2 * Copyright © 2014 Connor Abbott
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Connor Abbott (cwabbott0@gmail.com)
31 #include "util/hash_table.h"
32 #include "compiler/glsl/list.h"
33 #include "GL/gl.h" /* GLenum */
34 #include "util/list.h"
35 #include "util/ralloc.h"
37 #include "util/bitscan.h"
38 #include "util/bitset.h"
39 #include "util/macros.h"
40 #include "util/format/u_format.h"
41 #include "compiler/nir_types.h"
42 #include "compiler/shader_enums.h"
43 #include "compiler/shader_info.h"
44 #define XXH_INLINE_ALL
45 #include "util/xxhash.h"
49 #include "util/debug.h"
52 #include "nir_opcodes.h"
54 #if defined(_WIN32) && !defined(snprintf)
55 #define snprintf _snprintf
63 #define NIR_TRUE (~0u)
64 #define NIR_MAX_VEC_COMPONENTS 16
65 #define NIR_MAX_MATRIX_COLUMNS 4
66 #define NIR_STREAM_PACKED (1 << 8)
67 typedef uint16_t nir_component_mask_t
;
70 nir_num_components_valid(unsigned num_components
)
72 return (num_components
>= 1 &&
73 num_components
<= 4) ||
74 num_components
== 8 ||
78 /** Defines a cast function
80 * This macro defines a cast function from in_type to out_type where
81 * out_type is some structure type that contains a field of type out_type.
83 * Note that you have to be a bit careful as the generated cast function
86 #define NIR_DEFINE_CAST(name, in_type, out_type, field, \
87 type_field, type_value) \
88 static inline out_type * \
89 name(const in_type *parent) \
91 assert(parent && parent->type_field == type_value); \
92 return exec_node_data(out_type, parent, field); \
102 * Description of built-in state associated with a uniform
104 * \sa nir_variable::state_slots
107 gl_state_index16 tokens
[STATE_LENGTH
];
112 nir_var_shader_in
= (1 << 0),
113 nir_var_shader_out
= (1 << 1),
114 nir_var_shader_temp
= (1 << 2),
115 nir_var_function_temp
= (1 << 3),
116 nir_var_uniform
= (1 << 4),
117 nir_var_mem_ubo
= (1 << 5),
118 nir_var_system_value
= (1 << 6),
119 nir_var_mem_ssbo
= (1 << 7),
120 nir_var_mem_shared
= (1 << 8),
121 nir_var_mem_global
= (1 << 9),
122 nir_var_mem_push_const
= (1 << 10), /* not actually used for variables */
123 nir_num_variable_modes
= 11,
124 nir_var_all
= (1 << nir_num_variable_modes
) - 1,
131 nir_rounding_mode_undef
= 0,
132 nir_rounding_mode_rtne
= 1, /* round to nearest even */
133 nir_rounding_mode_ru
= 2, /* round up */
134 nir_rounding_mode_rd
= 3, /* round down */
135 nir_rounding_mode_rtz
= 4, /* round towards zero */
152 #define nir_const_value_to_array(arr, c, components, m) \
154 for (unsigned i = 0; i < components; ++i) \
158 static inline nir_const_value
159 nir_const_value_for_raw_uint(uint64_t x
, unsigned bit_size
)
162 memset(&v
, 0, sizeof(v
));
165 case 1: v
.b
= x
; break;
166 case 8: v
.u8
= x
; break;
167 case 16: v
.u16
= x
; break;
168 case 32: v
.u32
= x
; break;
169 case 64: v
.u64
= x
; break;
171 unreachable("Invalid bit size");
177 static inline nir_const_value
178 nir_const_value_for_int(int64_t i
, unsigned bit_size
)
181 memset(&v
, 0, sizeof(v
));
183 assert(bit_size
<= 64);
185 assert(i
>= (-(1ll << (bit_size
- 1))));
186 assert(i
< (1ll << (bit_size
- 1)));
189 return nir_const_value_for_raw_uint(i
, bit_size
);
192 static inline nir_const_value
193 nir_const_value_for_uint(uint64_t u
, unsigned bit_size
)
196 memset(&v
, 0, sizeof(v
));
198 assert(bit_size
<= 64);
200 assert(u
< (1ull << bit_size
));
202 return nir_const_value_for_raw_uint(u
, bit_size
);
205 static inline nir_const_value
206 nir_const_value_for_bool(bool b
, unsigned bit_size
)
208 /* Booleans use a 0/-1 convention */
209 return nir_const_value_for_int(-(int)b
, bit_size
);
212 /* This one isn't inline because it requires half-float conversion */
213 nir_const_value
nir_const_value_for_float(double b
, unsigned bit_size
);
215 static inline int64_t
216 nir_const_value_as_int(nir_const_value value
, unsigned bit_size
)
219 /* int1_t uses 0/-1 convention */
220 case 1: return -(int)value
.b
;
221 case 8: return value
.i8
;
222 case 16: return value
.i16
;
223 case 32: return value
.i32
;
224 case 64: return value
.i64
;
226 unreachable("Invalid bit size");
230 static inline uint64_t
231 nir_const_value_as_uint(nir_const_value value
, unsigned bit_size
)
234 case 1: return value
.b
;
235 case 8: return value
.u8
;
236 case 16: return value
.u16
;
237 case 32: return value
.u32
;
238 case 64: return value
.u64
;
240 unreachable("Invalid bit size");
245 nir_const_value_as_bool(nir_const_value value
, unsigned bit_size
)
247 int64_t i
= nir_const_value_as_int(value
, bit_size
);
249 /* Booleans of any size use 0/-1 convention */
250 assert(i
== 0 || i
== -1);
255 /* This one isn't inline because it requires half-float conversion */
256 double nir_const_value_as_float(nir_const_value value
, unsigned bit_size
);
258 typedef struct nir_constant
{
260 * Value of the constant.
262 * The field used to back the values supplied by the constant is determined
263 * by the type associated with the \c nir_variable. Constants may be
264 * scalars, vectors, or matrices.
266 nir_const_value values
[NIR_MAX_VEC_COMPONENTS
];
268 /* we could get this from the var->type but makes clone *much* easier to
269 * not have to care about the type.
271 unsigned num_elements
;
273 /* Array elements / Structure Fields */
274 struct nir_constant
**elements
;
278 * \brief Layout qualifiers for gl_FragDepth.
280 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
281 * with a layout qualifier.
284 nir_depth_layout_none
, /**< No depth layout is specified. */
285 nir_depth_layout_any
,
286 nir_depth_layout_greater
,
287 nir_depth_layout_less
,
288 nir_depth_layout_unchanged
292 * Enum keeping track of how a variable was declared.
296 * Normal declaration.
298 nir_var_declared_normally
= 0,
301 * Variable is implicitly generated by the compiler and should not be
302 * visible via the API.
305 } nir_var_declaration_type
;
308 * Either a uniform, global variable, shader input, or shader output. Based on
309 * ir_variable - it should be easy to translate between the two.
312 typedef struct nir_variable
{
313 struct exec_node node
;
316 * Declared type of the variable
318 const struct glsl_type
*type
;
321 * Declared name of the variable
325 struct nir_variable_data
{
327 * Storage class of the variable.
329 * \sa nir_variable_mode
331 nir_variable_mode mode
:11;
334 * Is the variable read-only?
336 * This is set for variables declared as \c const, shader inputs,
339 unsigned read_only
:1;
343 unsigned invariant
:1;
346 * Precision qualifier.
348 * In desktop GLSL we do not care about precision qualifiers at all, in
349 * fact, the spec says that precision qualifiers are ignored.
351 * To make things easy, we make it so that this field is always
352 * GLSL_PRECISION_NONE on desktop shaders. This way all the variables
353 * have the same precision value and the checks we add in the compiler
354 * for this field will never break a desktop shader compile.
356 unsigned precision
:2;
359 * Can this variable be coalesced with another?
361 * This is set by nir_lower_io_to_temporaries to say that any
362 * copies involving this variable should stay put. Propagating it can
363 * duplicate the resulting load/store, which is not wanted, and may
364 * result in a load/store of the variable with an indirect offset which
365 * the backend may not be able to handle.
367 unsigned cannot_coalesce
:1;
370 * When separate shader programs are enabled, only input/outputs between
371 * the stages of a multi-stage separate program can be safely removed
372 * from the shader interface. Other input/outputs must remains active.
374 * This is also used to make sure xfb varyings that are unused by the
375 * fragment shader are not removed.
377 unsigned always_active_io
:1;
380 * Interpolation mode for shader inputs / outputs
382 * \sa glsl_interp_mode
384 unsigned interpolation
:3;
387 * If non-zero, then this variable may be packed along with other variables
388 * into a single varying slot, so this offset should be applied when
389 * accessing components. For example, an offset of 1 means that the x
390 * component of this variable is actually stored in component y of the
391 * location specified by \c location.
393 unsigned location_frac
:2;
396 * If true, this variable represents an array of scalars that should
397 * be tightly packed. In other words, consecutive array elements
398 * should be stored one component apart, rather than one slot apart.
403 * Whether this is a fragment shader output implicitly initialized with
404 * the previous contents of the specified render target at the
405 * framebuffer location corresponding to this shader invocation.
407 unsigned fb_fetch_output
:1;
410 * Non-zero if this variable is considered bindless as defined by
411 * ARB_bindless_texture.
416 * Was an explicit binding set in the shader?
418 unsigned explicit_binding
:1;
421 * Was the location explicitly set in the shader?
423 * If the location is explicitly set in the shader, it \b cannot be changed
424 * by the linker or by the API (e.g., calls to \c glBindAttribLocation have
427 unsigned explicit_location
:1;
430 * Was a transfer feedback buffer set in the shader?
432 unsigned explicit_xfb_buffer
:1;
435 * Was a transfer feedback stride set in the shader?
437 unsigned explicit_xfb_stride
:1;
440 * Was an explicit offset set in the shader?
442 unsigned explicit_offset
:1;
445 * Layout of the matrix. Uses glsl_matrix_layout values.
447 unsigned matrix_layout
:2;
450 * Non-zero if this variable was created by lowering a named interface
453 unsigned from_named_ifc_block
:1;
456 * How the variable was declared. See nir_var_declaration_type.
458 * This is used to detect variables generated by the compiler, so should
459 * not be visible via the API.
461 unsigned how_declared
:2;
464 * Is this variable per-view? If so, we know it must be an array with
465 * size corresponding to the number of views.
470 * \brief Layout qualifier for gl_FragDepth.
472 * This is not equal to \c ir_depth_layout_none if and only if this
473 * variable is \c gl_FragDepth and a layout qualifier is specified.
475 nir_depth_layout depth_layout
:3;
478 * Vertex stream output identifier.
480 * For packed outputs, NIR_STREAM_PACKED is set and bits [2*i+1,2*i]
481 * indicate the stream of the i-th component.
486 * Access flags for memory variables (SSBO/global), image uniforms, and
487 * bindless images in uniforms/inputs/outputs.
489 enum gl_access_qualifier access
:8;
492 * Descriptor set binding for sampler or UBO.
494 unsigned descriptor_set
:5;
497 * output index for dual source blending.
502 * Initial binding point for a sampler or UBO.
504 * For array types, this represents the binding point for the first element.
509 * Storage location of the base of this variable
511 * The precise meaning of this field depends on the nature of the variable.
513 * - Vertex shader input: one of the values from \c gl_vert_attrib.
514 * - Vertex shader output: one of the values from \c gl_varying_slot.
515 * - Geometry shader input: one of the values from \c gl_varying_slot.
516 * - Geometry shader output: one of the values from \c gl_varying_slot.
517 * - Fragment shader input: one of the values from \c gl_varying_slot.
518 * - Fragment shader output: one of the values from \c gl_frag_result.
519 * - Uniforms: Per-stage uniform slot number for default uniform block.
520 * - Uniforms: Index within the uniform block definition for UBO members.
521 * - Non-UBO Uniforms: uniform slot number.
522 * - Other: This field is not currently used.
524 * If the variable is a uniform, shader input, or shader output, and the
525 * slot has not been assigned, the value will be -1.
530 * The actual location of the variable in the IR. Only valid for inputs,
531 * outputs, and uniforms (including samplers and images).
533 unsigned driver_location
;
536 * Location an atomic counter or transform feedback is stored at.
542 /** Image internal format if specified explicitly, otherwise PIPE_FORMAT_NONE. */
543 enum pipe_format format
;
548 * Transform feedback buffer.
553 * Transform feedback stride.
561 * Identifier for this variable generated by nir_index_vars() that is unique
562 * among other variables in the same exec_list.
566 /* Number of nir_variable_data members */
567 uint16_t num_members
;
570 * Built-in state that backs this uniform
572 * Once set at variable creation, \c state_slots must remain invariant.
573 * This is because, ideally, this array would be shared by all clones of
574 * this variable in the IR tree. In other words, we'd really like for it
575 * to be a fly-weight.
577 * If the variable is not a uniform, \c num_state_slots will be zero and
578 * \c state_slots will be \c NULL.
581 uint16_t num_state_slots
; /**< Number of state slots used */
582 nir_state_slot
*state_slots
; /**< State descriptors. */
586 * Constant expression assigned in the initializer of the variable
588 * This field should only be used temporarily by creators of NIR shaders
589 * and then lower_constant_initializers can be used to get rid of them.
590 * Most of the rest of NIR ignores this field or asserts that it's NULL.
592 nir_constant
*constant_initializer
;
595 * Global variable assigned in the initializer of the variable
596 * This field should only be used temporarily by creators of NIR shaders
597 * and then lower_constant_initializers can be used to get rid of them.
598 * Most of the rest of NIR ignores this field or asserts that it's NULL.
600 struct nir_variable
*pointer_initializer
;
603 * For variables that are in an interface block or are an instance of an
604 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
606 * \sa ir_variable::location
608 const struct glsl_type
*interface_type
;
611 * Description of per-member data for per-member struct variables
613 * This is used for variables which are actually an amalgamation of
614 * multiple entities such as a struct of built-in values or a struct of
615 * inputs each with their own layout specifier. This is only allowed on
616 * variables with a struct or array of array of struct type.
618 struct nir_variable_data
*members
;
621 #define nir_foreach_variable(var, var_list) \
622 foreach_list_typed(nir_variable, var, node, var_list)
624 #define nir_foreach_variable_safe(var, var_list) \
625 foreach_list_typed_safe(nir_variable, var, node, var_list)
627 #define nir_foreach_shader_in_variable(var, shader) \
628 nir_foreach_variable(var, &(shader)->inputs)
630 #define nir_foreach_shader_in_variable_safe(var, shader) \
631 nir_foreach_variable_safe(var, &(shader)->inputs)
633 #define nir_foreach_shader_out_variable(var, shader) \
634 nir_foreach_variable(var, &(shader)->outputs)
636 #define nir_foreach_shader_out_variable_safe(var, shader) \
637 nir_foreach_variable_safe(var, &(shader)->outputs)
640 nir_variable_is_global(const nir_variable
*var
)
642 return var
->data
.mode
!= nir_var_function_temp
;
645 typedef struct nir_register
{
646 struct exec_node node
;
648 unsigned num_components
; /** < number of vector components */
649 unsigned num_array_elems
; /** < size of array (0 for no array) */
651 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
654 /** generic register index. */
657 /** only for debug purposes, can be NULL */
660 /** set of nir_srcs where this register is used (read from) */
661 struct list_head uses
;
663 /** set of nir_dests where this register is defined (written to) */
664 struct list_head defs
;
666 /** set of nir_ifs where this register is used as a condition */
667 struct list_head if_uses
;
670 #define nir_foreach_register(reg, reg_list) \
671 foreach_list_typed(nir_register, reg, node, reg_list)
672 #define nir_foreach_register_safe(reg, reg_list) \
673 foreach_list_typed_safe(nir_register, reg, node, reg_list)
675 typedef enum PACKED
{
677 nir_instr_type_deref
,
680 nir_instr_type_intrinsic
,
681 nir_instr_type_load_const
,
683 nir_instr_type_ssa_undef
,
685 nir_instr_type_parallel_copy
,
688 typedef struct nir_instr
{
689 struct exec_node node
;
690 struct nir_block
*block
;
693 /* A temporary for optimization and analysis passes to use for storing
694 * flags. For instance, DCE uses this to store the "dead/live" info.
