nir/glsl: Add another way of doing lower_imul64 for gen8+
[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 typedef uint8_t nir_component_mask_t;
63
64 /** Defines a cast function
65 *
66 * This macro defines a cast function from in_type to out_type where
67 * out_type is some structure type that contains a field of type out_type.
68 *
69 * Note that you have to be a bit careful as the generated cast function
70 * destroys constness.
71 */
72 #define NIR_DEFINE_CAST(name, in_type, out_type, field, \
73 type_field, type_value) \
74 static inline out_type * \
75 name(const in_type *parent) \
76 { \
77 assert(parent && parent->type_field == type_value); \
78 return exec_node_data(out_type, parent, field); \
79 }
80
81 struct nir_function;
82 struct nir_shader;
83 struct nir_instr;
84 struct nir_builder;
85
86
87 /**
88 * Description of built-in state associated with a uniform
89 *
90 * \sa nir_variable::state_slots
91 */
92 typedef struct {
93 gl_state_index16 tokens[STATE_LENGTH];
94 int swizzle;
95 } nir_state_slot;
96
97 typedef enum {
98 nir_var_shader_in = (1 << 0),
99 nir_var_shader_out = (1 << 1),
100 nir_var_shader_temp = (1 << 2),
101 nir_var_function_temp = (1 << 3),
102 nir_var_uniform = (1 << 4),
103 nir_var_mem_ubo = (1 << 5),
104 nir_var_system_value = (1 << 6),
105 nir_var_mem_ssbo = (1 << 7),
106 nir_var_mem_shared = (1 << 8),
107 nir_var_mem_global = (1 << 9),
108 nir_var_all = ~0,
109 } nir_variable_mode;
110
111 /**
112 * Rounding modes.
113 */
114 typedef enum {
115 nir_rounding_mode_undef = 0,
116 nir_rounding_mode_rtne = 1, /* round to nearest even */
117 nir_rounding_mode_ru = 2, /* round up */
118 nir_rounding_mode_rd = 3, /* round down */
119 nir_rounding_mode_rtz = 4, /* round towards zero */
120 } nir_rounding_mode;
121
122 typedef union {
123 bool b[NIR_MAX_VEC_COMPONENTS];
124 float f32[NIR_MAX_VEC_COMPONENTS];
125 double f64[NIR_MAX_VEC_COMPONENTS];
126 int8_t i8[NIR_MAX_VEC_COMPONENTS];
127 uint8_t u8[NIR_MAX_VEC_COMPONENTS];
128 int16_t i16[NIR_MAX_VEC_COMPONENTS];
129 uint16_t u16[NIR_MAX_VEC_COMPONENTS];
130 int32_t i32[NIR_MAX_VEC_COMPONENTS];
131 uint32_t u32[NIR_MAX_VEC_COMPONENTS];
132 int64_t i64[NIR_MAX_VEC_COMPONENTS];
133 uint64_t u64[NIR_MAX_VEC_COMPONENTS];
134 } nir_const_value;
135
136 typedef struct nir_constant {
137 /**
138 * Value of the constant.
139 *
140 * The field used to back the values supplied by the constant is determined
141 * by the type associated with the \c nir_variable. Constants may be
142 * scalars, vectors, or matrices.
143 */
144 nir_const_value values[NIR_MAX_VEC_COMPONENTS];
145
146 /* we could get this from the var->type but makes clone *much* easier to
147 * not have to care about the type.
148 */
149 unsigned num_elements;
150
151 /* Array elements / Structure Fields */
152 struct nir_constant **elements;
153 } nir_constant;
154
155 /**
156 * \brief Layout qualifiers for gl_FragDepth.
157 *
158 * The AMD/ARB_conservative_depth extensions allow gl_FragDepth to be redeclared
159 * with a layout qualifier.
160 */
161 typedef enum {
162 nir_depth_layout_none, /**< No depth layout is specified. */
163 nir_depth_layout_any,
164 nir_depth_layout_greater,
165 nir_depth_layout_less,
166 nir_depth_layout_unchanged
167 } nir_depth_layout;
168
169 /**
170 * Enum keeping track of how a variable was declared.
171 */
172 typedef enum {
173 /**
174 * Normal declaration.
175 */
176 nir_var_declared_normally = 0,
177
178 /**
179 * Variable is implicitly generated by the compiler and should not be
180 * visible via the API.
181 */
182 nir_var_hidden,
183 } nir_var_declaration_type;
184
185 /**
186 * Either a uniform, global variable, shader input, or shader output. Based on
187 * ir_variable - it should be easy to translate between the two.
188 */
189
190 typedef struct nir_variable {
191 struct exec_node node;
192
193 /**
194 * Declared type of the variable
195 */
196 const struct glsl_type *type;
197
198 /**
199 * Declared name of the variable
200 */
201 char *name;
202
203 struct nir_variable_data {
204 /**
205 * Storage class of the variable.
206 *
207 * \sa nir_variable_mode
208 */
209 nir_variable_mode mode;
210
211 /**
212 * Is the variable read-only?
213 *
214 * This is set for variables declared as \c const, shader inputs,
215 * and uniforms.
216 */
217 unsigned read_only:1;
218 unsigned centroid:1;
219 unsigned sample:1;
220 unsigned patch:1;
221 unsigned invariant:1;
222
223 /**
224 * When separate shader programs are enabled, only input/outputs between
225 * the stages of a multi-stage separate program can be safely removed
226 * from the shader interface. Other input/outputs must remains active.
227 *
228 * This is also used to make sure xfb varyings that are unused by the
229 * fragment shader are not removed.
230 */
231 unsigned always_active_io:1;
232
233 /**
234 * Interpolation mode for shader inputs / outputs
235 *
236 * \sa glsl_interp_mode
237 */
238 unsigned interpolation:2;
239
240 /**
241 * If non-zero, then this variable may be packed along with other variables
242 * into a single varying slot, so this offset should be applied when
243 * accessing components. For example, an offset of 1 means that the x
244 * component of this variable is actually stored in component y of the
245 * location specified by \c location.
246 */
247 unsigned location_frac:2;
248
249 /**
250 * If true, this variable represents an array of scalars that should
251 * be tightly packed. In other words, consecutive array elements
252 * should be stored one component apart, rather than one slot apart.
253 */
254 unsigned compact:1;
255
256 /**
257 * Whether this is a fragment shader output implicitly initialized with
258 * the previous contents of the specified render target at the
259 * framebuffer location corresponding to this shader invocation.
260 */
261 unsigned fb_fetch_output:1;
262
263 /**
264 * Non-zero if this variable is considered bindless as defined by
265 * ARB_bindless_texture.
266 */
267 unsigned bindless:1;
268
269 /**
270 * Was an explicit binding set in the shader?
