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
;
622 _nir_shader_variable_has_mode(nir_variable
*var
, unsigned modes
)
624 /* This isn't a shader variable */
625 assert(!(modes
& nir_var_function_temp
));
626 return var
->data
.mode
& modes
;
629 #define nir_foreach_variable_in_list(var, var_list) \
630 foreach_list_typed(nir_variable, var, node, var_list)
632 #define nir_foreach_variable_in_list_safe(var, var_list) \
633 foreach_list_typed_safe(nir_variable, var, node, var_list)
635 #define nir_foreach_variable_in_shader(var, shader) \
636 nir_foreach_variable_in_list(var, &(shader)->variables)
638 #define nir_foreach_variable_in_shader_safe(var, shader) \
639 nir_foreach_variable_in_list_safe(var, &(shader)->variables)
641 #define nir_foreach_variable_with_modes(var, shader, modes) \
642 nir_foreach_variable_in_shader(var, shader) \
643 if (_nir_shader_variable_has_mode(var, modes))
645 #define nir_foreach_variable_with_modes_safe(var, shader, modes) \
646 nir_foreach_variable_in_shader_safe(var, shader) \
647 if (_nir_shader_variable_has_mode(var, modes))
649 #define nir_foreach_shader_in_variable(var, shader) \
650 nir_foreach_variable_with_modes(var, shader, nir_var_shader_in)
652 #define nir_foreach_shader_in_variable_safe(var, shader) \
653 nir_foreach_variable_with_modes_safe(var, shader, nir_var_shader_in)
655 #define nir_foreach_shader_out_variable(var, shader) \
656 nir_foreach_variable_with_modes(var, shader, nir_var_shader_out)
658 #define nir_foreach_shader_out_variable_safe(var, shader) \
659 nir_foreach_variable_with_modes_safe(var, shader, nir_var_shader_out)
661 #define nir_foreach_uniform_variable(var, shader) \
662 nir_foreach_variable_with_modes(var, shader, nir_var_uniform)
664 #define nir_foreach_uniform_variable_safe(var, shader) \
665 nir_foreach_variable_with_modes_safe(var, shader, nir_var_uniform)
668 nir_variable_is_global(const nir_variable
*var
)
670 return var
->data
.mode
!= nir_var_function_temp
;
673 typedef struct nir_register
{
674 struct exec_node node
;
676 unsigned num_components
; /** < number of vector components */
677 unsigned num_array_elems
; /** < size of array (0 for no array) */
679 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
682 /** generic register index. */
685 /** only for debug purposes, can be NULL */
688 /** set of nir_srcs where this register is used (read from) */
689 struct list_head uses
;
691 /** set of nir_dests where this register is defined (written to) */
692 struct list_head defs
;
694 /** set of nir_ifs where this register is used as a condition */
695 struct list_head if_uses
;
698 #define nir_foreach_register(reg, reg_list) \
699 foreach_list_typed(nir_register, reg, node, reg_list)
700 #define nir_foreach_register_safe(reg, reg_list) \
701 foreach_list_typed_safe(nir_register, reg, node, reg_list)
703 typedef enum PACKED
{
705 nir_instr_type_deref
,
708 nir_instr_type_intrinsic
,
709 nir_instr_type_load_const
,
711 nir_instr_type_ssa_undef
,
713 nir_instr_type_parallel_copy
,
716 typedef struct nir_instr
{
717 struct exec_node node
;
718 struct nir_block
*block
;
721 /* A temporary for optimization and analysis passes to use for storing
722 * flags. For instance, DCE uses this to store the "dead/live" info.
726 /** generic instruction index. */
730 static inline nir_instr
*
731 nir_instr_next(nir_instr
*instr
)
733 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
734 if (exec_node_is_tail_sentinel(next
))
737 return exec_node_data(nir_instr
, next
, node
);
740 static inline nir_instr
*
741 nir_instr_prev(nir_instr
*instr
)
743 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
744 if (exec_node_is_head_sentinel(prev
))
747 return exec_node_data(nir_instr
, prev
, node
);
751 nir_instr_is_first(const nir_instr
*instr
)
753 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr
->node
));
757 nir_instr_is_last(const nir_instr
*instr
)
759 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr
->node
));
762 typedef struct nir_ssa_def
{
763 /** for debugging only, can be NULL */
766 /** generic SSA definition index. */
769 /** Index into the live_in and live_out bitfields */
772 /** Instruction which produces this SSA value. */
773 nir_instr
*parent_instr
;
775 /** set of nir_instrs where this register is used (read from) */
776 struct list_head uses
;
778 /** set of nir_ifs where this register is used as a condition */
779 struct list_head if_uses
;
781 uint8_t num_components
;
783 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
787 * True if this SSA value may have different values in different SIMD
788 * invocations of the shader. This is set by nir_divergence_analysis.
797 struct nir_src
*indirect
; /** < NULL for no indirect offset */
798 unsigned base_offset
;
800 /* TODO use-def chain goes here */
804 nir_instr
*parent_instr
;
805 struct list_head def_link
;
808 struct nir_src
*indirect
; /** < NULL for no indirect offset */
809 unsigned base_offset
;
811 /* TODO def-use chain goes here */
816 typedef struct nir_src
{
818 /** Instruction that consumes this value as a source. */
819 nir_instr
*parent_instr
;
820 struct nir_if
*parent_if
;
823 struct list_head use_link
;
833 static inline nir_src
836 nir_src src
= { { NULL
} };
840 #define NIR_SRC_INIT nir_src_init()
842 #define nir_foreach_use(src, reg_or_ssa_def) \
843 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
845 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
846 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
848 #define nir_foreach_if_use(src, reg_or_ssa_def) \
849 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
851 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
852 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
863 static inline nir_dest
866 nir_dest dest
= { { { NULL
} } };
870 #define NIR_DEST_INIT nir_dest_init()
872 #define nir_foreach_def(dest, reg) \
873 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
875 #define nir_foreach_def_safe(dest, reg) \
876 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
878 static inline nir_src
879 nir_src_for_ssa(nir_ssa_def
*def
)
881 nir_src src
= NIR_SRC_INIT
;
889 static inline nir_src
890 nir_src_for_reg(nir_register
*reg
)
892 nir_src src
= NIR_SRC_INIT
;
896 src
.reg
.indirect
= NULL
;
897 src
.reg
.base_offset
= 0;
902 static inline nir_dest
903 nir_dest_for_reg(nir_register
*reg
)
905 nir_dest dest
= NIR_DEST_INIT
;
912 static inline unsigned
913 nir_src_bit_size(nir_src src
)
915 return src
.is_ssa
? src
.ssa
->bit_size
: src
.reg
.reg
->bit_size
;
918 static inline unsigned
919 nir_src_num_components(nir_src src
)
921 return src
.is_ssa
? src
.ssa
->num_components
: src
.reg
.reg
->num_components
;
925 nir_src_is_const(nir_src src
)
928 src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
;
932 nir_src_is_divergent(nir_src src
)
935 return src
.ssa
->divergent
;
938 static inline unsigned
939 nir_dest_bit_size(nir_dest dest
)
941 return dest
.is_ssa
? dest
.ssa
.bit_size
: dest
.reg
.reg
->bit_size
;
944 static inline unsigned
945 nir_dest_num_components(nir_dest dest
)
947 return dest
.is_ssa
? dest
.ssa
.num_components
: dest
.reg
.reg
->num_components
;
951 nir_dest_is_divergent(nir_dest dest
)
954 return dest
.ssa
.divergent
;
957 /* Are all components the same, ie. .xxxx */
959 nir_is_same_comp_swizzle(uint8_t *swiz
, unsigned nr_comp
)
961 for (unsigned i
= 1; i
< nr_comp
; i
++)
962 if (swiz
[i
] != swiz
[0])
967 /* Are all components sequential, ie. .yzw */
969 nir_is_sequential_comp_swizzle(uint8_t *swiz
, unsigned nr_comp
)
971 for (unsigned i
= 1; i
< nr_comp
; i
++)
972 if (swiz
[i
] != (swiz
[0] + i
))
977 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
978 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
984 * \name input modifiers
988 * For inputs interpreted as floating point, flips the sign bit. For
989 * inputs interpreted as integers, performs the two's complement negation.
994 * Clears the sign bit for floating point values, and computes the integer
995 * absolute value for integers. Note that the negate modifier acts after
996 * the absolute value modifier, therefore if both are set then all inputs
997 * will become negative.
1003 * For each input component, says which component of the register it is
1004 * chosen from. Note that which elements of the swizzle are used and which
1005 * are ignored are based on the write mask for most opcodes - for example,
1006 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
1007 * a swizzle of {2, x, 1, 0} where x means "don't care."
1009 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
1016 * \name saturate output modifier
1018 * Only valid for opcodes that output floating-point numbers. Clamps the
1019 * output to between 0.0 and 1.0 inclusive.
1024 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
1027 /** NIR sized and unsized types
1029 * The values in this enum are carefully chosen so that the sized type is
1030 * just the unsized type OR the number of bits.
1032 typedef enum PACKED
{
1033 nir_type_invalid
= 0, /* Not a valid type */
1037 nir_type_float
= 128,
1038 nir_type_bool1
= 1 | nir_type_bool
,
1039 nir_type_bool8
= 8 | nir_type_bool
,
1040 nir_type_bool16
= 16 | nir_type_bool
,
1041 nir_type_bool32
= 32 | nir_type_bool
,
1042 nir_type_int1
= 1 | nir_type_int
,
1043 nir_type_int8
= 8 | nir_type_int
,
1044 nir_type_int16
= 16 | nir_type_int
,
1045 nir_type_int32
= 32 | nir_type_int
,
1046 nir_type_int64
= 64 | nir_type_int
,
1047 nir_type_uint1
= 1 | nir_type_uint
,
1048 nir_type_uint8
= 8 | nir_type_uint
,
1049 nir_type_uint16
= 16 | nir_type_uint
,
1050 nir_type_uint32
= 32 | nir_type_uint
,
1051 nir_type_uint64
= 64 | nir_type_uint
,
1052 nir_type_float16
= 16 | nir_type_float
,
1053 nir_type_float32
= 32 | nir_type_float
,
1054 nir_type_float64
= 64 | nir_type_float
,
1057 #define NIR_ALU_TYPE_SIZE_MASK 0x79
1058 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
1060 static inline unsigned
1061 nir_alu_type_get_type_size(nir_alu_type type
)
1063 return type
& NIR_ALU_TYPE_SIZE_MASK
;
1066 static inline nir_alu_type
1067 nir_alu_type_get_base_type(nir_alu_type type
)
1069 return (nir_alu_type
)(type
& NIR_ALU_TYPE_BASE_TYPE_MASK
);
1072 static inline nir_alu_type
1073 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
1075 switch (base_type
) {
1076 case GLSL_TYPE_BOOL
:
1077 return nir_type_bool1
;
1079 case GLSL_TYPE_UINT
:
1080 return nir_type_uint32
;
1083 return nir_type_int32
;
1085 case GLSL_TYPE_UINT16
:
1086 return nir_type_uint16
;
1088 case GLSL_TYPE_INT16
:
1089 return nir_type_int16
;
1091 case GLSL_TYPE_UINT8
:
1092 return nir_type_uint8
;
1093 case GLSL_TYPE_INT8
:
1094 return nir_type_int8
;
1095 case GLSL_TYPE_UINT64
:
1096 return nir_type_uint64
;
1098 case GLSL_TYPE_INT64
:
1099 return nir_type_int64
;
1101 case GLSL_TYPE_FLOAT
:
1102 return nir_type_float32
;
1104 case GLSL_TYPE_FLOAT16
:
1105 return nir_type_float16
;
1107 case GLSL_TYPE_DOUBLE
:
1108 return nir_type_float64
;
1111 case GLSL_TYPE_SAMPLER
:
1112 case GLSL_TYPE_IMAGE
:
1113 case GLSL_TYPE_ATOMIC_UINT
:
1114 case GLSL_TYPE_STRUCT
:
1115 case GLSL_TYPE_INTERFACE
:
1116 case GLSL_TYPE_ARRAY
:
1117 case GLSL_TYPE_VOID
:
1118 case GLSL_TYPE_SUBROUTINE
:
1119 case GLSL_TYPE_FUNCTION
:
1120 case GLSL_TYPE_ERROR
:
1121 return nir_type_invalid
;
1124 unreachable("unknown type");
1127 static inline nir_alu_type
1128 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
1130 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
1133 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
1134 nir_rounding_mode rnd
);
1136 static inline nir_op
1137 nir_op_vec(unsigned components
)
1139 switch (components
) {
1140 case 1: return nir_op_mov
;
1141 case 2: return nir_op_vec2
;
1142 case 3: return nir_op_vec3
;
1143 case 4: return nir_op_vec4
;
1144 case 8: return nir_op_vec8
;
1145 case 16: return nir_op_vec16
;
1146 default: unreachable("bad component count");
1151 nir_op_is_vec(nir_op op
)
1167 nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode
, unsigned bit_size
)
1169 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
) ||
1170 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
) ||
1171 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
);
1175 nir_is_denorm_flush_to_zero(unsigned execution_mode
, unsigned bit_size
)
1177 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
) ||
1178 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
) ||
1179 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
);
1183 nir_is_denorm_preserve(unsigned execution_mode
, unsigned bit_size
)
1185 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP16
) ||
1186 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP32
) ||
1187 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP64
);
1191 nir_is_rounding_mode_rtne(unsigned execution_mode
, unsigned bit_size
)
1193 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1194 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1195 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1199 nir_is_rounding_mode_rtz(unsigned execution_mode
, unsigned bit_size
)
1201 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1202 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1203 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1207 nir_has_any_rounding_mode_rtz(unsigned execution_mode
)
1209 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1210 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1211 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1215 nir_has_any_rounding_mode_rtne(unsigned execution_mode
)
1217 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1218 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1219 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1222 static inline nir_rounding_mode
1223 nir_get_rounding_mode_from_float_controls(unsigned execution_mode
,
1226 if (nir_alu_type_get_base_type(type
) != nir_type_float
)
1227 return nir_rounding_mode_undef
;
1229 unsigned bit_size
= nir_alu_type_get_type_size(type
);
1231 if (nir_is_rounding_mode_rtz(execution_mode
, bit_size
))
1232 return nir_rounding_mode_rtz
;
1233 if (nir_is_rounding_mode_rtne(execution_mode
, bit_size
))
1234 return nir_rounding_mode_rtne
;
1235 return nir_rounding_mode_undef
;
1239 nir_has_any_rounding_mode_enabled(unsigned execution_mode
)
1242 nir_has_any_rounding_mode_rtne(execution_mode
) ||
1243 nir_has_any_rounding_mode_rtz(execution_mode
);
1249 * Operation where the first two sources are commutative.
