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"
47 #include "util/debug.h"
50 #include "nir_opcodes.h"
52 #if defined(_WIN32) && !defined(snprintf)
53 #define snprintf _snprintf
61 #define NIR_TRUE (~0u)
62 #define NIR_MAX_VEC_COMPONENTS 16
63 #define NIR_MAX_MATRIX_COLUMNS 4
64 #define NIR_STREAM_PACKED (1 << 8)
65 typedef uint16_t nir_component_mask_t
;
68 nir_num_components_valid(unsigned num_components
)
70 return (num_components
>= 1 &&
71 num_components
<= 4) ||
72 num_components
== 8 ||
76 /** Defines a cast function
78 * This macro defines a cast function from in_type to out_type where
79 * out_type is some structure type that contains a field of type out_type.
81 * Note that you have to be a bit careful as the generated cast function
84 #define NIR_DEFINE_CAST(name, in_type, out_type, field, \
85 type_field, type_value) \
86 static inline out_type * \
87 name(const in_type *parent) \
89 assert(parent && parent->type_field == type_value); \
90 return exec_node_data(out_type, parent, field); \
100 * Description of built-in state associated with a uniform
102 * \sa nir_variable::state_slots
105 gl_state_index16 tokens
[STATE_LENGTH
];
110 nir_var_shader_in
= (1 << 0),
111 nir_var_shader_out
= (1 << 1),
112 nir_var_shader_temp
= (1 << 2),
113 nir_var_function_temp
= (1 << 3),
114 nir_var_uniform
= (1 << 4),
115 nir_var_mem_ubo
= (1 << 5),
116 nir_var_system_value
= (1 << 6),
117 nir_var_mem_ssbo
= (1 << 7),
118 nir_var_mem_shared
= (1 << 8),
119 nir_var_mem_global
= (1 << 9),
120 nir_var_mem_push_const
= (1 << 10), /* not actually used for variables */
121 nir_num_variable_modes
= 11,
122 nir_var_all
= (1 << nir_num_variable_modes
) - 1,
129 nir_rounding_mode_undef
= 0,
130 nir_rounding_mode_rtne
= 1, /* round to nearest even */
131 nir_rounding_mode_ru
= 2, /* round up */
132 nir_rounding_mode_rd
= 3, /* round down */
133 nir_rounding_mode_rtz
= 4, /* round towards zero */
150 #define nir_const_value_to_array(arr, c, components, m) \
152 for (unsigned i = 0; i < components; ++i) \
156 static inline nir_const_value
157 nir_const_value_for_raw_uint(uint64_t x
, unsigned bit_size
)
160 memset(&v
, 0, sizeof(v
));
163 case 1: v
.b
= x
; break;
164 case 8: v
.u8
= x
; break;
165 case 16: v
.u16
= x
; break;
166 case 32: v
.u32
= x
; break;
167 case 64: v
.u64
= x
; break;
169 unreachable("Invalid bit size");
175 static inline nir_const_value
176 nir_const_value_for_int(int64_t i
, unsigned bit_size
)
179 memset(&v
, 0, sizeof(v
));
181 assert(bit_size
<= 64);
183 assert(i
>= (-(1ll << (bit_size
- 1))));
184 assert(i
< (1ll << (bit_size
- 1)));
187 return nir_const_value_for_raw_uint(i
, bit_size
);
190 static inline nir_const_value
191 nir_const_value_for_uint(uint64_t u
, unsigned bit_size
)
194 memset(&v
, 0, sizeof(v
));
196 assert(bit_size
<= 64);
198 assert(u
< (1ull << bit_size
));
200 return nir_const_value_for_raw_uint(u
, bit_size
);
203 static inline nir_const_value
204 nir_const_value_for_bool(bool b
, unsigned bit_size
)
206 /* Booleans use a 0/-1 convention */
207 return nir_const_value_for_int(-(int)b
, bit_size
);
210 /* This one isn't inline because it requires half-float conversion */
211 nir_const_value
nir_const_value_for_float(double b
, unsigned bit_size
);
213 static inline int64_t
214 nir_const_value_as_int(nir_const_value value
, unsigned bit_size
)
217 /* int1_t uses 0/-1 convention */
218 case 1: return -(int)value
.b
;
219 case 8: return value
.i8
;
220 case 16: return value
.i16
;
221 case 32: return value
.i32
;
222 case 64: return value
.i64
;
224 unreachable("Invalid bit size");
228 static inline uint64_t
229 nir_const_value_as_uint(nir_const_value value
, unsigned bit_size
)
232 case 1: return value
.b
;
233 case 8: return value
.u8
;
234 case 16: return value
.u16
;
235 case 32: return value
.u32
;
236 case 64: return value
.u64
;
238 unreachable("Invalid bit size");
243 nir_const_value_as_bool(nir_const_value value
, unsigned bit_size
)
245 int64_t i
= nir_const_value_as_int(value
, bit_size
);
247 /* Booleans of any size use 0/-1 convention */
248 assert(i
== 0 || i
== -1);
253 /* This one isn't inline because it requires half-float conversion */
254 double nir_const_value_as_float(nir_const_value value
, unsigned bit_size
);
256 typedef struct nir_constant
{
258 * Value of the constant.
260 * The field used to back the values supplied by the constant is determined
261 * by the type associated with the \c nir_variable. Constants may be
262 * scalars, vectors, or matrices.
264 nir_const_value values
[NIR_MAX_VEC_COMPONENTS
];
266 /* we could get this from the var->type but makes clone *much* easier to
267 * not have to care about the type.
269 unsigned num_elements
;
271 /* Array elements / Structure Fields */
272 struct nir_constant
**elements
;
276 * \brief Layout qualifiers for gl_FragDepth.
278 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
279 * with a layout qualifier.
282 nir_depth_layout_none
, /**< No depth layout is specified. */
283 nir_depth_layout_any
,
284 nir_depth_layout_greater
,
285 nir_depth_layout_less
,
286 nir_depth_layout_unchanged
290 * Enum keeping track of how a variable was declared.
294 * Normal declaration.
296 nir_var_declared_normally
= 0,
299 * Variable is implicitly generated by the compiler and should not be
300 * visible via the API.
303 } nir_var_declaration_type
;
306 * Either a uniform, global variable, shader input, or shader output. Based on
307 * ir_variable - it should be easy to translate between the two.
310 typedef struct nir_variable
{
311 struct exec_node node
;
314 * Declared type of the variable
316 const struct glsl_type
*type
;
319 * Declared name of the variable
323 struct nir_variable_data
{
325 * Storage class of the variable.
327 * \sa nir_variable_mode
329 nir_variable_mode mode
:11;
332 * Is the variable read-only?
334 * This is set for variables declared as \c const, shader inputs,
337 unsigned read_only
:1;
341 unsigned invariant
:1;
344 * Precision qualifier.
346 * In desktop GLSL we do not care about precision qualifiers at all, in
347 * fact, the spec says that precision qualifiers are ignored.
349 * To make things easy, we make it so that this field is always
350 * GLSL_PRECISION_NONE on desktop shaders. This way all the variables
351 * have the same precision value and the checks we add in the compiler
352 * for this field will never break a desktop shader compile.
354 unsigned precision
:2;
357 * Can this variable be coalesced with another?
359 * This is set by nir_lower_io_to_temporaries to say that any
360 * copies involving this variable should stay put. Propagating it can
361 * duplicate the resulting load/store, which is not wanted, and may
362 * result in a load/store of the variable with an indirect offset which
363 * the backend may not be able to handle.
365 unsigned cannot_coalesce
:1;
368 * When separate shader programs are enabled, only input/outputs between
369 * the stages of a multi-stage separate program can be safely removed
370 * from the shader interface. Other input/outputs must remains active.
372 * This is also used to make sure xfb varyings that are unused by the
373 * fragment shader are not removed.
375 unsigned always_active_io
:1;
378 * Interpolation mode for shader inputs / outputs
380 * \sa glsl_interp_mode
382 unsigned interpolation
:3;
385 * If non-zero, then this variable may be packed along with other variables
386 * into a single varying slot, so this offset should be applied when
387 * accessing components. For example, an offset of 1 means that the x
388 * component of this variable is actually stored in component y of the
389 * location specified by \c location.
391 unsigned location_frac
:2;
394 * If true, this variable represents an array of scalars that should
395 * be tightly packed. In other words, consecutive array elements
396 * should be stored one component apart, rather than one slot apart.
401 * Whether this is a fragment shader output implicitly initialized with
402 * the previous contents of the specified render target at the
403 * framebuffer location corresponding to this shader invocation.
405 unsigned fb_fetch_output
:1;
408 * Non-zero if this variable is considered bindless as defined by
409 * ARB_bindless_texture.
414 * Was an explicit binding set in the shader?
416 unsigned explicit_binding
:1;
419 * Was the location explicitly set in the shader?
421 * If the location is explicitly set in the shader, it \b cannot be changed
422 * by the linker or by the API (e.g., calls to \c glBindAttribLocation have
425 unsigned explicit_location
:1;
428 * Was a transfer feedback buffer set in the shader?
430 unsigned explicit_xfb_buffer
:1;
433 * Was a transfer feedback stride set in the shader?
435 unsigned explicit_xfb_stride
:1;
438 * Was an explicit offset set in the shader?
440 unsigned explicit_offset
:1;
443 * Non-zero if this variable was created by lowering a named interface
446 unsigned from_named_ifc_block
:1;
449 * How the variable was declared. See nir_var_declaration_type.
451 * This is used to detect variables generated by the compiler, so should
452 * not be visible via the API.
454 unsigned how_declared
:2;
457 * Is this variable per-view? If so, we know it must be an array with
458 * size corresponding to the number of views.
463 * \brief Layout qualifier for gl_FragDepth.
465 * This is not equal to \c ir_depth_layout_none if and only if this
466 * variable is \c gl_FragDepth and a layout qualifier is specified.
468 nir_depth_layout depth_layout
:3;
471 * Vertex stream output identifier.
473 * For packed outputs, NIR_STREAM_PACKED is set and bits [2*i+1,2*i]
474 * indicate the stream of the i-th component.
479 * Access flags for memory variables (SSBO/global), image uniforms, and
480 * bindless images in uniforms/inputs/outputs.
482 enum gl_access_qualifier access
:8;
485 * Descriptor set binding for sampler or UBO.
487 unsigned descriptor_set
:5;
490 * output index for dual source blending.
495 * Initial binding point for a sampler or UBO.
497 * For array types, this represents the binding point for the first element.
502 * Storage location of the base of this variable
504 * The precise meaning of this field depends on the nature of the variable.
506 * - Vertex shader input: one of the values from \c gl_vert_attrib.
507 * - Vertex shader output: one of the values from \c gl_varying_slot.
508 * - Geometry shader input: one of the values from \c gl_varying_slot.
509 * - Geometry shader output: one of the values from \c gl_varying_slot.
510 * - Fragment shader input: one of the values from \c gl_varying_slot.
511 * - Fragment shader output: one of the values from \c gl_frag_result.
512 * - Uniforms: Per-stage uniform slot number for default uniform block.
513 * - Uniforms: Index within the uniform block definition for UBO members.
514 * - Non-UBO Uniforms: uniform slot number.
515 * - Other: This field is not currently used.
517 * If the variable is a uniform, shader input, or shader output, and the
518 * slot has not been assigned, the value will be -1.
523 * The actual location of the variable in the IR. Only valid for inputs,
524 * outputs, and uniforms (including samplers and images).
526 unsigned driver_location
;
529 * Location an atomic counter or transform feedback is stored at.
535 /** Image internal format if specified explicitly, otherwise PIPE_FORMAT_NONE. */
536 enum pipe_format format
;
541 * Transform feedback buffer.
546 * Transform feedback stride.
554 * Identifier for this variable generated by nir_index_vars() that is unique
555 * among other variables in the same exec_list.
559 /* Number of nir_variable_data members */
560 uint16_t num_members
;
563 * Built-in state that backs this uniform
565 * Once set at variable creation, \c state_slots must remain invariant.
566 * This is because, ideally, this array would be shared by all clones of
567 * this variable in the IR tree. In other words, we'd really like for it
568 * to be a fly-weight.
570 * If the variable is not a uniform, \c num_state_slots will be zero and
571 * \c state_slots will be \c NULL.
574 uint16_t num_state_slots
; /**< Number of state slots used */
575 nir_state_slot
*state_slots
; /**< State descriptors. */
579 * Constant expression assigned in the initializer of the variable
581 * This field should only be used temporarily by creators of NIR shaders
582 * and then lower_constant_initializers can be used to get rid of them.
583 * Most of the rest of NIR ignores this field or asserts that it's NULL.
585 nir_constant
*constant_initializer
;
588 * Global variable assigned in the initializer of the variable
589 * This field should only be used temporarily by creators of NIR shaders
590 * and then lower_constant_initializers can be used to get rid of them.
591 * Most of the rest of NIR ignores this field or asserts that it's NULL.
593 struct nir_variable
*pointer_initializer
;
596 * For variables that are in an interface block or are an instance of an
597 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
599 * \sa ir_variable::location
601 const struct glsl_type
*interface_type
;
604 * Description of per-member data for per-member struct variables
606 * This is used for variables which are actually an amalgamation of
607 * multiple entities such as a struct of built-in values or a struct of
608 * inputs each with their own layout specifier. This is only allowed on
609 * variables with a struct or array of array of struct type.
