static inline unsigned
827 nir_dest_bit_size(nir_dest dest
)
829 return dest
.is_ssa
? dest
.ssa
.bit_size
: dest
.reg
.reg
->bit_size
;
832 static inline unsigned
833 nir_dest_num_components(nir_dest dest
)
835 return dest
.is_ssa
? dest
.ssa
.num_components
: dest
.reg
.reg
->num_components
;
838 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
839 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
845 * \name input modifiers
849 * For inputs interpreted as floating point, flips the sign bit. For
850 * inputs interpreted as integers, performs the two's complement negation.
855 * Clears the sign bit for floating point values, and computes the integer
856 * absolute value for integers. Note that the negate modifier acts after
857 * the absolute value modifier, therefore if both are set then all inputs
858 * will become negative.
864 * For each input component, says which component of the register it is
865 * chosen from. Note that which elements of the swizzle are used and which
866 * are ignored are based on the write mask for most opcodes - for example,
867 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
868 * a swizzle of {2, x, 1, 0} where x means "don't care."
870 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
877 * \name saturate output modifier
879 * Only valid for opcodes that output floating-point numbers. Clamps the
880 * output to between 0.0 and 1.0 inclusive.
885 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
888 /** NIR sized and unsized types
890 * The values in this enum are carefully chosen so that the sized type is
891 * just the unsized type OR the number of bits.
894 nir_type_invalid
= 0, /* Not a valid type */
898 nir_type_float
= 128,
899 nir_type_bool1
= 1 | nir_type_bool
,
900 nir_type_bool8
= 8 | nir_type_bool
,
901 nir_type_bool16
= 16 | nir_type_bool
,
902 nir_type_bool32
= 32 | nir_type_bool
,
903 nir_type_int1
= 1 | nir_type_int
,
904 nir_type_int8
= 8 | nir_type_int
,
905 nir_type_int16
= 16 | nir_type_int
,
906 nir_type_int32
= 32 | nir_type_int
,
907 nir_type_int64
= 64 | nir_type_int
,
908 nir_type_uint1
= 1 | nir_type_uint
,
909 nir_type_uint8
= 8 | nir_type_uint
,
910 nir_type_uint16
= 16 | nir_type_uint
,
911 nir_type_uint32
= 32 | nir_type_uint
,
912 nir_type_uint64
= 64 | nir_type_uint
,
913 nir_type_float16
= 16 | nir_type_float
,
914 nir_type_float32
= 32 | nir_type_float
,
915 nir_type_float64
= 64 | nir_type_float
,
918 #define NIR_ALU_TYPE_SIZE_MASK 0x79
919 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
921 static inline unsigned
922 nir_alu_type_get_type_size(nir_alu_type type
)
924 return type
& NIR_ALU_TYPE_SIZE_MASK
;
927 static inline unsigned
928 nir_alu_type_get_base_type(nir_alu_type type
)
930 return type
& NIR_ALU_TYPE_BASE_TYPE_MASK
;
933 static inline nir_alu_type
934 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
938 return nir_type_bool1
;
941 return nir_type_uint32
;
944 return nir_type_int32
;
946 case GLSL_TYPE_UINT16
:
947 return nir_type_uint16
;
949 case GLSL_TYPE_INT16
:
950 return nir_type_int16
;
952 case GLSL_TYPE_UINT8
:
953 return nir_type_uint8
;
955 return nir_type_int8
;
956 case GLSL_TYPE_UINT64
:
957 return nir_type_uint64
;
959 case GLSL_TYPE_INT64
:
960 return nir_type_int64
;
962 case GLSL_TYPE_FLOAT
:
963 return nir_type_float32
;
965 case GLSL_TYPE_FLOAT16
:
966 return nir_type_float16
;
968 case GLSL_TYPE_DOUBLE
:
969 return nir_type_float64
;
972 case GLSL_TYPE_SAMPLER
:
973 case GLSL_TYPE_IMAGE
:
974 case GLSL_TYPE_ATOMIC_UINT
:
975 case GLSL_TYPE_STRUCT
:
976 case GLSL_TYPE_INTERFACE
:
977 case GLSL_TYPE_ARRAY
:
979 case GLSL_TYPE_SUBROUTINE
:
980 case GLSL_TYPE_FUNCTION
:
981 case GLSL_TYPE_ERROR
:
982 return nir_type_invalid
;
985 unreachable("unknown type");
988 static inline nir_alu_type
989 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
991 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
994 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
995 nir_rounding_mode rnd
);
998 nir_op_vec(unsigned components
)
1000 switch (components
) {
1001 case 1: return nir_op_mov
;
1002 case 2: return nir_op_vec2
;
1003 case 3: return nir_op_vec3
;
1004 case 4: return nir_op_vec4
;
1005 default: unreachable("bad component count");
1010 nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode
, unsigned bit_size
)
1012 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
) ||
1013 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
) ||
1014 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
);
1018 nir_is_denorm_flush_to_zero(unsigned execution_mode
, unsigned bit_size
)
1020 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
) ||
1021 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
) ||
1022 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
);
1026 nir_is_denorm_preserve(unsigned execution_mode
, unsigned bit_size
)
1028 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP16
) ||
1029 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP32
) ||
1030 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP64
);
1034 nir_is_rounding_mode_rtne(unsigned execution_mode
, unsigned bit_size
)
1036 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1037 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1038 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1042 nir_is_rounding_mode_rtz(unsigned execution_mode
, unsigned bit_size
)
1044 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1045 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1046 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1050 nir_has_any_rounding_mode_rtz(unsigned execution_mode
)
1052 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1053 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1054 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1058 nir_has_any_rounding_mode_rtne(unsigned execution_mode
)
1060 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1061 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1062 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1065 static inline nir_rounding_mode
1066 nir_get_rounding_mode_from_float_controls(unsigned execution_mode
,
1069 if (nir_alu_type_get_base_type(type
) != nir_type_float
)
1070 return nir_rounding_mode_undef
;
1072 unsigned bit_size
= nir_alu_type_get_type_size(type
);
1074 if (nir_is_rounding_mode_rtz(execution_mode
, bit_size
))
1075 return nir_rounding_mode_rtz
;
1076 if (nir_is_rounding_mode_rtne(execution_mode
, bit_size
))
1077 return nir_rounding_mode_rtne
;
1078 return nir_rounding_mode_undef
;
1082 nir_has_any_rounding_mode_enabled(unsigned execution_mode
)
1085 nir_has_any_rounding_mode_rtne(execution_mode
) ||
1086 nir_has_any_rounding_mode_rtz(execution_mode
);
1092 * Operation where the first two sources are commutative.
1094 * For 2-source operations, this just mathematical commutativity. Some
1095 * 3-source operations, like ffma, are only commutative in the first two
1098 NIR_OP_IS_2SRC_COMMUTATIVE
= (1 << 0),
1099 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
1100 } nir_op_algebraic_property
;
1105 unsigned num_inputs
;
1108 * The number of components in the output
1110 * If non-zero, this is the size of the output and input sizes are
1111 * explicitly given; swizzle and writemask are still in effect, but if
1112 * the output component is masked out, then the input component may
1115 * If zero, the opcode acts in the standard, per-component manner; the
1116 * operation is performed on each component (except the ones that are
1117 * masked out) with the input being taken from the input swizzle for
1120 * The size of some of the inputs may be given (i.e. non-zero) even
1121 * though output_size is zero; in that case, the inputs with a zero
1122 * size act per-component, while the inputs with non-zero size don't.
1124 unsigned output_size
;
1127 * The type of vector that the instruction outputs. Note that the
1128 * staurate modifier is only allowed on outputs with the float type.
1131 nir_alu_type output_type
;
1134 * The number of components in each input
1136 unsigned input_sizes
[NIR_MAX_VEC_COMPONENTS
];
1139 * The type of vector that each input takes. Note that negate and
1140 * absolute value are only allowed on inputs with int or float type and
1141 * behave differently on the two.
1143 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
1145 nir_op_algebraic_property algebraic_properties
;
1147 /* Whether this represents a numeric conversion opcode */
1151 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
1153 typedef struct nir_alu_instr
{
1157 /** Indicates that this ALU instruction generates an exact value
1159 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1160 * really equivalent to either. This indicates that the value generated by
1161 * this operation is high-precision and any code transformations that touch
1162 * it must ensure that the resulting value is bit-for-bit identical to the
1168 * Indicates that this instruction do not cause wrapping to occur, in the
1169 * form of overflow or underflow.
1171 bool no_signed_wrap
:1;
1172 bool no_unsigned_wrap
:1;
1178 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
1179 nir_alu_instr
*instr
);
1180 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
1181 nir_alu_instr
*instr
);
1183 /* is this source channel used? */
1185 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
1188 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1189 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
1191 return (instr
->dest
.write_mask
>> channel
) & 1;
1194 static inline nir_component_mask_t
1195 nir_alu_instr_src_read_mask(const nir_alu_instr
*instr
, unsigned src
)
1197 nir_component_mask_t read_mask
= 0;
1198 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; c
++) {
1199 if (!nir_alu_instr_channel_used(instr
, src
, c
))
1202 read_mask
|= (1 << instr
->src
[src
].swizzle
[c
]);
1208 * Get the number of channels used for a source
1210 static inline unsigned
1211 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
1213 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1214 return nir_op_infos
[instr
->op
].input_sizes
[src
];
1216 return nir_dest_num_components(instr
->dest
.dest
);
1220 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1222 switch (instr
->op
) {
1242 bool nir_const_value_negative_equal(nir_const_value c1
, nir_const_value c2
,
1243 nir_alu_type full_type
);
1245 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
1246 unsigned src1
, unsigned src2
);
1248 bool nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
1249 const nir_alu_instr
*alu2
,
1250 unsigned src1
, unsigned src2
);
1254 nir_deref_type_array
,
1255 nir_deref_type_array_wildcard
,
1256 nir_deref_type_ptr_as_array
,
1257 nir_deref_type_struct
,
1258 nir_deref_type_cast
,
1264 /** The type of this deref instruction */
1265 nir_deref_type deref_type
;
1267 /** The mode of the underlying variable */
1268 nir_variable_mode mode
;
1270 /** The dereferenced type of the resulting pointer value */
1271 const struct glsl_type
*type
;
1274 /** Variable being dereferenced if deref_type is a deref_var */
1277 /** Parent deref if deref_type is not deref_var */
1281 /** Additional deref parameters */
1292 unsigned ptr_stride
;
1296 /** Destination to store the resulting "pointer" */
1300 static inline nir_deref_instr
*nir_src_as_deref(nir_src src
);
1302 static inline nir_deref_instr
*
1303 nir_deref_instr_parent(const nir_deref_instr
*instr
)
1305 if (instr
->deref_type
== nir_deref_type_var
)
1308 return nir_src_as_deref(instr
->parent
);
1311 static inline nir_variable
*
1312 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1314 while (instr
->deref_type
!= nir_deref_type_var
) {
1315 if (instr
->deref_type
== nir_deref_type_cast
)
1318 instr
= nir_deref_instr_parent(instr
);
1324 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1325 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr
*instr
);
1326 bool nir_deref_instr_has_complex_use(nir_deref_instr
*instr
);
1328 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1330 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*instr
);
1335 struct nir_function
*callee
;
1337 unsigned num_params
;
1341 #include "nir_intrinsics.h"
1343 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1345 /** Represents an intrinsic
1347 * An intrinsic is an instruction type for handling things that are
1348 * more-or-less regular operations but don't just consume and produce SSA
1349 * values like ALU operations do. Intrinsics are not for things that have
1350 * special semantic meaning such as phi nodes and parallel copies.
