nir: allow specifying filter callback in lower_alu_to_scalar
[mesa.git] / src / compiler / nir / nir.h
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 "util/hash_table.h"
32 #include "compiler/glsl/list.h"
33 #include "GL/gl.h" /* GLenum */
34 #include "util/list.h"
35 #include "util/ralloc.h"
36 #include "util/set.h"
37 #include "util/bitscan.h"
38 #include "util/bitset.h"
39 #include "util/macros.h"
40 #include "compiler/nir_types.h"
41 #include "compiler/shader_enums.h"
42 #include "compiler/shader_info.h"
43 #include <stdio.h>
44
45 #ifndef NDEBUG
46 #include "util/debug.h"
47 #endif /* NDEBUG */
48
49 #include "nir_opcodes.h"
50
51 #if defined(_WIN32) && !defined(snprintf)
52 #define snprintf _snprintf
53 #endif
54
55 #ifdef __cplusplus
56 extern "C" {
57 #endif
58
59 #define NIR_FALSE 0u
60 #define NIR_TRUE (~0u)
61 #define NIR_MAX_VEC_COMPONENTS 4
62 #define NIR_MAX_MATRIX_COLUMNS 4
63 typedef uint8_t nir_component_mask_t;
64
65 /** Defines a cast function
66 *
67 * This macro defines a cast function from in_type to out_type where
68 * out_type is some structure type that contains a field of type out_type.
69 *
70 * Note that you have to be a bit careful as the generated cast function
71 * destroys constness.
72 */
73 #define NIR_DEFINE_CAST(name, in_type, out_type, field, \
74 type_field, type_value) \
75 static inline out_type * \
76 name(const in_type *parent) \
77 { \
78 assert(parent && parent->type_field == type_value); \
79 return exec_node_data(out_type, parent, field); \
80 }
81
82 struct nir_function;
83 struct nir_shader;
84 struct nir_instr;
85 struct nir_builder;
86
87
88 /**
89 * Description of built-in state associated with a uniform
90 *
91 * \sa nir_variable::state_slots
92 */
93 typedef struct {
94 gl_state_index16 tokens[STATE_LENGTH];
95 int swizzle;
96 } nir_state_slot;
97
98 typedef enum {
99 nir_var_shader_in = (1 << 0),
100 nir_var_shader_out = (1 << 1),
101 nir_var_shader_temp = (1 << 2),
102 nir_var_function_temp = (1 << 3),
103 nir_var_uniform = (1 << 4),
104 nir_var_mem_ubo = (1 << 5),
105 nir_var_system_value = (1 << 6),
106 nir_var_mem_ssbo = (1 << 7),
107 nir_var_mem_shared = (1 << 8),
108 nir_var_mem_global = (1 << 9),
109 nir_var_all = ~0,
110 } nir_variable_mode;
111
112 /**
113 * Rounding modes.
114 */
115 typedef enum {
116 nir_rounding_mode_undef = 0,
117 nir_rounding_mode_rtne = 1, /* round to nearest even */
118 nir_rounding_mode_ru = 2, /* round up */
119 nir_rounding_mode_rd = 3, /* round down */
120 nir_rounding_mode_rtz = 4, /* round towards zero */
121 } nir_rounding_mode;
122
123 typedef union {
124 bool b;
125 float f32;
126 double f64;
127 int8_t i8;
128 uint8_t u8;
129 int16_t i16;
130 uint16_t u16;
131 int32_t i32;
132 uint32_t u32;
133 int64_t i64;
134 uint64_t u64;
135 } nir_const_value;
136
137 #define nir_const_value_to_array(arr, c, components, m) \
138 { \
139 for (unsigned i = 0; i < components; ++i) \
140 arr[i] = c[i].m; \
141 } while (false)
142
143 static inline nir_const_value
144 nir_const_value_for_raw_uint(uint64_t x, unsigned bit_size)
145 {
146 nir_const_value v;
147 memset(&v, 0, sizeof(v));
148
149 switch (bit_size) {
150 case 1: v.b = x; break;
151 case 8: v.u8 = x; break;
152 case 16: v.u16 = x; break;
153 case 32: v.u32 = x; break;
154 case 64: v.u64 = x; break;
155 default:
156 unreachable("Invalid bit size");
157 }
158
159 return v;
160 }
161
162 static inline nir_const_value
163 nir_const_value_for_int(int64_t i, unsigned bit_size)
164 {
165 nir_const_value v;
166 memset(&v, 0, sizeof(v));
167
168 assert(bit_size <= 64);
169 if (bit_size < 64) {
170 assert(i >= (-(1ll << (bit_size - 1))));
171 assert(i < (1ll << (bit_size - 1)));
172 }
173
174 return nir_const_value_for_raw_uint(i, bit_size);
175 }
176
177 static inline nir_const_value
178 nir_const_value_for_uint(uint64_t u, unsigned bit_size)
179 {
180 nir_const_value v;
181 memset(&v, 0, sizeof(v));
182
183 assert(bit_size <= 64);
184 if (bit_size < 64)
185 assert(u < (1ull << bit_size));
186
187 return nir_const_value_for_raw_uint(u, bit_size);
188 }
189
190 static inline nir_const_value
191 nir_const_value_for_bool(bool b, unsigned bit_size)
192 {
193 /* Booleans use a 0/-1 convention */
194 return nir_const_value_for_int(-(int)b, bit_size);
195 }
196
197 /* This one isn't inline because it requires half-float conversion */
198 nir_const_value nir_const_value_for_float(double b, unsigned bit_size);
199
200 static inline int64_t
201 nir_const_value_as_int(nir_const_value value, unsigned bit_size)
202 {
203 switch (bit_size) {
204 /* int1_t uses 0/-1 convention */
205 case 1: return -(int)value.b;
206 case 8: return value.i8;
207 case 16: return value.i16;
208 case 32: return value.i32;
209 case 64: return value.i64;
210 default:
211 unreachable("Invalid bit size");
212 }
213 }
214
215 static inline int64_t
216 nir_const_value_as_uint(nir_const_value value, unsigned bit_size)
217 {
218 switch (bit_size) {
219 case 1: return value.b;
220 case 8: return value.u8;
221 case 16: return value.u16;
222 case 32: return value.u32;
223 case 64: return value.u64;
224 default:
225 unreachable("Invalid bit size");
226 }
227 }
228
229 static inline bool
230 nir_const_value_as_bool(nir_const_value value, unsigned bit_size)
231 {
232 int64_t i = nir_const_value_as_int(value, bit_size);
233
234 /* Booleans of any size use 0/-1 convention */
235 assert(i == 0 || i == -1);
236
237 return i;
238 }
239
240 /* This one isn't inline because it requires half-float conversion */
241 double nir_const_value_as_float(nir_const_value value, unsigned bit_size);
242
243 typedef struct nir_constant {
244 /**
245 * Value of the constant.
246 *
247 * The field used to back the values supplied by the constant is determined
248 * by the type associated with the \c nir_variable. Constants may be
249 * scalars, vectors, or matrices.
250 */
251 nir_const_value values[NIR_MAX_VEC_COMPONENTS];
252
253 /* we could get this from the var->type but makes clone *much* easier to
254 * not have to care about the type.
255 */
256 unsigned num_elements;
257
258 /* Array elements / Structure Fields */
259 struct nir_constant **elements;
260 } nir_constant;
261
262 /**
263 * \brief Layout qualifiers for gl_FragDepth.
264 *
265 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
266 * with a layout qualifier.
267 */
268 typedef enum {
269 nir_depth_layout_none, /**< No depth layout is specified. */
270 nir_depth_layout_any,
271 nir_depth_layout_greater,
272 nir_depth_layout_less,
273 nir_depth_layout_unchanged
274 } nir_depth_layout;
275
276 /**
277 * Enum keeping track of how a variable was declared.
278 */
279 typedef enum {
280 /**
281 * Normal declaration.
282 */
283 nir_var_declared_normally = 0,
284
285 /**
286 * Variable is implicitly generated by the compiler and should not be
287 * visible via the API.
288 */
289 nir_var_hidden,
290 } nir_var_declaration_type;
291
292 /**
293 * Either a uniform, global variable, shader input, or shader output. Based on
294 * ir_variable - it should be easy to translate between the two.
295 */
296
297 typedef struct nir_variable {
298 struct exec_node node;
299
300 /**
301 * Declared type of the variable
302 */
303 const struct glsl_type *type;
304
305 /**
306 * Declared name of the variable
307 */
308 char *name;
309
310 struct nir_variable_data {
311 /**
312 * Storage class of the variable.
313 *
314 * \sa nir_variable_mode
315 */
316 nir_variable_mode mode;
317
318 /**
319 * Is the variable read-only?
320 *
321 * This is set for variables declared as \c const, shader inputs,
322 * and uniforms.
323 */
324 unsigned read_only:1;
325 unsigned centroid:1;
326 unsigned sample:1;
327 unsigned patch:1;
328 unsigned invariant:1;
329
330 /**
331 * Can this variable be coalesced with another?
332 *
333 * This is set by nir_lower_io_to_temporaries to say that any
334 * copies involving this variable should stay put. Propagating it can
335 * duplicate the resulting load/store, which is not wanted, and may
336 * result in a load/store of the variable with an indirect offset which
337 * the backend may not be able to handle.
338 */
339 unsigned cannot_coalesce:1;
340
341 /**
342 * When separate shader programs are enabled, only input/outputs between
343 * the stages of a multi-stage separate program can be safely removed
344 * from the shader interface. Other input/outputs must remains active.
345 *
346 * This is also used to make sure xfb varyings that are unused by the
347 * fragment shader are not removed.
348 */
349 unsigned always_active_io:1;
350
351 /**
352 * Interpolation mode for shader inputs / outputs
353 *
354 * \sa glsl_interp_mode
355 */
356 unsigned interpolation:2;
357
358 /**
359 * If non-zero, then this variable may be packed along with other variables
360 * into a single varying slot, so this offset should be applied when
361 * accessing components. For example, an offset of 1 means that the x
362 * component of this variable is actually stored in component y of the
363 * location specified by \c location.
364 */
365 unsigned location_frac:2;
366
367 /**
368 * If true, this variable represents an array of scalars that should
369 * be tightly packed. In other words, consecutive array elements
370 * should be stored one component apart, rather than one slot apart.
371 */
372 unsigned compact:1;
373
374 /**
375 * Whether this is a fragment shader output implicitly initialized with
376 * the previous contents of the specified render target at the
377 * framebuffer location corresponding to this shader invocation.
378 */
379 unsigned fb_fetch_output:1;
380
381 /**
382 * Non-zero if this variable is considered bindless as defined by
383 * ARB_bindless_texture.
384 */
385 unsigned bindless:1;
386
387 /**
388 * Was an explicit binding set in the shader?
389 */
390 unsigned explicit_binding:1;
391
392 /**
393 * Was a transfer feedback buffer set in the shader?
394 */
395 unsigned explicit_xfb_buffer:1;
396
397 /**
398 * Was a transfer feedback stride set in the shader?
399 */
400 unsigned explicit_xfb_stride:1;
401
402 /**
403 * Was an explicit offset set in the shader?
404 */
405 unsigned explicit_offset:1;
406
407 /**
408 * \brief Layout qualifier for gl_FragDepth.
409 *
410 * This is not equal to \c ir_depth_layout_none if and only if this
411 * variable is \c gl_FragDepth and a layout qualifier is specified.
412 */
413 nir_depth_layout depth_layout;
414
415 /**
416 * Storage location of the base of this variable
417 *
418 * The precise meaning of this field depends on the nature of the variable.
419 *
420 * - Vertex shader input: one of the values from \c gl_vert_attrib.
421 * - Vertex shader output: one of the values from \c gl_varying_slot.
422 * - Geometry shader input: one of the values from \c gl_varying_slot.
423 * - Geometry shader output: one of the values from \c gl_varying_slot.
424 * - Fragment shader input: one of the values from \c gl_varying_slot.
425 * - Fragment shader output: one of the values from \c gl_frag_result.
426 * - Uniforms: Per-stage uniform slot number for default uniform block.
427 * - Uniforms: Index within the uniform block definition for UBO members.
428 * - Non-UBO Uniforms: uniform slot number.
429 * - Other: This field is not currently used.
430 *
431 * If the variable is a uniform, shader input, or shader output, and the
432 * slot has not been assigned, the value will be -1.
433 */
434 int location;
435
436 /**
437 * The actual location of the variable in the IR. Only valid for inputs
438 * and outputs.
439 */
440 unsigned int driver_location;
441
442 /**
443 * Vertex stream output identifier.
444 *
445 * For packed outputs, bit 31 is set and bits [2*i+1,2*i] indicate the
446 * stream of the i-th component.
447 */
448 unsigned stream;
449
450 /**
451 * output index for dual source blending.
452 */
453 int index;
454
455 /**
456 * Descriptor set binding for sampler or UBO.
457 */
458 int descriptor_set;
459
460 /**
461 * Initial binding point for a sampler or UBO.
462 *
463 * For array types, this represents the binding point for the first element.
464 */
465 int binding;
466
467 /**
468 * Location an atomic counter or transform feedback is stored at.
469 */
470 unsigned offset;
471
472 /**
473 * Transform feedback buffer.
474 */
475 unsigned xfb_buffer;
476
477 /**
478 * Transform feedback stride.
479 */
480 unsigned xfb_stride;
481
482 /**
483 * How the variable was declared. See nir_var_declaration_type.
484 *
485 * This is used to detect variables generated by the compiler, so should
486 * not be visible via the API.
487 */
488 unsigned how_declared:2;
489
490 /**
491 * ARB_shader_image_load_store qualifiers.
492 */
493 struct {
494 enum gl_access_qualifier access;
495
496 /** Image internal format if specified explicitly, otherwise GL_NONE. */
497 GLenum format;
498 } image;
499 } data;
500
501 /**
502 * Built-in state that backs this uniform
503 *
504 * Once set at variable creation, \c state_slots must remain invariant.
505 * This is because, ideally, this array would be shared by all clones of
506 * this variable in the IR tree. In other words, we'd really like for it
507 * to be a fly-weight.
508 *
509 * If the variable is not a uniform, \c num_state_slots will be zero and
510 * \c state_slots will be \c NULL.
511 */
512 /*@{*/
513 unsigned num_state_slots; /**< Number of state slots used */
514 nir_state_slot *state_slots; /**< State descriptors. */
515 /*@}*/
516
517 /**
518 * Constant expression assigned in the initializer of the variable
519 *
520 * This field should only be used temporarily by creators of NIR shaders
521 * and then lower_constant_initializers can be used to get rid of them.
522 * Most of the rest of NIR ignores this field or asserts that it's NULL.
523 */
524 nir_constant *constant_initializer;
525
526 /**
527 * For variables that are in an interface block or are an instance of an
528 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
529 *
530 * \sa ir_variable::location
531 */
532 const struct glsl_type *interface_type;
533
534 /**
535 * Description of per-member data for per-member struct variables
536 *
537 * This is used for variables which are actually an amalgamation of
538 * multiple entities such as a struct of built-in values or a struct of
539 * inputs each with their own layout specifier. This is only allowed on
540 * variables with a struct or array of array of struct type.
