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 * \brief Layout qualifier for gl_FragDepth.
459 * This is not equal to \c ir_depth_layout_none if and only if this
460 * variable is \c gl_FragDepth and a layout qualifier is specified.
462 nir_depth_layout depth_layout
:3;
465 * Vertex stream output identifier.
467 * For packed outputs, NIR_STREAM_PACKED is set and bits [2*i+1,2*i]
468 * indicate the stream of the i-th component.
473 * Access flags for memory variables (SSBO/global), image uniforms, and
474 * bindless images in uniforms/inputs/outputs.
476 enum gl_access_qualifier access
:8;
479 * Descriptor set binding for sampler or UBO.
481 unsigned descriptor_set
:5;
484 * output index for dual source blending.
489 * Initial binding point for a sampler or UBO.
491 * For array types, this represents the binding point for the first element.
496 * Storage location of the base of this variable
498 * The precise meaning of this field depends on the nature of the variable.
500 * - Vertex shader input: one of the values from \c gl_vert_attrib.
501 * - Vertex shader output: one of the values from \c gl_varying_slot.
502 * - Geometry shader input: one of the values from \c gl_varying_slot.
503 * - Geometry shader output: one of the values from \c gl_varying_slot.
504 * - Fragment shader input: one of the values from \c gl_varying_slot.
505 * - Fragment shader output: one of the values from \c gl_frag_result.
506 * - Uniforms: Per-stage uniform slot number for default uniform block.
507 * - Uniforms: Index within the uniform block definition for UBO members.
508 * - Non-UBO Uniforms: uniform slot number.
509 * - Other: This field is not currently used.
511 * If the variable is a uniform, shader input, or shader output, and the
512 * slot has not been assigned, the value will be -1.
517 * The actual location of the variable in the IR. Only valid for inputs,
518 * outputs, and uniforms (including samplers and images).
520 unsigned driver_location
;
523 * Location an atomic counter or transform feedback is stored at.
529 /** Image internal format if specified explicitly, otherwise PIPE_FORMAT_NONE. */
530 enum pipe_format format
;
535 * Transform feedback buffer.
540 * Transform feedback stride.
548 * Identifier for this variable generated by nir_index_vars() that is unique
549 * among other variables in the same exec_list.
553 /* Number of nir_variable_data members */
554 uint16_t num_members
;
557 * Built-in state that backs this uniform
559 * Once set at variable creation, \c state_slots must remain invariant.
560 * This is because, ideally, this array would be shared by all clones of
561 * this variable in the IR tree. In other words, we'd really like for it
562 * to be a fly-weight.
564 * If the variable is not a uniform, \c num_state_slots will be zero and
565 * \c state_slots will be \c NULL.
568 uint16_t num_state_slots
; /**< Number of state slots used */
569 nir_state_slot
*state_slots
; /**< State descriptors. */
573 * Constant expression assigned in the initializer of the variable
575 * This field should only be used temporarily by creators of NIR shaders
576 * and then lower_constant_initializers can be used to get rid of them.
577 * Most of the rest of NIR ignores this field or asserts that it's NULL.
579 nir_constant
*constant_initializer
;
582 * Global variable assigned in the initializer of the variable
583 * This field should only be used temporarily by creators of NIR shaders
584 * and then lower_constant_initializers can be used to get rid of them.
585 * Most of the rest of NIR ignores this field or asserts that it's NULL.
587 struct nir_variable
*pointer_initializer
;
590 * For variables that are in an interface block or are an instance of an
591 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
593 * \sa ir_variable::location
595 const struct glsl_type
*interface_type
;
598 * Description of per-member data for per-member struct variables
600 * This is used for variables which are actually an amalgamation of
601 * multiple entities such as a struct of built-in values or a struct of
602 * inputs each with their own layout specifier. This is only allowed on
603 * variables with a struct or array of array of struct type.
605 struct nir_variable_data
*members
;
608 #define nir_foreach_variable(var, var_list) \
609 foreach_list_typed(nir_variable, var, node, var_list)
611 #define nir_foreach_variable_safe(var, var_list) \
612 foreach_list_typed_safe(nir_variable, var, node, var_list)
615 nir_variable_is_global(const nir_variable
*var
)
617 return var
->data
.mode
!= nir_var_function_temp
;
620 typedef struct nir_register
{
621 struct exec_node node
;
623 unsigned num_components
; /** < number of vector components */
624 unsigned num_array_elems
; /** < size of array (0 for no array) */
626 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
629 /** generic register index. */
632 /** only for debug purposes, can be NULL */
635 /** set of nir_srcs where this register is used (read from) */
636 struct list_head uses
;
638 /** set of nir_dests where this register is defined (written to) */
639 struct list_head defs
;
641 /** set of nir_ifs where this register is used as a condition */
642 struct list_head if_uses
;
645 #define nir_foreach_register(reg, reg_list) \
646 foreach_list_typed(nir_register, reg, node, reg_list)
647 #define nir_foreach_register_safe(reg, reg_list) \
648 foreach_list_typed_safe(nir_register, reg, node, reg_list)
650 typedef enum PACKED
{
652 nir_instr_type_deref
,
655 nir_instr_type_intrinsic
,
656 nir_instr_type_load_const
,
658 nir_instr_type_ssa_undef
,
660 nir_instr_type_parallel_copy
,
663 typedef struct nir_instr
{
664 struct exec_node node
;
665 struct nir_block
*block
;
668 /* A temporary for optimization and analysis passes to use for storing
669 * flags. For instance, DCE uses this to store the "dead/live" info.
673 /** generic instruction index. */
677 static inline nir_instr
*
678 nir_instr_next(nir_instr
*instr
)
680 struct exec_node
*next
= exec_node_get_next(&instr
->node
);
681 if (exec_node_is_tail_sentinel(next
))
684 return exec_node_data(nir_instr
, next
, node
);
687 static inline nir_instr
*
688 nir_instr_prev(nir_instr
*instr
)
690 struct exec_node
*prev
= exec_node_get_prev(&instr
->node
);
691 if (exec_node_is_head_sentinel(prev
))
694 return exec_node_data(nir_instr
, prev
, node
);
698 nir_instr_is_first(const nir_instr
*instr
)
700 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr
->node
));
704 nir_instr_is_last(const nir_instr
*instr
)
706 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr
->node
));
709 typedef struct nir_ssa_def
{
710 /** for debugging only, can be NULL */
713 /** generic SSA definition index. */
716 /** Index into the live_in and live_out bitfields */
719 /** Instruction which produces this SSA value. */
720 nir_instr
*parent_instr
;
722 /** set of nir_instrs where this register is used (read from) */
723 struct list_head uses
;
725 /** set of nir_ifs where this register is used as a condition */
726 struct list_head if_uses
;
728 uint8_t num_components
;
730 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
738 struct nir_src
*indirect
; /** < NULL for no indirect offset */
739 unsigned base_offset
;
741 /* TODO use-def chain goes here */
745 nir_instr
*parent_instr
;
746 struct list_head def_link
;
749 struct nir_src
*indirect
; /** < NULL for no indirect offset */
750 unsigned base_offset
;
752 /* TODO def-use chain goes here */
757 typedef struct nir_src
{
759 /** Instruction that consumes this value as a source. */
760 nir_instr
*parent_instr
;
761 struct nir_if
*parent_if
;
764 struct list_head use_link
;
774 static inline nir_src
777 nir_src src
= { { NULL
} };
781 #define NIR_SRC_INIT nir_src_init()
783 #define nir_foreach_use(src, reg_or_ssa_def) \
784 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
786 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
787 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
789 #define nir_foreach_if_use(src, reg_or_ssa_def) \
790 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
792 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
793 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
804 static inline nir_dest
807 nir_dest dest
= { { { NULL
} } };
811 #define NIR_DEST_INIT nir_dest_init()
813 #define nir_foreach_def(dest, reg) \
814 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
816 #define nir_foreach_def_safe(dest, reg) \
817 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
819 static inline nir_src
820 nir_src_for_ssa(nir_ssa_def
*def
)
822 nir_src src
= NIR_SRC_INIT
;
830 static inline nir_src
831 nir_src_for_reg(nir_register
*reg
)
833 nir_src src
= NIR_SRC_INIT
;
837 src
.reg
.indirect
= NULL
;
838 src
.reg
.base_offset
= 0;
843 static inline nir_dest
844 nir_dest_for_reg(nir_register
*reg
)
846 nir_dest dest
= NIR_DEST_INIT
;
853 static inline unsigned
854 nir_src_bit_size(nir_src src
)
856 return src
.is_ssa
? src
.ssa
->bit_size
: src
.reg
.reg
->bit_size
;
859 static inline unsigned
860 nir_src_num_components(nir_src src
)
862 return src
.is_ssa
? src
.ssa
->num_components
: src
.reg
.reg
->num_components
;
866 nir_src_is_const(nir_src src
)
869 src
.ssa
->parent_instr
->type
== nir_instr_type_load_const
;
872 static inline unsigned
873 nir_dest_bit_size(nir_dest dest
)
875 return dest
.is_ssa
? dest
.ssa
.bit_size
: dest
.reg
.reg
->bit_size
;
878 static inline unsigned
879 nir_dest_num_components(nir_dest dest
)
881 return dest
.is_ssa
? dest
.ssa
.num_components
: dest
.reg
.reg
->num_components
;
884 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
885 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
891 * \name input modifiers
895 * For inputs interpreted as floating point, flips the sign bit. For
896 * inputs interpreted as integers, performs the two's complement negation.
901 * Clears the sign bit for floating point values, and computes the integer
902 * absolute value for integers. Note that the negate modifier acts after
903 * the absolute value modifier, therefore if both are set then all inputs
904 * will become negative.
910 * For each input component, says which component of the register it is
911 * chosen from. Note that which elements of the swizzle are used and which
912 * are ignored are based on the write mask for most opcodes - for example,
913 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
914 * a swizzle of {2, x, 1, 0} where x means "don't care."
916 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
923 * \name saturate output modifier
925 * Only valid for opcodes that output floating-point numbers. Clamps the
926 * output to between 0.0 and 1.0 inclusive.
931 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
934 /** NIR sized and unsized types
936 * The values in this enum are carefully chosen so that the sized type is
937 * just the unsized type OR the number of bits.
