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
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. See nir_depth_layout.
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 unsigned 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 * See gl_access_qualifier.
488 * Access flags for memory variables (SSBO/global), image uniforms, and
489 * bindless images in uniforms/inputs/outputs.
494 * Descriptor set binding for sampler or UBO.
496 unsigned descriptor_set
:5;
499 * output index for dual source blending.
504 * Initial binding point for a sampler or UBO.
506 * For array types, this represents the binding point for the first element.
511 * Storage location of the base of this variable
513 * The precise meaning of this field depends on the nature of the variable.
515 * - Vertex shader input: one of the values from \c gl_vert_attrib.
516 * - Vertex shader output: one of the values from \c gl_varying_slot.
517 * - Geometry shader input: one of the values from \c gl_varying_slot.
518 * - Geometry shader output: one of the values from \c gl_varying_slot.
519 * - Fragment shader input: one of the values from \c gl_varying_slot.
520 * - Fragment shader output: one of the values from \c gl_frag_result.
521 * - Uniforms: Per-stage uniform slot number for default uniform block.
522 * - Uniforms: Index within the uniform block definition for UBO members.
523 * - Non-UBO Uniforms: uniform slot number.
524 * - Other: This field is not currently used.
526 * If the variable is a uniform, shader input, or shader output, and the
527 * slot has not been assigned, the value will be -1.
532 * The actual location of the variable in the IR. Only valid for inputs,
533 * outputs, and uniforms (including samplers and images).
535 unsigned driver_location
;
538 * Location an atomic counter or transform feedback is stored at.
544 /** Image internal format if specified explicitly, otherwise PIPE_FORMAT_NONE. */
545 enum pipe_format format
;
550 * Transform feedback buffer.
555 * Transform feedback stride.
563 * Identifier for this variable generated by nir_index_vars() that is unique
564 * among other variables in the same exec_list.
568 /* Number of nir_variable_data members */
569 uint16_t num_members
;
572 * Built-in state that backs this uniform
574 * Once set at variable creation, \c state_slots must remain invariant.
575 * This is because, ideally, this array would be shared by all clones of
576 * this variable in the IR tree. In other words, we'd really like for it
577 * to be a fly-weight.
579 * If the variable is not a uniform, \c num_state_slots will be zero and
580 * \c state_slots will be \c NULL.
583 uint16_t num_state_slots
; /**< Number of state slots used */
584 nir_state_slot
*state_slots
; /**< State descriptors. */
588 * Constant expression assigned in the initializer of the variable
590 * This field should only be used temporarily by creators of NIR shaders
591 * and then lower_constant_initializers can be used to get rid of them.
592 * Most of the rest of NIR ignores this field or asserts that it's NULL.
594 nir_constant
*constant_initializer
;
597 * Global variable assigned in the initializer of the variable
598 * This field should only be used temporarily by creators of NIR shaders
599 * and then lower_constant_initializers can be used to get rid of them.
600 * Most of the rest of NIR ignores this field or asserts that it's NULL.
602 struct nir_variable
*pointer_initializer
;
605 * For variables that are in an interface block or are an instance of an
606 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
608 * \sa ir_variable::location
610 const struct glsl_type
*interface_type
;
613 * Description of per-member data for per-member struct variables
615 * This is used for variables which are actually an amalgamation of
616 * multiple entities such as a struct of built-in values or a struct of
617 * inputs each with their own layout specifier. This is only allowed on
618 * variables with a struct or array of array of struct type.
620 struct nir_variable_data
*members
;
624 _nir_shader_variable_has_mode(nir_variable
*var
, unsigned modes
)
626 /* This isn't a shader variable */
627 assert(!(modes
& nir_var_function_temp
));
628 return var
->data
.mode
& modes
;
631 #define nir_foreach_variable_in_list(var, var_list) \
632 foreach_list_typed(nir_variable, var, node, var_list)
634 #define nir_foreach_variable_in_list_safe(var, var_list) \
635 foreach_list_typed_safe(nir_variable, var, node, var_list)
637 #define nir_foreach_variable_in_shader(var, shader) \
638 nir_foreach_variable_in_list(var, &(shader)->variables)
640 #define nir_foreach_variable_in_shader_safe(var, shader) \
641 nir_foreach_variable_in_list_safe(var, &(shader)->variables)
643 #define nir_foreach_variable_with_modes(var, shader, modes) \
644 nir_foreach_variable_in_shader(var, shader) \
645 if (_nir_shader_variable_has_mode(var, modes))
647 #define nir_foreach_variable_with_modes_safe(var, shader, modes) \
648 nir_foreach_variable_in_shader_safe(var, shader) \
649 if (_nir_shader_variable_has_mode(var, modes))
651 #define nir_foreach_shader_in_variable(var, shader) \
652 nir_foreach_variable_with_modes(var, shader, nir_var_shader_in)
654 #define nir_foreach_shader_in_variable_safe(var, shader) \
655 nir_foreach_variable_with_modes_safe(var, shader, nir_var_shader_in)
657 #define nir_foreach_shader_out_variable(var, shader) \
658 nir_foreach_variable_with_modes(var, shader, nir_var_shader_out)
660 #define nir_foreach_shader_out_variable_safe(var, shader) \
661 nir_foreach_variable_with_modes_safe(var, shader, nir_var_shader_out)
663 #define nir_foreach_uniform_variable(var, shader) \
664 nir_foreach_variable_with_modes(var, shader, nir_var_uniform)
666 #define nir_foreach_uniform_variable_safe(var, shader) \
667 nir_foreach_variable_with_modes_safe(var, shader, nir_var_uniform)
670 nir_variable_is_global(const nir_variable
*var
)
672 return var
->data
.mode
!= nir_var_function_temp
;
675 typedef struct nir_register
{
676 struct exec_node node
;
678 unsigned num_components
; /** < number of vector components */
679 unsigned num_array_elems
; /** < size of array (0 for no array) */
681 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
684 /** generic register index. */
687 /** only for debug purposes, can be NULL */
690 /** set of nir_srcs where this register is used (read from) */
691 struct list_head uses
;
693 /** set of nir_dests where this register is defined (written to) */
694 struct list_head defs
;
696 /** set of nir_ifs where this register is used as a condition */
697 struct list_head if_uses
;
700 #define nir_foreach_register(reg, reg_list) \
701 foreach_list_typed(nir_register, reg, node, reg_list)
702 #define nir_foreach_register_safe(reg, reg_list) \
703 foreach_list_typed_safe(nir_register, reg, node, reg_list)
705 typedef enum PACKED
{
707 nir_instr_type_deref
,
710 nir_instr_type_intrinsic
,
711 nir_instr_type_load_const
,
713 nir_instr_type_ssa_undef
,
715 nir_instr_type_parallel_copy
,
718 typedef struct nir_instr
{
719 struct exec_node node
;
720 struct nir_block
*block
;
723 /* A temporary for optimization and analysis passes to use for storing
724 * flags. For instance, DCE uses this to store the "dead/live" info.
728 /** generic instruction index. */
732 static inline nir_instr
*
733 nir_instr_next(nir_instr
*instr
)
735 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
736 if (exec_node_is_tail_sentinel(next
))
739 return exec_node_data(nir_instr
, next
, node
);
742 static inline nir_instr
*
743 nir_instr_prev(nir_instr
*instr
)
745 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
746 if (exec_node_is_head_sentinel(prev
))
749 return exec_node_data(nir_instr
, prev
, node
);
753 nir_instr_is_first(const nir_instr
*instr
)
755 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr
->node
));
759 nir_instr_is_last(const nir_instr
*instr
)
761 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr
->node
));
764 typedef struct nir_ssa_def
{
765 /** for debugging only, can be NULL */
768 /** generic SSA definition index. */
771 /** Index into the live_in and live_out bitfields */
774 /** Instruction which produces this SSA value. */
775 nir_instr
*parent_instr
;
777 /** set of nir_instrs where this register is used (read from) */
778 struct list_head uses
;
780 /** set of nir_ifs where this register is used as a condition */
781 struct list_head if_uses
;
783 uint8_t num_components
;
785 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
789 * True if this SSA value may have different values in different SIMD
790 * invocations of the shader. This is set by nir_divergence_analysis.
799 struct nir_src
*indirect
; /** < NULL for no indirect offset */
800 unsigned base_offset
;
802 /* TODO use-def chain goes here */
806 nir_instr
*parent_instr
;
807 struct list_head def_link
;
810 struct nir_src
*indirect
; /** < NULL for no indirect offset */
811 unsigned base_offset
;
813 /* TODO def-use chain goes here */
818 typedef struct nir_src
{
820 /** Instruction that consumes this value as a source. */
821 nir_instr
*parent_instr
;
822 struct nir_if
*parent_if
;
825 struct list_head use_link
;
835 static inline nir_src
838 nir_src src
= { { NULL
} };
842 #define NIR_SRC_INIT nir_src_init()
844 #define nir_foreach_use(src, reg_or_ssa_def) \
845 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
847 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
848 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
850 #define nir_foreach_if_use(src, reg_or_ssa_def) \
851 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
853 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
854 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
865 static inline nir_dest
868 nir_dest dest
= { { { NULL
} } };
872 #define NIR_DEST_INIT nir_dest_init()
874 #define nir_foreach_def(dest, reg) \
875 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
877 #define nir_foreach_def_safe(dest, reg) \
878 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
880 static inline nir_src
881 nir_src_for_ssa(nir_ssa_def
*def
)
883 nir_src src
= NIR_SRC_INIT
;
891 static inline nir_src
892 nir_src_for_reg(nir_register
*reg
)
894 nir_src src
= NIR_SRC_INIT
;
898 src
.reg
.indirect
= NULL
;
899 src
.reg
.base_offset
= 0;
904 static inline nir_dest
905 nir_dest_for_reg(nir_register
*reg
)
907 nir_dest dest
= NIR_DEST_INIT
;
914 static inline unsigned
915 nir_src_bit_size(nir_src src
)
917 return src
.is_ssa
? src
.ssa
->bit_size
: src
.reg
.reg
->bit_size
;
920 static inline unsigned
921 nir_src_num_components(nir_src src
)
923 return src
.is_ssa
? src
.ssa
->num_components
: src
.reg
.reg
->num_components
;
927 nir_src_is_const(nir_src src
)
930 src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
;
934 nir_src_is_divergent(nir_src src
)
937 return src
.ssa
->divergent
;
940 static inline unsigned
941 nir_dest_bit_size(nir_dest dest
)
943 return dest
.is_ssa
? dest
.ssa
.bit_size
: dest
.reg
.reg
->bit_size
;
946 static inline unsigned
947 nir_dest_num_components(nir_dest dest
)
949 return dest
.is_ssa
? dest
.ssa
.num_components
: dest
.reg
.reg
->num_components
;
953 nir_dest_is_divergent(nir_dest dest
)
956 return dest
.ssa
.divergent
;
959 /* Are all components the same, ie. .xxxx */
961 nir_is_same_comp_swizzle(uint8_t *swiz
, unsigned nr_comp
)
963 for (unsigned i
= 1; i
< nr_comp
; i
++)
964 if (swiz
[i
] != swiz
[0])
969 /* Are all components sequential, ie. .yzw */
971 nir_is_sequential_comp_swizzle(uint8_t *swiz
, unsigned nr_comp
)
973 for (unsigned i
= 1; i
< nr_comp
; i
++)
974 if (swiz
[i
] != (swiz
[0] + i
))
979 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
980 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
986 * \name input modifiers
990 * For inputs interpreted as floating point, flips the sign bit. For
991 * inputs interpreted as integers, performs the two's complement negation.
996 * Clears the sign bit for floating point values, and computes the integer
997 * absolute value for integers. Note that the negate modifier acts after
998 * the absolute value modifier, therefore if both are set then all inputs
999 * will become negative.
1005 * For each input component, says which component of the register it is
1006 * chosen from. Note that which elements of the swizzle are used and which
1007 * are ignored are based on the write mask for most opcodes - for example,
1008 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
1009 * a swizzle of {2, x, 1, 0} where x means "don't care."
1011 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
1018 * \name saturate output modifier
1020 * Only valid for opcodes that output floating-point numbers. Clamps the
1021 * output to between 0.0 and 1.0 inclusive.
1026 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
1029 /** NIR sized and unsized types
1031 * The values in this enum are carefully chosen so that the sized type is
1032 * just the unsized type OR the number of bits.
