2 * Copyright © 2014 Connor Abbott
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Connor Abbott (cwabbott0@gmail.com)
31 #include "util/hash_table.h"
32 #include "compiler/glsl/list.h"
33 #include "GL/gl.h" /* GLenum */
34 #include "util/list.h"
35 #include "util/ralloc.h"
37 #include "util/bitscan.h"
38 #include "util/bitset.h"
39 #include "util/macros.h"
40 #include "compiler/nir_types.h"
41 #include "compiler/shader_enums.h"
42 #include "compiler/shader_info.h"
46 #include "util/debug.h"
49 #include "nir_opcodes.h"
51 #if defined(_WIN32) && !defined(snprintf)
52 #define snprintf _snprintf
60 #define NIR_TRUE (~0u)
61 #define NIR_MAX_VEC_COMPONENTS 4
62 #define NIR_MAX_MATRIX_COLUMNS 4
63 typedef uint8_t nir_component_mask_t
;
65 /** Defines a cast function
67 * This macro defines a cast function from in_type to out_type where
68 * out_type is some structure type that contains a field of type out_type.
70 * Note that you have to be a bit careful as the generated cast function
73 #define NIR_DEFINE_CAST(name, in_type, out_type, field, \
74 type_field, type_value) \
75 static inline out_type * \
76 name(const in_type *parent) \
78 assert(parent && parent->type_field == type_value); \
79 return exec_node_data(out_type, parent, field); \
89 * Description of built-in state associated with a uniform
91 * \sa nir_variable::state_slots
94 gl_state_index16 tokens
[STATE_LENGTH
];
99 nir_var_shader_in
= (1 << 0),
100 nir_var_shader_out
= (1 << 1),
101 nir_var_shader_temp
= (1 << 2),
102 nir_var_function_temp
= (1 << 3),
103 nir_var_uniform
= (1 << 4),
104 nir_var_mem_ubo
= (1 << 5),
105 nir_var_system_value
= (1 << 6),
106 nir_var_mem_ssbo
= (1 << 7),
107 nir_var_mem_shared
= (1 << 8),
108 nir_var_mem_global
= (1 << 9),
116 nir_rounding_mode_undef
= 0,
117 nir_rounding_mode_rtne
= 1, /* round to nearest even */
118 nir_rounding_mode_ru
= 2, /* round up */
119 nir_rounding_mode_rd
= 3, /* round down */
120 nir_rounding_mode_rtz
= 4, /* round towards zero */
137 #define nir_const_value_to_array(arr, c, components, m) \
139 for (unsigned i = 0; i < components; ++i) \
143 static inline nir_const_value
144 nir_const_value_for_raw_uint(uint64_t x
, unsigned bit_size
)
147 memset(&v
, 0, sizeof(v
));
150 case 1: v
.b
= x
; break;
151 case 8: v
.u8
= x
; break;
152 case 16: v
.u16
= x
; break;
153 case 32: v
.u32
= x
; break;
154 case 64: v
.u64
= x
; break;
156 unreachable("Invalid bit size");
162 static inline nir_const_value
163 nir_const_value_for_int(int64_t i
, unsigned bit_size
)
166 memset(&v
, 0, sizeof(v
));
168 assert(bit_size
<= 64);
170 assert(i
>= (-(1ll << (bit_size
- 1))));
171 assert(i
< (1ll << (bit_size
- 1)));
174 return nir_const_value_for_raw_uint(i
, bit_size
);
177 static inline nir_const_value
178 nir_const_value_for_uint(uint64_t u
, unsigned bit_size
)
181 memset(&v
, 0, sizeof(v
));
183 assert(bit_size
<= 64);
185 assert(u
< (1ull << bit_size
));
187 return nir_const_value_for_raw_uint(u
, bit_size
);
190 static inline nir_const_value
191 nir_const_value_for_bool(bool b
, unsigned bit_size
)
193 /* Booleans use a 0/-1 convention */
194 return nir_const_value_for_int(-(int)b
, bit_size
);
197 /* This one isn't inline because it requires half-float conversion */
198 nir_const_value
nir_const_value_for_float(double b
, unsigned bit_size
);
200 static inline int64_t
201 nir_const_value_as_int(nir_const_value value
, unsigned bit_size
)
204 /* int1_t uses 0/-1 convention */
205 case 1: return -(int)value
.b
;
206 case 8: return value
.i8
;
207 case 16: return value
.i16
;
208 case 32: return value
.i32
;
209 case 64: return value
.i64
;
211 unreachable("Invalid bit size");
215 static inline int64_t
216 nir_const_value_as_uint(nir_const_value value
, unsigned bit_size
)
219 case 1: return value
.b
;
220 case 8: return value
.u8
;
221 case 16: return value
.u16
;
222 case 32: return value
.u32
;
223 case 64: return value
.u64
;
225 unreachable("Invalid bit size");
230 nir_const_value_as_bool(nir_const_value value
, unsigned bit_size
)
232 int64_t i
= nir_const_value_as_int(value
, bit_size
);
234 /* Booleans of any size use 0/-1 convention */
235 assert(i
== 0 || i
== -1);
240 /* This one isn't inline because it requires half-float conversion */
241 double nir_const_value_as_float(nir_const_value value
, unsigned bit_size
);
243 typedef struct nir_constant
{
245 * Value of the constant.
247 * The field used to back the values supplied by the constant is determined
248 * by the type associated with the \c nir_variable. Constants may be
249 * scalars, vectors, or matrices.
251 nir_const_value values
[NIR_MAX_VEC_COMPONENTS
];
253 /* we could get this from the var->type but makes clone *much* easier to
254 * not have to care about the type.
256 unsigned num_elements
;
258 /* Array elements / Structure Fields */
259 struct nir_constant
**elements
;
263 * \brief Layout qualifiers for gl_FragDepth.
265 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
266 * with a layout qualifier.
269 nir_depth_layout_none
, /**< No depth layout is specified. */
270 nir_depth_layout_any
,
271 nir_depth_layout_greater
,
272 nir_depth_layout_less
,
273 nir_depth_layout_unchanged
277 * Enum keeping track of how a variable was declared.
281 * Normal declaration.
283 nir_var_declared_normally
= 0,
286 * Variable is implicitly generated by the compiler and should not be
287 * visible via the API.
290 } nir_var_declaration_type
;
293 * Either a uniform, global variable, shader input, or shader output. Based on
294 * ir_variable - it should be easy to translate between the two.
297 typedef struct nir_variable
{
298 struct exec_node node
;
301 * Declared type of the variable
303 const struct glsl_type
*type
;
306 * Declared name of the variable
310 struct nir_variable_data
{
312 * Storage class of the variable.
314 * \sa nir_variable_mode
316 nir_variable_mode mode
;
319 * Is the variable read-only?
321 * This is set for variables declared as \c const, shader inputs,
324 unsigned read_only
:1;
328 unsigned invariant
:1;
331 * Can this variable be coalesced with another?
333 * This is set by nir_lower_io_to_temporaries to say that any
334 * copies involving this variable should stay put. Propagating it can
335 * duplicate the resulting load/store, which is not wanted, and may
336 * result in a load/store of the variable with an indirect offset which
337 * the backend may not be able to handle.
339 unsigned cannot_coalesce
:1;
342 * When separate shader programs are enabled, only input/outputs between
343 * the stages of a multi-stage separate program can be safely removed
344 * from the shader interface. Other input/outputs must remains active.
346 * This is also used to make sure xfb varyings that are unused by the
347 * fragment shader are not removed.
349 unsigned always_active_io
:1;
352 * Interpolation mode for shader inputs / outputs
354 * \sa glsl_interp_mode
356 unsigned interpolation
:2;
359 * If non-zero, then this variable may be packed along with other variables
360 * into a single varying slot, so this offset should be applied when
361 * accessing components. For example, an offset of 1 means that the x
362 * component of this variable is actually stored in component y of the
363 * location specified by \c location.
365 unsigned location_frac
:2;
368 * If true, this variable represents an array of scalars that should
369 * be tightly packed. In other words, consecutive array elements
370 * should be stored one component apart, rather than one slot apart.
375 * Whether this is a fragment shader output implicitly initialized with
376 * the previous contents of the specified render target at the
377 * framebuffer location corresponding to this shader invocation.
379 unsigned fb_fetch_output
:1;
382 * Non-zero if this variable is considered bindless as defined by
383 * ARB_bindless_texture.
388 * Was an explicit binding set in the shader?
390 unsigned explicit_binding
:1;
393 * Was a transfer feedback buffer set in the shader?
395 unsigned explicit_xfb_buffer
:1;
398 * Was a transfer feedback stride set in the shader?
400 unsigned explicit_xfb_stride
:1;
403 * Was an explicit offset set in the shader?
405 unsigned explicit_offset
:1;
408 * \brief Layout qualifier for gl_FragDepth.
410 * This is not equal to \c ir_depth_layout_none if and only if this
411 * variable is \c gl_FragDepth and a layout qualifier is specified.
413 nir_depth_layout depth_layout
;
416 * Storage location of the base of this variable
418 * The precise meaning of this field depends on the nature of the variable.
420 * - Vertex shader input: one of the values from \c gl_vert_attrib.
421 * - Vertex shader output: one of the values from \c gl_varying_slot.
422 * - Geometry shader input: one of the values from \c gl_varying_slot.
423 * - Geometry shader output: one of the values from \c gl_varying_slot.
424 * - Fragment shader input: one of the values from \c gl_varying_slot.
425 * - Fragment shader output: one of the values from \c gl_frag_result.
426 * - Uniforms: Per-stage uniform slot number for default uniform block.
427 * - Uniforms: Index within the uniform block definition for UBO members.
428 * - Non-UBO Uniforms: uniform slot number.
429 * - Other: This field is not currently used.
431 * If the variable is a uniform, shader input, or shader output, and the
432 * slot has not been assigned, the value will be -1.
437 * The actual location of the variable in the IR. Only valid for inputs
440 unsigned int driver_location
;
443 * Vertex stream output identifier.
445 * For packed outputs, bit 31 is set and bits [2*i+1,2*i] indicate the
446 * stream of the i-th component.
451 * output index for dual source blending.
456 * Descriptor set binding for sampler or UBO.
461 * Initial binding point for a sampler or UBO.
463 * For array types, this represents the binding point for the first element.
468 * Location an atomic counter or transform feedback is stored at.
473 * Transform feedback buffer.
478 * Transform feedback stride.
483 * How the variable was declared. See nir_var_declaration_type.
485 * This is used to detect variables generated by the compiler, so should
486 * not be visible via the API.
488 unsigned how_declared
:2;
491 * ARB_shader_image_load_store qualifiers.
494 enum gl_access_qualifier access
;
496 /** Image internal format if specified explicitly, otherwise GL_NONE. */
502 * Built-in state that backs this uniform
504 * Once set at variable creation, \c state_slots must remain invariant.
505 * This is because, ideally, this array would be shared by all clones of
506 * this variable in the IR tree. In other words, we'd really like for it
507 * to be a fly-weight.
509 * If the variable is not a uniform, \c num_state_slots will be zero and
510 * \c state_slots will be \c NULL.
513 unsigned num_state_slots
; /**< Number of state slots used */
514 nir_state_slot
*state_slots
; /**< State descriptors. */
518 * Constant expression assigned in the initializer of the variable
520 * This field should only be used temporarily by creators of NIR shaders
521 * and then lower_constant_initializers can be used to get rid of them.
522 * Most of the rest of NIR ignores this field or asserts that it's NULL.
524 nir_constant
*constant_initializer
;
527 * For variables that are in an interface block or are an instance of an
528 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
530 * \sa ir_variable::location
532 const struct glsl_type
*interface_type
;
535 * Description of per-member data for per-member struct variables
537 * This is used for variables which are actually an amalgamation of
538 * multiple entities such as a struct of built-in values or a struct of
539 * inputs each with their own layout specifier. This is only allowed on
540 * variables with a struct or array of array of struct type.
542 unsigned num_members
;
543 struct nir_variable_data
*members
;
546 #define nir_foreach_variable(var, var_list) \
547 foreach_list_typed(nir_variable, var, node, var_list)
549 #define nir_foreach_variable_safe(var, var_list) \
550 foreach_list_typed_safe(nir_variable, var, node, var_list)
553 nir_variable_is_global(const nir_variable
*var
)
555 return var
->data
.mode
!= nir_var_function_temp
;
558 typedef struct nir_register
{
559 struct exec_node node
;
561 unsigned num_components
; /** < number of vector components */
562 unsigned num_array_elems
; /** < size of array (0 for no array) */
564 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
567 /** generic register index. */
570 /** only for debug purposes, can be NULL */
573 /** set of nir_srcs where this register is used (read from) */
574 struct list_head uses
;
576 /** set of nir_dests where this register is defined (written to) */
577 struct list_head defs
;
579 /** set of nir_ifs where this register is used as a condition */
580 struct list_head if_uses
;
583 #define nir_foreach_register(reg, reg_list) \
584 foreach_list_typed(nir_register, reg, node, reg_list)
585 #define nir_foreach_register_safe(reg, reg_list) \
586 foreach_list_typed_safe(nir_register, reg, node, reg_list)
588 typedef enum PACKED
{
590 nir_instr_type_deref
,
593 nir_instr_type_intrinsic
,
594 nir_instr_type_load_const
,
596 nir_instr_type_ssa_undef
,
598 nir_instr_type_parallel_copy
,
601 typedef struct nir_instr
{
602 struct exec_node node
;
603 struct nir_block
*block
;
606 /* A temporary for optimization and analysis passes to use for storing
607 * flags. For instance, DCE uses this to store the "dead/live" info.
