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 "util/format/u_format.h"
41 #include "compiler/nir_types.h"
42 #include "compiler/shader_enums.h"
43 #include "compiler/shader_info.h"
44 #define XXH_INLINE_ALL
45 #include "util/xxhash.h"
49 #include "util/debug.h"
52 #include "nir_opcodes.h"
54 #if defined(_WIN32) && !defined(snprintf)
55 #define snprintf _snprintf
63 #define NIR_TRUE (~0u)
64 #define NIR_MAX_VEC_COMPONENTS 16
65 #define NIR_MAX_MATRIX_COLUMNS 4
66 #define NIR_STREAM_PACKED (1 << 8)
67 typedef uint16_t nir_component_mask_t
;
70 nir_num_components_valid(unsigned num_components
)
72 return (num_components
>= 1 &&
73 num_components
<= 4) ||
74 num_components
== 8 ||
78 /** Defines a cast function
80 * This macro defines a cast function from in_type to out_type where
81 * out_type is some structure type that contains a field of type out_type.
83 * Note that you have to be a bit careful as the generated cast function
86 #define NIR_DEFINE_CAST(name, in_type, out_type, field, \
87 type_field, type_value) \
88 static inline out_type * \
89 name(const in_type *parent) \
91 assert(parent && parent->type_field == type_value); \
92 return exec_node_data(out_type, parent, field); \
102 * Description of built-in state associated with a uniform
104 * \sa nir_variable::state_slots
107 gl_state_index16 tokens
[STATE_LENGTH
];
112 nir_var_shader_in
= (1 << 0),
113 nir_var_shader_out
= (1 << 1),
114 nir_var_shader_temp
= (1 << 2),
115 nir_var_function_temp
= (1 << 3),
116 nir_var_uniform
= (1 << 4),
117 nir_var_mem_ubo
= (1 << 5),
118 nir_var_system_value
= (1 << 6),
119 nir_var_mem_ssbo
= (1 << 7),
120 nir_var_mem_shared
= (1 << 8),
121 nir_var_mem_global
= (1 << 9),
122 nir_var_mem_push_const
= (1 << 10), /* not actually used for variables */
123 nir_num_variable_modes
= 11,
124 nir_var_all
= (1 << nir_num_variable_modes
) - 1,
131 nir_rounding_mode_undef
= 0,
132 nir_rounding_mode_rtne
= 1, /* round to nearest even */
133 nir_rounding_mode_ru
= 2, /* round up */
134 nir_rounding_mode_rd
= 3, /* round down */
135 nir_rounding_mode_rtz
= 4, /* round towards zero */
152 #define nir_const_value_to_array(arr, c, components, m) \
154 for (unsigned i = 0; i < components; ++i) \
158 static inline nir_const_value
159 nir_const_value_for_raw_uint(uint64_t x
, unsigned bit_size
)
162 memset(&v
, 0, sizeof(v
));
165 case 1: v
.b
= x
; break;
166 case 8: v
.u8
= x
; break;
167 case 16: v
.u16
= x
; break;
168 case 32: v
.u32
= x
; break;
169 case 64: v
.u64
= x
; break;
171 unreachable("Invalid bit size");
177 static inline nir_const_value
178 nir_const_value_for_int(int64_t i
, unsigned bit_size
)
181 memset(&v
, 0, sizeof(v
));
183 assert(bit_size
<= 64);
185 assert(i
>= (-(1ll << (bit_size
- 1))));
186 assert(i
< (1ll << (bit_size
- 1)));
189 return nir_const_value_for_raw_uint(i
, bit_size
);
192 static inline nir_const_value
193 nir_const_value_for_uint(uint64_t u
, unsigned bit_size
)
196 memset(&v
, 0, sizeof(v
));
198 assert(bit_size
<= 64);
200 assert(u
< (1ull << bit_size
));
202 return nir_const_value_for_raw_uint(u
, bit_size
);
205 static inline nir_const_value
206 nir_const_value_for_bool(bool b
, unsigned bit_size
)
208 /* Booleans use a 0/-1 convention */
209 return nir_const_value_for_int(-(int)b
, bit_size
);
212 /* This one isn't inline because it requires half-float conversion */
213 nir_const_value
nir_const_value_for_float(double b
, unsigned bit_size
);
215 static inline int64_t
216 nir_const_value_as_int(nir_const_value value
, unsigned bit_size
)
219 /* int1_t uses 0/-1 convention */
220 case 1: return -(int)value
.b
;
221 case 8: return value
.i8
;
222 case 16: return value
.i16
;
223 case 32: return value
.i32
;
224 case 64: return value
.i64
;
226 unreachable("Invalid bit size");
230 static inline uint64_t
231 nir_const_value_as_uint(nir_const_value value
, unsigned bit_size
)
234 case 1: return value
.b
;
235 case 8: return value
.u8
;
236 case 16: return value
.u16
;
237 case 32: return value
.u32
;
238 case 64: return value
.u64
;
240 unreachable("Invalid bit size");
245 nir_const_value_as_bool(nir_const_value value
, unsigned bit_size
)
247 int64_t i
= nir_const_value_as_int(value
, bit_size
);
249 /* Booleans of any size use 0/-1 convention */
250 assert(i
== 0 || i
== -1);
255 /* This one isn't inline because it requires half-float conversion */
256 double nir_const_value_as_float(nir_const_value value
, unsigned bit_size
);
258 typedef struct nir_constant
{
260 * Value of the constant.
262 * The field used to back the values supplied by the constant is determined
263 * by the type associated with the \c nir_variable. Constants may be
264 * scalars, vectors, or matrices.
266 nir_const_value values
[NIR_MAX_VEC_COMPONENTS
];
268 /* we could get this from the var->type but makes clone *much* easier to
269 * not have to care about the type.
271 unsigned num_elements
;
273 /* Array elements / Structure Fields */
274 struct nir_constant
**elements
;
278 * \brief Layout qualifiers for gl_FragDepth.
280 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
281 * with a layout qualifier.
284 nir_depth_layout_none
, /**< No depth layout is specified. */
285 nir_depth_layout_any
,
286 nir_depth_layout_greater
,
287 nir_depth_layout_less
,
288 nir_depth_layout_unchanged
292 * Enum keeping track of how a variable was declared.
296 * Normal declaration.
298 nir_var_declared_normally
= 0,
301 * Variable is implicitly generated by the compiler and should not be
302 * visible via the API.
305 } nir_var_declaration_type
;
308 * Either a uniform, global variable, shader input, or shader output. Based on
309 * ir_variable - it should be easy to translate between the two.
312 typedef struct nir_variable
{
313 struct exec_node node
;
316 * Declared type of the variable
318 const struct glsl_type
*type
;
321 * Declared name of the variable
325 struct nir_variable_data
{
327 * Storage class of the variable.
329 * \sa nir_variable_mode
331 nir_variable_mode mode
:11;
334 * Is the variable read-only?
336 * This is set for variables declared as \c const, shader inputs,
339 unsigned read_only
:1;
343 unsigned invariant
:1;
346 * Precision qualifier.
348 * In desktop GLSL we do not care about precision qualifiers at all, in
349 * fact, the spec says that precision qualifiers are ignored.
351 * To make things easy, we make it so that this field is always
352 * GLSL_PRECISION_NONE on desktop shaders. This way all the variables
353 * have the same precision value and the checks we add in the compiler
354 * for this field will never break a desktop shader compile.
356 unsigned precision
:2;
359 * Can this variable be coalesced with another?
361 * This is set by nir_lower_io_to_temporaries to say that any
362 * copies involving this variable should stay put. Propagating it can
363 * duplicate the resulting load/store, which is not wanted, and may
364 * result in a load/store of the variable with an indirect offset which
365 * the backend may not be able to handle.
367 unsigned cannot_coalesce
:1;
370 * When separate shader programs are enabled, only input/outputs between
371 * the stages of a multi-stage separate program can be safely removed
372 * from the shader interface. Other input/outputs must remains active.
374 * This is also used to make sure xfb varyings that are unused by the
375 * fragment shader are not removed.
377 unsigned always_active_io
:1;
380 * Interpolation mode for shader inputs / outputs
382 * \sa glsl_interp_mode
384 unsigned interpolation
:3;
387 * If non-zero, then this variable may be packed along with other variables
388 * into a single varying slot, so this offset should be applied when
389 * accessing components. For example, an offset of 1 means that the x
390 * component of this variable is actually stored in component y of the
391 * location specified by \c location.
393 unsigned location_frac
:2;
396 * If true, this variable represents an array of scalars that should
397 * be tightly packed. In other words, consecutive array elements
398 * should be stored one component apart, rather than one slot apart.
403 * Whether this is a fragment shader output implicitly initialized with
404 * the previous contents of the specified render target at the
405 * framebuffer location corresponding to this shader invocation.
407 unsigned fb_fetch_output
:1;
410 * Non-zero if this variable is considered bindless as defined by
411 * ARB_bindless_texture.
416 * Was an explicit binding set in the shader?
418 unsigned explicit_binding
:1;
421 * Was the location explicitly set in the shader?
423 * If the location is explicitly set in the shader, it \b cannot be changed
424 * by the linker or by the API (e.g., calls to \c glBindAttribLocation have
427 unsigned explicit_location
:1;
430 * Was a transfer feedback buffer set in the shader?
432 unsigned explicit_xfb_buffer
:1;
435 * Was a transfer feedback stride set in the shader?
437 unsigned explicit_xfb_stride
:1;
440 * Was an explicit offset set in the shader?
442 unsigned explicit_offset
:1;
445 * Layout of the matrix. Uses glsl_matrix_layout values.
447 unsigned matrix_layout
:2;
450 * Non-zero if this variable was created by lowering a named interface
453 unsigned from_named_ifc_block
:1;
456 * How the variable was declared. See nir_var_declaration_type.
458 * This is used to detect variables generated by the compiler, so should
459 * not be visible via the API.
461 unsigned how_declared
:2;
464 * Is this variable per-view? If so, we know it must be an array with
465 * size corresponding to the number of views.
470 * \brief Layout qualifier for gl_FragDepth.
472 * This is not equal to \c ir_depth_layout_none if and only if this
473 * variable is \c gl_FragDepth and a layout qualifier is specified.
475 nir_depth_layout depth_layout
:3;
478 * Vertex stream output identifier.
480 * For packed outputs, NIR_STREAM_PACKED is set and bits [2*i+1,2*i]
481 * indicate the stream of the i-th component.
486 * Access flags for memory variables (SSBO/global), image uniforms, and
487 * bindless images in uniforms/inputs/outputs.
489 enum gl_access_qualifier access
:8;
492 * Descriptor set binding for sampler or UBO.
494 unsigned descriptor_set
:5;
497 * output index for dual source blending.
502 * Initial binding point for a sampler or UBO.
504 * For array types, this represents the binding point for the first element.
509 * Storage location of the base of this variable
511 * The precise meaning of this field depends on the nature of the variable.
513 * - Vertex shader input: one of the values from \c gl_vert_attrib.
514 * - Vertex shader output: one of the values from \c gl_varying_slot.
515 * - Geometry shader input: one of the values from \c gl_varying_slot.
516 * - Geometry shader output: one of the values from \c gl_varying_slot.
517 * - Fragment shader input: one of the values from \c gl_varying_slot.
518 * - Fragment shader output: one of the values from \c gl_frag_result.
519 * - Uniforms: Per-stage uniform slot number for default uniform block.
520 * - Uniforms: Index within the uniform block definition for UBO members.
521 * - Non-UBO Uniforms: uniform slot number.
522 * - Other: This field is not currently used.
524 * If the variable is a uniform, shader input, or shader output, and the
525 * slot has not been assigned, the value will be -1.
530 * The actual location of the variable in the IR. Only valid for inputs,
531 * outputs, and uniforms (including samplers and images).
533 unsigned driver_location
;
536 * Location an atomic counter or transform feedback is stored at.
542 /** Image internal format if specified explicitly, otherwise PIPE_FORMAT_NONE. */
543 enum pipe_format format
;
548 * Transform feedback buffer.
553 * Transform feedback stride.
561 * Identifier for this variable generated by nir_index_vars() that is unique
562 * among other variables in the same exec_list.
566 /* Number of nir_variable_data members */
567 uint16_t num_members
;
570 * Built-in state that backs this uniform
572 * Once set at variable creation, \c state_slots must remain invariant.
573 * This is because, ideally, this array would be shared by all clones of
574 * this variable in the IR tree. In other words, we'd really like for it
575 * to be a fly-weight.
577 * If the variable is not a uniform, \c num_state_slots will be zero and
578 * \c state_slots will be \c NULL.
581 uint16_t num_state_slots
; /**< Number of state slots used */
582 nir_state_slot
*state_slots
; /**< State descriptors. */
586 * Constant expression assigned in the initializer of the variable
588 * This field should only be used temporarily by creators of NIR shaders
589 * and then lower_constant_initializers can be used to get rid of them.
590 * Most of the rest of NIR ignores this field or asserts that it's NULL.
592 nir_constant
*constant_initializer
;
595 * Global variable assigned in the initializer of the variable
596 * This field should only be used temporarily by creators of NIR shaders
597 * and then lower_constant_initializers can be used to get rid of them.
598 * Most of the rest of NIR ignores this field or asserts that it's NULL.
600 struct nir_variable
*pointer_initializer
;
603 * For variables that are in an interface block or are an instance of an
604 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
606 * \sa ir_variable::location
608 const struct glsl_type
*interface_type
;
611 * Description of per-member data for per-member struct variables
613 * This is used for variables which are actually an amalgamation of
614 * multiple entities such as a struct of built-in values or a struct of
615 * inputs each with their own layout specifier. This is only allowed on
616 * variables with a struct or array of array of struct type.
618 struct nir_variable_data
*members
;
621 #define nir_foreach_variable(var, var_list) \
622 foreach_list_typed(nir_variable, var, node, var_list)
624 #define nir_foreach_variable_safe(var, var_list) \
625 foreach_list_typed_safe(nir_variable, var, node, var_list)
628 nir_variable_is_global(const nir_variable
*var
)
630 return var
->data
.mode
!= nir_var_function_temp
;
633 typedef struct nir_register
{
634 struct exec_node node
;
636 unsigned num_components
; /** < number of vector components */
637 unsigned num_array_elems
; /** < size of array (0 for no array) */
639 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
642 /** generic register index. */
645 /** only for debug purposes, can be NULL */
648 /** set of nir_srcs where this register is used (read from) */
649 struct list_head uses
;
651 /** set of nir_dests where this register is defined (written to) */
652 struct list_head defs
;
654 /** set of nir_ifs where this register is used as a condition */
655 struct list_head if_uses
;
658 #define nir_foreach_register(reg, reg_list) \
659 foreach_list_typed(nir_register, reg, node, reg_list)
660 #define nir_foreach_register_safe(reg, reg_list) \
661 foreach_list_typed_safe(nir_register, reg, node, reg_list)
663 typedef enum PACKED
{
665 nir_instr_type_deref
,
668 nir_instr_type_intrinsic
,
669 nir_instr_type_load_const
,
671 nir_instr_type_ssa_undef
,
673 nir_instr_type_parallel_copy
,
676 typedef struct nir_instr
{
677 struct exec_node node
;
678 struct nir_block
*block
;
681 /* A temporary for optimization and analysis passes to use for storing
682 * flags. For instance, DCE uses this to store the "dead/live" info.