271 */
272 unsigned explicit_binding:1;
273
274 /**
275 * Was a transfer feedback buffer set in the shader?
276 */
277 unsigned explicit_xfb_buffer:1;
278
279 /**
280 * Was a transfer feedback stride set in the shader?
281 */
282 unsigned explicit_xfb_stride:1;
283
284 /**
285 * Was an explicit offset set in the shader?
286 */
287 unsigned explicit_offset:1;
288
289 /**
290 * \brief Layout qualifier for gl_FragDepth.
291 *
292 * This is not equal to \c ir_depth_layout_none if and only if this
293 * variable is \c gl_FragDepth and a layout qualifier is specified.
294 */
295 nir_depth_layout depth_layout;
296
297 /**
298 * Storage location of the base of this variable
299 *
300 * The precise meaning of this field depends on the nature of the variable.
301 *
302 * - Vertex shader input: one of the values from \c gl_vert_attrib.
303 * - Vertex shader output: one of the values from \c gl_varying_slot.
304 * - Geometry shader input: one of the values from \c gl_varying_slot.
305 * - Geometry shader output: one of the values from \c gl_varying_slot.
306 * - Fragment shader input: one of the values from \c gl_varying_slot.
307 * - Fragment shader output: one of the values from \c gl_frag_result.
308 * - Uniforms: Per-stage uniform slot number for default uniform block.
309 * - Uniforms: Index within the uniform block definition for UBO members.
310 * - Non-UBO Uniforms: uniform slot number.
311 * - Other: This field is not currently used.
312 *
313 * If the variable is a uniform, shader input, or shader output, and the
314 * slot has not been assigned, the value will be -1.
315 */
316 int location;
317
318 /**
319 * The actual location of the variable in the IR. Only valid for inputs
320 * and outputs.
321 */
322 unsigned int driver_location;
323
324 /**
325 * Vertex stream output identifier.
326 *
327 * For packed outputs, bit 31 is set and bits [2*i+1,2*i] indicate the
328 * stream of the i-th component.
329 */
330 unsigned stream;
331
332 /**
333 * output index for dual source blending.
334 */
335 int index;
336
337 /**
338 * Descriptor set binding for sampler or UBO.
339 */
340 int descriptor_set;
341
342 /**
343 * Initial binding point for a sampler or UBO.
344 *
345 * For array types, this represents the binding point for the first element.
346 */
347 int binding;
348
349 /**
350 * Location an atomic counter or transform feedback is stored at.
351 */
352 unsigned offset;
353
354 /**
355 * Transform feedback buffer.
356 */
357 unsigned xfb_buffer;
358
359 /**
360 * Transform feedback stride.
361 */
362 unsigned xfb_stride;
363
364 /**
365 * How the variable was declared. See nir_var_declaration_type.
366 *
367 * This is used to detect variables generated by the compiler, so should
368 * not be visible via the API.
369 */
370 unsigned how_declared:2;
371
372 /**
373 * ARB_shader_image_load_store qualifiers.
374 */
375 struct {
376 enum gl_access_qualifier access;
377
378 /** Image internal format if specified explicitly, otherwise GL_NONE. */
379 GLenum format;
380 } image;
381 } data;
382
383 /**
384 * Built-in state that backs this uniform
385 *
386 * Once set at variable creation, \c state_slots must remain invariant.
387 * This is because, ideally, this array would be shared by all clones of
388 * this variable in the IR tree. In other words, we'd really like for it
389 * to be a fly-weight.
390 *
391 * If the variable is not a uniform, \c num_state_slots will be zero and
392 * \c state_slots will be \c NULL.
393 */
394 /*@{*/
395 unsigned num_state_slots; /**< Number of state slots used */
396 nir_state_slot *state_slots; /**< State descriptors. */
397 /*@}*/
398
399 /**
400 * Constant expression assigned in the initializer of the variable
401 *
402 * This field should only be used temporarily by creators of NIR shaders
403 * and then lower_constant_initializers can be used to get rid of them.
404 * Most of the rest of NIR ignores this field or asserts that it's NULL.
405 */
406 nir_constant *constant_initializer;
407
408 /**
409 * For variables that are in an interface block or are an instance of an
410 * interface block, this is the \c GLSL_TYPE_INTERFACE type for that block.
411 *
412 * \sa ir_variable::location
413 */
414 const struct glsl_type *interface_type;
415
416 /**
417 * Description of per-member data for per-member struct variables
418 *
419 * This is used for variables which are actually an amalgamation of
420 * multiple entities such as a struct of built-in values or a struct of
421 * inputs each with their own layout specifier. This is only allowed on
422 * variables with a struct or array of array of struct type.
423 */
424 unsigned num_members;
425 struct nir_variable_data *members;
426 } nir_variable;
427
428 #define nir_foreach_variable(var, var_list) \
429 foreach_list_typed(nir_variable, var, node, var_list)
430
431 #define nir_foreach_variable_safe(var, var_list) \
432 foreach_list_typed_safe(nir_variable, var, node, var_list)
433
434 static inline bool
435 nir_variable_is_global(const nir_variable *var)
436 {
437 return var->data.mode != nir_var_function_temp;
438 }
439
440 typedef struct nir_register {
441 struct exec_node node;
442
443 unsigned num_components; /** < number of vector components */
444 unsigned num_array_elems; /** < size of array (0 for no array) */
445
446 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
447 uint8_t bit_size;
448
449 /** generic register index. */
450 unsigned index;
451
452 /** only for debug purposes, can be NULL */
453 const char *name;
454
455 /** whether this register is local (per-function) or global (per-shader) */
456 bool is_global;
457
458 /**
459 * If this flag is set to true, then accessing channels >= num_components
460 * is well-defined, and simply spills over to the next array element. This
461 * is useful for backends that can do per-component accessing, in
462 * particular scalar backends. By setting this flag and making
463 * num_components equal to 1, structures can be packed tightly into
464 * registers and then registers can be accessed per-component to get to
465 * each structure member, even if it crosses vec4 boundaries.
466 */
467 bool is_packed;
468
469 /** set of nir_srcs where this register is used (read from) */
470 struct list_head uses;
471
472 /** set of nir_dests where this register is defined (written to) */
473 struct list_head defs;
474
475 /** set of nir_ifs where this register is used as a condition */
476 struct list_head if_uses;
477 } nir_register;
478
479 #define nir_foreach_register(reg, reg_list) \
480 foreach_list_typed(nir_register, reg, node, reg_list)
481 #define nir_foreach_register_safe(reg, reg_list) \
482 foreach_list_typed_safe(nir_register, reg, node, reg_list)
483
484 typedef enum PACKED {
485 nir_instr_type_alu,
486 nir_instr_type_deref,
487 nir_instr_type_call,
488 nir_instr_type_tex,
489 nir_instr_type_intrinsic,
490 nir_instr_type_load_const,
491 nir_instr_type_jump,
492 nir_instr_type_ssa_undef,
493 nir_instr_type_phi,
494 nir_instr_type_parallel_copy,
495 } nir_instr_type;
496
497 typedef struct nir_instr {
498 struct exec_node node;
499 struct nir_block *block;
500 nir_instr_type type;
501
502 /* A temporary for optimization and analysis passes to use for storing
503 * flags. For instance, DCE uses this to store the "dead/live" info.