541 */
542 unsigned num_members;
543 struct nir_variable_data *members;
544 } nir_variable;
545
546 #define nir_foreach_variable(var, var_list) \
547 foreach_list_typed(nir_variable, var, node, var_list)
548
549 #define nir_foreach_variable_safe(var, var_list) \
550 foreach_list_typed_safe(nir_variable, var, node, var_list)
551
552 static inline bool
553 nir_variable_is_global(const nir_variable *var)
554 {
555 return var->data.mode != nir_var_function_temp;
556 }
557
558 typedef struct nir_register {
559 struct exec_node node;
560
561 unsigned num_components; /** < number of vector components */
562 unsigned num_array_elems; /** < size of array (0 for no array) */
563
564 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
565 uint8_t bit_size;
566
567 /** generic register index. */
568 unsigned index;
569
570 /** only for debug purposes, can be NULL */
571 const char *name;
572
573 /** set of nir_srcs where this register is used (read from) */
574 struct list_head uses;
575
576 /** set of nir_dests where this register is defined (written to) */
577 struct list_head defs;
578
579 /** set of nir_ifs where this register is used as a condition */
580 struct list_head if_uses;
581 } nir_register;
582
583 #define nir_foreach_register(reg, reg_list) \
584 foreach_list_typed(nir_register, reg, node, reg_list)
585 #define nir_foreach_register_safe(reg, reg_list) \
586 foreach_list_typed_safe(nir_register, reg, node, reg_list)
587
588 typedef enum PACKED {
589 nir_instr_type_alu,
590 nir_instr_type_deref,
591 nir_instr_type_call,
592 nir_instr_type_tex,
593 nir_instr_type_intrinsic,
594 nir_instr_type_load_const,
595 nir_instr_type_jump,
596 nir_instr_type_ssa_undef,
597 nir_instr_type_phi,
598 nir_instr_type_parallel_copy,
599 } nir_instr_type;
600
601 typedef struct nir_instr {
602 struct exec_node node;
603 struct nir_block *block;
604 nir_instr_type type;
605
606 /* A temporary for optimization and analysis passes to use for storing
607 * flags. For instance, DCE uses this to store the "dead/live" info.
608 */
609 uint8_t pass_flags;
610
611 /** generic instruction index. */
612 unsigned index;
613 } nir_instr;
614
615 static inline nir_instr *
616 nir_instr_next(nir_instr *instr)
617 {
618 struct exec_node *next = exec_node_get_next(&instr->node);
619 if (exec_node_is_tail_sentinel(next))
620 return NULL;
621 else
622 return exec_node_data(nir_instr, next, node);
623 }
624
625 static inline nir_instr *
626 nir_instr_prev(nir_instr *instr)
627 {
628 struct exec_node *prev = exec_node_get_prev(&instr->node);
629 if (exec_node_is_head_sentinel(prev))
630 return NULL;
631 else
632 return exec_node_data(nir_instr, prev, node);
633 }
634
635 static inline bool
636 nir_instr_is_first(const nir_instr *instr)
637 {
638 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr->node));
639 }
640
641 static inline bool
642 nir_instr_is_last(const nir_instr *instr)
643 {
644 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr->node));
645 }
646
647 typedef struct nir_ssa_def {
648 /** for debugging only, can be NULL */
649 const char* name;
650
651 /** generic SSA definition index. */
652 unsigned index;
653
654 /** Index into the live_in and live_out bitfields */
655 unsigned live_index;
656
657 /** Instruction which produces this SSA value. */
658 nir_instr *parent_instr;
659
660 /** set of nir_instrs where this register is used (read from) */
661 struct list_head uses;
662
663 /** set of nir_ifs where this register is used as a condition */
664 struct list_head if_uses;
665
666 uint8_t num_components;
667
668 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
669 uint8_t bit_size;
670 } nir_ssa_def;
671
672 struct nir_src;
673
674 typedef struct {
675 nir_register *reg;
676 struct nir_src *indirect; /** < NULL for no indirect offset */
677 unsigned base_offset;
678
679 /* TODO use-def chain goes here */
680 } nir_reg_src;
681
682 typedef struct {
683 nir_instr *parent_instr;
684 struct list_head def_link;
685
686 nir_register *reg;
687 struct nir_src *indirect; /** < NULL for no indirect offset */
688 unsigned base_offset;
689
690 /* TODO def-use chain goes here */
691 } nir_reg_dest;
692
693 struct nir_if;
694
695 typedef struct nir_src {
696 union {
697 /** Instruction that consumes this value as a source. */
698 nir_instr *parent_instr;
699 struct nir_if *parent_if;
700 };
701
702 struct list_head use_link;
703
704 union {
705 nir_reg_src reg;
706 nir_ssa_def *ssa;
707 };
708
709 bool is_ssa;
710 } nir_src;
711
712 static inline nir_src
713 nir_src_init(void)
714 {
715 nir_src src = { { NULL } };
716 return src;
717 }
718
719 #define NIR_SRC_INIT nir_src_init()
720
721 #define nir_foreach_use(src, reg_or_ssa_def) \
722 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
723
724 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
725 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
726
727 #define nir_foreach_if_use(src, reg_or_ssa_def) \
728 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
729
730 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
731 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
732
733 typedef struct {
734 union {
735 nir_reg_dest reg;
736 nir_ssa_def ssa;
737 };
738
739 bool is_ssa;
740 } nir_dest;
741
742 static inline nir_dest
743 nir_dest_init(void)
744 {
745 nir_dest dest = { { { NULL } } };
746 return dest;
747 }
748
749 #define NIR_DEST_INIT nir_dest_init()
750
751 #define nir_foreach_def(dest, reg) \
752 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
753
754 #define nir_foreach_def_safe(dest, reg) \
755 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
756
757 static inline nir_src
758 nir_src_for_ssa(nir_ssa_def *def)
759 {
760 nir_src src = NIR_SRC_INIT;
761
762 src.is_ssa = true;
763 src.ssa = def;
764
765 return src;
766 }
767
768 static inline nir_src
769 nir_src_for_reg(nir_register *reg)
770 {
771 nir_src src = NIR_SRC_INIT;
772
773 src.is_ssa = false;
774 src.reg.reg = reg;
775 src.reg.indirect = NULL;
776 src.reg.base_offset = 0;
777
778 return src;
779 }
780
781 static inline nir_dest
782 nir_dest_for_reg(nir_register *reg)
783 {
784 nir_dest dest = NIR_DEST_INIT;
785
786 dest.reg.reg = reg;
787
788 return dest;
789 }
790
791 static inline unsigned
792 nir_src_bit_size(nir_src src)
793 {
794 return src.is_ssa ? src.ssa->bit_size : src.reg.reg->bit_size;
795 }
796
797 static inline unsigned
798 nir_src_num_components(nir_src src)
799 {
800 return src.is_ssa ? src.ssa->num_components : src.reg.reg->num_components;
801 }
802
803 static inline bool
804 nir_src_is_const(nir_src src)
805 {
806 return src.is_ssa &&
807 src.ssa->parent_instr->type == nir_instr_type_load_const;
808 }
809
810 static inline unsigned
811 nir_dest_bit_size(nir_dest dest)
812 {
813 return dest.is_ssa ? dest.ssa.bit_size : dest.reg.reg->bit_size;
814 }
815
816 static inline unsigned
817 nir_dest_num_components(nir_dest dest)
818 {
819 return dest.is_ssa ? dest.ssa.num_components : dest.reg.reg->num_components;
820 }
821
822 void nir_src_copy(nir_src *dest, const nir_src *src, void *instr_or_if);
823 void nir_dest_copy(nir_dest *dest, const nir_dest *src, nir_instr *instr);
824
825 typedef struct {
826 nir_src src;
827
828 /**
829 * \name input modifiers
830 */
831 /*@{*/
832 /**
833 * For inputs interpreted as floating point, flips the sign bit. For
834 * inputs interpreted as integers, performs the two's complement negation.
835 */
836 bool negate;
837
838 /**
839 * Clears the sign bit for floating point values, and computes the integer
840 * absolute value for integers. Note that the negate modifier acts after
841 * the absolute value modifier, therefore if both are set then all inputs
842 * will become negative.
843 */
844 bool abs;
845 /*@}*/
846
847 /**
848 * For each input component, says which component of the register it is
849 * chosen from. Note that which elements of the swizzle are used and which
850 * are ignored are based on the write mask for most opcodes - for example,
851 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
852 * a swizzle of {2, x, 1, 0} where x means "don't care."
853 */
854 uint8_t swizzle[NIR_MAX_VEC_COMPONENTS];
855 } nir_alu_src;
856
857 typedef struct {
858 nir_dest dest;
859
860 /**
861 * \name saturate output modifier
862 *
863 * Only valid for opcodes that output floating-point numbers. Clamps the
864 * output to between 0.0 and 1.0 inclusive.
865 */
866
867 bool saturate;
868
869 unsigned write_mask : NIR_MAX_VEC_COMPONENTS; /* ignored if dest.is_ssa is true */
870 } nir_alu_dest;
871
872 /** NIR sized and unsized types
873 *
874 * The values in this enum are carefully chosen so that the sized type is
875 * just the unsized type OR the number of bits.
876 */
877 typedef enum {
878 nir_type_invalid = 0, /* Not a valid type */
879 nir_type_int = 2,
880 nir_type_uint = 4,
881 nir_type_bool = 6,
882 nir_type_float = 128,
883 nir_type_bool1 = 1 | nir_type_bool,
884 nir_type_bool32 = 32 | nir_type_bool,
885 nir_type_int1 = 1 | nir_type_int,
886 nir_type_int8 = 8 | nir_type_int,
887 nir_type_int16 = 16 | nir_type_int,
888 nir_type_int32 = 32 | nir_type_int,
889 nir_type_int64 = 64 | nir_type_int,
890 nir_type_uint1 = 1 | nir_type_uint,
891 nir_type_uint8 = 8 | nir_type_uint,
892 nir_type_uint16 = 16 | nir_type_uint,
893 nir_type_uint32 = 32 | nir_type_uint,
894 nir_type_uint64 = 64 | nir_type_uint,
895 nir_type_float16 = 16 | nir_type_float,
896 nir_type_float32 = 32 | nir_type_float,
897 nir_type_float64 = 64 | nir_type_float,
898 } nir_alu_type;
899
900 #define NIR_ALU_TYPE_SIZE_MASK 0x79
901 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
902
903 static inline unsigned
904 nir_alu_type_get_type_size(nir_alu_type type)
905 {
906 return type & NIR_ALU_TYPE_SIZE_MASK;
907 }
908
909 static inline unsigned
910 nir_alu_type_get_base_type(nir_alu_type type)
911 {
912 return type & NIR_ALU_TYPE_BASE_TYPE_MASK;
913 }
914
915 static inline nir_alu_type
916 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type)
917 {
918 switch (base_type) {
919 case GLSL_TYPE_BOOL:
920 return nir_type_bool1;
921 break;
922 case GLSL_TYPE_UINT:
923 return nir_type_uint32;
924 break;
925 case GLSL_TYPE_INT:
926 return nir_type_int32;
927 break;
928 case GLSL_TYPE_UINT16:
929 return nir_type_uint16;
930 break;
931 case GLSL_TYPE_INT16:
932 return nir_type_int16;
933 break;
934 case GLSL_TYPE_UINT8:
935 return nir_type_uint8;
936 case GLSL_TYPE_INT8:
937 return nir_type_int8;
938 case GLSL_TYPE_UINT64:
939 return nir_type_uint64;
940 break;
941 case GLSL_TYPE_INT64:
942 return nir_type_int64;
943 break;
944 case GLSL_TYPE_FLOAT:
945 return nir_type_float32;
946 break;
947 case GLSL_TYPE_FLOAT16:
948 return nir_type_float16;
949 break;
950 case GLSL_TYPE_DOUBLE:
951 return nir_type_float64;
952 break;
953
954 case GLSL_TYPE_SAMPLER:
955 case GLSL_TYPE_IMAGE:
956 case GLSL_TYPE_ATOMIC_UINT:
957 case GLSL_TYPE_STRUCT:
958 case GLSL_TYPE_INTERFACE:
959 case GLSL_TYPE_ARRAY:
960 case GLSL_TYPE_VOID:
961 case GLSL_TYPE_SUBROUTINE:
962 case GLSL_TYPE_FUNCTION:
963 case GLSL_TYPE_ERROR:
964 return nir_type_invalid;
965 }
966
967 unreachable("unknown type");
968 }
969
970 static inline nir_alu_type
971 nir_get_nir_type_for_glsl_type(const struct glsl_type *type)
972 {
973 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type));
974 }
975
976 nir_op nir_type_conversion_op(nir_alu_type src, nir_alu_type dst,
977 nir_rounding_mode rnd);
978
979 static inline nir_op
980 nir_op_vec(unsigned components)
981 {
982 switch (components) {
983 case 1: return nir_op_mov;
984 case 2: return nir_op_vec2;
985 case 3: return nir_op_vec3;
986 case 4: return nir_op_vec4;
987 default: unreachable("bad component count");
988 }
989 }
990
991 typedef enum {
992 /**
993 * Operation where the first two sources are commutative.
994 *
995 * For 2-source operations, this just mathematical commutativity. Some
996 * 3-source operations, like ffma, are only commutative in the first two
997 * sources.
998 */
999 NIR_OP_IS_2SRC_COMMUTATIVE = (1 << 0),
1000 NIR_OP_IS_ASSOCIATIVE = (1 << 1),
1001 } nir_op_algebraic_property;
1002
1003 typedef struct {
1004 const char *name;
1005
1006 unsigned num_inputs;
1007
1008 /**
1009 * The number of components in the output
1010 *
1011 * If non-zero, this is the size of the output and input sizes are
1012 * explicitly given; swizzle and writemask are still in effect, but if
1013 * the output component is masked out, then the input component may
1014 * still be in use.
1015 *
1016 * If zero, the opcode acts in the standard, per-component manner; the
1017 * operation is performed on each component (except the ones that are
1018 * masked out) with the input being taken from the input swizzle for
1019 * that component.
1020 *
1021 * The size of some of the inputs may be given (i.e. non-zero) even
1022 * though output_size is zero; in that case, the inputs with a zero
1023 * size act per-component, while the inputs with non-zero size don't.
1024 */
1025 unsigned output_size;
1026
1027 /**
1028 * The type of vector that the instruction outputs. Note that the
1029 * staurate modifier is only allowed on outputs with the float type.
1030 */
1031
1032 nir_alu_type output_type;
1033
1034 /**
1035 * The number of components in each input
1036 */
1037 unsigned input_sizes[NIR_MAX_VEC_COMPONENTS];
1038
1039 /**
1040 * The type of vector that each input takes. Note that negate and
1041 * absolute value are only allowed on inputs with int or float type and
1042 * behave differently on the two.
1043 */
1044 nir_alu_type input_types[NIR_MAX_VEC_COMPONENTS];
1045
1046 nir_op_algebraic_property algebraic_properties;
1047
1048 /* Whether this represents a numeric conversion opcode */
1049 bool is_conversion;
1050 } nir_op_info;
1051
1052 extern const nir_op_info nir_op_infos[nir_num_opcodes];
1053
1054 typedef struct nir_alu_instr {
1055 nir_instr instr;
1056 nir_op op;
1057
1058 /** Indicates that this ALU instruction generates an exact value
1059 *
1060 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1061 * really equivalent to either. This indicates that the value generated by
1062 * this operation is high-precision and any code transformations that touch
1063 * it must ensure that the resulting value is bit-for-bit identical to the
1064 * original.