611 /** generic instruction index. */
615 static inline nir_instr
*
616 nir_instr_next(nir_instr
*instr
)
618 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
619 if (exec_node_is_tail_sentinel(next
))
622 return exec_node_data(nir_instr
, next
, node
);
625 static inline nir_instr
*
626 nir_instr_prev(nir_instr
*instr
)
628 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
629 if (exec_node_is_head_sentinel(prev
))
632 return exec_node_data(nir_instr
, prev
, node
);
636 nir_instr_is_first(const nir_instr
*instr
)
638 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr
->node
));
642 nir_instr_is_last(const nir_instr
*instr
)
644 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr
->node
));
647 typedef struct nir_ssa_def
{
648 /** for debugging only, can be NULL */
651 /** generic SSA definition index. */
654 /** Index into the live_in and live_out bitfields */
657 /** Instruction which produces this SSA value. */
658 nir_instr
*parent_instr
;
660 /** set of nir_instrs where this register is used (read from) */
661 struct list_head uses
;
663 /** set of nir_ifs where this register is used as a condition */
664 struct list_head if_uses
;
666 uint8_t num_components
;
668 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
676 struct nir_src
*indirect
; /** < NULL for no indirect offset */
677 unsigned base_offset
;
679 /* TODO use-def chain goes here */
683 nir_instr
*parent_instr
;
684 struct list_head def_link
;
687 struct nir_src
*indirect
; /** < NULL for no indirect offset */
688 unsigned base_offset
;
690 /* TODO def-use chain goes here */
695 typedef struct nir_src
{
697 /** Instruction that consumes this value as a source. */
698 nir_instr
*parent_instr
;
699 struct nir_if
*parent_if
;
702 struct list_head use_link
;
712 static inline nir_src
715 nir_src src
= { { NULL
} };
719 #define NIR_SRC_INIT nir_src_init()
721 #define nir_foreach_use(src, reg_or_ssa_def) \
722 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
724 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
725 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
727 #define nir_foreach_if_use(src, reg_or_ssa_def) \
728 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
730 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
731 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
742 static inline nir_dest
745 nir_dest dest
= { { { NULL
} } };
749 #define NIR_DEST_INIT nir_dest_init()
751 #define nir_foreach_def(dest, reg) \
752 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
754 #define nir_foreach_def_safe(dest, reg) \
755 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
757 static inline nir_src
758 nir_src_for_ssa(nir_ssa_def
*def
)
760 nir_src src
= NIR_SRC_INIT
;
768 static inline nir_src
769 nir_src_for_reg(nir_register
*reg
)
771 nir_src src
= NIR_SRC_INIT
;
775 src
.reg
.indirect
= NULL
;
776 src
.reg
.base_offset
= 0;
781 static inline nir_dest
782 nir_dest_for_reg(nir_register
*reg
)
784 nir_dest dest
= NIR_DEST_INIT
;
791 static inline unsigned
792 nir_src_bit_size(nir_src src
)
794 return src
.is_ssa
? src
.ssa
->bit_size
: src
.reg
.reg
->bit_size
;
797 static inline unsigned
798 nir_src_num_components(nir_src src
)
800 return src
.is_ssa
? src
.ssa
->num_components
: src
.reg
.reg
->num_components
;
804 nir_src_is_const(nir_src src
)
807 src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
;
810 static inline unsigned
811 nir_dest_bit_size(nir_dest dest
)
813 return dest
.is_ssa
? dest
.ssa
.bit_size
: dest
.reg
.reg
->bit_size
;
816 static inline unsigned
817 nir_dest_num_components(nir_dest dest
)
819 return dest
.is_ssa
? dest
.ssa
.num_components
: dest
.reg
.reg
->num_components
;
822 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
823 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
829 * \name input modifiers
833 * For inputs interpreted as floating point, flips the sign bit. For
834 * inputs interpreted as integers, performs the two's complement negation.
839 * Clears the sign bit for floating point values, and computes the integer
840 * absolute value for integers. Note that the negate modifier acts after
841 * the absolute value modifier, therefore if both are set then all inputs
842 * will become negative.
848 * For each input component, says which component of the register it is
849 * chosen from. Note that which elements of the swizzle are used and which
850 * are ignored are based on the write mask for most opcodes - for example,
851 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
852 * a swizzle of {2, x, 1, 0} where x means "don't care."
854 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
861 * \name saturate output modifier
863 * Only valid for opcodes that output floating-point numbers. Clamps the
864 * output to between 0.0 and 1.0 inclusive.
869 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
872 /** NIR sized and unsized types
874 * The values in this enum are carefully chosen so that the sized type is
875 * just the unsized type OR the number of bits.
878 nir_type_invalid
= 0, /* Not a valid type */
882 nir_type_float
= 128,
883 nir_type_bool1
= 1 | nir_type_bool
,
884 nir_type_bool32
= 32 | nir_type_bool
,
885 nir_type_int1
= 1 | nir_type_int
,
886 nir_type_int8
= 8 | nir_type_int
,
887 nir_type_int16
= 16 | nir_type_int
,
888 nir_type_int32
= 32 | nir_type_int
,
889 nir_type_int64
= 64 | nir_type_int
,
890 nir_type_uint1
= 1 | nir_type_uint
,
891 nir_type_uint8
= 8 | nir_type_uint
,
892 nir_type_uint16
= 16 | nir_type_uint
,
893 nir_type_uint32
= 32 | nir_type_uint
,
894 nir_type_uint64
= 64 | nir_type_uint
,
895 nir_type_float16
= 16 | nir_type_float
,
896 nir_type_float32
= 32 | nir_type_float
,
897 nir_type_float64
= 64 | nir_type_float
,
900 #define NIR_ALU_TYPE_SIZE_MASK 0x79
901 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
903 static inline unsigned
904 nir_alu_type_get_type_size(nir_alu_type type
)
906 return type
& NIR_ALU_TYPE_SIZE_MASK
;
909 static inline unsigned
910 nir_alu_type_get_base_type(nir_alu_type type
)
912 return type
& NIR_ALU_TYPE_BASE_TYPE_MASK
;
915 static inline nir_alu_type
916 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
920 return nir_type_bool1
;
923 return nir_type_uint32
;
926 return nir_type_int32
;
928 case GLSL_TYPE_UINT16
:
929 return nir_type_uint16
;
931 case GLSL_TYPE_INT16
:
932 return nir_type_int16
;
934 case GLSL_TYPE_UINT8
:
935 return nir_type_uint8
;
937 return nir_type_int8
;
938 case GLSL_TYPE_UINT64
:
939 return nir_type_uint64
;
941 case GLSL_TYPE_INT64
:
942 return nir_type_int64
;
944 case GLSL_TYPE_FLOAT
:
945 return nir_type_float32
;
947 case GLSL_TYPE_FLOAT16
:
948 return nir_type_float16
;
950 case GLSL_TYPE_DOUBLE
:
951 return nir_type_float64
;
954 case GLSL_TYPE_SAMPLER
:
955 case GLSL_TYPE_IMAGE
:
956 case GLSL_TYPE_ATOMIC_UINT
:
957 case GLSL_TYPE_STRUCT
:
958 case GLSL_TYPE_INTERFACE
:
959 case GLSL_TYPE_ARRAY
:
961 case GLSL_TYPE_SUBROUTINE
:
962 case GLSL_TYPE_FUNCTION
:
963 case GLSL_TYPE_ERROR
:
964 return nir_type_invalid
;
967 unreachable("unknown type");
970 static inline nir_alu_type
971 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
973 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
976 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
977 nir_rounding_mode rnd
);
980 nir_op_vec(unsigned components
)
982 switch (components
) {
983 case 1: return nir_op_mov
;
984 case 2: return nir_op_vec2
;
985 case 3: return nir_op_vec3
;
986 case 4: return nir_op_vec4
;
987 default: unreachable("bad component count");
992 nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode
, unsigned bit_size
)
994 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
) ||
995 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
) ||
996 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
);
1000 nir_is_denorm_flush_to_zero(unsigned execution_mode
, unsigned bit_size
)
1002 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
) ||
1003 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
) ||
1004 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
);
1008 nir_is_denorm_preserve(unsigned execution_mode
, unsigned bit_size
)
1010 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP16
) ||
1011 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP32
) ||
1012 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP64
);
1016 nir_is_rounding_mode_rtne(unsigned execution_mode
, unsigned bit_size
)
1018 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1019 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1020 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1024 nir_is_rounding_mode_rtz(unsigned execution_mode
, unsigned bit_size
)
1026 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1027 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1028 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1032 nir_has_any_rounding_mode_rtz(unsigned execution_mode
)
1034 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1035 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1036 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1040 nir_has_any_rounding_mode_rtne(unsigned execution_mode
)
1042 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1043 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1044 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1047 static inline nir_rounding_mode
1048 nir_get_rounding_mode_from_float_controls(unsigned execution_mode
,
1051 if (nir_alu_type_get_base_type(type
) != nir_type_float
)
1052 return nir_rounding_mode_undef
;
1054 unsigned bit_size
= nir_alu_type_get_type_size(type
);
1056 if (nir_is_rounding_mode_rtz(execution_mode
, bit_size
))
1057 return nir_rounding_mode_rtz
;
1058 if (nir_is_rounding_mode_rtne(execution_mode
, bit_size
))
1059 return nir_rounding_mode_rtne
;
1060 return nir_rounding_mode_undef
;
1064 nir_has_any_rounding_mode_enabled(unsigned execution_mode
)
1067 nir_has_any_rounding_mode_rtne(execution_mode
) ||
1068 nir_has_any_rounding_mode_rtz(execution_mode
);
1074 * Operation where the first two sources are commutative.
1076 * For 2-source operations, this just mathematical commutativity. Some
1077 * 3-source operations, like ffma, are only commutative in the first two
1080 NIR_OP_IS_2SRC_COMMUTATIVE
= (1 << 0),
1081 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
1082 } nir_op_algebraic_property
;
1087 unsigned num_inputs
;
1090 * The number of components in the output
1092 * If non-zero, this is the size of the output and input sizes are
1093 * explicitly given; swizzle and writemask are still in effect, but if
1094 * the output component is masked out, then the input component may
1097 * If zero, the opcode acts in the standard, per-component manner; the
1098 * operation is performed on each component (except the ones that are
1099 * masked out) with the input being taken from the input swizzle for
1102 * The size of some of the inputs may be given (i.e. non-zero) even
1103 * though output_size is zero; in that case, the inputs with a zero
1104 * size act per-component, while the inputs with non-zero size don't.
1106 unsigned output_size
;
1109 * The type of vector that the instruction outputs. Note that the
1110 * staurate modifier is only allowed on outputs with the float type.
1113 nir_alu_type output_type
;
1116 * The number of components in each input
1118 unsigned input_sizes
[NIR_MAX_VEC_COMPONENTS
];
1121 * The type of vector that each input takes. Note that negate and
1122 * absolute value are only allowed on inputs with int or float type and
1123 * behave differently on the two.
1125 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
1127 nir_op_algebraic_property algebraic_properties
;
1129 /* Whether this represents a numeric conversion opcode */
1133 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
1135 typedef struct nir_alu_instr
{
1139 /** Indicates that this ALU instruction generates an exact value
1141 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1142 * really equivalent to either. This indicates that the value generated by
1143 * this operation is high-precision and any code transformations that touch
1144 * it must ensure that the resulting value is bit-for-bit identical to the
1150 * Indicates that this instruction do not cause wrapping to occur, in the
1151 * form of overflow or underflow.