504 */
505 uint8_t pass_flags;
506
507 /** generic instruction index. */
508 unsigned index;
509 } nir_instr;
510
511 static inline nir_instr *
512 nir_instr_next(nir_instr *instr)
513 {
514 struct exec_node *next = exec_node_get_next(&instr->node);
515 if (exec_node_is_tail_sentinel(next))
516 return NULL;
517 else
518 return exec_node_data(nir_instr, next, node);
519 }
520
521 static inline nir_instr *
522 nir_instr_prev(nir_instr *instr)
523 {
524 struct exec_node *prev = exec_node_get_prev(&instr->node);
525 if (exec_node_is_head_sentinel(prev))
526 return NULL;
527 else
528 return exec_node_data(nir_instr, prev, node);
529 }
530
531 static inline bool
532 nir_instr_is_first(const nir_instr *instr)
533 {
534 return exec_node_is_head_sentinel(exec_node_get_prev_const(&instr->node));
535 }
536
537 static inline bool
538 nir_instr_is_last(const nir_instr *instr)
539 {
540 return exec_node_is_tail_sentinel(exec_node_get_next_const(&instr->node));
541 }
542
543 typedef struct nir_ssa_def {
544 /** for debugging only, can be NULL */
545 const char* name;
546
547 /** generic SSA definition index. */
548 unsigned index;
549
550 /** Index into the live_in and live_out bitfields */
551 unsigned live_index;
552
553 /** Instruction which produces this SSA value. */
554 nir_instr *parent_instr;
555
556 /** set of nir_instrs where this register is used (read from) */
557 struct list_head uses;
558
559 /** set of nir_ifs where this register is used as a condition */
560 struct list_head if_uses;
561
562 uint8_t num_components;
563
564 /* The bit-size of each channel; must be one of 8, 16, 32, or 64 */
565 uint8_t bit_size;
566 } nir_ssa_def;
567
568 struct nir_src;
569
570 typedef struct {
571 nir_register *reg;
572 struct nir_src *indirect; /** < NULL for no indirect offset */
573 unsigned base_offset;
574
575 /* TODO use-def chain goes here */
576 } nir_reg_src;
577
578 typedef struct {
579 nir_instr *parent_instr;
580 struct list_head def_link;
581
582 nir_register *reg;
583 struct nir_src *indirect; /** < NULL for no indirect offset */
584 unsigned base_offset;
585
586 /* TODO def-use chain goes here */
587 } nir_reg_dest;
588
589 struct nir_if;
590
591 typedef struct nir_src {
592 union {
593 /** Instruction that consumes this value as a source. */
594 nir_instr *parent_instr;
595 struct nir_if *parent_if;
596 };
597
598 struct list_head use_link;
599
600 union {
601 nir_reg_src reg;
602 nir_ssa_def *ssa;
603 };
604
605 bool is_ssa;
606 } nir_src;
607
608 static inline nir_src
609 nir_src_init(void)
610 {
611 nir_src src = { { NULL } };
612 return src;
613 }
614
615 #define NIR_SRC_INIT nir_src_init()
616
617 #define nir_foreach_use(src, reg_or_ssa_def) \
618 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
619
620 #define nir_foreach_use_safe(src, reg_or_ssa_def) \
621 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->uses, use_link)
622
623 #define nir_foreach_if_use(src, reg_or_ssa_def) \
624 list_for_each_entry(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
625
626 #define nir_foreach_if_use_safe(src, reg_or_ssa_def) \
627 list_for_each_entry_safe(nir_src, src, &(reg_or_ssa_def)->if_uses, use_link)
628
629 typedef struct {
630 union {
631 nir_reg_dest reg;
632 nir_ssa_def ssa;
633 };
634
635 bool is_ssa;
636 } nir_dest;
637
638 static inline nir_dest
639 nir_dest_init(void)
640 {
641 nir_dest dest = { { { NULL } } };
642 return dest;
643 }
644
645 #define NIR_DEST_INIT nir_dest_init()
646
647 #define nir_foreach_def(dest, reg) \
648 list_for_each_entry(nir_dest, dest, &(reg)->defs, reg.def_link)
649
650 #define nir_foreach_def_safe(dest, reg) \
651 list_for_each_entry_safe(nir_dest, dest, &(reg)->defs, reg.def_link)
652
653 static inline nir_src
654 nir_src_for_ssa(nir_ssa_def *def)
655 {
656 nir_src src = NIR_SRC_INIT;
657
658 src.is_ssa = true;
659 src.ssa = def;
660
661 return src;
662 }
663
664 static inline nir_src
665 nir_src_for_reg(nir_register *reg)
666 {
667 nir_src src = NIR_SRC_INIT;
668
669 src.is_ssa = false;
670 src.reg.reg = reg;
671 src.reg.indirect = NULL;
672 src.reg.base_offset = 0;
673
674 return src;
675 }
676
677 static inline nir_dest
678 nir_dest_for_reg(nir_register *reg)
679 {
680 nir_dest dest = NIR_DEST_INIT;
681
682 dest.reg.reg = reg;
683
684 return dest;
685 }
686
687 static inline unsigned
688 nir_src_bit_size(nir_src src)
689 {
690 return src.is_ssa ? src.ssa->bit_size : src.reg.reg->bit_size;
691 }
692
693 static inline unsigned
694 nir_src_num_components(nir_src src)
695 {
696 return src.is_ssa ? src.ssa->num_components : src.reg.reg->num_components;
697 }
698
699 static inline bool
700 nir_src_is_const(nir_src src)
701 {
702 return src.is_ssa &&
703 src.ssa->parent_instr->type == nir_instr_type_load_const;
704 }
705
706 int64_t nir_src_as_int(nir_src src);
707 uint64_t nir_src_as_uint(nir_src src);
708 bool nir_src_as_bool(nir_src src);
709 double nir_src_as_float(nir_src src);
710 int64_t nir_src_comp_as_int(nir_src src, unsigned component);
711 uint64_t nir_src_comp_as_uint(nir_src src, unsigned component);
712 bool nir_src_comp_as_bool(nir_src src, unsigned component);
713 double nir_src_comp_as_float(nir_src src, unsigned component);
714
715 static inline unsigned
716 nir_dest_bit_size(nir_dest dest)
717 {
718 return dest.is_ssa ? dest.ssa.bit_size : dest.reg.reg->bit_size;
719 }
720
721 static inline unsigned
722 nir_dest_num_components(nir_dest dest)
723 {
724 return dest.is_ssa ? dest.ssa.num_components : dest.reg.reg->num_components;
725 }
726
727 void nir_src_copy(nir_src *dest, const nir_src *src, void *instr_or_if);
728 void nir_dest_copy(nir_dest *dest, const nir_dest *src, nir_instr *instr);
729
730 typedef struct {
731 nir_src src;
732
733 /**
734 * \name input modifiers
735 */
736 /*@{*/
737 /**
738 * For inputs interpreted as floating point, flips the sign bit. For
739 * inputs interpreted as integers, performs the two's complement negation.
