1 /*
2 * Copyright © 2014 Connor Abbott
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Connor Abbott (cwabbott0@gmail.com)
25 *
26 */
27
28 #ifndef NIR_H
29 #define NIR_H
30
31 #include <stdint.h>
32 #include "compiler/glsl_types.h"
33 #include "compiler/glsl/list.h"
34 #include "compiler/shader_enums.h"
35 #include "compiler/shader_info.h"
36 #include "util/bitscan.h"
37 #include "util/bitset.h"
38 #include "util/compiler.h"
39 #include "util/enum_operators.h"
40 #include "util/format/u_format.h"
41 #include "util/hash_table.h"
42 #include "util/list.h"
43 #include "util/log.h"
44 #include "util/macros.h"
45 #include "util/ralloc.h"
46 #include "util/set.h"
47 #include "util/u_math.h"
48 #include "util/u_printf.h"
49 #include "nir_defines.h"
50 #define XXH_INLINE_ALL
51 #include <stdio.h>
52 #include "util/xxhash.h"
53
54 #ifndef NDEBUG
55 #include "util/u_debug.h"
56 #endif /* NDEBUG */
57
58 #include "nir_opcodes.h"
59
60 #ifdef __cplusplus
61 extern "C" {
62 #endif
63
64 extern uint32_t nir_debug;
65 extern bool nir_debug_print_shader[MESA_SHADER_KERNEL + 1];
66
67 #ifndef NDEBUG
68 #define NIR_DEBUG(flag) unlikely(nir_debug &(NIR_DEBUG_##flag))
69 #else
70 #define NIR_DEBUG(flag) false
71 #endif
72
73 #define NIR_DEBUG_CLONE (1u << 0)
74 #define NIR_DEBUG_SERIALIZE (1u << 1)
75 #define NIR_DEBUG_NOVALIDATE (1u << 2)
76 #define NIR_DEBUG_TGSI (1u << 4)
77 #define NIR_DEBUG_PRINT_VS (1u << 5)
78 #define NIR_DEBUG_PRINT_TCS (1u << 6)
79 #define NIR_DEBUG_PRINT_TES (1u << 7)
80 #define NIR_DEBUG_PRINT_GS (1u << 8)
81 #define NIR_DEBUG_PRINT_FS (1u << 9)
82 #define NIR_DEBUG_PRINT_CS (1u << 10)
83 #define NIR_DEBUG_PRINT_TS (1u << 11)
84 #define NIR_DEBUG_PRINT_MS (1u << 12)
85 #define NIR_DEBUG_PRINT_RGS (1u << 13)
86 #define NIR_DEBUG_PRINT_AHS (1u << 14)
87 #define NIR_DEBUG_PRINT_CHS (1u << 15)
88 #define NIR_DEBUG_PRINT_MHS (1u << 16)
89 #define NIR_DEBUG_PRINT_IS (1u << 17)
90 #define NIR_DEBUG_PRINT_CBS (1u << 18)
91 #define NIR_DEBUG_PRINT_KS (1u << 19)
92 #define NIR_DEBUG_PRINT_NO_INLINE_CONSTS (1u << 20)
93 #define NIR_DEBUG_PRINT_INTERNAL (1u << 21)
94 #define NIR_DEBUG_PRINT_PASS_FLAGS (1u << 22)
95
96 #define NIR_DEBUG_PRINT (NIR_DEBUG_PRINT_VS | \
97 NIR_DEBUG_PRINT_TCS | \
98 NIR_DEBUG_PRINT_TES | \
99 NIR_DEBUG_PRINT_GS | \
100 NIR_DEBUG_PRINT_FS | \
101 NIR_DEBUG_PRINT_CS | \
102 NIR_DEBUG_PRINT_TS | \
103 NIR_DEBUG_PRINT_MS | \
104 NIR_DEBUG_PRINT_RGS | \
105 NIR_DEBUG_PRINT_AHS | \
106 NIR_DEBUG_PRINT_CHS | \
107 NIR_DEBUG_PRINT_MHS | \
108 NIR_DEBUG_PRINT_IS | \
109 NIR_DEBUG_PRINT_CBS | \
110 NIR_DEBUG_PRINT_KS)
111
112 #define NIR_FALSE 0u
113 #define NIR_TRUE (~0u)
114 #define NIR_MAX_VEC_COMPONENTS 16
115 #define NIR_MAX_MATRIX_COLUMNS 4
116 #define NIR_STREAM_PACKED (1 << 8)
117 typedef uint16_t nir_component_mask_t;
118
119 static inline bool
nir_num_components_valid(unsigned num_components)120 nir_num_components_valid(unsigned num_components)
121 {
122 return (num_components >= 1 &&
123 num_components <= 5) ||
124 num_components == 8 ||
125 num_components == 16;
126 }
127
128 /*
129 * Round up a vector size to a vector size that's valid in NIR. At present, NIR
130 * supports only vec2-5, vec8, and vec16. Attempting to generate other sizes
131 * will fail validation.
132 */
133 static inline unsigned
nir_round_up_components(unsigned n)134 nir_round_up_components(unsigned n)
135 {
136 return (n > 5) ? util_next_power_of_two(n) : n;
137 }
138
139 static inline nir_component_mask_t
nir_component_mask(unsigned num_components)140 nir_component_mask(unsigned num_components)
141 {
142 assert(nir_num_components_valid(num_components));
143 return (1u << num_components) - 1;
144 }
145
146 void
147 nir_process_debug_variable(void);
148
149 bool nir_component_mask_can_reinterpret(nir_component_mask_t mask,
150 unsigned old_bit_size,
151 unsigned new_bit_size);
152 nir_component_mask_t
153 nir_component_mask_reinterpret(nir_component_mask_t mask,
154 unsigned old_bit_size,
155 unsigned new_bit_size);
156
157 /** Defines a cast function
158 *
159 * This macro defines a cast function from in_type to out_type where
160 * out_type is some structure type that contains a field of type out_type.
161 *
162 * Note that you have to be a bit careful as the generated cast function
163 * destroys constness.
164 */
165 #define NIR_DEFINE_CAST(name, in_type, out_type, field, \
166 type_field, type_value) \
167 static inline out_type * \
168 name(const in_type *parent) \
169 { \
170 assert(parent && parent->type_field == type_value); \
171 return exec_node_data(out_type, parent, field); \
172 }
173
174 struct nir_function;
175 struct nir_shader;
176 struct nir_instr;
177 struct nir_builder;
178 struct nir_xfb_info;
179
180 /**
181 * Description of built-in state associated with a uniform
182 *
183 * :c:member:`nir_variable.state_slots`
184 */
185 typedef struct {
186 gl_state_index16 tokens[STATE_LENGTH];
187 } nir_state_slot;
188
189 /* clang-format off */
190 typedef enum {
191 nir_var_system_value = (1 << 0),
192 nir_var_uniform = (1 << 1),
193 nir_var_shader_in = (1 << 2),
194 nir_var_shader_out = (1 << 3),
195 nir_var_image = (1 << 4),
196 /** Incoming call or ray payload data for ray-tracing shaders */
197 nir_var_shader_call_data = (1 << 5),
198 /** Ray hit attributes */
199 nir_var_ray_hit_attrib = (1 << 6),
200
201 /* Modes named nir_var_mem_* have explicit data layout */
202 nir_var_mem_ubo = (1 << 7),
203 nir_var_mem_push_const = (1 << 8),
204 nir_var_mem_ssbo = (1 << 9),
205 nir_var_mem_constant = (1 << 10),
206 nir_var_mem_task_payload = (1 << 11),
207 nir_var_mem_node_payload = (1 << 12),
208 nir_var_mem_node_payload_in = (1 << 13),
209
210 nir_var_function_in = (1 << 14),
211 nir_var_function_out = (1 << 15),
212 nir_var_function_inout = (1 << 16),
213
214 /* Generic modes intentionally come last. See encode_dref_modes() in
215 * nir_serialize.c for more details.
216 */
217 nir_var_shader_temp = (1 << 17),
218 nir_var_function_temp = (1 << 18),
219 nir_var_mem_shared = (1 << 19),
220 nir_var_mem_global = (1 << 20),
221
222 nir_var_mem_generic = (nir_var_shader_temp |
223 nir_var_function_temp |
224 nir_var_mem_shared |
225 nir_var_mem_global),
226
227 nir_var_read_only_modes = nir_var_shader_in | nir_var_uniform |
228 nir_var_system_value | nir_var_mem_constant |
229 nir_var_mem_ubo,
230 /* Modes where vector derefs can be indexed as arrays. nir_var_shader_out
231 * is only for mesh stages. nir_var_system_value is only for kernel stages.
232 */
233 nir_var_vec_indexable_modes = nir_var_shader_temp | nir_var_function_temp |
234 nir_var_mem_ubo | nir_var_mem_ssbo |
235 nir_var_mem_shared | nir_var_mem_global |
236 nir_var_mem_push_const | nir_var_mem_task_payload |
237 nir_var_shader_out | nir_var_system_value,
238 nir_num_variable_modes = 21,
239 nir_var_all = (1 << nir_num_variable_modes) - 1,
240 } nir_variable_mode;
241 MESA_DEFINE_CPP_ENUM_BITFIELD_OPERATORS(nir_variable_mode)
242 /* clang-format on */
243
244 /**
245 * Rounding modes.
246 */
247 typedef enum {
248 nir_rounding_mode_undef = 0,
249 nir_rounding_mode_rtne = 1, /* round to nearest even */
250 nir_rounding_mode_ru = 2, /* round up */
251 nir_rounding_mode_rd = 3, /* round down */
252 nir_rounding_mode_rtz = 4, /* round towards zero */
253 } nir_rounding_mode;
254
255 /**
256 * Ray query values that can read from a RayQueryKHR object.
257 */
258 typedef enum {
259 nir_ray_query_value_intersection_type,
260 nir_ray_query_value_intersection_t,
261 nir_ray_query_value_intersection_instance_custom_index,
262 nir_ray_query_value_intersection_instance_id,
263 nir_ray_query_value_intersection_instance_sbt_index,
264 nir_ray_query_value_intersection_geometry_index,
265 nir_ray_query_value_intersection_primitive_index,
266 nir_ray_query_value_intersection_barycentrics,
267 nir_ray_query_value_intersection_front_face,
268 nir_ray_query_value_intersection_object_ray_direction,
269 nir_ray_query_value_intersection_object_ray_origin,
270 nir_ray_query_value_intersection_object_to_world,
271 nir_ray_query_value_intersection_world_to_object,
272 nir_ray_query_value_intersection_candidate_aabb_opaque,
273 nir_ray_query_value_tmin,
274 nir_ray_query_value_flags,
275 nir_ray_query_value_world_ray_direction,
276 nir_ray_query_value_world_ray_origin,
277 nir_ray_query_value_intersection_triangle_vertex_positions
278 } nir_ray_query_value;
279
280 /**
281 * Intel resource flags
282 */
283 typedef enum {
284 nir_resource_intel_bindless = 1u << 0,
285 nir_resource_intel_pushable = 1u << 1,
286 nir_resource_intel_sampler = 1u << 2,
287 nir_resource_intel_non_uniform = 1u << 3,
288 nir_resource_intel_sampler_embedded = 1u << 4,
289 } nir_resource_data_intel;
290
291 /**
292 * Which components to interpret as signed in cmat_muladd.
293 * See 'Cooperative Matrix Operands' in SPV_KHR_cooperative_matrix.
294 */
295 typedef enum {
296 NIR_CMAT_A_SIGNED = 1u << 0,
297 NIR_CMAT_B_SIGNED = 1u << 1,
298 NIR_CMAT_C_SIGNED = 1u << 2,
299 NIR_CMAT_RESULT_SIGNED = 1u << 3,
300 } nir_cmat_signed;
301
302 typedef union {
303 bool b;
304 float f32;
305 double f64;
306 int8_t i8;
307 uint8_t u8;
308 int16_t i16;
309 uint16_t u16;
310 int32_t i32;
311 uint32_t u32;
312 int64_t i64;
313 uint64_t u64;
314 } nir_const_value;
315
316 #define nir_const_value_to_array(arr, c, components, m) \
317 do { \
318 for (unsigned i = 0; i < components; ++i) \
319 arr[i] = c[i].m; \
320 } while (false)
321
322 static inline nir_const_value
nir_const_value_for_raw_uint(uint64_t x,unsigned bit_size)323 nir_const_value_for_raw_uint(uint64_t x, unsigned bit_size)
324 {
325 nir_const_value v;
326 memset(&v, 0, sizeof(v));
327
328 /* clang-format off */
329 switch (bit_size) {
330 case 1: v.b = x; break;
331 case 8: v.u8 = (uint8_t)x; break;
332 case 16: v.u16 = (uint16_t)x; break;
333 case 32: v.u32 = (uint32_t)x; break;
334 case 64: v.u64 = x; break;
335 default:
336 unreachable("Invalid bit size");
337 }
338 /* clang-format on */
339
340 return v;
341 }
342
343 static inline nir_const_value
nir_const_value_for_int(int64_t i,unsigned bit_size)344 nir_const_value_for_int(int64_t i, unsigned bit_size)
345 {
346 assert(bit_size <= 64);
347 if (bit_size < 64) {
348 assert(i >= (-(1ll << (bit_size - 1))));
349 assert(i < (1ll << (bit_size - 1)));
350 }
351
352 return nir_const_value_for_raw_uint(i, bit_size);
353 }
354
355 static inline nir_const_value
nir_const_value_for_uint(uint64_t u,unsigned bit_size)356 nir_const_value_for_uint(uint64_t u, unsigned bit_size)
357 {
358 assert(bit_size <= 64);
359 if (bit_size < 64)
360 assert(u < (1ull << bit_size));
361
362 return nir_const_value_for_raw_uint(u, bit_size);
363 }
364
365 static inline nir_const_value
nir_const_value_for_bool(bool b,unsigned bit_size)366 nir_const_value_for_bool(bool b, unsigned bit_size)
367 {
368 /* Booleans use a 0/-1 convention */
369 return nir_const_value_for_int(-(int)b, bit_size);
370 }
371
372 /* This one isn't inline because it requires half-float conversion */
373 nir_const_value nir_const_value_for_float(double b, unsigned bit_size);
374
375 static inline int64_t
nir_const_value_as_int(nir_const_value value,unsigned bit_size)376 nir_const_value_as_int(nir_const_value value, unsigned bit_size)
377 {
378 /* clang-format off */
379 switch (bit_size) {
380 /* int1_t uses 0/-1 convention */
381 case 1: return -(int)value.b;
382 case 8: return value.i8;
383 case 16: return value.i16;
384 case 32: return value.i32;
385 case 64: return value.i64;
386 default:
387 unreachable("Invalid bit size");
388 }
389 /* clang-format on */
390 }
391
392 static inline uint64_t
nir_const_value_as_uint(nir_const_value value,unsigned bit_size)393 nir_const_value_as_uint(nir_const_value value, unsigned bit_size)
394 {
395 /* clang-format off */
396 switch (bit_size) {
397 case 1: return value.b;
398 case 8: return value.u8;
399 case 16: return value.u16;
400 case 32: return value.u32;
401 case 64: return value.u64;
402 default:
403 unreachable("Invalid bit size");
404 }
405 /* clang-format on */
406 }
407
408 static inline bool
nir_const_value_as_bool(nir_const_value value,unsigned bit_size)409 nir_const_value_as_bool(nir_const_value value, unsigned bit_size)
410 {
411 int64_t i = nir_const_value_as_int(value, bit_size);
412
413 /* Booleans of any size use 0/-1 convention */
414 assert(i == 0 || i == -1);
415
416 return i;
417 }
418
419 /* This one isn't inline because it requires half-float conversion */
420 double nir_const_value_as_float(nir_const_value value, unsigned bit_size);
421
422 typedef struct nir_constant {
423 /**
424 * Value of the constant.
425 *
426 * The field used to back the values supplied by the constant is determined
427 * by the type associated with the ``nir_variable``. Constants may be
428 * scalars, vectors, or matrices.
429 */
430 nir_const_value values[NIR_MAX_VEC_COMPONENTS];
431
432 /* Indicates all the values are 0s which can enable some optimizations */
433 bool is_null_constant;
434
435 /* we could get this from the var->type but makes clone *much* easier to
436 * not have to care about the type.
437 */
438 unsigned num_elements;
439
440 /* Array elements / Structure Fields */
441 struct nir_constant **elements;
442 } nir_constant;
443
444 /**
445 * Layout qualifiers for gl_FragDepth.
446 *
447 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
448 * with a layout qualifier.
449 */
450 typedef enum {
451 /** No depth layout is specified. */
452 nir_depth_layout_none,
453 nir_depth_layout_any,
454 nir_depth_layout_greater,
455 nir_depth_layout_less,
456 nir_depth_layout_unchanged
457 } nir_depth_layout;
458
459 /**
460 * Enum keeping track of how a variable was declared.
461 */
462 typedef enum {
463 /**
464 * Normal declaration.
465 */
466 nir_var_declared_normally = 0,
467
468 /**
469 * Variable is an implicitly declared built-in that has not been explicitly
470 * re-declared by the shader.
471 */
472 nir_var_declared_implicitly,
473
474 /**
475 * Variable is implicitly generated by the compiler and should not be
476 * visible via the API.
477 */
478 nir_var_hidden,
479 } nir_var_declaration_type;
480
481 /**
482 * Either a uniform, global variable, shader input, or shader output. Based on
483 * ir_variable - it should be easy to translate between the two.
484 */
485
486 typedef struct nir_variable {
487 struct exec_node node;
488
489 /**
490 * Declared type of the variable
491 */
492 const struct glsl_type *type;
493
494 /**
495 * Declared name of the variable
496 */
497 char *name;
498
499 struct nir_variable_data {
500 /**
501 * Storage class of the variable.
502 *
503 * :c:struct:`nir_variable_mode`
504 */
505 unsigned mode : 21;
506
507 /**
508 * Is the variable read-only?
509 *
510 * This is set for variables declared as ``const``, shader inputs,
511 * and uniforms.
512 */
513 unsigned read_only : 1;
514 unsigned centroid : 1;
515 unsigned sample : 1;
516 unsigned patch : 1;
517 unsigned invariant : 1;
518
519 /**
520 * Was an 'invariant' qualifier explicitly set in the shader?
521 *
522 * This is used to cross validate glsl qualifiers.
523 */
524 unsigned explicit_invariant:1;
525
526 /**
527 * Is the variable a ray query?
528 */
529 unsigned ray_query : 1;
530
531 /**
532 * Precision qualifier.
533 *
534 * In desktop GLSL we do not care about precision qualifiers at all, in
535 * fact, the spec says that precision qualifiers are ignored.
536 *
537 * To make things easy, we make it so that this field is always
538 * GLSL_PRECISION_NONE on desktop shaders. This way all the variables
539 * have the same precision value and the checks we add in the compiler
540 * for this field will never break a desktop shader compile.
541 */
542 unsigned precision : 2;
543
544 /**
545 * Has this variable been statically assigned?
546 *
547 * This answers whether the variable was assigned in any path of
548 * the shader during ast_to_hir. This doesn't answer whether it is
549 * still written after dead code removal, nor is it maintained in
550 * non-ast_to_hir.cpp (GLSL parsing) paths.
551 */
552 unsigned assigned : 1;
553
554 /**
555 * Can this variable be coalesced with another?
556 *
557 * This is set by nir_lower_io_to_temporaries to say that any
558 * copies involving this variable should stay put. Propagating it can
559 * duplicate the resulting load/store, which is not wanted, and may
560 * result in a load/store of the variable with an indirect offset which
561 * the backend may not be able to handle.
562 */
563 unsigned cannot_coalesce : 1;
564
565 /**
566 * When separate shader programs are enabled, only input/outputs between
567 * the stages of a multi-stage separate program can be safely removed
568 * from the shader interface. Other input/outputs must remains active.
569 *
570 * This is also used to make sure xfb varyings that are unused by the
571 * fragment shader are not removed.
572 */
573 unsigned always_active_io : 1;
574
575 /**
576 * Interpolation mode for shader inputs / outputs
577 *
578 * :c:enum:`glsl_interp_mode`
579 */
580 unsigned interpolation : 3;
581
582 /**
583 * If non-zero, then this variable may be packed along with other variables
584 * into a single varying slot, so this offset should be applied when
585 * accessing components. For example, an offset of 1 means that the x
586 * component of this variable is actually stored in component y of the
587 * location specified by ``location``.
588 */
589 unsigned location_frac : 2;
590
591 /**
592 * If true, this variable represents an array of scalars that should
593 * be tightly packed. In other words, consecutive array elements
594 * should be stored one component apart, rather than one slot apart.
595 */
596 unsigned compact : 1;
597
598 /**
599 * Whether this is a fragment shader output implicitly initialized with
600 * the previous contents of the specified render target at the
601 * framebuffer location corresponding to this shader invocation.
602 */
603 unsigned fb_fetch_output : 1;
604
605 /**
606 * Non-zero if this variable is considered bindless as defined by
607 * ARB_bindless_texture.
608 */
609 unsigned bindless : 1;
610
611 /**
612 * Was an explicit binding set in the shader?
613 */
614 unsigned explicit_binding : 1;
615
616 /**
617 * Was the location explicitly set in the shader?
618 *
619 * If the location is explicitly set in the shader, it **cannot** be changed
620 * by the linker or by the API (e.g., calls to ``glBindAttribLocation`` have
621 * no effect).
622 */
623 unsigned explicit_location : 1;
624
625 /* Was the array implicitly sized during linking */
626 unsigned implicit_sized_array : 1;
627
628 /**
629 * Highest element accessed with a constant array index
630 *
631 * Not used for non-array variables. -1 is never accessed.
632 */
633 int max_array_access;
634
635 /**
636 * Does this variable have an initializer?
637 *
638 * This is used by the linker to cross-validiate initializers of global
639 * variables.
640 */
641 unsigned has_initializer:1;
642
643 /**
644 * Is the initializer created by the compiler (glsl_zero_init)
645 */
646 unsigned is_implicit_initializer:1;
647
648 /**
649 * Is this varying used by transform feedback?
650 *
651 * This is used by the linker to decide if it's safe to pack the varying.
652 */
653 unsigned is_xfb : 1;
654
655 /**
656 * Is this varying used only by transform feedback?
657 *
658 * This is used by the linker to decide if its safe to pack the varying.
659 */
660 unsigned is_xfb_only : 1;
661
662 /**
663 * Was a transfer feedback buffer set in the shader?
664 */
665 unsigned explicit_xfb_buffer : 1;
666
667 /**
668 * Was a transfer feedback stride set in the shader?
669 */
670 unsigned explicit_xfb_stride : 1;
671
672 /**
673 * Was an explicit offset set in the shader?
674 */
675 unsigned explicit_offset : 1;
676
677 /**
678 * Layout of the matrix. Uses glsl_matrix_layout values.
679 */
680 unsigned matrix_layout : 2;
681
682 /**
683 * Non-zero if this variable was created by lowering a named interface
684 * block.
685 */
686 unsigned from_named_ifc_block : 1;
687
688 /**
689 * Unsized array buffer variable.
690 */
691 unsigned from_ssbo_unsized_array : 1;
692
693 /**
694 * Non-zero if the variable must be a shader input. This is useful for
695 * constraints on function parameters.
696 */
697 unsigned must_be_shader_input : 1;
698
699 /**
700 * Has this variable been used for reading or writing?
701 *
702 * Several GLSL semantic checks require knowledge of whether or not a
703 * variable has been used. For example, it is an error to redeclare a
704 * variable as invariant after it has been used.
705 */
706 unsigned used:1;
707
708 /**
709 * How the variable was declared. See nir_var_declaration_type.
710 *
711 * This is used to detect variables generated by the compiler, so should
712 * not be visible via the API.
713 */
714 unsigned how_declared : 2;
715
716 /**
717 * Is this variable per-view? If so, we know it must be an array with
718 * size corresponding to the number of views.
719 */
720 unsigned per_view : 1;
721
722 /**
723 * Whether the variable is per-primitive.
724 * Can be use by Mesh Shader outputs and corresponding Fragment Shader inputs.
725 */
726 unsigned per_primitive : 1;
727
728 /**
729 * Whether the variable is declared to indicate that a fragment shader
730 * input will not have interpolated values.
731 */
732 unsigned per_vertex : 1;
733
734 /**
735 * Layout qualifier for gl_FragDepth. See nir_depth_layout.
736 *
737 * This is not equal to ``ir_depth_layout_none`` if and only if this
738 * variable is ``gl_FragDepth`` and a layout qualifier is specified.
739 */
740 unsigned depth_layout : 3;
741
742 /**
743 * Vertex stream output identifier.
744 *
745 * For packed outputs, NIR_STREAM_PACKED is set and bits [2*i+1,2*i]
746 * indicate the stream of the i-th component.
747 */
748 unsigned stream : 9;
749
750 /**
751 * See gl_access_qualifier.
752 *
753 * Access flags for memory variables (SSBO/global), image uniforms, and
754 * bindless images in uniforms/inputs/outputs.
755 */
756 unsigned access : 9;
757
758 /**
759 * Descriptor set binding for sampler or UBO.
760 */
761 unsigned descriptor_set : 5;
762
763 #define NIR_VARIABLE_NO_INDEX ~0
764
765 /**
766 * Output index for dual source blending or input attachment index. If
767 * it is not declared it is NIR_VARIABLE_NO_INDEX.
768 */
769 unsigned index;
770
771 /**
772 * Initial binding point for a sampler or UBO.
773 *
774 * For array types, this represents the binding point for the first element.
775 */
776 unsigned binding;
777
778 /**
779 * Storage location of the base of this variable
780 *
781 * The precise meaning of this field depends on the nature of the variable.
782 *
783 * - Vertex shader input: one of the values from ``gl_vert_attrib``.
784 * - Vertex shader output: one of the values from ``gl_varying_slot``.
785 * - Geometry shader input: one of the values from ``gl_varying_slot``.
786 * - Geometry shader output: one of the values from ``gl_varying_slot``.
787 * - Fragment shader input: one of the values from ``gl_varying_slot``.
788 * - Fragment shader output: one of the values from ``gl_frag_result``.
789 * - Task shader output: one of the values from ``gl_varying_slot``.
790 * - Mesh shader input: one of the values from ``gl_varying_slot``.
791 * - Mesh shader output: one of the values from ``gl_varying_slot``.
792 * - Uniforms: Per-stage uniform slot number for default uniform block.
793 * - Uniforms: Index within the uniform block definition for UBO members.
794 * - Non-UBO Uniforms: uniform slot number.
795 * - Other: This field is not currently used.
796 *
797 * If the variable is a uniform, shader input, or shader output, and the
798 * slot has not been assigned, the value will be -1.
799 */
800 int location;
801
802 /** Required alignment of this variable */
803 unsigned alignment;
804
805 /**
806 * The actual location of the variable in the IR. Only valid for inputs,
807 * outputs, uniforms (including samplers and images), and for UBO and SSBO
808 * variables in GLSL.
809 */
810 unsigned driver_location;
811
812 /**
813 * Location an atomic counter or transform feedback is stored at.
814 */
815 unsigned offset;
816
817 union {
818 struct {
819 /** Image internal format if specified explicitly, otherwise PIPE_FORMAT_NONE. */
820 enum pipe_format format;
821 } image;
822
823 struct {
824 /**
825 * For OpenCL inline samplers. See cl_sampler_addressing_mode and cl_sampler_filter_mode
826 */
827 unsigned is_inline_sampler : 1;
828 unsigned addressing_mode : 3;
829 unsigned normalized_coordinates : 1;
830 unsigned filter_mode : 1;
831 } sampler;
832
833 struct {
834 /**
835 * Transform feedback buffer.
836 */
837 uint16_t buffer : 2;
838
839 /**
840 * Transform feedback stride.
841 */
842 uint16_t stride;
843 } xfb;
844 };
845
846 /** Name of the node this payload will be enqueued to. */
847 const char *node_name;
848 } data;
849
850 /**
851 * Identifier for this variable generated by nir_index_vars() that is unique
852 * among other variables in the same exec_list.
853 */
854 unsigned index;
855
856 /* Number of nir_variable_data members */
857 uint16_t num_members;
858
859 /**
860 * For variables with non NULL interface_type, this points to an array of
861 * integers such that if the ith member of the interface block is an array,
862 * max_ifc_array_access[i] is the maximum array element of that member that
863 * has been accessed. If the ith member of the interface block is not an
864 * array, max_ifc_array_access[i] is unused.
865 *
866 * For variables whose type is not an interface block, this pointer is
867 * NULL.
868 */
869 int *max_ifc_array_access;
870
871 /**
872 * Built-in state that backs this uniform
873 *
874 * Once set at variable creation, ``state_slots`` must remain invariant.
875 * This is because, ideally, this array would be shared by all clones of
876 * this variable in the IR tree. In other words, we'd really like for it
877 * to be a fly-weight.
878 *
879 * If the variable is not a uniform, ``num_state_slots`` will be zero and
880 * ``state_slots`` will be ``NULL``.
881 *
882 * Number of state slots used.
883 */
884 uint16_t num_state_slots;
885 /** State descriptors. */
886 nir_state_slot *state_slots;
887
888 /**
889 * Constant expression assigned in the initializer of the variable
890 *
891 * This field should only be used temporarily by creators of NIR shaders
892 * and then nir_lower_variable_initializers can be used to get rid of them.
893 * Most of the rest of NIR ignores this field or asserts that it's NULL.
894 */
895 nir_constant *constant_initializer;
896
897 /**
898 * Global variable assigned in the initializer of the variable
899 * This field should only be used temporarily by creators of NIR shaders
900 * and then nir_lower_variable_initializers can be used to get rid of them.
901 * Most of the rest of NIR ignores this field or asserts that it's NULL.
902 */
903 struct nir_variable *pointer_initializer;
904
905 /**
906 * For variables that are in an interface block or are an instance of an
907 * interface block, this is the ``GLSL_TYPE_INTERFACE`` type for that block.
908 *
909 * ``ir_variable.location``
910 */
911 const struct glsl_type *interface_type;
912
913 /**
914 * Description of per-member data for per-member struct variables
915 *
916 * This is used for variables which are actually an amalgamation of
917 * multiple entities such as a struct of built-in values or a struct of
918 * inputs each with their own layout specifier. This is only allowed on
919 * variables with a struct or array of array of struct type.
920 */
921 struct nir_variable_data *members;
922 } nir_variable;
923
924 static inline bool
_nir_shader_variable_has_mode(nir_variable * var,unsigned modes)925 _nir_shader_variable_has_mode(nir_variable *var, unsigned modes)
926 {
927 /* This isn't a shader variable */
928 assert(!(modes & nir_var_function_temp));
929 return var->data.mode & modes;
930 }
931
932 #define nir_foreach_variable_in_list(var, var_list) \
933 foreach_list_typed(nir_variable, var, node, var_list)
934
935 #define nir_foreach_variable_in_list_safe(var, var_list) \
936 foreach_list_typed_safe(nir_variable, var, node, var_list)
937
938 #define nir_foreach_variable_in_shader(var, shader) \
939 nir_foreach_variable_in_list(var, &(shader)->variables)
940
941 #define nir_foreach_variable_in_shader_safe(var, shader) \
942 nir_foreach_variable_in_list_safe(var, &(shader)->variables)
943
944 #define nir_foreach_variable_with_modes(var, shader, modes) \
945 nir_foreach_variable_in_shader(var, shader) \
946 if (_nir_shader_variable_has_mode(var, modes))
947
948 #define nir_foreach_variable_with_modes_safe(var, shader, modes) \
949 nir_foreach_variable_in_shader_safe(var, shader) \
950 if (_nir_shader_variable_has_mode(var, modes))
951
952 #define nir_foreach_shader_in_variable(var, shader) \
953 nir_foreach_variable_with_modes(var, shader, nir_var_shader_in)
954
955 #define nir_foreach_shader_in_variable_safe(var, shader) \
956 nir_foreach_variable_with_modes_safe(var, shader, nir_var_shader_in)
957
958 #define nir_foreach_shader_out_variable(var, shader) \
959 nir_foreach_variable_with_modes(var, shader, nir_var_shader_out)
960
961 #define nir_foreach_shader_out_variable_safe(var, shader) \
962 nir_foreach_variable_with_modes_safe(var, shader, nir_var_shader_out)
963
964 #define nir_foreach_uniform_variable(var, shader) \
965 nir_foreach_variable_with_modes(var, shader, nir_var_uniform)
966
967 #define nir_foreach_uniform_variable_safe(var, shader) \
968 nir_foreach_variable_with_modes_safe(var, shader, nir_var_uniform)
969
970 #define nir_foreach_image_variable(var, shader) \
971 nir_foreach_variable_with_modes(var, shader, nir_var_image)
972
973 #define nir_foreach_image_variable_safe(var, shader) \
974 nir_foreach_variable_with_modes_safe(var, shader, nir_var_image)
975
976 static inline bool
nir_variable_is_global(const nir_variable * var)977 nir_variable_is_global(const nir_variable *var)
978 {
979 return var->data.mode != nir_var_function_temp;
980 }
981
982 typedef enum ENUM_PACKED {
983 nir_instr_type_alu,
984 nir_instr_type_deref,
985 nir_instr_type_call,
986 nir_instr_type_tex,
987 nir_instr_type_intrinsic,
988 nir_instr_type_load_const,
989 nir_instr_type_jump,
990 nir_instr_type_undef,
991 nir_instr_type_phi,
992 nir_instr_type_parallel_copy,
993 nir_instr_type_debug_info,
994 } nir_instr_type;
995
996 typedef struct nir_instr {
997 struct exec_node node;
998 struct nir_block *block;
999 nir_instr_type type;
1000
1001 /* A temporary for optimization and analysis passes to use for storing
1002 * flags. For instance, DCE uses this to store the "dead/live" info.
1003 */
1004 uint8_t pass_flags;
1005
1006 /** generic instruction index. */
1007 uint32_t index;
1008 } nir_instr;
1009
1010 static inline nir_instr *
nir_instr_next(nir_instr * instr)1011 nir_instr_next(nir_instr *instr)
1012 {
1013 struct exec_node *next = exec_node_get_next(&instr->node);
1014 if (exec_node_is_tail_sentinel(next))
1015 return NULL;
1016 else
1017 return exec_node_data(nir_instr, next, node);
1018 }
1019
1020 static inline nir_instr *
nir_instr_prev(nir_instr * instr)1021 nir_instr_prev(nir_instr *instr)
1022 {
1023 struct exec_node *prev = exec_node_get_prev(&instr->node);
1024 if (exec_node_is_head_sentinel(prev))
1025 return NULL;
1026 else
1027 return exec_node_data(nir_instr, prev, node);
1028 }
1029
1030 static inline bool
nir_instr_is_first(const nir_instr * instr)1031 nir_instr_is_first(const nir_instr *instr)
1032 {
1033 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr->node));
1034 }
1035
1036 static inline bool
nir_instr_is_last(const nir_instr * instr)1037 nir_instr_is_last(const nir_instr *instr)
1038 {
1039 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr->node));
1040 }
1041
1042 typedef struct nir_def {
1043 /** Instruction which produces this SSA value. */
1044 nir_instr *parent_instr;
1045
1046 /** set of nir_instrs where this register is used (read from) */
1047 struct list_head uses;
1048
1049 /** generic SSA definition index. */
1050 unsigned index;
1051
1052 uint8_t num_components;
1053
1054 /* The bit-size of each channel; must be one of 1, 8, 16, 32, or 64 */
1055 uint8_t bit_size;
1056
1057 /**
1058 * True if this SSA value may have different values in different SIMD
1059 * invocations of the shader. This is set by nir_divergence_analysis.
1060 */
1061 bool divergent;
1062
1063 /**
1064 * True if this SSA value is loop invariant w.r.t. the innermost parent
1065 * loop. This is set by nir_divergence_analysis and used to determine
1066 * the divergence of a nir_src.