686 /** generic instruction index. */
690 static inline nir_instr
*
691 nir_instr_next(nir_instr
*instr
)
693 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
694 if (exec_node_is_tail_sentinel(next
))
697 return exec_node_data(nir_instr
, next
, node
);
700 static inline nir_instr
*
701 nir_instr_prev(nir_instr
*instr
)
703 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
704 if (exec_node_is_head_sentinel(prev
))
707 return exec_node_data(nir_instr
, prev
, node
);
711 nir_instr_is_first(const nir_instr
*instr
)
713 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr
->node
));
717 nir_instr_is_last(const nir_instr
*instr
)
719 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr
->node
));
722 typedef struct nir_ssa_def
{
723 /** for debugging only, can be NULL */
726 /** generic SSA definition index. */
729 /** Index into the live_in and live_out bitfields */
732 /** Instruction which produces this SSA value. */
733 nir_instr
*parent_instr
;
735 /** set of nir_instrs where this register is used (read from) */
736 struct list_head uses
;
738 /** set of nir_ifs where this register is used as a condition */
739 struct list_head if_uses
;
741 uint8_t num_components
;
743 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
747 * True if this SSA value may have different values in different SIMD
748 * invocations of the shader. This is set by nir_divergence_analysis.
757 struct nir_src
*indirect
; /** < NULL for no indirect offset */
758 unsigned base_offset
;
760 /* TODO use-def chain goes here */
764 nir_instr
*parent_instr
;
765 struct list_head def_link
;
768 struct nir_src
*indirect
; /** < NULL for no indirect offset */
769 unsigned base_offset
;
771 /* TODO def-use chain goes here */
776 typedef struct nir_src
{
778 /** Instruction that consumes this value as a source. */
779 nir_instr
*parent_instr
;
780 struct nir_if
*parent_if
;
783 struct list_head use_link
;
793 static inline nir_src
796 nir_src src
= { { NULL
} };
800 #define NIR_SRC_INIT nir_src_init()
802 #define nir_foreach_use(src, reg_or_ssa_def) \
803 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
805 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
806 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
808 #define nir_foreach_if_use(src, reg_or_ssa_def) \
809 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
811 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
812 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
823 static inline nir_dest
826 nir_dest dest
= { { { NULL
} } };
830 #define NIR_DEST_INIT nir_dest_init()
832 #define nir_foreach_def(dest, reg) \
833 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
835 #define nir_foreach_def_safe(dest, reg) \
836 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
838 static inline nir_src
839 nir_src_for_ssa(nir_ssa_def
*def
)
841 nir_src src
= NIR_SRC_INIT
;
849 static inline nir_src
850 nir_src_for_reg(nir_register
*reg
)
852 nir_src src
= NIR_SRC_INIT
;
856 src
.reg
.indirect
= NULL
;
857 src
.reg
.base_offset
= 0;
862 static inline nir_dest
863 nir_dest_for_reg(nir_register
*reg
)
865 nir_dest dest
= NIR_DEST_INIT
;
872 static inline unsigned
873 nir_src_bit_size(nir_src src
)
875 return src
.is_ssa
? src
.ssa
->bit_size
: src
.reg
.reg
->bit_size
;
878 static inline unsigned
879 nir_src_num_components(nir_src src
)
881 return src
.is_ssa
? src
.ssa
->num_components
: src
.reg
.reg
->num_components
;
885 nir_src_is_const(nir_src src
)
888 src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
;
892 nir_src_is_divergent(nir_src src
)
895 return src
.ssa
->divergent
;
898 static inline unsigned
899 nir_dest_bit_size(nir_dest dest
)
901 return dest
.is_ssa
? dest
.ssa
.bit_size
: dest
.reg
.reg
->bit_size
;
904 static inline unsigned
905 nir_dest_num_components(nir_dest dest
)
907 return dest
.is_ssa
? dest
.ssa
.num_components
: dest
.reg
.reg
->num_components
;
911 nir_dest_is_divergent(nir_dest dest
)
914 return dest
.ssa
.divergent
;
917 /* Are all components the same, ie. .xxxx */
919 nir_is_same_comp_swizzle(uint8_t *swiz
, unsigned nr_comp
)
921 for (unsigned i
= 1; i
< nr_comp
; i
++)
922 if (swiz
[i
] != swiz
[0])
927 /* Are all components sequential, ie. .yzw */
929 nir_is_sequential_comp_swizzle(uint8_t *swiz
, unsigned nr_comp
)
931 for (unsigned i
= 1; i
< nr_comp
; i
++)
932 if (swiz
[i
] != (swiz
[0] + i
))
937 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
938 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
944 * \name input modifiers
948 * For inputs interpreted as floating point, flips the sign bit. For
949 * inputs interpreted as integers, performs the two's complement negation.
954 * Clears the sign bit for floating point values, and computes the integer
955 * absolute value for integers. Note that the negate modifier acts after
956 * the absolute value modifier, therefore if both are set then all inputs
957 * will become negative.
963 * For each input component, says which component of the register it is
964 * chosen from. Note that which elements of the swizzle are used and which
965 * are ignored are based on the write mask for most opcodes - for example,
966 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
967 * a swizzle of {2, x, 1, 0} where x means "don't care."
969 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
976 * \name saturate output modifier
978 * Only valid for opcodes that output floating-point numbers. Clamps the
979 * output to between 0.0 and 1.0 inclusive.
984 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
987 /** NIR sized and unsized types
989 * The values in this enum are carefully chosen so that the sized type is
990 * just the unsized type OR the number of bits.
992 typedef enum PACKED
{
993 nir_type_invalid
= 0, /* Not a valid type */
997 nir_type_float
= 128,
998 nir_type_bool1
= 1 | nir_type_bool
,
999 nir_type_bool8
= 8 | nir_type_bool
,
1000 nir_type_bool16
= 16 | nir_type_bool
,
1001 nir_type_bool32
= 32 | nir_type_bool
,
1002 nir_type_int1
= 1 | nir_type_int
,
1003 nir_type_int8
= 8 | nir_type_int
,
1004 nir_type_int16
= 16 | nir_type_int
,
1005 nir_type_int32
= 32 | nir_type_int
,
1006 nir_type_int64
= 64 | nir_type_int
,
1007 nir_type_uint1
= 1 | nir_type_uint
,
1008 nir_type_uint8
= 8 | nir_type_uint
,
1009 nir_type_uint16
= 16 | nir_type_uint
,
1010 nir_type_uint32
= 32 | nir_type_uint
,
1011 nir_type_uint64
= 64 | nir_type_uint
,
1012 nir_type_float16
= 16 | nir_type_float
,
1013 nir_type_float32
= 32 | nir_type_float
,
1014 nir_type_float64
= 64 | nir_type_float
,
1017 #define NIR_ALU_TYPE_SIZE_MASK 0x79
1018 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
1020 static inline unsigned
1021 nir_alu_type_get_type_size(nir_alu_type type
)
1023 return type
& NIR_ALU_TYPE_SIZE_MASK
;
1026 static inline nir_alu_type
1027 nir_alu_type_get_base_type(nir_alu_type type
)
1029 return (nir_alu_type
)(type
& NIR_ALU_TYPE_BASE_TYPE_MASK
);
1032 static inline nir_alu_type
1033 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
1035 switch (base_type
) {
1036 case GLSL_TYPE_BOOL
:
1037 return nir_type_bool1
;
1039 case GLSL_TYPE_UINT
:
1040 return nir_type_uint32
;
1043 return nir_type_int32
;
1045 case GLSL_TYPE_UINT16
:
1046 return nir_type_uint16
;
1048 case GLSL_TYPE_INT16
:
1049 return nir_type_int16
;
1051 case GLSL_TYPE_UINT8
:
1052 return nir_type_uint8
;
1053 case GLSL_TYPE_INT8
:
1054 return nir_type_int8
;
1055 case GLSL_TYPE_UINT64
:
1056 return nir_type_uint64
;
1058 case GLSL_TYPE_INT64
:
1059 return nir_type_int64
;
1061 case GLSL_TYPE_FLOAT
:
1062 return nir_type_float32
;
1064 case GLSL_TYPE_FLOAT16
:
1065 return nir_type_float16
;
1067 case GLSL_TYPE_DOUBLE
:
1068 return nir_type_float64
;
1071 case GLSL_TYPE_SAMPLER
:
1072 case GLSL_TYPE_IMAGE
:
1073 case GLSL_TYPE_ATOMIC_UINT
:
1074 case GLSL_TYPE_STRUCT
:
1075 case GLSL_TYPE_INTERFACE
:
1076 case GLSL_TYPE_ARRAY
:
1077 case GLSL_TYPE_VOID
:
1078 case GLSL_TYPE_SUBROUTINE
:
1079 case GLSL_TYPE_FUNCTION
:
1080 case GLSL_TYPE_ERROR
:
1081 return nir_type_invalid
;
1084 unreachable("unknown type");
1087 static inline nir_alu_type
1088 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
1090 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
1093 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
1094 nir_rounding_mode rnd
);
1096 static inline nir_op
1097 nir_op_vec(unsigned components
)
1099 switch (components
) {
1100 case 1: return nir_op_mov
;
1101 case 2: return nir_op_vec2
;
1102 case 3: return nir_op_vec3
;
1103 case 4: return nir_op_vec4
;
1104 case 8: return nir_op_vec8
;
1105 case 16: return nir_op_vec16
;
1106 default: unreachable("bad component count");
1111 nir_op_is_vec(nir_op op
)
1127 nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode
, unsigned bit_size
)
1129 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
) ||
1130 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
) ||
1131 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
);
1135 nir_is_denorm_flush_to_zero(unsigned execution_mode
, unsigned bit_size
)
1137 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
) ||
1138 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
) ||
1139 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
);
1143 nir_is_denorm_preserve(unsigned execution_mode
, unsigned bit_size
)
1145 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP16
) ||
1146 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP32
) ||
1147 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP64
);
1151 nir_is_rounding_mode_rtne(unsigned execution_mode
, unsigned bit_size
)
1153 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1154 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1155 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1159 nir_is_rounding_mode_rtz(unsigned execution_mode
, unsigned bit_size
)
1161 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1162 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1163 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1167 nir_has_any_rounding_mode_rtz(unsigned execution_mode
)
1169 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1170 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1171 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1175 nir_has_any_rounding_mode_rtne(unsigned execution_mode
)
1177 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1178 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1179 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1182 static inline nir_rounding_mode
1183 nir_get_rounding_mode_from_float_controls(unsigned execution_mode
,
1186 if (nir_alu_type_get_base_type(type
) != nir_type_float
)
1187 return nir_rounding_mode_undef
;
1189 unsigned bit_size
= nir_alu_type_get_type_size(type
);
1191 if (nir_is_rounding_mode_rtz(execution_mode
, bit_size
))
1192 return nir_rounding_mode_rtz
;
1193 if (nir_is_rounding_mode_rtne(execution_mode
, bit_size
))
1194 return nir_rounding_mode_rtne
;
1195 return nir_rounding_mode_undef
;
1199 nir_has_any_rounding_mode_enabled(unsigned execution_mode
)
1202 nir_has_any_rounding_mode_rtne(execution_mode
) ||
1203 nir_has_any_rounding_mode_rtz(execution_mode
);
1209 * Operation where the first two sources are commutative.
1211 * For 2-source operations, this just mathematical commutativity. Some
1212 * 3-source operations, like ffma, are only commutative in the first two
1215 NIR_OP_IS_2SRC_COMMUTATIVE
= (1 << 0),
1216 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
1217 } nir_op_algebraic_property
;
1225 * The number of components in the output
1227 * If non-zero, this is the size of the output and input sizes are
1228 * explicitly given; swizzle and writemask are still in effect, but if
1229 * the output component is masked out, then the input component may
1232 * If zero, the opcode acts in the standard, per-component manner; the
1233 * operation is performed on each component (except the ones that are
1234 * masked out) with the input being taken from the input swizzle for
1237 * The size of some of the inputs may be given (i.e. non-zero) even
1238 * though output_size is zero; in that case, the inputs with a zero
1239 * size act per-component, while the inputs with non-zero size don't.
1241 uint8_t output_size
;
1244 * The type of vector that the instruction outputs. Note that the
1245 * staurate modifier is only allowed on outputs with the float type.
1248 nir_alu_type output_type
;
1251 * The number of components in each input
1253 uint8_t input_sizes
[NIR_MAX_VEC_COMPONENTS
];
1256 * The type of vector that each input takes. Note that negate and
1257 * absolute value are only allowed on inputs with int or float type and
1258 * behave differently on the two.
1260 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
1262 nir_op_algebraic_property algebraic_properties
;
1264 /* Whether this represents a numeric conversion opcode */
1268 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
1270 typedef struct nir_alu_instr
{
1274 /** Indicates that this ALU instruction generates an exact value
1276 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1277 * really equivalent to either. This indicates that the value generated by
1278 * this operation is high-precision and any code transformations that touch
1279 * it must ensure that the resulting value is bit-for-bit identical to the
1285 * Indicates that this instruction do not cause wrapping to occur, in the
1286 * form of overflow or underflow.