611 struct nir_variable_data
*members
;
614 #define nir_foreach_variable(var, var_list) \
615 foreach_list_typed(nir_variable, var, node, var_list)
617 #define nir_foreach_variable_safe(var, var_list) \
618 foreach_list_typed_safe(nir_variable, var, node, var_list)
621 nir_variable_is_global(const nir_variable
*var
)
623 return var
->data
.mode
!= nir_var_function_temp
;
626 typedef struct nir_register
{
627 struct exec_node node
;
629 unsigned num_components
; /** < number of vector components */
630 unsigned num_array_elems
; /** < size of array (0 for no array) */
632 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
635 /** generic register index. */
638 /** only for debug purposes, can be NULL */
641 /** set of nir_srcs where this register is used (read from) */
642 struct list_head uses
;
644 /** set of nir_dests where this register is defined (written to) */
645 struct list_head defs
;
647 /** set of nir_ifs where this register is used as a condition */
648 struct list_head if_uses
;
651 #define nir_foreach_register(reg, reg_list) \
652 foreach_list_typed(nir_register, reg, node, reg_list)
653 #define nir_foreach_register_safe(reg, reg_list) \
654 foreach_list_typed_safe(nir_register, reg, node, reg_list)
656 typedef enum PACKED
{
658 nir_instr_type_deref
,
661 nir_instr_type_intrinsic
,
662 nir_instr_type_load_const
,
664 nir_instr_type_ssa_undef
,
666 nir_instr_type_parallel_copy
,
669 typedef struct nir_instr
{
670 struct exec_node node
;
671 struct nir_block
*block
;
674 /* A temporary for optimization and analysis passes to use for storing
675 * flags. For instance, DCE uses this to store the "dead/live" info.
679 /** generic instruction index. */
683 static inline nir_instr
*
684 nir_instr_next(nir_instr
*instr
)
686 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
687 if (exec_node_is_tail_sentinel(next
))
690 return exec_node_data(nir_instr
, next
, node
);
693 static inline nir_instr
*
694 nir_instr_prev(nir_instr
*instr
)
696 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
697 if (exec_node_is_head_sentinel(prev
))
700 return exec_node_data(nir_instr
, prev
, node
);
704 nir_instr_is_first(const nir_instr
*instr
)
706 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr
->node
));
710 nir_instr_is_last(const nir_instr
*instr
)
712 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr
->node
));
715 typedef struct nir_ssa_def
{
716 /** for debugging only, can be NULL */
719 /** generic SSA definition index. */
722 /** Index into the live_in and live_out bitfields */
725 /** Instruction which produces this SSA value. */
726 nir_instr
*parent_instr
;
728 /** set of nir_instrs where this register is used (read from) */
729 struct list_head uses
;
731 /** set of nir_ifs where this register is used as a condition */
732 struct list_head if_uses
;
734 uint8_t num_components
;
736 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
744 struct nir_src
*indirect
; /** < NULL for no indirect offset */
745 unsigned base_offset
;
747 /* TODO use-def chain goes here */
751 nir_instr
*parent_instr
;
752 struct list_head def_link
;
755 struct nir_src
*indirect
; /** < NULL for no indirect offset */
756 unsigned base_offset
;
758 /* TODO def-use chain goes here */
763 typedef struct nir_src
{
765 /** Instruction that consumes this value as a source. */
766 nir_instr
*parent_instr
;
767 struct nir_if
*parent_if
;
770 struct list_head use_link
;
780 static inline nir_src
783 nir_src src
= { { NULL
} };
787 #define NIR_SRC_INIT nir_src_init()
789 #define nir_foreach_use(src, reg_or_ssa_def) \
790 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
792 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
793 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
795 #define nir_foreach_if_use(src, reg_or_ssa_def) \
796 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
798 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
799 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
810 static inline nir_dest
813 nir_dest dest
= { { { NULL
} } };
817 #define NIR_DEST_INIT nir_dest_init()
819 #define nir_foreach_def(dest, reg) \
820 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
822 #define nir_foreach_def_safe(dest, reg) \
823 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
825 static inline nir_src
826 nir_src_for_ssa(nir_ssa_def
*def
)
828 nir_src src
= NIR_SRC_INIT
;
836 static inline nir_src
837 nir_src_for_reg(nir_register
*reg
)
839 nir_src src
= NIR_SRC_INIT
;
843 src
.reg
.indirect
= NULL
;
844 src
.reg
.base_offset
= 0;
849 static inline nir_dest
850 nir_dest_for_reg(nir_register
*reg
)
852 nir_dest dest
= NIR_DEST_INIT
;
859 static inline unsigned
860 nir_src_bit_size(nir_src src
)
862 return src
.is_ssa
? src
.ssa
->bit_size
: src
.reg
.reg
->bit_size
;
865 static inline unsigned
866 nir_src_num_components(nir_src src
)
868 return src
.is_ssa
? src
.ssa
->num_components
: src
.reg
.reg
->num_components
;
872 nir_src_is_const(nir_src src
)
875 src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
;
878 static inline unsigned
879 nir_dest_bit_size(nir_dest dest
)
881 return dest
.is_ssa
? dest
.ssa
.bit_size
: dest
.reg
.reg
->bit_size
;
884 static inline unsigned
885 nir_dest_num_components(nir_dest dest
)
887 return dest
.is_ssa
? dest
.ssa
.num_components
: dest
.reg
.reg
->num_components
;
890 /* Are all components the same, ie. .xxxx */
892 nir_is_same_comp_swizzle(uint8_t *swiz
, unsigned nr_comp
)
894 for (unsigned i
= 1; i
< nr_comp
; i
++)
895 if (swiz
[i
] != swiz
[0])
900 /* Are all components sequential, ie. .yzw */
902 nir_is_sequential_comp_swizzle(uint8_t *swiz
, unsigned nr_comp
)
904 for (unsigned i
= 1; i
< nr_comp
; i
++)
905 if (swiz
[i
] != (swiz
[0] + i
))
910 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
911 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
917 * \name input modifiers
921 * For inputs interpreted as floating point, flips the sign bit. For
922 * inputs interpreted as integers, performs the two's complement negation.
927 * Clears the sign bit for floating point values, and computes the integer
928 * absolute value for integers. Note that the negate modifier acts after
929 * the absolute value modifier, therefore if both are set then all inputs
930 * will become negative.
936 * For each input component, says which component of the register it is
937 * chosen from. Note that which elements of the swizzle are used and which
938 * are ignored are based on the write mask for most opcodes - for example,
939 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
940 * a swizzle of {2, x, 1, 0} where x means "don't care."
942 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
949 * \name saturate output modifier
951 * Only valid for opcodes that output floating-point numbers. Clamps the
952 * output to between 0.0 and 1.0 inclusive.
957 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
960 /** NIR sized and unsized types
962 * The values in this enum are carefully chosen so that the sized type is
963 * just the unsized type OR the number of bits.
966 nir_type_invalid
= 0, /* Not a valid type */
970 nir_type_float
= 128,
971 nir_type_bool1
= 1 | nir_type_bool
,
972 nir_type_bool8
= 8 | nir_type_bool
,
973 nir_type_bool16
= 16 | nir_type_bool
,
974 nir_type_bool32
= 32 | nir_type_bool
,
975 nir_type_int1
= 1 | nir_type_int
,
976 nir_type_int8
= 8 | nir_type_int
,
977 nir_type_int16
= 16 | nir_type_int
,
978 nir_type_int32
= 32 | nir_type_int
,
979 nir_type_int64
= 64 | nir_type_int
,
980 nir_type_uint1
= 1 | nir_type_uint
,
981 nir_type_uint8
= 8 | nir_type_uint
,
982 nir_type_uint16
= 16 | nir_type_uint
,
983 nir_type_uint32
= 32 | nir_type_uint
,
984 nir_type_uint64
= 64 | nir_type_uint
,
985 nir_type_float16
= 16 | nir_type_float
,
986 nir_type_float32
= 32 | nir_type_float
,
987 nir_type_float64
= 64 | nir_type_float
,
990 #define NIR_ALU_TYPE_SIZE_MASK 0x79
991 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
993 static inline unsigned
994 nir_alu_type_get_type_size(nir_alu_type type
)
996 return type
& NIR_ALU_TYPE_SIZE_MASK
;
999 static inline unsigned
1000 nir_alu_type_get_base_type(nir_alu_type type
)
1002 return type
& NIR_ALU_TYPE_BASE_TYPE_MASK
;
1005 static inline nir_alu_type
1006 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
1008 switch (base_type
) {
1009 case GLSL_TYPE_BOOL
:
1010 return nir_type_bool1
;
1012 case GLSL_TYPE_UINT
:
1013 return nir_type_uint32
;
1016 return nir_type_int32
;
1018 case GLSL_TYPE_UINT16
:
1019 return nir_type_uint16
;
1021 case GLSL_TYPE_INT16
:
1022 return nir_type_int16
;
1024 case GLSL_TYPE_UINT8
:
1025 return nir_type_uint8
;
1026 case GLSL_TYPE_INT8
:
1027 return nir_type_int8
;
1028 case GLSL_TYPE_UINT64
:
1029 return nir_type_uint64
;
1031 case GLSL_TYPE_INT64
:
1032 return nir_type_int64
;
1034 case GLSL_TYPE_FLOAT
:
1035 return nir_type_float32
;
1037 case GLSL_TYPE_FLOAT16
:
1038 return nir_type_float16
;
1040 case GLSL_TYPE_DOUBLE
:
1041 return nir_type_float64
;
1044 case GLSL_TYPE_SAMPLER
:
1045 case GLSL_TYPE_IMAGE
:
1046 case GLSL_TYPE_ATOMIC_UINT
:
1047 case GLSL_TYPE_STRUCT
:
1048 case GLSL_TYPE_INTERFACE
:
1049 case GLSL_TYPE_ARRAY
:
1050 case GLSL_TYPE_VOID
:
1051 case GLSL_TYPE_SUBROUTINE
:
1052 case GLSL_TYPE_FUNCTION
:
1053 case GLSL_TYPE_ERROR
:
1054 return nir_type_invalid
;
1057 unreachable("unknown type");
1060 static inline nir_alu_type
1061 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
1063 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
1066 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
1067 nir_rounding_mode rnd
);
1069 static inline nir_op
1070 nir_op_vec(unsigned components
)
1072 switch (components
) {
1073 case 1: return nir_op_mov
;
1074 case 2: return nir_op_vec2
;
1075 case 3: return nir_op_vec3
;
1076 case 4: return nir_op_vec4
;
1077 case 8: return nir_op_vec8
;
1078 case 16: return nir_op_vec16
;
1079 default: unreachable("bad component count");
1084 nir_op_is_vec(nir_op op
)
1100 nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode
, unsigned bit_size
)
1102 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
) ||
1103 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
) ||
1104 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
);
1108 nir_is_denorm_flush_to_zero(unsigned execution_mode
, unsigned bit_size
)
1110 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
) ||
1111 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
) ||
1112 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
);
1116 nir_is_denorm_preserve(unsigned execution_mode
, unsigned bit_size
)
1118 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP16
) ||
1119 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP32
) ||
1120 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP64
);
1124 nir_is_rounding_mode_rtne(unsigned execution_mode
, unsigned bit_size
)
1126 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1127 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1128 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1132 nir_is_rounding_mode_rtz(unsigned execution_mode
, unsigned bit_size
)
1134 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1135 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1136 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1140 nir_has_any_rounding_mode_rtz(unsigned execution_mode
)
1142 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1143 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1144 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1148 nir_has_any_rounding_mode_rtne(unsigned execution_mode
)
1150 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1151 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1152 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1155 static inline nir_rounding_mode
1156 nir_get_rounding_mode_from_float_controls(unsigned execution_mode
,
1159 if (nir_alu_type_get_base_type(type
) != nir_type_float
)
1160 return nir_rounding_mode_undef
;
1162 unsigned bit_size
= nir_alu_type_get_type_size(type
);
1164 if (nir_is_rounding_mode_rtz(execution_mode
, bit_size
))
1165 return nir_rounding_mode_rtz
;
1166 if (nir_is_rounding_mode_rtne(execution_mode
, bit_size
))
1167 return nir_rounding_mode_rtne
;
1168 return nir_rounding_mode_undef
;
1172 nir_has_any_rounding_mode_enabled(unsigned execution_mode
)
1175 nir_has_any_rounding_mode_rtne(execution_mode
) ||
1176 nir_has_any_rounding_mode_rtz(execution_mode
);
1182 * Operation where the first two sources are commutative.
1184 * For 2-source operations, this just mathematical commutativity. Some
1185 * 3-source operations, like ffma, are only commutative in the first two
1188 NIR_OP_IS_2SRC_COMMUTATIVE
= (1 << 0),
1189 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
1190 } nir_op_algebraic_property
;
1195 unsigned num_inputs
;
1198 * The number of components in the output
1200 * If non-zero, this is the size of the output and input sizes are
1201 * explicitly given; swizzle and writemask are still in effect, but if
1202 * the output component is masked out, then the input component may
1205 * If zero, the opcode acts in the standard, per-component manner; the
1206 * operation is performed on each component (except the ones that are
1207 * masked out) with the input being taken from the input swizzle for
1210 * The size of some of the inputs may be given (i.e. non-zero) even
1211 * though output_size is zero; in that case, the inputs with a zero
1212 * size act per-component, while the inputs with non-zero size don't.
1214 unsigned output_size
;
1217 * The type of vector that the instruction outputs. Note that the
1218 * staurate modifier is only allowed on outputs with the float type.
1221 nir_alu_type output_type
;
1224 * The number of components in each input
1226 unsigned input_sizes
[NIR_MAX_VEC_COMPONENTS
];
1229 * The type of vector that each input takes. Note that negate and
1230 * absolute value are only allowed on inputs with int or float type and
1231 * behave differently on the two.
1233 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
1235 nir_op_algebraic_property algebraic_properties
;
1237 /* Whether this represents a numeric conversion opcode */
1241 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
1243 typedef struct nir_alu_instr
{
1247 /** Indicates that this ALU instruction generates an exact value
1249 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1250 * really equivalent to either. This indicates that the value generated by
1251 * this operation is high-precision and any code transformations that touch
1252 * it must ensure that the resulting value is bit-for-bit identical to the
1258 * Indicates that this instruction do not cause wrapping to occur, in the
1259 * form of overflow or underflow.