1153 bool no_signed_wrap
:1;
1154 bool no_unsigned_wrap
:1;
1160 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
1161 nir_alu_instr
*instr
);
1162 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
1163 nir_alu_instr
*instr
);
1165 /* is this source channel used? */
1167 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
1170 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1171 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
1173 return (instr
->dest
.write_mask
>> channel
) & 1;
1176 static inline nir_component_mask_t
1177 nir_alu_instr_src_read_mask(const nir_alu_instr
*instr
, unsigned src
)
1179 nir_component_mask_t read_mask
= 0;
1180 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; c
++) {
1181 if (!nir_alu_instr_channel_used(instr
, src
, c
))
1184 read_mask
|= (1 << instr
->src
[src
].swizzle
[c
]);
1190 * Get the number of channels used for a source
1192 static inline unsigned
1193 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
1195 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1196 return nir_op_infos
[instr
->op
].input_sizes
[src
];
1198 return nir_dest_num_components(instr
->dest
.dest
);
1202 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1204 switch (instr
->op
) {
1224 bool nir_const_value_negative_equal(nir_const_value c1
, nir_const_value c2
,
1225 nir_alu_type full_type
);
1227 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
1228 unsigned src1
, unsigned src2
);
1230 bool nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
1231 const nir_alu_instr
*alu2
,
1232 unsigned src1
, unsigned src2
);
1236 nir_deref_type_array
,
1237 nir_deref_type_array_wildcard
,
1238 nir_deref_type_ptr_as_array
,
1239 nir_deref_type_struct
,
1240 nir_deref_type_cast
,
1246 /** The type of this deref instruction */
1247 nir_deref_type deref_type
;
1249 /** The mode of the underlying variable */
1250 nir_variable_mode mode
;
1252 /** The dereferenced type of the resulting pointer value */
1253 const struct glsl_type
*type
;
1256 /** Variable being dereferenced if deref_type is a deref_var */
1259 /** Parent deref if deref_type is not deref_var */
1263 /** Additional deref parameters */
1274 unsigned ptr_stride
;
1278 /** Destination to store the resulting "pointer" */
1282 static inline nir_deref_instr
*nir_src_as_deref(nir_src src
);
1284 static inline nir_deref_instr
*
1285 nir_deref_instr_parent(const nir_deref_instr
*instr
)
1287 if (instr
->deref_type
== nir_deref_type_var
)
1290 return nir_src_as_deref(instr
->parent
);
1293 static inline nir_variable
*
1294 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1296 while (instr
->deref_type
!= nir_deref_type_var
) {
1297 if (instr
->deref_type
== nir_deref_type_cast
)
1300 instr
= nir_deref_instr_parent(instr
);
1306 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1307 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr
*instr
);
1308 bool nir_deref_instr_has_complex_use(nir_deref_instr
*instr
);
1310 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1312 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*instr
);
1317 struct nir_function
*callee
;
1319 unsigned num_params
;
1323 #include "nir_intrinsics.h"
1325 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1327 /** Represents an intrinsic
1329 * An intrinsic is an instruction type for handling things that are
1330 * more-or-less regular operations but don't just consume and produce SSA
1331 * values like ALU operations do. Intrinsics are not for things that have
1332 * special semantic meaning such as phi nodes and parallel copies.
1333 * Examples of intrinsics include variable load/store operations, system
1334 * value loads, and the like. Even though texturing more-or-less falls
1335 * under this category, texturing is its own instruction type because
1336 * trying to represent texturing with intrinsics would lead to a
1337 * combinatorial explosion of intrinsic opcodes.
1339 * By having a single instruction type for handling a lot of different
1340 * cases, optimization passes can look for intrinsics and, for the most
1341 * part, completely ignore them. Each intrinsic type also has a few
1342 * possible flags that govern whether or not they can be reordered or
1343 * eliminated. That way passes like dead code elimination can still work
1344 * on intrisics without understanding the meaning of each.
1346 * Each intrinsic has some number of constant indices, some number of
1347 * variables, and some number of sources. What these sources, variables,
1348 * and indices mean depends on the intrinsic and is documented with the
1349 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1350 * instructions are the only types of instruction that can operate on
1356 nir_intrinsic_op intrinsic
;
1360 /** number of components if this is a vectorized intrinsic
1362 * Similarly to ALU operations, some intrinsics are vectorized.
1363 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1364 * For vectorized intrinsics, the num_components field specifies the
1365 * number of destination components and the number of source components
1366 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1368 uint8_t num_components
;
1370 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1373 } nir_intrinsic_instr
;
1375 static inline nir_variable
*
1376 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1378 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1382 * \name NIR intrinsics semantic flags
1384 * information about what the compiler can do with the intrinsics.
1386 * \sa nir_intrinsic_info::flags
1390 * whether the intrinsic can be safely eliminated if none of its output
1391 * value is not being used.
1393 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1396 * Whether the intrinsic can be reordered with respect to any other
1397 * intrinsic, i.e. whether the only reordering dependencies of the
1398 * intrinsic are due to the register reads/writes.
1400 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1401 } nir_intrinsic_semantic_flag
;
1404 * \name NIR intrinsics const-index flag
1406 * Indicates the usage of a const_index slot.
1408 * \sa nir_intrinsic_info::index_map
1412 * Generally instructions that take a offset src argument, can encode
1413 * a constant 'base' value which is added to the offset.
1415 NIR_INTRINSIC_BASE
= 1,
1418 * For store instructions, a writemask for the store.
1420 NIR_INTRINSIC_WRMASK
,
1423 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1425 NIR_INTRINSIC_STREAM_ID
,
1428 * The clip-plane id for load_user_clip_plane intrinsic.
1430 NIR_INTRINSIC_UCP_ID
,
1433 * The amount of data, starting from BASE, that this instruction may
1434 * access. This is used to provide bounds if the offset is not constant.
1436 NIR_INTRINSIC_RANGE
,
1439 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1441 NIR_INTRINSIC_DESC_SET
,
1444 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1446 NIR_INTRINSIC_BINDING
,
1451 NIR_INTRINSIC_COMPONENT
,
1454 * Interpolation mode (only meaningful for FS inputs).
1456 NIR_INTRINSIC_INTERP_MODE
,
1459 * A binary nir_op to use when performing a reduction or scan operation
1461 NIR_INTRINSIC_REDUCTION_OP
,
1464 * Cluster size for reduction operations
1466 NIR_INTRINSIC_CLUSTER_SIZE
,
1469 * Parameter index for a load_param intrinsic
1471 NIR_INTRINSIC_PARAM_IDX
,
1474 * Image dimensionality for image intrinsics
1476 * One of GLSL_SAMPLER_DIM_*
1478 NIR_INTRINSIC_IMAGE_DIM
,
1481 * Non-zero if we are accessing an array image
1483 NIR_INTRINSIC_IMAGE_ARRAY
,
1486 * Image format for image intrinsics
1488 NIR_INTRINSIC_FORMAT
,
1491 * Access qualifiers for image and memory access intrinsics
1493 NIR_INTRINSIC_ACCESS
,
1496 * Alignment for offsets and addresses
1498 * These two parameters, specify an alignment in terms of a multiplier and
1499 * an offset. The offset or address parameter X of the intrinsic is
1500 * guaranteed to satisfy the following:
1502 * (X - align_offset) % align_mul == 0
1504 NIR_INTRINSIC_ALIGN_MUL
,
1505 NIR_INTRINSIC_ALIGN_OFFSET
,
1508 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1510 NIR_INTRINSIC_DESC_TYPE
,
1513 * The nir_alu_type of a uniform/input/output
1518 * The swizzle mask for the instructions
1519 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1521 NIR_INTRINSIC_SWIZZLE_MASK
,
1523 /* Separate source/dest access flags for copies */
1524 NIR_INTRINSIC_SRC_ACCESS
,
1525 NIR_INTRINSIC_DST_ACCESS
,
1527 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1529 } nir_intrinsic_index_flag
;
1531 #define NIR_INTRINSIC_MAX_INPUTS 5
1536 unsigned num_srcs
; /** < number of register/SSA inputs */
1538 /** number of components of each input register
1540 * If this value is 0, the number of components is given by the
1541 * num_components field of nir_intrinsic_instr. If this value is -1, the
1542 * intrinsic consumes however many components are provided and it is not
1545 int src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1549 /** number of components of the output register
1551 * If this value is 0, the number of components is given by the
1552 * num_components field of nir_intrinsic_instr.
1554 unsigned dest_components
;
1556 /** bitfield of legal bit sizes */
1557 unsigned dest_bit_sizes
;
1559 /** the number of constant indices used by the intrinsic */
1560 unsigned num_indices
;
1562 /** indicates the usage of intr->const_index[n] */
1563 unsigned index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1565 /** semantic flags for calls to this intrinsic */
1566 nir_intrinsic_semantic_flag flags
;
1567 } nir_intrinsic_info
;
1569 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1571 static inline unsigned
1572 nir_intrinsic_src_components(nir_intrinsic_instr
*intr
, unsigned srcn
)
1574 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1575 assert(srcn
< info
->num_srcs
);
1576 if (info
->src_components
[srcn
] > 0)
1577 return info
->src_components
[srcn
];
1578 else if (info
->src_components
[srcn
] == 0)
1579 return intr
->num_components
;
1581 return nir_src_num_components(intr
->src
[srcn
]);
1584 static inline unsigned
1585 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1587 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1588 if (!info
->has_dest
)
1590 else if (info
->dest_components
)
1591 return info
->dest_components
;
1593 return intr
->num_components
;
1596 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1597 static inline type \
1598 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1600 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1601 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1602 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1604 static inline void \
1605 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1607 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1608 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1609 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1612 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1613 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1614 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1615 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1616 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1617 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1618 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1619 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1620 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1621 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1622 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1623 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1624 INTRINSIC_IDX_ACCESSORS(image_dim
, IMAGE_DIM
, enum glsl_sampler_dim
)
1625 INTRINSIC_IDX_ACCESSORS(image_array
, IMAGE_ARRAY
, bool)
1626 INTRINSIC_IDX_ACCESSORS(access
, ACCESS
, enum gl_access_qualifier
)
1627 INTRINSIC_IDX_ACCESSORS(src_access
, SRC_ACCESS
, enum gl_access_qualifier
)
1628 INTRINSIC_IDX_ACCESSORS(dst_access
, DST_ACCESS
, enum gl_access_qualifier
)
1629 INTRINSIC_IDX_ACCESSORS(format
, FORMAT
, unsigned)
1630 INTRINSIC_IDX_ACCESSORS(align_mul
, ALIGN_MUL
, unsigned)
1631 INTRINSIC_IDX_ACCESSORS(align_offset
, ALIGN_OFFSET
, unsigned)
1632 INTRINSIC_IDX_ACCESSORS(desc_type
, DESC_TYPE
, unsigned)
1633 INTRINSIC_IDX_ACCESSORS(type
, TYPE
, nir_alu_type
)
1634 INTRINSIC_IDX_ACCESSORS(swizzle_mask
, SWIZZLE_MASK
, unsigned)
1637 nir_intrinsic_set_align(nir_intrinsic_instr
*intrin
,
1638 unsigned align_mul
, unsigned align_offset
)
1640 assert(util_is_power_of_two_nonzero(align_mul
));
1641 assert(align_offset
< align_mul
);
1642 nir_intrinsic_set_align_mul(intrin
, align_mul
);
1643 nir_intrinsic_set_align_offset(intrin
, align_offset
);
1646 /** Returns a simple alignment for a load/store intrinsic offset
1648 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1649 * and ALIGN_OFFSET parameters, this helper takes both into account and
1650 * provides a single simple alignment parameter. The offset X is guaranteed
1651 * to satisfy X % align == 0.
1653 static inline unsigned
1654 nir_intrinsic_align(const nir_intrinsic_instr
*intrin
)
1656 const unsigned align_mul
= nir_intrinsic_align_mul(intrin
);
1657 const unsigned align_offset
= nir_intrinsic_align_offset(intrin
);
1658 assert(align_offset
< align_mul
);
1659 return align_offset
? 1 << (ffs(align_offset
) - 1) : align_mul
;
1662 /* Converts a image_deref_* intrinsic into a image_* one */
1663 void nir_rewrite_image_intrinsic(nir_intrinsic_instr
*instr
,
1664 nir_ssa_def
*handle
, bool bindless
);
1666 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1668 nir_intrinsic_can_reorder(nir_intrinsic_instr
*instr
)
1670 if (instr
->intrinsic
== nir_intrinsic_load_deref
||
1671 instr
->intrinsic
== nir_intrinsic_load_ssbo
||
1672 instr
->intrinsic
== nir_intrinsic_bindless_image_load
||
1673 instr
->intrinsic
== nir_intrinsic_image_deref_load
||
1674 instr
->intrinsic
== nir_intrinsic_image_load
) {
1675 return nir_intrinsic_access(instr
) & ACCESS_CAN_REORDER
;
1677 const nir_intrinsic_info
*info
=
1678 &nir_intrinsic_infos
[instr
->intrinsic
];
1679 return (info
->flags
& NIR_INTRINSIC_CAN_ELIMINATE
) &&
1680 (info
->flags
& NIR_INTRINSIC_CAN_REORDER
);
1685 * \group texture information
1687 * This gives semantic information about textures which is useful to the
1688 * frontend, the backend, and lowering passes, but not the optimizer.