698 /** generic instruction index. */
702 static inline nir_instr
*
703 nir_instr_next(nir_instr
*instr
)
705 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
706 if (exec_node_is_tail_sentinel(next
))
709 return exec_node_data(nir_instr
, next
, node
);
712 static inline nir_instr
*
713 nir_instr_prev(nir_instr
*instr
)
715 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
716 if (exec_node_is_head_sentinel(prev
))
719 return exec_node_data(nir_instr
, prev
, node
);
723 nir_instr_is_first(const nir_instr
*instr
)
725 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr
->node
));
729 nir_instr_is_last(const nir_instr
*instr
)
731 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr
->node
));
734 typedef struct nir_ssa_def
{
735 /** for debugging only, can be NULL */
738 /** generic SSA definition index. */
741 /** Index into the live_in and live_out bitfields */
744 /** Instruction which produces this SSA value. */
745 nir_instr
*parent_instr
;
747 /** set of nir_instrs where this register is used (read from) */
748 struct list_head uses
;
750 /** set of nir_ifs where this register is used as a condition */
751 struct list_head if_uses
;
753 uint8_t num_components
;
755 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
759 * True if this SSA value may have different values in different SIMD
760 * invocations of the shader. This is set by nir_divergence_analysis.
769 struct nir_src
*indirect
; /** < NULL for no indirect offset */
770 unsigned base_offset
;
772 /* TODO use-def chain goes here */
776 nir_instr
*parent_instr
;
777 struct list_head def_link
;
780 struct nir_src
*indirect
; /** < NULL for no indirect offset */
781 unsigned base_offset
;
783 /* TODO def-use chain goes here */
788 typedef struct nir_src
{
790 /** Instruction that consumes this value as a source. */
791 nir_instr
*parent_instr
;
792 struct nir_if
*parent_if
;
795 struct list_head use_link
;
805 static inline nir_src
808 nir_src src
= { { NULL
} };
812 #define NIR_SRC_INIT nir_src_init()
814 #define nir_foreach_use(src, reg_or_ssa_def) \
815 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
817 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
818 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
820 #define nir_foreach_if_use(src, reg_or_ssa_def) \
821 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
823 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
824 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
835 static inline nir_dest
838 nir_dest dest
= { { { NULL
} } };
842 #define NIR_DEST_INIT nir_dest_init()
844 #define nir_foreach_def(dest, reg) \
845 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
847 #define nir_foreach_def_safe(dest, reg) \
848 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
850 static inline nir_src
851 nir_src_for_ssa(nir_ssa_def
*def
)
853 nir_src src
= NIR_SRC_INIT
;
861 static inline nir_src
862 nir_src_for_reg(nir_register
*reg
)
864 nir_src src
= NIR_SRC_INIT
;
868 src
.reg
.indirect
= NULL
;
869 src
.reg
.base_offset
= 0;
874 static inline nir_dest
875 nir_dest_for_reg(nir_register
*reg
)
877 nir_dest dest
= NIR_DEST_INIT
;
884 static inline unsigned
885 nir_src_bit_size(nir_src src
)
887 return src
.is_ssa
? src
.ssa
->bit_size
: src
.reg
.reg
->bit_size
;
890 static inline unsigned
891 nir_src_num_components(nir_src src
)
893 return src
.is_ssa
? src
.ssa
->num_components
: src
.reg
.reg
->num_components
;
897 nir_src_is_const(nir_src src
)
900 src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
;
904 nir_src_is_divergent(nir_src src
)
907 return src
.ssa
->divergent
;
910 static inline unsigned
911 nir_dest_bit_size(nir_dest dest
)
913 return dest
.is_ssa
? dest
.ssa
.bit_size
: dest
.reg
.reg
->bit_size
;
916 static inline unsigned
917 nir_dest_num_components(nir_dest dest
)
919 return dest
.is_ssa
? dest
.ssa
.num_components
: dest
.reg
.reg
->num_components
;
923 nir_dest_is_divergent(nir_dest dest
)
926 return dest
.ssa
.divergent
;
929 /* Are all components the same, ie. .xxxx */
931 nir_is_same_comp_swizzle(uint8_t *swiz
, unsigned nr_comp
)
933 for (unsigned i
= 1; i
< nr_comp
; i
++)
934 if (swiz
[i
] != swiz
[0])
939 /* Are all components sequential, ie. .yzw */
941 nir_is_sequential_comp_swizzle(uint8_t *swiz
, unsigned nr_comp
)
943 for (unsigned i
= 1; i
< nr_comp
; i
++)
944 if (swiz
[i
] != (swiz
[0] + i
))
949 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
950 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
956 * \name input modifiers
960 * For inputs interpreted as floating point, flips the sign bit. For
961 * inputs interpreted as integers, performs the two's complement negation.
966 * Clears the sign bit for floating point values, and computes the integer
967 * absolute value for integers. Note that the negate modifier acts after
968 * the absolute value modifier, therefore if both are set then all inputs
969 * will become negative.
975 * For each input component, says which component of the register it is
976 * chosen from. Note that which elements of the swizzle are used and which
977 * are ignored are based on the write mask for most opcodes - for example,
978 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
979 * a swizzle of {2, x, 1, 0} where x means "don't care."
981 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
988 * \name saturate output modifier
990 * Only valid for opcodes that output floating-point numbers. Clamps the
991 * output to between 0.0 and 1.0 inclusive.
996 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
999 /** NIR sized and unsized types
1001 * The values in this enum are carefully chosen so that the sized type is
1002 * just the unsized type OR the number of bits.
1004 typedef enum PACKED
{
1005 nir_type_invalid
= 0, /* Not a valid type */
1009 nir_type_float
= 128,
1010 nir_type_bool1
= 1 | nir_type_bool
,
1011 nir_type_bool8
= 8 | nir_type_bool
,
1012 nir_type_bool16
= 16 | nir_type_bool
,
1013 nir_type_bool32
= 32 | nir_type_bool
,
1014 nir_type_int1
= 1 | nir_type_int
,
1015 nir_type_int8
= 8 | nir_type_int
,
1016 nir_type_int16
= 16 | nir_type_int
,
1017 nir_type_int32
= 32 | nir_type_int
,
1018 nir_type_int64
= 64 | nir_type_int
,
1019 nir_type_uint1
= 1 | nir_type_uint
,
1020 nir_type_uint8
= 8 | nir_type_uint
,
1021 nir_type_uint16
= 16 | nir_type_uint
,
1022 nir_type_uint32
= 32 | nir_type_uint
,
1023 nir_type_uint64
= 64 | nir_type_uint
,
1024 nir_type_float16
= 16 | nir_type_float
,
1025 nir_type_float32
= 32 | nir_type_float
,
1026 nir_type_float64
= 64 | nir_type_float
,
1029 #define NIR_ALU_TYPE_SIZE_MASK 0x79
1030 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
1032 static inline unsigned
1033 nir_alu_type_get_type_size(nir_alu_type type
)
1035 return type
& NIR_ALU_TYPE_SIZE_MASK
;
1038 static inline nir_alu_type
1039 nir_alu_type_get_base_type(nir_alu_type type
)
1041 return (nir_alu_type
)(type
& NIR_ALU_TYPE_BASE_TYPE_MASK
);
1044 static inline nir_alu_type
1045 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
1047 switch (base_type
) {
1048 case GLSL_TYPE_BOOL
:
1049 return nir_type_bool1
;
1051 case GLSL_TYPE_UINT
:
1052 return nir_type_uint32
;
1055 return nir_type_int32
;
1057 case GLSL_TYPE_UINT16
:
1058 return nir_type_uint16
;
1060 case GLSL_TYPE_INT16
:
1061 return nir_type_int16
;
1063 case GLSL_TYPE_UINT8
:
1064 return nir_type_uint8
;
1065 case GLSL_TYPE_INT8
:
1066 return nir_type_int8
;
1067 case GLSL_TYPE_UINT64
:
1068 return nir_type_uint64
;
1070 case GLSL_TYPE_INT64
:
1071 return nir_type_int64
;
1073 case GLSL_TYPE_FLOAT
:
1074 return nir_type_float32
;
1076 case GLSL_TYPE_FLOAT16
:
1077 return nir_type_float16
;
1079 case GLSL_TYPE_DOUBLE
:
1080 return nir_type_float64
;
1083 case GLSL_TYPE_SAMPLER
:
1084 case GLSL_TYPE_IMAGE
:
1085 case GLSL_TYPE_ATOMIC_UINT
:
1086 case GLSL_TYPE_STRUCT
:
1087 case GLSL_TYPE_INTERFACE
:
1088 case GLSL_TYPE_ARRAY
:
1089 case GLSL_TYPE_VOID
:
1090 case GLSL_TYPE_SUBROUTINE
:
1091 case GLSL_TYPE_FUNCTION
:
1092 case GLSL_TYPE_ERROR
:
1093 return nir_type_invalid
;
1096 unreachable("unknown type");
1099 static inline nir_alu_type
1100 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
1102 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
1105 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
1106 nir_rounding_mode rnd
);
1108 static inline nir_op
1109 nir_op_vec(unsigned components
)
1111 switch (components
) {
1112 case 1: return nir_op_mov
;
1113 case 2: return nir_op_vec2
;
1114 case 3: return nir_op_vec3
;
1115 case 4: return nir_op_vec4
;
1116 case 8: return nir_op_vec8
;
1117 case 16: return nir_op_vec16
;
1118 default: unreachable("bad component count");
1123 nir_op_is_vec(nir_op op
)
1139 nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode
, unsigned bit_size
)
1141 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
) ||
1142 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
) ||
1143 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
);
1147 nir_is_denorm_flush_to_zero(unsigned execution_mode
, unsigned bit_size
)
1149 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
) ||
1150 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
) ||
1151 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
);
1155 nir_is_denorm_preserve(unsigned execution_mode
, unsigned bit_size
)
1157 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP16
) ||
1158 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP32
) ||
1159 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP64
);
1163 nir_is_rounding_mode_rtne(unsigned execution_mode
, unsigned bit_size
)
1165 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1166 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1167 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1171 nir_is_rounding_mode_rtz(unsigned execution_mode
, unsigned bit_size
)
1173 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1174 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1175 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1179 nir_has_any_rounding_mode_rtz(unsigned execution_mode
)
1181 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1182 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1183 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1187 nir_has_any_rounding_mode_rtne(unsigned execution_mode
)
1189 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1190 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1191 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1194 static inline nir_rounding_mode
1195 nir_get_rounding_mode_from_float_controls(unsigned execution_mode
,
1198 if (nir_alu_type_get_base_type(type
) != nir_type_float
)
1199 return nir_rounding_mode_undef
;
1201 unsigned bit_size
= nir_alu_type_get_type_size(type
);
1203 if (nir_is_rounding_mode_rtz(execution_mode
, bit_size
))
1204 return nir_rounding_mode_rtz
;
1205 if (nir_is_rounding_mode_rtne(execution_mode
, bit_size
))
1206 return nir_rounding_mode_rtne
;
1207 return nir_rounding_mode_undef
;
1211 nir_has_any_rounding_mode_enabled(unsigned execution_mode
)
1214 nir_has_any_rounding_mode_rtne(execution_mode
) ||
1215 nir_has_any_rounding_mode_rtz(execution_mode
);
1221 * Operation where the first two sources are commutative.
1223 * For 2-source operations, this just mathematical commutativity. Some
1224 * 3-source operations, like ffma, are only commutative in the first two
1227 NIR_OP_IS_2SRC_COMMUTATIVE
= (1 << 0),
1228 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
1229 } nir_op_algebraic_property
;
1237 * The number of components in the output
1239 * If non-zero, this is the size of the output and input sizes are
1240 * explicitly given; swizzle and writemask are still in effect, but if
1241 * the output component is masked out, then the input component may
1244 * If zero, the opcode acts in the standard, per-component manner; the
1245 * operation is performed on each component (except the ones that are
1246 * masked out) with the input being taken from the input swizzle for
1249 * The size of some of the inputs may be given (i.e. non-zero) even
1250 * though output_size is zero; in that case, the inputs with a zero
1251 * size act per-component, while the inputs with non-zero size don't.
1253 uint8_t output_size
;
1256 * The type of vector that the instruction outputs. Note that the
1257 * staurate modifier is only allowed on outputs with the float type.
1260 nir_alu_type output_type
;
1263 * The number of components in each input
1265 uint8_t input_sizes
[NIR_MAX_VEC_COMPONENTS
];
1268 * The type of vector that each input takes. Note that negate and
1269 * absolute value are only allowed on inputs with int or float type and
1270 * behave differently on the two.
1272 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
1274 nir_op_algebraic_property algebraic_properties
;
1276 /* Whether this represents a numeric conversion opcode */
1280 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
1282 typedef struct nir_alu_instr
{
1286 /** Indicates that this ALU instruction generates an exact value
1288 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1289 * really equivalent to either. This indicates that the value generated by
1290 * this operation is high-precision and any code transformations that touch
1291 * it must ensure that the resulting value is bit-for-bit identical to the
1297 * Indicates that this instruction do not cause wrapping to occur, in the
1298 * form of overflow or underflow.