1067 */
1068 bool loop_invariant;
1069 } nir_def;
1070
1071 struct nir_src;
1072 struct nir_if;
1073
1074 typedef struct nir_src {
1075 /* Instruction or if-statement that consumes this value as a source. This
1076 * should only be accessed through nir_src_* helpers.
1077 *
1078 * Internally, it is a tagged pointer to a nir_instr or nir_if.
1079 */
1080 uintptr_t _parent;
1081
1082 struct list_head use_link;
1083 nir_def *ssa;
1084 } nir_src;
1085
1086 /* Layout of the _parent pointer. Bottom bit is set for nir_if parents (clear
1087 * for nir_instr parents). Remaining bits are the pointer.
1088 */
1089 #define NIR_SRC_PARENT_IS_IF (0x1)
1090 #define NIR_SRC_PARENT_MASK (~((uintptr_t) NIR_SRC_PARENT_IS_IF))
1091
1092 static inline bool
nir_src_is_if(const nir_src * src)1093 nir_src_is_if(const nir_src *src)
1094 {
1095 return src->_parent & NIR_SRC_PARENT_IS_IF;
1096 }
1097
1098 static inline nir_instr *
nir_src_parent_instr(const nir_src * src)1099 nir_src_parent_instr(const nir_src *src)
1100 {
1101 assert(!nir_src_is_if(src));
1102
1103 /* Because it is not an if, the tag is 0, therefore we do not need to mask */
1104 return (nir_instr *)(src->_parent);
1105 }
1106
1107 static inline struct nir_if *
nir_src_parent_if(const nir_src * src)1108 nir_src_parent_if(const nir_src *src)
1109 {
1110 assert(nir_src_is_if(src));
1111
1112 /* Because it is an if, the tag is 1, so we need to mask */
1113 return (struct nir_if *)(src->_parent & NIR_SRC_PARENT_MASK);
1114 }
1115
1116 static inline void
_nir_src_set_parent(nir_src * src,void * parent,bool is_if)1117 _nir_src_set_parent(nir_src *src, void *parent, bool is_if)
1118 {
1119 uintptr_t ptr = (uintptr_t) parent;
1120 assert((ptr & ~NIR_SRC_PARENT_MASK) == 0 && "pointer must be aligned");
1121
1122 if (is_if)
1123 ptr |= NIR_SRC_PARENT_IS_IF;
1124
1125 src->_parent = ptr;
1126 }
1127
1128 static inline void
nir_src_set_parent_instr(nir_src * src,nir_instr * parent_instr)1129 nir_src_set_parent_instr(nir_src *src, nir_instr *parent_instr)
1130 {
1131 _nir_src_set_parent(src, parent_instr, false);
1132 }
1133
1134 static inline void
nir_src_set_parent_if(nir_src * src,struct nir_if * parent_if)1135 nir_src_set_parent_if(nir_src *src, struct nir_if *parent_if)
1136 {
1137 _nir_src_set_parent(src, parent_if, true);
1138 }
1139
1140 static inline nir_src
nir_src_init(void)1141 nir_src_init(void)
1142 {
1143 nir_src src = { 0 };
1144 return src;
1145 }
1146
1147 #define NIR_SRC_INIT nir_src_init()
1148
1149 #define nir_foreach_use_including_if(src, reg_or_ssa_def) \
1150 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
1151
1152 #define nir_foreach_use_including_if_safe(src, reg_or_ssa_def) \
1153 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
1154
1155 #define nir_foreach_use(src, reg_or_ssa_def) \
1156 nir_foreach_use_including_if(src, reg_or_ssa_def) \
1157 if (!nir_src_is_if(src))
1158
1159 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
1160 nir_foreach_use_including_if_safe(src, reg_or_ssa_def) \
1161 if (!nir_src_is_if(src))
1162
1163 #define nir_foreach_if_use(src, reg_or_ssa_def) \
1164 nir_foreach_use_including_if(src, reg_or_ssa_def) \
1165 if (nir_src_is_if(src))
1166
1167 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
1168 nir_foreach_use_including_if_safe(src, reg_or_ssa_def) \
1169 if (nir_src_is_if(src))
1170
1171 static inline bool
nir_def_used_by_if(const nir_def * def)1172 nir_def_used_by_if(const nir_def *def)
1173 {
1174 nir_foreach_if_use(_, def)
1175 return true;
1176
1177 return false;
1178 }
1179
1180 static inline bool
nir_def_only_used_by_if(const nir_def * def)1181 nir_def_only_used_by_if(const nir_def *def)
1182 {
1183 nir_foreach_use(_, def)
1184 return false;
1185
1186 return true;
1187 }
1188
1189 static inline nir_src
nir_src_for_ssa(nir_def * def)1190 nir_src_for_ssa(nir_def *def)
1191 {
1192 nir_src src = NIR_SRC_INIT;
1193
1194 src.ssa = def;
1195
1196 return src;
1197 }
1198
1199 static inline unsigned
nir_src_bit_size(nir_src src)1200 nir_src_bit_size(nir_src src)
1201 {
1202 return src.ssa->bit_size;
1203 }
1204
1205 static inline unsigned
nir_src_num_components(nir_src src)1206 nir_src_num_components(nir_src src)
1207 {
1208 return src.ssa->num_components;
1209 }
1210
1211 static inline bool
nir_src_is_const(nir_src src)1212 nir_src_is_const(nir_src src)
1213 {
1214 return src.ssa->parent_instr->type == nir_instr_type_load_const;
1215 }
1216
1217 static inline bool
nir_src_is_undef(nir_src src)1218 nir_src_is_undef(nir_src src)
1219 {
1220 return src.ssa->parent_instr->type == nir_instr_type_undef;
1221 }
1222
1223 bool nir_src_is_divergent(nir_src *src);
1224
1225 /* Are all components the same, ie. .xxxx */
1226 static inline bool
nir_is_same_comp_swizzle(uint8_t * swiz,unsigned nr_comp)1227 nir_is_same_comp_swizzle(uint8_t *swiz, unsigned nr_comp)
1228 {
1229 for (unsigned i = 1; i < nr_comp; i++)
1230 if (swiz[i] != swiz[0])
1231 return false;
1232 return true;
1233 }
1234
1235 /* Are all components sequential, ie. .yzw */
1236 static inline bool
nir_is_sequential_comp_swizzle(uint8_t * swiz,unsigned nr_comp)1237 nir_is_sequential_comp_swizzle(uint8_t *swiz, unsigned nr_comp)
1238 {
1239 for (unsigned i = 1; i < nr_comp; i++)
1240 if (swiz[i] != (swiz[0] + i))
1241 return false;
1242 return true;
1243 }
1244
1245 /***/
1246 typedef struct nir_alu_src {
1247 /** Base source */
1248 nir_src src;
1249
1250 /**
1251 * For each input component, says which component of the register it is
1252 * chosen from.
1253 *
1254 * Note that which elements of the swizzle are used and which are ignored
1255 * are based on the write mask for most opcodes - for example, a statement
1256 * like "foo.xzw = bar.zyx" would have a writemask of 1101b and a swizzle
1257 * of {2, 1, x, 0} where x means "don't care."
1258 */
1259 uint8_t swizzle[NIR_MAX_VEC_COMPONENTS];
1260 } nir_alu_src;
1261
1262 nir_alu_type
1263 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type);
1264
1265 static inline nir_alu_type
nir_get_nir_type_for_glsl_type(const struct glsl_type * type)1266 nir_get_nir_type_for_glsl_type(const struct glsl_type *type)
1267 {
1268 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type));
1269 }
1270
1271 enum glsl_base_type
1272 nir_get_glsl_base_type_for_nir_type(nir_alu_type base_type);
1273
1274 nir_op nir_type_conversion_op(nir_alu_type src, nir_alu_type dst,
1275 nir_rounding_mode rnd);
1276
1277 /**
1278 * Atomic intrinsics perform different operations depending on the value of
1279 * their atomic_op constant index. nir_atomic_op defines the operations.
1280 */
1281 typedef enum {
1282 nir_atomic_op_iadd,
1283 nir_atomic_op_imin,
1284 nir_atomic_op_umin,
1285 nir_atomic_op_imax,
1286 nir_atomic_op_umax,
1287 nir_atomic_op_iand,
1288 nir_atomic_op_ior,
1289 nir_atomic_op_ixor,
1290 nir_atomic_op_xchg,
1291 nir_atomic_op_fadd,
1292 nir_atomic_op_fmin,
1293 nir_atomic_op_fmax,
1294 nir_atomic_op_cmpxchg,
1295 nir_atomic_op_fcmpxchg,
1296 nir_atomic_op_inc_wrap,
1297 nir_atomic_op_dec_wrap,
1298 nir_atomic_op_ordered_add_gfx12_amd,
1299 } nir_atomic_op;
1300
1301 static inline nir_alu_type
nir_atomic_op_type(nir_atomic_op op)1302 nir_atomic_op_type(nir_atomic_op op)
1303 {
1304 switch (op) {
1305 case nir_atomic_op_imin:
1306 case nir_atomic_op_imax:
1307 return nir_type_int;
1308
1309 case nir_atomic_op_fadd:
1310 case nir_atomic_op_fmin:
1311 case nir_atomic_op_fmax:
1312 case nir_atomic_op_fcmpxchg:
1313 return nir_type_float;
1314
1315 case nir_atomic_op_iadd:
1316 case nir_atomic_op_iand:
1317 case nir_atomic_op_ior:
1318 case nir_atomic_op_ixor:
1319 case nir_atomic_op_xchg:
1320 case nir_atomic_op_cmpxchg:
1321 case nir_atomic_op_umin:
1322 case nir_atomic_op_umax:
1323 case nir_atomic_op_inc_wrap:
1324 case nir_atomic_op_dec_wrap:
1325 case nir_atomic_op_ordered_add_gfx12_amd:
1326 return nir_type_uint;
1327 }
1328
1329 unreachable("Invalid nir_atomic_op");
1330 }
1331
1332 nir_op
1333 nir_atomic_op_to_alu(nir_atomic_op op);
1334
1335 /** Returns nir_op_vec<num_components> or nir_op_mov if num_components == 1
1336 *
1337 * This is subtly different from nir_op_is_vec() which returns false for
1338 * nir_op_mov. Returning nir_op_mov from nir_op_vec() when num_components == 1
1339 * makes sense under the assumption that the num_components of the resulting
1340 * nir_def will same as what is passed in here because a single-component mov
1341 * is effectively a vec1. However, if alu->def.num_components > 1, nir_op_mov
1342 * has different semantics from nir_op_vec* so so code which detects "is this
1343 * a vec?" typically needs to handle nir_op_mov separate from nir_op_vecN.
1344 *
1345 * In the unlikely case where you can handle nir_op_vecN and nir_op_mov
1346 * together, use nir_op_is_vec_or_mov().
1347 */
1348 nir_op
1349 nir_op_vec(unsigned num_components);
1350
1351 /** Returns true if this op is one of nir_op_vec*
1352 *
1353 * Returns false for nir_op_mov. See nir_op_vec() for more details.
1354 */
1355 bool
1356 nir_op_is_vec(nir_op op);
1357
1358 static inline bool
nir_op_is_vec_or_mov(nir_op op)1359 nir_op_is_vec_or_mov(nir_op op)
1360 {
1361 return op == nir_op_mov || nir_op_is_vec(op);
1362 }
1363
1364 static inline bool
nir_is_float_control_signed_zero_preserve(unsigned execution_mode,unsigned bit_size)1365 nir_is_float_control_signed_zero_preserve(unsigned execution_mode, unsigned bit_size)
1366 {
1367 return (16 == bit_size && execution_mode & FLOAT_CONTROLS_SIGNED_ZERO_PRESERVE_FP16) ||
1368 (32 == bit_size && execution_mode & FLOAT_CONTROLS_SIGNED_ZERO_PRESERVE_FP32) ||
1369 (64 == bit_size && execution_mode & FLOAT_CONTROLS_SIGNED_ZERO_PRESERVE_FP64);
1370 }
1371
1372 static inline bool
nir_is_float_control_inf_preserve(unsigned execution_mode,unsigned bit_size)1373 nir_is_float_control_inf_preserve(unsigned execution_mode, unsigned bit_size)
1374 {
1375 return (16 == bit_size && execution_mode & FLOAT_CONTROLS_INF_PRESERVE_FP16) ||
1376 (32 == bit_size && execution_mode & FLOAT_CONTROLS_INF_PRESERVE_FP32) ||
1377 (64 == bit_size && execution_mode & FLOAT_CONTROLS_INF_PRESERVE_FP64);
1378 }
1379
1380 static inline bool
nir_is_float_control_nan_preserve(unsigned execution_mode,unsigned bit_size)1381 nir_is_float_control_nan_preserve(unsigned execution_mode, unsigned bit_size)
1382 {
1383 return (16 == bit_size && execution_mode & FLOAT_CONTROLS_NAN_PRESERVE_FP16) ||
1384 (32 == bit_size && execution_mode & FLOAT_CONTROLS_NAN_PRESERVE_FP32) ||
1385 (64 == bit_size && execution_mode & FLOAT_CONTROLS_NAN_PRESERVE_FP64);
1386 }
1387
1388 static inline bool
nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode,unsigned bit_size)1389 nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode, unsigned bit_size)
1390 {
1391 return (16 == bit_size && execution_mode & FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16) ||
1392 (32 == bit_size && execution_mode & FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32) ||
1393 (64 == bit_size && execution_mode & FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64);
1394 }
1395
1396 static inline bool
nir_is_denorm_flush_to_zero(unsigned execution_mode,unsigned bit_size)1397 nir_is_denorm_flush_to_zero(unsigned execution_mode, unsigned bit_size)
1398 {
1399 return (16 == bit_size && execution_mode & FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16) ||
1400 (32 == bit_size && execution_mode & FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32) ||
1401 (64 == bit_size && execution_mode & FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64);
1402 }
1403
1404 static inline bool
nir_is_denorm_preserve(unsigned execution_mode,unsigned bit_size)1405 nir_is_denorm_preserve(unsigned execution_mode, unsigned bit_size)
1406 {
1407 return (16 == bit_size && execution_mode & FLOAT_CONTROLS_DENORM_PRESERVE_FP16) ||
1408 (32 == bit_size && execution_mode & FLOAT_CONTROLS_DENORM_PRESERVE_FP32) ||
1409 (64 == bit_size && execution_mode & FLOAT_CONTROLS_DENORM_PRESERVE_FP64);
1410 }
1411
1412 static inline bool
nir_is_rounding_mode_rtne(unsigned execution_mode,unsigned bit_size)1413 nir_is_rounding_mode_rtne(unsigned execution_mode, unsigned bit_size)
1414 {
1415 return (16 == bit_size && execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16) ||
1416 (32 == bit_size && execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32) ||
1417 (64 == bit_size && execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64);
1418 }
1419
1420 static inline bool
nir_is_rounding_mode_rtz(unsigned execution_mode,unsigned bit_size)1421 nir_is_rounding_mode_rtz(unsigned execution_mode, unsigned bit_size)
1422 {
1423 return (16 == bit_size && execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16) ||
1424 (32 == bit_size && execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32) ||
1425 (64 == bit_size && execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64);
1426 }
1427
1428 static inline bool
nir_has_any_rounding_mode_rtz(unsigned execution_mode)1429 nir_has_any_rounding_mode_rtz(unsigned execution_mode)
1430 {
1431 return (execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16) ||
1432 (execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32) ||
1433 (execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64);
1434 }
1435
1436 static inline bool
nir_has_any_rounding_mode_rtne(unsigned execution_mode)1437 nir_has_any_rounding_mode_rtne(unsigned execution_mode)
1438 {
1439 return (execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16) ||
1440 (execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32) ||
1441 (execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64);
1442 }
1443
1444 static inline nir_rounding_mode
nir_get_rounding_mode_from_float_controls(unsigned execution_mode,nir_alu_type type)1445 nir_get_rounding_mode_from_float_controls(unsigned execution_mode,
1446 nir_alu_type type)
1447 {
1448 if (nir_alu_type_get_base_type(type) != nir_type_float)
1449 return nir_rounding_mode_undef;
1450
1451 unsigned bit_size = nir_alu_type_get_type_size(type);
1452
1453 if (nir_is_rounding_mode_rtz(execution_mode, bit_size))
1454 return nir_rounding_mode_rtz;
1455 if (nir_is_rounding_mode_rtne(execution_mode, bit_size))
1456 return nir_rounding_mode_rtne;
1457 return nir_rounding_mode_undef;
1458 }
1459
1460 static inline bool
nir_has_any_rounding_mode_enabled(unsigned execution_mode)1461 nir_has_any_rounding_mode_enabled(unsigned execution_mode)
1462 {
1463 bool result =
1464 nir_has_any_rounding_mode_rtne(execution_mode) ||
1465 nir_has_any_rounding_mode_rtz(execution_mode);
1466 return result;
1467 }
1468
1469 typedef enum {
1470 /**
1471 * Operation where the first two sources are commutative.
1472 *
1473 * For 2-source operations, this just mathematical commutativity. Some
1474 * 3-source operations, like ffma, are only commutative in the first two
1475 * sources.
1476 */
1477 NIR_OP_IS_2SRC_COMMUTATIVE = (1 << 0),
1478
1479 /**
1480 * Operation is associative
1481 */
1482 NIR_OP_IS_ASSOCIATIVE = (1 << 1),
1483
1484 /**
1485 * Operation where src[0] is used to select src[1] on true or src[2] false.
1486 * src[0] may be Boolean, or it may be another type used in an implicit
1487 * comparison.
1488 */
1489 NIR_OP_IS_SELECTION = (1 << 2),
1490 } nir_op_algebraic_property;
1491
1492 /* vec16 is the widest ALU op in NIR, making the max number of input of ALU
1493 * instructions to be the same as NIR_MAX_VEC_COMPONENTS.
1494 */
1495 #define NIR_ALU_MAX_INPUTS NIR_MAX_VEC_COMPONENTS
1496
1497 /***/
1498 typedef struct nir_op_info {
1499 /** Name of the NIR ALU opcode */
1500 const char *name;
1501
1502 /** Number of inputs (sources) */
1503 uint8_t num_inputs;
1504
1505 /**
1506 * The number of components in the output
1507 *
1508 * If non-zero, this is the size of the output and input sizes are
1509 * explicitly given; swizzle and writemask are still in effect, but if
1510 * the output component is masked out, then the input component may
1511 * still be in use.
1512 *
1513 * If zero, the opcode acts in the standard, per-component manner; the
1514 * operation is performed on each component (except the ones that are
1515 * masked out) with the input being taken from the input swizzle for
1516 * that component.
1517 *
1518 * The size of some of the inputs may be given (i.e. non-zero) even
1519 * though output_size is zero; in that case, the inputs with a zero
1520 * size act per-component, while the inputs with non-zero size don't.
1521 */
1522 uint8_t output_size;
1523
1524 /**
1525 * The type of vector that the instruction outputs.
1526 */
1527 nir_alu_type output_type;
1528
1529 /**
1530 * The number of components in each input
1531 *
1532 * See nir_op_infos::output_size for more detail about the relationship
1533 * between input and output sizes.
1534 */
1535 uint8_t input_sizes[NIR_ALU_MAX_INPUTS];
1536
1537 /**
1538 * The type of vector that each input takes.
1539 */
1540 nir_alu_type input_types[NIR_ALU_MAX_INPUTS];
1541
1542 /** Algebraic properties of this opcode */
1543 nir_op_algebraic_property algebraic_properties;
1544
1545 /** Whether this represents a numeric conversion opcode */
1546 bool is_conversion;
1547 } nir_op_info;
1548
1549 /** Metadata for each nir_op, indexed by opcode */
1550 extern const nir_op_info nir_op_infos[nir_num_opcodes];
1551
1552 static inline bool
nir_op_is_selection(nir_op op)1553 nir_op_is_selection(nir_op op)
1554 {
1555 return (nir_op_infos[op].algebraic_properties & NIR_OP_IS_SELECTION) != 0;
1556 }
1557
1558 /***/
1559 typedef struct nir_alu_instr {
1560 /** Base instruction */
1561 nir_instr instr;
1562
1563 /** Opcode */
1564 nir_op op;
1565
1566 /** Indicates that this ALU instruction generates an exact value
1567 *
1568 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1569 * really equivalent to either. This indicates that the value generated by
1570 * this operation is high-precision and any code transformations that touch
1571 * it must ensure that the resulting value is bit-for-bit identical to the
1572 * original.
1573 */
1574 bool exact : 1;
1575
1576 /**
1577 * Indicates that this instruction doese not cause signed integer wrapping
1578 * to occur, in the form of overflow or underflow.
1579 */
1580 bool no_signed_wrap : 1;
1581
1582 /**
1583 * Indicates that this instruction does not cause unsigned integer wrapping
1584 * to occur, in the form of overflow or underflow.
1585 */
1586 bool no_unsigned_wrap : 1;
1587
1588 /**
1589 * The float controls bit float_controls2 cares about. That is,
1590 * NAN/INF/SIGNED_ZERO_PRESERVE only. Allow{Contract,Reassoc,Transform} are
1591 * still handled through the exact bit, and the other float controls bits
1592 * (rounding mode and denorm handling) remain in the execution mode only.
1593 */
1594 uint32_t fp_fast_math : 9;
1595
1596 /** Destination */
1597 nir_def def;
1598
1599 /** Sources
1600 *
1601 * The size of the array is given by :c:member:`nir_op_info.num_inputs`.
1602 */
1603 nir_alu_src src[];
1604 } nir_alu_instr;
1605
1606 static inline bool
nir_alu_instr_is_signed_zero_preserve(nir_alu_instr * alu)1607 nir_alu_instr_is_signed_zero_preserve(nir_alu_instr *alu)
1608 {
1609 return nir_is_float_control_signed_zero_preserve(alu->fp_fast_math, alu->def.bit_size);
1610 }
1611
1612 static inline bool
nir_alu_instr_is_inf_preserve(nir_alu_instr * alu)1613 nir_alu_instr_is_inf_preserve(nir_alu_instr *alu)
1614 {
1615 return nir_is_float_control_inf_preserve(alu->fp_fast_math, alu->def.bit_size);
1616 }
1617
1618 static inline bool
nir_alu_instr_is_nan_preserve(nir_alu_instr * alu)1619 nir_alu_instr_is_nan_preserve(nir_alu_instr *alu)
1620 {
1621 return nir_is_float_control_nan_preserve(alu->fp_fast_math, alu->def.bit_size);
1622 }
1623
1624 static inline bool
nir_alu_instr_is_signed_zero_inf_nan_preserve(nir_alu_instr * alu)1625 nir_alu_instr_is_signed_zero_inf_nan_preserve(nir_alu_instr *alu)
1626 {
1627 return nir_is_float_control_signed_zero_inf_nan_preserve(alu->fp_fast_math, alu->def.bit_size);
1628 }
1629
1630 void nir_alu_src_copy(nir_alu_src *dest, const nir_alu_src *src);
1631
1632 nir_component_mask_t
1633 nir_alu_instr_src_read_mask(const nir_alu_instr *instr, unsigned src);
1634 /**
1635 * Get the number of channels used for a source
1636 */
1637 unsigned
1638 nir_ssa_alu_instr_src_components(const nir_alu_instr *instr, unsigned src);
1639
1640 /* is this source channel used? */
1641 static inline bool
nir_alu_instr_channel_used(const nir_alu_instr * instr,unsigned src,unsigned channel)1642 nir_alu_instr_channel_used(const nir_alu_instr *instr, unsigned src,
1643 unsigned channel)
1644 {
1645 return channel < nir_ssa_alu_instr_src_components(instr, src);
1646 }
1647
1648 bool
1649 nir_alu_instr_is_comparison(const nir_alu_instr *instr);
1650
1651 bool nir_const_value_negative_equal(nir_const_value c1, nir_const_value c2,
1652 nir_alu_type full_type);
1653
1654 bool nir_alu_srcs_equal(const nir_alu_instr *alu1, const nir_alu_instr *alu2,
1655 unsigned src1, unsigned src2);
1656
1657 bool nir_alu_srcs_negative_equal_typed(const nir_alu_instr *alu1,
1658 const nir_alu_instr *alu2,
1659 unsigned src1, unsigned src2,
1660 nir_alu_type base_type);
1661 bool nir_alu_srcs_negative_equal(const nir_alu_instr *alu1,
1662 const nir_alu_instr *alu2,
1663 unsigned src1, unsigned src2);
1664
1665 bool nir_alu_src_is_trivial_ssa(const nir_alu_instr *alu, unsigned srcn);
1666
1667 typedef enum {
1668 nir_deref_type_var,
1669 nir_deref_type_array,
1670 nir_deref_type_array_wildcard,
1671 nir_deref_type_ptr_as_array,
1672 nir_deref_type_struct,
1673 nir_deref_type_cast,
1674 } nir_deref_type;
1675
1676 typedef struct {
1677 nir_instr instr;
1678
1679 /** The type of this deref instruction */
1680 nir_deref_type deref_type;
1681
1682 /** Bitmask what modes the underlying variable might be
1683 *
1684 * For OpenCL-style generic pointers, we may not know exactly what mode it
1685 * is at any given point in time in the compile process. This bitfield
1686 * contains the set of modes which it MAY be.
1687 *
1688 * Generally, this field should not be accessed directly. Use one of the
1689 * nir_deref_mode_ helpers instead.
1690 */
1691 nir_variable_mode modes;
1692
1693 /** The dereferenced type of the resulting pointer value */
1694 const struct glsl_type *type;
1695
1696 union {
1697 /** Variable being dereferenced if deref_type is a deref_var */
1698 nir_variable *var;
1699
1700 /** Parent deref if deref_type is not deref_var */
1701 nir_src parent;
1702 };
1703
1704 /** Additional deref parameters */
1705 union {
1706 struct {
1707 nir_src index;
1708 bool in_bounds;
1709 } arr;
1710
1711 struct {
1712 unsigned index;
1713 } strct;
1714
1715 struct {
1716 unsigned ptr_stride;
1717 unsigned align_mul;
1718 unsigned align_offset;
1719 } cast;
1720 };
1721
1722 /** Destination to store the resulting "pointer" */
1723 nir_def def;
1724 } nir_deref_instr;
1725
1726 /**
1727 * Returns true if the cast is trivial, i.e. the source and destination type is
1728 * the same.
1729 */
1730 bool nir_deref_cast_is_trivial(nir_deref_instr *cast);
1731
1732 /** Returns true if deref might have one of the given modes
1733 *
1734 * For multi-mode derefs, this returns true if any of the possible modes for
1735 * the deref to have any of the specified modes. This function returning true
1736 * does NOT mean that the deref definitely has one of those modes. It simply
1737 * means that, with the best information we have at the time, it might.
1738 */
1739 static inline bool
nir_deref_mode_may_be(const nir_deref_instr * deref,nir_variable_mode modes)1740 nir_deref_mode_may_be(const nir_deref_instr *deref, nir_variable_mode modes)
1741 {
1742 assert(!(modes & ~nir_var_all));
1743 assert(deref->modes != 0);
1744 return deref->modes & modes;
1745 }
1746
1747 /** Returns true if deref must have one of the given modes
1748 *
1749 * For multi-mode derefs, this returns true if NIR can prove that the given
1750 * deref has one of the specified modes. This function returning false does
1751 * NOT mean that deref doesn't have one of the given mode. It very well may
1752 * have one of those modes, we just don't have enough information to prove
1753 * that it does for sure.
1754 */
1755 static inline bool
nir_deref_mode_must_be(const nir_deref_instr * deref,nir_variable_mode modes)1756 nir_deref_mode_must_be(const nir_deref_instr *deref, nir_variable_mode modes)
1757 {
1758 assert(!(modes & ~nir_var_all));
1759 assert(deref->modes != 0);
1760 return !(deref->modes & ~modes);
1761 }
1762
1763 /** Returns true if deref has the given mode
1764 *
1765 * This returns true if the deref has exactly the mode specified. If the
1766 * deref may have that mode but may also have a different mode (i.e. modes has
1767 * multiple bits set), this will assert-fail.
1768 *
1769 * If you're confused about which nir_deref_mode_ helper to use, use this one
1770 * or nir_deref_mode_is_one_of below.
1771 */
1772 static inline bool
nir_deref_mode_is(const nir_deref_instr * deref,nir_variable_mode mode)1773 nir_deref_mode_is(const nir_deref_instr *deref, nir_variable_mode mode)
1774 {
1775 assert(util_bitcount(mode) == 1 && (mode & nir_var_all));
1776 assert(deref->modes != 0);
1777
1778 /* This is only for "simple" cases so, if modes might interact with this
1779 * deref then the deref has to have a single mode.
1780 */
1781 if (nir_deref_mode_may_be(deref, mode)) {
1782 assert(util_bitcount(deref->modes) == 1);
1783 assert(deref->modes == mode);
1784 }
1785
1786 return deref->modes == mode;
1787 }
1788
1789 /** Returns true if deref has one of the given modes
1790 *
1791 * This returns true if the deref has exactly one possible mode and that mode
1792 * is one of the modes specified. If the deref may have one of those modes
1793 * but may also have a different mode (i.e. modes has multiple bits set), this
1794 * will assert-fail.
1795 */
1796 static inline bool
nir_deref_mode_is_one_of(const nir_deref_instr * deref,nir_variable_mode modes)1797 nir_deref_mode_is_one_of(const nir_deref_instr *deref, nir_variable_mode modes)
1798 {
1799 /* This is only for "simple" cases so, if modes might interact with this
1800 * deref then the deref has to have a single mode.
1801 */
1802 if (nir_deref_mode_may_be(deref, modes)) {
1803 assert(util_bitcount(deref->modes) == 1);
1804 assert(nir_deref_mode_must_be(deref, modes));
1805 }
1806
1807 return nir_deref_mode_may_be(deref, modes);
1808 }
1809
1810 /** Returns true if deref's possible modes lie in the given set of modes
1811 *
1812 * This returns true if the deref's modes lie in the given set of modes. If
1813 * the deref's modes overlap with the specified modes but aren't entirely
1814 * contained in the specified set of modes, this will assert-fail. In
1815 * particular, if this is used in a generic pointers scenario, the specified
1816 * modes has to contain all or none of the possible generic pointer modes.
1817 *
1818 * This is intended mostly for mass-lowering of derefs which might have
1819 * generic pointers.
1820 */
1821 static inline bool
nir_deref_mode_is_in_set(const nir_deref_instr * deref,nir_variable_mode modes)1822 nir_deref_mode_is_in_set(const nir_deref_instr *deref, nir_variable_mode modes)
1823 {
1824 if (nir_deref_mode_may_be(deref, modes))
1825 assert(nir_deref_mode_must_be(deref, modes));
1826
1827 return nir_deref_mode_may_be(deref, modes);
1828 }
1829
1830 static inline nir_deref_instr *nir_src_as_deref(nir_src src);
1831
1832 static inline nir_deref_instr *
nir_deref_instr_parent(const nir_deref_instr * instr)1833 nir_deref_instr_parent(const nir_deref_instr *instr)
1834 {
1835 if (instr->deref_type == nir_deref_type_var)
1836 return NULL;
1837 else
1838 return nir_src_as_deref(instr->parent);
1839 }
1840
1841 static inline nir_variable *
nir_deref_instr_get_variable(const nir_deref_instr * instr)1842 nir_deref_instr_get_variable(const nir_deref_instr *instr)
1843 {
1844 while (instr->deref_type != nir_deref_type_var) {
1845 if (instr->deref_type == nir_deref_type_cast)
1846 return NULL;
1847
1848 instr = nir_deref_instr_parent(instr);
1849 }
1850
1851 return instr->var;
1852 }
1853
1854 bool nir_deref_instr_has_indirect(nir_deref_instr *instr);
1855 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr *instr);
1856
1857 typedef enum {
1858 nir_deref_instr_has_complex_use_allow_memcpy_src = (1 << 0),
1859 nir_deref_instr_has_complex_use_allow_memcpy_dst = (1 << 1),
1860 nir_deref_instr_has_complex_use_allow_atomics = (1 << 2),
1861 } nir_deref_instr_has_complex_use_options;
1862
1863 bool nir_deref_instr_has_complex_use(nir_deref_instr *instr,
1864 nir_deref_instr_has_complex_use_options opts);
1865
1866 bool nir_deref_instr_remove_if_unused(nir_deref_instr *instr);
1867
1868 unsigned nir_deref_instr_array_stride(nir_deref_instr *instr);
1869
1870 typedef struct {
1871 nir_instr instr;
1872
1873 struct nir_function *callee;
1874 /* If this function call is indirect, the function pointer to call.
1875 * Otherwise, null initialized.
1876 */
1877 nir_src indirect_callee;
1878
1879 unsigned num_params;
1880 nir_src params[];
1881 } nir_call_instr;
1882
1883 #include "nir_intrinsics.h"
1884
1885 #define NIR_INTRINSIC_MAX_CONST_INDEX 8
1886
1887 /** Represents an intrinsic
1888 *
1889 * An intrinsic is an instruction type for handling things that are
1890 * more-or-less regular operations but don't just consume and produce SSA
1891 * values like ALU operations do. Intrinsics are not for things that have
1892 * special semantic meaning such as phi nodes and parallel copies.
1893 * Examples of intrinsics include variable load/store operations, system
1894 * value loads, and the like. Even though texturing more-or-less falls
1895 * under this category, texturing is its own instruction type because
1896 * trying to represent texturing with intrinsics would lead to a
1897 * combinatorial explosion of intrinsic opcodes.
1898 *
1899 * By having a single instruction type for handling a lot of different
1900 * cases, optimization passes can look for intrinsics and, for the most
1901 * part, completely ignore them. Each intrinsic type also has a few
1902 * possible flags that govern whether or not they can be reordered or
1903 * eliminated. That way passes like dead code elimination can still work
1904 * on intrisics without understanding the meaning of each.
1905 *
1906 * Each intrinsic has some number of constant indices, some number of
1907 * variables, and some number of sources. What these sources, variables,
1908 * and indices mean depends on the intrinsic and is documented with the
1909 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1910 * instructions are the only types of instruction that can operate on
1911 * variables.
1912 */
1913 typedef struct {
1914 nir_instr instr;
1915
1916 nir_intrinsic_op intrinsic;
1917
1918 nir_def def;
1919
1920 /** number of components if this is a vectorized intrinsic
1921 *
1922 * Similarly to ALU operations, some intrinsics are vectorized.
1923 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1924 * For vectorized intrinsics, the num_components field specifies the
1925 * number of destination components and the number of source components
1926 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1927 */
1928 uint8_t num_components;
1929
1930 int const_index[NIR_INTRINSIC_MAX_CONST_INDEX];
1931
1932 /* a variable name associated with this instr; cannot be modified or freed */
1933 const char *name;
1934
1935 nir_src src[];
1936 } nir_intrinsic_instr;
1937
1938 static inline nir_variable *
nir_intrinsic_get_var(const nir_intrinsic_instr * intrin,unsigned i)1939 nir_intrinsic_get_var(const nir_intrinsic_instr *intrin, unsigned i)
1940 {
1941 return nir_deref_instr_get_variable(nir_src_as_deref(intrin->src[i]));
1942 }
1943
1944 typedef enum {
1945 /* Memory ordering. */
1946 NIR_MEMORY_ACQUIRE = 1 << 0,
1947 NIR_MEMORY_RELEASE = 1 << 1,
1948 NIR_MEMORY_ACQ_REL = NIR_MEMORY_ACQUIRE | NIR_MEMORY_RELEASE,
1949
1950 /* Memory visibility operations. */
1951 NIR_MEMORY_MAKE_AVAILABLE = 1 << 2,
1952 NIR_MEMORY_MAKE_VISIBLE = 1 << 3,
1953 } nir_memory_semantics;
1954
1955 /**
1956 * NIR intrinsics semantic flags
1957 *
1958 * information about what the compiler can do with the intrinsics.
1959 *
1960 * :c:member:`nir_intrinsic_info.flags`
1961 */
1962 typedef enum {
1963 /**
1964 * whether the intrinsic can be safely eliminated if none of its output
1965 * value is not being used.