1261 bool no_signed_wrap
:1;
1262 bool no_unsigned_wrap
:1;
1268 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
1269 nir_alu_instr
*instr
);
1270 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
1271 nir_alu_instr
*instr
);
1273 /* is this source channel used? */
1275 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
1278 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1279 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
1281 return (instr
->dest
.write_mask
>> channel
) & 1;
1284 static inline nir_component_mask_t
1285 nir_alu_instr_src_read_mask(const nir_alu_instr
*instr
, unsigned src
)
1287 nir_component_mask_t read_mask
= 0;
1288 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; c
++) {
1289 if (!nir_alu_instr_channel_used(instr
, src
, c
))
1292 read_mask
|= (1 << instr
->src
[src
].swizzle
[c
]);
1298 * Get the number of channels used for a source
1300 static inline unsigned
1301 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
1303 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1304 return nir_op_infos
[instr
->op
].input_sizes
[src
];
1306 return nir_dest_num_components(instr
->dest
.dest
);
1310 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1312 switch (instr
->op
) {
1332 bool nir_const_value_negative_equal(nir_const_value c1
, nir_const_value c2
,
1333 nir_alu_type full_type
);
1335 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
1336 unsigned src1
, unsigned src2
);
1338 bool nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
1339 const nir_alu_instr
*alu2
,
1340 unsigned src1
, unsigned src2
);
1344 nir_deref_type_array
,
1345 nir_deref_type_array_wildcard
,
1346 nir_deref_type_ptr_as_array
,
1347 nir_deref_type_struct
,
1348 nir_deref_type_cast
,
1354 /** The type of this deref instruction */
1355 nir_deref_type deref_type
;
1357 /** The mode of the underlying variable */
1358 nir_variable_mode mode
;
1360 /** The dereferenced type of the resulting pointer value */
1361 const struct glsl_type
*type
;
1364 /** Variable being dereferenced if deref_type is a deref_var */
1367 /** Parent deref if deref_type is not deref_var */
1371 /** Additional deref parameters */
1382 unsigned ptr_stride
;
1386 /** Destination to store the resulting "pointer" */
1390 static inline nir_deref_instr
*nir_src_as_deref(nir_src src
);
1392 static inline nir_deref_instr
*
1393 nir_deref_instr_parent(const nir_deref_instr
*instr
)
1395 if (instr
->deref_type
== nir_deref_type_var
)
1398 return nir_src_as_deref(instr
->parent
);
1401 static inline nir_variable
*
1402 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1404 while (instr
->deref_type
!= nir_deref_type_var
) {
1405 if (instr
->deref_type
== nir_deref_type_cast
)
1408 instr
= nir_deref_instr_parent(instr
);
1414 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1415 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr
*instr
);
1416 bool nir_deref_instr_has_complex_use(nir_deref_instr
*instr
);
1418 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1420 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*instr
);
1425 struct nir_function
*callee
;
1427 unsigned num_params
;
1431 #include "nir_intrinsics.h"
1433 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1435 /** Represents an intrinsic
1437 * An intrinsic is an instruction type for handling things that are
1438 * more-or-less regular operations but don't just consume and produce SSA
1439 * values like ALU operations do. Intrinsics are not for things that have
1440 * special semantic meaning such as phi nodes and parallel copies.
1441 * Examples of intrinsics include variable load/store operations, system
1442 * value loads, and the like. Even though texturing more-or-less falls
1443 * under this category, texturing is its own instruction type because
1444 * trying to represent texturing with intrinsics would lead to a
1445 * combinatorial explosion of intrinsic opcodes.
1447 * By having a single instruction type for handling a lot of different
1448 * cases, optimization passes can look for intrinsics and, for the most
1449 * part, completely ignore them. Each intrinsic type also has a few
1450 * possible flags that govern whether or not they can be reordered or
1451 * eliminated. That way passes like dead code elimination can still work
1452 * on intrisics without understanding the meaning of each.
1454 * Each intrinsic has some number of constant indices, some number of
1455 * variables, and some number of sources. What these sources, variables,
1456 * and indices mean depends on the intrinsic and is documented with the
1457 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1458 * instructions are the only types of instruction that can operate on
1464 nir_intrinsic_op intrinsic
;
1468 /** number of components if this is a vectorized intrinsic
1470 * Similarly to ALU operations, some intrinsics are vectorized.
1471 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1472 * For vectorized intrinsics, the num_components field specifies the
1473 * number of destination components and the number of source components
1474 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1476 uint8_t num_components
;
1478 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1481 } nir_intrinsic_instr
;
1483 static inline nir_variable
*
1484 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1486 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1490 /* Memory ordering. */
1491 NIR_MEMORY_ACQUIRE
= 1 << 0,
1492 NIR_MEMORY_RELEASE
= 1 << 1,
1493 NIR_MEMORY_ACQ_REL
= NIR_MEMORY_ACQUIRE
| NIR_MEMORY_RELEASE
,
1495 /* Memory visibility operations. */
1496 NIR_MEMORY_MAKE_AVAILABLE
= 1 << 2,
1497 NIR_MEMORY_MAKE_VISIBLE
= 1 << 3,
1498 } nir_memory_semantics
;
1501 NIR_SCOPE_INVOCATION
,
1503 NIR_SCOPE_WORKGROUP
,
1504 NIR_SCOPE_QUEUE_FAMILY
,
1509 * \name NIR intrinsics semantic flags
1511 * information about what the compiler can do with the intrinsics.
1513 * \sa nir_intrinsic_info::flags
1517 * whether the intrinsic can be safely eliminated if none of its output
1518 * value is not being used.
1520 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1523 * Whether the intrinsic can be reordered with respect to any other
1524 * intrinsic, i.e. whether the only reordering dependencies of the
1525 * intrinsic are due to the register reads/writes.
1527 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1528 } nir_intrinsic_semantic_flag
;
1531 * \name NIR intrinsics const-index flag
1533 * Indicates the usage of a const_index slot.
1535 * \sa nir_intrinsic_info::index_map
1539 * Generally instructions that take a offset src argument, can encode
1540 * a constant 'base' value which is added to the offset.
1542 NIR_INTRINSIC_BASE
= 1,
1545 * For store instructions, a writemask for the store.
1547 NIR_INTRINSIC_WRMASK
,
1550 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1552 NIR_INTRINSIC_STREAM_ID
,
1555 * The clip-plane id for load_user_clip_plane intrinsic.
1557 NIR_INTRINSIC_UCP_ID
,
1560 * The amount of data, starting from BASE, that this instruction may
1561 * access. This is used to provide bounds if the offset is not constant.
1563 NIR_INTRINSIC_RANGE
,
1566 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1568 NIR_INTRINSIC_DESC_SET
,
1571 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1573 NIR_INTRINSIC_BINDING
,
1578 NIR_INTRINSIC_COMPONENT
,
1581 * Interpolation mode (only meaningful for FS inputs).
1583 NIR_INTRINSIC_INTERP_MODE
,
1586 * A binary nir_op to use when performing a reduction or scan operation
1588 NIR_INTRINSIC_REDUCTION_OP
,
1591 * Cluster size for reduction operations
1593 NIR_INTRINSIC_CLUSTER_SIZE
,
1596 * Parameter index for a load_param intrinsic
1598 NIR_INTRINSIC_PARAM_IDX
,
1601 * Image dimensionality for image intrinsics
1603 * One of GLSL_SAMPLER_DIM_*
1605 NIR_INTRINSIC_IMAGE_DIM
,
1608 * Non-zero if we are accessing an array image
1610 NIR_INTRINSIC_IMAGE_ARRAY
,
1613 * Image format for image intrinsics
1615 NIR_INTRINSIC_FORMAT
,
1618 * Access qualifiers for image and memory access intrinsics
1620 NIR_INTRINSIC_ACCESS
,
1623 * Alignment for offsets and addresses
1625 * These two parameters, specify an alignment in terms of a multiplier and
1626 * an offset. The offset or address parameter X of the intrinsic is
1627 * guaranteed to satisfy the following:
1629 * (X - align_offset) % align_mul == 0
1631 NIR_INTRINSIC_ALIGN_MUL
,
1632 NIR_INTRINSIC_ALIGN_OFFSET
,
1635 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1637 NIR_INTRINSIC_DESC_TYPE
,
1640 * The nir_alu_type of a uniform/input/output
1645 * The swizzle mask for the instructions
1646 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1648 NIR_INTRINSIC_SWIZZLE_MASK
,
1650 /* Separate source/dest access flags for copies */
1651 NIR_INTRINSIC_SRC_ACCESS
,
1652 NIR_INTRINSIC_DST_ACCESS
,
1654 /* Driver location for nir_load_patch_location_ir3 */
1655 NIR_INTRINSIC_DRIVER_LOCATION
,
1658 * Mask of nir_memory_semantics, includes ordering and visibility.
1660 NIR_INTRINSIC_MEMORY_SEMANTICS
,
1663 * Mask of nir_variable_modes affected by the memory operation.
1665 NIR_INTRINSIC_MEMORY_MODES
,
1668 * Value of nir_scope.
1670 NIR_INTRINSIC_MEMORY_SCOPE
,
1672 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1674 } nir_intrinsic_index_flag
;
1676 #define NIR_INTRINSIC_MAX_INPUTS 5
1681 unsigned num_srcs
; /** < number of register/SSA inputs */
1683 /** number of components of each input register
1685 * If this value is 0, the number of components is given by the
1686 * num_components field of nir_intrinsic_instr. If this value is -1, the
1687 * intrinsic consumes however many components are provided and it is not
1690 int src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1694 /** number of components of the output register
1696 * If this value is 0, the number of components is given by the
1697 * num_components field of nir_intrinsic_instr.
1699 unsigned dest_components
;
1701 /** bitfield of legal bit sizes */
1702 unsigned dest_bit_sizes
;
1704 /** the number of constant indices used by the intrinsic */
1705 unsigned num_indices
;
1707 /** indicates the usage of intr->const_index[n] */
1708 unsigned index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1710 /** semantic flags for calls to this intrinsic */
1711 nir_intrinsic_semantic_flag flags
;
1712 } nir_intrinsic_info
;
1714 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1716 static inline unsigned
1717 nir_intrinsic_src_components(nir_intrinsic_instr
*intr
, unsigned srcn
)
1719 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1720 assert(srcn
< info
->num_srcs
);
1721 if (info
->src_components
[srcn
] > 0)
1722 return info
->src_components
[srcn
];
1723 else if (info
->src_components
[srcn
] == 0)
1724 return intr
->num_components
;
1726 return nir_src_num_components(intr
->src
[srcn
]);
1729 static inline unsigned
1730 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1732 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1733 if (!info
->has_dest
)
1735 else if (info
->dest_components
)
1736 return info
->dest_components
;
1738 return intr
->num_components
;
1741 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1742 static inline type \
1743 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1745 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1746 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1747 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1749 static inline void \
1750 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1752 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1753 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1754 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1757 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1758 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1759 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1760 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1761 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1762 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1763 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1764 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1765 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1766 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1767 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1768 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1769 INTRINSIC_IDX_ACCESSORS(image_dim
, IMAGE_DIM
, enum glsl_sampler_dim
)
1770 INTRINSIC_IDX_ACCESSORS(image_array
, IMAGE_ARRAY
, bool)
1771 INTRINSIC_IDX_ACCESSORS(access
, ACCESS
, enum gl_access_qualifier
)
1772 INTRINSIC_IDX_ACCESSORS(src_access
, SRC_ACCESS
, enum gl_access_qualifier
)
1773 INTRINSIC_IDX_ACCESSORS(dst_access
, DST_ACCESS
, enum gl_access_qualifier
)
1774 INTRINSIC_IDX_ACCESSORS(format
, FORMAT
, enum pipe_format
)
1775 INTRINSIC_IDX_ACCESSORS(align_mul
, ALIGN_MUL
, unsigned)
1776 INTRINSIC_IDX_ACCESSORS(align_offset
, ALIGN_OFFSET
, unsigned)
1777 INTRINSIC_IDX_ACCESSORS(desc_type
, DESC_TYPE
, unsigned)
1778 INTRINSIC_IDX_ACCESSORS(type
, TYPE
, nir_alu_type
)
1779 INTRINSIC_IDX_ACCESSORS(swizzle_mask
, SWIZZLE_MASK
, unsigned)
1780 INTRINSIC_IDX_ACCESSORS(driver_location
, DRIVER_LOCATION
, unsigned)
1781 INTRINSIC_IDX_ACCESSORS(memory_semantics
, MEMORY_SEMANTICS
, nir_memory_semantics
)
1782 INTRINSIC_IDX_ACCESSORS(memory_modes
, MEMORY_MODES
, nir_variable_mode
)
1783 INTRINSIC_IDX_ACCESSORS(memory_scope
, MEMORY_SCOPE
, nir_scope
)
1786 nir_intrinsic_set_align(nir_intrinsic_instr
*intrin
,
1787 unsigned align_mul
, unsigned align_offset
)
1789 assert(util_is_power_of_two_nonzero(align_mul
));
1790 assert(align_offset
< align_mul
);
1791 nir_intrinsic_set_align_mul(intrin
, align_mul
);
1792 nir_intrinsic_set_align_offset(intrin
, align_offset
);
1795 /** Returns a simple alignment for a load/store intrinsic offset
1797 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1798 * and ALIGN_OFFSET parameters, this helper takes both into account and
1799 * provides a single simple alignment parameter. The offset X is guaranteed
1800 * to satisfy X % align == 0.
1802 static inline unsigned
1803 nir_intrinsic_align(const nir_intrinsic_instr
*intrin
)
1805 const unsigned align_mul
= nir_intrinsic_align_mul(intrin
);
1806 const unsigned align_offset
= nir_intrinsic_align_offset(intrin
);
1807 assert(align_offset
< align_mul
);
1808 return align_offset
? 1 << (ffs(align_offset
) - 1) : align_mul
;
1812 nir_image_intrinsic_coord_components(const nir_intrinsic_instr
*instr
);
1814 /* Converts a image_deref_* intrinsic into a image_* one */
1815 void nir_rewrite_image_intrinsic(nir_intrinsic_instr
*instr
,
1816 nir_ssa_def
*handle
, bool bindless
);
1818 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1820 nir_intrinsic_can_reorder(nir_intrinsic_instr
*instr
)
1822 if (instr
->intrinsic
== nir_intrinsic_load_deref
||
1823 instr
->intrinsic
== nir_intrinsic_load_ssbo
||
1824 instr
->intrinsic
== nir_intrinsic_bindless_image_load
||
1825 instr
->intrinsic
== nir_intrinsic_image_deref_load
||
1826 instr
->intrinsic
== nir_intrinsic_image_load
) {
1827 return nir_intrinsic_access(instr
) & ACCESS_CAN_REORDER
;
1829 const nir_intrinsic_info
*info
=
1830 &nir_intrinsic_infos
[instr
->intrinsic
];
1831 return (info
->flags
& NIR_INTRINSIC_CAN_ELIMINATE
) &&
1832 (info
->flags
& NIR_INTRINSIC_CAN_REORDER
);
1837 * \group texture information
1839 * This gives semantic information about textures which is useful to the
1840 * frontend, the backend, and lowering passes, but not the optimizer.