1351 * Examples of intrinsics include variable load/store operations, system
1352 * value loads, and the like. Even though texturing more-or-less falls
1353 * under this category, texturing is its own instruction type because
1354 * trying to represent texturing with intrinsics would lead to a
1355 * combinatorial explosion of intrinsic opcodes.
1357 * By having a single instruction type for handling a lot of different
1358 * cases, optimization passes can look for intrinsics and, for the most
1359 * part, completely ignore them. Each intrinsic type also has a few
1360 * possible flags that govern whether or not they can be reordered or
1361 * eliminated. That way passes like dead code elimination can still work
1362 * on intrisics without understanding the meaning of each.
1364 * Each intrinsic has some number of constant indices, some number of
1365 * variables, and some number of sources. What these sources, variables,
1366 * and indices mean depends on the intrinsic and is documented with the
1367 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1368 * instructions are the only types of instruction that can operate on
1374 nir_intrinsic_op intrinsic
;
1378 /** number of components if this is a vectorized intrinsic
1380 * Similarly to ALU operations, some intrinsics are vectorized.
1381 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1382 * For vectorized intrinsics, the num_components field specifies the
1383 * number of destination components and the number of source components
1384 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1386 uint8_t num_components
;
1388 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1391 } nir_intrinsic_instr
;
1393 static inline nir_variable
*
1394 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1396 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1400 /* Memory ordering. */
1401 NIR_MEMORY_ACQUIRE
= 1 << 0,
1402 NIR_MEMORY_RELEASE
= 1 << 1,
1404 /* Memory visibility operations. */
1405 NIR_MEMORY_MAKE_AVAILABLE
= 1 << 3,
1406 NIR_MEMORY_MAKE_VISIBLE
= 1 << 4,
1407 } nir_memory_semantics
;
1411 NIR_SCOPE_QUEUE_FAMILY
,
1412 NIR_SCOPE_WORKGROUP
,
1414 NIR_SCOPE_INVOCATION
,
1418 * \name NIR intrinsics semantic flags
1420 * information about what the compiler can do with the intrinsics.
1422 * \sa nir_intrinsic_info::flags
1426 * whether the intrinsic can be safely eliminated if none of its output
1427 * value is not being used.
1429 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1432 * Whether the intrinsic can be reordered with respect to any other
1433 * intrinsic, i.e. whether the only reordering dependencies of the
1434 * intrinsic are due to the register reads/writes.
1436 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1437 } nir_intrinsic_semantic_flag
;
1440 * \name NIR intrinsics const-index flag
1442 * Indicates the usage of a const_index slot.
1444 * \sa nir_intrinsic_info::index_map
1448 * Generally instructions that take a offset src argument, can encode
1449 * a constant 'base' value which is added to the offset.
1451 NIR_INTRINSIC_BASE
= 1,
1454 * For store instructions, a writemask for the store.
1456 NIR_INTRINSIC_WRMASK
,
1459 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1461 NIR_INTRINSIC_STREAM_ID
,
1464 * The clip-plane id for load_user_clip_plane intrinsic.
1466 NIR_INTRINSIC_UCP_ID
,
1469 * The amount of data, starting from BASE, that this instruction may
1470 * access. This is used to provide bounds if the offset is not constant.
1472 NIR_INTRINSIC_RANGE
,
1475 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1477 NIR_INTRINSIC_DESC_SET
,
1480 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1482 NIR_INTRINSIC_BINDING
,
1487 NIR_INTRINSIC_COMPONENT
,
1490 * Interpolation mode (only meaningful for FS inputs).
1492 NIR_INTRINSIC_INTERP_MODE
,
1495 * A binary nir_op to use when performing a reduction or scan operation
1497 NIR_INTRINSIC_REDUCTION_OP
,
1500 * Cluster size for reduction operations
1502 NIR_INTRINSIC_CLUSTER_SIZE
,
1505 * Parameter index for a load_param intrinsic
1507 NIR_INTRINSIC_PARAM_IDX
,
1510 * Image dimensionality for image intrinsics
1512 * One of GLSL_SAMPLER_DIM_*
1514 NIR_INTRINSIC_IMAGE_DIM
,
1517 * Non-zero if we are accessing an array image
1519 NIR_INTRINSIC_IMAGE_ARRAY
,
1522 * Image format for image intrinsics
1524 NIR_INTRINSIC_FORMAT
,
1527 * Access qualifiers for image and memory access intrinsics
1529 NIR_INTRINSIC_ACCESS
,
1532 * Alignment for offsets and addresses
1534 * These two parameters, specify an alignment in terms of a multiplier and
1535 * an offset. The offset or address parameter X of the intrinsic is
1536 * guaranteed to satisfy the following:
1538 * (X - align_offset) % align_mul == 0
1540 NIR_INTRINSIC_ALIGN_MUL
,
1541 NIR_INTRINSIC_ALIGN_OFFSET
,
1544 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1546 NIR_INTRINSIC_DESC_TYPE
,
1549 * The nir_alu_type of a uniform/input/output
1554 * The swizzle mask for the instructions
1555 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1557 NIR_INTRINSIC_SWIZZLE_MASK
,
1559 /* Separate source/dest access flags for copies */
1560 NIR_INTRINSIC_SRC_ACCESS
,
1561 NIR_INTRINSIC_DST_ACCESS
,
1563 /* Driver location for nir_load_patch_location_ir3 */
1564 NIR_INTRINSIC_DRIVER_LOCATION
,
1567 * Mask of nir_memory_semantics, includes ordering and visibility.
1569 NIR_INTRINSIC_MEMORY_SEMANTICS
,
1572 * Mask of nir_variable_modes affected by the memory operation.
1574 NIR_INTRINSIC_MEMORY_MODES
,
1577 * Value of nir_scope.
1579 NIR_INTRINSIC_MEMORY_SCOPE
,
1581 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1583 } nir_intrinsic_index_flag
;
1585 #define NIR_INTRINSIC_MAX_INPUTS 5
1590 unsigned num_srcs
; /** < number of register/SSA inputs */
1592 /** number of components of each input register
1594 * If this value is 0, the number of components is given by the
1595 * num_components field of nir_intrinsic_instr. If this value is -1, the
1596 * intrinsic consumes however many components are provided and it is not
1599 int src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1603 /** number of components of the output register
1605 * If this value is 0, the number of components is given by the
1606 * num_components field of nir_intrinsic_instr.
1608 unsigned dest_components
;
1610 /** bitfield of legal bit sizes */
1611 unsigned dest_bit_sizes
;
1613 /** the number of constant indices used by the intrinsic */
1614 unsigned num_indices
;
1616 /** indicates the usage of intr->const_index[n] */
1617 unsigned index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1619 /** semantic flags for calls to this intrinsic */
1620 nir_intrinsic_semantic_flag flags
;
1621 } nir_intrinsic_info
;
1623 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1625 static inline unsigned
1626 nir_intrinsic_src_components(nir_intrinsic_instr
*intr
, unsigned srcn
)
1628 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1629 assert(srcn
< info
->num_srcs
);
1630 if (info
->src_components
[srcn
] > 0)
1631 return info
->src_components
[srcn
];
1632 else if (info
->src_components
[srcn
] == 0)
1633 return intr
->num_components
;
1635 return nir_src_num_components(intr
->src
[srcn
]);
1638 static inline unsigned
1639 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1641 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1642 if (!info
->has_dest
)
1644 else if (info
->dest_components
)
1645 return info
->dest_components
;
1647 return intr
->num_components
;
1650 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1651 static inline type \
1652 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1654 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1655 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1656 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1658 static inline void \
1659 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1661 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1662 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1663 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1666 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1667 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1668 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1669 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1670 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1671 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1672 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1673 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1674 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1675 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1676 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1677 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1678 INTRINSIC_IDX_ACCESSORS(image_dim
, IMAGE_DIM
, enum glsl_sampler_dim
)
1679 INTRINSIC_IDX_ACCESSORS(image_array
, IMAGE_ARRAY
, bool)
1680 INTRINSIC_IDX_ACCESSORS(access
, ACCESS
, enum gl_access_qualifier
)
1681 INTRINSIC_IDX_ACCESSORS(src_access
, SRC_ACCESS
, enum gl_access_qualifier
)
1682 INTRINSIC_IDX_ACCESSORS(dst_access
, DST_ACCESS
, enum gl_access_qualifier
)
1683 INTRINSIC_IDX_ACCESSORS(format
, FORMAT
, unsigned)
1684 INTRINSIC_IDX_ACCESSORS(align_mul
, ALIGN_MUL
, unsigned)
1685 INTRINSIC_IDX_ACCESSORS(align_offset
, ALIGN_OFFSET
, unsigned)
1686 INTRINSIC_IDX_ACCESSORS(desc_type
, DESC_TYPE
, unsigned)
1687 INTRINSIC_IDX_ACCESSORS(type
, TYPE
, nir_alu_type
)
1688 INTRINSIC_IDX_ACCESSORS(swizzle_mask
, SWIZZLE_MASK
, unsigned)
1689 INTRINSIC_IDX_ACCESSORS(driver_location
, DRIVER_LOCATION
, unsigned)
1690 INTRINSIC_IDX_ACCESSORS(memory_semantics
, MEMORY_SEMANTICS
, nir_memory_semantics
)
1691 INTRINSIC_IDX_ACCESSORS(memory_modes
, MEMORY_MODES
, nir_variable_mode
)
1692 INTRINSIC_IDX_ACCESSORS(memory_scope
, MEMORY_SCOPE
, nir_scope
)
1695 nir_intrinsic_set_align(nir_intrinsic_instr
*intrin
,
1696 unsigned align_mul
, unsigned align_offset
)
1698 assert(util_is_power_of_two_nonzero(align_mul
));
1699 assert(align_offset
< align_mul
);
1700 nir_intrinsic_set_align_mul(intrin
, align_mul
);
1701 nir_intrinsic_set_align_offset(intrin
, align_offset
);
1704 /** Returns a simple alignment for a load/store intrinsic offset
1706 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1707 * and ALIGN_OFFSET parameters, this helper takes both into account and
1708 * provides a single simple alignment parameter. The offset X is guaranteed
1709 * to satisfy X % align == 0.
1711 static inline unsigned
1712 nir_intrinsic_align(const nir_intrinsic_instr
*intrin
)
1714 const unsigned align_mul
= nir_intrinsic_align_mul(intrin
);
1715 const unsigned align_offset
= nir_intrinsic_align_offset(intrin
);
1716 assert(align_offset
< align_mul
);
1717 return align_offset
? 1 << (ffs(align_offset
) - 1) : align_mul
;
1720 /* Converts a image_deref_* intrinsic into a image_* one */
1721 void nir_rewrite_image_intrinsic(nir_intrinsic_instr
*instr
,
1722 nir_ssa_def
*handle
, bool bindless
);
1724 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1726 nir_intrinsic_can_reorder(nir_intrinsic_instr
*instr
)
1728 if (instr
->intrinsic
== nir_intrinsic_load_deref
||
1729 instr
->intrinsic
== nir_intrinsic_load_ssbo
||
1730 instr
->intrinsic
== nir_intrinsic_bindless_image_load
||
1731 instr
->intrinsic
== nir_intrinsic_image_deref_load
||
1732 instr
->intrinsic
== nir_intrinsic_image_load
) {
1733 return nir_intrinsic_access(instr
) & ACCESS_CAN_REORDER
;
1735 const nir_intrinsic_info
*info
=
1736 &nir_intrinsic_infos
[instr
->intrinsic
];
1737 return (info
->flags
& NIR_INTRINSIC_CAN_ELIMINATE
) &&
1738 (info
->flags
& NIR_INTRINSIC_CAN_REORDER
);
1743 * \group texture information
1745 * This gives semantic information about textures which is useful to the
1746 * frontend, the backend, and lowering passes, but not the optimizer.