940 nir_type_invalid
= 0, /* Not a valid type */
944 nir_type_float
= 128,
945 nir_type_bool1
= 1 | nir_type_bool
,
946 nir_type_bool8
= 8 | nir_type_bool
,
947 nir_type_bool16
= 16 | nir_type_bool
,
948 nir_type_bool32
= 32 | nir_type_bool
,
949 nir_type_int1
= 1 | nir_type_int
,
950 nir_type_int8
= 8 | nir_type_int
,
951 nir_type_int16
= 16 | nir_type_int
,
952 nir_type_int32
= 32 | nir_type_int
,
953 nir_type_int64
= 64 | nir_type_int
,
954 nir_type_uint1
= 1 | nir_type_uint
,
955 nir_type_uint8
= 8 | nir_type_uint
,
956 nir_type_uint16
= 16 | nir_type_uint
,
957 nir_type_uint32
= 32 | nir_type_uint
,
958 nir_type_uint64
= 64 | nir_type_uint
,
959 nir_type_float16
= 16 | nir_type_float
,
960 nir_type_float32
= 32 | nir_type_float
,
961 nir_type_float64
= 64 | nir_type_float
,
964 #define NIR_ALU_TYPE_SIZE_MASK 0x79
965 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
967 static inline unsigned
968 nir_alu_type_get_type_size(nir_alu_type type
)
970 return type
& NIR_ALU_TYPE_SIZE_MASK
;
973 static inline unsigned
974 nir_alu_type_get_base_type(nir_alu_type type
)
976 return type
& NIR_ALU_TYPE_BASE_TYPE_MASK
;
979 static inline nir_alu_type
980 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
984 return nir_type_bool1
;
987 return nir_type_uint32
;
990 return nir_type_int32
;
992 case GLSL_TYPE_UINT16
:
993 return nir_type_uint16
;
995 case GLSL_TYPE_INT16
:
996 return nir_type_int16
;
998 case GLSL_TYPE_UINT8
:
999 return nir_type_uint8
;
1000 case GLSL_TYPE_INT8
:
1001 return nir_type_int8
;
1002 case GLSL_TYPE_UINT64
:
1003 return nir_type_uint64
;
1005 case GLSL_TYPE_INT64
:
1006 return nir_type_int64
;
1008 case GLSL_TYPE_FLOAT
:
1009 return nir_type_float32
;
1011 case GLSL_TYPE_FLOAT16
:
1012 return nir_type_float16
;
1014 case GLSL_TYPE_DOUBLE
:
1015 return nir_type_float64
;
1018 case GLSL_TYPE_SAMPLER
:
1019 case GLSL_TYPE_IMAGE
:
1020 case GLSL_TYPE_ATOMIC_UINT
:
1021 case GLSL_TYPE_STRUCT
:
1022 case GLSL_TYPE_INTERFACE
:
1023 case GLSL_TYPE_ARRAY
:
1024 case GLSL_TYPE_VOID
:
1025 case GLSL_TYPE_SUBROUTINE
:
1026 case GLSL_TYPE_FUNCTION
:
1027 case GLSL_TYPE_ERROR
:
1028 return nir_type_invalid
;
1031 unreachable("unknown type");
1034 static inline nir_alu_type
1035 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
1037 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
1040 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
1041 nir_rounding_mode rnd
);
1043 static inline nir_op
1044 nir_op_vec(unsigned components
)
1046 switch (components
) {
1047 case 1: return nir_op_mov
;
1048 case 2: return nir_op_vec2
;
1049 case 3: return nir_op_vec3
;
1050 case 4: return nir_op_vec4
;
1051 case 8: return nir_op_vec8
;
1052 case 16: return nir_op_vec16
;
1053 default: unreachable("bad component count");
1058 nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode
, unsigned bit_size
)
1060 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
) ||
1061 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
) ||
1062 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
);
1066 nir_is_denorm_flush_to_zero(unsigned execution_mode
, unsigned bit_size
)
1068 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
) ||
1069 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
) ||
1070 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
);
1074 nir_is_denorm_preserve(unsigned execution_mode
, unsigned bit_size
)
1076 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP16
) ||
1077 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP32
) ||
1078 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP64
);
1082 nir_is_rounding_mode_rtne(unsigned execution_mode
, unsigned bit_size
)
1084 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1085 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1086 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1090 nir_is_rounding_mode_rtz(unsigned execution_mode
, unsigned bit_size
)
1092 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1093 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1094 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1098 nir_has_any_rounding_mode_rtz(unsigned execution_mode
)
1100 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1101 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1102 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1106 nir_has_any_rounding_mode_rtne(unsigned execution_mode
)
1108 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1109 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1110 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1113 static inline nir_rounding_mode
1114 nir_get_rounding_mode_from_float_controls(unsigned execution_mode
,
1117 if (nir_alu_type_get_base_type(type
) != nir_type_float
)
1118 return nir_rounding_mode_undef
;
1120 unsigned bit_size
= nir_alu_type_get_type_size(type
);
1122 if (nir_is_rounding_mode_rtz(execution_mode
, bit_size
))
1123 return nir_rounding_mode_rtz
;
1124 if (nir_is_rounding_mode_rtne(execution_mode
, bit_size
))
1125 return nir_rounding_mode_rtne
;
1126 return nir_rounding_mode_undef
;
1130 nir_has_any_rounding_mode_enabled(unsigned execution_mode
)
1133 nir_has_any_rounding_mode_rtne(execution_mode
) ||
1134 nir_has_any_rounding_mode_rtz(execution_mode
);
1140 * Operation where the first two sources are commutative.
1142 * For 2-source operations, this just mathematical commutativity. Some
1143 * 3-source operations, like ffma, are only commutative in the first two
1146 NIR_OP_IS_2SRC_COMMUTATIVE
= (1 << 0),
1147 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
1148 } nir_op_algebraic_property
;
1153 unsigned num_inputs
;
1156 * The number of components in the output
1158 * If non-zero, this is the size of the output and input sizes are
1159 * explicitly given; swizzle and writemask are still in effect, but if
1160 * the output component is masked out, then the input component may
1163 * If zero, the opcode acts in the standard, per-component manner; the
1164 * operation is performed on each component (except the ones that are
1165 * masked out) with the input being taken from the input swizzle for
1168 * The size of some of the inputs may be given (i.e. non-zero) even
1169 * though output_size is zero; in that case, the inputs with a zero
1170 * size act per-component, while the inputs with non-zero size don't.
1172 unsigned output_size
;
1175 * The type of vector that the instruction outputs. Note that the
1176 * staurate modifier is only allowed on outputs with the float type.
1179 nir_alu_type output_type
;
1182 * The number of components in each input
1184 unsigned input_sizes
[NIR_MAX_VEC_COMPONENTS
];
1187 * The type of vector that each input takes. Note that negate and
1188 * absolute value are only allowed on inputs with int or float type and
1189 * behave differently on the two.
1191 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
1193 nir_op_algebraic_property algebraic_properties
;
1195 /* Whether this represents a numeric conversion opcode */
1199 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
1201 typedef struct nir_alu_instr
{
1205 /** Indicates that this ALU instruction generates an exact value
1207 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1208 * really equivalent to either. This indicates that the value generated by
1209 * this operation is high-precision and any code transformations that touch
1210 * it must ensure that the resulting value is bit-for-bit identical to the
1216 * Indicates that this instruction do not cause wrapping to occur, in the
1217 * form of overflow or underflow.
1219 bool no_signed_wrap
:1;
1220 bool no_unsigned_wrap
:1;
1226 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
1227 nir_alu_instr
*instr
);
1228 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
1229 nir_alu_instr
*instr
);
1231 /* is this source channel used? */
1233 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
1236 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1237 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
1239 return (instr
->dest
.write_mask
>> channel
) & 1;
1242 static inline nir_component_mask_t
1243 nir_alu_instr_src_read_mask(const nir_alu_instr
*instr
, unsigned src
)
1245 nir_component_mask_t read_mask
= 0;
1246 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; c
++) {
1247 if (!nir_alu_instr_channel_used(instr
, src
, c
))
1250 read_mask
|= (1 << instr
->src
[src
].swizzle
[c
]);
1256 * Get the number of channels used for a source
1258 static inline unsigned
1259 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
1261 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1262 return nir_op_infos
[instr
->op
].input_sizes
[src
];
1264 return nir_dest_num_components(instr
->dest
.dest
);
1268 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1270 switch (instr
->op
) {
1290 bool nir_const_value_negative_equal(nir_const_value c1
, nir_const_value c2
,
1291 nir_alu_type full_type
);
1293 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
1294 unsigned src1
, unsigned src2
);
1296 bool nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
1297 const nir_alu_instr
*alu2
,
1298 unsigned src1
, unsigned src2
);
1302 nir_deref_type_array
,
1303 nir_deref_type_array_wildcard
,
1304 nir_deref_type_ptr_as_array
,
1305 nir_deref_type_struct
,
1306 nir_deref_type_cast
,
1312 /** The type of this deref instruction */
1313 nir_deref_type deref_type
;
1315 /** The mode of the underlying variable */
1316 nir_variable_mode mode
;
1318 /** The dereferenced type of the resulting pointer value */
1319 const struct glsl_type
*type
;
1322 /** Variable being dereferenced if deref_type is a deref_var */
1325 /** Parent deref if deref_type is not deref_var */
1329 /** Additional deref parameters */
1340 unsigned ptr_stride
;
1344 /** Destination to store the resulting "pointer" */
1348 static inline nir_deref_instr
*nir_src_as_deref(nir_src src
);
1350 static inline nir_deref_instr
*
1351 nir_deref_instr_parent(const nir_deref_instr
*instr
)
1353 if (instr
->deref_type
== nir_deref_type_var
)
1356 return nir_src_as_deref(instr
->parent
);
1359 static inline nir_variable
*
1360 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1362 while (instr
->deref_type
!= nir_deref_type_var
) {
1363 if (instr
->deref_type
== nir_deref_type_cast
)
1366 instr
= nir_deref_instr_parent(instr
);
1372 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1373 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr
*instr
);
1374 bool nir_deref_instr_has_complex_use(nir_deref_instr
*instr
);
1376 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1378 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*instr
);
1383 struct nir_function
*callee
;
1385 unsigned num_params
;
1389 #include "nir_intrinsics.h"
1391 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1393 /** Represents an intrinsic
1395 * An intrinsic is an instruction type for handling things that are
1396 * more-or-less regular operations but don't just consume and produce SSA
1397 * values like ALU operations do. Intrinsics are not for things that have
1398 * special semantic meaning such as phi nodes and parallel copies.
1399 * Examples of intrinsics include variable load/store operations, system
1400 * value loads, and the like. Even though texturing more-or-less falls
1401 * under this category, texturing is its own instruction type because
1402 * trying to represent texturing with intrinsics would lead to a
1403 * combinatorial explosion of intrinsic opcodes.
1405 * By having a single instruction type for handling a lot of different
1406 * cases, optimization passes can look for intrinsics and, for the most
1407 * part, completely ignore them. Each intrinsic type also has a few
1408 * possible flags that govern whether or not they can be reordered or
1409 * eliminated. That way passes like dead code elimination can still work
1410 * on intrisics without understanding the meaning of each.
1412 * Each intrinsic has some number of constant indices, some number of
1413 * variables, and some number of sources. What these sources, variables,
1414 * and indices mean depends on the intrinsic and is documented with the
1415 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1416 * instructions are the only types of instruction that can operate on
1422 nir_intrinsic_op intrinsic
;
1426 /** number of components if this is a vectorized intrinsic
1428 * Similarly to ALU operations, some intrinsics are vectorized.
1429 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1430 * For vectorized intrinsics, the num_components field specifies the
1431 * number of destination components and the number of source components
1432 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1434 uint8_t num_components
;
1436 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1439 } nir_intrinsic_instr
;
1441 static inline nir_variable
*
1442 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1444 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1448 /* Memory ordering. */
1449 NIR_MEMORY_ACQUIRE
= 1 << 0,
1450 NIR_MEMORY_RELEASE
= 1 << 1,
1451 NIR_MEMORY_ACQ_REL
= NIR_MEMORY_ACQUIRE
| NIR_MEMORY_RELEASE
,
1453 /* Memory visibility operations. */
1454 NIR_MEMORY_MAKE_AVAILABLE
= 1 << 3,
1455 NIR_MEMORY_MAKE_VISIBLE
= 1 << 4,
1456 } nir_memory_semantics
;
1460 NIR_SCOPE_QUEUE_FAMILY
,
1461 NIR_SCOPE_WORKGROUP
,
1463 NIR_SCOPE_INVOCATION
,
1467 * \name NIR intrinsics semantic flags
1469 * information about what the compiler can do with the intrinsics.
1471 * \sa nir_intrinsic_info::flags
1475 * whether the intrinsic can be safely eliminated if none of its output
1476 * value is not being used.
1478 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1481 * Whether the intrinsic can be reordered with respect to any other
1482 * intrinsic, i.e. whether the only reordering dependencies of the
1483 * intrinsic are due to the register reads/writes.
1485 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1486 } nir_intrinsic_semantic_flag
;
1489 * \name NIR intrinsics const-index flag
1491 * Indicates the usage of a const_index slot.
1493 * \sa nir_intrinsic_info::index_map
1497 * Generally instructions that take a offset src argument, can encode
1498 * a constant 'base' value which is added to the offset.
1500 NIR_INTRINSIC_BASE
= 1,
1503 * For store instructions, a writemask for the store.
1505 NIR_INTRINSIC_WRMASK
,
1508 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1510 NIR_INTRINSIC_STREAM_ID
,
1513 * The clip-plane id for load_user_clip_plane intrinsic.
1515 NIR_INTRINSIC_UCP_ID
,
1518 * The amount of data, starting from BASE, that this instruction may
1519 * access. This is used to provide bounds if the offset is not constant.
1521 NIR_INTRINSIC_RANGE
,
1524 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1526 NIR_INTRINSIC_DESC_SET
,
1529 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1531 NIR_INTRINSIC_BINDING
,
1536 NIR_INTRINSIC_COMPONENT
,
1539 * Interpolation mode (only meaningful for FS inputs).
1541 NIR_INTRINSIC_INTERP_MODE
,
1544 * A binary nir_op to use when performing a reduction or scan operation
1546 NIR_INTRINSIC_REDUCTION_OP
,
1549 * Cluster size for reduction operations
1551 NIR_INTRINSIC_CLUSTER_SIZE
,
1554 * Parameter index for a load_param intrinsic
1556 NIR_INTRINSIC_PARAM_IDX
,
1559 * Image dimensionality for image intrinsics
1561 * One of GLSL_SAMPLER_DIM_*
1563 NIR_INTRINSIC_IMAGE_DIM
,
1566 * Non-zero if we are accessing an array image
1568 NIR_INTRINSIC_IMAGE_ARRAY
,
1571 * Image format for image intrinsics
1573 NIR_INTRINSIC_FORMAT
,
1576 * Access qualifiers for image and memory access intrinsics
1578 NIR_INTRINSIC_ACCESS
,
1581 * Alignment for offsets and addresses
1583 * These two parameters, specify an alignment in terms of a multiplier and
1584 * an offset. The offset or address parameter X of the intrinsic is
1585 * guaranteed to satisfy the following:
1587 * (X - align_offset) % align_mul == 0
1589 NIR_INTRINSIC_ALIGN_MUL
,
1590 NIR_INTRINSIC_ALIGN_OFFSET
,
1593 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1595 NIR_INTRINSIC_DESC_TYPE
,
1598 * The nir_alu_type of a uniform/input/output
1603 * The swizzle mask for the instructions
1604 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1606 NIR_INTRINSIC_SWIZZLE_MASK
,
1608 /* Separate source/dest access flags for copies */
1609 NIR_INTRINSIC_SRC_ACCESS
,
1610 NIR_INTRINSIC_DST_ACCESS
,
1612 /* Driver location for nir_load_patch_location_ir3 */
1613 NIR_INTRINSIC_DRIVER_LOCATION
,
1616 * Mask of nir_memory_semantics, includes ordering and visibility.