1845 nir_tex_src_projector
,
1846 nir_tex_src_comparator
, /* shadow comparator */
1850 nir_tex_src_min_lod
,
1851 nir_tex_src_ms_index
, /* MSAA sample index */
1852 nir_tex_src_ms_mcs
, /* MSAA compression value */
1855 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
1856 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
1857 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
1858 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
1859 nir_tex_src_texture_handle
, /* < bindless texture handle */
1860 nir_tex_src_sampler_handle
, /* < bindless sampler handle */
1861 nir_tex_src_plane
, /* < selects plane for planar textures */
1862 nir_num_tex_src_types
1867 nir_tex_src_type src_type
;
1871 nir_texop_tex
, /**< Regular texture look-up */
1872 nir_texop_txb
, /**< Texture look-up with LOD bias */
1873 nir_texop_txl
, /**< Texture look-up with explicit LOD */
1874 nir_texop_txd
, /**< Texture look-up with partial derivatives */
1875 nir_texop_txf
, /**< Texel fetch with explicit LOD */
1876 nir_texop_txf_ms
, /**< Multisample texture fetch */
1877 nir_texop_txf_ms_fb
, /**< Multisample texture fetch from framebuffer */
1878 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
1879 nir_texop_txs
, /**< Texture size */
1880 nir_texop_lod
, /**< Texture lod query */
1881 nir_texop_tg4
, /**< Texture gather */
1882 nir_texop_query_levels
, /**< Texture levels query */
1883 nir_texop_texture_samples
, /**< Texture samples query */
1884 nir_texop_samples_identical
, /**< Query whether all samples are definitely
1887 nir_texop_tex_prefetch
, /**< Regular texture look-up, eligible for pre-dispatch */
1888 nir_texop_fragment_fetch
, /**< Multisample fragment color texture fetch */
1889 nir_texop_fragment_mask_fetch
,/**< Multisample fragment mask texture fetch */
1895 enum glsl_sampler_dim sampler_dim
;
1896 nir_alu_type dest_type
;
1901 unsigned num_srcs
, coord_components
;
1902 bool is_array
, is_shadow
;
1905 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1906 * components or the new-style shadow that outputs 1 component.
1908 bool is_new_style_shadow
;
1910 /* gather component selector */
1911 unsigned component
: 2;
1913 /* gather offsets */
1914 int8_t tg4_offsets
[4][2];
1916 /* True if the texture index or handle is not dynamically uniform */
1917 bool texture_non_uniform
;
1919 /* True if the sampler index or handle is not dynamically uniform */
1920 bool sampler_non_uniform
;
1922 /** The texture index
1924 * If this texture instruction has a nir_tex_src_texture_offset source,
1925 * then the texture index is given by texture_index + texture_offset.
1927 unsigned texture_index
;
1929 /** The sampler index
1931 * The following operations do not require a sampler and, as such, this
1932 * field should be ignored:
1934 * - nir_texop_txf_ms
1937 * - nir_texop_query_levels
1938 * - nir_texop_texture_samples
1939 * - nir_texop_samples_identical
1941 * If this texture instruction has a nir_tex_src_sampler_offset source,
1942 * then the sampler index is given by sampler_index + sampler_offset.
1944 unsigned sampler_index
;
1947 static inline unsigned
1948 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
1950 switch (instr
->op
) {
1951 case nir_texop_txs
: {
1953 switch (instr
->sampler_dim
) {
1954 case GLSL_SAMPLER_DIM_1D
:
1955 case GLSL_SAMPLER_DIM_BUF
:
1958 case GLSL_SAMPLER_DIM_2D
:
1959 case GLSL_SAMPLER_DIM_CUBE
:
1960 case GLSL_SAMPLER_DIM_MS
:
1961 case GLSL_SAMPLER_DIM_RECT
:
1962 case GLSL_SAMPLER_DIM_EXTERNAL
:
1963 case GLSL_SAMPLER_DIM_SUBPASS
:
1966 case GLSL_SAMPLER_DIM_3D
:
1970 unreachable("not reached");
1972 if (instr
->is_array
)
1980 case nir_texop_texture_samples
:
1981 case nir_texop_query_levels
:
1982 case nir_texop_samples_identical
:
1983 case nir_texop_fragment_mask_fetch
:
1987 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
1994 /* Returns true if this texture operation queries something about the texture
1995 * rather than actually sampling it.
1998 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
2000 switch (instr
->op
) {
2003 case nir_texop_texture_samples
:
2004 case nir_texop_query_levels
:
2005 case nir_texop_txf_ms_mcs
:
2012 case nir_texop_txf_ms
:
2013 case nir_texop_txf_ms_fb
:
2017 unreachable("Invalid texture opcode");
2022 nir_tex_instr_has_implicit_derivative(const nir_tex_instr
*instr
)
2024 switch (instr
->op
) {
2034 static inline nir_alu_type
2035 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
2037 switch (instr
->src
[src
].src_type
) {
2038 case nir_tex_src_coord
:
2039 switch (instr
->op
) {
2041 case nir_texop_txf_ms
:
2042 case nir_texop_txf_ms_fb
:
2043 case nir_texop_txf_ms_mcs
:
2044 case nir_texop_samples_identical
:
2045 return nir_type_int
;
2048 return nir_type_float
;
2051 case nir_tex_src_lod
:
2052 switch (instr
->op
) {
2055 return nir_type_int
;
2058 return nir_type_float
;
2061 case nir_tex_src_projector
:
2062 case nir_tex_src_comparator
:
2063 case nir_tex_src_bias
:
2064 case nir_tex_src_min_lod
:
2065 case nir_tex_src_ddx
:
2066 case nir_tex_src_ddy
:
2067 return nir_type_float
;
2069 case nir_tex_src_offset
:
2070 case nir_tex_src_ms_index
:
2071 case nir_tex_src_plane
:
2072 return nir_type_int
;
2074 case nir_tex_src_ms_mcs
:
2075 case nir_tex_src_texture_deref
:
2076 case nir_tex_src_sampler_deref
:
2077 case nir_tex_src_texture_offset
:
2078 case nir_tex_src_sampler_offset
:
2079 case nir_tex_src_texture_handle
:
2080 case nir_tex_src_sampler_handle
:
2081 return nir_type_uint
;
2083 case nir_num_tex_src_types
:
2084 unreachable("nir_num_tex_src_types is not a valid source type");
2087 unreachable("Invalid texture source type");
2090 static inline unsigned
2091 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
2093 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
2094 return instr
->coord_components
;
2096 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
2097 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
2100 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
2101 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
2102 if (instr
->is_array
)
2103 return instr
->coord_components
- 1;
2105 return instr
->coord_components
;
2108 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
2109 * the offset, since a cube maps to a single face.
2111 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
2112 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2114 else if (instr
->is_array
)
2115 return instr
->coord_components
- 1;
2117 return instr
->coord_components
;
2124 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
2126 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
2127 if (instr
->src
[i
].src_type
== type
)
2133 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
2134 nir_tex_src_type src_type
,
2137 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
2139 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
);
2146 nir_const_value value
[];
2147 } nir_load_const_instr
;
2160 /* creates a new SSA variable in an undefined state */
2165 } nir_ssa_undef_instr
;
2168 struct exec_node node
;
2170 /* The predecessor block corresponding to this source */
2171 struct nir_block
*pred
;
2176 #define nir_foreach_phi_src(phi_src, phi) \
2177 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
2178 #define nir_foreach_phi_src_safe(phi_src, phi) \
2179 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
2184 struct exec_list srcs
; /** < list of nir_phi_src */
2190 struct exec_node node
;
2193 } nir_parallel_copy_entry
;
2195 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
2196 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
2201 /* A list of nir_parallel_copy_entrys. The sources of all of the
2202 * entries are copied to the corresponding destinations "in parallel".
2203 * In other words, if we have two entries: a -> b and b -> a, the values
2206 struct exec_list entries
;
2207 } nir_parallel_copy_instr
;
2209 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
2210 type
, nir_instr_type_alu
)
2211 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
2212 type
, nir_instr_type_deref
)
2213 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
2214 type
, nir_instr_type_call
)
2215 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
2216 type
, nir_instr_type_jump
)
2217 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
2218 type
, nir_instr_type_tex
)
2219 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
2220 type
, nir_instr_type_intrinsic
)
2221 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
2222 type
, nir_instr_type_load_const
)
2223 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
2224 type
, nir_instr_type_ssa_undef
)
2225 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
2226 type
, nir_instr_type_phi
)
2227 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
2228 nir_parallel_copy_instr
, instr
,
2229 type
, nir_instr_type_parallel_copy
)
2232 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
2233 static inline type \
2234 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2236 assert(nir_src_is_const(src)); \
2237 nir_load_const_instr *load = \
2238 nir_instr_as_load_const(src.ssa->parent_instr); \
2239 assert(comp < load->def.num_components); \
2240 return nir_const_value_as_##suffix(load->value[comp], \
2241 load->def.bit_size); \
2244 static inline type \
2245 nir_src_as_##suffix(nir_src src) \
2247 assert(nir_src_num_components(src) == 1); \
2248 return nir_src_comp_as_##suffix(src, 0); \
2251 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2252 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint
)
2253 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2254 NIR_DEFINE_SRC_AS_CONST(double, float)
2256 #undef NIR_DEFINE_SRC_AS_CONST
2265 nir_ssa_scalar_is_const(nir_ssa_scalar s
)
2267 return s
.def
->parent_instr
->type
== nir_instr_type_load_const
;
2270 static inline nir_const_value
2271 nir_ssa_scalar_as_const_value(nir_ssa_scalar s
)
2273 assert(s
.comp
< s
.def
->num_components
);
2274 nir_load_const_instr
*load
= nir_instr_as_load_const(s
.def
->parent_instr
);
2275 return load
->value
[s
.comp
];
2278 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2279 static inline type \
2280 nir_ssa_scalar_as_##suffix(nir_ssa_scalar s) \
2282 return nir_const_value_as_##suffix( \
2283 nir_ssa_scalar_as_const_value(s), s.def->bit_size); \
2286 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2287 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint
)
2288 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2289 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2291 #undef NIR_DEFINE_SCALAR_AS_CONST
2294 nir_ssa_scalar_is_alu(nir_ssa_scalar s
)
2296 return s
.def
->parent_instr
->type
== nir_instr_type_alu
;
2299 static inline nir_op
2300 nir_ssa_scalar_alu_op(nir_ssa_scalar s
)
2302 return nir_instr_as_alu(s
.def
->parent_instr
)->op
;
2305 static inline nir_ssa_scalar
2306 nir_ssa_scalar_chase_alu_src(nir_ssa_scalar s
, unsigned alu_src_idx
)
2308 nir_ssa_scalar out
= { NULL
, 0 };
2310 nir_alu_instr
*alu
= nir_instr_as_alu(s
.def
->parent_instr
);
2311 assert(alu_src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
2313 /* Our component must be written */
2314 assert(s
.comp
< s
.def
->num_components
);
2315 assert(alu
->dest
.write_mask
& (1u << s
.comp
));
2317 assert(alu
->src
[alu_src_idx
].src
.is_ssa
);
2318 out
.def
= alu
->src
[alu_src_idx
].src
.ssa
;
2320 if (nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 0) {
2321 /* The ALU src is unsized so the source component follows the
2322 * destination component.
2324 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[s
.comp
];
2326 /* This is a sized source so all source components work together to
2327 * produce all the destination components. Since we need to return a
2328 * scalar, this only works if the source is a scalar.
2330 assert(nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 1);
2331 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[0];
2333 assert(out
.comp
< out
.def
->num_components
);
2342 * Control flow consists of a tree of control flow nodes, which include
2343 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2344 * instructions that always run start-to-finish. Each basic block also keeps
2345 * track of its successors (blocks which may run immediately after the current
2346 * block) and predecessors (blocks which could have run immediately before the
2347 * current block). Each function also has a start block and an end block which
2348 * all return statements point to (which is always empty). Together, all the
2349 * blocks with their predecessors and successors make up the control flow
2350 * graph (CFG) of the function. There are helpers that modify the tree of
2351 * control flow nodes while modifying the CFG appropriately; these should be
2352 * used instead of modifying the tree directly.
2359 nir_cf_node_function
2362 typedef struct nir_cf_node
{
2363 struct exec_node node
;
2364 nir_cf_node_type type
;
2365 struct nir_cf_node
*parent
;
2368 typedef struct nir_block
{
2369 nir_cf_node cf_node
;
2371 struct exec_list instr_list
; /** < list of nir_instr */
2373 /** generic block index; generated by nir_index_blocks */
2377 * Each block can only have up to 2 successors, so we put them in a simple
2378 * array - no need for anything more complicated.
2380 struct nir_block
*successors
[2];
2382 /* Set of nir_block predecessors in the CFG */
2383 struct set
*predecessors
;
2386 * this node's immediate dominator in the dominance tree - set to NULL for
2389 struct nir_block
*imm_dom
;
2391 /* This node's children in the dominance tree */
2392 unsigned num_dom_children
;
2393 struct nir_block
**dom_children
;
2395 /* Set of nir_blocks on the dominance frontier of this block */
2396 struct set
*dom_frontier
;
2399 * These two indices have the property that dom_{pre,post}_index for each
2400 * child of this block in the dominance tree will always be between
2401 * dom_pre_index and dom_post_index for this block, which makes testing if
2402 * a given block is dominated by another block an O(1) operation.