1251 * For 2-source operations, this just mathematical commutativity. Some
1252 * 3-source operations, like ffma, are only commutative in the first two
1255 NIR_OP_IS_2SRC_COMMUTATIVE
= (1 << 0),
1256 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
1257 } nir_op_algebraic_property
;
1265 * The number of components in the output
1267 * If non-zero, this is the size of the output and input sizes are
1268 * explicitly given; swizzle and writemask are still in effect, but if
1269 * the output component is masked out, then the input component may
1272 * If zero, the opcode acts in the standard, per-component manner; the
1273 * operation is performed on each component (except the ones that are
1274 * masked out) with the input being taken from the input swizzle for
1277 * The size of some of the inputs may be given (i.e. non-zero) even
1278 * though output_size is zero; in that case, the inputs with a zero
1279 * size act per-component, while the inputs with non-zero size don't.
1281 uint8_t output_size
;
1284 * The type of vector that the instruction outputs. Note that the
1285 * staurate modifier is only allowed on outputs with the float type.
1288 nir_alu_type output_type
;
1291 * The number of components in each input
1293 uint8_t input_sizes
[NIR_MAX_VEC_COMPONENTS
];
1296 * The type of vector that each input takes. Note that negate and
1297 * absolute value are only allowed on inputs with int or float type and
1298 * behave differently on the two.
1300 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
1302 nir_op_algebraic_property algebraic_properties
;
1304 /* Whether this represents a numeric conversion opcode */
1308 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
1310 typedef struct nir_alu_instr
{
1314 /** Indicates that this ALU instruction generates an exact value
1316 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1317 * really equivalent to either. This indicates that the value generated by
1318 * this operation is high-precision and any code transformations that touch
1319 * it must ensure that the resulting value is bit-for-bit identical to the
1325 * Indicates that this instruction do not cause wrapping to occur, in the
1326 * form of overflow or underflow.
1328 bool no_signed_wrap
:1;
1329 bool no_unsigned_wrap
:1;
1335 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
1336 nir_alu_instr
*instr
);
1337 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
1338 nir_alu_instr
*instr
);
1340 /* is this source channel used? */
1342 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
1345 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1346 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
1348 return (instr
->dest
.write_mask
>> channel
) & 1;
1351 static inline nir_component_mask_t
1352 nir_alu_instr_src_read_mask(const nir_alu_instr
*instr
, unsigned src
)
1354 nir_component_mask_t read_mask
= 0;
1355 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; c
++) {
1356 if (!nir_alu_instr_channel_used(instr
, src
, c
))
1359 read_mask
|= (1 << instr
->src
[src
].swizzle
[c
]);
1365 * Get the number of channels used for a source
1367 static inline unsigned
1368 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
1370 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1371 return nir_op_infos
[instr
->op
].input_sizes
[src
];
1373 return nir_dest_num_components(instr
->dest
.dest
);
1377 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1379 switch (instr
->op
) {
1399 bool nir_const_value_negative_equal(nir_const_value c1
, nir_const_value c2
,
1400 nir_alu_type full_type
);
1402 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
1403 unsigned src1
, unsigned src2
);
1405 bool nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
1406 const nir_alu_instr
*alu2
,
1407 unsigned src1
, unsigned src2
);
1411 nir_deref_type_array
,
1412 nir_deref_type_array_wildcard
,
1413 nir_deref_type_ptr_as_array
,
1414 nir_deref_type_struct
,
1415 nir_deref_type_cast
,
1421 /** The type of this deref instruction */
1422 nir_deref_type deref_type
;
1424 /** The mode of the underlying variable */
1425 nir_variable_mode mode
;
1427 /** The dereferenced type of the resulting pointer value */
1428 const struct glsl_type
*type
;
1431 /** Variable being dereferenced if deref_type is a deref_var */
1434 /** Parent deref if deref_type is not deref_var */
1438 /** Additional deref parameters */
1449 unsigned ptr_stride
;
1453 /** Destination to store the resulting "pointer" */
1457 static inline nir_deref_instr
*nir_src_as_deref(nir_src src
);
1459 static inline nir_deref_instr
*
1460 nir_deref_instr_parent(const nir_deref_instr
*instr
)
1462 if (instr
->deref_type
== nir_deref_type_var
)
1465 return nir_src_as_deref(instr
->parent
);
1468 static inline nir_variable
*
1469 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1471 while (instr
->deref_type
!= nir_deref_type_var
) {
1472 if (instr
->deref_type
== nir_deref_type_cast
)
1475 instr
= nir_deref_instr_parent(instr
);
1481 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1482 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr
*instr
);
1483 bool nir_deref_instr_has_complex_use(nir_deref_instr
*instr
);
1485 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1487 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*instr
);
1492 struct nir_function
*callee
;
1494 unsigned num_params
;
1498 #include "nir_intrinsics.h"
1500 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1502 /** Represents an intrinsic
1504 * An intrinsic is an instruction type for handling things that are
1505 * more-or-less regular operations but don't just consume and produce SSA
1506 * values like ALU operations do. Intrinsics are not for things that have
1507 * special semantic meaning such as phi nodes and parallel copies.
1508 * Examples of intrinsics include variable load/store operations, system
1509 * value loads, and the like. Even though texturing more-or-less falls
1510 * under this category, texturing is its own instruction type because
1511 * trying to represent texturing with intrinsics would lead to a
1512 * combinatorial explosion of intrinsic opcodes.
1514 * By having a single instruction type for handling a lot of different
1515 * cases, optimization passes can look for intrinsics and, for the most
1516 * part, completely ignore them. Each intrinsic type also has a few
1517 * possible flags that govern whether or not they can be reordered or
1518 * eliminated. That way passes like dead code elimination can still work
1519 * on intrisics without understanding the meaning of each.
1521 * Each intrinsic has some number of constant indices, some number of
1522 * variables, and some number of sources. What these sources, variables,
1523 * and indices mean depends on the intrinsic and is documented with the
1524 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1525 * instructions are the only types of instruction that can operate on
1531 nir_intrinsic_op intrinsic
;
1535 /** number of components if this is a vectorized intrinsic
1537 * Similarly to ALU operations, some intrinsics are vectorized.
1538 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1539 * For vectorized intrinsics, the num_components field specifies the
1540 * number of destination components and the number of source components
1541 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1543 uint8_t num_components
;
1545 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1548 } nir_intrinsic_instr
;
1550 static inline nir_variable
*
1551 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1553 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1557 /* Memory ordering. */
1558 NIR_MEMORY_ACQUIRE
= 1 << 0,
1559 NIR_MEMORY_RELEASE
= 1 << 1,
1560 NIR_MEMORY_ACQ_REL
= NIR_MEMORY_ACQUIRE
| NIR_MEMORY_RELEASE
,
1562 /* Memory visibility operations. */
1563 NIR_MEMORY_MAKE_AVAILABLE
= 1 << 2,
1564 NIR_MEMORY_MAKE_VISIBLE
= 1 << 3,
1565 } nir_memory_semantics
;
1569 NIR_SCOPE_INVOCATION
,
1571 NIR_SCOPE_WORKGROUP
,
1572 NIR_SCOPE_QUEUE_FAMILY
,
1577 * \name NIR intrinsics semantic flags
1579 * information about what the compiler can do with the intrinsics.
1581 * \sa nir_intrinsic_info::flags
1585 * whether the intrinsic can be safely eliminated if none of its output
1586 * value is not being used.
1588 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1591 * Whether the intrinsic can be reordered with respect to any other
1592 * intrinsic, i.e. whether the only reordering dependencies of the
1593 * intrinsic are due to the register reads/writes.
1595 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1596 } nir_intrinsic_semantic_flag
;
1599 * \name NIR intrinsics const-index flag
1601 * Indicates the usage of a const_index slot.
1603 * \sa nir_intrinsic_info::index_map
1607 * Generally instructions that take a offset src argument, can encode
1608 * a constant 'base' value which is added to the offset.
1610 NIR_INTRINSIC_BASE
= 1,
1613 * For store instructions, a writemask for the store.
1615 NIR_INTRINSIC_WRMASK
,
1618 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1620 NIR_INTRINSIC_STREAM_ID
,
1623 * The clip-plane id for load_user_clip_plane intrinsic.
1625 NIR_INTRINSIC_UCP_ID
,
1628 * The amount of data, starting from BASE, that this instruction may
1629 * access. This is used to provide bounds if the offset is not constant.
1631 NIR_INTRINSIC_RANGE
,
1634 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1636 NIR_INTRINSIC_DESC_SET
,
1639 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1641 NIR_INTRINSIC_BINDING
,
1646 NIR_INTRINSIC_COMPONENT
,
1649 * Interpolation mode (only meaningful for FS inputs).
1651 NIR_INTRINSIC_INTERP_MODE
,
1654 * A binary nir_op to use when performing a reduction or scan operation
1656 NIR_INTRINSIC_REDUCTION_OP
,
1659 * Cluster size for reduction operations
1661 NIR_INTRINSIC_CLUSTER_SIZE
,
1664 * Parameter index for a load_param intrinsic
1666 NIR_INTRINSIC_PARAM_IDX
,
1669 * Image dimensionality for image intrinsics
1671 * One of GLSL_SAMPLER_DIM_*
1673 NIR_INTRINSIC_IMAGE_DIM
,
1676 * Non-zero if we are accessing an array image
1678 NIR_INTRINSIC_IMAGE_ARRAY
,
1681 * Image format for image intrinsics
1683 NIR_INTRINSIC_FORMAT
,
1686 * Access qualifiers for image and memory access intrinsics
1688 NIR_INTRINSIC_ACCESS
,
1691 * Alignment for offsets and addresses
1693 * These two parameters, specify an alignment in terms of a multiplier and
1694 * an offset. The offset or address parameter X of the intrinsic is
1695 * guaranteed to satisfy the following:
1697 * (X - align_offset) % align_mul == 0
1699 NIR_INTRINSIC_ALIGN_MUL
,
1700 NIR_INTRINSIC_ALIGN_OFFSET
,
1703 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1705 NIR_INTRINSIC_DESC_TYPE
,
1708 * The nir_alu_type of a uniform/input/output
1713 * The swizzle mask for the instructions
1714 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1716 NIR_INTRINSIC_SWIZZLE_MASK
,
1718 /* Separate source/dest access flags for copies */
1719 NIR_INTRINSIC_SRC_ACCESS
,
1720 NIR_INTRINSIC_DST_ACCESS
,
1722 /* Driver location for nir_load_patch_location_ir3 */
1723 NIR_INTRINSIC_DRIVER_LOCATION
,
1726 * Mask of nir_memory_semantics, includes ordering and visibility.
1728 NIR_INTRINSIC_MEMORY_SEMANTICS
,
1731 * Mask of nir_variable_modes affected by the memory operation.
1733 NIR_INTRINSIC_MEMORY_MODES
,
1736 * Value of nir_scope.
1738 NIR_INTRINSIC_MEMORY_SCOPE
,
1741 * Value of nir_scope.
1743 NIR_INTRINSIC_EXECUTION_SCOPE
,
1745 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1747 } nir_intrinsic_index_flag
;
1749 #define NIR_INTRINSIC_MAX_INPUTS 5
1754 uint8_t num_srcs
; /** < number of register/SSA inputs */
1756 /** number of components of each input register
1758 * If this value is 0, the number of components is given by the
1759 * num_components field of nir_intrinsic_instr. If this value is -1, the
1760 * intrinsic consumes however many components are provided and it is not
1763 int8_t src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1767 /** number of components of the output register
1769 * If this value is 0, the number of components is given by the
1770 * num_components field of nir_intrinsic_instr.
