nir: support feeding state to nir_lower_clip_[vg]s
[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 static inline bool
992 nir_is_float_control_signed_zero_inf_nan_preserve(unsigned execution_mode, unsigned bit_size)
993 {
994 return (16 == bit_size && execution_mode & FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP16) ||
995 (32 == bit_size && execution_mode & FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP32) ||
996 (64 == bit_size && execution_mode & FLOAT_CONTROLS_SIGNED_ZERO_INF_NAN_PRESERVE_FP64);
997 }
998
999 static inline bool
1000 nir_is_denorm_flush_to_zero(unsigned execution_mode, unsigned bit_size)
1001 {
1002 return (16 == bit_size && execution_mode & FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP16) ||
1003 (32 == bit_size && execution_mode & FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP32) ||
1004 (64 == bit_size && execution_mode & FLOAT_CONTROLS_DENORM_FLUSH_TO_ZERO_FP64);
1005 }
1006
1007 static inline bool
1008 nir_is_denorm_preserve(unsigned execution_mode, unsigned bit_size)
1009 {
1010 return (16 == bit_size && execution_mode & FLOAT_CONTROLS_DENORM_PRESERVE_FP16) ||
1011 (32 == bit_size && execution_mode & FLOAT_CONTROLS_DENORM_PRESERVE_FP32) ||
1012 (64 == bit_size && execution_mode & FLOAT_CONTROLS_DENORM_PRESERVE_FP64);
1013 }
1014
1015 static inline bool
1016 nir_is_rounding_mode_rtne(unsigned execution_mode, unsigned bit_size)
1017 {
1018 return (16 == bit_size && execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16) ||
1019 (32 == bit_size && execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32) ||
1020 (64 == bit_size && execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64);
1021 }
1022
1023 static inline bool
1024 nir_is_rounding_mode_rtz(unsigned execution_mode, unsigned bit_size)
1025 {
1026 return (16 == bit_size && execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16) ||
1027 (32 == bit_size && execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32) ||
1028 (64 == bit_size && execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64);
1029 }
1030
1031 static inline bool
1032 nir_has_any_rounding_mode_rtz(unsigned execution_mode)
1033 {
1034 return (execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP16) ||
1035 (execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP32) ||
1036 (execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTZ_FP64);
1037 }
1038
1039 static inline bool
1040 nir_has_any_rounding_mode_rtne(unsigned execution_mode)
1041 {
1042 return (execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP16) ||
1043 (execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP32) ||
1044 (execution_mode & FLOAT_CONTROLS_ROUNDING_MODE_RTE_FP64);
1045 }
1046
1047 static inline nir_rounding_mode
1048 nir_get_rounding_mode_from_float_controls(unsigned execution_mode,
1049 nir_alu_type type)
1050 {
1051 if (nir_alu_type_get_base_type(type) != nir_type_float)
1052 return nir_rounding_mode_undef;
1053
1054 unsigned bit_size = nir_alu_type_get_type_size(type);
1055
1056 if (nir_is_rounding_mode_rtz(execution_mode, bit_size))
1057 return nir_rounding_mode_rtz;
1058 if (nir_is_rounding_mode_rtne(execution_mode, bit_size))
1059 return nir_rounding_mode_rtne;
1060 return nir_rounding_mode_undef;
1061 }
1062
1063 static inline bool
1064 nir_has_any_rounding_mode_enabled(unsigned execution_mode)
1065 {
1066 bool result =
1067 nir_has_any_rounding_mode_rtne(execution_mode) ||
1068 nir_has_any_rounding_mode_rtz(execution_mode);
1069 return result;
1070 }
1071
1072 typedef enum {
1073 /**
1074 * Operation where the first two sources are commutative.
1075 *
1076 * For 2-source operations, this just mathematical commutativity. Some
1077 * 3-source operations, like ffma, are only commutative in the first two
1078 * sources.
1079 */
1080 NIR_OP_IS_2SRC_COMMUTATIVE = (1 << 0),
1081 NIR_OP_IS_ASSOCIATIVE = (1 << 1),
1082 } nir_op_algebraic_property;
1083
1084 typedef struct {
1085 const char *name;
1086
1087 unsigned num_inputs;
1088
1089 /**
1090 * The number of components in the output
1091 *
1092 * If non-zero, this is the size of the output and input sizes are
1093 * explicitly given; swizzle and writemask are still in effect, but if
1094 * the output component is masked out, then the input component may
1095 * still be in use.
1096 *
1097 * If zero, the opcode acts in the standard, per-component manner; the
1098 * operation is performed on each component (except the ones that are
1099 * masked out) with the input being taken from the input swizzle for
1100 * that component.
1101 *
1102 * The size of some of the inputs may be given (i.e. non-zero) even
1103 * though output_size is zero; in that case, the inputs with a zero
1104 * size act per-component, while the inputs with non-zero size don't.
1105 */
1106 unsigned output_size;
1107
1108 /**
1109 * The type of vector that the instruction outputs. Note that the
1110 * staurate modifier is only allowed on outputs with the float type.
1111 */
1112
1113 nir_alu_type output_type;
1114
1115 /**
1116 * The number of components in each input
1117 */
1118 unsigned input_sizes[NIR_MAX_VEC_COMPONENTS];
1119
1120 /**
1121 * The type of vector that each input takes. Note that negate and
1122 * absolute value are only allowed on inputs with int or float type and
1123 * behave differently on the two.
1124 */
1125 nir_alu_type input_types[NIR_MAX_VEC_COMPONENTS];
1126
1127 nir_op_algebraic_property algebraic_properties;
1128
1129 /* Whether this represents a numeric conversion opcode */
1130 bool is_conversion;
1131 } nir_op_info;
1132
1133 extern const nir_op_info nir_op_infos[nir_num_opcodes];
1134
1135 typedef struct nir_alu_instr {
1136 nir_instr instr;
1137 nir_op op;
1138
1139 /** Indicates that this ALU instruction generates an exact value
1140 *
1141 * This is kind of a mixture of GLSL "precise" and "invariant" and not
1142 * really equivalent to either. This indicates that the value generated by
1143 * this operation is high-precision and any code transformations that touch
1144 * it must ensure that the resulting value is bit-for-bit identical to the
1145 * original.
1146 */
1147 bool exact:1;
1148
1149 /**
1150 * Indicates that this instruction do not cause wrapping to occur, in the
1151 * form of overflow or underflow.
1152 */
1153 bool no_signed_wrap:1;
1154 bool no_unsigned_wrap:1;
1155
1156 nir_alu_dest dest;
1157 nir_alu_src src[];
1158 } nir_alu_instr;
1159
1160 void nir_alu_src_copy(nir_alu_src *dest, const nir_alu_src *src,
1161 nir_alu_instr *instr);
1162 void nir_alu_dest_copy(nir_alu_dest *dest, const nir_alu_dest *src,
1163 nir_alu_instr *instr);
1164
1165 /* is this source channel used? */
1166 static inline bool
1167 nir_alu_instr_channel_used(const nir_alu_instr *instr, unsigned src,
1168 unsigned channel)
1169 {
1170 if (nir_op_infos[instr->op].input_sizes[src] > 0)
1171 return channel < nir_op_infos[instr->op].input_sizes[src];
1172
1173 return (instr->dest.write_mask >> channel) & 1;
1174 }
1175
1176 static inline nir_component_mask_t
1177 nir_alu_instr_src_read_mask(const nir_alu_instr *instr, unsigned src)
1178 {
1179 nir_component_mask_t read_mask = 0;
1180 for (unsigned c = 0; c < NIR_MAX_VEC_COMPONENTS; c++) {
1181 if (!nir_alu_instr_channel_used(instr, src, c))
1182 continue;
1183
1184 read_mask |= (1 << instr->src[src].swizzle[c]);
1185 }
1186 return read_mask;
1187 }
1188
1189 /**
1190 * Get the number of channels used for a source
1191 */
1192 static inline unsigned
1193 nir_ssa_alu_instr_src_components(const nir_alu_instr *instr, unsigned src)
1194 {
1195 if (nir_op_infos[instr->op].input_sizes[src] > 0)
1196 return nir_op_infos[instr->op].input_sizes[src];
1197
1198 return nir_dest_num_components(instr->dest.dest);
1199 }
1200
1201 static inline bool
1202 nir_alu_instr_is_comparison(const nir_alu_instr *instr)
1203 {
1204 switch (instr->op) {
1205 case nir_op_flt:
1206 case nir_op_fge:
1207 case nir_op_feq:
1208 case nir_op_fne:
1209 case nir_op_ilt:
1210 case nir_op_ult:
1211 case nir_op_ige:
1212 case nir_op_uge:
1213 case nir_op_ieq:
1214 case nir_op_ine:
1215 case nir_op_i2b1:
1216 case nir_op_f2b1:
1217 case nir_op_inot:
1218 return true;
1219 default:
1220 return false;
1221 }
1222 }
1223
1224 bool nir_const_value_negative_equal(nir_const_value c1, nir_const_value c2,
1225 nir_alu_type full_type);
1226
1227 bool nir_alu_srcs_equal(const nir_alu_instr *alu1, const nir_alu_instr *alu2,
1228 unsigned src1, unsigned src2);
1229
1230 bool nir_alu_srcs_negative_equal(const nir_alu_instr *alu1,
1231 const nir_alu_instr *alu2,
1232 unsigned src1, unsigned src2);
1233
1234 typedef enum {
1235 nir_deref_type_var,
1236 nir_deref_type_array,
1237 nir_deref_type_array_wildcard,
1238 nir_deref_type_ptr_as_array,
1239 nir_deref_type_struct,
1240 nir_deref_type_cast,
1241 } nir_deref_type;
1242
1243 typedef struct {
1244 nir_instr instr;
1245
1246 /** The type of this deref instruction */
1247 nir_deref_type deref_type;
1248
1249 /** The mode of the underlying variable */
1250 nir_variable_mode mode;
1251
1252 /** The dereferenced type of the resulting pointer value */
1253 const struct glsl_type *type;
1254
1255 union {
1256 /** Variable being dereferenced if deref_type is a deref_var */
1257 nir_variable *var;
1258
1259 /** Parent deref if deref_type is not deref_var */
1260 nir_src parent;
1261 };
1262
1263 /** Additional deref parameters */
1264 union {
1265 struct {
1266 nir_src index;
1267 } arr;
1268
1269 struct {
1270 unsigned index;
1271 } strct;
1272
1273 struct {
1274 unsigned ptr_stride;
1275 } cast;
1276 };
1277
1278 /** Destination to store the resulting "pointer" */
1279 nir_dest dest;
1280 } nir_deref_instr;
1281
1282 static inline nir_deref_instr *nir_src_as_deref(nir_src src);
1283
1284 static inline nir_deref_instr *
1285 nir_deref_instr_parent(const nir_deref_instr *instr)
1286 {
1287 if (instr->deref_type == nir_deref_type_var)
1288 return NULL;
1289 else
1290 return nir_src_as_deref(instr->parent);
1291 }
1292
1293 static inline nir_variable *
1294 nir_deref_instr_get_variable(const nir_deref_instr *instr)
1295 {
1296 while (instr->deref_type != nir_deref_type_var) {
1297 if (instr->deref_type == nir_deref_type_cast)
1298 return NULL;
1299
1300 instr = nir_deref_instr_parent(instr);
1301 }
1302
1303 return instr->var;
1304 }
1305
1306 bool nir_deref_instr_has_indirect(nir_deref_instr *instr);
1307 bool nir_deref_instr_is_known_out_of_bounds(nir_deref_instr *instr);
1308 bool nir_deref_instr_has_complex_use(nir_deref_instr *instr);
1309
1310 bool nir_deref_instr_remove_if_unused(nir_deref_instr *instr);
1311
1312 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr *instr);
1313
1314 typedef struct {
1315 nir_instr instr;
1316
1317 struct nir_function *callee;
1318
1319 unsigned num_params;
1320 nir_src params[];
1321 } nir_call_instr;
1322
1323 #include "nir_intrinsics.h"
1324
1325 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1326
1327 /** Represents an intrinsic
1328 *
1329 * An intrinsic is an instruction type for handling things that are
1330 * more-or-less regular operations but don't just consume and produce SSA
1331 * values like ALU operations do. Intrinsics are not for things that have
1332 * special semantic meaning such as phi nodes and parallel copies.
