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 /* Are all components the same, ie. .xxxx */
886 nir_is_same_comp_swizzle(uint8_t *swiz
, unsigned nr_comp
)
888 for (unsigned i
= 1; i
< nr_comp
; i
++)
889 if (swiz
[i
] != swiz
[0])
894 /* Are all components sequential, ie. .yzw */
896 nir_is_sequential_comp_swizzle(uint8_t *swiz
, unsigned nr_comp
)
898 for (unsigned i
= 1; i
< nr_comp
; i
++)
899 if (swiz
[i
] != (swiz
[0] + i
))
904 void nir_src_copy(nir_src
*dest
, const nir_src
*src
, void *instr_or_if
);
905 void nir_dest_copy(nir_dest
*dest
, const nir_dest
*src
, nir_instr
*instr
);
911 * \name input modifiers
915 * For inputs interpreted as floating point, flips the sign bit. For
916 * inputs interpreted as integers, performs the two's complement negation.
921 * Clears the sign bit for floating point values, and computes the integer
922 * absolute value for integers. Note that the negate modifier acts after
923 * the absolute value modifier, therefore if both are set then all inputs
924 * will become negative.
930 * For each input component, says which component of the register it is
931 * chosen from. Note that which elements of the swizzle are used and which
932 * are ignored are based on the write mask for most opcodes - for example,
933 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
934 * a swizzle of {2, x, 1, 0} where x means "don't care."
936 uint8_t swizzle
[NIR_MAX_VEC_COMPONENTS
];
943 * \name saturate output modifier
945 * Only valid for opcodes that output floating-point numbers. Clamps the
946 * output to between 0.0 and 1.0 inclusive.
951 unsigned write_mask
: NIR_MAX_VEC_COMPONENTS
; /* ignored if dest.is_ssa is true */
954 /** NIR sized and unsized types
956 * The values in this enum are carefully chosen so that the sized type is
957 * just the unsized type OR the number of bits.
960 nir_type_invalid
= 0, /* Not a valid type */
964 nir_type_float
= 128,
965 nir_type_bool1
= 1 | nir_type_bool
,
966 nir_type_bool8
= 8 | nir_type_bool
,
967 nir_type_bool16
= 16 | nir_type_bool
,
968 nir_type_bool32
= 32 | nir_type_bool
,
969 nir_type_int1
= 1 | nir_type_int
,
970 nir_type_int8
= 8 | nir_type_int
,
971 nir_type_int16
= 16 | nir_type_int
,
972 nir_type_int32
= 32 | nir_type_int
,
973 nir_type_int64
= 64 | nir_type_int
,
974 nir_type_uint1
= 1 | nir_type_uint
,
975 nir_type_uint8
= 8 | nir_type_uint
,
976 nir_type_uint16
= 16 | nir_type_uint
,
977 nir_type_uint32
= 32 | nir_type_uint
,
978 nir_type_uint64
= 64 | nir_type_uint
,
979 nir_type_float16
= 16 | nir_type_float
,
980 nir_type_float32
= 32 | nir_type_float
,
981 nir_type_float64
= 64 | nir_type_float
,
984 #define NIR_ALU_TYPE_SIZE_MASK 0x79
985 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
987 static inline unsigned
988 nir_alu_type_get_type_size(nir_alu_type type
)
990 return type
& NIR_ALU_TYPE_SIZE_MASK
;
993 static inline unsigned
994 nir_alu_type_get_base_type(nir_alu_type type
)
996 return type
& NIR_ALU_TYPE_BASE_TYPE_MASK
;
999 static inline nir_alu_type
1000 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type
)
1002 switch (base_type
) {
1003 case GLSL_TYPE_BOOL
:
1004 return nir_type_bool1
;
1006 case GLSL_TYPE_UINT
:
1007 return nir_type_uint32
;
1010 return nir_type_int32
;
1012 case GLSL_TYPE_UINT16
:
1013 return nir_type_uint16
;
1015 case GLSL_TYPE_INT16
:
1016 return nir_type_int16
;
1018 case GLSL_TYPE_UINT8
:
1019 return nir_type_uint8
;
1020 case GLSL_TYPE_INT8
:
1021 return nir_type_int8
;
1022 case GLSL_TYPE_UINT64
:
1023 return nir_type_uint64
;
1025 case GLSL_TYPE_INT64
:
1026 return nir_type_int64
;
1028 case GLSL_TYPE_FLOAT
:
1029 return nir_type_float32
;
1031 case GLSL_TYPE_FLOAT16
:
1032 return nir_type_float16
;
1034 case GLSL_TYPE_DOUBLE
:
1035 return nir_type_float64
;
1038 case GLSL_TYPE_SAMPLER
:
1039 case GLSL_TYPE_IMAGE
:
1040 case GLSL_TYPE_ATOMIC_UINT
:
1041 case GLSL_TYPE_STRUCT
:
1042 case GLSL_TYPE_INTERFACE
:
1043 case GLSL_TYPE_ARRAY
:
1044 case GLSL_TYPE_VOID
:
1045 case GLSL_TYPE_SUBROUTINE
:
1046 case GLSL_TYPE_FUNCTION
:
1047 case GLSL_TYPE_ERROR
:
1048 return nir_type_invalid
;
1051 unreachable("unknown type");
1054 static inline nir_alu_type
1055 nir_get_nir_type_for_glsl_type(const struct glsl_type
*type
)
1057 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type
));
1060 nir_op
nir_type_conversion_op(nir_alu_type src
, nir_alu_type dst
,
1061 nir_rounding_mode rnd
);
1063 static inline nir_op
1064 nir_op_vec(unsigned components
)
1066 switch (components
) {
1067 case 1: return nir_op_mov
;
1068 case 2: return nir_op_vec2
;
1069 case 3: return nir_op_vec3
;
1070 case 4: return nir_op_vec4
;
1071 case 8: return nir_op_vec8
;
1072 case 16: return nir_op_vec16
;
1073 default: unreachable("bad component count");
1078 nir_op_is_vec(nir_op op
)
1094 nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode
, unsigned bit_size
)
1096 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16
) ||
1097 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32
) ||
1098 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64
);
1102 nir_is_denorm_flush_to_zero(unsigned execution_mode
, unsigned bit_size
)
1104 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16
) ||
1105 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32
) ||
1106 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64
);
1110 nir_is_denorm_preserve(unsigned execution_mode
, unsigned bit_size
)
1112 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP16
) ||
1113 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP32
) ||
1114 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_DENORM_PRESERVE_FP64
);
1118 nir_is_rounding_mode_rtne(unsigned execution_mode
, unsigned bit_size
)
1120 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1121 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1122 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1126 nir_is_rounding_mode_rtz(unsigned execution_mode
, unsigned bit_size
)
1128 return (16 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1129 (32 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1130 (64 == bit_size
&& execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1134 nir_has_any_rounding_mode_rtz(unsigned execution_mode
)
1136 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16
) ||
1137 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32
) ||
1138 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64
);
1142 nir_has_any_rounding_mode_rtne(unsigned execution_mode
)
1144 return (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16
) ||
1145 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32
) ||
1146 (execution_mode
& FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64
);
1149 static inline nir_rounding_mode
1150 nir_get_rounding_mode_from_float_controls(unsigned execution_mode
,
1153 if (nir_alu_type_get_base_type(type
) != nir_type_float
)
1154 return nir_rounding_mode_undef
;
1156 unsigned bit_size
= nir_alu_type_get_type_size(type
);
1158 if (nir_is_rounding_mode_rtz(execution_mode
, bit_size
))
1159 return nir_rounding_mode_rtz
;
1160 if (nir_is_rounding_mode_rtne(execution_mode
, bit_size
))
1161 return nir_rounding_mode_rtne
;
1162 return nir_rounding_mode_undef
;
1166 nir_has_any_rounding_mode_enabled(unsigned execution_mode
)
1169 nir_has_any_rounding_mode_rtne(execution_mode
) ||
1170 nir_has_any_rounding_mode_rtz(execution_mode
);
1176 * Operation where the first two sources are commutative.
1178 * For 2-source operations, this just mathematical commutativity. Some
1179 * 3-source operations, like ffma, are only commutative in the first two
1182 NIR_OP_IS_2SRC_COMMUTATIVE
= (1 << 0),
1183 NIR_OP_IS_ASSOCIATIVE
= (1 << 1),
1184 } nir_op_algebraic_property
;
1189 unsigned num_inputs
;
1192 * The number of components in the output
1194 * If non-zero, this is the size of the output and input sizes are
1195 * explicitly given; swizzle and writemask are still in effect, but if
1196 * the output component is masked out, then the input component may
1199 * If zero, the opcode acts in the standard, per-component manner; the
1200 * operation is performed on each component (except the ones that are
1201 * masked out) with the input being taken from the input swizzle for
1204 * The size of some of the inputs may be given (i.e. non-zero) even
1205 * though output_size is zero; in that case, the inputs with a zero
1206 * size act per-component, while the inputs with non-zero size don't.
1208 unsigned output_size
;
1211 * The type of vector that the instruction outputs. Note that the
1212 * staurate modifier is only allowed on outputs with the float type.
1215 nir_alu_type output_type
;
1218 * The number of components in each input
1220 unsigned input_sizes
[NIR_MAX_VEC_COMPONENTS
];
1223 * The type of vector that each input takes. Note that negate and
1224 * absolute value are only allowed on inputs with int or float type and
1225 * behave differently on the two.
1227 nir_alu_type input_types
[NIR_MAX_VEC_COMPONENTS
];
1229 nir_op_algebraic_property algebraic_properties
;
1231 /* Whether this represents a numeric conversion opcode */
1235 extern const nir_op_info nir_op_infos
[nir_num_opcodes
];
1237 typedef struct nir_alu_instr
{
1241 /** Indicates that this ALU instruction generates an exact value
1243 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1244 * really equivalent to either. This indicates that the value generated by
1245 * this operation is high-precision and any code transformations that touch
1246 * it must ensure that the resulting value is bit-for-bit identical to the
1252 * Indicates that this instruction do not cause wrapping to occur, in the
1253 * form of overflow or underflow.
1255 bool no_signed_wrap
:1;
1256 bool no_unsigned_wrap
:1;
1262 void nir_alu_src_copy(nir_alu_src
*dest
, const nir_alu_src
*src
,
1263 nir_alu_instr
*instr
);
1264 void nir_alu_dest_copy(nir_alu_dest
*dest
, const nir_alu_dest
*src
,
1265 nir_alu_instr
*instr
);
1267 /* is this source channel used? */
1269 nir_alu_instr_channel_used(const nir_alu_instr
*instr
, unsigned src
,
1272 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1273 return channel
< nir_op_infos
[instr
->op
].input_sizes
[src
];
1275 return (instr
->dest
.write_mask
>> channel
) & 1;
1278 static inline nir_component_mask_t
1279 nir_alu_instr_src_read_mask(const nir_alu_instr
*instr
, unsigned src
)
1281 nir_component_mask_t read_mask
= 0;
1282 for (unsigned c
= 0; c
< NIR_MAX_VEC_COMPONENTS
; c
++) {
1283 if (!nir_alu_instr_channel_used(instr
, src
, c
))
1286 read_mask
|= (1 << instr
->src
[src
].swizzle
[c
]);
1292 * Get the number of channels used for a source
1294 static inline unsigned
1295 nir_ssa_alu_instr_src_components(const nir_alu_instr
*instr
, unsigned src
)
1297 if (nir_op_infos
[instr
->op
].input_sizes
[src
] > 0)
1298 return nir_op_infos
[instr
->op
].input_sizes
[src
];
1300 return nir_dest_num_components(instr
->dest
.dest
);
1304 nir_alu_instr_is_comparison(const nir_alu_instr
*instr
)
1306 switch (instr
->op
) {
1326 bool nir_const_value_negative_equal(nir_const_value c1
, nir_const_value c2
,
1327 nir_alu_type full_type
);
1329 bool nir_alu_srcs_equal(const nir_alu_instr
*alu1
, const nir_alu_instr
*alu2
,
1330 unsigned src1
, unsigned src2
);
1332 bool nir_alu_srcs_negative_equal(const nir_alu_instr
*alu1
,
1333 const nir_alu_instr
*alu2
,
1334 unsigned src1
, unsigned src2
);
1338 nir_deref_type_array
,
1339 nir_deref_type_array_wildcard
,
1340 nir_deref_type_ptr_as_array
,
1341 nir_deref_type_struct
,
1342 nir_deref_type_cast
,
1348 /** The type of this deref instruction */
1349 nir_deref_type deref_type
;
1351 /** The mode of the underlying variable */
1352 nir_variable_mode mode
;
1354 /** The dereferenced type of the resulting pointer value */
1355 const struct glsl_type
*type
;
1358 /** Variable being dereferenced if deref_type is a deref_var */
1361 /** Parent deref if deref_type is not deref_var */
1365 /** Additional deref parameters */
1376 unsigned ptr_stride
;
1380 /** Destination to store the resulting "pointer" */
1384 static inline nir_deref_instr
*nir_src_as_deref(nir_src src
);
1386 static inline nir_deref_instr
*
1387 nir_deref_instr_parent(const nir_deref_instr
*instr
)
1389 if (instr
->deref_type
== nir_deref_type_var
)
1392 return nir_src_as_deref(instr
->parent
);
1395 static inline nir_variable
*
1396 nir_deref_instr_get_variable(const nir_deref_instr
*instr
)
1398 while (instr
->deref_type
!= nir_deref_type_var
) {
1399 if (instr
->deref_type
== nir_deref_type_cast
)
1402 instr
= nir_deref_instr_parent(instr
);
1408 bool nir_deref_instr_has_indirect(nir_deref_instr
*instr
);
1409 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr
*instr
);
1410 bool nir_deref_instr_has_complex_use(nir_deref_instr
*instr
);
1412 bool nir_deref_instr_remove_if_unused(nir_deref_instr
*instr
);
1414 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr
*instr
);
1419 struct nir_function
*callee
;
1421 unsigned num_params
;
1425 #include "nir_intrinsics.h"
1427 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1429 /** Represents an intrinsic
1431 * An intrinsic is an instruction type for handling things that are
1432 * more-or-less regular operations but don't just consume and produce SSA
1433 * values like ALU operations do. Intrinsics are not for things that have
1434 * special semantic meaning such as phi nodes and parallel copies.