1751 nir_tex_src_projector
,
1752 nir_tex_src_comparator
, /* shadow comparator */
1756 nir_tex_src_min_lod
,
1757 nir_tex_src_ms_index
, /* MSAA sample index */
1758 nir_tex_src_ms_mcs
, /* MSAA compression value */
1761 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
1762 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
1763 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
1764 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
1765 nir_tex_src_texture_handle
, /* < bindless texture handle */
1766 nir_tex_src_sampler_handle
, /* < bindless sampler handle */
1767 nir_tex_src_plane
, /* < selects plane for planar textures */
1768 nir_num_tex_src_types
1773 nir_tex_src_type src_type
;
1777 nir_texop_tex
, /**< Regular texture look-up */
1778 nir_texop_txb
, /**< Texture look-up with LOD bias */
1779 nir_texop_txl
, /**< Texture look-up with explicit LOD */
1780 nir_texop_txd
, /**< Texture look-up with partial derivatives */
1781 nir_texop_txf
, /**< Texel fetch with explicit LOD */
1782 nir_texop_txf_ms
, /**< Multisample texture fetch */
1783 nir_texop_txf_ms_fb
, /**< Multisample texture fetch from framebuffer */
1784 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
1785 nir_texop_txs
, /**< Texture size */
1786 nir_texop_lod
, /**< Texture lod query */
1787 nir_texop_tg4
, /**< Texture gather */
1788 nir_texop_query_levels
, /**< Texture levels query */
1789 nir_texop_texture_samples
, /**< Texture samples query */
1790 nir_texop_samples_identical
, /**< Query whether all samples are definitely
1793 nir_texop_tex_prefetch
, /**< Regular texture look-up, eligible for pre-dispatch */
1799 enum glsl_sampler_dim sampler_dim
;
1800 nir_alu_type dest_type
;
1805 unsigned num_srcs
, coord_components
;
1806 bool is_array
, is_shadow
;
1809 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1810 * components or the new-style shadow that outputs 1 component.
1812 bool is_new_style_shadow
;
1814 /* gather component selector */
1815 unsigned component
: 2;
1817 /* gather offsets */
1818 int8_t tg4_offsets
[4][2];
1820 /* True if the texture index or handle is not dynamically uniform */
1821 bool texture_non_uniform
;
1823 /* True if the sampler index or handle is not dynamically uniform */
1824 bool sampler_non_uniform
;
1826 /** The texture index
1828 * If this texture instruction has a nir_tex_src_texture_offset source,
1829 * then the texture index is given by texture_index + texture_offset.
1831 unsigned texture_index
;
1833 /** The size of the texture array or 0 if it's not an array */
1834 unsigned texture_array_size
;
1836 /** The sampler index
1838 * The following operations do not require a sampler and, as such, this
1839 * field should be ignored:
1841 * - nir_texop_txf_ms
1844 * - nir_texop_query_levels
1845 * - nir_texop_texture_samples
1846 * - nir_texop_samples_identical
1848 * If this texture instruction has a nir_tex_src_sampler_offset source,
1849 * then the sampler index is given by sampler_index + sampler_offset.
1851 unsigned sampler_index
;
1854 static inline unsigned
1855 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
1857 switch (instr
->op
) {
1858 case nir_texop_txs
: {
1860 switch (instr
->sampler_dim
) {
1861 case GLSL_SAMPLER_DIM_1D
:
1862 case GLSL_SAMPLER_DIM_BUF
:
1865 case GLSL_SAMPLER_DIM_2D
:
1866 case GLSL_SAMPLER_DIM_CUBE
:
1867 case GLSL_SAMPLER_DIM_MS
:
1868 case GLSL_SAMPLER_DIM_RECT
:
1869 case GLSL_SAMPLER_DIM_EXTERNAL
:
1870 case GLSL_SAMPLER_DIM_SUBPASS
:
1873 case GLSL_SAMPLER_DIM_3D
:
1877 unreachable("not reached");
1879 if (instr
->is_array
)
1887 case nir_texop_texture_samples
:
1888 case nir_texop_query_levels
:
1889 case nir_texop_samples_identical
:
1893 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
1900 /* Returns true if this texture operation queries something about the texture
1901 * rather than actually sampling it.
1904 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
1906 switch (instr
->op
) {
1909 case nir_texop_texture_samples
:
1910 case nir_texop_query_levels
:
1911 case nir_texop_txf_ms_mcs
:
1918 case nir_texop_txf_ms
:
1919 case nir_texop_txf_ms_fb
:
1923 unreachable("Invalid texture opcode");
1928 nir_tex_instr_has_implicit_derivative(const nir_tex_instr
*instr
)
1930 switch (instr
->op
) {
1940 static inline nir_alu_type
1941 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
1943 switch (instr
->src
[src
].src_type
) {
1944 case nir_tex_src_coord
:
1945 switch (instr
->op
) {
1947 case nir_texop_txf_ms
:
1948 case nir_texop_txf_ms_fb
:
1949 case nir_texop_txf_ms_mcs
:
1950 case nir_texop_samples_identical
:
1951 return nir_type_int
;
1954 return nir_type_float
;
1957 case nir_tex_src_lod
:
1958 switch (instr
->op
) {
1961 return nir_type_int
;
1964 return nir_type_float
;
1967 case nir_tex_src_projector
:
1968 case nir_tex_src_comparator
:
1969 case nir_tex_src_bias
:
1970 case nir_tex_src_min_lod
:
1971 case nir_tex_src_ddx
:
1972 case nir_tex_src_ddy
:
1973 return nir_type_float
;
1975 case nir_tex_src_offset
:
1976 case nir_tex_src_ms_index
:
1977 case nir_tex_src_plane
:
1978 return nir_type_int
;
1980 case nir_tex_src_ms_mcs
:
1981 case nir_tex_src_texture_deref
:
1982 case nir_tex_src_sampler_deref
:
1983 case nir_tex_src_texture_offset
:
1984 case nir_tex_src_sampler_offset
:
1985 case nir_tex_src_texture_handle
:
1986 case nir_tex_src_sampler_handle
:
1987 return nir_type_uint
;
1989 case nir_num_tex_src_types
:
1990 unreachable("nir_num_tex_src_types is not a valid source type");
1993 unreachable("Invalid texture source type");
1996 static inline unsigned
1997 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
1999 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
2000 return instr
->coord_components
;
2002 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
2003 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
2006 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
2007 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
2008 if (instr
->is_array
)
2009 return instr
->coord_components
- 1;
2011 return instr
->coord_components
;
2014 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
2015 * the offset, since a cube maps to a single face.
2017 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
2018 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2020 else if (instr
->is_array
)
2021 return instr
->coord_components
- 1;
2023 return instr
->coord_components
;
2030 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
2032 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
2033 if (instr
->src
[i
].src_type
== type
)
2039 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
2040 nir_tex_src_type src_type
,
2043 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
2045 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
);
2052 nir_const_value value
[];
2053 } nir_load_const_instr
;
2066 /* creates a new SSA variable in an undefined state */
2071 } nir_ssa_undef_instr
;
2074 struct exec_node node
;
2076 /* The predecessor block corresponding to this source */
2077 struct nir_block
*pred
;
2082 #define nir_foreach_phi_src(phi_src, phi) \
2083 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
2084 #define nir_foreach_phi_src_safe(phi_src, phi) \
2085 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
2090 struct exec_list srcs
; /** < list of nir_phi_src */
2096 struct exec_node node
;
2099 } nir_parallel_copy_entry
;
2101 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
2102 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
2107 /* A list of nir_parallel_copy_entrys. The sources of all of the
2108 * entries are copied to the corresponding destinations "in parallel".
2109 * In other words, if we have two entries: a -> b and b -> a, the values
2112 struct exec_list entries
;
2113 } nir_parallel_copy_instr
;
2115 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
2116 type
, nir_instr_type_alu
)
2117 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
2118 type
, nir_instr_type_deref
)
2119 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
2120 type
, nir_instr_type_call
)
2121 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
2122 type
, nir_instr_type_jump
)
2123 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
2124 type
, nir_instr_type_tex
)
2125 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
2126 type
, nir_instr_type_intrinsic
)
2127 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
2128 type
, nir_instr_type_load_const
)
2129 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
2130 type
, nir_instr_type_ssa_undef
)
2131 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
2132 type
, nir_instr_type_phi
)
2133 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
2134 nir_parallel_copy_instr
, instr
,
2135 type
, nir_instr_type_parallel_copy
)
2138 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
2139 static inline type \
2140 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2142 assert(nir_src_is_const(src)); \
2143 nir_load_const_instr *load = \
2144 nir_instr_as_load_const(src.ssa->parent_instr); \
2145 assert(comp < load->def.num_components); \
2146 return nir_const_value_as_##suffix(load->value[comp], \
2147 load->def.bit_size); \
2150 static inline type \
2151 nir_src_as_##suffix(nir_src src) \
2153 assert(nir_src_num_components(src) == 1); \
2154 return nir_src_comp_as_##suffix(src, 0); \
2157 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2158 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint
)
2159 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2160 NIR_DEFINE_SRC_AS_CONST(double, float)
2162 #undef NIR_DEFINE_SRC_AS_CONST
2171 nir_ssa_scalar_is_const(nir_ssa_scalar s
)
2173 return s
.def
->parent_instr
->type
== nir_instr_type_load_const
;
2176 static inline nir_const_value
2177 nir_ssa_scalar_as_const_value(nir_ssa_scalar s
)
2179 assert(s
.comp
< s
.def
->num_components
);
2180 nir_load_const_instr
*load
= nir_instr_as_load_const(s
.def
->parent_instr
);
2181 return load
->value
[s
.comp
];
2184 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2185 static inline type \
2186 nir_ssa_scalar_as_##suffix(nir_ssa_scalar s) \
2188 return nir_const_value_as_##suffix( \
2189 nir_ssa_scalar_as_const_value(s), s.def->bit_size); \
2192 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2193 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint
)
2194 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2195 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2197 #undef NIR_DEFINE_SCALAR_AS_CONST
2200 nir_ssa_scalar_is_alu(nir_ssa_scalar s
)
2202 return s
.def
->parent_instr
->type
== nir_instr_type_alu
;
2205 static inline nir_op
2206 nir_ssa_scalar_alu_op(nir_ssa_scalar s
)
2208 return nir_instr_as_alu(s
.def
->parent_instr
)->op
;
2211 static inline nir_ssa_scalar
2212 nir_ssa_scalar_chase_alu_src(nir_ssa_scalar s
, unsigned alu_src_idx
)
2214 nir_ssa_scalar out
= { NULL
, 0 };
2216 nir_alu_instr
*alu
= nir_instr_as_alu(s
.def
->parent_instr
);
2217 assert(alu_src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
2219 /* Our component must be written */
2220 assert(s
.comp
< s
.def
->num_components
);
2221 assert(alu
->dest
.write_mask
& (1u << s
.comp
));
2223 assert(alu
->src
[alu_src_idx
].src
.is_ssa
);
2224 out
.def
= alu
->src
[alu_src_idx
].src
.ssa
;
2226 if (nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 0) {
2227 /* The ALU src is unsized so the source component follows the
2228 * destination component.