1034 typedef enum PACKED
{
1035 nir_type_invalid
= 0, /* Not a valid type */
1039 nir_type_float
= 128,
1040 nir_type_bool1
= 1 | nir_type_bool
,
1041 nir_type_bool8
= 8 | nir_type_bool
,
1042 nir_type_bool16
= 16 | nir_type_bool
,
1043 nir_type_bool32
= 32 | nir_type_bool
,
1044 nir_type_int1
= 1 | nir_type_int
,
1045 nir_type_int8
= 8 | nir_type_int
,
1046 nir_type_int16
= 16 | nir_type_int
,
1047 nir_type_int32
= 32 | nir_type_int
,
1048 nir_type_int64
= 64 | nir_type_int
,
1049 nir_type_uint1
= 1 | nir_type_uint
,
1050 nir_type_uint8
= 8 | nir_type_uint
,
1051 nir_type_uint16
= 16 | nir_type_uint
,
1052 nir_type_uint32
= 32 | nir_type_uint
,
1053 nir_type_uint64
= 64 | nir_type_uint
,
1054 nir_type_float16
= 16 | nir_type_float
,
1055 nir_type_float32
= 32 | nir_type_float
,
1056 nir_type_float64
= 64 | nir_type_float
,
1059 #define NIR_ALU_TYPE_SIZE_MASK 0x79
1060 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
1062 static inline unsigned
1063 nir_alu_type_get_type_size(nir_alu_type type
)
1065 return type
& NIR_ALU_TYPE_SIZE_MASK
;
1068 static inline nir_alu_type
1069 nir_alu_type_get_base_type(nir_alu_type type
)
1071 return (nir_alu_type
)(type
& NIR_ALU_TYPE_BASE_TYPE_MASK
);
1074 static inline nir_alu_type
1075 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
1077 switch (base_type
) {
1078 case GLSL_TYPE_BOOL
:
1079 return nir_type_bool1
;
1081 case GLSL_TYPE_UINT
:
1082 return nir_type_uint32
;
1085 return nir_type_int32
;
1087 case GLSL_TYPE_UINT16
:
1088 return nir_type_uint16
;
1090 case GLSL_TYPE_INT16
:
1091 return nir_type_int16
;
1093 case GLSL_TYPE_UINT8
:
1094 return nir_type_uint8
;
1095 case GLSL_TYPE_INT8
:
1096 return nir_type_int8
;
1097 case GLSL_TYPE_UINT64
:
1098 return nir_type_uint64
;
1100 case GLSL_TYPE_INT64
:
1101 return nir_type_int64
;
1103 case GLSL_TYPE_FLOAT
:
1104 return nir_type_float32
;
1106 case GLSL_TYPE_FLOAT16
:
1107 return nir_type_float16
;
1109 case GLSL_TYPE_DOUBLE
:
1110 return nir_type_float64
;
1113 case GLSL_TYPE_SAMPLER
:
1114 case GLSL_TYPE_IMAGE
:
1115 case GLSL_TYPE_ATOMIC_UINT
:
1116 case GLSL_TYPE_STRUCT
:
1117 case GLSL_TYPE_INTERFACE
:
1118 case GLSL_TYPE_ARRAY
:
1119 case GLSL_TYPE_VOID
:
1120 case GLSL_TYPE_SUBROUTINE
:
1121 case GLSL_TYPE_FUNCTION
:
1122 case GLSL_TYPE_ERROR
:
1123 return nir_type_invalid
;
1126 unreachable("unknown type");
1129 static inline nir_alu_type
1130 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
1132 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
1135 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
1136 nir_rounding_mode rnd
);
1138 static inline nir_op
1139 nir_op_vec(unsigned components
)
1141 switch (components
) {
1142 case 1: return nir_op_mov
;
1143 case 2: return nir_op_vec2
;
1144 case 3: return nir_op_vec3
;
1145 case 4: return nir_op_vec4
;
1146 case 8: return nir_op_vec8
;
1147 case 16: return nir_op_vec16
;
1148 default: unreachable("bad component count");
1153 nir_op_is_vec(nir_op op
)
1169 nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode
, unsigned bit_size
)
1171 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
) ||
1172 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
) ||
1173 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
);
1177 nir_is_denorm_flush_to_zero(unsigned execution_mode
, unsigned bit_size
)
1179 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
) ||
1180 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
) ||
1181 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
);
1185 nir_is_denorm_preserve(unsigned execution_mode
, unsigned bit_size
)
1187 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP16
) ||
1188 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP32
) ||
1189 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP64
);
1193 nir_is_rounding_mode_rtne(unsigned execution_mode
, unsigned bit_size
)
1195 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1196 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1197 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1201 nir_is_rounding_mode_rtz(unsigned execution_mode
, unsigned bit_size
)
1203 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1204 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1205 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1209 nir_has_any_rounding_mode_rtz(unsigned execution_mode
)
1211 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1212 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1213 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1217 nir_has_any_rounding_mode_rtne(unsigned execution_mode
)
1219 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1220 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1221 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1224 static inline nir_rounding_mode
1225 nir_get_rounding_mode_from_float_controls(unsigned execution_mode
,
1228 if (nir_alu_type_get_base_type(type
) != nir_type_float
)
1229 return nir_rounding_mode_undef
;
1231 unsigned bit_size
= nir_alu_type_get_type_size(type
);
1233 if (nir_is_rounding_mode_rtz(execution_mode
, bit_size
))
1234 return nir_rounding_mode_rtz
;
1235 if (nir_is_rounding_mode_rtne(execution_mode
, bit_size
))
1236 return nir_rounding_mode_rtne
;
1237 return nir_rounding_mode_undef
;
1241 nir_has_any_rounding_mode_enabled(unsigned execution_mode
)
1244 nir_has_any_rounding_mode_rtne(execution_mode
) ||
1245 nir_has_any_rounding_mode_rtz(execution_mode
);
1251 * Operation where the first two sources are commutative.
1253 * For 2-source operations, this just mathematical commutativity. Some
1254 * 3-source operations, like ffma, are only commutative in the first two
1257 NIR_OP_IS_2SRC_COMMUTATIVE
= (1 << 0),
1258 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
1259 } nir_op_algebraic_property
;
1267 * The number of components in the output
1269 * If non-zero, this is the size of the output and input sizes are
1270 * explicitly given; swizzle and writemask are still in effect, but if
1271 * the output component is masked out, then the input component may
1274 * If zero, the opcode acts in the standard, per-component manner; the
1275 * operation is performed on each component (except the ones that are
1276 * masked out) with the input being taken from the input swizzle for
1279 * The size of some of the inputs may be given (i.e. non-zero) even
1280 * though output_size is zero; in that case, the inputs with a zero
1281 * size act per-component, while the inputs with non-zero size don't.
1283 uint8_t output_size
;
1286 * The type of vector that the instruction outputs. Note that the
1287 * staurate modifier is only allowed on outputs with the float type.
1290 nir_alu_type output_type
;
1293 * The number of components in each input
1295 uint8_t input_sizes
[NIR_MAX_VEC_COMPONENTS
];
1298 * The type of vector that each input takes. Note that negate and
1299 * absolute value are only allowed on inputs with int or float type and
1300 * behave differently on the two.
1302 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
1304 nir_op_algebraic_property algebraic_properties
;
1306 /* Whether this represents a numeric conversion opcode */
1310 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
1312 typedef struct nir_alu_instr
{
1316 /** Indicates that this ALU instruction generates an exact value
1318 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1319 * really equivalent to either. This indicates that the value generated by
1320 * this operation is high-precision and any code transformations that touch
1321 * it must ensure that the resulting value is bit-for-bit identical to the
1327 * Indicates that this instruction do not cause wrapping to occur, in the
1328 * form of overflow or underflow.
1330 bool no_signed_wrap
:1;
1331 bool no_unsigned_wrap
:1;
1337 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
1338 nir_alu_instr
*instr
);
1339 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
1340 nir_alu_instr
*instr
);
1342 /* is this source channel used? */
1344 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
1347 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1348 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
1350 return (instr
->dest
.write_mask
>> channel
) & 1;
1353 static inline nir_component_mask_t
1354 nir_alu_instr_src_read_mask(const nir_alu_instr
*instr
, unsigned src
)
1356 nir_component_mask_t read_mask
= 0;
1357 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; c
++) {
1358 if (!nir_alu_instr_channel_used(instr
, src
, c
))
1361 read_mask
|= (1 << instr
->src
[src
].swizzle
[c
]);
1367 * Get the number of channels used for a source
1369 static inline unsigned
1370 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
1372 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1373 return nir_op_infos
[instr
->op
].input_sizes
[src
];
1375 return nir_dest_num_components(instr
->dest
.dest
);
1379 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1381 switch (instr
->op
) {
1401 bool nir_const_value_negative_equal(nir_const_value c1
, nir_const_value c2
,
1402 nir_alu_type full_type
);
1404 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
1405 unsigned src1
, unsigned src2
);
1407 bool nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
1408 const nir_alu_instr
*alu2
,
1409 unsigned src1
, unsigned src2
);
1413 nir_deref_type_array
,
1414 nir_deref_type_array_wildcard
,
1415 nir_deref_type_ptr_as_array
,
1416 nir_deref_type_struct
,
1417 nir_deref_type_cast
,
1423 /** The type of this deref instruction */
1424 nir_deref_type deref_type
;
1426 /** The mode of the underlying variable */
1427 nir_variable_mode mode
;
1429 /** The dereferenced type of the resulting pointer value */
1430 const struct glsl_type
*type
;
1433 /** Variable being dereferenced if deref_type is a deref_var */
1436 /** Parent deref if deref_type is not deref_var */
1440 /** Additional deref parameters */
1451 unsigned ptr_stride
;
1455 /** Destination to store the resulting "pointer" */
1459 static inline nir_deref_instr
*nir_src_as_deref(nir_src src
);
1461 static inline nir_deref_instr
*
1462 nir_deref_instr_parent(const nir_deref_instr
*instr
)
1464 if (instr
->deref_type
== nir_deref_type_var
)
1467 return nir_src_as_deref(instr
->parent
);
1470 static inline nir_variable
*
1471 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1473 while (instr
->deref_type
!= nir_deref_type_var
) {
1474 if (instr
->deref_type
== nir_deref_type_cast
)
1477 instr
= nir_deref_instr_parent(instr
);
1483 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1484 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr
*instr
);
1485 bool nir_deref_instr_has_complex_use(nir_deref_instr
*instr
);
1487 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1489 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*instr
);
1494 struct nir_function
*callee
;
1496 unsigned num_params
;
1500 #include "nir_intrinsics.h"
1502 #define NIR_INTRINSIC_MAX_CONST_INDEX 5
1504 /** Represents an intrinsic
1506 * An intrinsic is an instruction type for handling things that are
1507 * more-or-less regular operations but don't just consume and produce SSA
1508 * values like ALU operations do. Intrinsics are not for things that have
1509 * special semantic meaning such as phi nodes and parallel copies.
1510 * Examples of intrinsics include variable load/store operations, system
1511 * value loads, and the like. Even though texturing more-or-less falls
1512 * under this category, texturing is its own instruction type because
1513 * trying to represent texturing with intrinsics would lead to a
1514 * combinatorial explosion of intrinsic opcodes.
1516 * By having a single instruction type for handling a lot of different
1517 * cases, optimization passes can look for intrinsics and, for the most
1518 * part, completely ignore them. Each intrinsic type also has a few
1519 * possible flags that govern whether or not they can be reordered or
1520 * eliminated. That way passes like dead code elimination can still work
1521 * on intrisics without understanding the meaning of each.
1523 * Each intrinsic has some number of constant indices, some number of
1524 * variables, and some number of sources. What these sources, variables,
1525 * and indices mean depends on the intrinsic and is documented with the
1526 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1527 * instructions are the only types of instruction that can operate on
1533 nir_intrinsic_op intrinsic
;
1537 /** number of components if this is a vectorized intrinsic
1539 * Similarly to ALU operations, some intrinsics are vectorized.
1540 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1541 * For vectorized intrinsics, the num_components field specifies the
1542 * number of destination components and the number of source components
1543 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1545 uint8_t num_components
;
1547 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1550 } nir_intrinsic_instr
;
1552 static inline nir_variable
*
1553 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1555 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1559 /* Memory ordering. */
1560 NIR_MEMORY_ACQUIRE
= 1 << 0,
1561 NIR_MEMORY_RELEASE
= 1 << 1,
1562 NIR_MEMORY_ACQ_REL
= NIR_MEMORY_ACQUIRE
| NIR_MEMORY_RELEASE
,
1564 /* Memory visibility operations. */
1565 NIR_MEMORY_MAKE_AVAILABLE
= 1 << 2,
1566 NIR_MEMORY_MAKE_VISIBLE
= 1 << 3,
1567 } nir_memory_semantics
;
1571 NIR_SCOPE_INVOCATION
,
1573 NIR_SCOPE_WORKGROUP
,
1574 NIR_SCOPE_QUEUE_FAMILY
,
1579 * \name NIR intrinsics semantic flags
1581 * information about what the compiler can do with the intrinsics.
1583 * \sa nir_intrinsic_info::flags
1587 * whether the intrinsic can be safely eliminated if none of its output
1588 * value is not being used.
1590 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1593 * Whether the intrinsic can be reordered with respect to any other
1594 * intrinsic, i.e. whether the only reordering dependencies of the
1595 * intrinsic are due to the register reads/writes.
1597 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1598 } nir_intrinsic_semantic_flag
;
1601 * \name NIR intrinsics const-index flag
1603 * Indicates the usage of a const_index slot.
1605 * \sa nir_intrinsic_info::index_map
1609 * Generally instructions that take a offset src argument, can encode
1610 * a constant 'base' value which is added to the offset.
1612 NIR_INTRINSIC_BASE
= 1,
1615 * For store instructions, a writemask for the store.
1617 NIR_INTRINSIC_WRMASK
,
1620 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1622 NIR_INTRINSIC_STREAM_ID
,
1625 * The clip-plane id for load_user_clip_plane intrinsic.
1627 NIR_INTRINSIC_UCP_ID
,
1630 * The amount of data, starting from BASE, that this instruction may
1631 * access. This is used to provide bounds if the offset is not constant.
1633 NIR_INTRINSIC_RANGE
,
1636 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1638 NIR_INTRINSIC_DESC_SET
,
1641 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1643 NIR_INTRINSIC_BINDING
,
1648 NIR_INTRINSIC_COMPONENT
,
1651 * Interpolation mode (only meaningful for FS inputs).
1653 NIR_INTRINSIC_INTERP_MODE
,
1656 * A binary nir_op to use when performing a reduction or scan operation
1658 NIR_INTRINSIC_REDUCTION_OP
,
1661 * Cluster size for reduction operations
1663 NIR_INTRINSIC_CLUSTER_SIZE
,
1666 * Parameter index for a load_param intrinsic
1668 NIR_INTRINSIC_PARAM_IDX
,
1671 * Image dimensionality for image intrinsics
1673 * One of GLSL_SAMPLER_DIM_*
1675 NIR_INTRINSIC_IMAGE_DIM
,
1678 * Non-zero if we are accessing an array image
1680 NIR_INTRINSIC_IMAGE_ARRAY
,
1683 * Image format for image intrinsics
1685 NIR_INTRINSIC_FORMAT
,
1688 * Access qualifiers for image and memory access intrinsics
1690 NIR_INTRINSIC_ACCESS
,
1693 * Alignment for offsets and addresses
1695 * These two parameters, specify an alignment in terms of a multiplier and
1696 * an offset. The offset or address parameter X of the intrinsic is
1697 * guaranteed to satisfy the following:
1699 * (X - align_offset) % align_mul == 0
1701 NIR_INTRINSIC_ALIGN_MUL
,
1702 NIR_INTRINSIC_ALIGN_OFFSET
,
1705 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1707 NIR_INTRINSIC_DESC_TYPE
,
1710 * The nir_alu_type of a uniform/input/output
1715 * The swizzle mask for the instructions
1716 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1718 NIR_INTRINSIC_SWIZZLE_MASK
,
1720 /* Separate source/dest access flags for copies */
1721 NIR_INTRINSIC_SRC_ACCESS
,
1722 NIR_INTRINSIC_DST_ACCESS
,
1724 /* Driver location for nir_load_patch_location_ir3 */
1725 NIR_INTRINSIC_DRIVER_LOCATION
,
1728 * Mask of nir_memory_semantics, includes ordering and visibility.
1730 NIR_INTRINSIC_MEMORY_SEMANTICS
,
1733 * Mask of nir_variable_modes affected by the memory operation.
1735 NIR_INTRINSIC_MEMORY_MODES
,
1738 * Value of nir_scope.
1740 NIR_INTRINSIC_MEMORY_SCOPE
,
1743 * Value of nir_scope.
1745 NIR_INTRINSIC_EXECUTION_SCOPE
,
1748 * Value of nir_io_semantics.