1435 * Examples of intrinsics include variable load/store operations, system
1436 * value loads, and the like. Even though texturing more-or-less falls
1437 * under this category, texturing is its own instruction type because
1438 * trying to represent texturing with intrinsics would lead to a
1439 * combinatorial explosion of intrinsic opcodes.
1441 * By having a single instruction type for handling a lot of different
1442 * cases, optimization passes can look for intrinsics and, for the most
1443 * part, completely ignore them. Each intrinsic type also has a few
1444 * possible flags that govern whether or not they can be reordered or
1445 * eliminated. That way passes like dead code elimination can still work
1446 * on intrisics without understanding the meaning of each.
1448 * Each intrinsic has some number of constant indices, some number of
1449 * variables, and some number of sources. What these sources, variables,
1450 * and indices mean depends on the intrinsic and is documented with the
1451 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1452 * instructions are the only types of instruction that can operate on
1458 nir_intrinsic_op intrinsic
;
1462 /** number of components if this is a vectorized intrinsic
1464 * Similarly to ALU operations, some intrinsics are vectorized.
1465 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1466 * For vectorized intrinsics, the num_components field specifies the
1467 * number of destination components and the number of source components
1468 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1470 uint8_t num_components
;
1472 int const_index
[NIR_INTRINSIC_MAX_CONST_INDEX
];
1475 } nir_intrinsic_instr
;
1477 static inline nir_variable
*
1478 nir_intrinsic_get_var(nir_intrinsic_instr
*intrin
, unsigned i
)
1480 return nir_deref_instr_get_variable(nir_src_as_deref(intrin
->src
[i
]));
1484 /* Memory ordering. */
1485 NIR_MEMORY_ACQUIRE
= 1 << 0,
1486 NIR_MEMORY_RELEASE
= 1 << 1,
1487 NIR_MEMORY_ACQ_REL
= NIR_MEMORY_ACQUIRE
| NIR_MEMORY_RELEASE
,
1489 /* Memory visibility operations. */
1490 NIR_MEMORY_MAKE_AVAILABLE
= 1 << 2,
1491 NIR_MEMORY_MAKE_VISIBLE
= 1 << 3,
1492 } nir_memory_semantics
;
1495 NIR_SCOPE_INVOCATION
,
1497 NIR_SCOPE_WORKGROUP
,
1498 NIR_SCOPE_QUEUE_FAMILY
,
1503 * \name NIR intrinsics semantic flags
1505 * information about what the compiler can do with the intrinsics.
1507 * \sa nir_intrinsic_info::flags
1511 * whether the intrinsic can be safely eliminated if none of its output
1512 * value is not being used.
1514 NIR_INTRINSIC_CAN_ELIMINATE
= (1 << 0),
1517 * Whether the intrinsic can be reordered with respect to any other
1518 * intrinsic, i.e. whether the only reordering dependencies of the
1519 * intrinsic are due to the register reads/writes.
1521 NIR_INTRINSIC_CAN_REORDER
= (1 << 1),
1522 } nir_intrinsic_semantic_flag
;
1525 * \name NIR intrinsics const-index flag
1527 * Indicates the usage of a const_index slot.
1529 * \sa nir_intrinsic_info::index_map
1533 * Generally instructions that take a offset src argument, can encode
1534 * a constant 'base' value which is added to the offset.
1536 NIR_INTRINSIC_BASE
= 1,
1539 * For store instructions, a writemask for the store.
1541 NIR_INTRINSIC_WRMASK
,
1544 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1546 NIR_INTRINSIC_STREAM_ID
,
1549 * The clip-plane id for load_user_clip_plane intrinsic.
1551 NIR_INTRINSIC_UCP_ID
,
1554 * The amount of data, starting from BASE, that this instruction may
1555 * access. This is used to provide bounds if the offset is not constant.
1557 NIR_INTRINSIC_RANGE
,
1560 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1562 NIR_INTRINSIC_DESC_SET
,
1565 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1567 NIR_INTRINSIC_BINDING
,
1572 NIR_INTRINSIC_COMPONENT
,
1575 * Interpolation mode (only meaningful for FS inputs).
1577 NIR_INTRINSIC_INTERP_MODE
,
1580 * A binary nir_op to use when performing a reduction or scan operation
1582 NIR_INTRINSIC_REDUCTION_OP
,
1585 * Cluster size for reduction operations
1587 NIR_INTRINSIC_CLUSTER_SIZE
,
1590 * Parameter index for a load_param intrinsic
1592 NIR_INTRINSIC_PARAM_IDX
,
1595 * Image dimensionality for image intrinsics
1597 * One of GLSL_SAMPLER_DIM_*
1599 NIR_INTRINSIC_IMAGE_DIM
,
1602 * Non-zero if we are accessing an array image
1604 NIR_INTRINSIC_IMAGE_ARRAY
,
1607 * Image format for image intrinsics
1609 NIR_INTRINSIC_FORMAT
,
1612 * Access qualifiers for image and memory access intrinsics
1614 NIR_INTRINSIC_ACCESS
,
1617 * Alignment for offsets and addresses
1619 * These two parameters, specify an alignment in terms of a multiplier and
1620 * an offset. The offset or address parameter X of the intrinsic is
1621 * guaranteed to satisfy the following:
1623 * (X - align_offset) % align_mul == 0
1625 NIR_INTRINSIC_ALIGN_MUL
,
1626 NIR_INTRINSIC_ALIGN_OFFSET
,
1629 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1631 NIR_INTRINSIC_DESC_TYPE
,
1634 * The nir_alu_type of a uniform/input/output
1639 * The swizzle mask for the instructions
1640 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1642 NIR_INTRINSIC_SWIZZLE_MASK
,
1644 /* Separate source/dest access flags for copies */
1645 NIR_INTRINSIC_SRC_ACCESS
,
1646 NIR_INTRINSIC_DST_ACCESS
,
1648 /* Driver location for nir_load_patch_location_ir3 */
1649 NIR_INTRINSIC_DRIVER_LOCATION
,
1652 * Mask of nir_memory_semantics, includes ordering and visibility.
1654 NIR_INTRINSIC_MEMORY_SEMANTICS
,
1657 * Mask of nir_variable_modes affected by the memory operation.
1659 NIR_INTRINSIC_MEMORY_MODES
,
1662 * Value of nir_scope.
1664 NIR_INTRINSIC_MEMORY_SCOPE
,
1666 NIR_INTRINSIC_NUM_INDEX_FLAGS
,
1668 } nir_intrinsic_index_flag
;
1670 #define NIR_INTRINSIC_MAX_INPUTS 5
1675 unsigned num_srcs
; /** < number of register/SSA inputs */
1677 /** number of components of each input register
1679 * If this value is 0, the number of components is given by the
1680 * num_components field of nir_intrinsic_instr. If this value is -1, the
1681 * intrinsic consumes however many components are provided and it is not
1684 int src_components
[NIR_INTRINSIC_MAX_INPUTS
];
1688 /** number of components of the output register
1690 * If this value is 0, the number of components is given by the
1691 * num_components field of nir_intrinsic_instr.
1693 unsigned dest_components
;
1695 /** bitfield of legal bit sizes */
1696 unsigned dest_bit_sizes
;
1698 /** the number of constant indices used by the intrinsic */
1699 unsigned num_indices
;
1701 /** indicates the usage of intr->const_index[n] */
1702 unsigned index_map
[NIR_INTRINSIC_NUM_INDEX_FLAGS
];
1704 /** semantic flags for calls to this intrinsic */
1705 nir_intrinsic_semantic_flag flags
;
1706 } nir_intrinsic_info
;
1708 extern const nir_intrinsic_info nir_intrinsic_infos
[nir_num_intrinsics
];
1710 static inline unsigned
1711 nir_intrinsic_src_components(nir_intrinsic_instr
*intr
, unsigned srcn
)
1713 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1714 assert(srcn
< info
->num_srcs
);
1715 if (info
->src_components
[srcn
] > 0)
1716 return info
->src_components
[srcn
];
1717 else if (info
->src_components
[srcn
] == 0)
1718 return intr
->num_components
;
1720 return nir_src_num_components(intr
->src
[srcn
]);
1723 static inline unsigned
1724 nir_intrinsic_dest_components(nir_intrinsic_instr
*intr
)
1726 const nir_intrinsic_info
*info
= &nir_intrinsic_infos
[intr
->intrinsic
];
1727 if (!info
->has_dest
)
1729 else if (info
->dest_components
)
1730 return info
->dest_components
;
1732 return intr
->num_components
;
1735 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1736 static inline type \
1737 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1739 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1740 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1741 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1743 static inline void \
1744 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1746 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1747 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1748 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1751 INTRINSIC_IDX_ACCESSORS(write_mask
, WRMASK
, unsigned)
1752 INTRINSIC_IDX_ACCESSORS(base
, BASE
, int)
1753 INTRINSIC_IDX_ACCESSORS(stream_id
, STREAM_ID
, unsigned)
1754 INTRINSIC_IDX_ACCESSORS(ucp_id
, UCP_ID
, unsigned)
1755 INTRINSIC_IDX_ACCESSORS(range
, RANGE
, unsigned)
1756 INTRINSIC_IDX_ACCESSORS(desc_set
, DESC_SET
, unsigned)
1757 INTRINSIC_IDX_ACCESSORS(binding
, BINDING
, unsigned)
1758 INTRINSIC_IDX_ACCESSORS(component
, COMPONENT
, unsigned)
1759 INTRINSIC_IDX_ACCESSORS(interp_mode
, INTERP_MODE
, unsigned)
1760 INTRINSIC_IDX_ACCESSORS(reduction_op
, REDUCTION_OP
, unsigned)
1761 INTRINSIC_IDX_ACCESSORS(cluster_size
, CLUSTER_SIZE
, unsigned)
1762 INTRINSIC_IDX_ACCESSORS(param_idx
, PARAM_IDX
, unsigned)
1763 INTRINSIC_IDX_ACCESSORS(image_dim
, IMAGE_DIM
, enum glsl_sampler_dim
)
1764 INTRINSIC_IDX_ACCESSORS(image_array
, IMAGE_ARRAY
, bool)
1765 INTRINSIC_IDX_ACCESSORS(access
, ACCESS
, enum gl_access_qualifier
)
1766 INTRINSIC_IDX_ACCESSORS(src_access
, SRC_ACCESS
, enum gl_access_qualifier
)
1767 INTRINSIC_IDX_ACCESSORS(dst_access
, DST_ACCESS
, enum gl_access_qualifier
)
1768 INTRINSIC_IDX_ACCESSORS(format
, FORMAT
, enum pipe_format
)
1769 INTRINSIC_IDX_ACCESSORS(align_mul
, ALIGN_MUL
, unsigned)
1770 INTRINSIC_IDX_ACCESSORS(align_offset
, ALIGN_OFFSET
, unsigned)
1771 INTRINSIC_IDX_ACCESSORS(desc_type
, DESC_TYPE
, unsigned)
1772 INTRINSIC_IDX_ACCESSORS(type
, TYPE
, nir_alu_type
)
1773 INTRINSIC_IDX_ACCESSORS(swizzle_mask
, SWIZZLE_MASK
, unsigned)
1774 INTRINSIC_IDX_ACCESSORS(driver_location
, DRIVER_LOCATION
, unsigned)
1775 INTRINSIC_IDX_ACCESSORS(memory_semantics
, MEMORY_SEMANTICS
, nir_memory_semantics
)
1776 INTRINSIC_IDX_ACCESSORS(memory_modes
, MEMORY_MODES
, nir_variable_mode
)
1777 INTRINSIC_IDX_ACCESSORS(memory_scope
, MEMORY_SCOPE
, nir_scope
)
1780 nir_intrinsic_set_align(nir_intrinsic_instr
*intrin
,
1781 unsigned align_mul
, unsigned align_offset
)
1783 assert(util_is_power_of_two_nonzero(align_mul
));
1784 assert(align_offset
< align_mul
);
1785 nir_intrinsic_set_align_mul(intrin
, align_mul
);
1786 nir_intrinsic_set_align_offset(intrin
, align_offset
);
1789 /** Returns a simple alignment for a load/store intrinsic offset
1791 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1792 * and ALIGN_OFFSET parameters, this helper takes both into account and
1793 * provides a single simple alignment parameter. The offset X is guaranteed
1794 * to satisfy X % align == 0.
1796 static inline unsigned
1797 nir_intrinsic_align(const nir_intrinsic_instr
*intrin
)
1799 const unsigned align_mul
= nir_intrinsic_align_mul(intrin
);
1800 const unsigned align_offset
= nir_intrinsic_align_offset(intrin
);
1801 assert(align_offset
< align_mul
);
1802 return align_offset
? 1 << (ffs(align_offset
) - 1) : align_mul
;
1806 nir_image_intrinsic_coord_components(const nir_intrinsic_instr
*instr
);
1808 /* Converts a image_deref_* intrinsic into a image_* one */
1809 void nir_rewrite_image_intrinsic(nir_intrinsic_instr
*instr
,
1810 nir_ssa_def
*handle
, bool bindless
);
1812 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1814 nir_intrinsic_can_reorder(nir_intrinsic_instr
*instr
)
1816 if (instr
->intrinsic
== nir_intrinsic_load_deref
||
1817 instr
->intrinsic
== nir_intrinsic_load_ssbo
||
1818 instr
->intrinsic
== nir_intrinsic_bindless_image_load
||
1819 instr
->intrinsic
== nir_intrinsic_image_deref_load
||
1820 instr
->intrinsic
== nir_intrinsic_image_load
) {
1821 return nir_intrinsic_access(instr
) & ACCESS_CAN_REORDER
;
1823 const nir_intrinsic_info
*info
=
1824 &nir_intrinsic_infos
[instr
->intrinsic
];
1825 return (info
->flags
& NIR_INTRINSIC_CAN_ELIMINATE
) &&
1826 (info
->flags
& NIR_INTRINSIC_CAN_REORDER
);
1831 * \group texture information
1833 * This gives semantic information about textures which is useful to the
1834 * frontend, the backend, and lowering passes, but not the optimizer.