1333 * Examples of intrinsics include variable load/store operations, system
1334 * value loads, and the like. Even though texturing more-or-less falls
1335 * under this category, texturing is its own instruction type because
1336 * trying to represent texturing with intrinsics would lead to a
1337 * combinatorial explosion of intrinsic opcodes.
1338 *
1339 * By having a single instruction type for handling a lot of different
1340 * cases, optimization passes can look for intrinsics and, for the most
1341 * part, completely ignore them. Each intrinsic type also has a few
1342 * possible flags that govern whether or not they can be reordered or
1343 * eliminated. That way passes like dead code elimination can still work
1344 * on intrisics without understanding the meaning of each.
1345 *
1346 * Each intrinsic has some number of constant indices, some number of
1347 * variables, and some number of sources. What these sources, variables,
1348 * and indices mean depends on the intrinsic and is documented with the
1349 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1350 * instructions are the only types of instruction that can operate on
1351 * variables.
1352 */
1353 typedef struct {
1354 nir_instr instr;
1355
1356 nir_intrinsic_op intrinsic;
1357
1358 nir_dest dest;
1359
1360 /** number of components if this is a vectorized intrinsic
1361 *
1362 * Similarly to ALU operations, some intrinsics are vectorized.
1363 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1364 * For vectorized intrinsics, the num_components field specifies the
1365 * number of destination components and the number of source components
1366 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1367 */
1368 uint8_t num_components;
1369
1370 int const_index[NIR_INTRINSIC_MAX_CONST_INDEX];
1371
1372 nir_src src[];
1373 } nir_intrinsic_instr;
1374
1375 static inline nir_variable *
1376 nir_intrinsic_get_var(nir_intrinsic_instr *intrin, unsigned i)
1377 {
1378 return nir_deref_instr_get_variable(nir_src_as_deref(intrin->src[i]));
1379 }
1380
1381 /**
1382 * \name NIR intrinsics semantic flags
1383 *
1384 * information about what the compiler can do with the intrinsics.
1385 *
1386 * \sa nir_intrinsic_info::flags
1387 */
1388 typedef enum {
1389 /**
1390 * whether the intrinsic can be safely eliminated if none of its output
1391 * value is not being used.
1392 */
1393 NIR_INTRINSIC_CAN_ELIMINATE = (1 << 0),
1394
1395 /**
1396 * Whether the intrinsic can be reordered with respect to any other
1397 * intrinsic, i.e. whether the only reordering dependencies of the
1398 * intrinsic are due to the register reads/writes.
1399 */
1400 NIR_INTRINSIC_CAN_REORDER = (1 << 1),
1401 } nir_intrinsic_semantic_flag;
1402
1403 /**
1404 * \name NIR intrinsics const-index flag
1405 *
1406 * Indicates the usage of a const_index slot.
1407 *
1408 * \sa nir_intrinsic_info::index_map
1409 */
1410 typedef enum {
1411 /**
1412 * Generally instructions that take a offset src argument, can encode
1413 * a constant 'base' value which is added to the offset.
1414 */
1415 NIR_INTRINSIC_BASE = 1,
1416
1417 /**
1418 * For store instructions, a writemask for the store.
1419 */
1420 NIR_INTRINSIC_WRMASK,
1421
1422 /**
1423 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1424 */
1425 NIR_INTRINSIC_STREAM_ID,
1426
1427 /**
1428 * The clip-plane id for load_user_clip_plane intrinsic.
1429 */
1430 NIR_INTRINSIC_UCP_ID,
1431
1432 /**
1433 * The amount of data, starting from BASE, that this instruction may
1434 * access. This is used to provide bounds if the offset is not constant.
1435 */
1436 NIR_INTRINSIC_RANGE,
1437
1438 /**
1439 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1440 */
1441 NIR_INTRINSIC_DESC_SET,
1442
1443 /**
1444 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1445 */
1446 NIR_INTRINSIC_BINDING,
1447
1448 /**
1449 * Component offset.
1450 */
1451 NIR_INTRINSIC_COMPONENT,
1452
1453 /**
1454 * Interpolation mode (only meaningful for FS inputs).
1455 */
1456 NIR_INTRINSIC_INTERP_MODE,
1457
1458 /**
1459 * A binary nir_op to use when performing a reduction or scan operation
1460 */
1461 NIR_INTRINSIC_REDUCTION_OP,
1462
1463 /**
1464 * Cluster size for reduction operations
1465 */
1466 NIR_INTRINSIC_CLUSTER_SIZE,
1467
1468 /**
1469 * Parameter index for a load_param intrinsic
1470 */
1471 NIR_INTRINSIC_PARAM_IDX,
1472
1473 /**
1474 * Image dimensionality for image intrinsics
1475 *
1476 * One of GLSL_SAMPLER_DIM_*
1477 */
1478 NIR_INTRINSIC_IMAGE_DIM,
1479
1480 /**
1481 * Non-zero if we are accessing an array image
1482 */
1483 NIR_INTRINSIC_IMAGE_ARRAY,
1484
1485 /**
1486 * Image format for image intrinsics
1487 */
1488 NIR_INTRINSIC_FORMAT,
1489
1490 /**
1491 * Access qualifiers for image and memory access intrinsics
1492 */
1493 NIR_INTRINSIC_ACCESS,
1494
1495 /**
1496 * Alignment for offsets and addresses
1497 *
1498 * These two parameters, specify an alignment in terms of a multiplier and
1499 * an offset. The offset or address parameter X of the intrinsic is
1500 * guaranteed to satisfy the following:
1501 *
1502 * (X - align_offset) % align_mul == 0
1503 */
1504 NIR_INTRINSIC_ALIGN_MUL,
1505 NIR_INTRINSIC_ALIGN_OFFSET,
1506
1507 /**
1508 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1509 */
1510 NIR_INTRINSIC_DESC_TYPE,
1511
1512 /**
1513 * The nir_alu_type of a uniform/input/output
1514 */
1515 NIR_INTRINSIC_TYPE,
1516
1517 /**
1518 * The swizzle mask for the instructions
1519 * SwizzleInvocationsAMD and SwizzleInvocationsMaskedAMD
1520 */
1521 NIR_INTRINSIC_SWIZZLE_MASK,
1522
1523 /* Separate source/dest access flags for copies */
1524 NIR_INTRINSIC_SRC_ACCESS,
1525 NIR_INTRINSIC_DST_ACCESS,
1526
1527 NIR_INTRINSIC_NUM_INDEX_FLAGS,
1528
1529 } nir_intrinsic_index_flag;
1530
1531 #define NIR_INTRINSIC_MAX_INPUTS 5
1532
1533 typedef struct {
1534 const char *name;
1535
1536 unsigned num_srcs; /** < number of register/SSA inputs */
1537
1538 /** number of components of each input register
1539 *
1540 * If this value is 0, the number of components is given by the
1541 * num_components field of nir_intrinsic_instr. If this value is -1, the
1542 * intrinsic consumes however many components are provided and it is not
1543 * validated at all.
1544 */
1545 int src_components[NIR_INTRINSIC_MAX_INPUTS];
1546
1547 bool has_dest;
1548
1549 /** number of components of the output register
1550 *
1551 * If this value is 0, the number of components is given by the
1552 * num_components field of nir_intrinsic_instr.
1553 */
1554 unsigned dest_components;
1555
1556 /** bitfield of legal bit sizes */
1557 unsigned dest_bit_sizes;
1558
1559 /** the number of constant indices used by the intrinsic */
1560 unsigned num_indices;
1561
1562 /** indicates the usage of intr->const_index[n] */
1563 unsigned index_map[NIR_INTRINSIC_NUM_INDEX_FLAGS];
1564
1565 /** semantic flags for calls to this intrinsic */
1566 nir_intrinsic_semantic_flag flags;
1567 } nir_intrinsic_info;
1568
1569 extern const nir_intrinsic_info nir_intrinsic_infos[nir_num_intrinsics];
1570
1571 static inline unsigned
1572 nir_intrinsic_src_components(nir_intrinsic_instr *intr, unsigned srcn)
1573 {
1574 const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
1575 assert(srcn < info->num_srcs);
1576 if (info->src_components[srcn] > 0)
1577 return info->src_components[srcn];
1578 else if (info->src_components[srcn] == 0)
1579 return intr->num_components;
1580 else
1581 return nir_src_num_components(intr->src[srcn]);
1582 }
1583
1584 static inline unsigned
1585 nir_intrinsic_dest_components(nir_intrinsic_instr *intr)
1586 {
1587 const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
1588 if (!info->has_dest)
1589 return 0;
1590 else if (info->dest_components)
1591 return info->dest_components;
1592 else
1593 return intr->num_components;
1594 }
1595
1596 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1597 static inline type \
1598 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1599 { \
1600 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1601 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1602 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1603 } \
1604 static inline void \
1605 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1606 { \
1607 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1608 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1609 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1610 }
1611
1612 INTRINSIC_IDX_ACCESSORS(write_mask, WRMASK, unsigned)
1613 INTRINSIC_IDX_ACCESSORS(base, BASE, int)
1614 INTRINSIC_IDX_ACCESSORS(stream_id, STREAM_ID, unsigned)
1615 INTRINSIC_IDX_ACCESSORS(ucp_id, UCP_ID, unsigned)
1616 INTRINSIC_IDX_ACCESSORS(range, RANGE, unsigned)
1617 INTRINSIC_IDX_ACCESSORS(desc_set, DESC_SET, unsigned)
1618 INTRINSIC_IDX_ACCESSORS(binding, BINDING, unsigned)
1619 INTRINSIC_IDX_ACCESSORS(component, COMPONENT, unsigned)
1620 INTRINSIC_IDX_ACCESSORS(interp_mode, INTERP_MODE, unsigned)
1621 INTRINSIC_IDX_ACCESSORS(reduction_op, REDUCTION_OP, unsigned)
1622 INTRINSIC_IDX_ACCESSORS(cluster_size, CLUSTER_SIZE, unsigned)
1623 INTRINSIC_IDX_ACCESSORS(param_idx, PARAM_IDX, unsigned)
1624 INTRINSIC_IDX_ACCESSORS(image_dim, IMAGE_DIM, enum glsl_sampler_dim)
1625 INTRINSIC_IDX_ACCESSORS(image_array, IMAGE_ARRAY, bool)
1626 INTRINSIC_IDX_ACCESSORS(access, ACCESS, enum gl_access_qualifier)
1627 INTRINSIC_IDX_ACCESSORS(src_access, SRC_ACCESS, enum gl_access_qualifier)
1628 INTRINSIC_IDX_ACCESSORS(dst_access, DST_ACCESS, enum gl_access_qualifier)
1629 INTRINSIC_IDX_ACCESSORS(format, FORMAT, unsigned)
1630 INTRINSIC_IDX_ACCESSORS(align_mul, ALIGN_MUL, unsigned)
1631 INTRINSIC_IDX_ACCESSORS(align_offset, ALIGN_OFFSET, unsigned)
1632 INTRINSIC_IDX_ACCESSORS(desc_type, DESC_TYPE, unsigned)
1633 INTRINSIC_IDX_ACCESSORS(type, TYPE, nir_alu_type)
1634 INTRINSIC_IDX_ACCESSORS(swizzle_mask, SWIZZLE_MASK, unsigned)
1635
1636 static inline void
1637 nir_intrinsic_set_align(nir_intrinsic_instr *intrin,
1638 unsigned align_mul, unsigned align_offset)
1639 {
1640 assert(util_is_power_of_two_nonzero(align_mul));
1641 assert(align_offset < align_mul);
1642 nir_intrinsic_set_align_mul(intrin, align_mul);
1643 nir_intrinsic_set_align_offset(intrin, align_offset);
1644 }
1645
1646 /** Returns a simple alignment for a load/store intrinsic offset
1647 *
1648 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1649 * and ALIGN_OFFSET parameters, this helper takes both into account and
1650 * provides a single simple alignment parameter. The offset X is guaranteed
1651 * to satisfy X % align == 0.