1288 bool no_signed_wrap
:1;
1289 bool no_unsigned_wrap
:1;
1295 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
1296 nir_alu_instr
*instr
);
1297 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
1298 nir_alu_instr
*instr
);
1300 /* is this source channel used? */
1302 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
1305 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1306 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
1308 return (instr
->dest
.write_mask
>> channel
) & 1;
1311 static inline nir_component_mask_t
1312 nir_alu_instr_src_read_mask(const nir_alu_instr
*instr
, unsigned src
)
1314 nir_component_mask_t read_mask
= 0;
1315 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; c
++) {
1316 if (!nir_alu_instr_channel_used(instr
, src
, c
))
1319 read_mask
|= (1 << instr
->src
[src
].swizzle
[c
]);
1325 * Get the number of channels used for a source
1327 static inline unsigned
1328 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
1330 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1331 return nir_op_infos
[instr
->op
].input_sizes
[src
];
1333 return nir_dest_num_components(instr
->dest
.dest
);
1337 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1339 switch (instr
->op
) {
1359 bool nir_const_value_negative_equal(nir_const_value c1
, nir_const_value c2
,
1360 nir_alu_type full_type
);
1362 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
1363 unsigned src1
, unsigned src2
);
1365 bool nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
1366 const nir_alu_instr
*alu2
,
1367 unsigned src1
, unsigned src2
);
1371 nir_deref_type_array
,
1372 nir_deref_type_array_wildcard
,
1373 nir_deref_type_ptr_as_array
,
1374 nir_deref_type_struct
,
1375 nir_deref_type_cast
,
1381 /** The type of this deref instruction */
1382 nir_deref_type deref_type
;
1384 /** The mode of the underlying variable */
1385 nir_variable_mode mode
;
1387 /** The dereferenced type of the resulting pointer value */
1388 const struct glsl_type
*type
;
1391 /** Variable being dereferenced if deref_type is a deref_var */
1394 /** Parent deref if deref_type is not deref_var */
1398 /** Additional deref parameters */
1409 unsigned ptr_stride
;
1413 /** Destination to store the resulting "pointer" */
1417 static inline nir_deref_instr
*nir_src_as_deref(nir_src src
);
1419 static inline nir_deref_instr
*
1420 nir_deref_instr_parent(const nir_deref_instr
*instr
)
1422 if (instr
->deref_type
== nir_deref_type_var
)
1425 return nir_src_as_deref(instr
->parent
);
1428 static inline nir_variable
*
1429 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1431 while (instr
->deref_type
!= nir_deref_type_var
) {
1432 if (instr
->deref_type
== nir_deref_type_cast
)
1435 instr
= nir_deref_instr_parent(instr
);
1441 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1442 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr
*instr
);
1443 bool nir_deref_instr_has_complex_use(nir_deref_instr
*instr
);
1445 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1447 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*instr
);
1452 struct nir_function
*callee
;
1454 unsigned num_params
;
1458 #include "nir_intrinsics.h"
1460 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1462 /** Represents an intrinsic
1464 * An intrinsic is an instruction type for handling things that are
1465 * more-or-less regular operations but don't just consume and produce SSA
1466 * values like ALU operations do. Intrinsics are not for things that have
1467 * special semantic meaning such as phi nodes and parallel copies.
1468 * Examples of intrinsics include variable load/store operations, system
1469 * value loads, and the like. Even though texturing more-or-less falls
1470 * under this category, texturing is its own instruction type because
1471 * trying to represent texturing with intrinsics would lead to a
1472 * combinatorial explosion of intrinsic opcodes.
1474 * By having a single instruction type for handling a lot of different
1475 * cases, optimization passes can look for intrinsics and, for the most
1476 * part, completely ignore them. Each intrinsic type also has a few
1477 * possible flags that govern whether or not they can be reordered or
1478 * eliminated. That way passes like dead code elimination can still work
1479 * on intrisics without understanding the meaning of each.
1481 * Each intrinsic has some number of constant indices, some number of
1482 * variables, and some number of sources. What these sources, variables,
1483 * and indices mean depends on the intrinsic and is documented with the
1484 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1485 * instructions are the only types of instruction that can operate on
1491 nir_intrinsic_op intrinsic
;
1495 /** number of components if this is a vectorized intrinsic
1497 * Similarly to ALU operations, some intrinsics are vectorized.
1498 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1499 * For vectorized intrinsics, the num_components field specifies the
1500 * number of destination components and the number of source components
1501 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1503 uint8_t num_components
;
1505 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1508 } nir_intrinsic_instr
;
1510 static inline nir_variable
*
1511 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1513 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1517 /* Memory ordering. */
1518 NIR_MEMORY_ACQUIRE
= 1 << 0,
1519 NIR_MEMORY_RELEASE
= 1 << 1,
1520 NIR_MEMORY_ACQ_REL
= NIR_MEMORY_ACQUIRE
| NIR_MEMORY_RELEASE
,
1522 /* Memory visibility operations. */
1523 NIR_MEMORY_MAKE_AVAILABLE
= 1 << 2,
1524 NIR_MEMORY_MAKE_VISIBLE
= 1 << 3,
1525 } nir_memory_semantics
;
1529 NIR_SCOPE_INVOCATION
,
1531 NIR_SCOPE_WORKGROUP
,
1532 NIR_SCOPE_QUEUE_FAMILY
,
1537 * \name NIR intrinsics semantic flags
1539 * information about what the compiler can do with the intrinsics.
1541 * \sa nir_intrinsic_info::flags
1545 * whether the intrinsic can be safely eliminated if none of its output
1546 * value is not being used.
1548 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1551 * Whether the intrinsic can be reordered with respect to any other
1552 * intrinsic, i.e. whether the only reordering dependencies of the
1553 * intrinsic are due to the register reads/writes.
1555 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1556 } nir_intrinsic_semantic_flag
;
1559 * \name NIR intrinsics const-index flag
1561 * Indicates the usage of a const_index slot.
1563 * \sa nir_intrinsic_info::index_map
1567 * Generally instructions that take a offset src argument, can encode
1568 * a constant 'base' value which is added to the offset.
1570 NIR_INTRINSIC_BASE
= 1,
1573 * For store instructions, a writemask for the store.
1575 NIR_INTRINSIC_WRMASK
,
1578 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1580 NIR_INTRINSIC_STREAM_ID
,
1583 * The clip-plane id for load_user_clip_plane intrinsic.
1585 NIR_INTRINSIC_UCP_ID
,
1588 * The amount of data, starting from BASE, that this instruction may
1589 * access. This is used to provide bounds if the offset is not constant.
1591 NIR_INTRINSIC_RANGE
,
1594 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1596 NIR_INTRINSIC_DESC_SET
,
1599 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1601 NIR_INTRINSIC_BINDING
,
1606 NIR_INTRINSIC_COMPONENT
,
1609 * Interpolation mode (only meaningful for FS inputs).
1611 NIR_INTRINSIC_INTERP_MODE
,
1614 * A binary nir_op to use when performing a reduction or scan operation
1616 NIR_INTRINSIC_REDUCTION_OP
,
1619 * Cluster size for reduction operations
1621 NIR_INTRINSIC_CLUSTER_SIZE
,
1624 * Parameter index for a load_param intrinsic
1626 NIR_INTRINSIC_PARAM_IDX
,
1629 * Image dimensionality for image intrinsics
1631 * One of GLSL_SAMPLER_DIM_*
1633 NIR_INTRINSIC_IMAGE_DIM
,
1636 * Non-zero if we are accessing an array image
1638 NIR_INTRINSIC_IMAGE_ARRAY
,
1641 * Image format for image intrinsics
1643 NIR_INTRINSIC_FORMAT
,
1646 * Access qualifiers for image and memory access intrinsics
1648 NIR_INTRINSIC_ACCESS
,
1651 * Alignment for offsets and addresses
1653 * These two parameters, specify an alignment in terms of a multiplier and
1654 * an offset. The offset or address parameter X of the intrinsic is
1655 * guaranteed to satisfy the following:
1657 * (X - align_offset) % align_mul == 0
1659 NIR_INTRINSIC_ALIGN_MUL
,
1660 NIR_INTRINSIC_ALIGN_OFFSET
,
1663 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1665 NIR_INTRINSIC_DESC_TYPE
,
1668 * The nir_alu_type of a uniform/input/output
1673 * The swizzle mask for the instructions
1674 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1676 NIR_INTRINSIC_SWIZZLE_MASK
,
1678 /* Separate source/dest access flags for copies */
1679 NIR_INTRINSIC_SRC_ACCESS
,
1680 NIR_INTRINSIC_DST_ACCESS
,
1682 /* Driver location for nir_load_patch_location_ir3 */
1683 NIR_INTRINSIC_DRIVER_LOCATION
,
1686 * Mask of nir_memory_semantics, includes ordering and visibility.
1688 NIR_INTRINSIC_MEMORY_SEMANTICS
,
1691 * Mask of nir_variable_modes affected by the memory operation.
1693 NIR_INTRINSIC_MEMORY_MODES
,
1696 * Value of nir_scope.
1698 NIR_INTRINSIC_MEMORY_SCOPE
,
1701 * Value of nir_scope.
1703 NIR_INTRINSIC_EXECUTION_SCOPE
,
1705 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1707 } nir_intrinsic_index_flag
;
1709 #define NIR_INTRINSIC_MAX_INPUTS 5
1714 uint8_t num_srcs
; /** < number of register/SSA inputs */
1716 /** number of components of each input register
1718 * If this value is 0, the number of components is given by the
1719 * num_components field of nir_intrinsic_instr. If this value is -1, the
1720 * intrinsic consumes however many components are provided and it is not
1723 int8_t src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1727 /** number of components of the output register
1729 * If this value is 0, the number of components is given by the
1730 * num_components field of nir_intrinsic_instr.
1732 uint8_t dest_components
;
1734 /** bitfield of legal bit sizes */
1735 uint8_t dest_bit_sizes
;
1737 /** the number of constant indices used by the intrinsic */
1738 uint8_t num_indices
;
1740 /** indicates the usage of intr->const_index[n] */
1741 uint8_t index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1743 /** semantic flags for calls to this intrinsic */
1744 nir_intrinsic_semantic_flag flags
;
1745 } nir_intrinsic_info
;
1747 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1749 static inline unsigned
1750 nir_intrinsic_src_components(const nir_intrinsic_instr
*intr
, unsigned srcn
)
1752 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1753 assert(srcn
< info
->num_srcs
);
1754 if (info
->src_components
[srcn
] > 0)
1755 return info
->src_components
[srcn
];
1756 else if (info
->src_components
[srcn
] == 0)
1757 return intr
->num_components
;
1759 return nir_src_num_components(intr
->src
[srcn
]);
1762 static inline unsigned
1763 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1765 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1766 if (!info
->has_dest
)
1768 else if (info
->dest_components
)
1769 return info
->dest_components
;
1771 return intr
->num_components
;
1775 * Helper to copy const_index[] from src to dst, without assuming they
1779 nir_intrinsic_copy_const_indices(nir_intrinsic_instr
*dst
, nir_intrinsic_instr
*src
)
1781 if (src
->intrinsic
== dst
->intrinsic
) {
1782 memcpy(dst
->const_index
, src
->const_index
, sizeof(dst
->const_index
));
1786 const nir_intrinsic_info
*src_info
= &nir_intrinsic_infos
[src
->intrinsic
];
1787 const nir_intrinsic_info
*dst_info
= &nir_intrinsic_infos
[dst
->intrinsic
];
1789 for (unsigned i
= 0; i
< NIR_INTRINSIC_NUM_INDEX_FLAGS
; i
++) {
1790 if (src_info
->index_map
[i
] == 0)
1793 /* require that dst instruction also uses the same const_index[]: */
1794 assert(dst_info
->index_map
[i
] > 0);
1796 dst
->const_index
[dst_info
->index_map
[i
] - 1] =
1797 src
->const_index
[src_info
->index_map
[i
] - 1];
1801 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1802 static inline type \
1803 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1805 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1806 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1807 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1809 static inline void \
1810 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1812 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1813 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1814 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1817 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1818 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1819 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1820 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1821 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1822 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1823 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1824 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1825 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1826 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1827 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1828 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1829 INTRINSIC_IDX_ACCESSORS(image_dim
, IMAGE_DIM
, enum glsl_sampler_dim
)
1830 INTRINSIC_IDX_ACCESSORS(image_array
, IMAGE_ARRAY
, bool)
1831 INTRINSIC_IDX_ACCESSORS(access
, ACCESS
, enum gl_access_qualifier
)
1832 INTRINSIC_IDX_ACCESSORS(src_access
, SRC_ACCESS
, enum gl_access_qualifier
)
1833 INTRINSIC_IDX_ACCESSORS(dst_access
, DST_ACCESS
, enum gl_access_qualifier
)
1834 INTRINSIC_IDX_ACCESSORS(format
, FORMAT
, enum pipe_format
)
1835 INTRINSIC_IDX_ACCESSORS(align_mul
, ALIGN_MUL
, unsigned)
1836 INTRINSIC_IDX_ACCESSORS(align_offset
, ALIGN_OFFSET
, unsigned)
1837 INTRINSIC_IDX_ACCESSORS(desc_type
, DESC_TYPE
, unsigned)
1838 INTRINSIC_IDX_ACCESSORS(type
, TYPE
, nir_alu_type
)
1839 INTRINSIC_IDX_ACCESSORS(swizzle_mask
, SWIZZLE_MASK
, unsigned)
1840 INTRINSIC_IDX_ACCESSORS(driver_location
, DRIVER_LOCATION
, unsigned)
1841 INTRINSIC_IDX_ACCESSORS(memory_semantics
, MEMORY_SEMANTICS
, nir_memory_semantics
)
1842 INTRINSIC_IDX_ACCESSORS(memory_modes
, MEMORY_MODES
, nir_variable_mode
)
1843 INTRINSIC_IDX_ACCESSORS(memory_scope
, MEMORY_SCOPE
, nir_scope
)
1844 INTRINSIC_IDX_ACCESSORS(execution_scope
, EXECUTION_SCOPE
, nir_scope
)
1847 nir_intrinsic_set_align(nir_intrinsic_instr
*intrin
,
1848 unsigned align_mul
, unsigned align_offset
)
1850 assert(util_is_power_of_two_nonzero(align_mul
));
1851 assert(align_offset
< align_mul
);
1852 nir_intrinsic_set_align_mul(intrin
, align_mul
);
1853 nir_intrinsic_set_align_offset(intrin
, align_offset
);
1856 /** Returns a simple alignment for a load/store intrinsic offset
1858 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1859 * and ALIGN_OFFSET parameters, this helper takes both into account and
1860 * provides a single simple alignment parameter. The offset X is guaranteed
1861 * to satisfy X % align == 0.
1863 static inline unsigned
1864 nir_intrinsic_align(const nir_intrinsic_instr
*intrin
)
1866 const unsigned align_mul
= nir_intrinsic_align_mul(intrin
);
1867 const unsigned align_offset
= nir_intrinsic_align_offset(intrin
);
1868 assert(align_offset
< align_mul
);
1869 return align_offset
? 1 << (ffs(align_offset
) - 1) : align_mul
;
1873 nir_image_intrinsic_coord_components(const nir_intrinsic_instr
*instr
);
1875 /* Converts a image_deref_* intrinsic into a image_* one */
1876 void nir_rewrite_image_intrinsic(nir_intrinsic_instr
*instr
,
1877 nir_ssa_def
*handle
, bool bindless
);
1879 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1881 nir_intrinsic_can_reorder(nir_intrinsic_instr
*instr
)
1883 if (instr
->intrinsic
== nir_intrinsic_load_deref
||
1884 instr
->intrinsic
== nir_intrinsic_load_ssbo
||
1885 instr
->intrinsic
== nir_intrinsic_bindless_image_load
||
1886 instr
->intrinsic
== nir_intrinsic_image_deref_load
||
1887 instr
->intrinsic
== nir_intrinsic_image_load
) {
1888 return nir_intrinsic_access(instr
) & ACCESS_CAN_REORDER
;
1890 const nir_intrinsic_info
*info
=
1891 &nir_intrinsic_infos
[instr
->intrinsic
];
1892 return (info
->flags
& NIR_INTRINSIC_CAN_ELIMINATE
) &&
1893 (info
->flags
& NIR_INTRINSIC_CAN_REORDER
);
1898 * \group texture information
1900 * This gives semantic information about textures which is useful to the
1901 * frontend, the backend, and lowering passes, but not the optimizer.