1300 bool no_signed_wrap
:1;
1301 bool no_unsigned_wrap
:1;
1307 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
1308 nir_alu_instr
*instr
);
1309 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
1310 nir_alu_instr
*instr
);
1312 /* is this source channel used? */
1314 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
1317 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1318 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
1320 return (instr
->dest
.write_mask
>> channel
) & 1;
1323 static inline nir_component_mask_t
1324 nir_alu_instr_src_read_mask(const nir_alu_instr
*instr
, unsigned src
)
1326 nir_component_mask_t read_mask
= 0;
1327 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; c
++) {
1328 if (!nir_alu_instr_channel_used(instr
, src
, c
))
1331 read_mask
|= (1 << instr
->src
[src
].swizzle
[c
]);
1337 * Get the number of channels used for a source
1339 static inline unsigned
1340 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
1342 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1343 return nir_op_infos
[instr
->op
].input_sizes
[src
];
1345 return nir_dest_num_components(instr
->dest
.dest
);
1349 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1351 switch (instr
->op
) {
1371 bool nir_const_value_negative_equal(nir_const_value c1
, nir_const_value c2
,
1372 nir_alu_type full_type
);
1374 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
1375 unsigned src1
, unsigned src2
);
1377 bool nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
1378 const nir_alu_instr
*alu2
,
1379 unsigned src1
, unsigned src2
);
1383 nir_deref_type_array
,
1384 nir_deref_type_array_wildcard
,
1385 nir_deref_type_ptr_as_array
,
1386 nir_deref_type_struct
,
1387 nir_deref_type_cast
,
1393 /** The type of this deref instruction */
1394 nir_deref_type deref_type
;
1396 /** The mode of the underlying variable */
1397 nir_variable_mode mode
;
1399 /** The dereferenced type of the resulting pointer value */
1400 const struct glsl_type
*type
;
1403 /** Variable being dereferenced if deref_type is a deref_var */
1406 /** Parent deref if deref_type is not deref_var */
1410 /** Additional deref parameters */
1421 unsigned ptr_stride
;
1425 /** Destination to store the resulting "pointer" */
1429 static inline nir_deref_instr
*nir_src_as_deref(nir_src src
);
1431 static inline nir_deref_instr
*
1432 nir_deref_instr_parent(const nir_deref_instr
*instr
)
1434 if (instr
->deref_type
== nir_deref_type_var
)
1437 return nir_src_as_deref(instr
->parent
);
1440 static inline nir_variable
*
1441 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1443 while (instr
->deref_type
!= nir_deref_type_var
) {
1444 if (instr
->deref_type
== nir_deref_type_cast
)
1447 instr
= nir_deref_instr_parent(instr
);
1453 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1454 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr
*instr
);
1455 bool nir_deref_instr_has_complex_use(nir_deref_instr
*instr
);
1457 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1459 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*instr
);
1464 struct nir_function
*callee
;
1466 unsigned num_params
;
1470 #include "nir_intrinsics.h"
1472 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1474 /** Represents an intrinsic
1476 * An intrinsic is an instruction type for handling things that are
1477 * more-or-less regular operations but don't just consume and produce SSA
1478 * values like ALU operations do. Intrinsics are not for things that have
1479 * special semantic meaning such as phi nodes and parallel copies.
1480 * Examples of intrinsics include variable load/store operations, system
1481 * value loads, and the like. Even though texturing more-or-less falls
1482 * under this category, texturing is its own instruction type because
1483 * trying to represent texturing with intrinsics would lead to a
1484 * combinatorial explosion of intrinsic opcodes.
1486 * By having a single instruction type for handling a lot of different
1487 * cases, optimization passes can look for intrinsics and, for the most
1488 * part, completely ignore them. Each intrinsic type also has a few
1489 * possible flags that govern whether or not they can be reordered or
1490 * eliminated. That way passes like dead code elimination can still work
1491 * on intrisics without understanding the meaning of each.
1493 * Each intrinsic has some number of constant indices, some number of
1494 * variables, and some number of sources. What these sources, variables,
1495 * and indices mean depends on the intrinsic and is documented with the
1496 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1497 * instructions are the only types of instruction that can operate on
1503 nir_intrinsic_op intrinsic
;
1507 /** number of components if this is a vectorized intrinsic
1509 * Similarly to ALU operations, some intrinsics are vectorized.
1510 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1511 * For vectorized intrinsics, the num_components field specifies the
1512 * number of destination components and the number of source components
1513 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1515 uint8_t num_components
;
1517 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1520 } nir_intrinsic_instr
;
1522 static inline nir_variable
*
1523 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1525 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1529 /* Memory ordering. */
1530 NIR_MEMORY_ACQUIRE
= 1 << 0,
1531 NIR_MEMORY_RELEASE
= 1 << 1,
1532 NIR_MEMORY_ACQ_REL
= NIR_MEMORY_ACQUIRE
| NIR_MEMORY_RELEASE
,
1534 /* Memory visibility operations. */
1535 NIR_MEMORY_MAKE_AVAILABLE
= 1 << 2,
1536 NIR_MEMORY_MAKE_VISIBLE
= 1 << 3,
1537 } nir_memory_semantics
;
1541 NIR_SCOPE_INVOCATION
,
1543 NIR_SCOPE_WORKGROUP
,
1544 NIR_SCOPE_QUEUE_FAMILY
,
1549 * \name NIR intrinsics semantic flags
1551 * information about what the compiler can do with the intrinsics.
1553 * \sa nir_intrinsic_info::flags
1557 * whether the intrinsic can be safely eliminated if none of its output
1558 * value is not being used.
1560 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1563 * Whether the intrinsic can be reordered with respect to any other
1564 * intrinsic, i.e. whether the only reordering dependencies of the
1565 * intrinsic are due to the register reads/writes.
1567 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1568 } nir_intrinsic_semantic_flag
;
1571 * \name NIR intrinsics const-index flag
1573 * Indicates the usage of a const_index slot.
1575 * \sa nir_intrinsic_info::index_map
1579 * Generally instructions that take a offset src argument, can encode
1580 * a constant 'base' value which is added to the offset.
1582 NIR_INTRINSIC_BASE
= 1,
1585 * For store instructions, a writemask for the store.
1587 NIR_INTRINSIC_WRMASK
,
1590 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1592 NIR_INTRINSIC_STREAM_ID
,
1595 * The clip-plane id for load_user_clip_plane intrinsic.
1597 NIR_INTRINSIC_UCP_ID
,
1600 * The amount of data, starting from BASE, that this instruction may
1601 * access. This is used to provide bounds if the offset is not constant.
1603 NIR_INTRINSIC_RANGE
,
1606 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1608 NIR_INTRINSIC_DESC_SET
,
1611 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1613 NIR_INTRINSIC_BINDING
,
1618 NIR_INTRINSIC_COMPONENT
,
1621 * Interpolation mode (only meaningful for FS inputs).
1623 NIR_INTRINSIC_INTERP_MODE
,
1626 * A binary nir_op to use when performing a reduction or scan operation
1628 NIR_INTRINSIC_REDUCTION_OP
,
1631 * Cluster size for reduction operations
1633 NIR_INTRINSIC_CLUSTER_SIZE
,
1636 * Parameter index for a load_param intrinsic
1638 NIR_INTRINSIC_PARAM_IDX
,
1641 * Image dimensionality for image intrinsics
1643 * One of GLSL_SAMPLER_DIM_*
1645 NIR_INTRINSIC_IMAGE_DIM
,
1648 * Non-zero if we are accessing an array image
1650 NIR_INTRINSIC_IMAGE_ARRAY
,
1653 * Image format for image intrinsics
1655 NIR_INTRINSIC_FORMAT
,
1658 * Access qualifiers for image and memory access intrinsics
1660 NIR_INTRINSIC_ACCESS
,
1663 * Alignment for offsets and addresses
1665 * These two parameters, specify an alignment in terms of a multiplier and
1666 * an offset. The offset or address parameter X of the intrinsic is
1667 * guaranteed to satisfy the following:
1669 * (X - align_offset) % align_mul == 0
1671 NIR_INTRINSIC_ALIGN_MUL
,
1672 NIR_INTRINSIC_ALIGN_OFFSET
,
1675 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1677 NIR_INTRINSIC_DESC_TYPE
,
1680 * The nir_alu_type of a uniform/input/output
1685 * The swizzle mask for the instructions
1686 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1688 NIR_INTRINSIC_SWIZZLE_MASK
,
1690 /* Separate source/dest access flags for copies */
1691 NIR_INTRINSIC_SRC_ACCESS
,
1692 NIR_INTRINSIC_DST_ACCESS
,
1694 /* Driver location for nir_load_patch_location_ir3 */
1695 NIR_INTRINSIC_DRIVER_LOCATION
,
1698 * Mask of nir_memory_semantics, includes ordering and visibility.
1700 NIR_INTRINSIC_MEMORY_SEMANTICS
,
1703 * Mask of nir_variable_modes affected by the memory operation.
1705 NIR_INTRINSIC_MEMORY_MODES
,
1708 * Value of nir_scope.
1710 NIR_INTRINSIC_MEMORY_SCOPE
,
1713 * Value of nir_scope.
1715 NIR_INTRINSIC_EXECUTION_SCOPE
,
1717 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1719 } nir_intrinsic_index_flag
;
1721 #define NIR_INTRINSIC_MAX_INPUTS 5
1726 uint8_t num_srcs
; /** < number of register/SSA inputs */
1728 /** number of components of each input register
1730 * If this value is 0, the number of components is given by the
1731 * num_components field of nir_intrinsic_instr. If this value is -1, the
1732 * intrinsic consumes however many components are provided and it is not
1735 int8_t src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1739 /** number of components of the output register
1741 * If this value is 0, the number of components is given by the
1742 * num_components field of nir_intrinsic_instr.
1744 uint8_t dest_components
;
1746 /** bitfield of legal bit sizes */
1747 uint8_t dest_bit_sizes
;
1749 /** the number of constant indices used by the intrinsic */
1750 uint8_t num_indices
;
1752 /** indicates the usage of intr->const_index[n] */
1753 uint8_t index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1755 /** semantic flags for calls to this intrinsic */
1756 nir_intrinsic_semantic_flag flags
;
1757 } nir_intrinsic_info
;
1759 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1761 static inline unsigned
1762 nir_intrinsic_src_components(const nir_intrinsic_instr
*intr
, unsigned srcn
)
1764 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1765 assert(srcn
< info
->num_srcs
);
1766 if (info
->src_components
[srcn
] > 0)
1767 return info
->src_components
[srcn
];
1768 else if (info
->src_components
[srcn
] == 0)
1769 return intr
->num_components
;
1771 return nir_src_num_components(intr
->src
[srcn
]);
1774 static inline unsigned
1775 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1777 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1778 if (!info
->has_dest
)
1780 else if (info
->dest_components
)
1781 return info
->dest_components
;
1783 return intr
->num_components
;
1787 * Helper to copy const_index[] from src to dst, without assuming they
1791 nir_intrinsic_copy_const_indices(nir_intrinsic_instr
*dst
, nir_intrinsic_instr
*src
)
1793 if (src
->intrinsic
== dst
->intrinsic
) {
1794 memcpy(dst
->const_index
, src
->const_index
, sizeof(dst
->const_index
));
1798 const nir_intrinsic_info
*src_info
= &nir_intrinsic_infos
[src
->intrinsic
];
1799 const nir_intrinsic_info
*dst_info
= &nir_intrinsic_infos
[dst
->intrinsic
];
1801 for (unsigned i
= 0; i
< NIR_INTRINSIC_NUM_INDEX_FLAGS
; i
++) {
1802 if (src_info
->index_map
[i
] == 0)
1805 /* require that dst instruction also uses the same const_index[]: */
1806 assert(dst_info
->index_map
[i
] > 0);
1808 dst
->const_index
[dst_info
->index_map
[i
] - 1] =
1809 src
->const_index
[src_info
->index_map
[i
] - 1];
1813 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1814 static inline type \
1815 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1817 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1818 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1819 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1821 static inline void \
1822 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1824 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1825 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1826 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1829 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1830 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1831 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1832 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1833 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1834 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1835 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1836 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1837 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1838 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1839 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1840 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1841 INTRINSIC_IDX_ACCESSORS(image_dim
, IMAGE_DIM
, enum glsl_sampler_dim
)
1842 INTRINSIC_IDX_ACCESSORS(image_array
, IMAGE_ARRAY
, bool)
1843 INTRINSIC_IDX_ACCESSORS(access
, ACCESS
, enum gl_access_qualifier
)
1844 INTRINSIC_IDX_ACCESSORS(src_access
, SRC_ACCESS
, enum gl_access_qualifier
)
1845 INTRINSIC_IDX_ACCESSORS(dst_access
, DST_ACCESS
, enum gl_access_qualifier
)
1846 INTRINSIC_IDX_ACCESSORS(format
, FORMAT
, enum pipe_format
)
1847 INTRINSIC_IDX_ACCESSORS(align_mul
, ALIGN_MUL
, unsigned)
1848 INTRINSIC_IDX_ACCESSORS(align_offset
, ALIGN_OFFSET
, unsigned)
1849 INTRINSIC_IDX_ACCESSORS(desc_type
, DESC_TYPE
, unsigned)
1850 INTRINSIC_IDX_ACCESSORS(type
, TYPE
, nir_alu_type
)
1851 INTRINSIC_IDX_ACCESSORS(swizzle_mask
, SWIZZLE_MASK
, unsigned)
1852 INTRINSIC_IDX_ACCESSORS(driver_location
, DRIVER_LOCATION
, unsigned)
1853 INTRINSIC_IDX_ACCESSORS(memory_semantics
, MEMORY_SEMANTICS
, nir_memory_semantics
)
1854 INTRINSIC_IDX_ACCESSORS(memory_modes
, MEMORY_MODES
, nir_variable_mode
)
1855 INTRINSIC_IDX_ACCESSORS(memory_scope
, MEMORY_SCOPE
, nir_scope
)
1856 INTRINSIC_IDX_ACCESSORS(execution_scope
, EXECUTION_SCOPE
, nir_scope
)
1859 nir_intrinsic_set_align(nir_intrinsic_instr
*intrin
,
1860 unsigned align_mul
, unsigned align_offset
)
1862 assert(util_is_power_of_two_nonzero(align_mul
));
1863 assert(align_offset
< align_mul
);
1864 nir_intrinsic_set_align_mul(intrin
, align_mul
);
1865 nir_intrinsic_set_align_offset(intrin
, align_offset
);
1868 /** Returns a simple alignment for a load/store intrinsic offset
1870 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1871 * and ALIGN_OFFSET parameters, this helper takes both into account and
1872 * provides a single simple alignment parameter. The offset X is guaranteed
1873 * to satisfy X % align == 0.
1875 static inline unsigned
1876 nir_intrinsic_align(const nir_intrinsic_instr
*intrin
)
1878 const unsigned align_mul
= nir_intrinsic_align_mul(intrin
);
1879 const unsigned align_offset
= nir_intrinsic_align_offset(intrin
);
1880 assert(align_offset
< align_mul
);
1881 return align_offset
? 1 << (ffs(align_offset
) - 1) : align_mul
;
1885 nir_image_intrinsic_coord_components(const nir_intrinsic_instr
*instr
);
1887 /* Converts a image_deref_* intrinsic into a image_* one */
1888 void nir_rewrite_image_intrinsic(nir_intrinsic_instr
*instr
,
1889 nir_ssa_def
*handle
, bool bindless
);
1891 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1893 nir_intrinsic_can_reorder(nir_intrinsic_instr
*instr
)
1895 if (instr
->intrinsic
== nir_intrinsic_load_deref
||
1896 instr
->intrinsic
== nir_intrinsic_load_ssbo
||
1897 instr
->intrinsic
== nir_intrinsic_bindless_image_load
||
1898 instr
->intrinsic
== nir_intrinsic_image_deref_load
||
1899 instr
->intrinsic
== nir_intrinsic_image_load
) {
1900 return nir_intrinsic_access(instr
) & ACCESS_CAN_REORDER
;
1902 const nir_intrinsic_info
*info
=
1903 &nir_intrinsic_infos
[instr
->intrinsic
];
1904 return (info
->flags
& NIR_INTRINSIC_CAN_ELIMINATE
) &&
1905 (info
->flags
& NIR_INTRINSIC_CAN_REORDER
);
1910 * \group texture information
1912 * This gives semantic information about textures which is useful to the
1913 * frontend, the backend, and lowering passes, but not the optimizer.