1772 uint8_t dest_components
;
1774 /** bitfield of legal bit sizes */
1775 uint8_t dest_bit_sizes
;
1777 /** the number of constant indices used by the intrinsic */
1778 uint8_t num_indices
;
1780 /** indicates the usage of intr->const_index[n] */
1781 uint8_t index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1783 /** semantic flags for calls to this intrinsic */
1784 nir_intrinsic_semantic_flag flags
;
1785 } nir_intrinsic_info
;
1787 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1789 static inline unsigned
1790 nir_intrinsic_src_components(const nir_intrinsic_instr
*intr
, unsigned srcn
)
1792 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1793 assert(srcn
< info
->num_srcs
);
1794 if (info
->src_components
[srcn
] > 0)
1795 return info
->src_components
[srcn
];
1796 else if (info
->src_components
[srcn
] == 0)
1797 return intr
->num_components
;
1799 return nir_src_num_components(intr
->src
[srcn
]);
1802 static inline unsigned
1803 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1805 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1806 if (!info
->has_dest
)
1808 else if (info
->dest_components
)
1809 return info
->dest_components
;
1811 return intr
->num_components
;
1815 * Helper to copy const_index[] from src to dst, without assuming they
1819 nir_intrinsic_copy_const_indices(nir_intrinsic_instr
*dst
, nir_intrinsic_instr
*src
)
1821 if (src
->intrinsic
== dst
->intrinsic
) {
1822 memcpy(dst
->const_index
, src
->const_index
, sizeof(dst
->const_index
));
1826 const nir_intrinsic_info
*src_info
= &nir_intrinsic_infos
[src
->intrinsic
];
1827 const nir_intrinsic_info
*dst_info
= &nir_intrinsic_infos
[dst
->intrinsic
];
1829 for (unsigned i
= 0; i
< NIR_INTRINSIC_NUM_INDEX_FLAGS
; i
++) {
1830 if (src_info
->index_map
[i
] == 0)
1833 /* require that dst instruction also uses the same const_index[]: */
1834 assert(dst_info
->index_map
[i
] > 0);
1836 dst
->const_index
[dst_info
->index_map
[i
] - 1] =
1837 src
->const_index
[src_info
->index_map
[i
] - 1];
1841 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1842 static inline type \
1843 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1845 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1846 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1847 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1849 static inline void \
1850 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1852 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1853 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1854 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1857 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1858 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1859 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1860 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1861 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1862 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1863 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1864 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1865 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1866 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1867 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1868 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1869 INTRINSIC_IDX_ACCESSORS(image_dim
, IMAGE_DIM
, enum glsl_sampler_dim
)
1870 INTRINSIC_IDX_ACCESSORS(image_array
, IMAGE_ARRAY
, bool)
1871 INTRINSIC_IDX_ACCESSORS(access
, ACCESS
, enum gl_access_qualifier
)
1872 INTRINSIC_IDX_ACCESSORS(src_access
, SRC_ACCESS
, enum gl_access_qualifier
)
1873 INTRINSIC_IDX_ACCESSORS(dst_access
, DST_ACCESS
, enum gl_access_qualifier
)
1874 INTRINSIC_IDX_ACCESSORS(format
, FORMAT
, enum pipe_format
)
1875 INTRINSIC_IDX_ACCESSORS(align_mul
, ALIGN_MUL
, unsigned)
1876 INTRINSIC_IDX_ACCESSORS(align_offset
, ALIGN_OFFSET
, unsigned)
1877 INTRINSIC_IDX_ACCESSORS(desc_type
, DESC_TYPE
, unsigned)
1878 INTRINSIC_IDX_ACCESSORS(type
, TYPE
, nir_alu_type
)
1879 INTRINSIC_IDX_ACCESSORS(swizzle_mask
, SWIZZLE_MASK
, unsigned)
1880 INTRINSIC_IDX_ACCESSORS(driver_location
, DRIVER_LOCATION
, unsigned)
1881 INTRINSIC_IDX_ACCESSORS(memory_semantics
, MEMORY_SEMANTICS
, nir_memory_semantics
)
1882 INTRINSIC_IDX_ACCESSORS(memory_modes
, MEMORY_MODES
, nir_variable_mode
)
1883 INTRINSIC_IDX_ACCESSORS(memory_scope
, MEMORY_SCOPE
, nir_scope
)
1884 INTRINSIC_IDX_ACCESSORS(execution_scope
, EXECUTION_SCOPE
, nir_scope
)
1887 nir_intrinsic_set_align(nir_intrinsic_instr
*intrin
,
1888 unsigned align_mul
, unsigned align_offset
)
1890 assert(util_is_power_of_two_nonzero(align_mul
));
1891 assert(align_offset
< align_mul
);
1892 nir_intrinsic_set_align_mul(intrin
, align_mul
);
1893 nir_intrinsic_set_align_offset(intrin
, align_offset
);
1896 /** Returns a simple alignment for a load/store intrinsic offset
1898 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1899 * and ALIGN_OFFSET parameters, this helper takes both into account and
1900 * provides a single simple alignment parameter. The offset X is guaranteed
1901 * to satisfy X % align == 0.
1903 static inline unsigned
1904 nir_intrinsic_align(const nir_intrinsic_instr
*intrin
)
1906 const unsigned align_mul
= nir_intrinsic_align_mul(intrin
);
1907 const unsigned align_offset
= nir_intrinsic_align_offset(intrin
);
1908 assert(align_offset
< align_mul
);
1909 return align_offset
? 1 << (ffs(align_offset
) - 1) : align_mul
;
1913 nir_image_intrinsic_coord_components(const nir_intrinsic_instr
*instr
);
1915 /* Converts a image_deref_* intrinsic into a image_* one */
1916 void nir_rewrite_image_intrinsic(nir_intrinsic_instr
*instr
,
1917 nir_ssa_def
*handle
, bool bindless
);
1919 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1921 nir_intrinsic_can_reorder(nir_intrinsic_instr
*instr
)
1923 if (instr
->intrinsic
== nir_intrinsic_load_deref
||
1924 instr
->intrinsic
== nir_intrinsic_load_ssbo
||
1925 instr
->intrinsic
== nir_intrinsic_bindless_image_load
||
1926 instr
->intrinsic
== nir_intrinsic_image_deref_load
||
1927 instr
->intrinsic
== nir_intrinsic_image_load
) {
1928 return nir_intrinsic_access(instr
) & ACCESS_CAN_REORDER
;
1930 const nir_intrinsic_info
*info
=
1931 &nir_intrinsic_infos
[instr
->intrinsic
];
1932 return (info
->flags
& NIR_INTRINSIC_CAN_ELIMINATE
) &&
1933 (info
->flags
& NIR_INTRINSIC_CAN_REORDER
);
1938 * \group texture information
1940 * This gives semantic information about textures which is useful to the
1941 * frontend, the backend, and lowering passes, but not the optimizer.
1946 nir_tex_src_projector
,
1947 nir_tex_src_comparator
, /* shadow comparator */
1951 nir_tex_src_min_lod
,
1952 nir_tex_src_ms_index
, /* MSAA sample index */
1953 nir_tex_src_ms_mcs
, /* MSAA compression value */
1956 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
1957 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
1958 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
1959 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
1960 nir_tex_src_texture_handle
, /* < bindless texture handle */
1961 nir_tex_src_sampler_handle
, /* < bindless sampler handle */
1962 nir_tex_src_plane
, /* < selects plane for planar textures */
1963 nir_num_tex_src_types
1968 nir_tex_src_type src_type
;
1972 nir_texop_tex
, /**< Regular texture look-up */
1973 nir_texop_txb
, /**< Texture look-up with LOD bias */
1974 nir_texop_txl
, /**< Texture look-up with explicit LOD */
1975 nir_texop_txd
, /**< Texture look-up with partial derivatives */
1976 nir_texop_txf
, /**< Texel fetch with explicit LOD */
1977 nir_texop_txf_ms
, /**< Multisample texture fetch */
1978 nir_texop_txf_ms_fb
, /**< Multisample texture fetch from framebuffer */
1979 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
1980 nir_texop_txs
, /**< Texture size */
1981 nir_texop_lod
, /**< Texture lod query */
1982 nir_texop_tg4
, /**< Texture gather */
1983 nir_texop_query_levels
, /**< Texture levels query */
1984 nir_texop_texture_samples
, /**< Texture samples query */
1985 nir_texop_samples_identical
, /**< Query whether all samples are definitely
1988 nir_texop_tex_prefetch
, /**< Regular texture look-up, eligible for pre-dispatch */
1989 nir_texop_fragment_fetch
, /**< Multisample fragment color texture fetch */
1990 nir_texop_fragment_mask_fetch
,/**< Multisample fragment mask texture fetch */
1996 enum glsl_sampler_dim sampler_dim
;
1997 nir_alu_type dest_type
;
2002 unsigned num_srcs
, coord_components
;
2003 bool is_array
, is_shadow
;
2006 * If is_shadow is true, whether this is the old-style shadow that outputs 4
2007 * components or the new-style shadow that outputs 1 component.
2009 bool is_new_style_shadow
;
2011 /* gather component selector */
2012 unsigned component
: 2;
2014 /* gather offsets */
2015 int8_t tg4_offsets
[4][2];
2017 /* True if the texture index or handle is not dynamically uniform */
2018 bool texture_non_uniform
;
2020 /* True if the sampler index or handle is not dynamically uniform */
2021 bool sampler_non_uniform
;
2023 /** The texture index
2025 * If this texture instruction has a nir_tex_src_texture_offset source,
2026 * then the texture index is given by texture_index + texture_offset.
2028 unsigned texture_index
;
2030 /** The sampler index
2032 * The following operations do not require a sampler and, as such, this
2033 * field should be ignored:
2035 * - nir_texop_txf_ms
2038 * - nir_texop_query_levels
2039 * - nir_texop_texture_samples
2040 * - nir_texop_samples_identical
2042 * If this texture instruction has a nir_tex_src_sampler_offset source,
2043 * then the sampler index is given by sampler_index + sampler_offset.
2045 unsigned sampler_index
;
2049 * Returns true if the texture operation requires a sampler as a general rule,
2050 * see the documentation of sampler_index.
2052 * Note that the specific hw/driver backend could require to a sampler
2053 * object/configuration packet in any case, for some other reason.
2056 nir_tex_instr_need_sampler(const nir_tex_instr
*instr
)
2058 switch (instr
->op
) {
2060 case nir_texop_txf_ms
:
2063 case nir_texop_query_levels
:
2064 case nir_texop_texture_samples
:
2065 case nir_texop_samples_identical
:
2072 static inline unsigned
2073 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
2075 switch (instr
->op
) {
2076 case nir_texop_txs
: {
2078 switch (instr
->sampler_dim
) {
2079 case GLSL_SAMPLER_DIM_1D
:
2080 case GLSL_SAMPLER_DIM_BUF
:
2083 case GLSL_SAMPLER_DIM_2D
:
2084 case GLSL_SAMPLER_DIM_CUBE
:
2085 case GLSL_SAMPLER_DIM_MS
:
2086 case GLSL_SAMPLER_DIM_RECT
:
2087 case GLSL_SAMPLER_DIM_EXTERNAL
:
2088 case GLSL_SAMPLER_DIM_SUBPASS
:
2091 case GLSL_SAMPLER_DIM_3D
:
2095 unreachable("not reached");
2097 if (instr
->is_array
)
2105 case nir_texop_texture_samples
:
2106 case nir_texop_query_levels
:
2107 case nir_texop_samples_identical
:
2108 case nir_texop_fragment_mask_fetch
:
2112 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
2119 /* Returns true if this texture operation queries something about the texture
2120 * rather than actually sampling it.
2123 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
2125 switch (instr
->op
) {
2128 case nir_texop_texture_samples
:
2129 case nir_texop_query_levels
:
2130 case nir_texop_txf_ms_mcs
:
2137 case nir_texop_txf_ms
:
2138 case nir_texop_txf_ms_fb
:
2142 unreachable("Invalid texture opcode");
2147 nir_tex_instr_has_implicit_derivative(const nir_tex_instr
*instr
)
2149 switch (instr
->op
) {
2159 static inline nir_alu_type
2160 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
2162 switch (instr
->src
[src
].src_type
) {
2163 case nir_tex_src_coord
:
2164 switch (instr
->op
) {
2166 case nir_texop_txf_ms
:
2167 case nir_texop_txf_ms_fb
:
2168 case nir_texop_txf_ms_mcs
:
2169 case nir_texop_samples_identical
:
2170 return nir_type_int
;
2173 return nir_type_float
;
2176 case nir_tex_src_lod
:
2177 switch (instr
->op
) {
2180 return nir_type_int
;
2183 return nir_type_float
;
2186 case nir_tex_src_projector
:
2187 case nir_tex_src_comparator
:
2188 case nir_tex_src_bias
:
2189 case nir_tex_src_min_lod
:
2190 case nir_tex_src_ddx
:
2191 case nir_tex_src_ddy
:
2192 return nir_type_float
;
2194 case nir_tex_src_offset
:
2195 case nir_tex_src_ms_index
:
2196 case nir_tex_src_plane
:
2197 return nir_type_int
;
2199 case nir_tex_src_ms_mcs
:
2200 case nir_tex_src_texture_deref
:
2201 case nir_tex_src_sampler_deref
:
2202 case nir_tex_src_texture_offset
:
2203 case nir_tex_src_sampler_offset
:
2204 case nir_tex_src_texture_handle
:
2205 case nir_tex_src_sampler_handle
:
2206 return nir_type_uint
;
2208 case nir_num_tex_src_types
:
2209 unreachable("nir_num_tex_src_types is not a valid source type");
2212 unreachable("Invalid texture source type");
2215 static inline unsigned
2216 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
2218 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
2219 return instr
->coord_components
;
2221 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
2222 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
2225 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
2226 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
2227 if (instr
->is_array
)
2228 return instr
->coord_components
- 1;
2230 return instr
->coord_components
;
2233 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
2234 * the offset, since a cube maps to a single face.
2236 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
2237 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2239 else if (instr
->is_array
)
2240 return instr
->coord_components
- 1;
2242 return instr
->coord_components
;
2249 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
2251 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
2252 if (instr
->src
[i
].src_type
== type
)
2258 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
2259 nir_tex_src_type src_type
,
2262 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
2264 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
);
2271 nir_const_value value
[];
2272 } nir_load_const_instr
;
2275 /** Return from a function
2277 * This instruction is a classic function return. It jumps to
2278 * nir_function_impl::end_block. No return value is provided in this
2279 * instruction. Instead, the function is expected to write any return
2280 * data to a deref passed in from the caller.
2284 /** Break out of the inner-most loop
2286 * This has the same semantics as C's "break" statement.
2290 /** Jump back to the top of the inner-most loop
2292 * This has the same semantics as C's "continue" statement assuming that a
2293 * NIR loop is implemented as "while (1) { body }".