1906 nir_tex_src_projector
,
1907 nir_tex_src_comparator
, /* shadow comparator */
1911 nir_tex_src_min_lod
,
1912 nir_tex_src_ms_index
, /* MSAA sample index */
1913 nir_tex_src_ms_mcs
, /* MSAA compression value */
1916 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
1917 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
1918 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
1919 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
1920 nir_tex_src_texture_handle
, /* < bindless texture handle */
1921 nir_tex_src_sampler_handle
, /* < bindless sampler handle */
1922 nir_tex_src_plane
, /* < selects plane for planar textures */
1923 nir_num_tex_src_types
1928 nir_tex_src_type src_type
;
1932 nir_texop_tex
, /**< Regular texture look-up */
1933 nir_texop_txb
, /**< Texture look-up with LOD bias */
1934 nir_texop_txl
, /**< Texture look-up with explicit LOD */
1935 nir_texop_txd
, /**< Texture look-up with partial derivatives */
1936 nir_texop_txf
, /**< Texel fetch with explicit LOD */
1937 nir_texop_txf_ms
, /**< Multisample texture fetch */
1938 nir_texop_txf_ms_fb
, /**< Multisample texture fetch from framebuffer */
1939 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
1940 nir_texop_txs
, /**< Texture size */
1941 nir_texop_lod
, /**< Texture lod query */
1942 nir_texop_tg4
, /**< Texture gather */
1943 nir_texop_query_levels
, /**< Texture levels query */
1944 nir_texop_texture_samples
, /**< Texture samples query */
1945 nir_texop_samples_identical
, /**< Query whether all samples are definitely
1948 nir_texop_tex_prefetch
, /**< Regular texture look-up, eligible for pre-dispatch */
1949 nir_texop_fragment_fetch
, /**< Multisample fragment color texture fetch */
1950 nir_texop_fragment_mask_fetch
,/**< Multisample fragment mask texture fetch */
1956 enum glsl_sampler_dim sampler_dim
;
1957 nir_alu_type dest_type
;
1962 unsigned num_srcs
, coord_components
;
1963 bool is_array
, is_shadow
;
1966 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1967 * components or the new-style shadow that outputs 1 component.
1969 bool is_new_style_shadow
;
1971 /* gather component selector */
1972 unsigned component
: 2;
1974 /* gather offsets */
1975 int8_t tg4_offsets
[4][2];
1977 /* True if the texture index or handle is not dynamically uniform */
1978 bool texture_non_uniform
;
1980 /* True if the sampler index or handle is not dynamically uniform */
1981 bool sampler_non_uniform
;
1983 /** The texture index
1985 * If this texture instruction has a nir_tex_src_texture_offset source,
1986 * then the texture index is given by texture_index + texture_offset.
1988 unsigned texture_index
;
1990 /** The sampler index
1992 * The following operations do not require a sampler and, as such, this
1993 * field should be ignored:
1995 * - nir_texop_txf_ms
1998 * - nir_texop_query_levels
1999 * - nir_texop_texture_samples
2000 * - nir_texop_samples_identical
2002 * If this texture instruction has a nir_tex_src_sampler_offset source,
2003 * then the sampler index is given by sampler_index + sampler_offset.
2005 unsigned sampler_index
;
2009 * Returns true if the texture operation requires a sampler as a general rule,
2010 * see the documentation of sampler_index.
2012 * Note that the specific hw/driver backend could require to a sampler
2013 * object/configuration packet in any case, for some other reason.
2016 nir_tex_instr_need_sampler(const nir_tex_instr
*instr
)
2018 switch (instr
->op
) {
2020 case nir_texop_txf_ms
:
2023 case nir_texop_query_levels
:
2024 case nir_texop_texture_samples
:
2025 case nir_texop_samples_identical
:
2032 static inline unsigned
2033 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
2035 switch (instr
->op
) {
2036 case nir_texop_txs
: {
2038 switch (instr
->sampler_dim
) {
2039 case GLSL_SAMPLER_DIM_1D
:
2040 case GLSL_SAMPLER_DIM_BUF
:
2043 case GLSL_SAMPLER_DIM_2D
:
2044 case GLSL_SAMPLER_DIM_CUBE
:
2045 case GLSL_SAMPLER_DIM_MS
:
2046 case GLSL_SAMPLER_DIM_RECT
:
2047 case GLSL_SAMPLER_DIM_EXTERNAL
:
2048 case GLSL_SAMPLER_DIM_SUBPASS
:
2051 case GLSL_SAMPLER_DIM_3D
:
2055 unreachable("not reached");
2057 if (instr
->is_array
)
2065 case nir_texop_texture_samples
:
2066 case nir_texop_query_levels
:
2067 case nir_texop_samples_identical
:
2068 case nir_texop_fragment_mask_fetch
:
2072 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
2079 /* Returns true if this texture operation queries something about the texture
2080 * rather than actually sampling it.
2083 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
2085 switch (instr
->op
) {
2088 case nir_texop_texture_samples
:
2089 case nir_texop_query_levels
:
2090 case nir_texop_txf_ms_mcs
:
2097 case nir_texop_txf_ms
:
2098 case nir_texop_txf_ms_fb
:
2102 unreachable("Invalid texture opcode");
2107 nir_tex_instr_has_implicit_derivative(const nir_tex_instr
*instr
)
2109 switch (instr
->op
) {
2119 static inline nir_alu_type
2120 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
2122 switch (instr
->src
[src
].src_type
) {
2123 case nir_tex_src_coord
:
2124 switch (instr
->op
) {
2126 case nir_texop_txf_ms
:
2127 case nir_texop_txf_ms_fb
:
2128 case nir_texop_txf_ms_mcs
:
2129 case nir_texop_samples_identical
:
2130 return nir_type_int
;
2133 return nir_type_float
;
2136 case nir_tex_src_lod
:
2137 switch (instr
->op
) {
2140 return nir_type_int
;
2143 return nir_type_float
;
2146 case nir_tex_src_projector
:
2147 case nir_tex_src_comparator
:
2148 case nir_tex_src_bias
:
2149 case nir_tex_src_min_lod
:
2150 case nir_tex_src_ddx
:
2151 case nir_tex_src_ddy
:
2152 return nir_type_float
;
2154 case nir_tex_src_offset
:
2155 case nir_tex_src_ms_index
:
2156 case nir_tex_src_plane
:
2157 return nir_type_int
;
2159 case nir_tex_src_ms_mcs
:
2160 case nir_tex_src_texture_deref
:
2161 case nir_tex_src_sampler_deref
:
2162 case nir_tex_src_texture_offset
:
2163 case nir_tex_src_sampler_offset
:
2164 case nir_tex_src_texture_handle
:
2165 case nir_tex_src_sampler_handle
:
2166 return nir_type_uint
;
2168 case nir_num_tex_src_types
:
2169 unreachable("nir_num_tex_src_types is not a valid source type");
2172 unreachable("Invalid texture source type");
2175 static inline unsigned
2176 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
2178 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
2179 return instr
->coord_components
;
2181 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
2182 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
2185 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
2186 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
2187 if (instr
->is_array
)
2188 return instr
->coord_components
- 1;
2190 return instr
->coord_components
;
2193 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
2194 * the offset, since a cube maps to a single face.
2196 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
2197 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2199 else if (instr
->is_array
)
2200 return instr
->coord_components
- 1;
2202 return instr
->coord_components
;
2209 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
2211 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
2212 if (instr
->src
[i
].src_type
== type
)
2218 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
2219 nir_tex_src_type src_type
,
2222 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
2224 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
);
2231 nir_const_value value
[];
2232 } nir_load_const_instr
;
2235 /** Return from a function
2237 * This instruction is a classic function return. It jumps to
2238 * nir_function_impl::end_block. No return value is provided in this
2239 * instruction. Instead, the function is expected to write any return
2240 * data to a deref passed in from the caller.
2244 /** Break out of the inner-most loop
2246 * This has the same semantics as C's "break" statement.
2250 /** Jump back to the top of the inner-most loop
2252 * This has the same semantics as C's "continue" statement assuming that a
2253 * NIR loop is implemented as "while (1) { body }".
2263 /* creates a new SSA variable in an undefined state */
2268 } nir_ssa_undef_instr
;
2271 struct exec_node node
;
2273 /* The predecessor block corresponding to this source */
2274 struct nir_block
*pred
;
2279 #define nir_foreach_phi_src(phi_src, phi) \
2280 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
2281 #define nir_foreach_phi_src_safe(phi_src, phi) \
2282 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
2287 struct exec_list srcs
; /** < list of nir_phi_src */
2293 struct exec_node node
;
2296 } nir_parallel_copy_entry
;
2298 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
2299 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
2304 /* A list of nir_parallel_copy_entrys. The sources of all of the
2305 * entries are copied to the corresponding destinations "in parallel".
2306 * In other words, if we have two entries: a -> b and b -> a, the values
2309 struct exec_list entries
;
2310 } nir_parallel_copy_instr
;
2312 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
2313 type
, nir_instr_type_alu
)
2314 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
2315 type
, nir_instr_type_deref
)
2316 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
2317 type
, nir_instr_type_call
)
2318 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
2319 type
, nir_instr_type_jump
)
2320 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
2321 type
, nir_instr_type_tex
)
2322 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
2323 type
, nir_instr_type_intrinsic
)
2324 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
2325 type
, nir_instr_type_load_const
)
2326 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
2327 type
, nir_instr_type_ssa_undef
)
2328 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
2329 type
, nir_instr_type_phi
)
2330 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
2331 nir_parallel_copy_instr
, instr
,
2332 type
, nir_instr_type_parallel_copy
)
2335 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
2336 static inline type \
2337 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2339 assert(nir_src_is_const(src)); \
2340 nir_load_const_instr *load = \
2341 nir_instr_as_load_const(src.ssa->parent_instr); \
2342 assert(comp < load->def.num_components); \
2343 return nir_const_value_as_##suffix(load->value[comp], \
2344 load->def.bit_size); \
2347 static inline type \
2348 nir_src_as_##suffix(nir_src src) \
2350 assert(nir_src_num_components(src) == 1); \
2351 return nir_src_comp_as_##suffix(src, 0); \
2354 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2355 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint
)
2356 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2357 NIR_DEFINE_SRC_AS_CONST(double, float)
2359 #undef NIR_DEFINE_SRC_AS_CONST
2368 nir_ssa_scalar_is_const(nir_ssa_scalar s
)
2370 return s
.def
->parent_instr
->type
== nir_instr_type_load_const
;
2373 static inline nir_const_value
2374 nir_ssa_scalar_as_const_value(nir_ssa_scalar s
)
2376 assert(s
.comp
< s
.def
->num_components
);
2377 nir_load_const_instr
*load
= nir_instr_as_load_const(s
.def
->parent_instr
);
2378 return load
->value
[s
.comp
];
2381 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2382 static inline type \
2383 nir_ssa_scalar_as_##suffix(nir_ssa_scalar s) \
2385 return nir_const_value_as_##suffix( \
2386 nir_ssa_scalar_as_const_value(s), s.def->bit_size); \
2389 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2390 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint
)
2391 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2392 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2394 #undef NIR_DEFINE_SCALAR_AS_CONST
2397 nir_ssa_scalar_is_alu(nir_ssa_scalar s
)
2399 return s
.def
->parent_instr
->type
== nir_instr_type_alu
;
2402 static inline nir_op
2403 nir_ssa_scalar_alu_op(nir_ssa_scalar s
)
2405 return nir_instr_as_alu(s
.def
->parent_instr
)->op
;
2408 static inline nir_ssa_scalar
2409 nir_ssa_scalar_chase_alu_src(nir_ssa_scalar s
, unsigned alu_src_idx
)
2411 nir_ssa_scalar out
= { NULL
, 0 };
2413 nir_alu_instr
*alu
= nir_instr_as_alu(s
.def
->parent_instr
);
2414 assert(alu_src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
2416 /* Our component must be written */
2417 assert(s
.comp
< s
.def
->num_components
);
2418 assert(alu
->dest
.write_mask
& (1u << s
.comp
));
2420 assert(alu
->src
[alu_src_idx
].src
.is_ssa
);
2421 out
.def
= alu
->src
[alu_src_idx
].src
.ssa
;
2423 if (nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 0) {
2424 /* The ALU src is unsized so the source component follows the
2425 * destination component.
2427 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[s
.comp
];
2429 /* This is a sized source so all source components work together to
2430 * produce all the destination components. Since we need to return a
2431 * scalar, this only works if the source is a scalar.
2433 assert(nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 1);
2434 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[0];
2436 assert(out
.comp
< out
.def
->num_components
);
2445 * Control flow consists of a tree of control flow nodes, which include
2446 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2447 * instructions that always run start-to-finish. Each basic block also keeps
2448 * track of its successors (blocks which may run immediately after the current
2449 * block) and predecessors (blocks which could have run immediately before the
2450 * current block). Each function also has a start block and an end block which
2451 * all return statements point to (which is always empty). Together, all the
2452 * blocks with their predecessors and successors make up the control flow
2453 * graph (CFG) of the function. There are helpers that modify the tree of
2454 * control flow nodes while modifying the CFG appropriately; these should be
2455 * used instead of modifying the tree directly.
2462 nir_cf_node_function
2465 typedef struct nir_cf_node
{
2466 struct exec_node node
;
2467 nir_cf_node_type type
;
2468 struct nir_cf_node
*parent
;
2471 typedef struct nir_block
{
2472 nir_cf_node cf_node
;
2474 struct exec_list instr_list
; /** < list of nir_instr */
2476 /** generic block index; generated by nir_index_blocks */
2480 * Each block can only have up to 2 successors, so we put them in a simple
2481 * array - no need for anything more complicated.
2483 struct nir_block
*successors
[2];
2485 /* Set of nir_block predecessors in the CFG */
2486 struct set
*predecessors
;
2489 * this node's immediate dominator in the dominance tree - set to NULL for
2492 struct nir_block
*imm_dom
;
2494 /* This node's children in the dominance tree */
2495 unsigned num_dom_children
;
2496 struct nir_block
**dom_children
;
2498 /* Set of nir_blocks on the dominance frontier of this block */
2499 struct set
*dom_frontier
;
2502 * These two indices have the property that dom_{pre,post}_index for each
2503 * child of this block in the dominance tree will always be between
2504 * dom_pre_index and dom_post_index for this block, which makes testing if
2505 * a given block is dominated by another block an O(1) operation.