1693 nir_tex_src_projector
,
1694 nir_tex_src_comparator
, /* shadow comparator */
1698 nir_tex_src_min_lod
,
1699 nir_tex_src_ms_index
, /* MSAA sample index */
1700 nir_tex_src_ms_mcs
, /* MSAA compression value */
1703 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
1704 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
1705 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
1706 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
1707 nir_tex_src_texture_handle
, /* < bindless texture handle */
1708 nir_tex_src_sampler_handle
, /* < bindless sampler handle */
1709 nir_tex_src_plane
, /* < selects plane for planar textures */
1710 nir_num_tex_src_types
1715 nir_tex_src_type src_type
;
1719 nir_texop_tex
, /**< Regular texture look-up */
1720 nir_texop_txb
, /**< Texture look-up with LOD bias */
1721 nir_texop_txl
, /**< Texture look-up with explicit LOD */
1722 nir_texop_txd
, /**< Texture look-up with partial derivatives */
1723 nir_texop_txf
, /**< Texel fetch with explicit LOD */
1724 nir_texop_txf_ms
, /**< Multisample texture fetch */
1725 nir_texop_txf_ms_fb
, /**< Multisample texture fetch from framebuffer */
1726 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
1727 nir_texop_txs
, /**< Texture size */
1728 nir_texop_lod
, /**< Texture lod query */
1729 nir_texop_tg4
, /**< Texture gather */
1730 nir_texop_query_levels
, /**< Texture levels query */
1731 nir_texop_texture_samples
, /**< Texture samples query */
1732 nir_texop_samples_identical
, /**< Query whether all samples are definitely
1740 enum glsl_sampler_dim sampler_dim
;
1741 nir_alu_type dest_type
;
1746 unsigned num_srcs
, coord_components
;
1747 bool is_array
, is_shadow
;
1750 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1751 * components or the new-style shadow that outputs 1 component.
1753 bool is_new_style_shadow
;
1755 /* gather component selector */
1756 unsigned component
: 2;
1758 /* gather offsets */
1759 int8_t tg4_offsets
[4][2];
1761 /* True if the texture index or handle is not dynamically uniform */
1762 bool texture_non_uniform
;
1764 /* True if the sampler index or handle is not dynamically uniform */
1765 bool sampler_non_uniform
;
1767 /** The texture index
1769 * If this texture instruction has a nir_tex_src_texture_offset source,
1770 * then the texture index is given by texture_index + texture_offset.
1772 unsigned texture_index
;
1774 /** The size of the texture array or 0 if it's not an array */
1775 unsigned texture_array_size
;
1777 /** The sampler index
1779 * The following operations do not require a sampler and, as such, this
1780 * field should be ignored:
1782 * - nir_texop_txf_ms
1785 * - nir_texop_query_levels
1786 * - nir_texop_texture_samples
1787 * - nir_texop_samples_identical
1789 * If this texture instruction has a nir_tex_src_sampler_offset source,
1790 * then the sampler index is given by sampler_index + sampler_offset.
1792 unsigned sampler_index
;
1795 static inline unsigned
1796 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
1798 switch (instr
->op
) {
1799 case nir_texop_txs
: {
1801 switch (instr
->sampler_dim
) {
1802 case GLSL_SAMPLER_DIM_1D
:
1803 case GLSL_SAMPLER_DIM_BUF
:
1806 case GLSL_SAMPLER_DIM_2D
:
1807 case GLSL_SAMPLER_DIM_CUBE
:
1808 case GLSL_SAMPLER_DIM_MS
:
1809 case GLSL_SAMPLER_DIM_RECT
:
1810 case GLSL_SAMPLER_DIM_EXTERNAL
:
1811 case GLSL_SAMPLER_DIM_SUBPASS
:
1814 case GLSL_SAMPLER_DIM_3D
:
1818 unreachable("not reached");
1820 if (instr
->is_array
)
1828 case nir_texop_texture_samples
:
1829 case nir_texop_query_levels
:
1830 case nir_texop_samples_identical
:
1834 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
1841 /* Returns true if this texture operation queries something about the texture
1842 * rather than actually sampling it.
1845 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
1847 switch (instr
->op
) {
1850 case nir_texop_texture_samples
:
1851 case nir_texop_query_levels
:
1852 case nir_texop_txf_ms_mcs
:
1859 case nir_texop_txf_ms
:
1860 case nir_texop_txf_ms_fb
:
1864 unreachable("Invalid texture opcode");
1869 nir_tex_instr_has_implicit_derivative(const nir_tex_instr
*instr
)
1871 switch (instr
->op
) {
1881 static inline nir_alu_type
1882 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
1884 switch (instr
->src
[src
].src_type
) {
1885 case nir_tex_src_coord
:
1886 switch (instr
->op
) {
1888 case nir_texop_txf_ms
:
1889 case nir_texop_txf_ms_fb
:
1890 case nir_texop_txf_ms_mcs
:
1891 case nir_texop_samples_identical
:
1892 return nir_type_int
;
1895 return nir_type_float
;
1898 case nir_tex_src_lod
:
1899 switch (instr
->op
) {
1902 return nir_type_int
;
1905 return nir_type_float
;
1908 case nir_tex_src_projector
:
1909 case nir_tex_src_comparator
:
1910 case nir_tex_src_bias
:
1911 case nir_tex_src_min_lod
:
1912 case nir_tex_src_ddx
:
1913 case nir_tex_src_ddy
:
1914 return nir_type_float
;
1916 case nir_tex_src_offset
:
1917 case nir_tex_src_ms_index
:
1918 case nir_tex_src_plane
:
1919 return nir_type_int
;
1921 case nir_tex_src_ms_mcs
:
1922 case nir_tex_src_texture_deref
:
1923 case nir_tex_src_sampler_deref
:
1924 case nir_tex_src_texture_offset
:
1925 case nir_tex_src_sampler_offset
:
1926 case nir_tex_src_texture_handle
:
1927 case nir_tex_src_sampler_handle
:
1928 return nir_type_uint
;
1930 case nir_num_tex_src_types
:
1931 unreachable("nir_num_tex_src_types is not a valid source type");
1934 unreachable("Invalid texture source type");
1937 static inline unsigned
1938 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
1940 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
1941 return instr
->coord_components
;
1943 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
1944 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
1947 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
1948 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
1949 if (instr
->is_array
)
1950 return instr
->coord_components
- 1;
1952 return instr
->coord_components
;
1955 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
1956 * the offset, since a cube maps to a single face.
1958 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
1959 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
1961 else if (instr
->is_array
)
1962 return instr
->coord_components
- 1;
1964 return instr
->coord_components
;
1971 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
1973 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
1974 if (instr
->src
[i
].src_type
== type
)
1980 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
1981 nir_tex_src_type src_type
,
1984 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
1986 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
);
1993 nir_const_value value
[];
1994 } nir_load_const_instr
;
2007 /* creates a new SSA variable in an undefined state */
2012 } nir_ssa_undef_instr
;
2015 struct exec_node node
;
2017 /* The predecessor block corresponding to this source */
2018 struct nir_block
*pred
;
2023 #define nir_foreach_phi_src(phi_src, phi) \
2024 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
2025 #define nir_foreach_phi_src_safe(phi_src, phi) \
2026 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
2031 struct exec_list srcs
; /** < list of nir_phi_src */
2037 struct exec_node node
;
2040 } nir_parallel_copy_entry
;
2042 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
2043 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
2048 /* A list of nir_parallel_copy_entrys. The sources of all of the
2049 * entries are copied to the corresponding destinations "in parallel".
2050 * In other words, if we have two entries: a -> b and b -> a, the values
2053 struct exec_list entries
;
2054 } nir_parallel_copy_instr
;
2056 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
2057 type
, nir_instr_type_alu
)
2058 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
2059 type
, nir_instr_type_deref
)
2060 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
2061 type
, nir_instr_type_call
)
2062 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
2063 type
, nir_instr_type_jump
)
2064 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
2065 type
, nir_instr_type_tex
)
2066 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
2067 type
, nir_instr_type_intrinsic
)
2068 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
2069 type
, nir_instr_type_load_const
)
2070 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
2071 type
, nir_instr_type_ssa_undef
)
2072 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
2073 type
, nir_instr_type_phi
)
2074 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
2075 nir_parallel_copy_instr
, instr
,
2076 type
, nir_instr_type_parallel_copy
)
2079 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
2080 static inline type \
2081 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2083 assert(nir_src_is_const(src)); \
2084 nir_load_const_instr *load = \
2085 nir_instr_as_load_const(src.ssa->parent_instr); \
2086 assert(comp < load->def.num_components); \
2087 return nir_const_value_as_##suffix(load->value[comp], \
2088 load->def.bit_size); \
2091 static inline type \
2092 nir_src_as_##suffix(nir_src src) \
2094 assert(nir_src_num_components(src) == 1); \
2095 return nir_src_comp_as_##suffix(src, 0); \
2098 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2099 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint
)
2100 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2101 NIR_DEFINE_SRC_AS_CONST(double, float)
2103 #undef NIR_DEFINE_SRC_AS_CONST
2112 nir_ssa_scalar_is_const(nir_ssa_scalar s
)
2114 return s
.def
->parent_instr
->type
== nir_instr_type_load_const
;
2117 static inline nir_const_value
2118 nir_ssa_scalar_as_const_value(nir_ssa_scalar s
)
2120 assert(s
.comp
< s
.def
->num_components
);
2121 nir_load_const_instr
*load
= nir_instr_as_load_const(s
.def
->parent_instr
);
2122 return load
->value
[s
.comp
];
2125 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2126 static inline type \
2127 nir_ssa_scalar_as_##suffix(nir_ssa_scalar s) \
2129 return nir_const_value_as_##suffix( \
2130 nir_ssa_scalar_as_const_value(s), s.def->bit_size); \
2133 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2134 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint
)
2135 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2136 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2138 #undef NIR_DEFINE_SCALAR_AS_CONST
2141 nir_ssa_scalar_is_alu(nir_ssa_scalar s
)
2143 return s
.def
->parent_instr
->type
== nir_instr_type_alu
;
2146 static inline nir_op
2147 nir_ssa_scalar_alu_op(nir_ssa_scalar s
)
2149 return nir_instr_as_alu(s
.def
->parent_instr
)->op
;
2152 static inline nir_ssa_scalar
2153 nir_ssa_scalar_chase_alu_src(nir_ssa_scalar s
, unsigned alu_src_idx
)
2155 nir_ssa_scalar out
= { NULL
, 0 };
2157 nir_alu_instr
*alu
= nir_instr_as_alu(s
.def
->parent_instr
);
2158 assert(alu_src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
2160 /* Our component must be written */
2161 assert(s
.comp
< s
.def
->num_components
);
2162 assert(alu
->dest
.write_mask
& (1u << s
.comp
));
2164 assert(alu
->src
[alu_src_idx
].src
.is_ssa
);
2165 out
.def
= alu
->src
[alu_src_idx
].src
.ssa
;
2167 if (nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 0) {
2168 /* The ALU src is unsized so the source component follows the
2169 * destination component.
2171 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[s
.comp
];
2173 /* This is a sized source so all source components work together to
2174 * produce all the destination components. Since we need to return a
2175 * scalar, this only works if the source is a scalar.
2177 assert(nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 1);
2178 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[0];
2180 assert(out
.comp
< out
.def
->num_components
);
2189 * Control flow consists of a tree of control flow nodes, which include
2190 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2191 * instructions that always run start-to-finish. Each basic block also keeps
2192 * track of its successors (blocks which may run immediately after the current
2193 * block) and predecessors (blocks which could have run immediately before the
2194 * current block). Each function also has a start block and an end block which
2195 * all return statements point to (which is always empty). Together, all the
2196 * blocks with their predecessors and successors make up the control flow
2197 * graph (CFG) of the function. There are helpers that modify the tree of
2198 * control flow nodes while modifying the CFG appropriately; these should be
2199 * used instead of modifying the tree directly.
2206 nir_cf_node_function
2209 typedef struct nir_cf_node
{
2210 struct exec_node node
;
2211 nir_cf_node_type type
;
2212 struct nir_cf_node
*parent
;
2215 typedef struct nir_block
{
2216 nir_cf_node cf_node
;
2218 struct exec_list instr_list
; /** < list of nir_instr */
2220 /** generic block index; generated by nir_index_blocks */
2224 * Each block can only have up to 2 successors, so we put them in a simple
2225 * array - no need for anything more complicated.
2227 struct nir_block
*successors
[2];
2229 /* Set of nir_block predecessors in the CFG */
2230 struct set
*predecessors
;
2233 * this node's immediate dominator in the dominance tree - set to NULL for
2236 struct nir_block
*imm_dom
;
2238 /* This node's children in the dominance tree */
2239 unsigned num_dom_children
;
2240 struct nir_block
**dom_children
;
2242 /* Set of nir_blocks on the dominance frontier of this block */
2243 struct set
*dom_frontier
;
2246 * These two indices have the property that dom_{pre,post}_index for each
2247 * child of this block in the dominance tree will always be between
2248 * dom_pre_index and dom_post_index for this block, which makes testing if
2249 * a given block is dominated by another block an O(1) operation.