2404 unsigned dom_pre_index
, dom_post_index
;
2406 /* live in and out for this block; used for liveness analysis */
2407 BITSET_WORD
*live_in
;
2408 BITSET_WORD
*live_out
;
2411 static inline nir_instr
*
2412 nir_block_first_instr(nir_block
*block
)
2414 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
2415 return exec_node_data(nir_instr
, head
, node
);
2418 static inline nir_instr
*
2419 nir_block_last_instr(nir_block
*block
)
2421 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
2422 return exec_node_data(nir_instr
, tail
, node
);
2426 nir_block_ends_in_jump(nir_block
*block
)
2428 return !exec_list_is_empty(&block
->instr_list
) &&
2429 nir_block_last_instr(block
)->type
== nir_instr_type_jump
;
2432 #define nir_foreach_instr(instr, block) \
2433 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
2434 #define nir_foreach_instr_reverse(instr, block) \
2435 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
2436 #define nir_foreach_instr_safe(instr, block) \
2437 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
2438 #define nir_foreach_instr_reverse_safe(instr, block) \
2439 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
2442 nir_selection_control_none
= 0x0,
2443 nir_selection_control_flatten
= 0x1,
2444 nir_selection_control_dont_flatten
= 0x2,
2445 } nir_selection_control
;
2447 typedef struct nir_if
{
2448 nir_cf_node cf_node
;
2450 nir_selection_control control
;
2452 struct exec_list then_list
; /** < list of nir_cf_node */
2453 struct exec_list else_list
; /** < list of nir_cf_node */
2459 /** Instruction that generates nif::condition. */
2460 nir_instr
*conditional_instr
;
2462 /** Block within ::nif that has the break instruction. */
2463 nir_block
*break_block
;
2465 /** Last block for the then- or else-path that does not contain the break. */
2466 nir_block
*continue_from_block
;
2468 /** True when ::break_block is in the else-path of ::nif. */
2469 bool continue_from_then
;
2472 /* This is true if the terminators exact trip count is unknown. For
2475 * for (int i = 0; i < imin(x, 4); i++)
2478 * Here loop analysis would have set a max_trip_count of 4 however we dont
2479 * know for sure that this is the exact trip count.
2481 bool exact_trip_count_unknown
;
2483 struct list_head loop_terminator_link
;
2484 } nir_loop_terminator
;
2487 /* Estimated cost (in number of instructions) of the loop */
2488 unsigned instr_cost
;
2490 /* Guessed trip count based on array indexing */
2491 unsigned guessed_trip_count
;
2493 /* Maximum number of times the loop is run (if known) */
2494 unsigned max_trip_count
;
2496 /* Do we know the exact number of times the loop will be run */
2497 bool exact_trip_count_known
;
2499 /* Unroll the loop regardless of its size */
2502 /* Does the loop contain complex loop terminators, continues or other
2503 * complex behaviours? If this is true we can't rely on
2504 * loop_terminator_list to be complete or accurate.
2508 nir_loop_terminator
*limiting_terminator
;
2510 /* A list of loop_terminators terminating this loop. */
2511 struct list_head loop_terminator_list
;
2515 nir_loop_control_none
= 0x0,
2516 nir_loop_control_unroll
= 0x1,
2517 nir_loop_control_dont_unroll
= 0x2,
2521 nir_cf_node cf_node
;
2523 struct exec_list body
; /** < list of nir_cf_node */
2525 nir_loop_info
*info
;
2526 nir_loop_control control
;
2527 bool partially_unrolled
;
2531 * Various bits of metadata that can may be created or required by
2532 * optimization and analysis passes
2535 nir_metadata_none
= 0x0,
2536 nir_metadata_block_index
= 0x1,
2537 nir_metadata_dominance
= 0x2,
2538 nir_metadata_live_ssa_defs
= 0x4,
2539 nir_metadata_not_properly_reset
= 0x8,
2540 nir_metadata_loop_analysis
= 0x10,
2544 nir_cf_node cf_node
;
2546 /** pointer to the function of which this is an implementation */
2547 struct nir_function
*function
;
2549 struct exec_list body
; /** < list of nir_cf_node */
2551 nir_block
*end_block
;
2553 /** list for all local variables in the function */
2554 struct exec_list locals
;
2556 /** list of local registers in the function */
2557 struct exec_list registers
;
2559 /** next available local register index */
2562 /** next available SSA value index */
2565 /* total number of basic blocks, only valid when block_index_dirty = false */
2566 unsigned num_blocks
;
2568 nir_metadata valid_metadata
;
2569 } nir_function_impl
;
2571 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2572 nir_start_block(nir_function_impl
*impl
)
2574 return (nir_block
*) impl
->body
.head_sentinel
.next
;
2577 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2578 nir_impl_last_block(nir_function_impl
*impl
)
2580 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
2583 static inline nir_cf_node
*
2584 nir_cf_node_next(nir_cf_node
*node
)
2586 struct exec_node
*next
= exec_node_get_next(&node
->node
);
2587 if (exec_node_is_tail_sentinel(next
))
2590 return exec_node_data(nir_cf_node
, next
, node
);
2593 static inline nir_cf_node
*
2594 nir_cf_node_prev(nir_cf_node
*node
)
2596 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
2597 if (exec_node_is_head_sentinel(prev
))
2600 return exec_node_data(nir_cf_node
, prev
, node
);
2604 nir_cf_node_is_first(const nir_cf_node
*node
)
2606 return exec_node_is_head_sentinel(node
->node
.prev
);
2610 nir_cf_node_is_last(const nir_cf_node
*node
)
2612 return exec_node_is_tail_sentinel(node
->node
.next
);
2615 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
2616 type
, nir_cf_node_block
)
2617 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
2618 type
, nir_cf_node_if
)
2619 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
2620 type
, nir_cf_node_loop
)
2621 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
2622 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
2624 static inline nir_block
*
2625 nir_if_first_then_block(nir_if
*if_stmt
)
2627 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
2628 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2631 static inline nir_block
*
2632 nir_if_last_then_block(nir_if
*if_stmt
)
2634 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
2635 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2638 static inline nir_block
*
2639 nir_if_first_else_block(nir_if
*if_stmt
)
2641 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
2642 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2645 static inline nir_block
*
2646 nir_if_last_else_block(nir_if
*if_stmt
)
2648 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
2649 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2652 static inline nir_block
*
2653 nir_loop_first_block(nir_loop
*loop
)
2655 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
2656 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2659 static inline nir_block
*
2660 nir_loop_last_block(nir_loop
*loop
)
2662 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
2663 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2667 * Return true if this list of cf_nodes contains a single empty block.
2670 nir_cf_list_is_empty_block(struct exec_list
*cf_list
)
2672 if (exec_list_is_singular(cf_list
)) {
2673 struct exec_node
*head
= exec_list_get_head(cf_list
);
2675 nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2676 return exec_list_is_empty(&block
->instr_list
);
2682 uint8_t num_components
;
2686 typedef struct nir_function
{
2687 struct exec_node node
;
2690 struct nir_shader
*shader
;
2692 unsigned num_params
;
2693 nir_parameter
*params
;
2695 /** The implementation of this function.
2697 * If the function is only declared and not implemented, this is NULL.
2699 nir_function_impl
*impl
;
2705 nir_lower_imul64
= (1 << 0),
2706 nir_lower_isign64
= (1 << 1),
2707 /** Lower all int64 modulus and division opcodes */
2708 nir_lower_divmod64
= (1 << 2),
2709 /** Lower all 64-bit umul_high and imul_high opcodes */
2710 nir_lower_imul_high64
= (1 << 3),
2711 nir_lower_mov64
= (1 << 4),
2712 nir_lower_icmp64
= (1 << 5),
2713 nir_lower_iadd64
= (1 << 6),
2714 nir_lower_iabs64
= (1 << 7),
2715 nir_lower_ineg64
= (1 << 8),
2716 nir_lower_logic64
= (1 << 9),
2717 nir_lower_minmax64
= (1 << 10),
2718 nir_lower_shift64
= (1 << 11),
2719 nir_lower_imul_2x32_64
= (1 << 12),
2720 nir_lower_extract64
= (1 << 13),
2721 nir_lower_ufind_msb64
= (1 << 14),
2722 } nir_lower_int64_options
;
2725 nir_lower_drcp
= (1 << 0),
2726 nir_lower_dsqrt
= (1 << 1),
2727 nir_lower_drsq
= (1 << 2),
2728 nir_lower_dtrunc
= (1 << 3),
2729 nir_lower_dfloor
= (1 << 4),
2730 nir_lower_dceil
= (1 << 5),
2731 nir_lower_dfract
= (1 << 6),
2732 nir_lower_dround_even
= (1 << 7),
2733 nir_lower_dmod
= (1 << 8),
2734 nir_lower_dsub
= (1 << 9),
2735 nir_lower_ddiv
= (1 << 10),
2736 nir_lower_fp64_full_software
= (1 << 11),
2737 } nir_lower_doubles_options
;
2740 nir_divergence_single_prim_per_subgroup
= (1 << 0),
2741 nir_divergence_single_patch_per_tcs_subgroup
= (1 << 1),
2742 nir_divergence_single_patch_per_tes_subgroup
= (1 << 2),
2743 nir_divergence_view_index_uniform
= (1 << 3),
2744 } nir_divergence_options
;
2746 typedef struct nir_shader_compiler_options
{
2752 /** Lowers flrp when it does not support doubles */
2759 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2760 bool lower_bitfield_extract
;
2761 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
2762 bool lower_bitfield_extract_to_shifts
;
2763 /** Lowers bitfield_insert to bfi/bfm */
2764 bool lower_bitfield_insert
;
2765 /** Lowers bitfield_insert to compares, and shifts. */
2766 bool lower_bitfield_insert_to_shifts
;
2767 /** Lowers bitfield_insert to bfm/bitfield_select. */
2768 bool lower_bitfield_insert_to_bitfield_select
;
2769 /** Lowers bitfield_reverse to shifts. */
2770 bool lower_bitfield_reverse
;
2771 /** Lowers bit_count to shifts. */
2772 bool lower_bit_count
;
2773 /** Lowers ifind_msb to compare and ufind_msb */
2774 bool lower_ifind_msb
;
2775 /** Lowers find_lsb to ufind_msb and logic ops */
2776 bool lower_find_lsb
;
2777 bool lower_uadd_carry
;
2778 bool lower_usub_borrow
;
2779 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2780 bool lower_mul_high
;
2781 /** lowers fneg and ineg to fsub and isub. */
2783 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2786 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2789 /* lower fall_equalN/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
2790 bool lower_vector_cmp
;
2792 /** enables rules to lower idiv by power-of-two: */
2795 /** enable rules to avoid bit ops */
2798 /** enables rules to lower isign to imin+imax */
2801 /** enables rules to lower fsign to fsub and flt */
2804 /* lower fdph to fdot4 */
2807 /** lower fdot to fmul and fsum/fadd. */
2810 /* Does the native fdot instruction replicate its result for four
2811 * components? If so, then opt_algebraic_late will turn all fdotN
2812 * instructions into fdot_replicatedN instructions.
2814 bool fdot_replicates
;
2816 /** lowers ffloor to fsub+ffract: */
2819 /** lowers ffract to fsub+ffloor: */
2822 /** lowers fceil to fneg+ffloor+fneg: */
2829 bool lower_pack_half_2x16
;
2830 bool lower_pack_half_2x16_split
;
2831 bool lower_pack_unorm_2x16
;
2832 bool lower_pack_snorm_2x16
;
2833 bool lower_pack_unorm_4x8
;
2834 bool lower_pack_snorm_4x8
;
2835 bool lower_unpack_half_2x16
;
2836 bool lower_unpack_half_2x16_split
;
2837 bool lower_unpack_unorm_2x16
;
2838 bool lower_unpack_snorm_2x16
;
2839 bool lower_unpack_unorm_4x8
;
2840 bool lower_unpack_snorm_4x8
;
2842 bool lower_extract_byte
;
2843 bool lower_extract_word
;
2845 bool lower_all_io_to_temps
;
2846 bool lower_all_io_to_elements
;
2848 /* Indicates that the driver only has zero-based vertex id */
2849 bool vertex_id_zero_based
;
2852 * If enabled, gl_BaseVertex will be lowered as:
2853 * is_indexed_draw (~0/0) & firstvertex
2855 bool lower_base_vertex
;
2858 * If enabled, gl_HelperInvocation will be lowered as:
2860 * !((1 << sample_id) & sample_mask_in))
2862 * This depends on some possibly hw implementation details, which may
2863 * not be true for all hw. In particular that the FS is only executed
2864 * for covered samples or for helper invocations. So, do not blindly
2865 * enable this option.
2867 * Note: See also issue #22 in ARB_shader_image_load_store
2869 bool lower_helper_invocation
;
2872 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
2874 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
2875 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
2877 bool optimize_sample_mask_in
;
2879 bool lower_cs_local_index_from_id
;
2880 bool lower_cs_local_id_from_index
;
2882 bool lower_device_index_to_zero
;
2884 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
2885 bool lower_wpos_pntc
;
2888 * Set if nir_op_[iu]hadd and nir_op_[iu]rhadd instructions should be
2889 * lowered to simple arithmetic.
2891 * If this flag is set, the lowering will be applied to all bit-sizes of
2892 * these instructions.
2894 * \sa ::lower_hadd64
2899 * Set if only 64-bit nir_op_[iu]hadd and nir_op_[iu]rhadd instructions
2900 * should be lowered to simple arithmetic.
2902 * If this flag is set, the lowering will be applied to only 64-bit
2903 * versions of these instructions.
2910 * Set if nir_op_add_sat and nir_op_usub_sat should be lowered to simple
2913 * If this flag is set, the lowering will be applied to all bit-sizes of
2914 * these instructions.
2916 * \sa ::lower_usub_sat64
2921 * Set if only 64-bit nir_op_usub_sat should be lowered to simple
2924 * \sa ::lower_add_sat
2926 bool lower_usub_sat64
;
2929 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
2930 * for IO purposes and would prefer loads/stores be vectorized.
2933 bool lower_to_scalar
;
2936 * Should the linker unify inputs_read/outputs_written between adjacent
2937 * shader stages which are linked into a single program?
2939 bool unify_interfaces
;
2942 * Should nir_lower_io() create load_interpolated_input intrinsics?
2944 * If not, it generates regular load_input intrinsics and interpolation
2945 * information must be inferred from the list of input nir_variables.