1750 NIR_INTRINSIC_IO_SEMANTICS
,
1752 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1754 } nir_intrinsic_index_flag
;
1757 unsigned location
:7; /* gl_vert_attrib, gl_varying_slot, or gl_frag_result */
1758 unsigned num_slots
:6; /* max 32, may be pessimistic with const indexing */
1759 unsigned dual_source_blend_index
:1;
1760 unsigned fb_fetch_output
:1; /* for GL_KHR_blend_equation_advanced */
1761 unsigned gs_streams
:8; /* xxyyzzww: 2-bit stream index for each component */
1765 #define NIR_INTRINSIC_MAX_INPUTS 5
1770 uint8_t num_srcs
; /** < number of register/SSA inputs */
1772 /** number of components of each input register
1774 * If this value is 0, the number of components is given by the
1775 * num_components field of nir_intrinsic_instr. If this value is -1, the
1776 * intrinsic consumes however many components are provided and it is not
1779 int8_t src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1783 /** number of components of the output register
1785 * If this value is 0, the number of components is given by the
1786 * num_components field of nir_intrinsic_instr.
1788 uint8_t dest_components
;
1790 /** bitfield of legal bit sizes */
1791 uint8_t dest_bit_sizes
;
1793 /** the number of constant indices used by the intrinsic */
1794 uint8_t num_indices
;
1796 /** indicates the usage of intr->const_index[n] */
1797 uint8_t index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1799 /** semantic flags for calls to this intrinsic */
1800 nir_intrinsic_semantic_flag flags
;
1801 } nir_intrinsic_info
;
1803 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1805 static inline unsigned
1806 nir_intrinsic_src_components(const nir_intrinsic_instr
*intr
, unsigned srcn
)
1808 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1809 assert(srcn
< info
->num_srcs
);
1810 if (info
->src_components
[srcn
] > 0)
1811 return info
->src_components
[srcn
];
1812 else if (info
->src_components
[srcn
] == 0)
1813 return intr
->num_components
;
1815 return nir_src_num_components(intr
->src
[srcn
]);
1818 static inline unsigned
1819 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1821 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1822 if (!info
->has_dest
)
1824 else if (info
->dest_components
)
1825 return info
->dest_components
;
1827 return intr
->num_components
;
1831 * Helper to copy const_index[] from src to dst, without assuming they
1835 nir_intrinsic_copy_const_indices(nir_intrinsic_instr
*dst
, nir_intrinsic_instr
*src
)
1837 if (src
->intrinsic
== dst
->intrinsic
) {
1838 memcpy(dst
->const_index
, src
->const_index
, sizeof(dst
->const_index
));
1842 const nir_intrinsic_info
*src_info
= &nir_intrinsic_infos
[src
->intrinsic
];
1843 const nir_intrinsic_info
*dst_info
= &nir_intrinsic_infos
[dst
->intrinsic
];
1845 for (unsigned i
= 0; i
< NIR_INTRINSIC_NUM_INDEX_FLAGS
; i
++) {
1846 if (src_info
->index_map
[i
] == 0)
1849 /* require that dst instruction also uses the same const_index[]: */
1850 assert(dst_info
->index_map
[i
] > 0);
1852 dst
->const_index
[dst_info
->index_map
[i
] - 1] =
1853 src
->const_index
[src_info
->index_map
[i
] - 1];
1857 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1858 static inline type \
1859 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1861 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1862 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1863 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1865 static inline void \
1866 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1868 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1869 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1870 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1872 static inline bool \
1873 nir_intrinsic_has_##name(nir_intrinsic_instr *instr) \
1875 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1876 return info->index_map[NIR_INTRINSIC_##flag] > 0; \
1879 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1880 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1881 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1882 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1883 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1884 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1885 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1886 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1887 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1888 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1889 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1890 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1891 INTRINSIC_IDX_ACCESSORS(image_dim
, IMAGE_DIM
, enum glsl_sampler_dim
)
1892 INTRINSIC_IDX_ACCESSORS(image_array
, IMAGE_ARRAY
, bool)
1893 INTRINSIC_IDX_ACCESSORS(access
, ACCESS
, enum gl_access_qualifier
)
1894 INTRINSIC_IDX_ACCESSORS(src_access
, SRC_ACCESS
, enum gl_access_qualifier
)
1895 INTRINSIC_IDX_ACCESSORS(dst_access
, DST_ACCESS
, enum gl_access_qualifier
)
1896 INTRINSIC_IDX_ACCESSORS(format
, FORMAT
, enum pipe_format
)
1897 INTRINSIC_IDX_ACCESSORS(align_mul
, ALIGN_MUL
, unsigned)
1898 INTRINSIC_IDX_ACCESSORS(align_offset
, ALIGN_OFFSET
, unsigned)
1899 INTRINSIC_IDX_ACCESSORS(desc_type
, DESC_TYPE
, unsigned)
1900 INTRINSIC_IDX_ACCESSORS(type
, TYPE
, nir_alu_type
)
1901 INTRINSIC_IDX_ACCESSORS(swizzle_mask
, SWIZZLE_MASK
, unsigned)
1902 INTRINSIC_IDX_ACCESSORS(driver_location
, DRIVER_LOCATION
, unsigned)
1903 INTRINSIC_IDX_ACCESSORS(memory_semantics
, MEMORY_SEMANTICS
, nir_memory_semantics
)
1904 INTRINSIC_IDX_ACCESSORS(memory_modes
, MEMORY_MODES
, nir_variable_mode
)
1905 INTRINSIC_IDX_ACCESSORS(memory_scope
, MEMORY_SCOPE
, nir_scope
)
1906 INTRINSIC_IDX_ACCESSORS(execution_scope
, EXECUTION_SCOPE
, nir_scope
)
1909 nir_intrinsic_set_align(nir_intrinsic_instr
*intrin
,
1910 unsigned align_mul
, unsigned align_offset
)
1912 assert(util_is_power_of_two_nonzero(align_mul
));
1913 assert(align_offset
< align_mul
);
1914 nir_intrinsic_set_align_mul(intrin
, align_mul
);
1915 nir_intrinsic_set_align_offset(intrin
, align_offset
);
1918 /** Returns a simple alignment for a load/store intrinsic offset
1920 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1921 * and ALIGN_OFFSET parameters, this helper takes both into account and
1922 * provides a single simple alignment parameter. The offset X is guaranteed
1923 * to satisfy X % align == 0.
1925 static inline unsigned
1926 nir_intrinsic_align(const nir_intrinsic_instr
*intrin
)
1928 const unsigned align_mul
= nir_intrinsic_align_mul(intrin
);
1929 const unsigned align_offset
= nir_intrinsic_align_offset(intrin
);
1930 assert(align_offset
< align_mul
);
1931 return align_offset
? 1 << (ffs(align_offset
) - 1) : align_mul
;
1935 nir_intrinsic_set_io_semantics(nir_intrinsic_instr
*intrin
,
1936 nir_io_semantics semantics
)
1938 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intrin
->intrinsic
];
1939 assert(info
->index_map
[NIR_INTRINSIC_IO_SEMANTICS
] > 0);
1940 STATIC_ASSERT(sizeof(nir_io_semantics
) == sizeof(intrin
->const_index
[0]));
1941 semantics
._pad
= 0; /* clear padding bits */
1942 memcpy(&intrin
->const_index
[info
->index_map
[NIR_INTRINSIC_IO_SEMANTICS
] - 1],
1943 &semantics
, sizeof(semantics
));
1946 static inline nir_io_semantics
1947 nir_intrinsic_io_semantics(const nir_intrinsic_instr
*intrin
)
1949 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intrin
->intrinsic
];
1950 assert(info
->index_map
[NIR_INTRINSIC_IO_SEMANTICS
] > 0);
1951 nir_io_semantics semantics
;
1953 &intrin
->const_index
[info
->index_map
[NIR_INTRINSIC_IO_SEMANTICS
] - 1],
1959 nir_image_intrinsic_coord_components(const nir_intrinsic_instr
*instr
);
1961 /* Converts a image_deref_* intrinsic into a image_* one */
1962 void nir_rewrite_image_intrinsic(nir_intrinsic_instr
*instr
,
1963 nir_ssa_def
*handle
, bool bindless
);
1965 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1967 nir_intrinsic_can_reorder(nir_intrinsic_instr
*instr
)
1969 if (instr
->intrinsic
== nir_intrinsic_load_deref
||
1970 instr
->intrinsic
== nir_intrinsic_load_ssbo
||
1971 instr
->intrinsic
== nir_intrinsic_bindless_image_load
||
1972 instr
->intrinsic
== nir_intrinsic_image_deref_load
||
1973 instr
->intrinsic
== nir_intrinsic_image_load
) {
1974 return nir_intrinsic_access(instr
) & ACCESS_CAN_REORDER
;
1976 const nir_intrinsic_info
*info
=
1977 &nir_intrinsic_infos
[instr
->intrinsic
];
1978 return (info
->flags
& NIR_INTRINSIC_CAN_ELIMINATE
) &&
1979 (info
->flags
& NIR_INTRINSIC_CAN_REORDER
);
1984 * \group texture information
1986 * This gives semantic information about textures which is useful to the
1987 * frontend, the backend, and lowering passes, but not the optimizer.
1992 nir_tex_src_projector
,
1993 nir_tex_src_comparator
, /* shadow comparator */
1997 nir_tex_src_min_lod
,
1998 nir_tex_src_ms_index
, /* MSAA sample index */
1999 nir_tex_src_ms_mcs
, /* MSAA compression value */
2002 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
2003 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
2004 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
2005 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
2006 nir_tex_src_texture_handle
, /* < bindless texture handle */
2007 nir_tex_src_sampler_handle
, /* < bindless sampler handle */
2008 nir_tex_src_plane
, /* < selects plane for planar textures */
2009 nir_num_tex_src_types
2014 nir_tex_src_type src_type
;
2018 nir_texop_tex
, /**< Regular texture look-up */
2019 nir_texop_txb
, /**< Texture look-up with LOD bias */
2020 nir_texop_txl
, /**< Texture look-up with explicit LOD */
2021 nir_texop_txd
, /**< Texture look-up with partial derivatives */
2022 nir_texop_txf
, /**< Texel fetch with explicit LOD */
2023 nir_texop_txf_ms
, /**< Multisample texture fetch */
2024 nir_texop_txf_ms_fb
, /**< Multisample texture fetch from framebuffer */
2025 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
2026 nir_texop_txs
, /**< Texture size */
2027 nir_texop_lod
, /**< Texture lod query */
2028 nir_texop_tg4
, /**< Texture gather */
2029 nir_texop_query_levels
, /**< Texture levels query */
2030 nir_texop_texture_samples
, /**< Texture samples query */
2031 nir_texop_samples_identical
, /**< Query whether all samples are definitely
2034 nir_texop_tex_prefetch
, /**< Regular texture look-up, eligible for pre-dispatch */
2035 nir_texop_fragment_fetch
, /**< Multisample fragment color texture fetch */
2036 nir_texop_fragment_mask_fetch
,/**< Multisample fragment mask texture fetch */
2042 enum glsl_sampler_dim sampler_dim
;
2043 nir_alu_type dest_type
;
2048 unsigned num_srcs
, coord_components
;
2049 bool is_array
, is_shadow
;
2052 * If is_shadow is true, whether this is the old-style shadow that outputs 4
2053 * components or the new-style shadow that outputs 1 component.
2055 bool is_new_style_shadow
;
2057 /* gather component selector */
2058 unsigned component
: 2;
2060 /* gather offsets */
2061 int8_t tg4_offsets
[4][2];
2063 /* True if the texture index or handle is not dynamically uniform */
2064 bool texture_non_uniform
;
2066 /* True if the sampler index or handle is not dynamically uniform */
2067 bool sampler_non_uniform
;
2069 /** The texture index
2071 * If this texture instruction has a nir_tex_src_texture_offset source,
2072 * then the texture index is given by texture_index + texture_offset.
2074 unsigned texture_index
;
2076 /** The sampler index
2078 * The following operations do not require a sampler and, as such, this
2079 * field should be ignored:
2081 * - nir_texop_txf_ms
2084 * - nir_texop_query_levels
2085 * - nir_texop_texture_samples
2086 * - nir_texop_samples_identical
2088 * If this texture instruction has a nir_tex_src_sampler_offset source,
2089 * then the sampler index is given by sampler_index + sampler_offset.
2091 unsigned sampler_index
;
2095 * Returns true if the texture operation requires a sampler as a general rule,
2096 * see the documentation of sampler_index.
2098 * Note that the specific hw/driver backend could require to a sampler
2099 * object/configuration packet in any case, for some other reason.
2102 nir_tex_instr_need_sampler(const nir_tex_instr
*instr
)
2104 switch (instr
->op
) {
2106 case nir_texop_txf_ms
:
2109 case nir_texop_query_levels
:
2110 case nir_texop_texture_samples
:
2111 case nir_texop_samples_identical
:
2118 static inline unsigned
2119 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
2121 switch (instr
->op
) {
2122 case nir_texop_txs
: {
2124 switch (instr
->sampler_dim
) {
2125 case GLSL_SAMPLER_DIM_1D
:
2126 case GLSL_SAMPLER_DIM_BUF
:
2129 case GLSL_SAMPLER_DIM_2D
:
2130 case GLSL_SAMPLER_DIM_CUBE
:
2131 case GLSL_SAMPLER_DIM_MS
:
2132 case GLSL_SAMPLER_DIM_RECT
:
2133 case GLSL_SAMPLER_DIM_EXTERNAL
:
2134 case GLSL_SAMPLER_DIM_SUBPASS
:
2137 case GLSL_SAMPLER_DIM_3D
:
2141 unreachable("not reached");
2143 if (instr
->is_array
)
2151 case nir_texop_texture_samples
:
2152 case nir_texop_query_levels
:
2153 case nir_texop_samples_identical
:
2154 case nir_texop_fragment_mask_fetch
:
2158 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
2165 /* Returns true if this texture operation queries something about the texture
2166 * rather than actually sampling it.