2251 unsigned dom_pre_index
, dom_post_index
;
2253 /* live in and out for this block; used for liveness analysis */
2254 BITSET_WORD
*live_in
;
2255 BITSET_WORD
*live_out
;
2258 static inline nir_instr
*
2259 nir_block_first_instr(nir_block
*block
)
2261 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
2262 return exec_node_data(nir_instr
, head
, node
);
2265 static inline nir_instr
*
2266 nir_block_last_instr(nir_block
*block
)
2268 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
2269 return exec_node_data(nir_instr
, tail
, node
);
2273 nir_block_ends_in_jump(nir_block
*block
)
2275 return !exec_list_is_empty(&block
->instr_list
) &&
2276 nir_block_last_instr(block
)->type
== nir_instr_type_jump
;
2279 #define nir_foreach_instr(instr, block) \
2280 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
2281 #define nir_foreach_instr_reverse(instr, block) \
2282 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
2283 #define nir_foreach_instr_safe(instr, block) \
2284 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
2285 #define nir_foreach_instr_reverse_safe(instr, block) \
2286 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
2289 nir_selection_control_none
= 0x0,
2290 nir_selection_control_flatten
= 0x1,
2291 nir_selection_control_dont_flatten
= 0x2,
2292 } nir_selection_control
;
2294 typedef struct nir_if
{
2295 nir_cf_node cf_node
;
2297 nir_selection_control control
;
2299 struct exec_list then_list
; /** < list of nir_cf_node */
2300 struct exec_list else_list
; /** < list of nir_cf_node */
2306 /** Instruction that generates nif::condition. */
2307 nir_instr
*conditional_instr
;
2309 /** Block within ::nif that has the break instruction. */
2310 nir_block
*break_block
;
2312 /** Last block for the then- or else-path that does not contain the break. */
2313 nir_block
*continue_from_block
;
2315 /** True when ::break_block is in the else-path of ::nif. */
2316 bool continue_from_then
;
2319 /* This is true if the terminators exact trip count is unknown. For
2322 * for (int i = 0; i < imin(x, 4); i++)
2325 * Here loop analysis would have set a max_trip_count of 4 however we dont
2326 * know for sure that this is the exact trip count.
2328 bool exact_trip_count_unknown
;
2330 struct list_head loop_terminator_link
;
2331 } nir_loop_terminator
;
2334 /* Estimated cost (in number of instructions) of the loop */
2335 unsigned instr_cost
;
2337 /* Guessed trip count based on array indexing */
2338 unsigned guessed_trip_count
;
2340 /* Maximum number of times the loop is run (if known) */
2341 unsigned max_trip_count
;
2343 /* Do we know the exact number of times the loop will be run */
2344 bool exact_trip_count_known
;
2346 /* Unroll the loop regardless of its size */
2349 /* Does the loop contain complex loop terminators, continues or other
2350 * complex behaviours? If this is true we can't rely on
2351 * loop_terminator_list to be complete or accurate.
2355 nir_loop_terminator
*limiting_terminator
;
2357 /* A list of loop_terminators terminating this loop. */
2358 struct list_head loop_terminator_list
;
2362 nir_loop_control_none
= 0x0,
2363 nir_loop_control_unroll
= 0x1,
2364 nir_loop_control_dont_unroll
= 0x2,
2368 nir_cf_node cf_node
;
2370 struct exec_list body
; /** < list of nir_cf_node */
2372 nir_loop_info
*info
;
2373 nir_loop_control control
;
2374 bool partially_unrolled
;
2378 * Various bits of metadata that can may be created or required by
2379 * optimization and analysis passes
2382 nir_metadata_none
= 0x0,
2383 nir_metadata_block_index
= 0x1,
2384 nir_metadata_dominance
= 0x2,
2385 nir_metadata_live_ssa_defs
= 0x4,
2386 nir_metadata_not_properly_reset
= 0x8,
2387 nir_metadata_loop_analysis
= 0x10,
2391 nir_cf_node cf_node
;
2393 /** pointer to the function of which this is an implementation */
2394 struct nir_function
*function
;
2396 struct exec_list body
; /** < list of nir_cf_node */
2398 nir_block
*end_block
;
2400 /** list for all local variables in the function */
2401 struct exec_list locals
;
2403 /** list of local registers in the function */
2404 struct exec_list registers
;
2406 /** next available local register index */
2409 /** next available SSA value index */
2412 /* total number of basic blocks, only valid when block_index_dirty = false */
2413 unsigned num_blocks
;
2415 nir_metadata valid_metadata
;
2416 } nir_function_impl
;
2418 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2419 nir_start_block(nir_function_impl
*impl
)
2421 return (nir_block
*) impl
->body
.head_sentinel
.next
;
2424 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2425 nir_impl_last_block(nir_function_impl
*impl
)
2427 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
2430 static inline nir_cf_node
*
2431 nir_cf_node_next(nir_cf_node
*node
)
2433 struct exec_node
*next
= exec_node_get_next(&node
->node
);
2434 if (exec_node_is_tail_sentinel(next
))
2437 return exec_node_data(nir_cf_node
, next
, node
);
2440 static inline nir_cf_node
*
2441 nir_cf_node_prev(nir_cf_node
*node
)
2443 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
2444 if (exec_node_is_head_sentinel(prev
))
2447 return exec_node_data(nir_cf_node
, prev
, node
);
2451 nir_cf_node_is_first(const nir_cf_node
*node
)
2453 return exec_node_is_head_sentinel(node
->node
.prev
);
2457 nir_cf_node_is_last(const nir_cf_node
*node
)
2459 return exec_node_is_tail_sentinel(node
->node
.next
);
2462 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
2463 type
, nir_cf_node_block
)
2464 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
2465 type
, nir_cf_node_if
)
2466 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
2467 type
, nir_cf_node_loop
)
2468 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
2469 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
2471 static inline nir_block
*
2472 nir_if_first_then_block(nir_if
*if_stmt
)
2474 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
2475 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2478 static inline nir_block
*
2479 nir_if_last_then_block(nir_if
*if_stmt
)
2481 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
2482 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2485 static inline nir_block
*
2486 nir_if_first_else_block(nir_if
*if_stmt
)
2488 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
2489 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2492 static inline nir_block
*
2493 nir_if_last_else_block(nir_if
*if_stmt
)
2495 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
2496 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2499 static inline nir_block
*
2500 nir_loop_first_block(nir_loop
*loop
)
2502 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
2503 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2506 static inline nir_block
*
2507 nir_loop_last_block(nir_loop
*loop
)
2509 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
2510 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2514 * Return true if this list of cf_nodes contains a single empty block.
2517 nir_cf_list_is_empty_block(struct exec_list
*cf_list
)
2519 if (exec_list_is_singular(cf_list
)) {
2520 struct exec_node
*head
= exec_list_get_head(cf_list
);
2522 nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2523 return exec_list_is_empty(&block
->instr_list
);
2529 uint8_t num_components
;
2533 typedef struct nir_function
{
2534 struct exec_node node
;
2537 struct nir_shader
*shader
;
2539 unsigned num_params
;
2540 nir_parameter
*params
;
2542 /** The implementation of this function.
2544 * If the function is only declared and not implemented, this is NULL.
2546 nir_function_impl
*impl
;
2552 nir_lower_imul64
= (1 << 0),
2553 nir_lower_isign64
= (1 << 1),
2554 /** Lower all int64 modulus and division opcodes */
2555 nir_lower_divmod64
= (1 << 2),
2556 /** Lower all 64-bit umul_high and imul_high opcodes */
2557 nir_lower_imul_high64
= (1 << 3),
2558 nir_lower_mov64
= (1 << 4),
2559 nir_lower_icmp64
= (1 << 5),
2560 nir_lower_iadd64
= (1 << 6),
2561 nir_lower_iabs64
= (1 << 7),
2562 nir_lower_ineg64
= (1 << 8),
2563 nir_lower_logic64
= (1 << 9),
2564 nir_lower_minmax64
= (1 << 10),
2565 nir_lower_shift64
= (1 << 11),
2566 nir_lower_imul_2x32_64
= (1 << 12),
2567 nir_lower_extract64
= (1 << 13),
2568 } nir_lower_int64_options
;
2571 nir_lower_drcp
= (1 << 0),
2572 nir_lower_dsqrt
= (1 << 1),
2573 nir_lower_drsq
= (1 << 2),
2574 nir_lower_dtrunc
= (1 << 3),
2575 nir_lower_dfloor
= (1 << 4),
2576 nir_lower_dceil
= (1 << 5),
2577 nir_lower_dfract
= (1 << 6),
2578 nir_lower_dround_even
= (1 << 7),
2579 nir_lower_dmod
= (1 << 8),
2580 nir_lower_dsub
= (1 << 9),
2581 nir_lower_ddiv
= (1 << 10),
2582 nir_lower_fp64_full_software
= (1 << 11),
2583 } nir_lower_doubles_options
;
2586 nir_divergence_single_prim_per_subgroup
= (1 << 0),
2587 nir_divergence_single_patch_per_tcs_subgroup
= (1 << 1),
2588 nir_divergence_single_patch_per_tes_subgroup
= (1 << 2),
2589 nir_divergence_view_index_uniform
= (1 << 3),
2590 } nir_divergence_options
;
2592 typedef struct nir_shader_compiler_options
{
2598 /** Lowers flrp when it does not support doubles */
2605 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2606 bool lower_bitfield_extract
;
2607 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
2608 bool lower_bitfield_extract_to_shifts
;
2609 /** Lowers bitfield_insert to bfi/bfm */
2610 bool lower_bitfield_insert
;
2611 /** Lowers bitfield_insert to compares, and shifts. */
2612 bool lower_bitfield_insert_to_shifts
;
2613 /** Lowers bitfield_insert to bfm/bitfield_select. */
2614 bool lower_bitfield_insert_to_bitfield_select
;
2615 /** Lowers bitfield_reverse to shifts. */
2616 bool lower_bitfield_reverse
;
2617 /** Lowers bit_count to shifts. */
2618 bool lower_bit_count
;
2619 /** Lowers ifind_msb to compare and ufind_msb */
2620 bool lower_ifind_msb
;
2621 /** Lowers find_lsb to ufind_msb and logic ops */
2622 bool lower_find_lsb
;
2623 bool lower_uadd_carry
;
2624 bool lower_usub_borrow
;
2625 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2626 bool lower_mul_high
;
2627 /** lowers fneg and ineg to fsub and isub. */
2629 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2632 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2635 /* lower fall_equalN/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
2636 bool lower_vector_cmp
;
2638 /** enables rules to lower idiv by power-of-two: */
2641 /** enable rules to avoid bit ops */
2644 /** enables rules to lower isign to imin+imax */
2647 /** enables rules to lower fsign to fsub and flt */
2650 /* lower fdph to fdot4 */
2653 /** lower fdot to fmul and fsum/fadd. */
2656 /* Does the native fdot instruction replicate its result for four
2657 * components? If so, then opt_algebraic_late will turn all fdotN
2658 * instructions into fdot_replicatedN instructions.
2660 bool fdot_replicates
;
2662 /** lowers ffloor to fsub+ffract: */
2665 /** lowers ffract to fsub+ffloor: */
2668 /** lowers fceil to fneg+ffloor+fneg: */
2675 bool lower_pack_half_2x16
;
2676 bool lower_pack_unorm_2x16
;
2677 bool lower_pack_snorm_2x16
;
2678 bool lower_pack_unorm_4x8
;
2679 bool lower_pack_snorm_4x8
;
2680 bool lower_unpack_half_2x16
;
2681 bool lower_unpack_unorm_2x16
;
2682 bool lower_unpack_snorm_2x16
;
2683 bool lower_unpack_unorm_4x8
;
2684 bool lower_unpack_snorm_4x8
;
2686 bool lower_extract_byte
;
2687 bool lower_extract_word
;
2689 bool lower_all_io_to_temps
;
2690 bool lower_all_io_to_elements
;
2692 /* Indicates that the driver only has zero-based vertex id */
2693 bool vertex_id_zero_based
;
2696 * If enabled, gl_BaseVertex will be lowered as:
2697 * is_indexed_draw (~0/0) & firstvertex
2699 bool lower_base_vertex
;
2702 * If enabled, gl_HelperInvocation will be lowered as:
2704 * !((1 << sample_id) & sample_mask_in))
2706 * This depends on some possibly hw implementation details, which may
2707 * not be true for all hw. In particular that the FS is only executed
2708 * for covered samples or for helper invocations. So, do not blindly
2709 * enable this option.
2711 * Note: See also issue #22 in ARB_shader_image_load_store
2713 bool lower_helper_invocation
;
2716 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
2718 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
2719 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
2721 bool optimize_sample_mask_in
;
2723 bool lower_cs_local_index_from_id
;
2724 bool lower_cs_local_id_from_index
;
2726 bool lower_device_index_to_zero
;
2728 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
2729 bool lower_wpos_pntc
;
2735 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
2736 * for IO purposes and would prefer loads/stores be vectorized.
2739 bool lower_to_scalar
;
2742 * Should nir_lower_io() create load_interpolated_input intrinsics?
2744 * If not, it generates regular load_input intrinsics and interpolation
2745 * information must be inferred from the list of input nir_variables.
2747 bool use_interpolated_input_intrinsics
;
2749 /* Lowers when 32x32->64 bit multiplication is not supported */
2750 bool lower_mul_2x32_64
;
2752 /* Lowers when rotate instruction is not supported */
2756 * Is this the Intel vec4 backend?