1918 nir_tex_src_projector
,
1919 nir_tex_src_comparator
, /* shadow comparator */
1923 nir_tex_src_min_lod
,
1924 nir_tex_src_ms_index
, /* MSAA sample index */
1925 nir_tex_src_ms_mcs
, /* MSAA compression value */
1928 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
1929 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
1930 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
1931 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
1932 nir_tex_src_texture_handle
, /* < bindless texture handle */
1933 nir_tex_src_sampler_handle
, /* < bindless sampler handle */
1934 nir_tex_src_plane
, /* < selects plane for planar textures */
1935 nir_num_tex_src_types
1940 nir_tex_src_type src_type
;
1944 nir_texop_tex
, /**< Regular texture look-up */
1945 nir_texop_txb
, /**< Texture look-up with LOD bias */
1946 nir_texop_txl
, /**< Texture look-up with explicit LOD */
1947 nir_texop_txd
, /**< Texture look-up with partial derivatives */
1948 nir_texop_txf
, /**< Texel fetch with explicit LOD */
1949 nir_texop_txf_ms
, /**< Multisample texture fetch */
1950 nir_texop_txf_ms_fb
, /**< Multisample texture fetch from framebuffer */
1951 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
1952 nir_texop_txs
, /**< Texture size */
1953 nir_texop_lod
, /**< Texture lod query */
1954 nir_texop_tg4
, /**< Texture gather */
1955 nir_texop_query_levels
, /**< Texture levels query */
1956 nir_texop_texture_samples
, /**< Texture samples query */
1957 nir_texop_samples_identical
, /**< Query whether all samples are definitely
1960 nir_texop_tex_prefetch
, /**< Regular texture look-up, eligible for pre-dispatch */
1961 nir_texop_fragment_fetch
, /**< Multisample fragment color texture fetch */
1962 nir_texop_fragment_mask_fetch
,/**< Multisample fragment mask texture fetch */
1968 enum glsl_sampler_dim sampler_dim
;
1969 nir_alu_type dest_type
;
1974 unsigned num_srcs
, coord_components
;
1975 bool is_array
, is_shadow
;
1978 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1979 * components or the new-style shadow that outputs 1 component.
1981 bool is_new_style_shadow
;
1983 /* gather component selector */
1984 unsigned component
: 2;
1986 /* gather offsets */
1987 int8_t tg4_offsets
[4][2];
1989 /* True if the texture index or handle is not dynamically uniform */
1990 bool texture_non_uniform
;
1992 /* True if the sampler index or handle is not dynamically uniform */
1993 bool sampler_non_uniform
;
1995 /** The texture index
1997 * If this texture instruction has a nir_tex_src_texture_offset source,
1998 * then the texture index is given by texture_index + texture_offset.
2000 unsigned texture_index
;
2002 /** The sampler index
2004 * The following operations do not require a sampler and, as such, this
2005 * field should be ignored:
2007 * - nir_texop_txf_ms
2010 * - nir_texop_query_levels
2011 * - nir_texop_texture_samples
2012 * - nir_texop_samples_identical
2014 * If this texture instruction has a nir_tex_src_sampler_offset source,
2015 * then the sampler index is given by sampler_index + sampler_offset.
2017 unsigned sampler_index
;
2021 * Returns true if the texture operation requires a sampler as a general rule,
2022 * see the documentation of sampler_index.
2024 * Note that the specific hw/driver backend could require to a sampler
2025 * object/configuration packet in any case, for some other reason.
2028 nir_tex_instr_need_sampler(const nir_tex_instr
*instr
)
2030 switch (instr
->op
) {
2032 case nir_texop_txf_ms
:
2035 case nir_texop_query_levels
:
2036 case nir_texop_texture_samples
:
2037 case nir_texop_samples_identical
:
2044 static inline unsigned
2045 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
2047 switch (instr
->op
) {
2048 case nir_texop_txs
: {
2050 switch (instr
->sampler_dim
) {
2051 case GLSL_SAMPLER_DIM_1D
:
2052 case GLSL_SAMPLER_DIM_BUF
:
2055 case GLSL_SAMPLER_DIM_2D
:
2056 case GLSL_SAMPLER_DIM_CUBE
:
2057 case GLSL_SAMPLER_DIM_MS
:
2058 case GLSL_SAMPLER_DIM_RECT
:
2059 case GLSL_SAMPLER_DIM_EXTERNAL
:
2060 case GLSL_SAMPLER_DIM_SUBPASS
:
2063 case GLSL_SAMPLER_DIM_3D
:
2067 unreachable("not reached");
2069 if (instr
->is_array
)
2077 case nir_texop_texture_samples
:
2078 case nir_texop_query_levels
:
2079 case nir_texop_samples_identical
:
2080 case nir_texop_fragment_mask_fetch
:
2084 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
2091 /* Returns true if this texture operation queries something about the texture
2092 * rather than actually sampling it.
2095 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
2097 switch (instr
->op
) {
2100 case nir_texop_texture_samples
:
2101 case nir_texop_query_levels
:
2102 case nir_texop_txf_ms_mcs
:
2109 case nir_texop_txf_ms
:
2110 case nir_texop_txf_ms_fb
:
2114 unreachable("Invalid texture opcode");
2119 nir_tex_instr_has_implicit_derivative(const nir_tex_instr
*instr
)
2121 switch (instr
->op
) {
2131 static inline nir_alu_type
2132 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
2134 switch (instr
->src
[src
].src_type
) {
2135 case nir_tex_src_coord
:
2136 switch (instr
->op
) {
2138 case nir_texop_txf_ms
:
2139 case nir_texop_txf_ms_fb
:
2140 case nir_texop_txf_ms_mcs
:
2141 case nir_texop_samples_identical
:
2142 return nir_type_int
;
2145 return nir_type_float
;
2148 case nir_tex_src_lod
:
2149 switch (instr
->op
) {
2152 return nir_type_int
;
2155 return nir_type_float
;
2158 case nir_tex_src_projector
:
2159 case nir_tex_src_comparator
:
2160 case nir_tex_src_bias
:
2161 case nir_tex_src_min_lod
:
2162 case nir_tex_src_ddx
:
2163 case nir_tex_src_ddy
:
2164 return nir_type_float
;
2166 case nir_tex_src_offset
:
2167 case nir_tex_src_ms_index
:
2168 case nir_tex_src_plane
:
2169 return nir_type_int
;
2171 case nir_tex_src_ms_mcs
:
2172 case nir_tex_src_texture_deref
:
2173 case nir_tex_src_sampler_deref
:
2174 case nir_tex_src_texture_offset
:
2175 case nir_tex_src_sampler_offset
:
2176 case nir_tex_src_texture_handle
:
2177 case nir_tex_src_sampler_handle
:
2178 return nir_type_uint
;
2180 case nir_num_tex_src_types
:
2181 unreachable("nir_num_tex_src_types is not a valid source type");
2184 unreachable("Invalid texture source type");
2187 static inline unsigned
2188 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
2190 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
2191 return instr
->coord_components
;
2193 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
2194 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
2197 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
2198 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
2199 if (instr
->is_array
)
2200 return instr
->coord_components
- 1;
2202 return instr
->coord_components
;
2205 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
2206 * the offset, since a cube maps to a single face.
2208 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
2209 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2211 else if (instr
->is_array
)
2212 return instr
->coord_components
- 1;
2214 return instr
->coord_components
;
2221 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
2223 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
2224 if (instr
->src
[i
].src_type
== type
)
2230 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
2231 nir_tex_src_type src_type
,
2234 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
2236 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
);
2243 nir_const_value value
[];
2244 } nir_load_const_instr
;
2247 /** Return from a function
2249 * This instruction is a classic function return. It jumps to
2250 * nir_function_impl::end_block. No return value is provided in this
2251 * instruction. Instead, the function is expected to write any return
2252 * data to a deref passed in from the caller.
2256 /** Break out of the inner-most loop
2258 * This has the same semantics as C's "break" statement.
2262 /** Jump back to the top of the inner-most loop
2264 * This has the same semantics as C's "continue" statement assuming that a
2265 * NIR loop is implemented as "while (1) { body }".
2275 /* creates a new SSA variable in an undefined state */
2280 } nir_ssa_undef_instr
;
2283 struct exec_node node
;
2285 /* The predecessor block corresponding to this source */
2286 struct nir_block
*pred
;
2291 #define nir_foreach_phi_src(phi_src, phi) \
2292 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
2293 #define nir_foreach_phi_src_safe(phi_src, phi) \
2294 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
2299 struct exec_list srcs
; /** < list of nir_phi_src */
2305 struct exec_node node
;
2308 } nir_parallel_copy_entry
;
2310 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
2311 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
2316 /* A list of nir_parallel_copy_entrys. The sources of all of the
2317 * entries are copied to the corresponding destinations "in parallel".
2318 * In other words, if we have two entries: a -> b and b -> a, the values
2321 struct exec_list entries
;
2322 } nir_parallel_copy_instr
;
2324 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
2325 type
, nir_instr_type_alu
)
2326 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
2327 type
, nir_instr_type_deref
)
2328 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
2329 type
, nir_instr_type_call
)
2330 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
2331 type
, nir_instr_type_jump
)
2332 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
2333 type
, nir_instr_type_tex
)
2334 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
2335 type
, nir_instr_type_intrinsic
)
2336 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
2337 type
, nir_instr_type_load_const
)
2338 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
2339 type
, nir_instr_type_ssa_undef
)
2340 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
2341 type
, nir_instr_type_phi
)
2342 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
2343 nir_parallel_copy_instr
, instr
,
2344 type
, nir_instr_type_parallel_copy
)
2347 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
2348 static inline type \
2349 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2351 assert(nir_src_is_const(src)); \
2352 nir_load_const_instr *load = \
2353 nir_instr_as_load_const(src.ssa->parent_instr); \
2354 assert(comp < load->def.num_components); \
2355 return nir_const_value_as_##suffix(load->value[comp], \
2356 load->def.bit_size); \
2359 static inline type \
2360 nir_src_as_##suffix(nir_src src) \
2362 assert(nir_src_num_components(src) == 1); \
2363 return nir_src_comp_as_##suffix(src, 0); \
2366 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2367 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint
)
2368 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2369 NIR_DEFINE_SRC_AS_CONST(double, float)
2371 #undef NIR_DEFINE_SRC_AS_CONST
2380 nir_ssa_scalar_is_const(nir_ssa_scalar s
)
2382 return s
.def
->parent_instr
->type
== nir_instr_type_load_const
;
2385 static inline nir_const_value
2386 nir_ssa_scalar_as_const_value(nir_ssa_scalar s
)
2388 assert(s
.comp
< s
.def
->num_components
);
2389 nir_load_const_instr
*load
= nir_instr_as_load_const(s
.def
->parent_instr
);
2390 return load
->value
[s
.comp
];
2393 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2394 static inline type \
2395 nir_ssa_scalar_as_##suffix(nir_ssa_scalar s) \
2397 return nir_const_value_as_##suffix( \
2398 nir_ssa_scalar_as_const_value(s), s.def->bit_size); \
2401 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2402 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint
)
2403 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2404 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2406 #undef NIR_DEFINE_SCALAR_AS_CONST
2409 nir_ssa_scalar_is_alu(nir_ssa_scalar s
)
2411 return s
.def
->parent_instr
->type
== nir_instr_type_alu
;
2414 static inline nir_op
2415 nir_ssa_scalar_alu_op(nir_ssa_scalar s
)
2417 return nir_instr_as_alu(s
.def
->parent_instr
)->op
;
2420 static inline nir_ssa_scalar
2421 nir_ssa_scalar_chase_alu_src(nir_ssa_scalar s
, unsigned alu_src_idx
)
2423 nir_ssa_scalar out
= { NULL
, 0 };
2425 nir_alu_instr
*alu
= nir_instr_as_alu(s
.def
->parent_instr
);
2426 assert(alu_src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
2428 /* Our component must be written */
2429 assert(s
.comp
< s
.def
->num_components
);
2430 assert(alu
->dest
.write_mask
& (1u << s
.comp
));
2432 assert(alu
->src
[alu_src_idx
].src
.is_ssa
);
2433 out
.def
= alu
->src
[alu_src_idx
].src
.ssa
;
2435 if (nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 0) {
2436 /* The ALU src is unsized so the source component follows the
2437 * destination component.
2439 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[s
.comp
];
2441 /* This is a sized source so all source components work together to
2442 * produce all the destination components. Since we need to return a
2443 * scalar, this only works if the source is a scalar.
2445 assert(nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 1);
2446 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[0];
2448 assert(out
.comp
< out
.def
->num_components
);
2457 * Control flow consists of a tree of control flow nodes, which include
2458 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2459 * instructions that always run start-to-finish. Each basic block also keeps
2460 * track of its successors (blocks which may run immediately after the current
2461 * block) and predecessors (blocks which could have run immediately before the
2462 * current block). Each function also has a start block and an end block which
2463 * all return statements point to (which is always empty). Together, all the
2464 * blocks with their predecessors and successors make up the control flow
2465 * graph (CFG) of the function. There are helpers that modify the tree of
2466 * control flow nodes while modifying the CFG appropriately; these should be
2467 * used instead of modifying the tree directly.
2474 nir_cf_node_function
2477 typedef struct nir_cf_node
{
2478 struct exec_node node
;
2479 nir_cf_node_type type
;
2480 struct nir_cf_node
*parent
;
2483 typedef struct nir_block
{
2484 nir_cf_node cf_node
;
2486 struct exec_list instr_list
; /** < list of nir_instr */
2488 /** generic block index; generated by nir_index_blocks */
2492 * Each block can only have up to 2 successors, so we put them in a simple
2493 * array - no need for anything more complicated.
2495 struct nir_block
*successors
[2];
2497 /* Set of nir_block predecessors in the CFG */
2498 struct set
*predecessors
;
2501 * this node's immediate dominator in the dominance tree - set to NULL for
2504 struct nir_block
*imm_dom
;
2506 /* This node's children in the dominance tree */
2507 unsigned num_dom_children
;
2508 struct nir_block
**dom_children
;
2510 /* Set of nir_blocks on the dominance frontier of this block */
2511 struct set
*dom_frontier
;
2514 * These two indices have the property that dom_{pre,post}_index for each
2515 * child of this block in the dominance tree will always be between
2516 * dom_pre_index and dom_post_index for this block, which makes testing if
2517 * a given block is dominated by another block an O(1) operation.