740 */
741 bool negate;
742
743 /**
744 * Clears the sign bit for floating point values, and computes the integer
745 * absolute value for integers. Note that the negate modifier acts after
746 * the absolute value modifier, therefore if both are set then all inputs
747 * will become negative.
748 */
749 bool abs;
750 /*@}*/
751
752 /**
753 * For each input component, says which component of the register it is
754 * chosen from. Note that which elements of the swizzle are used and which
755 * are ignored are based on the write mask for most opcodes - for example,
756 * a statement like "foo.xzw = bar.zyx" would have a writemask of 1101b and
757 * a swizzle of {2, x, 1, 0} where x means "don't care."
758 */
759 uint8_t swizzle[NIR_MAX_VEC_COMPONENTS];
760 } nir_alu_src;
761
762 typedef struct {
763 nir_dest dest;
764
765 /**
766 * \name saturate output modifier
767 *
768 * Only valid for opcodes that output floating-point numbers. Clamps the
769 * output to between 0.0 and 1.0 inclusive.
770 */
771
772 bool saturate;
773
774 unsigned write_mask : NIR_MAX_VEC_COMPONENTS; /* ignored if dest.is_ssa is true */
775 } nir_alu_dest;
776
777 /** NIR sized and unsized types
778 *
779 * The values in this enum are carefully chosen so that the sized type is
780 * just the unsized type OR the number of bits.
781 */
782 typedef enum {
783 nir_type_invalid = 0, /* Not a valid type */
784 nir_type_int = 2,
785 nir_type_uint = 4,
786 nir_type_bool = 6,
787 nir_type_float = 128,
788 nir_type_bool1 = 1 | nir_type_bool,
789 nir_type_bool32 = 32 | nir_type_bool,
790 nir_type_int1 = 1 | nir_type_int,
791 nir_type_int8 = 8 | nir_type_int,
792 nir_type_int16 = 16 | nir_type_int,
793 nir_type_int32 = 32 | nir_type_int,
794 nir_type_int64 = 64 | nir_type_int,
795 nir_type_uint1 = 1 | nir_type_uint,
796 nir_type_uint8 = 8 | nir_type_uint,
797 nir_type_uint16 = 16 | nir_type_uint,
798 nir_type_uint32 = 32 | nir_type_uint,
799 nir_type_uint64 = 64 | nir_type_uint,
800 nir_type_float16 = 16 | nir_type_float,
801 nir_type_float32 = 32 | nir_type_float,
802 nir_type_float64 = 64 | nir_type_float,
803 } nir_alu_type;
804
805 #define NIR_ALU_TYPE_SIZE_MASK 0x79
806 #define NIR_ALU_TYPE_BASE_TYPE_MASK 0x86
807
808 static inline unsigned
809 nir_alu_type_get_type_size(nir_alu_type type)
810 {
811 return type & NIR_ALU_TYPE_SIZE_MASK;
812 }
813
814 static inline unsigned
815 nir_alu_type_get_base_type(nir_alu_type type)
816 {
817 return type & NIR_ALU_TYPE_BASE_TYPE_MASK;
818 }
819
820 static inline nir_alu_type
821 nir_get_nir_type_for_glsl_base_type(enum glsl_base_type base_type)
822 {
823 switch (base_type) {
824 case GLSL_TYPE_BOOL:
825 return nir_type_bool1;
826 break;
827 case GLSL_TYPE_UINT:
828 return nir_type_uint32;
829 break;
830 case GLSL_TYPE_INT:
831 return nir_type_int32;
832 break;
833 case GLSL_TYPE_UINT16:
834 return nir_type_uint16;
835 break;
836 case GLSL_TYPE_INT16:
837 return nir_type_int16;
838 break;
839 case GLSL_TYPE_UINT8:
840 return nir_type_uint8;
841 case GLSL_TYPE_INT8:
842 return nir_type_int8;
843 case GLSL_TYPE_UINT64:
844 return nir_type_uint64;
845 break;
846 case GLSL_TYPE_INT64:
847 return nir_type_int64;
848 break;
849 case GLSL_TYPE_FLOAT:
850 return nir_type_float32;
851 break;
852 case GLSL_TYPE_FLOAT16:
853 return nir_type_float16;
854 break;
855 case GLSL_TYPE_DOUBLE:
856 return nir_type_float64;
857 break;
858 default:
859 unreachable("unknown type");
860 }
861 }
862
863 static inline nir_alu_type
864 nir_get_nir_type_for_glsl_type(const struct glsl_type *type)
865 {
866 return nir_get_nir_type_for_glsl_base_type(glsl_get_base_type(type));
867 }
868
869 nir_op nir_type_conversion_op(nir_alu_type src, nir_alu_type dst,
870 nir_rounding_mode rnd);
871
872 typedef enum {
873 NIR_OP_IS_COMMUTATIVE = (1 << 0),
874 NIR_OP_IS_ASSOCIATIVE = (1 << 1),
875 } nir_op_algebraic_property;
876
877 typedef struct {
878 const char *name;
879
880 unsigned num_inputs;
881
882 /**
883 * The number of components in the output
884 *
885 * If non-zero, this is the size of the output and input sizes are
886 * explicitly given; swizzle and writemask are still in effect, but if
887 * the output component is masked out, then the input component may
888 * still be in use.
889 *
890 * If zero, the opcode acts in the standard, per-component manner; the
891 * operation is performed on each component (except the ones that are
892 * masked out) with the input being taken from the input swizzle for
893 * that component.
894 *
895 * The size of some of the inputs may be given (i.e. non-zero) even
896 * though output_size is zero; in that case, the inputs with a zero
897 * size act per-component, while the inputs with non-zero size don't.
898 */
899 unsigned output_size;
900
901 /**
902 * The type of vector that the instruction outputs. Note that the
903 * staurate modifier is only allowed on outputs with the float type.
904 */
905
906 nir_alu_type output_type;
907
908 /**
909 * The number of components in each input
910 */
911 unsigned input_sizes[NIR_MAX_VEC_COMPONENTS];
912
913 /**
914 * The type of vector that each input takes. Note that negate and
915 * absolute value are only allowed on inputs with int or float type and
916 * behave differently on the two.