2169 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
2171 switch (instr
->op
) {
2174 case nir_texop_texture_samples
:
2175 case nir_texop_query_levels
:
2176 case nir_texop_txf_ms_mcs
:
2183 case nir_texop_txf_ms
:
2184 case nir_texop_txf_ms_fb
:
2188 unreachable("Invalid texture opcode");
2193 nir_tex_instr_has_implicit_derivative(const nir_tex_instr
*instr
)
2195 switch (instr
->op
) {
2205 static inline nir_alu_type
2206 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
2208 switch (instr
->src
[src
].src_type
) {
2209 case nir_tex_src_coord
:
2210 switch (instr
->op
) {
2212 case nir_texop_txf_ms
:
2213 case nir_texop_txf_ms_fb
:
2214 case nir_texop_txf_ms_mcs
:
2215 case nir_texop_samples_identical
:
2216 return nir_type_int
;
2219 return nir_type_float
;
2222 case nir_tex_src_lod
:
2223 switch (instr
->op
) {
2226 return nir_type_int
;
2229 return nir_type_float
;
2232 case nir_tex_src_projector
:
2233 case nir_tex_src_comparator
:
2234 case nir_tex_src_bias
:
2235 case nir_tex_src_min_lod
:
2236 case nir_tex_src_ddx
:
2237 case nir_tex_src_ddy
:
2238 return nir_type_float
;
2240 case nir_tex_src_offset
:
2241 case nir_tex_src_ms_index
:
2242 case nir_tex_src_plane
:
2243 return nir_type_int
;
2245 case nir_tex_src_ms_mcs
:
2246 case nir_tex_src_texture_deref
:
2247 case nir_tex_src_sampler_deref
:
2248 case nir_tex_src_texture_offset
:
2249 case nir_tex_src_sampler_offset
:
2250 case nir_tex_src_texture_handle
:
2251 case nir_tex_src_sampler_handle
:
2252 return nir_type_uint
;
2254 case nir_num_tex_src_types
:
2255 unreachable("nir_num_tex_src_types is not a valid source type");
2258 unreachable("Invalid texture source type");
2261 static inline unsigned
2262 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
2264 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
2265 return instr
->coord_components
;
2267 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
2268 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
2271 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
2272 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
2273 if (instr
->is_array
)
2274 return instr
->coord_components
- 1;
2276 return instr
->coord_components
;
2279 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
2280 * the offset, since a cube maps to a single face.
2282 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
2283 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2285 else if (instr
->is_array
)
2286 return instr
->coord_components
- 1;
2288 return instr
->coord_components
;
2295 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
2297 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
2298 if (instr
->src
[i
].src_type
== type
)
2304 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
2305 nir_tex_src_type src_type
,
2308 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
2310 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
);
2317 nir_const_value value
[];
2318 } nir_load_const_instr
;
2321 /** Return from a function
2323 * This instruction is a classic function return. It jumps to
2324 * nir_function_impl::end_block. No return value is provided in this
2325 * instruction. Instead, the function is expected to write any return
2326 * data to a deref passed in from the caller.
2330 /** Break out of the inner-most loop
2332 * This has the same semantics as C's "break" statement.
2336 /** Jump back to the top of the inner-most loop
2338 * This has the same semantics as C's "continue" statement assuming that a
2339 * NIR loop is implemented as "while (1) { body }".
2343 /** Jumps for unstructured CFG.
2345 * As within an unstructured CFG we can't rely on block ordering we need to
2346 * place explicit jumps at the end of every block.
2356 struct nir_block
*target
;
2357 struct nir_block
*else_target
;
2360 /* creates a new SSA variable in an undefined state */
2365 } nir_ssa_undef_instr
;
2368 struct exec_node node
;
2370 /* The predecessor block corresponding to this source */
2371 struct nir_block
*pred
;
2376 #define nir_foreach_phi_src(phi_src, phi) \
2377 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
2378 #define nir_foreach_phi_src_safe(phi_src, phi) \
2379 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
2384 struct exec_list srcs
; /** < list of nir_phi_src */
2390 struct exec_node node
;
2393 } nir_parallel_copy_entry
;
2395 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
2396 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
2401 /* A list of nir_parallel_copy_entrys. The sources of all of the
2402 * entries are copied to the corresponding destinations "in parallel".
2403 * In other words, if we have two entries: a -> b and b -> a, the values
2406 struct exec_list entries
;
2407 } nir_parallel_copy_instr
;
2409 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
2410 type
, nir_instr_type_alu
)
2411 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
2412 type
, nir_instr_type_deref
)
2413 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
2414 type
, nir_instr_type_call
)
2415 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
2416 type
, nir_instr_type_jump
)
2417 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
2418 type
, nir_instr_type_tex
)
2419 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
2420 type
, nir_instr_type_intrinsic
)
2421 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
2422 type
, nir_instr_type_load_const
)
2423 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
2424 type
, nir_instr_type_ssa_undef
)
2425 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
2426 type
, nir_instr_type_phi
)
2427 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
2428 nir_parallel_copy_instr
, instr
,
2429 type
, nir_instr_type_parallel_copy
)
2432 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
2433 static inline type \
2434 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2436 assert(nir_src_is_const(src)); \
2437 nir_load_const_instr *load = \
2438 nir_instr_as_load_const(src.ssa->parent_instr); \
2439 assert(comp < load->def.num_components); \
2440 return nir_const_value_as_##suffix(load->value[comp], \
2441 load->def.bit_size); \
2444 static inline type \
2445 nir_src_as_##suffix(nir_src src) \
2447 assert(nir_src_num_components(src) == 1); \
2448 return nir_src_comp_as_##suffix(src, 0); \
2451 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2452 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint
)
2453 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2454 NIR_DEFINE_SRC_AS_CONST(double, float)
2456 #undef NIR_DEFINE_SRC_AS_CONST
2465 nir_ssa_scalar_is_const(nir_ssa_scalar s
)
2467 return s
.def
->parent_instr
->type
== nir_instr_type_load_const
;
2470 static inline nir_const_value
2471 nir_ssa_scalar_as_const_value(nir_ssa_scalar s
)
2473 assert(s
.comp
< s
.def
->num_components
);
2474 nir_load_const_instr
*load
= nir_instr_as_load_const(s
.def
->parent_instr
);
2475 return load
->value
[s
.comp
];
2478 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2479 static inline type \
2480 nir_ssa_scalar_as_##suffix(nir_ssa_scalar s) \
2482 return nir_const_value_as_##suffix( \
2483 nir_ssa_scalar_as_const_value(s), s.def->bit_size); \
2486 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2487 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint
)
2488 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2489 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2491 #undef NIR_DEFINE_SCALAR_AS_CONST
2494 nir_ssa_scalar_is_alu(nir_ssa_scalar s
)
2496 return s
.def
->parent_instr
->type
== nir_instr_type_alu
;
2499 static inline nir_op
2500 nir_ssa_scalar_alu_op(nir_ssa_scalar s
)
2502 return nir_instr_as_alu(s
.def
->parent_instr
)->op
;
2505 static inline nir_ssa_scalar
2506 nir_ssa_scalar_chase_alu_src(nir_ssa_scalar s
, unsigned alu_src_idx
)
2508 nir_ssa_scalar out
= { NULL
, 0 };
2510 nir_alu_instr
*alu
= nir_instr_as_alu(s
.def
->parent_instr
);
2511 assert(alu_src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
2513 /* Our component must be written */
2514 assert(s
.comp
< s
.def
->num_components
);
2515 assert(alu
->dest
.write_mask
& (1u << s
.comp
));
2517 assert(alu
->src
[alu_src_idx
].src
.is_ssa
);
2518 out
.def
= alu
->src
[alu_src_idx
].src
.ssa
;
2520 if (nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 0) {
2521 /* The ALU src is unsized so the source component follows the
2522 * destination component.
2524 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[s
.comp
];
2526 /* This is a sized source so all source components work together to
2527 * produce all the destination components. Since we need to return a
2528 * scalar, this only works if the source is a scalar.
2530 assert(nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 1);
2531 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[0];
2533 assert(out
.comp
< out
.def
->num_components
);
2542 * Control flow consists of a tree of control flow nodes, which include
2543 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2544 * instructions that always run start-to-finish. Each basic block also keeps
2545 * track of its successors (blocks which may run immediately after the current
2546 * block) and predecessors (blocks which could have run immediately before the
2547 * current block). Each function also has a start block and an end block which
2548 * all return statements point to (which is always empty). Together, all the
2549 * blocks with their predecessors and successors make up the control flow
2550 * graph (CFG) of the function. There are helpers that modify the tree of
2551 * control flow nodes while modifying the CFG appropriately; these should be
2552 * used instead of modifying the tree directly.
2559 nir_cf_node_function
2562 typedef struct nir_cf_node
{
2563 struct exec_node node
;
2564 nir_cf_node_type type
;
2565 struct nir_cf_node
*parent
;
2568 typedef struct nir_block
{
2569 nir_cf_node cf_node
;
2571 struct exec_list instr_list
; /** < list of nir_instr */
2573 /** generic block index; generated by nir_index_blocks */
2577 * Each block can only have up to 2 successors, so we put them in a simple
2578 * array - no need for anything more complicated.
2580 struct nir_block
*successors
[2];
2582 /* Set of nir_block predecessors in the CFG */
2583 struct set
*predecessors
;
2586 * this node's immediate dominator in the dominance tree - set to NULL for
2589 struct nir_block
*imm_dom
;
2591 /* This node's children in the dominance tree */
2592 unsigned num_dom_children
;
2593 struct nir_block
**dom_children
;
2595 /* Set of nir_blocks on the dominance frontier of this block */
2596 struct set
*dom_frontier
;
2599 * These two indices have the property that dom_{pre,post}_index for each
2600 * child of this block in the dominance tree will always be between
2601 * dom_pre_index and dom_post_index for this block, which makes testing if
2602 * a given block is dominated by another block an O(1) operation.
2604 int16_t dom_pre_index
, dom_post_index
;
2606 /* live in and out for this block; used for liveness analysis */
2607 BITSET_WORD
*live_in
;
2608 BITSET_WORD
*live_out
;
2612 nir_block_is_reachable(nir_block
*b
)
2614 /* See also nir_block_dominates */
2615 return b
->dom_post_index
!= -1;
2618 static inline nir_instr
*
2619 nir_block_first_instr(nir_block
*block
)
2621 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
2622 return exec_node_data(nir_instr
, head
, node
);
2625 static inline nir_instr
*
2626 nir_block_last_instr(nir_block
*block
)
2628 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
2629 return exec_node_data(nir_instr
, tail
, node
);
2633 nir_block_ends_in_jump(nir_block
*block
)
2635 return !exec_list_is_empty(&block
->instr_list
) &&
2636 nir_block_last_instr(block
)->type
== nir_instr_type_jump
;
2639 #define nir_foreach_instr(instr, block) \
2640 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
2641 #define nir_foreach_instr_reverse(instr, block) \
2642 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
2643 #define nir_foreach_instr_safe(instr, block) \
2644 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
2645 #define nir_foreach_instr_reverse_safe(instr, block) \
2646 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
2649 nir_selection_control_none
= 0x0,
2650 nir_selection_control_flatten
= 0x1,
2651 nir_selection_control_dont_flatten
= 0x2,
2652 } nir_selection_control
;
2654 typedef struct nir_if
{
2655 nir_cf_node cf_node
;
2657 nir_selection_control control
;
2659 struct exec_list then_list
; /** < list of nir_cf_node */
2660 struct exec_list else_list
; /** < list of nir_cf_node */
2666 /** Instruction that generates nif::condition. */
2667 nir_instr
*conditional_instr
;
2669 /** Block within ::nif that has the break instruction. */
2670 nir_block
*break_block
;
2672 /** Last block for the then- or else-path that does not contain the break. */
2673 nir_block
*continue_from_block
;
2675 /** True when ::break_block is in the else-path of ::nif. */
2676 bool continue_from_then
;
2679 /* This is true if the terminators exact trip count is unknown. For
2682 * for (int i = 0; i < imin(x, 4); i++)
2685 * Here loop analysis would have set a max_trip_count of 4 however we dont
2686 * know for sure that this is the exact trip count.
2688 bool exact_trip_count_unknown
;
2690 struct list_head loop_terminator_link
;
2691 } nir_loop_terminator
;
2694 /* Estimated cost (in number of instructions) of the loop */
2695 unsigned instr_cost
;
2697 /* Guessed trip count based on array indexing */
2698 unsigned guessed_trip_count
;
2700 /* Maximum number of times the loop is run (if known) */
2701 unsigned max_trip_count
;
2703 /* Do we know the exact number of times the loop will be run */
2704 bool exact_trip_count_known
;
2706 /* Unroll the loop regardless of its size */
2709 /* Does the loop contain complex loop terminators, continues or other
2710 * complex behaviours? If this is true we can't rely on
2711 * loop_terminator_list to be complete or accurate.
2715 nir_loop_terminator
*limiting_terminator
;
2717 /* A list of loop_terminators terminating this loop. */
2718 struct list_head loop_terminator_list
;
2722 nir_loop_control_none
= 0x0,
2723 nir_loop_control_unroll
= 0x1,
2724 nir_loop_control_dont_unroll
= 0x2,
2728 nir_cf_node cf_node
;
2730 struct exec_list body
; /** < list of nir_cf_node */
2732 nir_loop_info
*info
;
2733 nir_loop_control control
;
2734 bool partially_unrolled
;
2738 * Various bits of metadata that can may be created or required by
2739 * optimization and analysis passes
2742 nir_metadata_none
= 0x0,
2744 /** Indicates that nir_block::index values are valid.
2746 * The start block has index 0 and they increase through a natural walk of
2747 * the CFG. nir_function_impl::num_blocks is the number of blocks and
2748 * every block index is in the range [0, nir_function_impl::num_blocks].
2750 * A pass can preserve this metadata type if it doesn't touch the CFG.
2752 nir_metadata_block_index
= 0x1,
2754 /** Indicates that block dominance information is valid
2758 * - nir_block::num_dom_children
2759 * - nir_block::dom_children
2760 * - nir_block::dom_frontier
2761 * - nir_block::dom_pre_index
2762 * - nir_block::dom_post_index
2764 * A pass can preserve this metadata type if it doesn't touch the CFG.
2766 nir_metadata_dominance
= 0x2,
2768 /** Indicates that SSA def data-flow liveness information is valid
2772 * - nir_ssa_def::live_index
2773 * - nir_block::live_in
2774 * - nir_block::live_out
2776 * A pass can preserve this metadata type if it never adds or removes any
2777 * SSA defs (most passes shouldn't preserve this metadata type).
2779 nir_metadata_live_ssa_defs
= 0x4,
2781 /** A dummy metadata value to track when a pass forgot to call
2782 * nir_metadata_preserve.
2784 * A pass should always clear this value even if it doesn't make any
2785 * progress to indicate that it thought about preserving metadata.
2787 nir_metadata_not_properly_reset
= 0x8,
2789 /** Indicates that loop analysis information is valid.
2791 * This includes everything pointed to by nir_loop::info.
2793 * A pass can preserve this metadata type if it is guaranteed to not affect
2794 * any loop metadata. However, since loop metadata includes things like
2795 * loop counts which depend on arithmetic in the loop, this is very hard to
2796 * determine. Most passes shouldn't preserve this metadata type.