2519 int16_t dom_pre_index
, dom_post_index
;
2521 /* live in and out for this block; used for liveness analysis */
2522 BITSET_WORD
*live_in
;
2523 BITSET_WORD
*live_out
;
2527 nir_block_is_reachable(nir_block
*b
)
2529 /* See also nir_block_dominates */
2530 return b
->dom_post_index
!= -1;
2533 static inline nir_instr
*
2534 nir_block_first_instr(nir_block
*block
)
2536 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
2537 return exec_node_data(nir_instr
, head
, node
);
2540 static inline nir_instr
*
2541 nir_block_last_instr(nir_block
*block
)
2543 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
2544 return exec_node_data(nir_instr
, tail
, node
);
2548 nir_block_ends_in_jump(nir_block
*block
)
2550 return !exec_list_is_empty(&block
->instr_list
) &&
2551 nir_block_last_instr(block
)->type
== nir_instr_type_jump
;
2554 #define nir_foreach_instr(instr, block) \
2555 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
2556 #define nir_foreach_instr_reverse(instr, block) \
2557 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
2558 #define nir_foreach_instr_safe(instr, block) \
2559 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
2560 #define nir_foreach_instr_reverse_safe(instr, block) \
2561 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
2564 nir_selection_control_none
= 0x0,
2565 nir_selection_control_flatten
= 0x1,
2566 nir_selection_control_dont_flatten
= 0x2,
2567 } nir_selection_control
;
2569 typedef struct nir_if
{
2570 nir_cf_node cf_node
;
2572 nir_selection_control control
;
2574 struct exec_list then_list
; /** < list of nir_cf_node */
2575 struct exec_list else_list
; /** < list of nir_cf_node */
2581 /** Instruction that generates nif::condition. */
2582 nir_instr
*conditional_instr
;
2584 /** Block within ::nif that has the break instruction. */
2585 nir_block
*break_block
;
2587 /** Last block for the then- or else-path that does not contain the break. */
2588 nir_block
*continue_from_block
;
2590 /** True when ::break_block is in the else-path of ::nif. */
2591 bool continue_from_then
;
2594 /* This is true if the terminators exact trip count is unknown. For
2597 * for (int i = 0; i < imin(x, 4); i++)
2600 * Here loop analysis would have set a max_trip_count of 4 however we dont
2601 * know for sure that this is the exact trip count.
2603 bool exact_trip_count_unknown
;
2605 struct list_head loop_terminator_link
;
2606 } nir_loop_terminator
;
2609 /* Estimated cost (in number of instructions) of the loop */
2610 unsigned instr_cost
;
2612 /* Guessed trip count based on array indexing */
2613 unsigned guessed_trip_count
;
2615 /* Maximum number of times the loop is run (if known) */
2616 unsigned max_trip_count
;
2618 /* Do we know the exact number of times the loop will be run */
2619 bool exact_trip_count_known
;
2621 /* Unroll the loop regardless of its size */
2624 /* Does the loop contain complex loop terminators, continues or other
2625 * complex behaviours? If this is true we can't rely on
2626 * loop_terminator_list to be complete or accurate.
2630 nir_loop_terminator
*limiting_terminator
;
2632 /* A list of loop_terminators terminating this loop. */
2633 struct list_head loop_terminator_list
;
2637 nir_loop_control_none
= 0x0,
2638 nir_loop_control_unroll
= 0x1,
2639 nir_loop_control_dont_unroll
= 0x2,
2643 nir_cf_node cf_node
;
2645 struct exec_list body
; /** < list of nir_cf_node */
2647 nir_loop_info
*info
;
2648 nir_loop_control control
;
2649 bool partially_unrolled
;
2653 * Various bits of metadata that can may be created or required by
2654 * optimization and analysis passes
2657 nir_metadata_none
= 0x0,
2659 /** Indicates that nir_block::index values are valid.
2661 * The start block has index 0 and they increase through a natural walk of
2662 * the CFG. nir_function_impl::num_blocks is the number of blocks and
2663 * every block index is in the range [0, nir_function_impl::num_blocks].
2665 * A pass can preserve this metadata type if it doesn't touch the CFG.
2667 nir_metadata_block_index
= 0x1,
2669 /** Indicates that block dominance information is valid
2673 * - nir_block::num_dom_children
2674 * - nir_block::dom_children
2675 * - nir_block::dom_frontier
2676 * - nir_block::dom_pre_index
2677 * - nir_block::dom_post_index
2679 * A pass can preserve this metadata type if it doesn't touch the CFG.
2681 nir_metadata_dominance
= 0x2,
2683 /** Indicates that SSA def data-flow liveness information is valid
2687 * - nir_ssa_def::live_index
2688 * - nir_block::live_in
2689 * - nir_block::live_out
2691 * A pass can preserve this metadata type if it never adds or removes any
2692 * SSA defs (most passes shouldn't preserve this metadata type).
2694 nir_metadata_live_ssa_defs
= 0x4,
2696 /** A dummy metadata value to track when a pass forgot to call
2697 * nir_metadata_preserve.
2699 * A pass should always clear this value even if it doesn't make any
2700 * progress to indicate that it thought about preserving metadata.
2702 nir_metadata_not_properly_reset
= 0x8,
2704 /** Indicates that loop analysis information is valid.
2706 * This includes everything pointed to by nir_loop::info.
2708 * A pass can preserve this metadata type if it is guaranteed to not affect
2709 * any loop metadata. However, since loop metadata includes things like
2710 * loop counts which depend on arithmetic in the loop, this is very hard to
2711 * determine. Most passes shouldn't preserve this metadata type.
2713 nir_metadata_loop_analysis
= 0x10,
2717 * This includes all nir_metadata flags except not_properly_reset. Passes
2718 * which do not change the shader in any way should call
2720 * nir_metadata_preserve(impl, nir_metadata_all);
2722 nir_metadata_all
= ~nir_metadata_not_properly_reset
,
2726 nir_cf_node cf_node
;
2728 /** pointer to the function of which this is an implementation */
2729 struct nir_function
*function
;
2731 struct exec_list body
; /** < list of nir_cf_node */
2733 nir_block
*end_block
;
2735 /** list for all local variables in the function */
2736 struct exec_list locals
;
2738 /** list of local registers in the function */
2739 struct exec_list registers
;
2741 /** next available local register index */
2744 /** next available SSA value index */
2747 /* total number of basic blocks, only valid when block_index_dirty = false */
2748 unsigned num_blocks
;
2750 nir_metadata valid_metadata
;
2751 } nir_function_impl
;
2753 #define nir_foreach_function_temp_variable(var, impl) \
2754 foreach_list_typed(nir_variable, var, node, &(impl)->locals)
2756 #define nir_foreach_function_temp_variable_safe(var, impl) \
2757 foreach_list_typed_safe(nir_variable, var, node, &(impl)->locals)
2759 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2760 nir_start_block(nir_function_impl
*impl
)
2762 return (nir_block
*) impl
->body
.head_sentinel
.next
;
2765 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2766 nir_impl_last_block(nir_function_impl
*impl
)
2768 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
2771 static inline nir_cf_node
*
2772 nir_cf_node_next(nir_cf_node
*node
)
2774 struct exec_node
*next
= exec_node_get_next(&node
->node
);
2775 if (exec_node_is_tail_sentinel(next
))
2778 return exec_node_data(nir_cf_node
, next
, node
);
2781 static inline nir_cf_node
*
2782 nir_cf_node_prev(nir_cf_node
*node
)
2784 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
2785 if (exec_node_is_head_sentinel(prev
))
2788 return exec_node_data(nir_cf_node
, prev
, node
);
2792 nir_cf_node_is_first(const nir_cf_node
*node
)
2794 return exec_node_is_head_sentinel(node
->node
.prev
);
2798 nir_cf_node_is_last(const nir_cf_node
*node
)
2800 return exec_node_is_tail_sentinel(node
->node
.next
);
2803 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
2804 type
, nir_cf_node_block
)
2805 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
2806 type
, nir_cf_node_if
)
2807 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
2808 type
, nir_cf_node_loop
)
2809 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
2810 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
2812 static inline nir_block
*
2813 nir_if_first_then_block(nir_if
*if_stmt
)
2815 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
2816 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2819 static inline nir_block
*
2820 nir_if_last_then_block(nir_if
*if_stmt
)
2822 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
2823 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2826 static inline nir_block
*
2827 nir_if_first_else_block(nir_if
*if_stmt
)
2829 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
2830 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2833 static inline nir_block
*
2834 nir_if_last_else_block(nir_if
*if_stmt
)
2836 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
2837 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2840 static inline nir_block
*
2841 nir_loop_first_block(nir_loop
*loop
)
2843 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
2844 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2847 static inline nir_block
*
2848 nir_loop_last_block(nir_loop
*loop
)
2850 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
2851 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2855 * Return true if this list of cf_nodes contains a single empty block.
2858 nir_cf_list_is_empty_block(struct exec_list
*cf_list
)
2860 if (exec_list_is_singular(cf_list
)) {
2861 struct exec_node
*head
= exec_list_get_head(cf_list
);
2863 nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2864 return exec_list_is_empty(&block
->instr_list
);
2870 uint8_t num_components
;
2874 typedef struct nir_function
{
2875 struct exec_node node
;
2878 struct nir_shader
*shader
;
2880 unsigned num_params
;
2881 nir_parameter
*params
;
2883 /** The implementation of this function.
2885 * If the function is only declared and not implemented, this is NULL.
2887 nir_function_impl
*impl
;
2893 nir_lower_imul64
= (1 << 0),
2894 nir_lower_isign64
= (1 << 1),
2895 /** Lower all int64 modulus and division opcodes */
2896 nir_lower_divmod64
= (1 << 2),
2897 /** Lower all 64-bit umul_high and imul_high opcodes */
2898 nir_lower_imul_high64
= (1 << 3),
2899 nir_lower_mov64
= (1 << 4),
2900 nir_lower_icmp64
= (1 << 5),
2901 nir_lower_iadd64
= (1 << 6),
2902 nir_lower_iabs64
= (1 << 7),
2903 nir_lower_ineg64
= (1 << 8),
2904 nir_lower_logic64
= (1 << 9),
2905 nir_lower_minmax64
= (1 << 10),
2906 nir_lower_shift64
= (1 << 11),
2907 nir_lower_imul_2x32_64
= (1 << 12),
2908 nir_lower_extract64
= (1 << 13),
2909 nir_lower_ufind_msb64
= (1 << 14),
2910 } nir_lower_int64_options
;
2913 nir_lower_drcp
= (1 << 0),
2914 nir_lower_dsqrt
= (1 << 1),
2915 nir_lower_drsq
= (1 << 2),
2916 nir_lower_dtrunc
= (1 << 3),
2917 nir_lower_dfloor
= (1 << 4),
2918 nir_lower_dceil
= (1 << 5),
2919 nir_lower_dfract
= (1 << 6),
2920 nir_lower_dround_even
= (1 << 7),
2921 nir_lower_dmod
= (1 << 8),
2922 nir_lower_dsub
= (1 << 9),
2923 nir_lower_ddiv
= (1 << 10),
2924 nir_lower_fp64_full_software
= (1 << 11),
2925 } nir_lower_doubles_options
;
2928 nir_divergence_single_prim_per_subgroup
= (1 << 0),
2929 nir_divergence_single_patch_per_tcs_subgroup
= (1 << 1),
2930 nir_divergence_single_patch_per_tes_subgroup
= (1 << 2),
2931 nir_divergence_view_index_uniform
= (1 << 3),
2932 } nir_divergence_options
;
2934 typedef struct nir_shader_compiler_options
{
2940 /** Lowers flrp when it does not support doubles */
2947 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2948 bool lower_bitfield_extract
;
2949 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
2950 bool lower_bitfield_extract_to_shifts
;
2951 /** Lowers bitfield_insert to bfi/bfm */
2952 bool lower_bitfield_insert
;
2953 /** Lowers bitfield_insert to compares, and shifts. */
2954 bool lower_bitfield_insert_to_shifts
;
2955 /** Lowers bitfield_insert to bfm/bitfield_select. */
2956 bool lower_bitfield_insert_to_bitfield_select
;
2957 /** Lowers bitfield_reverse to shifts. */
2958 bool lower_bitfield_reverse
;
2959 /** Lowers bit_count to shifts. */
2960 bool lower_bit_count
;
2961 /** Lowers ifind_msb to compare and ufind_msb */
2962 bool lower_ifind_msb
;
2963 /** Lowers find_lsb to ufind_msb and logic ops */
2964 bool lower_find_lsb
;
2965 bool lower_uadd_carry
;
2966 bool lower_usub_borrow
;
2967 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2968 bool lower_mul_high
;
2969 /** lowers fneg and ineg to fsub and isub. */
2971 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2974 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2977 /* lower fall_equalN/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
2978 bool lower_vector_cmp
;
2980 /** enables rules to lower idiv by power-of-two: */
2983 /** enable rules to avoid bit ops */
2986 /** enables rules to lower isign to imin+imax */
2989 /** enables rules to lower fsign to fsub and flt */
2992 /* lower fdph to fdot4 */
2995 /** lower fdot to fmul and fsum/fadd. */
2998 /* Does the native fdot instruction replicate its result for four
2999 * components? If so, then opt_algebraic_late will turn all fdotN
3000 * instructions into fdot_replicatedN instructions.
3002 bool fdot_replicates
;
3004 /** lowers ffloor to fsub+ffract: */
3007 /** lowers ffract to fsub+ffloor: */
3010 /** lowers fceil to fneg+ffloor+fneg: */
3017 bool lower_pack_half_2x16
;
3018 bool lower_pack_unorm_2x16
;
3019 bool lower_pack_snorm_2x16
;
3020 bool lower_pack_unorm_4x8
;
3021 bool lower_pack_snorm_4x8
;
3022 bool lower_unpack_half_2x16
;
3023 bool lower_unpack_unorm_2x16
;
3024 bool lower_unpack_snorm_2x16
;
3025 bool lower_unpack_unorm_4x8
;
3026 bool lower_unpack_snorm_4x8
;
3028 bool lower_pack_split
;
3030 bool lower_extract_byte
;
3031 bool lower_extract_word
;
3033 bool lower_all_io_to_temps
;
3034 bool lower_all_io_to_elements
;
3036 /* Indicates that the driver only has zero-based vertex id */
3037 bool vertex_id_zero_based
;
3040 * If enabled, gl_BaseVertex will be lowered as:
3041 * is_indexed_draw (~0/0) & firstvertex
3043 bool lower_base_vertex
;
3046 * If enabled, gl_HelperInvocation will be lowered as:
3048 * !((1 << sample_id) & sample_mask_in))
3050 * This depends on some possibly hw implementation details, which may
3051 * not be true for all hw. In particular that the FS is only executed
3052 * for covered samples or for helper invocations. So, do not blindly
3053 * enable this option.
3055 * Note: See also issue #22 in ARB_shader_image_load_store
3057 bool lower_helper_invocation
;
3060 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
3062 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
3063 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
3065 bool optimize_sample_mask_in
;
3067 bool lower_cs_local_index_from_id
;
3068 bool lower_cs_local_id_from_index
;
3070 bool lower_device_index_to_zero
;
3072 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
3073 bool lower_wpos_pntc
;
3076 * Set if nir_op_[iu]hadd and nir_op_[iu]rhadd instructions should be
3077 * lowered to simple arithmetic.
3079 * If this flag is set, the lowering will be applied to all bit-sizes of
3080 * these instructions.
3082 * \sa ::lower_hadd64
3087 * Set if only 64-bit nir_op_[iu]hadd and nir_op_[iu]rhadd instructions
3088 * should be lowered to simple arithmetic.
3090 * If this flag is set, the lowering will be applied to only 64-bit
3091 * versions of these instructions.
3098 * Set if nir_op_add_sat and nir_op_usub_sat should be lowered to simple
3101 * If this flag is set, the lowering will be applied to all bit-sizes of
3102 * these instructions.