2230 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[s
.comp
];
2232 /* This is a sized source so all source components work together to
2233 * produce all the destination components. Since we need to return a
2234 * scalar, this only works if the source is a scalar.
2236 assert(nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 1);
2237 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[0];
2239 assert(out
.comp
< out
.def
->num_components
);
2248 * Control flow consists of a tree of control flow nodes, which include
2249 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2250 * instructions that always run start-to-finish. Each basic block also keeps
2251 * track of its successors (blocks which may run immediately after the current
2252 * block) and predecessors (blocks which could have run immediately before the
2253 * current block). Each function also has a start block and an end block which
2254 * all return statements point to (which is always empty). Together, all the
2255 * blocks with their predecessors and successors make up the control flow
2256 * graph (CFG) of the function. There are helpers that modify the tree of
2257 * control flow nodes while modifying the CFG appropriately; these should be
2258 * used instead of modifying the tree directly.
2265 nir_cf_node_function
2268 typedef struct nir_cf_node
{
2269 struct exec_node node
;
2270 nir_cf_node_type type
;
2271 struct nir_cf_node
*parent
;
2274 typedef struct nir_block
{
2275 nir_cf_node cf_node
;
2277 struct exec_list instr_list
; /** < list of nir_instr */
2279 /** generic block index; generated by nir_index_blocks */
2283 * Each block can only have up to 2 successors, so we put them in a simple
2284 * array - no need for anything more complicated.
2286 struct nir_block
*successors
[2];
2288 /* Set of nir_block predecessors in the CFG */
2289 struct set
*predecessors
;
2292 * this node's immediate dominator in the dominance tree - set to NULL for
2295 struct nir_block
*imm_dom
;
2297 /* This node's children in the dominance tree */
2298 unsigned num_dom_children
;
2299 struct nir_block
**dom_children
;
2301 /* Set of nir_blocks on the dominance frontier of this block */
2302 struct set
*dom_frontier
;
2305 * These two indices have the property that dom_{pre,post}_index for each
2306 * child of this block in the dominance tree will always be between
2307 * dom_pre_index and dom_post_index for this block, which makes testing if
2308 * a given block is dominated by another block an O(1) operation.
2310 unsigned dom_pre_index
, dom_post_index
;
2312 /* live in and out for this block; used for liveness analysis */
2313 BITSET_WORD
*live_in
;
2314 BITSET_WORD
*live_out
;
2317 static inline nir_instr
*
2318 nir_block_first_instr(nir_block
*block
)
2320 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
2321 return exec_node_data(nir_instr
, head
, node
);
2324 static inline nir_instr
*
2325 nir_block_last_instr(nir_block
*block
)
2327 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
2328 return exec_node_data(nir_instr
, tail
, node
);
2332 nir_block_ends_in_jump(nir_block
*block
)
2334 return !exec_list_is_empty(&block
->instr_list
) &&
2335 nir_block_last_instr(block
)->type
== nir_instr_type_jump
;
2338 #define nir_foreach_instr(instr, block) \
2339 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
2340 #define nir_foreach_instr_reverse(instr, block) \
2341 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
2342 #define nir_foreach_instr_safe(instr, block) \
2343 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
2344 #define nir_foreach_instr_reverse_safe(instr, block) \
2345 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
2348 nir_selection_control_none
= 0x0,
2349 nir_selection_control_flatten
= 0x1,
2350 nir_selection_control_dont_flatten
= 0x2,
2351 } nir_selection_control
;
2353 typedef struct nir_if
{
2354 nir_cf_node cf_node
;
2356 nir_selection_control control
;
2358 struct exec_list then_list
; /** < list of nir_cf_node */
2359 struct exec_list else_list
; /** < list of nir_cf_node */
2365 /** Instruction that generates nif::condition. */
2366 nir_instr
*conditional_instr
;
2368 /** Block within ::nif that has the break instruction. */
2369 nir_block
*break_block
;
2371 /** Last block for the then- or else-path that does not contain the break. */
2372 nir_block
*continue_from_block
;
2374 /** True when ::break_block is in the else-path of ::nif. */
2375 bool continue_from_then
;
2378 /* This is true if the terminators exact trip count is unknown. For
2381 * for (int i = 0; i < imin(x, 4); i++)
2384 * Here loop analysis would have set a max_trip_count of 4 however we dont
2385 * know for sure that this is the exact trip count.
2387 bool exact_trip_count_unknown
;
2389 struct list_head loop_terminator_link
;
2390 } nir_loop_terminator
;
2393 /* Estimated cost (in number of instructions) of the loop */
2394 unsigned instr_cost
;
2396 /* Guessed trip count based on array indexing */
2397 unsigned guessed_trip_count
;
2399 /* Maximum number of times the loop is run (if known) */
2400 unsigned max_trip_count
;
2402 /* Do we know the exact number of times the loop will be run */
2403 bool exact_trip_count_known
;
2405 /* Unroll the loop regardless of its size */
2408 /* Does the loop contain complex loop terminators, continues or other
2409 * complex behaviours? If this is true we can't rely on
2410 * loop_terminator_list to be complete or accurate.
2414 nir_loop_terminator
*limiting_terminator
;
2416 /* A list of loop_terminators terminating this loop. */
2417 struct list_head loop_terminator_list
;
2421 nir_loop_control_none
= 0x0,
2422 nir_loop_control_unroll
= 0x1,
2423 nir_loop_control_dont_unroll
= 0x2,
2427 nir_cf_node cf_node
;
2429 struct exec_list body
; /** < list of nir_cf_node */
2431 nir_loop_info
*info
;
2432 nir_loop_control control
;
2433 bool partially_unrolled
;
2437 * Various bits of metadata that can may be created or required by
2438 * optimization and analysis passes
2441 nir_metadata_none
= 0x0,
2442 nir_metadata_block_index
= 0x1,
2443 nir_metadata_dominance
= 0x2,
2444 nir_metadata_live_ssa_defs
= 0x4,
2445 nir_metadata_not_properly_reset
= 0x8,
2446 nir_metadata_loop_analysis
= 0x10,
2450 nir_cf_node cf_node
;
2452 /** pointer to the function of which this is an implementation */
2453 struct nir_function
*function
;
2455 struct exec_list body
; /** < list of nir_cf_node */
2457 nir_block
*end_block
;
2459 /** list for all local variables in the function */
2460 struct exec_list locals
;
2462 /** list of local registers in the function */
2463 struct exec_list registers
;
2465 /** next available local register index */
2468 /** next available SSA value index */
2471 /* total number of basic blocks, only valid when block_index_dirty = false */
2472 unsigned num_blocks
;
2474 nir_metadata valid_metadata
;
2475 } nir_function_impl
;
2477 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2478 nir_start_block(nir_function_impl
*impl
)
2480 return (nir_block
*) impl
->body
.head_sentinel
.next
;
2483 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2484 nir_impl_last_block(nir_function_impl
*impl
)
2486 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
2489 static inline nir_cf_node
*
2490 nir_cf_node_next(nir_cf_node
*node
)
2492 struct exec_node
*next
= exec_node_get_next(&node
->node
);
2493 if (exec_node_is_tail_sentinel(next
))
2496 return exec_node_data(nir_cf_node
, next
, node
);
2499 static inline nir_cf_node
*
2500 nir_cf_node_prev(nir_cf_node
*node
)
2502 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
2503 if (exec_node_is_head_sentinel(prev
))
2506 return exec_node_data(nir_cf_node
, prev
, node
);
2510 nir_cf_node_is_first(const nir_cf_node
*node
)
2512 return exec_node_is_head_sentinel(node
->node
.prev
);
2516 nir_cf_node_is_last(const nir_cf_node
*node
)
2518 return exec_node_is_tail_sentinel(node
->node
.next
);
2521 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
2522 type
, nir_cf_node_block
)
2523 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
2524 type
, nir_cf_node_if
)
2525 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
2526 type
, nir_cf_node_loop
)
2527 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
2528 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
2530 static inline nir_block
*
2531 nir_if_first_then_block(nir_if
*if_stmt
)
2533 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
2534 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2537 static inline nir_block
*
2538 nir_if_last_then_block(nir_if
*if_stmt
)
2540 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
2541 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2544 static inline nir_block
*
2545 nir_if_first_else_block(nir_if
*if_stmt
)
2547 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
2548 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2551 static inline nir_block
*
2552 nir_if_last_else_block(nir_if
*if_stmt
)
2554 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
2555 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2558 static inline nir_block
*
2559 nir_loop_first_block(nir_loop
*loop
)
2561 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
2562 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2565 static inline nir_block
*
2566 nir_loop_last_block(nir_loop
*loop
)
2568 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
2569 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2573 * Return true if this list of cf_nodes contains a single empty block.
2576 nir_cf_list_is_empty_block(struct exec_list
*cf_list
)
2578 if (exec_list_is_singular(cf_list
)) {
2579 struct exec_node
*head
= exec_list_get_head(cf_list
);
2581 nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2582 return exec_list_is_empty(&block
->instr_list
);
2588 uint8_t num_components
;
2592 typedef struct nir_function
{
2593 struct exec_node node
;
2596 struct nir_shader
*shader
;
2598 unsigned num_params
;
2599 nir_parameter
*params
;
2601 /** The implementation of this function.
2603 * If the function is only declared and not implemented, this is NULL.