1839 nir_tex_src_projector
,
1840 nir_tex_src_comparator
, /* shadow comparator */
1844 nir_tex_src_min_lod
,
1845 nir_tex_src_ms_index
, /* MSAA sample index */
1846 nir_tex_src_ms_mcs
, /* MSAA compression value */
1849 nir_tex_src_texture_deref
, /* < deref pointing to the texture */
1850 nir_tex_src_sampler_deref
, /* < deref pointing to the sampler */
1851 nir_tex_src_texture_offset
, /* < dynamically uniform indirect offset */
1852 nir_tex_src_sampler_offset
, /* < dynamically uniform indirect offset */
1853 nir_tex_src_texture_handle
, /* < bindless texture handle */
1854 nir_tex_src_sampler_handle
, /* < bindless sampler handle */
1855 nir_tex_src_plane
, /* < selects plane for planar textures */
1856 nir_num_tex_src_types
1861 nir_tex_src_type src_type
;
1865 nir_texop_tex
, /**< Regular texture look-up */
1866 nir_texop_txb
, /**< Texture look-up with LOD bias */
1867 nir_texop_txl
, /**< Texture look-up with explicit LOD */
1868 nir_texop_txd
, /**< Texture look-up with partial derivatives */
1869 nir_texop_txf
, /**< Texel fetch with explicit LOD */
1870 nir_texop_txf_ms
, /**< Multisample texture fetch */
1871 nir_texop_txf_ms_fb
, /**< Multisample texture fetch from framebuffer */
1872 nir_texop_txf_ms_mcs
, /**< Multisample compression value fetch */
1873 nir_texop_txs
, /**< Texture size */
1874 nir_texop_lod
, /**< Texture lod query */
1875 nir_texop_tg4
, /**< Texture gather */
1876 nir_texop_query_levels
, /**< Texture levels query */
1877 nir_texop_texture_samples
, /**< Texture samples query */
1878 nir_texop_samples_identical
, /**< Query whether all samples are definitely
1881 nir_texop_tex_prefetch
, /**< Regular texture look-up, eligible for pre-dispatch */
1882 nir_texop_fragment_fetch
, /**< Multisample fragment color texture fetch */
1883 nir_texop_fragment_mask_fetch
,/**< Multisample fragment mask texture fetch */
1889 enum glsl_sampler_dim sampler_dim
;
1890 nir_alu_type dest_type
;
1895 unsigned num_srcs
, coord_components
;
1896 bool is_array
, is_shadow
;
1899 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1900 * components or the new-style shadow that outputs 1 component.
1902 bool is_new_style_shadow
;
1904 /* gather component selector */
1905 unsigned component
: 2;
1907 /* gather offsets */
1908 int8_t tg4_offsets
[4][2];
1910 /* True if the texture index or handle is not dynamically uniform */
1911 bool texture_non_uniform
;
1913 /* True if the sampler index or handle is not dynamically uniform */
1914 bool sampler_non_uniform
;
1916 /** The texture index
1918 * If this texture instruction has a nir_tex_src_texture_offset source,
1919 * then the texture index is given by texture_index + texture_offset.
1921 unsigned texture_index
;
1923 /** The sampler index
1925 * The following operations do not require a sampler and, as such, this
1926 * field should be ignored:
1928 * - nir_texop_txf_ms
1931 * - nir_texop_query_levels
1932 * - nir_texop_texture_samples
1933 * - nir_texop_samples_identical
1935 * If this texture instruction has a nir_tex_src_sampler_offset source,
1936 * then the sampler index is given by sampler_index + sampler_offset.
1938 unsigned sampler_index
;
1941 static inline unsigned
1942 nir_tex_instr_dest_size(const nir_tex_instr
*instr
)
1944 switch (instr
->op
) {
1945 case nir_texop_txs
: {
1947 switch (instr
->sampler_dim
) {
1948 case GLSL_SAMPLER_DIM_1D
:
1949 case GLSL_SAMPLER_DIM_BUF
:
1952 case GLSL_SAMPLER_DIM_2D
:
1953 case GLSL_SAMPLER_DIM_CUBE
:
1954 case GLSL_SAMPLER_DIM_MS
:
1955 case GLSL_SAMPLER_DIM_RECT
:
1956 case GLSL_SAMPLER_DIM_EXTERNAL
:
1957 case GLSL_SAMPLER_DIM_SUBPASS
:
1960 case GLSL_SAMPLER_DIM_3D
:
1964 unreachable("not reached");
1966 if (instr
->is_array
)
1974 case nir_texop_texture_samples
:
1975 case nir_texop_query_levels
:
1976 case nir_texop_samples_identical
:
1977 case nir_texop_fragment_mask_fetch
:
1981 if (instr
->is_shadow
&& instr
->is_new_style_shadow
)
1988 /* Returns true if this texture operation queries something about the texture
1989 * rather than actually sampling it.
1992 nir_tex_instr_is_query(const nir_tex_instr
*instr
)
1994 switch (instr
->op
) {
1997 case nir_texop_texture_samples
:
1998 case nir_texop_query_levels
:
1999 case nir_texop_txf_ms_mcs
:
2006 case nir_texop_txf_ms
:
2007 case nir_texop_txf_ms_fb
:
2011 unreachable("Invalid texture opcode");
2016 nir_tex_instr_has_implicit_derivative(const nir_tex_instr
*instr
)
2018 switch (instr
->op
) {
2028 static inline nir_alu_type
2029 nir_tex_instr_src_type(const nir_tex_instr
*instr
, unsigned src
)
2031 switch (instr
->src
[src
].src_type
) {
2032 case nir_tex_src_coord
:
2033 switch (instr
->op
) {
2035 case nir_texop_txf_ms
:
2036 case nir_texop_txf_ms_fb
:
2037 case nir_texop_txf_ms_mcs
:
2038 case nir_texop_samples_identical
:
2039 return nir_type_int
;
2042 return nir_type_float
;
2045 case nir_tex_src_lod
:
2046 switch (instr
->op
) {
2049 return nir_type_int
;
2052 return nir_type_float
;
2055 case nir_tex_src_projector
:
2056 case nir_tex_src_comparator
:
2057 case nir_tex_src_bias
:
2058 case nir_tex_src_min_lod
:
2059 case nir_tex_src_ddx
:
2060 case nir_tex_src_ddy
:
2061 return nir_type_float
;
2063 case nir_tex_src_offset
:
2064 case nir_tex_src_ms_index
:
2065 case nir_tex_src_plane
:
2066 return nir_type_int
;
2068 case nir_tex_src_ms_mcs
:
2069 case nir_tex_src_texture_deref
:
2070 case nir_tex_src_sampler_deref
:
2071 case nir_tex_src_texture_offset
:
2072 case nir_tex_src_sampler_offset
:
2073 case nir_tex_src_texture_handle
:
2074 case nir_tex_src_sampler_handle
:
2075 return nir_type_uint
;
2077 case nir_num_tex_src_types
:
2078 unreachable("nir_num_tex_src_types is not a valid source type");
2081 unreachable("Invalid texture source type");
2084 static inline unsigned
2085 nir_tex_instr_src_size(const nir_tex_instr
*instr
, unsigned src
)
2087 if (instr
->src
[src
].src_type
== nir_tex_src_coord
)
2088 return instr
->coord_components
;
2090 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
2091 if (instr
->src
[src
].src_type
== nir_tex_src_ms_mcs
)
2094 if (instr
->src
[src
].src_type
== nir_tex_src_ddx
||
2095 instr
->src
[src
].src_type
== nir_tex_src_ddy
) {
2096 if (instr
->is_array
)
2097 return instr
->coord_components
- 1;
2099 return instr
->coord_components
;
2102 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
2103 * the offset, since a cube maps to a single face.
2105 if (instr
->src
[src
].src_type
== nir_tex_src_offset
) {
2106 if (instr
->sampler_dim
== GLSL_SAMPLER_DIM_CUBE
)
2108 else if (instr
->is_array
)
2109 return instr
->coord_components
- 1;
2111 return instr
->coord_components
;
2118 nir_tex_instr_src_index(const nir_tex_instr
*instr
, nir_tex_src_type type
)
2120 for (unsigned i
= 0; i
< instr
->num_srcs
; i
++)
2121 if (instr
->src
[i
].src_type
== type
)
2127 void nir_tex_instr_add_src(nir_tex_instr
*tex
,
2128 nir_tex_src_type src_type
,
2131 void nir_tex_instr_remove_src(nir_tex_instr
*tex
, unsigned src_idx
);
2133 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr
*tex
);
2140 nir_const_value value
[];
2141 } nir_load_const_instr
;
2154 /* creates a new SSA variable in an undefined state */
2159 } nir_ssa_undef_instr
;
2162 struct exec_node node
;
2164 /* The predecessor block corresponding to this source */
2165 struct nir_block
*pred
;
2170 #define nir_foreach_phi_src(phi_src, phi) \
2171 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
2172 #define nir_foreach_phi_src_safe(phi_src, phi) \
2173 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
2178 struct exec_list srcs
; /** < list of nir_phi_src */
2184 struct exec_node node
;
2187 } nir_parallel_copy_entry
;
2189 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
2190 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
2195 /* A list of nir_parallel_copy_entrys. The sources of all of the
2196 * entries are copied to the corresponding destinations "in parallel".
2197 * In other words, if we have two entries: a -> b and b -> a, the values
2200 struct exec_list entries
;
2201 } nir_parallel_copy_instr
;
2203 NIR_DEFINE_CAST(nir_instr_as_alu
, nir_instr
, nir_alu_instr
, instr
,
2204 type
, nir_instr_type_alu
)
2205 NIR_DEFINE_CAST(nir_instr_as_deref
, nir_instr
, nir_deref_instr
, instr
,
2206 type
, nir_instr_type_deref
)
2207 NIR_DEFINE_CAST(nir_instr_as_call
, nir_instr
, nir_call_instr
, instr
,
2208 type
, nir_instr_type_call
)
2209 NIR_DEFINE_CAST(nir_instr_as_jump
, nir_instr
, nir_jump_instr
, instr
,
2210 type
, nir_instr_type_jump
)
2211 NIR_DEFINE_CAST(nir_instr_as_tex
, nir_instr
, nir_tex_instr
, instr
,
2212 type
, nir_instr_type_tex
)
2213 NIR_DEFINE_CAST(nir_instr_as_intrinsic
, nir_instr
, nir_intrinsic_instr
, instr
,
2214 type
, nir_instr_type_intrinsic
)
2215 NIR_DEFINE_CAST(nir_instr_as_load_const
, nir_instr
, nir_load_const_instr
, instr
,
2216 type
, nir_instr_type_load_const
)
2217 NIR_DEFINE_CAST(nir_instr_as_ssa_undef
, nir_instr
, nir_ssa_undef_instr
, instr
,
2218 type
, nir_instr_type_ssa_undef
)
2219 NIR_DEFINE_CAST(nir_instr_as_phi
, nir_instr
, nir_phi_instr
, instr
,
2220 type
, nir_instr_type_phi
)
2221 NIR_DEFINE_CAST(nir_instr_as_parallel_copy
, nir_instr
,
2222 nir_parallel_copy_instr
, instr
,
2223 type
, nir_instr_type_parallel_copy
)
2226 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
2227 static inline type \
2228 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2230 assert(nir_src_is_const(src)); \
2231 nir_load_const_instr *load = \
2232 nir_instr_as_load_const(src.ssa->parent_instr); \
2233 assert(comp < load->def.num_components); \
2234 return nir_const_value_as_##suffix(load->value[comp], \
2235 load->def.bit_size); \
2238 static inline type \
2239 nir_src_as_##suffix(nir_src src) \
2241 assert(nir_src_num_components(src) == 1); \
2242 return nir_src_comp_as_##suffix(src, 0); \
2245 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2246 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint
)
2247 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2248 NIR_DEFINE_SRC_AS_CONST(double, float)
2250 #undef NIR_DEFINE_SRC_AS_CONST
2259 nir_ssa_scalar_is_const(nir_ssa_scalar s
)
2261 return s
.def
->parent_instr
->type
== nir_instr_type_load_const
;
2264 static inline nir_const_value
2265 nir_ssa_scalar_as_const_value(nir_ssa_scalar s
)
2267 assert(s
.comp
< s
.def
->num_components
);
2268 nir_load_const_instr
*load
= nir_instr_as_load_const(s
.def
->parent_instr
);
2269 return load
->value
[s
.comp
];
2272 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2273 static inline type \
2274 nir_ssa_scalar_as_##suffix(nir_ssa_scalar s) \
2276 return nir_const_value_as_##suffix( \
2277 nir_ssa_scalar_as_const_value(s), s.def->bit_size); \
2280 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2281 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint
)
2282 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2283 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2285 #undef NIR_DEFINE_SCALAR_AS_CONST
2288 nir_ssa_scalar_is_alu(nir_ssa_scalar s
)
2290 return s
.def
->parent_instr
->type
== nir_instr_type_alu
;
2293 static inline nir_op
2294 nir_ssa_scalar_alu_op(nir_ssa_scalar s
)
2296 return nir_instr_as_alu(s
.def
->parent_instr
)->op
;
2299 static inline nir_ssa_scalar
2300 nir_ssa_scalar_chase_alu_src(nir_ssa_scalar s
, unsigned alu_src_idx
)
2302 nir_ssa_scalar out
= { NULL
, 0 };
2304 nir_alu_instr
*alu
= nir_instr_as_alu(s
.def
->parent_instr
);
2305 assert(alu_src_idx
< nir_op_infos
[alu
->op
].num_inputs
);
2307 /* Our component must be written */
2308 assert(s
.comp
< s
.def
->num_components
);
2309 assert(alu
->dest
.write_mask
& (1u << s
.comp
));
2311 assert(alu
->src
[alu_src_idx
].src
.is_ssa
);
2312 out
.def
= alu
->src
[alu_src_idx
].src
.ssa
;
2314 if (nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 0) {
2315 /* The ALU src is unsized so the source component follows the
2316 * destination component.
2318 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[s
.comp
];
2320 /* This is a sized source so all source components work together to
2321 * produce all the destination components. Since we need to return a
2322 * scalar, this only works if the source is a scalar.