1065 */
1066 bool exact:1;
1067
1068 /**
1069 * Indicates that this instruction do not cause wrapping to occur, in the
1070 * form of overflow or underflow.
1071 */
1072 bool no_signed_wrap:1;
1073 bool no_unsigned_wrap:1;
1074
1075 nir_alu_dest dest;
1076 nir_alu_src src[];
1077 } nir_alu_instr;
1078
1079 void nir_alu_src_copy(nir_alu_src *dest, const nir_alu_src *src,
1080 nir_alu_instr *instr);
1081 void nir_alu_dest_copy(nir_alu_dest *dest, const nir_alu_dest *src,
1082 nir_alu_instr *instr);
1083
1084 /* is this source channel used? */
1085 static inline bool
1086 nir_alu_instr_channel_used(const nir_alu_instr *instr, unsigned src,
1087 unsigned channel)
1088 {
1089 if (nir_op_infos[instr->op].input_sizes[src] > 0)
1090 return channel < nir_op_infos[instr->op].input_sizes[src];
1091
1092 return (instr->dest.write_mask >> channel) & 1;
1093 }
1094
1095 static inline nir_component_mask_t
1096 nir_alu_instr_src_read_mask(const nir_alu_instr *instr, unsigned src)
1097 {
1098 nir_component_mask_t read_mask = 0;
1099 for (unsigned c = 0; c < NIR_MAX_VEC_COMPONENTS; c++) {
1100 if (!nir_alu_instr_channel_used(instr, src, c))
1101 continue;
1102
1103 read_mask |= (1 << instr->src[src].swizzle[c]);
1104 }
1105 return read_mask;
1106 }
1107
1108 /**
1109 * Get the number of channels used for a source
1110 */
1111 static inline unsigned
1112 nir_ssa_alu_instr_src_components(const nir_alu_instr *instr, unsigned src)
1113 {
1114 if (nir_op_infos[instr->op].input_sizes[src] > 0)
1115 return nir_op_infos[instr->op].input_sizes[src];
1116
1117 return nir_dest_num_components(instr->dest.dest);
1118 }
1119
1120 static inline bool
1121 nir_alu_instr_is_comparison(const nir_alu_instr *instr)
1122 {
1123 switch (instr->op) {
1124 case nir_op_flt:
1125 case nir_op_fge:
1126 case nir_op_feq:
1127 case nir_op_fne:
1128 case nir_op_ilt:
1129 case nir_op_ult:
1130 case nir_op_ige:
1131 case nir_op_uge:
1132 case nir_op_ieq:
1133 case nir_op_ine:
1134 case nir_op_i2b1:
1135 case nir_op_f2b1:
1136 case nir_op_inot:
1137 return true;
1138 default:
1139 return false;
1140 }
1141 }
1142
1143 bool nir_const_value_negative_equal(nir_const_value c1, nir_const_value c2,
1144 nir_alu_type full_type);
1145
1146 bool nir_alu_srcs_equal(const nir_alu_instr *alu1, const nir_alu_instr *alu2,
1147 unsigned src1, unsigned src2);
1148
1149 bool nir_alu_srcs_negative_equal(const nir_alu_instr *alu1,
1150 const nir_alu_instr *alu2,
1151 unsigned src1, unsigned src2);
1152
1153 typedef enum {
1154 nir_deref_type_var,
1155 nir_deref_type_array,
1156 nir_deref_type_array_wildcard,
1157 nir_deref_type_ptr_as_array,
1158 nir_deref_type_struct,
1159 nir_deref_type_cast,
1160 } nir_deref_type;
1161
1162 typedef struct {
1163 nir_instr instr;
1164
1165 /** The type of this deref instruction */
1166 nir_deref_type deref_type;
1167
1168 /** The mode of the underlying variable */
1169 nir_variable_mode mode;
1170
1171 /** The dereferenced type of the resulting pointer value */
1172 const struct glsl_type *type;
1173
1174 union {
1175 /** Variable being dereferenced if deref_type is a deref_var */
1176 nir_variable *var;
1177
1178 /** Parent deref if deref_type is not deref_var */
1179 nir_src parent;
1180 };
1181
1182 /** Additional deref parameters */
1183 union {
1184 struct {
1185 nir_src index;
1186 } arr;
1187
1188 struct {
1189 unsigned index;
1190 } strct;
1191
1192 struct {
1193 unsigned ptr_stride;
1194 } cast;
1195 };
1196
1197 /** Destination to store the resulting "pointer" */
1198 nir_dest dest;
1199 } nir_deref_instr;
1200
1201 static inline nir_deref_instr *nir_src_as_deref(nir_src src);
1202
1203 static inline nir_deref_instr *
1204 nir_deref_instr_parent(const nir_deref_instr *instr)
1205 {
1206 if (instr->deref_type == nir_deref_type_var)
1207 return NULL;
1208 else
1209 return nir_src_as_deref(instr->parent);
1210 }
1211
1212 static inline nir_variable *
1213 nir_deref_instr_get_variable(const nir_deref_instr *instr)
1214 {
1215 while (instr->deref_type != nir_deref_type_var) {
1216 if (instr->deref_type == nir_deref_type_cast)
1217 return NULL;
1218
1219 instr = nir_deref_instr_parent(instr);
1220 }
1221
1222 return instr->var;
1223 }
1224
1225 bool nir_deref_instr_has_indirect(nir_deref_instr *instr);
1226 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr *instr);
1227 bool nir_deref_instr_has_complex_use(nir_deref_instr *instr);
1228
1229 bool nir_deref_instr_remove_if_unused(nir_deref_instr *instr);
1230
1231 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr *instr);
1232
1233 typedef struct {
1234 nir_instr instr;
1235
1236 struct nir_function *callee;
1237
1238 unsigned num_params;
1239 nir_src params[];
1240 } nir_call_instr;
1241
1242 #include "nir_intrinsics.h"
1243
1244 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1245
1246 /** Represents an intrinsic
1247 *
1248 * An intrinsic is an instruction type for handling things that are
1249 * more-or-less regular operations but don't just consume and produce SSA
1250 * values like ALU operations do. Intrinsics are not for things that have
1251 * special semantic meaning such as phi nodes and parallel copies.
1252 * Examples of intrinsics include variable load/store operations, system
1253 * value loads, and the like. Even though texturing more-or-less falls
1254 * under this category, texturing is its own instruction type because
1255 * trying to represent texturing with intrinsics would lead to a
1256 * combinatorial explosion of intrinsic opcodes.
1257 *
1258 * By having a single instruction type for handling a lot of different
1259 * cases, optimization passes can look for intrinsics and, for the most
1260 * part, completely ignore them. Each intrinsic type also has a few
1261 * possible flags that govern whether or not they can be reordered or
1262 * eliminated. That way passes like dead code elimination can still work
1263 * on intrisics without understanding the meaning of each.
1264 *
1265 * Each intrinsic has some number of constant indices, some number of
1266 * variables, and some number of sources. What these sources, variables,
1267 * and indices mean depends on the intrinsic and is documented with the
1268 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1269 * instructions are the only types of instruction that can operate on
1270 * variables.
1271 */
1272 typedef struct {
1273 nir_instr instr;
1274
1275 nir_intrinsic_op intrinsic;
1276
1277 nir_dest dest;
1278
1279 /** number of components if this is a vectorized intrinsic
1280 *
1281 * Similarly to ALU operations, some intrinsics are vectorized.
1282 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1283 * For vectorized intrinsics, the num_components field specifies the
1284 * number of destination components and the number of source components
1285 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1286 */
1287 uint8_t num_components;
1288
1289 int const_index[NIR_INTRINSIC_MAX_CONST_INDEX];
1290
1291 nir_src src[];
1292 } nir_intrinsic_instr;
1293
1294 static inline nir_variable *
1295 nir_intrinsic_get_var(nir_intrinsic_instr *intrin, unsigned i)
1296 {
1297 return nir_deref_instr_get_variable(nir_src_as_deref(intrin->src[i]));
1298 }
1299
1300 /**
1301 * \name NIR intrinsics semantic flags
1302 *
1303 * information about what the compiler can do with the intrinsics.
1304 *
1305 * \sa nir_intrinsic_info::flags
1306 */
1307 typedef enum {
1308 /**
1309 * whether the intrinsic can be safely eliminated if none of its output
1310 * value is not being used.
1311 */
1312 NIR_INTRINSIC_CAN_ELIMINATE = (1 << 0),
1313
1314 /**
1315 * Whether the intrinsic can be reordered with respect to any other
1316 * intrinsic, i.e. whether the only reordering dependencies of the
1317 * intrinsic are due to the register reads/writes.
1318 */
1319 NIR_INTRINSIC_CAN_REORDER = (1 << 1),
1320 } nir_intrinsic_semantic_flag;
1321
1322 /**
1323 * \name NIR intrinsics const-index flag
1324 *
1325 * Indicates the usage of a const_index slot.
1326 *
1327 * \sa nir_intrinsic_info::index_map
1328 */
1329 typedef enum {
1330 /**
1331 * Generally instructions that take a offset src argument, can encode
1332 * a constant 'base' value which is added to the offset.
1333 */
1334 NIR_INTRINSIC_BASE = 1,
1335
1336 /**
1337 * For store instructions, a writemask for the store.
1338 */
1339 NIR_INTRINSIC_WRMASK,
1340
1341 /**
1342 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1343 */
1344 NIR_INTRINSIC_STREAM_ID,
1345
1346 /**
1347 * The clip-plane id for load_user_clip_plane intrinsic.
1348 */
1349 NIR_INTRINSIC_UCP_ID,
1350
1351 /**
1352 * The amount of data, starting from BASE, that this instruction may
1353 * access. This is used to provide bounds if the offset is not constant.
1354 */
1355 NIR_INTRINSIC_RANGE,
1356
1357 /**
1358 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1359 */
1360 NIR_INTRINSIC_DESC_SET,
1361
1362 /**
1363 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1364 */
1365 NIR_INTRINSIC_BINDING,
1366
1367 /**
1368 * Component offset.
1369 */
1370 NIR_INTRINSIC_COMPONENT,
1371
1372 /**
1373 * Interpolation mode (only meaningful for FS inputs).
1374 */
1375 NIR_INTRINSIC_INTERP_MODE,
1376
1377 /**
1378 * A binary nir_op to use when performing a reduction or scan operation
1379 */
1380 NIR_INTRINSIC_REDUCTION_OP,
1381
1382 /**
1383 * Cluster size for reduction operations
1384 */
1385 NIR_INTRINSIC_CLUSTER_SIZE,
1386
1387 /**
1388 * Parameter index for a load_param intrinsic
1389 */
1390 NIR_INTRINSIC_PARAM_IDX,
1391
1392 /**
1393 * Image dimensionality for image intrinsics
1394 *
1395 * One of GLSL_SAMPLER_DIM_*
1396 */
1397 NIR_INTRINSIC_IMAGE_DIM,
1398
1399 /**
1400 * Non-zero if we are accessing an array image
1401 */
1402 NIR_INTRINSIC_IMAGE_ARRAY,
1403
1404 /**
1405 * Image format for image intrinsics
1406 */
1407 NIR_INTRINSIC_FORMAT,
1408
1409 /**
1410 * Access qualifiers for image and memory access intrinsics
1411 */
1412 NIR_INTRINSIC_ACCESS,
1413
1414 /**
1415 * Alignment for offsets and addresses
1416 *
1417 * These two parameters, specify an alignment in terms of a multiplier and
1418 * an offset. The offset or address parameter X of the intrinsic is
1419 * guaranteed to satisfy the following:
1420 *
1421 * (X - align_offset) % align_mul == 0
1422 */
1423 NIR_INTRINSIC_ALIGN_MUL,
1424 NIR_INTRINSIC_ALIGN_OFFSET,
1425
1426 /**
1427 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1428 */
1429 NIR_INTRINSIC_DESC_TYPE,
1430
1431 /**
1432 * The nir_alu_type of a uniform/input/output
1433 */
1434 NIR_INTRINSIC_TYPE,
1435
1436 /**
1437 * The swizzle mask for the instructions
1438 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1439 */
1440 NIR_INTRINSIC_SWIZZLE_MASK,
1441
1442 /* Separate source/dest access flags for copies */
1443 NIR_INTRINSIC_SRC_ACCESS,
1444 NIR_INTRINSIC_DST_ACCESS,
1445
1446 NIR_INTRINSIC_NUM_INDEX_FLAGS,
1447
1448 } nir_intrinsic_index_flag;
1449
1450 #define NIR_INTRINSIC_MAX_INPUTS 5
1451
1452 typedef struct {
1453 const char *name;
1454
1455 unsigned num_srcs; /** < number of register/SSA inputs */
1456
1457 /** number of components of each input register
1458 *
1459 * If this value is 0, the number of components is given by the
1460 * num_components field of nir_intrinsic_instr. If this value is -1, the
1461 * intrinsic consumes however many components are provided and it is not
1462 * validated at all.
1463 */
1464 int src_components[NIR_INTRINSIC_MAX_INPUTS];
1465
1466 bool has_dest;
1467
1468 /** number of components of the output register
1469 *
1470 * If this value is 0, the number of components is given by the
1471 * num_components field of nir_intrinsic_instr.