2605 nir_function_impl
*impl
;
2611 nir_lower_imul64
= (1 << 0),
2612 nir_lower_isign64
= (1 << 1),
2613 /** Lower all int64 modulus and division opcodes */
2614 nir_lower_divmod64
= (1 << 2),
2615 /** Lower all 64-bit umul_high and imul_high opcodes */
2616 nir_lower_imul_high64
= (1 << 3),
2617 nir_lower_mov64
= (1 << 4),
2618 nir_lower_icmp64
= (1 << 5),
2619 nir_lower_iadd64
= (1 << 6),
2620 nir_lower_iabs64
= (1 << 7),
2621 nir_lower_ineg64
= (1 << 8),
2622 nir_lower_logic64
= (1 << 9),
2623 nir_lower_minmax64
= (1 << 10),
2624 nir_lower_shift64
= (1 << 11),
2625 nir_lower_imul_2x32_64
= (1 << 12),
2626 nir_lower_extract64
= (1 << 13),
2627 nir_lower_ufind_msb64
= (1 << 14),
2628 } nir_lower_int64_options
;
2631 nir_lower_drcp
= (1 << 0),
2632 nir_lower_dsqrt
= (1 << 1),
2633 nir_lower_drsq
= (1 << 2),
2634 nir_lower_dtrunc
= (1 << 3),
2635 nir_lower_dfloor
= (1 << 4),
2636 nir_lower_dceil
= (1 << 5),
2637 nir_lower_dfract
= (1 << 6),
2638 nir_lower_dround_even
= (1 << 7),
2639 nir_lower_dmod
= (1 << 8),
2640 nir_lower_dsub
= (1 << 9),
2641 nir_lower_ddiv
= (1 << 10),
2642 nir_lower_fp64_full_software
= (1 << 11),
2643 } nir_lower_doubles_options
;
2646 nir_divergence_single_prim_per_subgroup
= (1 << 0),
2647 nir_divergence_single_patch_per_tcs_subgroup
= (1 << 1),
2648 nir_divergence_single_patch_per_tes_subgroup
= (1 << 2),
2649 nir_divergence_view_index_uniform
= (1 << 3),
2650 } nir_divergence_options
;
2652 typedef struct nir_shader_compiler_options
{
2658 /** Lowers flrp when it does not support doubles */
2665 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2666 bool lower_bitfield_extract
;
2667 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
2668 bool lower_bitfield_extract_to_shifts
;
2669 /** Lowers bitfield_insert to bfi/bfm */
2670 bool lower_bitfield_insert
;
2671 /** Lowers bitfield_insert to compares, and shifts. */
2672 bool lower_bitfield_insert_to_shifts
;
2673 /** Lowers bitfield_insert to bfm/bitfield_select. */
2674 bool lower_bitfield_insert_to_bitfield_select
;
2675 /** Lowers bitfield_reverse to shifts. */
2676 bool lower_bitfield_reverse
;
2677 /** Lowers bit_count to shifts. */
2678 bool lower_bit_count
;
2679 /** Lowers ifind_msb to compare and ufind_msb */
2680 bool lower_ifind_msb
;
2681 /** Lowers find_lsb to ufind_msb and logic ops */
2682 bool lower_find_lsb
;
2683 bool lower_uadd_carry
;
2684 bool lower_usub_borrow
;
2685 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2686 bool lower_mul_high
;
2687 /** lowers fneg and ineg to fsub and isub. */
2689 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2692 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2695 /* lower fall_equalN/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
2696 bool lower_vector_cmp
;
2698 /** enables rules to lower idiv by power-of-two: */
2701 /** enable rules to avoid bit ops */
2704 /** enables rules to lower isign to imin+imax */
2707 /** enables rules to lower fsign to fsub and flt */
2710 /* lower fdph to fdot4 */
2713 /** lower fdot to fmul and fsum/fadd. */
2716 /* Does the native fdot instruction replicate its result for four
2717 * components? If so, then opt_algebraic_late will turn all fdotN
2718 * instructions into fdot_replicatedN instructions.
2720 bool fdot_replicates
;
2722 /** lowers ffloor to fsub+ffract: */
2725 /** lowers ffract to fsub+ffloor: */
2728 /** lowers fceil to fneg+ffloor+fneg: */
2735 bool lower_pack_half_2x16
;
2736 bool lower_pack_unorm_2x16
;
2737 bool lower_pack_snorm_2x16
;
2738 bool lower_pack_unorm_4x8
;
2739 bool lower_pack_snorm_4x8
;
2740 bool lower_unpack_half_2x16
;
2741 bool lower_unpack_unorm_2x16
;
2742 bool lower_unpack_snorm_2x16
;
2743 bool lower_unpack_unorm_4x8
;
2744 bool lower_unpack_snorm_4x8
;
2746 bool lower_extract_byte
;
2747 bool lower_extract_word
;
2749 bool lower_all_io_to_temps
;
2750 bool lower_all_io_to_elements
;
2752 /* Indicates that the driver only has zero-based vertex id */
2753 bool vertex_id_zero_based
;
2756 * If enabled, gl_BaseVertex will be lowered as:
2757 * is_indexed_draw (~0/0) & firstvertex
2759 bool lower_base_vertex
;
2762 * If enabled, gl_HelperInvocation will be lowered as:
2764 * !((1 << sample_id) & sample_mask_in))
2766 * This depends on some possibly hw implementation details, which may
2767 * not be true for all hw. In particular that the FS is only executed
2768 * for covered samples or for helper invocations. So, do not blindly
2769 * enable this option.
2771 * Note: See also issue #22 in ARB_shader_image_load_store
2773 bool lower_helper_invocation
;
2776 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
2778 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
2779 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
2781 bool optimize_sample_mask_in
;
2783 bool lower_cs_local_index_from_id
;
2784 bool lower_cs_local_id_from_index
;
2786 bool lower_device_index_to_zero
;
2788 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
2789 bool lower_wpos_pntc
;
2795 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
2796 * for IO purposes and would prefer loads/stores be vectorized.
2799 bool lower_to_scalar
;
2802 * Should nir_lower_io() create load_interpolated_input intrinsics?
2804 * If not, it generates regular load_input intrinsics and interpolation
2805 * information must be inferred from the list of input nir_variables.
2807 bool use_interpolated_input_intrinsics
;
2809 /* Lowers when 32x32->64 bit multiplication is not supported */
2810 bool lower_mul_2x32_64
;
2812 /* Lowers when rotate instruction is not supported */
2816 * Backend supports imul24, and would like to use it (when possible)
2817 * for address/offset calculation. If true, driver should call
2818 * nir_lower_amul(). (If not set, amul will automatically be lowered
2824 * Is this the Intel vec4 backend?
2826 * Used to inhibit algebraic optimizations that are known to be harmful on
2827 * the Intel vec4 backend. This is generally applicable to any
2828 * optimization that might cause more immediate values to be used in
2829 * 3-source (e.g., ffma and flrp) instructions.
2833 unsigned max_unroll_iterations
;
2835 nir_lower_int64_options lower_int64_options
;
2836 nir_lower_doubles_options lower_doubles_options
;
2837 } nir_shader_compiler_options
;
2839 typedef struct nir_shader
{
2840 /** list of uniforms (nir_variable) */
2841 struct exec_list uniforms
;
2843 /** list of inputs (nir_variable) */
2844 struct exec_list inputs
;
2846 /** list of outputs (nir_variable) */
2847 struct exec_list outputs
;
2849 /** list of shared compute variables (nir_variable) */
2850 struct exec_list shared
;
2852 /** Set of driver-specific options for the shader.
2854 * The memory for the options is expected to be kept in a single static
2855 * copy by the driver.
2857 const struct nir_shader_compiler_options
*options
;
2859 /** Various bits of compile-time information about a given shader */
2860 struct shader_info info
;
2862 /** list of global variables in the shader (nir_variable) */
2863 struct exec_list globals
;
2865 /** list of system value variables in the shader (nir_variable) */
2866 struct exec_list system_values
;
2868 struct exec_list functions
; /** < list of nir_function */
2871 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
2874 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
2876 /** Size in bytes of required scratch space */
2877 unsigned scratch_size
;
2879 /** Constant data associated with this shader.
2881 * Constant data is loaded through load_constant intrinsics. See also
2882 * nir_opt_large_constants.
2884 void *constant_data
;
2885 unsigned constant_data_size
;
2888 #define nir_foreach_function(func, shader) \
2889 foreach_list_typed(nir_function, func, node, &(shader)->functions)
2891 static inline nir_function_impl
*
2892 nir_shader_get_entrypoint(nir_shader
*shader
)
2894 nir_function
*func
= NULL
;
2896 nir_foreach_function(function
, shader
) {
2897 assert(func
== NULL
);
2898 if (function
->is_entrypoint
) {
2909 assert(func
->num_params
== 0);
2914 nir_shader
*nir_shader_create(void *mem_ctx
,
2915 gl_shader_stage stage
,
2916 const nir_shader_compiler_options
*options
,
2919 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
2921 void nir_reg_remove(nir_register
*reg
);
2923 /** Adds a variable to the appropriate list in nir_shader */
2924 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
2927 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
2929 assert(var
->data
.mode
== nir_var_function_temp
);
2930 exec_list_push_tail(&impl
->locals
, &var
->node
);
2933 /** creates a variable, sets a few defaults, and adds it to the list */
2934 nir_variable
*nir_variable_create(nir_shader
*shader
,
2935 nir_variable_mode mode
,
2936 const struct glsl_type
*type
,
2938 /** creates a local variable and adds it to the list */
2939 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
2940 const struct glsl_type
*type
,
2943 /** creates a function and adds it to the shader's list of functions */
2944 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
2946 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
2947 /** creates a function_impl that isn't tied to any particular function */
2948 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
2950 nir_block
*nir_block_create(nir_shader
*shader
);
2951 nir_if
*nir_if_create(nir_shader
*shader
);
2952 nir_loop
*nir_loop_create(nir_shader
*shader
);
2954 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
2956 /** requests that the given pieces of metadata be generated */
2957 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
2958 /** dirties all but the preserved metadata */
2959 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
2961 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
2962 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
2964 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
2965 nir_deref_type deref_type
);
2967 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
2969 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
2970 unsigned num_components
,
2973 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
2974 nir_intrinsic_op op
);
2976 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
2977 nir_function
*callee
);
2979 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
2981 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
2983 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
2985 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
2986 unsigned num_components
,
2989 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
2992 * NIR Cursors and Instruction Insertion API
2995 * A tiny struct representing a point to insert/extract instructions or
2996 * control flow nodes. Helps reduce the combinatorial explosion of possible
2997 * points to insert/extract.
2999 * \sa nir_control_flow.h
3002 nir_cursor_before_block
,
3003 nir_cursor_after_block
,
3004 nir_cursor_before_instr
,
3005 nir_cursor_after_instr
,
3006 } nir_cursor_option
;
3009 nir_cursor_option option
;
3016 static inline nir_block
*
3017 nir_cursor_current_block(nir_cursor cursor
)
3019 if (cursor
.option
== nir_cursor_before_instr
||
3020 cursor
.option
== nir_cursor_after_instr
) {
3021 return cursor
.instr
->block
;
3023 return cursor
.block
;
3027 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
3029 static inline nir_cursor
3030 nir_before_block(nir_block
*block
)
3033 cursor
.option
= nir_cursor_before_block
;
3034 cursor
.block
= block
;
3038 static inline nir_cursor
3039 nir_after_block(nir_block
*block
)
3042 cursor
.option
= nir_cursor_after_block
;
3043 cursor
.block
= block
;
3047 static inline nir_cursor
3048 nir_before_instr(nir_instr
*instr
)
3051 cursor
.option
= nir_cursor_before_instr
;
3052 cursor
.instr
= instr
;
3056 static inline nir_cursor
3057 nir_after_instr(nir_instr
*instr
)
3060 cursor
.option
= nir_cursor_after_instr
;
3061 cursor
.instr
= instr
;
3065 static inline nir_cursor
3066 nir_after_block_before_jump(nir_block
*block
)
3068 nir_instr
*last_instr
= nir_block_last_instr(block
);
3069 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
3070 return nir_before_instr(last_instr
);
3072 return nir_after_block(block
);
3076 static inline nir_cursor
3077 nir_before_src(nir_src
*src
, bool is_if_condition
)
3079 if (is_if_condition
) {
3080 nir_block
*prev_block
=
3081 nir_cf_node_as_block(nir_cf_node_prev(&src
->parent_if
->cf_node
));
3082 assert(!nir_block_ends_in_jump(prev_block
));
3083 return nir_after_block(prev_block
);
3084 } else if (src
->parent_instr
->type
== nir_instr_type_phi
) {
3086 nir_phi_instr
*cond_phi
= nir_instr_as_phi(src
->parent_instr
);
3088 nir_foreach_phi_src(phi_src
, cond_phi
) {
3089 if (phi_src
->src
.ssa
== src
->ssa
) {
3096 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
3097 * to have a more specific name.