917 */
918 nir_alu_type input_types[NIR_MAX_VEC_COMPONENTS];
919
920 nir_op_algebraic_property algebraic_properties;
921 } nir_op_info;
922
923 extern const nir_op_info nir_op_infos[nir_num_opcodes];
924
925 typedef struct nir_alu_instr {
926 nir_instr instr;
927 nir_op op;
928
929 /** Indicates that this ALU instruction generates an exact value
930 *
931 * This is kind of a mixture of GLSL "precise" and "invariant" and not
932 * really equivalent to either. This indicates that the value generated by
933 * this operation is high-precision and any code transformations that touch
934 * it must ensure that the resulting value is bit-for-bit identical to the
935 * original.
936 */
937 bool exact;
938
939 nir_alu_dest dest;
940 nir_alu_src src[];
941 } nir_alu_instr;
942
943 void nir_alu_src_copy(nir_alu_src *dest, const nir_alu_src *src,
944 nir_alu_instr *instr);
945 void nir_alu_dest_copy(nir_alu_dest *dest, const nir_alu_dest *src,
946 nir_alu_instr *instr);
947
948 /* is this source channel used? */
949 static inline bool
950 nir_alu_instr_channel_used(const nir_alu_instr *instr, unsigned src,
951 unsigned channel)
952 {
953 if (nir_op_infos[instr->op].input_sizes[src] > 0)
954 return channel < nir_op_infos[instr->op].input_sizes[src];
955
956 return (instr->dest.write_mask >> channel) & 1;
957 }
958
959 static inline nir_component_mask_t
960 nir_alu_instr_src_read_mask(const nir_alu_instr *instr, unsigned src)
961 {
962 nir_component_mask_t read_mask = 0;
963 for (unsigned c = 0; c < NIR_MAX_VEC_COMPONENTS; c++) {
964 if (!nir_alu_instr_channel_used(instr, src, c))
965 continue;
966
967 read_mask |= (1 << instr->src[src].swizzle[c]);
968 }
969 return read_mask;
970 }
971
972 /*
973 * For instructions whose destinations are SSA, get the number of channels
974 * used for a source
975 */
976 static inline unsigned
977 nir_ssa_alu_instr_src_components(const nir_alu_instr *instr, unsigned src)
978 {
979 assert(instr->dest.dest.is_ssa);
980
981 if (nir_op_infos[instr->op].input_sizes[src] > 0)
982 return nir_op_infos[instr->op].input_sizes[src];
983
984 return instr->dest.dest.ssa.num_components;
985 }
986
987 bool nir_alu_srcs_equal(const nir_alu_instr *alu1, const nir_alu_instr *alu2,
988 unsigned src1, unsigned src2);
989
990 typedef enum {
991 nir_deref_type_var,
992 nir_deref_type_array,
993 nir_deref_type_array_wildcard,
994 nir_deref_type_ptr_as_array,
995 nir_deref_type_struct,
996 nir_deref_type_cast,
997 } nir_deref_type;
998
999 typedef struct {
1000 nir_instr instr;
1001
1002 /** The type of this deref instruction */
1003 nir_deref_type deref_type;
1004
1005 /** The mode of the underlying variable */
1006 nir_variable_mode mode;
1007
1008 /** The dereferenced type of the resulting pointer value */
1009 const struct glsl_type *type;
1010
1011 union {
1012 /** Variable being dereferenced if deref_type is a deref_var */
1013 nir_variable *var;
1014
1015 /** Parent deref if deref_type is not deref_var */
1016 nir_src parent;
1017 };
1018
1019 /** Additional deref parameters */
1020 union {
1021 struct {
1022 nir_src index;
1023 } arr;
1024
1025 struct {
1026 unsigned index;
1027 } strct;
1028
1029 struct {
1030 unsigned ptr_stride;
1031 } cast;
1032 };
1033
1034 /** Destination to store the resulting "pointer" */
1035 nir_dest dest;
1036 } nir_deref_instr;
1037
1038 NIR_DEFINE_CAST(nir_instr_as_deref, nir_instr, nir_deref_instr, instr,
1039 type, nir_instr_type_deref)
1040
1041 static inline nir_deref_instr *
1042 nir_src_as_deref(nir_src src)
1043 {
1044 if (!src.is_ssa)
1045 return NULL;
1046
1047 if (src.ssa->parent_instr->type != nir_instr_type_deref)
1048 return NULL;
1049
1050 return nir_instr_as_deref(src.ssa->parent_instr);
1051 }
1052
1053 static inline nir_deref_instr *
1054 nir_deref_instr_parent(const nir_deref_instr *instr)
1055 {
1056 if (instr->deref_type == nir_deref_type_var)
1057 return NULL;
1058 else
1059 return nir_src_as_deref(instr->parent);
1060 }
1061
1062 static inline nir_variable *
1063 nir_deref_instr_get_variable(const nir_deref_instr *instr)
1064 {
1065 while (instr->deref_type != nir_deref_type_var) {
1066 if (instr->deref_type == nir_deref_type_cast)
1067 return NULL;
1068
1069 instr = nir_deref_instr_parent(instr);
1070 }
1071
1072 return instr->var;
1073 }
1074
1075 bool nir_deref_instr_has_indirect(nir_deref_instr *instr);
1076
1077 bool nir_deref_instr_remove_if_unused(nir_deref_instr *instr);
1078
1079 unsigned nir_deref_instr_ptr_as_array_stride(nir_deref_instr *instr);
1080
1081 typedef struct {
1082 nir_instr instr;
1083
1084 struct nir_function *callee;
1085
1086 unsigned num_params;
1087 nir_src params[];
1088 } nir_call_instr;
1089
1090 #include "nir_intrinsics.h"
1091
1092 #define NIR_INTRINSIC_MAX_CONST_INDEX 4
1093
1094 /** Represents an intrinsic
1095 *
1096 * An intrinsic is an instruction type for handling things that are
1097 * more-or-less regular operations but don't just consume and produce SSA
1098 * values like ALU operations do. Intrinsics are not for things that have
1099 * special semantic meaning such as phi nodes and parallel copies.
1100 * Examples of intrinsics include variable load/store operations, system
1101 * value loads, and the like. Even though texturing more-or-less falls
1102 * under this category, texturing is its own instruction type because
1103 * trying to represent texturing with intrinsics would lead to a
1104 * combinatorial explosion of intrinsic opcodes.
1105 *
1106 * By having a single instruction type for handling a lot of different
1107 * cases, optimization passes can look for intrinsics and, for the most
1108 * part, completely ignore them. Each intrinsic type also has a few
1109 * possible flags that govern whether or not they can be reordered or
1110 * eliminated. That way passes like dead code elimination can still work
1111 * on intrisics without understanding the meaning of each.