1652 */
1653 static inline unsigned
1654 nir_intrinsic_align(const nir_intrinsic_instr *intrin)
1655 {
1656 const unsigned align_mul = nir_intrinsic_align_mul(intrin);
1657 const unsigned align_offset = nir_intrinsic_align_offset(intrin);
1658 assert(align_offset < align_mul);
1659 return align_offset ? 1 << (ffs(align_offset) - 1) : align_mul;
1660 }
1661
1662 /* Converts a image_deref_* intrinsic into a image_* one */
1663 void nir_rewrite_image_intrinsic(nir_intrinsic_instr *instr,
1664 nir_ssa_def *handle, bool bindless);
1665
1666 /* Determine if an intrinsic can be arbitrarily reordered and eliminated. */
1667 static inline bool
1668 nir_intrinsic_can_reorder(nir_intrinsic_instr *instr)
1669 {
1670 if (instr->intrinsic == nir_intrinsic_load_deref ||
1671 instr->intrinsic == nir_intrinsic_load_ssbo ||
1672 instr->intrinsic == nir_intrinsic_bindless_image_load ||
1673 instr->intrinsic == nir_intrinsic_image_deref_load ||
1674 instr->intrinsic == nir_intrinsic_image_load) {
1675 return nir_intrinsic_access(instr) & ACCESS_CAN_REORDER;
1676 } else {
1677 const nir_intrinsic_info *info =
1678 &nir_intrinsic_infos[instr->intrinsic];
1679 return (info->flags & NIR_INTRINSIC_CAN_ELIMINATE) &&
1680 (info->flags & NIR_INTRINSIC_CAN_REORDER);
1681 }
1682 }
1683
1684 /**
1685 * \group texture information
1686 *
1687 * This gives semantic information about textures which is useful to the
1688 * frontend, the backend, and lowering passes, but not the optimizer.
1689 */
1690
1691 typedef enum {
1692 nir_tex_src_coord,
1693 nir_tex_src_projector,
1694 nir_tex_src_comparator, /* shadow comparator */
1695 nir_tex_src_offset,
1696 nir_tex_src_bias,
1697 nir_tex_src_lod,
1698 nir_tex_src_min_lod,
1699 nir_tex_src_ms_index, /* MSAA sample index */
1700 nir_tex_src_ms_mcs, /* MSAA compression value */
1701 nir_tex_src_ddx,
1702 nir_tex_src_ddy,
1703 nir_tex_src_texture_deref, /* < deref pointing to the texture */
1704 nir_tex_src_sampler_deref, /* < deref pointing to the sampler */
1705 nir_tex_src_texture_offset, /* < dynamically uniform indirect offset */
1706 nir_tex_src_sampler_offset, /* < dynamically uniform indirect offset */
1707 nir_tex_src_texture_handle, /* < bindless texture handle */
1708 nir_tex_src_sampler_handle, /* < bindless sampler handle */
1709 nir_tex_src_plane, /* < selects plane for planar textures */
1710 nir_num_tex_src_types
1711 } nir_tex_src_type;
1712
1713 typedef struct {
1714 nir_src src;
1715 nir_tex_src_type src_type;
1716 } nir_tex_src;
1717
1718 typedef enum {
1719 nir_texop_tex, /**< Regular texture look-up */
1720 nir_texop_txb, /**< Texture look-up with LOD bias */
1721 nir_texop_txl, /**< Texture look-up with explicit LOD */
1722 nir_texop_txd, /**< Texture look-up with partial derivatives */
1723 nir_texop_txf, /**< Texel fetch with explicit LOD */
1724 nir_texop_txf_ms, /**< Multisample texture fetch */
1725 nir_texop_txf_ms_fb, /**< Multisample texture fetch from framebuffer */
1726 nir_texop_txf_ms_mcs, /**< Multisample compression value fetch */
1727 nir_texop_txs, /**< Texture size */
1728 nir_texop_lod, /**< Texture lod query */
1729 nir_texop_tg4, /**< Texture gather */
1730 nir_texop_query_levels, /**< Texture levels query */
1731 nir_texop_texture_samples, /**< Texture samples query */
1732 nir_texop_samples_identical, /**< Query whether all samples are definitely
1733 * identical.
1734 */
1735 } nir_texop;
1736
1737 typedef struct {
1738 nir_instr instr;
1739
1740 enum glsl_sampler_dim sampler_dim;
1741 nir_alu_type dest_type;
1742
1743 nir_texop op;
1744 nir_dest dest;
1745 nir_tex_src *src;
1746 unsigned num_srcs, coord_components;
1747 bool is_array, is_shadow;
1748
1749 /**
1750 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1751 * components or the new-style shadow that outputs 1 component.
1752 */
1753 bool is_new_style_shadow;
1754
1755 /* gather component selector */
1756 unsigned component : 2;
1757
1758 /* gather offsets */
1759 int8_t tg4_offsets[4][2];
1760
1761 /* True if the texture index or handle is not dynamically uniform */
1762 bool texture_non_uniform;
1763
1764 /* True if the sampler index or handle is not dynamically uniform */
1765 bool sampler_non_uniform;
1766
1767 /** The texture index
1768 *
1769 * If this texture instruction has a nir_tex_src_texture_offset source,
1770 * then the texture index is given by texture_index + texture_offset.
1771 */
1772 unsigned texture_index;
1773
1774 /** The size of the texture array or 0 if it's not an array */
1775 unsigned texture_array_size;
1776
1777 /** The sampler index
1778 *
1779 * The following operations do not require a sampler and, as such, this
1780 * field should be ignored:
1781 * - nir_texop_txf
1782 * - nir_texop_txf_ms
1783 * - nir_texop_txs
1784 * - nir_texop_lod
1785 * - nir_texop_query_levels
1786 * - nir_texop_texture_samples
1787 * - nir_texop_samples_identical
1788 *
1789 * If this texture instruction has a nir_tex_src_sampler_offset source,
1790 * then the sampler index is given by sampler_index + sampler_offset.
1791 */
1792 unsigned sampler_index;
1793 } nir_tex_instr;
1794
1795 static inline unsigned
1796 nir_tex_instr_dest_size(const nir_tex_instr *instr)
1797 {
1798 switch (instr->op) {
1799 case nir_texop_txs: {
1800 unsigned ret;
1801 switch (instr->sampler_dim) {
1802 case GLSL_SAMPLER_DIM_1D:
1803 case GLSL_SAMPLER_DIM_BUF:
1804 ret = 1;
1805 break;
1806 case GLSL_SAMPLER_DIM_2D:
1807 case GLSL_SAMPLER_DIM_CUBE:
1808 case GLSL_SAMPLER_DIM_MS:
1809 case GLSL_SAMPLER_DIM_RECT:
1810 case GLSL_SAMPLER_DIM_EXTERNAL:
1811 case GLSL_SAMPLER_DIM_SUBPASS:
1812 ret = 2;
1813 break;
1814 case GLSL_SAMPLER_DIM_3D:
1815 ret = 3;
1816 break;
1817 default:
1818 unreachable("not reached");
1819 }
1820 if (instr->is_array)
1821 ret++;
1822 return ret;
1823 }
1824
1825 case nir_texop_lod:
1826 return 2;
1827
1828 case nir_texop_texture_samples:
1829 case nir_texop_query_levels:
1830 case nir_texop_samples_identical:
1831 return 1;
1832
1833 default:
1834 if (instr->is_shadow && instr->is_new_style_shadow)
1835 return 1;
1836
1837 return 4;
1838 }
1839 }
1840
1841 /* Returns true if this texture operation queries something about the texture
1842 * rather than actually sampling it.
1843 */
1844 static inline bool
1845 nir_tex_instr_is_query(const nir_tex_instr *instr)
1846 {
1847 switch (instr->op) {
1848 case nir_texop_txs:
1849 case nir_texop_lod:
1850 case nir_texop_texture_samples:
1851 case nir_texop_query_levels:
1852 case nir_texop_txf_ms_mcs:
1853 return true;
1854 case nir_texop_tex:
1855 case nir_texop_txb:
1856 case nir_texop_txl:
1857 case nir_texop_txd:
1858 case nir_texop_txf:
1859 case nir_texop_txf_ms:
1860 case nir_texop_txf_ms_fb:
1861 case nir_texop_tg4:
1862 return false;
1863 default:
1864 unreachable("Invalid texture opcode");
1865 }
1866 }
1867
1868 static inline bool
1869 nir_tex_instr_has_implicit_derivative(const nir_tex_instr *instr)
1870 {
1871 switch (instr->op) {
1872 case nir_texop_tex:
1873 case nir_texop_txb:
1874 case nir_texop_lod:
1875 return true;
1876 default:
1877 return false;
1878 }
1879 }
1880
1881 static inline nir_alu_type
1882 nir_tex_instr_src_type(const nir_tex_instr *instr, unsigned src)
1883 {
1884 switch (instr->src[src].src_type) {
1885 case nir_tex_src_coord:
1886 switch (instr->op) {
1887 case nir_texop_txf:
1888 case nir_texop_txf_ms:
1889 case nir_texop_txf_ms_fb:
1890 case nir_texop_txf_ms_mcs:
1891 case nir_texop_samples_identical:
1892 return nir_type_int;
1893
1894 default:
1895 return nir_type_float;
1896 }
1897
1898 case nir_tex_src_lod:
1899 switch (instr->op) {
1900 case nir_texop_txs:
1901 case nir_texop_txf:
1902 return nir_type_int;
1903
1904 default:
1905 return nir_type_float;
1906 }
1907
1908 case nir_tex_src_projector:
1909 case nir_tex_src_comparator:
1910 case nir_tex_src_bias:
1911 case nir_tex_src_min_lod:
1912 case nir_tex_src_ddx:
1913 case nir_tex_src_ddy:
1914 return nir_type_float;
1915
1916 case nir_tex_src_offset:
1917 case nir_tex_src_ms_index:
1918 case nir_tex_src_plane:
1919 return nir_type_int;
1920
1921 case nir_tex_src_ms_mcs:
1922 case nir_tex_src_texture_deref:
1923 case nir_tex_src_sampler_deref:
1924 case nir_tex_src_texture_offset:
1925 case nir_tex_src_sampler_offset:
1926 case nir_tex_src_texture_handle:
1927 case nir_tex_src_sampler_handle:
1928 return nir_type_uint;
1929
1930 case nir_num_tex_src_types:
1931 unreachable("nir_num_tex_src_types is not a valid source type");
1932 }
1933
1934 unreachable("Invalid texture source type");
1935 }
1936
1937 static inline unsigned
1938 nir_tex_instr_src_size(const nir_tex_instr *instr, unsigned src)
1939 {
1940 if (instr->src[src].src_type == nir_tex_src_coord)
1941 return instr->coord_components;
1942
1943 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
1944 if (instr->src[src].src_type == nir_tex_src_ms_mcs)
1945 return 4;
1946
1947 if (instr->src[src].src_type == nir_tex_src_ddx ||
1948 instr->src[src].src_type == nir_tex_src_ddy) {
1949 if (instr->is_array)
1950 return instr->coord_components - 1;
1951 else
1952 return instr->coord_components;
1953 }
1954
1955 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
1956 * the offset, since a cube maps to a single face.