3099 nir_phi_src
*phi_src
= LIST_ENTRY(nir_phi_src
, src
, src
);
3100 return nir_after_block_before_jump(phi_src
->pred
);
3102 return nir_before_instr(src
->parent_instr
);
3106 static inline nir_cursor
3107 nir_before_cf_node(nir_cf_node
*node
)
3109 if (node
->type
== nir_cf_node_block
)
3110 return nir_before_block(nir_cf_node_as_block(node
));
3112 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
3115 static inline nir_cursor
3116 nir_after_cf_node(nir_cf_node
*node
)
3118 if (node
->type
== nir_cf_node_block
)
3119 return nir_after_block(nir_cf_node_as_block(node
));
3121 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
3124 static inline nir_cursor
3125 nir_after_phis(nir_block
*block
)
3127 nir_foreach_instr(instr
, block
) {
3128 if (instr
->type
!= nir_instr_type_phi
)
3129 return nir_before_instr(instr
);
3131 return nir_after_block(block
);
3134 static inline nir_cursor
3135 nir_after_cf_node_and_phis(nir_cf_node
*node
)
3137 if (node
->type
== nir_cf_node_block
)
3138 return nir_after_block(nir_cf_node_as_block(node
));
3140 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
3142 return nir_after_phis(block
);
3145 static inline nir_cursor
3146 nir_before_cf_list(struct exec_list
*cf_list
)
3148 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
3149 exec_list_get_head(cf_list
), node
);
3150 return nir_before_cf_node(first_node
);
3153 static inline nir_cursor
3154 nir_after_cf_list(struct exec_list
*cf_list
)
3156 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
3157 exec_list_get_tail(cf_list
), node
);
3158 return nir_after_cf_node(last_node
);
3162 * Insert a NIR instruction at the given cursor.
3164 * Note: This does not update the cursor.
3166 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
3169 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
3171 nir_instr_insert(nir_before_instr(instr
), before
);
3175 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
3177 nir_instr_insert(nir_after_instr(instr
), after
);
3181 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
3183 nir_instr_insert(nir_before_block(block
), before
);
3187 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
3189 nir_instr_insert(nir_after_block(block
), after
);
3193 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
3195 nir_instr_insert(nir_before_cf_node(node
), before
);
3199 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
3201 nir_instr_insert(nir_after_cf_node(node
), after
);
3205 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
3207 nir_instr_insert(nir_before_cf_list(list
), before
);
3211 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
3213 nir_instr_insert(nir_after_cf_list(list
), after
);
3216 void nir_instr_remove_v(nir_instr
*instr
);
3218 static inline nir_cursor
3219 nir_instr_remove(nir_instr
*instr
)
3222 nir_instr
*prev
= nir_instr_prev(instr
);
3224 cursor
= nir_after_instr(prev
);
3226 cursor
= nir_before_block(instr
->block
);
3228 nir_instr_remove_v(instr
);
3234 nir_ssa_def
*nir_instr_ssa_def(nir_instr
*instr
);
3236 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
3237 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
3238 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
3239 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
3241 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
3242 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
3244 nir_const_value
*nir_src_as_const_value(nir_src src
);
3246 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
3247 static inline c_type * \
3248 nir_src_as_ ## name (nir_src src) \
3250 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
3251 ? cast_macro(src.ssa->parent_instr) : NULL; \
3254 NIR_SRC_AS_(alu_instr
, nir_alu_instr
, nir_instr_type_alu
, nir_instr_as_alu
)
3255 NIR_SRC_AS_(intrinsic
, nir_intrinsic_instr
,
3256 nir_instr_type_intrinsic
, nir_instr_as_intrinsic
)
3257 NIR_SRC_AS_(deref
, nir_deref_instr
, nir_instr_type_deref
, nir_instr_as_deref
)
3259 bool nir_src_is_dynamically_uniform(nir_src src
);
3260 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
3261 bool nir_instrs_equal(const nir_instr
*instr1
, const nir_instr
*instr2
);
3262 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
3263 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
3264 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
3265 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
3268 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
3269 unsigned num_components
, unsigned bit_size
,
3271 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
3272 unsigned num_components
, unsigned bit_size
,
3275 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
3276 const struct glsl_type
*type
,
3279 assert(glsl_type_is_vector_or_scalar(type
));
3280 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
3281 glsl_get_bit_size(type
), name
);
3283 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
3284 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
3285 nir_instr
*after_me
);
3287 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
3290 * finds the next basic block in source-code order, returns NULL if there is
3294 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
3296 /* Performs the opposite of nir_block_cf_tree_next() */
3298 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
3300 /* Gets the first block in a CF node in source-code order */
3302 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
3304 /* Gets the last block in a CF node in source-code order */
3306 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
3308 /* Gets the next block after a CF node in source-code order */
3310 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
3312 /* Macros for loops that visit blocks in source-code order */
3314 #define nir_foreach_block(block, impl) \
3315 for (nir_block *block = nir_start_block(impl); block != NULL; \
3316 block = nir_block_cf_tree_next(block))
3318 #define nir_foreach_block_safe(block, impl) \
3319 for (nir_block *block = nir_start_block(impl), \
3320 *next = nir_block_cf_tree_next(block); \
3322 block = next, next = nir_block_cf_tree_next(block))
3324 #define nir_foreach_block_reverse(block, impl) \
3325 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
3326 block = nir_block_cf_tree_prev(block))
3328 #define nir_foreach_block_reverse_safe(block, impl) \
3329 for (nir_block *block = nir_impl_last_block(impl), \
3330 *prev = nir_block_cf_tree_prev(block); \
3332 block = prev, prev = nir_block_cf_tree_prev(block))
3334 #define nir_foreach_block_in_cf_node(block, node) \
3335 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
3336 block != nir_cf_node_cf_tree_next(node); \
3337 block = nir_block_cf_tree_next(block))
3339 /* If the following CF node is an if, this function returns that if.
3340 * Otherwise, it returns NULL.
3342 nir_if
*nir_block_get_following_if(nir_block
*block
);
3344 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
3346 void nir_index_local_regs(nir_function_impl
*impl
);
3347 void nir_index_ssa_defs(nir_function_impl
*impl
);
3348 unsigned nir_index_instrs(nir_function_impl
*impl
);
3350 void nir_index_blocks(nir_function_impl
*impl
);
3352 void nir_index_vars(nir_shader
*shader
, nir_function_impl
*impl
, nir_variable_mode modes
);
3354 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
3355 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
3356 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
3357 void nir_print_deref(const nir_deref_instr
*deref
, FILE *fp
);
3359 /** Shallow clone of a single ALU instruction. */
3360 nir_alu_instr
*nir_alu_instr_clone(nir_shader
*s
, const nir_alu_instr
*orig
);
3362 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
3363 nir_function_impl
*nir_function_impl_clone(nir_shader
*shader
,
3364 const nir_function_impl
*fi
);
3365 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
3366 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
3368 void nir_shader_replace(nir_shader
*dest
, nir_shader
*src
);
3370 void nir_shader_serialize_deserialize(nir_shader
*s
);
3373 void nir_validate_shader(nir_shader
*shader
, const char *when
);
3374 void nir_metadata_set_validation_flag(nir_shader
*shader
);
3375 void nir_metadata_check_validation_flag(nir_shader
*shader
);
3378 should_skip_nir(const char *name
)
3380 static const char *list
= NULL
;
3382 /* Comma separated list of names to skip. */
3383 list
= getenv("NIR_SKIP");
3391 return comma_separated_list_contains(list
, name
);
3395 should_clone_nir(void)
3397 static int should_clone
= -1;
3398 if (should_clone
< 0)
3399 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
3401 return should_clone
;
3405 should_serialize_deserialize_nir(void)
3407 static int test_serialize
= -1;
3408 if (test_serialize
< 0)
3409 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
3411 return test_serialize
;
3415 should_print_nir(void)
3417 static int should_print
= -1;
3418 if (should_print
< 0)
3419 should_print
= env_var_as_boolean("NIR_PRINT", false);
3421 return should_print
;
3424 static inline void nir_validate_shader(nir_shader
*shader
, const char *when
) { (void) shader
; (void)when
; }
3425 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
3426 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
3427 static inline bool should_skip_nir(UNUSED
const char *pass_name
) { return false; }
3428 static inline bool should_clone_nir(void) { return false; }
3429 static inline bool should_serialize_deserialize_nir(void) { return false; }
3430 static inline bool should_print_nir(void) { return false; }
3433 #define _PASS(pass, nir, do_pass) do { \
3434 if (should_skip_nir(#pass)) { \
3435 printf("skipping %s\n", #pass); \
3439 nir_validate_shader(nir, "after " #pass); \
3440 if (should_clone_nir()) { \
3441 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3442 nir_shader_replace(nir, clone); \
3444 if (should_serialize_deserialize_nir()) { \
3445 nir_shader_serialize_deserialize(nir); \
3449 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3450 nir_metadata_set_validation_flag(nir); \
3451 if (should_print_nir()) \
3452 printf("%s\n", #pass); \
3453 if (pass(nir, ##__VA_ARGS__)) { \
3455 if (should_print_nir()) \
3456 nir_print_shader(nir, stdout); \
3457 nir_metadata_check_validation_flag(nir); \
3461 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3462 if (should_print_nir()) \
3463 printf("%s\n", #pass); \
3464 pass(nir, ##__VA_ARGS__); \
3465 if (should_print_nir()) \
3466 nir_print_shader(nir, stdout); \
3469 #define NIR_SKIP(name) should_skip_nir(#name)
3471 /** An instruction filtering callback
3473 * Returns true if the instruction should be processed and false otherwise.
3475 typedef bool (*nir_instr_filter_cb
)(const nir_instr
*, const void *);
3477 /** A simple instruction lowering callback
3479 * Many instruction lowering passes can be written as a simple function which
3480 * takes an instruction as its input and returns a sequence of instructions
3481 * that implement the consumed instruction. This function type represents
3482 * such a lowering function. When called, a function with this prototype
3483 * should either return NULL indicating that no lowering needs to be done or
3484 * emit a sequence of instructions using the provided builder (whose cursor
3485 * will already be placed after the instruction to be lowered) and return the
3486 * resulting nir_ssa_def.
3488 typedef nir_ssa_def
*(*nir_lower_instr_cb
)(struct nir_builder
*,
3489 nir_instr
*, void *);
3492 * Special return value for nir_lower_instr_cb when some progress occurred
3493 * (like changing an input to the instr) that didn't result in a replacement
3494 * SSA def being generated.
3496 #define NIR_LOWER_INSTR_PROGRESS ((nir_ssa_def *)(uintptr_t)1)
3498 /** Iterate over all the instructions in a nir_function_impl and lower them
3499 * using the provided callbacks
3501 * This function implements the guts of a standard lowering pass for you. It
3502 * iterates over all of the instructions in a nir_function_impl and calls the
3503 * filter callback on each one. If the filter callback returns true, it then
3504 * calls the lowering call back on the instruction. (Splitting it this way
3505 * allows us to avoid some save/restore work for instructions we know won't be
3506 * lowered.) If the instruction is dead after the lowering is complete, it
3507 * will be removed. If new instructions are added, the lowering callback will
3508 * also be called on them in case multiple lowerings are required.