3104 * \sa ::lower_usub_sat64
3109 * Set if only 64-bit nir_op_usub_sat should be lowered to simple
3112 * \sa ::lower_add_sat
3114 bool lower_usub_sat64
;
3117 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
3118 * for IO purposes and would prefer loads/stores be vectorized.
3121 bool lower_to_scalar
;
3124 * Whether nir_opt_vectorize should only create 16-bit 2D vectors.
3126 bool vectorize_vec2_16bit
;
3129 * Should the linker unify inputs_read/outputs_written between adjacent
3130 * shader stages which are linked into a single program?
3132 bool unify_interfaces
;
3135 * Should nir_lower_io() create load_interpolated_input intrinsics?
3137 * If not, it generates regular load_input intrinsics and interpolation
3138 * information must be inferred from the list of input nir_variables.
3140 bool use_interpolated_input_intrinsics
;
3142 /* Lowers when 32x32->64 bit multiplication is not supported */
3143 bool lower_mul_2x32_64
;
3145 /* Lowers when rotate instruction is not supported */
3149 * Backend supports imul24, and would like to use it (when possible)
3150 * for address/offset calculation. If true, driver should call
3151 * nir_lower_amul(). (If not set, amul will automatically be lowered
3156 /** Backend supports umul24, if not set umul24 will automatically be lowered
3157 * to imul with masked inputs */
3160 /** Backend supports umad24, if not set umad24 will automatically be lowered
3161 * to imul with masked inputs and iadd */
3164 /* Whether to generate only scoped_barrier intrinsics instead of the set of
3165 * memory and control barrier intrinsics based on GLSL.
3167 bool use_scoped_barrier
;
3170 * Is this the Intel vec4 backend?
3172 * Used to inhibit algebraic optimizations that are known to be harmful on
3173 * the Intel vec4 backend. This is generally applicable to any
3174 * optimization that might cause more immediate values to be used in
3175 * 3-source (e.g., ffma and flrp) instructions.
3179 /** Lower nir_op_ibfe and nir_op_ubfe that have two constant sources. */
3180 bool lower_bfe_with_two_constants
;
3182 /** Whether 8-bit ALU is supported. */
3183 bool support_8bit_alu
;
3185 /** Whether 16-bit ALU is supported. */
3186 bool support_16bit_alu
;
3188 unsigned max_unroll_iterations
;
3190 nir_lower_int64_options lower_int64_options
;
3191 nir_lower_doubles_options lower_doubles_options
;
3192 } nir_shader_compiler_options
;
3194 typedef struct nir_shader
{
3195 /** list of uniforms (nir_variable) */
3196 struct exec_list uniforms
;
3198 /** list of inputs (nir_variable) */
3199 struct exec_list inputs
;
3201 /** list of outputs (nir_variable) */
3202 struct exec_list outputs
;
3204 /** list of shared compute variables (nir_variable) */
3205 struct exec_list shared
;
3207 /** Set of driver-specific options for the shader.
3209 * The memory for the options is expected to be kept in a single static
3210 * copy by the driver.
3212 const struct nir_shader_compiler_options
*options
;
3214 /** Various bits of compile-time information about a given shader */
3215 struct shader_info info
;
3217 /** list of global variables in the shader (nir_variable) */
3218 struct exec_list globals
;
3220 /** list of system value variables in the shader (nir_variable) */
3221 struct exec_list system_values
;
3223 struct exec_list functions
; /** < list of nir_function */
3226 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
3229 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
3231 /** Size in bytes of required scratch space */
3232 unsigned scratch_size
;
3234 /** Constant data associated with this shader.
3236 * Constant data is loaded through load_constant intrinsics. See also
3237 * nir_opt_large_constants.
3239 void *constant_data
;
3240 unsigned constant_data_size
;
3243 #define nir_foreach_function(func, shader) \
3244 foreach_list_typed(nir_function, func, node, &(shader)->functions)
3246 static inline nir_function_impl
*
3247 nir_shader_get_entrypoint(nir_shader
*shader
)
3249 nir_function
*func
= NULL
;
3251 nir_foreach_function(function
, shader
) {
3252 assert(func
== NULL
);
3253 if (function
->is_entrypoint
) {
3264 assert(func
->num_params
== 0);
3269 nir_shader
*nir_shader_create(void *mem_ctx
,
3270 gl_shader_stage stage
,
3271 const nir_shader_compiler_options
*options
,
3274 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
3276 void nir_reg_remove(nir_register
*reg
);
3278 /** Adds a variable to the appropriate list in nir_shader */
3279 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
3282 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
3284 assert(var
->data
.mode
== nir_var_function_temp
);
3285 exec_list_push_tail(&impl
->locals
, &var
->node
);
3288 /** creates a variable, sets a few defaults, and adds it to the list */
3289 nir_variable
*nir_variable_create(nir_shader
*shader
,
3290 nir_variable_mode mode
,
3291 const struct glsl_type
*type
,
3293 /** creates a local variable and adds it to the list */
3294 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
3295 const struct glsl_type
*type
,
3298 /** creates a function and adds it to the shader's list of functions */
3299 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
3301 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
3302 /** creates a function_impl that isn't tied to any particular function */
3303 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
3305 nir_block
*nir_block_create(nir_shader
*shader
);
3306 nir_if
*nir_if_create(nir_shader
*shader
);
3307 nir_loop
*nir_loop_create(nir_shader
*shader
);
3309 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
3311 /** requests that the given pieces of metadata be generated */
3312 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
3313 /** dirties all but the preserved metadata */
3314 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
3315 /** Preserves all metadata for the given shader */
3316 void nir_shader_preserve_all_metadata(nir_shader
*shader
);
3318 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
3319 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
3321 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
3322 nir_deref_type deref_type
);
3324 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
3326 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
3327 unsigned num_components
,
3330 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
3331 nir_intrinsic_op op
);
3333 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
3334 nir_function
*callee
);
3336 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
3338 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
3340 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
3342 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
3343 unsigned num_components
,
3346 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
3349 * NIR Cursors and Instruction Insertion API
3352 * A tiny struct representing a point to insert/extract instructions or
3353 * control flow nodes. Helps reduce the combinatorial explosion of possible
3354 * points to insert/extract.
3356 * \sa nir_control_flow.h
3359 nir_cursor_before_block
,
3360 nir_cursor_after_block
,
3361 nir_cursor_before_instr
,
3362 nir_cursor_after_instr
,
3363 } nir_cursor_option
;
3366 nir_cursor_option option
;
3373 static inline nir_block
*
3374 nir_cursor_current_block(nir_cursor cursor
)
3376 if (cursor
.option
== nir_cursor_before_instr
||
3377 cursor
.option
== nir_cursor_after_instr
) {
3378 return cursor
.instr
->block
;
3380 return cursor
.block
;
3384 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
3386 static inline nir_cursor
3387 nir_before_block(nir_block
*block
)
3390 cursor
.option
= nir_cursor_before_block
;
3391 cursor
.block
= block
;
3395 static inline nir_cursor
3396 nir_after_block(nir_block
*block
)
3399 cursor
.option
= nir_cursor_after_block
;
3400 cursor
.block
= block
;
3404 static inline nir_cursor
3405 nir_before_instr(nir_instr
*instr
)
3408 cursor
.option
= nir_cursor_before_instr
;
3409 cursor
.instr
= instr
;
3413 static inline nir_cursor
3414 nir_after_instr(nir_instr
*instr
)
3417 cursor
.option
= nir_cursor_after_instr
;
3418 cursor
.instr
= instr
;
3422 static inline nir_cursor
3423 nir_after_block_before_jump(nir_block
*block
)
3425 nir_instr
*last_instr
= nir_block_last_instr(block
);
3426 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
3427 return nir_before_instr(last_instr
);
3429 return nir_after_block(block
);
3433 static inline nir_cursor
3434 nir_before_src(nir_src
*src
, bool is_if_condition
)
3436 if (is_if_condition
) {
3437 nir_block
*prev_block
=
3438 nir_cf_node_as_block(nir_cf_node_prev(&src
->parent_if
->cf_node
));
3439 assert(!nir_block_ends_in_jump(prev_block
));
3440 return nir_after_block(prev_block
);
3441 } else if (src
->parent_instr
->type
== nir_instr_type_phi
) {
3443 nir_phi_instr
*cond_phi
= nir_instr_as_phi(src
->parent_instr
);
3445 nir_foreach_phi_src(phi_src
, cond_phi
) {
3446 if (phi_src
->src
.ssa
== src
->ssa
) {
3453 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
3454 * to have a more specific name.
3456 nir_phi_src
*phi_src
= LIST_ENTRY(nir_phi_src
, src
, src
);
3457 return nir_after_block_before_jump(phi_src
->pred
);
3459 return nir_before_instr(src
->parent_instr
);
3463 static inline nir_cursor
3464 nir_before_cf_node(nir_cf_node
*node
)
3466 if (node
->type
== nir_cf_node_block
)
3467 return nir_before_block(nir_cf_node_as_block(node
));
3469 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
3472 static inline nir_cursor
3473 nir_after_cf_node(nir_cf_node
*node
)
3475 if (node
->type
== nir_cf_node_block
)
3476 return nir_after_block(nir_cf_node_as_block(node
));
3478 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
3481 static inline nir_cursor
3482 nir_after_phis(nir_block
*block
)
3484 nir_foreach_instr(instr
, block
) {
3485 if (instr
->type
!= nir_instr_type_phi
)
3486 return nir_before_instr(instr
);
3488 return nir_after_block(block
);
3491 static inline nir_cursor
3492 nir_after_cf_node_and_phis(nir_cf_node
*node
)
3494 if (node
->type
== nir_cf_node_block
)
3495 return nir_after_block(nir_cf_node_as_block(node
));
3497 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
3499 return nir_after_phis(block
);
3502 static inline nir_cursor
3503 nir_before_cf_list(struct exec_list
*cf_list
)
3505 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
3506 exec_list_get_head(cf_list
), node
);
3507 return nir_before_cf_node(first_node
);
3510 static inline nir_cursor
3511 nir_after_cf_list(struct exec_list
*cf_list
)
3513 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
3514 exec_list_get_tail(cf_list
), node
);
3515 return nir_after_cf_node(last_node
);
3519 * Insert a NIR instruction at the given cursor.
3521 * Note: This does not update the cursor.
3523 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
3526 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
3528 nir_instr_insert(nir_before_instr(instr
), before
);
3532 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
3534 nir_instr_insert(nir_after_instr(instr
), after
);
3538 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
3540 nir_instr_insert(nir_before_block(block
), before
);
3544 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
3546 nir_instr_insert(nir_after_block(block
), after
);
3550 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
3552 nir_instr_insert(nir_before_cf_node(node
), before
);
3556 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
3558 nir_instr_insert(nir_after_cf_node(node
), after
);
3562 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
3564 nir_instr_insert(nir_before_cf_list(list
), before
);
3568 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
3570 nir_instr_insert(nir_after_cf_list(list
), after
);
3573 void nir_instr_remove_v(nir_instr
*instr
);
3575 static inline nir_cursor
3576 nir_instr_remove(nir_instr
*instr
)
3579 nir_instr
*prev
= nir_instr_prev(instr
);
3581 cursor
= nir_after_instr(prev
);
3583 cursor
= nir_before_block(instr
->block
);
3585 nir_instr_remove_v(instr
);
3591 nir_ssa_def
*nir_instr_ssa_def(nir_instr
*instr
);
3593 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
3594 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
3595 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
3596 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
3598 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
3599 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
3600 bool nir_foreach_phi_src_leaving_block(nir_block
*instr
,
3601 nir_foreach_src_cb cb
,
3604 nir_const_value
*nir_src_as_const_value(nir_src src
);
3606 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
3607 static inline c_type * \
3608 nir_src_as_ ## name (nir_src src) \
3610 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
3611 ? cast_macro(src.ssa->parent_instr) : NULL; \
3614 NIR_SRC_AS_(alu_instr
, nir_alu_instr
, nir_instr_type_alu
, nir_instr_as_alu
)
3615 NIR_SRC_AS_(intrinsic
, nir_intrinsic_instr
,
3616 nir_instr_type_intrinsic
, nir_instr_as_intrinsic
)
3617 NIR_SRC_AS_(deref
, nir_deref_instr
, nir_instr_type_deref
, nir_instr_as_deref
)
3619 bool nir_src_is_dynamically_uniform(nir_src src
);
3620 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
3621 bool nir_instrs_equal(const nir_instr
*instr1
, const nir_instr
*instr2
);
3622 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
3623 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
3624 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
3625 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
3628 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
3629 unsigned num_components
, unsigned bit_size
,
3631 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
3632 unsigned num_components
, unsigned bit_size
,
3635 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
3636 const struct glsl_type
*type
,
3639 assert(glsl_type_is_vector_or_scalar(type
));
3640 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
3641 glsl_get_bit_size(type
), name
);
3643 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
3644 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
3645 nir_instr
*after_me
);
3647 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
3650 * finds the next basic block in source-code order, returns NULL if there is
3654 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
3656 /* Performs the opposite of nir_block_cf_tree_next() */
3658 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
3660 /* Gets the first block in a CF node in source-code order */
3662 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
3664 /* Gets the last block in a CF node in source-code order */
3666 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
3668 /* Gets the next block after a CF node in source-code order */
3670 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
3672 /* Macros for loops that visit blocks in source-code order */
3674 #define nir_foreach_block(block, impl) \
3675 for (nir_block *block = nir_start_block(impl); block != NULL; \
3676 block = nir_block_cf_tree_next(block))
3678 #define nir_foreach_block_safe(block, impl) \
3679 for (nir_block *block = nir_start_block(impl), \
3680 *next = nir_block_cf_tree_next(block); \
3682 block = next, next = nir_block_cf_tree_next(block))
3684 #define nir_foreach_block_reverse(block, impl) \
3685 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
3686 block = nir_block_cf_tree_prev(block))
3688 #define nir_foreach_block_reverse_safe(block, impl) \
3689 for (nir_block *block = nir_impl_last_block(impl), \
3690 *prev = nir_block_cf_tree_prev(block); \
3692 block = prev, prev = nir_block_cf_tree_prev(block))
3694 #define nir_foreach_block_in_cf_node(block, node) \
3695 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
3696 block != nir_cf_node_cf_tree_next(node); \
3697 block = nir_block_cf_tree_next(block))
3699 /* If the following CF node is an if, this function returns that if.
3700 * Otherwise, it returns NULL.