1957 */
1958 if (instr->src[src].src_type == nir_tex_src_offset) {
1959 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE)
1960 return 2;
1961 else if (instr->is_array)
1962 return instr->coord_components - 1;
1963 else
1964 return instr->coord_components;
1965 }
1966
1967 return 1;
1968 }
1969
1970 static inline int
1971 nir_tex_instr_src_index(const nir_tex_instr *instr, nir_tex_src_type type)
1972 {
1973 for (unsigned i = 0; i < instr->num_srcs; i++)
1974 if (instr->src[i].src_type == type)
1975 return (int) i;
1976
1977 return -1;
1978 }
1979
1980 void nir_tex_instr_add_src(nir_tex_instr *tex,
1981 nir_tex_src_type src_type,
1982 nir_src src);
1983
1984 void nir_tex_instr_remove_src(nir_tex_instr *tex, unsigned src_idx);
1985
1986 bool nir_tex_instr_has_explicit_tg4_offsets(nir_tex_instr *tex);
1987
1988 typedef struct {
1989 nir_instr instr;
1990
1991 nir_ssa_def def;
1992
1993 nir_const_value value[];
1994 } nir_load_const_instr;
1995
1996 typedef enum {
1997 nir_jump_return,
1998 nir_jump_break,
1999 nir_jump_continue,
2000 } nir_jump_type;
2001
2002 typedef struct {
2003 nir_instr instr;
2004 nir_jump_type type;
2005 } nir_jump_instr;
2006
2007 /* creates a new SSA variable in an undefined state */
2008
2009 typedef struct {
2010 nir_instr instr;
2011 nir_ssa_def def;
2012 } nir_ssa_undef_instr;
2013
2014 typedef struct {
2015 struct exec_node node;
2016
2017 /* The predecessor block corresponding to this source */
2018 struct nir_block *pred;
2019
2020 nir_src src;
2021 } nir_phi_src;
2022
2023 #define nir_foreach_phi_src(phi_src, phi) \
2024 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
2025 #define nir_foreach_phi_src_safe(phi_src, phi) \
2026 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
2027
2028 typedef struct {
2029 nir_instr instr;
2030
2031 struct exec_list srcs; /** < list of nir_phi_src */
2032
2033 nir_dest dest;
2034 } nir_phi_instr;
2035
2036 typedef struct {
2037 struct exec_node node;
2038 nir_src src;
2039 nir_dest dest;
2040 } nir_parallel_copy_entry;
2041
2042 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
2043 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
2044
2045 typedef struct {
2046 nir_instr instr;
2047
2048 /* A list of nir_parallel_copy_entrys. The sources of all of the
2049 * entries are copied to the corresponding destinations "in parallel".
2050 * In other words, if we have two entries: a -> b and b -> a, the values
2051 * get swapped.
2052 */
2053 struct exec_list entries;
2054 } nir_parallel_copy_instr;
2055
2056 NIR_DEFINE_CAST(nir_instr_as_alu, nir_instr, nir_alu_instr, instr,
2057 type, nir_instr_type_alu)
2058 NIR_DEFINE_CAST(nir_instr_as_deref, nir_instr, nir_deref_instr, instr,
2059 type, nir_instr_type_deref)
2060 NIR_DEFINE_CAST(nir_instr_as_call, nir_instr, nir_call_instr, instr,
2061 type, nir_instr_type_call)
2062 NIR_DEFINE_CAST(nir_instr_as_jump, nir_instr, nir_jump_instr, instr,
2063 type, nir_instr_type_jump)
2064 NIR_DEFINE_CAST(nir_instr_as_tex, nir_instr, nir_tex_instr, instr,
2065 type, nir_instr_type_tex)
2066 NIR_DEFINE_CAST(nir_instr_as_intrinsic, nir_instr, nir_intrinsic_instr, instr,
2067 type, nir_instr_type_intrinsic)
2068 NIR_DEFINE_CAST(nir_instr_as_load_const, nir_instr, nir_load_const_instr, instr,
2069 type, nir_instr_type_load_const)
2070 NIR_DEFINE_CAST(nir_instr_as_ssa_undef, nir_instr, nir_ssa_undef_instr, instr,
2071 type, nir_instr_type_ssa_undef)
2072 NIR_DEFINE_CAST(nir_instr_as_phi, nir_instr, nir_phi_instr, instr,
2073 type, nir_instr_type_phi)
2074 NIR_DEFINE_CAST(nir_instr_as_parallel_copy, nir_instr,
2075 nir_parallel_copy_instr, instr,
2076 type, nir_instr_type_parallel_copy)
2077
2078
2079 #define NIR_DEFINE_SRC_AS_CONST(type, suffix) \
2080 static inline type \
2081 nir_src_comp_as_##suffix(nir_src src, unsigned comp) \
2082 { \
2083 assert(nir_src_is_const(src)); \
2084 nir_load_const_instr *load = \
2085 nir_instr_as_load_const(src.ssa->parent_instr); \
2086 assert(comp < load->def.num_components); \
2087 return nir_const_value_as_##suffix(load->value[comp], \
2088 load->def.bit_size); \
2089 } \
2090 \
2091 static inline type \
2092 nir_src_as_##suffix(nir_src src) \
2093 { \
2094 assert(nir_src_num_components(src) == 1); \
2095 return nir_src_comp_as_##suffix(src, 0); \
2096 }
2097
2098 NIR_DEFINE_SRC_AS_CONST(int64_t, int)
2099 NIR_DEFINE_SRC_AS_CONST(uint64_t, uint)
2100 NIR_DEFINE_SRC_AS_CONST(bool, bool)
2101 NIR_DEFINE_SRC_AS_CONST(double, float)
2102
2103 #undef NIR_DEFINE_SRC_AS_CONST
2104
2105
2106 typedef struct {
2107 nir_ssa_def *def;
2108 unsigned comp;
2109 } nir_ssa_scalar;
2110
2111 static inline bool
2112 nir_ssa_scalar_is_const(nir_ssa_scalar s)
2113 {
2114 return s.def->parent_instr->type == nir_instr_type_load_const;
2115 }
2116
2117 static inline nir_const_value
2118 nir_ssa_scalar_as_const_value(nir_ssa_scalar s)
2119 {
2120 assert(s.comp < s.def->num_components);
2121 nir_load_const_instr *load = nir_instr_as_load_const(s.def->parent_instr);
2122 return load->value[s.comp];
2123 }
2124
2125 #define NIR_DEFINE_SCALAR_AS_CONST(type, suffix) \
2126 static inline type \
2127 nir_ssa_scalar_as_##suffix(nir_ssa_scalar s) \
2128 { \
2129 return nir_const_value_as_##suffix( \
2130 nir_ssa_scalar_as_const_value(s), s.def->bit_size); \
2131 }
2132
2133 NIR_DEFINE_SCALAR_AS_CONST(int64_t, int)
2134 NIR_DEFINE_SCALAR_AS_CONST(uint64_t, uint)
2135 NIR_DEFINE_SCALAR_AS_CONST(bool, bool)
2136 NIR_DEFINE_SCALAR_AS_CONST(double, float)
2137
2138 #undef NIR_DEFINE_SCALAR_AS_CONST
2139
2140 static inline bool
2141 nir_ssa_scalar_is_alu(nir_ssa_scalar s)
2142 {
2143 return s.def->parent_instr->type == nir_instr_type_alu;
2144 }
2145
2146 static inline nir_op
2147 nir_ssa_scalar_alu_op(nir_ssa_scalar s)
2148 {
2149 return nir_instr_as_alu(s.def->parent_instr)->op;
2150 }
2151
2152 static inline nir_ssa_scalar
2153 nir_ssa_scalar_chase_alu_src(nir_ssa_scalar s, unsigned alu_src_idx)
2154 {
2155 nir_ssa_scalar out = { NULL, 0 };
2156
2157 nir_alu_instr *alu = nir_instr_as_alu(s.def->parent_instr);
2158 assert(alu_src_idx < nir_op_infos[alu->op].num_inputs);
2159
2160 /* Our component must be written */
2161 assert(s.comp < s.def->num_components);
2162 assert(alu->dest.write_mask & (1u << s.comp));
2163
2164 assert(alu->src[alu_src_idx].src.is_ssa);
2165 out.def = alu->src[alu_src_idx].src.ssa;
2166
2167 if (nir_op_infos[alu->op].input_sizes[alu_src_idx] == 0) {
2168 /* The ALU src is unsized so the source component follows the
2169 * destination component.
2170 */
2171 out.comp = alu->src[alu_src_idx].swizzle[s.comp];
2172 } else {
2173 /* This is a sized source so all source components work together to
2174 * produce all the destination components. Since we need to return a
2175 * scalar, this only works if the source is a scalar.
2176 */
2177 assert(nir_op_infos[alu->op].input_sizes[alu_src_idx] == 1);
2178 out.comp = alu->src[alu_src_idx].swizzle[0];
2179 }
2180 assert(out.comp < out.def->num_components);
2181
2182 return out;
2183 }
2184
2185
2186 /*
2187 * Control flow
2188 *
2189 * Control flow consists of a tree of control flow nodes, which include
2190 * if-statements and loops. The leaves of the tree are basic blocks, lists of
2191 * instructions that always run start-to-finish. Each basic block also keeps
2192 * track of its successors (blocks which may run immediately after the current
2193 * block) and predecessors (blocks which could have run immediately before the
2194 * current block). Each function also has a start block and an end block which
2195 * all return statements point to (which is always empty). Together, all the
2196 * blocks with their predecessors and successors make up the control flow
2197 * graph (CFG) of the function. There are helpers that modify the tree of
2198 * control flow nodes while modifying the CFG appropriately; these should be
2199 * used instead of modifying the tree directly.
2200 */
2201
2202 typedef enum {
2203 nir_cf_node_block,
2204 nir_cf_node_if,
2205 nir_cf_node_loop,
2206 nir_cf_node_function
2207 } nir_cf_node_type;
2208
2209 typedef struct nir_cf_node {
2210 struct exec_node node;
2211 nir_cf_node_type type;
2212 struct nir_cf_node *parent;
2213 } nir_cf_node;
2214
2215 typedef struct nir_block {
2216 nir_cf_node cf_node;
2217
2218 struct exec_list instr_list; /** < list of nir_instr */
2219
2220 /** generic block index; generated by nir_index_blocks */
2221 unsigned index;
2222
2223 /*
2224 * Each block can only have up to 2 successors, so we put them in a simple
2225 * array - no need for anything more complicated.
2226 */
2227 struct nir_block *successors[2];
2228
2229 /* Set of nir_block predecessors in the CFG */
2230 struct set *predecessors;
2231
2232 /*
2233 * this node's immediate dominator in the dominance tree - set to NULL for
2234 * the start block.
2235 */
2236 struct nir_block *imm_dom;
2237
2238 /* This node's children in the dominance tree */
2239 unsigned num_dom_children;
2240 struct nir_block **dom_children;
2241
2242 /* Set of nir_blocks on the dominance frontier of this block */
2243 struct set *dom_frontier;
2244
2245 /*
2246 * These two indices have the property that dom_{pre,post}_index for each
2247 * child of this block in the dominance tree will always be between
2248 * dom_pre_index and dom_post_index for this block, which makes testing if
2249 * a given block is dominated by another block an O(1) operation.
2250 */
2251 unsigned dom_pre_index, dom_post_index;
2252
2253 /* live in and out for this block; used for liveness analysis */
2254 BITSET_WORD *live_in;
2255 BITSET_WORD *live_out;
2256 } nir_block;
2257
2258 static inline nir_instr *
2259 nir_block_first_instr(nir_block *block)
2260 {
2261 struct exec_node *head = exec_list_get_head(&block->instr_list);
2262 return exec_node_data(nir_instr, head, node);
2263 }
2264
2265 static inline nir_instr *
2266 nir_block_last_instr(nir_block *block)
2267 {
2268 struct exec_node *tail = exec_list_get_tail(&block->instr_list);
2269 return exec_node_data(nir_instr, tail, node);
2270 }
2271
2272 static inline bool
2273 nir_block_ends_in_jump(nir_block *block)
2274 {
2275 return !exec_list_is_empty(&block->instr_list) &&
2276 nir_block_last_instr(block)->type == nir_instr_type_jump;
2277 }
2278
2279 #define nir_foreach_instr(instr, block) \
2280 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
2281 #define nir_foreach_instr_reverse(instr, block) \
2282 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
2283 #define nir_foreach_instr_safe(instr, block) \
2284 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
2285 #define nir_foreach_instr_reverse_safe(instr, block) \
2286 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
2287
2288 typedef enum {
2289 nir_selection_control_none = 0x0,
2290 nir_selection_control_flatten = 0x1,
2291 nir_selection_control_dont_flatten = 0x2,
2292 } nir_selection_control;
2293
2294 typedef struct nir_if {
2295 nir_cf_node cf_node;
2296 nir_src condition;
2297 nir_selection_control control;
2298
2299 struct exec_list then_list; /** < list of nir_cf_node */
2300 struct exec_list else_list; /** < list of nir_cf_node */
2301 } nir_if;
2302
2303 typedef struct {
2304 nir_if *nif;
2305
2306 /** Instruction that generates nif::condition. */
2307 nir_instr *conditional_instr;
2308
2309 /** Block within ::nif that has the break instruction. */
2310 nir_block *break_block;
2311
2312 /** Last block for the then- or else-path that does not contain the break. */
2313 nir_block *continue_from_block;
2314
2315 /** True when ::break_block is in the else-path of ::nif. */
2316 bool continue_from_then;
2317 bool induction_rhs;
2318
2319 /* This is true if the terminators exact trip count is unknown. For
2320 * example:
2321 *
2322 * for (int i = 0; i < imin(x, 4); i++)
2323 * ...