3510 * The metadata for the nir_function_impl will also be updated. If any blocks
3511 * are added (they cannot be removed), dominance and block indices will be
3514 bool nir_function_impl_lower_instructions(nir_function_impl
*impl
,
3515 nir_instr_filter_cb filter
,
3516 nir_lower_instr_cb lower
,
3518 bool nir_shader_lower_instructions(nir_shader
*shader
,
3519 nir_instr_filter_cb filter
,
3520 nir_lower_instr_cb lower
,
3523 void nir_calc_dominance_impl(nir_function_impl
*impl
);
3524 void nir_calc_dominance(nir_shader
*shader
);
3526 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
3527 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
3528 bool nir_block_is_unreachable(nir_block
*block
);
3530 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
3531 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
3533 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
3534 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
3536 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
3537 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
3539 int nir_gs_count_vertices(const nir_shader
*shader
);
3541 bool nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3542 bool nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3543 bool nir_split_var_copies(nir_shader
*shader
);
3544 bool nir_split_per_member_structs(nir_shader
*shader
);
3545 bool nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
);
3547 bool nir_lower_returns_impl(nir_function_impl
*impl
);
3548 bool nir_lower_returns(nir_shader
*shader
);
3550 void nir_inline_function_impl(struct nir_builder
*b
,
3551 const nir_function_impl
*impl
,
3552 nir_ssa_def
**params
);
3553 bool nir_inline_functions(nir_shader
*shader
);
3555 bool nir_propagate_invariant(nir_shader
*shader
);
3557 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
3558 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
3559 nir_intrinsic_instr
*copy
);
3560 bool nir_lower_var_copies(nir_shader
*shader
);
3562 void nir_fixup_deref_modes(nir_shader
*shader
);
3564 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
3567 nir_lower_direct_array_deref_of_vec_load
= (1 << 0),
3568 nir_lower_indirect_array_deref_of_vec_load
= (1 << 1),
3569 nir_lower_direct_array_deref_of_vec_store
= (1 << 2),
3570 nir_lower_indirect_array_deref_of_vec_store
= (1 << 3),
3571 } nir_lower_array_deref_of_vec_options
;
3573 bool nir_lower_array_deref_of_vec(nir_shader
*shader
, nir_variable_mode modes
,
3574 nir_lower_array_deref_of_vec_options options
);
3576 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
3578 bool nir_lower_locals_to_regs(nir_shader
*shader
);
3580 void nir_lower_io_to_temporaries(nir_shader
*shader
,
3581 nir_function_impl
*entrypoint
,
3582 bool outputs
, bool inputs
);
3584 bool nir_lower_vars_to_scratch(nir_shader
*shader
,
3585 nir_variable_mode modes
,
3587 glsl_type_size_align_func size_align
);
3589 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
3591 void nir_gather_ssa_types(nir_function_impl
*impl
,
3592 BITSET_WORD
*float_types
,
3593 BITSET_WORD
*int_types
);
3595 void nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
3596 int (*type_size
)(const struct glsl_type
*, bool));
3598 /* Some helpers to do very simple linking */
3599 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3600 bool nir_remove_unused_io_vars(nir_shader
*shader
, struct exec_list
*var_list
,
3601 uint64_t *used_by_other_stage
,
3602 uint64_t *used_by_other_stage_patches
);
3603 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
3604 bool default_to_smooth_interp
);
3605 void nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3606 bool nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3608 bool nir_lower_amul(nir_shader
*shader
,
3609 int (*type_size
)(const struct glsl_type
*, bool));
3611 void nir_assign_io_var_locations(struct exec_list
*var_list
,
3613 gl_shader_stage stage
);
3616 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
3617 * to 32-bit operations. This is only valid for nir_var_shader_in/out
3620 nir_lower_io_lower_64bit_to_32
= (1 << 0),
3622 /* If set, this forces all non-flat fragment shader inputs to be
3623 * interpolated as if with the "sample" qualifier. This requires
3624 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
3626 nir_lower_io_force_sample_interpolation
= (1 << 1),
3627 } nir_lower_io_options
;
3628 bool nir_lower_io(nir_shader
*shader
,
3629 nir_variable_mode modes
,
3630 int (*type_size
)(const struct glsl_type
*, bool),
3631 nir_lower_io_options
);
3633 bool nir_io_add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
);
3636 nir_lower_vars_to_explicit_types(nir_shader
*shader
,
3637 nir_variable_mode modes
,
3638 glsl_type_size_align_func type_info
);
3642 * An address format which is a simple 32-bit global GPU address.
3644 nir_address_format_32bit_global
,
3647 * An address format which is a simple 64-bit global GPU address.
3649 nir_address_format_64bit_global
,
3652 * An address format which is a bounds-checked 64-bit global GPU address.
3654 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
3655 * address stored with the low bits in .x and high bits in .y, .z is a
3656 * size, and .w is an offset. When the final I/O operation is lowered, .w
3657 * is checked against .z and the operation is predicated on the result.
3659 nir_address_format_64bit_bounded_global
,
3662 * An address format which is comprised of a vec2 where the first
3663 * component is a buffer index and the second is an offset.
3665 nir_address_format_32bit_index_offset
,
3668 * An address format which is a simple 32-bit offset.
3670 nir_address_format_32bit_offset
,
3673 * An address format representing a purely logical addressing model. In
3674 * this model, all deref chains must be complete from the dereference
3675 * operation to the variable. Cast derefs are not allowed. These
3676 * addresses will be 32-bit scalars but the format is immaterial because
3677 * you can always chase the chain.
3679 nir_address_format_logical
,
3680 } nir_address_format
;
3682 static inline unsigned
3683 nir_address_format_bit_size(nir_address_format addr_format
)
3685 switch (addr_format
) {
3686 case nir_address_format_32bit_global
: return 32;
3687 case nir_address_format_64bit_global
: return 64;
3688 case nir_address_format_64bit_bounded_global
: return 32;
3689 case nir_address_format_32bit_index_offset
: return 32;
3690 case nir_address_format_32bit_offset
: return 32;
3691 case nir_address_format_logical
: return 32;
3693 unreachable("Invalid address format");
3696 static inline unsigned
3697 nir_address_format_num_components(nir_address_format addr_format
)
3699 switch (addr_format
) {
3700 case nir_address_format_32bit_global
: return 1;
3701 case nir_address_format_64bit_global
: return 1;
3702 case nir_address_format_64bit_bounded_global
: return 4;
3703 case nir_address_format_32bit_index_offset
: return 2;
3704 case nir_address_format_32bit_offset
: return 1;
3705 case nir_address_format_logical
: return 1;
3707 unreachable("Invalid address format");
3710 static inline const struct glsl_type
*
3711 nir_address_format_to_glsl_type(nir_address_format addr_format
)
3713 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
3714 assert(bit_size
== 32 || bit_size
== 64);
3715 return glsl_vector_type(bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
,
3716 nir_address_format_num_components(addr_format
));
3719 const nir_const_value
*nir_address_format_null_value(nir_address_format addr_format
);
3721 nir_ssa_def
*nir_build_addr_ieq(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3722 nir_address_format addr_format
);
3724 nir_ssa_def
*nir_build_addr_isub(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3725 nir_address_format addr_format
);
3727 nir_ssa_def
* nir_explicit_io_address_from_deref(struct nir_builder
*b
,
3728 nir_deref_instr
*deref
,
3729 nir_ssa_def
*base_addr
,
3730 nir_address_format addr_format
);
3731 void nir_lower_explicit_io_instr(struct nir_builder
*b
,
3732 nir_intrinsic_instr
*io_instr
,
3734 nir_address_format addr_format
);
3736 bool nir_lower_explicit_io(nir_shader
*shader
,
3737 nir_variable_mode modes
,
3738 nir_address_format
);
3740 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
3741 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
3743 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
3745 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
3746 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
3747 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
3749 bool nir_remove_dead_derefs(nir_shader
*shader
);
3750 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
3751 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
);
3752 bool nir_lower_constant_initializers(nir_shader
*shader
,
3753 nir_variable_mode modes
);
3755 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
3756 bool nir_lower_vec_to_movs(nir_shader
*shader
);
3757 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
3759 const gl_state_index16
*alpha_ref_state_tokens
);
3760 bool nir_lower_alu(nir_shader
*shader
);
3762 bool nir_lower_flrp(nir_shader
*shader
, unsigned lowering_mask
,
3763 bool always_precise
, bool have_ffma
);
3765 bool nir_lower_alu_to_scalar(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
3766 bool nir_lower_bool_to_float(nir_shader
*shader
);
3767 bool nir_lower_bool_to_int32(nir_shader
*shader
);
3768 bool nir_lower_int_to_float(nir_shader
*shader
);
3769 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
3770 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
3771 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
3772 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
3773 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
3775 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
3776 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
3777 bool nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode mask
);
3779 void nir_lower_fragcoord_wtrans(nir_shader
*shader
);
3780 void nir_lower_viewport_transform(nir_shader
*shader
);
3781 bool nir_lower_uniforms_to_ubo(nir_shader
*shader
, int multiplier
);
3783 typedef struct nir_lower_subgroups_options
{
3784 uint8_t subgroup_size
;
3785 uint8_t ballot_bit_size
;
3786 bool lower_to_scalar
:1;
3787 bool lower_vote_trivial
:1;
3788 bool lower_vote_eq_to_ballot
:1;
3789 bool lower_subgroup_masks
:1;
3790 bool lower_shuffle
:1;
3791 bool lower_shuffle_to_32bit
:1;
3793 } nir_lower_subgroups_options
;
3795 bool nir_lower_subgroups(nir_shader
*shader
,
3796 const nir_lower_subgroups_options
*options
);
3798 bool nir_lower_system_values(nir_shader
*shader
);
3800 enum PACKED nir_lower_tex_packing
{
3801 nir_lower_tex_packing_none
= 0,
3802 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
3803 * or unsigned ints based on the sampler type
3805 nir_lower_tex_packing_16
,
3806 /* The sampler returns 1 32-bit word of 4x8 unorm */
3807 nir_lower_tex_packing_8
,
3810 typedef struct nir_lower_tex_options
{
3812 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
3813 * sampler types a texture projector is lowered.
3818 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
3820 bool lower_txf_offset
;
3823 * If true, lower away nir_tex_src_offset for all rect textures.
3825 bool lower_rect_offset
;
3828 * If true, lower rect textures to 2D, using txs to fetch the
3829 * texture dimensions and dividing the texture coords by the
3830 * texture dims to normalize.
3835 * If true, convert yuv to rgb.
3837 unsigned lower_y_uv_external
;
3838 unsigned lower_y_u_v_external
;
3839 unsigned lower_yx_xuxv_external
;
3840 unsigned lower_xy_uxvx_external
;
3841 unsigned lower_ayuv_external
;
3842 unsigned lower_xyuv_external
;
3845 * To emulate certain texture wrap modes, this can be used
3846 * to saturate the specified tex coord to [0.0, 1.0]. The
3847 * bits are according to sampler #, ie. if, for example:
3849 * (conf->saturate_s & (1 << n))
3851 * is true, then the s coord for sampler n is saturated.
3853 * Note that clamping must happen *after* projector lowering
3854 * so any projected texture sample instruction with a clamped
3855 * coordinate gets automatically lowered, regardless of the
3856 * 'lower_txp' setting.
3858 unsigned saturate_s
;
3859 unsigned saturate_t
;
3860 unsigned saturate_r
;
3862 /* Bitmask of textures that need swizzling.
3864 * If (swizzle_result & (1 << texture_index)), then the swizzle in
3865 * swizzles[texture_index] is applied to the result of the texturing
3868 unsigned swizzle_result
;
3870 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
3871 * while 4 and 5 represent 0 and 1 respectively.