1472 */
1473 unsigned dest_components;
1474
1475 /** bitfield of legal bit sizes */
1476 unsigned dest_bit_sizes;
1477
1478 /** the number of constant indices used by the intrinsic */
1479 unsigned num_indices;
1480
1481 /** indicates the usage of intr->const_index[n] */
1482 unsigned index_map[NIR_INTRINSIC_NUM_INDEX_FLAGS];
1483
1484 /** semantic flags for calls to this intrinsic */
1485 nir_intrinsic_semantic_flag flags;
1486 } nir_intrinsic_info;
1487
1488 extern const nir_intrinsic_info nir_intrinsic_infos[nir_num_intrinsics];
1489
1490 static inline unsigned
1491 nir_intrinsic_src_components(nir_intrinsic_instr *intr, unsigned srcn)
1492 {
1493 const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
1494 assert(srcn < info->num_srcs);
1495 if (info->src_components[srcn] > 0)
1496 return info->src_components[srcn];
1497 else if (info->src_components[srcn] == 0)
1498 return intr->num_components;
1499 else
1500 return nir_src_num_components(intr->src[srcn]);
1501 }
1502
1503 static inline unsigned
1504 nir_intrinsic_dest_components(nir_intrinsic_instr *intr)
1505 {
1506 const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
1507 if (!info->has_dest)
1508 return 0;
1509 else if (info->dest_components)
1510 return info->dest_components;
1511 else
1512 return intr->num_components;
1513 }
1514
1515 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1516 static inline type \
1517 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1518 { \
1519 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1520 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1521 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1522 } \
1523 static inline void \
1524 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1525 { \
1526 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1527 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1528 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1529 }
1530
1531 INTRINSIC_IDX_ACCESSORS(write_mask, WRMASK, unsigned)
1532 INTRINSIC_IDX_ACCESSORS(base, BASE, int)
1533 INTRINSIC_IDX_ACCESSORS(stream_id, STREAM_ID, unsigned)
1534 INTRINSIC_IDX_ACCESSORS(ucp_id, UCP_ID, unsigned)
1535 INTRINSIC_IDX_ACCESSORS(range, RANGE, unsigned)
1536 INTRINSIC_IDX_ACCESSORS(desc_set, DESC_SET, unsigned)
1537 INTRINSIC_IDX_ACCESSORS(binding, BINDING, unsigned)
1538 INTRINSIC_IDX_ACCESSORS(component, COMPONENT, unsigned)
1539 INTRINSIC_IDX_ACCESSORS(interp_mode, INTERP_MODE, unsigned)
1540 INTRINSIC_IDX_ACCESSORS(reduction_op, REDUCTION_OP, unsigned)
1541 INTRINSIC_IDX_ACCESSORS(cluster_size, CLUSTER_SIZE, unsigned)
1542 INTRINSIC_IDX_ACCESSORS(param_idx, PARAM_IDX, unsigned)
1543 INTRINSIC_IDX_ACCESSORS(image_dim, IMAGE_DIM, enum glsl_sampler_dim)
1544 INTRINSIC_IDX_ACCESSORS(image_array, IMAGE_ARRAY, bool)
1545 INTRINSIC_IDX_ACCESSORS(access, ACCESS, enum gl_access_qualifier)
1546 INTRINSIC_IDX_ACCESSORS(src_access, SRC_ACCESS, enum gl_access_qualifier)
1547 INTRINSIC_IDX_ACCESSORS(dst_access, DST_ACCESS, enum gl_access_qualifier)
1548 INTRINSIC_IDX_ACCESSORS(format, FORMAT, unsigned)
1549 INTRINSIC_IDX_ACCESSORS(align_mul, ALIGN_MUL, unsigned)
1550 INTRINSIC_IDX_ACCESSORS(align_offset, ALIGN_OFFSET, unsigned)
1551 INTRINSIC_IDX_ACCESSORS(desc_type, DESC_TYPE, unsigned)
1552 INTRINSIC_IDX_ACCESSORS(type, TYPE, nir_alu_type)
1553 INTRINSIC_IDX_ACCESSORS(swizzle_mask, SWIZZLE_MASK, unsigned)
1554
1555 static inline void
1556 nir_intrinsic_set_align(nir_intrinsic_instr *intrin,
1557 unsigned align_mul, unsigned align_offset)
1558 {
1559 assert(util_is_power_of_two_nonzero(align_mul));
1560 assert(align_offset < align_mul);
1561 nir_intrinsic_set_align_mul(intrin, align_mul);
1562 nir_intrinsic_set_align_offset(intrin, align_offset);
1563 }
1564
1565 /** Returns a simple alignment for a load/store intrinsic offset
1566 *
1567 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1568 * and ALIGN_OFFSET parameters, this helper takes both into account and
1569 * provides a single simple alignment parameter. The offset X is guaranteed
1570 * to satisfy X % align == 0.
1571 */
1572 static inline unsigned
1573 nir_intrinsic_align(const nir_intrinsic_instr *intrin)
1574 {
1575 const unsigned align_mul = nir_intrinsic_align_mul(intrin);
1576 const unsigned align_offset = nir_intrinsic_align_offset(intrin);
1577 assert(align_offset < align_mul);
1578 return align_offset ? 1 << (ffs(align_offset) - 1) : align_mul;
1579 }
1580
1581 /* Converts a image_deref_* intrinsic into a image_* one */
1582 void nir_rewrite_image_intrinsic(nir_intrinsic_instr *instr,
1583 nir_ssa_def *handle, bool bindless);
1584
1585 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1586 static inline bool
1587 nir_intrinsic_can_reorder(nir_intrinsic_instr *instr)
1588 {
1589 if (instr->intrinsic == nir_intrinsic_load_deref ||
1590 instr->intrinsic == nir_intrinsic_load_ssbo ||
1591 instr->intrinsic == nir_intrinsic_bindless_image_load ||
1592 instr->intrinsic == nir_intrinsic_image_deref_load ||
1593 instr->intrinsic == nir_intrinsic_image_load) {
1594 return nir_intrinsic_access(instr) & ACCESS_CAN_REORDER;
1595 } else {
1596 const nir_intrinsic_info *info =
1597 &nir_intrinsic_infos[instr->intrinsic];
1598 return (info->flags & NIR_INTRINSIC_CAN_ELIMINATE) &&
1599 (info->flags & NIR_INTRINSIC_CAN_REORDER);
1600 }
1601 }
1602
1603 /**
1604 * \group texture information
1605 *
1606 * This gives semantic information about textures which is useful to the
1607 * frontend, the backend, and lowering passes, but not the optimizer.
1608 */
1609
1610 typedef enum {
1611 nir_tex_src_coord,
1612 nir_tex_src_projector,
1613 nir_tex_src_comparator, /* shadow comparator */
1614 nir_tex_src_offset,
1615 nir_tex_src_bias,
1616 nir_tex_src_lod,
1617 nir_tex_src_min_lod,
1618 nir_tex_src_ms_index, /* MSAA sample index */
1619 nir_tex_src_ms_mcs, /* MSAA compression value */
1620 nir_tex_src_ddx,
1621 nir_tex_src_ddy,
1622 nir_tex_src_texture_deref, /* < deref pointing to the texture */
1623 nir_tex_src_sampler_deref, /* < deref pointing to the sampler */
1624 nir_tex_src_texture_offset, /* < dynamically uniform indirect offset */
1625 nir_tex_src_sampler_offset, /* < dynamically uniform indirect offset */
1626 nir_tex_src_texture_handle, /* < bindless texture handle */
1627 nir_tex_src_sampler_handle, /* < bindless sampler handle */
1628 nir_tex_src_plane, /* < selects plane for planar textures */
1629 nir_num_tex_src_types
1630 } nir_tex_src_type;
1631
1632 typedef struct {
1633 nir_src src;
1634 nir_tex_src_type src_type;
1635 } nir_tex_src;
1636
1637 typedef enum {
1638 nir_texop_tex, /**< Regular texture look-up */
1639 nir_texop_txb, /**< Texture look-up with LOD bias */
1640 nir_texop_txl, /**< Texture look-up with explicit LOD */
1641 nir_texop_txd, /**< Texture look-up with partial derivatives */
1642 nir_texop_txf, /**< Texel fetch with explicit LOD */
1643 nir_texop_txf_ms, /**< Multisample texture fetch */
1644 nir_texop_txf_ms_fb, /**< Multisample texture fetch from framebuffer */
1645 nir_texop_txf_ms_mcs, /**< Multisample compression value fetch */
1646 nir_texop_txs, /**< Texture size */
1647 nir_texop_lod, /**< Texture lod query */
1648 nir_texop_tg4, /**< Texture gather */
1649 nir_texop_query_levels, /**< Texture levels query */
1650 nir_texop_texture_samples, /**< Texture samples query */
1651 nir_texop_samples_identical, /**< Query whether all samples are definitely
1652 * identical.
1653 */
1654 } nir_texop;
1655
1656 typedef struct {
1657 nir_instr instr;
1658
1659 enum glsl_sampler_dim sampler_dim;
1660 nir_alu_type dest_type;
1661
1662 nir_texop op;
1663 nir_dest dest;
1664 nir_tex_src *src;
1665 unsigned num_srcs, coord_components;
1666 bool is_array, is_shadow;
1667
1668 /**
1669 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1670 * components or the new-style shadow that outputs 1 component.
1671 */
1672 bool is_new_style_shadow;
1673
1674 /* gather component selector */
1675 unsigned component : 2;
1676
1677 /* gather offsets */
1678 int8_t tg4_offsets[4][2];
1679
1680 /* True if the texture index or handle is not dynamically uniform */
1681 bool texture_non_uniform;
1682
1683 /* True if the sampler index or handle is not dynamically uniform */
1684 bool sampler_non_uniform;
1685
1686 /** The texture index
1687 *
1688 * If this texture instruction has a nir_tex_src_texture_offset source,
1689 * then the texture index is given by texture_index + texture_offset.
1690 */
1691 unsigned texture_index;
1692
1693 /** The size of the texture array or 0 if it's not an array */
1694 unsigned texture_array_size;
1695
1696 /** The sampler index
1697 *
1698 * The following operations do not require a sampler and, as such, this
1699 * field should be ignored:
1700 * - nir_texop_txf
1701 * - nir_texop_txf_ms
1702 * - nir_texop_txs
1703 * - nir_texop_lod
1704 * - nir_texop_query_levels
1705 * - nir_texop_texture_samples
1706 * - nir_texop_samples_identical
1707 *
1708 * If this texture instruction has a nir_tex_src_sampler_offset source,
1709 * then the sampler index is given by sampler_index + sampler_offset.
1710 */
1711 unsigned sampler_index;
1712 } nir_tex_instr;
1713
1714 static inline unsigned
1715 nir_tex_instr_dest_size(const nir_tex_instr *instr)
1716 {
1717 switch (instr->op) {
1718 case nir_texop_txs: {
1719 unsigned ret;
1720 switch (instr->sampler_dim) {
1721 case GLSL_SAMPLER_DIM_1D:
1722 case GLSL_SAMPLER_DIM_BUF:
1723 ret = 1;
1724 break;
1725 case GLSL_SAMPLER_DIM_2D:
1726 case GLSL_SAMPLER_DIM_CUBE:
1727 case GLSL_SAMPLER_DIM_MS:
1728 case GLSL_SAMPLER_DIM_RECT:
1729 case GLSL_SAMPLER_DIM_EXTERNAL:
1730 case GLSL_SAMPLER_DIM_SUBPASS:
1731 ret = 2;
1732 break;
1733 case GLSL_SAMPLER_DIM_3D:
1734 ret = 3;
1735 break;
1736 default:
1737 unreachable("not reached");
1738 }
1739 if (instr->is_array)
1740 ret++;
1741 return ret;
1742 }
1743
1744 case nir_texop_lod:
1745 return 2;
1746
1747 case nir_texop_texture_samples:
1748 case nir_texop_query_levels:
1749 case nir_texop_samples_identical:
1750 return 1;
1751
1752 default:
1753 if (instr->is_shadow && instr->is_new_style_shadow)
1754 return 1;
1755
1756 return 4;
1757 }
1758 }
1759
1760 /* Returns true if this texture operation queries something about the texture
1761 * rather than actually sampling it.
1762 */
1763 static inline bool
1764 nir_tex_instr_is_query(const nir_tex_instr *instr)
1765 {
1766 switch (instr->op) {
1767 case nir_texop_txs:
1768 case nir_texop_lod:
1769 case nir_texop_texture_samples:
1770 case nir_texop_query_levels:
1771 case nir_texop_txf_ms_mcs:
1772 return true;
1773 case nir_texop_tex:
1774 case nir_texop_txb:
1775 case nir_texop_txl:
1776 case nir_texop_txd:
1777 case nir_texop_txf:
1778 case nir_texop_txf_ms:
1779 case nir_texop_txf_ms_fb:
1780 case nir_texop_tg4:
1781 return false;
1782 default:
1783 unreachable("Invalid texture opcode");
1784 }
1785 }
1786
1787 static inline bool
1788 nir_tex_instr_has_implicit_derivative(const nir_tex_instr *instr)
1789 {
1790 switch (instr->op) {
1791 case nir_texop_tex:
1792 case nir_texop_txb:
1793 case nir_texop_lod:
1794 return true;
1795 default:
1796 return false;
1797 }
1798 }
1799
1800 static inline nir_alu_type
1801 nir_tex_instr_src_type(const nir_tex_instr *instr, unsigned src)
1802 {
1803 switch (instr->src[src].src_type) {
1804 case nir_tex_src_coord:
1805 switch (instr->op) {
1806 case nir_texop_txf:
1807 case nir_texop_txf_ms:
1808 case nir_texop_txf_ms_fb:
1809 case nir_texop_txf_ms_mcs:
1810 case nir_texop_samples_identical:
1811 return nir_type_int;
1812
1813 default:
1814 return nir_type_float;
1815 }
1816
1817 case nir_tex_src_lod:
1818 switch (instr->op) {
1819 case nir_texop_txs:
1820 case nir_texop_txf:
1821 return nir_type_int;
1822
1823 default:
1824 return nir_type_float;
1825 }
1826
1827 case nir_tex_src_projector:
1828 case nir_tex_src_comparator:
1829 case nir_tex_src_bias:
1830 case nir_tex_src_min_lod:
1831 case nir_tex_src_ddx:
1832 case nir_tex_src_ddy:
1833 return nir_type_float;
1834
1835 case nir_tex_src_offset:
1836 case nir_tex_src_ms_index:
1837 case nir_tex_src_plane:
1838 return nir_type_int;
1839
1840 case nir_tex_src_ms_mcs:
1841 case nir_tex_src_texture_deref:
1842 case nir_tex_src_sampler_deref:
1843 case nir_tex_src_texture_offset:
1844 case nir_tex_src_sampler_offset:
1845 case nir_tex_src_texture_handle:
1846 case nir_tex_src_sampler_handle:
1847 return nir_type_uint;
1848
1849 case nir_num_tex_src_types:
1850 unreachable("nir_num_tex_src_types is not a valid source type");
1851 }
1852
1853 unreachable("Invalid texture source type");
1854 }
1855
1856 static inline unsigned
1857 nir_tex_instr_src_size(const nir_tex_instr *instr, unsigned src)
1858 {
1859 if (instr->src[src].src_type == nir_tex_src_coord)
1860 return instr->coord_components;
1861
1862 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
1863 if (instr->src[src].src_type == nir_tex_src_ms_mcs)
1864 return 4;
1865
1866 if (instr->src[src].src_type == nir_tex_src_ddx ||
1867 instr->src[src].src_type == nir_tex_src_ddy) {
1868 if (instr->is_array)
1869 return instr->coord_components - 1;
1870 else
1871 return instr->coord_components;
1872 }
1873
1874 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
1875 * the offset, since a cube maps to a single face.
1876 */
1877 if (instr->src[src].src_type == nir_tex_src_offset) {
1878 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE)
1879 return 2;
1880 else if (instr->is_array)
1881 return instr->coord_components - 1;
1882 else
1883 return instr->coord_components;
1884 }
1885
1886 return 1;
1887 }
1888
1889 static inline int
1890 nir_tex_instr_src_index(const nir_tex_instr *instr, nir_tex_src_type type)
1891 {
1892 for (unsigned i = 0; i < instr->num_srcs; i++)
1893 if (instr->src[i].src_type == type)
1894 return (int) i;
1895
1896 return -1;
1897 }
1898
1899 void nir_tex_instr_add_src(nir_tex_instr *tex,
1900 nir_tex_src_type src_type,
1901 nir_src src);
1902
1903 void nir_tex_instr_remove_src(nir_tex_instr *tex, unsigned src_idx);
1904
1905 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr *tex);
1906
1907 typedef struct {
1908 nir_instr instr;
1909
1910 nir_ssa_def def;
1911
1912 nir_const_value value[];
1913 } nir_load_const_instr;
1914
1915 typedef enum {
1916 nir_jump_return,
1917 nir_jump_break,
1918 nir_jump_continue,
1919 } nir_jump_type;
1920
1921 typedef struct {
1922 nir_instr instr;
1923 nir_jump_type type;
1924 } nir_jump_instr;
1925
1926 /* creates a new SSA variable in an undefined state */
1927
1928 typedef struct {
1929 nir_instr instr;
1930 nir_ssa_def def;
1931 } nir_ssa_undef_instr;
1932
1933 typedef struct {
1934 struct exec_node node;
1935
1936 /* The predecessor block corresponding to this source */
1937 struct nir_block *pred;
1938
1939 nir_src src;
1940 } nir_phi_src;
1941
1942 #define nir_foreach_phi_src(phi_src, phi) \
1943 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
1944 #define nir_foreach_phi_src_safe(phi_src, phi) \
1945 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
1946
1947 typedef struct {
1948 nir_instr instr;
1949
1950 struct exec_list srcs; /** < list of nir_phi_src */
1951
1952 nir_dest dest;
1953 } nir_phi_instr;
1954
1955 typedef struct {
1956 struct exec_node node;
1957 nir_src src;
1958 nir_dest dest;
1959 } nir_parallel_copy_entry;
1960
1961 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
1962 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
1963
1964 typedef struct {
1965 nir_instr instr;
1966
1967 /* A list of nir_parallel_copy_entrys. The sources of all of the
1968 * entries are copied to the corresponding destinations "in parallel".