2798 nir_metadata_loop_analysis
= 0x10,
2802 * This includes all nir_metadata flags except not_properly_reset. Passes
2803 * which do not change the shader in any way should call
2805 * nir_metadata_preserve(impl, nir_metadata_all);
2807 nir_metadata_all
= ~nir_metadata_not_properly_reset
,
2811 nir_cf_node cf_node
;
2813 /** pointer to the function of which this is an implementation */
2814 struct nir_function
*function
;
2816 struct exec_list body
; /** < list of nir_cf_node */
2818 nir_block
*end_block
;
2820 /** list for all local variables in the function */
2821 struct exec_list locals
;
2823 /** list of local registers in the function */
2824 struct exec_list registers
;
2826 /** next available local register index */
2829 /** next available SSA value index */
2832 /* total number of basic blocks, only valid when block_index_dirty = false */
2833 unsigned num_blocks
;
2835 /** True if this nir_function_impl uses structured control-flow
2837 * Structured nir_function_impls have different validation rules.
2841 nir_metadata valid_metadata
;
2842 } nir_function_impl
;
2844 #define nir_foreach_function_temp_variable(var, impl) \
2845 foreach_list_typed(nir_variable, var, node, &(impl)->locals)
2847 #define nir_foreach_function_temp_variable_safe(var, impl) \
2848 foreach_list_typed_safe(nir_variable, var, node, &(impl)->locals)
2850 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2851 nir_start_block(nir_function_impl
*impl
)
2853 return (nir_block
*) impl
->body
.head_sentinel
.next
;
2856 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2857 nir_impl_last_block(nir_function_impl
*impl
)
2859 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
2862 static inline nir_cf_node
*
2863 nir_cf_node_next(nir_cf_node
*node
)
2865 struct exec_node
*next
= exec_node_get_next(&node
->node
);
2866 if (exec_node_is_tail_sentinel(next
))
2869 return exec_node_data(nir_cf_node
, next
, node
);
2872 static inline nir_cf_node
*
2873 nir_cf_node_prev(nir_cf_node
*node
)
2875 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
2876 if (exec_node_is_head_sentinel(prev
))
2879 return exec_node_data(nir_cf_node
, prev
, node
);
2883 nir_cf_node_is_first(const nir_cf_node
*node
)
2885 return exec_node_is_head_sentinel(node
->node
.prev
);
2889 nir_cf_node_is_last(const nir_cf_node
*node
)
2891 return exec_node_is_tail_sentinel(node
->node
.next
);
2894 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
2895 type
, nir_cf_node_block
)
2896 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
2897 type
, nir_cf_node_if
)
2898 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
2899 type
, nir_cf_node_loop
)
2900 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
2901 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
2903 static inline nir_block
*
2904 nir_if_first_then_block(nir_if
*if_stmt
)
2906 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
2907 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2910 static inline nir_block
*
2911 nir_if_last_then_block(nir_if
*if_stmt
)
2913 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
2914 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2917 static inline nir_block
*
2918 nir_if_first_else_block(nir_if
*if_stmt
)
2920 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
2921 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2924 static inline nir_block
*
2925 nir_if_last_else_block(nir_if
*if_stmt
)
2927 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
2928 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2931 static inline nir_block
*
2932 nir_loop_first_block(nir_loop
*loop
)
2934 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
2935 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2938 static inline nir_block
*
2939 nir_loop_last_block(nir_loop
*loop
)
2941 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
2942 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2946 * Return true if this list of cf_nodes contains a single empty block.
2949 nir_cf_list_is_empty_block(struct exec_list
*cf_list
)
2951 if (exec_list_is_singular(cf_list
)) {
2952 struct exec_node
*head
= exec_list_get_head(cf_list
);
2954 nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2955 return exec_list_is_empty(&block
->instr_list
);
2961 uint8_t num_components
;
2965 typedef struct nir_function
{
2966 struct exec_node node
;
2969 struct nir_shader
*shader
;
2971 unsigned num_params
;
2972 nir_parameter
*params
;
2974 /** The implementation of this function.
2976 * If the function is only declared and not implemented, this is NULL.
2978 nir_function_impl
*impl
;
2984 nir_lower_imul64
= (1 << 0),
2985 nir_lower_isign64
= (1 << 1),
2986 /** Lower all int64 modulus and division opcodes */
2987 nir_lower_divmod64
= (1 << 2),
2988 /** Lower all 64-bit umul_high and imul_high opcodes */
2989 nir_lower_imul_high64
= (1 << 3),
2990 nir_lower_mov64
= (1 << 4),
2991 nir_lower_icmp64
= (1 << 5),
2992 nir_lower_iadd64
= (1 << 6),
2993 nir_lower_iabs64
= (1 << 7),
2994 nir_lower_ineg64
= (1 << 8),
2995 nir_lower_logic64
= (1 << 9),
2996 nir_lower_minmax64
= (1 << 10),
2997 nir_lower_shift64
= (1 << 11),
2998 nir_lower_imul_2x32_64
= (1 << 12),
2999 nir_lower_extract64
= (1 << 13),
3000 nir_lower_ufind_msb64
= (1 << 14),
3001 } nir_lower_int64_options
;
3004 nir_lower_drcp
= (1 << 0),
3005 nir_lower_dsqrt
= (1 << 1),
3006 nir_lower_drsq
= (1 << 2),
3007 nir_lower_dtrunc
= (1 << 3),
3008 nir_lower_dfloor
= (1 << 4),
3009 nir_lower_dceil
= (1 << 5),
3010 nir_lower_dfract
= (1 << 6),
3011 nir_lower_dround_even
= (1 << 7),
3012 nir_lower_dmod
= (1 << 8),
3013 nir_lower_dsub
= (1 << 9),
3014 nir_lower_ddiv
= (1 << 10),
3015 nir_lower_fp64_full_software
= (1 << 11),
3016 } nir_lower_doubles_options
;
3019 nir_divergence_single_prim_per_subgroup
= (1 << 0),
3020 nir_divergence_single_patch_per_tcs_subgroup
= (1 << 1),
3021 nir_divergence_single_patch_per_tes_subgroup
= (1 << 2),
3022 nir_divergence_view_index_uniform
= (1 << 3),
3023 } nir_divergence_options
;
3025 typedef struct nir_shader_compiler_options
{
3031 /** Lowers flrp when it does not support doubles */
3038 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
3039 bool lower_bitfield_extract
;
3040 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
3041 bool lower_bitfield_extract_to_shifts
;
3042 /** Lowers bitfield_insert to bfi/bfm */
3043 bool lower_bitfield_insert
;
3044 /** Lowers bitfield_insert to compares, and shifts. */
3045 bool lower_bitfield_insert_to_shifts
;
3046 /** Lowers bitfield_insert to bfm/bitfield_select. */
3047 bool lower_bitfield_insert_to_bitfield_select
;
3048 /** Lowers bitfield_reverse to shifts. */
3049 bool lower_bitfield_reverse
;
3050 /** Lowers bit_count to shifts. */
3051 bool lower_bit_count
;
3052 /** Lowers ifind_msb to compare and ufind_msb */
3053 bool lower_ifind_msb
;
3054 /** Lowers find_lsb to ufind_msb and logic ops */
3055 bool lower_find_lsb
;
3056 bool lower_uadd_carry
;
3057 bool lower_usub_borrow
;
3058 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
3059 bool lower_mul_high
;
3060 /** lowers fneg and ineg to fsub and isub. */
3062 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
3065 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fneu} + b2f: */
3068 /* lower fall_equalN/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
3069 bool lower_vector_cmp
;
3071 /** enables rules to lower idiv by power-of-two: */
3074 /** enable rules to avoid bit ops */
3077 /** enables rules to lower isign to imin+imax */
3080 /** enables rules to lower fsign to fsub and flt */
3083 /** enables rules to lower iabs to ineg+imax */
3086 /* lower fdph to fdot4 */
3089 /** lower fdot to fmul and fsum/fadd. */
3092 /* Does the native fdot instruction replicate its result for four
3093 * components? If so, then opt_algebraic_late will turn all fdotN
3094 * instructions into fdot_replicatedN instructions.
3096 bool fdot_replicates
;
3098 /** lowers ffloor to fsub+ffract: */
3101 /** lowers ffract to fsub+ffloor: */
3104 /** lowers fceil to fneg+ffloor+fneg: */
3111 bool lower_pack_half_2x16
;
3112 bool lower_pack_unorm_2x16
;
3113 bool lower_pack_snorm_2x16
;
3114 bool lower_pack_unorm_4x8
;
3115 bool lower_pack_snorm_4x8
;
3116 bool lower_pack_64_2x32_split
;
3117 bool lower_pack_32_2x16_split
;
3118 bool lower_unpack_half_2x16
;
3119 bool lower_unpack_unorm_2x16
;
3120 bool lower_unpack_snorm_2x16
;
3121 bool lower_unpack_unorm_4x8
;
3122 bool lower_unpack_snorm_4x8
;
3123 bool lower_unpack_64_2x32_split
;
3124 bool lower_unpack_32_2x16_split
;
3126 bool lower_pack_split
;
3128 bool lower_extract_byte
;
3129 bool lower_extract_word
;
3131 bool lower_all_io_to_temps
;
3132 bool lower_all_io_to_elements
;
3134 /* Indicates that the driver only has zero-based vertex id */
3135 bool vertex_id_zero_based
;
3138 * If enabled, gl_BaseVertex will be lowered as:
3139 * is_indexed_draw (~0/0) & firstvertex
3141 bool lower_base_vertex
;
3144 * If enabled, gl_HelperInvocation will be lowered as:
3146 * !((1 << sample_id) & sample_mask_in))
3148 * This depends on some possibly hw implementation details, which may
3149 * not be true for all hw. In particular that the FS is only executed
3150 * for covered samples or for helper invocations. So, do not blindly
3151 * enable this option.
3153 * Note: See also issue #22 in ARB_shader_image_load_store
3155 bool lower_helper_invocation
;
3158 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
3160 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
3161 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
3163 bool optimize_sample_mask_in
;
3165 bool lower_cs_local_index_from_id
;
3166 bool lower_cs_local_id_from_index
;
3168 /* Prevents lowering global_invocation_id to be in terms of work_group_id */
3169 bool has_cs_global_id
;
3171 bool lower_device_index_to_zero
;
3173 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
3174 bool lower_wpos_pntc
;
3177 * Set if nir_op_[iu]hadd and nir_op_[iu]rhadd instructions should be
3178 * lowered to simple arithmetic.
3180 * If this flag is set, the lowering will be applied to all bit-sizes of
3181 * these instructions.
3183 * \sa ::lower_hadd64
3188 * Set if only 64-bit nir_op_[iu]hadd and nir_op_[iu]rhadd instructions
3189 * should be lowered to simple arithmetic.
3191 * If this flag is set, the lowering will be applied to only 64-bit
3192 * versions of these instructions.
3199 * Set if nir_op_add_sat and nir_op_usub_sat should be lowered to simple
3202 * If this flag is set, the lowering will be applied to all bit-sizes of
3203 * these instructions.
3205 * \sa ::lower_usub_sat64
3210 * Set if only 64-bit nir_op_usub_sat should be lowered to simple
3213 * \sa ::lower_add_sat
3215 bool lower_usub_sat64
;
3218 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
3219 * for IO purposes and would prefer loads/stores be vectorized.
3222 bool lower_to_scalar
;
3225 * Whether nir_opt_vectorize should only create 16-bit 2D vectors.
3227 bool vectorize_vec2_16bit
;
3230 * Should the linker unify inputs_read/outputs_written between adjacent
3231 * shader stages which are linked into a single program?
3233 bool unify_interfaces
;
3236 * Should nir_lower_io() create load_interpolated_input intrinsics?
3238 * If not, it generates regular load_input intrinsics and interpolation
3239 * information must be inferred from the list of input nir_variables.
3241 bool use_interpolated_input_intrinsics
;
3243 /* Lowers when 32x32->64 bit multiplication is not supported */
3244 bool lower_mul_2x32_64
;
3246 /* Lowers when rotate instruction is not supported */
3250 * Backend supports imul24, and would like to use it (when possible)
3251 * for address/offset calculation. If true, driver should call
3252 * nir_lower_amul(). (If not set, amul will automatically be lowered
3257 /** Backend supports umul24, if not set umul24 will automatically be lowered
3258 * to imul with masked inputs */
3261 /** Backend supports umad24, if not set umad24 will automatically be lowered
3262 * to imul with masked inputs and iadd */
3265 /* Whether to generate only scoped_barrier intrinsics instead of the set of
3266 * memory and control barrier intrinsics based on GLSL.
3268 bool use_scoped_barrier
;
3271 * Is this the Intel vec4 backend?
3273 * Used to inhibit algebraic optimizations that are known to be harmful on
3274 * the Intel vec4 backend. This is generally applicable to any
3275 * optimization that might cause more immediate values to be used in
3276 * 3-source (e.g., ffma and flrp) instructions.
3280 /** Lower nir_op_ibfe and nir_op_ubfe that have two constant sources. */
3281 bool lower_bfe_with_two_constants
;
3283 /** Whether 8-bit ALU is supported. */
3284 bool support_8bit_alu
;
3286 /** Whether 16-bit ALU is supported. */
3287 bool support_16bit_alu
;
3289 unsigned max_unroll_iterations
;
3291 nir_lower_int64_options lower_int64_options
;
3292 nir_lower_doubles_options lower_doubles_options
;
3293 } nir_shader_compiler_options
;
3295 typedef struct nir_shader
{
3296 /** list of uniforms (nir_variable) */
3297 struct exec_list variables
;
3299 /** Set of driver-specific options for the shader.
3301 * The memory for the options is expected to be kept in a single static
3302 * copy by the driver.
3304 const struct nir_shader_compiler_options
*options
;
3306 /** Various bits of compile-time information about a given shader */
3307 struct shader_info info
;
3309 struct exec_list functions
; /** < list of nir_function */
3312 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
3315 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
3317 /** Size in bytes of required scratch space */
3318 unsigned scratch_size
;
3320 /** Constant data associated with this shader.
3322 * Constant data is loaded through load_constant intrinsics (as compared to
3323 * the NIR load_const instructions which have the constant value inlined
3324 * into them). This is usually generated by nir_opt_large_constants (so
3325 * shaders don't have to load_const into a temporary array when they want
3326 * to indirect on a const array).