2947 bool use_interpolated_input_intrinsics
;
2949 /* Lowers when 32x32->64 bit multiplication is not supported */
2950 bool lower_mul_2x32_64
;
2952 /* Lowers when rotate instruction is not supported */
2956 * Backend supports imul24, and would like to use it (when possible)
2957 * for address/offset calculation. If true, driver should call
2958 * nir_lower_amul(). (If not set, amul will automatically be lowered
2963 /* Whether to generate only scoped_memory_barrier intrinsics instead of the
2964 * set of memory barrier intrinsics based on GLSL.
2966 bool use_scoped_memory_barrier
;
2969 * Is this the Intel vec4 backend?
2971 * Used to inhibit algebraic optimizations that are known to be harmful on
2972 * the Intel vec4 backend. This is generally applicable to any
2973 * optimization that might cause more immediate values to be used in
2974 * 3-source (e.g., ffma and flrp) instructions.
2978 unsigned max_unroll_iterations
;
2980 nir_lower_int64_options lower_int64_options
;
2981 nir_lower_doubles_options lower_doubles_options
;
2982 } nir_shader_compiler_options
;
2984 typedef struct nir_shader
{
2985 /** list of uniforms (nir_variable) */
2986 struct exec_list uniforms
;
2988 /** list of inputs (nir_variable) */
2989 struct exec_list inputs
;
2991 /** list of outputs (nir_variable) */
2992 struct exec_list outputs
;
2994 /** list of shared compute variables (nir_variable) */
2995 struct exec_list shared
;
2997 /** Set of driver-specific options for the shader.
2999 * The memory for the options is expected to be kept in a single static
3000 * copy by the driver.
3002 const struct nir_shader_compiler_options
*options
;
3004 /** Various bits of compile-time information about a given shader */
3005 struct shader_info info
;
3007 /** list of global variables in the shader (nir_variable) */
3008 struct exec_list globals
;
3010 /** list of system value variables in the shader (nir_variable) */
3011 struct exec_list system_values
;
3013 struct exec_list functions
; /** < list of nir_function */
3016 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
3019 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
3021 /** Size in bytes of required scratch space */
3022 unsigned scratch_size
;
3024 /** Constant data associated with this shader.
3026 * Constant data is loaded through load_constant intrinsics. See also
3027 * nir_opt_large_constants.
3029 void *constant_data
;
3030 unsigned constant_data_size
;
3033 #define nir_foreach_function(func, shader) \
3034 foreach_list_typed(nir_function, func, node, &(shader)->functions)
3036 static inline nir_function_impl
*
3037 nir_shader_get_entrypoint(nir_shader
*shader
)
3039 nir_function
*func
= NULL
;
3041 nir_foreach_function(function
, shader
) {
3042 assert(func
== NULL
);
3043 if (function
->is_entrypoint
) {
3054 assert(func
->num_params
== 0);
3059 nir_shader
*nir_shader_create(void *mem_ctx
,
3060 gl_shader_stage stage
,
3061 const nir_shader_compiler_options
*options
,
3064 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
3066 void nir_reg_remove(nir_register
*reg
);
3068 /** Adds a variable to the appropriate list in nir_shader */
3069 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
3072 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
3074 assert(var
->data
.mode
== nir_var_function_temp
);
3075 exec_list_push_tail(&impl
->locals
, &var
->node
);
3078 /** creates a variable, sets a few defaults, and adds it to the list */
3079 nir_variable
*nir_variable_create(nir_shader
*shader
,
3080 nir_variable_mode mode
,
3081 const struct glsl_type
*type
,
3083 /** creates a local variable and adds it to the list */
3084 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
3085 const struct glsl_type
*type
,
3088 /** creates a function and adds it to the shader's list of functions */
3089 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
3091 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
3092 /** creates a function_impl that isn't tied to any particular function */
3093 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
3095 nir_block
*nir_block_create(nir_shader
*shader
);
3096 nir_if
*nir_if_create(nir_shader
*shader
);
3097 nir_loop
*nir_loop_create(nir_shader
*shader
);
3099 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
3101 /** requests that the given pieces of metadata be generated */
3102 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
3103 /** dirties all but the preserved metadata */
3104 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
3106 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
3107 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
3109 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
3110 nir_deref_type deref_type
);
3112 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
3114 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
3115 unsigned num_components
,
3118 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
3119 nir_intrinsic_op op
);
3121 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
3122 nir_function
*callee
);
3124 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
3126 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
3128 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
3130 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
3131 unsigned num_components
,
3134 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
3137 * NIR Cursors and Instruction Insertion API
3140 * A tiny struct representing a point to insert/extract instructions or
3141 * control flow nodes. Helps reduce the combinatorial explosion of possible
3142 * points to insert/extract.
3144 * \sa nir_control_flow.h
3147 nir_cursor_before_block
,
3148 nir_cursor_after_block
,
3149 nir_cursor_before_instr
,
3150 nir_cursor_after_instr
,
3151 } nir_cursor_option
;
3154 nir_cursor_option option
;
3161 static inline nir_block
*
3162 nir_cursor_current_block(nir_cursor cursor
)
3164 if (cursor
.option
== nir_cursor_before_instr
||
3165 cursor
.option
== nir_cursor_after_instr
) {
3166 return cursor
.instr
->block
;
3168 return cursor
.block
;
3172 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
3174 static inline nir_cursor
3175 nir_before_block(nir_block
*block
)
3178 cursor
.option
= nir_cursor_before_block
;
3179 cursor
.block
= block
;
3183 static inline nir_cursor
3184 nir_after_block(nir_block
*block
)
3187 cursor
.option
= nir_cursor_after_block
;
3188 cursor
.block
= block
;
3192 static inline nir_cursor
3193 nir_before_instr(nir_instr
*instr
)
3196 cursor
.option
= nir_cursor_before_instr
;
3197 cursor
.instr
= instr
;
3201 static inline nir_cursor
3202 nir_after_instr(nir_instr
*instr
)
3205 cursor
.option
= nir_cursor_after_instr
;
3206 cursor
.instr
= instr
;
3210 static inline nir_cursor
3211 nir_after_block_before_jump(nir_block
*block
)
3213 nir_instr
*last_instr
= nir_block_last_instr(block
);
3214 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
3215 return nir_before_instr(last_instr
);
3217 return nir_after_block(block
);
3221 static inline nir_cursor
3222 nir_before_src(nir_src
*src
, bool is_if_condition
)
3224 if (is_if_condition
) {
3225 nir_block
*prev_block
=
3226 nir_cf_node_as_block(nir_cf_node_prev(&src
->parent_if
->cf_node
));
3227 assert(!nir_block_ends_in_jump(prev_block
));
3228 return nir_after_block(prev_block
);
3229 } else if (src
->parent_instr
->type
== nir_instr_type_phi
) {
3231 nir_phi_instr
*cond_phi
= nir_instr_as_phi(src
->parent_instr
);
3233 nir_foreach_phi_src(phi_src
, cond_phi
) {
3234 if (phi_src
->src
.ssa
== src
->ssa
) {
3241 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
3242 * to have a more specific name.
3244 nir_phi_src
*phi_src
= LIST_ENTRY(nir_phi_src
, src
, src
);
3245 return nir_after_block_before_jump(phi_src
->pred
);
3247 return nir_before_instr(src
->parent_instr
);
3251 static inline nir_cursor
3252 nir_before_cf_node(nir_cf_node
*node
)
3254 if (node
->type
== nir_cf_node_block
)
3255 return nir_before_block(nir_cf_node_as_block(node
));
3257 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
3260 static inline nir_cursor
3261 nir_after_cf_node(nir_cf_node
*node
)
3263 if (node
->type
== nir_cf_node_block
)
3264 return nir_after_block(nir_cf_node_as_block(node
));
3266 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
3269 static inline nir_cursor
3270 nir_after_phis(nir_block
*block
)
3272 nir_foreach_instr(instr
, block
) {
3273 if (instr
->type
!= nir_instr_type_phi
)
3274 return nir_before_instr(instr
);
3276 return nir_after_block(block
);
3279 static inline nir_cursor
3280 nir_after_cf_node_and_phis(nir_cf_node
*node
)
3282 if (node
->type
== nir_cf_node_block
)
3283 return nir_after_block(nir_cf_node_as_block(node
));
3285 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
3287 return nir_after_phis(block
);
3290 static inline nir_cursor
3291 nir_before_cf_list(struct exec_list
*cf_list
)
3293 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
3294 exec_list_get_head(cf_list
), node
);
3295 return nir_before_cf_node(first_node
);
3298 static inline nir_cursor
3299 nir_after_cf_list(struct exec_list
*cf_list
)
3301 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
3302 exec_list_get_tail(cf_list
), node
);
3303 return nir_after_cf_node(last_node
);
3307 * Insert a NIR instruction at the given cursor.
3309 * Note: This does not update the cursor.
3311 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
3314 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
3316 nir_instr_insert(nir_before_instr(instr
), before
);
3320 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
3322 nir_instr_insert(nir_after_instr(instr
), after
);
3326 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
3328 nir_instr_insert(nir_before_block(block
), before
);
3332 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
3334 nir_instr_insert(nir_after_block(block
), after
);
3338 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
3340 nir_instr_insert(nir_before_cf_node(node
), before
);
3344 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
3346 nir_instr_insert(nir_after_cf_node(node
), after
);
3350 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
3352 nir_instr_insert(nir_before_cf_list(list
), before
);
3356 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
3358 nir_instr_insert(nir_after_cf_list(list
), after
);
3361 void nir_instr_remove_v(nir_instr
*instr
);
3363 static inline nir_cursor
3364 nir_instr_remove(nir_instr
*instr
)
3367 nir_instr
*prev
= nir_instr_prev(instr
);
3369 cursor
= nir_after_instr(prev
);
3371 cursor
= nir_before_block(instr
->block
);
3373 nir_instr_remove_v(instr
);
3379 nir_ssa_def
*nir_instr_ssa_def(nir_instr
*instr
);
3381 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
3382 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
3383 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
3384 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
3386 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
3387 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
3389 nir_const_value
*nir_src_as_const_value(nir_src src
);
3391 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
3392 static inline c_type * \
3393 nir_src_as_ ## name (nir_src src) \
3395 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
3396 ? cast_macro(src.ssa->parent_instr) : NULL; \
3399 NIR_SRC_AS_(alu_instr
, nir_alu_instr
, nir_instr_type_alu
, nir_instr_as_alu
)
3400 NIR_SRC_AS_(intrinsic
, nir_intrinsic_instr
,
3401 nir_instr_type_intrinsic
, nir_instr_as_intrinsic
)
3402 NIR_SRC_AS_(deref
, nir_deref_instr
, nir_instr_type_deref
, nir_instr_as_deref
)
3404 bool nir_src_is_dynamically_uniform(nir_src src
);
3405 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
3406 bool nir_instrs_equal(const nir_instr
*instr1
, const nir_instr
*instr2
);
3407 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
3408 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
3409 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
3410 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
3413 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
3414 unsigned num_components
, unsigned bit_size
,
3416 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
3417 unsigned num_components
, unsigned bit_size
,
3420 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
3421 const struct glsl_type
*type
,
3424 assert(glsl_type_is_vector_or_scalar(type
));
3425 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
3426 glsl_get_bit_size(type
), name
);
3428 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
3429 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
3430 nir_instr
*after_me
);
3432 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
3435 * finds the next basic block in source-code order, returns NULL if there is
3439 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
3441 /* Performs the opposite of nir_block_cf_tree_next() */
3443 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
3445 /* Gets the first block in a CF node in source-code order */
3447 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
3449 /* Gets the last block in a CF node in source-code order */
3451 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
3453 /* Gets the next block after a CF node in source-code order */
3455 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
3457 /* Macros for loops that visit blocks in source-code order */
3459 #define nir_foreach_block(block, impl) \
3460 for (nir_block *block = nir_start_block(impl); block != NULL; \
3461 block = nir_block_cf_tree_next(block))
3463 #define nir_foreach_block_safe(block, impl) \
3464 for (nir_block *block = nir_start_block(impl), \
3465 *next = nir_block_cf_tree_next(block); \
3467 block = next, next = nir_block_cf_tree_next(block))
3469 #define nir_foreach_block_reverse(block, impl) \
3470 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
3471 block = nir_block_cf_tree_prev(block))
3473 #define nir_foreach_block_reverse_safe(block, impl) \
3474 for (nir_block *block = nir_impl_last_block(impl), \
3475 *prev = nir_block_cf_tree_prev(block); \
3477 block = prev, prev = nir_block_cf_tree_prev(block))
3479 #define nir_foreach_block_in_cf_node(block, node) \
3480 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
3481 block != nir_cf_node_cf_tree_next(node); \
3482 block = nir_block_cf_tree_next(block))
3484 /* If the following CF node is an if, this function returns that if.
3485 * Otherwise, it returns NULL.