3702 nir_if
*nir_block_get_following_if(nir_block
*block
);
3704 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
3706 void nir_index_local_regs(nir_function_impl
*impl
);
3707 void nir_index_ssa_defs(nir_function_impl
*impl
);
3708 unsigned nir_index_instrs(nir_function_impl
*impl
);
3710 void nir_index_blocks(nir_function_impl
*impl
);
3712 void nir_index_vars(nir_shader
*shader
, nir_function_impl
*impl
, nir_variable_mode modes
);
3714 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
3715 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
3716 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
3717 void nir_print_deref(const nir_deref_instr
*deref
, FILE *fp
);
3719 /** Shallow clone of a single ALU instruction. */
3720 nir_alu_instr
*nir_alu_instr_clone(nir_shader
*s
, const nir_alu_instr
*orig
);
3722 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
3723 nir_function_impl
*nir_function_impl_clone(nir_shader
*shader
,
3724 const nir_function_impl
*fi
);
3725 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
3726 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
3728 void nir_shader_replace(nir_shader
*dest
, nir_shader
*src
);
3730 void nir_shader_serialize_deserialize(nir_shader
*s
);
3733 void nir_validate_shader(nir_shader
*shader
, const char *when
);
3734 void nir_metadata_set_validation_flag(nir_shader
*shader
);
3735 void nir_metadata_check_validation_flag(nir_shader
*shader
);
3738 should_skip_nir(const char *name
)
3740 static const char *list
= NULL
;
3742 /* Comma separated list of names to skip. */
3743 list
= getenv("NIR_SKIP");
3751 return comma_separated_list_contains(list
, name
);
3755 should_clone_nir(void)
3757 static int should_clone
= -1;
3758 if (should_clone
< 0)
3759 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
3761 return should_clone
;
3765 should_serialize_deserialize_nir(void)
3767 static int test_serialize
= -1;
3768 if (test_serialize
< 0)
3769 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
3771 return test_serialize
;
3775 should_print_nir(void)
3777 static int should_print
= -1;
3778 if (should_print
< 0)
3779 should_print
= env_var_as_boolean("NIR_PRINT", false);
3781 return should_print
;
3784 static inline void nir_validate_shader(nir_shader
*shader
, const char *when
) { (void) shader
; (void)when
; }
3785 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
3786 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
3787 static inline bool should_skip_nir(UNUSED
const char *pass_name
) { return false; }
3788 static inline bool should_clone_nir(void) { return false; }
3789 static inline bool should_serialize_deserialize_nir(void) { return false; }
3790 static inline bool should_print_nir(void) { return false; }
3793 #define _PASS(pass, nir, do_pass) do { \
3794 if (should_skip_nir(#pass)) { \
3795 printf("skipping %s\n", #pass); \
3799 nir_validate_shader(nir, "after " #pass); \
3800 if (should_clone_nir()) { \
3801 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3802 nir_shader_replace(nir, clone); \
3804 if (should_serialize_deserialize_nir()) { \
3805 nir_shader_serialize_deserialize(nir); \
3809 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3810 nir_metadata_set_validation_flag(nir); \
3811 if (should_print_nir()) \
3812 printf("%s\n", #pass); \
3813 if (pass(nir, ##__VA_ARGS__)) { \
3815 if (should_print_nir()) \
3816 nir_print_shader(nir, stdout); \
3817 nir_metadata_check_validation_flag(nir); \
3821 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3822 if (should_print_nir()) \
3823 printf("%s\n", #pass); \
3824 pass(nir, ##__VA_ARGS__); \
3825 if (should_print_nir()) \
3826 nir_print_shader(nir, stdout); \
3829 #define NIR_SKIP(name) should_skip_nir(#name)
3831 /** An instruction filtering callback
3833 * Returns true if the instruction should be processed and false otherwise.
3835 typedef bool (*nir_instr_filter_cb
)(const nir_instr
*, const void *);
3837 /** A simple instruction lowering callback
3839 * Many instruction lowering passes can be written as a simple function which
3840 * takes an instruction as its input and returns a sequence of instructions
3841 * that implement the consumed instruction. This function type represents
3842 * such a lowering function. When called, a function with this prototype
3843 * should either return NULL indicating that no lowering needs to be done or
3844 * emit a sequence of instructions using the provided builder (whose cursor
3845 * will already be placed after the instruction to be lowered) and return the
3846 * resulting nir_ssa_def.
3848 typedef nir_ssa_def
*(*nir_lower_instr_cb
)(struct nir_builder
*,
3849 nir_instr
*, void *);
3852 * Special return value for nir_lower_instr_cb when some progress occurred
3853 * (like changing an input to the instr) that didn't result in a replacement
3854 * SSA def being generated.
3856 #define NIR_LOWER_INSTR_PROGRESS ((nir_ssa_def *)(uintptr_t)1)
3858 /** Iterate over all the instructions in a nir_function_impl and lower them
3859 * using the provided callbacks
3861 * This function implements the guts of a standard lowering pass for you. It
3862 * iterates over all of the instructions in a nir_function_impl and calls the
3863 * filter callback on each one. If the filter callback returns true, it then
3864 * calls the lowering call back on the instruction. (Splitting it this way
3865 * allows us to avoid some save/restore work for instructions we know won't be
3866 * lowered.) If the instruction is dead after the lowering is complete, it
3867 * will be removed. If new instructions are added, the lowering callback will
3868 * also be called on them in case multiple lowerings are required.
3870 * The metadata for the nir_function_impl will also be updated. If any blocks
3871 * are added (they cannot be removed), dominance and block indices will be
3874 bool nir_function_impl_lower_instructions(nir_function_impl
*impl
,
3875 nir_instr_filter_cb filter
,
3876 nir_lower_instr_cb lower
,
3878 bool nir_shader_lower_instructions(nir_shader
*shader
,
3879 nir_instr_filter_cb filter
,
3880 nir_lower_instr_cb lower
,
3883 void nir_calc_dominance_impl(nir_function_impl
*impl
);
3884 void nir_calc_dominance(nir_shader
*shader
);
3886 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
3887 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
3888 bool nir_block_is_unreachable(nir_block
*block
);
3890 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
3891 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
3893 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
3894 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
3896 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
3897 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
3899 int nir_gs_count_vertices(const nir_shader
*shader
);
3901 bool nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3902 bool nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3903 bool nir_split_var_copies(nir_shader
*shader
);
3904 bool nir_split_per_member_structs(nir_shader
*shader
);
3905 bool nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
);
3907 bool nir_lower_returns_impl(nir_function_impl
*impl
);
3908 bool nir_lower_returns(nir_shader
*shader
);
3910 void nir_inline_function_impl(struct nir_builder
*b
,
3911 const nir_function_impl
*impl
,
3912 nir_ssa_def
**params
);
3913 bool nir_inline_functions(nir_shader
*shader
);
3915 bool nir_propagate_invariant(nir_shader
*shader
);
3917 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
3918 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
3919 nir_intrinsic_instr
*copy
);
3920 bool nir_lower_var_copies(nir_shader
*shader
);
3922 void nir_fixup_deref_modes(nir_shader
*shader
);
3924 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
3927 nir_lower_direct_array_deref_of_vec_load
= (1 << 0),
3928 nir_lower_indirect_array_deref_of_vec_load
= (1 << 1),
3929 nir_lower_direct_array_deref_of_vec_store
= (1 << 2),
3930 nir_lower_indirect_array_deref_of_vec_store
= (1 << 3),
3931 } nir_lower_array_deref_of_vec_options
;
3933 bool nir_lower_array_deref_of_vec(nir_shader
*shader
, nir_variable_mode modes
,
3934 nir_lower_array_deref_of_vec_options options
);
3936 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
3938 bool nir_lower_locals_to_regs(nir_shader
*shader
);
3940 void nir_lower_io_to_temporaries(nir_shader
*shader
,
3941 nir_function_impl
*entrypoint
,
3942 bool outputs
, bool inputs
);
3944 bool nir_lower_vars_to_scratch(nir_shader
*shader
,
3945 nir_variable_mode modes
,
3947 glsl_type_size_align_func size_align
);
3949 void nir_lower_clip_halfz(nir_shader
*shader
);
3951 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
3953 void nir_gather_ssa_types(nir_function_impl
*impl
,
3954 BITSET_WORD
*float_types
,
3955 BITSET_WORD
*int_types
);
3957 void nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
3958 int (*type_size
)(const struct glsl_type
*, bool));
3960 /* Some helpers to do very simple linking */
3961 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3962 bool nir_remove_unused_io_vars(nir_shader
*shader
, nir_variable_mode mode
,
3963 uint64_t *used_by_other_stage
,
3964 uint64_t *used_by_other_stage_patches
);
3965 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
3966 bool default_to_smooth_interp
);
3967 void nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3968 bool nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3970 bool nir_lower_amul(nir_shader
*shader
,
3971 int (*type_size
)(const struct glsl_type
*, bool));
3973 void nir_assign_io_var_locations(struct exec_list
*var_list
,
3975 gl_shader_stage stage
);
3978 uint8_t num_linked_io_vars
;
3979 uint8_t num_linked_patch_io_vars
;
3980 } nir_linked_io_var_info
;
3982 nir_linked_io_var_info
3983 nir_assign_linked_io_var_locations(nir_shader
*producer
,
3984 nir_shader
*consumer
);
3987 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
3988 * to 32-bit operations. This is only valid for nir_var_shader_in/out
3991 nir_lower_io_lower_64bit_to_32
= (1 << 0),
3993 /* If set, this forces all non-flat fragment shader inputs to be
3994 * interpolated as if with the "sample" qualifier. This requires
3995 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
3997 nir_lower_io_force_sample_interpolation
= (1 << 1),
3998 } nir_lower_io_options
;
3999 bool nir_lower_io(nir_shader
*shader
,
4000 nir_variable_mode modes
,
4001 int (*type_size
)(const struct glsl_type
*, bool),
4002 nir_lower_io_options
);
4004 bool nir_io_add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
);
4007 nir_lower_vars_to_explicit_types(nir_shader
*shader
,
4008 nir_variable_mode modes
,
4009 glsl_type_size_align_func type_info
);
4013 * An address format which is a simple 32-bit global GPU address.
4015 nir_address_format_32bit_global
,
4018 * An address format which is a simple 64-bit global GPU address.
4020 nir_address_format_64bit_global
,
4023 * An address format which is a bounds-checked 64-bit global GPU address.
4025 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
4026 * address stored with the low bits in .x and high bits in .y, .z is a
4027 * size, and .w is an offset. When the final I/O operation is lowered, .w
4028 * is checked against .z and the operation is predicated on the result.
4030 nir_address_format_64bit_bounded_global
,
4033 * An address format which is comprised of a vec2 where the first
4034 * component is a buffer index and the second is an offset.
4036 nir_address_format_32bit_index_offset
,
4039 * An address format which is comprised of a vec3 where the first two
4040 * components specify the buffer and the third is an offset.
4042 nir_address_format_vec2_index_32bit_offset
,
4045 * An address format which is a simple 32-bit offset.
4047 nir_address_format_32bit_offset
,
4050 * An address format representing a purely logical addressing model. In
4051 * this model, all deref chains must be complete from the dereference
4052 * operation to the variable. Cast derefs are not allowed. These
4053 * addresses will be 32-bit scalars but the format is immaterial because
4054 * you can always chase the chain.
4056 nir_address_format_logical
,
4057 } nir_address_format
;
4059 static inline unsigned
4060 nir_address_format_bit_size(nir_address_format addr_format
)
4062 switch (addr_format
) {
4063 case nir_address_format_32bit_global
: return 32;
4064 case nir_address_format_64bit_global
: return 64;
4065 case nir_address_format_64bit_bounded_global
: return 32;
4066 case nir_address_format_32bit_index_offset
: return 32;
4067 case nir_address_format_vec2_index_32bit_offset
: return 32;
4068 case nir_address_format_32bit_offset
: return 32;
4069 case nir_address_format_logical
: return 32;
4071 unreachable("Invalid address format");
4074 static inline unsigned
4075 nir_address_format_num_components(nir_address_format addr_format
)
4077 switch (addr_format
) {
4078 case nir_address_format_32bit_global
: return 1;
4079 case nir_address_format_64bit_global
: return 1;
4080 case nir_address_format_64bit_bounded_global
: return 4;
4081 case nir_address_format_32bit_index_offset
: return 2;
4082 case nir_address_format_vec2_index_32bit_offset
: return 3;
4083 case nir_address_format_32bit_offset
: return 1;
4084 case nir_address_format_logical
: return 1;
4086 unreachable("Invalid address format");
4089 static inline const struct glsl_type
*
4090 nir_address_format_to_glsl_type(nir_address_format addr_format
)
4092 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
4093 assert(bit_size
== 32 || bit_size
== 64);
4094 return glsl_vector_type(bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
,
4095 nir_address_format_num_components(addr_format
));
4098 const nir_const_value
*nir_address_format_null_value(nir_address_format addr_format
);
4100 nir_ssa_def
*nir_build_addr_ieq(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
4101 nir_address_format addr_format
);
4103 nir_ssa_def
*nir_build_addr_isub(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
4104 nir_address_format addr_format
);
4106 nir_ssa_def
* nir_explicit_io_address_from_deref(struct nir_builder
*b
,
4107 nir_deref_instr
*deref
,
4108 nir_ssa_def
*base_addr
,
4109 nir_address_format addr_format
);
4110 void nir_lower_explicit_io_instr(struct nir_builder
*b
,
4111 nir_intrinsic_instr
*io_instr
,
4113 nir_address_format addr_format
);
4115 bool nir_lower_explicit_io(nir_shader
*shader
,
4116 nir_variable_mode modes
,
4117 nir_address_format
);
4119 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
4120 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
4122 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
4124 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
4125 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
4126 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
4128 bool nir_remove_dead_derefs(nir_shader
*shader
);
4129 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
4130 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
,
4131 bool (*can_remove_var
)(nir_variable
*var
));
4132 bool nir_lower_variable_initializers(nir_shader
*shader
,
4133 nir_variable_mode modes
);
4135 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
4136 bool nir_lower_vec_to_movs(nir_shader
*shader
);
4137 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
4139 const gl_state_index16
*alpha_ref_state_tokens
);
4140 bool nir_lower_alu(nir_shader
*shader
);
4142 bool nir_lower_flrp(nir_shader
*shader
, unsigned lowering_mask
,
4143 bool always_precise
, bool have_ffma
);
4145 bool nir_lower_alu_to_scalar(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
4146 bool nir_lower_bool_to_bitsize(nir_shader
*shader
);
4147 bool nir_lower_bool_to_float(nir_shader
*shader
);
4148 bool nir_lower_bool_to_int32(nir_shader
*shader
);
4149 bool nir_lower_int_to_float(nir_shader
*shader
);
4150 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
4151 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
4152 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
4153 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
4154 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
4156 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
4157 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
4158 bool nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode mask
);
4160 bool nir_lower_fragcolor(nir_shader
*shader
);
4161 void nir_lower_fragcoord_wtrans(nir_shader
*shader
);
4162 void nir_lower_viewport_transform(nir_shader
*shader
);
4163 bool nir_lower_uniforms_to_ubo(nir_shader
*shader
, int multiplier
);
4165 typedef struct nir_lower_subgroups_options
{
4166 uint8_t subgroup_size
;
4167 uint8_t ballot_bit_size
;
4168 bool lower_to_scalar
:1;
4169 bool lower_vote_trivial
:1;
4170 bool lower_vote_eq_to_ballot
:1;
4171 bool lower_subgroup_masks
:1;
4172 bool lower_shuffle
:1;
4173 bool lower_shuffle_to_32bit
:1;
4174 bool lower_shuffle_to_swizzle_amd
:1;
4176 bool lower_quad_broadcast_dynamic
:1;
4177 bool lower_quad_broadcast_dynamic_to_const
:1;
4178 } nir_lower_subgroups_options
;
4180 bool nir_lower_subgroups(nir_shader
*shader
,
4181 const nir_lower_subgroups_options
*options
);
4183 bool nir_lower_system_values(nir_shader
*shader
);
4185 enum PACKED nir_lower_tex_packing
{
4186 nir_lower_tex_packing_none
= 0,
4187 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
4188 * or unsigned ints based on the sampler type
4190 nir_lower_tex_packing_16
,
4191 /* The sampler returns 1 32-bit word of 4x8 unorm */
4192 nir_lower_tex_packing_8
,
4195 typedef struct nir_lower_tex_options
{
4197 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
4198 * sampler types a texture projector is lowered.