2324 *
2325 * Here loop analysis would have set a max_trip_count of 4 however we dont
2326 * know for sure that this is the exact trip count.
2327 */
2328 bool exact_trip_count_unknown;
2329
2330 struct list_head loop_terminator_link;
2331 } nir_loop_terminator;
2332
2333 typedef struct {
2334 /* Estimated cost (in number of instructions) of the loop */
2335 unsigned instr_cost;
2336
2337 /* Guessed trip count based on array indexing */
2338 unsigned guessed_trip_count;
2339
2340 /* Maximum number of times the loop is run (if known) */
2341 unsigned max_trip_count;
2342
2343 /* Do we know the exact number of times the loop will be run */
2344 bool exact_trip_count_known;
2345
2346 /* Unroll the loop regardless of its size */
2347 bool force_unroll;
2348
2349 /* Does the loop contain complex loop terminators, continues or other
2350 * complex behaviours? If this is true we can't rely on
2351 * loop_terminator_list to be complete or accurate.
2352 */
2353 bool complex_loop;
2354
2355 nir_loop_terminator *limiting_terminator;
2356
2357 /* A list of loop_terminators terminating this loop. */
2358 struct list_head loop_terminator_list;
2359 } nir_loop_info;
2360
2361 typedef enum {
2362 nir_loop_control_none = 0x0,
2363 nir_loop_control_unroll = 0x1,
2364 nir_loop_control_dont_unroll = 0x2,
2365 } nir_loop_control;
2366
2367 typedef struct {
2368 nir_cf_node cf_node;
2369
2370 struct exec_list body; /** < list of nir_cf_node */
2371
2372 nir_loop_info *info;
2373 nir_loop_control control;
2374 bool partially_unrolled;
2375 } nir_loop;
2376
2377 /**
2378 * Various bits of metadata that can may be created or required by
2379 * optimization and analysis passes
2380 */
2381 typedef enum {
2382 nir_metadata_none = 0x0,
2383 nir_metadata_block_index = 0x1,
2384 nir_metadata_dominance = 0x2,
2385 nir_metadata_live_ssa_defs = 0x4,
2386 nir_metadata_not_properly_reset = 0x8,
2387 nir_metadata_loop_analysis = 0x10,
2388 } nir_metadata;
2389
2390 typedef struct {
2391 nir_cf_node cf_node;
2392
2393 /** pointer to the function of which this is an implementation */
2394 struct nir_function *function;
2395
2396 struct exec_list body; /** < list of nir_cf_node */
2397
2398 nir_block *end_block;
2399
2400 /** list for all local variables in the function */
2401 struct exec_list locals;
2402
2403 /** list of local registers in the function */
2404 struct exec_list registers;
2405
2406 /** next available local register index */
2407 unsigned reg_alloc;
2408
2409 /** next available SSA value index */
2410 unsigned ssa_alloc;
2411
2412 /* total number of basic blocks, only valid when block_index_dirty = false */
2413 unsigned num_blocks;
2414
2415 nir_metadata valid_metadata;
2416 } nir_function_impl;
2417
2418 ATTRIBUTE_RETURNS_NONNULL static inline nir_block *
2419 nir_start_block(nir_function_impl *impl)
2420 {
2421 return (nir_block *) impl->body.head_sentinel.next;
2422 }
2423
2424 ATTRIBUTE_RETURNS_NONNULL static inline nir_block *
2425 nir_impl_last_block(nir_function_impl *impl)
2426 {
2427 return (nir_block *) impl->body.tail_sentinel.prev;
2428 }
2429
2430 static inline nir_cf_node *
2431 nir_cf_node_next(nir_cf_node *node)
2432 {
2433 struct exec_node *next = exec_node_get_next(&node->node);
2434 if (exec_node_is_tail_sentinel(next))
2435 return NULL;
2436 else
2437 return exec_node_data(nir_cf_node, next, node);
2438 }
2439
2440 static inline nir_cf_node *
2441 nir_cf_node_prev(nir_cf_node *node)
2442 {
2443 struct exec_node *prev = exec_node_get_prev(&node->node);
2444 if (exec_node_is_head_sentinel(prev))
2445 return NULL;
2446 else
2447 return exec_node_data(nir_cf_node, prev, node);
2448 }
2449
2450 static inline bool
2451 nir_cf_node_is_first(const nir_cf_node *node)
2452 {
2453 return exec_node_is_head_sentinel(node->node.prev);
2454 }
2455
2456 static inline bool
2457 nir_cf_node_is_last(const nir_cf_node *node)
2458 {
2459 return exec_node_is_tail_sentinel(node->node.next);
2460 }
2461
2462 NIR_DEFINE_CAST(nir_cf_node_as_block, nir_cf_node, nir_block, cf_node,
2463 type, nir_cf_node_block)
2464 NIR_DEFINE_CAST(nir_cf_node_as_if, nir_cf_node, nir_if, cf_node,
2465 type, nir_cf_node_if)
2466 NIR_DEFINE_CAST(nir_cf_node_as_loop, nir_cf_node, nir_loop, cf_node,
2467 type, nir_cf_node_loop)
2468 NIR_DEFINE_CAST(nir_cf_node_as_function, nir_cf_node,
2469 nir_function_impl, cf_node, type, nir_cf_node_function)
2470
2471 static inline nir_block *
2472 nir_if_first_then_block(nir_if *if_stmt)
2473 {
2474 struct exec_node *head = exec_list_get_head(&if_stmt->then_list);
2475 return nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
2476 }
2477
2478 static inline nir_block *
2479 nir_if_last_then_block(nir_if *if_stmt)
2480 {
2481 struct exec_node *tail = exec_list_get_tail(&if_stmt->then_list);
2482 return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
2483 }
2484
2485 static inline nir_block *
2486 nir_if_first_else_block(nir_if *if_stmt)
2487 {
2488 struct exec_node *head = exec_list_get_head(&if_stmt->else_list);
2489 return nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
2490 }
2491
2492 static inline nir_block *
2493 nir_if_last_else_block(nir_if *if_stmt)
2494 {
2495 struct exec_node *tail = exec_list_get_tail(&if_stmt->else_list);
2496 return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
2497 }
2498
2499 static inline nir_block *
2500 nir_loop_first_block(nir_loop *loop)
2501 {
2502 struct exec_node *head = exec_list_get_head(&loop->body);
2503 return nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
2504 }
2505
2506 static inline nir_block *
2507 nir_loop_last_block(nir_loop *loop)
2508 {
2509 struct exec_node *tail = exec_list_get_tail(&loop->body);
2510 return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
2511 }
2512
2513 /**
2514 * Return true if this list of cf_nodes contains a single empty block.
2515 */
2516 static inline bool
2517 nir_cf_list_is_empty_block(struct exec_list *cf_list)
2518 {
2519 if (exec_list_is_singular(cf_list)) {
2520 struct exec_node *head = exec_list_get_head(cf_list);
2521 nir_block *block =
2522 nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
2523 return exec_list_is_empty(&block->instr_list);
2524 }
2525 return false;
2526 }
2527
2528 typedef struct {
2529 uint8_t num_components;
2530 uint8_t bit_size;
2531 } nir_parameter;
2532
2533 typedef struct nir_function {
2534 struct exec_node node;
2535
2536 const char *name;
2537 struct nir_shader *shader;
2538
2539 unsigned num_params;
2540 nir_parameter *params;
2541
2542 /** The implementation of this function.
2543 *
2544 * If the function is only declared and not implemented, this is NULL.
2545 */
2546 nir_function_impl *impl;
2547
2548 bool is_entrypoint;
2549 } nir_function;
2550
2551 typedef enum {
2552 nir_lower_imul64 = (1 << 0),
2553 nir_lower_isign64 = (1 << 1),
2554 /** Lower all int64 modulus and division opcodes */
2555 nir_lower_divmod64 = (1 << 2),
2556 /** Lower all 64-bit umul_high and imul_high opcodes */
2557 nir_lower_imul_high64 = (1 << 3),
2558 nir_lower_mov64 = (1 << 4),
2559 nir_lower_icmp64 = (1 << 5),
2560 nir_lower_iadd64 = (1 << 6),
2561 nir_lower_iabs64 = (1 << 7),
2562 nir_lower_ineg64 = (1 << 8),
2563 nir_lower_logic64 = (1 << 9),
2564 nir_lower_minmax64 = (1 << 10),
2565 nir_lower_shift64 = (1 << 11),
2566 nir_lower_imul_2x32_64 = (1 << 12),
2567 nir_lower_extract64 = (1 << 13),
2568 } nir_lower_int64_options;
2569
2570 typedef enum {
2571 nir_lower_drcp = (1 << 0),
2572 nir_lower_dsqrt = (1 << 1),
2573 nir_lower_drsq = (1 << 2),
2574 nir_lower_dtrunc = (1 << 3),
2575 nir_lower_dfloor = (1 << 4),
2576 nir_lower_dceil = (1 << 5),
2577 nir_lower_dfract = (1 << 6),
2578 nir_lower_dround_even = (1 << 7),
2579 nir_lower_dmod = (1 << 8),
2580 nir_lower_dsub = (1 << 9),
2581 nir_lower_ddiv = (1 << 10),
2582 nir_lower_fp64_full_software = (1 << 11),
2583 } nir_lower_doubles_options;
2584
2585 typedef enum {
2586 nir_divergence_single_prim_per_subgroup = (1 << 0),
2587 nir_divergence_single_patch_per_tcs_subgroup = (1 << 1),
2588 nir_divergence_single_patch_per_tes_subgroup = (1 << 2),
2589 nir_divergence_view_index_uniform = (1 << 3),
2590 } nir_divergence_options;
2591
2592 typedef struct nir_shader_compiler_options {
2593 bool lower_fdiv;
2594 bool lower_ffma;
2595 bool fuse_ffma;
2596 bool lower_flrp16;
2597 bool lower_flrp32;
2598 /** Lowers flrp when it does not support doubles */
2599 bool lower_flrp64;
2600 bool lower_fpow;
2601 bool lower_fsat;
2602 bool lower_fsqrt;
2603 bool lower_sincos;
2604 bool lower_fmod;
2605 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2606 bool lower_bitfield_extract;
2607 /** Lowers ibitfield_extract/ubitfield_extract to compares, shifts. */
2608 bool lower_bitfield_extract_to_shifts;
2609 /** Lowers bitfield_insert to bfi/bfm */
2610 bool lower_bitfield_insert;
2611 /** Lowers bitfield_insert to compares, and shifts. */
2612 bool lower_bitfield_insert_to_shifts;
2613 /** Lowers bitfield_insert to bfm/bitfield_select. */
2614 bool lower_bitfield_insert_to_bitfield_select;
2615 /** Lowers bitfield_reverse to shifts. */
2616 bool lower_bitfield_reverse;
2617 /** Lowers bit_count to shifts. */
2618 bool lower_bit_count;
2619 /** Lowers ifind_msb to compare and ufind_msb */
2620 bool lower_ifind_msb;
2621 /** Lowers find_lsb to ufind_msb and logic ops */
2622 bool lower_find_lsb;
2623 bool lower_uadd_carry;
2624 bool lower_usub_borrow;
2625 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2626 bool lower_mul_high;
2627 /** lowers fneg and ineg to fsub and isub. */
2628 bool lower_negate;
2629 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2630 bool lower_sub;
2631
2632 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2633 bool lower_scmp;
2634
2635 /* lower fall_equalN/fany_nequalN (ex:fany_nequal4 to sne+fdot4+fsat) */
2636 bool lower_vector_cmp;
2637
2638 /** enables rules to lower idiv by power-of-two: */
2639 bool lower_idiv;
2640
2641 /** enable rules to avoid bit ops */
2642 bool lower_bitops;
2643
2644 /** enables rules to lower isign to imin+imax */
2645 bool lower_isign;
2646
2647 /** enables rules to lower fsign to fsub and flt */
2648 bool lower_fsign;
2649
2650 /* lower fdph to fdot4 */
2651 bool lower_fdph;
2652
2653 /** lower fdot to fmul and fsum/fadd. */
2654 bool lower_fdot;
2655
2656 /* Does the native fdot instruction replicate its result for four
2657 * components? If so, then opt_algebraic_late will turn all fdotN
2658 * instructions into fdot_replicatedN instructions.