2303 /* creates a new SSA variable in an undefined state */
2308 } nir_ssa_undef_instr
;
2311 struct exec_node node
;
2313 /* The predecessor block corresponding to this source */
2314 struct nir_block
*pred
;
2319 #define nir_foreach_phi_src(phi_src, phi) \
2320 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
2321 #define nir_foreach_phi_src_safe(phi_src, phi) \
2322 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
2327 struct exec_list srcs
; /** < list of nir_phi_src */
2333 struct exec_node node
;
2336 } nir_parallel_copy_entry
;
2338 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
2339 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
2344 /* A list of nir_parallel_copy_entrys. The sources of all of the
2345 * entries are copied to the corresponding destinations "in parallel".
2346 * In other words, if we have two entries: a -> b and b -> a, the values
2349 struct exec_list entries
;
2350 } nir_parallel_copy_instr
;
2352 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
2353 type
, nir_instr_type_alu
)
2354 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
2355 type
, nir_instr_type_deref
)
2356 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
2357 type
, nir_instr_type_call
)
2358 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
2359 type
, nir_instr_type_jump
)
2360 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
2361 type
, nir_instr_type_tex
)
2362 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
2363 type
, nir_instr_type_intrinsic
)
2364 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
2365 type
, nir_instr_type_load_const
)
2366 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
2367 type
, nir_instr_type_ssa_undef
)
2368 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
2369 type
, nir_instr_type_phi
)
2370 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
2371 nir_parallel_copy_instr
, instr
,
2372 type
, nir_instr_type_parallel_copy
)
2375 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
2376 static inline type \
2377 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2379 assert(nir_src_is_const(src)); \
2380 nir_load_const_instr *load = \
2381 nir_instr_as_load_const(src.ssa->parent_instr); \
2382 assert(comp < load->def.num_components); \
2383 return nir_const_value_as_##suffix(load->value[comp], \
2384 load->def.bit_size); \
2387 static inline type \
2388 nir_src_as_##suffix(nir_src src) \
2390 assert(nir_src_num_components(src) == 1); \
2391 return nir_src_comp_as_##suffix(src, 0); \
2394 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2395 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint
)
2396 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2397 NIR_DEFINE_SRC_AS_CONST(double, float)
2399 #undef NIR_DEFINE_SRC_AS_CONST
2408 nir_ssa_scalar_is_const(nir_ssa_scalar s
)
2410 return s
.def
->parent_instr
->type
== nir_instr_type_load_const
;
2413 static inline nir_const_value
2414 nir_ssa_scalar_as_const_value(nir_ssa_scalar s
)
2416 assert(s
.comp
< s
.def
->num_components
);
2417 nir_load_const_instr
*load
= nir_instr_as_load_const(s
.def
->parent_instr
);
2418 return load
->value
[s
.comp
];
2421 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2422 static inline type \
2423 nir_ssa_scalar_as_##suffix(nir_ssa_scalar s) \
2425 return nir_const_value_as_##suffix( \
2426 nir_ssa_scalar_as_const_value(s), s.def->bit_size); \
2429 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2430 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint
)
2431 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2432 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2434 #undef NIR_DEFINE_SCALAR_AS_CONST
2437 nir_ssa_scalar_is_alu(nir_ssa_scalar s
)
2439 return s
.def
->parent_instr
->type
== nir_instr_type_alu
;
2442 static inline nir_op
2443 nir_ssa_scalar_alu_op(nir_ssa_scalar s
)
2445 return nir_instr_as_alu(s
.def
->parent_instr
)->op
;
2448 static inline nir_ssa_scalar
2449 nir_ssa_scalar_chase_alu_src(nir_ssa_scalar s
, unsigned alu_src_idx
)
2451 nir_ssa_scalar out
= { NULL
, 0 };
2453 nir_alu_instr
*alu
= nir_instr_as_alu(s
.def
->parent_instr
);
2454 assert(alu_src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
2456 /* Our component must be written */
2457 assert(s
.comp
< s
.def
->num_components
);
2458 assert(alu
->dest
.write_mask
& (1u << s
.comp
));
2460 assert(alu
->src
[alu_src_idx
].src
.is_ssa
);
2461 out
.def
= alu
->src
[alu_src_idx
].src
.ssa
;
2463 if (nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 0) {
2464 /* The ALU src is unsized so the source component follows the
2465 * destination component.
2467 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[s
.comp
];
2469 /* This is a sized source so all source components work together to
2470 * produce all the destination components. Since we need to return a
2471 * scalar, this only works if the source is a scalar.
2473 assert(nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 1);
2474 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[0];
2476 assert(out
.comp
< out
.def
->num_components
);
2485 * Control flow consists of a tree of control flow nodes, which include
2486 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2487 * instructions that always run start-to-finish. Each basic block also keeps
2488 * track of its successors (blocks which may run immediately after the current
2489 * block) and predecessors (blocks which could have run immediately before the
2490 * current block). Each function also has a start block and an end block which
2491 * all return statements point to (which is always empty). Together, all the
2492 * blocks with their predecessors and successors make up the control flow
2493 * graph (CFG) of the function. There are helpers that modify the tree of
2494 * control flow nodes while modifying the CFG appropriately; these should be
2495 * used instead of modifying the tree directly.
2502 nir_cf_node_function
2505 typedef struct nir_cf_node
{
2506 struct exec_node node
;
2507 nir_cf_node_type type
;
2508 struct nir_cf_node
*parent
;
2511 typedef struct nir_block
{
2512 nir_cf_node cf_node
;
2514 struct exec_list instr_list
; /** < list of nir_instr */
2516 /** generic block index; generated by nir_index_blocks */
2520 * Each block can only have up to 2 successors, so we put them in a simple
2521 * array - no need for anything more complicated.
2523 struct nir_block
*successors
[2];
2525 /* Set of nir_block predecessors in the CFG */
2526 struct set
*predecessors
;
2529 * this node's immediate dominator in the dominance tree - set to NULL for
2532 struct nir_block
*imm_dom
;
2534 /* This node's children in the dominance tree */
2535 unsigned num_dom_children
;
2536 struct nir_block
**dom_children
;
2538 /* Set of nir_blocks on the dominance frontier of this block */
2539 struct set
*dom_frontier
;
2542 * These two indices have the property that dom_{pre,post}_index for each
2543 * child of this block in the dominance tree will always be between
2544 * dom_pre_index and dom_post_index for this block, which makes testing if
2545 * a given block is dominated by another block an O(1) operation.
2547 int16_t dom_pre_index
, dom_post_index
;
2549 /* live in and out for this block; used for liveness analysis */
2550 BITSET_WORD
*live_in
;
2551 BITSET_WORD
*live_out
;
2555 nir_block_is_reachable(nir_block
*b
)
2557 /* See also nir_block_dominates */
2558 return b
->dom_post_index
!= -1;
2561 static inline nir_instr
*
2562 nir_block_first_instr(nir_block
*block
)
2564 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
2565 return exec_node_data(nir_instr
, head
, node
);
2568 static inline nir_instr
*
2569 nir_block_last_instr(nir_block
*block
)
2571 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
2572 return exec_node_data(nir_instr
, tail
, node
);
2576 nir_block_ends_in_jump(nir_block
*block
)
2578 return !exec_list_is_empty(&block
->instr_list
) &&
2579 nir_block_last_instr(block
)->type
== nir_instr_type_jump
;
2582 #define nir_foreach_instr(instr, block) \
2583 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
2584 #define nir_foreach_instr_reverse(instr, block) \
2585 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
2586 #define nir_foreach_instr_safe(instr, block) \
2587 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
2588 #define nir_foreach_instr_reverse_safe(instr, block) \
2589 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
2592 nir_selection_control_none
= 0x0,
2593 nir_selection_control_flatten
= 0x1,
2594 nir_selection_control_dont_flatten
= 0x2,
2595 } nir_selection_control
;
2597 typedef struct nir_if
{
2598 nir_cf_node cf_node
;
2600 nir_selection_control control
;
2602 struct exec_list then_list
; /** < list of nir_cf_node */
2603 struct exec_list else_list
; /** < list of nir_cf_node */
2609 /** Instruction that generates nif::condition. */
2610 nir_instr
*conditional_instr
;
2612 /** Block within ::nif that has the break instruction. */
2613 nir_block
*break_block
;
2615 /** Last block for the then- or else-path that does not contain the break. */
2616 nir_block
*continue_from_block
;
2618 /** True when ::break_block is in the else-path of ::nif. */
2619 bool continue_from_then
;
2622 /* This is true if the terminators exact trip count is unknown. For
2625 * for (int i = 0; i < imin(x, 4); i++)
2628 * Here loop analysis would have set a max_trip_count of 4 however we dont
2629 * know for sure that this is the exact trip count.
2631 bool exact_trip_count_unknown
;
2633 struct list_head loop_terminator_link
;
2634 } nir_loop_terminator
;
2637 /* Estimated cost (in number of instructions) of the loop */
2638 unsigned instr_cost
;
2640 /* Guessed trip count based on array indexing */
2641 unsigned guessed_trip_count
;
2643 /* Maximum number of times the loop is run (if known) */
2644 unsigned max_trip_count
;
2646 /* Do we know the exact number of times the loop will be run */
2647 bool exact_trip_count_known
;
2649 /* Unroll the loop regardless of its size */
2652 /* Does the loop contain complex loop terminators, continues or other
2653 * complex behaviours? If this is true we can't rely on
2654 * loop_terminator_list to be complete or accurate.
2658 nir_loop_terminator
*limiting_terminator
;
2660 /* A list of loop_terminators terminating this loop. */
2661 struct list_head loop_terminator_list
;
2665 nir_loop_control_none
= 0x0,
2666 nir_loop_control_unroll
= 0x1,
2667 nir_loop_control_dont_unroll
= 0x2,
2671 nir_cf_node cf_node
;
2673 struct exec_list body
; /** < list of nir_cf_node */
2675 nir_loop_info
*info
;
2676 nir_loop_control control
;
2677 bool partially_unrolled
;
2681 * Various bits of metadata that can may be created or required by
2682 * optimization and analysis passes
2685 nir_metadata_none
= 0x0,
2687 /** Indicates that nir_block::index values are valid.
2689 * The start block has index 0 and they increase through a natural walk of
2690 * the CFG. nir_function_impl::num_blocks is the number of blocks and
2691 * every block index is in the range [0, nir_function_impl::num_blocks].
2693 * A pass can preserve this metadata type if it doesn't touch the CFG.
2695 nir_metadata_block_index
= 0x1,
2697 /** Indicates that block dominance information is valid
2701 * - nir_block::num_dom_children
2702 * - nir_block::dom_children
2703 * - nir_block::dom_frontier
2704 * - nir_block::dom_pre_index
2705 * - nir_block::dom_post_index
2707 * A pass can preserve this metadata type if it doesn't touch the CFG.
2709 nir_metadata_dominance
= 0x2,
2711 /** Indicates that SSA def data-flow liveness information is valid
2715 * - nir_ssa_def::live_index
2716 * - nir_block::live_in
2717 * - nir_block::live_out
2719 * A pass can preserve this metadata type if it never adds or removes any
2720 * SSA defs (most passes shouldn't preserve this metadata type).
2722 nir_metadata_live_ssa_defs
= 0x4,
2724 /** A dummy metadata value to track when a pass forgot to call
2725 * nir_metadata_preserve.
2727 * A pass should always clear this value even if it doesn't make any
2728 * progress to indicate that it thought about preserving metadata.
2730 nir_metadata_not_properly_reset
= 0x8,
2732 /** Indicates that loop analysis information is valid.
2734 * This includes everything pointed to by nir_loop::info.
2736 * A pass can preserve this metadata type if it is guaranteed to not affect
2737 * any loop metadata. However, since loop metadata includes things like
2738 * loop counts which depend on arithmetic in the loop, this is very hard to
2739 * determine. Most passes shouldn't preserve this metadata type.
2741 nir_metadata_loop_analysis
= 0x10,
2745 * This includes all nir_metadata flags except not_properly_reset. Passes
2746 * which do not change the shader in any way should call
2748 * nir_metadata_preserve(impl, nir_metadata_all);
2750 nir_metadata_all
= ~nir_metadata_not_properly_reset
,
2754 nir_cf_node cf_node
;
2756 /** pointer to the function of which this is an implementation */
2757 struct nir_function
*function
;
2759 struct exec_list body
; /** < list of nir_cf_node */
2761 nir_block
*end_block
;
2763 /** list for all local variables in the function */
2764 struct exec_list locals
;
2766 /** list of local registers in the function */
2767 struct exec_list registers
;
2769 /** next available local register index */
2772 /** next available SSA value index */
2775 /* total number of basic blocks, only valid when block_index_dirty = false */
2776 unsigned num_blocks
;
2778 /** True if this nir_function_impl uses structured control-flow
2780 * Structured nir_function_impls have different validation rules.