2507 int16_t dom_pre_index
, dom_post_index
;
2509 /* live in and out for this block; used for liveness analysis */
2510 BITSET_WORD
*live_in
;
2511 BITSET_WORD
*live_out
;
2515 nir_block_is_reachable(nir_block
*b
)
2517 /* See also nir_block_dominates */
2518 return b
->dom_post_index
!= -1;
2521 static inline nir_instr
*
2522 nir_block_first_instr(nir_block
*block
)
2524 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
2525 return exec_node_data(nir_instr
, head
, node
);
2528 static inline nir_instr
*
2529 nir_block_last_instr(nir_block
*block
)
2531 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
2532 return exec_node_data(nir_instr
, tail
, node
);
2536 nir_block_ends_in_jump(nir_block
*block
)
2538 return !exec_list_is_empty(&block
->instr_list
) &&
2539 nir_block_last_instr(block
)->type
== nir_instr_type_jump
;
2542 #define nir_foreach_instr(instr, block) \
2543 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
2544 #define nir_foreach_instr_reverse(instr, block) \
2545 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
2546 #define nir_foreach_instr_safe(instr, block) \
2547 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
2548 #define nir_foreach_instr_reverse_safe(instr, block) \
2549 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
2552 nir_selection_control_none
= 0x0,
2553 nir_selection_control_flatten
= 0x1,
2554 nir_selection_control_dont_flatten
= 0x2,
2555 } nir_selection_control
;
2557 typedef struct nir_if
{
2558 nir_cf_node cf_node
;
2560 nir_selection_control control
;
2562 struct exec_list then_list
; /** < list of nir_cf_node */
2563 struct exec_list else_list
; /** < list of nir_cf_node */
2569 /** Instruction that generates nif::condition. */
2570 nir_instr
*conditional_instr
;
2572 /** Block within ::nif that has the break instruction. */
2573 nir_block
*break_block
;
2575 /** Last block for the then- or else-path that does not contain the break. */
2576 nir_block
*continue_from_block
;
2578 /** True when ::break_block is in the else-path of ::nif. */
2579 bool continue_from_then
;
2582 /* This is true if the terminators exact trip count is unknown. For
2585 * for (int i = 0; i < imin(x, 4); i++)
2588 * Here loop analysis would have set a max_trip_count of 4 however we dont
2589 * know for sure that this is the exact trip count.
2591 bool exact_trip_count_unknown
;
2593 struct list_head loop_terminator_link
;
2594 } nir_loop_terminator
;
2597 /* Estimated cost (in number of instructions) of the loop */
2598 unsigned instr_cost
;
2600 /* Guessed trip count based on array indexing */
2601 unsigned guessed_trip_count
;
2603 /* Maximum number of times the loop is run (if known) */
2604 unsigned max_trip_count
;
2606 /* Do we know the exact number of times the loop will be run */
2607 bool exact_trip_count_known
;
2609 /* Unroll the loop regardless of its size */
2612 /* Does the loop contain complex loop terminators, continues or other
2613 * complex behaviours? If this is true we can't rely on
2614 * loop_terminator_list to be complete or accurate.
2618 nir_loop_terminator
*limiting_terminator
;
2620 /* A list of loop_terminators terminating this loop. */
2621 struct list_head loop_terminator_list
;
2625 nir_loop_control_none
= 0x0,
2626 nir_loop_control_unroll
= 0x1,
2627 nir_loop_control_dont_unroll
= 0x2,
2631 nir_cf_node cf_node
;
2633 struct exec_list body
; /** < list of nir_cf_node */
2635 nir_loop_info
*info
;
2636 nir_loop_control control
;
2637 bool partially_unrolled
;
2641 * Various bits of metadata that can may be created or required by
2642 * optimization and analysis passes
2645 nir_metadata_none
= 0x0,
2647 /** Indicates that nir_block::index values are valid.
2649 * The start block has index 0 and they increase through a natural walk of
2650 * the CFG. nir_function_impl::num_blocks is the number of blocks and
2651 * every block index is in the range [0, nir_function_impl::num_blocks].
2653 * A pass can preserve this metadata type if it doesn't touch the CFG.
2655 nir_metadata_block_index
= 0x1,
2657 /** Indicates that block dominance information is valid
2661 * - nir_block::num_dom_children
2662 * - nir_block::dom_children
2663 * - nir_block::dom_frontier
2664 * - nir_block::dom_pre_index
2665 * - nir_block::dom_post_index
2667 * A pass can preserve this metadata type if it doesn't touch the CFG.
2669 nir_metadata_dominance
= 0x2,
2671 /** Indicates that SSA def data-flow liveness information is valid
2675 * - nir_ssa_def::live_index
2676 * - nir_block::live_in
2677 * - nir_block::live_out
2679 * A pass can preserve this metadata type if it never adds or removes any
2680 * SSA defs (most passes shouldn't preserve this metadata type).
2682 nir_metadata_live_ssa_defs
= 0x4,
2684 /** A dummy metadata value to track when a pass forgot to call
2685 * nir_metadata_preserve.
2687 * A pass should always clear this value even if it doesn't make any
2688 * progress to indicate that it thought about preserving metadata.
2690 nir_metadata_not_properly_reset
= 0x8,
2692 /** Indicates that loop analysis information is valid.
2694 * This includes everything pointed to by nir_loop::info.
2696 * A pass can preserve this metadata type if it is guaranteed to not affect
2697 * any loop metadata. However, since loop metadata includes things like
2698 * loop counts which depend on arithmetic in the loop, this is very hard to
2699 * determine. Most passes shouldn't preserve this metadata type.
2701 nir_metadata_loop_analysis
= 0x10,
2705 nir_cf_node cf_node
;
2707 /** pointer to the function of which this is an implementation */
2708 struct nir_function
*function
;
2710 struct exec_list body
; /** < list of nir_cf_node */
2712 nir_block
*end_block
;
2714 /** list for all local variables in the function */
2715 struct exec_list locals
;
2717 /** list of local registers in the function */
2718 struct exec_list registers
;
2720 /** next available local register index */
2723 /** next available SSA value index */
2726 /* total number of basic blocks, only valid when block_index_dirty = false */
2727 unsigned num_blocks
;
2729 nir_metadata valid_metadata
;
2730 } nir_function_impl
;
2732 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2733 nir_start_block(nir_function_impl
*impl
)
2735 return (nir_block
*) impl
->body
.head_sentinel
.next
;
2738 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2739 nir_impl_last_block(nir_function_impl
*impl
)
2741 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
2744 static inline nir_cf_node
*
2745 nir_cf_node_next(nir_cf_node
*node
)
2747 struct exec_node
*next
= exec_node_get_next(&node
->node
);
2748 if (exec_node_is_tail_sentinel(next
))
2751 return exec_node_data(nir_cf_node
, next
, node
);
2754 static inline nir_cf_node
*
2755 nir_cf_node_prev(nir_cf_node
*node
)
2757 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
2758 if (exec_node_is_head_sentinel(prev
))
2761 return exec_node_data(nir_cf_node
, prev
, node
);
2765 nir_cf_node_is_first(const nir_cf_node
*node
)
2767 return exec_node_is_head_sentinel(node
->node
.prev
);
2771 nir_cf_node_is_last(const nir_cf_node
*node
)
2773 return exec_node_is_tail_sentinel(node
->node
.next
);
2776 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
2777 type
, nir_cf_node_block
)
2778 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
2779 type
, nir_cf_node_if
)
2780 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
2781 type
, nir_cf_node_loop
)
2782 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
2783 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
2785 static inline nir_block
*
2786 nir_if_first_then_block(nir_if
*if_stmt
)
2788 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
2789 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2792 static inline nir_block
*
2793 nir_if_last_then_block(nir_if
*if_stmt
)
2795 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
2796 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2799 static inline nir_block
*
2800 nir_if_first_else_block(nir_if
*if_stmt
)
2802 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
2803 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2806 static inline nir_block
*
2807 nir_if_last_else_block(nir_if
*if_stmt
)
2809 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
2810 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2813 static inline nir_block
*
2814 nir_loop_first_block(nir_loop
*loop
)
2816 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
2817 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2820 static inline nir_block
*
2821 nir_loop_last_block(nir_loop
*loop
)
2823 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
2824 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2828 * Return true if this list of cf_nodes contains a single empty block.
2831 nir_cf_list_is_empty_block(struct exec_list
*cf_list
)
2833 if (exec_list_is_singular(cf_list
)) {
2834 struct exec_node
*head
= exec_list_get_head(cf_list
);
2836 nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2837 return exec_list_is_empty(&block
->instr_list
);
2843 uint8_t num_components
;
2847 typedef struct nir_function
{
2848 struct exec_node node
;
2851 struct nir_shader
*shader
;
2853 unsigned num_params
;
2854 nir_parameter
*params
;
2856 /** The implementation of this function.
2858 * If the function is only declared and not implemented, this is NULL.
2860 nir_function_impl
*impl
;
2866 nir_lower_imul64
= (1 << 0),
2867 nir_lower_isign64
= (1 << 1),
2868 /** Lower all int64 modulus and division opcodes */
2869 nir_lower_divmod64
= (1 << 2),
2870 /** Lower all 64-bit umul_high and imul_high opcodes */
2871 nir_lower_imul_high64
= (1 << 3),
2872 nir_lower_mov64
= (1 << 4),
2873 nir_lower_icmp64
= (1 << 5),
2874 nir_lower_iadd64
= (1 << 6),
2875 nir_lower_iabs64
= (1 << 7),
2876 nir_lower_ineg64
= (1 << 8),
2877 nir_lower_logic64
= (1 << 9),
2878 nir_lower_minmax64
= (1 << 10),
2879 nir_lower_shift64
= (1 << 11),
2880 nir_lower_imul_2x32_64
= (1 << 12),
2881 nir_lower_extract64
= (1 << 13),
2882 nir_lower_ufind_msb64
= (1 << 14),
2883 } nir_lower_int64_options
;
2886 nir_lower_drcp
= (1 << 0),
2887 nir_lower_dsqrt
= (1 << 1),
2888 nir_lower_drsq
= (1 << 2),
2889 nir_lower_dtrunc
= (1 << 3),
2890 nir_lower_dfloor
= (1 << 4),
2891 nir_lower_dceil
= (1 << 5),
2892 nir_lower_dfract
= (1 << 6),
2893 nir_lower_dround_even
= (1 << 7),
2894 nir_lower_dmod
= (1 << 8),
2895 nir_lower_dsub
= (1 << 9),
2896 nir_lower_ddiv
= (1 << 10),
2897 nir_lower_fp64_full_software
= (1 << 11),
2898 } nir_lower_doubles_options
;
2901 nir_divergence_single_prim_per_subgroup
= (1 << 0),
2902 nir_divergence_single_patch_per_tcs_subgroup
= (1 << 1),
2903 nir_divergence_single_patch_per_tes_subgroup
= (1 << 2),
2904 nir_divergence_view_index_uniform
= (1 << 3),
2905 } nir_divergence_options
;
2907 typedef struct nir_shader_compiler_options
{
2913 /** Lowers flrp when it does not support doubles */
2920 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2921 bool lower_bitfield_extract
;
2922 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
2923 bool lower_bitfield_extract_to_shifts
;
2924 /** Lowers bitfield_insert to bfi/bfm */
2925 bool lower_bitfield_insert
;
2926 /** Lowers bitfield_insert to compares, and shifts. */
2927 bool lower_bitfield_insert_to_shifts
;
2928 /** Lowers bitfield_insert to bfm/bitfield_select. */
2929 bool lower_bitfield_insert_to_bitfield_select
;
2930 /** Lowers bitfield_reverse to shifts. */
2931 bool lower_bitfield_reverse
;
2932 /** Lowers bit_count to shifts. */
2933 bool lower_bit_count
;
2934 /** Lowers ifind_msb to compare and ufind_msb */
2935 bool lower_ifind_msb
;
2936 /** Lowers find_lsb to ufind_msb and logic ops */
2937 bool lower_find_lsb
;
2938 bool lower_uadd_carry
;
2939 bool lower_usub_borrow
;
2940 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2941 bool lower_mul_high
;
2942 /** lowers fneg and ineg to fsub and isub. */
2944 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2947 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2950 /* lower fall_equalN/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
2951 bool lower_vector_cmp
;
2953 /** enables rules to lower idiv by power-of-two: */
2956 /** enable rules to avoid bit ops */
2959 /** enables rules to lower isign to imin+imax */
2962 /** enables rules to lower fsign to fsub and flt */
2965 /* lower fdph to fdot4 */
2968 /** lower fdot to fmul and fsum/fadd. */
2971 /* Does the native fdot instruction replicate its result for four
2972 * components? If so, then opt_algebraic_late will turn all fdotN
2973 * instructions into fdot_replicatedN instructions.
2975 bool fdot_replicates
;
2977 /** lowers ffloor to fsub+ffract: */
2980 /** lowers ffract to fsub+ffloor: */
2983 /** lowers fceil to fneg+ffloor+fneg: */
2990 bool lower_pack_half_2x16
;
2991 bool lower_pack_unorm_2x16
;
2992 bool lower_pack_snorm_2x16
;
2993 bool lower_pack_unorm_4x8
;
2994 bool lower_pack_snorm_4x8
;
2995 bool lower_unpack_half_2x16
;
2996 bool lower_unpack_unorm_2x16
;
2997 bool lower_unpack_snorm_2x16
;
2998 bool lower_unpack_unorm_4x8
;
2999 bool lower_unpack_snorm_4x8
;
3001 bool lower_pack_split
;
3003 bool lower_extract_byte
;
3004 bool lower_extract_word
;
3006 bool lower_all_io_to_temps
;
3007 bool lower_all_io_to_elements
;
3009 /* Indicates that the driver only has zero-based vertex id */
3010 bool vertex_id_zero_based
;
3013 * If enabled, gl_BaseVertex will be lowered as:
3014 * is_indexed_draw (~0/0) & firstvertex
3016 bool lower_base_vertex
;
3019 * If enabled, gl_HelperInvocation will be lowered as:
3021 * !((1 << sample_id) & sample_mask_in))
3023 * This depends on some possibly hw implementation details, which may
3024 * not be true for all hw. In particular that the FS is only executed
3025 * for covered samples or for helper invocations. So, do not blindly
3026 * enable this option.
3028 * Note: See also issue #22 in ARB_shader_image_load_store
3030 bool lower_helper_invocation
;
3033 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
3035 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
3036 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
3038 bool optimize_sample_mask_in
;
3040 bool lower_cs_local_index_from_id
;
3041 bool lower_cs_local_id_from_index
;
3043 bool lower_device_index_to_zero
;
3045 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
3046 bool lower_wpos_pntc
;
3049 * Set if nir_op_[iu]hadd and nir_op_[iu]rhadd instructions should be
3050 * lowered to simple arithmetic.
3052 * If this flag is set, the lowering will be applied to all bit-sizes of
3053 * these instructions.
3055 * \sa ::lower_hadd64
3060 * Set if only 64-bit nir_op_[iu]hadd and nir_op_[iu]rhadd instructions
3061 * should be lowered to simple arithmetic.