3873 uint8_t swizzles
[32][4];
3875 /* Can be used to scale sampled values in range required by the format. */
3876 float scale_factors
[32];
3879 * Bitmap of textures that need srgb to linear conversion. If
3880 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
3881 * of the texture are lowered to linear.
3883 unsigned lower_srgb
;
3886 * If true, lower nir_texop_tex on shaders that doesn't support implicit
3887 * LODs to nir_texop_txl.
3889 bool lower_tex_without_implicit_lod
;
3892 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
3894 bool lower_txd_cube_map
;
3897 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
3902 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
3903 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
3904 * with lower_txd_cube_map.
3906 bool lower_txd_shadow
;
3909 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
3910 * Implies lower_txd_cube_map and lower_txd_shadow.
3915 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
3916 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
3918 bool lower_txb_shadow_clamp
;
3921 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
3922 * with nir_texop_txl. This includes cube maps.
3924 bool lower_txd_shadow_clamp
;
3927 * If true, lower nir_texop_txd on when it uses both offset and min_lod
3928 * with nir_texop_txl. This includes cube maps.
3930 bool lower_txd_offset_clamp
;
3933 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
3934 * sampler is bindless.
3936 bool lower_txd_clamp_bindless_sampler
;
3939 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
3940 * sampler index is not statically determinable to be less than 16.
3942 bool lower_txd_clamp_if_sampler_index_not_lt_16
;
3945 * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
3946 * 0-lod followed by a nir_ishr.
3951 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
3952 * mixed-up tg4 locations.
3954 bool lower_tg4_broadcom_swizzle
;
3957 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
3959 bool lower_tg4_offsets
;
3961 enum nir_lower_tex_packing lower_tex_packing
[32];
3962 } nir_lower_tex_options
;
3964 bool nir_lower_tex(nir_shader
*shader
,
3965 const nir_lower_tex_options
*options
);
3967 enum nir_lower_non_uniform_access_type
{
3968 nir_lower_non_uniform_ubo_access
= (1 << 0),
3969 nir_lower_non_uniform_ssbo_access
= (1 << 1),
3970 nir_lower_non_uniform_texture_access
= (1 << 2),
3971 nir_lower_non_uniform_image_access
= (1 << 3),
3974 bool nir_lower_non_uniform_access(nir_shader
*shader
,
3975 enum nir_lower_non_uniform_access_type
);
3977 enum nir_lower_idiv_path
{
3978 /* This path is based on NV50LegalizeSSA::handleDIV(). It is the faster of
3979 * the two but it is not exact in some cases (for example, 1091317713u /
3980 * 1034u gives 5209173 instead of 1055432) */
3981 nir_lower_idiv_fast
,
3982 /* This path is based on AMDGPUTargetLowering::LowerUDIVREM() and
3983 * AMDGPUTargetLowering::LowerSDIVREM(). It requires more instructions than
3984 * the nv50 path and many of them are integer multiplications, so it is
3985 * probably slower. It should always return the correct result, though. */
3986 nir_lower_idiv_precise
,
3989 bool nir_lower_idiv(nir_shader
*shader
, enum nir_lower_idiv_path path
);
3991 bool nir_lower_input_attachments(nir_shader
*shader
, bool use_fragcoord_sysval
);
3993 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
,
3995 bool use_clipdist_array
,
3996 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
3997 bool nir_lower_clip_gs(nir_shader
*shader
, unsigned ucp_enables
,
3998 bool use_clipdist_array
,
3999 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4000 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
,
4001 bool use_clipdist_array
);
4002 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
4004 void nir_lower_point_size_mov(nir_shader
*shader
,
4005 const gl_state_index16
*pointsize_state_tokens
);
4007 bool nir_lower_frexp(nir_shader
*nir
);
4009 void nir_lower_two_sided_color(nir_shader
*shader
);
4011 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
4013 bool nir_lower_flatshade(nir_shader
*shader
);
4015 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
4016 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
4017 const gl_state_index16
*uniform_state_tokens
);
4019 typedef struct nir_lower_wpos_ytransform_options
{
4020 gl_state_index16 state_tokens
[STATE_LENGTH
];
4021 bool fs_coord_origin_upper_left
:1;
4022 bool fs_coord_origin_lower_left
:1;
4023 bool fs_coord_pixel_center_integer
:1;
4024 bool fs_coord_pixel_center_half_integer
:1;
4025 } nir_lower_wpos_ytransform_options
;
4027 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
4028 const nir_lower_wpos_ytransform_options
*options
);
4029 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
4031 bool nir_lower_fb_read(nir_shader
*shader
);
4033 typedef struct nir_lower_drawpixels_options
{
4034 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
4035 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
4036 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
4037 unsigned drawpix_sampler
;
4038 unsigned pixelmap_sampler
;
4040 bool scale_and_bias
:1;
4041 } nir_lower_drawpixels_options
;
4043 void nir_lower_drawpixels(nir_shader
*shader
,
4044 const nir_lower_drawpixels_options
*options
);
4046 typedef struct nir_lower_bitmap_options
{
4049 } nir_lower_bitmap_options
;
4051 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
4053 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
, unsigned ssbo_offset
);
4056 nir_lower_int_source_mods
= 1 << 0,
4057 nir_lower_float_source_mods
= 1 << 1,
4058 nir_lower_triop_abs
= 1 << 2,
4059 nir_lower_all_source_mods
= (1 << 3) - 1
4060 } nir_lower_to_source_mods_flags
;
4063 bool nir_lower_to_source_mods(nir_shader
*shader
, nir_lower_to_source_mods_flags options
);
4065 bool nir_lower_gs_intrinsics(nir_shader
*shader
);
4067 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
4069 bool nir_lower_bit_size(nir_shader
*shader
,
4070 nir_lower_bit_size_callback callback
,
4071 void *callback_data
);
4073 nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode
);
4074 bool nir_lower_int64(nir_shader
*shader
, nir_lower_int64_options options
);
4076 nir_lower_doubles_options
nir_lower_doubles_op_to_options_mask(nir_op opcode
);
4077 bool nir_lower_doubles(nir_shader
*shader
, const nir_shader
*softfp64
,
4078 nir_lower_doubles_options options
);
4079 bool nir_lower_pack(nir_shader
*shader
);
4081 bool nir_lower_point_size(nir_shader
*shader
, float min
, float max
);
4084 nir_lower_interpolation_at_sample
= (1 << 1),
4085 nir_lower_interpolation_at_offset
= (1 << 2),
4086 nir_lower_interpolation_centroid
= (1 << 3),
4087 nir_lower_interpolation_pixel
= (1 << 4),
4088 nir_lower_interpolation_sample
= (1 << 5),
4089 } nir_lower_interpolation_options
;
4091 bool nir_lower_interpolation(nir_shader
*shader
,
4092 nir_lower_interpolation_options options
);
4094 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
4096 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
4098 void nir_loop_analyze_impl(nir_function_impl
*impl
,
4099 nir_variable_mode indirect_mask
);
4101 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
4103 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
4104 bool nir_repair_ssa(nir_shader
*shader
);
4106 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
4107 bool nir_convert_to_lcssa(nir_shader
*shader
, bool skip_invariants
, bool skip_bool_invariants
);
4108 bool* nir_divergence_analysis(nir_shader
*shader
, nir_divergence_options options
);
4110 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
4111 * registers. If false, convert all values (even those not involved in a phi
4112 * node) to registers.
4114 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
4116 bool nir_lower_phis_to_regs_block(nir_block
*block
);
4117 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
4118 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
);
4120 bool nir_lower_samplers(nir_shader
*shader
);
4122 /* This is here for unit tests. */
4123 bool nir_opt_comparison_pre_impl(nir_function_impl
*impl
);
4125 bool nir_opt_comparison_pre(nir_shader
*shader
);
4127 bool nir_opt_access(nir_shader
*shader
);
4128 bool nir_opt_algebraic(nir_shader
*shader
);
4129 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
4130 bool nir_opt_algebraic_late(nir_shader
*shader
);
4131 bool nir_opt_constant_folding(nir_shader
*shader
);
4133 bool nir_opt_combine_stores(nir_shader
*shader
, nir_variable_mode modes
);
4135 bool nir_copy_prop(nir_shader
*shader
);
4137 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
4139 bool nir_opt_cse(nir_shader
*shader
);
4141 bool nir_opt_dce(nir_shader
*shader
);
4143 bool nir_opt_dead_cf(nir_shader
*shader
);
4145 bool nir_opt_dead_write_vars(nir_shader
*shader
);
4147 bool nir_opt_deref_impl(nir_function_impl
*impl
);
4148 bool nir_opt_deref(nir_shader
*shader
);
4150 bool nir_opt_find_array_copies(nir_shader
*shader
);
4152 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
4154 bool nir_opt_idiv_const(nir_shader
*shader
, unsigned min_bit_size
);
4156 bool nir_opt_if(nir_shader
*shader
, bool aggressive_last_continue
);
4158 bool nir_opt_intrinsics(nir_shader
*shader
);
4160 bool nir_opt_large_constants(nir_shader
*shader
,
4161 glsl_type_size_align_func size_align
,
4162 unsigned threshold
);
4164 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
4167 nir_move_const_undef
= (1 << 0),
4168 nir_move_load_ubo
= (1 << 1),
4169 nir_move_load_input
= (1 << 2),
4170 nir_move_comparisons
= (1 << 3),
4173 bool nir_can_move_instr(nir_instr
*instr
, nir_move_options options
);
4175 bool nir_opt_sink(nir_shader
*shader
, nir_move_options options
);
4177 bool nir_opt_move(nir_shader
*shader
, nir_move_options options
);
4179 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
,
4180 bool indirect_load_ok
, bool expensive_alu_ok
);
4182 bool nir_opt_rematerialize_compares(nir_shader
*shader
);
4184 bool nir_opt_remove_phis(nir_shader
*shader
);
4185 bool nir_opt_remove_phis_block(nir_block
*block
);
4187 bool nir_opt_shrink_load(nir_shader
*shader
);
4189 bool nir_opt_trivial_continues(nir_shader
*shader
);
4191 bool nir_opt_undef(nir_shader
*shader
);
4193 bool nir_opt_vectorize(nir_shader
*shader
);
4195 bool nir_opt_conditional_discard(nir_shader
*shader
);
4197 typedef bool (*nir_should_vectorize_mem_func
)(unsigned align
, unsigned bit_size
,
4198 unsigned num_components
, unsigned high_offset
,
4199 nir_intrinsic_instr
*low
, nir_intrinsic_instr
*high
);
4201 bool nir_opt_load_store_vectorize(nir_shader
*shader
, nir_variable_mode modes
,
4202 nir_should_vectorize_mem_func callback
);
4204 void nir_sweep(nir_shader
*shader
);
4206 void nir_remap_dual_slot_attributes(nir_shader
*shader
,
4207 uint64_t *dual_slot_inputs
);
4208 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
);
4210 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
4211 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);
4214 nir_variable_is_in_ubo(const nir_variable
*var
)
4216 return (var
->data
.mode
== nir_var_mem_ubo
&&
4217 var
->interface_type
!= NULL
);
4221 nir_variable_is_in_ssbo(const nir_variable
*var
)
4223 return (var
->data
.mode
== nir_var_mem_ssbo
&&
4224 var
->interface_type
!= NULL
);
4228 nir_variable_is_in_block(const nir_variable
*var
)
4230 return nir_variable_is_in_ubo(var
) || nir_variable_is_in_ssbo(var
);