1969 * In other words, if we have two entries: a -> b and b -> a, the values
1970 * get swapped.
1971 */
1972 struct exec_list entries;
1973 } nir_parallel_copy_instr;
1974
1975 NIR_DEFINE_CAST(nir_instr_as_alu, nir_instr, nir_alu_instr, instr,
1976 type, nir_instr_type_alu)
1977 NIR_DEFINE_CAST(nir_instr_as_deref, nir_instr, nir_deref_instr, instr,
1978 type, nir_instr_type_deref)
1979 NIR_DEFINE_CAST(nir_instr_as_call, nir_instr, nir_call_instr, instr,
1980 type, nir_instr_type_call)
1981 NIR_DEFINE_CAST(nir_instr_as_jump, nir_instr, nir_jump_instr, instr,
1982 type, nir_instr_type_jump)
1983 NIR_DEFINE_CAST(nir_instr_as_tex, nir_instr, nir_tex_instr, instr,
1984 type, nir_instr_type_tex)
1985 NIR_DEFINE_CAST(nir_instr_as_intrinsic, nir_instr, nir_intrinsic_instr, instr,
1986 type, nir_instr_type_intrinsic)
1987 NIR_DEFINE_CAST(nir_instr_as_load_const, nir_instr, nir_load_const_instr, instr,
1988 type, nir_instr_type_load_const)
1989 NIR_DEFINE_CAST(nir_instr_as_ssa_undef, nir_instr, nir_ssa_undef_instr, instr,
1990 type, nir_instr_type_ssa_undef)
1991 NIR_DEFINE_CAST(nir_instr_as_phi, nir_instr, nir_phi_instr, instr,
1992 type, nir_instr_type_phi)
1993 NIR_DEFINE_CAST(nir_instr_as_parallel_copy, nir_instr,
1994 nir_parallel_copy_instr, instr,
1995 type, nir_instr_type_parallel_copy)
1996
1997
1998 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
1999 static inline type \
2000 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2001 { \
2002 assert(nir_src_is_const(src)); \
2003 nir_load_const_instr *load = \
2004 nir_instr_as_load_const(src.ssa->parent_instr); \
2005 assert(comp < load->def.num_components); \
2006 return nir_const_value_as_##suffix(load->value[comp], \
2007 load->def.bit_size); \
2008 } \
2009 \
2010 static inline type \
2011 nir_src_as_##suffix(nir_src src) \
2012 { \
2013 assert(nir_src_num_components(src) == 1); \
2014 return nir_src_comp_as_##suffix(src, 0); \
2015 }
2016
2017 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2018 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint)
2019 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2020 NIR_DEFINE_SRC_AS_CONST(double, float)
2021
2022 #undef NIR_DEFINE_SRC_AS_CONST
2023
2024
2025 typedef struct {
2026 nir_ssa_def *def;
2027 unsigned comp;
2028 } nir_ssa_scalar;
2029
2030 static inline bool
2031 nir_ssa_scalar_is_const(nir_ssa_scalar s)
2032 {
2033 return s.def->parent_instr->type == nir_instr_type_load_const;
2034 }
2035
2036 static inline nir_const_value
2037 nir_ssa_scalar_as_const_value(nir_ssa_scalar s)
2038 {
2039 assert(s.comp < s.def->num_components);
2040 nir_load_const_instr *load = nir_instr_as_load_const(s.def->parent_instr);
2041 return load->value[s.comp];
2042 }
2043
2044 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2045 static inline type \
2046 nir_ssa_scalar_as_##suffix(nir_ssa_scalar s) \
2047 { \
2048 return nir_const_value_as_##suffix( \
2049 nir_ssa_scalar_as_const_value(s), s.def->bit_size); \
2050 }
2051
2052 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2053 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint)
2054 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2055 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2056
2057 #undef NIR_DEFINE_SCALAR_AS_CONST
2058
2059 static inline bool
2060 nir_ssa_scalar_is_alu(nir_ssa_scalar s)
2061 {
2062 return s.def->parent_instr->type == nir_instr_type_alu;
2063 }
2064
2065 static inline nir_op
2066 nir_ssa_scalar_alu_op(nir_ssa_scalar s)
2067 {
2068 return nir_instr_as_alu(s.def->parent_instr)->op;
2069 }
2070
2071 static inline nir_ssa_scalar
2072 nir_ssa_scalar_chase_alu_src(nir_ssa_scalar s, unsigned alu_src_idx)
2073 {
2074 nir_ssa_scalar out = { NULL, 0 };
2075
2076 nir_alu_instr *alu = nir_instr_as_alu(s.def->parent_instr);
2077 assert(alu_src_idx < nir_op_infos[alu->op].num_inputs);
2078
2079 /* Our component must be written */
2080 assert(s.comp < s.def->num_components);
2081 assert(alu->dest.write_mask & (1u << s.comp));
2082
2083 assert(alu->src[alu_src_idx].src.is_ssa);
2084 out.def = alu->src[alu_src_idx].src.ssa;
2085
2086 if (nir_op_infos[alu->op].input_sizes[alu_src_idx] == 0) {
2087 /* The ALU src is unsized so the source component follows the
2088 * destination component.
2089 */
2090 out.comp = alu->src[alu_src_idx].swizzle[s.comp];
2091 } else {
2092 /* This is a sized source so all source components work together to
2093 * produce all the destination components. Since we need to return a
2094 * scalar, this only works if the source is a scalar.
2095 */
2096 assert(nir_op_infos[alu->op].input_sizes[alu_src_idx] == 1);
2097 out.comp = alu->src[alu_src_idx].swizzle[0];
2098 }
2099 assert(out.comp < out.def->num_components);
2100
2101 return out;
2102 }
2103
2104
2105 /*
2106 * Control flow
2107 *
2108 * Control flow consists of a tree of control flow nodes, which include
2109 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2110 * instructions that always run start-to-finish. Each basic block also keeps
2111 * track of its successors (blocks which may run immediately after the current
2112 * block) and predecessors (blocks which could have run immediately before the
2113 * current block). Each function also has a start block and an end block which
2114 * all return statements point to (which is always empty). Together, all the
2115 * blocks with their predecessors and successors make up the control flow
2116 * graph (CFG) of the function. There are helpers that modify the tree of
2117 * control flow nodes while modifying the CFG appropriately; these should be
2118 * used instead of modifying the tree directly.
2119 */
2120
2121 typedef enum {
2122 nir_cf_node_block,
2123 nir_cf_node_if,
2124 nir_cf_node_loop,
2125 nir_cf_node_function
2126 } nir_cf_node_type;
2127
2128 typedef struct nir_cf_node {
2129 struct exec_node node;
2130 nir_cf_node_type type;
2131 struct nir_cf_node *parent;
2132 } nir_cf_node;
2133
2134 typedef struct nir_block {
2135 nir_cf_node cf_node;
2136
2137 struct exec_list instr_list; /** < list of nir_instr */
2138
2139 /** generic block index; generated by nir_index_blocks */
2140 unsigned index;
2141
2142 /*
2143 * Each block can only have up to 2 successors, so we put them in a simple
2144 * array - no need for anything more complicated.
2145 */
2146 struct nir_block *successors[2];
2147
2148 /* Set of nir_block predecessors in the CFG */
2149 struct set *predecessors;
2150
2151 /*
2152 * this node's immediate dominator in the dominance tree - set to NULL for
2153 * the start block.
2154 */
2155 struct nir_block *imm_dom;
2156
2157 /* This node's children in the dominance tree */
2158 unsigned num_dom_children;
2159 struct nir_block **dom_children;
2160
2161 /* Set of nir_blocks on the dominance frontier of this block */
2162 struct set *dom_frontier;
2163
2164 /*
2165 * These two indices have the property that dom_{pre,post}_index for each
2166 * child of this block in the dominance tree will always be between
2167 * dom_pre_index and dom_post_index for this block, which makes testing if
2168 * a given block is dominated by another block an O(1) operation.
2169 */
2170 unsigned dom_pre_index, dom_post_index;
2171
2172 /* live in and out for this block; used for liveness analysis */
2173 BITSET_WORD *live_in;
2174 BITSET_WORD *live_out;
2175 } nir_block;
2176
2177 static inline nir_instr *
2178 nir_block_first_instr(nir_block *block)
2179 {
2180 struct exec_node *head = exec_list_get_head(&block->instr_list);
2181 return exec_node_data(nir_instr, head, node);
2182 }
2183
2184 static inline nir_instr *
2185 nir_block_last_instr(nir_block *block)
2186 {
2187 struct exec_node *tail = exec_list_get_tail(&block->instr_list);
2188 return exec_node_data(nir_instr, tail, node);
2189 }
2190
2191 static inline bool
2192 nir_block_ends_in_jump(nir_block *block)
2193 {
2194 return !exec_list_is_empty(&block->instr_list) &&
2195 nir_block_last_instr(block)->type == nir_instr_type_jump;
2196 }
2197
2198 #define nir_foreach_instr(instr, block) \
2199 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
2200 #define nir_foreach_instr_reverse(instr, block) \
2201 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
2202 #define nir_foreach_instr_safe(instr, block) \
2203 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
2204 #define nir_foreach_instr_reverse_safe(instr, block) \
2205 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
2206
2207 typedef enum {
2208 nir_selection_control_none = 0x0,
2209 nir_selection_control_flatten = 0x1,
2210 nir_selection_control_dont_flatten = 0x2,
2211 } nir_selection_control;
2212
2213 typedef struct nir_if {
2214 nir_cf_node cf_node;
2215 nir_src condition;
2216 nir_selection_control control;
2217
2218 struct exec_list then_list; /** < list of nir_cf_node */
2219 struct exec_list else_list; /** < list of nir_cf_node */
2220 } nir_if;
2221
2222 typedef struct {
2223 nir_if *nif;
2224
2225 /** Instruction that generates nif::condition. */
2226 nir_instr *conditional_instr;
2227
2228 /** Block within ::nif that has the break instruction. */
2229 nir_block *break_block;
2230
2231 /** Last block for the then- or else-path that does not contain the break. */
2232 nir_block *continue_from_block;
2233
2234 /** True when ::break_block is in the else-path of ::nif. */
2235 bool continue_from_then;
2236 bool induction_rhs;
2237
2238 /* This is true if the terminators exact trip count is unknown. For
2239 * example:
2240 *
2241 * for (int i = 0; i < imin(x, 4); i++)
2242 * ...
2243 *
2244 * Here loop analysis would have set a max_trip_count of 4 however we dont
2245 * know for sure that this is the exact trip count.
2246 */
2247 bool exact_trip_count_unknown;
2248
2249 struct list_head loop_terminator_link;
2250 } nir_loop_terminator;
2251
2252 typedef struct {
2253 /* Estimated cost (in number of instructions) of the loop */
2254 unsigned instr_cost;
2255
2256 /* Guessed trip count based on array indexing */
2257 unsigned guessed_trip_count;
2258
2259 /* Maximum number of times the loop is run (if known) */
2260 unsigned max_trip_count;
2261
2262 /* Do we know the exact number of times the loop will be run */
2263 bool exact_trip_count_known;
2264
2265 /* Unroll the loop regardless of its size */
2266 bool force_unroll;
2267
2268 /* Does the loop contain complex loop terminators, continues or other
2269 * complex behaviours? If this is true we can't rely on
2270 * loop_terminator_list to be complete or accurate.
2271 */
2272 bool complex_loop;
2273
2274 nir_loop_terminator *limiting_terminator;
2275
2276 /* A list of loop_terminators terminating this loop. */
2277 struct list_head loop_terminator_list;
2278 } nir_loop_info;
2279
2280 typedef enum {
2281 nir_loop_control_none = 0x0,
2282 nir_loop_control_unroll = 0x1,
2283 nir_loop_control_dont_unroll = 0x2,
2284 } nir_loop_control;
2285
2286 typedef struct {
2287 nir_cf_node cf_node;
2288
2289 struct exec_list body; /** < list of nir_cf_node */
2290
2291 nir_loop_info *info;
2292 nir_loop_control control;
2293 bool partially_unrolled;
2294 } nir_loop;
2295
2296 /**
2297 * Various bits of metadata that can may be created or required by
2298 * optimization and analysis passes
2299 */
2300 typedef enum {
2301 nir_metadata_none = 0x0,
2302 nir_metadata_block_index = 0x1,
2303 nir_metadata_dominance = 0x2,
2304 nir_metadata_live_ssa_defs = 0x4,
2305 nir_metadata_not_properly_reset = 0x8,
2306 nir_metadata_loop_analysis = 0x10,
2307 } nir_metadata;
2308
2309 typedef struct {
2310 nir_cf_node cf_node;
2311
2312 /** pointer to the function of which this is an implementation */
2313 struct nir_function *function;
2314
2315 struct exec_list body; /** < list of nir_cf_node */
2316
2317 nir_block *end_block;
2318
2319 /** list for all local variables in the function */
2320 struct exec_list locals;
2321
2322 /** list of local registers in the function */
2323 struct exec_list registers;
2324
2325 /** next available local register index */
2326 unsigned reg_alloc;
2327
2328 /** next available SSA value index */
2329 unsigned ssa_alloc;
2330
2331 /* total number of basic blocks, only valid when block_index_dirty = false */
2332 unsigned num_blocks;
2333
2334 nir_metadata valid_metadata;
2335 } nir_function_impl;
2336
2337 ATTRIBUTE_RETURNS_NONNULL static inline nir_block *
2338 nir_start_block(nir_function_impl *impl)
2339 {
2340 return (nir_block *) impl->body.head_sentinel.next;
2341 }
2342
2343 ATTRIBUTE_RETURNS_NONNULL static inline nir_block *
2344 nir_impl_last_block(nir_function_impl *impl)
2345 {
2346 return (nir_block *) impl->body.tail_sentinel.prev;
2347 }
2348
2349 static inline nir_cf_node *
2350 nir_cf_node_next(nir_cf_node *node)
2351 {
2352 struct exec_node *next = exec_node_get_next(&node->node);
2353 if (exec_node_is_tail_sentinel(next))
2354 return NULL;
2355 else
2356 return exec_node_data(nir_cf_node, next, node);
2357 }
2358
2359 static inline nir_cf_node *
2360 nir_cf_node_prev(nir_cf_node *node)
2361 {
2362 struct exec_node *prev = exec_node_get_prev(&node->node);
2363 if (exec_node_is_head_sentinel(prev))
2364 return NULL;
2365 else
2366 return exec_node_data(nir_cf_node, prev, node);
2367 }
2368
2369 static inline bool
2370 nir_cf_node_is_first(const nir_cf_node *node)
2371 {
2372 return exec_node_is_head_sentinel(node->node.prev);
2373 }
2374
2375 static inline bool
2376 nir_cf_node_is_last(const nir_cf_node *node)
2377 {
2378 return exec_node_is_tail_sentinel(node->node.next);
2379 }
2380
2381 NIR_DEFINE_CAST(nir_cf_node_as_block, nir_cf_node, nir_block, cf_node,
2382 type, nir_cf_node_block)
2383 NIR_DEFINE_CAST(nir_cf_node_as_if, nir_cf_node, nir_if, cf_node,
2384 type, nir_cf_node_if)
2385 NIR_DEFINE_CAST(nir_cf_node_as_loop, nir_cf_node, nir_loop, cf_node,
2386 type, nir_cf_node_loop)
2387 NIR_DEFINE_CAST(nir_cf_node_as_function, nir_cf_node,
2388 nir_function_impl, cf_node, type, nir_cf_node_function)
2389
2390 static inline nir_block *
2391 nir_if_first_then_block(nir_if *if_stmt)
2392 {
2393 struct exec_node *head = exec_list_get_head(&if_stmt->then_list);
2394 return nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
2395 }
2396
2397 static inline nir_block *
2398 nir_if_last_then_block(nir_if *if_stmt)
2399 {
2400 struct exec_node *tail = exec_list_get_tail(&if_stmt->then_list);
2401 return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
2402 }
2403
2404 static inline nir_block *
2405 nir_if_first_else_block(nir_if *if_stmt)
2406 {
2407 struct exec_node *head = exec_list_get_head(&if_stmt->else_list);
2408 return nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
2409 }
2410
2411 static inline nir_block *
2412 nir_if_last_else_block(nir_if *if_stmt)
2413 {
2414 struct exec_node *tail = exec_list_get_tail(&if_stmt->else_list);
2415 return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
2416 }
2417
2418 static inline nir_block *
2419 nir_loop_first_block(nir_loop *loop)
2420 {
2421 struct exec_node *head = exec_list_get_head(&loop->body);
2422 return nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
2423 }
2424
2425 static inline nir_block *
2426 nir_loop_last_block(nir_loop *loop)
2427 {
2428 struct exec_node *tail = exec_list_get_tail(&loop->body);
2429 return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
2430 }
2431
2432 /**
2433 * Return true if this list of cf_nodes contains a single empty block.