1966 */
1967 NIR_INTRINSIC_CAN_ELIMINATE = (1 << 0),
1968
1969 /**
1970 * Whether the intrinsic can be reordered with respect to any other
1971 * intrinsic, i.e. whether the only reordering dependencies of the
1972 * intrinsic are due to the register reads/writes.
1973 */
1974 NIR_INTRINSIC_CAN_REORDER = (1 << 1),
1975 } nir_intrinsic_semantic_flag;
1976
1977 /**
1978 * Maximum valid value for a nir align_mul value (in intrinsics or derefs).
1979 *
1980 * Offsets can be signed, so this is the largest power of two in int32_t.
1981 */
1982 #define NIR_ALIGN_MUL_MAX 0x40000000
1983
1984 typedef struct nir_io_semantics {
1985 unsigned location : 7; /* gl_vert_attrib, gl_varying_slot, or gl_frag_result */
1986 unsigned num_slots : 6; /* max 32, may be pessimistic with const indexing */
1987 unsigned dual_source_blend_index : 1;
1988 unsigned fb_fetch_output : 1; /* for GL_KHR_blend_equation_advanced */
1989 unsigned fb_fetch_output_coherent : 1;
1990 unsigned gs_streams : 8; /* xxyyzzww: 2-bit stream index for each component */
1991 unsigned medium_precision : 1; /* GLSL mediump qualifier */
1992 unsigned per_view : 1;
1993 unsigned high_16bits : 1; /* whether accessing low or high half of the slot */
1994 unsigned invariant : 1; /* The variable has the invariant flag set */
1995 unsigned high_dvec2 : 1; /* whether accessing the high half of dvec3/dvec4 */
1996 /* CLIP_DISTn, LAYER, VIEWPORT, and TESS_LEVEL_* have up to 3 uses:
1997 * - an output consumed by the next stage
1998 * - a system value output affecting fixed-func hardware, e.g. the clipper
1999 * - a transform feedback output written to memory
2000 * The following fields disable the first two. Transform feedback is disabled
2001 * by transform feedback info.
2002 */
2003 unsigned no_varying : 1; /* whether this output isn't consumed by the next stage */
2004 unsigned no_sysval_output : 1; /* whether this system value output has no
2005 effect due to current pipeline states */
2006 unsigned interp_explicit_strict : 1; /* preserve original vertex order */
2007 } nir_io_semantics;
2008
2009 /* Transform feedback info for 2 outputs. nir_intrinsic_store_output contains
2010 * this structure twice to support up to 4 outputs. The structure is limited
2011 * to 32 bits because it's stored in nir_intrinsic_instr::const_index[].
2012 */
2013 typedef struct nir_io_xfb {
2014 struct {
2015 /* start_component is equal to the index of out[]; add 2 for io_xfb2 */
2016 /* start_component is not relative to nir_intrinsic_component */
2017 /* get the stream index from nir_io_semantics */
2018 uint8_t num_components : 4; /* max 4; if this is 0, xfb is disabled */
2019 uint8_t buffer : 4; /* buffer index, max 3 */
2020 uint8_t offset; /* transform feedback buffer offset in dwords,
2021 max (1K - 4) bytes */
2022 } out[2];
2023 } nir_io_xfb;
2024
2025 unsigned
2026 nir_instr_xfb_write_mask(nir_intrinsic_instr *instr);
2027
2028 #define NIR_INTRINSIC_MAX_INPUTS 11
2029
2030 typedef struct {
2031 const char *name;
2032
2033 /** number of register/SSA inputs */
2034 uint8_t num_srcs;
2035
2036 /** number of components of each input register
2037 *
2038 * If this value is 0, the number of components is given by the
2039 * num_components field of nir_intrinsic_instr. If this value is -1, the
2040 * intrinsic consumes however many components are provided and it is not
2041 * validated at all.
2042 */
2043 int8_t src_components[NIR_INTRINSIC_MAX_INPUTS];
2044
2045 bool has_dest;
2046
2047 /** number of components of the output register
2048 *
2049 * If this value is 0, the number of components is given by the
2050 * num_components field of nir_intrinsic_instr.
2051 */
2052 uint8_t dest_components;
2053
2054 /** bitfield of legal bit sizes */
2055 uint8_t dest_bit_sizes;
2056
2057 /** source which the destination bit size must match
2058 *
2059 * Some intrinsics, such as subgroup intrinsics, are data manipulation
2060 * intrinsics and they have similar bit-size rules to ALU ops. This enables
2061 * validation to validate a bit more and enables auto-generated builder code
2062 * to properly determine destination bit sizes automatically.
2063 */
2064 int8_t bit_size_src;
2065
2066 /** the number of constant indices used by the intrinsic */
2067 uint8_t num_indices;
2068
2069 /** list of indices */
2070 uint8_t indices[NIR_INTRINSIC_MAX_CONST_INDEX];
2071
2072 /** indicates the usage of intr->const_index[n] */
2073 uint8_t index_map[NIR_INTRINSIC_NUM_INDEX_FLAGS];
2074
2075 /** semantic flags for calls to this intrinsic */
2076 nir_intrinsic_semantic_flag flags;
2077 } nir_intrinsic_info;
2078
2079 extern const nir_intrinsic_info nir_intrinsic_infos[nir_num_intrinsics];
2080
2081 unsigned
2082 nir_intrinsic_src_components(const nir_intrinsic_instr *intr, unsigned srcn);
2083
2084 unsigned
2085 nir_intrinsic_dest_components(nir_intrinsic_instr *intr);
2086
2087 nir_alu_type
2088 nir_intrinsic_instr_src_type(const nir_intrinsic_instr *intrin, unsigned src);
2089
2090 nir_alu_type
2091 nir_intrinsic_instr_dest_type(const nir_intrinsic_instr *intrin);
2092
2093 /**
2094 * Helper to copy const_index[] from src to dst, without assuming they
2095 * match in order.
2096 */
2097 void nir_intrinsic_copy_const_indices(nir_intrinsic_instr *dst, nir_intrinsic_instr *src);
2098
2099 #include "nir_intrinsics_indices.h"
2100
2101 static inline void
nir_intrinsic_set_align(nir_intrinsic_instr * intrin,unsigned align_mul,unsigned align_offset)2102 nir_intrinsic_set_align(nir_intrinsic_instr *intrin,
2103 unsigned align_mul, unsigned align_offset)
2104 {
2105 assert(util_is_power_of_two_nonzero(align_mul));
2106 assert(align_offset < align_mul);
2107 nir_intrinsic_set_align_mul(intrin, align_mul);
2108 nir_intrinsic_set_align_offset(intrin, align_offset);
2109 }
2110
2111 /** Returns a simple alignment for an align_mul/offset pair
2112 *
2113 * This helper converts from the full mul+offset alignment scheme used by
2114 * most NIR intrinsics to a simple alignment. The returned value is the
2115 * largest power of two which divides both align_mul and align_offset.
2116 * For any offset X which satisfies the complex alignment described by
2117 * align_mul/offset, X % align == 0.
2118 */
2119 static inline uint32_t
nir_combined_align(uint32_t align_mul,uint32_t align_offset)2120 nir_combined_align(uint32_t align_mul, uint32_t align_offset)
2121 {
2122 assert(util_is_power_of_two_nonzero(align_mul));
2123 assert(align_offset < align_mul);
2124 return align_offset ? 1 << (ffs(align_offset) - 1) : align_mul;
2125 }
2126
2127 /** Returns a simple alignment for a load/store intrinsic offset
2128 *
2129 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
2130 * and ALIGN_OFFSET parameters, this helper takes both into account and
2131 * provides a single simple alignment parameter. The offset X is guaranteed
2132 * to satisfy X % align == 0.
2133 */
2134 static inline unsigned
nir_intrinsic_align(const nir_intrinsic_instr * intrin)2135 nir_intrinsic_align(const nir_intrinsic_instr *intrin)
2136 {
2137 return nir_combined_align(nir_intrinsic_align_mul(intrin),
2138 nir_intrinsic_align_offset(intrin));
2139 }
2140
2141 static inline bool
nir_intrinsic_has_align(const nir_intrinsic_instr * intrin)2142 nir_intrinsic_has_align(const nir_intrinsic_instr *intrin)
2143 {
2144 return nir_intrinsic_has_align_mul(intrin) &&
2145 nir_intrinsic_has_align_offset(intrin);
2146 }
2147
2148 unsigned
2149 nir_image_intrinsic_coord_components(const nir_intrinsic_instr *instr);
2150
2151 /* Converts a image_deref_* intrinsic into a image_* one */
2152 void nir_rewrite_image_intrinsic(nir_intrinsic_instr *instr,
2153 nir_def *handle, bool bindless);
2154
2155 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
2156 bool nir_intrinsic_can_reorder(nir_intrinsic_instr *instr);
2157
2158 bool nir_intrinsic_writes_external_memory(const nir_intrinsic_instr *instr);
2159
2160 static inline bool
nir_intrinsic_is_ray_query(nir_intrinsic_op intrinsic)2161 nir_intrinsic_is_ray_query(nir_intrinsic_op intrinsic)
2162 {
2163 switch (intrinsic) {
2164 case nir_intrinsic_rq_confirm_intersection:
2165 case nir_intrinsic_rq_generate_intersection:
2166 case nir_intrinsic_rq_initialize:
2167 case nir_intrinsic_rq_load:
2168 case nir_intrinsic_rq_proceed:
2169 case nir_intrinsic_rq_terminate:
2170 return true;
2171 default:
2172 return false;
2173 }
2174 }
2175
2176 /** Texture instruction source type */
2177 typedef enum nir_tex_src_type {
2178 /** Texture coordinate
2179 *
2180 * Must have :c:member:`nir_tex_instr.coord_components` components.
2181 */
2182 nir_tex_src_coord,
2183
2184 /** Projector
2185 *
2186 * The texture coordinate (except for the array component, if any) is
2187 * divided by this value before LOD computation and sampling.
2188 *
2189 * Must be a float scalar.
2190 */
2191 nir_tex_src_projector,
2192
2193 /** Shadow comparator
2194 *
2195 * For shadow sampling, the fetched texel values are compared against the
2196 * shadow comparator using the compare op specified by the sampler object
2197 * and converted to 1.0 if the comparison succeeds and 0.0 if it fails.
2198 * Interpolation happens after this conversion so the actual result may be
2199 * anywhere in the range [0.0, 1.0].
2200 *
2201 * Only valid if :c:member:`nir_tex_instr.is_shadow` and must be a float
2202 * scalar.
2203 */
2204 nir_tex_src_comparator,
2205
2206 /** Coordinate offset
2207 *
2208 * An integer value that is added to the texel address before sampling.
2209 * This is only allowed with operations that take an explicit LOD as it is
2210 * applied in integer texel space after LOD selection and not normalized
2211 * coordinate space.
2212 */
2213 nir_tex_src_offset,
2214
2215 /** LOD bias
2216 *
2217 * This value is added to the computed LOD before mip-mapping.
2218 */
2219 nir_tex_src_bias,
2220
2221 /** Explicit LOD */
2222 nir_tex_src_lod,
2223
2224 /** Min LOD
2225 *
2226 * The computed LOD is clamped to be at least as large as min_lod before
2227 * mip-mapping.
2228 */
2229 nir_tex_src_min_lod,
2230
2231 /** MSAA sample index */
2232 nir_tex_src_ms_index,
2233
2234 /** Intel-specific MSAA compression data */
2235 nir_tex_src_ms_mcs_intel,
2236
2237 /** Explicit horizontal (X-major) coordinate derivative */
2238 nir_tex_src_ddx,
2239
2240 /** Explicit vertical (Y-major) coordinate derivative */
2241 nir_tex_src_ddy,
2242
2243 /** Texture variable dereference */
2244 nir_tex_src_texture_deref,
2245
2246 /** Sampler variable dereference */
2247 nir_tex_src_sampler_deref,
2248
2249 /** Texture index offset
2250 *
2251 * This is added to :c:member:`nir_tex_instr.texture_index`. Unless
2252 * :c:member:`nir_tex_instr.texture_non_uniform` is set, this is guaranteed
2253 * to be dynamically uniform.
2254 */
2255 nir_tex_src_texture_offset,
2256
2257 /** Dynamically uniform sampler index offset
2258 *
2259 * This is added to :c:member:`nir_tex_instr.sampler_index`. Unless
2260 * :c:member:`nir_tex_instr.sampler_non_uniform` is set, this is guaranteed to be
2261 * dynamically uniform. This should not be present until GLSL ES 3.20, GLSL
2262 * 4.00, or ARB_gpu_shader5, because in ES 3.10 and GL 3.30 samplers said
2263 * "When aggregated into arrays within a shader, samplers can only be indexed
2264 * with a constant integral expression."
2265 */
2266 nir_tex_src_sampler_offset,
2267
2268 /** Bindless texture handle
2269 *
2270 * This is, unfortunately, a bit overloaded at the moment. There are
2271 * generally two types of bindless handles:
2272 *
2273 * 1. For GL_ARB_bindless bindless handles. These are part of the
2274 * GL/Gallium-level API and are always a 64-bit integer.
2275 *
2276 * 2. HW-specific handles. GL_ARB_bindless handles may be lowered to
2277 * these. Also, these are used by many Vulkan drivers to implement
2278 * descriptor sets, especially for UPDATE_AFTER_BIND descriptors.
2279 * The details of hardware handles (bit size, format, etc.) is
2280 * HW-specific.
2281 *
2282 * Because of this overloading and the resulting ambiguity, we currently
2283 * don't validate anything for these.
2284 */
2285 nir_tex_src_texture_handle,
2286
2287 /** Bindless sampler handle
2288 *
2289 * See nir_tex_src_texture_handle,
2290 */
2291 nir_tex_src_sampler_handle,
2292
2293 /** Tex src intrinsic
2294 *
2295 * This is an intrinsic used before function inlining i.e. before we know
2296 * if a bindless value has been given as function param for use as a tex
2297 * src.
2298 */
2299 nir_tex_src_sampler_deref_intrinsic,
2300 nir_tex_src_texture_deref_intrinsic,
2301
2302 /** Plane index for multi-plane YCbCr textures */
2303 nir_tex_src_plane,
2304
2305 /**
2306 * Backend-specific vec4 tex src argument.
2307 *
2308 * Can be used to have NIR optimization (copy propagation, lower_vec_to_regs)
2309 * apply to the packing of the tex srcs. This lowering must only happen
2310 * after nir_lower_tex().
2311 *
2312 * The nir_tex_instr_src_type() of this argument is float, so no lowering
2313 * will happen if nir_lower_int_to_float is used.
2314 */
2315 nir_tex_src_backend1,
2316
2317 /** Second backend-specific vec4 tex src argument, see nir_tex_src_backend1. */
2318 nir_tex_src_backend2,
2319
2320 nir_num_tex_src_types
2321 } nir_tex_src_type;
2322
2323 /** A texture instruction source */
2324 typedef struct nir_tex_src {
2325 /** Base source */
2326 nir_src src;
2327
2328 /** Type of this source */
2329 nir_tex_src_type src_type;
2330 } nir_tex_src;
2331
2332 /** Texture instruction opcode */
2333 typedef enum nir_texop {
2334 /** Regular texture look-up */
2335 nir_texop_tex,
2336 /** Texture look-up with LOD bias */
2337 nir_texop_txb,
2338 /** Texture look-up with explicit LOD */
2339 nir_texop_txl,
2340 /** Texture look-up with partial derivatives */
2341 nir_texop_txd,
2342 /** Texel fetch with explicit LOD */
2343 nir_texop_txf,
2344 /** Multisample texture fetch */
2345 nir_texop_txf_ms,
2346 /** Multisample texture fetch from framebuffer */
2347 nir_texop_txf_ms_fb,
2348 /** Multisample compression value fetch */
2349 nir_texop_txf_ms_mcs_intel,
2350 /** Texture size */
2351 nir_texop_txs,
2352 /** Texture lod query */
2353 nir_texop_lod,
2354 /** Texture gather */
2355 nir_texop_tg4,
2356 /** Texture levels query */
2357 nir_texop_query_levels,
2358 /** Texture samples query */
2359 nir_texop_texture_samples,
2360 /** Query whether all samples are definitely identical. */
2361 nir_texop_samples_identical,
2362 /** Regular texture look-up, eligible for pre-dispatch */
2363 nir_texop_tex_prefetch,
2364 /** Multisample fragment color texture fetch */
2365 nir_texop_fragment_fetch_amd,
2366 /** Multisample fragment mask texture fetch */
2367 nir_texop_fragment_mask_fetch_amd,
2368 /** Returns a buffer or image descriptor. */
2369 nir_texop_descriptor_amd,
2370 /** Returns a sampler descriptor. */
2371 nir_texop_sampler_descriptor_amd,
2372 /** Returns the sampler's LOD bias */
2373 nir_texop_lod_bias_agx,
2374 /** Returns a bool indicating that the sampler uses a custom border colour */
2375 nir_texop_has_custom_border_color_agx,
2376 /** Returns the sampler's custom border colour (if has_custom_border_agx) */
2377 nir_texop_custom_border_color_agx,
2378 /** Maps to TXQ.DIMENSION */
2379 nir_texop_hdr_dim_nv,
2380 /** Maps to TXQ.TEXTURE_TYPE */
2381 nir_texop_tex_type_nv,
2382 } nir_texop;
2383
2384 /** Represents a texture instruction */
2385 typedef struct nir_tex_instr {
2386 /** Base instruction */
2387 nir_instr instr;
2388
2389 /** Dimensionality of the texture operation
2390 *
2391 * This will typically match the dimensionality of the texture deref type
2392 * if a nir_tex_src_texture_deref is present. However, it may not if
2393 * texture lowering has occurred.
2394 */
2395 enum glsl_sampler_dim sampler_dim;
2396
2397 /** ALU type of the destination
2398 *
2399 * This is the canonical sampled type for this texture operation and may
2400 * not exactly match the sampled type of the deref type when a
2401 * nir_tex_src_texture_deref is present. For OpenCL, the sampled type of
2402 * the texture deref will be GLSL_TYPE_VOID and this is allowed to be
2403 * anything. With SPIR-V, the signedness of integer types is allowed to
2404 * differ. For all APIs, the bit size may differ if the driver has done
2405 * any sort of mediump or similar lowering since texture types always have
2406 * 32-bit sampled types.
2407 */
2408 nir_alu_type dest_type;
2409
2410 /** Texture opcode */
2411 nir_texop op;
2412
2413 /** Destination */
2414 nir_def def;
2415
2416 /** Array of sources
2417 *
2418 * This array has :c:member:`nir_tex_instr.num_srcs` elements
2419 */
2420 nir_tex_src *src;
2421
2422 /** Number of sources */
2423 unsigned num_srcs;
2424
2425 /** Number of components in the coordinate, if any */
2426 unsigned coord_components;
2427
2428 /** True if the texture instruction acts on an array texture */
2429 bool is_array;
2430
2431 /** True if the texture instruction performs a shadow comparison
2432 *
2433 * If this is true, the texture instruction must have a
2434 * nir_tex_src_comparator.
2435 */
2436 bool is_shadow;
2437
2438 /**
2439 * If is_shadow is true, whether this is the old-style shadow that outputs
2440 * 4 components or the new-style shadow that outputs 1 component.
2441 */
2442 bool is_new_style_shadow;
2443
2444 /**
2445 * True if this texture instruction should return a sparse residency code.
2446 * The code is in the last component of the result.
2447 */
2448 bool is_sparse;
2449
2450 /** nir_texop_tg4 component selector
2451 *
2452 * This determines which RGBA component is gathered.
2453 */
2454 unsigned component : 2;
2455
2456 /** Validation needs to know this for gradient component count */
2457 unsigned array_is_lowered_cube : 1;
2458
2459 /** True if this tg4 instruction has an implicit LOD or LOD bias, instead of using level 0 */
2460 unsigned is_gather_implicit_lod : 1;
2461
2462 /** Gather offsets */
2463 int8_t tg4_offsets[4][2];
2464
2465 /** True if the texture index or handle is not dynamically uniform */
2466 bool texture_non_uniform;
2467
2468 /** True if the sampler index or handle is not dynamically uniform.
2469 *
2470 * This may be set when VK_EXT_descriptor_indexing is supported and the
2471 * appropriate capability is enabled.
2472 *
2473 * This should always be false in GLSL (GLSL ES 3.20 says "When aggregated
2474 * into arrays within a shader, opaque types can only be indexed with a
2475 * dynamically uniform integral expression", and GLSL 4.60 says "When
2476 * aggregated into arrays within a shader, [texture, sampler, and
2477 * samplerShadow] types can only be indexed with a dynamically uniform
2478 * expression, or texture lookup will result in undefined values.").
2479 */
2480 bool sampler_non_uniform;
2481
2482 /** The texture index
2483 *
2484 * If this texture instruction has a nir_tex_src_texture_offset source,
2485 * then the texture index is given by texture_index + texture_offset.
2486 */
2487 unsigned texture_index;
2488
2489 /** The sampler index
2490 *
2491 * The following operations do not require a sampler and, as such, this
2492 * field should be ignored:
2493 *
2494 * - nir_texop_txf
2495 * - nir_texop_txf_ms
2496 * - nir_texop_txs
2497 * - nir_texop_query_levels
2498 * - nir_texop_texture_samples
2499 * - nir_texop_samples_identical
2500 *
2501 * If this texture instruction has a nir_tex_src_sampler_offset source,
2502 * then the sampler index is given by sampler_index + sampler_offset.
2503 */
2504 unsigned sampler_index;
2505
2506 /* Back-end specific flags, intended to be used in combination with
2507 * nir_tex_src_backend1/2 to provide additional hw-specific information
2508 * to the back-end compiler.
2509 */
2510 uint32_t backend_flags;
2511 } nir_tex_instr;
2512
2513 /**
2514 * Returns true if the texture operation requires a sampler as a general rule
2515 *
2516 * Note that the specific hw/driver backend could require to a sampler
2517 * object/configuration packet in any case, for some other reason.
2518 *
2519 * See also :c:member:`nir_tex_instr.sampler_index`.
2520 */
2521 bool nir_tex_instr_need_sampler(const nir_tex_instr *instr);
2522
2523 /** Returns the number of components returned by this nir_tex_instr
2524 *
2525 * Useful for code building texture instructions when you don't want to think
2526 * about how many components a particular texture op returns. This does not
2527 * include the sparse residency code.
2528 */
2529 unsigned
2530 nir_tex_instr_result_size(const nir_tex_instr *instr);
2531
2532 /**
2533 * Returns the destination size of this nir_tex_instr including the sparse
2534 * residency code, if any.
2535 */
2536 static inline unsigned
nir_tex_instr_dest_size(const nir_tex_instr * instr)2537 nir_tex_instr_dest_size(const nir_tex_instr *instr)
2538 {
2539 /* One more component is needed for the residency code. */
2540 return nir_tex_instr_result_size(instr) + instr->is_sparse;
2541 }
2542
2543 /**
2544 * Returns true if this texture operation queries something about the texture
2545 * rather than actually sampling it.
2546 */
2547 bool
2548 nir_tex_instr_is_query(const nir_tex_instr *instr);
2549
2550 /** Returns true if this texture instruction does implicit derivatives
2551 *
2552 * This is important as there are extra control-flow rules around derivatives
2553 * and texture instructions which perform them implicitly.
2554 */
2555 bool
2556 nir_tex_instr_has_implicit_derivative(const nir_tex_instr *instr);
2557
2558 /** Returns the ALU type of the given texture instruction source */
2559 nir_alu_type
2560 nir_tex_instr_src_type(const nir_tex_instr *instr, unsigned src);
2561
2562 /**
2563 * Returns the number of components required by the given texture instruction
2564 * source
2565 */
2566 unsigned
2567 nir_tex_instr_src_size(const nir_tex_instr *instr, unsigned src);
2568
2569 /**
2570 * Returns the index of the texture instruction source with the given
2571 * nir_tex_src_type or -1 if no such source exists.
2572 */
2573 static inline int
nir_tex_instr_src_index(const nir_tex_instr * instr,nir_tex_src_type type)2574 nir_tex_instr_src_index(const nir_tex_instr *instr, nir_tex_src_type type)
2575 {
2576 for (unsigned i = 0; i < instr->num_srcs; i++)
2577 if (instr->src[i].src_type == type)
2578 return (int)i;
2579
2580 return -1;
2581 }
2582
2583 /** Adds a source to a texture instruction */
2584 void nir_tex_instr_add_src(nir_tex_instr *tex,
2585 nir_tex_src_type src_type,
2586 nir_def *src);
2587
2588 /** Removes a source from a texture instruction */
2589 void nir_tex_instr_remove_src(nir_tex_instr *tex, unsigned src_idx);
2590
2591 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr *tex);
2592
2593 typedef struct {
2594 nir_instr instr;
2595
2596 nir_def def;
2597
2598 nir_const_value value[];
2599 } nir_load_const_instr;
2600
2601 typedef enum {
2602 /** Return from a function
2603 *
2604 * This instruction is a classic function return. It jumps to
2605 * nir_function_impl::end_block. No return value is provided in this
2606 * instruction. Instead, the function is expected to write any return
2607 * data to a deref passed in from the caller.
2608 */
2609 nir_jump_return,
2610
2611 /** Immediately exit the current shader
2612 *
2613 * This instruction is roughly the equivalent of C's "exit()" in that it
2614 * immediately terminates the current shader invocation. From a CFG
2615 * perspective, it looks like a jump to nir_function_impl::end_block but
2616 * it actually jumps to the end block of the shader entrypoint. A halt
2617 * instruction in the shader entrypoint itself is semantically identical
2618 * to a return.
2619 *
2620 * For shaders with built-in I/O, any outputs written prior to a halt
2621 * instruction remain written and any outputs not written prior to the
2622 * halt have undefined values. It does NOT cause an implicit discard of
2623 * written results. If one wants discard results in a fragment shader,
2624 * for instance, a discard or demote intrinsic is required.
2625 */
2626 nir_jump_halt,
2627
2628 /** Break out of the inner-most loop
2629 *
2630 * This has the same semantics as C's "break" statement.
2631 */
2632 nir_jump_break,
2633
2634 /** Jump back to the top of the inner-most loop
2635 *
2636 * This has the same semantics as C's "continue" statement assuming that a
2637 * NIR loop is implemented as "while (1) { body }".
2638 */
2639 nir_jump_continue,
2640
2641 /** Jumps for unstructured CFG.
2642 *
2643 * As within an unstructured CFG we can't rely on block ordering we need to
2644 * place explicit jumps at the end of every block.
2645 */
2646 nir_jump_goto,
2647 nir_jump_goto_if,
2648 } nir_jump_type;
2649
2650 typedef struct {
2651 nir_instr instr;
2652 nir_jump_type type;
2653 nir_src condition;
2654 struct nir_block *target;
2655 struct nir_block *else_target;
2656 } nir_jump_instr;
2657
2658 /* creates a new SSA variable in an undefined state */
2659
2660 typedef struct {
2661 nir_instr instr;
2662 nir_def def;
2663 } nir_undef_instr;
2664
2665 typedef struct {
2666 struct exec_node node;
2667
2668 /* The predecessor block corresponding to this source */
2669 struct nir_block *pred;
2670
2671 nir_src src;
2672 } nir_phi_src;
2673
2674 #define nir_foreach_phi_src(phi_src, phi) \
2675 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
2676 #define nir_foreach_phi_src_safe(phi_src, phi) \
2677 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
2678
2679 typedef struct {
2680 nir_instr instr;
2681
2682 /** list of nir_phi_src */
2683 struct exec_list srcs;
2684
2685 nir_def def;
2686 } nir_phi_instr;
2687
2688 static inline nir_phi_src *
nir_phi_get_src_from_block(nir_phi_instr * phi,struct nir_block * block)2689 nir_phi_get_src_from_block(nir_phi_instr *phi, struct nir_block *block)
2690 {
2691 nir_foreach_phi_src(src, phi) {
2692 if (src->pred == block)
2693 return src;
2694 }
2695
2696 assert(!"Block is not a predecessor of phi.");
2697 return NULL;
2698 }
2699
2700 typedef struct {
2701 struct exec_node node;
2702 bool src_is_reg;
2703 bool dest_is_reg;
2704 nir_src src;
2705 union {
2706 nir_def def;
2707 nir_src reg;
2708 } dest;
2709 } nir_parallel_copy_entry;
2710
2711 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
2712 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
2713
2714 typedef struct {
2715 nir_instr instr;
2716
2717 /* A list of nir_parallel_copy_entrys. The sources of all of the
2718 * entries are copied to the corresponding destinations "in parallel".
2719 * In other words, if we have two entries: a -> b and b -> a, the values
2720 * get swapped.
2721 */
2722 struct exec_list entries;
2723 } nir_parallel_copy_instr;
2724
2725 typedef enum nir_debug_info_type {
2726 nir_debug_info_src_loc,
2727 nir_debug_info_string,
2728 } nir_debug_info_type;
2729
2730 typedef enum nir_debug_info_source {
2731 nir_debug_info_spirv,
2732 nir_debug_info_nir,
2733 } nir_debug_info_source;
2734
2735 typedef struct nir_debug_info_instr {
2736 nir_instr instr;
2737
2738 nir_debug_info_type type;
2739
2740 union {
2741 struct {
2742 nir_src filename;
2743 /* 0 if only the spirv_offset is available. */
2744 uint32_t line;
2745 uint32_t column;
2746
2747 uint32_t spirv_offset;
2748
2749 nir_debug_info_source source;
2750 } src_loc;
2751
2752 uint16_t string_length;
2753 };
2754
2755 nir_def def;
2756
2757 char string[];
2758 } nir_debug_info_instr;
2759
2760 NIR_DEFINE_CAST(nir_instr_as_alu, nir_instr, nir_alu_instr, instr,
2761 type, nir_instr_type_alu)
2762 NIR_DEFINE_CAST(nir_instr_as_deref, nir_instr, nir_deref_instr, instr,
2763 type, nir_instr_type_deref)
2764 NIR_DEFINE_CAST(nir_instr_as_call, nir_instr, nir_call_instr, instr,
2765 type, nir_instr_type_call)
2766 NIR_DEFINE_CAST(nir_instr_as_jump, nir_instr, nir_jump_instr, instr,
2767 type, nir_instr_type_jump)
2768 NIR_DEFINE_CAST(nir_instr_as_tex, nir_instr, nir_tex_instr, instr,
2769 type, nir_instr_type_tex)
2770 NIR_DEFINE_CAST(nir_instr_as_intrinsic, nir_instr, nir_intrinsic_instr, instr,
2771 type, nir_instr_type_intrinsic)
2772 NIR_DEFINE_CAST(nir_instr_as_load_const, nir_instr, nir_load_const_instr, instr,
2773 type, nir_instr_type_load_const)
2774 NIR_DEFINE_CAST(nir_instr_as_undef, nir_instr, nir_undef_instr, instr,
2775 type, nir_instr_type_undef)
2776 NIR_DEFINE_CAST(nir_instr_as_phi, nir_instr, nir_phi_instr, instr,
2777 type, nir_instr_type_phi)
2778 NIR_DEFINE_CAST(nir_instr_as_parallel_copy, nir_instr,
2779 nir_parallel_copy_instr, instr,
2780 type, nir_instr_type_parallel_copy)
2781 NIR_DEFINE_CAST(nir_instr_as_debug_info, nir_instr,
2782 nir_debug_info_instr, instr,
2783 type, nir_instr_type_debug_info)
2784
2785 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
2786 static inline type \
2787 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2788 { \
2789 assert(nir_src_is_const(src)); \
2790 nir_load_const_instr *load = \
2791 nir_instr_as_load_const(src.ssa->parent_instr); \
2792 assert(comp < load->def.num_components); \
2793 return nir_const_value_as_##suffix(load->value[comp], \
2794 load->def.bit_size); \
2795 } \
2796 \
2797 static inline type \
2798 nir_src_as_##suffix(nir_src src) \
2799 { \
2800 assert(nir_src_num_components(src) == 1); \
2801 return nir_src_comp_as_##suffix(src, 0); \
2802 }
2803
2804 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2805 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint)
2806 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2807 NIR_DEFINE_SRC_AS_CONST(double, float)
2808
2809 #undef NIR_DEFINE_SRC_AS_CONST
2810
2811 typedef struct {
2812 nir_def *def;
2813 unsigned comp;
2814 } nir_scalar;
2815
2816 static inline bool
nir_scalar_is_const(nir_scalar s)2817 nir_scalar_is_const(nir_scalar s)
2818 {
2819 return s.def->parent_instr->type == nir_instr_type_load_const;
2820 }
2821
2822 static inline bool
nir_scalar_is_undef(nir_scalar s)2823 nir_scalar_is_undef(nir_scalar s)
2824 {
2825 return s.def->parent_instr->type == nir_instr_type_undef;
2826 }
2827
2828 static inline nir_const_value
nir_scalar_as_const_value(nir_scalar s)2829 nir_scalar_as_const_value(nir_scalar s)
2830 {
2831 assert(s.comp < s.def->num_components);
2832 nir_load_const_instr *load = nir_instr_as_load_const(s.def->parent_instr);
2833 return load->value[s.comp];
2834 }
2835
2836 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2837 static inline type \
2838 nir_scalar_as_##suffix(nir_scalar s) \
2839 { \
2840 return nir_const_value_as_##suffix( \
2841 nir_scalar_as_const_value(s), s.def->bit_size); \
2842 }
2843
NIR_DEFINE_SCALAR_AS_CONST(int64_t,int)2844 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2845 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint)
2846 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2847 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2848
2849 #undef NIR_DEFINE_SCALAR_AS_CONST
2850
2851 static inline bool
2852 nir_scalar_is_alu(nir_scalar s)
2853 {
2854 return s.def->parent_instr->type == nir_instr_type_alu;
2855 }
2856
2857 static inline nir_op
nir_scalar_alu_op(nir_scalar s)2858 nir_scalar_alu_op(nir_scalar s)
2859 {
2860 return nir_instr_as_alu(s.def->parent_instr)->op;
2861 }
2862
2863 static inline bool
nir_scalar_is_intrinsic(nir_scalar s)2864 nir_scalar_is_intrinsic(nir_scalar s)
2865 {
2866 return s.def->parent_instr->type == nir_instr_type_intrinsic;
2867 }
2868
2869 static inline nir_intrinsic_op
nir_scalar_intrinsic_op(nir_scalar s)2870 nir_scalar_intrinsic_op(nir_scalar s)
2871 {
2872 return nir_instr_as_intrinsic(s.def->parent_instr)->intrinsic;
2873 }
2874
2875 static inline nir_scalar
nir_scalar_chase_alu_src(nir_scalar s,unsigned alu_src_idx)2876 nir_scalar_chase_alu_src(nir_scalar s, unsigned alu_src_idx)
2877 {
2878 nir_scalar out = { NULL, 0 };
2879
2880 nir_alu_instr *alu = nir_instr_as_alu(s.def->parent_instr);
2881 assert(alu_src_idx < nir_op_infos[alu->op].num_inputs);
2882
2883 /* Our component must be written */
2884 assert(s.comp < s.def->num_components);
2885
2886 out.def = alu->src[alu_src_idx].src.ssa;
2887
2888 if (nir_op_infos[alu->op].input_sizes[alu_src_idx] == 0) {
2889 /* The ALU src is unsized so the source component follows the
2890 * destination component.
2891 */
2892 out.comp = alu->src[alu_src_idx].swizzle[s.comp];
2893 } else {
2894 /* This is a sized source so all source components work together to
2895 * produce all the destination components. Since we need to return a
2896 * scalar, this only works if the source is a scalar.