2784 nir_metadata valid_metadata
;
2785 } nir_function_impl
;
2787 #define nir_foreach_function_temp_variable(var, impl) \
2788 foreach_list_typed(nir_variable, var, node, &(impl)->locals)
2790 #define nir_foreach_function_temp_variable_safe(var, impl) \
2791 foreach_list_typed_safe(nir_variable, var, node, &(impl)->locals)
2793 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2794 nir_start_block(nir_function_impl
*impl
)
2796 return (nir_block
*) impl
->body
.head_sentinel
.next
;
2799 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2800 nir_impl_last_block(nir_function_impl
*impl
)
2802 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
2805 static inline nir_cf_node
*
2806 nir_cf_node_next(nir_cf_node
*node
)
2808 struct exec_node
*next
= exec_node_get_next(&node
->node
);
2809 if (exec_node_is_tail_sentinel(next
))
2812 return exec_node_data(nir_cf_node
, next
, node
);
2815 static inline nir_cf_node
*
2816 nir_cf_node_prev(nir_cf_node
*node
)
2818 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
2819 if (exec_node_is_head_sentinel(prev
))
2822 return exec_node_data(nir_cf_node
, prev
, node
);
2826 nir_cf_node_is_first(const nir_cf_node
*node
)
2828 return exec_node_is_head_sentinel(node
->node
.prev
);
2832 nir_cf_node_is_last(const nir_cf_node
*node
)
2834 return exec_node_is_tail_sentinel(node
->node
.next
);
2837 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
2838 type
, nir_cf_node_block
)
2839 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
2840 type
, nir_cf_node_if
)
2841 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
2842 type
, nir_cf_node_loop
)
2843 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
2844 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
2846 static inline nir_block
*
2847 nir_if_first_then_block(nir_if
*if_stmt
)
2849 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
2850 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2853 static inline nir_block
*
2854 nir_if_last_then_block(nir_if
*if_stmt
)
2856 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
2857 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2860 static inline nir_block
*
2861 nir_if_first_else_block(nir_if
*if_stmt
)
2863 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
2864 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2867 static inline nir_block
*
2868 nir_if_last_else_block(nir_if
*if_stmt
)
2870 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
2871 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2874 static inline nir_block
*
2875 nir_loop_first_block(nir_loop
*loop
)
2877 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
2878 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2881 static inline nir_block
*
2882 nir_loop_last_block(nir_loop
*loop
)
2884 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
2885 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2889 * Return true if this list of cf_nodes contains a single empty block.
2892 nir_cf_list_is_empty_block(struct exec_list
*cf_list
)
2894 if (exec_list_is_singular(cf_list
)) {
2895 struct exec_node
*head
= exec_list_get_head(cf_list
);
2897 nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2898 return exec_list_is_empty(&block
->instr_list
);
2904 uint8_t num_components
;
2908 typedef struct nir_function
{
2909 struct exec_node node
;
2912 struct nir_shader
*shader
;
2914 unsigned num_params
;
2915 nir_parameter
*params
;
2917 /** The implementation of this function.
2919 * If the function is only declared and not implemented, this is NULL.
2921 nir_function_impl
*impl
;
2927 nir_lower_imul64
= (1 << 0),
2928 nir_lower_isign64
= (1 << 1),
2929 /** Lower all int64 modulus and division opcodes */
2930 nir_lower_divmod64
= (1 << 2),
2931 /** Lower all 64-bit umul_high and imul_high opcodes */
2932 nir_lower_imul_high64
= (1 << 3),
2933 nir_lower_mov64
= (1 << 4),
2934 nir_lower_icmp64
= (1 << 5),
2935 nir_lower_iadd64
= (1 << 6),
2936 nir_lower_iabs64
= (1 << 7),
2937 nir_lower_ineg64
= (1 << 8),
2938 nir_lower_logic64
= (1 << 9),
2939 nir_lower_minmax64
= (1 << 10),
2940 nir_lower_shift64
= (1 << 11),
2941 nir_lower_imul_2x32_64
= (1 << 12),
2942 nir_lower_extract64
= (1 << 13),
2943 nir_lower_ufind_msb64
= (1 << 14),
2944 } nir_lower_int64_options
;
2947 nir_lower_drcp
= (1 << 0),
2948 nir_lower_dsqrt
= (1 << 1),
2949 nir_lower_drsq
= (1 << 2),
2950 nir_lower_dtrunc
= (1 << 3),
2951 nir_lower_dfloor
= (1 << 4),
2952 nir_lower_dceil
= (1 << 5),
2953 nir_lower_dfract
= (1 << 6),
2954 nir_lower_dround_even
= (1 << 7),
2955 nir_lower_dmod
= (1 << 8),
2956 nir_lower_dsub
= (1 << 9),
2957 nir_lower_ddiv
= (1 << 10),
2958 nir_lower_fp64_full_software
= (1 << 11),
2959 } nir_lower_doubles_options
;
2962 nir_divergence_single_prim_per_subgroup
= (1 << 0),
2963 nir_divergence_single_patch_per_tcs_subgroup
= (1 << 1),
2964 nir_divergence_single_patch_per_tes_subgroup
= (1 << 2),
2965 nir_divergence_view_index_uniform
= (1 << 3),
2966 } nir_divergence_options
;
2968 typedef struct nir_shader_compiler_options
{
2974 /** Lowers flrp when it does not support doubles */
2981 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2982 bool lower_bitfield_extract
;
2983 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
2984 bool lower_bitfield_extract_to_shifts
;
2985 /** Lowers bitfield_insert to bfi/bfm */
2986 bool lower_bitfield_insert
;
2987 /** Lowers bitfield_insert to compares, and shifts. */
2988 bool lower_bitfield_insert_to_shifts
;
2989 /** Lowers bitfield_insert to bfm/bitfield_select. */
2990 bool lower_bitfield_insert_to_bitfield_select
;
2991 /** Lowers bitfield_reverse to shifts. */
2992 bool lower_bitfield_reverse
;
2993 /** Lowers bit_count to shifts. */
2994 bool lower_bit_count
;
2995 /** Lowers ifind_msb to compare and ufind_msb */
2996 bool lower_ifind_msb
;
2997 /** Lowers find_lsb to ufind_msb and logic ops */
2998 bool lower_find_lsb
;
2999 bool lower_uadd_carry
;
3000 bool lower_usub_borrow
;
3001 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
3002 bool lower_mul_high
;
3003 /** lowers fneg and ineg to fsub and isub. */
3005 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
3008 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
3011 /* lower fall_equalN/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
3012 bool lower_vector_cmp
;
3014 /** enables rules to lower idiv by power-of-two: */
3017 /** enable rules to avoid bit ops */
3020 /** enables rules to lower isign to imin+imax */
3023 /** enables rules to lower fsign to fsub and flt */
3026 /* lower fdph to fdot4 */
3029 /** lower fdot to fmul and fsum/fadd. */
3032 /* Does the native fdot instruction replicate its result for four
3033 * components? If so, then opt_algebraic_late will turn all fdotN
3034 * instructions into fdot_replicatedN instructions.
3036 bool fdot_replicates
;
3038 /** lowers ffloor to fsub+ffract: */
3041 /** lowers ffract to fsub+ffloor: */
3044 /** lowers fceil to fneg+ffloor+fneg: */
3051 bool lower_pack_half_2x16
;
3052 bool lower_pack_unorm_2x16
;
3053 bool lower_pack_snorm_2x16
;
3054 bool lower_pack_unorm_4x8
;
3055 bool lower_pack_snorm_4x8
;
3056 bool lower_unpack_half_2x16
;
3057 bool lower_unpack_unorm_2x16
;
3058 bool lower_unpack_snorm_2x16
;
3059 bool lower_unpack_unorm_4x8
;
3060 bool lower_unpack_snorm_4x8
;
3062 bool lower_pack_split
;
3064 bool lower_extract_byte
;
3065 bool lower_extract_word
;
3067 bool lower_all_io_to_temps
;
3068 bool lower_all_io_to_elements
;
3070 /* Indicates that the driver only has zero-based vertex id */
3071 bool vertex_id_zero_based
;
3074 * If enabled, gl_BaseVertex will be lowered as:
3075 * is_indexed_draw (~0/0) & firstvertex
3077 bool lower_base_vertex
;
3080 * If enabled, gl_HelperInvocation will be lowered as:
3082 * !((1 << sample_id) & sample_mask_in))
3084 * This depends on some possibly hw implementation details, which may
3085 * not be true for all hw. In particular that the FS is only executed
3086 * for covered samples or for helper invocations. So, do not blindly
3087 * enable this option.
3089 * Note: See also issue #22 in ARB_shader_image_load_store
3091 bool lower_helper_invocation
;
3094 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
3096 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
3097 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
3099 bool optimize_sample_mask_in
;
3101 bool lower_cs_local_index_from_id
;
3102 bool lower_cs_local_id_from_index
;
3104 bool lower_device_index_to_zero
;
3106 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
3107 bool lower_wpos_pntc
;
3110 * Set if nir_op_[iu]hadd and nir_op_[iu]rhadd instructions should be
3111 * lowered to simple arithmetic.
3113 * If this flag is set, the lowering will be applied to all bit-sizes of
3114 * these instructions.
3116 * \sa ::lower_hadd64
3121 * Set if only 64-bit nir_op_[iu]hadd and nir_op_[iu]rhadd instructions
3122 * should be lowered to simple arithmetic.
3124 * If this flag is set, the lowering will be applied to only 64-bit
3125 * versions of these instructions.
3132 * Set if nir_op_add_sat and nir_op_usub_sat should be lowered to simple
3135 * If this flag is set, the lowering will be applied to all bit-sizes of
3136 * these instructions.
3138 * \sa ::lower_usub_sat64
3143 * Set if only 64-bit nir_op_usub_sat should be lowered to simple
3146 * \sa ::lower_add_sat
3148 bool lower_usub_sat64
;
3151 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
3152 * for IO purposes and would prefer loads/stores be vectorized.
3155 bool lower_to_scalar
;
3158 * Whether nir_opt_vectorize should only create 16-bit 2D vectors.
3160 bool vectorize_vec2_16bit
;
3163 * Should the linker unify inputs_read/outputs_written between adjacent
3164 * shader stages which are linked into a single program?
3166 bool unify_interfaces
;
3169 * Should nir_lower_io() create load_interpolated_input intrinsics?
3171 * If not, it generates regular load_input intrinsics and interpolation
3172 * information must be inferred from the list of input nir_variables.
3174 bool use_interpolated_input_intrinsics
;
3176 /* Lowers when 32x32->64 bit multiplication is not supported */
3177 bool lower_mul_2x32_64
;
3179 /* Lowers when rotate instruction is not supported */
3183 * Backend supports imul24, and would like to use it (when possible)
3184 * for address/offset calculation. If true, driver should call
3185 * nir_lower_amul(). (If not set, amul will automatically be lowered
3190 /** Backend supports umul24, if not set umul24 will automatically be lowered
3191 * to imul with masked inputs */
3194 /** Backend supports umad24, if not set umad24 will automatically be lowered
3195 * to imul with masked inputs and iadd */
3198 /* Whether to generate only scoped_barrier intrinsics instead of the set of
3199 * memory and control barrier intrinsics based on GLSL.
3201 bool use_scoped_barrier
;
3204 * Is this the Intel vec4 backend?
3206 * Used to inhibit algebraic optimizations that are known to be harmful on
3207 * the Intel vec4 backend. This is generally applicable to any
3208 * optimization that might cause more immediate values to be used in
3209 * 3-source (e.g., ffma and flrp) instructions.
3213 /** Lower nir_op_ibfe and nir_op_ubfe that have two constant sources. */
3214 bool lower_bfe_with_two_constants
;
3216 /** Whether 8-bit ALU is supported. */
3217 bool support_8bit_alu
;
3219 /** Whether 16-bit ALU is supported. */
3220 bool support_16bit_alu
;
3222 unsigned max_unroll_iterations
;
3224 nir_lower_int64_options lower_int64_options
;
3225 nir_lower_doubles_options lower_doubles_options
;
3226 } nir_shader_compiler_options
;
3228 typedef struct nir_shader
{
3229 /** list of uniforms (nir_variable) */
3230 struct exec_list variables
;
3232 /** Set of driver-specific options for the shader.
3234 * The memory for the options is expected to be kept in a single static
3235 * copy by the driver.
3237 const struct nir_shader_compiler_options
*options
;
3239 /** Various bits of compile-time information about a given shader */
3240 struct shader_info info
;
3242 struct exec_list functions
; /** < list of nir_function */
3245 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
3248 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
3250 /** Size in bytes of required scratch space */
3251 unsigned scratch_size
;
3253 /** Constant data associated with this shader.
3255 * Constant data is loaded through load_constant intrinsics (as compared to
3256 * the NIR load_const instructions which have the constant value inlined
3257 * into them). This is usually generated by nir_opt_large_constants (so
3258 * shaders don't have to load_const into a temporary array when they want
3259 * to indirect on a const array).
3261 void *constant_data
;
3262 /** Size of the constant data associated with the shader, in bytes */
3263 unsigned constant_data_size
;
3266 #define nir_foreach_function(func, shader) \
3267 foreach_list_typed(nir_function, func, node, &(shader)->functions)
3269 static inline nir_function_impl
*
3270 nir_shader_get_entrypoint(nir_shader
*shader
)
3272 nir_function
*func
= NULL
;
3274 nir_foreach_function(function
, shader
) {
3275 assert(func
== NULL
);
3276 if (function
->is_entrypoint
) {
3287 assert(func
->num_params
== 0);
3292 nir_shader
*nir_shader_create(void *mem_ctx
,
3293 gl_shader_stage stage
,
3294 const nir_shader_compiler_options
*options
,
3297 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
3299 void nir_reg_remove(nir_register
*reg
);
3301 /** Adds a variable to the appropriate list in nir_shader */
3302 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
3305 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
3307 assert(var
->data
.mode
== nir_var_function_temp
);
3308 exec_list_push_tail(&impl
->locals
, &var
->node
);
3311 /** creates a variable, sets a few defaults, and adds it to the list */
3312 nir_variable
*nir_variable_create(nir_shader
*shader
,
3313 nir_variable_mode mode
,
3314 const struct glsl_type
*type
,
3316 /** creates a local variable and adds it to the list */
3317 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
3318 const struct glsl_type
*type
,
3321 nir_variable
*nir_find_variable_with_location(nir_shader
*shader
,
3322 nir_variable_mode mode
,
3325 nir_variable
*nir_find_variable_with_driver_location(nir_shader
*shader
,
3326 nir_variable_mode mode
,
3329 /** creates a function and adds it to the shader's list of functions */
3330 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
3332 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
3333 /** creates a function_impl that isn't tied to any particular function */
3334 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
3336 nir_block
*nir_block_create(nir_shader
*shader
);
3337 nir_if
*nir_if_create(nir_shader
*shader
);
3338 nir_loop
*nir_loop_create(nir_shader
*shader
);
3340 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
3342 /** requests that the given pieces of metadata be generated */
3343 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
3344 /** dirties all but the preserved metadata */
3345 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
3346 /** Preserves all metadata for the given shader */
3347 void nir_shader_preserve_all_metadata(nir_shader
*shader
);
3349 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
3350 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
3352 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
3353 nir_deref_type deref_type
);
3355 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
3357 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
3358 unsigned num_components
,
3361 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
3362 nir_intrinsic_op op
);
3364 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
3365 nir_function
*callee
);
3367 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
3369 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
3371 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
3373 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
3374 unsigned num_components
,
3377 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
3380 * NIR Cursors and Instruction Insertion API
3383 * A tiny struct representing a point to insert/extract instructions or
3384 * control flow nodes. Helps reduce the combinatorial explosion of possible
3385 * points to insert/extract.