2758 * Used to inhibit algebraic optimizations that are known to be harmful on
2759 * the Intel vec4 backend. This is generally applicable to any
2760 * optimization that might cause more immediate values to be used in
2761 * 3-source (e.g., ffma and flrp) instructions.
2765 unsigned max_unroll_iterations
;
2767 nir_lower_int64_options lower_int64_options
;
2768 nir_lower_doubles_options lower_doubles_options
;
2769 } nir_shader_compiler_options
;
2771 typedef struct nir_shader
{
2772 /** list of uniforms (nir_variable) */
2773 struct exec_list uniforms
;
2775 /** list of inputs (nir_variable) */
2776 struct exec_list inputs
;
2778 /** list of outputs (nir_variable) */
2779 struct exec_list outputs
;
2781 /** list of shared compute variables (nir_variable) */
2782 struct exec_list shared
;
2784 /** Set of driver-specific options for the shader.
2786 * The memory for the options is expected to be kept in a single static
2787 * copy by the driver.
2789 const struct nir_shader_compiler_options
*options
;
2791 /** Various bits of compile-time information about a given shader */
2792 struct shader_info info
;
2794 /** list of global variables in the shader (nir_variable) */
2795 struct exec_list globals
;
2797 /** list of system value variables in the shader (nir_variable) */
2798 struct exec_list system_values
;
2800 struct exec_list functions
; /** < list of nir_function */
2803 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
2806 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
2808 /** Size in bytes of required scratch space */
2809 unsigned scratch_size
;
2811 /** Constant data associated with this shader.
2813 * Constant data is loaded through load_constant intrinsics. See also
2814 * nir_opt_large_constants.
2816 void *constant_data
;
2817 unsigned constant_data_size
;
2820 #define nir_foreach_function(func, shader) \
2821 foreach_list_typed(nir_function, func, node, &(shader)->functions)
2823 static inline nir_function_impl
*
2824 nir_shader_get_entrypoint(nir_shader
*shader
)
2826 nir_function
*func
= NULL
;
2828 nir_foreach_function(function
, shader
) {
2829 assert(func
== NULL
);
2830 if (function
->is_entrypoint
) {
2841 assert(func
->num_params
== 0);
2846 nir_shader
*nir_shader_create(void *mem_ctx
,
2847 gl_shader_stage stage
,
2848 const nir_shader_compiler_options
*options
,
2851 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
2853 void nir_reg_remove(nir_register
*reg
);
2855 /** Adds a variable to the appropriate list in nir_shader */
2856 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
2859 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
2861 assert(var
->data
.mode
== nir_var_function_temp
);
2862 exec_list_push_tail(&impl
->locals
, &var
->node
);
2865 /** creates a variable, sets a few defaults, and adds it to the list */
2866 nir_variable
*nir_variable_create(nir_shader
*shader
,
2867 nir_variable_mode mode
,
2868 const struct glsl_type
*type
,
2870 /** creates a local variable and adds it to the list */
2871 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
2872 const struct glsl_type
*type
,
2875 /** creates a function and adds it to the shader's list of functions */
2876 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
2878 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
2879 /** creates a function_impl that isn't tied to any particular function */
2880 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
2882 nir_block
*nir_block_create(nir_shader
*shader
);
2883 nir_if
*nir_if_create(nir_shader
*shader
);
2884 nir_loop
*nir_loop_create(nir_shader
*shader
);
2886 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
2888 /** requests that the given pieces of metadata be generated */
2889 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
2890 /** dirties all but the preserved metadata */
2891 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
2893 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
2894 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
2896 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
2897 nir_deref_type deref_type
);
2899 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
2901 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
2902 unsigned num_components
,
2905 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
2906 nir_intrinsic_op op
);
2908 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
2909 nir_function
*callee
);
2911 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
2913 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
2915 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
2917 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
2918 unsigned num_components
,
2921 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
2924 * NIR Cursors and Instruction Insertion API
2927 * A tiny struct representing a point to insert/extract instructions or
2928 * control flow nodes. Helps reduce the combinatorial explosion of possible
2929 * points to insert/extract.
2931 * \sa nir_control_flow.h
2934 nir_cursor_before_block
,
2935 nir_cursor_after_block
,
2936 nir_cursor_before_instr
,
2937 nir_cursor_after_instr
,
2938 } nir_cursor_option
;
2941 nir_cursor_option option
;
2948 static inline nir_block
*
2949 nir_cursor_current_block(nir_cursor cursor
)
2951 if (cursor
.option
== nir_cursor_before_instr
||
2952 cursor
.option
== nir_cursor_after_instr
) {
2953 return cursor
.instr
->block
;
2955 return cursor
.block
;
2959 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
2961 static inline nir_cursor
2962 nir_before_block(nir_block
*block
)
2965 cursor
.option
= nir_cursor_before_block
;
2966 cursor
.block
= block
;
2970 static inline nir_cursor
2971 nir_after_block(nir_block
*block
)
2974 cursor
.option
= nir_cursor_after_block
;
2975 cursor
.block
= block
;
2979 static inline nir_cursor
2980 nir_before_instr(nir_instr
*instr
)
2983 cursor
.option
= nir_cursor_before_instr
;
2984 cursor
.instr
= instr
;
2988 static inline nir_cursor
2989 nir_after_instr(nir_instr
*instr
)
2992 cursor
.option
= nir_cursor_after_instr
;
2993 cursor
.instr
= instr
;
2997 static inline nir_cursor
2998 nir_after_block_before_jump(nir_block
*block
)
3000 nir_instr
*last_instr
= nir_block_last_instr(block
);
3001 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
3002 return nir_before_instr(last_instr
);
3004 return nir_after_block(block
);
3008 static inline nir_cursor
3009 nir_before_src(nir_src
*src
, bool is_if_condition
)
3011 if (is_if_condition
) {
3012 nir_block
*prev_block
=
3013 nir_cf_node_as_block(nir_cf_node_prev(&src
->parent_if
->cf_node
));
3014 assert(!nir_block_ends_in_jump(prev_block
));
3015 return nir_after_block(prev_block
);
3016 } else if (src
->parent_instr
->type
== nir_instr_type_phi
) {
3018 nir_phi_instr
*cond_phi
= nir_instr_as_phi(src
->parent_instr
);
3020 nir_foreach_phi_src(phi_src
, cond_phi
) {
3021 if (phi_src
->src
.ssa
== src
->ssa
) {
3028 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
3029 * to have a more specific name.
3031 nir_phi_src
*phi_src
= LIST_ENTRY(nir_phi_src
, src
, src
);
3032 return nir_after_block_before_jump(phi_src
->pred
);
3034 return nir_before_instr(src
->parent_instr
);
3038 static inline nir_cursor
3039 nir_before_cf_node(nir_cf_node
*node
)
3041 if (node
->type
== nir_cf_node_block
)
3042 return nir_before_block(nir_cf_node_as_block(node
));
3044 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
3047 static inline nir_cursor
3048 nir_after_cf_node(nir_cf_node
*node
)
3050 if (node
->type
== nir_cf_node_block
)
3051 return nir_after_block(nir_cf_node_as_block(node
));
3053 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
3056 static inline nir_cursor
3057 nir_after_phis(nir_block
*block
)
3059 nir_foreach_instr(instr
, block
) {
3060 if (instr
->type
!= nir_instr_type_phi
)
3061 return nir_before_instr(instr
);
3063 return nir_after_block(block
);
3066 static inline nir_cursor
3067 nir_after_cf_node_and_phis(nir_cf_node
*node
)
3069 if (node
->type
== nir_cf_node_block
)
3070 return nir_after_block(nir_cf_node_as_block(node
));
3072 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
3074 return nir_after_phis(block
);
3077 static inline nir_cursor
3078 nir_before_cf_list(struct exec_list
*cf_list
)
3080 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
3081 exec_list_get_head(cf_list
), node
);
3082 return nir_before_cf_node(first_node
);
3085 static inline nir_cursor
3086 nir_after_cf_list(struct exec_list
*cf_list
)
3088 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
3089 exec_list_get_tail(cf_list
), node
);
3090 return nir_after_cf_node(last_node
);
3094 * Insert a NIR instruction at the given cursor.
3096 * Note: This does not update the cursor.
3098 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
3101 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
3103 nir_instr_insert(nir_before_instr(instr
), before
);
3107 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
3109 nir_instr_insert(nir_after_instr(instr
), after
);
3113 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
3115 nir_instr_insert(nir_before_block(block
), before
);
3119 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
3121 nir_instr_insert(nir_after_block(block
), after
);
3125 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
3127 nir_instr_insert(nir_before_cf_node(node
), before
);
3131 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
3133 nir_instr_insert(nir_after_cf_node(node
), after
);
3137 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
3139 nir_instr_insert(nir_before_cf_list(list
), before
);
3143 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
3145 nir_instr_insert(nir_after_cf_list(list
), after
);
3148 void nir_instr_remove_v(nir_instr
*instr
);
3150 static inline nir_cursor
3151 nir_instr_remove(nir_instr
*instr
)
3154 nir_instr
*prev
= nir_instr_prev(instr
);
3156 cursor
= nir_after_instr(prev
);
3158 cursor
= nir_before_block(instr
->block
);
3160 nir_instr_remove_v(instr
);
3166 nir_ssa_def
*nir_instr_ssa_def(nir_instr
*instr
);
3168 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
3169 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
3170 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
3171 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
3173 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
3174 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
3176 nir_const_value
*nir_src_as_const_value(nir_src src
);
3178 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
3179 static inline c_type * \
3180 nir_src_as_ ## name (nir_src src) \
3182 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
3183 ? cast_macro(src.ssa->parent_instr) : NULL; \
3186 NIR_SRC_AS_(alu_instr
, nir_alu_instr
, nir_instr_type_alu
, nir_instr_as_alu
)
3187 NIR_SRC_AS_(intrinsic
, nir_intrinsic_instr
,
3188 nir_instr_type_intrinsic
, nir_instr_as_intrinsic
)
3189 NIR_SRC_AS_(deref
, nir_deref_instr
, nir_instr_type_deref
, nir_instr_as_deref
)
3191 bool nir_src_is_dynamically_uniform(nir_src src
);
3192 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
3193 bool nir_instrs_equal(const nir_instr
*instr1
, const nir_instr
*instr2
);
3194 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
3195 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
3196 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
3197 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
3200 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
3201 unsigned num_components
, unsigned bit_size
,
3203 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
3204 unsigned num_components
, unsigned bit_size
,
3207 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
3208 const struct glsl_type
*type
,
3211 assert(glsl_type_is_vector_or_scalar(type
));
3212 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
3213 glsl_get_bit_size(type
), name
);
3215 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
3216 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
3217 nir_instr
*after_me
);
3219 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
3222 * finds the next basic block in source-code order, returns NULL if there is
3226 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
3228 /* Performs the opposite of nir_block_cf_tree_next() */
3230 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
3232 /* Gets the first block in a CF node in source-code order */
3234 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
3236 /* Gets the last block in a CF node in source-code order */
3238 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
3240 /* Gets the next block after a CF node in source-code order */
3242 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
3244 /* Macros for loops that visit blocks in source-code order */
3246 #define nir_foreach_block(block, impl) \
3247 for (nir_block *block = nir_start_block(impl); block != NULL; \
3248 block = nir_block_cf_tree_next(block))
3250 #define nir_foreach_block_safe(block, impl) \
3251 for (nir_block *block = nir_start_block(impl), \
3252 *next = nir_block_cf_tree_next(block); \
3254 block = next, next = nir_block_cf_tree_next(block))
3256 #define nir_foreach_block_reverse(block, impl) \
3257 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
3258 block = nir_block_cf_tree_prev(block))
3260 #define nir_foreach_block_reverse_safe(block, impl) \
3261 for (nir_block *block = nir_impl_last_block(impl), \
3262 *prev = nir_block_cf_tree_prev(block); \
3264 block = prev, prev = nir_block_cf_tree_prev(block))
3266 #define nir_foreach_block_in_cf_node(block, node) \
3267 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
3268 block != nir_cf_node_cf_tree_next(node); \
3269 block = nir_block_cf_tree_next(block))
3271 /* If the following CF node is an if, this function returns that if.
3272 * Otherwise, it returns NULL.