2324 assert(nir_op_infos
[alu
->op
].input_sizes
[alu_src_idx
] == 1);
2325 out
.comp
= alu
->src
[alu_src_idx
].swizzle
[0];
2327 assert(out
.comp
< out
.def
->num_components
);
2336 * Control flow consists of a tree of control flow nodes, which include
2337 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2338 * instructions that always run start-to-finish. Each basic block also keeps
2339 * track of its successors (blocks which may run immediately after the current
2340 * block) and predecessors (blocks which could have run immediately before the
2341 * current block). Each function also has a start block and an end block which
2342 * all return statements point to (which is always empty). Together, all the
2343 * blocks with their predecessors and successors make up the control flow
2344 * graph (CFG) of the function. There are helpers that modify the tree of
2345 * control flow nodes while modifying the CFG appropriately; these should be
2346 * used instead of modifying the tree directly.
2353 nir_cf_node_function
2356 typedef struct nir_cf_node
{
2357 struct exec_node node
;
2358 nir_cf_node_type type
;
2359 struct nir_cf_node
*parent
;
2362 typedef struct nir_block
{
2363 nir_cf_node cf_node
;
2365 struct exec_list instr_list
; /** < list of nir_instr */
2367 /** generic block index; generated by nir_index_blocks */
2371 * Each block can only have up to 2 successors, so we put them in a simple
2372 * array - no need for anything more complicated.
2374 struct nir_block
*successors
[2];
2376 /* Set of nir_block predecessors in the CFG */
2377 struct set
*predecessors
;
2380 * this node's immediate dominator in the dominance tree - set to NULL for
2383 struct nir_block
*imm_dom
;
2385 /* This node's children in the dominance tree */
2386 unsigned num_dom_children
;
2387 struct nir_block
**dom_children
;
2389 /* Set of nir_blocks on the dominance frontier of this block */
2390 struct set
*dom_frontier
;
2393 * These two indices have the property that dom_{pre,post}_index for each
2394 * child of this block in the dominance tree will always be between
2395 * dom_pre_index and dom_post_index for this block, which makes testing if
2396 * a given block is dominated by another block an O(1) operation.
2398 unsigned dom_pre_index
, dom_post_index
;
2400 /* live in and out for this block; used for liveness analysis */
2401 BITSET_WORD
*live_in
;
2402 BITSET_WORD
*live_out
;
2405 static inline nir_instr
*
2406 nir_block_first_instr(nir_block
*block
)
2408 struct exec_node
*head
= exec_list_get_head(&block
->instr_list
);
2409 return exec_node_data(nir_instr
, head
, node
);
2412 static inline nir_instr
*
2413 nir_block_last_instr(nir_block
*block
)
2415 struct exec_node
*tail
= exec_list_get_tail(&block
->instr_list
);
2416 return exec_node_data(nir_instr
, tail
, node
);
2420 nir_block_ends_in_jump(nir_block
*block
)
2422 return !exec_list_is_empty(&block
->instr_list
) &&
2423 nir_block_last_instr(block
)->type
== nir_instr_type_jump
;
2426 #define nir_foreach_instr(instr, block) \
2427 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
2428 #define nir_foreach_instr_reverse(instr, block) \
2429 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
2430 #define nir_foreach_instr_safe(instr, block) \
2431 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
2432 #define nir_foreach_instr_reverse_safe(instr, block) \
2433 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
2436 nir_selection_control_none
= 0x0,
2437 nir_selection_control_flatten
= 0x1,
2438 nir_selection_control_dont_flatten
= 0x2,
2439 } nir_selection_control
;
2441 typedef struct nir_if
{
2442 nir_cf_node cf_node
;
2444 nir_selection_control control
;
2446 struct exec_list then_list
; /** < list of nir_cf_node */
2447 struct exec_list else_list
; /** < list of nir_cf_node */
2453 /** Instruction that generates nif::condition. */
2454 nir_instr
*conditional_instr
;
2456 /** Block within ::nif that has the break instruction. */
2457 nir_block
*break_block
;
2459 /** Last block for the then- or else-path that does not contain the break. */
2460 nir_block
*continue_from_block
;
2462 /** True when ::break_block is in the else-path of ::nif. */
2463 bool continue_from_then
;
2466 /* This is true if the terminators exact trip count is unknown. For
2469 * for (int i = 0; i < imin(x, 4); i++)
2472 * Here loop analysis would have set a max_trip_count of 4 however we dont
2473 * know for sure that this is the exact trip count.
2475 bool exact_trip_count_unknown
;
2477 struct list_head loop_terminator_link
;
2478 } nir_loop_terminator
;
2481 /* Estimated cost (in number of instructions) of the loop */
2482 unsigned instr_cost
;
2484 /* Guessed trip count based on array indexing */
2485 unsigned guessed_trip_count
;
2487 /* Maximum number of times the loop is run (if known) */
2488 unsigned max_trip_count
;
2490 /* Do we know the exact number of times the loop will be run */
2491 bool exact_trip_count_known
;
2493 /* Unroll the loop regardless of its size */
2496 /* Does the loop contain complex loop terminators, continues or other
2497 * complex behaviours? If this is true we can't rely on
2498 * loop_terminator_list to be complete or accurate.
2502 nir_loop_terminator
*limiting_terminator
;
2504 /* A list of loop_terminators terminating this loop. */
2505 struct list_head loop_terminator_list
;
2509 nir_loop_control_none
= 0x0,
2510 nir_loop_control_unroll
= 0x1,
2511 nir_loop_control_dont_unroll
= 0x2,
2515 nir_cf_node cf_node
;
2517 struct exec_list body
; /** < list of nir_cf_node */
2519 nir_loop_info
*info
;
2520 nir_loop_control control
;
2521 bool partially_unrolled
;
2525 * Various bits of metadata that can may be created or required by
2526 * optimization and analysis passes
2529 nir_metadata_none
= 0x0,
2530 nir_metadata_block_index
= 0x1,
2531 nir_metadata_dominance
= 0x2,
2532 nir_metadata_live_ssa_defs
= 0x4,
2533 nir_metadata_not_properly_reset
= 0x8,
2534 nir_metadata_loop_analysis
= 0x10,
2538 nir_cf_node cf_node
;
2540 /** pointer to the function of which this is an implementation */
2541 struct nir_function
*function
;
2543 struct exec_list body
; /** < list of nir_cf_node */
2545 nir_block
*end_block
;
2547 /** list for all local variables in the function */
2548 struct exec_list locals
;
2550 /** list of local registers in the function */
2551 struct exec_list registers
;
2553 /** next available local register index */
2556 /** next available SSA value index */
2559 /* total number of basic blocks, only valid when block_index_dirty = false */
2560 unsigned num_blocks
;
2562 nir_metadata valid_metadata
;
2563 } nir_function_impl
;
2565 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2566 nir_start_block(nir_function_impl
*impl
)
2568 return (nir_block
*) impl
->body
.head_sentinel
.next
;
2571 ATTRIBUTE_RETURNS_NONNULL
static inline nir_block
*
2572 nir_impl_last_block(nir_function_impl
*impl
)
2574 return (nir_block
*) impl
->body
.tail_sentinel
.prev
;
2577 static inline nir_cf_node
*
2578 nir_cf_node_next(nir_cf_node
*node
)
2580 struct exec_node
*next
= exec_node_get_next(&node
->node
);
2581 if (exec_node_is_tail_sentinel(next
))
2584 return exec_node_data(nir_cf_node
, next
, node
);
2587 static inline nir_cf_node
*
2588 nir_cf_node_prev(nir_cf_node
*node
)
2590 struct exec_node
*prev
= exec_node_get_prev(&node
->node
);
2591 if (exec_node_is_head_sentinel(prev
))
2594 return exec_node_data(nir_cf_node
, prev
, node
);
2598 nir_cf_node_is_first(const nir_cf_node
*node
)
2600 return exec_node_is_head_sentinel(node
->node
.prev
);
2604 nir_cf_node_is_last(const nir_cf_node
*node
)
2606 return exec_node_is_tail_sentinel(node
->node
.next
);
2609 NIR_DEFINE_CAST(nir_cf_node_as_block
, nir_cf_node
, nir_block
, cf_node
,
2610 type
, nir_cf_node_block
)
2611 NIR_DEFINE_CAST(nir_cf_node_as_if
, nir_cf_node
, nir_if
, cf_node
,
2612 type
, nir_cf_node_if
)
2613 NIR_DEFINE_CAST(nir_cf_node_as_loop
, nir_cf_node
, nir_loop
, cf_node
,
2614 type
, nir_cf_node_loop
)
2615 NIR_DEFINE_CAST(nir_cf_node_as_function
, nir_cf_node
,
2616 nir_function_impl
, cf_node
, type
, nir_cf_node_function
)
2618 static inline nir_block
*
2619 nir_if_first_then_block(nir_if
*if_stmt
)
2621 struct exec_node
*head
= exec_list_get_head(&if_stmt
->then_list
);
2622 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2625 static inline nir_block
*
2626 nir_if_last_then_block(nir_if
*if_stmt
)
2628 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->then_list
);
2629 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2632 static inline nir_block
*
2633 nir_if_first_else_block(nir_if
*if_stmt
)
2635 struct exec_node
*head
= exec_list_get_head(&if_stmt
->else_list
);
2636 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2639 static inline nir_block
*
2640 nir_if_last_else_block(nir_if
*if_stmt
)
2642 struct exec_node
*tail
= exec_list_get_tail(&if_stmt
->else_list
);
2643 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2646 static inline nir_block
*
2647 nir_loop_first_block(nir_loop
*loop
)
2649 struct exec_node
*head
= exec_list_get_head(&loop
->body
);
2650 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2653 static inline nir_block
*
2654 nir_loop_last_block(nir_loop
*loop
)
2656 struct exec_node
*tail
= exec_list_get_tail(&loop
->body
);
2657 return nir_cf_node_as_block(exec_node_data(nir_cf_node
, tail
, node
));
2661 * Return true if this list of cf_nodes contains a single empty block.
2664 nir_cf_list_is_empty_block(struct exec_list
*cf_list
)
2666 if (exec_list_is_singular(cf_list
)) {
2667 struct exec_node
*head
= exec_list_get_head(cf_list
);
2669 nir_cf_node_as_block(exec_node_data(nir_cf_node
, head
, node
));
2670 return exec_list_is_empty(&block
->instr_list
);
2676 uint8_t num_components
;
2680 typedef struct nir_function
{
2681 struct exec_node node
;
2684 struct nir_shader
*shader
;
2686 unsigned num_params
;
2687 nir_parameter
*params
;
2689 /** The implementation of this function.
2691 * If the function is only declared and not implemented, this is NULL.
2693 nir_function_impl
*impl
;
2699 nir_lower_imul64
= (1 << 0),
2700 nir_lower_isign64
= (1 << 1),
2701 /** Lower all int64 modulus and division opcodes */
2702 nir_lower_divmod64
= (1 << 2),
2703 /** Lower all 64-bit umul_high and imul_high opcodes */
2704 nir_lower_imul_high64
= (1 << 3),
2705 nir_lower_mov64
= (1 << 4),
2706 nir_lower_icmp64
= (1 << 5),
2707 nir_lower_iadd64
= (1 << 6),
2708 nir_lower_iabs64
= (1 << 7),
2709 nir_lower_ineg64
= (1 << 8),
2710 nir_lower_logic64
= (1 << 9),
2711 nir_lower_minmax64
= (1 << 10),
2712 nir_lower_shift64
= (1 << 11),
2713 nir_lower_imul_2x32_64
= (1 << 12),
2714 nir_lower_extract64
= (1 << 13),
2715 nir_lower_ufind_msb64
= (1 << 14),
2716 } nir_lower_int64_options
;
2719 nir_lower_drcp
= (1 << 0),
2720 nir_lower_dsqrt
= (1 << 1),
2721 nir_lower_drsq
= (1 << 2),
2722 nir_lower_dtrunc
= (1 << 3),
2723 nir_lower_dfloor
= (1 << 4),
2724 nir_lower_dceil
= (1 << 5),
2725 nir_lower_dfract
= (1 << 6),
2726 nir_lower_dround_even
= (1 << 7),
2727 nir_lower_dmod
= (1 << 8),
2728 nir_lower_dsub
= (1 << 9),
2729 nir_lower_ddiv
= (1 << 10),
2730 nir_lower_fp64_full_software
= (1 << 11),
2731 } nir_lower_doubles_options
;
2734 nir_divergence_single_prim_per_subgroup
= (1 << 0),
2735 nir_divergence_single_patch_per_tcs_subgroup
= (1 << 1),
2736 nir_divergence_single_patch_per_tes_subgroup
= (1 << 2),
2737 nir_divergence_view_index_uniform
= (1 << 3),
2738 } nir_divergence_options
;
2740 typedef struct nir_shader_compiler_options
{
2746 /** Lowers flrp when it does not support doubles */
2753 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2754 bool lower_bitfield_extract
;
2755 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
2756 bool lower_bitfield_extract_to_shifts
;
2757 /** Lowers bitfield_insert to bfi/bfm */
2758 bool lower_bitfield_insert
;
2759 /** Lowers bitfield_insert to compares, and shifts. */
2760 bool lower_bitfield_insert_to_shifts
;
2761 /** Lowers bitfield_insert to bfm/bitfield_select. */
2762 bool lower_bitfield_insert_to_bitfield_select
;
2763 /** Lowers bitfield_reverse to shifts. */
2764 bool lower_bitfield_reverse
;
2765 /** Lowers bit_count to shifts. */
2766 bool lower_bit_count
;
2767 /** Lowers ifind_msb to compare and ufind_msb */
2768 bool lower_ifind_msb
;
2769 /** Lowers find_lsb to ufind_msb and logic ops */
2770 bool lower_find_lsb
;
2771 bool lower_uadd_carry
;
2772 bool lower_usub_borrow
;
2773 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2774 bool lower_mul_high
;
2775 /** lowers fneg and ineg to fsub and isub. */
2777 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2780 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2783 /* lower fall_equalN/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
2784 bool lower_vector_cmp
;
2786 /** enables rules to lower idiv by power-of-two: */
2789 /** enable rules to avoid bit ops */
2792 /** enables rules to lower isign to imin+imax */
2795 /** enables rules to lower fsign to fsub and flt */
2798 /* lower fdph to fdot4 */
2801 /** lower fdot to fmul and fsum/fadd. */
2804 /* Does the native fdot instruction replicate its result for four
2805 * components? If so, then opt_algebraic_late will turn all fdotN
2806 * instructions into fdot_replicatedN instructions.