2897 */
2898 assert(nir_op_infos[alu->op].input_sizes[alu_src_idx] == 1);
2899 out.comp = alu->src[alu_src_idx].swizzle[0];
2900 }
2901 assert(out.comp < out.def->num_components);
2902
2903 return out;
2904 }
2905
2906 nir_scalar nir_scalar_chase_movs(nir_scalar s);
2907
2908 static inline nir_scalar
nir_get_scalar(nir_def * def,unsigned channel)2909 nir_get_scalar(nir_def *def, unsigned channel)
2910 {
2911 nir_scalar s = { def, channel };
2912 return s;
2913 }
2914
2915 /** Returns a nir_scalar where we've followed the bit-exact mov/vec use chain to the original definition */
2916 static inline nir_scalar
nir_scalar_resolved(nir_def * def,unsigned channel)2917 nir_scalar_resolved(nir_def *def, unsigned channel)
2918 {
2919 return nir_scalar_chase_movs(nir_get_scalar(def, channel));
2920 }
2921
2922 static inline bool
nir_scalar_equal(nir_scalar s1,nir_scalar s2)2923 nir_scalar_equal(nir_scalar s1, nir_scalar s2)
2924 {
2925 return s1.def == s2.def && s1.comp == s2.comp;
2926 }
2927
2928 static inline uint64_t
nir_alu_src_as_uint(nir_alu_src src)2929 nir_alu_src_as_uint(nir_alu_src src)
2930 {
2931 nir_scalar scalar = nir_get_scalar(src.src.ssa, src.swizzle[0]);
2932 return nir_scalar_as_uint(scalar);
2933 }
2934
2935 typedef struct {
2936 bool success;
2937
2938 nir_variable *var;
2939 unsigned desc_set;
2940 unsigned binding;
2941 unsigned num_indices;
2942 nir_src indices[4];
2943 bool read_first_invocation;
2944 } nir_binding;
2945
2946 nir_binding nir_chase_binding(nir_src rsrc);
2947 nir_variable *nir_get_binding_variable(struct nir_shader *shader, nir_binding binding);
2948
2949 /*
2950 * Control flow
2951 *
2952 * Control flow consists of a tree of control flow nodes, which include
2953 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2954 * instructions that always run start-to-finish. Each basic block also keeps
2955 * track of its successors (blocks which may run immediately after the current
2956 * block) and predecessors (blocks which could have run immediately before the
2957 * current block). Each function also has a start block and an end block which
2958 * all return statements point to (which is always empty). Together, all the
2959 * blocks with their predecessors and successors make up the control flow
2960 * graph (CFG) of the function. There are helpers that modify the tree of
2961 * control flow nodes while modifying the CFG appropriately; these should be
2962 * used instead of modifying the tree directly.
2963 */
2964
2965 typedef enum {
2966 nir_cf_node_block,
2967 nir_cf_node_if,
2968 nir_cf_node_loop,
2969 nir_cf_node_function
2970 } nir_cf_node_type;
2971
2972 typedef struct nir_cf_node {
2973 struct exec_node node;
2974 nir_cf_node_type type;
2975 struct nir_cf_node *parent;
2976 } nir_cf_node;
2977
2978 typedef struct nir_block {
2979 nir_cf_node cf_node;
2980
2981 /** list of nir_instr */
2982 struct exec_list instr_list;
2983
2984 /** generic block index; generated by nir_index_blocks */
2985 unsigned index;
2986
2987 /* This indicates whether the block or any parent block is executed
2988 * conditionally and whether the condition uses a divergent value.
2989 */
2990 bool divergent;
2991
2992 /*
2993 * Each block can only have up to 2 successors, so we put them in a simple
2994 * array - no need for anything more complicated.
2995 */
2996 struct nir_block *successors[2];
2997
2998 /* Set of nir_block predecessors in the CFG */
2999 struct set *predecessors;
3000
3001 /*
3002 * this node's immediate dominator in the dominance tree - set to NULL for
3003 * the start block and any unreachable blocks.
3004 */
3005 struct nir_block *imm_dom;
3006
3007 /* This node's children in the dominance tree */
3008 unsigned num_dom_children;
3009 struct nir_block **dom_children;
3010
3011 /* Set of nir_blocks on the dominance frontier of this block */
3012 struct set *dom_frontier;
3013
3014 /*
3015 * These two indices have the property that dom_{pre,post}_index for each
3016 * child of this block in the dominance tree will always be between
3017 * dom_pre_index and dom_post_index for this block, which makes testing if
3018 * a given block is dominated by another block an O(1) operation.
3019 */
3020 uint32_t dom_pre_index, dom_post_index;
3021
3022 /**
3023 * Value just before the first nir_instr->index in the block, but after
3024 * end_ip that of any predecessor block.
3025 */
3026 uint32_t start_ip;
3027 /**
3028 * Value just after the last nir_instr->index in the block, but before the
3029 * start_ip of any successor block.
3030 */
3031 uint32_t end_ip;
3032
3033 /* SSA def live in and out for this block; used for liveness analysis.
3034 * Indexed by ssa_def->index
3035 */
3036 BITSET_WORD *live_in;
3037 BITSET_WORD *live_out;
3038 } nir_block;
3039
3040 static inline bool
nir_block_is_reachable(nir_block * b)3041 nir_block_is_reachable(nir_block *b)
3042 {
3043 /* See also nir_block_dominates */
3044 return b->dom_post_index != 0;
3045 }
3046
3047 static inline nir_instr *
nir_block_first_instr(nir_block * block)3048 nir_block_first_instr(nir_block *block)
3049 {
3050 struct exec_node *head = exec_list_get_head(&block->instr_list);
3051 return exec_node_data(nir_instr, head, node);
3052 }
3053
3054 static inline nir_instr *
nir_block_last_instr(nir_block * block)3055 nir_block_last_instr(nir_block *block)
3056 {
3057 struct exec_node *tail = exec_list_get_tail(&block->instr_list);
3058 return exec_node_data(nir_instr, tail, node);
3059 }
3060
3061 static inline bool
nir_block_ends_in_jump(nir_block * block)3062 nir_block_ends_in_jump(nir_block *block)
3063 {
3064 return !exec_list_is_empty(&block->instr_list) &&
3065 nir_block_last_instr(block)->type == nir_instr_type_jump;
3066 }
3067
3068 static inline bool
nir_block_ends_in_return_or_halt(nir_block * block)3069 nir_block_ends_in_return_or_halt(nir_block *block)
3070 {
3071 if (exec_list_is_empty(&block->instr_list))
3072 return false;
3073
3074 nir_instr *instr = nir_block_last_instr(block);
3075 if (instr->type != nir_instr_type_jump)
3076 return false;
3077
3078 nir_jump_instr *jump_instr = nir_instr_as_jump(instr);
3079 return jump_instr->type == nir_jump_return ||
3080 jump_instr->type == nir_jump_halt;
3081 }
3082
3083 static inline bool
nir_block_ends_in_break(nir_block * block)3084 nir_block_ends_in_break(nir_block *block)
3085 {
3086 if (exec_list_is_empty(&block->instr_list))
3087 return false;
3088
3089 nir_instr *instr = nir_block_last_instr(block);
3090 return instr->type == nir_instr_type_jump &&
3091 nir_instr_as_jump(instr)->type == nir_jump_break;
3092 }
3093
3094 bool nir_block_contains_work(nir_block *block);
3095
3096 #define nir_foreach_instr(instr, block) \
3097 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
3098 #define nir_foreach_instr_reverse(instr, block) \
3099 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
3100 #define nir_foreach_instr_safe(instr, block) \
3101 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
3102 #define nir_foreach_instr_reverse_safe(instr, block) \
3103 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
3104
3105 /* Phis come first in the block */
3106 static inline nir_phi_instr *
nir_first_phi_in_block(nir_block * block)3107 nir_first_phi_in_block(nir_block *block)
3108 {
3109 nir_foreach_instr(instr, block) {
3110 if (instr->type == nir_instr_type_phi)
3111 return nir_instr_as_phi(instr);
3112 else
3113 return NULL;
3114 }
3115
3116 return NULL;
3117 }
3118
3119 static inline nir_phi_instr *
nir_next_phi(nir_phi_instr * phi)3120 nir_next_phi(nir_phi_instr *phi)
3121 {
3122 nir_instr *next = nir_instr_next(&phi->instr);
3123
3124 if (next && next->type == nir_instr_type_phi)
3125 return nir_instr_as_phi(next);
3126 else
3127 return NULL;
3128 }
3129
3130 #define nir_foreach_phi(instr, block) \
3131 for (nir_phi_instr *instr = nir_first_phi_in_block(block); instr != NULL; \
3132 instr = nir_next_phi(instr))
3133
3134 #define nir_foreach_phi_safe(instr, block) \
3135 for (nir_phi_instr *instr = nir_first_phi_in_block(block), \
3136 *__next = instr ? nir_next_phi(instr) : NULL; \
3137 instr != NULL; \
3138 instr = __next, __next = instr ? nir_next_phi(instr) : NULL)
3139
3140 static inline nir_phi_instr *
nir_block_last_phi_instr(nir_block * block)3141 nir_block_last_phi_instr(nir_block *block)
3142 {
3143 nir_phi_instr *last_phi = NULL;
3144 nir_foreach_phi(instr, block)
3145 last_phi = instr;
3146
3147 return last_phi;
3148 }
3149
3150 typedef enum {
3151 nir_selection_control_none = 0x0,
3152
3153 /**
3154 * Defined by SPIR-V spec 3.22 "Selection Control".
3155 * The application prefers to remove control flow.
3156 */
3157 nir_selection_control_flatten = 0x1,
3158
3159 /**
3160 * Defined by SPIR-V spec 3.22 "Selection Control".
3161 * The application prefers to keep control flow.
3162 */
3163 nir_selection_control_dont_flatten = 0x2,
3164
3165 /**
3166 * May be applied by the compiler stack when it knows
3167 * that a branch is divergent, and:
3168 * - either both the if and else are always taken
3169 * - the if or else is empty and the other is always taken
3170 */
3171 nir_selection_control_divergent_always_taken = 0x3,
3172 } nir_selection_control;
3173
3174 typedef struct nir_if {
3175 nir_cf_node cf_node;
3176 nir_src condition;
3177 nir_selection_control control;
3178
3179 /** list of nir_cf_node */
3180 struct exec_list then_list;
3181
3182 /** list of nir_cf_node */
3183 struct exec_list else_list;
3184 } nir_if;
3185
3186 typedef struct {
3187 nir_if *nif;
3188
3189 /** Condition instruction that contains the induction variable */
3190 nir_instr *conditional_instr;
3191
3192 /** Block within ::nif that has the break instruction. */
3193 nir_block *break_block;
3194
3195 /** Last block for the then- or else-path that does not contain the break. */
3196 nir_block *continue_from_block;
3197
3198 /** True when ::break_block is in the else-path of ::nif. */
3199 bool continue_from_then;
3200 bool induction_rhs;
3201
3202 /* This is true if the terminators exact trip count is unknown. For
3203 * example:
3204 *
3205 * for (int i = 0; i < imin(x, 4); i++)
3206 * ...
3207 *
3208 * Here loop analysis would have set a max_trip_count of 4 however we dont
3209 * know for sure that this is the exact trip count.
3210 */
3211 bool exact_trip_count_unknown;
3212
3213 struct list_head loop_terminator_link;
3214 } nir_loop_terminator;
3215
3216 typedef struct {
3217 /* SSA def of the phi-node associated with this induction variable. */
3218 nir_def *basis;
3219
3220 /* SSA def of the increment of the induction variable. */
3221 nir_def *def;
3222
3223 /* Init statement */
3224 nir_src *init_src;
3225
3226 /* Update statement */
3227 nir_alu_src *update_src;
3228 } nir_loop_induction_variable;
3229
3230 typedef struct {
3231 /* Estimated cost (in number of instructions) of the loop */
3232 unsigned instr_cost;
3233
3234 /* Contains fp64 ops that will be lowered */
3235 bool has_soft_fp64;
3236
3237 /* Guessed trip count based on array indexing */
3238 unsigned guessed_trip_count;
3239
3240 /* Maximum number of times the loop is run (if known) */
3241 unsigned max_trip_count;
3242
3243 /* Do we know the exact number of times the loop will be run */
3244 bool exact_trip_count_known;
3245
3246 /* Unroll the loop regardless of its size */
3247 bool force_unroll;
3248
3249 /* Does the loop contain complex loop terminators, continues or other
3250 * complex behaviours? If this is true we can't rely on
3251 * loop_terminator_list to be complete or accurate.
3252 */
3253 bool complex_loop;
3254
3255 nir_loop_terminator *limiting_terminator;
3256
3257 /* A list of loop_terminators terminating this loop. */
3258 struct list_head loop_terminator_list;
3259
3260 /* hash table of induction variables for this loop */
3261 struct hash_table *induction_vars;
3262 } nir_loop_info;
3263
3264 typedef enum {
3265 nir_loop_control_none = 0x0,
3266 nir_loop_control_unroll = 0x1,
3267 nir_loop_control_dont_unroll = 0x2,
3268 } nir_loop_control;
3269
3270 typedef struct {
3271 nir_cf_node cf_node;
3272
3273 /** list of nir_cf_node */
3274 struct exec_list body;
3275
3276 /** (optional) list of nir_cf_node */
3277 struct exec_list continue_list;
3278
3279 nir_loop_info *info;
3280 nir_loop_control control;
3281 bool partially_unrolled;
3282
3283 /**
3284 * Whether some loop-active invocations might take a different control-flow path:
3285 * divergent_continue indicates that a continue statement might be taken by
3286 * only some of the loop-active invocations. A subsequent break is always
3287 * considered divergent.
3288 */
3289 bool divergent_continue;
3290 bool divergent_break;
3291 } nir_loop;
3292
3293 static inline bool
nir_loop_is_divergent(nir_loop * loop)3294 nir_loop_is_divergent(nir_loop *loop)
3295 {
3296 return loop->divergent_continue || loop->divergent_break;
3297 }
3298
3299 /**
3300 * Various bits of metadata that can may be created or required by
3301 * optimization and analysis passes
3302 */
3303 typedef enum {
3304 nir_metadata_none = 0x0,
3305
3306 /** Indicates that nir_block::index values are valid.
3307 *
3308 * The start block has index 0 and they increase through a natural walk of
3309 * the CFG. nir_function_impl::num_blocks is the number of blocks and
3310 * every block index is in the range [0, nir_function_impl::num_blocks].
3311 *
3312 * A pass can preserve this metadata type if it doesn't touch the CFG.
3313 */
3314 nir_metadata_block_index = 0x1,
3315
3316 /** Indicates that block dominance information is valid
3317 *
3318 * This includes:
3319 *
3320 * - nir_block::num_dom_children
3321 * - nir_block::dom_children
3322 * - nir_block::dom_frontier
3323 * - nir_block::dom_pre_index
3324 * - nir_block::dom_post_index
3325 *
3326 * A pass can preserve this metadata type if it doesn't touch the CFG.
3327 */
3328 nir_metadata_dominance = 0x2,
3329
3330 /** Indicates that SSA def data-flow liveness information is valid
3331 *
3332 * This includes:
3333 *
3334 * - nir_block::live_in
3335 * - nir_block::live_out
3336 *
3337 * A pass can preserve this metadata type if it never adds or removes any
3338 * SSA defs or uses of SSA defs (most passes shouldn't preserve this
3339 * metadata type).
3340 */
3341 nir_metadata_live_defs = 0x4,
3342
3343 /** A dummy metadata value to track when a pass forgot to call
3344 * nir_metadata_preserve.
3345 *
3346 * A pass should always clear this value even if it doesn't make any
3347 * progress to indicate that it thought about preserving metadata.
3348 */
3349 nir_metadata_not_properly_reset = 0x8,
3350
3351 /** Indicates that loop analysis information is valid.
3352 *
3353 * This includes everything pointed to by nir_loop::info.
3354 *
3355 * A pass can preserve this metadata type if it is guaranteed to not affect
3356 * any loop metadata. However, since loop metadata includes things like
3357 * loop counts which depend on arithmetic in the loop, this is very hard to
3358 * determine. Most passes shouldn't preserve this metadata type.
3359 */
3360 nir_metadata_loop_analysis = 0x10,
3361
3362 /** Indicates that nir_instr::index values are valid.
3363 *
3364 * The start instruction has index 0 and they increase through a natural
3365 * walk of instructions in blocks in the CFG. The indices my have holes
3366 * after passes such as DCE.
3367 *
3368 * A pass can preserve this metadata type if it never adds or moves any
3369 * instructions (most passes shouldn't preserve this metadata type), but
3370 * can preserve it if it only removes instructions.
3371 */
3372 nir_metadata_instr_index = 0x20,
3373
3374 /** All control flow metadata
3375 *
3376 * This includes all metadata preserved by a pass that preserves control flow
3377 * but modifies instructions. For example, a pass using
3378 * nir_shader_instructions_pass will typically preserve this if it does not
3379 * insert control flow.
3380 *
3381 * This is the most common metadata set to preserve, so it has its own alias.
3382 */
3383 nir_metadata_control_flow = nir_metadata_block_index |
3384 nir_metadata_dominance,
3385
3386 /** All metadata
3387 *
3388 * This includes all nir_metadata flags except not_properly_reset. Passes
3389 * which do not change the shader in any way should call
3390 *
3391 * nir_metadata_preserve(impl, nir_metadata_all);
3392 */
3393 nir_metadata_all = ~nir_metadata_not_properly_reset,
3394 } nir_metadata;
3395 MESA_DEFINE_CPP_ENUM_BITFIELD_OPERATORS(nir_metadata)
3396
3397 typedef struct {
3398 nir_cf_node cf_node;
3399
3400 /** pointer to the function of which this is an implementation */
3401 struct nir_function *function;
3402
3403 /**
3404 * For entrypoints, a pointer to a nir_function_impl which runs before
3405 * it, once per draw or dispatch, communicating via store_preamble and
3406 * load_preamble intrinsics. If NULL then there is no preamble.
3407 */
3408 struct nir_function *preamble;
3409
3410 /** list of nir_cf_node */
3411 struct exec_list body;
3412
3413 nir_block *end_block;
3414
3415 /** list for all local variables in the function */
3416 struct exec_list locals;
3417
3418 /** next available SSA value index */
3419 unsigned ssa_alloc;
3420
3421 /* total number of basic blocks, only valid when block_index_dirty = false */
3422 unsigned num_blocks;
3423
3424 /** True if this nir_function_impl uses structured control-flow
3425 *
3426 * Structured nir_function_impls have different validation rules.
3427 */
3428 bool structured;
3429
3430 nir_metadata valid_metadata;
3431 nir_variable_mode loop_analysis_indirect_mask;
3432 bool loop_analysis_force_unroll_sampler_indirect;
3433 } nir_function_impl;
3434
3435 #define nir_foreach_function_temp_variable(var, impl) \
3436 foreach_list_typed(nir_variable, var, node, &(impl)->locals)
3437
3438 #define nir_foreach_function_temp_variable_safe(var, impl) \
3439 foreach_list_typed_safe(nir_variable, var, node, &(impl)->locals)
3440
3441 ATTRIBUTE_RETURNS_NONNULL static inline nir_block *
nir_start_block(nir_function_impl * impl)3442 nir_start_block(nir_function_impl *impl)
3443 {
3444 return (nir_block *)impl->body.head_sentinel.next;
3445 }
3446
3447 ATTRIBUTE_RETURNS_NONNULL static inline nir_block *
nir_impl_last_block(nir_function_impl * impl)3448 nir_impl_last_block(nir_function_impl *impl)
3449 {
3450 return (nir_block *)impl->body.tail_sentinel.prev;
3451 }
3452
3453 static inline nir_cf_node *
nir_cf_node_next(nir_cf_node * node)3454 nir_cf_node_next(nir_cf_node *node)
3455 {
3456 struct exec_node *next = exec_node_get_next(&node->node);
3457 if (exec_node_is_tail_sentinel(next))
3458 return NULL;
3459 else
3460 return exec_node_data(nir_cf_node, next, node);
3461 }
3462
3463 static inline nir_cf_node *
nir_cf_node_prev(nir_cf_node * node)3464 nir_cf_node_prev(nir_cf_node *node)
3465 {
3466 struct exec_node *prev = exec_node_get_prev(&node->node);
3467 if (exec_node_is_head_sentinel(prev))
3468 return NULL;
3469 else
3470 return exec_node_data(nir_cf_node, prev, node);
3471 }
3472
3473 static inline bool
nir_cf_node_is_first(const nir_cf_node * node)3474 nir_cf_node_is_first(const nir_cf_node *node)
3475 {
3476 return exec_node_is_head_sentinel(node->node.prev);
3477 }
3478
3479 static inline bool
nir_cf_node_is_last(const nir_cf_node * node)3480 nir_cf_node_is_last(const nir_cf_node *node)
3481 {
3482 return exec_node_is_tail_sentinel(node->node.next);
3483 }
3484
NIR_DEFINE_CAST(nir_cf_node_as_block,nir_cf_node,nir_block,cf_node,type,nir_cf_node_block)3485 NIR_DEFINE_CAST(nir_cf_node_as_block, nir_cf_node, nir_block, cf_node,
3486 type, nir_cf_node_block)
3487 NIR_DEFINE_CAST(nir_cf_node_as_if, nir_cf_node, nir_if, cf_node,
3488 type, nir_cf_node_if)
3489 NIR_DEFINE_CAST(nir_cf_node_as_loop, nir_cf_node, nir_loop, cf_node,
3490 type, nir_cf_node_loop)
3491 NIR_DEFINE_CAST(nir_cf_node_as_function, nir_cf_node,
3492 nir_function_impl, cf_node, type, nir_cf_node_function)
3493
3494 static inline nir_block *
3495 nir_if_first_then_block(nir_if *if_stmt)
3496 {
3497 struct exec_node *head = exec_list_get_head(&if_stmt->then_list);
3498 return nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
3499 }
3500
3501 static inline nir_block *
nir_if_last_then_block(nir_if * if_stmt)3502 nir_if_last_then_block(nir_if *if_stmt)
3503 {
3504 struct exec_node *tail = exec_list_get_tail(&if_stmt->then_list);
3505 return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
3506 }
3507
3508 static inline nir_block *
nir_if_first_else_block(nir_if * if_stmt)3509 nir_if_first_else_block(nir_if *if_stmt)
3510 {
3511 struct exec_node *head = exec_list_get_head(&if_stmt->else_list);
3512 return nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
3513 }
3514
3515 static inline nir_block *
nir_if_last_else_block(nir_if * if_stmt)3516 nir_if_last_else_block(nir_if *if_stmt)
3517 {
3518 struct exec_node *tail = exec_list_get_tail(&if_stmt->else_list);
3519 return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
3520 }
3521
3522 static inline nir_block *
nir_loop_first_block(nir_loop * loop)3523 nir_loop_first_block(nir_loop *loop)
3524 {
3525 struct exec_node *head = exec_list_get_head(&loop->body);
3526 return nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
3527 }
3528
3529 static inline nir_block *
nir_loop_last_block(nir_loop * loop)3530 nir_loop_last_block(nir_loop *loop)
3531 {
3532 struct exec_node *tail = exec_list_get_tail(&loop->body);
3533 return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
3534 }
3535
3536 static inline bool
nir_loop_has_continue_construct(const nir_loop * loop)3537 nir_loop_has_continue_construct(const nir_loop *loop)
3538 {
3539 return !exec_list_is_empty(&loop->continue_list);
3540 }
3541
3542 static inline nir_block *
nir_loop_first_continue_block(nir_loop * loop)3543 nir_loop_first_continue_block(nir_loop *loop)
3544 {
3545 assert(nir_loop_has_continue_construct(loop));
3546 struct exec_node *head = exec_list_get_head(&loop->continue_list);
3547 return nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
3548 }
3549
3550 static inline nir_block *
nir_loop_last_continue_block(nir_loop * loop)3551 nir_loop_last_continue_block(nir_loop *loop)
3552 {
3553 assert(nir_loop_has_continue_construct(loop));
3554 struct exec_node *tail = exec_list_get_tail(&loop->continue_list);
3555 return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
3556 }
3557
3558 /**
3559 * Return the target block of a nir_jump_continue statement
3560 */
3561 static inline nir_block *
nir_loop_continue_target(nir_loop * loop)3562 nir_loop_continue_target(nir_loop *loop)
3563 {
3564 if (nir_loop_has_continue_construct(loop))
3565 return nir_loop_first_continue_block(loop);
3566 else
3567 return nir_loop_first_block(loop);
3568 }
3569
3570 /**
3571 * Return true if this list of cf_nodes contains a single empty block.
3572 */
3573 static inline bool
nir_cf_list_is_empty_block(struct exec_list * cf_list)3574 nir_cf_list_is_empty_block(struct exec_list *cf_list)
3575 {
3576 if (exec_list_is_singular(cf_list)) {
3577 struct exec_node *head = exec_list_get_head(cf_list);
3578 nir_block *block =
3579 nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
3580 return exec_list_is_empty(&block->instr_list);
3581 }
3582 return false;
3583 }
3584
3585 typedef struct {
3586 uint8_t num_components;
3587 uint8_t bit_size;
3588
3589 /* True if this parameter is a deref used for returning values */
3590 bool is_return;
3591
3592 bool implicit_conversion_prohibited;
3593
3594 /* True if this parameter is not divergent. This is inverted to make
3595 * parameters divergent by default unless explicitly specified
3596 * otherwise.
3597 */
3598 bool is_uniform;
3599
3600 nir_variable_mode mode;
3601
3602 /* Drivers may optionally stash flags here describing the parameter.
3603 * For example, this might encode whether the driver expects the value
3604 * to be uniform or divergent, if the driver handles divergent parameters
3605 * differently from uniform ones.
3606 *
3607 * NIR will preserve this value but does not interpret it in any way.
3608 */
3609 uint32_t driver_attributes;
3610
3611 /* The type of the function param */
3612 const struct glsl_type *type;
3613
3614 /* Name if known, null if unknown */
3615 const char *name;
3616 } nir_parameter;
3617
3618 typedef struct nir_function {
3619 struct exec_node node;
3620
3621 const char *name;
3622 struct nir_shader *shader;
3623
3624 unsigned num_params;
3625 nir_parameter *params;
3626
3627 /** The implementation of this function.
3628 *
3629 * If the function is only declared and not implemented, this is NULL.
3630 *
3631 * Unless setting to NULL or NIR_SERIALIZE_FUNC_HAS_IMPL, set with
3632 * nir_function_set_impl to maintain IR invariants.
3633 */
3634 nir_function_impl *impl;
3635
3636 /* Drivers may optionally stash flags here describing the function call.
3637 * For example, this might encode the ABI used for the call if a driver
3638 * supports multiple ABIs.
3639 *
3640 * NIR will preserve this value but does not interpret it in any way.
3641 */
3642 uint32_t driver_attributes;
3643
3644 bool is_entrypoint;
3645 /* from SPIR-V linkage, only for libraries */
3646 bool is_exported;
3647 bool is_preamble;
3648 /* from SPIR-V function control */
3649 bool should_inline;
3650 bool dont_inline; /* from SPIR-V */
3651
3652 /* Static workgroup size, if this is a kernel function in a library of OpenCL
3653 * kernels. Normally, the size in the shader info is used instead.
3654 */
3655 unsigned workgroup_size[3];
3656
3657 /**
3658 * Is this function a subroutine type declaration
3659 * e.g. subroutine void type1(float arg1);
3660 */
3661 bool is_subroutine;
3662
3663 /* Temporary function created to wrap global instructions before they can
3664 * be inlined into the main function.
3665 */
3666 bool is_tmp_globals_wrapper;
3667
3668 /**
3669 * Is this function associated to a subroutine type
3670 * e.g. subroutine (type1, type2) function_name { function_body };
3671 * would have num_subroutine_types 2,
3672 * and pointers to the type1 and type2 types.
3673 */
3674 int num_subroutine_types;
3675 const struct glsl_type **subroutine_types;
3676
3677 int subroutine_index;
3678
3679 /* A temporary for passes to use for storing flags. */
3680 uint32_t pass_flags;
3681 } nir_function;
3682
3683 typedef enum {
3684 nir_lower_imul64 = (1 << 0),
3685 nir_lower_isign64 = (1 << 1),
3686 /** Lower all int64 modulus and division opcodes */
3687 nir_lower_divmod64 = (1 << 2),
3688 /** Lower all 64-bit umul_high and imul_high opcodes */
3689 nir_lower_imul_high64 = (1 << 3),
3690 nir_lower_bcsel64 = (1 << 4),
3691 nir_lower_icmp64 = (1 << 5),
3692 nir_lower_iadd64 = (1 << 6),
3693 nir_lower_iabs64 = (1 << 7),
3694 nir_lower_ineg64 = (1 << 8),
3695 nir_lower_logic64 = (1 << 9),
3696 nir_lower_minmax64 = (1 << 10),
3697 nir_lower_shift64 = (1 << 11),
3698 nir_lower_imul_2x32_64 = (1 << 12),
3699 nir_lower_extract64 = (1 << 13),
3700 nir_lower_ufind_msb64 = (1 << 14),
3701 nir_lower_bit_count64 = (1 << 15),
3702 nir_lower_subgroup_shuffle64 = (1 << 16),
3703 nir_lower_scan_reduce_bitwise64 = (1 << 17),
3704 nir_lower_scan_reduce_iadd64 = (1 << 18),
3705 nir_lower_vote_ieq64 = (1 << 19),
3706 nir_lower_usub_sat64 = (1 << 20),
3707 nir_lower_iadd_sat64 = (1 << 21),
3708 nir_lower_find_lsb64 = (1 << 22),
3709 nir_lower_conv64 = (1 << 23),
3710 nir_lower_uadd_sat64 = (1 << 24),
3711 nir_lower_iadd3_64 = (1 << 25),
3712 } nir_lower_int64_options;
3713
3714 typedef enum {
3715 nir_lower_drcp = (1 << 0),
3716 nir_lower_dsqrt = (1 << 1),
3717 nir_lower_drsq = (1 << 2),
3718 nir_lower_dtrunc = (1 << 3),
3719 nir_lower_dfloor = (1 << 4),
3720 nir_lower_dceil = (1 << 5),
3721 nir_lower_dfract = (1 << 6),
3722 nir_lower_dround_even = (1 << 7),
3723 nir_lower_dmod = (1 << 8),
3724 nir_lower_dsub = (1 << 9),
3725 nir_lower_ddiv = (1 << 10),
3726 nir_lower_dsign = (1 << 11),
3727 nir_lower_dminmax = (1 << 12),
3728 nir_lower_dsat = (1 << 13),
3729 nir_lower_fp64_full_software = (1 << 14),
3730 } nir_lower_doubles_options;
3731
3732 typedef enum {
3733 nir_divergence_single_prim_per_subgroup = (1 << 0),
3734 nir_divergence_single_patch_per_tcs_subgroup = (1 << 1),
3735 nir_divergence_single_patch_per_tes_subgroup = (1 << 2),
3736 nir_divergence_view_index_uniform = (1 << 3),
3737 nir_divergence_single_frag_shading_rate_per_subgroup = (1 << 4),
3738 nir_divergence_multiple_workgroup_per_compute_subgroup = (1 << 5),
3739 nir_divergence_shader_record_ptr_uniform = (1 << 6),
3740 nir_divergence_uniform_load_tears = (1 << 7),
3741 /* If used, this allows phis for divergent merges with undef and a uniform source to be considered uniform */
3742 nir_divergence_ignore_undef_if_phi_srcs = (1 << 8),
3743 } nir_divergence_options;
3744
3745 typedef enum {
3746 /**
3747 * Whether a fragment shader can interpolate the same input multiple times
3748 * with different modes (smooth, noperspective) and locations (pixel,
3749 * centroid, sample, at_offset, at_sample), excluding the flat mode.
3750 *
3751 * This matches AMD GPU flexibility and limitations and is a superset of
3752 * the GL4 requirement that each input can be interpolated at its specified
3753 * location, and then also as centroid, at_offset, and at_sample.
3754 */
3755 nir_io_has_flexible_input_interpolation_except_flat = BITFIELD_BIT(0),
3756
3757 /**
3758 * nir_opt_varyings compacts (relocates) components of varyings by
3759 * rewriting their locations completely, effectively moving components of
3760 * varyings between slots. This option forces nir_opt_varyings to make
3761 * VARYING_SLOT_POS unused by moving its contents to VARn if the consumer
3762 * is not FS. If this option is not set and POS is unused, it moves
3763 * components of VARn to POS until it's fully used.
3764 */
3765 nir_io_dont_use_pos_for_non_fs_varyings = BITFIELD_BIT(1),
3766
3767 nir_io_16bit_input_output_support = BITFIELD_BIT(2),
3768
3769 /**
3770 * Implement mediump inputs and outputs as normal 32-bit IO.
3771 * Causes the mediump flag to be not set for IO semantics, essentially
3772 * destroying any mediump-related IO information in the shader.
3773 */
3774 nir_io_mediump_is_32bit = BITFIELD_BIT(3),
3775
3776 /**
3777 * Whether nir_opt_vectorize_io should ignore FS inputs.
3778 */
3779 nir_io_prefer_scalar_fs_inputs = BITFIELD_BIT(4),
3780
3781 /**
3782 * Whether interpolated fragment shader vec4 slots can use load_input for
3783 * a subset of its components to skip interpolation for those components.
3784 * The result of such load_input is a value from a random (not necessarily
3785 * provoking) vertex. If a value from the provoking vertex is required,
3786 * the vec4 slot should have no load_interpolated_input instructions.
3787 *
3788 * This exposes the AMD capability that allows packing flat inputs with
3789 * interpolated inputs in a limited number of cases. Normally, flat
3790 * components must be in a separate vec4 slot to get the value from
3791 * the provoking vertex. If the compiler can prove that all per-vertex
3792 * values are equal (convergent, i.e. the provoking vertex doesn't matter),
3793 * it can put such flat components into any interpolated vec4 slot.
3794 *
3795 * It should also be set if the hw can mix flat and interpolated components
3796 * in the same vec4 slot.
3797 *
3798 * This causes nir_opt_varyings to skip interpolation for all varyings
3799 * that are convergent, and enables better compaction and inter-shader code
3800 * motion for convergent varyings.
3801 */
3802 nir_io_mix_convergent_flat_with_interpolated = BITFIELD_BIT(5),
3803
3804 /**
3805 * Whether src_type and dest_type of IO intrinsics are irrelevant and
3806 * should be ignored by nir_opt_vectorize_io. All drivers that always treat
3807 * load_input and store_output as untyped and load_interpolated_input as
3808 * float##bit_size should set this.
3809 */
3810 nir_io_vectorizer_ignores_types = BITFIELD_BIT(6),
3811
3812 /**
3813 * Whether nir_opt_varyings should never promote convergent FS inputs
3814 * to flat.
3815 */
3816 nir_io_always_interpolate_convergent_fs_inputs = BITFIELD_BIT(7),
3817
3818 /**
3819 * Whether the first assigned color channel component should be equal to
3820 * the first unused VARn component.
3821 *
3822 * For example, if the first unused VARn channel is VAR0.z, color channels
3823 * are assigned in this order:
3824 * COL0.z, COL0.w, COL0.x, COL0.y, COL1.z, COL1.w, COL1.x, COL1.y
3825 *
3826 * This allows certain drivers to merge outputs if each output sets
3827 * different components, for example 2 outputs writing VAR0.xy and COL0.z
3828 * will only use 1 HW output.
3829 */
3830 nir_io_compaction_rotates_color_channels = BITFIELD_BIT(8),
3831
3832 /* Options affecting the GLSL compiler or Gallium are below. */
3833
3834 /**
3835 * Lower load_deref/store_deref to load_input/store_output/etc. intrinsics.
3836 * This is only affects GLSL compilation and Gallium.
3837 */
3838 nir_io_has_intrinsics = BITFIELD_BIT(16),
3839
3840 /**
3841 * Don't run nir_opt_varyings and nir_opt_vectorize_io.
3842 *
3843 * This option is deprecated and is a hack. DO NOT USE.
3844 * Use MESA_GLSL_DISABLE_IO_OPT=1 instead.
3845 */
3846 nir_io_dont_optimize = BITFIELD_BIT(17),
3847
3848 /**
3849 * Whether clip and cull distance arrays should be separate. If this is not
3850 * set, cull distances will be moved into VARYING_SLOT_CLIP_DISTn after clip
3851 * distances, and shader_info::clip_distance_array_size will be the index
3852 * of the first cull distance. nir_lower_clip_cull_distance_arrays does
3853 * that.