1618 NIR_INTRINSIC_MEMORY_SEMANTICS
,
1621 * Mask of nir_variable_modes affected by the memory operation.
1623 NIR_INTRINSIC_MEMORY_MODES
,
1626 * Value of nir_scope.
1628 NIR_INTRINSIC_MEMORY_SCOPE
,
1630 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1632 } nir_intrinsic_index_flag
;
1634 #define NIR_INTRINSIC_MAX_INPUTS 5
1639 unsigned num_srcs
; /** < number of register/SSA inputs */
1641 /** number of components of each input register
1643 * If this value is 0, the number of components is given by the
1644 * num_components field of nir_intrinsic_instr. If this value is -1, the
1645 * intrinsic consumes however many components are provided and it is not
1648 int src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1652 /** number of components of the output register
1654 * If this value is 0, the number of components is given by the
1655 * num_components field of nir_intrinsic_instr.
1657 unsigned dest_components
;
1659 /** bitfield of legal bit sizes */
1660 unsigned dest_bit_sizes
;
1662 /** the number of constant indices used by the intrinsic */
1663 unsigned num_indices
;
1665 /** indicates the usage of intr->const_index[n] */
1666 unsigned index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1668 /** semantic flags for calls to this intrinsic */
1669 nir_intrinsic_semantic_flag flags
;
1670 } nir_intrinsic_info
;
1672 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1674 static inline unsigned
1675 nir_intrinsic_src_components(nir_intrinsic_instr
*intr
, unsigned srcn
)
1677 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1678 assert(srcn
< info
->num_srcs
);
1679 if (info
->src_components
[srcn
] > 0)
1680 return info
->src_components
[srcn
];
1681 else if (info
->src_components
[srcn
] == 0)
1682 return intr
->num_components
;
1684 return nir_src_num_components(intr
->src
[srcn
]);
1687 static inline unsigned
1688 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1690 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1691 if (!info
->has_dest
)
1693 else if (info
->dest_components
)
1694 return info
->dest_components
;
1696 return intr
->num_components
;
1699 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1700 static inline type \
1701 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1703 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1704 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1705 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1707 static inline void \
1708 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1710 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1711 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1712 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1715 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1716 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1717 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1718 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1719 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1720 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1721 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1722 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1723 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1724 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1725 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1726 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1727 INTRINSIC_IDX_ACCESSORS(image_dim
, IMAGE_DIM
, enum glsl_sampler_dim
)
1728 INTRINSIC_IDX_ACCESSORS(image_array
, IMAGE_ARRAY
, bool)
1729 INTRINSIC_IDX_ACCESSORS(access
, ACCESS
, enum gl_access_qualifier
)
1730 INTRINSIC_IDX_ACCESSORS(src_access
, SRC_ACCESS
, enum gl_access_qualifier
)
1731 INTRINSIC_IDX_ACCESSORS(dst_access
, DST_ACCESS
, enum gl_access_qualifier
)
1732 INTRINSIC_IDX_ACCESSORS(format
, FORMAT
, enum pipe_format
)
1733 INTRINSIC_IDX_ACCESSORS(align_mul
, ALIGN_MUL
, unsigned)
1734 INTRINSIC_IDX_ACCESSORS(align_offset
, ALIGN_OFFSET
, unsigned)
1735 INTRINSIC_IDX_ACCESSORS(desc_type
, DESC_TYPE
, unsigned)
1736 INTRINSIC_IDX_ACCESSORS(type
, TYPE
, nir_alu_type
)
1737 INTRINSIC_IDX_ACCESSORS(swizzle_mask
, SWIZZLE_MASK
, unsigned)
1738 INTRINSIC_IDX_ACCESSORS(driver_location
, DRIVER_LOCATION
, unsigned)
1739 INTRINSIC_IDX_ACCESSORS(memory_semantics
, MEMORY_SEMANTICS
, nir_memory_semantics
)
1740 INTRINSIC_IDX_ACCESSORS(memory_modes
, MEMORY_MODES
, nir_variable_mode
)
1741 INTRINSIC_IDX_ACCESSORS(memory_scope
, MEMORY_SCOPE
, nir_scope
)
1744 nir_intrinsic_set_align(nir_intrinsic_instr
*intrin
,
1745 unsigned align_mul
, unsigned align_offset
)
1747 assert(util_is_power_of_two_nonzero(align_mul
));
1748 assert(align_offset
< align_mul
);
1749 nir_intrinsic_set_align_mul(intrin
, align_mul
);
1750 nir_intrinsic_set_align_offset(intrin
, align_offset
);
1753 /** Returns a simple alignment for a load/store intrinsic offset
1755 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1756 * and ALIGN_OFFSET parameters, this helper takes both into account and
1757 * provides a single simple alignment parameter. The offset X is guaranteed
1758 * to satisfy X % align == 0.
1760 static inline unsigned
1761 nir_intrinsic_align(const nir_intrinsic_instr
*intrin
)
1763 const unsigned align_mul
= nir_intrinsic_align_mul(intrin
);
1764 const unsigned align_offset
= nir_intrinsic_align_offset(intrin
);
1765 assert(align_offset
< align_mul
);
1766 return align_offset
? 1 << (ffs(align_offset
) - 1) : align_mul
;
1770 nir_image_intrinsic_coord_components(const nir_intrinsic_instr
*instr
);
1772 /* Converts a image_deref_* intrinsic into a image_* one */
1773 void nir_rewrite_image_intrinsic(nir_intrinsic_instr
*instr
,
1774 nir_ssa_def
*handle
, bool bindless
);
1776 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1778 nir_intrinsic_can_reorder(nir_intrinsic_instr
*instr
)
1780 if (instr
->intrinsic
== nir_intrinsic_load_deref
||
1781 instr
->intrinsic
== nir_intrinsic_load_ssbo
||
1782 instr
->intrinsic
== nir_intrinsic_bindless_image_load
||
1783 instr
->intrinsic
== nir_intrinsic_image_deref_load
||
1784 instr
->intrinsic
== nir_intrinsic_image_load
) {
1785 return nir_intrinsic_access(instr
) & ACCESS_CAN_REORDER
;
1787 const nir_intrinsic_info
*info
=
1788 &nir_intrinsic_infos
[instr
->intrinsic
];
1789 return (info
->flags
& NIR_INTRINSIC_CAN_ELIMINATE
) &&
1790 (info
->flags
& NIR_INTRINSIC_CAN_REORDER
);
1795 * \group texture information
1797 * This gives semantic information about textures which is useful to the
1798 * frontend, the backend, and lowering passes, but not the optimizer.
1803 nir_tex_src_projector
,
1804 nir_tex_src_comparator
, /* shadow comparator */
1808 nir_tex_src_min_lod
,
1809 nir_tex_src_ms_index
, /* MSAA sample index */
1810 nir_tex_src_ms_mcs
, /* MSAA compression value */
1813 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
1814 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
1815 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
1816 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
1817 nir_tex_src_texture_handle
, /* < bindless texture handle */
1818 nir_tex_src_sampler_handle
, /* < bindless sampler handle */
1819 nir_tex_src_plane
, /* < selects plane for planar textures */
1820 nir_num_tex_src_types
1825 nir_tex_src_type src_type
;
1829 nir_texop_tex
, /**< Regular texture look-up */
1830 nir_texop_txb
, /**< Texture look-up with LOD bias */
1831 nir_texop_txl
, /**< Texture look-up with explicit LOD */
1832 nir_texop_txd
, /**< Texture look-up with partial derivatives */
1833 nir_texop_txf
, /**< Texel fetch with explicit LOD */
1834 nir_texop_txf_ms
, /**< Multisample texture fetch */
1835 nir_texop_txf_ms_fb
, /**< Multisample texture fetch from framebuffer */
1836 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
1837 nir_texop_txs
, /**< Texture size */
1838 nir_texop_lod
, /**< Texture lod query */
1839 nir_texop_tg4
, /**< Texture gather */
1840 nir_texop_query_levels
, /**< Texture levels query */
1841 nir_texop_texture_samples
, /**< Texture samples query */
1842 nir_texop_samples_identical
, /**< Query whether all samples are definitely
1845 nir_texop_tex_prefetch
, /**< Regular texture look-up, eligible for pre-dispatch */
1846 nir_texop_fragment_fetch
, /**< Multisample fragment color texture fetch */
1847 nir_texop_fragment_mask_fetch
,/**< Multisample fragment mask texture fetch */
1853 enum glsl_sampler_dim sampler_dim
;
1854 nir_alu_type dest_type
;
1859 unsigned num_srcs
, coord_components
;
1860 bool is_array
, is_shadow
;
1863 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1864 * components or the new-style shadow that outputs 1 component.
1866 bool is_new_style_shadow
;
1868 /* gather component selector */
1869 unsigned component
: 2;
1871 /* gather offsets */
1872 int8_t tg4_offsets
[4][2];
1874 /* True if the texture index or handle is not dynamically uniform */
1875 bool texture_non_uniform
;
1877 /* True if the sampler index or handle is not dynamically uniform */
1878 bool sampler_non_uniform
;
1880 /** The texture index
1882 * If this texture instruction has a nir_tex_src_texture_offset source,
1883 * then the texture index is given by texture_index + texture_offset.
1885 unsigned texture_index
;
1887 /** The size of the texture array or 0 if it's not an array */
1888 unsigned texture_array_size
;
1890 /** The sampler index
1892 * The following operations do not require a sampler and, as such, this
1893 * field should be ignored:
1895 * - nir_texop_txf_ms
1898 * - nir_texop_query_levels
1899 * - nir_texop_texture_samples
1900 * - nir_texop_samples_identical
1902 * If this texture instruction has a nir_tex_src_sampler_offset source,
1903 * then the sampler index is given by sampler_index + sampler_offset.
1905 unsigned sampler_index
;
1908 static inline unsigned
1909 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
1911 switch (instr
->op
) {
1912 case nir_texop_txs
: {
1914 switch (instr
->sampler_dim
) {
1915 case GLSL_SAMPLER_DIM_1D
:
1916 case GLSL_SAMPLER_DIM_BUF
:
1919 case GLSL_SAMPLER_DIM_2D
:
1920 case GLSL_SAMPLER_DIM_CUBE
:
1921 case GLSL_SAMPLER_DIM_MS
:
1922 case GLSL_SAMPLER_DIM_RECT
:
1923 case GLSL_SAMPLER_DIM_EXTERNAL
:
1924 case GLSL_SAMPLER_DIM_SUBPASS
:
1927 case GLSL_SAMPLER_DIM_3D
:
1931 unreachable("not reached");
1933 if (instr
->is_array
)
1941 case nir_texop_texture_samples
:
1942 case nir_texop_query_levels
:
1943 case nir_texop_samples_identical
:
1944 case nir_texop_fragment_mask_fetch
:
1948 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
1955 /* Returns true if this texture operation queries something about the texture
1956 * rather than actually sampling it.