2659 */
2660 bool fdot_replicates;
2661
2662 /** lowers ffloor to fsub+ffract: */
2663 bool lower_ffloor;
2664
2665 /** lowers ffract to fsub+ffloor: */
2666 bool lower_ffract;
2667
2668 /** lowers fceil to fneg+ffloor+fneg: */
2669 bool lower_fceil;
2670
2671 bool lower_ftrunc;
2672
2673 bool lower_ldexp;
2674
2675 bool lower_pack_half_2x16;
2676 bool lower_pack_unorm_2x16;
2677 bool lower_pack_snorm_2x16;
2678 bool lower_pack_unorm_4x8;
2679 bool lower_pack_snorm_4x8;
2680 bool lower_unpack_half_2x16;
2681 bool lower_unpack_unorm_2x16;
2682 bool lower_unpack_snorm_2x16;
2683 bool lower_unpack_unorm_4x8;
2684 bool lower_unpack_snorm_4x8;
2685
2686 bool lower_extract_byte;
2687 bool lower_extract_word;
2688
2689 bool lower_all_io_to_temps;
2690 bool lower_all_io_to_elements;
2691
2692 /* Indicates that the driver only has zero-based vertex id */
2693 bool vertex_id_zero_based;
2694
2695 /**
2696 * If enabled, gl_BaseVertex will be lowered as:
2697 * is_indexed_draw (~0/0) & firstvertex
2698 */
2699 bool lower_base_vertex;
2700
2701 /**
2702 * If enabled, gl_HelperInvocation will be lowered as:
2703 *
2704 * !((1 << sample_id) & sample_mask_in))
2705 *
2706 * This depends on some possibly hw implementation details, which may
2707 * not be true for all hw. In particular that the FS is only executed
2708 * for covered samples or for helper invocations. So, do not blindly
2709 * enable this option.
2710 *
2711 * Note: See also issue #22 in ARB_shader_image_load_store
2712 */
2713 bool lower_helper_invocation;
2714
2715 /**
2716 * Convert gl_SampleMaskIn to gl_HelperInvocation as follows:
2717 *
2718 * gl_SampleMaskIn == 0 ---> gl_HelperInvocation
2719 * gl_SampleMaskIn != 0 ---> !gl_HelperInvocation
2720 */
2721 bool optimize_sample_mask_in;
2722
2723 bool lower_cs_local_index_from_id;
2724 bool lower_cs_local_id_from_index;
2725
2726 bool lower_device_index_to_zero;
2727
2728 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
2729 bool lower_wpos_pntc;
2730
2731 bool lower_hadd;
2732 bool lower_add_sat;
2733
2734 /**
2735 * Should IO be re-vectorized? Some scalar ISAs still operate on vec4's
2736 * for IO purposes and would prefer loads/stores be vectorized.
2737 */
2738 bool vectorize_io;
2739 bool lower_to_scalar;
2740
2741 /**
2742 * Should nir_lower_io() create load_interpolated_input intrinsics?
2743 *
2744 * If not, it generates regular load_input intrinsics and interpolation
2745 * information must be inferred from the list of input nir_variables.
2746 */
2747 bool use_interpolated_input_intrinsics;
2748
2749 /* Lowers when 32x32->64 bit multiplication is not supported */
2750 bool lower_mul_2x32_64;
2751
2752 /* Lowers when rotate instruction is not supported */
2753 bool lower_rotate;
2754
2755 /**
2756 * Is this the Intel vec4 backend?
2757 *
2758 * Used to inhibit algebraic optimizations that are known to be harmful on
2759 * the Intel vec4 backend. This is generally applicable to any
2760 * optimization that might cause more immediate values to be used in
2761 * 3-source (e.g., ffma and flrp) instructions.
2762 */
2763 bool intel_vec4;
2764
2765 unsigned max_unroll_iterations;
2766
2767 nir_lower_int64_options lower_int64_options;
2768 nir_lower_doubles_options lower_doubles_options;
2769 } nir_shader_compiler_options;
2770
2771 typedef struct nir_shader {
2772 /** list of uniforms (nir_variable) */
2773 struct exec_list uniforms;
2774
2775 /** list of inputs (nir_variable) */
2776 struct exec_list inputs;
2777
2778 /** list of outputs (nir_variable) */
2779 struct exec_list outputs;
2780
2781 /** list of shared compute variables (nir_variable) */
2782 struct exec_list shared;
2783
2784 /** Set of driver-specific options for the shader.
2785 *
2786 * The memory for the options is expected to be kept in a single static
2787 * copy by the driver.
2788 */
2789 const struct nir_shader_compiler_options *options;
2790
2791 /** Various bits of compile-time information about a given shader */
2792 struct shader_info info;
2793
2794 /** list of global variables in the shader (nir_variable) */
2795 struct exec_list globals;
2796
2797 /** list of system value variables in the shader (nir_variable) */
2798 struct exec_list system_values;
2799
2800 struct exec_list functions; /** < list of nir_function */
2801
2802 /**
2803 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
2804 * access plus one
2805 */
2806 unsigned num_inputs, num_uniforms, num_outputs, num_shared;
2807
2808 /** Size in bytes of required scratch space */
2809 unsigned scratch_size;
2810
2811 /** Constant data associated with this shader.
2812 *
2813 * Constant data is loaded through load_constant intrinsics. See also
2814 * nir_opt_large_constants.
2815 */
2816 void *constant_data;
2817 unsigned constant_data_size;
2818 } nir_shader;
2819
2820 #define nir_foreach_function(func, shader) \
2821 foreach_list_typed(nir_function, func, node, &(shader)->functions)
2822
2823 static inline nir_function_impl *
2824 nir_shader_get_entrypoint(nir_shader *shader)
2825 {
2826 nir_function *func = NULL;
2827
2828 nir_foreach_function(function, shader) {
2829 assert(func == NULL);
2830 if (function->is_entrypoint) {
2831 func = function;
2832 #ifndef NDEBUG
2833 break;
2834 #endif
2835 }
2836 }
2837
2838 if (!func)
2839 return NULL;
2840
2841 assert(func->num_params == 0);
2842 assert(func->impl);
2843 return func->impl;
2844 }
2845
2846 nir_shader *nir_shader_create(void *mem_ctx,
2847 gl_shader_stage stage,
2848 const nir_shader_compiler_options *options,
2849 shader_info *si);
2850
2851 nir_register *nir_local_reg_create(nir_function_impl *impl);
2852
2853 void nir_reg_remove(nir_register *reg);
2854
2855 /** Adds a variable to the appropriate list in nir_shader */
2856 void nir_shader_add_variable(nir_shader *shader, nir_variable *var);
2857
2858 static inline void
2859 nir_function_impl_add_variable(nir_function_impl *impl, nir_variable *var)
2860 {
2861 assert(var->data.mode == nir_var_function_temp);
2862 exec_list_push_tail(&impl->locals, &var->node);
2863 }
2864
2865 /** creates a variable, sets a few defaults, and adds it to the list */
2866 nir_variable *nir_variable_create(nir_shader *shader,
2867 nir_variable_mode mode,
2868 const struct glsl_type *type,
2869 const char *name);
2870 /** creates a local variable and adds it to the list */
2871 nir_variable *nir_local_variable_create(nir_function_impl *impl,
2872 const struct glsl_type *type,
2873 const char *name);
2874
2875 /** creates a function and adds it to the shader's list of functions */
2876 nir_function *nir_function_create(nir_shader *shader, const char *name);
2877
2878 nir_function_impl *nir_function_impl_create(nir_function *func);
2879 /** creates a function_impl that isn't tied to any particular function */
2880 nir_function_impl *nir_function_impl_create_bare(nir_shader *shader);
2881
2882 nir_block *nir_block_create(nir_shader *shader);
2883 nir_if *nir_if_create(nir_shader *shader);
2884 nir_loop *nir_loop_create(nir_shader *shader);
2885
2886 nir_function_impl *nir_cf_node_get_function(nir_cf_node *node);
2887
2888 /** requests that the given pieces of metadata be generated */
2889 void nir_metadata_require(nir_function_impl *impl, nir_metadata required, ...);
2890 /** dirties all but the preserved metadata */
2891 void nir_metadata_preserve(nir_function_impl *impl, nir_metadata preserved);
2892
2893 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
2894 nir_alu_instr *nir_alu_instr_create(nir_shader *shader, nir_op op);
2895
2896 nir_deref_instr *nir_deref_instr_create(nir_shader *shader,
2897 nir_deref_type deref_type);
2898
2899 nir_jump_instr *nir_jump_instr_create(nir_shader *shader, nir_jump_type type);
2900
2901 nir_load_const_instr *nir_load_const_instr_create(nir_shader *shader,
2902 unsigned num_components,
2903 unsigned bit_size);
2904
2905 nir_intrinsic_instr *nir_intrinsic_instr_create(nir_shader *shader,
2906 nir_intrinsic_op op);
2907
2908 nir_call_instr *nir_call_instr_create(nir_shader *shader,
2909 nir_function *callee);
2910
2911 nir_tex_instr *nir_tex_instr_create(nir_shader *shader, unsigned num_srcs);
2912
2913 nir_phi_instr *nir_phi_instr_create(nir_shader *shader);
2914
2915 nir_parallel_copy_instr *nir_parallel_copy_instr_create(nir_shader *shader);
2916
2917 nir_ssa_undef_instr *nir_ssa_undef_instr_create(nir_shader *shader,
2918 unsigned num_components,
2919 unsigned bit_size);
2920
2921 nir_const_value nir_alu_binop_identity(nir_op binop, unsigned bit_size);
2922
2923 /**
2924 * NIR Cursors and Instruction Insertion API
2925 * @{
2926 *
2927 * A tiny struct representing a point to insert/extract instructions or
2928 * control flow nodes. Helps reduce the combinatorial explosion of possible
2929 * points to insert/extract.