3063 * If this flag is set, the lowering will be applied to only 64-bit
3064 * versions of these instructions.
3071 * Set if nir_op_add_sat and nir_op_usub_sat should be lowered to simple
3074 * If this flag is set, the lowering will be applied to all bit-sizes of
3075 * these instructions.
3077 * \sa ::lower_usub_sat64
3082 * Set if only 64-bit nir_op_usub_sat should be lowered to simple
3085 * \sa ::lower_add_sat
3087 bool lower_usub_sat64
;
3090 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
3091 * for IO purposes and would prefer loads/stores be vectorized.
3094 bool lower_to_scalar
;
3097 * Whether nir_opt_vectorize should only create 16-bit 2D vectors.
3099 bool vectorize_vec2_16bit
;
3102 * Should the linker unify inputs_read/outputs_written between adjacent
3103 * shader stages which are linked into a single program?
3105 bool unify_interfaces
;
3108 * Should nir_lower_io() create load_interpolated_input intrinsics?
3110 * If not, it generates regular load_input intrinsics and interpolation
3111 * information must be inferred from the list of input nir_variables.
3113 bool use_interpolated_input_intrinsics
;
3115 /* Lowers when 32x32->64 bit multiplication is not supported */
3116 bool lower_mul_2x32_64
;
3118 /* Lowers when rotate instruction is not supported */
3122 * Backend supports imul24, and would like to use it (when possible)
3123 * for address/offset calculation. If true, driver should call
3124 * nir_lower_amul(). (If not set, amul will automatically be lowered
3129 /** Backend supports umul24, if not set umul24 will automatically be lowered
3130 * to imul with masked inputs */
3133 /** Backend supports umad24, if not set umad24 will automatically be lowered
3134 * to imul with masked inputs and iadd */
3137 /* Whether to generate only scoped_barrier intrinsics instead of the set of
3138 * memory and control barrier intrinsics based on GLSL.
3140 bool use_scoped_barrier
;
3143 * Is this the Intel vec4 backend?
3145 * Used to inhibit algebraic optimizations that are known to be harmful on
3146 * the Intel vec4 backend. This is generally applicable to any
3147 * optimization that might cause more immediate values to be used in
3148 * 3-source (e.g., ffma and flrp) instructions.
3152 /** Lower nir_op_ibfe and nir_op_ubfe that have two constant sources. */
3153 bool lower_bfe_with_two_constants
;
3155 /** Whether 8-bit ALU is supported. */
3156 bool support_8bit_alu
;
3158 /** Whether 16-bit ALU is supported. */
3159 bool support_16bit_alu
;
3161 unsigned max_unroll_iterations
;
3163 nir_lower_int64_options lower_int64_options
;
3164 nir_lower_doubles_options lower_doubles_options
;
3165 } nir_shader_compiler_options
;
3167 typedef struct nir_shader
{
3168 /** list of uniforms (nir_variable) */
3169 struct exec_list uniforms
;
3171 /** list of inputs (nir_variable) */
3172 struct exec_list inputs
;
3174 /** list of outputs (nir_variable) */
3175 struct exec_list outputs
;
3177 /** list of shared compute variables (nir_variable) */
3178 struct exec_list shared
;
3180 /** Set of driver-specific options for the shader.
3182 * The memory for the options is expected to be kept in a single static
3183 * copy by the driver.
3185 const struct nir_shader_compiler_options
*options
;
3187 /** Various bits of compile-time information about a given shader */
3188 struct shader_info info
;
3190 /** list of global variables in the shader (nir_variable) */
3191 struct exec_list globals
;
3193 /** list of system value variables in the shader (nir_variable) */
3194 struct exec_list system_values
;
3196 struct exec_list functions
; /** < list of nir_function */
3199 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
3202 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
3204 /** Size in bytes of required scratch space */
3205 unsigned scratch_size
;
3207 /** Constant data associated with this shader.
3209 * Constant data is loaded through load_constant intrinsics. See also
3210 * nir_opt_large_constants.
3212 void *constant_data
;
3213 unsigned constant_data_size
;
3216 #define nir_foreach_function(func, shader) \
3217 foreach_list_typed(nir_function, func, node, &(shader)->functions)
3219 static inline nir_function_impl
*
3220 nir_shader_get_entrypoint(nir_shader
*shader
)
3222 nir_function
*func
= NULL
;
3224 nir_foreach_function(function
, shader
) {
3225 assert(func
== NULL
);
3226 if (function
->is_entrypoint
) {
3237 assert(func
->num_params
== 0);
3242 nir_shader
*nir_shader_create(void *mem_ctx
,
3243 gl_shader_stage stage
,
3244 const nir_shader_compiler_options
*options
,
3247 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
3249 void nir_reg_remove(nir_register
*reg
);
3251 /** Adds a variable to the appropriate list in nir_shader */
3252 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
3255 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
3257 assert(var
->data
.mode
== nir_var_function_temp
);
3258 exec_list_push_tail(&impl
->locals
, &var
->node
);
3261 /** creates a variable, sets a few defaults, and adds it to the list */
3262 nir_variable
*nir_variable_create(nir_shader
*shader
,
3263 nir_variable_mode mode
,
3264 const struct glsl_type
*type
,
3266 /** creates a local variable and adds it to the list */
3267 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
3268 const struct glsl_type
*type
,
3271 /** creates a function and adds it to the shader's list of functions */
3272 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
3274 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
3275 /** creates a function_impl that isn't tied to any particular function */
3276 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
3278 nir_block
*nir_block_create(nir_shader
*shader
);
3279 nir_if
*nir_if_create(nir_shader
*shader
);
3280 nir_loop
*nir_loop_create(nir_shader
*shader
);
3282 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
3284 /** requests that the given pieces of metadata be generated */
3285 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
3286 /** dirties all but the preserved metadata */
3287 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
3289 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
3290 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
3292 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
3293 nir_deref_type deref_type
);
3295 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
3297 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
3298 unsigned num_components
,
3301 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
3302 nir_intrinsic_op op
);
3304 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
3305 nir_function
*callee
);
3307 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
3309 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
3311 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
3313 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
3314 unsigned num_components
,
3317 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
3320 * NIR Cursors and Instruction Insertion API
3323 * A tiny struct representing a point to insert/extract instructions or
3324 * control flow nodes. Helps reduce the combinatorial explosion of possible
3325 * points to insert/extract.
3327 * \sa nir_control_flow.h
3330 nir_cursor_before_block
,
3331 nir_cursor_after_block
,
3332 nir_cursor_before_instr
,
3333 nir_cursor_after_instr
,
3334 } nir_cursor_option
;
3337 nir_cursor_option option
;
3344 static inline nir_block
*
3345 nir_cursor_current_block(nir_cursor cursor
)
3347 if (cursor
.option
== nir_cursor_before_instr
||
3348 cursor
.option
== nir_cursor_after_instr
) {
3349 return cursor
.instr
->block
;
3351 return cursor
.block
;
3355 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
3357 static inline nir_cursor
3358 nir_before_block(nir_block
*block
)
3361 cursor
.option
= nir_cursor_before_block
;
3362 cursor
.block
= block
;
3366 static inline nir_cursor
3367 nir_after_block(nir_block
*block
)
3370 cursor
.option
= nir_cursor_after_block
;
3371 cursor
.block
= block
;
3375 static inline nir_cursor
3376 nir_before_instr(nir_instr
*instr
)
3379 cursor
.option
= nir_cursor_before_instr
;
3380 cursor
.instr
= instr
;
3384 static inline nir_cursor
3385 nir_after_instr(nir_instr
*instr
)
3388 cursor
.option
= nir_cursor_after_instr
;
3389 cursor
.instr
= instr
;
3393 static inline nir_cursor
3394 nir_after_block_before_jump(nir_block
*block
)
3396 nir_instr
*last_instr
= nir_block_last_instr(block
);
3397 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
3398 return nir_before_instr(last_instr
);
3400 return nir_after_block(block
);
3404 static inline nir_cursor
3405 nir_before_src(nir_src
*src
, bool is_if_condition
)
3407 if (is_if_condition
) {
3408 nir_block
*prev_block
=
3409 nir_cf_node_as_block(nir_cf_node_prev(&src
->parent_if
->cf_node
));
3410 assert(!nir_block_ends_in_jump(prev_block
));
3411 return nir_after_block(prev_block
);
3412 } else if (src
->parent_instr
->type
== nir_instr_type_phi
) {
3414 nir_phi_instr
*cond_phi
= nir_instr_as_phi(src
->parent_instr
);
3416 nir_foreach_phi_src(phi_src
, cond_phi
) {
3417 if (phi_src
->src
.ssa
== src
->ssa
) {
3424 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
3425 * to have a more specific name.
3427 nir_phi_src
*phi_src
= LIST_ENTRY(nir_phi_src
, src
, src
);
3428 return nir_after_block_before_jump(phi_src
->pred
);
3430 return nir_before_instr(src
->parent_instr
);
3434 static inline nir_cursor
3435 nir_before_cf_node(nir_cf_node
*node
)
3437 if (node
->type
== nir_cf_node_block
)
3438 return nir_before_block(nir_cf_node_as_block(node
));
3440 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
3443 static inline nir_cursor
3444 nir_after_cf_node(nir_cf_node
*node
)
3446 if (node
->type
== nir_cf_node_block
)
3447 return nir_after_block(nir_cf_node_as_block(node
));
3449 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
3452 static inline nir_cursor
3453 nir_after_phis(nir_block
*block
)
3455 nir_foreach_instr(instr
, block
) {
3456 if (instr
->type
!= nir_instr_type_phi
)
3457 return nir_before_instr(instr
);
3459 return nir_after_block(block
);
3462 static inline nir_cursor
3463 nir_after_cf_node_and_phis(nir_cf_node
*node
)
3465 if (node
->type
== nir_cf_node_block
)
3466 return nir_after_block(nir_cf_node_as_block(node
));
3468 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
3470 return nir_after_phis(block
);
3473 static inline nir_cursor
3474 nir_before_cf_list(struct exec_list
*cf_list
)
3476 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
3477 exec_list_get_head(cf_list
), node
);
3478 return nir_before_cf_node(first_node
);
3481 static inline nir_cursor
3482 nir_after_cf_list(struct exec_list
*cf_list
)
3484 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
3485 exec_list_get_tail(cf_list
), node
);
3486 return nir_after_cf_node(last_node
);
3490 * Insert a NIR instruction at the given cursor.
3492 * Note: This does not update the cursor.
3494 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
3497 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
3499 nir_instr_insert(nir_before_instr(instr
), before
);
3503 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
3505 nir_instr_insert(nir_after_instr(instr
), after
);
3509 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
3511 nir_instr_insert(nir_before_block(block
), before
);
3515 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
3517 nir_instr_insert(nir_after_block(block
), after
);
3521 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
3523 nir_instr_insert(nir_before_cf_node(node
), before
);
3527 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
3529 nir_instr_insert(nir_after_cf_node(node
), after
);
3533 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
3535 nir_instr_insert(nir_before_cf_list(list
), before
);
3539 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
3541 nir_instr_insert(nir_after_cf_list(list
), after
);
3544 void nir_instr_remove_v(nir_instr
*instr
);
3546 static inline nir_cursor
3547 nir_instr_remove(nir_instr
*instr
)
3550 nir_instr
*prev
= nir_instr_prev(instr
);
3552 cursor
= nir_after_instr(prev
);
3554 cursor
= nir_before_block(instr
->block
);
3556 nir_instr_remove_v(instr
);
3562 nir_ssa_def
*nir_instr_ssa_def(nir_instr
*instr
);
3564 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
3565 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
3566 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
3567 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
3569 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
3570 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
3571 bool nir_foreach_phi_src_leaving_block(nir_block
*instr
,
3572 nir_foreach_src_cb cb
,
3575 nir_const_value
*nir_src_as_const_value(nir_src src
);
3577 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
3578 static inline c_type * \
3579 nir_src_as_ ## name (nir_src src) \
3581 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
3582 ? cast_macro(src.ssa->parent_instr) : NULL; \
3585 NIR_SRC_AS_(alu_instr
, nir_alu_instr
, nir_instr_type_alu
, nir_instr_as_alu
)
3586 NIR_SRC_AS_(intrinsic
, nir_intrinsic_instr
,
3587 nir_instr_type_intrinsic
, nir_instr_as_intrinsic
)
3588 NIR_SRC_AS_(deref
, nir_deref_instr
, nir_instr_type_deref
, nir_instr_as_deref
)
3590 bool nir_src_is_dynamically_uniform(nir_src src
);
3591 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
3592 bool nir_instrs_equal(const nir_instr
*instr1
, const nir_instr
*instr2
);
3593 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
3594 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
3595 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
3596 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
3599 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
3600 unsigned num_components
, unsigned bit_size
,
3602 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
3603 unsigned num_components
, unsigned bit_size
,
3606 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
3607 const struct glsl_type
*type
,
3610 assert(glsl_type_is_vector_or_scalar(type
));
3611 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
3612 glsl_get_bit_size(type
), name
);
3614 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
3615 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
3616 nir_instr
*after_me
);
3618 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
3621 * finds the next basic block in source-code order, returns NULL if there is
3625 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
3627 /* Performs the opposite of nir_block_cf_tree_next() */
3629 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
3631 /* Gets the first block in a CF node in source-code order */
3633 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
3635 /* Gets the last block in a CF node in source-code order */
3637 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
3639 /* Gets the next block after a CF node in source-code order */
3641 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
3643 /* Macros for loops that visit blocks in source-code order */
3645 #define nir_foreach_block(block, impl) \
3646 for (nir_block *block = nir_start_block(impl); block != NULL; \
3647 block = nir_block_cf_tree_next(block))
3649 #define nir_foreach_block_safe(block, impl) \
3650 for (nir_block *block = nir_start_block(impl), \
3651 *next = nir_block_cf_tree_next(block); \
3653 block = next, next = nir_block_cf_tree_next(block))
3655 #define nir_foreach_block_reverse(block, impl) \
3656 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
3657 block = nir_block_cf_tree_prev(block))
3659 #define nir_foreach_block_reverse_safe(block, impl) \
3660 for (nir_block *block = nir_impl_last_block(impl), \
3661 *prev = nir_block_cf_tree_prev(block); \
3663 block = prev, prev = nir_block_cf_tree_prev(block))
3665 #define nir_foreach_block_in_cf_node(block, node) \
3666 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
3667 block != nir_cf_node_cf_tree_next(node); \
3668 block = nir_block_cf_tree_next(block))
3670 /* If the following CF node is an if, this function returns that if.
3671 * Otherwise, it returns NULL.