1959 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
1961 switch (instr
->op
) {
1964 case nir_texop_texture_samples
:
1965 case nir_texop_query_levels
:
1966 case nir_texop_txf_ms_mcs
:
1973 case nir_texop_txf_ms
:
1974 case nir_texop_txf_ms_fb
:
1978 unreachable("Invalid texture opcode");
1983 nir_tex_instr_has_implicit_derivative(const nir_tex_instr
*instr
)
1985 switch (instr
->op
) {
1995 static inline nir_alu_type
1996 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
1998 switch (instr
->src
[src
].src_type
) {
1999 case nir_tex_src_coord
:
2000 switch (instr
->op
) {
2002 case nir_texop_txf_ms
:
2003 case nir_texop_txf_ms_fb
:
2004 case nir_texop_txf_ms_mcs
:
2005 case nir_texop_samples_identical
:
2006 return nir_type_int
;
2009 return nir_type_float
;
2012 case nir_tex_src_lod
:
2013 switch (instr
->op
) {
2016 return nir_type_int
;
2019 return nir_type_float
;
2022 case nir_tex_src_projector
:
2023 case nir_tex_src_comparator
:
2024 case nir_tex_src_bias
:
2025 case nir_tex_src_min_lod
:
2026 case nir_tex_src_ddx
:
2027 case nir_tex_src_ddy
:
2028 return nir_type_float
;
2030 case nir_tex_src_offset
:
2031 case nir_tex_src_ms_index
:
2032 case nir_tex_src_plane
:
2033 return nir_type_int
;
2035 case nir_tex_src_ms_mcs
:
2036 case nir_tex_src_texture_deref
:
2037 case nir_tex_src_sampler_deref
:
2038 case nir_tex_src_texture_offset
:
2039 case nir_tex_src_sampler_offset
:
2040 case nir_tex_src_texture_handle
:
2041 case nir_tex_src_sampler_handle
:
2042 return nir_type_uint
;
2044 case nir_num_tex_src_types
:
2045 unreachable("nir_num_tex_src_types is not a valid source type");
2048 unreachable("Invalid texture source type");
2051 static inline unsigned
2052 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
2054 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
2055 return instr
->coord_components
;
2057 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
2058 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
2061 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
2062 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
2063 if (instr
->is_array
)
2064 return instr
->coord_components
- 1;
2066 return instr
->coord_components
;
2069 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
2070 * the offset, since a cube maps to a single face.
2072 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
2073 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2075 else if (instr
->is_array
)
2076 return instr
->coord_components
- 1;
2078 return instr
->coord_components
;
2085 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
2087 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
2088 if (instr
->src
[i
].src_type
== type
)
2094 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
2095 nir_tex_src_type src_type
,
2098 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
2100 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
);
2107 nir_const_value value
[];
2108 } nir_load_const_instr
;
2121 /* creates a new SSA variable in an undefined state */
2126 } nir_ssa_undef_instr
;
2129 struct exec_node node
;
2131 /* The predecessor block corresponding to this source */
2132 struct nir_block
*pred
;
2137 #define nir_foreach_phi_src(phi_src, phi) \
2138 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
2139 #define nir_foreach_phi_src_safe(phi_src, phi) \
2140 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
2145 struct exec_list srcs
; /** < list of nir_phi_src */
2151 struct exec_node node
;
2154 } nir_parallel_copy_entry
;
2156 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
2157 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
2162 /* A list of nir_parallel_copy_entrys. The sources of all of the
2163 * entries are copied to the corresponding destinations "in parallel".
2164 * In other words, if we have two entries: a -> b and b -> a, the values
2167 struct exec_list entries
;
2168 } nir_parallel_copy_instr
;
2170 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
2171 type
, nir_instr_type_alu
)
2172 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
2173 type
, nir_instr_type_deref
)
2174 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
2175 type
, nir_instr_type_call
)
2176 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
2177 type
, nir_instr_type_jump
)
2178 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
2179 type
, nir_instr_type_tex
)
2180 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
2181 type
, nir_instr_type_intrinsic
)
2182 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
2183 type
, nir_instr_type_load_const
)
2184 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
2185 type
, nir_instr_type_ssa_undef
)
2186 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
2187 type
, nir_instr_type_phi
)
2188 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
2189 nir_parallel_copy_instr
, instr
,
2190 type
, nir_instr_type_parallel_copy
)
2193 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
2194 static inline type \
2195 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2197 assert(nir_src_is_const(src)); \
2198 nir_load_const_instr *load = \
2199 nir_instr_as_load_const(src.ssa->parent_instr); \
2200 assert(comp < load->def.num_components); \
2201 return nir_const_value_as_##suffix(load->value[comp], \
2202 load->def.bit_size); \
2205 static inline type \
2206 nir_src_as_##suffix(nir_src src) \
2208 assert(nir_src_num_components(src) == 1); \
2209 return nir_src_comp_as_##suffix(src, 0); \
2212 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2213 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint
)
2214 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2215 NIR_DEFINE_SRC_AS_CONST(double, float)
2217 #undef NIR_DEFINE_SRC_AS_CONST
2226 nir_ssa_scalar_is_const(nir_ssa_scalar s
)
2228 return s
.def
->parent_instr
->type
== nir_instr_type_load_const
;
2231 static inline nir_const_value
2232 nir_ssa_scalar_as_const_value(nir_ssa_scalar s
)
2234 assert(s
.comp
< s
.def
->num_components
);
2235 nir_load_const_instr
*load
= nir_instr_as_load_const(s
.def
->parent_instr
);
2236 return load
->value
[s
.comp
];
2239 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2240 static inline type \
2241 nir_ssa_scalar_as_##suffix(nir_ssa_scalar s) \
2243 return nir_const_value_as_##suffix( \
2244 nir_ssa_scalar_as_const_value(s), s.def->bit_size); \
2247 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2248 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint
)
2249 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2250 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2252 #undef NIR_DEFINE_SCALAR_AS_CONST
2255 nir_ssa_scalar_is_alu(nir_ssa_scalar s
)
2257 return s
.def
->parent_instr
->type
== nir_instr_type_alu
;
2260 static inline nir_op
2261 nir_ssa_scalar_alu_op(nir_ssa_scalar s
)
2263 return nir_instr_as_alu(s
.def
->parent_instr
)->op
;
2266 static inline nir_ssa_scalar
2267 nir_ssa_scalar_chase_alu_src(nir_ssa_scalar s
, unsigned alu_src_idx
)
2269 nir_ssa_scalar out
= { NULL
, 0 };
2271 nir_alu_instr
*alu
= nir_instr_as_alu(s
.def
->parent_instr
);
2272 assert(alu_src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
2274 /* Our component must be written */
2275 assert(s
.comp
< s
.def
->num_components
);
2276 assert(alu
->dest
.write_mask
& (1u << s
.comp
));
2278 assert(alu
->src
[alu_src_idx
].src
.is_ssa
);
2279 out
.def
= alu
->src
[alu_src_idx
].src
.ssa
;
2281 if (nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 0) {
2282 /* The ALU src is unsized so the source component follows the
2283 * destination component.
2285 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[s
.comp
];
2287 /* This is a sized source so all source components work together to
2288 * produce all the destination components. Since we need to return a
2289 * scalar, this only works if the source is a scalar.
2291 assert(nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 1);
2292 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[0];
2294 assert(out
.comp
< out
.def
->num_components
);
2303 * Control flow consists of a tree of control flow nodes, which include
2304 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2305 * instructions that always run start-to-finish. Each basic block also keeps
2306 * track of its successors (blocks which may run immediately after the current
2307 * block) and predecessors (blocks which could have run immediately before the
2308 * current block). Each function also has a start block and an end block which
2309 * all return statements point to (which is always empty). Together, all the
2310 * blocks with their predecessors and successors make up the control flow
2311 * graph (CFG) of the function. There are helpers that modify the tree of
2312 * control flow nodes while modifying the CFG appropriately; these should be
2313 * used instead of modifying the tree directly.
2320 nir_cf_node_function
2323 typedef struct nir_cf_node
{
2324 struct exec_node node
;
2325 nir_cf_node_type type
;
2326 struct nir_cf_node
*parent
;
2329 typedef struct nir_block
{
2330 nir_cf_node cf_node
;
2332 struct exec_list instr_list
; /** < list of nir_instr */
2334 /** generic block index; generated by nir_index_blocks */
2338 * Each block can only have up to 2 successors, so we put them in a simple
2339 * array - no need for anything more complicated.
2341 struct nir_block
*successors
[2];
2343 /* Set of nir_block predecessors in the CFG */
2344 struct set
*predecessors
;
2347 * this node's immediate dominator in the dominance tree - set to NULL for
2350 struct nir_block
*imm_dom
;
2352 /* This node's children in the dominance tree */
2353 unsigned num_dom_children
;
2354 struct nir_block
**dom_children
;
2356 /* Set of nir_blocks on the dominance frontier of this block */
2357 struct set
*dom_frontier
;
2360 * These two indices have the property that dom_{pre,post}_index for each
2361 * child of this block in the dominance tree will always be between
2362 * dom_pre_index and dom_post_index for this block, which makes testing if
2363 * a given block is dominated by another block an O(1) operation.
2365 unsigned dom_pre_index
, dom_post_index
;
2367 /* live in and out for this block; used for liveness analysis */
2368 BITSET_WORD
*live_in
;
2369 BITSET_WORD
*live_out
;
2372 static inline nir_instr
*
2373 nir_block_first_instr(nir_block
*block
)
2375 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
2376 return exec_node_data(nir_instr
, head
, node
);
2379 static inline nir_instr
*
2380 nir_block_last_instr(nir_block
*block
)
2382 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
2383 return exec_node_data(nir_instr
, tail
, node
);
2387 nir_block_ends_in_jump(nir_block
*block
)
2389 return !exec_list_is_empty(&block
->instr_list
) &&
2390 nir_block_last_instr(block
)->type
== nir_instr_type_jump
;
2393 #define nir_foreach_instr(instr, block) \
2394 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
2395 #define nir_foreach_instr_reverse(instr, block) \
2396 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
2397 #define nir_foreach_instr_safe(instr, block) \
2398 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
2399 #define nir_foreach_instr_reverse_safe(instr, block) \
2400 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
2403 nir_selection_control_none
= 0x0,
2404 nir_selection_control_flatten
= 0x1,
2405 nir_selection_control_dont_flatten
= 0x2,
2406 } nir_selection_control
;
2408 typedef struct nir_if
{
2409 nir_cf_node cf_node
;
2411 nir_selection_control control
;
2413 struct exec_list then_list
; /** < list of nir_cf_node */
2414 struct exec_list else_list
; /** < list of nir_cf_node */
2420 /** Instruction that generates nif::condition. */
2421 nir_instr
*conditional_instr
;
2423 /** Block within ::nif that has the break instruction. */
2424 nir_block
*break_block
;
2426 /** Last block for the then- or else-path that does not contain the break. */
2427 nir_block
*continue_from_block
;
2429 /** True when ::break_block is in the else-path of ::nif. */
2430 bool continue_from_then
;
2433 /* This is true if the terminators exact trip count is unknown. For
2436 * for (int i = 0; i < imin(x, 4); i++)
2439 * Here loop analysis would have set a max_trip_count of 4 however we dont
2440 * know for sure that this is the exact trip count.
2442 bool exact_trip_count_unknown
;
2444 struct list_head loop_terminator_link
;
2445 } nir_loop_terminator
;
2448 /* Estimated cost (in number of instructions) of the loop */
2449 unsigned instr_cost
;
2451 /* Guessed trip count based on array indexing */
2452 unsigned guessed_trip_count
;
2454 /* Maximum number of times the loop is run (if known) */
2455 unsigned max_trip_count
;
2457 /* Do we know the exact number of times the loop will be run */
2458 bool exact_trip_count_known
;
2460 /* Unroll the loop regardless of its size */
2463 /* Does the loop contain complex loop terminators, continues or other
2464 * complex behaviours? If this is true we can't rely on
2465 * loop_terminator_list to be complete or accurate.
2469 nir_loop_terminator
*limiting_terminator
;
2471 /* A list of loop_terminators terminating this loop. */
2472 struct list_head loop_terminator_list
;
2476 nir_loop_control_none
= 0x0,
2477 nir_loop_control_unroll
= 0x1,
2478 nir_loop_control_dont_unroll
= 0x2,
2482 nir_cf_node cf_node
;
2484 struct exec_list body
; /** < list of nir_cf_node */
2486 nir_loop_info
*info
;
2487 nir_loop_control control
;
2488 bool partially_unrolled
;
2492 * Various bits of metadata that can may be created or required by
2493 * optimization and analysis passes
2496 nir_metadata_none
= 0x0,
2497 nir_metadata_block_index
= 0x1,
2498 nir_metadata_dominance
= 0x2,
2499 nir_metadata_live_ssa_defs
= 0x4,
2500 nir_metadata_not_properly_reset
= 0x8,
2501 nir_metadata_loop_analysis
= 0x10,
2505 nir_cf_node cf_node
;
2507 /** pointer to the function of which this is an implementation */
2508 struct nir_function
*function
;
2510 struct exec_list body
; /** < list of nir_cf_node */
2512 nir_block
*end_block
;
2514 /** list for all local variables in the function */
2515 struct exec_list locals
;
2517 /** list of local registers in the function */
2518 struct exec_list registers
;
2520 /** next available local register index */
2523 /** next available SSA value index */
2526 /* total number of basic blocks, only valid when block_index_dirty = false */
2527 unsigned num_blocks
;
2529 nir_metadata valid_metadata
;
2530 } nir_function_impl
;
2532 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2533 nir_start_block(nir_function_impl
*impl
)
2535 return (nir_block
*) impl
->body
.head_sentinel
.next
;
2538 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2539 nir_impl_last_block(nir_function_impl
*impl
)
2541 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
2544 static inline nir_cf_node
*
2545 nir_cf_node_next(nir_cf_node
*node
)
2547 struct exec_node
*next
= exec_node_get_next(&node
->node
);
2548 if (exec_node_is_tail_sentinel(next
))
2551 return exec_node_data(nir_cf_node
, next
, node
);
2554 static inline nir_cf_node
*
2555 nir_cf_node_prev(nir_cf_node
*node
)
2557 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
2558 if (exec_node_is_head_sentinel(prev
))
2561 return exec_node_data(nir_cf_node
, prev
, node
);
2565 nir_cf_node_is_first(const nir_cf_node
*node
)
2567 return exec_node_is_head_sentinel(node
->node
.prev
);
2571 nir_cf_node_is_last(const nir_cf_node
*node
)
2573 return exec_node_is_tail_sentinel(node
->node
.next
);
2576 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
2577 type
, nir_cf_node_block
)
2578 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
2579 type
, nir_cf_node_if
)
2580 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
2581 type
, nir_cf_node_loop
)
2582 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
2583 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
2585 static inline nir_block
*
2586 nir_if_first_then_block(nir_if
*if_stmt
)
2588 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
2589 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2592 static inline nir_block
*
2593 nir_if_last_then_block(nir_if
*if_stmt
)
2595 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
2596 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2599 static inline nir_block
*
2600 nir_if_first_else_block(nir_if
*if_stmt
)
2602 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
2603 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2606 static inline nir_block
*
2607 nir_if_last_else_block(nir_if
*if_stmt
)
2609 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
2610 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2613 static inline nir_block
*
2614 nir_loop_first_block(nir_loop
*loop
)
2616 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
2617 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2620 static inline nir_block
*
2621 nir_loop_last_block(nir_loop
*loop
)
2623 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
2624 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2628 * Return true if this list of cf_nodes contains a single empty block.