4203 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
4205 bool lower_txf_offset
;
4208 * If true, lower away nir_tex_src_offset for all rect textures.
4210 bool lower_rect_offset
;
4213 * If true, lower rect textures to 2D, using txs to fetch the
4214 * texture dimensions and dividing the texture coords by the
4215 * texture dims to normalize.
4220 * If true, convert yuv to rgb.
4222 unsigned lower_y_uv_external
;
4223 unsigned lower_y_u_v_external
;
4224 unsigned lower_yx_xuxv_external
;
4225 unsigned lower_xy_uxvx_external
;
4226 unsigned lower_ayuv_external
;
4227 unsigned lower_xyuv_external
;
4230 * To emulate certain texture wrap modes, this can be used
4231 * to saturate the specified tex coord to [0.0, 1.0]. The
4232 * bits are according to sampler #, ie. if, for example:
4234 * (conf->saturate_s & (1 << n))
4236 * is true, then the s coord for sampler n is saturated.
4238 * Note that clamping must happen *after* projector lowering
4239 * so any projected texture sample instruction with a clamped
4240 * coordinate gets automatically lowered, regardless of the
4241 * 'lower_txp' setting.
4243 unsigned saturate_s
;
4244 unsigned saturate_t
;
4245 unsigned saturate_r
;
4247 /* Bitmask of textures that need swizzling.
4249 * If (swizzle_result & (1 << texture_index)), then the swizzle in
4250 * swizzles[texture_index] is applied to the result of the texturing
4253 unsigned swizzle_result
;
4255 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
4256 * while 4 and 5 represent 0 and 1 respectively.
4258 uint8_t swizzles
[32][4];
4260 /* Can be used to scale sampled values in range required by the format. */
4261 float scale_factors
[32];
4264 * Bitmap of textures that need srgb to linear conversion. If
4265 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
4266 * of the texture are lowered to linear.
4268 unsigned lower_srgb
;
4271 * If true, lower nir_texop_tex on shaders that doesn't support implicit
4272 * LODs to nir_texop_txl.
4274 bool lower_tex_without_implicit_lod
;
4277 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
4279 bool lower_txd_cube_map
;
4282 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
4287 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
4288 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
4289 * with lower_txd_cube_map.
4291 bool lower_txd_shadow
;
4294 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
4295 * Implies lower_txd_cube_map and lower_txd_shadow.
4300 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
4301 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
4303 bool lower_txb_shadow_clamp
;
4306 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
4307 * with nir_texop_txl. This includes cube maps.
4309 bool lower_txd_shadow_clamp
;
4312 * If true, lower nir_texop_txd on when it uses both offset and min_lod
4313 * with nir_texop_txl. This includes cube maps.
4315 bool lower_txd_offset_clamp
;
4318 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4319 * sampler is bindless.
4321 bool lower_txd_clamp_bindless_sampler
;
4324 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4325 * sampler index is not statically determinable to be less than 16.
4327 bool lower_txd_clamp_if_sampler_index_not_lt_16
;
4330 * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
4331 * 0-lod followed by a nir_ishr.
4336 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
4337 * mixed-up tg4 locations.
4339 bool lower_tg4_broadcom_swizzle
;
4342 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
4344 bool lower_tg4_offsets
;
4346 enum nir_lower_tex_packing lower_tex_packing
[32];
4347 } nir_lower_tex_options
;
4349 bool nir_lower_tex(nir_shader
*shader
,
4350 const nir_lower_tex_options
*options
);
4352 enum nir_lower_non_uniform_access_type
{
4353 nir_lower_non_uniform_ubo_access
= (1 << 0),
4354 nir_lower_non_uniform_ssbo_access
= (1 << 1),
4355 nir_lower_non_uniform_texture_access
= (1 << 2),
4356 nir_lower_non_uniform_image_access
= (1 << 3),
4359 bool nir_lower_non_uniform_access(nir_shader
*shader
,
4360 enum nir_lower_non_uniform_access_type
);
4362 enum nir_lower_idiv_path
{
4363 /* This path is based on NV50LegalizeSSA::handleDIV(). It is the faster of
4364 * the two but it is not exact in some cases (for example, 1091317713u /
4365 * 1034u gives 5209173 instead of 1055432) */
4366 nir_lower_idiv_fast
,
4367 /* This path is based on AMDGPUTargetLowering::LowerUDIVREM() and
4368 * AMDGPUTargetLowering::LowerSDIVREM(). It requires more instructions than
4369 * the nv50 path and many of them are integer multiplications, so it is
4370 * probably slower. It should always return the correct result, though. */
4371 nir_lower_idiv_precise
,
4374 bool nir_lower_idiv(nir_shader
*shader
, enum nir_lower_idiv_path path
);
4376 bool nir_lower_input_attachments(nir_shader
*shader
, bool use_fragcoord_sysval
);
4378 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
,
4380 bool use_clipdist_array
,
4381 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4382 bool nir_lower_clip_gs(nir_shader
*shader
, unsigned ucp_enables
,
4383 bool use_clipdist_array
,
4384 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4385 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
,
4386 bool use_clipdist_array
);
4387 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
4388 bool nir_lower_clip_disable(nir_shader
*shader
, unsigned clip_plane_enable
);
4390 void nir_lower_point_size_mov(nir_shader
*shader
,
4391 const gl_state_index16
*pointsize_state_tokens
);
4393 bool nir_lower_frexp(nir_shader
*nir
);
4395 void nir_lower_two_sided_color(nir_shader
*shader
, bool face_sysval
);
4397 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
4399 bool nir_lower_flatshade(nir_shader
*shader
);
4401 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
4402 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
4403 const gl_state_index16
*uniform_state_tokens
);
4405 typedef struct nir_lower_wpos_ytransform_options
{
4406 gl_state_index16 state_tokens
[STATE_LENGTH
];
4407 bool fs_coord_origin_upper_left
:1;
4408 bool fs_coord_origin_lower_left
:1;
4409 bool fs_coord_pixel_center_integer
:1;
4410 bool fs_coord_pixel_center_half_integer
:1;
4411 } nir_lower_wpos_ytransform_options
;
4413 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
4414 const nir_lower_wpos_ytransform_options
*options
);
4415 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
4417 bool nir_lower_wrmasks(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
4419 bool nir_lower_fb_read(nir_shader
*shader
);
4421 typedef struct nir_lower_drawpixels_options
{
4422 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
4423 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
4424 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
4425 unsigned drawpix_sampler
;
4426 unsigned pixelmap_sampler
;
4428 bool scale_and_bias
:1;
4429 } nir_lower_drawpixels_options
;
4431 void nir_lower_drawpixels(nir_shader
*shader
,
4432 const nir_lower_drawpixels_options
*options
);
4434 typedef struct nir_lower_bitmap_options
{
4437 } nir_lower_bitmap_options
;
4439 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
4441 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
);
4444 nir_lower_int_source_mods
= 1 << 0,
4445 nir_lower_float_source_mods
= 1 << 1,
4446 nir_lower_triop_abs
= 1 << 2,
4447 nir_lower_all_source_mods
= (1 << 3) - 1
4448 } nir_lower_to_source_mods_flags
;
4451 bool nir_lower_to_source_mods(nir_shader
*shader
, nir_lower_to_source_mods_flags options
);
4453 bool nir_lower_gs_intrinsics(nir_shader
*shader
, bool per_stream
);
4455 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
4457 bool nir_lower_bit_size(nir_shader
*shader
,
4458 nir_lower_bit_size_callback callback
,
4459 void *callback_data
);
4461 nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode
);
4462 bool nir_lower_int64(nir_shader
*shader
, nir_lower_int64_options options
);
4464 nir_lower_doubles_options
nir_lower_doubles_op_to_options_mask(nir_op opcode
);
4465 bool nir_lower_doubles(nir_shader
*shader
, const nir_shader
*softfp64
,
4466 nir_lower_doubles_options options
);
4467 bool nir_lower_pack(nir_shader
*shader
);
4469 void nir_lower_mediump_outputs(nir_shader
*nir
);
4471 bool nir_lower_point_size(nir_shader
*shader
, float min
, float max
);
4474 nir_lower_interpolation_at_sample
= (1 << 1),
4475 nir_lower_interpolation_at_offset
= (1 << 2),
4476 nir_lower_interpolation_centroid
= (1 << 3),
4477 nir_lower_interpolation_pixel
= (1 << 4),
4478 nir_lower_interpolation_sample
= (1 << 5),
4479 } nir_lower_interpolation_options
;
4481 bool nir_lower_interpolation(nir_shader
*shader
,
4482 nir_lower_interpolation_options options
);
4484 bool nir_lower_discard_to_demote(nir_shader
*shader
);
4486 bool nir_lower_memory_model(nir_shader
*shader
);
4488 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
4490 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
4492 void nir_loop_analyze_impl(nir_function_impl
*impl
,
4493 nir_variable_mode indirect_mask
);
4495 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
4497 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
4498 bool nir_repair_ssa(nir_shader
*shader
);
4500 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
4501 bool nir_convert_to_lcssa(nir_shader
*shader
, bool skip_invariants
, bool skip_bool_invariants
);
4502 void nir_divergence_analysis(nir_shader
*shader
, nir_divergence_options options
);
4504 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
4505 * registers. If false, convert all values (even those not involved in a phi
4506 * node) to registers.
4508 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
4510 bool nir_lower_phis_to_regs_block(nir_block
*block
);
4511 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
4512 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
);
4514 bool nir_lower_samplers(nir_shader
*shader
);
4515 bool nir_lower_ssbo(nir_shader
*shader
);
4517 /* This is here for unit tests. */
4518 bool nir_opt_comparison_pre_impl(nir_function_impl
*impl
);
4520 bool nir_opt_comparison_pre(nir_shader
*shader
);
4522 bool nir_opt_access(nir_shader
*shader
);
4523 bool nir_opt_algebraic(nir_shader
*shader
);
4524 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
4525 bool nir_opt_algebraic_late(nir_shader
*shader
);
4526 bool nir_opt_algebraic_distribute_src_mods(nir_shader
*shader
);
4527 bool nir_opt_constant_folding(nir_shader
*shader
);
4529 /* Try to combine a and b into a. Return true if combination was possible,
4530 * which will result in b being removed by the pass. Return false if
4531 * combination wasn't possible.
4533 typedef bool (*nir_combine_memory_barrier_cb
)(
4534 nir_intrinsic_instr
*a
, nir_intrinsic_instr
*b
, void *data
);
4536 bool nir_opt_combine_memory_barriers(nir_shader
*shader
,
4537 nir_combine_memory_barrier_cb combine_cb
,
4540 bool nir_opt_combine_stores(nir_shader
*shader
, nir_variable_mode modes
);
4542 bool nir_copy_prop(nir_shader
*shader
);
4544 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
4546 bool nir_opt_cse(nir_shader
*shader
);
4548 bool nir_opt_dce(nir_shader
*shader
);
4550 bool nir_opt_dead_cf(nir_shader
*shader
);
4552 bool nir_opt_dead_write_vars(nir_shader
*shader
);
4554 bool nir_opt_deref_impl(nir_function_impl
*impl
);
4555 bool nir_opt_deref(nir_shader
*shader
);
4557 bool nir_opt_find_array_copies(nir_shader
*shader
);
4559 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
4561 bool nir_opt_idiv_const(nir_shader
*shader
, unsigned min_bit_size
);
4563 bool nir_opt_if(nir_shader
*shader
, bool aggressive_last_continue
);
4565 bool nir_opt_intrinsics(nir_shader
*shader
);
4567 bool nir_opt_large_constants(nir_shader
*shader
,
4568 glsl_type_size_align_func size_align
,
4569 unsigned threshold
);
4571 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
4574 nir_move_const_undef
= (1 << 0),
4575 nir_move_load_ubo
= (1 << 1),
4576 nir_move_load_input
= (1 << 2),
4577 nir_move_comparisons
= (1 << 3),
4578 nir_move_copies
= (1 << 4),
4581 bool nir_can_move_instr(nir_instr
*instr
, nir_move_options options
);
4583 bool nir_opt_sink(nir_shader
*shader
, nir_move_options options
);
4585 bool nir_opt_move(nir_shader
*shader
, nir_move_options options
);
4587 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
,
4588 bool indirect_load_ok
, bool expensive_alu_ok
);
4590 bool nir_opt_rematerialize_compares(nir_shader
*shader
);
4592 bool nir_opt_remove_phis(nir_shader
*shader
);
4593 bool nir_opt_remove_phis_block(nir_block
*block
);
4595 bool nir_opt_shrink_load(nir_shader
*shader
);
4597 bool nir_opt_trivial_continues(nir_shader
*shader
);
4599 bool nir_opt_undef(nir_shader
*shader
);
4601 bool nir_opt_vectorize(nir_shader
*shader
);
4603 bool nir_opt_conditional_discard(nir_shader
*shader
);
4605 typedef bool (*nir_should_vectorize_mem_func
)(unsigned align
, unsigned bit_size
,
4606 unsigned num_components
, unsigned high_offset
,
4607 nir_intrinsic_instr
*low
, nir_intrinsic_instr
*high
);
4609 bool nir_opt_load_store_vectorize(nir_shader
*shader
, nir_variable_mode modes
,
4610 nir_should_vectorize_mem_func callback
,
4611 nir_variable_mode robust_modes
);
4613 void nir_strip(nir_shader
*shader
);
4615 void nir_sweep(nir_shader
*shader
);
4617 void nir_remap_dual_slot_attributes(nir_shader
*shader
,
4618 uint64_t *dual_slot_inputs
);
4619 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
);
4621 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
4622 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);
4625 nir_variable_is_in_ubo(const nir_variable
*var
)
4627 return (var
->data
.mode
== nir_var_mem_ubo
&&
4628 var
->interface_type
!= NULL
);
4632 nir_variable_is_in_ssbo(const nir_variable
*var
)
4634 return (var
->data
.mode
== nir_var_mem_ssbo
&&
4635 var
->interface_type
!= NULL
);
4639 nir_variable_is_in_block(const nir_variable
*var
)
4641 return nir_variable_is_in_ubo(var
) || nir_variable_is_in_ssbo(var
);
4644 typedef struct nir_unsigned_upper_bound_config
{
4645 unsigned min_subgroup_size
;
4646 unsigned max_subgroup_size
;
4647 unsigned max_work_group_invocations
;
4648 unsigned max_work_group_count
[3];
4649 unsigned max_work_group_size
[3];
4651 uint32_t vertex_attrib_max
[32];
4652 } nir_unsigned_upper_bound_config
;
4655 nir_unsigned_upper_bound(nir_shader
*shader
, struct hash_table
*range_ht
,
4656 nir_ssa_scalar scalar
,
4657 const nir_unsigned_upper_bound_config
*config
);
4660 nir_addition_might_overflow(nir_shader
*shader
, struct hash_table
*range_ht
,
4661 nir_ssa_scalar ssa
, unsigned const_val
,
4662 const nir_unsigned_upper_bound_config
*config
);