3673 nir_if
*nir_block_get_following_if(nir_block
*block
);
3675 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
3677 void nir_index_local_regs(nir_function_impl
*impl
);
3678 void nir_index_ssa_defs(nir_function_impl
*impl
);
3679 unsigned nir_index_instrs(nir_function_impl
*impl
);
3681 void nir_index_blocks(nir_function_impl
*impl
);
3683 void nir_index_vars(nir_shader
*shader
, nir_function_impl
*impl
, nir_variable_mode modes
);
3685 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
3686 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
3687 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
3688 void nir_print_deref(const nir_deref_instr
*deref
, FILE *fp
);
3690 /** Shallow clone of a single ALU instruction. */
3691 nir_alu_instr
*nir_alu_instr_clone(nir_shader
*s
, const nir_alu_instr
*orig
);
3693 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
3694 nir_function_impl
*nir_function_impl_clone(nir_shader
*shader
,
3695 const nir_function_impl
*fi
);
3696 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
3697 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
3699 void nir_shader_replace(nir_shader
*dest
, nir_shader
*src
);
3701 void nir_shader_serialize_deserialize(nir_shader
*s
);
3704 void nir_validate_shader(nir_shader
*shader
, const char *when
);
3705 void nir_metadata_set_validation_flag(nir_shader
*shader
);
3706 void nir_metadata_check_validation_flag(nir_shader
*shader
);
3709 should_skip_nir(const char *name
)
3711 static const char *list
= NULL
;
3713 /* Comma separated list of names to skip. */
3714 list
= getenv("NIR_SKIP");
3722 return comma_separated_list_contains(list
, name
);
3726 should_clone_nir(void)
3728 static int should_clone
= -1;
3729 if (should_clone
< 0)
3730 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
3732 return should_clone
;
3736 should_serialize_deserialize_nir(void)
3738 static int test_serialize
= -1;
3739 if (test_serialize
< 0)
3740 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
3742 return test_serialize
;
3746 should_print_nir(void)
3748 static int should_print
= -1;
3749 if (should_print
< 0)
3750 should_print
= env_var_as_boolean("NIR_PRINT", false);
3752 return should_print
;
3755 static inline void nir_validate_shader(nir_shader
*shader
, const char *when
) { (void) shader
; (void)when
; }
3756 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
3757 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
3758 static inline bool should_skip_nir(UNUSED
const char *pass_name
) { return false; }
3759 static inline bool should_clone_nir(void) { return false; }
3760 static inline bool should_serialize_deserialize_nir(void) { return false; }
3761 static inline bool should_print_nir(void) { return false; }
3764 #define _PASS(pass, nir, do_pass) do { \
3765 if (should_skip_nir(#pass)) { \
3766 printf("skipping %s\n", #pass); \
3770 nir_validate_shader(nir, "after " #pass); \
3771 if (should_clone_nir()) { \
3772 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3773 nir_shader_replace(nir, clone); \
3775 if (should_serialize_deserialize_nir()) { \
3776 nir_shader_serialize_deserialize(nir); \
3780 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3781 nir_metadata_set_validation_flag(nir); \
3782 if (should_print_nir()) \
3783 printf("%s\n", #pass); \
3784 if (pass(nir, ##__VA_ARGS__)) { \
3786 if (should_print_nir()) \
3787 nir_print_shader(nir, stdout); \
3788 nir_metadata_check_validation_flag(nir); \
3792 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3793 if (should_print_nir()) \
3794 printf("%s\n", #pass); \
3795 pass(nir, ##__VA_ARGS__); \
3796 if (should_print_nir()) \
3797 nir_print_shader(nir, stdout); \
3800 #define NIR_SKIP(name) should_skip_nir(#name)
3802 /** An instruction filtering callback
3804 * Returns true if the instruction should be processed and false otherwise.
3806 typedef bool (*nir_instr_filter_cb
)(const nir_instr
*, const void *);
3808 /** A simple instruction lowering callback
3810 * Many instruction lowering passes can be written as a simple function which
3811 * takes an instruction as its input and returns a sequence of instructions
3812 * that implement the consumed instruction. This function type represents
3813 * such a lowering function. When called, a function with this prototype
3814 * should either return NULL indicating that no lowering needs to be done or
3815 * emit a sequence of instructions using the provided builder (whose cursor
3816 * will already be placed after the instruction to be lowered) and return the
3817 * resulting nir_ssa_def.
3819 typedef nir_ssa_def
*(*nir_lower_instr_cb
)(struct nir_builder
*,
3820 nir_instr
*, void *);
3823 * Special return value for nir_lower_instr_cb when some progress occurred
3824 * (like changing an input to the instr) that didn't result in a replacement
3825 * SSA def being generated.
3827 #define NIR_LOWER_INSTR_PROGRESS ((nir_ssa_def *)(uintptr_t)1)
3829 /** Iterate over all the instructions in a nir_function_impl and lower them
3830 * using the provided callbacks
3832 * This function implements the guts of a standard lowering pass for you. It
3833 * iterates over all of the instructions in a nir_function_impl and calls the
3834 * filter callback on each one. If the filter callback returns true, it then
3835 * calls the lowering call back on the instruction. (Splitting it this way
3836 * allows us to avoid some save/restore work for instructions we know won't be
3837 * lowered.) If the instruction is dead after the lowering is complete, it
3838 * will be removed. If new instructions are added, the lowering callback will
3839 * also be called on them in case multiple lowerings are required.
3841 * The metadata for the nir_function_impl will also be updated. If any blocks
3842 * are added (they cannot be removed), dominance and block indices will be
3845 bool nir_function_impl_lower_instructions(nir_function_impl
*impl
,
3846 nir_instr_filter_cb filter
,
3847 nir_lower_instr_cb lower
,
3849 bool nir_shader_lower_instructions(nir_shader
*shader
,
3850 nir_instr_filter_cb filter
,
3851 nir_lower_instr_cb lower
,
3854 void nir_calc_dominance_impl(nir_function_impl
*impl
);
3855 void nir_calc_dominance(nir_shader
*shader
);
3857 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
3858 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
3859 bool nir_block_is_unreachable(nir_block
*block
);
3861 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
3862 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
3864 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
3865 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
3867 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
3868 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
3870 int nir_gs_count_vertices(const nir_shader
*shader
);
3872 bool nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3873 bool nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3874 bool nir_split_var_copies(nir_shader
*shader
);
3875 bool nir_split_per_member_structs(nir_shader
*shader
);
3876 bool nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
);
3878 bool nir_lower_returns_impl(nir_function_impl
*impl
);
3879 bool nir_lower_returns(nir_shader
*shader
);
3881 void nir_inline_function_impl(struct nir_builder
*b
,
3882 const nir_function_impl
*impl
,
3883 nir_ssa_def
**params
);
3884 bool nir_inline_functions(nir_shader
*shader
);
3886 bool nir_propagate_invariant(nir_shader
*shader
);
3888 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
3889 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
3890 nir_intrinsic_instr
*copy
);
3891 bool nir_lower_var_copies(nir_shader
*shader
);
3893 void nir_fixup_deref_modes(nir_shader
*shader
);
3895 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
3898 nir_lower_direct_array_deref_of_vec_load
= (1 << 0),
3899 nir_lower_indirect_array_deref_of_vec_load
= (1 << 1),
3900 nir_lower_direct_array_deref_of_vec_store
= (1 << 2),
3901 nir_lower_indirect_array_deref_of_vec_store
= (1 << 3),
3902 } nir_lower_array_deref_of_vec_options
;
3904 bool nir_lower_array_deref_of_vec(nir_shader
*shader
, nir_variable_mode modes
,
3905 nir_lower_array_deref_of_vec_options options
);
3907 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
3909 bool nir_lower_locals_to_regs(nir_shader
*shader
);
3911 void nir_lower_io_to_temporaries(nir_shader
*shader
,
3912 nir_function_impl
*entrypoint
,
3913 bool outputs
, bool inputs
);
3915 bool nir_lower_vars_to_scratch(nir_shader
*shader
,
3916 nir_variable_mode modes
,
3918 glsl_type_size_align_func size_align
);
3920 void nir_lower_clip_halfz(nir_shader
*shader
);
3922 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
3924 void nir_gather_ssa_types(nir_function_impl
*impl
,
3925 BITSET_WORD
*float_types
,
3926 BITSET_WORD
*int_types
);
3928 void nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
3929 int (*type_size
)(const struct glsl_type
*, bool));
3931 /* Some helpers to do very simple linking */
3932 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3933 bool nir_remove_unused_io_vars(nir_shader
*shader
, struct exec_list
*var_list
,
3934 uint64_t *used_by_other_stage
,
3935 uint64_t *used_by_other_stage_patches
);
3936 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
3937 bool default_to_smooth_interp
);
3938 void nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3939 bool nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3941 bool nir_lower_amul(nir_shader
*shader
,
3942 int (*type_size
)(const struct glsl_type
*, bool));
3944 void nir_assign_io_var_locations(struct exec_list
*var_list
,
3946 gl_shader_stage stage
);
3949 uint8_t num_linked_io_vars
;
3950 uint8_t num_linked_patch_io_vars
;
3951 } nir_linked_io_var_info
;
3953 nir_linked_io_var_info
3954 nir_assign_linked_io_var_locations(nir_shader
*producer
,
3955 nir_shader
*consumer
);
3958 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
3959 * to 32-bit operations. This is only valid for nir_var_shader_in/out
3962 nir_lower_io_lower_64bit_to_32
= (1 << 0),
3964 /* If set, this forces all non-flat fragment shader inputs to be
3965 * interpolated as if with the "sample" qualifier. This requires
3966 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
3968 nir_lower_io_force_sample_interpolation
= (1 << 1),
3969 } nir_lower_io_options
;
3970 bool nir_lower_io(nir_shader
*shader
,
3971 nir_variable_mode modes
,
3972 int (*type_size
)(const struct glsl_type
*, bool),
3973 nir_lower_io_options
);
3975 bool nir_io_add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
);
3978 nir_lower_vars_to_explicit_types(nir_shader
*shader
,
3979 nir_variable_mode modes
,
3980 glsl_type_size_align_func type_info
);
3984 * An address format which is a simple 32-bit global GPU address.
3986 nir_address_format_32bit_global
,
3989 * An address format which is a simple 64-bit global GPU address.
3991 nir_address_format_64bit_global
,
3994 * An address format which is a bounds-checked 64-bit global GPU address.
3996 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
3997 * address stored with the low bits in .x and high bits in .y, .z is a
3998 * size, and .w is an offset. When the final I/O operation is lowered, .w
3999 * is checked against .z and the operation is predicated on the result.
4001 nir_address_format_64bit_bounded_global
,
4004 * An address format which is comprised of a vec2 where the first
4005 * component is a buffer index and the second is an offset.
4007 nir_address_format_32bit_index_offset
,
4010 * An address format which is a simple 32-bit offset.
4012 nir_address_format_32bit_offset
,
4015 * An address format representing a purely logical addressing model. In
4016 * this model, all deref chains must be complete from the dereference
4017 * operation to the variable. Cast derefs are not allowed. These
4018 * addresses will be 32-bit scalars but the format is immaterial because
4019 * you can always chase the chain.
4021 nir_address_format_logical
,
4022 } nir_address_format
;
4024 static inline unsigned
4025 nir_address_format_bit_size(nir_address_format addr_format
)
4027 switch (addr_format
) {
4028 case nir_address_format_32bit_global
: return 32;
4029 case nir_address_format_64bit_global
: return 64;
4030 case nir_address_format_64bit_bounded_global
: return 32;
4031 case nir_address_format_32bit_index_offset
: return 32;
4032 case nir_address_format_32bit_offset
: return 32;
4033 case nir_address_format_logical
: return 32;
4035 unreachable("Invalid address format");
4038 static inline unsigned
4039 nir_address_format_num_components(nir_address_format addr_format
)
4041 switch (addr_format
) {
4042 case nir_address_format_32bit_global
: return 1;
4043 case nir_address_format_64bit_global
: return 1;
4044 case nir_address_format_64bit_bounded_global
: return 4;
4045 case nir_address_format_32bit_index_offset
: return 2;
4046 case nir_address_format_32bit_offset
: return 1;
4047 case nir_address_format_logical
: return 1;
4049 unreachable("Invalid address format");
4052 static inline const struct glsl_type
*
4053 nir_address_format_to_glsl_type(nir_address_format addr_format
)
4055 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
4056 assert(bit_size
== 32 || bit_size
== 64);
4057 return glsl_vector_type(bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
,
4058 nir_address_format_num_components(addr_format
));
4061 const nir_const_value
*nir_address_format_null_value(nir_address_format addr_format
);
4063 nir_ssa_def
*nir_build_addr_ieq(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
4064 nir_address_format addr_format
);
4066 nir_ssa_def
*nir_build_addr_isub(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
4067 nir_address_format addr_format
);
4069 nir_ssa_def
* nir_explicit_io_address_from_deref(struct nir_builder
*b
,
4070 nir_deref_instr
*deref
,
4071 nir_ssa_def
*base_addr
,
4072 nir_address_format addr_format
);
4073 void nir_lower_explicit_io_instr(struct nir_builder
*b
,
4074 nir_intrinsic_instr
*io_instr
,
4076 nir_address_format addr_format
);
4078 bool nir_lower_explicit_io(nir_shader
*shader
,
4079 nir_variable_mode modes
,
4080 nir_address_format
);
4082 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
4083 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
4085 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
4087 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
4088 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
4089 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
4091 bool nir_remove_dead_derefs(nir_shader
*shader
);
4092 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
4093 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
,
4094 bool (*can_remove_var
)(nir_variable
*var
));
4095 bool nir_lower_variable_initializers(nir_shader
*shader
,
4096 nir_variable_mode modes
);
4098 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
4099 bool nir_lower_vec_to_movs(nir_shader
*shader
);
4100 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
4102 const gl_state_index16
*alpha_ref_state_tokens
);
4103 bool nir_lower_alu(nir_shader
*shader
);
4105 bool nir_lower_flrp(nir_shader
*shader
, unsigned lowering_mask
,
4106 bool always_precise
, bool have_ffma
);
4108 bool nir_lower_alu_to_scalar(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
4109 bool nir_lower_bool_to_bitsize(nir_shader
*shader
);
4110 bool nir_lower_bool_to_float(nir_shader
*shader
);
4111 bool nir_lower_bool_to_int32(nir_shader
*shader
);
4112 bool nir_lower_int_to_float(nir_shader
*shader
);
4113 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
4114 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
4115 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
4116 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
4117 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
4119 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
4120 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
4121 bool nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode mask
);
4123 void nir_lower_fragcoord_wtrans(nir_shader
*shader
);
4124 void nir_lower_viewport_transform(nir_shader
*shader
);
4125 bool nir_lower_uniforms_to_ubo(nir_shader
*shader
, int multiplier
);
4127 typedef struct nir_lower_subgroups_options
{
4128 uint8_t subgroup_size
;
4129 uint8_t ballot_bit_size
;
4130 bool lower_to_scalar
:1;
4131 bool lower_vote_trivial
:1;
4132 bool lower_vote_eq_to_ballot
:1;
4133 bool lower_subgroup_masks
:1;
4134 bool lower_shuffle
:1;
4135 bool lower_shuffle_to_32bit
:1;
4137 bool lower_quad_broadcast_dynamic
:1;
4138 bool lower_quad_broadcast_dynamic_to_const
:1;
4139 } nir_lower_subgroups_options
;
4141 bool nir_lower_subgroups(nir_shader
*shader
,
4142 const nir_lower_subgroups_options
*options
);
4144 bool nir_lower_system_values(nir_shader
*shader
);
4146 enum PACKED nir_lower_tex_packing
{
4147 nir_lower_tex_packing_none
= 0,
4148 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
4149 * or unsigned ints based on the sampler type
4151 nir_lower_tex_packing_16
,
4152 /* The sampler returns 1 32-bit word of 4x8 unorm */
4153 nir_lower_tex_packing_8
,
4156 typedef struct nir_lower_tex_options
{
4158 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
4159 * sampler types a texture projector is lowered.