2808 bool fdot_replicates
;
2810 /** lowers ffloor to fsub+ffract: */
2813 /** lowers ffract to fsub+ffloor: */
2816 /** lowers fceil to fneg+ffloor+fneg: */
2823 bool lower_pack_half_2x16
;
2824 bool lower_pack_half_2x16_split
;
2825 bool lower_pack_unorm_2x16
;
2826 bool lower_pack_snorm_2x16
;
2827 bool lower_pack_unorm_4x8
;
2828 bool lower_pack_snorm_4x8
;
2829 bool lower_unpack_half_2x16
;
2830 bool lower_unpack_half_2x16_split
;
2831 bool lower_unpack_unorm_2x16
;
2832 bool lower_unpack_snorm_2x16
;
2833 bool lower_unpack_unorm_4x8
;
2834 bool lower_unpack_snorm_4x8
;
2836 bool lower_extract_byte
;
2837 bool lower_extract_word
;
2839 bool lower_all_io_to_temps
;
2840 bool lower_all_io_to_elements
;
2842 /* Indicates that the driver only has zero-based vertex id */
2843 bool vertex_id_zero_based
;
2846 * If enabled, gl_BaseVertex will be lowered as:
2847 * is_indexed_draw (~0/0) & firstvertex
2849 bool lower_base_vertex
;
2852 * If enabled, gl_HelperInvocation will be lowered as:
2854 * !((1 << sample_id) & sample_mask_in))
2856 * This depends on some possibly hw implementation details, which may
2857 * not be true for all hw. In particular that the FS is only executed
2858 * for covered samples or for helper invocations. So, do not blindly
2859 * enable this option.
2861 * Note: See also issue #22 in ARB_shader_image_load_store
2863 bool lower_helper_invocation
;
2866 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
2868 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
2869 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
2871 bool optimize_sample_mask_in
;
2873 bool lower_cs_local_index_from_id
;
2874 bool lower_cs_local_id_from_index
;
2876 bool lower_device_index_to_zero
;
2878 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
2879 bool lower_wpos_pntc
;
2882 * Set if nir_op_[iu]hadd and nir_op_[iu]rhadd instructions should be
2883 * lowered to simple arithmetic.
2885 * If this flag is set, the lowering will be applied to all bit-sizes of
2886 * these instructions.
2888 * \sa ::lower_hadd64
2893 * Set if only 64-bit nir_op_[iu]hadd and nir_op_[iu]rhadd instructions
2894 * should be lowered to simple arithmetic.
2896 * If this flag is set, the lowering will be applied to only 64-bit
2897 * versions of these instructions.
2904 * Set if nir_op_add_sat and nir_op_usub_sat should be lowered to simple
2907 * If this flag is set, the lowering will be applied to all bit-sizes of
2908 * these instructions.
2910 * \sa ::lower_usub_sat64
2915 * Set if only 64-bit nir_op_usub_sat should be lowered to simple
2918 * \sa ::lower_add_sat
2920 bool lower_usub_sat64
;
2923 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
2924 * for IO purposes and would prefer loads/stores be vectorized.
2927 bool lower_to_scalar
;
2930 * Should the linker unify inputs_read/outputs_written between adjacent
2931 * shader stages which are linked into a single program?
2933 bool unify_interfaces
;
2936 * Should nir_lower_io() create load_interpolated_input intrinsics?
2938 * If not, it generates regular load_input intrinsics and interpolation
2939 * information must be inferred from the list of input nir_variables.
2941 bool use_interpolated_input_intrinsics
;
2943 /* Lowers when 32x32->64 bit multiplication is not supported */
2944 bool lower_mul_2x32_64
;
2946 /* Lowers when rotate instruction is not supported */
2950 * Backend supports imul24, and would like to use it (when possible)
2951 * for address/offset calculation. If true, driver should call
2952 * nir_lower_amul(). (If not set, amul will automatically be lowered
2957 /* Whether to generate only scoped_memory_barrier intrinsics instead of the
2958 * set of memory barrier intrinsics based on GLSL.
2960 bool use_scoped_memory_barrier
;
2963 * Is this the Intel vec4 backend?
2965 * Used to inhibit algebraic optimizations that are known to be harmful on
2966 * the Intel vec4 backend. This is generally applicable to any
2967 * optimization that might cause more immediate values to be used in
2968 * 3-source (e.g., ffma and flrp) instructions.
2972 unsigned max_unroll_iterations
;
2974 nir_lower_int64_options lower_int64_options
;
2975 nir_lower_doubles_options lower_doubles_options
;
2976 } nir_shader_compiler_options
;
2978 typedef struct nir_shader
{
2979 /** list of uniforms (nir_variable) */
2980 struct exec_list uniforms
;
2982 /** list of inputs (nir_variable) */
2983 struct exec_list inputs
;
2985 /** list of outputs (nir_variable) */
2986 struct exec_list outputs
;
2988 /** list of shared compute variables (nir_variable) */
2989 struct exec_list shared
;
2991 /** Set of driver-specific options for the shader.
2993 * The memory for the options is expected to be kept in a single static
2994 * copy by the driver.
2996 const struct nir_shader_compiler_options
*options
;
2998 /** Various bits of compile-time information about a given shader */
2999 struct shader_info info
;
3001 /** list of global variables in the shader (nir_variable) */
3002 struct exec_list globals
;
3004 /** list of system value variables in the shader (nir_variable) */
3005 struct exec_list system_values
;
3007 struct exec_list functions
; /** < list of nir_function */
3010 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
3013 unsigned num_inputs
, num_uniforms
, num_outputs
, num_shared
;
3015 /** Size in bytes of required scratch space */
3016 unsigned scratch_size
;
3018 /** Constant data associated with this shader.
3020 * Constant data is loaded through load_constant intrinsics. See also
3021 * nir_opt_large_constants.
3023 void *constant_data
;
3024 unsigned constant_data_size
;
3027 #define nir_foreach_function(func, shader) \
3028 foreach_list_typed(nir_function, func, node, &(shader)->functions)
3030 static inline nir_function_impl
*
3031 nir_shader_get_entrypoint(nir_shader
*shader
)
3033 nir_function
*func
= NULL
;
3035 nir_foreach_function(function
, shader
) {
3036 assert(func
== NULL
);
3037 if (function
->is_entrypoint
) {
3048 assert(func
->num_params
== 0);
3053 nir_shader
*nir_shader_create(void *mem_ctx
,
3054 gl_shader_stage stage
,
3055 const nir_shader_compiler_options
*options
,
3058 nir_register
*nir_local_reg_create(nir_function_impl
*impl
);
3060 void nir_reg_remove(nir_register
*reg
);
3062 /** Adds a variable to the appropriate list in nir_shader */
3063 void nir_shader_add_variable(nir_shader
*shader
, nir_variable
*var
);
3066 nir_function_impl_add_variable(nir_function_impl
*impl
, nir_variable
*var
)
3068 assert(var
->data
.mode
== nir_var_function_temp
);
3069 exec_list_push_tail(&impl
->locals
, &var
->node
);
3072 /** creates a variable, sets a few defaults, and adds it to the list */
3073 nir_variable
*nir_variable_create(nir_shader
*shader
,
3074 nir_variable_mode mode
,
3075 const struct glsl_type
*type
,
3077 /** creates a local variable and adds it to the list */
3078 nir_variable
*nir_local_variable_create(nir_function_impl
*impl
,
3079 const struct glsl_type
*type
,
3082 /** creates a function and adds it to the shader's list of functions */
3083 nir_function
*nir_function_create(nir_shader
*shader
, const char *name
);
3085 nir_function_impl
*nir_function_impl_create(nir_function
*func
);
3086 /** creates a function_impl that isn't tied to any particular function */
3087 nir_function_impl
*nir_function_impl_create_bare(nir_shader
*shader
);
3089 nir_block
*nir_block_create(nir_shader
*shader
);
3090 nir_if
*nir_if_create(nir_shader
*shader
);
3091 nir_loop
*nir_loop_create(nir_shader
*shader
);
3093 nir_function_impl
*nir_cf_node_get_function(nir_cf_node
*node
);
3095 /** requests that the given pieces of metadata be generated */
3096 void nir_metadata_require(nir_function_impl
*impl
, nir_metadata required
, ...);
3097 /** dirties all but the preserved metadata */
3098 void nir_metadata_preserve(nir_function_impl
*impl
, nir_metadata preserved
);
3100 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
3101 nir_alu_instr
*nir_alu_instr_create(nir_shader
*shader
, nir_op op
);
3103 nir_deref_instr
*nir_deref_instr_create(nir_shader
*shader
,
3104 nir_deref_type deref_type
);
3106 nir_jump_instr
*nir_jump_instr_create(nir_shader
*shader
, nir_jump_type type
);
3108 nir_load_const_instr
*nir_load_const_instr_create(nir_shader
*shader
,
3109 unsigned num_components
,
3112 nir_intrinsic_instr
*nir_intrinsic_instr_create(nir_shader
*shader
,
3113 nir_intrinsic_op op
);
3115 nir_call_instr
*nir_call_instr_create(nir_shader
*shader
,
3116 nir_function
*callee
);
3118 nir_tex_instr
*nir_tex_instr_create(nir_shader
*shader
, unsigned num_srcs
);
3120 nir_phi_instr
*nir_phi_instr_create(nir_shader
*shader
);
3122 nir_parallel_copy_instr
*nir_parallel_copy_instr_create(nir_shader
*shader
);
3124 nir_ssa_undef_instr
*nir_ssa_undef_instr_create(nir_shader
*shader
,
3125 unsigned num_components
,
3128 nir_const_value
nir_alu_binop_identity(nir_op binop
, unsigned bit_size
);
3131 * NIR Cursors and Instruction Insertion API
3134 * A tiny struct representing a point to insert/extract instructions or
3135 * control flow nodes. Helps reduce the combinatorial explosion of possible
3136 * points to insert/extract.
3138 * \sa nir_control_flow.h
3141 nir_cursor_before_block
,
3142 nir_cursor_after_block
,
3143 nir_cursor_before_instr
,
3144 nir_cursor_after_instr
,
3145 } nir_cursor_option
;
3148 nir_cursor_option option
;
3155 static inline nir_block
*
3156 nir_cursor_current_block(nir_cursor cursor
)
3158 if (cursor
.option
== nir_cursor_before_instr
||
3159 cursor
.option
== nir_cursor_after_instr
) {
3160 return cursor
.instr
->block
;
3162 return cursor
.block
;
3166 bool nir_cursors_equal(nir_cursor a
, nir_cursor b
);
3168 static inline nir_cursor
3169 nir_before_block(nir_block
*block
)
3172 cursor
.option
= nir_cursor_before_block
;
3173 cursor
.block
= block
;
3177 static inline nir_cursor
3178 nir_after_block(nir_block
*block
)
3181 cursor
.option
= nir_cursor_after_block
;
3182 cursor
.block
= block
;
3186 static inline nir_cursor
3187 nir_before_instr(nir_instr
*instr
)
3190 cursor
.option
= nir_cursor_before_instr
;
3191 cursor
.instr
= instr
;
3195 static inline nir_cursor
3196 nir_after_instr(nir_instr
*instr
)
3199 cursor
.option
= nir_cursor_after_instr
;
3200 cursor
.instr
= instr
;
3204 static inline nir_cursor
3205 nir_after_block_before_jump(nir_block
*block
)
3207 nir_instr
*last_instr
= nir_block_last_instr(block
);
3208 if (last_instr
&& last_instr
->type
== nir_instr_type_jump
) {
3209 return nir_before_instr(last_instr
);
3211 return nir_after_block(block
);
3215 static inline nir_cursor
3216 nir_before_src(nir_src
*src
, bool is_if_condition
)
3218 if (is_if_condition
) {
3219 nir_block
*prev_block
=
3220 nir_cf_node_as_block(nir_cf_node_prev(&src
->parent_if
->cf_node
));
3221 assert(!nir_block_ends_in_jump(prev_block
));
3222 return nir_after_block(prev_block
);
3223 } else if (src
->parent_instr
->type
== nir_instr_type_phi
) {
3225 nir_phi_instr
*cond_phi
= nir_instr_as_phi(src
->parent_instr
);
3227 nir_foreach_phi_src(phi_src
, cond_phi
) {
3228 if (phi_src
->src
.ssa
== src
->ssa
) {
3235 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
3236 * to have a more specific name.
3238 nir_phi_src
*phi_src
= LIST_ENTRY(nir_phi_src
, src
, src
);
3239 return nir_after_block_before_jump(phi_src
->pred
);
3241 return nir_before_instr(src
->parent_instr
);
3245 static inline nir_cursor
3246 nir_before_cf_node(nir_cf_node
*node
)
3248 if (node
->type
== nir_cf_node_block
)
3249 return nir_before_block(nir_cf_node_as_block(node
));
3251 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node
)));
3254 static inline nir_cursor
3255 nir_after_cf_node(nir_cf_node
*node
)
3257 if (node
->type
== nir_cf_node_block
)
3258 return nir_after_block(nir_cf_node_as_block(node
));
3260 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node
)));
3263 static inline nir_cursor
3264 nir_after_phis(nir_block
*block
)
3266 nir_foreach_instr(instr
, block
) {
3267 if (instr
->type
!= nir_instr_type_phi
)
3268 return nir_before_instr(instr
);
3270 return nir_after_block(block
);
3273 static inline nir_cursor
3274 nir_after_cf_node_and_phis(nir_cf_node
*node
)
3276 if (node
->type
== nir_cf_node_block
)
3277 return nir_after_block(nir_cf_node_as_block(node
));
3279 nir_block
*block
= nir_cf_node_as_block(nir_cf_node_next(node
));
3281 return nir_after_phis(block
);
3284 static inline nir_cursor
3285 nir_before_cf_list(struct exec_list
*cf_list
)
3287 nir_cf_node
*first_node
= exec_node_data(nir_cf_node
,
3288 exec_list_get_head(cf_list
), node
);
3289 return nir_before_cf_node(first_node
);
3292 static inline nir_cursor
3293 nir_after_cf_list(struct exec_list
*cf_list
)
3295 nir_cf_node
*last_node
= exec_node_data(nir_cf_node
,
3296 exec_list_get_tail(cf_list
), node
);
3297 return nir_after_cf_node(last_node
);
3301 * Insert a NIR instruction at the given cursor.
3303 * Note: This does not update the cursor.