3387 * \sa nir_control_flow.h
3390 nir_cursor_before_block
,
3391 nir_cursor_after_block
,
3392 nir_cursor_before_instr
,
3393 nir_cursor_after_instr
,
3394 } nir_cursor_option
;
3397 nir_cursor_option option
;
3404 static inline nir_block
*
3405 nir_cursor_current_block(nir_cursor cursor
)
3407 if (cursor
.option
== nir_cursor_before_instr
||
3408 cursor
.option
== nir_cursor_after_instr
) {
3409 return cursor
.instr
->block
;
3411 return cursor
.block
;
3415 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
3417 static inline nir_cursor
3418 nir_before_block(nir_block
*block
)
3421 cursor
.option
= nir_cursor_before_block
;
3422 cursor
.block
= block
;
3426 static inline nir_cursor
3427 nir_after_block(nir_block
*block
)
3430 cursor
.option
= nir_cursor_after_block
;
3431 cursor
.block
= block
;
3435 static inline nir_cursor
3436 nir_before_instr(nir_instr
*instr
)
3439 cursor
.option
= nir_cursor_before_instr
;
3440 cursor
.instr
= instr
;
3444 static inline nir_cursor
3445 nir_after_instr(nir_instr
*instr
)
3448 cursor
.option
= nir_cursor_after_instr
;
3449 cursor
.instr
= instr
;
3453 static inline nir_cursor
3454 nir_after_block_before_jump(nir_block
*block
)
3456 nir_instr
*last_instr
= nir_block_last_instr(block
);
3457 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
3458 return nir_before_instr(last_instr
);
3460 return nir_after_block(block
);
3464 static inline nir_cursor
3465 nir_before_src(nir_src
*src
, bool is_if_condition
)
3467 if (is_if_condition
) {
3468 nir_block
*prev_block
=
3469 nir_cf_node_as_block(nir_cf_node_prev(&src
->parent_if
->cf_node
));
3470 assert(!nir_block_ends_in_jump(prev_block
));
3471 return nir_after_block(prev_block
);
3472 } else if (src
->parent_instr
->type
== nir_instr_type_phi
) {
3474 nir_phi_instr
*cond_phi
= nir_instr_as_phi(src
->parent_instr
);
3476 nir_foreach_phi_src(phi_src
, cond_phi
) {
3477 if (phi_src
->src
.ssa
== src
->ssa
) {
3484 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
3485 * to have a more specific name.
3487 nir_phi_src
*phi_src
= LIST_ENTRY(nir_phi_src
, src
, src
);
3488 return nir_after_block_before_jump(phi_src
->pred
);
3490 return nir_before_instr(src
->parent_instr
);
3494 static inline nir_cursor
3495 nir_before_cf_node(nir_cf_node
*node
)
3497 if (node
->type
== nir_cf_node_block
)
3498 return nir_before_block(nir_cf_node_as_block(node
));
3500 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
3503 static inline nir_cursor
3504 nir_after_cf_node(nir_cf_node
*node
)
3506 if (node
->type
== nir_cf_node_block
)
3507 return nir_after_block(nir_cf_node_as_block(node
));
3509 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
3512 static inline nir_cursor
3513 nir_after_phis(nir_block
*block
)
3515 nir_foreach_instr(instr
, block
) {
3516 if (instr
->type
!= nir_instr_type_phi
)
3517 return nir_before_instr(instr
);
3519 return nir_after_block(block
);
3522 static inline nir_cursor
3523 nir_after_cf_node_and_phis(nir_cf_node
*node
)
3525 if (node
->type
== nir_cf_node_block
)
3526 return nir_after_block(nir_cf_node_as_block(node
));
3528 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
3530 return nir_after_phis(block
);
3533 static inline nir_cursor
3534 nir_before_cf_list(struct exec_list
*cf_list
)
3536 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
3537 exec_list_get_head(cf_list
), node
);
3538 return nir_before_cf_node(first_node
);
3541 static inline nir_cursor
3542 nir_after_cf_list(struct exec_list
*cf_list
)
3544 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
3545 exec_list_get_tail(cf_list
), node
);
3546 return nir_after_cf_node(last_node
);
3550 * Insert a NIR instruction at the given cursor.
3552 * Note: This does not update the cursor.
3554 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
3557 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
3559 nir_instr_insert(nir_before_instr(instr
), before
);
3563 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
3565 nir_instr_insert(nir_after_instr(instr
), after
);
3569 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
3571 nir_instr_insert(nir_before_block(block
), before
);
3575 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
3577 nir_instr_insert(nir_after_block(block
), after
);
3581 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
3583 nir_instr_insert(nir_before_cf_node(node
), before
);
3587 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
3589 nir_instr_insert(nir_after_cf_node(node
), after
);
3593 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
3595 nir_instr_insert(nir_before_cf_list(list
), before
);
3599 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
3601 nir_instr_insert(nir_after_cf_list(list
), after
);
3604 void nir_instr_remove_v(nir_instr
*instr
);
3606 static inline nir_cursor
3607 nir_instr_remove(nir_instr
*instr
)
3610 nir_instr
*prev
= nir_instr_prev(instr
);
3612 cursor
= nir_after_instr(prev
);
3614 cursor
= nir_before_block(instr
->block
);
3616 nir_instr_remove_v(instr
);
3622 nir_ssa_def
*nir_instr_ssa_def(nir_instr
*instr
);
3624 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
3625 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
3626 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
3627 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
3629 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
3630 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
3631 bool nir_foreach_phi_src_leaving_block(nir_block
*instr
,
3632 nir_foreach_src_cb cb
,
3635 nir_const_value
*nir_src_as_const_value(nir_src src
);
3637 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
3638 static inline c_type * \
3639 nir_src_as_ ## name (nir_src src) \
3641 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
3642 ? cast_macro(src.ssa->parent_instr) : NULL; \
3645 NIR_SRC_AS_(alu_instr
, nir_alu_instr
, nir_instr_type_alu
, nir_instr_as_alu
)
3646 NIR_SRC_AS_(intrinsic
, nir_intrinsic_instr
,
3647 nir_instr_type_intrinsic
, nir_instr_as_intrinsic
)
3648 NIR_SRC_AS_(deref
, nir_deref_instr
, nir_instr_type_deref
, nir_instr_as_deref
)
3650 bool nir_src_is_dynamically_uniform(nir_src src
);
3651 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
3652 bool nir_instrs_equal(const nir_instr
*instr1
, const nir_instr
*instr2
);
3653 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
3654 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
3655 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
3656 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
3659 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
3660 unsigned num_components
, unsigned bit_size
,
3662 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
3663 unsigned num_components
, unsigned bit_size
,
3666 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
3667 const struct glsl_type
*type
,
3670 assert(glsl_type_is_vector_or_scalar(type
));
3671 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
3672 glsl_get_bit_size(type
), name
);
3674 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
3675 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
3676 nir_instr
*after_me
);
3678 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
3681 * finds the next basic block in source-code order, returns NULL if there is
3685 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
3687 /* Performs the opposite of nir_block_cf_tree_next() */
3689 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
3691 /* Gets the first block in a CF node in source-code order */
3693 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
3695 /* Gets the last block in a CF node in source-code order */
3697 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
3699 /* Gets the next block after a CF node in source-code order */
3701 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
3703 /* Macros for loops that visit blocks in source-code order */
3705 #define nir_foreach_block(block, impl) \
3706 for (nir_block *block = nir_start_block(impl); block != NULL; \
3707 block = nir_block_cf_tree_next(block))
3709 #define nir_foreach_block_safe(block, impl) \
3710 for (nir_block *block = nir_start_block(impl), \
3711 *next = nir_block_cf_tree_next(block); \
3713 block = next, next = nir_block_cf_tree_next(block))
3715 #define nir_foreach_block_reverse(block, impl) \
3716 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
3717 block = nir_block_cf_tree_prev(block))
3719 #define nir_foreach_block_reverse_safe(block, impl) \
3720 for (nir_block *block = nir_impl_last_block(impl), \
3721 *prev = nir_block_cf_tree_prev(block); \
3723 block = prev, prev = nir_block_cf_tree_prev(block))
3725 #define nir_foreach_block_in_cf_node(block, node) \
3726 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
3727 block != nir_cf_node_cf_tree_next(node); \
3728 block = nir_block_cf_tree_next(block))
3730 /* If the following CF node is an if, this function returns that if.
3731 * Otherwise, it returns NULL.
3733 nir_if
*nir_block_get_following_if(nir_block
*block
);
3735 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
3737 void nir_index_local_regs(nir_function_impl
*impl
);
3738 void nir_index_ssa_defs(nir_function_impl
*impl
);
3739 unsigned nir_index_instrs(nir_function_impl
*impl
);
3741 void nir_index_blocks(nir_function_impl
*impl
);
3743 unsigned nir_shader_index_vars(nir_shader
*shader
, nir_variable_mode modes
);
3744 unsigned nir_function_impl_index_vars(nir_function_impl
*impl
);
3746 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
3747 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
3748 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
3749 void nir_print_deref(const nir_deref_instr
*deref
, FILE *fp
);
3751 /** Shallow clone of a single ALU instruction. */
3752 nir_alu_instr
*nir_alu_instr_clone(nir_shader
*s
, const nir_alu_instr
*orig
);
3754 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
3755 nir_function_impl
*nir_function_impl_clone(nir_shader
*shader
,
3756 const nir_function_impl
*fi
);
3757 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
3758 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
3760 void nir_shader_replace(nir_shader
*dest
, nir_shader
*src
);
3762 void nir_shader_serialize_deserialize(nir_shader
*s
);
3765 void nir_validate_shader(nir_shader
*shader
, const char *when
);
3766 void nir_metadata_set_validation_flag(nir_shader
*shader
);
3767 void nir_metadata_check_validation_flag(nir_shader
*shader
);
3770 should_skip_nir(const char *name
)
3772 static const char *list
= NULL
;
3774 /* Comma separated list of names to skip. */
3775 list
= getenv("NIR_SKIP");
3783 return comma_separated_list_contains(list
, name
);
3787 should_clone_nir(void)
3789 static int should_clone
= -1;
3790 if (should_clone
< 0)
3791 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
3793 return should_clone
;
3797 should_serialize_deserialize_nir(void)
3799 static int test_serialize
= -1;
3800 if (test_serialize
< 0)
3801 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
3803 return test_serialize
;
3807 should_print_nir(void)
3809 static int should_print
= -1;
3810 if (should_print
< 0)
3811 should_print
= env_var_as_boolean("NIR_PRINT", false);
3813 return should_print
;
3816 static inline void nir_validate_shader(nir_shader
*shader
, const char *when
) { (void) shader
; (void)when
; }
3817 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
3818 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
3819 static inline bool should_skip_nir(UNUSED
const char *pass_name
) { return false; }
3820 static inline bool should_clone_nir(void) { return false; }
3821 static inline bool should_serialize_deserialize_nir(void) { return false; }
3822 static inline bool should_print_nir(void) { return false; }
3825 #define _PASS(pass, nir, do_pass) do { \
3826 if (should_skip_nir(#pass)) { \
3827 printf("skipping %s\n", #pass); \
3831 nir_validate_shader(nir, "after " #pass); \
3832 if (should_clone_nir()) { \
3833 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3834 nir_shader_replace(nir, clone); \
3836 if (should_serialize_deserialize_nir()) { \
3837 nir_shader_serialize_deserialize(nir); \
3841 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3842 nir_metadata_set_validation_flag(nir); \
3843 if (should_print_nir()) \
3844 printf("%s\n", #pass); \
3845 if (pass(nir, ##__VA_ARGS__)) { \
3847 if (should_print_nir()) \
3848 nir_print_shader(nir, stdout); \
3849 nir_metadata_check_validation_flag(nir); \
3853 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3854 if (should_print_nir()) \
3855 printf("%s\n", #pass); \
3856 pass(nir, ##__VA_ARGS__); \
3857 if (should_print_nir()) \
3858 nir_print_shader(nir, stdout); \
3861 #define NIR_SKIP(name) should_skip_nir(#name)
3863 /** An instruction filtering callback
3865 * Returns true if the instruction should be processed and false otherwise.
3867 typedef bool (*nir_instr_filter_cb
)(const nir_instr
*, const void *);
3869 /** A simple instruction lowering callback
3871 * Many instruction lowering passes can be written as a simple function which
3872 * takes an instruction as its input and returns a sequence of instructions
3873 * that implement the consumed instruction. This function type represents
3874 * such a lowering function. When called, a function with this prototype
3875 * should either return NULL indicating that no lowering needs to be done or
3876 * emit a sequence of instructions using the provided builder (whose cursor
3877 * will already be placed after the instruction to be lowered) and return the
3878 * resulting nir_ssa_def.