3328 void *constant_data
;
3329 /** Size of the constant data associated with the shader, in bytes */
3330 unsigned constant_data_size
;
3333 #define nir_foreach_function(func, shader) \
3334 foreach_list_typed(nir_function, func, node, &(shader)->functions)
3336 static inline nir_function_impl
*
3337 nir_shader_get_entrypoint(nir_shader
*shader
)
3339 nir_function
*func
= NULL
;
3341 nir_foreach_function(function
, shader
) {
3342 assert(func
== NULL
);
3343 if (function
->is_entrypoint
) {
3354 assert(func
->num_params
== 0);
3359 nir_shader
*nir_shader_create(void *mem_ctx
,
3360 gl_shader_stage stage
,
3361 const nir_shader_compiler_options
*options
,
3364 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
3366 void nir_reg_remove(nir_register
*reg
);
3368 /** Adds a variable to the appropriate list in nir_shader */
3369 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
3372 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
3374 assert(var
->data
.mode
== nir_var_function_temp
);
3375 exec_list_push_tail(&impl
->locals
, &var
->node
);
3378 /** creates a variable, sets a few defaults, and adds it to the list */
3379 nir_variable
*nir_variable_create(nir_shader
*shader
,
3380 nir_variable_mode mode
,
3381 const struct glsl_type
*type
,
3383 /** creates a local variable and adds it to the list */
3384 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
3385 const struct glsl_type
*type
,
3388 nir_variable
*nir_find_variable_with_location(nir_shader
*shader
,
3389 nir_variable_mode mode
,
3392 nir_variable
*nir_find_variable_with_driver_location(nir_shader
*shader
,
3393 nir_variable_mode mode
,
3396 /** creates a function and adds it to the shader's list of functions */
3397 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
3399 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
3400 /** creates a function_impl that isn't tied to any particular function */
3401 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
3403 nir_block
*nir_block_create(nir_shader
*shader
);
3404 nir_if
*nir_if_create(nir_shader
*shader
);
3405 nir_loop
*nir_loop_create(nir_shader
*shader
);
3407 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
3409 /** requests that the given pieces of metadata be generated */
3410 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
3411 /** dirties all but the preserved metadata */
3412 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
3413 /** Preserves all metadata for the given shader */
3414 void nir_shader_preserve_all_metadata(nir_shader
*shader
);
3416 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
3417 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
3419 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
3420 nir_deref_type deref_type
);
3422 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
3424 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
3425 unsigned num_components
,
3428 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
3429 nir_intrinsic_op op
);
3431 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
3432 nir_function
*callee
);
3434 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
3436 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
3438 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
3440 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
3441 unsigned num_components
,
3444 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
3447 * NIR Cursors and Instruction Insertion API
3450 * A tiny struct representing a point to insert/extract instructions or
3451 * control flow nodes. Helps reduce the combinatorial explosion of possible
3452 * points to insert/extract.
3454 * \sa nir_control_flow.h
3457 nir_cursor_before_block
,
3458 nir_cursor_after_block
,
3459 nir_cursor_before_instr
,
3460 nir_cursor_after_instr
,
3461 } nir_cursor_option
;
3464 nir_cursor_option option
;
3471 static inline nir_block
*
3472 nir_cursor_current_block(nir_cursor cursor
)
3474 if (cursor
.option
== nir_cursor_before_instr
||
3475 cursor
.option
== nir_cursor_after_instr
) {
3476 return cursor
.instr
->block
;
3478 return cursor
.block
;
3482 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
3484 static inline nir_cursor
3485 nir_before_block(nir_block
*block
)
3488 cursor
.option
= nir_cursor_before_block
;
3489 cursor
.block
= block
;
3493 static inline nir_cursor
3494 nir_after_block(nir_block
*block
)
3497 cursor
.option
= nir_cursor_after_block
;
3498 cursor
.block
= block
;
3502 static inline nir_cursor
3503 nir_before_instr(nir_instr
*instr
)
3506 cursor
.option
= nir_cursor_before_instr
;
3507 cursor
.instr
= instr
;
3511 static inline nir_cursor
3512 nir_after_instr(nir_instr
*instr
)
3515 cursor
.option
= nir_cursor_after_instr
;
3516 cursor
.instr
= instr
;
3520 static inline nir_cursor
3521 nir_after_block_before_jump(nir_block
*block
)
3523 nir_instr
*last_instr
= nir_block_last_instr(block
);
3524 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
3525 return nir_before_instr(last_instr
);
3527 return nir_after_block(block
);
3531 static inline nir_cursor
3532 nir_before_src(nir_src
*src
, bool is_if_condition
)
3534 if (is_if_condition
) {
3535 nir_block
*prev_block
=
3536 nir_cf_node_as_block(nir_cf_node_prev(&src
->parent_if
->cf_node
));
3537 assert(!nir_block_ends_in_jump(prev_block
));
3538 return nir_after_block(prev_block
);
3539 } else if (src
->parent_instr
->type
== nir_instr_type_phi
) {
3541 nir_phi_instr
*cond_phi
= nir_instr_as_phi(src
->parent_instr
);
3543 nir_foreach_phi_src(phi_src
, cond_phi
) {
3544 if (phi_src
->src
.ssa
== src
->ssa
) {
3551 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
3552 * to have a more specific name.
3554 nir_phi_src
*phi_src
= LIST_ENTRY(nir_phi_src
, src
, src
);
3555 return nir_after_block_before_jump(phi_src
->pred
);
3557 return nir_before_instr(src
->parent_instr
);
3561 static inline nir_cursor
3562 nir_before_cf_node(nir_cf_node
*node
)
3564 if (node
->type
== nir_cf_node_block
)
3565 return nir_before_block(nir_cf_node_as_block(node
));
3567 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
3570 static inline nir_cursor
3571 nir_after_cf_node(nir_cf_node
*node
)
3573 if (node
->type
== nir_cf_node_block
)
3574 return nir_after_block(nir_cf_node_as_block(node
));
3576 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
3579 static inline nir_cursor
3580 nir_after_phis(nir_block
*block
)
3582 nir_foreach_instr(instr
, block
) {
3583 if (instr
->type
!= nir_instr_type_phi
)
3584 return nir_before_instr(instr
);
3586 return nir_after_block(block
);
3589 static inline nir_cursor
3590 nir_after_cf_node_and_phis(nir_cf_node
*node
)
3592 if (node
->type
== nir_cf_node_block
)
3593 return nir_after_block(nir_cf_node_as_block(node
));
3595 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
3597 return nir_after_phis(block
);
3600 static inline nir_cursor
3601 nir_before_cf_list(struct exec_list
*cf_list
)
3603 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
3604 exec_list_get_head(cf_list
), node
);
3605 return nir_before_cf_node(first_node
);
3608 static inline nir_cursor
3609 nir_after_cf_list(struct exec_list
*cf_list
)
3611 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
3612 exec_list_get_tail(cf_list
), node
);
3613 return nir_after_cf_node(last_node
);
3617 * Insert a NIR instruction at the given cursor.
3619 * Note: This does not update the cursor.
3621 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
3624 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
3626 nir_instr_insert(nir_before_instr(instr
), before
);
3630 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
3632 nir_instr_insert(nir_after_instr(instr
), after
);
3636 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
3638 nir_instr_insert(nir_before_block(block
), before
);
3642 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
3644 nir_instr_insert(nir_after_block(block
), after
);
3648 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
3650 nir_instr_insert(nir_before_cf_node(node
), before
);
3654 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
3656 nir_instr_insert(nir_after_cf_node(node
), after
);
3660 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
3662 nir_instr_insert(nir_before_cf_list(list
), before
);
3666 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
3668 nir_instr_insert(nir_after_cf_list(list
), after
);
3671 void nir_instr_remove_v(nir_instr
*instr
);
3673 static inline nir_cursor
3674 nir_instr_remove(nir_instr
*instr
)
3677 nir_instr
*prev
= nir_instr_prev(instr
);
3679 cursor
= nir_after_instr(prev
);
3681 cursor
= nir_before_block(instr
->block
);
3683 nir_instr_remove_v(instr
);
3689 nir_ssa_def
*nir_instr_ssa_def(nir_instr
*instr
);
3691 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
3692 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
3693 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
3694 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
3696 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
3697 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
3698 bool nir_foreach_phi_src_leaving_block(nir_block
*instr
,
3699 nir_foreach_src_cb cb
,
3702 nir_const_value
*nir_src_as_const_value(nir_src src
);
3704 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
3705 static inline c_type * \
3706 nir_src_as_ ## name (nir_src src) \
3708 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
3709 ? cast_macro(src.ssa->parent_instr) : NULL; \
3712 NIR_SRC_AS_(alu_instr
, nir_alu_instr
, nir_instr_type_alu
, nir_instr_as_alu
)
3713 NIR_SRC_AS_(intrinsic
, nir_intrinsic_instr
,
3714 nir_instr_type_intrinsic
, nir_instr_as_intrinsic
)
3715 NIR_SRC_AS_(deref
, nir_deref_instr
, nir_instr_type_deref
, nir_instr_as_deref
)
3717 bool nir_src_is_dynamically_uniform(nir_src src
);
3718 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
3719 bool nir_instrs_equal(const nir_instr
*instr1
, const nir_instr
*instr2
);
3720 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
3721 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
3722 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
3723 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
3726 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
3727 unsigned num_components
, unsigned bit_size
,
3729 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
3730 unsigned num_components
, unsigned bit_size
,
3733 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
3734 const struct glsl_type
*type
,
3737 assert(glsl_type_is_vector_or_scalar(type
));
3738 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
3739 glsl_get_bit_size(type
), name
);
3741 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
3742 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
3743 nir_instr
*after_me
);
3745 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
3748 /** Returns the next block, disregarding structure
3750 * The ordering is deterministic but has no guarantees beyond that. In
3751 * particular, it is not guaranteed to be dominance-preserving.
3753 nir_block
*nir_block_unstructured_next(nir_block
*block
);
3754 nir_block
*nir_unstructured_start_block(nir_function_impl
*impl
);
3756 #define nir_foreach_block_unstructured(block, impl) \
3757 for (nir_block *block = nir_unstructured_start_block(impl); block != NULL; \
3758 block = nir_block_unstructured_next(block))
3760 #define nir_foreach_block_unstructured_safe(block, impl) \
3761 for (nir_block *block = nir_unstructured_start_block(impl), \
3762 *next = nir_block_unstructured_next(block); \
3764 block = next, next = nir_block_unstructured_next(block))
3767 * finds the next basic block in source-code order, returns NULL if there is
3771 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
3773 /* Performs the opposite of nir_block_cf_tree_next() */
3775 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
3777 /* Gets the first block in a CF node in source-code order */
3779 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
3781 /* Gets the last block in a CF node in source-code order */
3783 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
3785 /* Gets the next block after a CF node in source-code order */
3787 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
3789 /* Macros for loops that visit blocks in source-code order */
3791 #define nir_foreach_block(block, impl) \
3792 for (nir_block *block = nir_start_block(impl); block != NULL; \
3793 block = nir_block_cf_tree_next(block))
3795 #define nir_foreach_block_safe(block, impl) \
3796 for (nir_block *block = nir_start_block(impl), \
3797 *next = nir_block_cf_tree_next(block); \
3799 block = next, next = nir_block_cf_tree_next(block))
3801 #define nir_foreach_block_reverse(block, impl) \
3802 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
3803 block = nir_block_cf_tree_prev(block))
3805 #define nir_foreach_block_reverse_safe(block, impl) \
3806 for (nir_block *block = nir_impl_last_block(impl), \
3807 *prev = nir_block_cf_tree_prev(block); \
3809 block = prev, prev = nir_block_cf_tree_prev(block))
3811 #define nir_foreach_block_in_cf_node(block, node) \
3812 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
3813 block != nir_cf_node_cf_tree_next(node); \
3814 block = nir_block_cf_tree_next(block))
3816 /* If the following CF node is an if, this function returns that if.
3817 * Otherwise, it returns NULL.
3819 nir_if
*nir_block_get_following_if(nir_block
*block
);
3821 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
3823 void nir_index_local_regs(nir_function_impl
*impl
);
3824 void nir_index_ssa_defs(nir_function_impl
*impl
);
3825 unsigned nir_index_instrs(nir_function_impl
*impl
);
3827 void nir_index_blocks(nir_function_impl
*impl
);
3829 unsigned nir_shader_index_vars(nir_shader
*shader
, nir_variable_mode modes
);
3830 unsigned nir_function_impl_index_vars(nir_function_impl
*impl
);
3832 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
3833 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
3834 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
3835 void nir_print_deref(const nir_deref_instr
*deref
, FILE *fp
);
3837 /** Shallow clone of a single ALU instruction. */
3838 nir_alu_instr
*nir_alu_instr_clone(nir_shader
*s
, const nir_alu_instr
*orig
);
3840 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
3841 nir_function_impl
*nir_function_impl_clone(nir_shader
*shader
,
3842 const nir_function_impl
*fi
);
3843 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
3844 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
3846 void nir_shader_replace(nir_shader
*dest
, nir_shader
*src
);
3848 void nir_shader_serialize_deserialize(nir_shader
*s
);
3851 void nir_validate_shader(nir_shader
*shader
, const char *when
);
3852 void nir_metadata_set_validation_flag(nir_shader
*shader
);
3853 void nir_metadata_check_validation_flag(nir_shader
*shader
);
3856 should_skip_nir(const char *name
)
3858 static const char *list
= NULL
;
3860 /* Comma separated list of names to skip. */
3861 list
= getenv("NIR_SKIP");
3869 return comma_separated_list_contains(list
, name
);
3873 should_clone_nir(void)
3875 static int should_clone
= -1;
3876 if (should_clone
< 0)
3877 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
3879 return should_clone
;
3883 should_serialize_deserialize_nir(void)
3885 static int test_serialize
= -1;
3886 if (test_serialize
< 0)
3887 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
3889 return test_serialize
;
3893 should_print_nir(void)
3895 static int should_print
= -1;
3896 if (should_print
< 0)
3897 should_print
= env_var_as_boolean("NIR_PRINT", false);
3899 return should_print
;
3902 static inline void nir_validate_shader(nir_shader
*shader
, const char *when
) { (void) shader
; (void)when
; }
3903 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
3904 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
3905 static inline bool should_skip_nir(UNUSED
const char *pass_name
) { return false; }
3906 static inline bool should_clone_nir(void) { return false; }
3907 static inline bool should_serialize_deserialize_nir(void) { return false; }
3908 static inline bool should_print_nir(void) { return false; }
3911 #define _PASS(pass, nir, do_pass) do { \
3912 if (should_skip_nir(#pass)) { \
3913 printf("skipping %s\n", #pass); \
3917 nir_validate_shader(nir, "after " #pass); \
3918 if (should_clone_nir()) { \
3919 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3920 nir_shader_replace(nir, clone); \
3922 if (should_serialize_deserialize_nir()) { \
3923 nir_shader_serialize_deserialize(nir); \
3927 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3928 nir_metadata_set_validation_flag(nir); \
3929 if (should_print_nir()) \
3930 printf("%s\n", #pass); \
3931 if (pass(nir, ##__VA_ARGS__)) { \
3933 if (should_print_nir()) \
3934 nir_print_shader(nir, stdout); \
3935 nir_metadata_check_validation_flag(nir); \
3939 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3940 if (should_print_nir()) \
3941 printf("%s\n", #pass); \
3942 pass(nir, ##__VA_ARGS__); \
3943 if (should_print_nir()) \
3944 nir_print_shader(nir, stdout); \
3947 #define NIR_SKIP(name) should_skip_nir(#name)
3949 /** An instruction filtering callback
3951 * Returns true if the instruction should be processed and false otherwise.