1112 *
1113 * Each intrinsic has some number of constant indices, some number of
1114 * variables, and some number of sources. What these sources, variables,
1115 * and indices mean depends on the intrinsic and is documented with the
1116 * intrinsic declaration in nir_intrinsics.h. Intrinsics and texture
1117 * instructions are the only types of instruction that can operate on
1118 * variables.
1119 */
1120 typedef struct {
1121 nir_instr instr;
1122
1123 nir_intrinsic_op intrinsic;
1124
1125 nir_dest dest;
1126
1127 /** number of components if this is a vectorized intrinsic
1128 *
1129 * Similarly to ALU operations, some intrinsics are vectorized.
1130 * An intrinsic is vectorized if nir_intrinsic_infos.dest_components == 0.
1131 * For vectorized intrinsics, the num_components field specifies the
1132 * number of destination components and the number of source components
1133 * for all sources with nir_intrinsic_infos.src_components[i] == 0.
1134 */
1135 uint8_t num_components;
1136
1137 int const_index[NIR_INTRINSIC_MAX_CONST_INDEX];
1138
1139 nir_src src[];
1140 } nir_intrinsic_instr;
1141
1142 static inline nir_variable *
1143 nir_intrinsic_get_var(nir_intrinsic_instr *intrin, unsigned i)
1144 {
1145 return nir_deref_instr_get_variable(nir_src_as_deref(intrin->src[i]));
1146 }
1147
1148 /**
1149 * \name NIR intrinsics semantic flags
1150 *
1151 * information about what the compiler can do with the intrinsics.
1152 *
1153 * \sa nir_intrinsic_info::flags
1154 */
1155 typedef enum {
1156 /**
1157 * whether the intrinsic can be safely eliminated if none of its output
1158 * value is not being used.
1159 */
1160 NIR_INTRINSIC_CAN_ELIMINATE = (1 << 0),
1161
1162 /**
1163 * Whether the intrinsic can be reordered with respect to any other
1164 * intrinsic, i.e. whether the only reordering dependencies of the
1165 * intrinsic are due to the register reads/writes.
1166 */
1167 NIR_INTRINSIC_CAN_REORDER = (1 << 1),
1168 } nir_intrinsic_semantic_flag;
1169
1170 /**
1171 * \name NIR intrinsics const-index flag
1172 *
1173 * Indicates the usage of a const_index slot.
1174 *
1175 * \sa nir_intrinsic_info::index_map
1176 */
1177 typedef enum {
1178 /**
1179 * Generally instructions that take a offset src argument, can encode
1180 * a constant 'base' value which is added to the offset.
1181 */
1182 NIR_INTRINSIC_BASE = 1,
1183
1184 /**
1185 * For store instructions, a writemask for the store.
1186 */
1187 NIR_INTRINSIC_WRMASK = 2,
1188
1189 /**
1190 * The stream-id for GS emit_vertex/end_primitive intrinsics.
1191 */
1192 NIR_INTRINSIC_STREAM_ID = 3,
1193
1194 /**
1195 * The clip-plane id for load_user_clip_plane intrinsic.
1196 */
1197 NIR_INTRINSIC_UCP_ID = 4,
1198
1199 /**
1200 * The amount of data, starting from BASE, that this instruction may
1201 * access. This is used to provide bounds if the offset is not constant.
1202 */
1203 NIR_INTRINSIC_RANGE = 5,
1204
1205 /**
1206 * The Vulkan descriptor set for vulkan_resource_index intrinsic.
1207 */
1208 NIR_INTRINSIC_DESC_SET = 6,
1209
1210 /**
1211 * The Vulkan descriptor set binding for vulkan_resource_index intrinsic.
1212 */
1213 NIR_INTRINSIC_BINDING = 7,
1214
1215 /**
1216 * Component offset.
1217 */
1218 NIR_INTRINSIC_COMPONENT = 8,
1219
1220 /**
1221 * Interpolation mode (only meaningful for FS inputs).
1222 */
1223 NIR_INTRINSIC_INTERP_MODE = 9,
1224
1225 /**
1226 * A binary nir_op to use when performing a reduction or scan operation
1227 */
1228 NIR_INTRINSIC_REDUCTION_OP = 10,
1229
1230 /**
1231 * Cluster size for reduction operations
1232 */
1233 NIR_INTRINSIC_CLUSTER_SIZE = 11,
1234
1235 /**
1236 * Parameter index for a load_param intrinsic
1237 */
1238 NIR_INTRINSIC_PARAM_IDX = 12,
1239
1240 /**
1241 * Image dimensionality for image intrinsics
1242 *
1243 * One of GLSL_SAMPLER_DIM_*
1244 */
1245 NIR_INTRINSIC_IMAGE_DIM = 13,
1246
1247 /**
1248 * Non-zero if we are accessing an array image
1249 */
1250 NIR_INTRINSIC_IMAGE_ARRAY = 14,
1251
1252 /**
1253 * Image format for image intrinsics
1254 */
1255 NIR_INTRINSIC_FORMAT = 15,
1256
1257 /**
1258 * Access qualifiers for image and memory access intrinsics
1259 */
1260 NIR_INTRINSIC_ACCESS = 16,
1261
1262 /**
1263 * Alignment for offsets and addresses
1264 *
1265 * These two parameters, specify an alignment in terms of a multiplier and
1266 * an offset. The offset or address parameter X of the intrinsic is
1267 * guaranteed to satisfy the following:
1268 *
1269 * (X - align_offset) % align_mul == 0
1270 */
1271 NIR_INTRINSIC_ALIGN_MUL = 17,
1272 NIR_INTRINSIC_ALIGN_OFFSET = 18,
1273
1274 /**
1275 * The Vulkan descriptor type for a vulkan_resource_[re]index intrinsic.
1276 */
1277 NIR_INTRINSIC_DESC_TYPE = 19,
1278
1279 NIR_INTRINSIC_NUM_INDEX_FLAGS,
1280
1281 } nir_intrinsic_index_flag;
1282
1283 #define NIR_INTRINSIC_MAX_INPUTS 5
1284
1285 typedef struct {
1286 const char *name;
1287
1288 unsigned num_srcs; /** < number of register/SSA inputs */
1289
1290 /** number of components of each input register
1291 *
1292 * If this value is 0, the number of components is given by the
1293 * num_components field of nir_intrinsic_instr. If this value is -1, the
1294 * intrinsic consumes however many components are provided and it is not
1295 * validated at all.
1296 */
1297 int src_components[NIR_INTRINSIC_MAX_INPUTS];
1298
1299 bool has_dest;
1300
1301 /** number of components of the output register
1302 *
1303 * If this value is 0, the number of components is given by the
1304 * num_components field of nir_intrinsic_instr.