3880 typedef nir_ssa_def
*(*nir_lower_instr_cb
)(struct nir_builder
*,
3881 nir_instr
*, void *);
3884 * Special return value for nir_lower_instr_cb when some progress occurred
3885 * (like changing an input to the instr) that didn't result in a replacement
3886 * SSA def being generated.
3888 #define NIR_LOWER_INSTR_PROGRESS ((nir_ssa_def *)(uintptr_t)1)
3890 /** Iterate over all the instructions in a nir_function_impl and lower them
3891 * using the provided callbacks
3893 * This function implements the guts of a standard lowering pass for you. It
3894 * iterates over all of the instructions in a nir_function_impl and calls the
3895 * filter callback on each one. If the filter callback returns true, it then
3896 * calls the lowering call back on the instruction. (Splitting it this way
3897 * allows us to avoid some save/restore work for instructions we know won't be
3898 * lowered.) If the instruction is dead after the lowering is complete, it
3899 * will be removed. If new instructions are added, the lowering callback will
3900 * also be called on them in case multiple lowerings are required.
3902 * The metadata for the nir_function_impl will also be updated. If any blocks
3903 * are added (they cannot be removed), dominance and block indices will be
3906 bool nir_function_impl_lower_instructions(nir_function_impl
*impl
,
3907 nir_instr_filter_cb filter
,
3908 nir_lower_instr_cb lower
,
3910 bool nir_shader_lower_instructions(nir_shader
*shader
,
3911 nir_instr_filter_cb filter
,
3912 nir_lower_instr_cb lower
,
3915 void nir_calc_dominance_impl(nir_function_impl
*impl
);
3916 void nir_calc_dominance(nir_shader
*shader
);
3918 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
3919 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
3920 bool nir_block_is_unreachable(nir_block
*block
);
3922 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
3923 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
3925 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
3926 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
3928 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
3929 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
3931 int nir_gs_count_vertices(const nir_shader
*shader
);
3933 bool nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3934 bool nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3935 bool nir_split_var_copies(nir_shader
*shader
);
3936 bool nir_split_per_member_structs(nir_shader
*shader
);
3937 bool nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
);
3939 bool nir_lower_returns_impl(nir_function_impl
*impl
);
3940 bool nir_lower_returns(nir_shader
*shader
);
3942 void nir_inline_function_impl(struct nir_builder
*b
,
3943 const nir_function_impl
*impl
,
3944 nir_ssa_def
**params
);
3945 bool nir_inline_functions(nir_shader
*shader
);
3947 bool nir_propagate_invariant(nir_shader
*shader
);
3949 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
3950 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
3951 nir_intrinsic_instr
*copy
);
3952 bool nir_lower_var_copies(nir_shader
*shader
);
3954 void nir_fixup_deref_modes(nir_shader
*shader
);
3956 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
3959 nir_lower_direct_array_deref_of_vec_load
= (1 << 0),
3960 nir_lower_indirect_array_deref_of_vec_load
= (1 << 1),
3961 nir_lower_direct_array_deref_of_vec_store
= (1 << 2),
3962 nir_lower_indirect_array_deref_of_vec_store
= (1 << 3),
3963 } nir_lower_array_deref_of_vec_options
;
3965 bool nir_lower_array_deref_of_vec(nir_shader
*shader
, nir_variable_mode modes
,
3966 nir_lower_array_deref_of_vec_options options
);
3968 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
3970 bool nir_lower_locals_to_regs(nir_shader
*shader
);
3972 void nir_lower_io_to_temporaries(nir_shader
*shader
,
3973 nir_function_impl
*entrypoint
,
3974 bool outputs
, bool inputs
);
3976 bool nir_lower_vars_to_scratch(nir_shader
*shader
,
3977 nir_variable_mode modes
,
3979 glsl_type_size_align_func size_align
);
3981 void nir_lower_clip_halfz(nir_shader
*shader
);
3983 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
3985 void nir_gather_ssa_types(nir_function_impl
*impl
,
3986 BITSET_WORD
*float_types
,
3987 BITSET_WORD
*int_types
);
3989 void nir_assign_var_locations(nir_shader
*shader
, nir_variable_mode mode
,
3991 int (*type_size
)(const struct glsl_type
*, bool));
3993 /* Some helpers to do very simple linking */
3994 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3995 bool nir_remove_unused_io_vars(nir_shader
*shader
, nir_variable_mode mode
,
3996 uint64_t *used_by_other_stage
,
3997 uint64_t *used_by_other_stage_patches
);
3998 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
3999 bool default_to_smooth_interp
);
4000 void nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
);
4001 bool nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
);
4003 bool nir_lower_amul(nir_shader
*shader
,
4004 int (*type_size
)(const struct glsl_type
*, bool));
4006 void nir_assign_io_var_locations(nir_shader
*shader
,
4007 nir_variable_mode mode
,
4009 gl_shader_stage stage
);
4012 uint8_t num_linked_io_vars
;
4013 uint8_t num_linked_patch_io_vars
;
4014 } nir_linked_io_var_info
;
4016 nir_linked_io_var_info
4017 nir_assign_linked_io_var_locations(nir_shader
*producer
,
4018 nir_shader
*consumer
);
4021 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
4022 * to 32-bit operations. This is only valid for nir_var_shader_in/out
4025 nir_lower_io_lower_64bit_to_32
= (1 << 0),
4027 /* If set, this forces all non-flat fragment shader inputs to be
4028 * interpolated as if with the "sample" qualifier. This requires
4029 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
4031 nir_lower_io_force_sample_interpolation
= (1 << 1),
4032 } nir_lower_io_options
;
4033 bool nir_lower_io(nir_shader
*shader
,
4034 nir_variable_mode modes
,
4035 int (*type_size
)(const struct glsl_type
*, bool),
4036 nir_lower_io_options
);
4038 bool nir_io_add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
);
4041 nir_lower_vars_to_explicit_types(nir_shader
*shader
,
4042 nir_variable_mode modes
,
4043 glsl_type_size_align_func type_info
);
4047 * An address format which is a simple 32-bit global GPU address.
4049 nir_address_format_32bit_global
,
4052 * An address format which is a simple 64-bit global GPU address.
4054 nir_address_format_64bit_global
,
4057 * An address format which is a bounds-checked 64-bit global GPU address.
4059 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
4060 * address stored with the low bits in .x and high bits in .y, .z is a
4061 * size, and .w is an offset. When the final I/O operation is lowered, .w
4062 * is checked against .z and the operation is predicated on the result.
4064 nir_address_format_64bit_bounded_global
,
4067 * An address format which is comprised of a vec2 where the first
4068 * component is a buffer index and the second is an offset.
4070 nir_address_format_32bit_index_offset
,
4073 * An address format which is comprised of a vec3 where the first two
4074 * components specify the buffer and the third is an offset.
4076 nir_address_format_vec2_index_32bit_offset
,
4079 * An address format which is a simple 32-bit offset.
4081 nir_address_format_32bit_offset
,
4084 * An address format representing a purely logical addressing model. In
4085 * this model, all deref chains must be complete from the dereference
4086 * operation to the variable. Cast derefs are not allowed. These
4087 * addresses will be 32-bit scalars but the format is immaterial because
4088 * you can always chase the chain.
4090 nir_address_format_logical
,
4091 } nir_address_format
;
4093 static inline unsigned
4094 nir_address_format_bit_size(nir_address_format addr_format
)
4096 switch (addr_format
) {
4097 case nir_address_format_32bit_global
: return 32;
4098 case nir_address_format_64bit_global
: return 64;
4099 case nir_address_format_64bit_bounded_global
: return 32;
4100 case nir_address_format_32bit_index_offset
: return 32;
4101 case nir_address_format_vec2_index_32bit_offset
: return 32;
4102 case nir_address_format_32bit_offset
: return 32;
4103 case nir_address_format_logical
: return 32;
4105 unreachable("Invalid address format");
4108 static inline unsigned
4109 nir_address_format_num_components(nir_address_format addr_format
)
4111 switch (addr_format
) {
4112 case nir_address_format_32bit_global
: return 1;
4113 case nir_address_format_64bit_global
: return 1;
4114 case nir_address_format_64bit_bounded_global
: return 4;
4115 case nir_address_format_32bit_index_offset
: return 2;
4116 case nir_address_format_vec2_index_32bit_offset
: return 3;
4117 case nir_address_format_32bit_offset
: return 1;
4118 case nir_address_format_logical
: return 1;
4120 unreachable("Invalid address format");
4123 static inline const struct glsl_type
*
4124 nir_address_format_to_glsl_type(nir_address_format addr_format
)
4126 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
4127 assert(bit_size
== 32 || bit_size
== 64);
4128 return glsl_vector_type(bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
,
4129 nir_address_format_num_components(addr_format
));
4132 const nir_const_value
*nir_address_format_null_value(nir_address_format addr_format
);
4134 nir_ssa_def
*nir_build_addr_ieq(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
4135 nir_address_format addr_format
);
4137 nir_ssa_def
*nir_build_addr_isub(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
4138 nir_address_format addr_format
);
4140 nir_ssa_def
* nir_explicit_io_address_from_deref(struct nir_builder
*b
,
4141 nir_deref_instr
*deref
,
4142 nir_ssa_def
*base_addr
,
4143 nir_address_format addr_format
);
4144 void nir_lower_explicit_io_instr(struct nir_builder
*b
,
4145 nir_intrinsic_instr
*io_instr
,
4147 nir_address_format addr_format
);
4149 bool nir_lower_explicit_io(nir_shader
*shader
,
4150 nir_variable_mode modes
,
4151 nir_address_format
);
4153 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
4154 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
4156 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
4158 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
4159 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
4160 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
4162 bool nir_remove_dead_derefs(nir_shader
*shader
);
4163 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
4164 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
,
4165 bool (*can_remove_var
)(nir_variable
*var
));
4166 bool nir_lower_variable_initializers(nir_shader
*shader
,
4167 nir_variable_mode modes
);
4169 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
4170 bool nir_lower_vec_to_movs(nir_shader
*shader
);
4171 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
4173 const gl_state_index16
*alpha_ref_state_tokens
);
4174 bool nir_lower_alu(nir_shader
*shader
);
4176 bool nir_lower_flrp(nir_shader
*shader
, unsigned lowering_mask
,
4177 bool always_precise
, bool have_ffma
);
4179 bool nir_lower_alu_to_scalar(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
4180 bool nir_lower_bool_to_bitsize(nir_shader
*shader
);
4181 bool nir_lower_bool_to_float(nir_shader
*shader
);
4182 bool nir_lower_bool_to_int32(nir_shader
*shader
);
4183 bool nir_lower_int_to_float(nir_shader
*shader
);
4184 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
4185 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
4186 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
4187 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
4188 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
4190 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
4191 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
4192 bool nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode mask
);
4194 bool nir_lower_fragcolor(nir_shader
*shader
);
4195 void nir_lower_fragcoord_wtrans(nir_shader
*shader
);
4196 void nir_lower_viewport_transform(nir_shader
*shader
);
4197 bool nir_lower_uniforms_to_ubo(nir_shader
*shader
, int multiplier
);
4199 typedef struct nir_lower_subgroups_options
{
4200 uint8_t subgroup_size
;
4201 uint8_t ballot_bit_size
;
4202 bool lower_to_scalar
:1;
4203 bool lower_vote_trivial
:1;
4204 bool lower_vote_eq_to_ballot
:1;
4205 bool lower_subgroup_masks
:1;
4206 bool lower_shuffle
:1;
4207 bool lower_shuffle_to_32bit
:1;
4208 bool lower_shuffle_to_swizzle_amd
:1;
4210 bool lower_quad_broadcast_dynamic
:1;
4211 bool lower_quad_broadcast_dynamic_to_const
:1;
4212 } nir_lower_subgroups_options
;
4214 bool nir_lower_subgroups(nir_shader
*shader
,
4215 const nir_lower_subgroups_options
*options
);
4217 bool nir_lower_system_values(nir_shader
*shader
);
4219 enum PACKED nir_lower_tex_packing
{
4220 nir_lower_tex_packing_none
= 0,
4221 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
4222 * or unsigned ints based on the sampler type
4224 nir_lower_tex_packing_16
,
4225 /* The sampler returns 1 32-bit word of 4x8 unorm */
4226 nir_lower_tex_packing_8
,
4229 typedef struct nir_lower_tex_options
{
4231 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
4232 * sampler types a texture projector is lowered.
4237 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
4239 bool lower_txf_offset
;
4242 * If true, lower away nir_tex_src_offset for all rect textures.
4244 bool lower_rect_offset
;
4247 * If true, lower rect textures to 2D, using txs to fetch the
4248 * texture dimensions and dividing the texture coords by the
4249 * texture dims to normalize.
4254 * If true, convert yuv to rgb.
4256 unsigned lower_y_uv_external
;
4257 unsigned lower_y_u_v_external
;
4258 unsigned lower_yx_xuxv_external
;
4259 unsigned lower_xy_uxvx_external
;
4260 unsigned lower_ayuv_external
;
4261 unsigned lower_xyuv_external
;
4262 unsigned bt709_external
;
4263 unsigned bt2020_external
;
4266 * To emulate certain texture wrap modes, this can be used
4267 * to saturate the specified tex coord to [0.0, 1.0]. The
4268 * bits are according to sampler #, ie. if, for example:
4270 * (conf->saturate_s & (1 << n))
4272 * is true, then the s coord for sampler n is saturated.
4274 * Note that clamping must happen *after* projector lowering
4275 * so any projected texture sample instruction with a clamped
4276 * coordinate gets automatically lowered, regardless of the
4277 * 'lower_txp' setting.