3487 nir_if
*nir_block_get_following_if(nir_block
*block
);
3489 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
3491 void nir_index_local_regs(nir_function_impl
*impl
);
3492 void nir_index_ssa_defs(nir_function_impl
*impl
);
3493 unsigned nir_index_instrs(nir_function_impl
*impl
);
3495 void nir_index_blocks(nir_function_impl
*impl
);
3497 void nir_index_vars(nir_shader
*shader
, nir_function_impl
*impl
, nir_variable_mode modes
);
3499 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
3500 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
3501 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
3502 void nir_print_deref(const nir_deref_instr
*deref
, FILE *fp
);
3504 /** Shallow clone of a single ALU instruction. */
3505 nir_alu_instr
*nir_alu_instr_clone(nir_shader
*s
, const nir_alu_instr
*orig
);
3507 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
3508 nir_function_impl
*nir_function_impl_clone(nir_shader
*shader
,
3509 const nir_function_impl
*fi
);
3510 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
3511 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
3513 void nir_shader_replace(nir_shader
*dest
, nir_shader
*src
);
3515 void nir_shader_serialize_deserialize(nir_shader
*s
);
3518 void nir_validate_shader(nir_shader
*shader
, const char *when
);
3519 void nir_metadata_set_validation_flag(nir_shader
*shader
);
3520 void nir_metadata_check_validation_flag(nir_shader
*shader
);
3523 should_skip_nir(const char *name
)
3525 static const char *list
= NULL
;
3527 /* Comma separated list of names to skip. */
3528 list
= getenv("NIR_SKIP");
3536 return comma_separated_list_contains(list
, name
);
3540 should_clone_nir(void)
3542 static int should_clone
= -1;
3543 if (should_clone
< 0)
3544 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
3546 return should_clone
;
3550 should_serialize_deserialize_nir(void)
3552 static int test_serialize
= -1;
3553 if (test_serialize
< 0)
3554 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
3556 return test_serialize
;
3560 should_print_nir(void)
3562 static int should_print
= -1;
3563 if (should_print
< 0)
3564 should_print
= env_var_as_boolean("NIR_PRINT", false);
3566 return should_print
;
3569 static inline void nir_validate_shader(nir_shader
*shader
, const char *when
) { (void) shader
; (void)when
; }
3570 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
3571 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
3572 static inline bool should_skip_nir(UNUSED
const char *pass_name
) { return false; }
3573 static inline bool should_clone_nir(void) { return false; }
3574 static inline bool should_serialize_deserialize_nir(void) { return false; }
3575 static inline bool should_print_nir(void) { return false; }
3578 #define _PASS(pass, nir, do_pass) do { \
3579 if (should_skip_nir(#pass)) { \
3580 printf("skipping %s\n", #pass); \
3584 nir_validate_shader(nir, "after " #pass); \
3585 if (should_clone_nir()) { \
3586 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3587 nir_shader_replace(nir, clone); \
3589 if (should_serialize_deserialize_nir()) { \
3590 nir_shader_serialize_deserialize(nir); \
3594 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3595 nir_metadata_set_validation_flag(nir); \
3596 if (should_print_nir()) \
3597 printf("%s\n", #pass); \
3598 if (pass(nir, ##__VA_ARGS__)) { \
3600 if (should_print_nir()) \
3601 nir_print_shader(nir, stdout); \
3602 nir_metadata_check_validation_flag(nir); \
3606 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3607 if (should_print_nir()) \
3608 printf("%s\n", #pass); \
3609 pass(nir, ##__VA_ARGS__); \
3610 if (should_print_nir()) \
3611 nir_print_shader(nir, stdout); \
3614 #define NIR_SKIP(name) should_skip_nir(#name)
3616 /** An instruction filtering callback
3618 * Returns true if the instruction should be processed and false otherwise.
3620 typedef bool (*nir_instr_filter_cb
)(const nir_instr
*, const void *);
3622 /** A simple instruction lowering callback
3624 * Many instruction lowering passes can be written as a simple function which
3625 * takes an instruction as its input and returns a sequence of instructions
3626 * that implement the consumed instruction. This function type represents
3627 * such a lowering function. When called, a function with this prototype
3628 * should either return NULL indicating that no lowering needs to be done or
3629 * emit a sequence of instructions using the provided builder (whose cursor
3630 * will already be placed after the instruction to be lowered) and return the
3631 * resulting nir_ssa_def.
3633 typedef nir_ssa_def
*(*nir_lower_instr_cb
)(struct nir_builder
*,
3634 nir_instr
*, void *);
3637 * Special return value for nir_lower_instr_cb when some progress occurred
3638 * (like changing an input to the instr) that didn't result in a replacement
3639 * SSA def being generated.
3641 #define NIR_LOWER_INSTR_PROGRESS ((nir_ssa_def *)(uintptr_t)1)
3643 /** Iterate over all the instructions in a nir_function_impl and lower them
3644 * using the provided callbacks
3646 * This function implements the guts of a standard lowering pass for you. It
3647 * iterates over all of the instructions in a nir_function_impl and calls the
3648 * filter callback on each one. If the filter callback returns true, it then
3649 * calls the lowering call back on the instruction. (Splitting it this way
3650 * allows us to avoid some save/restore work for instructions we know won't be
3651 * lowered.) If the instruction is dead after the lowering is complete, it
3652 * will be removed. If new instructions are added, the lowering callback will
3653 * also be called on them in case multiple lowerings are required.
3655 * The metadata for the nir_function_impl will also be updated. If any blocks
3656 * are added (they cannot be removed), dominance and block indices will be
3659 bool nir_function_impl_lower_instructions(nir_function_impl
*impl
,
3660 nir_instr_filter_cb filter
,
3661 nir_lower_instr_cb lower
,
3663 bool nir_shader_lower_instructions(nir_shader
*shader
,
3664 nir_instr_filter_cb filter
,
3665 nir_lower_instr_cb lower
,
3668 void nir_calc_dominance_impl(nir_function_impl
*impl
);
3669 void nir_calc_dominance(nir_shader
*shader
);
3671 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
3672 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
3673 bool nir_block_is_unreachable(nir_block
*block
);
3675 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
3676 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
3678 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
3679 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
3681 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
3682 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
3684 int nir_gs_count_vertices(const nir_shader
*shader
);
3686 bool nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3687 bool nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3688 bool nir_split_var_copies(nir_shader
*shader
);
3689 bool nir_split_per_member_structs(nir_shader
*shader
);
3690 bool nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
);
3692 bool nir_lower_returns_impl(nir_function_impl
*impl
);
3693 bool nir_lower_returns(nir_shader
*shader
);
3695 void nir_inline_function_impl(struct nir_builder
*b
,
3696 const nir_function_impl
*impl
,
3697 nir_ssa_def
**params
);
3698 bool nir_inline_functions(nir_shader
*shader
);
3700 bool nir_propagate_invariant(nir_shader
*shader
);
3702 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
3703 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
3704 nir_intrinsic_instr
*copy
);
3705 bool nir_lower_var_copies(nir_shader
*shader
);
3707 void nir_fixup_deref_modes(nir_shader
*shader
);
3709 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
3712 nir_lower_direct_array_deref_of_vec_load
= (1 << 0),
3713 nir_lower_indirect_array_deref_of_vec_load
= (1 << 1),
3714 nir_lower_direct_array_deref_of_vec_store
= (1 << 2),
3715 nir_lower_indirect_array_deref_of_vec_store
= (1 << 3),
3716 } nir_lower_array_deref_of_vec_options
;
3718 bool nir_lower_array_deref_of_vec(nir_shader
*shader
, nir_variable_mode modes
,
3719 nir_lower_array_deref_of_vec_options options
);
3721 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
3723 bool nir_lower_locals_to_regs(nir_shader
*shader
);
3725 void nir_lower_io_to_temporaries(nir_shader
*shader
,
3726 nir_function_impl
*entrypoint
,
3727 bool outputs
, bool inputs
);
3729 bool nir_lower_vars_to_scratch(nir_shader
*shader
,
3730 nir_variable_mode modes
,
3732 glsl_type_size_align_func size_align
);
3734 void nir_lower_clip_halfz(nir_shader
*shader
);
3736 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
3738 void nir_gather_ssa_types(nir_function_impl
*impl
,
3739 BITSET_WORD
*float_types
,
3740 BITSET_WORD
*int_types
);
3742 void nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
3743 int (*type_size
)(const struct glsl_type
*, bool));
3745 /* Some helpers to do very simple linking */
3746 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3747 bool nir_remove_unused_io_vars(nir_shader
*shader
, struct exec_list
*var_list
,
3748 uint64_t *used_by_other_stage
,
3749 uint64_t *used_by_other_stage_patches
);
3750 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
3751 bool default_to_smooth_interp
);
3752 void nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3753 bool nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3755 bool nir_lower_amul(nir_shader
*shader
,
3756 int (*type_size
)(const struct glsl_type
*, bool));
3758 void nir_assign_io_var_locations(struct exec_list
*var_list
,
3760 gl_shader_stage stage
);
3763 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
3764 * to 32-bit operations. This is only valid for nir_var_shader_in/out
3767 nir_lower_io_lower_64bit_to_32
= (1 << 0),
3769 /* If set, this forces all non-flat fragment shader inputs to be
3770 * interpolated as if with the "sample" qualifier. This requires
3771 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
3773 nir_lower_io_force_sample_interpolation
= (1 << 1),
3774 } nir_lower_io_options
;
3775 bool nir_lower_io(nir_shader
*shader
,
3776 nir_variable_mode modes
,
3777 int (*type_size
)(const struct glsl_type
*, bool),
3778 nir_lower_io_options
);
3780 bool nir_io_add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
);
3783 nir_lower_vars_to_explicit_types(nir_shader
*shader
,
3784 nir_variable_mode modes
,
3785 glsl_type_size_align_func type_info
);
3789 * An address format which is a simple 32-bit global GPU address.
3791 nir_address_format_32bit_global
,
3794 * An address format which is a simple 64-bit global GPU address.
3796 nir_address_format_64bit_global
,
3799 * An address format which is a bounds-checked 64-bit global GPU address.
3801 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
3802 * address stored with the low bits in .x and high bits in .y, .z is a
3803 * size, and .w is an offset. When the final I/O operation is lowered, .w
3804 * is checked against .z and the operation is predicated on the result.
3806 nir_address_format_64bit_bounded_global
,
3809 * An address format which is comprised of a vec2 where the first
3810 * component is a buffer index and the second is an offset.
3812 nir_address_format_32bit_index_offset
,
3815 * An address format which is a simple 32-bit offset.
3817 nir_address_format_32bit_offset
,
3820 * An address format representing a purely logical addressing model. In
3821 * this model, all deref chains must be complete from the dereference
3822 * operation to the variable. Cast derefs are not allowed. These
3823 * addresses will be 32-bit scalars but the format is immaterial because
3824 * you can always chase the chain.
3826 nir_address_format_logical
,
3827 } nir_address_format
;
3829 static inline unsigned
3830 nir_address_format_bit_size(nir_address_format addr_format
)
3832 switch (addr_format
) {
3833 case nir_address_format_32bit_global
: return 32;
3834 case nir_address_format_64bit_global
: return 64;
3835 case nir_address_format_64bit_bounded_global
: return 32;
3836 case nir_address_format_32bit_index_offset
: return 32;
3837 case nir_address_format_32bit_offset
: return 32;
3838 case nir_address_format_logical
: return 32;
3840 unreachable("Invalid address format");
3843 static inline unsigned
3844 nir_address_format_num_components(nir_address_format addr_format
)
3846 switch (addr_format
) {
3847 case nir_address_format_32bit_global
: return 1;
3848 case nir_address_format_64bit_global
: return 1;
3849 case nir_address_format_64bit_bounded_global
: return 4;
3850 case nir_address_format_32bit_index_offset
: return 2;
3851 case nir_address_format_32bit_offset
: return 1;
3852 case nir_address_format_logical
: return 1;
3854 unreachable("Invalid address format");
3857 static inline const struct glsl_type
*
3858 nir_address_format_to_glsl_type(nir_address_format addr_format
)
3860 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
3861 assert(bit_size
== 32 || bit_size
== 64);
3862 return glsl_vector_type(bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
,
3863 nir_address_format_num_components(addr_format
));
3866 const nir_const_value
*nir_address_format_null_value(nir_address_format addr_format
);
3868 nir_ssa_def
*nir_build_addr_ieq(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3869 nir_address_format addr_format
);
3871 nir_ssa_def
*nir_build_addr_isub(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3872 nir_address_format addr_format
);
3874 nir_ssa_def
* nir_explicit_io_address_from_deref(struct nir_builder
*b
,
3875 nir_deref_instr
*deref
,
3876 nir_ssa_def
*base_addr
,
3877 nir_address_format addr_format
);
3878 void nir_lower_explicit_io_instr(struct nir_builder
*b
,
3879 nir_intrinsic_instr
*io_instr
,
3881 nir_address_format addr_format
);
3883 bool nir_lower_explicit_io(nir_shader
*shader
,
3884 nir_variable_mode modes
,
3885 nir_address_format
);
3887 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
3888 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
3890 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
3892 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
3893 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
3894 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
3896 bool nir_remove_dead_derefs(nir_shader
*shader
);
3897 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
3898 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
);
3899 bool nir_lower_variable_initializers(nir_shader
*shader
,
3900 nir_variable_mode modes
);
3902 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
3903 bool nir_lower_vec_to_movs(nir_shader
*shader
);
3904 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
3906 const gl_state_index16
*alpha_ref_state_tokens
);
3907 bool nir_lower_alu(nir_shader
*shader
);
3909 bool nir_lower_flrp(nir_shader
*shader
, unsigned lowering_mask
,
3910 bool always_precise
, bool have_ffma
);
3912 bool nir_lower_alu_to_scalar(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
3913 bool nir_lower_bool_to_bitsize(nir_shader
*shader
);
3914 bool nir_lower_bool_to_float(nir_shader
*shader
);
3915 bool nir_lower_bool_to_int32(nir_shader
*shader
);
3916 bool nir_lower_int_to_float(nir_shader
*shader
);
3917 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
3918 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
3919 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
3920 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
3921 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
3923 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
3924 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
3925 bool nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode mask
);
3927 void nir_lower_fragcoord_wtrans(nir_shader
*shader
);
3928 void nir_lower_viewport_transform(nir_shader
*shader
);
3929 bool nir_lower_uniforms_to_ubo(nir_shader
*shader
, int multiplier
);
3931 typedef struct nir_lower_subgroups_options
{
3932 uint8_t subgroup_size
;
3933 uint8_t ballot_bit_size
;
3934 bool lower_to_scalar
:1;
3935 bool lower_vote_trivial
:1;
3936 bool lower_vote_eq_to_ballot
:1;
3937 bool lower_subgroup_masks
:1;
3938 bool lower_shuffle
:1;
3939 bool lower_shuffle_to_32bit
:1;
3941 bool lower_quad_broadcast_dynamic
:1;
3942 bool lower_quad_broadcast_dynamic_to_const
:1;
3943 } nir_lower_subgroups_options
;
3945 bool nir_lower_subgroups(nir_shader
*shader
,
3946 const nir_lower_subgroups_options
*options
);
3948 bool nir_lower_system_values(nir_shader
*shader
);
3950 enum PACKED nir_lower_tex_packing
{
3951 nir_lower_tex_packing_none
= 0,
3952 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
3953 * or unsigned ints based on the sampler type
3955 nir_lower_tex_packing_16
,
3956 /* The sampler returns 1 32-bit word of 4x8 unorm */
3957 nir_lower_tex_packing_8
,
3960 typedef struct nir_lower_tex_options
{
3962 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
3963 * sampler types a texture projector is lowered.