3305 void nir_instr_insert(nir_cursor cursor
, nir_instr
*instr
);
3308 nir_instr_insert_before(nir_instr
*instr
, nir_instr
*before
)
3310 nir_instr_insert(nir_before_instr(instr
), before
);
3314 nir_instr_insert_after(nir_instr
*instr
, nir_instr
*after
)
3316 nir_instr_insert(nir_after_instr(instr
), after
);
3320 nir_instr_insert_before_block(nir_block
*block
, nir_instr
*before
)
3322 nir_instr_insert(nir_before_block(block
), before
);
3326 nir_instr_insert_after_block(nir_block
*block
, nir_instr
*after
)
3328 nir_instr_insert(nir_after_block(block
), after
);
3332 nir_instr_insert_before_cf(nir_cf_node
*node
, nir_instr
*before
)
3334 nir_instr_insert(nir_before_cf_node(node
), before
);
3338 nir_instr_insert_after_cf(nir_cf_node
*node
, nir_instr
*after
)
3340 nir_instr_insert(nir_after_cf_node(node
), after
);
3344 nir_instr_insert_before_cf_list(struct exec_list
*list
, nir_instr
*before
)
3346 nir_instr_insert(nir_before_cf_list(list
), before
);
3350 nir_instr_insert_after_cf_list(struct exec_list
*list
, nir_instr
*after
)
3352 nir_instr_insert(nir_after_cf_list(list
), after
);
3355 void nir_instr_remove_v(nir_instr
*instr
);
3357 static inline nir_cursor
3358 nir_instr_remove(nir_instr
*instr
)
3361 nir_instr
*prev
= nir_instr_prev(instr
);
3363 cursor
= nir_after_instr(prev
);
3365 cursor
= nir_before_block(instr
->block
);
3367 nir_instr_remove_v(instr
);
3373 nir_ssa_def
*nir_instr_ssa_def(nir_instr
*instr
);
3375 typedef bool (*nir_foreach_ssa_def_cb
)(nir_ssa_def
*def
, void *state
);
3376 typedef bool (*nir_foreach_dest_cb
)(nir_dest
*dest
, void *state
);
3377 typedef bool (*nir_foreach_src_cb
)(nir_src
*src
, void *state
);
3378 bool nir_foreach_ssa_def(nir_instr
*instr
, nir_foreach_ssa_def_cb cb
,
3380 bool nir_foreach_dest(nir_instr
*instr
, nir_foreach_dest_cb cb
, void *state
);
3381 bool nir_foreach_src(nir_instr
*instr
, nir_foreach_src_cb cb
, void *state
);
3383 nir_const_value
*nir_src_as_const_value(nir_src src
);
3385 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
3386 static inline c_type * \
3387 nir_src_as_ ## name (nir_src src) \
3389 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
3390 ? cast_macro(src.ssa->parent_instr) : NULL; \
3393 NIR_SRC_AS_(alu_instr
, nir_alu_instr
, nir_instr_type_alu
, nir_instr_as_alu
)
3394 NIR_SRC_AS_(intrinsic
, nir_intrinsic_instr
,
3395 nir_instr_type_intrinsic
, nir_instr_as_intrinsic
)
3396 NIR_SRC_AS_(deref
, nir_deref_instr
, nir_instr_type_deref
, nir_instr_as_deref
)
3398 bool nir_src_is_dynamically_uniform(nir_src src
);
3399 bool nir_srcs_equal(nir_src src1
, nir_src src2
);
3400 bool nir_instrs_equal(const nir_instr
*instr1
, const nir_instr
*instr2
);
3401 void nir_instr_rewrite_src(nir_instr
*instr
, nir_src
*src
, nir_src new_src
);
3402 void nir_instr_move_src(nir_instr
*dest_instr
, nir_src
*dest
, nir_src
*src
);
3403 void nir_if_rewrite_condition(nir_if
*if_stmt
, nir_src new_src
);
3404 void nir_instr_rewrite_dest(nir_instr
*instr
, nir_dest
*dest
,
3407 void nir_ssa_dest_init(nir_instr
*instr
, nir_dest
*dest
,
3408 unsigned num_components
, unsigned bit_size
,
3410 void nir_ssa_def_init(nir_instr
*instr
, nir_ssa_def
*def
,
3411 unsigned num_components
, unsigned bit_size
,
3414 nir_ssa_dest_init_for_type(nir_instr
*instr
, nir_dest
*dest
,
3415 const struct glsl_type
*type
,
3418 assert(glsl_type_is_vector_or_scalar(type
));
3419 nir_ssa_dest_init(instr
, dest
, glsl_get_components(type
),
3420 glsl_get_bit_size(type
), name
);
3422 void nir_ssa_def_rewrite_uses(nir_ssa_def
*def
, nir_src new_src
);
3423 void nir_ssa_def_rewrite_uses_after(nir_ssa_def
*def
, nir_src new_src
,
3424 nir_instr
*after_me
);
3426 nir_component_mask_t
nir_ssa_def_components_read(const nir_ssa_def
*def
);
3429 * finds the next basic block in source-code order, returns NULL if there is
3433 nir_block
*nir_block_cf_tree_next(nir_block
*block
);
3435 /* Performs the opposite of nir_block_cf_tree_next() */
3437 nir_block
*nir_block_cf_tree_prev(nir_block
*block
);
3439 /* Gets the first block in a CF node in source-code order */
3441 nir_block
*nir_cf_node_cf_tree_first(nir_cf_node
*node
);
3443 /* Gets the last block in a CF node in source-code order */
3445 nir_block
*nir_cf_node_cf_tree_last(nir_cf_node
*node
);
3447 /* Gets the next block after a CF node in source-code order */
3449 nir_block
*nir_cf_node_cf_tree_next(nir_cf_node
*node
);
3451 /* Macros for loops that visit blocks in source-code order */
3453 #define nir_foreach_block(block, impl) \
3454 for (nir_block *block = nir_start_block(impl); block != NULL; \
3455 block = nir_block_cf_tree_next(block))
3457 #define nir_foreach_block_safe(block, impl) \
3458 for (nir_block *block = nir_start_block(impl), \
3459 *next = nir_block_cf_tree_next(block); \
3461 block = next, next = nir_block_cf_tree_next(block))
3463 #define nir_foreach_block_reverse(block, impl) \
3464 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
3465 block = nir_block_cf_tree_prev(block))
3467 #define nir_foreach_block_reverse_safe(block, impl) \
3468 for (nir_block *block = nir_impl_last_block(impl), \
3469 *prev = nir_block_cf_tree_prev(block); \
3471 block = prev, prev = nir_block_cf_tree_prev(block))
3473 #define nir_foreach_block_in_cf_node(block, node) \
3474 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
3475 block != nir_cf_node_cf_tree_next(node); \
3476 block = nir_block_cf_tree_next(block))
3478 /* If the following CF node is an if, this function returns that if.
3479 * Otherwise, it returns NULL.
3481 nir_if
*nir_block_get_following_if(nir_block
*block
);
3483 nir_loop
*nir_block_get_following_loop(nir_block
*block
);
3485 void nir_index_local_regs(nir_function_impl
*impl
);
3486 void nir_index_ssa_defs(nir_function_impl
*impl
);
3487 unsigned nir_index_instrs(nir_function_impl
*impl
);
3489 void nir_index_blocks(nir_function_impl
*impl
);
3491 void nir_index_vars(nir_shader
*shader
, nir_function_impl
*impl
, nir_variable_mode modes
);
3493 void nir_print_shader(nir_shader
*shader
, FILE *fp
);
3494 void nir_print_shader_annotated(nir_shader
*shader
, FILE *fp
, struct hash_table
*errors
);
3495 void nir_print_instr(const nir_instr
*instr
, FILE *fp
);
3496 void nir_print_deref(const nir_deref_instr
*deref
, FILE *fp
);
3498 /** Shallow clone of a single ALU instruction. */
3499 nir_alu_instr
*nir_alu_instr_clone(nir_shader
*s
, const nir_alu_instr
*orig
);
3501 nir_shader
*nir_shader_clone(void *mem_ctx
, const nir_shader
*s
);
3502 nir_function_impl
*nir_function_impl_clone(nir_shader
*shader
,
3503 const nir_function_impl
*fi
);
3504 nir_constant
*nir_constant_clone(const nir_constant
*c
, nir_variable
*var
);
3505 nir_variable
*nir_variable_clone(const nir_variable
*c
, nir_shader
*shader
);
3507 void nir_shader_replace(nir_shader
*dest
, nir_shader
*src
);
3509 void nir_shader_serialize_deserialize(nir_shader
*s
);
3512 void nir_validate_shader(nir_shader
*shader
, const char *when
);
3513 void nir_metadata_set_validation_flag(nir_shader
*shader
);
3514 void nir_metadata_check_validation_flag(nir_shader
*shader
);
3517 should_skip_nir(const char *name
)
3519 static const char *list
= NULL
;
3521 /* Comma separated list of names to skip. */
3522 list
= getenv("NIR_SKIP");
3530 return comma_separated_list_contains(list
, name
);
3534 should_clone_nir(void)
3536 static int should_clone
= -1;
3537 if (should_clone
< 0)
3538 should_clone
= env_var_as_boolean("NIR_TEST_CLONE", false);
3540 return should_clone
;
3544 should_serialize_deserialize_nir(void)
3546 static int test_serialize
= -1;
3547 if (test_serialize
< 0)
3548 test_serialize
= env_var_as_boolean("NIR_TEST_SERIALIZE", false);
3550 return test_serialize
;
3554 should_print_nir(void)
3556 static int should_print
= -1;
3557 if (should_print
< 0)
3558 should_print
= env_var_as_boolean("NIR_PRINT", false);
3560 return should_print
;
3563 static inline void nir_validate_shader(nir_shader
*shader
, const char *when
) { (void) shader
; (void)when
; }
3564 static inline void nir_metadata_set_validation_flag(nir_shader
*shader
) { (void) shader
; }
3565 static inline void nir_metadata_check_validation_flag(nir_shader
*shader
) { (void) shader
; }
3566 static inline bool should_skip_nir(UNUSED
const char *pass_name
) { return false; }
3567 static inline bool should_clone_nir(void) { return false; }
3568 static inline bool should_serialize_deserialize_nir(void) { return false; }
3569 static inline bool should_print_nir(void) { return false; }
3572 #define _PASS(pass, nir, do_pass) do { \
3573 if (should_skip_nir(#pass)) { \
3574 printf("skipping %s\n", #pass); \
3578 nir_validate_shader(nir, "after " #pass); \
3579 if (should_clone_nir()) { \
3580 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3581 nir_shader_replace(nir, clone); \
3583 if (should_serialize_deserialize_nir()) { \
3584 nir_shader_serialize_deserialize(nir); \
3588 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3589 nir_metadata_set_validation_flag(nir); \
3590 if (should_print_nir()) \
3591 printf("%s\n", #pass); \
3592 if (pass(nir, ##__VA_ARGS__)) { \
3594 if (should_print_nir()) \
3595 nir_print_shader(nir, stdout); \
3596 nir_metadata_check_validation_flag(nir); \
3600 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3601 if (should_print_nir()) \
3602 printf("%s\n", #pass); \
3603 pass(nir, ##__VA_ARGS__); \
3604 if (should_print_nir()) \
3605 nir_print_shader(nir, stdout); \
3608 #define NIR_SKIP(name) should_skip_nir(#name)
3610 /** An instruction filtering callback
3612 * Returns true if the instruction should be processed and false otherwise.
3614 typedef bool (*nir_instr_filter_cb
)(const nir_instr
*, const void *);
3616 /** A simple instruction lowering callback
3618 * Many instruction lowering passes can be written as a simple function which
3619 * takes an instruction as its input and returns a sequence of instructions
3620 * that implement the consumed instruction. This function type represents
3621 * such a lowering function. When called, a function with this prototype
3622 * should either return NULL indicating that no lowering needs to be done or
3623 * emit a sequence of instructions using the provided builder (whose cursor
3624 * will already be placed after the instruction to be lowered) and return the
3625 * resulting nir_ssa_def.
3627 typedef nir_ssa_def
*(*nir_lower_instr_cb
)(struct nir_builder
*,
3628 nir_instr
*, void *);
3631 * Special return value for nir_lower_instr_cb when some progress occurred
3632 * (like changing an input to the instr) that didn't result in a replacement
3633 * SSA def being generated.
3635 #define NIR_LOWER_INSTR_PROGRESS ((nir_ssa_def *)(uintptr_t)1)
3637 /** Iterate over all the instructions in a nir_function_impl and lower them
3638 * using the provided callbacks
3640 * This function implements the guts of a standard lowering pass for you. It
3641 * iterates over all of the instructions in a nir_function_impl and calls the
3642 * filter callback on each one. If the filter callback returns true, it then
3643 * calls the lowering call back on the instruction. (Splitting it this way
3644 * allows us to avoid some save/restore work for instructions we know won't be
3645 * lowered.) If the instruction is dead after the lowering is complete, it
3646 * will be removed. If new instructions are added, the lowering callback will
3647 * also be called on them in case multiple lowerings are required.