3953 typedef bool (*nir_instr_filter_cb
)(const nir_instr
*, const void *);
3955 /** A simple instruction lowering callback
3957 * Many instruction lowering passes can be written as a simple function which
3958 * takes an instruction as its input and returns a sequence of instructions
3959 * that implement the consumed instruction. This function type represents
3960 * such a lowering function. When called, a function with this prototype
3961 * should either return NULL indicating that no lowering needs to be done or
3962 * emit a sequence of instructions using the provided builder (whose cursor
3963 * will already be placed after the instruction to be lowered) and return the
3964 * resulting nir_ssa_def.
3966 typedef nir_ssa_def
*(*nir_lower_instr_cb
)(struct nir_builder
*,
3967 nir_instr
*, void *);
3970 * Special return value for nir_lower_instr_cb when some progress occurred
3971 * (like changing an input to the instr) that didn't result in a replacement
3972 * SSA def being generated.
3974 #define NIR_LOWER_INSTR_PROGRESS ((nir_ssa_def *)(uintptr_t)1)
3976 /** Iterate over all the instructions in a nir_function_impl and lower them
3977 * using the provided callbacks
3979 * This function implements the guts of a standard lowering pass for you. It
3980 * iterates over all of the instructions in a nir_function_impl and calls the
3981 * filter callback on each one. If the filter callback returns true, it then
3982 * calls the lowering call back on the instruction. (Splitting it this way
3983 * allows us to avoid some save/restore work for instructions we know won't be
3984 * lowered.) If the instruction is dead after the lowering is complete, it
3985 * will be removed. If new instructions are added, the lowering callback will
3986 * also be called on them in case multiple lowerings are required.
3988 * The metadata for the nir_function_impl will also be updated. If any blocks
3989 * are added (they cannot be removed), dominance and block indices will be
3992 bool nir_function_impl_lower_instructions(nir_function_impl
*impl
,
3993 nir_instr_filter_cb filter
,
3994 nir_lower_instr_cb lower
,
3996 bool nir_shader_lower_instructions(nir_shader
*shader
,
3997 nir_instr_filter_cb filter
,
3998 nir_lower_instr_cb lower
,
4001 void nir_calc_dominance_impl(nir_function_impl
*impl
);
4002 void nir_calc_dominance(nir_shader
*shader
);
4004 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
4005 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
4006 bool nir_block_is_unreachable(nir_block
*block
);
4008 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
4009 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
4011 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
4012 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
4014 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
4015 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
4017 int nir_gs_count_vertices(const nir_shader
*shader
);
4019 bool nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
4020 bool nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
4021 bool nir_split_var_copies(nir_shader
*shader
);
4022 bool nir_split_per_member_structs(nir_shader
*shader
);
4023 bool nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
);
4025 bool nir_lower_returns_impl(nir_function_impl
*impl
);
4026 bool nir_lower_returns(nir_shader
*shader
);
4028 void nir_inline_function_impl(struct nir_builder
*b
,
4029 const nir_function_impl
*impl
,
4030 nir_ssa_def
**params
);
4031 bool nir_inline_functions(nir_shader
*shader
);
4033 bool nir_propagate_invariant(nir_shader
*shader
);
4035 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
4036 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
4037 nir_intrinsic_instr
*copy
);
4038 bool nir_lower_var_copies(nir_shader
*shader
);
4040 void nir_fixup_deref_modes(nir_shader
*shader
);
4042 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
4045 nir_lower_direct_array_deref_of_vec_load
= (1 << 0),
4046 nir_lower_indirect_array_deref_of_vec_load
= (1 << 1),
4047 nir_lower_direct_array_deref_of_vec_store
= (1 << 2),
4048 nir_lower_indirect_array_deref_of_vec_store
= (1 << 3),
4049 } nir_lower_array_deref_of_vec_options
;
4051 bool nir_lower_array_deref_of_vec(nir_shader
*shader
, nir_variable_mode modes
,
4052 nir_lower_array_deref_of_vec_options options
);
4054 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
4056 bool nir_lower_locals_to_regs(nir_shader
*shader
);
4058 void nir_lower_io_to_temporaries(nir_shader
*shader
,
4059 nir_function_impl
*entrypoint
,
4060 bool outputs
, bool inputs
);
4062 bool nir_lower_vars_to_scratch(nir_shader
*shader
,
4063 nir_variable_mode modes
,
4065 glsl_type_size_align_func size_align
);
4067 void nir_lower_clip_halfz(nir_shader
*shader
);
4069 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
4071 void nir_gather_ssa_types(nir_function_impl
*impl
,
4072 BITSET_WORD
*float_types
,
4073 BITSET_WORD
*int_types
);
4075 void nir_assign_var_locations(nir_shader
*shader
, nir_variable_mode mode
,
4077 int (*type_size
)(const struct glsl_type
*, bool));
4079 /* Some helpers to do very simple linking */
4080 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
4081 bool nir_remove_unused_io_vars(nir_shader
*shader
, nir_variable_mode mode
,
4082 uint64_t *used_by_other_stage
,
4083 uint64_t *used_by_other_stage_patches
);
4084 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
4085 bool default_to_smooth_interp
);
4086 void nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
);
4087 bool nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
);
4089 bool nir_lower_amul(nir_shader
*shader
,
4090 int (*type_size
)(const struct glsl_type
*, bool));
4092 bool nir_lower_ubo_vec4(nir_shader
*shader
);
4094 void nir_assign_io_var_locations(nir_shader
*shader
,
4095 nir_variable_mode mode
,
4097 gl_shader_stage stage
);
4100 uint8_t num_linked_io_vars
;
4101 uint8_t num_linked_patch_io_vars
;
4102 } nir_linked_io_var_info
;
4104 nir_linked_io_var_info
4105 nir_assign_linked_io_var_locations(nir_shader
*producer
,
4106 nir_shader
*consumer
);
4109 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
4110 * to 32-bit operations. This is only valid for nir_var_shader_in/out
4113 nir_lower_io_lower_64bit_to_32
= (1 << 0),
4115 /* If set, this forces all non-flat fragment shader inputs to be
4116 * interpolated as if with the "sample" qualifier. This requires
4117 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
4119 nir_lower_io_force_sample_interpolation
= (1 << 1),
4120 } nir_lower_io_options
;
4121 bool nir_lower_io(nir_shader
*shader
,
4122 nir_variable_mode modes
,
4123 int (*type_size
)(const struct glsl_type
*, bool),
4124 nir_lower_io_options
);
4126 bool nir_io_add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
);
4129 nir_lower_vars_to_explicit_types(nir_shader
*shader
,
4130 nir_variable_mode modes
,
4131 glsl_type_size_align_func type_info
);
4135 * An address format which is a simple 32-bit global GPU address.
4137 nir_address_format_32bit_global
,
4140 * An address format which is a simple 64-bit global GPU address.
4142 nir_address_format_64bit_global
,
4145 * An address format which is a bounds-checked 64-bit global GPU address.
4147 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
4148 * address stored with the low bits in .x and high bits in .y, .z is a
4149 * size, and .w is an offset. When the final I/O operation is lowered, .w
4150 * is checked against .z and the operation is predicated on the result.
4152 nir_address_format_64bit_bounded_global
,
4155 * An address format which is comprised of a vec2 where the first
4156 * component is a buffer index and the second is an offset.
4158 nir_address_format_32bit_index_offset
,
4161 * An address format which is a 64-bit value, where the high 32 bits
4162 * are a buffer index, and the low 32 bits are an offset.
4164 nir_address_format_32bit_index_offset_pack64
,
4167 * An address format which is comprised of a vec3 where the first two
4168 * components specify the buffer and the third is an offset.
4170 nir_address_format_vec2_index_32bit_offset
,
4173 * An address format which is a simple 32-bit offset.
4175 nir_address_format_32bit_offset
,
4178 * An address format which is a simple 32-bit offset cast to 64-bit.
4180 nir_address_format_32bit_offset_as_64bit
,
4183 * An address format representing a purely logical addressing model. In
4184 * this model, all deref chains must be complete from the dereference
4185 * operation to the variable. Cast derefs are not allowed. These
4186 * addresses will be 32-bit scalars but the format is immaterial because
4187 * you can always chase the chain.
4189 nir_address_format_logical
,
4190 } nir_address_format
;
4192 static inline unsigned
4193 nir_address_format_bit_size(nir_address_format addr_format
)
4195 switch (addr_format
) {
4196 case nir_address_format_32bit_global
: return 32;
4197 case nir_address_format_64bit_global
: return 64;
4198 case nir_address_format_64bit_bounded_global
: return 32;
4199 case nir_address_format_32bit_index_offset
: return 32;
4200 case nir_address_format_32bit_index_offset_pack64
: return 64;
4201 case nir_address_format_vec2_index_32bit_offset
: return 32;
4202 case nir_address_format_32bit_offset
: return 32;
4203 case nir_address_format_32bit_offset_as_64bit
: return 64;
4204 case nir_address_format_logical
: return 32;
4206 unreachable("Invalid address format");
4209 static inline unsigned
4210 nir_address_format_num_components(nir_address_format addr_format
)
4212 switch (addr_format
) {
4213 case nir_address_format_32bit_global
: return 1;
4214 case nir_address_format_64bit_global
: return 1;
4215 case nir_address_format_64bit_bounded_global
: return 4;
4216 case nir_address_format_32bit_index_offset
: return 2;
4217 case nir_address_format_32bit_index_offset_pack64
: return 1;
4218 case nir_address_format_vec2_index_32bit_offset
: return 3;
4219 case nir_address_format_32bit_offset
: return 1;
4220 case nir_address_format_32bit_offset_as_64bit
: return 1;
4221 case nir_address_format_logical
: return 1;
4223 unreachable("Invalid address format");
4226 static inline const struct glsl_type
*
4227 nir_address_format_to_glsl_type(nir_address_format addr_format
)
4229 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
4230 assert(bit_size
== 32 || bit_size
== 64);
4231 return glsl_vector_type(bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
,
4232 nir_address_format_num_components(addr_format
));
4235 const nir_const_value
*nir_address_format_null_value(nir_address_format addr_format
);
4237 nir_ssa_def
*nir_build_addr_ieq(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
4238 nir_address_format addr_format
);
4240 nir_ssa_def
*nir_build_addr_isub(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
4241 nir_address_format addr_format
);
4243 nir_ssa_def
* nir_explicit_io_address_from_deref(struct nir_builder
*b
,
4244 nir_deref_instr
*deref
,
4245 nir_ssa_def
*base_addr
,
4246 nir_address_format addr_format
);
4247 void nir_lower_explicit_io_instr(struct nir_builder
*b
,
4248 nir_intrinsic_instr
*io_instr
,
4250 nir_address_format addr_format
);
4252 bool nir_lower_explicit_io(nir_shader
*shader
,
4253 nir_variable_mode modes
,
4254 nir_address_format
);
4256 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
4257 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
4259 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
4261 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
4262 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
4263 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
4265 bool nir_remove_dead_derefs(nir_shader
*shader
);
4266 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
4267 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
,
4268 bool (*can_remove_var
)(nir_variable
*var
));
4269 bool nir_lower_variable_initializers(nir_shader
*shader
,
4270 nir_variable_mode modes
);
4272 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
4273 bool nir_lower_vec_to_movs(nir_shader
*shader
);
4274 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
4276 const gl_state_index16
*alpha_ref_state_tokens
);
4277 bool nir_lower_alu(nir_shader
*shader
);
4279 bool nir_lower_flrp(nir_shader
*shader
, unsigned lowering_mask
,
4280 bool always_precise
, bool have_ffma
);
4282 bool nir_lower_alu_to_scalar(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
4283 bool nir_lower_bool_to_bitsize(nir_shader
*shader
);
4284 bool nir_lower_bool_to_float(nir_shader
*shader
);
4285 bool nir_lower_bool_to_int32(nir_shader
*shader
);
4286 bool nir_lower_int_to_float(nir_shader
*shader
);
4287 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
4288 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
4289 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
4290 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
4291 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
4293 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
4294 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
4295 bool nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode mask
);
4297 bool nir_lower_fragcolor(nir_shader
*shader
);
4298 bool nir_lower_fragcoord_wtrans(nir_shader
*shader
);
4299 void nir_lower_viewport_transform(nir_shader
*shader
);
4300 bool nir_lower_uniforms_to_ubo(nir_shader
*shader
, int multiplier
);
4302 typedef struct nir_lower_subgroups_options
{
4303 uint8_t subgroup_size
;
4304 uint8_t ballot_bit_size
;
4305 bool lower_to_scalar
:1;
4306 bool lower_vote_trivial
:1;
4307 bool lower_vote_eq_to_ballot
:1;
4308 bool lower_subgroup_masks
:1;
4309 bool lower_shuffle
:1;
4310 bool lower_shuffle_to_32bit
:1;
4311 bool lower_shuffle_to_swizzle_amd
:1;
4313 bool lower_quad_broadcast_dynamic
:1;
4314 bool lower_quad_broadcast_dynamic_to_const
:1;
4315 } nir_lower_subgroups_options
;
4317 bool nir_lower_subgroups(nir_shader
*shader
,
4318 const nir_lower_subgroups_options
*options
);
4320 bool nir_lower_system_values(nir_shader
*shader
);
4322 typedef struct nir_lower_compute_system_values_options
{
4323 bool has_base_global_invocation_id
:1;
4324 bool has_base_work_group_id
:1;
4325 } nir_lower_compute_system_values_options
;
4327 bool nir_lower_compute_system_values(nir_shader
*shader
,
4328 const nir_lower_compute_system_values_options
*options
);
4330 enum PACKED nir_lower_tex_packing
{
4331 nir_lower_tex_packing_none
= 0,
4332 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
4333 * or unsigned ints based on the sampler type
4335 nir_lower_tex_packing_16
,
4336 /* The sampler returns 1 32-bit word of 4x8 unorm */
4337 nir_lower_tex_packing_8
,
4340 typedef struct nir_lower_tex_options
{
4342 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
4343 * sampler types a texture projector is lowered.
4348 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
4350 bool lower_txf_offset
;
4353 * If true, lower away nir_tex_src_offset for all rect textures.
4355 bool lower_rect_offset
;
4358 * If true, lower rect textures to 2D, using txs to fetch the
4359 * texture dimensions and dividing the texture coords by the
4360 * texture dims to normalize.
4365 * If true, convert yuv to rgb.