3968 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
3970 bool lower_txf_offset
;
3973 * If true, lower away nir_tex_src_offset for all rect textures.
3975 bool lower_rect_offset
;
3978 * If true, lower rect textures to 2D, using txs to fetch the
3979 * texture dimensions and dividing the texture coords by the
3980 * texture dims to normalize.
3985 * If true, convert yuv to rgb.
3987 unsigned lower_y_uv_external
;
3988 unsigned lower_y_u_v_external
;
3989 unsigned lower_yx_xuxv_external
;
3990 unsigned lower_xy_uxvx_external
;
3991 unsigned lower_ayuv_external
;
3992 unsigned lower_xyuv_external
;
3995 * To emulate certain texture wrap modes, this can be used
3996 * to saturate the specified tex coord to [0.0, 1.0]. The
3997 * bits are according to sampler #, ie. if, for example:
3999 * (conf->saturate_s & (1 << n))
4001 * is true, then the s coord for sampler n is saturated.
4003 * Note that clamping must happen *after* projector lowering
4004 * so any projected texture sample instruction with a clamped
4005 * coordinate gets automatically lowered, regardless of the
4006 * 'lower_txp' setting.
4008 unsigned saturate_s
;
4009 unsigned saturate_t
;
4010 unsigned saturate_r
;
4012 /* Bitmask of textures that need swizzling.
4014 * If (swizzle_result & (1 << texture_index)), then the swizzle in
4015 * swizzles[texture_index] is applied to the result of the texturing
4018 unsigned swizzle_result
;
4020 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
4021 * while 4 and 5 represent 0 and 1 respectively.
4023 uint8_t swizzles
[32][4];
4025 /* Can be used to scale sampled values in range required by the format. */
4026 float scale_factors
[32];
4029 * Bitmap of textures that need srgb to linear conversion. If
4030 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
4031 * of the texture are lowered to linear.
4033 unsigned lower_srgb
;
4036 * If true, lower nir_texop_tex on shaders that doesn't support implicit
4037 * LODs to nir_texop_txl.
4039 bool lower_tex_without_implicit_lod
;
4042 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
4044 bool lower_txd_cube_map
;
4047 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
4052 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
4053 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
4054 * with lower_txd_cube_map.
4056 bool lower_txd_shadow
;
4059 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
4060 * Implies lower_txd_cube_map and lower_txd_shadow.
4065 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
4066 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
4068 bool lower_txb_shadow_clamp
;
4071 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
4072 * with nir_texop_txl. This includes cube maps.
4074 bool lower_txd_shadow_clamp
;
4077 * If true, lower nir_texop_txd on when it uses both offset and min_lod
4078 * with nir_texop_txl. This includes cube maps.
4080 bool lower_txd_offset_clamp
;
4083 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4084 * sampler is bindless.
4086 bool lower_txd_clamp_bindless_sampler
;
4089 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4090 * sampler index is not statically determinable to be less than 16.
4092 bool lower_txd_clamp_if_sampler_index_not_lt_16
;
4095 * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
4096 * 0-lod followed by a nir_ishr.
4101 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
4102 * mixed-up tg4 locations.
4104 bool lower_tg4_broadcom_swizzle
;
4107 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
4109 bool lower_tg4_offsets
;
4111 enum nir_lower_tex_packing lower_tex_packing
[32];
4112 } nir_lower_tex_options
;
4114 bool nir_lower_tex(nir_shader
*shader
,
4115 const nir_lower_tex_options
*options
);
4117 enum nir_lower_non_uniform_access_type
{
4118 nir_lower_non_uniform_ubo_access
= (1 << 0),
4119 nir_lower_non_uniform_ssbo_access
= (1 << 1),
4120 nir_lower_non_uniform_texture_access
= (1 << 2),
4121 nir_lower_non_uniform_image_access
= (1 << 3),
4124 bool nir_lower_non_uniform_access(nir_shader
*shader
,
4125 enum nir_lower_non_uniform_access_type
);
4127 enum nir_lower_idiv_path
{
4128 /* This path is based on NV50LegalizeSSA::handleDIV(). It is the faster of
4129 * the two but it is not exact in some cases (for example, 1091317713u /
4130 * 1034u gives 5209173 instead of 1055432) */
4131 nir_lower_idiv_fast
,
4132 /* This path is based on AMDGPUTargetLowering::LowerUDIVREM() and
4133 * AMDGPUTargetLowering::LowerSDIVREM(). It requires more instructions than
4134 * the nv50 path and many of them are integer multiplications, so it is
4135 * probably slower. It should always return the correct result, though. */
4136 nir_lower_idiv_precise
,
4139 bool nir_lower_idiv(nir_shader
*shader
, enum nir_lower_idiv_path path
);
4141 bool nir_lower_input_attachments(nir_shader
*shader
, bool use_fragcoord_sysval
);
4143 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
,
4145 bool use_clipdist_array
,
4146 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4147 bool nir_lower_clip_gs(nir_shader
*shader
, unsigned ucp_enables
,
4148 bool use_clipdist_array
,
4149 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4150 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
,
4151 bool use_clipdist_array
);
4152 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
4154 void nir_lower_point_size_mov(nir_shader
*shader
,
4155 const gl_state_index16
*pointsize_state_tokens
);
4157 bool nir_lower_frexp(nir_shader
*nir
);
4159 void nir_lower_two_sided_color(nir_shader
*shader
);
4161 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
4163 bool nir_lower_flatshade(nir_shader
*shader
);
4165 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
4166 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
4167 const gl_state_index16
*uniform_state_tokens
);
4169 typedef struct nir_lower_wpos_ytransform_options
{
4170 gl_state_index16 state_tokens
[STATE_LENGTH
];
4171 bool fs_coord_origin_upper_left
:1;
4172 bool fs_coord_origin_lower_left
:1;
4173 bool fs_coord_pixel_center_integer
:1;
4174 bool fs_coord_pixel_center_half_integer
:1;
4175 } nir_lower_wpos_ytransform_options
;
4177 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
4178 const nir_lower_wpos_ytransform_options
*options
);
4179 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
4181 bool nir_lower_fb_read(nir_shader
*shader
);
4183 typedef struct nir_lower_drawpixels_options
{
4184 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
4185 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
4186 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
4187 unsigned drawpix_sampler
;
4188 unsigned pixelmap_sampler
;
4190 bool scale_and_bias
:1;
4191 } nir_lower_drawpixels_options
;
4193 void nir_lower_drawpixels(nir_shader
*shader
,
4194 const nir_lower_drawpixels_options
*options
);
4196 typedef struct nir_lower_bitmap_options
{
4199 } nir_lower_bitmap_options
;
4201 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
4203 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
);
4206 nir_lower_int_source_mods
= 1 << 0,
4207 nir_lower_float_source_mods
= 1 << 1,
4208 nir_lower_triop_abs
= 1 << 2,
4209 nir_lower_all_source_mods
= (1 << 3) - 1
4210 } nir_lower_to_source_mods_flags
;
4213 bool nir_lower_to_source_mods(nir_shader
*shader
, nir_lower_to_source_mods_flags options
);
4215 bool nir_lower_gs_intrinsics(nir_shader
*shader
, bool per_stream
);
4217 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
4219 bool nir_lower_bit_size(nir_shader
*shader
,
4220 nir_lower_bit_size_callback callback
,
4221 void *callback_data
);
4223 nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode
);
4224 bool nir_lower_int64(nir_shader
*shader
, nir_lower_int64_options options
);
4226 nir_lower_doubles_options
nir_lower_doubles_op_to_options_mask(nir_op opcode
);
4227 bool nir_lower_doubles(nir_shader
*shader
, const nir_shader
*softfp64
,
4228 nir_lower_doubles_options options
);
4229 bool nir_lower_pack(nir_shader
*shader
);
4231 bool nir_lower_point_size(nir_shader
*shader
, float min
, float max
);
4234 nir_lower_interpolation_at_sample
= (1 << 1),
4235 nir_lower_interpolation_at_offset
= (1 << 2),
4236 nir_lower_interpolation_centroid
= (1 << 3),
4237 nir_lower_interpolation_pixel
= (1 << 4),
4238 nir_lower_interpolation_sample
= (1 << 5),
4239 } nir_lower_interpolation_options
;
4241 bool nir_lower_interpolation(nir_shader
*shader
,
4242 nir_lower_interpolation_options options
);
4244 bool nir_lower_discard_to_demote(nir_shader
*shader
);
4246 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
4248 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
4250 void nir_loop_analyze_impl(nir_function_impl
*impl
,
4251 nir_variable_mode indirect_mask
);
4253 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
4255 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
4256 bool nir_repair_ssa(nir_shader
*shader
);
4258 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
4259 bool nir_convert_to_lcssa(nir_shader
*shader
, bool skip_invariants
, bool skip_bool_invariants
);
4260 bool* nir_divergence_analysis(nir_shader
*shader
, nir_divergence_options options
);
4262 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
4263 * registers. If false, convert all values (even those not involved in a phi
4264 * node) to registers.
4266 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
4268 bool nir_lower_phis_to_regs_block(nir_block
*block
);
4269 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
4270 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
);
4272 bool nir_lower_samplers(nir_shader
*shader
);
4273 bool nir_lower_ssbo(nir_shader
*shader
);
4275 /* This is here for unit tests. */
4276 bool nir_opt_comparison_pre_impl(nir_function_impl
*impl
);
4278 bool nir_opt_comparison_pre(nir_shader
*shader
);
4280 bool nir_opt_access(nir_shader
*shader
);
4281 bool nir_opt_algebraic(nir_shader
*shader
);
4282 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
4283 bool nir_opt_algebraic_late(nir_shader
*shader
);
4284 bool nir_opt_algebraic_distribute_src_mods(nir_shader
*shader
);
4285 bool nir_opt_constant_folding(nir_shader
*shader
);
4287 /* Try to combine a and b into a. Return true if combination was possible,
4288 * which will result in b being removed by the pass. Return false if
4289 * combination wasn't possible.
4291 typedef bool (*nir_combine_memory_barrier_cb
)(
4292 nir_intrinsic_instr
*a
, nir_intrinsic_instr
*b
, void *data
);
4294 bool nir_opt_combine_memory_barriers(nir_shader
*shader
,
4295 nir_combine_memory_barrier_cb combine_cb
,
4298 bool nir_opt_combine_stores(nir_shader
*shader
, nir_variable_mode modes
);
4300 bool nir_copy_prop(nir_shader
*shader
);
4302 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
4304 bool nir_opt_cse(nir_shader
*shader
);
4306 bool nir_opt_dce(nir_shader
*shader
);
4308 bool nir_opt_dead_cf(nir_shader
*shader
);
4310 bool nir_opt_dead_write_vars(nir_shader
*shader
);
4312 bool nir_opt_deref_impl(nir_function_impl
*impl
);
4313 bool nir_opt_deref(nir_shader
*shader
);
4315 bool nir_opt_find_array_copies(nir_shader
*shader
);
4317 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
4319 bool nir_opt_idiv_const(nir_shader
*shader
, unsigned min_bit_size
);
4321 bool nir_opt_if(nir_shader
*shader
, bool aggressive_last_continue
);
4323 bool nir_opt_intrinsics(nir_shader
*shader
);
4325 bool nir_opt_large_constants(nir_shader
*shader
,
4326 glsl_type_size_align_func size_align
,
4327 unsigned threshold
);
4329 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
4332 nir_move_const_undef
= (1 << 0),
4333 nir_move_load_ubo
= (1 << 1),
4334 nir_move_load_input
= (1 << 2),
4335 nir_move_comparisons
= (1 << 3),
4336 nir_move_copies
= (1 << 4),
4339 bool nir_can_move_instr(nir_instr
*instr
, nir_move_options options
);
4341 bool nir_opt_sink(nir_shader
*shader
, nir_move_options options
);
4343 bool nir_opt_move(nir_shader
*shader
, nir_move_options options
);
4345 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
,
4346 bool indirect_load_ok
, bool expensive_alu_ok
);
4348 bool nir_opt_rematerialize_compares(nir_shader
*shader
);
4350 bool nir_opt_remove_phis(nir_shader
*shader
);
4351 bool nir_opt_remove_phis_block(nir_block
*block
);
4353 bool nir_opt_shrink_load(nir_shader
*shader
);
4355 bool nir_opt_trivial_continues(nir_shader
*shader
);
4357 bool nir_opt_undef(nir_shader
*shader
);
4359 bool nir_opt_vectorize(nir_shader
*shader
);
4361 bool nir_opt_conditional_discard(nir_shader
*shader
);
4363 typedef bool (*nir_should_vectorize_mem_func
)(unsigned align
, unsigned bit_size
,
4364 unsigned num_components
, unsigned high_offset
,
4365 nir_intrinsic_instr
*low
, nir_intrinsic_instr
*high
);
4367 bool nir_opt_load_store_vectorize(nir_shader
*shader
, nir_variable_mode modes
,
4368 nir_should_vectorize_mem_func callback
);
4370 void nir_schedule(nir_shader
*shader
, int threshold
);
4372 void nir_strip(nir_shader
*shader
);
4374 void nir_sweep(nir_shader
*shader
);
4376 void nir_remap_dual_slot_attributes(nir_shader
*shader
,
4377 uint64_t *dual_slot_inputs
);
4378 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
);
4380 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
4381 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);
4384 nir_variable_is_in_ubo(const nir_variable
*var
)
4386 return (var
->data
.mode
== nir_var_mem_ubo
&&
4387 var
->interface_type
!= NULL
);
4391 nir_variable_is_in_ssbo(const nir_variable
*var
)
4393 return (var
->data
.mode
== nir_var_mem_ssbo
&&
4394 var
->interface_type
!= NULL
);
4398 nir_variable_is_in_block(const nir_variable
*var
)
4400 return nir_variable_is_in_ubo(var
) || nir_variable_is_in_ssbo(var
);