3649 * The metadata for the nir_function_impl will also be updated. If any blocks
3650 * are added (they cannot be removed), dominance and block indices will be
3653 bool nir_function_impl_lower_instructions(nir_function_impl
*impl
,
3654 nir_instr_filter_cb filter
,
3655 nir_lower_instr_cb lower
,
3657 bool nir_shader_lower_instructions(nir_shader
*shader
,
3658 nir_instr_filter_cb filter
,
3659 nir_lower_instr_cb lower
,
3662 void nir_calc_dominance_impl(nir_function_impl
*impl
);
3663 void nir_calc_dominance(nir_shader
*shader
);
3665 nir_block
*nir_dominance_lca(nir_block
*b1
, nir_block
*b2
);
3666 bool nir_block_dominates(nir_block
*parent
, nir_block
*child
);
3667 bool nir_block_is_unreachable(nir_block
*block
);
3669 void nir_dump_dom_tree_impl(nir_function_impl
*impl
, FILE *fp
);
3670 void nir_dump_dom_tree(nir_shader
*shader
, FILE *fp
);
3672 void nir_dump_dom_frontier_impl(nir_function_impl
*impl
, FILE *fp
);
3673 void nir_dump_dom_frontier(nir_shader
*shader
, FILE *fp
);
3675 void nir_dump_cfg_impl(nir_function_impl
*impl
, FILE *fp
);
3676 void nir_dump_cfg(nir_shader
*shader
, FILE *fp
);
3678 int nir_gs_count_vertices(const nir_shader
*shader
);
3680 bool nir_shrink_vec_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3681 bool nir_split_array_vars(nir_shader
*shader
, nir_variable_mode modes
);
3682 bool nir_split_var_copies(nir_shader
*shader
);
3683 bool nir_split_per_member_structs(nir_shader
*shader
);
3684 bool nir_split_struct_vars(nir_shader
*shader
, nir_variable_mode modes
);
3686 bool nir_lower_returns_impl(nir_function_impl
*impl
);
3687 bool nir_lower_returns(nir_shader
*shader
);
3689 void nir_inline_function_impl(struct nir_builder
*b
,
3690 const nir_function_impl
*impl
,
3691 nir_ssa_def
**params
);
3692 bool nir_inline_functions(nir_shader
*shader
);
3694 bool nir_propagate_invariant(nir_shader
*shader
);
3696 void nir_lower_var_copy_instr(nir_intrinsic_instr
*copy
, nir_shader
*shader
);
3697 void nir_lower_deref_copy_instr(struct nir_builder
*b
,
3698 nir_intrinsic_instr
*copy
);
3699 bool nir_lower_var_copies(nir_shader
*shader
);
3701 void nir_fixup_deref_modes(nir_shader
*shader
);
3703 bool nir_lower_global_vars_to_local(nir_shader
*shader
);
3706 nir_lower_direct_array_deref_of_vec_load
= (1 << 0),
3707 nir_lower_indirect_array_deref_of_vec_load
= (1 << 1),
3708 nir_lower_direct_array_deref_of_vec_store
= (1 << 2),
3709 nir_lower_indirect_array_deref_of_vec_store
= (1 << 3),
3710 } nir_lower_array_deref_of_vec_options
;
3712 bool nir_lower_array_deref_of_vec(nir_shader
*shader
, nir_variable_mode modes
,
3713 nir_lower_array_deref_of_vec_options options
);
3715 bool nir_lower_indirect_derefs(nir_shader
*shader
, nir_variable_mode modes
);
3717 bool nir_lower_locals_to_regs(nir_shader
*shader
);
3719 void nir_lower_io_to_temporaries(nir_shader
*shader
,
3720 nir_function_impl
*entrypoint
,
3721 bool outputs
, bool inputs
);
3723 bool nir_lower_vars_to_scratch(nir_shader
*shader
,
3724 nir_variable_mode modes
,
3726 glsl_type_size_align_func size_align
);
3728 void nir_lower_clip_halfz(nir_shader
*shader
);
3730 void nir_shader_gather_info(nir_shader
*shader
, nir_function_impl
*entrypoint
);
3732 void nir_gather_ssa_types(nir_function_impl
*impl
,
3733 BITSET_WORD
*float_types
,
3734 BITSET_WORD
*int_types
);
3736 void nir_assign_var_locations(struct exec_list
*var_list
, unsigned *size
,
3737 int (*type_size
)(const struct glsl_type
*, bool));
3739 /* Some helpers to do very simple linking */
3740 bool nir_remove_unused_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3741 bool nir_remove_unused_io_vars(nir_shader
*shader
, struct exec_list
*var_list
,
3742 uint64_t *used_by_other_stage
,
3743 uint64_t *used_by_other_stage_patches
);
3744 void nir_compact_varyings(nir_shader
*producer
, nir_shader
*consumer
,
3745 bool default_to_smooth_interp
);
3746 void nir_link_xfb_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3747 bool nir_link_opt_varyings(nir_shader
*producer
, nir_shader
*consumer
);
3749 bool nir_lower_amul(nir_shader
*shader
,
3750 int (*type_size
)(const struct glsl_type
*, bool));
3752 void nir_assign_io_var_locations(struct exec_list
*var_list
,
3754 gl_shader_stage stage
);
3757 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
3758 * to 32-bit operations. This is only valid for nir_var_shader_in/out
3761 nir_lower_io_lower_64bit_to_32
= (1 << 0),
3763 /* If set, this forces all non-flat fragment shader inputs to be
3764 * interpolated as if with the "sample" qualifier. This requires
3765 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
3767 nir_lower_io_force_sample_interpolation
= (1 << 1),
3768 } nir_lower_io_options
;
3769 bool nir_lower_io(nir_shader
*shader
,
3770 nir_variable_mode modes
,
3771 int (*type_size
)(const struct glsl_type
*, bool),
3772 nir_lower_io_options
);
3774 bool nir_io_add_const_offset_to_base(nir_shader
*nir
, nir_variable_mode mode
);
3777 nir_lower_vars_to_explicit_types(nir_shader
*shader
,
3778 nir_variable_mode modes
,
3779 glsl_type_size_align_func type_info
);
3783 * An address format which is a simple 32-bit global GPU address.
3785 nir_address_format_32bit_global
,
3788 * An address format which is a simple 64-bit global GPU address.
3790 nir_address_format_64bit_global
,
3793 * An address format which is a bounds-checked 64-bit global GPU address.
3795 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
3796 * address stored with the low bits in .x and high bits in .y, .z is a
3797 * size, and .w is an offset. When the final I/O operation is lowered, .w
3798 * is checked against .z and the operation is predicated on the result.
3800 nir_address_format_64bit_bounded_global
,
3803 * An address format which is comprised of a vec2 where the first
3804 * component is a buffer index and the second is an offset.
3806 nir_address_format_32bit_index_offset
,
3809 * An address format which is a simple 32-bit offset.
3811 nir_address_format_32bit_offset
,
3814 * An address format representing a purely logical addressing model. In
3815 * this model, all deref chains must be complete from the dereference
3816 * operation to the variable. Cast derefs are not allowed. These
3817 * addresses will be 32-bit scalars but the format is immaterial because
3818 * you can always chase the chain.
3820 nir_address_format_logical
,
3821 } nir_address_format
;
3823 static inline unsigned
3824 nir_address_format_bit_size(nir_address_format addr_format
)
3826 switch (addr_format
) {
3827 case nir_address_format_32bit_global
: return 32;
3828 case nir_address_format_64bit_global
: return 64;
3829 case nir_address_format_64bit_bounded_global
: return 32;
3830 case nir_address_format_32bit_index_offset
: return 32;
3831 case nir_address_format_32bit_offset
: return 32;
3832 case nir_address_format_logical
: return 32;
3834 unreachable("Invalid address format");
3837 static inline unsigned
3838 nir_address_format_num_components(nir_address_format addr_format
)
3840 switch (addr_format
) {
3841 case nir_address_format_32bit_global
: return 1;
3842 case nir_address_format_64bit_global
: return 1;
3843 case nir_address_format_64bit_bounded_global
: return 4;
3844 case nir_address_format_32bit_index_offset
: return 2;
3845 case nir_address_format_32bit_offset
: return 1;
3846 case nir_address_format_logical
: return 1;
3848 unreachable("Invalid address format");
3851 static inline const struct glsl_type
*
3852 nir_address_format_to_glsl_type(nir_address_format addr_format
)
3854 unsigned bit_size
= nir_address_format_bit_size(addr_format
);
3855 assert(bit_size
== 32 || bit_size
== 64);
3856 return glsl_vector_type(bit_size
== 32 ? GLSL_TYPE_UINT
: GLSL_TYPE_UINT64
,
3857 nir_address_format_num_components(addr_format
));
3860 const nir_const_value
*nir_address_format_null_value(nir_address_format addr_format
);
3862 nir_ssa_def
*nir_build_addr_ieq(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3863 nir_address_format addr_format
);
3865 nir_ssa_def
*nir_build_addr_isub(struct nir_builder
*b
, nir_ssa_def
*addr0
, nir_ssa_def
*addr1
,
3866 nir_address_format addr_format
);
3868 nir_ssa_def
* nir_explicit_io_address_from_deref(struct nir_builder
*b
,
3869 nir_deref_instr
*deref
,
3870 nir_ssa_def
*base_addr
,
3871 nir_address_format addr_format
);
3872 void nir_lower_explicit_io_instr(struct nir_builder
*b
,
3873 nir_intrinsic_instr
*io_instr
,
3875 nir_address_format addr_format
);
3877 bool nir_lower_explicit_io(nir_shader
*shader
,
3878 nir_variable_mode modes
,
3879 nir_address_format
);
3881 nir_src
*nir_get_io_offset_src(nir_intrinsic_instr
*instr
);
3882 nir_src
*nir_get_io_vertex_index_src(nir_intrinsic_instr
*instr
);
3884 bool nir_is_per_vertex_io(const nir_variable
*var
, gl_shader_stage stage
);
3886 bool nir_lower_regs_to_ssa_impl(nir_function_impl
*impl
);
3887 bool nir_lower_regs_to_ssa(nir_shader
*shader
);
3888 bool nir_lower_vars_to_ssa(nir_shader
*shader
);
3890 bool nir_remove_dead_derefs(nir_shader
*shader
);
3891 bool nir_remove_dead_derefs_impl(nir_function_impl
*impl
);
3892 bool nir_remove_dead_variables(nir_shader
*shader
, nir_variable_mode modes
);
3893 bool nir_lower_variable_initializers(nir_shader
*shader
,
3894 nir_variable_mode modes
);
3896 bool nir_move_vec_src_uses_to_dest(nir_shader
*shader
);
3897 bool nir_lower_vec_to_movs(nir_shader
*shader
);
3898 void nir_lower_alpha_test(nir_shader
*shader
, enum compare_func func
,
3900 const gl_state_index16
*alpha_ref_state_tokens
);
3901 bool nir_lower_alu(nir_shader
*shader
);
3903 bool nir_lower_flrp(nir_shader
*shader
, unsigned lowering_mask
,
3904 bool always_precise
, bool have_ffma
);
3906 bool nir_lower_alu_to_scalar(nir_shader
*shader
, nir_instr_filter_cb cb
, const void *data
);
3907 bool nir_lower_bool_to_bitsize(nir_shader
*shader
);
3908 bool nir_lower_bool_to_float(nir_shader
*shader
);
3909 bool nir_lower_bool_to_int32(nir_shader
*shader
);
3910 bool nir_lower_int_to_float(nir_shader
*shader
);
3911 bool nir_lower_load_const_to_scalar(nir_shader
*shader
);
3912 bool nir_lower_read_invocation_to_scalar(nir_shader
*shader
);
3913 bool nir_lower_phis_to_scalar(nir_shader
*shader
);
3914 void nir_lower_io_arrays_to_elements(nir_shader
*producer
, nir_shader
*consumer
);
3915 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader
*shader
,
3917 void nir_lower_io_to_scalar(nir_shader
*shader
, nir_variable_mode mask
);
3918 void nir_lower_io_to_scalar_early(nir_shader
*shader
, nir_variable_mode mask
);
3919 bool nir_lower_io_to_vector(nir_shader
*shader
, nir_variable_mode mask
);
3921 void nir_lower_fragcoord_wtrans(nir_shader
*shader
);
3922 void nir_lower_viewport_transform(nir_shader
*shader
);
3923 bool nir_lower_uniforms_to_ubo(nir_shader
*shader
, int multiplier
);
3925 typedef struct nir_lower_subgroups_options
{
3926 uint8_t subgroup_size
;
3927 uint8_t ballot_bit_size
;
3928 bool lower_to_scalar
:1;
3929 bool lower_vote_trivial
:1;
3930 bool lower_vote_eq_to_ballot
:1;
3931 bool lower_subgroup_masks
:1;
3932 bool lower_shuffle
:1;
3933 bool lower_shuffle_to_32bit
:1;
3935 bool lower_quad_broadcast_dynamic
:1;
3936 bool lower_quad_broadcast_dynamic_to_const
:1;
3937 } nir_lower_subgroups_options
;
3939 bool nir_lower_subgroups(nir_shader
*shader
,
3940 const nir_lower_subgroups_options
*options
);
3942 bool nir_lower_system_values(nir_shader
*shader
);
3944 enum PACKED nir_lower_tex_packing
{
3945 nir_lower_tex_packing_none
= 0,
3946 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
3947 * or unsigned ints based on the sampler type
3949 nir_lower_tex_packing_16
,
3950 /* The sampler returns 1 32-bit word of 4x8 unorm */
3951 nir_lower_tex_packing_8
,
3954 typedef struct nir_lower_tex_options
{
3956 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
3957 * sampler types a texture projector is lowered.
3962 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
3964 bool lower_txf_offset
;
3967 * If true, lower away nir_tex_src_offset for all rect textures.
3969 bool lower_rect_offset
;
3972 * If true, lower rect textures to 2D, using txs to fetch the
3973 * texture dimensions and dividing the texture coords by the
3974 * texture dims to normalize.
3979 * If true, convert yuv to rgb.
3981 unsigned lower_y_uv_external
;
3982 unsigned lower_y_u_v_external
;
3983 unsigned lower_yx_xuxv_external
;
3984 unsigned lower_xy_uxvx_external
;
3985 unsigned lower_ayuv_external
;
3986 unsigned lower_xyuv_external
;
3989 * To emulate certain texture wrap modes, this can be used
3990 * to saturate the specified tex coord to [0.0, 1.0]. The
3991 * bits are according to sampler #, ie. if, for example:
3993 * (conf->saturate_s & (1 << n))
3995 * is true, then the s coord for sampler n is saturated.
3997 * Note that clamping must happen *after* projector lowering
3998 * so any projected texture sample instruction with a clamped
3999 * coordinate gets automatically lowered, regardless of the
4000 * 'lower_txp' setting.
4002 unsigned saturate_s
;
4003 unsigned saturate_t
;
4004 unsigned saturate_r
;
4006 /* Bitmask of textures that need swizzling.
4008 * If (swizzle_result & (1 << texture_index)), then the swizzle in
4009 * swizzles[texture_index] is applied to the result of the texturing
4012 unsigned swizzle_result
;
4014 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
4015 * while 4 and 5 represent 0 and 1 respectively.
4017 uint8_t swizzles
[32][4];
4019 /* Can be used to scale sampled values in range required by the format. */
4020 float scale_factors
[32];
4023 * Bitmap of textures that need srgb to linear conversion. If
4024 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
4025 * of the texture are lowered to linear.
4027 unsigned lower_srgb
;
4030 * If true, lower nir_texop_tex on shaders that doesn't support implicit
4031 * LODs to nir_texop_txl.