2930 *
2931 * \sa nir_control_flow.h
2932 */
2933 typedef enum {
2934 nir_cursor_before_block,
2935 nir_cursor_after_block,
2936 nir_cursor_before_instr,
2937 nir_cursor_after_instr,
2938 } nir_cursor_option;
2939
2940 typedef struct {
2941 nir_cursor_option option;
2942 union {
2943 nir_block *block;
2944 nir_instr *instr;
2945 };
2946 } nir_cursor;
2947
2948 static inline nir_block *
2949 nir_cursor_current_block(nir_cursor cursor)
2950 {
2951 if (cursor.option == nir_cursor_before_instr ||
2952 cursor.option == nir_cursor_after_instr) {
2953 return cursor.instr->block;
2954 } else {
2955 return cursor.block;
2956 }
2957 }
2958
2959 bool nir_cursors_equal(nir_cursor a, nir_cursor b);
2960
2961 static inline nir_cursor
2962 nir_before_block(nir_block *block)
2963 {
2964 nir_cursor cursor;
2965 cursor.option = nir_cursor_before_block;
2966 cursor.block = block;
2967 return cursor;
2968 }
2969
2970 static inline nir_cursor
2971 nir_after_block(nir_block *block)
2972 {
2973 nir_cursor cursor;
2974 cursor.option = nir_cursor_after_block;
2975 cursor.block = block;
2976 return cursor;
2977 }
2978
2979 static inline nir_cursor
2980 nir_before_instr(nir_instr *instr)
2981 {
2982 nir_cursor cursor;
2983 cursor.option = nir_cursor_before_instr;
2984 cursor.instr = instr;
2985 return cursor;
2986 }
2987
2988 static inline nir_cursor
2989 nir_after_instr(nir_instr *instr)
2990 {
2991 nir_cursor cursor;
2992 cursor.option = nir_cursor_after_instr;
2993 cursor.instr = instr;
2994 return cursor;
2995 }
2996
2997 static inline nir_cursor
2998 nir_after_block_before_jump(nir_block *block)
2999 {
3000 nir_instr *last_instr = nir_block_last_instr(block);
3001 if (last_instr && last_instr->type == nir_instr_type_jump) {
3002 return nir_before_instr(last_instr);
3003 } else {
3004 return nir_after_block(block);
3005 }
3006 }
3007
3008 static inline nir_cursor
3009 nir_before_src(nir_src *src, bool is_if_condition)
3010 {
3011 if (is_if_condition) {
3012 nir_block *prev_block =
3013 nir_cf_node_as_block(nir_cf_node_prev(&src->parent_if->cf_node));
3014 assert(!nir_block_ends_in_jump(prev_block));
3015 return nir_after_block(prev_block);
3016 } else if (src->parent_instr->type == nir_instr_type_phi) {
3017 #ifndef NDEBUG
3018 nir_phi_instr *cond_phi = nir_instr_as_phi(src->parent_instr);
3019 bool found = false;
3020 nir_foreach_phi_src(phi_src, cond_phi) {
3021 if (phi_src->src.ssa == src->ssa) {
3022 found = true;
3023 break;
3024 }
3025 }
3026 assert(found);
3027 #endif
3028 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
3029 * to have a more specific name.
3030 */
3031 nir_phi_src *phi_src = LIST_ENTRY(nir_phi_src, src, src);
3032 return nir_after_block_before_jump(phi_src->pred);
3033 } else {
3034 return nir_before_instr(src->parent_instr);
3035 }
3036 }
3037
3038 static inline nir_cursor
3039 nir_before_cf_node(nir_cf_node *node)
3040 {
3041 if (node->type == nir_cf_node_block)
3042 return nir_before_block(nir_cf_node_as_block(node));
3043
3044 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node)));
3045 }
3046
3047 static inline nir_cursor
3048 nir_after_cf_node(nir_cf_node *node)
3049 {
3050 if (node->type == nir_cf_node_block)
3051 return nir_after_block(nir_cf_node_as_block(node));
3052
3053 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node)));
3054 }
3055
3056 static inline nir_cursor
3057 nir_after_phis(nir_block *block)
3058 {
3059 nir_foreach_instr(instr, block) {
3060 if (instr->type != nir_instr_type_phi)
3061 return nir_before_instr(instr);
3062 }
3063 return nir_after_block(block);
3064 }
3065
3066 static inline nir_cursor
3067 nir_after_cf_node_and_phis(nir_cf_node *node)
3068 {
3069 if (node->type == nir_cf_node_block)
3070 return nir_after_block(nir_cf_node_as_block(node));
3071
3072 nir_block *block = nir_cf_node_as_block(nir_cf_node_next(node));
3073
3074 return nir_after_phis(block);
3075 }
3076
3077 static inline nir_cursor
3078 nir_before_cf_list(struct exec_list *cf_list)
3079 {
3080 nir_cf_node *first_node = exec_node_data(nir_cf_node,
3081 exec_list_get_head(cf_list), node);
3082 return nir_before_cf_node(first_node);
3083 }
3084
3085 static inline nir_cursor
3086 nir_after_cf_list(struct exec_list *cf_list)
3087 {
3088 nir_cf_node *last_node = exec_node_data(nir_cf_node,
3089 exec_list_get_tail(cf_list), node);
3090 return nir_after_cf_node(last_node);
3091 }
3092
3093 /**
3094 * Insert a NIR instruction at the given cursor.
3095 *
3096 * Note: This does not update the cursor.
3097 */
3098 void nir_instr_insert(nir_cursor cursor, nir_instr *instr);
3099
3100 static inline void
3101 nir_instr_insert_before(nir_instr *instr, nir_instr *before)
3102 {
3103 nir_instr_insert(nir_before_instr(instr), before);
3104 }
3105
3106 static inline void
3107 nir_instr_insert_after(nir_instr *instr, nir_instr *after)
3108 {
3109 nir_instr_insert(nir_after_instr(instr), after);
3110 }
3111
3112 static inline void
3113 nir_instr_insert_before_block(nir_block *block, nir_instr *before)
3114 {
3115 nir_instr_insert(nir_before_block(block), before);
3116 }
3117
3118 static inline void
3119 nir_instr_insert_after_block(nir_block *block, nir_instr *after)
3120 {
3121 nir_instr_insert(nir_after_block(block), after);
3122 }
3123
3124 static inline void
3125 nir_instr_insert_before_cf(nir_cf_node *node, nir_instr *before)
3126 {
3127 nir_instr_insert(nir_before_cf_node(node), before);
3128 }
3129
3130 static inline void
3131 nir_instr_insert_after_cf(nir_cf_node *node, nir_instr *after)
3132 {
3133 nir_instr_insert(nir_after_cf_node(node), after);
3134 }
3135
3136 static inline void
3137 nir_instr_insert_before_cf_list(struct exec_list *list, nir_instr *before)
3138 {
3139 nir_instr_insert(nir_before_cf_list(list), before);
3140 }
3141
3142 static inline void
3143 nir_instr_insert_after_cf_list(struct exec_list *list, nir_instr *after)
3144 {
3145 nir_instr_insert(nir_after_cf_list(list), after);
3146 }
3147
3148 void nir_instr_remove_v(nir_instr *instr);
3149
3150 static inline nir_cursor
3151 nir_instr_remove(nir_instr *instr)
3152 {
3153 nir_cursor cursor;
3154 nir_instr *prev = nir_instr_prev(instr);
3155 if (prev) {
3156 cursor = nir_after_instr(prev);
3157 } else {
3158 cursor = nir_before_block(instr->block);
3159 }
3160 nir_instr_remove_v(instr);
3161 return cursor;
3162 }
3163
3164 /** @} */
3165
3166 nir_ssa_def *nir_instr_ssa_def(nir_instr *instr);
3167
3168 typedef bool (*nir_foreach_ssa_def_cb)(nir_ssa_def *def, void *state);
3169 typedef bool (*nir_foreach_dest_cb)(nir_dest *dest, void *state);
3170 typedef bool (*nir_foreach_src_cb)(nir_src *src, void *state);
3171 bool nir_foreach_ssa_def(nir_instr *instr, nir_foreach_ssa_def_cb cb,
3172 void *state);
3173 bool nir_foreach_dest(nir_instr *instr, nir_foreach_dest_cb cb, void *state);
3174 bool nir_foreach_src(nir_instr *instr, nir_foreach_src_cb cb, void *state);
3175
3176 nir_const_value *nir_src_as_const_value(nir_src src);
3177
3178 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
3179 static inline c_type * \
3180 nir_src_as_ ## name (nir_src src) \
3181 { \
3182 return src.is_ssa && src.ssa->parent_instr->type == type_enum \
3183 ? cast_macro(src.ssa->parent_instr) : NULL; \
3184 }
3185
3186 NIR_SRC_AS_(alu_instr, nir_alu_instr, nir_instr_type_alu, nir_instr_as_alu)
3187 NIR_SRC_AS_(intrinsic, nir_intrinsic_instr,
3188 nir_instr_type_intrinsic, nir_instr_as_intrinsic)
3189 NIR_SRC_AS_(deref, nir_deref_instr, nir_instr_type_deref, nir_instr_as_deref)
3190
3191 bool nir_src_is_dynamically_uniform(nir_src src);
3192 bool nir_srcs_equal(nir_src src1, nir_src src2);
3193 bool nir_instrs_equal(const nir_instr *instr1, const nir_instr *instr2);
3194 void nir_instr_rewrite_src(nir_instr *instr, nir_src *src, nir_src new_src);
3195 void nir_instr_move_src(nir_instr *dest_instr, nir_src *dest, nir_src *src);
3196 void nir_if_rewrite_condition(nir_if *if_stmt, nir_src new_src);
3197 void nir_instr_rewrite_dest(nir_instr *instr, nir_dest *dest,
3198 nir_dest new_dest);
3199
3200 void nir_ssa_dest_init(nir_instr *instr, nir_dest *dest,
3201 unsigned num_components, unsigned bit_size,
3202 const char *name);
3203 void nir_ssa_def_init(nir_instr *instr, nir_ssa_def *def,
3204 unsigned num_components, unsigned bit_size,
3205 const char *name);
3206 static inline void
3207 nir_ssa_dest_init_for_type(nir_instr *instr, nir_dest *dest,
3208 const struct glsl_type *type,
3209 const char *name)
3210 {
3211 assert(glsl_type_is_vector_or_scalar(type));
3212 nir_ssa_dest_init(instr, dest, glsl_get_components(type),
3213 glsl_get_bit_size(type), name);
3214 }
3215 void nir_ssa_def_rewrite_uses(nir_ssa_def *def, nir_src new_src);
3216 void nir_ssa_def_rewrite_uses_after(nir_ssa_def *def, nir_src new_src,
3217 nir_instr *after_me);
3218
3219 nir_component_mask_t nir_ssa_def_components_read(const nir_ssa_def *def);
3220
3221 /*
3222 * finds the next basic block in source-code order, returns NULL if there is
3223 * none
3224 */
3225
3226 nir_block *nir_block_cf_tree_next(nir_block *block);
3227
3228 /* Performs the opposite of nir_block_cf_tree_next() */
3229
3230 nir_block *nir_block_cf_tree_prev(nir_block *block);
3231
3232 /* Gets the first block in a CF node in source-code order */
3233
3234 nir_block *nir_cf_node_cf_tree_first(nir_cf_node *node);
3235
3236 /* Gets the last block in a CF node in source-code order */
3237
3238 nir_block *nir_cf_node_cf_tree_last(nir_cf_node *node);
3239
3240 /* Gets the next block after a CF node in source-code order */
3241
3242 nir_block *nir_cf_node_cf_tree_next(nir_cf_node *node);
3243
3244 /* Macros for loops that visit blocks in source-code order */
3245
3246 #define nir_foreach_block(block, impl) \
3247 for (nir_block *block = nir_start_block(impl); block != NULL; \
3248 block = nir_block_cf_tree_next(block))
3249
3250 #define nir_foreach_block_safe(block, impl) \
3251 for (nir_block *block = nir_start_block(impl), \
3252 *next = nir_block_cf_tree_next(block); \
3253 block != NULL; \
3254 block = next, next = nir_block_cf_tree_next(block))
3255
3256 #define nir_foreach_block_reverse(block, impl) \
3257 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
3258 block = nir_block_cf_tree_prev(block))
3259
3260 #define nir_foreach_block_reverse_safe(block, impl) \
3261 for (nir_block *block = nir_impl_last_block(impl), \
3262 *prev = nir_block_cf_tree_prev(block); \
3263 block != NULL; \
3264 block = prev, prev = nir_block_cf_tree_prev(block))
3265
3266 #define nir_foreach_block_in_cf_node(block, node) \
3267 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
3268 block != nir_cf_node_cf_tree_next(node); \
3269 block = nir_block_cf_tree_next(block))
3270
3271 /* If the following CF node is an if, this function returns that if.
3272 * Otherwise, it returns NULL.