2434 */
2435 static inline bool
2436 nir_cf_list_is_empty_block(struct exec_list *cf_list)
2437 {
2438 if (exec_list_is_singular(cf_list)) {
2439 struct exec_node *head = exec_list_get_head(cf_list);
2440 nir_block *block =
2441 nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
2442 return exec_list_is_empty(&block->instr_list);
2443 }
2444 return false;
2445 }
2446
2447 typedef struct {
2448 uint8_t num_components;
2449 uint8_t bit_size;
2450 } nir_parameter;
2451
2452 typedef struct nir_function {
2453 struct exec_node node;
2454
2455 const char *name;
2456 struct nir_shader *shader;
2457
2458 unsigned num_params;
2459 nir_parameter *params;
2460
2461 /** The implementation of this function.
2462 *
2463 * If the function is only declared and not implemented, this is NULL.
2464 */
2465 nir_function_impl *impl;
2466
2467 bool is_entrypoint;
2468 } nir_function;
2469
2470 typedef enum {
2471 nir_lower_imul64 = (1 << 0),
2472 nir_lower_isign64 = (1 << 1),
2473 /** Lower all int64 modulus and division opcodes */
2474 nir_lower_divmod64 = (1 << 2),
2475 /** Lower all 64-bit umul_high and imul_high opcodes */
2476 nir_lower_imul_high64 = (1 << 3),
2477 nir_lower_mov64 = (1 << 4),
2478 nir_lower_icmp64 = (1 << 5),
2479 nir_lower_iadd64 = (1 << 6),
2480 nir_lower_iabs64 = (1 << 7),
2481 nir_lower_ineg64 = (1 << 8),
2482 nir_lower_logic64 = (1 << 9),
2483 nir_lower_minmax64 = (1 << 10),
2484 nir_lower_shift64 = (1 << 11),
2485 nir_lower_imul_2x32_64 = (1 << 12),
2486 nir_lower_extract64 = (1 << 13),
2487 } nir_lower_int64_options;
2488
2489 typedef enum {
2490 nir_lower_drcp = (1 << 0),
2491 nir_lower_dsqrt = (1 << 1),
2492 nir_lower_drsq = (1 << 2),
2493 nir_lower_dtrunc = (1 << 3),
2494 nir_lower_dfloor = (1 << 4),
2495 nir_lower_dceil = (1 << 5),
2496 nir_lower_dfract = (1 << 6),
2497 nir_lower_dround_even = (1 << 7),
2498 nir_lower_dmod = (1 << 8),
2499 nir_lower_dsub = (1 << 9),
2500 nir_lower_ddiv = (1 << 10),
2501 nir_lower_fp64_full_software = (1 << 11),
2502 } nir_lower_doubles_options;
2503
2504 typedef enum {
2505 nir_divergence_single_prim_per_subgroup = (1 << 0),
2506 nir_divergence_single_patch_per_tcs_subgroup = (1 << 1),
2507 nir_divergence_single_patch_per_tes_subgroup = (1 << 2),
2508 nir_divergence_view_index_uniform = (1 << 3),
2509 } nir_divergence_options;
2510
2511 typedef struct nir_shader_compiler_options {
2512 bool lower_fdiv;
2513 bool lower_ffma;
2514 bool fuse_ffma;
2515 bool lower_flrp16;
2516 bool lower_flrp32;
2517 /** Lowers flrp when it does not support doubles */
2518 bool lower_flrp64;
2519 bool lower_fpow;
2520 bool lower_fsat;
2521 bool lower_fsqrt;
2522 bool lower_sincos;
2523 bool lower_fmod;
2524 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2525 bool lower_bitfield_extract;
2526 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
2527 bool lower_bitfield_extract_to_shifts;
2528 /** Lowers bitfield_insert to bfi/bfm */
2529 bool lower_bitfield_insert;
2530 /** Lowers bitfield_insert to compares, and shifts. */
2531 bool lower_bitfield_insert_to_shifts;
2532 /** Lowers bitfield_insert to bfm/bitfield_select. */
2533 bool lower_bitfield_insert_to_bitfield_select;
2534 /** Lowers bitfield_reverse to shifts. */
2535 bool lower_bitfield_reverse;
2536 /** Lowers bit_count to shifts. */
2537 bool lower_bit_count;
2538 /** Lowers ifind_msb to compare and ufind_msb */
2539 bool lower_ifind_msb;
2540 /** Lowers find_lsb to ufind_msb and logic ops */
2541 bool lower_find_lsb;
2542 bool lower_uadd_carry;
2543 bool lower_usub_borrow;
2544 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2545 bool lower_mul_high;
2546 /** lowers fneg and ineg to fsub and isub. */
2547 bool lower_negate;
2548 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2549 bool lower_sub;
2550
2551 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2552 bool lower_scmp;
2553
2554 /* lower fall_equalN/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
2555 bool lower_vector_cmp;
2556
2557 /** enables rules to lower idiv by power-of-two: */
2558 bool lower_idiv;
2559
2560 /** enable rules to avoid bit ops */
2561 bool lower_bitops;
2562
2563 /** enables rules to lower isign to imin+imax */
2564 bool lower_isign;
2565
2566 /** enables rules to lower fsign to fsub and flt */
2567 bool lower_fsign;
2568
2569 /* lower fdph to fdot4 */
2570 bool lower_fdph;
2571
2572 /** lower fdot to fmul and fsum/fadd. */
2573 bool lower_fdot;
2574
2575 /* Does the native fdot instruction replicate its result for four
2576 * components? If so, then opt_algebraic_late will turn all fdotN
2577 * instructions into fdot_replicatedN instructions.
2578 */
2579 bool fdot_replicates;
2580
2581 /** lowers ffloor to fsub+ffract: */
2582 bool lower_ffloor;
2583
2584 /** lowers ffract to fsub+ffloor: */
2585 bool lower_ffract;
2586
2587 /** lowers fceil to fneg+ffloor+fneg: */
2588 bool lower_fceil;
2589
2590 bool lower_ftrunc;
2591
2592 bool lower_ldexp;
2593
2594 bool lower_pack_half_2x16;
2595 bool lower_pack_unorm_2x16;
2596 bool lower_pack_snorm_2x16;
2597 bool lower_pack_unorm_4x8;
2598 bool lower_pack_snorm_4x8;
2599 bool lower_unpack_half_2x16;
2600 bool lower_unpack_unorm_2x16;
2601 bool lower_unpack_snorm_2x16;
2602 bool lower_unpack_unorm_4x8;
2603 bool lower_unpack_snorm_4x8;
2604
2605 bool lower_extract_byte;
2606 bool lower_extract_word;
2607
2608 bool lower_all_io_to_temps;
2609 bool lower_all_io_to_elements;
2610
2611 /* Indicates that the driver only has zero-based vertex id */
2612 bool vertex_id_zero_based;
2613
2614 /**
2615 * If enabled, gl_BaseVertex will be lowered as:
2616 * is_indexed_draw (~0/0) & firstvertex
2617 */
2618 bool lower_base_vertex;
2619
2620 /**
2621 * If enabled, gl_HelperInvocation will be lowered as:
2622 *
2623 * !((1 << sample_id) & sample_mask_in))
2624 *
2625 * This depends on some possibly hw implementation details, which may
2626 * not be true for all hw. In particular that the FS is only executed
2627 * for covered samples or for helper invocations. So, do not blindly
2628 * enable this option.
2629 *
2630 * Note: See also issue #22 in ARB_shader_image_load_store
2631 */
2632 bool lower_helper_invocation;
2633
2634 /**
2635 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
2636 *
2637 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
2638 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
2639 */
2640 bool optimize_sample_mask_in;
2641
2642 bool lower_cs_local_index_from_id;
2643 bool lower_cs_local_id_from_index;
2644
2645 bool lower_device_index_to_zero;
2646
2647 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
2648 bool lower_wpos_pntc;
2649
2650 bool lower_hadd;
2651 bool lower_add_sat;
2652
2653 /**
2654 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
2655 * for IO purposes and would prefer loads/stores be vectorized.
2656 */
2657 bool vectorize_io;
2658
2659 /**
2660 * Should nir_lower_io() create load_interpolated_input intrinsics?
2661 *
2662 * If not, it generates regular load_input intrinsics and interpolation
2663 * information must be inferred from the list of input nir_variables.
2664 */
2665 bool use_interpolated_input_intrinsics;
2666
2667 /* Lowers when 32x32->64 bit multiplication is not supported */
2668 bool lower_mul_2x32_64;
2669
2670 /* Lowers when rotate instruction is not supported */
2671 bool lower_rotate;
2672
2673 /**
2674 * Is this the Intel vec4 backend?
2675 *
2676 * Used to inhibit algebraic optimizations that are known to be harmful on
2677 * the Intel vec4 backend. This is generally applicable to any
2678 * optimization that might cause more immediate values to be used in
2679 * 3-source (e.g., ffma and flrp) instructions.
2680 */
2681 bool intel_vec4;
2682
2683 unsigned max_unroll_iterations;
2684
2685 nir_lower_int64_options lower_int64_options;
2686 nir_lower_doubles_options lower_doubles_options;
2687 } nir_shader_compiler_options;
2688
2689 typedef struct nir_shader {
2690 /** list of uniforms (nir_variable) */
2691 struct exec_list uniforms;
2692
2693 /** list of inputs (nir_variable) */
2694 struct exec_list inputs;
2695
2696 /** list of outputs (nir_variable) */
2697 struct exec_list outputs;
2698
2699 /** list of shared compute variables (nir_variable) */
2700 struct exec_list shared;
2701
2702 /** Set of driver-specific options for the shader.
2703 *
2704 * The memory for the options is expected to be kept in a single static
2705 * copy by the driver.
2706 */
2707 const struct nir_shader_compiler_options *options;
2708
2709 /** Various bits of compile-time information about a given shader */
2710 struct shader_info info;
2711
2712 /** list of global variables in the shader (nir_variable) */
2713 struct exec_list globals;
2714
2715 /** list of system value variables in the shader (nir_variable) */
2716 struct exec_list system_values;
2717
2718 struct exec_list functions; /** < list of nir_function */
2719
2720 /**
2721 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
2722 * access plus one
2723 */
2724 unsigned num_inputs, num_uniforms, num_outputs, num_shared;
2725
2726 /** Size in bytes of required scratch space */
2727 unsigned scratch_size;
2728
2729 /** Constant data associated with this shader.
2730 *
2731 * Constant data is loaded through load_constant intrinsics. See also
2732 * nir_opt_large_constants.
2733 */
2734 void *constant_data;
2735 unsigned constant_data_size;
2736 } nir_shader;
2737
2738 #define nir_foreach_function(func, shader) \
2739 foreach_list_typed(nir_function, func, node, &(shader)->functions)
2740
2741 static inline nir_function_impl *
2742 nir_shader_get_entrypoint(nir_shader *shader)
2743 {
2744 nir_function *func = NULL;
2745
2746 nir_foreach_function(function, shader) {
2747 assert(func == NULL);
2748 if (function->is_entrypoint) {
2749 func = function;
2750 #ifndef NDEBUG
2751 break;
2752 #endif
2753 }
2754 }
2755
2756 if (!func)
2757 return NULL;
2758
2759 assert(func->num_params == 0);
2760 assert(func->impl);
2761 return func->impl;
2762 }
2763
2764 nir_shader *nir_shader_create(void *mem_ctx,
2765 gl_shader_stage stage,
2766 const nir_shader_compiler_options *options,
2767 shader_info *si);
2768
2769 nir_register *nir_local_reg_create(nir_function_impl *impl);
2770
2771 void nir_reg_remove(nir_register *reg);
2772
2773 /** Adds a variable to the appropriate list in nir_shader */
2774 void nir_shader_add_variable(nir_shader *shader, nir_variable *var);
2775
2776 static inline void
2777 nir_function_impl_add_variable(nir_function_impl *impl, nir_variable *var)
2778 {
2779 assert(var->data.mode == nir_var_function_temp);
2780 exec_list_push_tail(&impl->locals, &var->node);
2781 }
2782
2783 /** creates a variable, sets a few defaults, and adds it to the list */
2784 nir_variable *nir_variable_create(nir_shader *shader,
2785 nir_variable_mode mode,
2786 const struct glsl_type *type,
2787 const char *name);
2788 /** creates a local variable and adds it to the list */
2789 nir_variable *nir_local_variable_create(nir_function_impl *impl,
2790 const struct glsl_type *type,
2791 const char *name);
2792
2793 /** creates a function and adds it to the shader's list of functions */
2794 nir_function *nir_function_create(nir_shader *shader, const char *name);
2795
2796 nir_function_impl *nir_function_impl_create(nir_function *func);
2797 /** creates a function_impl that isn't tied to any particular function */
2798 nir_function_impl *nir_function_impl_create_bare(nir_shader *shader);
2799
2800 nir_block *nir_block_create(nir_shader *shader);
2801 nir_if *nir_if_create(nir_shader *shader);
2802 nir_loop *nir_loop_create(nir_shader *shader);
2803
2804 nir_function_impl *nir_cf_node_get_function(nir_cf_node *node);
2805
2806 /** requests that the given pieces of metadata be generated */
2807 void nir_metadata_require(nir_function_impl *impl, nir_metadata required, ...);
2808 /** dirties all but the preserved metadata */
2809 void nir_metadata_preserve(nir_function_impl *impl, nir_metadata preserved);
2810
2811 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
2812 nir_alu_instr *nir_alu_instr_create(nir_shader *shader, nir_op op);
2813
2814 nir_deref_instr *nir_deref_instr_create(nir_shader *shader,
2815 nir_deref_type deref_type);
2816
2817 nir_jump_instr *nir_jump_instr_create(nir_shader *shader, nir_jump_type type);
2818
2819 nir_load_const_instr *nir_load_const_instr_create(nir_shader *shader,
2820 unsigned num_components,
2821 unsigned bit_size);
2822
2823 nir_intrinsic_instr *nir_intrinsic_instr_create(nir_shader *shader,
2824 nir_intrinsic_op op);
2825
2826 nir_call_instr *nir_call_instr_create(nir_shader *shader,
2827 nir_function *callee);
2828
2829 nir_tex_instr *nir_tex_instr_create(nir_shader *shader, unsigned num_srcs);
2830
2831 nir_phi_instr *nir_phi_instr_create(nir_shader *shader);
2832
2833 nir_parallel_copy_instr *nir_parallel_copy_instr_create(nir_shader *shader);
2834
2835 nir_ssa_undef_instr *nir_ssa_undef_instr_create(nir_shader *shader,
2836 unsigned num_components,
2837 unsigned bit_size);
2838
2839 nir_const_value nir_alu_binop_identity(nir_op binop, unsigned bit_size);
2840
2841 /**
2842 * NIR Cursors and Instruction Insertion API
2843 * @{
2844 *
2845 * A tiny struct representing a point to insert/extract instructions or
2846 * control flow nodes. Helps reduce the combinatorial explosion of possible
2847 * points to insert/extract.