4033 bool lower_tex_without_implicit_lod
;
4036 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
4038 bool lower_txd_cube_map
;
4041 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
4046 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
4047 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
4048 * with lower_txd_cube_map.
4050 bool lower_txd_shadow
;
4053 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
4054 * Implies lower_txd_cube_map and lower_txd_shadow.
4059 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
4060 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
4062 bool lower_txb_shadow_clamp
;
4065 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
4066 * with nir_texop_txl. This includes cube maps.
4068 bool lower_txd_shadow_clamp
;
4071 * If true, lower nir_texop_txd on when it uses both offset and min_lod
4072 * with nir_texop_txl. This includes cube maps.
4074 bool lower_txd_offset_clamp
;
4077 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4078 * sampler is bindless.
4080 bool lower_txd_clamp_bindless_sampler
;
4083 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
4084 * sampler index is not statically determinable to be less than 16.
4086 bool lower_txd_clamp_if_sampler_index_not_lt_16
;
4089 * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
4090 * 0-lod followed by a nir_ishr.
4095 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
4096 * mixed-up tg4 locations.
4098 bool lower_tg4_broadcom_swizzle
;
4101 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
4103 bool lower_tg4_offsets
;
4105 enum nir_lower_tex_packing lower_tex_packing
[32];
4106 } nir_lower_tex_options
;
4108 bool nir_lower_tex(nir_shader
*shader
,
4109 const nir_lower_tex_options
*options
);
4111 enum nir_lower_non_uniform_access_type
{
4112 nir_lower_non_uniform_ubo_access
= (1 << 0),
4113 nir_lower_non_uniform_ssbo_access
= (1 << 1),
4114 nir_lower_non_uniform_texture_access
= (1 << 2),
4115 nir_lower_non_uniform_image_access
= (1 << 3),
4118 bool nir_lower_non_uniform_access(nir_shader
*shader
,
4119 enum nir_lower_non_uniform_access_type
);
4121 enum nir_lower_idiv_path
{
4122 /* This path is based on NV50LegalizeSSA::handleDIV(). It is the faster of
4123 * the two but it is not exact in some cases (for example, 1091317713u /
4124 * 1034u gives 5209173 instead of 1055432) */
4125 nir_lower_idiv_fast
,
4126 /* This path is based on AMDGPUTargetLowering::LowerUDIVREM() and
4127 * AMDGPUTargetLowering::LowerSDIVREM(). It requires more instructions than
4128 * the nv50 path and many of them are integer multiplications, so it is
4129 * probably slower. It should always return the correct result, though. */
4130 nir_lower_idiv_precise
,
4133 bool nir_lower_idiv(nir_shader
*shader
, enum nir_lower_idiv_path path
);
4135 bool nir_lower_input_attachments(nir_shader
*shader
, bool use_fragcoord_sysval
);
4137 bool nir_lower_clip_vs(nir_shader
*shader
, unsigned ucp_enables
,
4139 bool use_clipdist_array
,
4140 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4141 bool nir_lower_clip_gs(nir_shader
*shader
, unsigned ucp_enables
,
4142 bool use_clipdist_array
,
4143 const gl_state_index16 clipplane_state_tokens
[][STATE_LENGTH
]);
4144 bool nir_lower_clip_fs(nir_shader
*shader
, unsigned ucp_enables
,
4145 bool use_clipdist_array
);
4146 bool nir_lower_clip_cull_distance_arrays(nir_shader
*nir
);
4148 void nir_lower_point_size_mov(nir_shader
*shader
,
4149 const gl_state_index16
*pointsize_state_tokens
);
4151 bool nir_lower_frexp(nir_shader
*nir
);
4153 void nir_lower_two_sided_color(nir_shader
*shader
);
4155 bool nir_lower_clamp_color_outputs(nir_shader
*shader
);
4157 bool nir_lower_flatshade(nir_shader
*shader
);
4159 void nir_lower_passthrough_edgeflags(nir_shader
*shader
);
4160 bool nir_lower_patch_vertices(nir_shader
*nir
, unsigned static_count
,
4161 const gl_state_index16
*uniform_state_tokens
);
4163 typedef struct nir_lower_wpos_ytransform_options
{
4164 gl_state_index16 state_tokens
[STATE_LENGTH
];
4165 bool fs_coord_origin_upper_left
:1;
4166 bool fs_coord_origin_lower_left
:1;
4167 bool fs_coord_pixel_center_integer
:1;
4168 bool fs_coord_pixel_center_half_integer
:1;
4169 } nir_lower_wpos_ytransform_options
;
4171 bool nir_lower_wpos_ytransform(nir_shader
*shader
,
4172 const nir_lower_wpos_ytransform_options
*options
);
4173 bool nir_lower_wpos_center(nir_shader
*shader
, const bool for_sample_shading
);
4175 bool nir_lower_fb_read(nir_shader
*shader
);
4177 typedef struct nir_lower_drawpixels_options
{
4178 gl_state_index16 texcoord_state_tokens
[STATE_LENGTH
];
4179 gl_state_index16 scale_state_tokens
[STATE_LENGTH
];
4180 gl_state_index16 bias_state_tokens
[STATE_LENGTH
];
4181 unsigned drawpix_sampler
;
4182 unsigned pixelmap_sampler
;
4184 bool scale_and_bias
:1;
4185 } nir_lower_drawpixels_options
;
4187 void nir_lower_drawpixels(nir_shader
*shader
,
4188 const nir_lower_drawpixels_options
*options
);
4190 typedef struct nir_lower_bitmap_options
{
4193 } nir_lower_bitmap_options
;
4195 void nir_lower_bitmap(nir_shader
*shader
, const nir_lower_bitmap_options
*options
);
4197 bool nir_lower_atomics_to_ssbo(nir_shader
*shader
);
4200 nir_lower_int_source_mods
= 1 << 0,
4201 nir_lower_float_source_mods
= 1 << 1,
4202 nir_lower_triop_abs
= 1 << 2,
4203 nir_lower_all_source_mods
= (1 << 3) - 1
4204 } nir_lower_to_source_mods_flags
;
4207 bool nir_lower_to_source_mods(nir_shader
*shader
, nir_lower_to_source_mods_flags options
);
4209 bool nir_lower_gs_intrinsics(nir_shader
*shader
, bool per_stream
);
4211 typedef unsigned (*nir_lower_bit_size_callback
)(const nir_alu_instr
*, void *);
4213 bool nir_lower_bit_size(nir_shader
*shader
,
4214 nir_lower_bit_size_callback callback
,
4215 void *callback_data
);
4217 nir_lower_int64_options
nir_lower_int64_op_to_options_mask(nir_op opcode
);
4218 bool nir_lower_int64(nir_shader
*shader
, nir_lower_int64_options options
);
4220 nir_lower_doubles_options
nir_lower_doubles_op_to_options_mask(nir_op opcode
);
4221 bool nir_lower_doubles(nir_shader
*shader
, const nir_shader
*softfp64
,
4222 nir_lower_doubles_options options
);
4223 bool nir_lower_pack(nir_shader
*shader
);
4225 bool nir_lower_point_size(nir_shader
*shader
, float min
, float max
);
4228 nir_lower_interpolation_at_sample
= (1 << 1),
4229 nir_lower_interpolation_at_offset
= (1 << 2),
4230 nir_lower_interpolation_centroid
= (1 << 3),
4231 nir_lower_interpolation_pixel
= (1 << 4),
4232 nir_lower_interpolation_sample
= (1 << 5),
4233 } nir_lower_interpolation_options
;
4235 bool nir_lower_interpolation(nir_shader
*shader
,
4236 nir_lower_interpolation_options options
);
4238 bool nir_lower_discard_to_demote(nir_shader
*shader
);
4240 bool nir_normalize_cubemap_coords(nir_shader
*shader
);
4242 void nir_live_ssa_defs_impl(nir_function_impl
*impl
);
4244 void nir_loop_analyze_impl(nir_function_impl
*impl
,
4245 nir_variable_mode indirect_mask
);
4247 bool nir_ssa_defs_interfere(nir_ssa_def
*a
, nir_ssa_def
*b
);
4249 bool nir_repair_ssa_impl(nir_function_impl
*impl
);
4250 bool nir_repair_ssa(nir_shader
*shader
);
4252 void nir_convert_loop_to_lcssa(nir_loop
*loop
);
4253 bool nir_convert_to_lcssa(nir_shader
*shader
, bool skip_invariants
, bool skip_bool_invariants
);
4254 bool* nir_divergence_analysis(nir_shader
*shader
, nir_divergence_options options
);
4256 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
4257 * registers. If false, convert all values (even those not involved in a phi
4258 * node) to registers.
4260 bool nir_convert_from_ssa(nir_shader
*shader
, bool phi_webs_only
);
4262 bool nir_lower_phis_to_regs_block(nir_block
*block
);
4263 bool nir_lower_ssa_defs_to_regs_block(nir_block
*block
);
4264 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl
*impl
);
4266 bool nir_lower_samplers(nir_shader
*shader
);
4267 bool nir_lower_ssbo(nir_shader
*shader
);
4269 /* This is here for unit tests. */
4270 bool nir_opt_comparison_pre_impl(nir_function_impl
*impl
);
4272 bool nir_opt_comparison_pre(nir_shader
*shader
);
4274 bool nir_opt_access(nir_shader
*shader
);
4275 bool nir_opt_algebraic(nir_shader
*shader
);
4276 bool nir_opt_algebraic_before_ffma(nir_shader
*shader
);
4277 bool nir_opt_algebraic_late(nir_shader
*shader
);
4278 bool nir_opt_algebraic_distribute_src_mods(nir_shader
*shader
);
4279 bool nir_opt_constant_folding(nir_shader
*shader
);
4281 /* Try to combine a and b into a. Return true if combination was possible,
4282 * which will result in b being removed by the pass. Return false if
4283 * combination wasn't possible.
4285 typedef bool (*nir_combine_memory_barrier_cb
)(
4286 nir_intrinsic_instr
*a
, nir_intrinsic_instr
*b
, void *data
);
4288 bool nir_opt_combine_memory_barriers(nir_shader
*shader
,
4289 nir_combine_memory_barrier_cb combine_cb
,
4292 bool nir_opt_combine_stores(nir_shader
*shader
, nir_variable_mode modes
);
4294 bool nir_copy_prop(nir_shader
*shader
);
4296 bool nir_opt_copy_prop_vars(nir_shader
*shader
);
4298 bool nir_opt_cse(nir_shader
*shader
);
4300 bool nir_opt_dce(nir_shader
*shader
);
4302 bool nir_opt_dead_cf(nir_shader
*shader
);
4304 bool nir_opt_dead_write_vars(nir_shader
*shader
);
4306 bool nir_opt_deref_impl(nir_function_impl
*impl
);
4307 bool nir_opt_deref(nir_shader
*shader
);
4309 bool nir_opt_find_array_copies(nir_shader
*shader
);
4311 bool nir_opt_gcm(nir_shader
*shader
, bool value_number
);
4313 bool nir_opt_idiv_const(nir_shader
*shader
, unsigned min_bit_size
);
4315 bool nir_opt_if(nir_shader
*shader
, bool aggressive_last_continue
);
4317 bool nir_opt_intrinsics(nir_shader
*shader
);
4319 bool nir_opt_large_constants(nir_shader
*shader
,
4320 glsl_type_size_align_func size_align
,
4321 unsigned threshold
);
4323 bool nir_opt_loop_unroll(nir_shader
*shader
, nir_variable_mode indirect_mask
);
4326 nir_move_const_undef
= (1 << 0),
4327 nir_move_load_ubo
= (1 << 1),
4328 nir_move_load_input
= (1 << 2),
4329 nir_move_comparisons
= (1 << 3),
4330 nir_move_copies
= (1 << 4),
4333 bool nir_can_move_instr(nir_instr
*instr
, nir_move_options options
);
4335 bool nir_opt_sink(nir_shader
*shader
, nir_move_options options
);
4337 bool nir_opt_move(nir_shader
*shader
, nir_move_options options
);
4339 bool nir_opt_peephole_select(nir_shader
*shader
, unsigned limit
,
4340 bool indirect_load_ok
, bool expensive_alu_ok
);
4342 bool nir_opt_rematerialize_compares(nir_shader
*shader
);
4344 bool nir_opt_remove_phis(nir_shader
*shader
);
4345 bool nir_opt_remove_phis_block(nir_block
*block
);
4347 bool nir_opt_shrink_load(nir_shader
*shader
);
4349 bool nir_opt_trivial_continues(nir_shader
*shader
);
4351 bool nir_opt_undef(nir_shader
*shader
);
4353 bool nir_opt_vectorize(nir_shader
*shader
);
4355 bool nir_opt_conditional_discard(nir_shader
*shader
);
4357 typedef bool (*nir_should_vectorize_mem_func
)(unsigned align
, unsigned bit_size
,
4358 unsigned num_components
, unsigned high_offset
,
4359 nir_intrinsic_instr
*low
, nir_intrinsic_instr
*high
);
4361 bool nir_opt_load_store_vectorize(nir_shader
*shader
, nir_variable_mode modes
,
4362 nir_should_vectorize_mem_func callback
);
4364 void nir_schedule(nir_shader
*shader
, int threshold
);
4366 void nir_strip(nir_shader
*shader
);
4368 void nir_sweep(nir_shader
*shader
);
4370 void nir_remap_dual_slot_attributes(nir_shader
*shader
,
4371 uint64_t *dual_slot_inputs
);
4372 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs
, uint64_t dual_slot
);
4374 nir_intrinsic_op
nir_intrinsic_from_system_value(gl_system_value val
);
4375 gl_system_value
nir_system_value_from_intrinsic(nir_intrinsic_op intrin
);
4378 nir_variable_is_in_ubo(const nir_variable
*var
)
4380 return (var
->data
.mode
== nir_var_mem_ubo
&&
4381 var
->interface_type
!= NULL
);
4385 nir_variable_is_in_ssbo(const nir_variable
*var
)
4387 return (var
->data
.mode
== nir_var_mem_ssbo
&&
4388 var
->interface_type
!= NULL
);
4392 nir_variable_is_in_block(const nir_variable
*var
)
4394 return nir_variable_is_in_ubo(var
) || nir_variable_is_in_ssbo(var
);