2631 nir_cf_list_is_empty_block(struct exec_list
*cf_list
)
2633 if (exec_list_is_singular(cf_list
)) {
2634 struct exec_node
*head
= exec_list_get_head(cf_list
);
2636 nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2637 return exec_list_is_empty(&block
->instr_list
);
2643 uint8_t num_components
;
2647 typedef struct nir_function
{
2648 struct exec_node node
;
2651 struct nir_shader
*shader
;
2653 unsigned num_params
;
2654 nir_parameter
*params
;
2656 /** The implementation of this function.
2658 * If the function is only declared and not implemented, this is NULL.
2660 nir_function_impl
*impl
;
2666 nir_lower_imul64
= (1 << 0),
2667 nir_lower_isign64
= (1 << 1),
2668 /** Lower all int64 modulus and division opcodes */
2669 nir_lower_divmod64
= (1 << 2),
2670 /** Lower all 64-bit umul_high and imul_high opcodes */
2671 nir_lower_imul_high64
= (1 << 3),
2672 nir_lower_mov64
= (1 << 4),
2673 nir_lower_icmp64
= (1 << 5),
2674 nir_lower_iadd64
= (1 << 6),
2675 nir_lower_iabs64
= (1 << 7),
2676 nir_lower_ineg64
= (1 << 8),
2677 nir_lower_logic64
= (1 << 9),
2678 nir_lower_minmax64
= (1 << 10),
2679 nir_lower_shift64
= (1 << 11),
2680 nir_lower_imul_2x32_64
= (1 << 12),
2681 nir_lower_extract64
= (1 << 13),
2682 nir_lower_ufind_msb64
= (1 << 14),
2683 } nir_lower_int64_options
;
2686 nir_lower_drcp
= (1 << 0),
2687 nir_lower_dsqrt
= (1 << 1),
2688 nir_lower_drsq
= (1 << 2),
2689 nir_lower_dtrunc
= (1 << 3),
2690 nir_lower_dfloor
= (1 << 4),
2691 nir_lower_dceil
= (1 << 5),
2692 nir_lower_dfract
= (1 << 6),
2693 nir_lower_dround_even
= (1 << 7),
2694 nir_lower_dmod
= (1 << 8),
2695 nir_lower_dsub
= (1 << 9),
2696 nir_lower_ddiv
= (1 << 10),
2697 nir_lower_fp64_full_software
= (1 << 11),
2698 } nir_lower_doubles_options
;
2701 nir_divergence_single_prim_per_subgroup
= (1 << 0),
2702 nir_divergence_single_patch_per_tcs_subgroup
= (1 << 1),
2703 nir_divergence_single_patch_per_tes_subgroup
= (1 << 2),
2704 nir_divergence_view_index_uniform
= (1 << 3),
2705 } nir_divergence_options
;
2707 typedef struct nir_shader_compiler_options
{
2713 /** Lowers flrp when it does not support doubles */
2720 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2721 bool lower_bitfield_extract
;
2722 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
2723 bool lower_bitfield_extract_to_shifts
;
2724 /** Lowers bitfield_insert to bfi/bfm */
2725 bool lower_bitfield_insert
;
2726 /** Lowers bitfield_insert to compares, and shifts. */
2727 bool lower_bitfield_insert_to_shifts
;
2728 /** Lowers bitfield_insert to bfm/bitfield_select. */
2729 bool lower_bitfield_insert_to_bitfield_select
;
2730 /** Lowers bitfield_reverse to shifts. */
2731 bool lower_bitfield_reverse
;
2732 /** Lowers bit_count to shifts. */
2733 bool lower_bit_count
;
2734 /** Lowers ifind_msb to compare and ufind_msb */
2735 bool lower_ifind_msb
;
2736 /** Lowers find_lsb to ufind_msb and logic ops */
2737 bool lower_find_lsb
;
2738 bool lower_uadd_carry
;
2739 bool lower_usub_borrow
;
2740 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2741 bool lower_mul_high
;
2742 /** lowers fneg and ineg to fsub and isub. */
2744 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2747 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2750 /* lower fall_equalN/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
2751 bool lower_vector_cmp
;
2753 /** enables rules to lower idiv by power-of-two: */
2756 /** enable rules to avoid bit ops */
2759 /** enables rules to lower isign to imin+imax */
2762 /** enables rules to lower fsign to fsub and flt */
2765 /* lower fdph to fdot4 */
2768 /** lower fdot to fmul and fsum/fadd. */
2771 /* Does the native fdot instruction replicate its result for four
2772 * components? If so, then opt_algebraic_late will turn all fdotN
2773 * instructions into fdot_replicatedN instructions.
2775 bool fdot_replicates
;
2777 /** lowers ffloor to fsub+ffract: */
2780 /** lowers ffract to fsub+ffloor: */
2783 /** lowers fceil to fneg+ffloor+fneg: */
2790 bool lower_pack_half_2x16
;
2791 bool lower_pack_half_2x16_split
;
2792 bool lower_pack_unorm_2x16
;
2793 bool lower_pack_snorm_2x16
;
2794 bool lower_pack_unorm_4x8
;
2795 bool lower_pack_snorm_4x8
;
2796 bool lower_unpack_half_2x16
;
2797 bool lower_unpack_half_2x16_split
;
2798 bool lower_unpack_unorm_2x16
;
2799 bool lower_unpack_snorm_2x16
;
2800 bool lower_unpack_unorm_4x8
;
2801 bool lower_unpack_snorm_4x8
;
2803 bool lower_extract_byte
;
2804 bool lower_extract_word
;
2806 bool lower_all_io_to_temps
;
2807 bool lower_all_io_to_elements
;
2809 /* Indicates that the driver only has zero-based vertex id */
2810 bool vertex_id_zero_based
;
2813 * If enabled, gl_BaseVertex will be lowered as:
2814 * is_indexed_draw (~0/0) & firstvertex
2816 bool lower_base_vertex
;
2819 * If enabled, gl_HelperInvocation will be lowered as:
2821 * !((1 << sample_id) & sample_mask_in))
2823 * This depends on some possibly hw implementation details, which may
2824 * not be true for all hw. In particular that the FS is only executed
2825 * for covered samples or for helper invocations. So, do not blindly
2826 * enable this option.
2828 * Note: See also issue #22 in ARB_shader_image_load_store
2830 bool lower_helper_invocation
;
2833 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
2835 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
2836 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
2838 bool optimize_sample_mask_in
;
2840 bool lower_cs_local_index_from_id
;
2841 bool lower_cs_local_id_from_index
;
2843 bool lower_device_index_to_zero
;
2845 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
2846 bool lower_wpos_pntc
;
2849 * Set if nir_op_[iu]hadd and nir_op_[iu]rhadd instructions should be
2850 * lowered to simple arithmetic.
2852 * If this flag is set, the lowering will be applied to all bit-sizes of
2853 * these instructions.
2855 * \sa ::lower_hadd64
2860 * Set if only 64-bit nir_op_[iu]hadd and nir_op_[iu]rhadd instructions
2861 * should be lowered to simple arithmetic.
2863 * If this flag is set, the lowering will be applied to only 64-bit
2864 * versions of these instructions.
2871 * Set if nir_op_add_sat and nir_op_usub_sat should be lowered to simple
2874 * If this flag is set, the lowering will be applied to all bit-sizes of
2875 * these instructions.
2877 * \sa ::lower_usub_sat64
2882 * Set if only 64-bit nir_op_usub_sat should be lowered to simple
2885 * \sa ::lower_add_sat
2887 bool lower_usub_sat64
;
2890 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
2891 * for IO purposes and would prefer loads/stores be vectorized.
2894 bool lower_to_scalar
;
2897 * Should the linker unify inputs_read/outputs_written between adjacent
2898 * shader stages which are linked into a single program?
2900 bool unify_interfaces
;
2903 * Should nir_lower_io() create load_interpolated_input intrinsics?
2905 * If not, it generates regular load_input intrinsics and interpolation
2906 * information must be inferred from the list of input nir_variables.
2908 bool use_interpolated_input_intrinsics
;
2910 /* Lowers when 32x32->64 bit multiplication is not supported */
2911 bool lower_mul_2x32_64
;
2913 /* Lowers when rotate instruction is not supported */
2917 * Backend supports imul24, and would like to use it (when possible)
2918 * for address/offset calculation. If true, driver should call
2919 * nir_lower_amul(). (If not set, amul will automatically be lowered
2925 * Is this the Intel vec4 backend?
2927 * Used to inhibit algebraic optimizations that are known to be harmful on
2928 * the Intel vec4 backend. This is generally applicable to any
2929 * optimization that might cause more immediate values to be used in
2930 * 3-source (e.g., ffma and flrp) instructions.
2934 unsigned max_unroll_iterations
;
2936 nir_lower_int64_options lower_int64_options
;
2937 nir_lower_doubles_options lower_doubles_options
;
2938 } nir_shader_compiler_options
;
2940 typedef struct nir_shader
{
2941 /** list of uniforms (nir_variable) */
2942 struct exec_list uniforms
;
2944 /** list of inputs (nir_variable) */
2945 struct exec_list inputs
;
2947 /** list of outputs (nir_variable) */
2948 struct exec_list outputs
;
2950 /** list of shared compute variables (nir_variable) */
2951 struct exec_list shared
;
2953 /** Set of driver-specific options for the shader.
2955 * The memory for the options is expected to be kept in a single static
2956 * copy by the driver.
2958 const struct nir_shader_compiler_options
*options
;
2960 /** Various bits of compile-time information about a given shader */
2961 struct shader_info info
;
2963 /** list of global variables in the shader (nir_variable) */
2964 struct exec_list globals
;
2966 /** list of system value variables in the shader (nir_variable) */
2967 struct exec_list system_values
;
2969 struct exec_list functions
; /** < list of nir_function */
2972 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
2975 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
2977 /** Size in bytes of required scratch space */
2978 unsigned scratch_size
;
2980 /** Constant data associated with this shader.
2982 * Constant data is loaded through load_constant intrinsics. See also
2983 * nir_opt_large_constants.