3274 nir_if
*nir_block_get_following_if(nir_block
*block
);
3276 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
3278 void nir_index_local_regs(nir_function_impl
*impl
);
3279 void nir_index_ssa_defs(nir_function_impl
*impl
);
3280 unsigned nir_index_instrs(nir_function_impl
*impl
);
3282 void nir_index_blocks(nir_function_impl
*impl
);
3284 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
3285 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
3286 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
3287 void nir_print_deref(const nir_deref_instr
*deref
, FILE *fp
);
3289 /** Shallow clone of a single ALU instruction. */
3290 nir_alu_instr
*nir_alu_instr_clone(nir_shader
*s
, const nir_alu_instr
*orig
);
3292 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
3293 nir_function_impl
*nir_function_impl_clone(nir_shader
*shader
,
3294 const nir_function_impl
*fi
);
3295 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
3296 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
3298 void nir_shader_replace(nir_shader
*dest
, nir_shader
*src
);
3300 void nir_shader_serialize_deserialize(nir_shader
*s
);
3303 void nir_validate_shader(nir_shader
*shader
, const char *when
);
3304 void nir_metadata_set_validation_flag(nir_shader
*shader
);
3305 void nir_metadata_check_validation_flag(nir_shader
*shader
);
3308 should_skip_nir(const char *name
)
3310 static const char *list
= NULL
;
3312 /* Comma separated list of names to skip. */
3313 list
= getenv("NIR_SKIP");
3321 return comma_separated_list_contains(list
, name
);
3325 should_clone_nir(void)
3327 static int should_clone
= -1;
3328 if (should_clone
< 0)
3329 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
3331 return should_clone
;
3335 should_serialize_deserialize_nir(void)
3337 static int test_serialize
= -1;
3338 if (test_serialize
< 0)
3339 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
3341 return test_serialize
;
3345 should_print_nir(void)
3347 static int should_print
= -1;
3348 if (should_print
< 0)
3349 should_print
= env_var_as_boolean("NIR_PRINT", false);
3351 return should_print
;
3354 static inline void nir_validate_shader(nir_shader
*shader
, const char *when
) { (void) shader
; (void)when
; }
3355 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
3356 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
3357 static inline bool should_skip_nir(UNUSED
const char *pass_name
) { return false; }
3358 static inline bool should_clone_nir(void) { return false; }
3359 static inline bool should_serialize_deserialize_nir(void) { return false; }
3360 static inline bool should_print_nir(void) { return false; }
3363 #define _PASS(pass, nir, do_pass) do { \
3364 if (should_skip_nir(#pass)) { \
3365 printf("skipping %s\n", #pass); \
3369 nir_validate_shader(nir, "after " #pass); \
3370 if (should_clone_nir()) { \
3371 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3372 nir_shader_replace(nir, clone); \
3374 if (should_serialize_deserialize_nir()) { \
3375 nir_shader_serialize_deserialize(nir); \
3379 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3380 nir_metadata_set_validation_flag(nir); \
3381 if (should_print_nir()) \
3382 printf("%s\n", #pass); \
3383 if (pass(nir, ##__VA_ARGS__)) { \
3385 if (should_print_nir()) \
3386 nir_print_shader(nir, stdout); \
3387 nir_metadata_check_validation_flag(nir); \
3391 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3392 if (should_print_nir()) \
3393 printf("%s\n", #pass); \
3394 pass(nir, ##__VA_ARGS__); \
3395 if (should_print_nir()) \
3396 nir_print_shader(nir, stdout); \
3399 #define NIR_SKIP(name) should_skip_nir(#name)
3401 /** An instruction filtering callback
3403 * Returns true if the instruction should be processed and false otherwise.
3405 typedef bool (*nir_instr_filter_cb
)(const nir_instr
*, const void *);
3407 /** A simple instruction lowering callback
3409 * Many instruction lowering passes can be written as a simple function which
3410 * takes an instruction as its input and returns a sequence of instructions
3411 * that implement the consumed instruction. This function type represents
3412 * such a lowering function. When called, a function with this prototype
3413 * should either return NULL indicating that no lowering needs to be done or
3414 * emit a sequence of instructions using the provided builder (whose cursor
3415 * will already be placed after the instruction to be lowered) and return the
3416 * resulting nir_ssa_def.
3418 typedef nir_ssa_def
*(*nir_lower_instr_cb
)(struct nir_builder
*,
3419 nir_instr
*, void *);
3422 * Special return value for nir_lower_instr_cb when some progress occurred
3423 * (like changing an input to the instr) that didn't result in a replacement
3424 * SSA def being generated.
3426 #define NIR_LOWER_INSTR_PROGRESS ((nir_ssa_def *)(uintptr_t)1)
3428 /** Iterate over all the instructions in a nir_function_impl and lower them
3429 * using the provided callbacks
3431 * This function implements the guts of a standard lowering pass for you. It
3432 * iterates over all of the instructions in a nir_function_impl and calls the
3433 * filter callback on each one. If the filter callback returns true, it then
3434 * calls the lowering call back on the instruction. (Splitting it this way
3435 * allows us to avoid some save/restore work for instructions we know won't be
3436 * lowered.) If the instruction is dead after the lowering is complete, it
3437 * will be removed. If new instructions are added, the lowering callback will
3438 * also be called on them in case multiple lowerings are required.
3440 * The metadata for the nir_function_impl will also be updated. If any blocks
3441 * are added (they cannot be removed), dominance and block indices will be
3444 bool nir_function_impl_lower_instructions(nir_function_impl
*impl
,
3445 nir_instr_filter_cb filter
,
3446 nir_lower_instr_cb lower
,
3448 bool nir_shader_lower_instructions(nir_shader
*shader
,
3449 nir_instr_filter_cb filter
,
3450 nir_lower_instr_cb lower
,
3453 void nir_calc_dominance_impl(nir_function_impl
*impl
);
3454 void nir_calc_dominance(nir_shader
*shader
);
3456 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
3457 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
3458 bool nir_block_is_unreachable(nir_block
*block
);
3460 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
3461 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
3463 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
3464 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
3466 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
3467 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
3469 int nir_gs_count_vertices(const nir_shader
*shader
);
3471 bool nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3472 bool nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3473 bool nir_split_var_copies(nir_shader
*shader
);
3474 bool nir_split_per_member_structs(nir_shader
*shader
);
3475 bool nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
);
3477 bool nir_lower_returns_impl(nir_function_impl
*impl
);
3478 bool nir_lower_returns(nir_shader
*shader
);
3480 void nir_inline_function_impl(struct nir_builder
*b
,
3481 const nir_function_impl
*impl
,
3482 nir_ssa_def
**params
);
3483 bool nir_inline_functions(nir_shader
*shader
);
3485 bool nir_propagate_invariant(nir_shader
*shader
);
3487 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
3488 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
3489 nir_intrinsic_instr
*copy
);
3490 bool nir_lower_var_copies(nir_shader
*shader
);
3492 void nir_fixup_deref_modes(nir_shader
*shader
);
3494 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
3497 nir_lower_direct_array_deref_of_vec_load
= (1 << 0),
3498 nir_lower_indirect_array_deref_of_vec_load
= (1 << 1),
3499 nir_lower_direct_array_deref_of_vec_store
= (1 << 2),
3500 nir_lower_indirect_array_deref_of_vec_store
= (1 << 3),
3501 } nir_lower_array_deref_of_vec_options
;
3503 bool nir_lower_array_deref_of_vec(nir_shader
*shader
, nir_variable_mode modes
,
3504 nir_lower_array_deref_of_vec_options options
);
3506 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
3508 bool nir_lower_locals_to_regs(nir_shader
*shader
);
3510 void nir_lower_io_to_temporaries(nir_shader
*shader
,
3511 nir_function_impl
*entrypoint
,
3512 bool outputs
, bool inputs
);
3514 bool nir_lower_vars_to_scratch(nir_shader
*shader
,
3515 nir_variable_mode modes
,
3517 glsl_type_size_align_func size_align
);
3519 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
3521 void nir_gather_ssa_types(nir_function_impl
*impl
,
3522 BITSET_WORD
*float_types
,
3523 BITSET_WORD
*int_types
);
3525 void nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
3526 int (*type_size
)(const struct glsl_type
*, bool));
3528 /* Some helpers to do very simple linking */
3529 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3530 bool nir_remove_unused_io_vars(nir_shader
*shader
, struct exec_list
*var_list
,
3531 uint64_t *used_by_other_stage
,
3532 uint64_t *used_by_other_stage_patches
);
3533 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
3534 bool default_to_smooth_interp
);
3535 void nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3536 bool nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3539 void nir_assign_io_var_locations(struct exec_list
*var_list
,
3541 gl_shader_stage stage
);
3544 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
3545 * to 32-bit operations. This is only valid for nir_var_shader_in/out
3548 nir_lower_io_lower_64bit_to_32
= (1 << 0),
3550 /* If set, this forces all non-flat fragment shader inputs to be
3551 * interpolated as if with the "sample" qualifier. This requires
3552 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
3554 nir_lower_io_force_sample_interpolation
= (1 << 1),
3555 } nir_lower_io_options
;
3556 bool nir_lower_io(nir_shader
*shader
,
3557 nir_variable_mode modes
,
3558 int (*type_size
)(const struct glsl_type
*, bool),
3559 nir_lower_io_options
);
3561 bool nir_io_add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
);
3564 nir_lower_vars_to_explicit_types(nir_shader
*shader
,
3565 nir_variable_mode modes
,
3566 glsl_type_size_align_func type_info
);
3570 * An address format which is a simple 32-bit global GPU address.
3572 nir_address_format_32bit_global
,
3575 * An address format which is a simple 64-bit global GPU address.
3577 nir_address_format_64bit_global
,
3580 * An address format which is a bounds-checked 64-bit global GPU address.
3582 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
3583 * address stored with the low bits in .x and high bits in .y, .z is a
3584 * size, and .w is an offset. When the final I/O operation is lowered, .w
3585 * is checked against .z and the operation is predicated on the result.
3587 nir_address_format_64bit_bounded_global
,
3590 * An address format which is comprised of a vec2 where the first
3591 * component is a buffer index and the second is an offset.
3593 nir_address_format_32bit_index_offset
,
3596 * An address format which is a simple 32-bit offset.
3598 nir_address_format_32bit_offset
,
3601 * An address format representing a purely logical addressing model. In
3602 * this model, all deref chains must be complete from the dereference
3603 * operation to the variable. Cast derefs are not allowed. These
3604 * addresses will be 32-bit scalars but the format is immaterial because
3605 * you can always chase the chain.
3607 nir_address_format_logical
,
3608 } nir_address_format
;
3610 static inline unsigned
3611 nir_address_format_bit_size(nir_address_format addr_format
)
3613 switch (addr_format
) {
3614 case nir_address_format_32bit_global
: return 32;
3615 case nir_address_format_64bit_global
: return 64;
3616 case nir_address_format_64bit_bounded_global
: return 32;
3617 case nir_address_format_32bit_index_offset
: return 32;
3618 case nir_address_format_32bit_offset
: return 32;
3619 case nir_address_format_logical
: return 32;
3621 unreachable("Invalid address format");
3624 static inline unsigned
3625 nir_address_format_num_components(nir_address_format addr_format
)
3627 switch (addr_format
) {
3628 case nir_address_format_32bit_global
: return 1;
3629 case nir_address_format_64bit_global
: return 1;
3630 case nir_address_format_64bit_bounded_global
: return 4;
3631 case nir_address_format_32bit_index_offset
: return 2;
3632 case nir_address_format_32bit_offset
: return 1;
3633 case nir_address_format_logical
: return 1;
3635 unreachable("Invalid address format");
3638 static inline const struct glsl_type
*
3639 nir_address_format_to_glsl_type(nir_address_format addr_format
)
3641 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
3642 assert(bit_size
== 32 || bit_size
== 64);
3643 return glsl_vector_type(bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
,
3644 nir_address_format_num_components(addr_format
));
3647 const nir_const_value
*nir_address_format_null_value(nir_address_format addr_format
);
3649 nir_ssa_def
*nir_build_addr_ieq(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3650 nir_address_format addr_format
);
3652 nir_ssa_def
*nir_build_addr_isub(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3653 nir_address_format addr_format
);
3655 nir_ssa_def
* nir_explicit_io_address_from_deref(struct nir_builder
*b
,
3656 nir_deref_instr
*deref
,
3657 nir_ssa_def
*base_addr
,
3658 nir_address_format addr_format
);
3659 void nir_lower_explicit_io_instr(struct nir_builder
*b
,
3660 nir_intrinsic_instr
*io_instr
,
3662 nir_address_format addr_format
);
3664 bool nir_lower_explicit_io(nir_shader
*shader
,
3665 nir_variable_mode modes
,
3666 nir_address_format
);
3668 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
3669 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
3671 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
3673 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
3674 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
3675 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
3677 bool nir_remove_dead_derefs(nir_shader
*shader
);
3678 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
3679 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
);
3680 bool nir_lower_constant_initializers(nir_shader
*shader
,
3681 nir_variable_mode modes
);
3683 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
3684 bool nir_lower_vec_to_movs(nir_shader
*shader
);
3685 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
3687 const gl_state_index16
*alpha_ref_state_tokens
);
3688 bool nir_lower_alu(nir_shader
*shader
);
3690 bool nir_lower_flrp(nir_shader
*shader
, unsigned lowering_mask
,
3691 bool always_precise
, bool have_ffma
);
3693 bool nir_lower_alu_to_scalar(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
3694 bool nir_lower_bool_to_float(nir_shader
*shader
);
3695 bool nir_lower_bool_to_int32(nir_shader
*shader
);
3696 bool nir_lower_int_to_float(nir_shader
*shader
);
3697 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
3698 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
3699 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
3700 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
3701 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
3703 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
3704 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
3705 bool nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode mask
);
3707 void nir_lower_fragcoord_wtrans(nir_shader
*shader
);
3708 void nir_lower_viewport_transform(nir_shader
*shader
);
3709 bool nir_lower_uniforms_to_ubo(nir_shader
*shader
, int multiplier
);
3711 typedef struct nir_lower_subgroups_options
{
3712 uint8_t subgroup_size
;
3713 uint8_t ballot_bit_size
;
3714 bool lower_to_scalar
:1;
3715 bool lower_vote_trivial
:1;
3716 bool lower_vote_eq_to_ballot
:1;
3717 bool lower_subgroup_masks
:1;
3718 bool lower_shuffle
:1;
3719 bool lower_shuffle_to_32bit
:1;
3721 } nir_lower_subgroups_options
;
3723 bool nir_lower_subgroups(nir_shader
*shader
,
3724 const nir_lower_subgroups_options
*options
);
3726 bool nir_lower_system_values(nir_shader
*shader
);
3728 enum PACKED nir_lower_tex_packing
{
3729 nir_lower_tex_packing_none
= 0,
3730 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
3731 * or unsigned ints based on the sampler type
3733 nir_lower_tex_packing_16
,
3734 /* The sampler returns 1 32-bit word of 4x8 unorm */
3735 nir_lower_tex_packing_8
,
3738 typedef struct nir_lower_tex_options
{
3740 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
3741 * sampler types a texture projector is lowered.