2848 *
2849 * \sa nir_control_flow.h
2850 */
2851 typedef enum {
2852 nir_cursor_before_block,
2853 nir_cursor_after_block,
2854 nir_cursor_before_instr,
2855 nir_cursor_after_instr,
2856 } nir_cursor_option;
2857
2858 typedef struct {
2859 nir_cursor_option option;
2860 union {
2861 nir_block *block;
2862 nir_instr *instr;
2863 };
2864 } nir_cursor;
2865
2866 static inline nir_block *
2867 nir_cursor_current_block(nir_cursor cursor)
2868 {
2869 if (cursor.option == nir_cursor_before_instr ||
2870 cursor.option == nir_cursor_after_instr) {
2871 return cursor.instr->block;
2872 } else {
2873 return cursor.block;
2874 }
2875 }
2876
2877 bool nir_cursors_equal(nir_cursor a, nir_cursor b);
2878
2879 static inline nir_cursor
2880 nir_before_block(nir_block *block)
2881 {
2882 nir_cursor cursor;
2883 cursor.option = nir_cursor_before_block;
2884 cursor.block = block;
2885 return cursor;
2886 }
2887
2888 static inline nir_cursor
2889 nir_after_block(nir_block *block)
2890 {
2891 nir_cursor cursor;
2892 cursor.option = nir_cursor_after_block;
2893 cursor.block = block;
2894 return cursor;
2895 }
2896
2897 static inline nir_cursor
2898 nir_before_instr(nir_instr *instr)
2899 {
2900 nir_cursor cursor;
2901 cursor.option = nir_cursor_before_instr;
2902 cursor.instr = instr;
2903 return cursor;
2904 }
2905
2906 static inline nir_cursor
2907 nir_after_instr(nir_instr *instr)
2908 {
2909 nir_cursor cursor;
2910 cursor.option = nir_cursor_after_instr;
2911 cursor.instr = instr;
2912 return cursor;
2913 }
2914
2915 static inline nir_cursor
2916 nir_after_block_before_jump(nir_block *block)
2917 {
2918 nir_instr *last_instr = nir_block_last_instr(block);
2919 if (last_instr && last_instr->type == nir_instr_type_jump) {
2920 return nir_before_instr(last_instr);
2921 } else {
2922 return nir_after_block(block);
2923 }
2924 }
2925
2926 static inline nir_cursor
2927 nir_before_src(nir_src *src, bool is_if_condition)
2928 {
2929 if (is_if_condition) {
2930 nir_block *prev_block =
2931 nir_cf_node_as_block(nir_cf_node_prev(&src->parent_if->cf_node));
2932 assert(!nir_block_ends_in_jump(prev_block));
2933 return nir_after_block(prev_block);
2934 } else if (src->parent_instr->type == nir_instr_type_phi) {
2935 #ifndef NDEBUG
2936 nir_phi_instr *cond_phi = nir_instr_as_phi(src->parent_instr);
2937 bool found = false;
2938 nir_foreach_phi_src(phi_src, cond_phi) {
2939 if (phi_src->src.ssa == src->ssa) {
2940 found = true;
2941 break;
2942 }
2943 }
2944 assert(found);
2945 #endif
2946 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
2947 * to have a more specific name.
2948 */
2949 nir_phi_src *phi_src = LIST_ENTRY(nir_phi_src, src, src);
2950 return nir_after_block_before_jump(phi_src->pred);
2951 } else {
2952 return nir_before_instr(src->parent_instr);
2953 }
2954 }
2955
2956 static inline nir_cursor
2957 nir_before_cf_node(nir_cf_node *node)
2958 {
2959 if (node->type == nir_cf_node_block)
2960 return nir_before_block(nir_cf_node_as_block(node));
2961
2962 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node)));
2963 }
2964
2965 static inline nir_cursor
2966 nir_after_cf_node(nir_cf_node *node)
2967 {
2968 if (node->type == nir_cf_node_block)
2969 return nir_after_block(nir_cf_node_as_block(node));
2970
2971 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node)));
2972 }
2973
2974 static inline nir_cursor
2975 nir_after_phis(nir_block *block)
2976 {
2977 nir_foreach_instr(instr, block) {
2978 if (instr->type != nir_instr_type_phi)
2979 return nir_before_instr(instr);
2980 }
2981 return nir_after_block(block);
2982 }
2983
2984 static inline nir_cursor
2985 nir_after_cf_node_and_phis(nir_cf_node *node)
2986 {
2987 if (node->type == nir_cf_node_block)
2988 return nir_after_block(nir_cf_node_as_block(node));
2989
2990 nir_block *block = nir_cf_node_as_block(nir_cf_node_next(node));
2991
2992 return nir_after_phis(block);
2993 }
2994
2995 static inline nir_cursor
2996 nir_before_cf_list(struct exec_list *cf_list)
2997 {
2998 nir_cf_node *first_node = exec_node_data(nir_cf_node,
2999 exec_list_get_head(cf_list), node);
3000 return nir_before_cf_node(first_node);
3001 }
3002
3003 static inline nir_cursor
3004 nir_after_cf_list(struct exec_list *cf_list)
3005 {
3006 nir_cf_node *last_node = exec_node_data(nir_cf_node,
3007 exec_list_get_tail(cf_list), node);
3008 return nir_after_cf_node(last_node);
3009 }
3010
3011 /**
3012 * Insert a NIR instruction at the given cursor.
3013 *
3014 * Note: This does not update the cursor.
3015 */
3016 void nir_instr_insert(nir_cursor cursor, nir_instr *instr);
3017
3018 static inline void
3019 nir_instr_insert_before(nir_instr *instr, nir_instr *before)
3020 {
3021 nir_instr_insert(nir_before_instr(instr), before);
3022 }
3023
3024 static inline void
3025 nir_instr_insert_after(nir_instr *instr, nir_instr *after)
3026 {
3027 nir_instr_insert(nir_after_instr(instr), after);
3028 }
3029
3030 static inline void
3031 nir_instr_insert_before_block(nir_block *block, nir_instr *before)
3032 {
3033 nir_instr_insert(nir_before_block(block), before);
3034 }
3035
3036 static inline void
3037 nir_instr_insert_after_block(nir_block *block, nir_instr *after)
3038 {
3039 nir_instr_insert(nir_after_block(block), after);
3040 }
3041
3042 static inline void
3043 nir_instr_insert_before_cf(nir_cf_node *node, nir_instr *before)
3044 {
3045 nir_instr_insert(nir_before_cf_node(node), before);
3046 }
3047
3048 static inline void
3049 nir_instr_insert_after_cf(nir_cf_node *node, nir_instr *after)
3050 {
3051 nir_instr_insert(nir_after_cf_node(node), after);
3052 }
3053
3054 static inline void
3055 nir_instr_insert_before_cf_list(struct exec_list *list, nir_instr *before)
3056 {
3057 nir_instr_insert(nir_before_cf_list(list), before);
3058 }
3059
3060 static inline void
3061 nir_instr_insert_after_cf_list(struct exec_list *list, nir_instr *after)
3062 {
3063 nir_instr_insert(nir_after_cf_list(list), after);
3064 }
3065
3066 void nir_instr_remove_v(nir_instr *instr);
3067
3068 static inline nir_cursor
3069 nir_instr_remove(nir_instr *instr)
3070 {
3071 nir_cursor cursor;
3072 nir_instr *prev = nir_instr_prev(instr);
3073 if (prev) {
3074 cursor = nir_after_instr(prev);
3075 } else {
3076 cursor = nir_before_block(instr->block);
3077 }
3078 nir_instr_remove_v(instr);
3079 return cursor;
3080 }
3081
3082 /** @} */
3083
3084 nir_ssa_def *nir_instr_ssa_def(nir_instr *instr);
3085
3086 typedef bool (*nir_foreach_ssa_def_cb)(nir_ssa_def *def, void *state);
3087 typedef bool (*nir_foreach_dest_cb)(nir_dest *dest, void *state);
3088 typedef bool (*nir_foreach_src_cb)(nir_src *src, void *state);
3089 bool nir_foreach_ssa_def(nir_instr *instr, nir_foreach_ssa_def_cb cb,
3090 void *state);
3091 bool nir_foreach_dest(nir_instr *instr, nir_foreach_dest_cb cb, void *state);
3092 bool nir_foreach_src(nir_instr *instr, nir_foreach_src_cb cb, void *state);
3093
3094 nir_const_value *nir_src_as_const_value(nir_src src);
3095
3096 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
3097 static inline c_type * \
3098 nir_src_as_ ## name (nir_src src) \
3099 { \
3100 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
3101 ? cast_macro(src.ssa->parent_instr) : NULL; \
3102 }
3103
3104 NIR_SRC_AS_(alu_instr, nir_alu_instr, nir_instr_type_alu, nir_instr_as_alu)
3105 NIR_SRC_AS_(intrinsic, nir_intrinsic_instr,
3106 nir_instr_type_intrinsic, nir_instr_as_intrinsic)
3107 NIR_SRC_AS_(deref, nir_deref_instr, nir_instr_type_deref, nir_instr_as_deref)
3108
3109 bool nir_src_is_dynamically_uniform(nir_src src);
3110 bool nir_srcs_equal(nir_src src1, nir_src src2);
3111 bool nir_instrs_equal(const nir_instr *instr1, const nir_instr *instr2);
3112 void nir_instr_rewrite_src(nir_instr *instr, nir_src *src, nir_src new_src);
3113 void nir_instr_move_src(nir_instr *dest_instr, nir_src *dest, nir_src *src);
3114 void nir_if_rewrite_condition(nir_if *if_stmt, nir_src new_src);
3115 void nir_instr_rewrite_dest(nir_instr *instr, nir_dest *dest,
3116 nir_dest new_dest);
3117
3118 void nir_ssa_dest_init(nir_instr *instr, nir_dest *dest,
3119 unsigned num_components, unsigned bit_size,
3120 const char *name);
3121 void nir_ssa_def_init(nir_instr *instr, nir_ssa_def *def,
3122 unsigned num_components, unsigned bit_size,
3123 const char *name);
3124 static inline void
3125 nir_ssa_dest_init_for_type(nir_instr *instr, nir_dest *dest,
3126 const struct glsl_type *type,
3127 const char *name)
3128 {
3129 assert(glsl_type_is_vector_or_scalar(type));
3130 nir_ssa_dest_init(instr, dest, glsl_get_components(type),
3131 glsl_get_bit_size(type), name);
3132 }
3133 void nir_ssa_def_rewrite_uses(nir_ssa_def *def, nir_src new_src);
3134 void nir_ssa_def_rewrite_uses_after(nir_ssa_def *def, nir_src new_src,
3135 nir_instr *after_me);
3136
3137 nir_component_mask_t nir_ssa_def_components_read(const nir_ssa_def *def);
3138
3139 /*
3140 * finds the next basic block in source-code order, returns NULL if there is
3141 * none
3142 */
3143
3144 nir_block *nir_block_cf_tree_next(nir_block *block);
3145
3146 /* Performs the opposite of nir_block_cf_tree_next() */
3147
3148 nir_block *nir_block_cf_tree_prev(nir_block *block);
3149
3150 /* Gets the first block in a CF node in source-code order */
3151
3152 nir_block *nir_cf_node_cf_tree_first(nir_cf_node *node);
3153
3154 /* Gets the last block in a CF node in source-code order */
3155
3156 nir_block *nir_cf_node_cf_tree_last(nir_cf_node *node);
3157
3158 /* Gets the next block after a CF node in source-code order */
3159
3160 nir_block *nir_cf_node_cf_tree_next(nir_cf_node *node);
3161
3162 /* Macros for loops that visit blocks in source-code order */
3163
3164 #define nir_foreach_block(block, impl) \
3165 for (nir_block *block = nir_start_block(impl); block != NULL; \
3166 block = nir_block_cf_tree_next(block))
3167
3168 #define nir_foreach_block_safe(block, impl) \
3169 for (nir_block *block = nir_start_block(impl), \
3170 *next = nir_block_cf_tree_next(block); \
3171 block != NULL; \
3172 block = next, next = nir_block_cf_tree_next(block))
3173
3174 #define nir_foreach_block_reverse(block, impl) \
3175 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
3176 block = nir_block_cf_tree_prev(block))
3177
3178 #define nir_foreach_block_reverse_safe(block, impl) \
3179 for (nir_block *block = nir_impl_last_block(impl), \
3180 *prev = nir_block_cf_tree_prev(block); \
3181 block != NULL; \
3182 block = prev, prev = nir_block_cf_tree_prev(block))
3183
3184 #define nir_foreach_block_in_cf_node(block, node) \
3185 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
3186 block != nir_cf_node_cf_tree_next(node); \
3187 block = nir_block_cf_tree_next(block))
3188
3189 /* If the following CF node is an if, this function returns that if.
3190 * Otherwise, it returns NULL.