1305 */
1306 unsigned dest_components;
1307
1308 /** bitfield of legal bit sizes */
1309 unsigned dest_bit_sizes;
1310
1311 /** the number of constant indices used by the intrinsic */
1312 unsigned num_indices;
1313
1314 /** indicates the usage of intr->const_index[n] */
1315 unsigned index_map[NIR_INTRINSIC_NUM_INDEX_FLAGS];
1316
1317 /** semantic flags for calls to this intrinsic */
1318 nir_intrinsic_semantic_flag flags;
1319 } nir_intrinsic_info;
1320
1321 extern const nir_intrinsic_info nir_intrinsic_infos[nir_num_intrinsics];
1322
1323 static inline unsigned
1324 nir_intrinsic_src_components(nir_intrinsic_instr *intr, unsigned srcn)
1325 {
1326 const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
1327 assert(srcn < info->num_srcs);
1328 if (info->src_components[srcn] > 0)
1329 return info->src_components[srcn];
1330 else if (info->src_components[srcn] == 0)
1331 return intr->num_components;
1332 else
1333 return nir_src_num_components(intr->src[srcn]);
1334 }
1335
1336 static inline unsigned
1337 nir_intrinsic_dest_components(nir_intrinsic_instr *intr)
1338 {
1339 const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
1340 if (!info->has_dest)
1341 return 0;
1342 else if (info->dest_components)
1343 return info->dest_components;
1344 else
1345 return intr->num_components;
1346 }
1347
1348 #define INTRINSIC_IDX_ACCESSORS(name, flag, type) \
1349 static inline type \
1350 nir_intrinsic_##name(const nir_intrinsic_instr *instr) \
1351 { \
1352 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1353 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1354 return (type)instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1]; \
1355 } \
1356 static inline void \
1357 nir_intrinsic_set_##name(nir_intrinsic_instr *instr, type val) \
1358 { \
1359 const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; \
1360 assert(info->index_map[NIR_INTRINSIC_##flag] > 0); \
1361 instr->const_index[info->index_map[NIR_INTRINSIC_##flag] - 1] = val; \
1362 }
1363
1364 INTRINSIC_IDX_ACCESSORS(write_mask, WRMASK, unsigned)
1365 INTRINSIC_IDX_ACCESSORS(base, BASE, int)
1366 INTRINSIC_IDX_ACCESSORS(stream_id, STREAM_ID, unsigned)
1367 INTRINSIC_IDX_ACCESSORS(ucp_id, UCP_ID, unsigned)
1368 INTRINSIC_IDX_ACCESSORS(range, RANGE, unsigned)
1369 INTRINSIC_IDX_ACCESSORS(desc_set, DESC_SET, unsigned)
1370 INTRINSIC_IDX_ACCESSORS(binding, BINDING, unsigned)
1371 INTRINSIC_IDX_ACCESSORS(component, COMPONENT, unsigned)
1372 INTRINSIC_IDX_ACCESSORS(interp_mode, INTERP_MODE, unsigned)
1373 INTRINSIC_IDX_ACCESSORS(reduction_op, REDUCTION_OP, unsigned)
1374 INTRINSIC_IDX_ACCESSORS(cluster_size, CLUSTER_SIZE, unsigned)
1375 INTRINSIC_IDX_ACCESSORS(param_idx, PARAM_IDX, unsigned)
1376 INTRINSIC_IDX_ACCESSORS(image_dim, IMAGE_DIM, enum glsl_sampler_dim)
1377 INTRINSIC_IDX_ACCESSORS(image_array, IMAGE_ARRAY, bool)
1378 INTRINSIC_IDX_ACCESSORS(access, ACCESS, enum gl_access_qualifier)
1379 INTRINSIC_IDX_ACCESSORS(format, FORMAT, unsigned)
1380 INTRINSIC_IDX_ACCESSORS(align_mul, ALIGN_MUL, unsigned)
1381 INTRINSIC_IDX_ACCESSORS(align_offset, ALIGN_OFFSET, unsigned)
1382 INTRINSIC_IDX_ACCESSORS(desc_type, DESC_TYPE, unsigned)
1383
1384 static inline void
1385 nir_intrinsic_set_align(nir_intrinsic_instr *intrin,
1386 unsigned align_mul, unsigned align_offset)
1387 {
1388 assert(util_is_power_of_two_nonzero(align_mul));
1389 assert(align_offset < align_mul);
1390 nir_intrinsic_set_align_mul(intrin, align_mul);
1391 nir_intrinsic_set_align_offset(intrin, align_offset);
1392 }
1393
1394 /** Returns a simple alignment for a load/store intrinsic offset
1395 *
1396 * Instead of the full mul+offset alignment scheme provided by the ALIGN_MUL
1397 * and ALIGN_OFFSET parameters, this helper takes both into account and
1398 * provides a single simple alignment parameter. The offset X is guaranteed
1399 * to satisfy X % align == 0.
1400 */
1401 static inline unsigned
1402 nir_intrinsic_align(const nir_intrinsic_instr *intrin)
1403 {
1404 const unsigned align_mul = nir_intrinsic_align_mul(intrin);
1405 const unsigned align_offset = nir_intrinsic_align_offset(intrin);
1406 assert(align_offset < align_mul);
1407 return align_offset ? 1 << (ffs(align_offset) - 1) : align_mul;
1408 }
1409
1410 /**
1411 * \group texture information
1412 *
1413 * This gives semantic information about textures which is useful to the
1414 * frontend, the backend, and lowering passes, but not the optimizer.
1415 */
1416
1417 typedef enum {
1418 nir_tex_src_coord,
1419 nir_tex_src_projector,
1420 nir_tex_src_comparator, /* shadow comparator */
1421 nir_tex_src_offset,
1422 nir_tex_src_bias,
1423 nir_tex_src_lod,
1424 nir_tex_src_min_lod,
1425 nir_tex_src_ms_index, /* MSAA sample index */
1426 nir_tex_src_ms_mcs, /* MSAA compression value */
1427 nir_tex_src_ddx,
1428 nir_tex_src_ddy,
1429 nir_tex_src_texture_deref, /* < deref pointing to the texture */
1430 nir_tex_src_sampler_deref, /* < deref pointing to the sampler */
1431 nir_tex_src_texture_offset, /* < dynamically uniform indirect offset */
1432 nir_tex_src_sampler_offset, /* < dynamically uniform indirect offset */
1433 nir_tex_src_plane, /* < selects plane for planar textures */
1434 nir_num_tex_src_types
1435 } nir_tex_src_type;
1436
1437 typedef struct {
1438 nir_src src;
1439 nir_tex_src_type src_type;
1440 } nir_tex_src;
1441
1442 typedef enum {
1443 nir_texop_tex, /**< Regular texture look-up */
1444 nir_texop_txb, /**< Texture look-up with LOD bias */
1445 nir_texop_txl, /**< Texture look-up with explicit LOD */
1446 nir_texop_txd, /**< Texture look-up with partial derivatives */
1447 nir_texop_txf, /**< Texel fetch with explicit LOD */
1448 nir_texop_txf_ms, /**< Multisample texture fetch */
1449 nir_texop_txf_ms_mcs, /**< Multisample compression value fetch */
1450 nir_texop_txs, /**< Texture size */
1451 nir_texop_lod, /**< Texture lod query */
1452 nir_texop_tg4, /**< Texture gather */
1453 nir_texop_query_levels, /**< Texture levels query */
1454 nir_texop_texture_samples, /**< Texture samples query */
1455 nir_texop_samples_identical, /**< Query whether all samples are definitely
1456 * identical.