3746 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
3748 bool lower_txf_offset
;
3751 * If true, lower away nir_tex_src_offset for all rect textures.
3753 bool lower_rect_offset
;
3756 * If true, lower rect textures to 2D, using txs to fetch the
3757 * texture dimensions and dividing the texture coords by the
3758 * texture dims to normalize.
3763 * If true, convert yuv to rgb.
3765 unsigned lower_y_uv_external
;
3766 unsigned lower_y_u_v_external
;
3767 unsigned lower_yx_xuxv_external
;
3768 unsigned lower_xy_uxvx_external
;
3769 unsigned lower_ayuv_external
;
3770 unsigned lower_xyuv_external
;
3773 * To emulate certain texture wrap modes, this can be used
3774 * to saturate the specified tex coord to [0.0, 1.0]. The
3775 * bits are according to sampler #, ie. if, for example:
3777 * (conf->saturate_s & (1 << n))
3779 * is true, then the s coord for sampler n is saturated.
3781 * Note that clamping must happen *after* projector lowering
3782 * so any projected texture sample instruction with a clamped
3783 * coordinate gets automatically lowered, regardless of the
3784 * 'lower_txp' setting.
3786 unsigned saturate_s
;
3787 unsigned saturate_t
;
3788 unsigned saturate_r
;
3790 /* Bitmask of textures that need swizzling.
3792 * If (swizzle_result & (1 << texture_index)), then the swizzle in
3793 * swizzles[texture_index] is applied to the result of the texturing
3796 unsigned swizzle_result
;
3798 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
3799 * while 4 and 5 represent 0 and 1 respectively.
3801 uint8_t swizzles
[32][4];
3803 /* Can be used to scale sampled values in range required by the format. */
3804 float scale_factors
[32];
3807 * Bitmap of textures that need srgb to linear conversion. If
3808 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
3809 * of the texture are lowered to linear.
3811 unsigned lower_srgb
;
3814 * If true, lower nir_texop_tex on shaders that doesn't support implicit
3815 * LODs to nir_texop_txl.
3817 bool lower_tex_without_implicit_lod
;
3820 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
3822 bool lower_txd_cube_map
;
3825 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
3830 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
3831 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
3832 * with lower_txd_cube_map.
3834 bool lower_txd_shadow
;
3837 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
3838 * Implies lower_txd_cube_map and lower_txd_shadow.
3843 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
3844 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
3846 bool lower_txb_shadow_clamp
;
3849 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
3850 * with nir_texop_txl. This includes cube maps.
3852 bool lower_txd_shadow_clamp
;
3855 * If true, lower nir_texop_txd on when it uses both offset and min_lod
3856 * with nir_texop_txl. This includes cube maps.
3858 bool lower_txd_offset_clamp
;
3861 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
3862 * sampler is bindless.
3864 bool lower_txd_clamp_bindless_sampler
;
3867 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
3868 * sampler index is not statically determinable to be less than 16.
3870 bool lower_txd_clamp_if_sampler_index_not_lt_16
;
3873 * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
3874 * 0-lod followed by a nir_ishr.
3879 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
3880 * mixed-up tg4 locations.
3882 bool lower_tg4_broadcom_swizzle
;
3885 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
3887 bool lower_tg4_offsets
;
3889 enum nir_lower_tex_packing lower_tex_packing
[32];
3890 } nir_lower_tex_options
;
3892 bool nir_lower_tex(nir_shader
*shader
,
3893 const nir_lower_tex_options
*options
);
3895 enum nir_lower_non_uniform_access_type
{
3896 nir_lower_non_uniform_ubo_access
= (1 << 0),
3897 nir_lower_non_uniform_ssbo_access
= (1 << 1),
3898 nir_lower_non_uniform_texture_access
= (1 << 2),
3899 nir_lower_non_uniform_image_access
= (1 << 3),
3902 bool nir_lower_non_uniform_access(nir_shader
*shader
,
3903 enum nir_lower_non_uniform_access_type
);
3905 bool nir_lower_idiv(nir_shader
*shader
);
3907 bool nir_lower_input_attachments(nir_shader
*shader
, bool use_fragcoord_sysval
);
3909 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
, bool use_vars
);
3910 bool nir_lower_clip_gs(nir_shader
*shader
, unsigned ucp_enables
);
3911 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
);
3912 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
3914 bool nir_lower_frexp(nir_shader
*nir
);
3916 void nir_lower_two_sided_color(nir_shader
*shader
);
3918 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
3920 bool nir_lower_flatshade(nir_shader
*shader
);
3922 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
3923 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
3924 const gl_state_index16
*uniform_state_tokens
);
3926 typedef struct nir_lower_wpos_ytransform_options
{
3927 gl_state_index16 state_tokens
[STATE_LENGTH
];
3928 bool fs_coord_origin_upper_left
:1;
3929 bool fs_coord_origin_lower_left
:1;
3930 bool fs_coord_pixel_center_integer
:1;
3931 bool fs_coord_pixel_center_half_integer
:1;
3932 } nir_lower_wpos_ytransform_options
;
3934 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
3935 const nir_lower_wpos_ytransform_options
*options
);
3936 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
3938 bool nir_lower_fb_read(nir_shader
*shader
);
3940 typedef struct nir_lower_drawpixels_options
{
3941 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
3942 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
3943 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
3944 unsigned drawpix_sampler
;
3945 unsigned pixelmap_sampler
;
3947 bool scale_and_bias
:1;
3948 } nir_lower_drawpixels_options
;
3950 void nir_lower_drawpixels(nir_shader
*shader
,
3951 const nir_lower_drawpixels_options
*options
);
3953 typedef struct nir_lower_bitmap_options
{
3956 } nir_lower_bitmap_options
;
3958 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
3960 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
, unsigned ssbo_offset
);
3963 nir_lower_int_source_mods
= 1 << 0,
3964 nir_lower_float_source_mods
= 1 << 1,
3965 nir_lower_triop_abs
= 1 << 2,
3966 nir_lower_all_source_mods
= (1 << 3) - 1
3967 } nir_lower_to_source_mods_flags
;
3970 bool nir_lower_to_source_mods(nir_shader
*shader
, nir_lower_to_source_mods_flags options
);
3972 bool nir_lower_gs_intrinsics(nir_shader
*shader
);
3974 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
3976 bool nir_lower_bit_size(nir_shader
*shader
,
3977 nir_lower_bit_size_callback callback
,
3978 void *callback_data
);
3980 nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode
);
3981 bool nir_lower_int64(nir_shader
*shader
, nir_lower_int64_options options
);
3983 nir_lower_doubles_options
nir_lower_doubles_op_to_options_mask(nir_op opcode
);
3984 bool nir_lower_doubles(nir_shader
*shader
, const nir_shader
*softfp64
,
3985 nir_lower_doubles_options options
);
3986 bool nir_lower_pack(nir_shader
*shader
);
3988 bool nir_lower_point_size(nir_shader
*shader
, float min
, float max
);
3991 nir_lower_interpolation_at_sample
= (1 << 1),
3992 nir_lower_interpolation_at_offset
= (1 << 2),
3993 nir_lower_interpolation_centroid
= (1 << 3),
3994 nir_lower_interpolation_pixel
= (1 << 4),
3995 nir_lower_interpolation_sample
= (1 << 5),
3996 } nir_lower_interpolation_options
;
3998 bool nir_lower_interpolation(nir_shader
*shader
,
3999 nir_lower_interpolation_options options
);
4001 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
4003 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
4005 void nir_loop_analyze_impl(nir_function_impl
*impl
,
4006 nir_variable_mode indirect_mask
);
4008 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
4010 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
4011 bool nir_repair_ssa(nir_shader
*shader
);
4013 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
4014 bool nir_convert_to_lcssa(nir_shader
*shader
, bool skip_invariants
, bool skip_bool_invariants
);
4015 bool* nir_divergence_analysis(nir_shader
*shader
, nir_divergence_options options
);
4017 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
4018 * registers. If false, convert all values (even those not involved in a phi
4019 * node) to registers.
4021 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
4023 bool nir_lower_phis_to_regs_block(nir_block
*block
);
4024 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
4025 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
);
4027 bool nir_lower_samplers(nir_shader
*shader
);
4029 /* This is here for unit tests. */
4030 bool nir_opt_comparison_pre_impl(nir_function_impl
*impl
);
4032 bool nir_opt_comparison_pre(nir_shader
*shader
);
4034 bool nir_opt_access(nir_shader
*shader
);
4035 bool nir_opt_algebraic(nir_shader
*shader
);
4036 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
4037 bool nir_opt_algebraic_late(nir_shader
*shader
);
4038 bool nir_opt_constant_folding(nir_shader
*shader
);
4040 bool nir_opt_combine_stores(nir_shader
*shader
, nir_variable_mode modes
);
4042 bool nir_copy_prop(nir_shader
*shader
);
4044 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
4046 bool nir_opt_cse(nir_shader
*shader
);
4048 bool nir_opt_dce(nir_shader
*shader
);
4050 bool nir_opt_dead_cf(nir_shader
*shader
);
4052 bool nir_opt_dead_write_vars(nir_shader
*shader
);
4054 bool nir_opt_deref_impl(nir_function_impl
*impl
);
4055 bool nir_opt_deref(nir_shader
*shader
);
4057 bool nir_opt_find_array_copies(nir_shader
*shader
);
4059 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
4061 bool nir_opt_idiv_const(nir_shader
*shader
, unsigned min_bit_size
);
4063 bool nir_opt_if(nir_shader
*shader
, bool aggressive_last_continue
);
4065 bool nir_opt_intrinsics(nir_shader
*shader
);
4067 bool nir_opt_large_constants(nir_shader
*shader
,
4068 glsl_type_size_align_func size_align
,
4069 unsigned threshold
);
4071 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
4074 nir_move_const_undef
= (1 << 0),
4075 nir_move_load_ubo
= (1 << 1),
4076 nir_move_load_input
= (1 << 2),
4077 nir_move_comparisons
= (1 << 3),
4080 bool nir_can_move_instr(nir_instr
*instr
, nir_move_options options
);
4082 bool nir_opt_sink(nir_shader
*shader
, nir_move_options options
);
4084 bool nir_opt_move(nir_shader
*shader
, nir_move_options options
);
4086 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
,
4087 bool indirect_load_ok
, bool expensive_alu_ok
);
4089 bool nir_opt_rematerialize_compares(nir_shader
*shader
);
4091 bool nir_opt_remove_phis(nir_shader
*shader
);
4092 bool nir_opt_remove_phis_block(nir_block
*block
);
4094 bool nir_opt_shrink_load(nir_shader
*shader
);
4096 bool nir_opt_trivial_continues(nir_shader
*shader
);
4098 bool nir_opt_undef(nir_shader
*shader
);
4100 bool nir_opt_vectorize(nir_shader
*shader
);
4102 bool nir_opt_conditional_discard(nir_shader
*shader
);
4104 void nir_strip(nir_shader
*shader
);
4106 void nir_sweep(nir_shader
*shader
);
4108 void nir_remap_dual_slot_attributes(nir_shader
*shader
,
4109 uint64_t *dual_slot_inputs
);
4110 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
);
4112 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
4113 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);
4116 nir_variable_is_in_ubo(const nir_variable
*var
)
4118 return (var
->data
.mode
== nir_var_mem_ubo
&&
4119 var
->interface_type
!= NULL
);
4123 nir_variable_is_in_ssbo(const nir_variable
*var
)
4125 return (var
->data
.mode
== nir_var_mem_ssbo
&&
4126 var
->interface_type
!= NULL
);
4130 nir_variable_is_in_block(const nir_variable
*var
)
4132 return nir_variable_is_in_ubo(var
) || nir_variable_is_in_ssbo(var
);