3854 */
3855 nir_io_separate_clip_cull_distance_arrays = BITFIELD_BIT(18),
3856 } nir_io_options;
3857
3858 typedef enum {
3859 nir_lower_packing_op_pack_64_2x32,
3860 nir_lower_packing_op_unpack_64_2x32,
3861 nir_lower_packing_op_pack_64_4x16,
3862 nir_lower_packing_op_unpack_64_4x16,
3863 nir_lower_packing_op_pack_32_2x16,
3864 nir_lower_packing_op_unpack_32_2x16,
3865 nir_lower_packing_op_pack_32_4x8,
3866 nir_lower_packing_op_unpack_32_4x8,
3867 nir_lower_packing_num_ops,
3868 } nir_lower_packing_op;
3869
3870 /** An instruction filtering callback
3871 *
3872 * Returns true if the instruction should be processed and false otherwise.
3873 */
3874 typedef bool (*nir_instr_filter_cb)(const nir_instr *, const void *);
3875
3876 /* Like nir_instr_filter_cb but specialized to intrinsics */
3877 typedef bool (*nir_intrin_filter_cb)(const nir_intrinsic_instr *, const void *);
3878
3879 /** A vectorization width callback
3880 *
3881 * Returns the maximum vectorization width per instruction.
3882 * 0, if the instruction must not be modified.
3883 *
3884 * The vectorization width must be a power of 2.
3885 */
3886 typedef uint8_t (*nir_vectorize_cb)(const nir_instr *, const void *);
3887
3888 typedef struct nir_shader_compiler_options {
3889 bool lower_fdiv;
3890 bool lower_ffma16;
3891 bool lower_ffma32;
3892 bool lower_ffma64;
3893 bool fuse_ffma16;
3894 bool fuse_ffma32;
3895 bool fuse_ffma64;
3896 bool lower_flrp16;
3897 bool lower_flrp32;
3898 /** Lowers flrp when it does not support doubles */
3899 bool lower_flrp64;
3900 bool lower_fpow;
3901 bool lower_fsat;
3902 bool lower_fsqrt;
3903 bool lower_sincos;
3904 bool lower_fmod;
3905 /** Lowers ibitfield_extract/ubitfield_extract. */
3906 bool lower_bitfield_extract;
3907 /** Lowers bitfield_insert. */
3908 bool lower_bitfield_insert;
3909 /** Lowers bitfield_reverse to shifts. */
3910 bool lower_bitfield_reverse;
3911 /** Lowers bit_count to shifts. */
3912 bool lower_bit_count;
3913 /** Lowers ifind_msb. */
3914 bool lower_ifind_msb;
3915 /** Lowers ufind_msb. */
3916 bool lower_ufind_msb;
3917 /** Lowers find_lsb to ufind_msb and logic ops */
3918 bool lower_find_lsb;
3919 bool lower_uadd_carry;
3920 bool lower_usub_borrow;
3921 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
3922 bool lower_mul_high;
3923 /** lowers fneg to fmul(x, -1.0). Driver must call nir_opt_algebraic_late() */
3924 bool lower_fneg;
3925 /** lowers ineg to isub. Driver must call nir_opt_algebraic_late(). */
3926 bool lower_ineg;
3927 /** lowers fisnormal to alu ops. */
3928 bool lower_fisnormal;
3929
3930 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fneu} + b2f: */
3931 bool lower_scmp;
3932
3933 /* lower b/fall_equalN/b/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
3934 bool lower_vector_cmp;
3935
3936 /** enable rules to avoid bit ops */
3937 bool lower_bitops;
3938
3939 /** enables rules to lower isign to imin+imax */
3940 bool lower_isign;
3941
3942 /** enables rules to lower fsign to fsub and flt */
3943 bool lower_fsign;
3944
3945 /** enables rules to lower iabs to ineg+imax */
3946 bool lower_iabs;
3947
3948 /** enable rules that avoid generating umax from signed integer ops */
3949 bool lower_umax;
3950
3951 /** enable rules that avoid generating umin from signed integer ops */
3952 bool lower_umin;
3953
3954 /* lower fmin/fmax with signed zero preserve to fmin/fmax with
3955 * no_signed_zero, for backends whose fmin/fmax implementations do not
3956 * implement IEEE-754-2019 semantics for signed zero.
3957 */
3958 bool lower_fminmax_signed_zero;
3959
3960 /* lower fdph to fdot4 */
3961 bool lower_fdph;
3962
3963 /** lower fdot to fmul and fsum/fadd. */
3964 bool lower_fdot;
3965
3966 /* Does the native fdot instruction replicate its result for four
3967 * components? If so, then opt_algebraic_late will turn all fdotN
3968 * instructions into fdotN_replicated instructions.
3969 */
3970 bool fdot_replicates;
3971
3972 /** lowers ffloor to fsub+ffract: */
3973 bool lower_ffloor;
3974
3975 /** lowers ffract to fsub+ffloor: */
3976 bool lower_ffract;
3977
3978 /** lowers fceil to fneg+ffloor+fneg: */
3979 bool lower_fceil;
3980
3981 bool lower_ftrunc;
3982
3983 /** Lowers fround_even to ffract+feq+csel.
3984 *
3985 * Not correct in that it doesn't correctly handle the "_even" part of the
3986 * rounding, but good enough for DX9 array indexing handling on DX9-class
3987 * hardware.
3988 */
3989 bool lower_fround_even;
3990
3991 bool lower_ldexp;
3992
3993 bool lower_pack_half_2x16;
3994 bool lower_pack_unorm_2x16;
3995 bool lower_pack_snorm_2x16;
3996 bool lower_pack_unorm_4x8;
3997 bool lower_pack_snorm_4x8;
3998 bool lower_pack_64_2x32;
3999 bool lower_pack_64_4x16;
4000 bool lower_pack_32_2x16;
4001 bool lower_pack_64_2x32_split;
4002 bool lower_pack_32_2x16_split;
4003 bool lower_unpack_half_2x16;
4004 bool lower_unpack_unorm_2x16;
4005 bool lower_unpack_snorm_2x16;
4006 bool lower_unpack_unorm_4x8;
4007 bool lower_unpack_snorm_4x8;
4008 bool lower_unpack_64_2x32_split;
4009 bool lower_unpack_32_2x16_split;
4010
4011 bool lower_pack_split;
4012
4013 bool lower_extract_byte;
4014 bool lower_extract_word;
4015 bool lower_insert_byte;
4016 bool lower_insert_word;
4017
4018 /* TODO: this flag is potentially useless, remove? */
4019 bool lower_all_io_to_temps;
4020
4021 /* Indicates that the driver only has zero-based vertex id */
4022 bool vertex_id_zero_based;
4023
4024 /**
4025 * If enabled, gl_BaseVertex will be lowered as:
4026 * is_indexed_draw (~0/0) & firstvertex
4027 */
4028 bool lower_base_vertex;
4029
4030 /**
4031 * If enabled, gl_HelperInvocation will be lowered as:
4032 *
4033 * !((1 << sample_id) & sample_mask_in))
4034 *
4035 * This depends on some possibly hw implementation details, which may
4036 * not be true for all hw. In particular that the FS is only executed
4037 * for covered samples or for helper invocations. So, do not blindly
4038 * enable this option.
4039 *
4040 * Note: See also issue #22 in ARB_shader_image_load_store
4041 */
4042 bool lower_helper_invocation;
4043
4044 /**
4045 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
4046 *
4047 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
4048 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
4049 */
4050 bool optimize_sample_mask_in;
4051
4052 /**
4053 * Optimize load_front_face ? a : -a to load_front_face_fsign * a
4054 */
4055 bool optimize_load_front_face_fsign;
4056
4057 /**
4058 * Optimize boolean reductions of quad broadcasts. This should only be enabled if
4059 * nir_intrinsic_reduce supports INCLUDE_HELPERS.
4060 */
4061 bool optimize_quad_vote_to_reduce;
4062
4063 bool lower_cs_local_index_to_id;
4064 bool lower_cs_local_id_to_index;
4065
4066 /* Prevents lowering global_invocation_id to be in terms of workgroup_id */
4067 bool has_cs_global_id;
4068
4069 bool lower_device_index_to_zero;
4070
4071 /* Set if nir_lower_pntc_ytransform() should invert gl_PointCoord.
4072 * Either when frame buffer is flipped or GL_POINT_SPRITE_COORD_ORIGIN
4073 * is GL_LOWER_LEFT.
4074 */
4075 bool lower_wpos_pntc;
4076
4077 /**
4078 * Set if nir_op_[iu]hadd and nir_op_[iu]rhadd instructions should be
4079 * lowered to simple arithmetic.
4080 *
4081 * If this flag is set, the lowering will be applied to all bit-sizes of
4082 * these instructions.
4083 *
4084 * :c:member:`lower_hadd64`
4085 */
4086 bool lower_hadd;
4087
4088 /**
4089 * Set if only 64-bit nir_op_[iu]hadd and nir_op_[iu]rhadd instructions
4090 * should be lowered to simple arithmetic.
4091 *
4092 * If this flag is set, the lowering will be applied to only 64-bit
4093 * versions of these instructions.
4094 *
4095 * :c:member:`lower_hadd`
4096 */
4097 bool lower_hadd64;
4098
4099 /**
4100 * Set if nir_op_uadd_sat should be lowered to simple arithmetic.
4101 *
4102 * If this flag is set, the lowering will be applied to all bit-sizes of
4103 * these instructions.
4104 */
4105 bool lower_uadd_sat;
4106
4107 /**
4108 * Set if nir_op_usub_sat should be lowered to simple arithmetic.
4109 *
4110 * If this flag is set, the lowering will be applied to all bit-sizes of
4111 * these instructions.
4112 */
4113 bool lower_usub_sat;
4114
4115 /**
4116 * Set if nir_op_iadd_sat and nir_op_isub_sat should be lowered to simple
4117 * arithmetic.
4118 *
4119 * If this flag is set, the lowering will be applied to all bit-sizes of
4120 * these instructions.
4121 */
4122 bool lower_iadd_sat;
4123
4124 /**
4125 * Set if imul_32x16 and umul_32x16 should be lowered to simple
4126 * arithmetic.
4127 */
4128 bool lower_mul_32x16;
4129
4130 bool vectorize_tess_levels;
4131 bool lower_to_scalar;
4132 nir_instr_filter_cb lower_to_scalar_filter;
4133
4134 /**
4135 * Disables potentially harmful algebraic transformations for architectures
4136 * with SIMD-within-a-register semantics.
4137 *
4138 * Note, to actually vectorize 16bit instructions, use nir_opt_vectorize()
4139 * with a suitable callback function.
4140 */
4141 bool vectorize_vec2_16bit;
4142
4143 /**
4144 * Should the linker unify inputs_read/outputs_written between adjacent
4145 * shader stages which are linked into a single program?
4146 */
4147 bool unify_interfaces;
4148
4149 /**
4150 * Whether nir_lower_io() will lower interpolateAt functions to
4151 * load_interpolated_input intrinsics.
4152 *
4153 * Unlike nir_lower_io_use_interpolated_input_intrinsics this will only
4154 * lower these functions and leave input load intrinsics untouched.
4155 */
4156 bool lower_interpolate_at;
4157
4158 /* Lowers when 32x32->64 bit multiplication is not supported */
4159 bool lower_mul_2x32_64;
4160
4161 /* Indicates that urol and uror are supported */
4162 bool has_rotate8;
4163 bool has_rotate16;
4164 bool has_rotate32;
4165
4166 /** Backend supports shfr */
4167 bool has_shfr32;
4168
4169 /** Backend supports ternary addition */
4170 bool has_iadd3;
4171
4172 /**
4173 * Backend supports amul and would like them generated whenever
4174 * possible. This is stronger than has_imul24 for amul, but does not imply
4175 * support for imul24.
4176 */
4177 bool has_amul;
4178
4179 /**
4180 * Backend supports imul24, and would like to use it (when possible)
4181 * for address/offset calculation. If true, driver should call
4182 * nir_lower_amul(). (If not set, amul will automatically be lowered
4183 * to imul.)
4184 */
4185 bool has_imul24;
4186
4187 /** Backend supports umul24, if not set umul24 will automatically be lowered
4188 * to imul with masked inputs */
4189 bool has_umul24;
4190
4191 /** Backend supports 32-bit imad */
4192 bool has_imad32;
4193
4194 /** Backend supports umad24, if not set umad24 will automatically be lowered
4195 * to imul with masked inputs and iadd */
4196 bool has_umad24;
4197
4198 /* Backend supports fused compare against zero and csel */
4199 bool has_fused_comp_and_csel;
4200 /* Backend supports fused int eq/ne against zero and csel. */
4201 bool has_icsel_eqz64;
4202 bool has_icsel_eqz32;
4203 bool has_icsel_eqz16;
4204
4205 /* Backend supports fneo, fequ, fltu, fgeu. */
4206 bool has_fneo_fcmpu;
4207
4208 /* Backend supports ford and funord. */
4209 bool has_ford_funord;
4210
4211 /** Backend supports fsub, if not set fsub will automatically be lowered to
4212 * fadd(x, fneg(y)). If true, driver should call nir_opt_algebraic_late(). */
4213 bool has_fsub;
4214
4215 /** Backend supports isub, if not set isub will automatically be lowered to
4216 * iadd(x, ineg(y)). If true, driver should call nir_opt_algebraic_late(). */
4217 bool has_isub;
4218
4219 /** Backend supports pack_32_4x8 or pack_32_4x8_split. */
4220 bool has_pack_32_4x8;
4221
4222 /** Backend supports nir_load_texture_scale and prefers it over txs for nir
4223 * lowerings. */
4224 bool has_texture_scaling;
4225
4226 /** Backend supports sdot_4x8_iadd. */
4227 bool has_sdot_4x8;
4228
4229 /** Backend supports udot_4x8_uadd. */
4230 bool has_udot_4x8;
4231
4232 /** Backend supports sudot_4x8_iadd. */
4233 bool has_sudot_4x8;
4234
4235 /** Backend supports sdot_4x8_iadd_sat. */
4236 bool has_sdot_4x8_sat;
4237
4238 /** Backend supports udot_4x8_uadd_sat. */
4239 bool has_udot_4x8_sat;
4240
4241 /** Backend supports sudot_4x8_iadd_sat. */
4242 bool has_sudot_4x8_sat;
4243
4244 /** Backend supports sdot_2x16 and udot_2x16 opcodes. */
4245 bool has_dot_2x16;
4246
4247 /** Backend supports fmulz (and ffmaz if lower_ffma32=false) */
4248 bool has_fmulz;
4249
4250 /**
4251 * Backend supports fmulz (and ffmaz if lower_ffma32=false) but only if
4252 * FLOAT_CONTROLS_DENORM_PRESERVE_FP32 is not set
4253 */
4254 bool has_fmulz_no_denorms;
4255
4256 /** Backend supports 32bit ufind_msb_rev and ifind_msb_rev. */
4257 bool has_find_msb_rev;
4258
4259 /** Backend supports pack_half_2x16_rtz_split. */
4260 bool has_pack_half_2x16_rtz;
4261
4262 /** Backend supports bitz/bitnz. */
4263 bool has_bit_test;
4264
4265 /** Backend supports ubfe/ibfe. */
4266 bool has_bfe;
4267
4268 /** Backend supports bfm. */
4269 bool has_bfm;
4270
4271 /** Backend supports bfi. */
4272 bool has_bfi;
4273
4274 /** Backend supports bitfield_select. */
4275 bool has_bitfield_select;
4276
4277 /** Backend supports uclz. */
4278 bool has_uclz;
4279
4280 /** Backend support msad_u4x8. */
4281 bool has_msad;
4282
4283 /**
4284 * Is this the Intel vec4 backend?
4285 *
4286 * Used to inhibit algebraic optimizations that are known to be harmful on
4287 * the Intel vec4 backend. This is generally applicable to any
4288 * optimization that might cause more immediate values to be used in
4289 * 3-source (e.g., ffma and flrp) instructions.
4290 */
4291 bool intel_vec4;
4292
4293 /**
4294 * For most Intel GPUs, all ternary operations such as FMA and BFE cannot
4295 * have immediates, so two to three instructions may eventually be needed.
4296 */
4297 bool avoid_ternary_with_two_constants;
4298
4299 /** Whether 8-bit ALU is supported. */
4300 bool support_8bit_alu;
4301
4302 /** Whether 16-bit ALU is supported. */
4303 bool support_16bit_alu;
4304
4305 unsigned max_unroll_iterations;
4306 unsigned max_unroll_iterations_aggressive;
4307 unsigned max_unroll_iterations_fp64;
4308
4309 bool lower_uniforms_to_ubo;
4310
4311 /* Specifies if indirect sampler array access will trigger forced loop
4312 * unrolling.
4313 */
4314 bool force_indirect_unrolling_sampler;
4315
4316 /* Some older drivers don't support GLSL versions with the concept of flat
4317 * varyings and also don't support integers. This setting helps us avoid
4318 * marking varyings as flat and potentially having them changed to ints via
4319 * varying packing.
4320 */
4321 bool no_integers;
4322
4323 /**
4324 * Specifies which type of indirectly accessed variables should force
4325 * loop unrolling.
4326 */
4327 nir_variable_mode force_indirect_unrolling;
4328
4329 bool driver_functions;
4330
4331 /**
4332 * If true, the driver will call nir_lower_int64 itself and the frontend
4333 * should not do so. This may enable better optimization around address
4334 * modes.
4335 */
4336 bool late_lower_int64;
4337 nir_lower_int64_options lower_int64_options;
4338 nir_lower_doubles_options lower_doubles_options;
4339 nir_divergence_options divergence_analysis_options;
4340
4341 /**
4342 * The masks of shader stages that support indirect indexing with
4343 * load_input and store_output intrinsics. It's used by
4344 * nir_lower_io_passes.
4345 */
4346 uint8_t support_indirect_inputs;
4347 uint8_t support_indirect_outputs;
4348
4349 /** store the variable offset into the instrinsic range_base instead
4350 * of adding it to the image index.
4351 */
4352 bool lower_image_offset_to_range_base;
4353
4354 /** store the variable offset into the instrinsic range_base instead
4355 * of adding it to the atomic source
4356 */
4357 bool lower_atomic_offset_to_range_base;
4358
4359 /** Don't convert medium-precision casts (e.g. f2fmp) into concrete
4360 * type casts (e.g. f2f16).
4361 */
4362 bool preserve_mediump;
4363
4364 /** lowers fquantize2f16 to alu ops. */
4365 bool lower_fquantize2f16;
4366
4367 /** Lower f2f16 to f2f16_rtz when execution mode is not rtne. */
4368 bool force_f2f16_rtz;
4369
4370 /** Lower VARYING_SLOT_LAYER in FS to SYSTEM_VALUE_LAYER_ID. */
4371 bool lower_layer_fs_input_to_sysval;
4372
4373 /** clip/cull distance and tess level arrays use compact semantics */
4374 bool compact_arrays;
4375
4376 /**
4377 * Whether discard gets emitted as nir_intrinsic_demote.
4378 * Otherwise, nir_intrinsic_terminate is being used.
4379 */
4380 bool discard_is_demote;
4381
4382 /**
4383 * Whether the new-style derivative intrinsics are supported. If false,
4384 * legacy ALU derivative ops will be emitted. This transitional option will
4385 * be removed once all drivers are converted to derivative intrinsics.
4386 */
4387 bool has_ddx_intrinsics;
4388
4389 /** Whether derivative intrinsics must be scalarized. */
4390 bool scalarize_ddx;
4391
4392 /**
4393 * Assign a range of driver locations to per-view outputs, with unique
4394 * slots for each view. If unset, per-view outputs will be treated
4395 * similarly to other arrayed IO, and only slots for one view will be
4396 * assigned. Regardless of this setting, per-view outputs are only assigned
4397 * slots for one value in var->data.location.
4398 */
4399 bool per_view_unique_driver_locations;
4400
4401 /**
4402 * Emit nir_intrinsic_store_per_view_output with compacted view indices
4403 * rather than absolute view indices. When using compacted indices, the Nth
4404 * index refers to the Nth enabled view, not the Nth absolute view. For
4405 * example, with view mask 0b1010, compacted index 0 is absolute index 1,
4406 * and compacted index 1 is absolute index 3. Note that compacted view
4407 * indices do not correspond directly to gl_ViewIndex.
4408 *
4409 * If compact_view_index is unset, per-view indices must be constant before
4410 * nir_lower_io. This can be guaranteed by calling nir_lower_io_temporaries
4411 * first.
4412 */
4413 bool compact_view_index;
4414
4415 /** Options determining lowering and behavior of inputs and outputs. */
4416 nir_io_options io_options;
4417
4418 /**
4419 * Bit mask of nir_lower_packing_op to skip lowering some nir ops in
4420 * nir_lower_packing().
4421 */
4422 unsigned skip_lower_packing_ops;
4423
4424 /** Driver callback where drivers can define how to lower mediump.
4425 * Used by nir_lower_io_passes.
4426 */
4427 void (*lower_mediump_io)(struct nir_shader *nir);
4428
4429 /**
4430 * Return the maximum cost of an expression that's written to a shader
4431 * output that can be moved into the next shader to remove that output.
4432 *
4433 * Currently only uniform expressions are moved. A uniform expression is
4434 * any ALU expression sourcing only constants, uniforms, and UBO loads.
4435 *
4436 * Set to NULL or return 0 if you only want to propagate constants from
4437 * outputs to inputs.
4438 *
4439 * Drivers can set the maximum cost based on the types of consecutive
4440 * shaders or shader SHA1s.
4441 *
4442 * Drivers should also set "varying_estimate_instr_cost".
4443 */
4444 unsigned (*varying_expression_max_cost)(struct nir_shader *consumer,
4445 struct nir_shader *producer);
4446
4447 /**
4448 * Return the cost of an instruction that could be moved into the next
4449 * shader. If the cost of all instructions in an expression is <=
4450 * varying_expression_max_cost(), the instruction is moved.
4451 *
4452 * When this callback isn't set, nir_opt_varyings uses its own version.
4453 */
4454 unsigned (*varying_estimate_instr_cost)(struct nir_instr *instr);
4455
4456 /**
4457 * When the varying_expression_max_cost callback isn't set, this specifies
4458 * the maximum cost of a uniform expression that is allowed to be moved
4459 * from output stores into the next shader stage to eliminate those output
4460 * stores and corresponding inputs.
4461 *
4462 * 0 only allows propagating constants written to output stores to
4463 * the next shader.
4464 *
4465 * At least 2 is required for moving a uniform stored in an output into
4466 * the next shader according to default_varying_estimate_instr_cost.
4467 */
4468 unsigned max_varying_expression_cost;
4469 } nir_shader_compiler_options;
4470
4471 typedef struct nir_shader {
4472 gc_ctx *gctx;
4473
4474 /** list of uniforms (nir_variable) */
4475 struct exec_list variables;
4476
4477 /** Set of driver-specific options for the shader.
4478 *
4479 * The memory for the options is expected to be kept in a single static
4480 * copy by the driver.
4481 */
4482 const struct nir_shader_compiler_options *options;
4483
4484 /** Various bits of compile-time information about a given shader */
4485 struct shader_info info;
4486
4487 /** list of nir_function */
4488 struct exec_list functions;
4489
4490 /**
4491 * The size of the variable space for load_input_*, load_uniform_*, etc.
4492 * intrinsics. This is in back-end specific units which is likely one of
4493 * bytes, dwords, or vec4s depending on context and back-end.
4494 */
4495 unsigned num_inputs, num_uniforms, num_outputs;
4496
4497 /** Size in bytes of required implicitly bound global memory */
4498 unsigned global_mem_size;
4499
4500 /** Size in bytes of required scratch space */
4501 unsigned scratch_size;
4502
4503 /** Constant data associated with this shader.
4504 *
4505 * Constant data is loaded through load_constant intrinsics (as compared to
4506 * the NIR load_const instructions which have the constant value inlined
4507 * into them). This is usually generated by nir_opt_large_constants (so
4508 * shaders don't have to load_const into a temporary array when they want
4509 * to indirect on a const array).
4510 */
4511 void *constant_data;
4512 /** Size of the constant data associated with the shader, in bytes */
4513 unsigned constant_data_size;
4514
4515 struct nir_xfb_info *xfb_info;
4516
4517 unsigned printf_info_count;
4518 u_printf_info *printf_info;
4519 } nir_shader;
4520
4521 #define nir_foreach_function(func, shader) \
4522 foreach_list_typed(nir_function, func, node, &(shader)->functions)
4523
4524 #define nir_foreach_function_safe(func, shader) \
4525 foreach_list_typed_safe(nir_function, func, node, &(shader)->functions)
4526
4527 #define nir_foreach_entrypoint(func, lib) \
4528 nir_foreach_function(func, lib) \
4529 if (func->is_entrypoint)
4530
4531 #define nir_foreach_entrypoint_safe(func, lib) \
4532 nir_foreach_function_safe(func, lib) \
4533 if (func->is_entrypoint)
4534
4535 static inline nir_function *
nir_foreach_function_with_impl_first(const nir_shader * shader)4536 nir_foreach_function_with_impl_first(const nir_shader *shader)
4537 {
4538 foreach_list_typed(nir_function, func, node, &shader->functions) {
4539 if (func->impl != NULL)
4540 return func;
4541 }
4542
4543 return NULL;
4544 }
4545
4546 static inline nir_function_impl *
nir_foreach_function_with_impl_next(nir_function ** it)4547 nir_foreach_function_with_impl_next(nir_function **it)
4548 {
4549 foreach_list_typed_from(nir_function, func, node, _, (*it)->node.next) {
4550 if (func->impl != NULL) {
4551 *it = func;
4552 return func->impl;
4553 }
4554 }
4555
4556 return NULL;
4557 }
4558
4559 #define nir_foreach_function_with_impl(it, impl_it, shader) \
4560 for (nir_function *it = nir_foreach_function_with_impl_first(shader); \
4561 it != NULL; \
4562 it = NULL) \
4563 \
4564 for (nir_function_impl *impl_it = it->impl; \
4565 impl_it != NULL; \
4566 impl_it = nir_foreach_function_with_impl_next(&it))
4567
4568 /* Equivalent to
4569 *
4570 * nir_foreach_function(func, shader) {
4571 * if (func->impl != NULL) {
4572 * ...
4573 * }
4574 * }
4575 *
4576 * Carefully written to ensure break/continue work in the user code.
4577 */
4578
4579 #define nir_foreach_function_impl(it, shader) \
4580 nir_foreach_function_with_impl(_func_##it, it, shader)
4581
4582 static inline nir_function_impl *
nir_shader_get_entrypoint(const nir_shader * shader)4583 nir_shader_get_entrypoint(const nir_shader *shader)
4584 {
4585 nir_function *func = NULL;
4586
4587 nir_foreach_function(function, shader) {
4588 assert(func == NULL);
4589 if (function->is_entrypoint) {
4590 func = function;
4591 #ifndef NDEBUG
4592 break;
4593 #endif
4594 }
4595 }
4596
4597 if (!func)
4598 return NULL;
4599
4600 assert(func->num_params == 0);
4601 assert(func->impl);
4602 return func->impl;
4603 }
4604
4605 static inline nir_function *
nir_shader_get_function_for_name(const nir_shader * shader,const char * name)4606 nir_shader_get_function_for_name(const nir_shader *shader, const char *name)
4607 {
4608 nir_foreach_function(func, shader) {
4609 if (func->name && strcmp(func->name, name) == 0)
4610 return func;
4611 }
4612
4613 return NULL;
4614 }
4615
4616 /*
4617 * After all functions are forcibly inlined, these passes remove redundant
4618 * functions from a shader and library respectively.
4619 */
4620 void nir_remove_non_entrypoints(nir_shader *shader);
4621 void nir_remove_non_exported(nir_shader *shader);
4622 void nir_remove_entrypoints(nir_shader *shader);
4623 void nir_fixup_is_exported(nir_shader *shader);
4624
4625 nir_shader *nir_shader_create(void *mem_ctx,
4626 gl_shader_stage stage,
4627 const nir_shader_compiler_options *options,
4628 shader_info *si);
4629
4630 /** Adds a variable to the appropriate list in nir_shader */
4631 void nir_shader_add_variable(nir_shader *shader, nir_variable *var);
4632
4633 static inline void
nir_function_impl_add_variable(nir_function_impl * impl,nir_variable * var)4634 nir_function_impl_add_variable(nir_function_impl *impl, nir_variable *var)
4635 {
4636 assert(var->data.mode == nir_var_function_temp);
4637 exec_list_push_tail(&impl->locals, &var->node);
4638 }
4639
4640 /** creates a variable, sets a few defaults, and adds it to the list */
4641 nir_variable *nir_variable_create(nir_shader *shader,
4642 nir_variable_mode mode,
4643 const struct glsl_type *type,
4644 const char *name);
4645 /** creates a local variable and adds it to the list */
4646 nir_variable *nir_local_variable_create(nir_function_impl *impl,
4647 const struct glsl_type *type,
4648 const char *name);
4649
4650 /** Creates a uniform builtin state variable. */
4651 nir_variable *
4652 nir_state_variable_create(nir_shader *shader,
4653 const struct glsl_type *type,
4654 const char *name,
4655 const gl_state_index16 tokens[STATE_LENGTH]);
4656
4657 /* Gets the variable for the given mode and location, creating it (with the given
4658 * type) if necessary.
4659 */
4660 nir_variable *
4661 nir_get_variable_with_location(nir_shader *shader, nir_variable_mode mode, int location,
4662 const struct glsl_type *type);
4663
4664 /* Creates a variable for the given mode and location.
4665 */
4666 nir_variable *
4667 nir_create_variable_with_location(nir_shader *shader, nir_variable_mode mode, int location,
4668 const struct glsl_type *type);
4669
4670 nir_variable *nir_find_variable_with_location(nir_shader *shader,
4671 nir_variable_mode mode,
4672 unsigned location);
4673
4674 nir_variable *nir_find_variable_with_driver_location(nir_shader *shader,
4675 nir_variable_mode mode,
4676 unsigned location);
4677
4678 nir_variable *nir_find_state_variable(nir_shader *s,
4679 gl_state_index16 tokens[STATE_LENGTH]);
4680
4681 nir_variable *nir_find_sampler_variable_with_tex_index(nir_shader *shader,
4682 unsigned texture_index);
4683
4684 void nir_sort_variables_with_modes(nir_shader *shader,
4685 int (*compar)(const nir_variable *,
4686 const nir_variable *),
4687 nir_variable_mode modes);
4688
4689 /** creates a function and adds it to the shader's list of functions */
4690 nir_function *nir_function_create(nir_shader *shader, const char *name);
4691
4692 static inline void
nir_function_set_impl(nir_function * func,nir_function_impl * impl)4693 nir_function_set_impl(nir_function *func, nir_function_impl *impl)
4694 {
4695 func->impl = impl;
4696 impl->function = func;
4697 }
4698
4699 nir_function_impl *nir_function_impl_create(nir_function *func);
4700 /** creates a function_impl that isn't tied to any particular function */
4701 nir_function_impl *nir_function_impl_create_bare(nir_shader *shader);
4702
4703 nir_block *nir_block_create(nir_shader *shader);
4704 nir_if *nir_if_create(nir_shader *shader);
4705 nir_loop *nir_loop_create(nir_shader *shader);
4706
4707 nir_function_impl *nir_cf_node_get_function(nir_cf_node *node);
4708
4709 /** requests that the given pieces of metadata be generated */
4710 void nir_metadata_require(nir_function_impl *impl, nir_metadata required, ...);
4711 /** dirties all but the preserved metadata */
4712 void nir_metadata_preserve(nir_function_impl *impl, nir_metadata preserved);
4713 /** Preserves all metadata for the given shader */
4714 void nir_shader_preserve_all_metadata(nir_shader *shader);
4715
4716 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
4717 nir_alu_instr *nir_alu_instr_create(nir_shader *shader, nir_op op);
4718
4719 nir_deref_instr *nir_deref_instr_create(nir_shader *shader,
4720 nir_deref_type deref_type);
4721
4722 nir_jump_instr *nir_jump_instr_create(nir_shader *shader, nir_jump_type type);
4723
4724 nir_load_const_instr *nir_load_const_instr_create(nir_shader *shader,
4725 unsigned num_components,
4726 unsigned bit_size);
4727
4728 nir_intrinsic_instr *nir_intrinsic_instr_create(nir_shader *shader,
4729 nir_intrinsic_op op);
4730
4731 nir_call_instr *nir_call_instr_create(nir_shader *shader,
4732 nir_function *callee);
4733
4734 /** Creates a NIR texture instruction */
4735 nir_tex_instr *nir_tex_instr_create(nir_shader *shader, unsigned num_srcs);
4736
4737 nir_phi_instr *nir_phi_instr_create(nir_shader *shader);
4738 nir_phi_src *nir_phi_instr_add_src(nir_phi_instr *instr,
4739 nir_block *pred, nir_def *src);
4740
4741 nir_parallel_copy_instr *nir_parallel_copy_instr_create(nir_shader *shader);
4742
4743 nir_debug_info_instr *nir_debug_info_instr_create(nir_shader *shader,
4744 nir_debug_info_type type,
4745 uint32_t string_length);
4746
4747 nir_undef_instr *nir_undef_instr_create(nir_shader *shader,
4748 unsigned num_components,
4749 unsigned bit_size);
4750
4751 nir_const_value nir_alu_binop_identity(nir_op binop, unsigned bit_size);
4752
4753 /**
4754 * NIR Cursors and Instruction Insertion API
4755 * @{
4756 *
4757 * A tiny struct representing a point to insert/extract instructions or
4758 * control flow nodes. Helps reduce the combinatorial explosion of possible
4759 * points to insert/extract.
4760 *
4761 * \sa nir_control_flow.h
4762 */
4763 typedef enum {
4764 nir_cursor_before_block,
4765 nir_cursor_after_block,
4766 nir_cursor_before_instr,
4767 nir_cursor_after_instr,
4768 } nir_cursor_option;
4769
4770 typedef struct {
4771 nir_cursor_option option;
4772 union {
4773 nir_block *block;
4774 nir_instr *instr;
4775 };
4776 } nir_cursor;
4777
4778 static inline nir_block *
nir_cursor_current_block(nir_cursor cursor)4779 nir_cursor_current_block(nir_cursor cursor)
4780 {
4781 if (cursor.option == nir_cursor_before_instr ||
4782 cursor.option == nir_cursor_after_instr) {
4783 return cursor.instr->block;
4784 } else {
4785 return cursor.block;
4786 }
4787 }
4788
4789 bool nir_cursors_equal(nir_cursor a, nir_cursor b);
4790
4791 static inline nir_cursor
nir_before_block(nir_block * block)4792 nir_before_block(nir_block *block)
4793 {
4794 nir_cursor cursor;
4795 cursor.option = nir_cursor_before_block;
4796 cursor.block = block;
4797 return cursor;
4798 }
4799
4800 static inline nir_cursor
nir_after_block(nir_block * block)4801 nir_after_block(nir_block *block)
4802 {
4803 nir_cursor cursor;
4804 cursor.option = nir_cursor_after_block;
4805 cursor.block = block;
4806 return cursor;
4807 }
4808
4809 static inline nir_cursor
nir_before_instr(nir_instr * instr)4810 nir_before_instr(nir_instr *instr)
4811 {
4812 nir_cursor cursor;
4813 cursor.option = nir_cursor_before_instr;
4814 cursor.instr = instr;
4815 return cursor;
4816 }
4817
4818 static inline nir_cursor
nir_after_instr(nir_instr * instr)4819 nir_after_instr(nir_instr *instr)
4820 {
4821 nir_cursor cursor;
4822 cursor.option = nir_cursor_after_instr;
4823 cursor.instr = instr;
4824 return cursor;
4825 }
4826
4827 static inline nir_cursor
nir_before_block_after_phis(nir_block * block)4828 nir_before_block_after_phis(nir_block *block)
4829 {
4830 nir_phi_instr *last_phi = nir_block_last_phi_instr(block);
4831 if (last_phi)
4832 return nir_after_instr(&last_phi->instr);
4833 else
4834 return nir_before_block(block);
4835 }
4836
4837 static inline nir_cursor
nir_after_block_before_jump(nir_block * block)4838 nir_after_block_before_jump(nir_block *block)
4839 {
4840 nir_instr *last_instr = nir_block_last_instr(block);
4841 if (last_instr && last_instr->type == nir_instr_type_jump) {
4842 return nir_before_instr(last_instr);
4843 } else {
4844 return nir_after_block(block);
4845 }
4846 }
4847
4848 static inline nir_cursor
nir_before_src(nir_src * src)4849 nir_before_src(nir_src *src)
4850 {
4851 if (nir_src_is_if(src)) {
4852 nir_block *prev_block =
4853 nir_cf_node_as_block(nir_cf_node_prev(&nir_src_parent_if(src)->cf_node));
4854 return nir_after_block(prev_block);
4855 } else if (nir_src_parent_instr(src)->type == nir_instr_type_phi) {
4856 #ifndef NDEBUG
4857 nir_phi_instr *cond_phi = nir_instr_as_phi(nir_src_parent_instr(src));
4858 bool found = false;
4859 nir_foreach_phi_src(phi_src, cond_phi) {
4860 if (phi_src->src.ssa == src->ssa) {
4861 found = true;
4862 break;
4863 }
4864 }
4865 assert(found);
4866 #endif
4867 /* The list_entry() macro is a generic container-of macro, it just happens
4868 * to have a more specific name.