2985 void *constant_data
;
2986 unsigned constant_data_size
;
2989 #define nir_foreach_function(func, shader) \
2990 foreach_list_typed(nir_function, func, node, &(shader)->functions)
2992 static inline nir_function_impl
*
2993 nir_shader_get_entrypoint(nir_shader
*shader
)
2995 nir_function
*func
= NULL
;
2997 nir_foreach_function(function
, shader
) {
2998 assert(func
== NULL
);
2999 if (function
->is_entrypoint
) {
3010 assert(func
->num_params
== 0);
3015 nir_shader
*nir_shader_create(void *mem_ctx
,
3016 gl_shader_stage stage
,
3017 const nir_shader_compiler_options
*options
,
3020 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
3022 void nir_reg_remove(nir_register
*reg
);
3024 /** Adds a variable to the appropriate list in nir_shader */
3025 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
3028 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
3030 assert(var
->data
.mode
== nir_var_function_temp
);
3031 exec_list_push_tail(&impl
->locals
, &var
->node
);
3034 /** creates a variable, sets a few defaults, and adds it to the list */
3035 nir_variable
*nir_variable_create(nir_shader
*shader
,
3036 nir_variable_mode mode
,
3037 const struct glsl_type
*type
,
3039 /** creates a local variable and adds it to the list */
3040 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
3041 const struct glsl_type
*type
,
3044 /** creates a function and adds it to the shader's list of functions */
3045 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
3047 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
3048 /** creates a function_impl that isn't tied to any particular function */
3049 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
3051 nir_block
*nir_block_create(nir_shader
*shader
);
3052 nir_if
*nir_if_create(nir_shader
*shader
);
3053 nir_loop
*nir_loop_create(nir_shader
*shader
);
3055 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
3057 /** requests that the given pieces of metadata be generated */
3058 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
3059 /** dirties all but the preserved metadata */
3060 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
3062 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
3063 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
3065 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
3066 nir_deref_type deref_type
);
3068 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
3070 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
3071 unsigned num_components
,
3074 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
3075 nir_intrinsic_op op
);
3077 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
3078 nir_function
*callee
);
3080 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
3082 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
3084 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
3086 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
3087 unsigned num_components
,
3090 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
3093 * NIR Cursors and Instruction Insertion API
3096 * A tiny struct representing a point to insert/extract instructions or
3097 * control flow nodes. Helps reduce the combinatorial explosion of possible
3098 * points to insert/extract.
3100 * \sa nir_control_flow.h
3103 nir_cursor_before_block
,
3104 nir_cursor_after_block
,
3105 nir_cursor_before_instr
,
3106 nir_cursor_after_instr
,
3107 } nir_cursor_option
;
3110 nir_cursor_option option
;
3117 static inline nir_block
*
3118 nir_cursor_current_block(nir_cursor cursor
)
3120 if (cursor
.option
== nir_cursor_before_instr
||
3121 cursor
.option
== nir_cursor_after_instr
) {
3122 return cursor
.instr
->block
;
3124 return cursor
.block
;
3128 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
3130 static inline nir_cursor
3131 nir_before_block(nir_block
*block
)
3134 cursor
.option
= nir_cursor_before_block
;
3135 cursor
.block
= block
;
3139 static inline nir_cursor
3140 nir_after_block(nir_block
*block
)
3143 cursor
.option
= nir_cursor_after_block
;
3144 cursor
.block
= block
;
3148 static inline nir_cursor
3149 nir_before_instr(nir_instr
*instr
)
3152 cursor
.option
= nir_cursor_before_instr
;
3153 cursor
.instr
= instr
;
3157 static inline nir_cursor
3158 nir_after_instr(nir_instr
*instr
)
3161 cursor
.option
= nir_cursor_after_instr
;
3162 cursor
.instr
= instr
;
3166 static inline nir_cursor
3167 nir_after_block_before_jump(nir_block
*block
)
3169 nir_instr
*last_instr
= nir_block_last_instr(block
);
3170 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
3171 return nir_before_instr(last_instr
);
3173 return nir_after_block(block
);
3177 static inline nir_cursor
3178 nir_before_src(nir_src
*src
, bool is_if_condition
)
3180 if (is_if_condition
) {
3181 nir_block
*prev_block
=
3182 nir_cf_node_as_block(nir_cf_node_prev(&src
->parent_if
->cf_node
));
3183 assert(!nir_block_ends_in_jump(prev_block
));
3184 return nir_after_block(prev_block
);
3185 } else if (src
->parent_instr
->type
== nir_instr_type_phi
) {
3187 nir_phi_instr
*cond_phi
= nir_instr_as_phi(src
->parent_instr
);
3189 nir_foreach_phi_src(phi_src
, cond_phi
) {
3190 if (phi_src
->src
.ssa
== src
->ssa
) {
3197 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
3198 * to have a more specific name.
3200 nir_phi_src
*phi_src
= LIST_ENTRY(nir_phi_src
, src
, src
);
3201 return nir_after_block_before_jump(phi_src
->pred
);
3203 return nir_before_instr(src
->parent_instr
);
3207 static inline nir_cursor
3208 nir_before_cf_node(nir_cf_node
*node
)
3210 if (node
->type
== nir_cf_node_block
)
3211 return nir_before_block(nir_cf_node_as_block(node
));
3213 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
3216 static inline nir_cursor
3217 nir_after_cf_node(nir_cf_node
*node
)
3219 if (node
->type
== nir_cf_node_block
)
3220 return nir_after_block(nir_cf_node_as_block(node
));
3222 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
3225 static inline nir_cursor
3226 nir_after_phis(nir_block
*block
)
3228 nir_foreach_instr(instr
, block
) {
3229 if (instr
->type
!= nir_instr_type_phi
)
3230 return nir_before_instr(instr
);
3232 return nir_after_block(block
);
3235 static inline nir_cursor
3236 nir_after_cf_node_and_phis(nir_cf_node
*node
)
3238 if (node
->type
== nir_cf_node_block
)
3239 return nir_after_block(nir_cf_node_as_block(node
));
3241 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
3243 return nir_after_phis(block
);
3246 static inline nir_cursor
3247 nir_before_cf_list(struct exec_list
*cf_list
)
3249 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
3250 exec_list_get_head(cf_list
), node
);
3251 return nir_before_cf_node(first_node
);
3254 static inline nir_cursor
3255 nir_after_cf_list(struct exec_list
*cf_list
)
3257 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
3258 exec_list_get_tail(cf_list
), node
);
3259 return nir_after_cf_node(last_node
);
3263 * Insert a NIR instruction at the given cursor.
3265 * Note: This does not update the cursor.
3267 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
3270 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
3272 nir_instr_insert(nir_before_instr(instr
), before
);
3276 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
3278 nir_instr_insert(nir_after_instr(instr
), after
);
3282 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
3284 nir_instr_insert(nir_before_block(block
), before
);
3288 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
3290 nir_instr_insert(nir_after_block(block
), after
);
3294 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
3296 nir_instr_insert(nir_before_cf_node(node
), before
);
3300 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
3302 nir_instr_insert(nir_after_cf_node(node
), after
);
3306 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
3308 nir_instr_insert(nir_before_cf_list(list
), before
);
3312 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
3314 nir_instr_insert(nir_after_cf_list(list
), after
);
3317 void nir_instr_remove_v(nir_instr
*instr
);
3319 static inline nir_cursor
3320 nir_instr_remove(nir_instr
*instr
)
3323 nir_instr
*prev
= nir_instr_prev(instr
);
3325 cursor
= nir_after_instr(prev
);
3327 cursor
= nir_before_block(instr
->block
);
3329 nir_instr_remove_v(instr
);
3335 nir_ssa_def
*nir_instr_ssa_def(nir_instr
*instr
);
3337 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
3338 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
3339 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
3340 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
3342 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
3343 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
3345 nir_const_value
*nir_src_as_const_value(nir_src src
);
3347 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
3348 static inline c_type * \
3349 nir_src_as_ ## name (nir_src src) \
3351 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
3352 ? cast_macro(src.ssa->parent_instr) : NULL; \
3355 NIR_SRC_AS_(alu_instr
, nir_alu_instr
, nir_instr_type_alu
, nir_instr_as_alu
)
3356 NIR_SRC_AS_(intrinsic
, nir_intrinsic_instr
,
3357 nir_instr_type_intrinsic
, nir_instr_as_intrinsic
)
3358 NIR_SRC_AS_(deref
, nir_deref_instr
, nir_instr_type_deref
, nir_instr_as_deref
)
3360 bool nir_src_is_dynamically_uniform(nir_src src
);
3361 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
3362 bool nir_instrs_equal(const nir_instr
*instr1
, const nir_instr
*instr2
);
3363 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
3364 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
3365 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
3366 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
3369 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
3370 unsigned num_components
, unsigned bit_size
,
3372 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
3373 unsigned num_components
, unsigned bit_size
,
3376 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
3377 const struct glsl_type
*type
,
3380 assert(glsl_type_is_vector_or_scalar(type
));
3381 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
3382 glsl_get_bit_size(type
), name
);
3384 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
3385 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
3386 nir_instr
*after_me
);
3388 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
3391 * finds the next basic block in source-code order, returns NULL if there is
3395 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
3397 /* Performs the opposite of nir_block_cf_tree_next() */
3399 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
3401 /* Gets the first block in a CF node in source-code order */
3403 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
3405 /* Gets the last block in a CF node in source-code order */
3407 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
3409 /* Gets the next block after a CF node in source-code order */
3411 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
3413 /* Macros for loops that visit blocks in source-code order */
3415 #define nir_foreach_block(block, impl) \
3416 for (nir_block *block = nir_start_block(impl); block != NULL; \
3417 block = nir_block_cf_tree_next(block))
3419 #define nir_foreach_block_safe(block, impl) \
3420 for (nir_block *block = nir_start_block(impl), \
3421 *next = nir_block_cf_tree_next(block); \
3423 block = next, next = nir_block_cf_tree_next(block))
3425 #define nir_foreach_block_reverse(block, impl) \
3426 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
3427 block = nir_block_cf_tree_prev(block))
3429 #define nir_foreach_block_reverse_safe(block, impl) \
3430 for (nir_block *block = nir_impl_last_block(impl), \
3431 *prev = nir_block_cf_tree_prev(block); \
3433 block = prev, prev = nir_block_cf_tree_prev(block))
3435 #define nir_foreach_block_in_cf_node(block, node) \
3436 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
3437 block != nir_cf_node_cf_tree_next(node); \
3438 block = nir_block_cf_tree_next(block))
3440 /* If the following CF node is an if, this function returns that if.
3441 * Otherwise, it returns NULL.
3443 nir_if
*nir_block_get_following_if(nir_block
*block
);
3445 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
3447 void nir_index_local_regs(nir_function_impl
*impl
);
3448 void nir_index_ssa_defs(nir_function_impl
*impl
);
3449 unsigned nir_index_instrs(nir_function_impl
*impl
);
3451 void nir_index_blocks(nir_function_impl
*impl
);
3453 void nir_index_vars(nir_shader
*shader
, nir_function_impl
*impl
, nir_variable_mode modes
);
3455 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
3456 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
3457 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
3458 void nir_print_deref(const nir_deref_instr
*deref
, FILE *fp
);
3460 /** Shallow clone of a single ALU instruction. */
3461 nir_alu_instr
*nir_alu_instr_clone(nir_shader
*s
, const nir_alu_instr
*orig
);
3463 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
3464 nir_function_impl
*nir_function_impl_clone(nir_shader
*shader
,
3465 const nir_function_impl
*fi
);
3466 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
3467 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
3469 void nir_shader_replace(nir_shader
*dest
, nir_shader
*src
);
3471 void nir_shader_serialize_deserialize(nir_shader
*s
);
3474 void nir_validate_shader(nir_shader
*shader
, const char *when
);
3475 void nir_metadata_set_validation_flag(nir_shader
*shader
);
3476 void nir_metadata_check_validation_flag(nir_shader
*shader
);
3479 should_skip_nir(const char *name
)
3481 static const char *list
= NULL
;
3483 /* Comma separated list of names to skip. */
3484 list
= getenv("NIR_SKIP");
3492 return comma_separated_list_contains(list
, name
);
3496 should_clone_nir(void)
3498 static int should_clone
= -1;
3499 if (should_clone
< 0)
3500 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
3502 return should_clone
;
3506 should_serialize_deserialize_nir(void)
3508 static int test_serialize
= -1;
3509 if (test_serialize
< 0)
3510 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
3512 return test_serialize
;
3516 should_print_nir(void)
3518 static int should_print
= -1;
3519 if (should_print
< 0)
3520 should_print
= env_var_as_boolean("NIR_PRINT", false);
3522 return should_print
;
3525 static inline void nir_validate_shader(nir_shader
*shader
, const char *when
) { (void) shader
; (void)when
; }
3526 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
3527 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
3528 static inline bool should_skip_nir(UNUSED
const char *pass_name
) { return false; }
3529 static inline bool should_clone_nir(void) { return false; }
3530 static inline bool should_serialize_deserialize_nir(void) { return false; }
3531 static inline bool should_print_nir(void) { return false; }
3534 #define _PASS(pass, nir, do_pass) do { \
3535 if (should_skip_nir(#pass)) { \
3536 printf("skipping %s\n", #pass); \
3540 nir_validate_shader(nir, "after " #pass); \
3541 if (should_clone_nir()) { \
3542 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3543 nir_shader_replace(nir, clone); \
3545 if (should_serialize_deserialize_nir()) { \
3546 nir_shader_serialize_deserialize(nir); \
3550 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3551 nir_metadata_set_validation_flag(nir); \
3552 if (should_print_nir()) \
3553 printf("%s\n", #pass); \
3554 if (pass(nir, ##__VA_ARGS__)) { \
3556 if (should_print_nir()) \
3557 nir_print_shader(nir, stdout); \
3558 nir_metadata_check_validation_flag(nir); \
3562 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3563 if (should_print_nir()) \
3564 printf("%s\n", #pass); \
3565 pass(nir, ##__VA_ARGS__); \
3566 if (should_print_nir()) \
3567 nir_print_shader(nir, stdout); \
3570 #define NIR_SKIP(name) should_skip_nir(#name)
3572 /** An instruction filtering callback
3574 * Returns true if the instruction should be processed and false otherwise.