3273 */
3274 nir_if *nir_block_get_following_if(nir_block *block);
3275
3276 nir_loop *nir_block_get_following_loop(nir_block *block);
3277
3278 void nir_index_local_regs(nir_function_impl *impl);
3279 void nir_index_ssa_defs(nir_function_impl *impl);
3280 unsigned nir_index_instrs(nir_function_impl *impl);
3281
3282 void nir_index_blocks(nir_function_impl *impl);
3283
3284 void nir_print_shader(nir_shader *shader, FILE *fp);
3285 void nir_print_shader_annotated(nir_shader *shader, FILE *fp, struct hash_table *errors);
3286 void nir_print_instr(const nir_instr *instr, FILE *fp);
3287 void nir_print_deref(const nir_deref_instr *deref, FILE *fp);
3288
3289 /** Shallow clone of a single ALU instruction. */
3290 nir_alu_instr *nir_alu_instr_clone(nir_shader *s, const nir_alu_instr *orig);
3291
3292 nir_shader *nir_shader_clone(void *mem_ctx, const nir_shader *s);
3293 nir_function_impl *nir_function_impl_clone(nir_shader *shader,
3294 const nir_function_impl *fi);
3295 nir_constant *nir_constant_clone(const nir_constant *c, nir_variable *var);
3296 nir_variable *nir_variable_clone(const nir_variable *c, nir_shader *shader);
3297
3298 void nir_shader_replace(nir_shader *dest, nir_shader *src);
3299
3300 void nir_shader_serialize_deserialize(nir_shader *s);
3301
3302 #ifndef NDEBUG
3303 void nir_validate_shader(nir_shader *shader, const char *when);
3304 void nir_metadata_set_validation_flag(nir_shader *shader);
3305 void nir_metadata_check_validation_flag(nir_shader *shader);
3306
3307 static inline bool
3308 should_skip_nir(const char *name)
3309 {
3310 static const char *list = NULL;
3311 if (!list) {
3312 /* Comma separated list of names to skip. */
3313 list = getenv("NIR_SKIP");
3314 if (!list)
3315 list = "";
3316 }
3317
3318 if (!list[0])
3319 return false;
3320
3321 return comma_separated_list_contains(list, name);
3322 }
3323
3324 static inline bool
3325 should_clone_nir(void)
3326 {
3327 static int should_clone = -1;
3328 if (should_clone < 0)
3329 should_clone = env_var_as_boolean("NIR_TEST_CLONE", false);
3330
3331 return should_clone;
3332 }
3333
3334 static inline bool
3335 should_serialize_deserialize_nir(void)
3336 {
3337 static int test_serialize = -1;
3338 if (test_serialize < 0)
3339 test_serialize = env_var_as_boolean("NIR_TEST_SERIALIZE", false);
3340
3341 return test_serialize;
3342 }
3343
3344 static inline bool
3345 should_print_nir(void)
3346 {
3347 static int should_print = -1;
3348 if (should_print < 0)
3349 should_print = env_var_as_boolean("NIR_PRINT", false);
3350
3351 return should_print;
3352 }
3353 #else
3354 static inline void nir_validate_shader(nir_shader *shader, const char *when) { (void) shader; (void)when; }
3355 static inline void nir_metadata_set_validation_flag(nir_shader *shader) { (void) shader; }
3356 static inline void nir_metadata_check_validation_flag(nir_shader *shader) { (void) shader; }
3357 static inline bool should_skip_nir(UNUSED const char *pass_name) { return false; }
3358 static inline bool should_clone_nir(void) { return false; }
3359 static inline bool should_serialize_deserialize_nir(void) { return false; }
3360 static inline bool should_print_nir(void) { return false; }
3361 #endif /* NDEBUG */
3362
3363 #define _PASS(pass, nir, do_pass) do { \
3364 if (should_skip_nir(#pass)) { \
3365 printf("skipping %s\n", #pass); \
3366 break; \
3367 } \
3368 do_pass \
3369 nir_validate_shader(nir, "after " #pass); \
3370 if (should_clone_nir()) { \
3371 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
3372 nir_shader_replace(nir, clone); \
3373 } \
3374 if (should_serialize_deserialize_nir()) { \
3375 nir_shader_serialize_deserialize(nir); \
3376 } \
3377 } while (0)
3378
3379 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
3380 nir_metadata_set_validation_flag(nir); \
3381 if (should_print_nir()) \
3382 printf("%s\n", #pass); \
3383 if (pass(nir, ##__VA_ARGS__)) { \
3384 progress = true; \
3385 if (should_print_nir()) \
3386 nir_print_shader(nir, stdout); \
3387 nir_metadata_check_validation_flag(nir); \
3388 } \
3389 )
3390
3391 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
3392 if (should_print_nir()) \
3393 printf("%s\n", #pass); \
3394 pass(nir, ##__VA_ARGS__); \
3395 if (should_print_nir()) \
3396 nir_print_shader(nir, stdout); \
3397 )
3398
3399 #define NIR_SKIP(name) should_skip_nir(#name)
3400
3401 /** An instruction filtering callback
3402 *
3403 * Returns true if the instruction should be processed and false otherwise.
3404 */
3405 typedef bool (*nir_instr_filter_cb)(const nir_instr *, const void *);
3406
3407 /** A simple instruction lowering callback
3408 *
3409 * Many instruction lowering passes can be written as a simple function which
3410 * takes an instruction as its input and returns a sequence of instructions
3411 * that implement the consumed instruction. This function type represents
3412 * such a lowering function. When called, a function with this prototype
3413 * should either return NULL indicating that no lowering needs to be done or
3414 * emit a sequence of instructions using the provided builder (whose cursor
3415 * will already be placed after the instruction to be lowered) and return the
3416 * resulting nir_ssa_def.
3417 */
3418 typedef nir_ssa_def *(*nir_lower_instr_cb)(struct nir_builder *,
3419 nir_instr *, void *);
3420
3421 /**
3422 * Special return value for nir_lower_instr_cb when some progress occurred
3423 * (like changing an input to the instr) that didn't result in a replacement
3424 * SSA def being generated.
3425 */
3426 #define NIR_LOWER_INSTR_PROGRESS ((nir_ssa_def *)(uintptr_t)1)
3427
3428 /** Iterate over all the instructions in a nir_function_impl and lower them
3429 * using the provided callbacks
3430 *
3431 * This function implements the guts of a standard lowering pass for you. It
3432 * iterates over all of the instructions in a nir_function_impl and calls the
3433 * filter callback on each one. If the filter callback returns true, it then
3434 * calls the lowering call back on the instruction. (Splitting it this way
3435 * allows us to avoid some save/restore work for instructions we know won't be
3436 * lowered.) If the instruction is dead after the lowering is complete, it
3437 * will be removed. If new instructions are added, the lowering callback will
3438 * also be called on them in case multiple lowerings are required.
3439 *
3440 * The metadata for the nir_function_impl will also be updated. If any blocks
3441 * are added (they cannot be removed), dominance and block indices will be
3442 * invalidated.
3443 */
3444 bool nir_function_impl_lower_instructions(nir_function_impl *impl,
3445 nir_instr_filter_cb filter,
3446 nir_lower_instr_cb lower,
3447 void *cb_data);
3448 bool nir_shader_lower_instructions(nir_shader *shader,
3449 nir_instr_filter_cb filter,
3450 nir_lower_instr_cb lower,
3451 void *cb_data);
3452
3453 void nir_calc_dominance_impl(nir_function_impl *impl);
3454 void nir_calc_dominance(nir_shader *shader);
3455
3456 nir_block *nir_dominance_lca(nir_block *b1, nir_block *b2);
3457 bool nir_block_dominates(nir_block *parent, nir_block *child);
3458 bool nir_block_is_unreachable(nir_block *block);
3459
3460 void nir_dump_dom_tree_impl(nir_function_impl *impl, FILE *fp);
3461 void nir_dump_dom_tree(nir_shader *shader, FILE *fp);
3462
3463 void nir_dump_dom_frontier_impl(nir_function_impl *impl, FILE *fp);
3464 void nir_dump_dom_frontier(nir_shader *shader, FILE *fp);
3465
3466 void nir_dump_cfg_impl(nir_function_impl *impl, FILE *fp);
3467 void nir_dump_cfg(nir_shader *shader, FILE *fp);
3468
3469 int nir_gs_count_vertices(const nir_shader *shader);
3470
3471 bool nir_shrink_vec_array_vars(nir_shader *shader, nir_variable_mode modes);
3472 bool nir_split_array_vars(nir_shader *shader, nir_variable_mode modes);
3473 bool nir_split_var_copies(nir_shader *shader);
3474 bool nir_split_per_member_structs(nir_shader *shader);
3475 bool nir_split_struct_vars(nir_shader *shader, nir_variable_mode modes);
3476
3477 bool nir_lower_returns_impl(nir_function_impl *impl);
3478 bool nir_lower_returns(nir_shader *shader);
3479
3480 void nir_inline_function_impl(struct nir_builder *b,
3481 const nir_function_impl *impl,
3482 nir_ssa_def **params);
3483 bool nir_inline_functions(nir_shader *shader);
3484
3485 bool nir_propagate_invariant(nir_shader *shader);
3486
3487 void nir_lower_var_copy_instr(nir_intrinsic_instr *copy, nir_shader *shader);
3488 void nir_lower_deref_copy_instr(struct nir_builder *b,
3489 nir_intrinsic_instr *copy);
3490 bool nir_lower_var_copies(nir_shader *shader);
3491
3492 void nir_fixup_deref_modes(nir_shader *shader);
3493
3494 bool nir_lower_global_vars_to_local(nir_shader *shader);
3495
3496 typedef enum {
3497 nir_lower_direct_array_deref_of_vec_load = (1 << 0),
3498 nir_lower_indirect_array_deref_of_vec_load = (1 << 1),
3499 nir_lower_direct_array_deref_of_vec_store = (1 << 2),
3500 nir_lower_indirect_array_deref_of_vec_store = (1 << 3),
3501 } nir_lower_array_deref_of_vec_options;
3502
3503 bool nir_lower_array_deref_of_vec(nir_shader *shader, nir_variable_mode modes,
3504 nir_lower_array_deref_of_vec_options options);
3505
3506 bool nir_lower_indirect_derefs(nir_shader *shader, nir_variable_mode modes);
3507
3508 bool nir_lower_locals_to_regs(nir_shader *shader);
3509
3510 void nir_lower_io_to_temporaries(nir_shader *shader,
3511 nir_function_impl *entrypoint,
3512 bool outputs, bool inputs);
3513
3514 bool nir_lower_vars_to_scratch(nir_shader *shader,
3515 nir_variable_mode modes,
3516 int size_threshold,
3517 glsl_type_size_align_func size_align);
3518
3519 void nir_shader_gather_info(nir_shader *shader, nir_function_impl *entrypoint);
3520
3521 void nir_gather_ssa_types(nir_function_impl *impl,
3522 BITSET_WORD *float_types,
3523 BITSET_WORD *int_types);
3524
3525 void nir_assign_var_locations(struct exec_list *var_list, unsigned *size,
3526 int (*type_size)(const struct glsl_type *, bool));
3527
3528 /* Some helpers to do very simple linking */
3529 bool nir_remove_unused_varyings(nir_shader *producer, nir_shader *consumer);
3530 bool nir_remove_unused_io_vars(nir_shader *shader, struct exec_list *var_list,
3531 uint64_t *used_by_other_stage,
3532 uint64_t *used_by_other_stage_patches);
3533 void nir_compact_varyings(nir_shader *producer, nir_shader *consumer,
3534 bool default_to_smooth_interp);
3535 void nir_link_xfb_varyings(nir_shader *producer, nir_shader *consumer);
3536 bool nir_link_opt_varyings(nir_shader *producer, nir_shader *consumer);
3537
3538
3539 void nir_assign_io_var_locations(struct exec_list *var_list,
3540 unsigned *size,
3541 gl_shader_stage stage);
3542
3543 typedef enum {
3544 /* If set, this causes all 64-bit IO operations to be lowered on-the-fly
3545 * to 32-bit operations. This is only valid for nir_var_shader_in/out
3546 * modes.
3547 */
3548 nir_lower_io_lower_64bit_to_32 = (1 << 0),
3549
3550 /* If set, this forces all non-flat fragment shader inputs to be
3551 * interpolated as if with the "sample" qualifier. This requires
3552 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
3553 */
3554 nir_lower_io_force_sample_interpolation = (1 << 1),
3555 } nir_lower_io_options;
3556 bool nir_lower_io(nir_shader *shader,
3557 nir_variable_mode modes,
3558 int (*type_size)(const struct glsl_type *, bool),