4869 */
4870 nir_phi_src *phi_src = list_entry(src, nir_phi_src, src);
4871 return nir_after_block_before_jump(phi_src->pred);
4872 } else {
4873 return nir_before_instr(nir_src_parent_instr(src));
4874 }
4875 }
4876
4877 static inline nir_cursor
nir_before_cf_node(nir_cf_node * node)4878 nir_before_cf_node(nir_cf_node *node)
4879 {
4880 if (node->type == nir_cf_node_block)
4881 return nir_before_block(nir_cf_node_as_block(node));
4882
4883 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node)));
4884 }
4885
4886 static inline nir_cursor
nir_after_cf_node(nir_cf_node * node)4887 nir_after_cf_node(nir_cf_node *node)
4888 {
4889 if (node->type == nir_cf_node_block)
4890 return nir_after_block(nir_cf_node_as_block(node));
4891
4892 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node)));
4893 }
4894
4895 static inline nir_cursor
nir_after_phis(nir_block * block)4896 nir_after_phis(nir_block *block)
4897 {
4898 nir_foreach_instr(instr, block) {
4899 if (instr->type != nir_instr_type_phi)
4900 return nir_before_instr(instr);
4901 }
4902 return nir_after_block(block);
4903 }
4904
4905 static inline nir_cursor
nir_after_instr_and_phis(nir_instr * instr)4906 nir_after_instr_and_phis(nir_instr *instr)
4907 {
4908 if (instr->type == nir_instr_type_phi)
4909 return nir_after_phis(instr->block);
4910 else
4911 return nir_after_instr(instr);
4912 }
4913
4914 static inline nir_cursor
nir_after_cf_node_and_phis(nir_cf_node * node)4915 nir_after_cf_node_and_phis(nir_cf_node *node)
4916 {
4917 if (node->type == nir_cf_node_block)
4918 return nir_after_block(nir_cf_node_as_block(node));
4919
4920 nir_block *block = nir_cf_node_as_block(nir_cf_node_next(node));
4921
4922 return nir_after_phis(block);
4923 }
4924
4925 static inline nir_cursor
nir_before_cf_list(struct exec_list * cf_list)4926 nir_before_cf_list(struct exec_list *cf_list)
4927 {
4928 nir_cf_node *first_node = exec_node_data(nir_cf_node,
4929 exec_list_get_head(cf_list), node);
4930 return nir_before_cf_node(first_node);
4931 }
4932
4933 static inline nir_cursor
nir_after_cf_list(struct exec_list * cf_list)4934 nir_after_cf_list(struct exec_list *cf_list)
4935 {
4936 nir_cf_node *last_node = exec_node_data(nir_cf_node,
4937 exec_list_get_tail(cf_list), node);
4938 return nir_after_cf_node(last_node);
4939 }
4940
4941 static inline nir_cursor
nir_before_impl(nir_function_impl * impl)4942 nir_before_impl(nir_function_impl *impl)
4943 {
4944 return nir_before_cf_list(&impl->body);
4945 }
4946
4947 static inline nir_cursor
nir_after_impl(nir_function_impl * impl)4948 nir_after_impl(nir_function_impl *impl)
4949 {
4950 return nir_after_cf_list(&impl->body);
4951 }
4952
4953 /**
4954 * Insert a NIR instruction at the given cursor.
4955 *
4956 * Note: This does not update the cursor.
4957 */
4958 void nir_instr_insert(nir_cursor cursor, nir_instr *instr);
4959
4960 bool nir_instr_move(nir_cursor cursor, nir_instr *instr);
4961
4962 static inline void
nir_instr_insert_before(nir_instr * instr,nir_instr * before)4963 nir_instr_insert_before(nir_instr *instr, nir_instr *before)
4964 {
4965 nir_instr_insert(nir_before_instr(instr), before);
4966 }
4967
4968 static inline void
nir_instr_insert_after(nir_instr * instr,nir_instr * after)4969 nir_instr_insert_after(nir_instr *instr, nir_instr *after)
4970 {
4971 nir_instr_insert(nir_after_instr(instr), after);
4972 }
4973
4974 static inline void
nir_instr_insert_before_block(nir_block * block,nir_instr * before)4975 nir_instr_insert_before_block(nir_block *block, nir_instr *before)
4976 {
4977 nir_instr_insert(nir_before_block(block), before);
4978 }
4979
4980 static inline void
nir_instr_insert_after_block(nir_block * block,nir_instr * after)4981 nir_instr_insert_after_block(nir_block *block, nir_instr *after)
4982 {
4983 nir_instr_insert(nir_after_block(block), after);
4984 }
4985
4986 static inline void
nir_instr_insert_before_cf(nir_cf_node * node,nir_instr * before)4987 nir_instr_insert_before_cf(nir_cf_node *node, nir_instr *before)
4988 {
4989 nir_instr_insert(nir_before_cf_node(node), before);
4990 }
4991
4992 static inline void
nir_instr_insert_after_cf(nir_cf_node * node,nir_instr * after)4993 nir_instr_insert_after_cf(nir_cf_node *node, nir_instr *after)
4994 {
4995 nir_instr_insert(nir_after_cf_node(node), after);
4996 }
4997
4998 static inline void
nir_instr_insert_before_cf_list(struct exec_list * list,nir_instr * before)4999 nir_instr_insert_before_cf_list(struct exec_list *list, nir_instr *before)
5000 {
5001 nir_instr_insert(nir_before_cf_list(list), before);
5002 }
5003
5004 static inline void
nir_instr_insert_after_cf_list(struct exec_list * list,nir_instr * after)5005 nir_instr_insert_after_cf_list(struct exec_list *list, nir_instr *after)
5006 {
5007 nir_instr_insert(nir_after_cf_list(list), after);
5008 }
5009
5010 void nir_instr_remove_v(nir_instr *instr);
5011 void nir_instr_free(nir_instr *instr);
5012 void nir_instr_free_list(struct exec_list *list);
5013
5014 static inline nir_cursor
nir_instr_remove(nir_instr * instr)5015 nir_instr_remove(nir_instr *instr)
5016 {
5017 nir_cursor cursor;
5018 nir_instr *prev = nir_instr_prev(instr);
5019 if (prev) {
5020 cursor = nir_after_instr(prev);
5021 } else {
5022 cursor = nir_before_block(instr->block);
5023 }
5024 nir_instr_remove_v(instr);
5025 return cursor;
5026 }
5027
5028 nir_cursor nir_instr_free_and_dce(nir_instr *instr);
5029
5030 /** @} */
5031
5032 nir_def *nir_instr_def(nir_instr *instr);
5033
5034 typedef bool (*nir_foreach_def_cb)(nir_def *def, void *state);
5035 typedef bool (*nir_foreach_src_cb)(nir_src *src, void *state);
5036 static inline bool nir_foreach_src(nir_instr *instr, nir_foreach_src_cb cb, void *state);
5037 bool nir_foreach_phi_src_leaving_block(nir_block *instr,
5038 nir_foreach_src_cb cb,
5039 void *state);
5040
5041 nir_const_value *nir_src_as_const_value(nir_src src);
5042
5043 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
5044 static inline c_type * \
5045 nir_src_as_##name(nir_src src) \
5046 { \
5047 return src.ssa->parent_instr->type == type_enum \
5048 ? cast_macro(src.ssa->parent_instr) \
5049 : NULL; \
5050 }
5051
5052 NIR_SRC_AS_(alu_instr, nir_alu_instr, nir_instr_type_alu, nir_instr_as_alu)
5053 NIR_SRC_AS_(intrinsic, nir_intrinsic_instr,
5054 nir_instr_type_intrinsic, nir_instr_as_intrinsic)
5055 NIR_SRC_AS_(deref, nir_deref_instr, nir_instr_type_deref, nir_instr_as_deref)
5056 NIR_SRC_AS_(debug_info, nir_debug_info_instr, nir_instr_type_debug_info, nir_instr_as_debug_info)
5057
5058 const char *nir_src_as_string(nir_src src);
5059
5060 bool nir_src_is_always_uniform(nir_src src);
5061 bool nir_srcs_equal(nir_src src1, nir_src src2);
5062 bool nir_instrs_equal(const nir_instr *instr1, const nir_instr *instr2);
5063 nir_block *nir_src_get_block(nir_src *src);
5064
5065 static inline void
nir_src_rewrite(nir_src * src,nir_def * new_ssa)5066 nir_src_rewrite(nir_src *src, nir_def *new_ssa)
5067 {
5068 assert(src->ssa);
5069 assert(nir_src_is_if(src) ? (nir_src_parent_if(src) != NULL) : (nir_src_parent_instr(src) != NULL));
5070 list_del(&src->use_link);
5071 src->ssa = new_ssa;
5072 list_addtail(&src->use_link, &new_ssa->uses);
5073 }
5074
5075 /** Initialize a nir_src
5076 *
5077 * This is almost never the helper you want to use. This helper assumes that
5078 * the source is uninitialized garbage and blasts over it without doing any
5079 * tear-down the existing source, including removing it from uses lists.
5080 * Using this helper on a source that currently exists in any uses list will
5081 * result in linked list corruption. It also assumes that the instruction is
5082 * currently live in the IR and adds the source to the uses list for the given
5083 * nir_def as part of setup.
5084 *
5085 * This is pretty much only useful for adding sources to extant instructions
5086 * or manipulating parallel copy instructions as part of out-of-SSA.
5087 *
5088 * When in doubt, use nir_src_rewrite() instead.
5089 */
5090 void nir_instr_init_src(nir_instr *instr, nir_src *src, nir_def *def);
5091
5092 /** Clear a nir_src
5093 *
5094 * This helper clears a nir_src by removing it from any uses lists and
5095 * resetting its contents to NIR_SRC_INIT. This is typically used as a
5096 * precursor to removing the source from the instruction by adjusting a
5097 * num_srcs parameter somewhere or overwriting it with nir_instr_move_src().
5098 */
5099 void nir_instr_clear_src(nir_instr *instr, nir_src *src);
5100
5101 void nir_instr_move_src(nir_instr *dest_instr, nir_src *dest, nir_src *src);
5102
5103 /** Returns true if first comes before second in a block. */
5104 bool nir_instr_is_before(nir_instr *first, nir_instr *second);
5105
5106 void nir_def_init(nir_instr *instr, nir_def *def,
5107 unsigned num_components, unsigned bit_size);
5108 static inline void
nir_def_init_for_type(nir_instr * instr,nir_def * def,const struct glsl_type * type)5109 nir_def_init_for_type(nir_instr *instr, nir_def *def,
5110 const struct glsl_type *type)
5111 {
5112 assert(glsl_type_is_vector_or_scalar(type));
5113 nir_def_init(instr, def, glsl_get_components(type),
5114 glsl_get_bit_size(type));
5115 }
5116 void nir_def_rewrite_uses(nir_def *def, nir_def *new_ssa);
5117 void nir_def_rewrite_uses_src(nir_def *def, nir_src new_src);
5118 void nir_def_rewrite_uses_after(nir_def *def, nir_def *new_ssa,
5119 nir_instr *after_me);
5120
5121 static inline void
nir_def_replace(nir_def * def,nir_def * new_ssa)5122 nir_def_replace(nir_def *def, nir_def *new_ssa)
5123 {
5124 nir_def_rewrite_uses(def, new_ssa);
5125 nir_instr_remove(def->parent_instr);
5126 }
5127
5128 nir_component_mask_t nir_src_components_read(const nir_src *src);
5129 nir_component_mask_t nir_def_components_read(const nir_def *def);
5130 bool nir_def_all_uses_are_fsat(const nir_def *def);
5131 bool nir_def_all_uses_ignore_sign_bit(const nir_def *def);
5132
5133 static inline int
nir_def_first_component_read(nir_def * def)5134 nir_def_first_component_read(nir_def *def)
5135 {
5136 return (int)ffs(nir_def_components_read(def)) - 1;
5137 }
5138
5139 static inline int
nir_def_last_component_read(nir_def * def)5140 nir_def_last_component_read(nir_def *def)
5141 {
5142 return (int)util_last_bit(nir_def_components_read(def)) - 1;
5143 }
5144
5145 static inline bool
nir_def_is_unused(nir_def * ssa)5146 nir_def_is_unused(nir_def *ssa)
5147 {
5148 return list_is_empty(&ssa->uses);
5149 }
5150
5151 /** Sorts unstructured blocks
5152 *
5153 * NIR requires that unstructured blocks be sorted in reverse post
5154 * depth-first-search order. This is the standard ordering used in the
5155 * compiler literature which guarantees dominance. In particular, reverse
5156 * post-DFS order guarantees that dominators occur in the list before the
5157 * blocks they dominate.
5158 *
5159 * NOTE: This function also implicitly deletes any unreachable blocks.
5160 */
5161 void nir_sort_unstructured_blocks(nir_function_impl *impl);
5162
5163 /** Returns the next block
5164 *
5165 * For structured control-flow, this follows the same order as
5166 * nir_block_cf_tree_next(). For unstructured control-flow the blocks are in
5167 * reverse post-DFS order. (See nir_sort_unstructured_blocks() above.)
5168 */
5169 nir_block *nir_block_unstructured_next(nir_block *block);
5170 nir_block *nir_unstructured_start_block(nir_function_impl *impl);
5171
5172 #define nir_foreach_block_unstructured(block, impl) \
5173 for (nir_block *block = nir_unstructured_start_block(impl); block != NULL; \
5174 block = nir_block_unstructured_next(block))
5175
5176 #define nir_foreach_block_unstructured_safe(block, impl) \
5177 for (nir_block *block = nir_unstructured_start_block(impl), \
5178 *next = nir_block_unstructured_next(block); \
5179 block != NULL; \
5180 block = next, next = nir_block_unstructured_next(block))
5181
5182 /*
5183 * finds the next basic block in source-code order, returns NULL if there is
5184 * none
5185 */
5186
5187 nir_block *nir_block_cf_tree_next(nir_block *block);
5188
5189 /* Performs the opposite of nir_block_cf_tree_next() */
5190
5191 nir_block *nir_block_cf_tree_prev(nir_block *block);
5192
5193 /* Gets the first block in a CF node in source-code order */
5194
5195 nir_block *nir_cf_node_cf_tree_first(nir_cf_node *node);
5196
5197 /* Gets the last block in a CF node in source-code order */
5198
5199 nir_block *nir_cf_node_cf_tree_last(nir_cf_node *node);
5200
5201 /* Gets the next block after a CF node in source-code order */
5202
5203 nir_block *nir_cf_node_cf_tree_next(nir_cf_node *node);
5204
5205 /* Gets the block before a CF node in source-code order */
5206
5207 nir_block *nir_cf_node_cf_tree_prev(nir_cf_node *node);
5208
5209 /* Macros for loops that visit blocks in source-code order */
5210
5211 #define nir_foreach_block(block, impl) \
5212 for (nir_block *block = nir_start_block(impl); block != NULL; \
5213 block = nir_block_cf_tree_next(block))
5214
5215 #define nir_foreach_block_safe(block, impl) \
5216 for (nir_block *block = nir_start_block(impl), \
5217 *next = nir_block_cf_tree_next(block); \
5218 block != NULL; \
5219 block = next, next = nir_block_cf_tree_next(block))
5220
5221 #define nir_foreach_block_reverse(block, impl) \
5222 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
5223 block = nir_block_cf_tree_prev(block))
5224
5225 #define nir_foreach_block_reverse_safe(block, impl) \
5226 for (nir_block *block = nir_impl_last_block(impl), \
5227 *prev = nir_block_cf_tree_prev(block); \
5228 block != NULL; \
5229 block = prev, prev = nir_block_cf_tree_prev(block))
5230
5231 #define nir_foreach_block_in_cf_node(block, node) \
5232 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
5233 block != nir_cf_node_cf_tree_next(node); \
5234 block = nir_block_cf_tree_next(block))
5235
5236 #define nir_foreach_block_in_cf_node_safe(block, node) \
5237 for (nir_block *block = nir_cf_node_cf_tree_first(node), \
5238 *next = nir_block_cf_tree_next(block); \
5239 block != nir_cf_node_cf_tree_next(node); \
5240 block = next, next = nir_block_cf_tree_next(block))
5241
5242 #define nir_foreach_block_in_cf_node_reverse(block, node) \
5243 for (nir_block *block = nir_cf_node_cf_tree_last(node); \
5244 block != nir_cf_node_cf_tree_prev(node); \
5245 block = nir_block_cf_tree_prev(block))
5246
5247 #define nir_foreach_block_in_cf_node_reverse_safe(block, node) \
5248 for (nir_block *block = nir_cf_node_cf_tree_last(node), \
5249 *prev = nir_block_cf_tree_prev(block); \
5250 block != nir_cf_node_cf_tree_prev(node); \
5251 block = prev, prev = nir_block_cf_tree_prev(block))
5252
5253 /* If the following CF node is an if, this function returns that if.
5254 * Otherwise, it returns NULL.
5255 */
5256 nir_if *nir_block_get_following_if(nir_block *block);
5257
5258 nir_loop *nir_block_get_following_loop(nir_block *block);
5259
5260 nir_block **nir_block_get_predecessors_sorted(const nir_block *block, void *mem_ctx);
5261
5262 void nir_index_ssa_defs(nir_function_impl *impl);
5263 unsigned nir_index_instrs(nir_function_impl *impl);
5264
5265 void nir_index_blocks(nir_function_impl *impl);
5266
5267 void nir_shader_clear_pass_flags(nir_shader *shader);
5268
5269 unsigned nir_shader_index_vars(nir_shader *shader, nir_variable_mode modes);
5270 unsigned nir_function_impl_index_vars(nir_function_impl *impl);
5271
5272 void nir_print_shader(nir_shader *shader, FILE *fp);
5273 void nir_print_shader_annotated(nir_shader *shader, FILE *fp, struct hash_table *errors);
5274 void nir_print_instr(const nir_instr *instr, FILE *fp);
5275 void nir_print_deref(const nir_deref_instr *deref, FILE *fp);
5276 void nir_log_shader_annotated_tagged(enum mesa_log_level level, const char *tag, nir_shader *shader, struct hash_table *annotations);
5277 #define nir_log_shadere(s) nir_log_shader_annotated_tagged(MESA_LOG_ERROR, (MESA_LOG_TAG), (s), NULL)
5278 #define nir_log_shaderw(s) nir_log_shader_annotated_tagged(MESA_LOG_WARN, (MESA_LOG_TAG), (s), NULL)
5279 #define nir_log_shaderi(s) nir_log_shader_annotated_tagged(MESA_LOG_INFO, (MESA_LOG_TAG), (s), NULL)
5280 #define nir_log_shader_annotated(s, annotations) nir_log_shader_annotated_tagged(MESA_LOG_ERROR, (MESA_LOG_TAG), (s), annotations)
5281
5282 char *nir_shader_as_str(nir_shader *nir, void *mem_ctx);
5283 char *nir_shader_as_str_annotated(nir_shader *nir, struct hash_table *annotations, void *mem_ctx);
5284 char *nir_instr_as_str(const nir_instr *instr, void *mem_ctx);
5285
5286 /** Adds debug information to the shader. The line numbers point to
5287 * the corresponding lines in the printed NIR, starting first_line;
5288 */
5289 char *nir_shader_gather_debug_info(nir_shader *shader, const char *filename, uint32_t first_line);
5290
5291 /** Shallow clone of a single instruction. */
5292 nir_instr *nir_instr_clone(nir_shader *s, const nir_instr *orig);
5293
5294 /** Clone a single instruction, including a remap table to rewrite sources. */
5295 nir_instr *nir_instr_clone_deep(nir_shader *s, const nir_instr *orig,
5296 struct hash_table *remap_table);
5297
5298 /** Shallow clone of a single ALU instruction. */
5299 nir_alu_instr *nir_alu_instr_clone(nir_shader *s, const nir_alu_instr *orig);
5300
5301 nir_shader *nir_shader_clone(void *mem_ctx, const nir_shader *s);
5302 nir_function *nir_function_clone(nir_shader *ns, const nir_function *fxn);
5303 nir_function_impl *nir_function_impl_clone(nir_shader *shader,
5304 const nir_function_impl *fi);
5305 nir_function_impl *
5306 nir_function_impl_clone_remap_globals(nir_shader *shader,
5307 const nir_function_impl *fi,
5308 struct hash_table *remap_table);
5309 nir_constant *nir_constant_clone(const nir_constant *c, nir_variable *var);
5310 nir_variable *nir_variable_clone(const nir_variable *c, nir_shader *shader);
5311
5312 void nir_shader_replace(nir_shader *dest, nir_shader *src);
5313
5314 void nir_shader_serialize_deserialize(nir_shader *s);
5315
5316 #ifndef NDEBUG
5317 void nir_validate_shader(nir_shader *shader, const char *when);
5318 void nir_validate_ssa_dominance(nir_shader *shader, const char *when);
5319 void nir_metadata_set_validation_flag(nir_shader *shader);
5320 void nir_metadata_check_validation_flag(nir_shader *shader);
5321
5322 static inline bool
should_skip_nir(const char * name)5323 should_skip_nir(const char *name)
5324 {
5325 static const char *list = NULL;
5326 if (!list) {
5327 /* Comma separated list of names to skip. */
5328 list = getenv("NIR_SKIP");
5329 if (!list)
5330 list = "";
5331 }
5332
5333 if (!list[0])
5334 return false;
5335
5336 return comma_separated_list_contains(list, name);
5337 }
5338
5339 static inline bool
should_print_nir(nir_shader * shader)5340 should_print_nir(nir_shader *shader)
5341 {
5342 if ((shader->info.internal && !NIR_DEBUG(PRINT_INTERNAL)) ||
5343 shader->info.stage < 0 ||
5344 shader->info.stage > MESA_SHADER_KERNEL)
5345 return false;
5346
5347 return unlikely(nir_debug_print_shader[shader->info.stage]);
5348 }
5349 #else
5350 static inline void
nir_validate_shader(nir_shader * shader,const char * when)5351 nir_validate_shader(nir_shader *shader, const char *when)
5352 {
5353 (void)shader;
5354 (void)when;
5355 }
5356 static inline void
nir_validate_ssa_dominance(nir_shader * shader,const char * when)5357 nir_validate_ssa_dominance(nir_shader *shader, const char *when)
5358 {
5359 (void)shader;
5360 (void)when;
5361 }
5362 static inline void
nir_metadata_set_validation_flag(nir_shader * shader)5363 nir_metadata_set_validation_flag(nir_shader *shader)
5364 {
5365 (void)shader;
5366 }
5367 static inline void
nir_metadata_check_validation_flag(nir_shader * shader)5368 nir_metadata_check_validation_flag(nir_shader *shader)
5369 {
5370 (void)shader;
5371 }
5372 static inline bool
should_skip_nir(UNUSED const char * pass_name)5373 should_skip_nir(UNUSED const char *pass_name)
5374 {
5375 return false;
5376 }
5377 static inline bool
should_print_nir(UNUSED nir_shader * shader)5378 should_print_nir(UNUSED nir_shader *shader)
5379 {
5380 return false;
5381 }
5382 #endif /* NDEBUG */
5383
5384 #define _PASS(pass, nir, do_pass) \
5385 do { \
5386 if (should_skip_nir(#pass)) { \
5387 printf("skipping %s\n", #pass); \
5388 break; \
5389 } \
5390 do_pass if (NIR_DEBUG(CLONE)) \
5391 { \
5392 nir_shader *_clone = nir_shader_clone(ralloc_parent(nir), nir);\
5393 nir_shader_replace(nir, _clone); \
5394 } \
5395 if (NIR_DEBUG(SERIALIZE)) { \
5396 nir_shader_serialize_deserialize(nir); \
5397 } \
5398 } while (0)
5399
5400 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, { \
5401 nir_metadata_set_validation_flag(nir); \
5402 if (should_print_nir(nir)) \
5403 printf("%s\n", #pass); \
5404 if (pass(nir, ##__VA_ARGS__)) { \
5405 nir_validate_shader(nir, "after " #pass " in " __FILE__); \
5406 UNUSED bool _; \
5407 progress = true; \
5408 if (should_print_nir(nir)) \
5409 nir_print_shader(nir, stdout); \
5410 nir_metadata_check_validation_flag(nir); \
5411 } \
5412 })
5413
5414 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, { \
5415 if (should_print_nir(nir)) \
5416 printf("%s\n", #pass); \
5417 pass(nir, ##__VA_ARGS__); \
5418 nir_validate_shader(nir, "after " #pass " in " __FILE__); \
5419 if (should_print_nir(nir)) \
5420 nir_print_shader(nir, stdout); \
5421 })
5422
5423 #define _NIR_LOOP_PASS(progress, idempotent, skip, nir, pass, ...) \
5424 do { \
5425 bool nir_loop_pass_progress = false; \
5426 if (!_mesa_set_search(skip, (void (*)())&pass)) \
5427 NIR_PASS(nir_loop_pass_progress, nir, pass, ##__VA_ARGS__); \
5428 if (nir_loop_pass_progress) \
5429 _mesa_set_clear(skip, NULL); \
5430 if (idempotent || !nir_loop_pass_progress) \
5431 _mesa_set_add(skip, (void (*)())&pass); \
5432 UNUSED bool _ = false; \
5433 progress |= nir_loop_pass_progress; \
5434 } while (0)
5435
5436 /* Helper to skip a pass if no different passes have made progress since it was
5437 * previously run. Note that two passes are considered the same if they have
5438 * the same function pointer, even if they used different options.
5439 *
5440 * The usage of this is mostly identical to NIR_PASS. "skip" is a "struct set *"
5441 * (created by _mesa_pointer_set_create) which the macro uses to keep track of
5442 * already run passes.
5443 *
5444 * Example:
5445 * bool progress = true;
5446 * struct set *skip = _mesa_pointer_set_create(NULL);
5447 * while (progress) {
5448 * progress = false;
5449 * NIR_LOOP_PASS(progress, skip, nir, pass1);
5450 * NIR_LOOP_PASS_NOT_IDEMPOTENT(progress, skip, nir, nir_opt_algebraic);
5451 * NIR_LOOP_PASS(progress, skip, nir, pass2);
5452 * ...
5453 * }
5454 * _mesa_set_destroy(skip, NULL);
5455 *
5456 * You shouldn't mix usage of this with the NIR_PASS set of helpers, without
5457 * using a new "skip" in-between.
5458 */
5459 #define NIR_LOOP_PASS(progress, skip, nir, pass, ...) \
5460 _NIR_LOOP_PASS(progress, true, skip, nir, pass, ##__VA_ARGS__)
5461
5462 /* Like NIR_LOOP_PASS, but use this for passes which may make further progress
5463 * when repeated.
5464 */
5465 #define NIR_LOOP_PASS_NOT_IDEMPOTENT(progress, skip, nir, pass, ...) \
5466 _NIR_LOOP_PASS(progress, false, skip, nir, pass, ##__VA_ARGS__)
5467
5468 #define NIR_SKIP(name) should_skip_nir(#name)
5469
5470 /** An instruction filtering callback with writemask
5471 *
5472 * Returns true if the instruction should be processed with the associated
5473 * writemask and false otherwise.
5474 */
5475 typedef bool (*nir_instr_writemask_filter_cb)(const nir_instr *,
5476 unsigned writemask, const void *);
5477
5478 /** A simple instruction lowering callback
5479 *
5480 * Many instruction lowering passes can be written as a simple function which
5481 * takes an instruction as its input and returns a sequence of instructions
5482 * that implement the consumed instruction. This function type represents
5483 * such a lowering function. When called, a function with this prototype
5484 * should either return NULL indicating that no lowering needs to be done or
5485 * emit a sequence of instructions using the provided builder (whose cursor
5486 * will already be placed after the instruction to be lowered) and return the
5487 * resulting nir_def.
5488 */
5489 typedef nir_def *(*nir_lower_instr_cb)(struct nir_builder *,
5490 nir_instr *, void *);
5491
5492 /**
5493 * Special return value for nir_lower_instr_cb when some progress occurred
5494 * (like changing an input to the instr) that didn't result in a replacement
5495 * SSA def being generated.
5496 */
5497 #define NIR_LOWER_INSTR_PROGRESS ((nir_def *)(uintptr_t)1)
5498
5499 /**
5500 * Special return value for nir_lower_instr_cb when some progress occurred
5501 * that should remove the current instruction that doesn't create an output
5502 * (like a store)
5503 */
5504
5505 #define NIR_LOWER_INSTR_PROGRESS_REPLACE ((nir_def *)(uintptr_t)2)
5506
5507 /** Iterate over all the instructions in a nir_function_impl and lower them
5508 * using the provided callbacks
5509 *
5510 * This function implements the guts of a standard lowering pass for you. It
5511 * iterates over all of the instructions in a nir_function_impl and calls the
5512 * filter callback on each one. If the filter callback returns true, it then
5513 * calls the lowering call back on the instruction. (Splitting it this way
5514 * allows us to avoid some save/restore work for instructions we know won't be
5515 * lowered.) If the instruction is dead after the lowering is complete, it
5516 * will be removed. If new instructions are added, the lowering callback will
5517 * also be called on them in case multiple lowerings are required.
5518 *
5519 * If the callback indicates that the original instruction is replaced (either
5520 * through a new SSA def or NIR_LOWER_INSTR_PROGRESS_REPLACE), then the
5521 * instruction is removed along with any now-dead SSA defs it used.
5522 *
5523 * The metadata for the nir_function_impl will also be updated. If any blocks
5524 * are added (they cannot be removed), dominance and block indices will be
5525 * invalidated.
5526 */
5527 bool nir_function_impl_lower_instructions(nir_function_impl *impl,
5528 nir_instr_filter_cb filter,
5529 nir_lower_instr_cb lower,
5530 void *cb_data);
5531 bool nir_shader_lower_instructions(nir_shader *shader,
5532 nir_instr_filter_cb filter,
5533 nir_lower_instr_cb lower,
5534 void *cb_data);
5535
5536 void nir_calc_dominance_impl(nir_function_impl *impl);
5537 void nir_calc_dominance(nir_shader *shader);
5538
5539 nir_block *nir_dominance_lca(nir_block *b1, nir_block *b2);
5540 bool nir_block_dominates(nir_block *parent, nir_block *child);
5541 bool nir_block_is_unreachable(nir_block *block);
5542
5543 void nir_dump_dom_tree_impl(nir_function_impl *impl, FILE *fp);
5544 void nir_dump_dom_tree(nir_shader *shader, FILE *fp);
5545
5546 void nir_dump_dom_frontier_impl(nir_function_impl *impl, FILE *fp);
5547 void nir_dump_dom_frontier(nir_shader *shader, FILE *fp);
5548
5549 void nir_dump_cfg_impl(nir_function_impl *impl, FILE *fp);
5550 void nir_dump_cfg(nir_shader *shader, FILE *fp);
5551
5552 void nir_gs_count_vertices_and_primitives(const nir_shader *shader,
5553 int *out_vtxcnt,
5554 int *out_prmcnt,
5555 int *out_decomposed_prmcnt,
5556 unsigned num_streams);
5557
5558 typedef enum {
5559 nir_group_all,
5560 nir_group_same_resource_only,
5561 } nir_load_grouping;
5562
5563 void nir_group_loads(nir_shader *shader, nir_load_grouping grouping,
5564 unsigned max_distance);
5565
5566 bool nir_shrink_vec_array_vars(nir_shader *shader, nir_variable_mode modes);
5567 bool nir_split_array_vars(nir_shader *shader, nir_variable_mode modes);
5568 bool nir_split_var_copies(nir_shader *shader);
5569 bool nir_split_per_member_structs(nir_shader *shader);
5570 bool nir_split_struct_vars(nir_shader *shader, nir_variable_mode modes);
5571
5572 bool nir_lower_returns_impl(nir_function_impl *impl);
5573 bool nir_lower_returns(nir_shader *shader);
5574
5575 void nir_inline_function_impl(struct nir_builder *b,
5576 const nir_function_impl *impl,
5577 nir_def **params,
5578 struct hash_table *shader_var_remap);
5579 bool nir_inline_functions(nir_shader *shader);
5580 void nir_cleanup_functions(nir_shader *shader);
5581 bool nir_link_shader_functions(nir_shader *shader,
5582 const nir_shader *link_shader);
5583 bool nir_lower_calls_to_builtins(nir_shader *s);
5584
5585 void nir_find_inlinable_uniforms(nir_shader *shader);
5586 void nir_inline_uniforms(nir_shader *shader, unsigned num_uniforms,
5587 const uint32_t *uniform_values,
5588 const uint16_t *uniform_dw_offsets);
5589 bool nir_collect_src_uniforms(const nir_src *src, int component,
5590 uint32_t *uni_offsets, uint8_t *num_offsets,
5591 unsigned max_num_bo, unsigned max_offset);
5592 void nir_add_inlinable_uniforms(const nir_src *cond, nir_loop_info *info,
5593 uint32_t *uni_offsets, uint8_t *num_offsets,
5594 unsigned max_num_bo, unsigned max_offset);
5595
5596 bool nir_propagate_invariant(nir_shader *shader, bool invariant_prim);
5597
5598 void nir_lower_var_copy_instr(nir_intrinsic_instr *copy, nir_shader *shader);
5599 void nir_lower_deref_copy_instr(struct nir_builder *b,
5600 nir_intrinsic_instr *copy);
5601 bool nir_lower_var_copies(nir_shader *shader);
5602
5603 bool nir_opt_memcpy(nir_shader *shader);
5604 bool nir_lower_memcpy(nir_shader *shader);
5605
5606 void nir_fixup_deref_modes(nir_shader *shader);
5607 void nir_fixup_deref_types(nir_shader *shader);
5608
5609 bool nir_lower_global_vars_to_local(nir_shader *shader);
5610 void nir_lower_constant_to_temp(nir_shader *shader);
5611
5612 typedef enum {
5613 nir_lower_direct_array_deref_of_vec_load = (1 << 0),
5614 nir_lower_indirect_array_deref_of_vec_load = (1 << 1),
5615 nir_lower_direct_array_deref_of_vec_store = (1 << 2),
5616 nir_lower_indirect_array_deref_of_vec_store = (1 << 3),
5617 } nir_lower_array_deref_of_vec_options;
5618
5619 bool nir_lower_array_deref_of_vec(nir_shader *shader, nir_variable_mode modes,
5620 bool (*filter)(nir_variable *),
5621 nir_lower_array_deref_of_vec_options options);
5622
5623 bool nir_lower_indirect_derefs(nir_shader *shader, nir_variable_mode modes,
5624 uint32_t max_lower_array_len);
5625
5626 bool nir_lower_indirect_var_derefs(nir_shader *shader,
5627 const struct set *vars);
5628
5629 bool nir_lower_locals_to_regs(nir_shader *shader, uint8_t bool_bitsize);
5630
5631 bool nir_lower_io_to_temporaries(nir_shader *shader,
5632 nir_function_impl *entrypoint,
5633 bool outputs, bool inputs);
5634
5635 bool nir_lower_vars_to_scratch(nir_shader *shader,
5636 nir_variable_mode modes,
5637 int size_threshold,
5638 glsl_type_size_align_func variable_size_align,
5639 glsl_type_size_align_func scratch_layout_size_align);
5640
5641 bool nir_lower_scratch_to_var(nir_shader *nir);
5642
5643 void nir_lower_clip_halfz(nir_shader *shader);
5644
5645 void nir_shader_gather_info(nir_shader *shader, nir_function_impl *entrypoint);
5646
5647 void nir_gather_types(nir_function_impl *impl,
5648 BITSET_WORD *float_types,
5649 BITSET_WORD *int_types);
5650
5651 typedef struct {
5652 /* Whether all invocations write tess level outputs.