3576 typedef bool (*nir_instr_filter_cb
)(const nir_instr
*, const void *);
3578 /** A simple instruction lowering callback
3580 * Many instruction lowering passes can be written as a simple function which
3581 * takes an instruction as its input and returns a sequence of instructions
3582 * that implement the consumed instruction. This function type represents
3583 * such a lowering function. When called, a function with this prototype
3584 * should either return NULL indicating that no lowering needs to be done or
3585 * emit a sequence of instructions using the provided builder (whose cursor
3586 * will already be placed after the instruction to be lowered) and return the
3587 * resulting nir_ssa_def.
3589 typedef nir_ssa_def
*(*nir_lower_instr_cb
)(struct nir_builder
*,
3590 nir_instr
*, void *);
3593 * Special return value for nir_lower_instr_cb when some progress occurred
3594 * (like changing an input to the instr) that didn't result in a replacement
3595 * SSA def being generated.
3597 #define NIR_LOWER_INSTR_PROGRESS ((nir_ssa_def *)(uintptr_t)1)
3599 /** Iterate over all the instructions in a nir_function_impl and lower them
3600 * using the provided callbacks
3602 * This function implements the guts of a standard lowering pass for you. It
3603 * iterates over all of the instructions in a nir_function_impl and calls the
3604 * filter callback on each one. If the filter callback returns true, it then
3605 * calls the lowering call back on the instruction. (Splitting it this way
3606 * allows us to avoid some save/restore work for instructions we know won't be
3607 * lowered.) If the instruction is dead after the lowering is complete, it
3608 * will be removed. If new instructions are added, the lowering callback will
3609 * also be called on them in case multiple lowerings are required.
3611 * The metadata for the nir_function_impl will also be updated. If any blocks
3612 * are added (they cannot be removed), dominance and block indices will be
3615 bool nir_function_impl_lower_instructions(nir_function_impl
*impl
,
3616 nir_instr_filter_cb filter
,
3617 nir_lower_instr_cb lower
,
3619 bool nir_shader_lower_instructions(nir_shader
*shader
,
3620 nir_instr_filter_cb filter
,
3621 nir_lower_instr_cb lower
,
3624 void nir_calc_dominance_impl(nir_function_impl
*impl
);
3625 void nir_calc_dominance(nir_shader
*shader
);
3627 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
3628 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
3629 bool nir_block_is_unreachable(nir_block
*block
);
3631 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
3632 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
3634 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
3635 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
3637 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
3638 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
3640 int nir_gs_count_vertices(const nir_shader
*shader
);
3642 bool nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3643 bool nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3644 bool nir_split_var_copies(nir_shader
*shader
);
3645 bool nir_split_per_member_structs(nir_shader
*shader
);
3646 bool nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
);
3648 bool nir_lower_returns_impl(nir_function_impl
*impl
);
3649 bool nir_lower_returns(nir_shader
*shader
);
3651 void nir_inline_function_impl(struct nir_builder
*b
,
3652 const nir_function_impl
*impl
,
3653 nir_ssa_def
**params
);
3654 bool nir_inline_functions(nir_shader
*shader
);
3656 bool nir_propagate_invariant(nir_shader
*shader
);
3658 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
3659 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
3660 nir_intrinsic_instr
*copy
);
3661 bool nir_lower_var_copies(nir_shader
*shader
);
3663 void nir_fixup_deref_modes(nir_shader
*shader
);
3665 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
3668 nir_lower_direct_array_deref_of_vec_load
= (1 << 0),
3669 nir_lower_indirect_array_deref_of_vec_load
= (1 << 1),
3670 nir_lower_direct_array_deref_of_vec_store
= (1 << 2),
3671 nir_lower_indirect_array_deref_of_vec_store
= (1 << 3),
3672 } nir_lower_array_deref_of_vec_options
;
3674 bool nir_lower_array_deref_of_vec(nir_shader
*shader
, nir_variable_mode modes
,
3675 nir_lower_array_deref_of_vec_options options
);
3677 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
3679 bool nir_lower_locals_to_regs(nir_shader
*shader
);
3681 void nir_lower_io_to_temporaries(nir_shader
*shader
,
3682 nir_function_impl
*entrypoint
,
3683 bool outputs
, bool inputs
);
3685 bool nir_lower_vars_to_scratch(nir_shader
*shader
,
3686 nir_variable_mode modes
,
3688 glsl_type_size_align_func size_align
);
3690 void nir_lower_clip_halfz(nir_shader
*shader
);
3692 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
3694 void nir_gather_ssa_types(nir_function_impl
*impl
,
3695 BITSET_WORD
*float_types
,
3696 BITSET_WORD
*int_types
);
3698 void nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
3699 int (*type_size
)(const struct glsl_type
*, bool));
3701 /* Some helpers to do very simple linking */
3702 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3703 bool nir_remove_unused_io_vars(nir_shader
*shader
, struct exec_list
*var_list
,
3704 uint64_t *used_by_other_stage
,
3705 uint64_t *used_by_other_stage_patches
);
3706 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
3707 bool default_to_smooth_interp
);
3708 void nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3709 bool nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3711 bool nir_lower_amul(nir_shader
*shader
,
3712 int (*type_size
)(const struct glsl_type
*, bool));
3714 void nir_assign_io_var_locations(struct exec_list
*var_list
,
3716 gl_shader_stage stage
);
3719 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
3720 * to 32-bit operations. This is only valid for nir_var_shader_in/out
3723 nir_lower_io_lower_64bit_to_32
= (1 << 0),
3725 /* If set, this forces all non-flat fragment shader inputs to be
3726 * interpolated as if with the "sample" qualifier. This requires
3727 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
3729 nir_lower_io_force_sample_interpolation
= (1 << 1),
3730 } nir_lower_io_options
;
3731 bool nir_lower_io(nir_shader
*shader
,
3732 nir_variable_mode modes
,
3733 int (*type_size
)(const struct glsl_type
*, bool),
3734 nir_lower_io_options
);
3736 bool nir_io_add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
);
3739 nir_lower_vars_to_explicit_types(nir_shader
*shader
,
3740 nir_variable_mode modes
,
3741 glsl_type_size_align_func type_info
);
3745 * An address format which is a simple 32-bit global GPU address.
3747 nir_address_format_32bit_global
,
3750 * An address format which is a simple 64-bit global GPU address.
3752 nir_address_format_64bit_global
,
3755 * An address format which is a bounds-checked 64-bit global GPU address.
3757 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
3758 * address stored with the low bits in .x and high bits in .y, .z is a
3759 * size, and .w is an offset. When the final I/O operation is lowered, .w
3760 * is checked against .z and the operation is predicated on the result.
3762 nir_address_format_64bit_bounded_global
,
3765 * An address format which is comprised of a vec2 where the first
3766 * component is a buffer index and the second is an offset.
3768 nir_address_format_32bit_index_offset
,
3771 * An address format which is a simple 32-bit offset.
3773 nir_address_format_32bit_offset
,
3776 * An address format representing a purely logical addressing model. In
3777 * this model, all deref chains must be complete from the dereference
3778 * operation to the variable. Cast derefs are not allowed. These
3779 * addresses will be 32-bit scalars but the format is immaterial because
3780 * you can always chase the chain.
3782 nir_address_format_logical
,
3783 } nir_address_format
;
3785 static inline unsigned
3786 nir_address_format_bit_size(nir_address_format addr_format
)
3788 switch (addr_format
) {
3789 case nir_address_format_32bit_global
: return 32;
3790 case nir_address_format_64bit_global
: return 64;
3791 case nir_address_format_64bit_bounded_global
: return 32;
3792 case nir_address_format_32bit_index_offset
: return 32;
3793 case nir_address_format_32bit_offset
: return 32;
3794 case nir_address_format_logical
: return 32;
3796 unreachable("Invalid address format");
3799 static inline unsigned
3800 nir_address_format_num_components(nir_address_format addr_format
)
3802 switch (addr_format
) {
3803 case nir_address_format_32bit_global
: return 1;
3804 case nir_address_format_64bit_global
: return 1;
3805 case nir_address_format_64bit_bounded_global
: return 4;
3806 case nir_address_format_32bit_index_offset
: return 2;
3807 case nir_address_format_32bit_offset
: return 1;
3808 case nir_address_format_logical
: return 1;
3810 unreachable("Invalid address format");
3813 static inline const struct glsl_type
*
3814 nir_address_format_to_glsl_type(nir_address_format addr_format
)
3816 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
3817 assert(bit_size
== 32 || bit_size
== 64);
3818 return glsl_vector_type(bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
,
3819 nir_address_format_num_components(addr_format
));
3822 const nir_const_value
*nir_address_format_null_value(nir_address_format addr_format
);
3824 nir_ssa_def
*nir_build_addr_ieq(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3825 nir_address_format addr_format
);
3827 nir_ssa_def
*nir_build_addr_isub(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3828 nir_address_format addr_format
);
3830 nir_ssa_def
* nir_explicit_io_address_from_deref(struct nir_builder
*b
,
3831 nir_deref_instr
*deref
,
3832 nir_ssa_def
*base_addr
,
3833 nir_address_format addr_format
);
3834 void nir_lower_explicit_io_instr(struct nir_builder
*b
,
3835 nir_intrinsic_instr
*io_instr
,
3837 nir_address_format addr_format
);
3839 bool nir_lower_explicit_io(nir_shader
*shader
,
3840 nir_variable_mode modes
,
3841 nir_address_format
);
3843 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
3844 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
3846 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
3848 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
3849 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
3850 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
3852 bool nir_remove_dead_derefs(nir_shader
*shader
);
3853 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
3854 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
);
3855 bool nir_lower_variable_initializers(nir_shader
*shader
,
3856 nir_variable_mode modes
);
3858 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
3859 bool nir_lower_vec_to_movs(nir_shader
*shader
);
3860 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
3862 const gl_state_index16
*alpha_ref_state_tokens
);
3863 bool nir_lower_alu(nir_shader
*shader
);
3865 bool nir_lower_flrp(nir_shader
*shader
, unsigned lowering_mask
,
3866 bool always_precise
, bool have_ffma
);
3868 bool nir_lower_alu_to_scalar(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
3869 bool nir_lower_bool_to_float(nir_shader
*shader
);
3870 bool nir_lower_bool_to_int32(nir_shader
*shader
);
3871 bool nir_lower_int_to_float(nir_shader
*shader
);
3872 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
3873 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
3874 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
3875 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
3876 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
3878 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
3879 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
3880 bool nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode mask
);
3882 void nir_lower_fragcoord_wtrans(nir_shader
*shader
);
3883 void nir_lower_viewport_transform(nir_shader
*shader
);
3884 bool nir_lower_uniforms_to_ubo(nir_shader
*shader
, int multiplier
);
3886 typedef struct nir_lower_subgroups_options
{
3887 uint8_t subgroup_size
;
3888 uint8_t ballot_bit_size
;
3889 bool lower_to_scalar
:1;
3890 bool lower_vote_trivial
:1;
3891 bool lower_vote_eq_to_ballot
:1;
3892 bool lower_subgroup_masks
:1;
3893 bool lower_shuffle
:1;
3894 bool lower_shuffle_to_32bit
:1;
3896 bool lower_quad_broadcast_dynamic
:1;
3897 } nir_lower_subgroups_options
;
3899 bool nir_lower_subgroups(nir_shader
*shader
,
3900 const nir_lower_subgroups_options
*options
);
3902 bool nir_lower_system_values(nir_shader
*shader
);
3904 enum PACKED nir_lower_tex_packing
{
3905 nir_lower_tex_packing_none
= 0,
3906 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
3907 * or unsigned ints based on the sampler type
3909 nir_lower_tex_packing_16
,
3910 /* The sampler returns 1 32-bit word of 4x8 unorm */
3911 nir_lower_tex_packing_8
,
3914 typedef struct nir_lower_tex_options
{
3916 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
3917 * sampler types a texture projector is lowered.
3922 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
3924 bool lower_txf_offset
;
3927 * If true, lower away nir_tex_src_offset for all rect textures.
3929 bool lower_rect_offset
;
3932 * If true, lower rect textures to 2D, using txs to fetch the
3933 * texture dimensions and dividing the texture coords by the
3934 * texture dims to normalize.
3939 * If true, convert yuv to rgb.