3191 */
3192 nir_if *nir_block_get_following_if(nir_block *block);
3193
3194 nir_loop *nir_block_get_following_loop(nir_block *block);
3195
3196 void nir_index_local_regs(nir_function_impl *impl);
3197 void nir_index_ssa_defs(nir_function_impl *impl);
3198 unsigned nir_index_instrs(nir_function_impl *impl);
3199
3200 void nir_index_blocks(nir_function_impl *impl);
3201
3202 void nir_print_shader(nir_shader *shader, FILE *fp);
3203 void nir_print_shader_annotated(nir_shader *shader, FILE *fp, struct hash_table *errors);
3204 void nir_print_instr(const nir_instr *instr, FILE *fp);
3205 void nir_print_deref(const nir_deref_instr *deref, FILE *fp);
3206
3207 /** Shallow clone of a single ALU instruction. */
3208 nir_alu_instr *nir_alu_instr_clone(nir_shader *s, const nir_alu_instr *orig);
3209
3210 nir_shader *nir_shader_clone(void *mem_ctx, const nir_shader *s);
3211 nir_function_impl *nir_function_impl_clone(nir_shader *shader,
3212 const nir_function_impl *fi);
3213 nir_constant *nir_constant_clone(const nir_constant *c, nir_variable *var);
3214 nir_variable *nir_variable_clone(const nir_variable *c, nir_shader *shader);
3215
3216 void nir_shader_replace(nir_shader *dest, nir_shader *src);
3217
3218 void nir_shader_serialize_deserialize(nir_shader *s);
3219
3220 #ifndef NDEBUG
3221 void nir_validate_shader(nir_shader *shader, const char *when);
3222 void nir_metadata_set_validation_flag(nir_shader *shader);
3223 void nir_metadata_check_validation_flag(nir_shader *shader);
3224
3225 static inline bool
3226 should_skip_nir(const char *name)
3227 {
3228 static const char *list = NULL;
3229 if (!list) {
3230 /* Comma separated list of names to skip. */
3231 list = getenv("NIR_SKIP");
3232 if (!list)
3233 list = "";
3234 }
3235
3236 if (!list[0])
3237 return false;
3238
3239 return comma_separated_list_contains(list, name);
3240 }
3241
3242 static inline bool
3243 should_clone_nir(void)
3244 {
3245 static int should_clone = -1;
3246 if (should_clone < 0)
3247 should_clone = env_var_as_boolean("NIR_TEST_CLONE", false);
3248
3249 return should_clone;
3250 }
3251
3252 static inline bool
3253 should_serialize_deserialize_nir(void)
3254 {
3255 static int test_serialize = -1;
3256 if (test_serialize < 0)
3257 test_serialize = env_var_as_boolean("NIR_TEST_SERIALIZE", false);
3258
3259 return test_serialize;
3260 }
3261
3262 static inline bool
3263 should_print_nir(void)
3264 {
3265 static int should_print = -1;
3266 if (should_print < 0)
3267 should_print = env_var_as_boolean("NIR_PRINT", false);
3268
3269 return should_print;
3270 }
3271 #else
3272 static inline void nir_validate_shader(nir_shader *shader, const char *when) { (void) shader; (void)when; }
3273 static inline void nir_metadata_set_validation_flag(nir_shader *shader) { (void) shader; }
3274 static inline void nir_metadata_check_validation_flag(nir_shader *shader) { (void) shader; }
3275 static inline bool should_skip_nir(UNUSED const char *pass_name) { return false; }
3276 static inline bool should_clone_nir(void) { return false; }
3277 static inline bool should_serialize_deserialize_nir(void) { return false; }
3278 static inline bool should_print_nir(void) { return false; }
3279 #endif /* NDEBUG */
3280
3281 #define _PASS(pass, nir, do_pass) do { \
3282 if (should_skip_nir(#pass)) { \
3283 printf("skipping %s\n", #pass); \
3284 break; \
3285 } \
3286 do_pass \
3287 nir_validate_shader(nir, "after " #pass); \
3288 if (should_clone_nir()) { \
3289 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3290 nir_shader_replace(nir, clone); \
3291 } \
3292 if (should_serialize_deserialize_nir()) { \
3293 nir_shader_serialize_deserialize(nir); \
3294 } \
3295 } while (0)
3296
3297 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3298 nir_metadata_set_validation_flag(nir); \
3299 if (should_print_nir()) \
3300 printf("%s\n", #pass); \
3301 if (pass(nir, ##__VA_ARGS__)) { \
3302 progress = true; \
3303 if (should_print_nir()) \
3304 nir_print_shader(nir, stdout); \
3305 nir_metadata_check_validation_flag(nir); \
3306 } \
3307 )
3308
3309 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3310 if (should_print_nir()) \
3311 printf("%s\n", #pass); \
3312 pass(nir, ##__VA_ARGS__); \
3313 if (should_print_nir()) \
3314 nir_print_shader(nir, stdout); \
3315 )
3316
3317 #define NIR_SKIP(name) should_skip_nir(#name)
3318
3319 /** An instruction filtering callback
3320 *
3321 * Returns true if the instruction should be processed and false otherwise.
3322 */
3323 typedef bool (*nir_instr_filter_cb)(const nir_instr *, const void *);
3324
3325 /** A simple instruction lowering callback
3326 *
3327 * Many instruction lowering passes can be written as a simple function which
3328 * takes an instruction as its input and returns a sequence of instructions
3329 * that implement the consumed instruction. This function type represents
3330 * such a lowering function. When called, a function with this prototype
3331 * should either return NULL indicating that no lowering needs to be done or
3332 * emit a sequence of instructions using the provided builder (whose cursor
3333 * will already be placed after the instruction to be lowered) and return the
3334 * resulting nir_ssa_def.
3335 */
3336 typedef nir_ssa_def *(*nir_lower_instr_cb)(struct nir_builder *,
3337 nir_instr *, void *);
3338
3339 /**
3340 * Special return value for nir_lower_instr_cb when some progress occurred
3341 * (like changing an input to the instr) that didn't result in a replacement
3342 * SSA def being generated.
3343 */
3344 #define NIR_LOWER_INSTR_PROGRESS ((nir_ssa_def *)(uintptr_t)1)
3345
3346 /** Iterate over all the instructions in a nir_function_impl and lower them
3347 * using the provided callbacks
3348 *
3349 * This function implements the guts of a standard lowering pass for you. It
3350 * iterates over all of the instructions in a nir_function_impl and calls the
3351 * filter callback on each one. If the filter callback returns true, it then
3352 * calls the lowering call back on the instruction. (Splitting it this way
3353 * allows us to avoid some save/restore work for instructions we know won't be
3354 * lowered.) If the instruction is dead after the lowering is complete, it
3355 * will be removed. If new instructions are added, the lowering callback will
3356 * also be called on them in case multiple lowerings are required.
3357 *
3358 * The metadata for the nir_function_impl will also be updated. If any blocks
3359 * are added (they cannot be removed), dominance and block indices will be
3360 * invalidated.
3361 */
3362 bool nir_function_impl_lower_instructions(nir_function_impl *impl,
3363 nir_instr_filter_cb filter,
3364 nir_lower_instr_cb lower,
3365 void *cb_data);
3366 bool nir_shader_lower_instructions(nir_shader *shader,
3367 nir_instr_filter_cb filter,
3368 nir_lower_instr_cb lower,
3369 void *cb_data);
3370
3371 void nir_calc_dominance_impl(nir_function_impl *impl);
3372 void nir_calc_dominance(nir_shader *shader);
3373
3374 nir_block *nir_dominance_lca(nir_block *b1, nir_block *b2);
3375 bool nir_block_dominates(nir_block *parent, nir_block *child);
3376
3377 void nir_dump_dom_tree_impl(nir_function_impl *impl, FILE *fp);
3378 void nir_dump_dom_tree(nir_shader *shader, FILE *fp);
3379
3380 void nir_dump_dom_frontier_impl(nir_function_impl *impl, FILE *fp);
3381 void nir_dump_dom_frontier(nir_shader *shader, FILE *fp);
3382
3383 void nir_dump_cfg_impl(nir_function_impl *impl, FILE *fp);
3384 void nir_dump_cfg(nir_shader *shader, FILE *fp);
3385
3386 int nir_gs_count_vertices(const nir_shader *shader);
3387
3388 bool nir_shrink_vec_array_vars(nir_shader *shader, nir_variable_mode modes);
3389 bool nir_split_array_vars(nir_shader *shader, nir_variable_mode modes);
3390 bool nir_split_var_copies(nir_shader *shader);
3391 bool nir_split_per_member_structs(nir_shader *shader);
3392 bool nir_split_struct_vars(nir_shader *shader, nir_variable_mode modes);
3393
3394 bool nir_lower_returns_impl(nir_function_impl *impl);
3395 bool nir_lower_returns(nir_shader *shader);
3396
3397 void nir_inline_function_impl(struct nir_builder *b,
3398 const nir_function_impl *impl,
3399 nir_ssa_def **params);
3400 bool nir_inline_functions(nir_shader *shader);
3401
3402 bool nir_propagate_invariant(nir_shader *shader);
3403
3404 void nir_lower_var_copy_instr(nir_intrinsic_instr *copy, nir_shader *shader);
3405 void nir_lower_deref_copy_instr(struct nir_builder *b,
3406 nir_intrinsic_instr *copy);
3407 bool nir_lower_var_copies(nir_shader *shader);
3408
3409 void nir_fixup_deref_modes(nir_shader *shader);
3410
3411 bool nir_lower_global_vars_to_local(nir_shader *shader);
3412
3413 typedef enum {
3414 nir_lower_direct_array_deref_of_vec_load = (1 << 0),
3415 nir_lower_indirect_array_deref_of_vec_load = (1 << 1),
3416 nir_lower_direct_array_deref_of_vec_store = (1 << 2),
3417 nir_lower_indirect_array_deref_of_vec_store = (1 << 3),
3418 } nir_lower_array_deref_of_vec_options;
3419
3420 bool nir_lower_array_deref_of_vec(nir_shader *shader, nir_variable_mode modes,
3421 nir_lower_array_deref_of_vec_options options);
3422
3423 bool nir_lower_indirect_derefs(nir_shader *shader, nir_variable_mode modes);
3424
3425 bool nir_lower_locals_to_regs(nir_shader *shader);
3426
3427 void nir_lower_io_to_temporaries(nir_shader *shader,
3428 nir_function_impl *entrypoint,
3429 bool outputs, bool inputs);
3430
3431 bool nir_lower_vars_to_scratch(nir_shader *shader,
3432 nir_variable_mode modes,
3433 int size_threshold,
3434 glsl_type_size_align_func size_align);
3435
3436 void nir_shader_gather_info(nir_shader *shader, nir_function_impl *entrypoint);
3437
3438 void nir_gather_ssa_types(nir_function_impl *impl,
3439 BITSET_WORD *float_types,
3440 BITSET_WORD *int_types);
3441
3442 void nir_assign_var_locations(struct exec_list *var_list, unsigned *size,
3443 int (*type_size)(const struct glsl_type *, bool));
3444
3445 /* Some helpers to do very simple linking */
3446 bool nir_remove_unused_varyings(nir_shader *producer, nir_shader *consumer);
3447 bool nir_remove_unused_io_vars(nir_shader *shader, struct exec_list *var_list,
3448 uint64_t *used_by_other_stage,
3449 uint64_t *used_by_other_stage_patches);
3450 void nir_compact_varyings(nir_shader *producer, nir_shader *consumer,
3451 bool default_to_smooth_interp);
3452 void nir_link_xfb_varyings(nir_shader *producer, nir_shader *consumer);
3453 bool nir_link_opt_varyings(nir_shader *producer, nir_shader *consumer);
3454
3455
3456 void nir_assign_io_var_locations(struct exec_list *var_list,
3457 unsigned *size,
3458 gl_shader_stage stage);
3459
3460 typedef enum {
3461 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
3462 * to 32-bit operations. This is only valid for nir_var_shader_in/out
3463 * modes.
3464 */
3465 nir_lower_io_lower_64bit_to_32 = (1 << 0),
3466
3467 /* If set, this forces all non-flat fragment shader inputs to be
3468 * interpolated as if with the "sample" qualifier. This requires
3469 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
3470 */
3471 nir_lower_io_force_sample_interpolation = (1 << 1),
3472 } nir_lower_io_options;
3473 bool nir_lower_io(nir_shader *shader,
3474 nir_variable_mode modes,
3475 int (*type_size)(const struct glsl_type *, bool),
3476 nir_lower_io_options);
3477
3478 bool nir_io_add_const_offset_to_base(nir_shader *nir, nir_variable_mode mode);
3479
3480 bool
3481 nir_lower_vars_to_explicit_types(nir_shader *shader,
3482 nir_variable_mode modes,
3483 glsl_type_size_align_func type_info);
3484
3485 typedef enum {
3486 /**
3487 * An address format which is a simple 32-bit global GPU address.
3488 */
3489 nir_address_format_32bit_global,
3490
3491 /**
3492 * An address format which is a simple 64-bit global GPU address.
3493 */
3494 nir_address_format_64bit_global,
3495
3496 /**
3497 * An address format which is a bounds-checked 64-bit global GPU address.
3498 *
3499 * The address is comprised as a 32-bit vec4 where .xy are a uint64_t base
3500 * address stored with the low bits in .x and high bits in .y, .z is a
3501 * size, and .w is an offset. When the final I/O operation is lowered, .w
3502 * is checked against .z and the operation is predicated on the result.
3503 */
3504 nir_address_format_64bit_bounded_global,
3505
3506 /**
3507 * An address format which is comprised of a vec2 where the first
3508 * component is a buffer index and the second is an offset.
3509 */
3510 nir_address_format_32bit_index_offset,
3511
3512 /**
3513 * An address format which is a simple 32-bit offset.
3514 */
3515 nir_address_format_32bit_offset,
3516
3517 /**
3518 * An address format representing a purely logical addressing model. In
3519 * this model, all deref chains must be complete from the dereference
3520 * operation to the variable. Cast derefs are not allowed. These
3521 * addresses will be 32-bit scalars but the format is immaterial because
3522 * you can always chase the chain.
3523 */
3524 nir_address_format_logical,
3525 } nir_address_format;
3526
3527 static inline unsigned
3528 nir_address_format_bit_size(nir_address_format addr_format)
3529 {
3530 switch (addr_format) {
3531 case nir_address_format_32bit_global: return 32;
3532 case nir_address_format_64bit_global: return 64;
3533 case nir_address_format_64bit_bounded_global: return 32;
3534 case nir_address_format_32bit_index_offset: return 32;
3535 case nir_address_format_32bit_offset: return 32;
3536 case nir_address_format_logical: return 32;
3537 }
3538 unreachable("Invalid address format");
3539 }
3540
3541 static inline unsigned
3542 nir_address_format_num_components(nir_address_format addr_format)
3543 {
3544 switch (addr_format) {
3545 case nir_address_format_32bit_global: return 1;
3546 case nir_address_format_64bit_global: return 1;
3547 case nir_address_format_64bit_bounded_global: return 4;
3548 case nir_address_format_32bit_index_offset: return 2;
3549 case nir_address_format_32bit_offset: return 1;
3550 case nir_address_format_logical: return 1;
3551 }
3552 unreachable("Invalid address format");
3553 }
3554
3555 static inline const struct glsl_type *
3556 nir_address_format_to_glsl_type(nir_address_format addr_format)
3557 {
3558 unsigned bit_size = nir_address_format_bit_size(addr_format);
3559 assert(bit_size == 32 || bit_size == 64);
3560 return glsl_vector_type(bit_size == 32 ? GLSL_TYPE_UINT : GLSL_TYPE_UINT64,
3561 nir_address_format_num_components(addr_format));
3562 }
3563
3564 const nir_const_value *nir_address_format_null_value(nir_address_format addr_format);
3565
3566 nir_ssa_def *nir_build_addr_ieq(struct nir_builder *b, nir_ssa_def *addr0, nir_ssa_def *addr1,