5653 *
5654 * This is useful when a pass wants to read tess level values at the end
5655 * of the shader. If this is true, the pass doesn't have to insert a barrier
5656 * and use output loads, it can just use the SSA defs that are being stored
5657 * (or phis thereof) to get the tess level output values.
5658 */
5659 bool all_invocations_define_tess_levels;
5660
5661 /* Whether any of the outer tess level components is effectively 0, meaning
5662 * that the shader discards the patch. NaNs and negative values are included
5663 * in this. If the patch is discarded, inner tess levels have no effect.
5664 */
5665 bool all_tess_levels_are_effectively_zero;
5666
5667 /* Whether all tess levels are effectively 1, meaning that the tessellator
5668 * behaves as if they were 1. There is a range of values that lead to that
5669 * behavior depending on the tessellation spacing.
5670 */
5671 bool all_tess_levels_are_effectively_one;
5672
5673 /* Whether the shader uses a barrier synchronizing TCS output stores.
5674 * For example, passes that write an output at the beginning of the shader
5675 * and load it at the end can use this to determine whether they have to
5676 * insert a barrier or whether the shader already contains a barrier.
5677 */
5678 bool always_executes_barrier;
5679
5680 /* Whether outer tess levels <= 0 are written anywhere in the shader. */
5681 bool discards_patches;
5682 } nir_tcs_info;
5683
5684 void
5685 nir_gather_tcs_info(const nir_shader *nir, nir_tcs_info *info,
5686 enum tess_primitive_mode prim,
5687 enum gl_tess_spacing spacing);
5688
5689 void nir_assign_var_locations(nir_shader *shader, nir_variable_mode mode,
5690 unsigned *size,
5691 int (*type_size)(const struct glsl_type *, bool));
5692
5693 /* Some helpers to do very simple linking */
5694 bool nir_remove_unused_varyings(nir_shader *producer, nir_shader *consumer);
5695 bool nir_remove_unused_io_vars(nir_shader *shader, nir_variable_mode mode,
5696 uint64_t *used_by_other_stage,
5697 uint64_t *used_by_other_stage_patches);
5698 void nir_compact_varyings(nir_shader *producer, nir_shader *consumer,
5699 bool default_to_smooth_interp);
5700 void nir_link_xfb_varyings(nir_shader *producer, nir_shader *consumer);
5701 bool nir_link_opt_varyings(nir_shader *producer, nir_shader *consumer);
5702 void nir_link_varying_precision(nir_shader *producer, nir_shader *consumer);
5703 nir_variable *nir_clone_uniform_variable(nir_shader *nir,
5704 nir_variable *uniform, bool spirv);
5705 nir_deref_instr *nir_clone_deref_instr(struct nir_builder *b,
5706 nir_variable *var,
5707 nir_deref_instr *deref);
5708
5709
5710 /* Return status from nir_opt_varyings. */
5711 typedef enum {
5712 /* Whether the IR changed such that NIR optimizations should be run, such
5713 * as due to removal of loads and stores. IO semantic changes such as
5714 * compaction don't count as IR changes because they don't affect NIR
5715 * optimizations.
5716 */
5717 nir_progress_producer = BITFIELD_BIT(0),
5718 nir_progress_consumer = BITFIELD_BIT(1),
5719 } nir_opt_varyings_progress;
5720
5721 nir_opt_varyings_progress
5722 nir_opt_varyings(nir_shader *producer, nir_shader *consumer, bool spirv,
5723 unsigned max_uniform_components, unsigned max_ubos_per_stage);
5724
5725 bool nir_slot_is_sysval_output(gl_varying_slot slot,
5726 gl_shader_stage next_shader);
5727 bool nir_slot_is_varying(gl_varying_slot slot, gl_shader_stage next_shader);
5728 bool nir_slot_is_sysval_output_and_varying(gl_varying_slot slot,
5729 gl_shader_stage next_shader);
5730 bool nir_remove_varying(nir_intrinsic_instr *intr, gl_shader_stage next_shader);
5731 bool nir_remove_sysval_output(nir_intrinsic_instr *intr, gl_shader_stage next_shader);
5732
5733 bool nir_lower_amul(nir_shader *shader,
5734 int (*type_size)(const struct glsl_type *, bool));
5735
5736 bool nir_lower_ubo_vec4(nir_shader *shader);
5737
5738 void nir_sort_variables_by_location(nir_shader *shader, nir_variable_mode mode);
5739 void nir_assign_io_var_locations(nir_shader *shader,
5740 nir_variable_mode mode,
5741 unsigned *size,
5742 gl_shader_stage stage);
5743
5744 bool nir_opt_clip_cull_const(nir_shader *shader);
5745
5746 typedef enum {
5747 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
5748 * to 32-bit operations. This is only valid for nir_var_shader_in/out
5749 * modes.
5750 *
5751 * Note that this destroys dual-slot information i.e. whether an input
5752 * occupies the low or high half of dvec4. Instead, it adds an offset of 1
5753 * to the load (which is ambiguous) and expects driver locations of inputs
5754 * to be final, which prevents any further optimizations.
5755 *
5756 * TODO: remove this in favor of nir_lower_io_lower_64bit_to_32_new.
5757 */
5758 nir_lower_io_lower_64bit_to_32 = (1 << 0),
5759
5760 /* If set, this causes the subset of 64-bit IO operations involving floats to be lowered on-the-fly
5761 * to 32-bit operations. This is only valid for nir_var_shader_in/out
5762 * modes.
5763 */
5764 nir_lower_io_lower_64bit_float_to_32 = (1 << 1),
5765
5766 /* This causes all 64-bit IO operations to be lowered to 32-bit operations.
5767 * This is only valid for nir_var_shader_in/out modes.
5768 *
5769 * Only VS inputs: Dual slot information is preserved as nir_io_semantics::
5770 * high_dvec2 and gathered into shader_info::dual_slot_inputs, so that
5771 * the shader can be arbitrarily optimized and the low or high half of
5772 * dvec4 can be DCE'd independently without affecting the other half.
5773 */
5774 nir_lower_io_lower_64bit_to_32_new = (1 << 2),
5775
5776 /**
5777 * Should nir_lower_io() create load_interpolated_input intrinsics?
5778 *
5779 * If not, it generates regular load_input intrinsics and interpolation
5780 * information must be inferred from the list of input nir_variables.
5781 */
5782 nir_lower_io_use_interpolated_input_intrinsics = (1 << 3),
5783 } nir_lower_io_options;
5784 bool nir_lower_io(nir_shader *shader,
5785 nir_variable_mode modes,
5786 int (*type_size)(const struct glsl_type *, bool),
5787 nir_lower_io_options);
5788
5789 bool nir_io_add_const_offset_to_base(nir_shader *nir, nir_variable_mode modes);
5790 bool nir_lower_color_inputs(nir_shader *nir);
5791 void nir_lower_io_passes(nir_shader *nir, bool renumber_vs_inputs);
5792 bool nir_io_add_intrinsic_xfb_info(nir_shader *nir);
5793
5794 bool
5795 nir_lower_vars_to_explicit_types(nir_shader *shader,
5796 nir_variable_mode modes,
5797 glsl_type_size_align_func type_info);
5798 void
5799 nir_gather_explicit_io_initializers(nir_shader *shader,
5800 void *dst, size_t dst_size,
5801 nir_variable_mode mode);
5802
5803 bool nir_lower_vec3_to_vec4(nir_shader *shader, nir_variable_mode modes);
5804
5805 typedef enum {
5806 /**
5807 * An address format which is a simple 32-bit global GPU address.
5808 */
5809 nir_address_format_32bit_global,
5810
5811 /**
5812 * An address format which is a simple 64-bit global GPU address.
5813 */
5814 nir_address_format_64bit_global,
5815
5816 /**
5817 * An address format which is a 64-bit global GPU address encoded as a
5818 * 2x32-bit vector.
5819 */
5820 nir_address_format_2x32bit_global,
5821
5822 /**
5823 * An address format which is a 64-bit global base address and a 32-bit
5824 * offset.
5825 *
5826 * This is identical to 64bit_bounded_global except that bounds checking
5827 * is not applied when lowering to global access. Even though the size is
5828 * never used for an actual bounds check, it needs to be valid so we can
5829 * lower deref_buffer_array_length properly.
5830 */
5831 nir_address_format_64bit_global_32bit_offset,
5832
5833 /**
5834 * An address format which is a bounds-checked 64-bit global GPU address.
5835 *
5836 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
5837 * address stored with the low bits in .x and high bits in .y, .z is a
5838 * size, and .w is an offset. When the final I/O operation is lowered, .w
5839 * is checked against .z and the operation is predicated on the result.
5840 */
5841 nir_address_format_64bit_bounded_global,
5842
5843 /**
5844 * An address format which is comprised of a vec2 where the first
5845 * component is a buffer index and the second is an offset.
5846 */
5847 nir_address_format_32bit_index_offset,
5848
5849 /**
5850 * An address format which is a 64-bit value, where the high 32 bits
5851 * are a buffer index, and the low 32 bits are an offset.
5852 */
5853 nir_address_format_32bit_index_offset_pack64,
5854
5855 /**
5856 * An address format which is comprised of a vec3 where the first two
5857 * components specify the buffer and the third is an offset.
5858 */
5859 nir_address_format_vec2_index_32bit_offset,
5860
5861 /**
5862 * An address format which represents generic pointers with a 62-bit
5863 * pointer and a 2-bit enum in the top two bits. The top two bits have
5864 * the following meanings:
5865 *
5866 * - 0x0: Global memory
5867 * - 0x1: Shared memory
5868 * - 0x2: Scratch memory
5869 * - 0x3: Global memory
5870 *
5871 * The redundancy between 0x0 and 0x3 is because of Intel sign-extension of
5872 * addresses. Valid global memory addresses may naturally have either 0 or
5873 * ~0 as their high bits.
5874 *
5875 * Shared and scratch pointers are represented as 32-bit offsets with the
5876 * top 32 bits only being used for the enum. This allows us to avoid
5877 * 64-bit address calculations in a bunch of cases.
5878 */
5879 nir_address_format_62bit_generic,
5880
5881 /**
5882 * An address format which is a simple 32-bit offset.
5883 */
5884 nir_address_format_32bit_offset,
5885
5886 /**
5887 * An address format which is a simple 32-bit offset cast to 64-bit.
5888 */
5889 nir_address_format_32bit_offset_as_64bit,
5890
5891 /**
5892 * An address format representing a purely logical addressing model. In
5893 * this model, all deref chains must be complete from the dereference
5894 * operation to the variable. Cast derefs are not allowed. These
5895 * addresses will be 32-bit scalars but the format is immaterial because
5896 * you can always chase the chain.
5897 */
5898 nir_address_format_logical,
5899 } nir_address_format;
5900
5901 unsigned
5902 nir_address_format_bit_size(nir_address_format addr_format);
5903
5904 unsigned
5905 nir_address_format_num_components(nir_address_format addr_format);
5906
5907 static inline const struct glsl_type *
nir_address_format_to_glsl_type(nir_address_format addr_format)5908 nir_address_format_to_glsl_type(nir_address_format addr_format)
5909 {
5910 unsigned bit_size = nir_address_format_bit_size(addr_format);
5911 assert(bit_size == 32 || bit_size == 64);
5912 return glsl_vector_type(bit_size == 32 ? GLSL_TYPE_UINT : GLSL_TYPE_UINT64,
5913 nir_address_format_num_components(addr_format));
5914 }
5915
5916 const nir_const_value *nir_address_format_null_value(nir_address_format addr_format);
5917
5918 nir_def *nir_build_addr_iadd(struct nir_builder *b, nir_def *addr,
5919 nir_address_format addr_format,
5920 nir_variable_mode modes,
5921 nir_def *offset);
5922
5923 nir_def *nir_build_addr_iadd_imm(struct nir_builder *b, nir_def *addr,
5924 nir_address_format addr_format,
5925 nir_variable_mode modes,
5926 int64_t offset);
5927
5928 nir_def *nir_build_addr_ieq(struct nir_builder *b, nir_def *addr0, nir_def *addr1,
5929 nir_address_format addr_format);
5930
5931 nir_def *nir_build_addr_isub(struct nir_builder *b, nir_def *addr0, nir_def *addr1,
5932 nir_address_format addr_format);
5933
5934 nir_def *nir_explicit_io_address_from_deref(struct nir_builder *b,
5935 nir_deref_instr *deref,
5936 nir_def *base_addr,
5937 nir_address_format addr_format);
5938
5939 bool nir_get_explicit_deref_align(nir_deref_instr *deref,
5940 bool default_to_type_align,
5941 uint32_t *align_mul,
5942 uint32_t *align_offset);
5943
5944 void nir_lower_explicit_io_instr(struct nir_builder *b,
5945 nir_intrinsic_instr *io_instr,
5946 nir_def *addr,
5947 nir_address_format addr_format);
5948
5949 bool nir_lower_explicit_io(nir_shader *shader,
5950 nir_variable_mode modes,
5951 nir_address_format);
5952
5953 typedef enum {
5954 /* Use open-coded funnel shifts for each component. */
5955 nir_mem_access_shift_method_scalar,
5956 /* Prefer to use 64-bit shifts to do the same with less instructions. Useful
5957 * if 64-bit shifts are cheap.
5958 */
5959 nir_mem_access_shift_method_shift64,
5960 /* If nir_op_alignbyte_amd can be used, this is the best option with just a
5961 * single nir_op_alignbyte_amd for each 32-bit components.
5962 */
5963 nir_mem_access_shift_method_bytealign_amd,
5964 } nir_mem_access_shift_method;
5965
5966 typedef struct {
5967 uint8_t num_components;
5968 uint8_t bit_size;
5969 uint16_t align;
5970 /* If a load's alignment is increased, this specifies how the data should be
5971 * shifted before converting to the original bit size.
5972 */
5973 nir_mem_access_shift_method shift;
5974 } nir_mem_access_size_align;
5975
5976 /* clang-format off */
5977 typedef nir_mem_access_size_align
5978 (*nir_lower_mem_access_bit_sizes_cb)(nir_intrinsic_op intrin,
5979 uint8_t bytes,
5980 uint8_t bit_size,
5981 uint32_t align_mul,
5982 uint32_t align_offset,
5983 bool offset_is_const,
5984 enum gl_access_qualifier,
5985 const void *cb_data);
5986 /* clang-format on */
5987
5988 typedef struct {
5989 nir_lower_mem_access_bit_sizes_cb callback;
5990 nir_variable_mode modes;
5991 bool may_lower_unaligned_stores_to_atomics;
5992 void *cb_data;
5993 } nir_lower_mem_access_bit_sizes_options;
5994
5995 bool nir_lower_mem_access_bit_sizes(nir_shader *shader,
5996 const nir_lower_mem_access_bit_sizes_options *options);
5997
5998 bool nir_lower_robust_access(nir_shader *s,
5999 nir_intrin_filter_cb filter, const void *data);
6000
6001 /* clang-format off */
6002 typedef bool (*nir_should_vectorize_mem_func)(unsigned align_mul,
6003 unsigned align_offset,
6004 unsigned bit_size,
6005 unsigned num_components,
6006 /* The hole between low and
6007 * high if they are not adjacent. */
6008 int64_t hole_size,
6009 nir_intrinsic_instr *low,
6010 nir_intrinsic_instr *high,
6011 void *data);
6012 /* clang-format on */
6013
6014 typedef struct {
6015 nir_should_vectorize_mem_func callback;
6016 nir_variable_mode modes;
6017 nir_variable_mode robust_modes;
6018 void *cb_data;
6019 bool has_shared2_amd;
6020 } nir_load_store_vectorize_options;
6021
6022 bool nir_opt_load_store_vectorize(nir_shader *shader, const nir_load_store_vectorize_options *options);
6023 bool nir_opt_load_store_update_alignments(nir_shader *shader);
6024
6025 typedef bool (*nir_lower_shader_calls_should_remat_func)(nir_instr *instr, void *data);
6026
6027 typedef struct nir_lower_shader_calls_options {
6028 /* Address format used for load/store operations on the call stack. */
6029 nir_address_format address_format;
6030
6031 /* Stack alignment */
6032 unsigned stack_alignment;
6033
6034 /* Put loads from the stack as close as possible from where they're needed.
6035 * You might want to disable combined_loads for best effects.
6036 */
6037 bool localized_loads;
6038
6039 /* If this function pointer is not NULL, lower_shader_calls will run
6040 * nir_opt_load_store_vectorize for stack load/store operations. Otherwise
6041 * the optimizaion is not run.
6042 */
6043 nir_should_vectorize_mem_func vectorizer_callback;
6044
6045 /* Data passed to vectorizer_callback */
6046 void *vectorizer_data;
6047
6048 /* If this function pointer is not NULL, lower_shader_calls will call this
6049 * function on instructions that require spill/fill/rematerialization of
6050 * their value. If this function returns true, lower_shader_calls will
6051 * ensure that the instruction is rematerialized, adding the sources of the
6052 * instruction to be spilled/filled.
6053 */
6054 nir_lower_shader_calls_should_remat_func should_remat_callback;
6055
6056 /* Data passed to should_remat_callback */
6057 void *should_remat_data;
6058 } nir_lower_shader_calls_options;
6059
6060 bool
6061 nir_lower_shader_calls(nir_shader *shader,
6062 const nir_lower_shader_calls_options *options,
6063 nir_shader ***resume_shaders_out,
6064 uint32_t *num_resume_shaders_out,
6065 void *mem_ctx);
6066
6067 int nir_get_io_offset_src_number(const nir_intrinsic_instr *instr);
6068 int nir_get_io_arrayed_index_src_number(const nir_intrinsic_instr *instr);
6069
6070 nir_src *nir_get_io_offset_src(nir_intrinsic_instr *instr);
6071 nir_src *nir_get_io_arrayed_index_src(nir_intrinsic_instr *instr);
6072 nir_src *nir_get_shader_call_payload_src(nir_intrinsic_instr *call);
6073
6074 bool nir_is_arrayed_io(const nir_variable *var, gl_shader_stage stage);
6075
6076 bool nir_lower_reg_intrinsics_to_ssa_impl(nir_function_impl *impl);
6077 bool nir_lower_reg_intrinsics_to_ssa(nir_shader *shader);
6078 bool nir_lower_vars_to_ssa(nir_shader *shader);
6079
6080 bool nir_remove_dead_derefs(nir_shader *shader);
6081 bool nir_remove_dead_derefs_impl(nir_function_impl *impl);
6082
6083 typedef struct nir_remove_dead_variables_options {
6084 bool (*can_remove_var)(nir_variable *var, void *data);
6085 void *can_remove_var_data;
6086 } nir_remove_dead_variables_options;
6087
6088 bool nir_remove_dead_variables(nir_shader *shader, nir_variable_mode modes,
6089 const nir_remove_dead_variables_options *options);
6090
6091 bool nir_lower_variable_initializers(nir_shader *shader,
6092 nir_variable_mode modes);
6093 bool nir_zero_initialize_shared_memory(nir_shader *shader,
6094 const unsigned shared_size,
6095 const unsigned chunk_size);
6096 bool nir_clear_shared_memory(nir_shader *shader,
6097 const unsigned shared_size,
6098 const unsigned chunk_size);
6099
6100 bool nir_move_vec_src_uses_to_dest(nir_shader *shader, bool skip_const_srcs);
6101 bool nir_move_output_stores_to_end(nir_shader *nir);
6102 bool nir_lower_vec_to_regs(nir_shader *shader, nir_instr_writemask_filter_cb cb,
6103 const void *_data);
6104 bool nir_lower_alpha_test(nir_shader *shader, enum compare_func func,
6105 bool alpha_to_one,
6106 const gl_state_index16 *alpha_ref_state_tokens);
6107 bool nir_lower_alu(nir_shader *shader);
6108
6109 bool nir_lower_flrp(nir_shader *shader, unsigned lowering_mask,
6110 bool always_precise);
6111
6112 bool nir_scale_fdiv(nir_shader *shader);
6113
6114 bool nir_lower_alu_to_scalar(nir_shader *shader, nir_instr_filter_cb cb, const void *data);
6115 bool nir_lower_alu_width(nir_shader *shader, nir_vectorize_cb cb, const void *data);
6116 bool nir_lower_alu_vec8_16_srcs(nir_shader *shader);
6117 bool nir_lower_bool_to_bitsize(nir_shader *shader);
6118 bool nir_lower_bool_to_float(nir_shader *shader, bool has_fcsel_ne);
6119 bool nir_lower_bool_to_int32(nir_shader *shader);
6120 bool nir_opt_simplify_convert_alu_types(nir_shader *shader);
6121 bool nir_lower_const_arrays_to_uniforms(nir_shader *shader,
6122 unsigned max_uniform_components);
6123 bool nir_lower_convert_alu_types(nir_shader *shader,
6124 bool (*should_lower)(nir_intrinsic_instr *));
6125 bool nir_lower_constant_convert_alu_types(nir_shader *shader);
6126 bool nir_lower_alu_conversion_to_intrinsic(nir_shader *shader);
6127 bool nir_lower_int_to_float(nir_shader *shader);
6128 bool nir_lower_load_const_to_scalar(nir_shader *shader);
6129 bool nir_lower_read_invocation_to_scalar(nir_shader *shader);
6130 bool nir_lower_phis_to_scalar(nir_shader *shader, bool lower_all);
6131 void nir_lower_io_arrays_to_elements(nir_shader *producer, nir_shader *consumer);
6132 bool nir_lower_io_arrays_to_elements_no_indirects(nir_shader *shader,
6133 bool outputs_only);
6134 bool nir_lower_io_to_scalar(nir_shader *shader, nir_variable_mode mask, nir_instr_filter_cb filter, void *filter_data);
6135 bool nir_lower_io_to_scalar_early(nir_shader *shader, nir_variable_mode mask);
6136 bool nir_lower_io_to_vector(nir_shader *shader, nir_variable_mode mask);
6137 bool nir_vectorize_tess_levels(nir_shader *shader);
6138 nir_shader *nir_create_passthrough_tcs_impl(const nir_shader_compiler_options *options,
6139 unsigned *locations, unsigned num_locations,
6140 uint8_t patch_vertices);
6141 nir_shader *nir_create_passthrough_tcs(const nir_shader_compiler_options *options,
6142 const nir_shader *vs, uint8_t patch_vertices);
6143 nir_shader *nir_create_passthrough_gs(const nir_shader_compiler_options *options,
6144 const nir_shader *prev_stage,
6145 enum mesa_prim primitive_type,
6146 enum mesa_prim output_primitive_type,
6147 bool emulate_edgeflags,
6148 bool force_line_strip_out,
6149 bool passthrough_prim_id);
6150
6151 bool nir_lower_fragcolor(nir_shader *shader, unsigned max_cbufs);
6152 bool nir_lower_fragcoord_wtrans(nir_shader *shader);
6153 bool nir_opt_frag_coord_to_pixel_coord(nir_shader *shader);
6154 bool nir_lower_frag_coord_to_pixel_coord(nir_shader *shader);
6155 bool nir_lower_viewport_transform(nir_shader *shader);
6156 bool nir_lower_uniforms_to_ubo(nir_shader *shader, bool dword_packed, bool load_vec4);
6157
6158 bool nir_lower_is_helper_invocation(nir_shader *shader);
6159
6160 bool nir_lower_single_sampled(nir_shader *shader);
6161
6162 bool nir_lower_atomics(nir_shader *shader, nir_instr_filter_cb filter);
6163
6164 typedef struct nir_lower_subgroups_options {
6165 /* In addition to the boolean lowering options below, this optional callback
6166 * will filter instructions for lowering if non-NULL. The data passed will be
6167 * filter_data.
6168 */
6169 nir_instr_filter_cb filter;
6170
6171 /* Extra data passed to the filter. */
6172 const void *filter_data;
6173
6174 /* In case the exact subgroup size is not known, subgroup_size should be
6175 * set to 0. In that case, the maximum subgroup size will be calculated by
6176 * ballot_components * ballot_bit_size.
6177 */
6178 uint8_t subgroup_size;
6179 uint8_t ballot_bit_size;
6180 uint8_t ballot_components;
6181 bool lower_to_scalar : 1;
6182 bool lower_vote_trivial : 1;
6183 bool lower_vote_eq : 1;
6184 bool lower_vote_bool_eq : 1;
6185 bool lower_first_invocation_to_ballot : 1;
6186 bool lower_read_first_invocation : 1;
6187 bool lower_subgroup_masks : 1;
6188 bool lower_relative_shuffle : 1;
6189 bool lower_shuffle_to_32bit : 1;
6190 bool lower_shuffle_to_swizzle_amd : 1;
6191 bool lower_shuffle : 1;
6192 bool lower_quad : 1;
6193 bool lower_quad_broadcast_dynamic : 1;
6194 bool lower_quad_broadcast_dynamic_to_const : 1;
6195 bool lower_quad_vote : 1;
6196 bool lower_elect : 1;
6197 bool lower_read_invocation_to_cond : 1;
6198 bool lower_rotate_to_shuffle : 1;
6199 bool lower_rotate_clustered_to_shuffle : 1;
6200 bool lower_ballot_bit_count_to_mbcnt_amd : 1;
6201 bool lower_inverse_ballot : 1;
6202 bool lower_reduce : 1;
6203 bool lower_boolean_reduce : 1;
6204 bool lower_boolean_shuffle : 1;
6205 } nir_lower_subgroups_options;
6206
6207 bool nir_lower_subgroups(nir_shader *shader,
6208 const nir_lower_subgroups_options *options);
6209
6210 bool nir_lower_system_values(nir_shader *shader);
6211
6212 nir_def *
6213 nir_build_lowered_load_helper_invocation(struct nir_builder *b);
6214
6215 typedef struct nir_lower_compute_system_values_options {
6216 bool has_base_global_invocation_id : 1;
6217 bool has_base_workgroup_id : 1;
6218 bool has_global_size : 1;
6219 bool shuffle_local_ids_for_quad_derivatives : 1;
6220 bool lower_local_invocation_index : 1;
6221 bool lower_cs_local_id_to_index : 1;
6222 bool lower_workgroup_id_to_index : 1;
6223 bool global_id_is_32bit : 1;
6224 /* At shader execution time, check if WorkGroupId should be 1D
6225 * and compute it quickly. Fall back to slow computation if not.
6226 */
6227 bool shortcut_1d_workgroup_id : 1;
6228 uint32_t num_workgroups[3]; /* Compile-time-known dispatch sizes, or 0 if unknown. */
6229 } nir_lower_compute_system_values_options;
6230
6231 bool nir_lower_compute_system_values(nir_shader *shader,
6232 const nir_lower_compute_system_values_options *options);
6233
6234 struct nir_lower_sysvals_to_varyings_options {
6235 bool frag_coord : 1;
6236 bool front_face : 1;
6237 bool point_coord : 1;
6238 };
6239
6240 bool
6241 nir_lower_sysvals_to_varyings(nir_shader *shader,
6242 const struct nir_lower_sysvals_to_varyings_options *options);
6243
6244 /***/
6245 enum ENUM_PACKED nir_lower_tex_packing {
6246 /** No packing */
6247 nir_lower_tex_packing_none = 0,
6248 /**
6249 * The sampler returns up to 2 32-bit words of half floats or 16-bit signed
6250 * or unsigned ints based on the sampler type
6251 */
6252 nir_lower_tex_packing_16,
6253 /** The sampler returns 1 32-bit word of 4x8 unorm */
6254 nir_lower_tex_packing_8,
6255 };
6256
6257 /***/
6258 typedef struct nir_lower_tex_options {
6259 /**
6260 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
6261 * sampler types a texture projector is lowered.
6262 */
6263 unsigned lower_txp;
6264
6265 /**
6266 * If true, lower texture projector for any array sampler dims
6267 */
6268 bool lower_txp_array;
6269
6270 /**
6271 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
6272 */
6273 bool lower_txf_offset;
6274
6275 /**
6276 * If true, lower away nir_tex_src_offset for all rect textures.
6277 */
6278 bool lower_rect_offset;
6279
6280 /**
6281 * If not NULL, this filter will return true for tex instructions that
6282 * should lower away nir_tex_src_offset.
6283 */
6284 nir_instr_filter_cb lower_offset_filter;
6285
6286 /**
6287 * If true, lower rect textures to 2D, using txs to fetch the
6288 * texture dimensions and dividing the texture coords by the
6289 * texture dims to normalize.
6290 */
6291 bool lower_rect;
6292
6293 /**
6294 * If true, lower 1D textures to 2D. This requires the GL/VK driver to map 1D
6295 * textures to 2D textures with height=1.
6296 *
6297 * lower_1d_shadow does this lowering for shadow textures only.
6298 */
6299 bool lower_1d;
6300 bool lower_1d_shadow;
6301
6302 /**
6303 * If true, convert yuv to rgb.
6304 */
6305 unsigned lower_y_uv_external;
6306 unsigned lower_y_vu_external;
6307 unsigned lower_y_u_v_external;
6308 unsigned lower_yx_xuxv_external;
6309 unsigned lower_yx_xvxu_external;
6310 unsigned lower_xy_uxvx_external;
6311 unsigned lower_xy_vxux_external;
6312 unsigned lower_ayuv_external;
6313 unsigned lower_xyuv_external;
6314 unsigned lower_yuv_external;
6315 unsigned lower_yu_yv_external;
6316 unsigned lower_yv_yu_external;
6317 unsigned lower_y41x_external;
6318 unsigned bt709_external;
6319 unsigned bt2020_external;
6320 unsigned yuv_full_range_external;
6321
6322 /**
6323 * To emulate certain texture wrap modes, this can be used
6324 * to saturate the specified tex coord to [0.0, 1.0]. The
6325 * bits are according to sampler #, ie. if, for example:
6326 *
6327 * (conf->saturate_s & (1 << n))
6328 *
6329 * is true, then the s coord for sampler n is saturated.
6330 *
6331 * Note that clamping must happen *after* projector lowering
6332 * so any projected texture sample instruction with a clamped
6333 * coordinate gets automatically lowered, regardless of the
6334 * 'lower_txp' setting.
6335 */
6336 unsigned saturate_s;
6337 unsigned saturate_t;
6338 unsigned saturate_r;
6339
6340 /* Bitmask of textures that need swizzling.
6341 *
6342 * If (swizzle_result & (1 << texture_index)), then the swizzle in
6343 * swizzles[texture_index] is applied to the result of the texturing
6344 * operation.
6345 */
6346 unsigned swizzle_result;
6347
6348 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
6349 * while 4 and 5 represent 0 and 1 respectively.
6350 *
6351 * Indexed by texture-id.
6352 */
6353 uint8_t swizzles[32][4];
6354
6355 /* Can be used to scale sampled values in range required by the
6356 * format.
6357 *
6358 * Indexed by texture-id.
6359 */
6360 float scale_factors[32];
6361
6362 /**
6363 * Bitmap of textures that need srgb to linear conversion. If
6364 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
6365 * of the texture are lowered to linear.
6366 */
6367 unsigned lower_srgb;
6368
6369 /**
6370 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
6371 */
6372 bool lower_txd_cube_map;
6373
6374 /**
6375 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
6376 */
6377 bool lower_txd_3d;
6378
6379 /**
6380 * If true, lower nir_texop_txd any array surfaces with nir_texop_txl.
6381 */
6382 bool lower_txd_array;
6383
6384 /**
6385 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
6386 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
6387 * with lower_txd_cube_map.
6388 */
6389 bool lower_txd_shadow;
6390
6391 /**
6392 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
6393 * Implies lower_txd_cube_map and lower_txd_shadow.
6394 */
6395 bool lower_txd;
6396
6397 /**
6398 * If true, lower nir_texop_txd when it uses min_lod.
6399 */
6400 bool lower_txd_clamp;
6401
6402 /**
6403 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
6404 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
6405 */
6406 bool lower_txb_shadow_clamp;
6407
6408 /**
6409 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
6410 * with nir_texop_txl. This includes cube maps.
6411 */
6412 bool lower_txd_shadow_clamp;
6413
6414 /**
6415 * If true, lower nir_texop_txd on when it uses both offset and min_lod
6416 * with nir_texop_txl. This includes cube maps.
6417 */
6418 bool lower_txd_offset_clamp;
6419
6420 /**
6421 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
6422 * sampler is bindless.
6423 */
6424 bool lower_txd_clamp_bindless_sampler;
6425
6426 /**
6427 * If true, lower nir_texop_txd with min_lod to a nir_texop_txl if the
6428 * sampler index is not statically determinable to be less than 16.
6429 */
6430 bool lower_txd_clamp_if_sampler_index_not_lt_16;
6431
6432 /**
6433 * If true, lower nir_texop_txs with a non-0-lod into nir_texop_txs with
6434 * 0-lod followed by a nir_ishr.
6435 */
6436 bool lower_txs_lod;
6437
6438 /**
6439 * If true, lower nir_texop_txs for cube arrays to a nir_texop_txs with a
6440 * 2D array type followed by a nir_idiv by 6.
6441 */
6442 bool lower_txs_cube_array;
6443
6444 /**
6445 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
6446 * mixed-up tg4 locations.
6447 */
6448 bool lower_tg4_broadcom_swizzle;
6449
6450 /**
6451 * If true, lowers tg4 with 4 constant offsets to 4 tg4 calls
6452 */
6453 bool lower_tg4_offsets;
6454
6455 /**
6456 * Lower txf_ms to fragment_mask_fetch and fragment_fetch and samples_identical to
6457 * fragment_mask_fetch.
6458 */
6459 bool lower_to_fragment_fetch_amd;
6460
6461 /**
6462 * To lower packed sampler return formats. This will be called for all
6463 * tex instructions.
6464 */
6465 enum nir_lower_tex_packing (*lower_tex_packing_cb)(const nir_tex_instr *tex, const void *data);
6466 const void *lower_tex_packing_data;
6467
6468 /**
6469 * If true, lower nir_texop_lod to return -FLT_MAX if the sum of the
6470 * absolute values of derivatives is 0 for all coordinates.
6471 */
6472 bool lower_lod_zero_width;
6473
6474 /* Turns nir_op_tex and other ops with an implicit derivative, in stages
6475 * without implicit derivatives (like the vertex shader) to have an explicit
6476 * LOD with a value of 0.
6477 */
6478 bool lower_invalid_implicit_lod;
6479
6480 /* If true, texture_index (sampler_index) will be zero if a texture_offset
6481 * (sampler_offset) source is present. This is convenient for backends that
6482 * support indirect indexing of textures (samplers) but not offsetting it.
6483 */
6484 bool lower_index_to_offset;
6485
6486 /**
6487 * Payload data to be sent to callback / filter functions.