3941 unsigned lower_y_uv_external
;
3942 unsigned lower_y_u_v_external
;
3943 unsigned lower_yx_xuxv_external
;
3944 unsigned lower_xy_uxvx_external
;
3945 unsigned lower_ayuv_external
;
3946 unsigned lower_xyuv_external
;
3949 * To emulate certain texture wrap modes, this can be used
3950 * to saturate the specified tex coord to [0.0, 1.0]. The
3951 * bits are according to sampler #, ie. if, for example:
3953 * (conf->saturate_s & (1 << n))
3955 * is true, then the s coord for sampler n is saturated.
3957 * Note that clamping must happen *after* projector lowering
3958 * so any projected texture sample instruction with a clamped
3959 * coordinate gets automatically lowered, regardless of the
3960 * 'lower_txp' setting.
3962 unsigned saturate_s
;
3963 unsigned saturate_t
;
3964 unsigned saturate_r
;
3966 /* Bitmask of textures that need swizzling.
3968 * If (swizzle_result & (1 << texture_index)), then the swizzle in
3969 * swizzles[texture_index] is applied to the result of the texturing
3972 unsigned swizzle_result
;
3974 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
3975 * while 4 and 5 represent 0 and 1 respectively.
3977 uint8_t swizzles
[32][4];
3979 /* Can be used to scale sampled values in range required by the format. */
3980 float scale_factors
[32];
3983 * Bitmap of textures that need srgb to linear conversion. If
3984 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
3985 * of the texture are lowered to linear.
3987 unsigned lower_srgb
;
3990 * If true, lower nir_texop_tex on shaders that doesn't support implicit
3991 * LODs to nir_texop_txl.
3993 bool lower_tex_without_implicit_lod
;
3996 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
3998 bool lower_txd_cube_map
;
4001 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
4006 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
4007 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
4008 * with lower_txd_cube_map.
4010 bool lower_txd_shadow
;
4013 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
4014 * Implies lower_txd_cube_map and lower_txd_shadow.
4019 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
4020 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
4022 bool lower_txb_shadow_clamp
;
4025 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
4026 * with nir_texop_txl. This includes cube maps.
4028 bool lower_txd_shadow_clamp
;
4031 * If true, lower nir_texop_txd on when it uses both offset and min_lod
4032 * with nir_texop_txl. This includes cube maps.
4034 bool lower_txd_offset_clamp
;
4037 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4038 * sampler is bindless.
4040 bool lower_txd_clamp_bindless_sampler
;
4043 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4044 * sampler index is not statically determinable to be less than 16.
4046 bool lower_txd_clamp_if_sampler_index_not_lt_16
;
4049 * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
4050 * 0-lod followed by a nir_ishr.
4055 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
4056 * mixed-up tg4 locations.
4058 bool lower_tg4_broadcom_swizzle
;
4061 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
4063 bool lower_tg4_offsets
;
4065 enum nir_lower_tex_packing lower_tex_packing
[32];
4066 } nir_lower_tex_options
;
4068 bool nir_lower_tex(nir_shader
*shader
,
4069 const nir_lower_tex_options
*options
);
4071 enum nir_lower_non_uniform_access_type
{
4072 nir_lower_non_uniform_ubo_access
= (1 << 0),
4073 nir_lower_non_uniform_ssbo_access
= (1 << 1),
4074 nir_lower_non_uniform_texture_access
= (1 << 2),
4075 nir_lower_non_uniform_image_access
= (1 << 3),
4078 bool nir_lower_non_uniform_access(nir_shader
*shader
,
4079 enum nir_lower_non_uniform_access_type
);
4081 enum nir_lower_idiv_path
{
4082 /* This path is based on NV50LegalizeSSA::handleDIV(). It is the faster of
4083 * the two but it is not exact in some cases (for example, 1091317713u /
4084 * 1034u gives 5209173 instead of 1055432) */
4085 nir_lower_idiv_fast
,
4086 /* This path is based on AMDGPUTargetLowering::LowerUDIVREM() and
4087 * AMDGPUTargetLowering::LowerSDIVREM(). It requires more instructions than
4088 * the nv50 path and many of them are integer multiplications, so it is
4089 * probably slower. It should always return the correct result, though. */
4090 nir_lower_idiv_precise
,
4093 bool nir_lower_idiv(nir_shader
*shader
, enum nir_lower_idiv_path path
);
4095 bool nir_lower_input_attachments(nir_shader
*shader
, bool use_fragcoord_sysval
);
4097 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
,
4099 bool use_clipdist_array
,
4100 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4101 bool nir_lower_clip_gs(nir_shader
*shader
, unsigned ucp_enables
,
4102 bool use_clipdist_array
,
4103 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4104 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
,
4105 bool use_clipdist_array
);
4106 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
4108 void nir_lower_point_size_mov(nir_shader
*shader
,
4109 const gl_state_index16
*pointsize_state_tokens
);
4111 bool nir_lower_frexp(nir_shader
*nir
);
4113 void nir_lower_two_sided_color(nir_shader
*shader
);
4115 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
4117 bool nir_lower_flatshade(nir_shader
*shader
);
4119 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
4120 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
4121 const gl_state_index16
*uniform_state_tokens
);
4123 typedef struct nir_lower_wpos_ytransform_options
{
4124 gl_state_index16 state_tokens
[STATE_LENGTH
];
4125 bool fs_coord_origin_upper_left
:1;
4126 bool fs_coord_origin_lower_left
:1;
4127 bool fs_coord_pixel_center_integer
:1;
4128 bool fs_coord_pixel_center_half_integer
:1;
4129 } nir_lower_wpos_ytransform_options
;
4131 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
4132 const nir_lower_wpos_ytransform_options
*options
);
4133 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
4135 bool nir_lower_fb_read(nir_shader
*shader
);
4137 typedef struct nir_lower_drawpixels_options
{
4138 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
4139 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
4140 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
4141 unsigned drawpix_sampler
;
4142 unsigned pixelmap_sampler
;
4144 bool scale_and_bias
:1;
4145 } nir_lower_drawpixels_options
;
4147 void nir_lower_drawpixels(nir_shader
*shader
,
4148 const nir_lower_drawpixels_options
*options
);
4150 typedef struct nir_lower_bitmap_options
{
4153 } nir_lower_bitmap_options
;
4155 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
4157 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
);
4160 nir_lower_int_source_mods
= 1 << 0,
4161 nir_lower_float_source_mods
= 1 << 1,
4162 nir_lower_triop_abs
= 1 << 2,
4163 nir_lower_all_source_mods
= (1 << 3) - 1
4164 } nir_lower_to_source_mods_flags
;
4167 bool nir_lower_to_source_mods(nir_shader
*shader
, nir_lower_to_source_mods_flags options
);
4169 bool nir_lower_gs_intrinsics(nir_shader
*shader
, bool per_stream
);
4171 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
4173 bool nir_lower_bit_size(nir_shader
*shader
,
4174 nir_lower_bit_size_callback callback
,
4175 void *callback_data
);
4177 nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode
);
4178 bool nir_lower_int64(nir_shader
*shader
, nir_lower_int64_options options
);
4180 nir_lower_doubles_options
nir_lower_doubles_op_to_options_mask(nir_op opcode
);
4181 bool nir_lower_doubles(nir_shader
*shader
, const nir_shader
*softfp64
,
4182 nir_lower_doubles_options options
);
4183 bool nir_lower_pack(nir_shader
*shader
);
4185 bool nir_lower_point_size(nir_shader
*shader
, float min
, float max
);
4188 nir_lower_interpolation_at_sample
= (1 << 1),
4189 nir_lower_interpolation_at_offset
= (1 << 2),
4190 nir_lower_interpolation_centroid
= (1 << 3),
4191 nir_lower_interpolation_pixel
= (1 << 4),
4192 nir_lower_interpolation_sample
= (1 << 5),
4193 } nir_lower_interpolation_options
;
4195 bool nir_lower_interpolation(nir_shader
*shader
,
4196 nir_lower_interpolation_options options
);
4198 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
4200 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
4202 void nir_loop_analyze_impl(nir_function_impl
*impl
,
4203 nir_variable_mode indirect_mask
);
4205 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
4207 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
4208 bool nir_repair_ssa(nir_shader
*shader
);
4210 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
4211 bool nir_convert_to_lcssa(nir_shader
*shader
, bool skip_invariants
, bool skip_bool_invariants
);
4212 bool* nir_divergence_analysis(nir_shader
*shader
, nir_divergence_options options
);
4214 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
4215 * registers. If false, convert all values (even those not involved in a phi
4216 * node) to registers.
4218 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
4220 bool nir_lower_phis_to_regs_block(nir_block
*block
);
4221 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
4222 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
);
4224 bool nir_lower_samplers(nir_shader
*shader
);
4225 bool nir_lower_ssbo(nir_shader
*shader
);
4227 /* This is here for unit tests. */
4228 bool nir_opt_comparison_pre_impl(nir_function_impl
*impl
);
4230 bool nir_opt_comparison_pre(nir_shader
*shader
);
4232 bool nir_opt_access(nir_shader
*shader
);
4233 bool nir_opt_algebraic(nir_shader
*shader
);
4234 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
4235 bool nir_opt_algebraic_late(nir_shader
*shader
);
4236 bool nir_opt_constant_folding(nir_shader
*shader
);
4238 bool nir_opt_combine_stores(nir_shader
*shader
, nir_variable_mode modes
);
4240 bool nir_copy_prop(nir_shader
*shader
);
4242 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
4244 bool nir_opt_cse(nir_shader
*shader
);
4246 bool nir_opt_dce(nir_shader
*shader
);
4248 bool nir_opt_dead_cf(nir_shader
*shader
);
4250 bool nir_opt_dead_write_vars(nir_shader
*shader
);
4252 bool nir_opt_deref_impl(nir_function_impl
*impl
);
4253 bool nir_opt_deref(nir_shader
*shader
);
4255 bool nir_opt_find_array_copies(nir_shader
*shader
);
4257 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
4259 bool nir_opt_idiv_const(nir_shader
*shader
, unsigned min_bit_size
);
4261 bool nir_opt_if(nir_shader
*shader
, bool aggressive_last_continue
);
4263 bool nir_opt_intrinsics(nir_shader
*shader
);
4265 bool nir_opt_large_constants(nir_shader
*shader
,
4266 glsl_type_size_align_func size_align
,
4267 unsigned threshold
);
4269 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
4272 nir_move_const_undef
= (1 << 0),
4273 nir_move_load_ubo
= (1 << 1),
4274 nir_move_load_input
= (1 << 2),
4275 nir_move_comparisons
= (1 << 3),
4276 nir_move_copies
= (1 << 4),
4279 bool nir_can_move_instr(nir_instr
*instr
, nir_move_options options
);
4281 bool nir_opt_sink(nir_shader
*shader
, nir_move_options options
);
4283 bool nir_opt_move(nir_shader
*shader
, nir_move_options options
);
4285 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
,
4286 bool indirect_load_ok
, bool expensive_alu_ok
);
4288 bool nir_opt_rematerialize_compares(nir_shader
*shader
);
4290 bool nir_opt_remove_phis(nir_shader
*shader
);
4291 bool nir_opt_remove_phis_block(nir_block
*block
);
4293 bool nir_opt_shrink_load(nir_shader
*shader
);
4295 bool nir_opt_trivial_continues(nir_shader
*shader
);
4297 bool nir_opt_undef(nir_shader
*shader
);
4299 bool nir_opt_vectorize(nir_shader
*shader
);
4301 bool nir_opt_conditional_discard(nir_shader
*shader
);
4303 typedef bool (*nir_should_vectorize_mem_func
)(unsigned align
, unsigned bit_size
,
4304 unsigned num_components
, unsigned high_offset
,
4305 nir_intrinsic_instr
*low
, nir_intrinsic_instr
*high
);
4307 bool nir_opt_load_store_vectorize(nir_shader
*shader
, nir_variable_mode modes
,
4308 nir_should_vectorize_mem_func callback
);
4310 void nir_schedule(nir_shader
*shader
, int threshold
);
4312 void nir_strip(nir_shader
*shader
);
4314 void nir_sweep(nir_shader
*shader
);
4316 void nir_remap_dual_slot_attributes(nir_shader
*shader
,
4317 uint64_t *dual_slot_inputs
);
4318 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
);
4320 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
4321 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);
4324 nir_variable_is_in_ubo(const nir_variable
*var
)
4326 return (var
->data
.mode
== nir_var_mem_ubo
&&
4327 var
->interface_type
!= NULL
);
4331 nir_variable_is_in_ssbo(const nir_variable
*var
)
4333 return (var
->data
.mode
== nir_var_mem_ssbo
&&
4334 var
->interface_type
!= NULL
);
4338 nir_variable_is_in_block(const nir_variable
*var
)
4340 return nir_variable_is_in_ubo(var
) || nir_variable_is_in_ssbo(var
);