4279 unsigned saturate_s
;
4280 unsigned saturate_t
;
4281 unsigned saturate_r
;
4283 /* Bitmask of textures that need swizzling.
4285 * If (swizzle_result & (1 << texture_index)), then the swizzle in
4286 * swizzles[texture_index] is applied to the result of the texturing
4289 unsigned swizzle_result
;
4291 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
4292 * while 4 and 5 represent 0 and 1 respectively.
4294 uint8_t swizzles
[32][4];
4296 /* Can be used to scale sampled values in range required by the format. */
4297 float scale_factors
[32];
4300 * Bitmap of textures that need srgb to linear conversion. If
4301 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
4302 * of the texture are lowered to linear.
4304 unsigned lower_srgb
;
4307 * If true, lower nir_texop_tex on shaders that doesn't support implicit
4308 * LODs to nir_texop_txl.
4310 bool lower_tex_without_implicit_lod
;
4313 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
4315 bool lower_txd_cube_map
;
4318 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
4323 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
4324 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
4325 * with lower_txd_cube_map.
4327 bool lower_txd_shadow
;
4330 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
4331 * Implies lower_txd_cube_map and lower_txd_shadow.
4336 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
4337 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
4339 bool lower_txb_shadow_clamp
;
4342 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
4343 * with nir_texop_txl. This includes cube maps.
4345 bool lower_txd_shadow_clamp
;
4348 * If true, lower nir_texop_txd on when it uses both offset and min_lod
4349 * with nir_texop_txl. This includes cube maps.
4351 bool lower_txd_offset_clamp
;
4354 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4355 * sampler is bindless.
4357 bool lower_txd_clamp_bindless_sampler
;
4360 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4361 * sampler index is not statically determinable to be less than 16.
4363 bool lower_txd_clamp_if_sampler_index_not_lt_16
;
4366 * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
4367 * 0-lod followed by a nir_ishr.
4372 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
4373 * mixed-up tg4 locations.
4375 bool lower_tg4_broadcom_swizzle
;
4378 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
4380 bool lower_tg4_offsets
;
4382 enum nir_lower_tex_packing lower_tex_packing
[32];
4383 } nir_lower_tex_options
;
4385 bool nir_lower_tex(nir_shader
*shader
,
4386 const nir_lower_tex_options
*options
);
4388 enum nir_lower_non_uniform_access_type
{
4389 nir_lower_non_uniform_ubo_access
= (1 << 0),
4390 nir_lower_non_uniform_ssbo_access
= (1 << 1),
4391 nir_lower_non_uniform_texture_access
= (1 << 2),
4392 nir_lower_non_uniform_image_access
= (1 << 3),
4395 bool nir_lower_non_uniform_access(nir_shader
*shader
,
4396 enum nir_lower_non_uniform_access_type
);
4398 enum nir_lower_idiv_path
{
4399 /* This path is based on NV50LegalizeSSA::handleDIV(). It is the faster of
4400 * the two but it is not exact in some cases (for example, 1091317713u /
4401 * 1034u gives 5209173 instead of 1055432) */
4402 nir_lower_idiv_fast
,
4403 /* This path is based on AMDGPUTargetLowering::LowerUDIVREM() and
4404 * AMDGPUTargetLowering::LowerSDIVREM(). It requires more instructions than
4405 * the nv50 path and many of them are integer multiplications, so it is
4406 * probably slower. It should always return the correct result, though. */
4407 nir_lower_idiv_precise
,
4410 bool nir_lower_idiv(nir_shader
*shader
, enum nir_lower_idiv_path path
);
4412 bool nir_lower_input_attachments(nir_shader
*shader
, bool use_fragcoord_sysval
);
4414 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
,
4416 bool use_clipdist_array
,
4417 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4418 bool nir_lower_clip_gs(nir_shader
*shader
, unsigned ucp_enables
,
4419 bool use_clipdist_array
,
4420 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4421 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
,
4422 bool use_clipdist_array
);
4423 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
4424 bool nir_lower_clip_disable(nir_shader
*shader
, unsigned clip_plane_enable
);
4426 void nir_lower_point_size_mov(nir_shader
*shader
,
4427 const gl_state_index16
*pointsize_state_tokens
);
4429 bool nir_lower_frexp(nir_shader
*nir
);
4431 void nir_lower_two_sided_color(nir_shader
*shader
, bool face_sysval
);
4433 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
4435 bool nir_lower_flatshade(nir_shader
*shader
);
4437 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
4438 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
4439 const gl_state_index16
*uniform_state_tokens
);
4441 typedef struct nir_lower_wpos_ytransform_options
{
4442 gl_state_index16 state_tokens
[STATE_LENGTH
];
4443 bool fs_coord_origin_upper_left
:1;
4444 bool fs_coord_origin_lower_left
:1;
4445 bool fs_coord_pixel_center_integer
:1;
4446 bool fs_coord_pixel_center_half_integer
:1;
4447 } nir_lower_wpos_ytransform_options
;
4449 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
4450 const nir_lower_wpos_ytransform_options
*options
);
4451 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
4453 bool nir_lower_wrmasks(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
4455 bool nir_lower_fb_read(nir_shader
*shader
);
4457 typedef struct nir_lower_drawpixels_options
{
4458 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
4459 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
4460 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
4461 unsigned drawpix_sampler
;
4462 unsigned pixelmap_sampler
;
4464 bool scale_and_bias
:1;
4465 } nir_lower_drawpixels_options
;
4467 void nir_lower_drawpixels(nir_shader
*shader
,
4468 const nir_lower_drawpixels_options
*options
);
4470 typedef struct nir_lower_bitmap_options
{
4473 } nir_lower_bitmap_options
;
4475 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
4477 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
);
4480 nir_lower_int_source_mods
= 1 << 0,
4481 nir_lower_float_source_mods
= 1 << 1,
4482 nir_lower_triop_abs
= 1 << 2,
4483 nir_lower_all_source_mods
= (1 << 3) - 1
4484 } nir_lower_to_source_mods_flags
;
4487 bool nir_lower_to_source_mods(nir_shader
*shader
, nir_lower_to_source_mods_flags options
);
4489 bool nir_lower_gs_intrinsics(nir_shader
*shader
, bool per_stream
);
4491 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
4493 bool nir_lower_bit_size(nir_shader
*shader
,
4494 nir_lower_bit_size_callback callback
,
4495 void *callback_data
);
4497 nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode
);
4498 bool nir_lower_int64(nir_shader
*shader
);
4500 nir_lower_doubles_options
nir_lower_doubles_op_to_options_mask(nir_op opcode
);
4501 bool nir_lower_doubles(nir_shader
*shader
, const nir_shader
*softfp64
,
4502 nir_lower_doubles_options options
);
4503 bool nir_lower_pack(nir_shader
*shader
);
4505 void nir_lower_mediump_outputs(nir_shader
*nir
);
4507 bool nir_lower_point_size(nir_shader
*shader
, float min
, float max
);
4510 nir_lower_interpolation_at_sample
= (1 << 1),
4511 nir_lower_interpolation_at_offset
= (1 << 2),
4512 nir_lower_interpolation_centroid
= (1 << 3),
4513 nir_lower_interpolation_pixel
= (1 << 4),
4514 nir_lower_interpolation_sample
= (1 << 5),
4515 } nir_lower_interpolation_options
;
4517 bool nir_lower_interpolation(nir_shader
*shader
,
4518 nir_lower_interpolation_options options
);
4520 bool nir_lower_discard_to_demote(nir_shader
*shader
);
4522 bool nir_lower_memory_model(nir_shader
*shader
);
4524 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
4526 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
4528 void nir_loop_analyze_impl(nir_function_impl
*impl
,
4529 nir_variable_mode indirect_mask
);
4531 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
4533 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
4534 bool nir_repair_ssa(nir_shader
*shader
);
4536 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
4537 bool nir_convert_to_lcssa(nir_shader
*shader
, bool skip_invariants
, bool skip_bool_invariants
);
4538 void nir_divergence_analysis(nir_shader
*shader
, nir_divergence_options options
);
4540 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
4541 * registers. If false, convert all values (even those not involved in a phi
4542 * node) to registers.
4544 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
4546 bool nir_lower_phis_to_regs_block(nir_block
*block
);
4547 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
4548 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
);
4550 bool nir_lower_samplers(nir_shader
*shader
);
4551 bool nir_lower_ssbo(nir_shader
*shader
);
4553 /* This is here for unit tests. */
4554 bool nir_opt_comparison_pre_impl(nir_function_impl
*impl
);
4556 bool nir_opt_comparison_pre(nir_shader
*shader
);
4558 bool nir_opt_access(nir_shader
*shader
);
4559 bool nir_opt_algebraic(nir_shader
*shader
);
4560 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
4561 bool nir_opt_algebraic_late(nir_shader
*shader
);
4562 bool nir_opt_algebraic_distribute_src_mods(nir_shader
*shader
);
4563 bool nir_opt_constant_folding(nir_shader
*shader
);
4565 /* Try to combine a and b into a. Return true if combination was possible,
4566 * which will result in b being removed by the pass. Return false if
4567 * combination wasn't possible.
4569 typedef bool (*nir_combine_memory_barrier_cb
)(
4570 nir_intrinsic_instr
*a
, nir_intrinsic_instr
*b
, void *data
);
4572 bool nir_opt_combine_memory_barriers(nir_shader
*shader
,
4573 nir_combine_memory_barrier_cb combine_cb
,
4576 bool nir_opt_combine_stores(nir_shader
*shader
, nir_variable_mode modes
);
4578 bool nir_copy_prop(nir_shader
*shader
);
4580 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
4582 bool nir_opt_cse(nir_shader
*shader
);
4584 bool nir_opt_dce(nir_shader
*shader
);
4586 bool nir_opt_dead_cf(nir_shader
*shader
);
4588 bool nir_opt_dead_write_vars(nir_shader
*shader
);
4590 bool nir_opt_deref_impl(nir_function_impl
*impl
);
4591 bool nir_opt_deref(nir_shader
*shader
);
4593 bool nir_opt_find_array_copies(nir_shader
*shader
);
4595 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
4597 bool nir_opt_idiv_const(nir_shader
*shader
, unsigned min_bit_size
);
4599 bool nir_opt_if(nir_shader
*shader
, bool aggressive_last_continue
);
4601 bool nir_opt_intrinsics(nir_shader
*shader
);
4603 bool nir_opt_large_constants(nir_shader
*shader
,
4604 glsl_type_size_align_func size_align
,
4605 unsigned threshold
);
4607 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
4610 nir_move_const_undef
= (1 << 0),
4611 nir_move_load_ubo
= (1 << 1),
4612 nir_move_load_input
= (1 << 2),
4613 nir_move_comparisons
= (1 << 3),
4614 nir_move_copies
= (1 << 4),
4617 bool nir_can_move_instr(nir_instr
*instr
, nir_move_options options
);
4619 bool nir_opt_sink(nir_shader
*shader
, nir_move_options options
);
4621 bool nir_opt_move(nir_shader
*shader
, nir_move_options options
);
4623 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
,
4624 bool indirect_load_ok
, bool expensive_alu_ok
);
4626 bool nir_opt_rematerialize_compares(nir_shader
*shader
);
4628 bool nir_opt_remove_phis(nir_shader
*shader
);
4629 bool nir_opt_remove_phis_block(nir_block
*block
);
4631 bool nir_opt_shrink_vectors(nir_shader
*shader
);
4633 bool nir_opt_trivial_continues(nir_shader
*shader
);
4635 bool nir_opt_undef(nir_shader
*shader
);
4637 bool nir_opt_vectorize(nir_shader
*shader
);
4639 bool nir_opt_conditional_discard(nir_shader
*shader
);
4641 typedef bool (*nir_should_vectorize_mem_func
)(unsigned align
, unsigned bit_size
,
4642 unsigned num_components
, unsigned high_offset
,
4643 nir_intrinsic_instr
*low
, nir_intrinsic_instr
*high
);
4645 bool nir_opt_load_store_vectorize(nir_shader
*shader
, nir_variable_mode modes
,
4646 nir_should_vectorize_mem_func callback
,
4647 nir_variable_mode robust_modes
);
4649 void nir_strip(nir_shader
*shader
);
4651 void nir_sweep(nir_shader
*shader
);
4653 void nir_remap_dual_slot_attributes(nir_shader
*shader
,
4654 uint64_t *dual_slot_inputs
);
4655 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
);
4657 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
4658 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);
4661 nir_variable_is_in_ubo(const nir_variable
*var
)
4663 return (var
->data
.mode
== nir_var_mem_ubo
&&
4664 var
->interface_type
!= NULL
);
4668 nir_variable_is_in_ssbo(const nir_variable
*var
)
4670 return (var
->data
.mode
== nir_var_mem_ssbo
&&
4671 var
->interface_type
!= NULL
);
4675 nir_variable_is_in_block(const nir_variable
*var
)
4677 return nir_variable_is_in_ubo(var
) || nir_variable_is_in_ssbo(var
);
4680 typedef struct nir_unsigned_upper_bound_config
{
4681 unsigned min_subgroup_size
;
4682 unsigned max_subgroup_size
;
4683 unsigned max_work_group_invocations
;
4684 unsigned max_work_group_count
[3];
4685 unsigned max_work_group_size
[3];
4687 uint32_t vertex_attrib_max
[32];
4688 } nir_unsigned_upper_bound_config
;
4691 nir_unsigned_upper_bound(nir_shader
*shader
, struct hash_table
*range_ht
,
4692 nir_ssa_scalar scalar
,
4693 const nir_unsigned_upper_bound_config
*config
);
4696 nir_addition_might_overflow(nir_shader
*shader
, struct hash_table
*range_ht
,
4697 nir_ssa_scalar ssa
, unsigned const_val
,
4698 const nir_unsigned_upper_bound_config
*config
);