4164 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
4166 bool lower_txf_offset
;
4169 * If true, lower away nir_tex_src_offset for all rect textures.
4171 bool lower_rect_offset
;
4174 * If true, lower rect textures to 2D, using txs to fetch the
4175 * texture dimensions and dividing the texture coords by the
4176 * texture dims to normalize.
4181 * If true, convert yuv to rgb.
4183 unsigned lower_y_uv_external
;
4184 unsigned lower_y_u_v_external
;
4185 unsigned lower_yx_xuxv_external
;
4186 unsigned lower_xy_uxvx_external
;
4187 unsigned lower_ayuv_external
;
4188 unsigned lower_xyuv_external
;
4191 * To emulate certain texture wrap modes, this can be used
4192 * to saturate the specified tex coord to [0.0, 1.0]. The
4193 * bits are according to sampler #, ie. if, for example:
4195 * (conf->saturate_s & (1 << n))
4197 * is true, then the s coord for sampler n is saturated.
4199 * Note that clamping must happen *after* projector lowering
4200 * so any projected texture sample instruction with a clamped
4201 * coordinate gets automatically lowered, regardless of the
4202 * 'lower_txp' setting.
4204 unsigned saturate_s
;
4205 unsigned saturate_t
;
4206 unsigned saturate_r
;
4208 /* Bitmask of textures that need swizzling.
4210 * If (swizzle_result & (1 << texture_index)), then the swizzle in
4211 * swizzles[texture_index] is applied to the result of the texturing
4214 unsigned swizzle_result
;
4216 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
4217 * while 4 and 5 represent 0 and 1 respectively.
4219 uint8_t swizzles
[32][4];
4221 /* Can be used to scale sampled values in range required by the format. */
4222 float scale_factors
[32];
4225 * Bitmap of textures that need srgb to linear conversion. If
4226 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
4227 * of the texture are lowered to linear.
4229 unsigned lower_srgb
;
4232 * If true, lower nir_texop_tex on shaders that doesn't support implicit
4233 * LODs to nir_texop_txl.
4235 bool lower_tex_without_implicit_lod
;
4238 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
4240 bool lower_txd_cube_map
;
4243 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
4248 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
4249 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
4250 * with lower_txd_cube_map.
4252 bool lower_txd_shadow
;
4255 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
4256 * Implies lower_txd_cube_map and lower_txd_shadow.
4261 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
4262 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
4264 bool lower_txb_shadow_clamp
;
4267 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
4268 * with nir_texop_txl. This includes cube maps.
4270 bool lower_txd_shadow_clamp
;
4273 * If true, lower nir_texop_txd on when it uses both offset and min_lod
4274 * with nir_texop_txl. This includes cube maps.
4276 bool lower_txd_offset_clamp
;
4279 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4280 * sampler is bindless.
4282 bool lower_txd_clamp_bindless_sampler
;
4285 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4286 * sampler index is not statically determinable to be less than 16.
4288 bool lower_txd_clamp_if_sampler_index_not_lt_16
;
4291 * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
4292 * 0-lod followed by a nir_ishr.
4297 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
4298 * mixed-up tg4 locations.
4300 bool lower_tg4_broadcom_swizzle
;
4303 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
4305 bool lower_tg4_offsets
;
4307 enum nir_lower_tex_packing lower_tex_packing
[32];
4308 } nir_lower_tex_options
;
4310 bool nir_lower_tex(nir_shader
*shader
,
4311 const nir_lower_tex_options
*options
);
4313 enum nir_lower_non_uniform_access_type
{
4314 nir_lower_non_uniform_ubo_access
= (1 << 0),
4315 nir_lower_non_uniform_ssbo_access
= (1 << 1),
4316 nir_lower_non_uniform_texture_access
= (1 << 2),
4317 nir_lower_non_uniform_image_access
= (1 << 3),
4320 bool nir_lower_non_uniform_access(nir_shader
*shader
,
4321 enum nir_lower_non_uniform_access_type
);
4323 enum nir_lower_idiv_path
{
4324 /* This path is based on NV50LegalizeSSA::handleDIV(). It is the faster of
4325 * the two but it is not exact in some cases (for example, 1091317713u /
4326 * 1034u gives 5209173 instead of 1055432) */
4327 nir_lower_idiv_fast
,
4328 /* This path is based on AMDGPUTargetLowering::LowerUDIVREM() and
4329 * AMDGPUTargetLowering::LowerSDIVREM(). It requires more instructions than
4330 * the nv50 path and many of them are integer multiplications, so it is
4331 * probably slower. It should always return the correct result, though. */
4332 nir_lower_idiv_precise
,
4335 bool nir_lower_idiv(nir_shader
*shader
, enum nir_lower_idiv_path path
);
4337 bool nir_lower_input_attachments(nir_shader
*shader
, bool use_fragcoord_sysval
);
4339 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
,
4341 bool use_clipdist_array
,
4342 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4343 bool nir_lower_clip_gs(nir_shader
*shader
, unsigned ucp_enables
,
4344 bool use_clipdist_array
,
4345 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4346 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
,
4347 bool use_clipdist_array
);
4348 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
4350 void nir_lower_point_size_mov(nir_shader
*shader
,
4351 const gl_state_index16
*pointsize_state_tokens
);
4353 bool nir_lower_frexp(nir_shader
*nir
);
4355 void nir_lower_two_sided_color(nir_shader
*shader
);
4357 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
4359 bool nir_lower_flatshade(nir_shader
*shader
);
4361 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
4362 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
4363 const gl_state_index16
*uniform_state_tokens
);
4365 typedef struct nir_lower_wpos_ytransform_options
{
4366 gl_state_index16 state_tokens
[STATE_LENGTH
];
4367 bool fs_coord_origin_upper_left
:1;
4368 bool fs_coord_origin_lower_left
:1;
4369 bool fs_coord_pixel_center_integer
:1;
4370 bool fs_coord_pixel_center_half_integer
:1;
4371 } nir_lower_wpos_ytransform_options
;
4373 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
4374 const nir_lower_wpos_ytransform_options
*options
);
4375 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
4377 bool nir_lower_wrmasks(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
4379 bool nir_lower_fb_read(nir_shader
*shader
);
4381 typedef struct nir_lower_drawpixels_options
{
4382 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
4383 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
4384 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
4385 unsigned drawpix_sampler
;
4386 unsigned pixelmap_sampler
;
4388 bool scale_and_bias
:1;
4389 } nir_lower_drawpixels_options
;
4391 void nir_lower_drawpixels(nir_shader
*shader
,
4392 const nir_lower_drawpixels_options
*options
);
4394 typedef struct nir_lower_bitmap_options
{
4397 } nir_lower_bitmap_options
;
4399 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
4401 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
);
4404 nir_lower_int_source_mods
= 1 << 0,
4405 nir_lower_float_source_mods
= 1 << 1,
4406 nir_lower_triop_abs
= 1 << 2,
4407 nir_lower_all_source_mods
= (1 << 3) - 1
4408 } nir_lower_to_source_mods_flags
;
4411 bool nir_lower_to_source_mods(nir_shader
*shader
, nir_lower_to_source_mods_flags options
);
4413 bool nir_lower_gs_intrinsics(nir_shader
*shader
, bool per_stream
);
4415 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
4417 bool nir_lower_bit_size(nir_shader
*shader
,
4418 nir_lower_bit_size_callback callback
,
4419 void *callback_data
);
4421 nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode
);
4422 bool nir_lower_int64(nir_shader
*shader
, nir_lower_int64_options options
);
4424 nir_lower_doubles_options
nir_lower_doubles_op_to_options_mask(nir_op opcode
);
4425 bool nir_lower_doubles(nir_shader
*shader
, const nir_shader
*softfp64
,
4426 nir_lower_doubles_options options
);
4427 bool nir_lower_pack(nir_shader
*shader
);
4429 void nir_lower_mediump_outputs(nir_shader
*nir
);
4431 bool nir_lower_point_size(nir_shader
*shader
, float min
, float max
);
4434 nir_lower_interpolation_at_sample
= (1 << 1),
4435 nir_lower_interpolation_at_offset
= (1 << 2),
4436 nir_lower_interpolation_centroid
= (1 << 3),
4437 nir_lower_interpolation_pixel
= (1 << 4),
4438 nir_lower_interpolation_sample
= (1 << 5),
4439 } nir_lower_interpolation_options
;
4441 bool nir_lower_interpolation(nir_shader
*shader
,
4442 nir_lower_interpolation_options options
);
4444 bool nir_lower_discard_to_demote(nir_shader
*shader
);
4446 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
4448 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
4450 void nir_loop_analyze_impl(nir_function_impl
*impl
,
4451 nir_variable_mode indirect_mask
);
4453 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
4455 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
4456 bool nir_repair_ssa(nir_shader
*shader
);
4458 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
4459 bool nir_convert_to_lcssa(nir_shader
*shader
, bool skip_invariants
, bool skip_bool_invariants
);
4460 void nir_divergence_analysis(nir_shader
*shader
, nir_divergence_options options
);
4462 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
4463 * registers. If false, convert all values (even those not involved in a phi
4464 * node) to registers.
4466 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
4468 bool nir_lower_phis_to_regs_block(nir_block
*block
);
4469 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
4470 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
);
4472 bool nir_lower_samplers(nir_shader
*shader
);
4473 bool nir_lower_ssbo(nir_shader
*shader
);
4475 /* This is here for unit tests. */
4476 bool nir_opt_comparison_pre_impl(nir_function_impl
*impl
);
4478 bool nir_opt_comparison_pre(nir_shader
*shader
);
4480 bool nir_opt_access(nir_shader
*shader
);
4481 bool nir_opt_algebraic(nir_shader
*shader
);
4482 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
4483 bool nir_opt_algebraic_late(nir_shader
*shader
);
4484 bool nir_opt_algebraic_distribute_src_mods(nir_shader
*shader
);
4485 bool nir_opt_constant_folding(nir_shader
*shader
);
4487 /* Try to combine a and b into a. Return true if combination was possible,
4488 * which will result in b being removed by the pass. Return false if
4489 * combination wasn't possible.
4491 typedef bool (*nir_combine_memory_barrier_cb
)(
4492 nir_intrinsic_instr
*a
, nir_intrinsic_instr
*b
, void *data
);
4494 bool nir_opt_combine_memory_barriers(nir_shader
*shader
,
4495 nir_combine_memory_barrier_cb combine_cb
,
4498 bool nir_opt_combine_stores(nir_shader
*shader
, nir_variable_mode modes
);
4500 bool nir_copy_prop(nir_shader
*shader
);
4502 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
4504 bool nir_opt_cse(nir_shader
*shader
);
4506 bool nir_opt_dce(nir_shader
*shader
);
4508 bool nir_opt_dead_cf(nir_shader
*shader
);
4510 bool nir_opt_dead_write_vars(nir_shader
*shader
);
4512 bool nir_opt_deref_impl(nir_function_impl
*impl
);
4513 bool nir_opt_deref(nir_shader
*shader
);
4515 bool nir_opt_find_array_copies(nir_shader
*shader
);
4517 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
4519 bool nir_opt_idiv_const(nir_shader
*shader
, unsigned min_bit_size
);
4521 bool nir_opt_if(nir_shader
*shader
, bool aggressive_last_continue
);
4523 bool nir_opt_intrinsics(nir_shader
*shader
);
4525 bool nir_opt_large_constants(nir_shader
*shader
,
4526 glsl_type_size_align_func size_align
,
4527 unsigned threshold
);
4529 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
4532 nir_move_const_undef
= (1 << 0),
4533 nir_move_load_ubo
= (1 << 1),
4534 nir_move_load_input
= (1 << 2),
4535 nir_move_comparisons
= (1 << 3),
4536 nir_move_copies
= (1 << 4),
4539 bool nir_can_move_instr(nir_instr
*instr
, nir_move_options options
);
4541 bool nir_opt_sink(nir_shader
*shader
, nir_move_options options
);
4543 bool nir_opt_move(nir_shader
*shader
, nir_move_options options
);
4545 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
,
4546 bool indirect_load_ok
, bool expensive_alu_ok
);
4548 bool nir_opt_rematerialize_compares(nir_shader
*shader
);
4550 bool nir_opt_remove_phis(nir_shader
*shader
);
4551 bool nir_opt_remove_phis_block(nir_block
*block
);
4553 bool nir_opt_shrink_load(nir_shader
*shader
);
4555 bool nir_opt_trivial_continues(nir_shader
*shader
);
4557 bool nir_opt_undef(nir_shader
*shader
);
4559 bool nir_opt_vectorize(nir_shader
*shader
);
4561 bool nir_opt_conditional_discard(nir_shader
*shader
);
4563 typedef bool (*nir_should_vectorize_mem_func
)(unsigned align
, unsigned bit_size
,
4564 unsigned num_components
, unsigned high_offset
,
4565 nir_intrinsic_instr
*low
, nir_intrinsic_instr
*high
);
4567 bool nir_opt_load_store_vectorize(nir_shader
*shader
, nir_variable_mode modes
,
4568 nir_should_vectorize_mem_func callback
,
4569 nir_variable_mode robust_modes
);
4571 void nir_schedule(nir_shader
*shader
, int threshold
);
4573 void nir_strip(nir_shader
*shader
);
4575 void nir_sweep(nir_shader
*shader
);
4577 void nir_remap_dual_slot_attributes(nir_shader
*shader
,
4578 uint64_t *dual_slot_inputs
);
4579 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
);
4581 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
4582 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);
4585 nir_variable_is_in_ubo(const nir_variable
*var
)
4587 return (var
->data
.mode
== nir_var_mem_ubo
&&
4588 var
->interface_type
!= NULL
);
4592 nir_variable_is_in_ssbo(const nir_variable
*var
)
4594 return (var
->data
.mode
== nir_var_mem_ssbo
&&
4595 var
->interface_type
!= NULL
);
4599 nir_variable_is_in_block(const nir_variable
*var
)
4601 return nir_variable_is_in_ubo(var
) || nir_variable_is_in_ssbo(var
);