6488 */
6489 void *callback_data;
6490 } nir_lower_tex_options;
6491
6492 /** Lowers complex texture instructions to simpler ones */
6493 bool nir_lower_tex(nir_shader *shader,
6494 const nir_lower_tex_options *options);
6495
6496 typedef struct nir_lower_tex_shadow_swizzle {
6497 unsigned swizzle_r : 3;
6498 unsigned swizzle_g : 3;
6499 unsigned swizzle_b : 3;
6500 unsigned swizzle_a : 3;
6501 } nir_lower_tex_shadow_swizzle;
6502
6503 bool
6504 nir_lower_tex_shadow(nir_shader *s,
6505 unsigned n_states,
6506 enum compare_func *compare_func,
6507 nir_lower_tex_shadow_swizzle *tex_swizzles,
6508 bool is_fixed_point_format);
6509
6510 typedef struct nir_lower_image_options {
6511 /**
6512 * If true, lower cube size operations.
6513 */
6514 bool lower_cube_size;
6515
6516 /**
6517 * Lower multi sample image load and samples_identical to use fragment_mask_load.
6518 */
6519 bool lower_to_fragment_mask_load_amd;
6520
6521 /**
6522 * Lower image_samples to a constant in case the driver doesn't support multisampled
6523 * images.
6524 */
6525 bool lower_image_samples_to_one;
6526 } nir_lower_image_options;
6527
6528 bool nir_lower_image(nir_shader *nir,
6529 const nir_lower_image_options *options);
6530
6531 bool
6532 nir_lower_image_atomics_to_global(nir_shader *s);
6533
6534 bool nir_lower_readonly_images_to_tex(nir_shader *shader, bool per_variable);
6535
6536 enum nir_lower_non_uniform_access_type {
6537 nir_lower_non_uniform_ubo_access = (1 << 0),
6538 nir_lower_non_uniform_ssbo_access = (1 << 1),
6539 nir_lower_non_uniform_texture_access = (1 << 2),
6540 nir_lower_non_uniform_image_access = (1 << 3),
6541 nir_lower_non_uniform_get_ssbo_size = (1 << 4),
6542 nir_lower_non_uniform_access_type_count = 5,
6543 };
6544
6545 /* Given the nir_src used for the resource, return the channels which might be non-uniform. */
6546 typedef nir_component_mask_t (*nir_lower_non_uniform_access_callback)(const nir_src *, void *);
6547
6548 typedef struct nir_lower_non_uniform_access_options {
6549 enum nir_lower_non_uniform_access_type types;
6550 nir_lower_non_uniform_access_callback callback;
6551 void *callback_data;
6552 } nir_lower_non_uniform_access_options;
6553
6554 bool nir_has_non_uniform_access(nir_shader *shader, enum nir_lower_non_uniform_access_type types);
6555 bool nir_opt_non_uniform_access(nir_shader *shader);
6556 bool nir_lower_non_uniform_access(nir_shader *shader,
6557 const nir_lower_non_uniform_access_options *options);
6558
6559 typedef struct {
6560 /* Whether 16-bit floating point arithmetic should be allowed in 8-bit
6561 * division lowering
6562 */
6563 bool allow_fp16;
6564 } nir_lower_idiv_options;
6565
6566 bool nir_lower_idiv(nir_shader *shader, const nir_lower_idiv_options *options);
6567
6568 typedef struct nir_input_attachment_options {
6569 bool use_fragcoord_sysval;
6570 bool use_layer_id_sysval;
6571 bool use_view_id_for_layer;
6572 bool unscaled_depth_stencil_ir3;
6573 uint32_t unscaled_input_attachment_ir3;
6574 } nir_input_attachment_options;
6575
6576 bool nir_lower_input_attachments(nir_shader *shader,
6577 const nir_input_attachment_options *options);
6578
6579 bool nir_lower_clip_vs(nir_shader *shader, unsigned ucp_enables,
6580 bool use_vars,
6581 bool use_clipdist_array,
6582 const gl_state_index16 clipplane_state_tokens[][STATE_LENGTH]);
6583 bool nir_lower_clip_gs(nir_shader *shader, unsigned ucp_enables,
6584 bool use_clipdist_array,
6585 const gl_state_index16 clipplane_state_tokens[][STATE_LENGTH]);
6586 bool nir_lower_clip_fs(nir_shader *shader, unsigned ucp_enables,
6587 bool use_clipdist_array, bool use_load_interp);
6588
6589 bool nir_lower_clip_cull_distance_to_vec4s(nir_shader *shader);
6590 bool nir_lower_clip_cull_distance_arrays(nir_shader *nir);
6591 bool nir_lower_clip_disable(nir_shader *shader, unsigned clip_plane_enable);
6592
6593 bool nir_lower_point_size_mov(nir_shader *shader,
6594 const gl_state_index16 *pointsize_state_tokens);
6595
6596 bool nir_lower_frexp(nir_shader *nir);
6597
6598 bool nir_lower_two_sided_color(nir_shader *shader, bool face_sysval);
6599
6600 bool nir_lower_clamp_color_outputs(nir_shader *shader);
6601
6602 bool nir_lower_flatshade(nir_shader *shader);
6603
6604 bool nir_lower_passthrough_edgeflags(nir_shader *shader);
6605 bool nir_lower_patch_vertices(nir_shader *nir, unsigned static_count,
6606 const gl_state_index16 *uniform_state_tokens);
6607
6608 typedef struct nir_lower_wpos_ytransform_options {
6609 gl_state_index16 state_tokens[STATE_LENGTH];
6610 bool fs_coord_origin_upper_left : 1;
6611 bool fs_coord_origin_lower_left : 1;
6612 bool fs_coord_pixel_center_integer : 1;
6613 bool fs_coord_pixel_center_half_integer : 1;
6614 } nir_lower_wpos_ytransform_options;
6615
6616 bool nir_lower_wpos_ytransform(nir_shader *shader,
6617 const nir_lower_wpos_ytransform_options *options);
6618 bool nir_lower_wpos_center(nir_shader *shader);
6619
6620 bool nir_lower_pntc_ytransform(nir_shader *shader,
6621 const gl_state_index16 clipplane_state_tokens[][STATE_LENGTH]);
6622
6623 bool nir_lower_wrmasks(nir_shader *shader, nir_instr_filter_cb cb, const void *data);
6624
6625 bool nir_lower_fb_read(nir_shader *shader);
6626
6627 typedef struct nir_lower_drawpixels_options {
6628 gl_state_index16 texcoord_state_tokens[STATE_LENGTH];
6629 gl_state_index16 scale_state_tokens[STATE_LENGTH];
6630 gl_state_index16 bias_state_tokens[STATE_LENGTH];
6631 unsigned drawpix_sampler;
6632 unsigned pixelmap_sampler;
6633 bool pixel_maps : 1;
6634 bool scale_and_bias : 1;
6635 } nir_lower_drawpixels_options;
6636
6637 bool nir_lower_drawpixels(nir_shader *shader,
6638 const nir_lower_drawpixels_options *options);
6639
6640 typedef struct nir_lower_bitmap_options {
6641 unsigned sampler;
6642 bool swizzle_xxxx;
6643 } nir_lower_bitmap_options;
6644
6645 bool nir_lower_bitmap(nir_shader *shader, const nir_lower_bitmap_options *options);
6646
6647 bool nir_lower_atomics_to_ssbo(nir_shader *shader, unsigned offset_align_state);
6648
6649 typedef enum {
6650 nir_lower_gs_intrinsics_per_stream = 1 << 0,
6651 nir_lower_gs_intrinsics_count_primitives = 1 << 1,
6652 nir_lower_gs_intrinsics_count_vertices_per_primitive = 1 << 2,
6653 nir_lower_gs_intrinsics_overwrite_incomplete = 1 << 3,
6654 nir_lower_gs_intrinsics_always_end_primitive = 1 << 4,
6655 nir_lower_gs_intrinsics_count_decomposed_primitives = 1 << 5,
6656 } nir_lower_gs_intrinsics_flags;
6657
6658 bool nir_lower_gs_intrinsics(nir_shader *shader, nir_lower_gs_intrinsics_flags options);
6659
6660 bool nir_lower_tess_coord_z(nir_shader *shader, bool triangles);
6661
6662 typedef struct {
6663 bool payload_to_shared_for_atomics : 1;
6664 bool payload_to_shared_for_small_types : 1;
6665 uint32_t payload_offset_in_bytes;
6666 } nir_lower_task_shader_options;
6667
6668 bool nir_lower_task_shader(nir_shader *shader, nir_lower_task_shader_options options);
6669
6670 typedef unsigned (*nir_lower_bit_size_callback)(const nir_instr *, void *);
6671
6672 bool nir_lower_bit_size(nir_shader *shader,
6673 nir_lower_bit_size_callback callback,
6674 void *callback_data);
6675 bool nir_lower_64bit_phis(nir_shader *shader);
6676
6677 bool nir_split_64bit_vec3_and_vec4(nir_shader *shader);
6678
6679 nir_lower_int64_options nir_lower_int64_op_to_options_mask(nir_op opcode);
6680 bool nir_lower_int64(nir_shader *shader);
6681 bool nir_lower_int64_float_conversions(nir_shader *shader);
6682
6683 nir_lower_doubles_options nir_lower_doubles_op_to_options_mask(nir_op opcode);
6684 bool nir_lower_doubles(nir_shader *shader, const nir_shader *softfp64,
6685 nir_lower_doubles_options options);
6686 bool nir_lower_pack(nir_shader *shader);
6687
6688 bool nir_recompute_io_bases(nir_shader *nir, nir_variable_mode modes);
6689 bool nir_lower_mediump_vars(nir_shader *nir, nir_variable_mode modes);
6690 bool nir_lower_mediump_io(nir_shader *nir, nir_variable_mode modes,
6691 uint64_t varying_mask, bool use_16bit_slots);
6692 bool nir_force_mediump_io(nir_shader *nir, nir_variable_mode modes,
6693 nir_alu_type types);
6694 bool nir_unpack_16bit_varying_slots(nir_shader *nir, nir_variable_mode modes);
6695
6696 struct nir_opt_tex_srcs_options {
6697 unsigned sampler_dims;
6698 unsigned src_types;
6699 };
6700
6701 struct nir_opt_16bit_tex_image_options {
6702 nir_rounding_mode rounding_mode;
6703 nir_alu_type opt_tex_dest_types;
6704 nir_alu_type opt_image_dest_types;
6705 bool integer_dest_saturates;
6706 bool opt_image_store_data;
6707 bool opt_image_srcs;
6708 unsigned opt_srcs_options_count;
6709 struct nir_opt_tex_srcs_options *opt_srcs_options;
6710 };
6711
6712 bool nir_opt_16bit_tex_image(nir_shader *nir,
6713 struct nir_opt_16bit_tex_image_options *options);
6714
6715 typedef struct {
6716 bool legalize_type; /* whether this src should be legalized */
6717 uint8_t bit_size; /* bit_size to enforce */
6718 nir_tex_src_type match_src; /* if bit_size is 0, match bit size of this */
6719 } nir_tex_src_type_constraint, nir_tex_src_type_constraints[nir_num_tex_src_types];
6720
6721 bool nir_legalize_16bit_sampler_srcs(nir_shader *nir,
6722 nir_tex_src_type_constraints constraints);
6723
6724 bool nir_lower_point_size(nir_shader *shader, float min, float max);
6725
6726 void nir_lower_texcoord_replace(nir_shader *s, unsigned coord_replace,
6727 bool point_coord_is_sysval, bool yinvert);
6728
6729 bool nir_lower_texcoord_replace_late(nir_shader *s, unsigned coord_replace,
6730 bool point_coord_is_sysval);
6731
6732 typedef enum {
6733 nir_lower_interpolation_at_sample = (1 << 1),
6734 nir_lower_interpolation_at_offset = (1 << 2),
6735 nir_lower_interpolation_centroid = (1 << 3),
6736 nir_lower_interpolation_pixel = (1 << 4),
6737 nir_lower_interpolation_sample = (1 << 5),
6738 } nir_lower_interpolation_options;
6739
6740 bool nir_lower_interpolation(nir_shader *shader,
6741 nir_lower_interpolation_options options);
6742
6743 typedef enum {
6744 nir_lower_discard_if_to_cf = (1 << 0),
6745 nir_lower_demote_if_to_cf = (1 << 1),
6746 nir_lower_terminate_if_to_cf = (1 << 2),
6747 } nir_lower_discard_if_options;
6748
6749 bool nir_lower_discard_if(nir_shader *shader, nir_lower_discard_if_options options);
6750
6751 bool nir_lower_terminate_to_demote(nir_shader *nir);
6752
6753 bool nir_lower_memory_model(nir_shader *shader);
6754
6755 bool nir_lower_goto_ifs(nir_shader *shader);
6756 bool nir_lower_continue_constructs(nir_shader *shader);
6757
6758 typedef struct nir_lower_multiview_options {
6759 uint32_t view_mask;
6760
6761 /**
6762 * Bitfield of output locations that may be converted to a per-view array.
6763 *
6764 * If a variable exists in an allowed location, it will be converted to an
6765 * array even if its value does not depend on the view index.
6766 */
6767 uint64_t allowed_per_view_outputs;
6768 } nir_lower_multiview_options;
6769
6770 bool nir_shader_uses_view_index(nir_shader *shader);
6771 bool nir_can_lower_multiview(nir_shader *shader, nir_lower_multiview_options options);
6772 bool nir_lower_multiview(nir_shader *shader, nir_lower_multiview_options options);
6773
6774 bool nir_lower_view_index_to_device_index(nir_shader *shader);
6775
6776 typedef enum {
6777 nir_lower_fp16_rtz = (1 << 0),
6778 nir_lower_fp16_rtne = (1 << 1),
6779 nir_lower_fp16_ru = (1 << 2),
6780 nir_lower_fp16_rd = (1 << 3),
6781 nir_lower_fp16_all = 0xf,
6782 nir_lower_fp16_split_fp64 = (1 << 4),
6783 } nir_lower_fp16_cast_options;
6784 bool nir_lower_fp16_casts(nir_shader *shader, nir_lower_fp16_cast_options options);
6785 bool nir_normalize_cubemap_coords(nir_shader *shader);
6786
6787 bool nir_shader_supports_implicit_lod(nir_shader *shader);
6788
6789 void nir_live_defs_impl(nir_function_impl *impl);
6790
6791 const BITSET_WORD *nir_get_live_defs(nir_cursor cursor, void *mem_ctx);
6792
6793 void nir_loop_analyze_impl(nir_function_impl *impl,
6794 nir_variable_mode indirect_mask,
6795 bool force_unroll_sampler_indirect);
6796
6797 /* This requires both nir_metadata_live_defs and nir_metadata_instr_index. */
6798 bool nir_defs_interfere(nir_def *a, nir_def *b);
6799
6800 bool nir_repair_ssa_impl(nir_function_impl *impl);
6801 bool nir_repair_ssa(nir_shader *shader);
6802
6803 void nir_convert_loop_to_lcssa(nir_loop *loop);
6804 bool nir_convert_to_lcssa(nir_shader *shader, bool skip_invariants, bool skip_bool_invariants);
6805 void nir_divergence_analysis_impl(nir_function_impl *impl, nir_divergence_options options);
6806 void nir_divergence_analysis(nir_shader *shader);
6807 void nir_vertex_divergence_analysis(nir_shader *shader);
6808 bool nir_has_divergent_loop(nir_shader *shader);
6809
6810 void
6811 nir_rewrite_uses_to_load_reg(struct nir_builder *b, nir_def *old,
6812 nir_def *reg);
6813
6814 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
6815 * registers. If false, convert all values (even those not involved in a phi
6816 * node) to registers.
6817 * If consider_divergence is true, this pass will use divergence information
6818 * in order to not coalesce copies from uniform to divergent registers.
6819 */
6820 bool nir_convert_from_ssa(nir_shader *shader,
6821 bool phi_webs_only, bool consider_divergence);
6822
6823 bool nir_lower_phis_to_regs_block(nir_block *block);
6824 bool nir_lower_ssa_defs_to_regs_block(nir_block *block);
6825
6826 bool nir_rematerialize_deref_in_use_blocks(nir_deref_instr *instr);
6827 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl *impl);
6828
6829 bool nir_lower_samplers(nir_shader *shader);
6830 bool nir_lower_cl_images(nir_shader *shader, bool lower_image_derefs, bool lower_sampler_derefs);
6831 bool nir_dedup_inline_samplers(nir_shader *shader);
6832
6833 typedef struct nir_lower_ssbo_options {
6834 bool native_loads;
6835 bool native_offset;
6836 } nir_lower_ssbo_options;
6837
6838 bool nir_lower_ssbo(nir_shader *shader, const nir_lower_ssbo_options *opts);
6839
6840 bool nir_lower_helper_writes(nir_shader *shader, bool lower_plain_stores);
6841
6842 typedef struct nir_lower_printf_options {
6843 unsigned max_buffer_size;
6844 unsigned ptr_bit_size;
6845 bool use_printf_base_identifier;
6846 bool hash_format_strings;
6847 } nir_lower_printf_options;
6848
6849 bool nir_lower_printf(nir_shader *nir, const nir_lower_printf_options *options);
6850 bool nir_lower_printf_buffer(nir_shader *nir, uint64_t address, uint32_t size);
6851
6852 /* This is here for unit tests. */
6853 bool nir_opt_comparison_pre_impl(nir_function_impl *impl);
6854
6855 bool nir_opt_comparison_pre(nir_shader *shader);
6856
6857 typedef struct nir_opt_access_options {
6858 bool is_vulkan;
6859 } nir_opt_access_options;
6860
6861 bool nir_opt_access(nir_shader *shader, const nir_opt_access_options *options);
6862 bool nir_opt_algebraic(nir_shader *shader);
6863 bool nir_opt_algebraic_before_ffma(nir_shader *shader);
6864 bool nir_opt_algebraic_before_lower_int64(nir_shader *shader);
6865 bool nir_opt_algebraic_late(nir_shader *shader);
6866 bool nir_opt_algebraic_distribute_src_mods(nir_shader *shader);
6867 bool nir_opt_constant_folding(nir_shader *shader);
6868
6869 /* Try to combine a and b into a. Return true if combination was possible,
6870 * which will result in b being removed by the pass. Return false if
6871 * combination wasn't possible.
6872 */
6873 typedef bool (*nir_combine_barrier_cb)(
6874 nir_intrinsic_instr *a, nir_intrinsic_instr *b, void *data);
6875
6876 bool nir_opt_combine_barriers(nir_shader *shader,
6877 nir_combine_barrier_cb combine_cb,
6878 void *data);
6879 bool nir_opt_barrier_modes(nir_shader *shader);
6880
6881 bool nir_minimize_call_live_states(nir_shader *shader);
6882
6883 bool nir_opt_combine_stores(nir_shader *shader, nir_variable_mode modes);
6884
6885 bool nir_copy_prop_impl(nir_function_impl *impl);
6886 bool nir_copy_prop(nir_shader *shader);
6887
6888 bool nir_opt_copy_prop_vars(nir_shader *shader);
6889
6890 bool nir_opt_cse(nir_shader *shader);
6891
6892 bool nir_opt_dce(nir_shader *shader);
6893
6894 bool nir_opt_dead_cf(nir_shader *shader);
6895
6896 bool nir_opt_dead_write_vars(nir_shader *shader);
6897
6898 bool nir_opt_deref_impl(nir_function_impl *impl);
6899 bool nir_opt_deref(nir_shader *shader);
6900
6901 bool nir_opt_find_array_copies(nir_shader *shader);
6902
6903 bool nir_def_is_frag_coord_z(nir_def *def);
6904 bool nir_opt_fragdepth(nir_shader *shader);
6905
6906 bool nir_opt_gcm(nir_shader *shader, bool value_number);
6907
6908 bool nir_opt_generate_bfi(nir_shader *shader);
6909
6910 bool nir_opt_idiv_const(nir_shader *shader, unsigned min_bit_size);
6911
6912 bool nir_opt_mqsad(nir_shader *shader);
6913
6914 typedef enum {
6915 nir_opt_if_optimize_phi_true_false = (1 << 0),
6916 nir_opt_if_avoid_64bit_phis = (1 << 1),
6917 } nir_opt_if_options;
6918
6919 bool nir_opt_if(nir_shader *shader, nir_opt_if_options options);
6920
6921 bool nir_opt_intrinsics(nir_shader *shader);
6922
6923 bool nir_opt_large_constants(nir_shader *shader,
6924 glsl_type_size_align_func size_align,
6925 unsigned threshold);
6926
6927 bool nir_opt_licm(nir_shader *shader);
6928 bool nir_opt_loop(nir_shader *shader);
6929
6930 bool nir_opt_loop_unroll(nir_shader *shader);
6931
6932 typedef enum {
6933 nir_move_const_undef = (1 << 0),
6934 nir_move_load_ubo = (1 << 1),
6935 nir_move_load_input = (1 << 2),
6936 nir_move_comparisons = (1 << 3),
6937 nir_move_copies = (1 << 4),
6938 nir_move_load_ssbo = (1 << 5),
6939 nir_move_load_uniform = (1 << 6),
6940 nir_move_alu = (1 << 7),
6941 } nir_move_options;
6942
6943 bool nir_can_move_instr(nir_instr *instr, nir_move_options options);
6944
6945 bool nir_opt_sink(nir_shader *shader, nir_move_options options);
6946
6947 bool nir_opt_move(nir_shader *shader, nir_move_options options);
6948
6949 typedef struct {
6950 /** nir_load_uniform max base offset */
6951 uint32_t uniform_max;
6952
6953 /** nir_load_ubo_vec4 max base offset */
6954 uint32_t ubo_vec4_max;
6955
6956 /** nir_var_mem_shared max base offset */
6957 uint32_t shared_max;
6958
6959 /** nir_var_mem_shared atomic max base offset */
6960 uint32_t shared_atomic_max;
6961
6962 /** nir_load/store_buffer_amd max base offset */
6963 uint32_t buffer_max;
6964
6965 /**
6966 * Callback to get the max base offset for instructions for which the
6967 * corresponding value above is zero.
6968 */
6969 uint32_t (*max_offset_cb)(nir_intrinsic_instr *intr, const void *data);
6970
6971 /** Data to pass to max_offset_cb. */
6972 const void *max_offset_data;
6973
6974 /**
6975 * Allow the offset calculation to wrap. If false, constant additions that
6976 * might wrap will not be folded into the offset.
6977 */
6978 bool allow_offset_wrap;
6979 } nir_opt_offsets_options;
6980
6981 bool nir_opt_offsets(nir_shader *shader, const nir_opt_offsets_options *options);
6982
6983 bool nir_opt_peephole_select(nir_shader *shader, unsigned limit,
6984 bool indirect_load_ok, bool expensive_alu_ok);
6985
6986 bool nir_opt_reassociate_bfi(nir_shader *shader);
6987
6988 bool nir_opt_rematerialize_compares(nir_shader *shader);
6989
6990 bool nir_opt_remove_phis(nir_shader *shader);
6991 bool nir_remove_single_src_phis_block(nir_block *block);
6992
6993 bool nir_opt_phi_precision(nir_shader *shader);
6994
6995 bool nir_opt_shrink_stores(nir_shader *shader, bool shrink_image_store);
6996
6997 bool nir_opt_shrink_vectors(nir_shader *shader, bool shrink_start);
6998
6999 bool nir_opt_undef(nir_shader *shader);
7000
7001 bool nir_lower_undef_to_zero(nir_shader *shader);
7002
7003 bool nir_opt_uniform_atomics(nir_shader *shader, bool fs_atomics_predicated);
7004
7005 bool nir_opt_uniform_subgroup(nir_shader *shader,
7006 const nir_lower_subgroups_options *);
7007
7008 bool nir_opt_vectorize(nir_shader *shader, nir_vectorize_cb filter,
7009 void *data);
7010 bool nir_opt_vectorize_io(nir_shader *shader, nir_variable_mode modes);
7011
7012 bool nir_opt_conditional_discard(nir_shader *shader);
7013 bool nir_opt_move_discards_to_top(nir_shader *shader);
7014
7015 bool nir_opt_ray_queries(nir_shader *shader);
7016
7017 bool nir_opt_ray_query_ranges(nir_shader *shader);
7018
7019 void nir_sweep(nir_shader *shader);
7020
7021 nir_intrinsic_op nir_intrinsic_from_system_value(gl_system_value val);
7022 gl_system_value nir_system_value_from_intrinsic(nir_intrinsic_op intrin);
7023
7024 static inline bool
nir_variable_is_in_ubo(const nir_variable * var)7025 nir_variable_is_in_ubo(const nir_variable *var)
7026 {
7027 return (var->data.mode == nir_var_mem_ubo &&
7028 var->interface_type != NULL);
7029 }
7030
7031 static inline bool
nir_variable_is_in_ssbo(const nir_variable * var)7032 nir_variable_is_in_ssbo(const nir_variable *var)
7033 {
7034 return (var->data.mode == nir_var_mem_ssbo &&
7035 var->interface_type != NULL);
7036 }
7037
7038 static inline bool
nir_variable_is_in_block(const nir_variable * var)7039 nir_variable_is_in_block(const nir_variable *var)
7040 {
7041 return nir_variable_is_in_ubo(var) || nir_variable_is_in_ssbo(var);
7042 }
7043
7044 static inline unsigned
nir_variable_count_slots(const nir_variable * var,const struct glsl_type * type)7045 nir_variable_count_slots(const nir_variable *var, const struct glsl_type *type)
7046 {
7047 return var->data.compact ? DIV_ROUND_UP(var->data.location_frac + glsl_get_length(type), 4) : glsl_count_attribute_slots(type, false);
7048 }
7049
7050 static inline unsigned
nir_deref_count_slots(nir_deref_instr * deref,nir_variable * var)7051 nir_deref_count_slots(nir_deref_instr *deref, nir_variable *var)
7052 {
7053 if (var->data.compact) {
7054 switch (deref->deref_type) {
7055 case nir_deref_type_array:
7056 return 1;
7057 case nir_deref_type_var:
7058 return nir_variable_count_slots(var, deref->type);
7059 default:
7060 unreachable("illegal deref type");
7061 }
7062 }
7063 return glsl_count_attribute_slots(deref->type, false);
7064 }
7065
7066 /* See default_ub_config in nir_range_analysis.c for documentation. */
7067 typedef struct nir_unsigned_upper_bound_config {
7068 unsigned min_subgroup_size;
7069 unsigned max_subgroup_size;
7070 unsigned max_workgroup_invocations;
7071 unsigned max_workgroup_count[3];
7072 unsigned max_workgroup_size[3];
7073
7074 uint32_t vertex_attrib_max[32];
7075 } nir_unsigned_upper_bound_config;
7076
7077 uint32_t
7078 nir_unsigned_upper_bound(nir_shader *shader, struct hash_table *range_ht,
7079 nir_scalar scalar,
7080 const nir_unsigned_upper_bound_config *config);
7081
7082 bool
7083 nir_addition_might_overflow(nir_shader *shader, struct hash_table *range_ht,
7084 nir_scalar ssa, unsigned const_val,
7085 const nir_unsigned_upper_bound_config *config);
7086
7087 typedef struct {
7088 /* True if gl_DrawID is considered uniform, i.e. if the preamble is run
7089 * at least once per "internal" draw rather than per user-visible draw.
7090 */
7091 bool drawid_uniform;
7092
7093 /* True if the subgroup size is uniform. */
7094 bool subgroup_size_uniform;
7095
7096 /* True if load_workgroup_size is supported in the preamble. */
7097 bool load_workgroup_size_allowed;
7098
7099 /* size/align for load/store_preamble. */
7100 void (*def_size)(nir_def *def, unsigned *size, unsigned *align);
7101
7102 /* Total available size for load/store_preamble storage, in units
7103 * determined by def_size.
7104 */
7105 unsigned preamble_storage_size;
7106
7107 /* Give the cost for an instruction. nir_opt_preamble will prioritize
7108 * instructions with higher costs. Instructions with cost 0 may still be
7109 * lifted, but only when required to lift other instructions with non-0
7110 * cost (e.g. a load_const source of an expression).
7111 */
7112 float (*instr_cost_cb)(nir_instr *instr, const void *data);
7113
7114 /* Give the cost of rewriting the instruction to use load_preamble. This
7115 * may happen from inserting move instructions, etc. If the benefit doesn't
7116 * exceed the cost here then we won't rewrite it.
7117 */
7118 float (*rewrite_cost_cb)(nir_def *def, const void *data);
7119
7120 /* Instructions whose definitions should not be rewritten. These could
7121 * still be moved to the preamble, but they shouldn't be the root of a
7122 * replacement expression. Instructions with cost 0 and derefs are
7123 * automatically included by the pass.
7124 */
7125 nir_instr_filter_cb avoid_instr_cb;
7126
7127 const void *cb_data;
7128 } nir_opt_preamble_options;
7129
7130 bool
7131 nir_opt_preamble(nir_shader *shader,
7132 const nir_opt_preamble_options *options,
7133 unsigned *size);
7134
7135 nir_function_impl *nir_shader_get_preamble(nir_shader *shader);
7136
7137 bool nir_lower_point_smooth(nir_shader *shader, bool set_barycentrics);
7138 bool nir_lower_poly_line_smooth(nir_shader *shader, unsigned num_smooth_aa_sample);
7139
7140 bool nir_mod_analysis(nir_scalar val, nir_alu_type val_type, unsigned div, unsigned *mod);
7141
7142 bool
7143 nir_remove_tex_shadow(nir_shader *shader, unsigned textures_bitmask);
7144
7145 void
7146 nir_trivialize_registers(nir_shader *s);
7147
7148 unsigned
7149 nir_static_workgroup_size(const nir_shader *s);
7150
7151 static inline nir_intrinsic_instr *
nir_reg_get_decl(nir_def * reg)7152 nir_reg_get_decl(nir_def *reg)
7153 {
7154 assert(reg->parent_instr->type == nir_instr_type_intrinsic);
7155 nir_intrinsic_instr *decl = nir_instr_as_intrinsic(reg->parent_instr);
7156 assert(decl->intrinsic == nir_intrinsic_decl_reg);
7157
7158 return decl;
7159 }
7160
7161 static inline nir_intrinsic_instr *
nir_next_decl_reg(nir_intrinsic_instr * prev,nir_function_impl * impl)7162 nir_next_decl_reg(nir_intrinsic_instr *prev, nir_function_impl *impl)
7163 {
7164 nir_instr *start;
7165 if (prev != NULL)
7166 start = nir_instr_next(&prev->instr);
7167 else if (impl != NULL)
7168 start = nir_block_first_instr(nir_start_block(impl));
7169 else
7170 return NULL;
7171
7172 for (nir_instr *instr = start; instr; instr = nir_instr_next(instr)) {
7173 if (instr->type != nir_instr_type_intrinsic)
7174 continue;
7175
7176 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
7177 if (intrin->intrinsic == nir_intrinsic_decl_reg)
7178 return intrin;
7179 }
7180
7181 return NULL;
7182 }
7183
7184 #define nir_foreach_reg_decl(reg, impl) \
7185 for (nir_intrinsic_instr *reg = nir_next_decl_reg(NULL, impl); \
7186 reg; reg = nir_next_decl_reg(reg, NULL))
7187
7188 #define nir_foreach_reg_decl_safe(reg, impl) \
7189 for (nir_intrinsic_instr *reg = nir_next_decl_reg(NULL, impl), \
7190 *next_ = nir_next_decl_reg(reg, NULL); \
7191 reg; reg = next_, next_ = nir_next_decl_reg(next_, NULL))
7192
7193 static inline nir_cursor
nir_after_reg_decls(nir_function_impl * impl)7194 nir_after_reg_decls(nir_function_impl *impl)
7195 {
7196 nir_intrinsic_instr *last_reg_decl = NULL;
7197 nir_foreach_reg_decl(reg_decl, impl)
7198 last_reg_decl = reg_decl;
7199
7200 if (last_reg_decl != NULL)
7201 return nir_after_instr(&last_reg_decl->instr);
7202 return nir_before_impl(impl);
7203 }
7204
7205 static inline bool
nir_is_load_reg(nir_intrinsic_instr * intr)7206 nir_is_load_reg(nir_intrinsic_instr *intr)
7207 {
7208 return intr->intrinsic == nir_intrinsic_load_reg ||
7209 intr->intrinsic == nir_intrinsic_load_reg_indirect;
7210 }
7211
7212 static inline bool
nir_is_store_reg(nir_intrinsic_instr * intr)7213 nir_is_store_reg(nir_intrinsic_instr *intr)
7214 {
7215 return intr->intrinsic == nir_intrinsic_store_reg ||
7216 intr->intrinsic == nir_intrinsic_store_reg_indirect;
7217 }
7218
7219 #define nir_foreach_reg_load(load, reg) \
7220 assert(reg->intrinsic == nir_intrinsic_decl_reg); \
7221 \
7222 nir_foreach_use(load, ®->def) \
7223 if (nir_is_load_reg(nir_instr_as_intrinsic(nir_src_parent_instr(load))))
7224
7225 #define nir_foreach_reg_load_safe(load, reg) \
7226 assert(reg->intrinsic == nir_intrinsic_decl_reg); \
7227 \
7228 nir_foreach_use_safe(load, ®->def) \
7229 if (nir_is_load_reg(nir_instr_as_intrinsic(nir_src_parent_instr(load))))
7230
7231 #define nir_foreach_reg_store(store, reg) \
7232 assert(reg->intrinsic == nir_intrinsic_decl_reg); \
7233 \
7234 nir_foreach_use(store, ®->def) \
7235 if (nir_is_store_reg(nir_instr_as_intrinsic(nir_src_parent_instr(store))))
7236
7237 #define nir_foreach_reg_store_safe(store, reg) \
7238 assert(reg->intrinsic == nir_intrinsic_decl_reg); \
7239 \
7240 nir_foreach_use_safe(store, ®->def) \
7241 if (nir_is_store_reg(nir_instr_as_intrinsic(nir_src_parent_instr(store))))
7242
7243 static inline nir_intrinsic_instr *
nir_load_reg_for_def(const nir_def * def)7244 nir_load_reg_for_def(const nir_def *def)
7245 {
7246 if (def->parent_instr->type != nir_instr_type_intrinsic)
7247 return NULL;
7248
7249 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(def->parent_instr);
7250 if (!nir_is_load_reg(intr))
7251 return NULL;
7252
7253 return intr;
7254 }
7255
7256 static inline nir_intrinsic_instr *
nir_store_reg_for_def(const nir_def * def)7257 nir_store_reg_for_def(const nir_def *def)
7258 {
7259 /* Look for the trivial store: single use of our destination by a
7260 * store_register intrinsic.
7261 */
7262 if (!list_is_singular(&def->uses))
7263 return NULL;
7264
7265 nir_src *src = list_first_entry(&def->uses, nir_src, use_link);
7266 if (nir_src_is_if(src))
7267 return NULL;
7268
7269 nir_instr *parent = nir_src_parent_instr(src);
7270 if (parent->type != nir_instr_type_intrinsic)
7271 return NULL;
7272
7273 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(parent);
7274 if (!nir_is_store_reg(intr))
7275 return NULL;
7276
7277 /* The first value is data. Third is indirect index, ignore that one. */
7278 if (&intr->src[0] != src)
7279 return NULL;
7280
7281 return intr;
7282 }
7283
7284 struct nir_use_dominance_state;
7285
7286 struct nir_use_dominance_state *
7287 nir_calc_use_dominance_impl(nir_function_impl *impl, bool post_dominance);
7288
7289 nir_instr *
7290 nir_get_immediate_use_dominator(struct nir_use_dominance_state *state,
7291 nir_instr *instr);
7292 nir_instr *nir_use_dominance_lca(struct nir_use_dominance_state *state,
7293 nir_instr *i1, nir_instr *i2);
7294 bool nir_instr_dominates_use(struct nir_use_dominance_state *state,
7295 nir_instr *parent, nir_instr *child);
7296 void nir_print_use_dominators(struct nir_use_dominance_state *state,
7297 nir_instr **instructions,
7298 unsigned num_instructions);
7299
7300 #include "nir_inline_helpers.h"
7301
7302 #ifdef __cplusplus
7303 } /* extern "C" */
7304 #endif
7305
7306 #endif /* NIR_H */
7307