4367 unsigned lower_y_uv_external
;
4368 unsigned lower_y_u_v_external
;
4369 unsigned lower_yx_xuxv_external
;
4370 unsigned lower_xy_uxvx_external
;
4371 unsigned lower_ayuv_external
;
4372 unsigned lower_xyuv_external
;
4373 unsigned bt709_external
;
4374 unsigned bt2020_external
;
4377 * To emulate certain texture wrap modes, this can be used
4378 * to saturate the specified tex coord to [0.0, 1.0]. The
4379 * bits are according to sampler #, ie. if, for example:
4381 * (conf->saturate_s & (1 << n))
4383 * is true, then the s coord for sampler n is saturated.
4385 * Note that clamping must happen *after* projector lowering
4386 * so any projected texture sample instruction with a clamped
4387 * coordinate gets automatically lowered, regardless of the
4388 * 'lower_txp' setting.
4390 unsigned saturate_s
;
4391 unsigned saturate_t
;
4392 unsigned saturate_r
;
4394 /* Bitmask of textures that need swizzling.
4396 * If (swizzle_result & (1 << texture_index)), then the swizzle in
4397 * swizzles[texture_index] is applied to the result of the texturing
4400 unsigned swizzle_result
;
4402 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
4403 * while 4 and 5 represent 0 and 1 respectively.
4405 uint8_t swizzles
[32][4];
4407 /* Can be used to scale sampled values in range required by the format. */
4408 float scale_factors
[32];
4411 * Bitmap of textures that need srgb to linear conversion. If
4412 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
4413 * of the texture are lowered to linear.
4415 unsigned lower_srgb
;
4418 * If true, lower nir_texop_tex on shaders that doesn't support implicit
4419 * LODs to nir_texop_txl.
4421 bool lower_tex_without_implicit_lod
;
4424 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
4426 bool lower_txd_cube_map
;
4429 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
4434 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
4435 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
4436 * with lower_txd_cube_map.
4438 bool lower_txd_shadow
;
4441 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
4442 * Implies lower_txd_cube_map and lower_txd_shadow.
4447 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
4448 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
4450 bool lower_txb_shadow_clamp
;
4453 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
4454 * with nir_texop_txl. This includes cube maps.
4456 bool lower_txd_shadow_clamp
;
4459 * If true, lower nir_texop_txd on when it uses both offset and min_lod
4460 * with nir_texop_txl. This includes cube maps.
4462 bool lower_txd_offset_clamp
;
4465 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4466 * sampler is bindless.
4468 bool lower_txd_clamp_bindless_sampler
;
4471 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4472 * sampler index is not statically determinable to be less than 16.
4474 bool lower_txd_clamp_if_sampler_index_not_lt_16
;
4477 * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
4478 * 0-lod followed by a nir_ishr.
4483 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
4484 * mixed-up tg4 locations.
4486 bool lower_tg4_broadcom_swizzle
;
4489 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
4491 bool lower_tg4_offsets
;
4493 enum nir_lower_tex_packing lower_tex_packing
[32];
4494 } nir_lower_tex_options
;
4496 bool nir_lower_tex(nir_shader
*shader
,
4497 const nir_lower_tex_options
*options
);
4499 enum nir_lower_non_uniform_access_type
{
4500 nir_lower_non_uniform_ubo_access
= (1 << 0),
4501 nir_lower_non_uniform_ssbo_access
= (1 << 1),
4502 nir_lower_non_uniform_texture_access
= (1 << 2),
4503 nir_lower_non_uniform_image_access
= (1 << 3),
4506 bool nir_lower_non_uniform_access(nir_shader
*shader
,
4507 enum nir_lower_non_uniform_access_type
);
4509 enum nir_lower_idiv_path
{
4510 /* This path is based on NV50LegalizeSSA::handleDIV(). It is the faster of
4511 * the two but it is not exact in some cases (for example, 1091317713u /
4512 * 1034u gives 5209173 instead of 1055432) */
4513 nir_lower_idiv_fast
,
4514 /* This path is based on AMDGPUTargetLowering::LowerUDIVREM() and
4515 * AMDGPUTargetLowering::LowerSDIVREM(). It requires more instructions than
4516 * the nv50 path and many of them are integer multiplications, so it is
4517 * probably slower. It should always return the correct result, though. */
4518 nir_lower_idiv_precise
,
4521 bool nir_lower_idiv(nir_shader
*shader
, enum nir_lower_idiv_path path
);
4523 typedef struct nir_input_attachment_options
{
4524 bool use_fragcoord_sysval
;
4525 bool use_layer_id_sysval
;
4526 bool use_view_id_for_layer
;
4527 } nir_input_attachment_options
;
4529 bool nir_lower_input_attachments(nir_shader
*shader
,
4530 const nir_input_attachment_options
*options
);
4532 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
,
4534 bool use_clipdist_array
,
4535 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4536 bool nir_lower_clip_gs(nir_shader
*shader
, unsigned ucp_enables
,
4537 bool use_clipdist_array
,
4538 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4539 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
,
4540 bool use_clipdist_array
);
4541 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
4542 bool nir_lower_clip_disable(nir_shader
*shader
, unsigned clip_plane_enable
);
4544 void nir_lower_point_size_mov(nir_shader
*shader
,
4545 const gl_state_index16
*pointsize_state_tokens
);
4547 bool nir_lower_frexp(nir_shader
*nir
);
4549 void nir_lower_two_sided_color(nir_shader
*shader
, bool face_sysval
);
4551 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
4553 bool nir_lower_flatshade(nir_shader
*shader
);
4555 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
4556 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
4557 const gl_state_index16
*uniform_state_tokens
);
4559 typedef struct nir_lower_wpos_ytransform_options
{
4560 gl_state_index16 state_tokens
[STATE_LENGTH
];
4561 bool fs_coord_origin_upper_left
:1;
4562 bool fs_coord_origin_lower_left
:1;
4563 bool fs_coord_pixel_center_integer
:1;
4564 bool fs_coord_pixel_center_half_integer
:1;
4565 } nir_lower_wpos_ytransform_options
;
4567 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
4568 const nir_lower_wpos_ytransform_options
*options
);
4569 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
4571 bool nir_lower_wrmasks(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
4573 bool nir_lower_fb_read(nir_shader
*shader
);
4575 typedef struct nir_lower_drawpixels_options
{
4576 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
4577 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
4578 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
4579 unsigned drawpix_sampler
;
4580 unsigned pixelmap_sampler
;
4582 bool scale_and_bias
:1;
4583 } nir_lower_drawpixels_options
;
4585 void nir_lower_drawpixels(nir_shader
*shader
,
4586 const nir_lower_drawpixels_options
*options
);
4588 typedef struct nir_lower_bitmap_options
{
4591 } nir_lower_bitmap_options
;
4593 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
4595 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
);
4598 nir_lower_int_source_mods
= 1 << 0,
4599 nir_lower_float_source_mods
= 1 << 1,
4600 nir_lower_triop_abs
= 1 << 2,
4601 nir_lower_all_source_mods
= (1 << 3) - 1
4602 } nir_lower_to_source_mods_flags
;
4605 bool nir_lower_to_source_mods(nir_shader
*shader
, nir_lower_to_source_mods_flags options
);
4607 bool nir_lower_gs_intrinsics(nir_shader
*shader
, bool per_stream
);
4609 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
4611 bool nir_lower_bit_size(nir_shader
*shader
,
4612 nir_lower_bit_size_callback callback
,
4613 void *callback_data
);
4615 nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode
);
4616 bool nir_lower_int64(nir_shader
*shader
);
4618 nir_lower_doubles_options
nir_lower_doubles_op_to_options_mask(nir_op opcode
);
4619 bool nir_lower_doubles(nir_shader
*shader
, const nir_shader
*softfp64
,
4620 nir_lower_doubles_options options
);
4621 bool nir_lower_pack(nir_shader
*shader
);
4623 void nir_lower_mediump_outputs(nir_shader
*nir
);
4625 bool nir_lower_point_size(nir_shader
*shader
, float min
, float max
);
4628 nir_lower_interpolation_at_sample
= (1 << 1),
4629 nir_lower_interpolation_at_offset
= (1 << 2),
4630 nir_lower_interpolation_centroid
= (1 << 3),
4631 nir_lower_interpolation_pixel
= (1 << 4),
4632 nir_lower_interpolation_sample
= (1 << 5),
4633 } nir_lower_interpolation_options
;
4635 bool nir_lower_interpolation(nir_shader
*shader
,
4636 nir_lower_interpolation_options options
);
4638 bool nir_lower_discard_to_demote(nir_shader
*shader
);
4640 bool nir_lower_memory_model(nir_shader
*shader
);
4642 bool nir_lower_goto_ifs(nir_shader
*shader
);
4644 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
4646 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
4648 void nir_loop_analyze_impl(nir_function_impl
*impl
,
4649 nir_variable_mode indirect_mask
);
4651 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
4653 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
4654 bool nir_repair_ssa(nir_shader
*shader
);
4656 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
4657 bool nir_convert_to_lcssa(nir_shader
*shader
, bool skip_invariants
, bool skip_bool_invariants
);
4658 void nir_divergence_analysis(nir_shader
*shader
, nir_divergence_options options
);
4660 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
4661 * registers. If false, convert all values (even those not involved in a phi
4662 * node) to registers.
4664 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
4666 bool nir_lower_phis_to_regs_block(nir_block
*block
);
4667 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
4668 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
);
4670 bool nir_lower_samplers(nir_shader
*shader
);
4671 bool nir_lower_ssbo(nir_shader
*shader
);
4673 /* This is here for unit tests. */
4674 bool nir_opt_comparison_pre_impl(nir_function_impl
*impl
);
4676 bool nir_opt_comparison_pre(nir_shader
*shader
);
4678 bool nir_opt_access(nir_shader
*shader
);
4679 bool nir_opt_algebraic(nir_shader
*shader
);
4680 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
4681 bool nir_opt_algebraic_late(nir_shader
*shader
);
4682 bool nir_opt_algebraic_distribute_src_mods(nir_shader
*shader
);
4683 bool nir_opt_constant_folding(nir_shader
*shader
);
4685 /* Try to combine a and b into a. Return true if combination was possible,
4686 * which will result in b being removed by the pass. Return false if
4687 * combination wasn't possible.
4689 typedef bool (*nir_combine_memory_barrier_cb
)(
4690 nir_intrinsic_instr
*a
, nir_intrinsic_instr
*b
, void *data
);
4692 bool nir_opt_combine_memory_barriers(nir_shader
*shader
,
4693 nir_combine_memory_barrier_cb combine_cb
,
4696 bool nir_opt_combine_stores(nir_shader
*shader
, nir_variable_mode modes
);
4698 bool nir_copy_prop(nir_shader
*shader
);
4700 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
4702 bool nir_opt_cse(nir_shader
*shader
);
4704 bool nir_opt_dce(nir_shader
*shader
);
4706 bool nir_opt_dead_cf(nir_shader
*shader
);
4708 bool nir_opt_dead_write_vars(nir_shader
*shader
);
4710 bool nir_opt_deref_impl(nir_function_impl
*impl
);
4711 bool nir_opt_deref(nir_shader
*shader
);
4713 bool nir_opt_find_array_copies(nir_shader
*shader
);
4715 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
4717 bool nir_opt_idiv_const(nir_shader
*shader
, unsigned min_bit_size
);
4719 bool nir_opt_if(nir_shader
*shader
, bool aggressive_last_continue
);
4721 bool nir_opt_intrinsics(nir_shader
*shader
);
4723 bool nir_opt_large_constants(nir_shader
*shader
,
4724 glsl_type_size_align_func size_align
,
4725 unsigned threshold
);
4727 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
4730 nir_move_const_undef
= (1 << 0),
4731 nir_move_load_ubo
= (1 << 1),
4732 nir_move_load_input
= (1 << 2),
4733 nir_move_comparisons
= (1 << 3),
4734 nir_move_copies
= (1 << 4),
4737 bool nir_can_move_instr(nir_instr
*instr
, nir_move_options options
);
4739 bool nir_opt_sink(nir_shader
*shader
, nir_move_options options
);
4741 bool nir_opt_move(nir_shader
*shader
, nir_move_options options
);
4743 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
,
4744 bool indirect_load_ok
, bool expensive_alu_ok
);
4746 bool nir_opt_rematerialize_compares(nir_shader
*shader
);
4748 bool nir_opt_remove_phis(nir_shader
*shader
);
4749 bool nir_opt_remove_phis_block(nir_block
*block
);
4751 bool nir_opt_shrink_vectors(nir_shader
*shader
);
4753 bool nir_opt_trivial_continues(nir_shader
*shader
);
4755 bool nir_opt_undef(nir_shader
*shader
);
4757 bool nir_opt_vectorize(nir_shader
*shader
);
4759 bool nir_opt_conditional_discard(nir_shader
*shader
);
4761 typedef bool (*nir_should_vectorize_mem_func
)(unsigned align
, unsigned bit_size
,
4762 unsigned num_components
, unsigned high_offset
,
4763 nir_intrinsic_instr
*low
, nir_intrinsic_instr
*high
);
4765 bool nir_opt_load_store_vectorize(nir_shader
*shader
, nir_variable_mode modes
,
4766 nir_should_vectorize_mem_func callback
,
4767 nir_variable_mode robust_modes
);
4769 void nir_sweep(nir_shader
*shader
);
4771 void nir_remap_dual_slot_attributes(nir_shader
*shader
,
4772 uint64_t *dual_slot_inputs
);
4773 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
);
4775 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
4776 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);
4779 nir_variable_is_in_ubo(const nir_variable
*var
)
4781 return (var
->data
.mode
== nir_var_mem_ubo
&&
4782 var
->interface_type
!= NULL
);
4786 nir_variable_is_in_ssbo(const nir_variable
*var
)
4788 return (var
->data
.mode
== nir_var_mem_ssbo
&&
4789 var
->interface_type
!= NULL
);
4793 nir_variable_is_in_block(const nir_variable
*var
)
4795 return nir_variable_is_in_ubo(var
) || nir_variable_is_in_ssbo(var
);
4798 typedef struct nir_unsigned_upper_bound_config
{
4799 unsigned min_subgroup_size
;
4800 unsigned max_subgroup_size
;
4801 unsigned max_work_group_invocations
;
4802 unsigned max_work_group_count
[3];
4803 unsigned max_work_group_size
[3];
4805 uint32_t vertex_attrib_max
[32];
4806 } nir_unsigned_upper_bound_config
;
4809 nir_unsigned_upper_bound(nir_shader
*shader
, struct hash_table
*range_ht
,
4810 nir_ssa_scalar scalar
,
4811 const nir_unsigned_upper_bound_config
*config
);
4814 nir_addition_might_overflow(nir_shader
*shader
, struct hash_table
*range_ht
,
4815 nir_ssa_scalar ssa
, unsigned const_val
,
4816 const nir_unsigned_upper_bound_config
*config
);