1457 */
1458 } nir_texop;
1459
1460 typedef struct {
1461 nir_instr instr;
1462
1463 enum glsl_sampler_dim sampler_dim;
1464 nir_alu_type dest_type;
1465
1466 nir_texop op;
1467 nir_dest dest;
1468 nir_tex_src *src;
1469 unsigned num_srcs, coord_components;
1470 bool is_array, is_shadow;
1471
1472 /**
1473 * If is_shadow is true, whether this is the old-style shadow that outputs 4
1474 * components or the new-style shadow that outputs 1 component.
1475 */
1476 bool is_new_style_shadow;
1477
1478 /* gather component selector */
1479 unsigned component : 2;
1480
1481 /** The texture index
1482 *
1483 * If this texture instruction has a nir_tex_src_texture_offset source,
1484 * then the texture index is given by texture_index + texture_offset.
1485 */
1486 unsigned texture_index;
1487
1488 /** The size of the texture array or 0 if it's not an array */
1489 unsigned texture_array_size;
1490
1491 /** The sampler index
1492 *
1493 * The following operations do not require a sampler and, as such, this
1494 * field should be ignored:
1495 * - nir_texop_txf
1496 * - nir_texop_txf_ms
1497 * - nir_texop_txs
1498 * - nir_texop_lod
1499 * - nir_texop_query_levels
1500 * - nir_texop_texture_samples
1501 * - nir_texop_samples_identical
1502 *
1503 * If this texture instruction has a nir_tex_src_sampler_offset source,
1504 * then the sampler index is given by sampler_index + sampler_offset.
1505 */
1506 unsigned sampler_index;
1507 } nir_tex_instr;
1508
1509 static inline unsigned
1510 nir_tex_instr_dest_size(const nir_tex_instr *instr)
1511 {
1512 switch (instr->op) {
1513 case nir_texop_txs: {
1514 unsigned ret;
1515 switch (instr->sampler_dim) {
1516 case GLSL_SAMPLER_DIM_1D:
1517 case GLSL_SAMPLER_DIM_BUF:
1518 ret = 1;
1519 break;
1520 case GLSL_SAMPLER_DIM_2D:
1521 case GLSL_SAMPLER_DIM_CUBE:
1522 case GLSL_SAMPLER_DIM_MS:
1523 case GLSL_SAMPLER_DIM_RECT:
1524 case GLSL_SAMPLER_DIM_EXTERNAL:
1525 case GLSL_SAMPLER_DIM_SUBPASS:
1526 ret = 2;
1527 break;
1528 case GLSL_SAMPLER_DIM_3D:
1529 ret = 3;
1530 break;
1531 default:
1532 unreachable("not reached");
1533 }
1534 if (instr->is_array)
1535 ret++;
1536 return ret;
1537 }
1538
1539 case nir_texop_lod:
1540 return 2;
1541
1542 case nir_texop_texture_samples:
1543 case nir_texop_query_levels:
1544 case nir_texop_samples_identical:
1545 return 1;
1546
1547 default:
1548 if (instr->is_shadow && instr->is_new_style_shadow)
1549 return 1;
1550
1551 return 4;
1552 }
1553 }
1554
1555 /* Returns true if this texture operation queries something about the texture
1556 * rather than actually sampling it.
1557 */
1558 static inline bool
1559 nir_tex_instr_is_query(const nir_tex_instr *instr)
1560 {
1561 switch (instr->op) {
1562 case nir_texop_txs:
1563 case nir_texop_lod:
1564 case nir_texop_texture_samples:
1565 case nir_texop_query_levels:
1566 case nir_texop_txf_ms_mcs:
1567 return true;
1568 case nir_texop_tex:
1569 case nir_texop_txb:
1570 case nir_texop_txl:
1571 case nir_texop_txd:
1572 case nir_texop_txf:
1573 case nir_texop_txf_ms:
1574 case nir_texop_tg4:
1575 return false;
1576 default:
1577 unreachable("Invalid texture opcode");
1578 }
1579 }
1580
1581 static inline bool
1582 nir_alu_instr_is_comparison(const nir_alu_instr *instr)
1583 {
1584 switch (instr->op) {
1585 case nir_op_flt:
1586 case nir_op_fge:
1587 case nir_op_feq:
1588 case nir_op_fne:
1589 case nir_op_ilt:
1590 case nir_op_ult:
1591 case nir_op_ige:
1592 case nir_op_uge:
1593 case nir_op_ieq:
1594 case nir_op_ine:
1595 case nir_op_i2b1:
1596 case nir_op_f2b1:
1597 case nir_op_inot:
1598 case nir_op_fnot:
1599 return true;
1600 default:
1601 return false;
1602 }
1603 }
1604
1605 static inline nir_alu_type
1606 nir_tex_instr_src_type(const nir_tex_instr *instr, unsigned src)
1607 {
1608 switch (instr->src[src].src_type) {
1609 case nir_tex_src_coord:
1610 switch (instr->op) {
1611 case nir_texop_txf:
1612 case nir_texop_txf_ms:
1613 case nir_texop_txf_ms_mcs:
1614 case nir_texop_samples_identical:
1615 return nir_type_int;
1616
1617 default:
1618 return nir_type_float;
1619 }
1620
1621 case nir_tex_src_lod:
1622 switch (instr->op) {
1623 case nir_texop_txs:
1624 case nir_texop_txf:
1625 return nir_type_int;
1626
1627 default:
1628 return nir_type_float;
1629 }
1630
1631 case nir_tex_src_projector:
1632 case nir_tex_src_comparator:
1633 case nir_tex_src_bias:
1634 case nir_tex_src_ddx:
1635 case nir_tex_src_ddy:
1636 return nir_type_float;
1637
1638 case nir_tex_src_offset:
1639 case nir_tex_src_ms_index:
1640 case nir_tex_src_texture_offset:
1641 case nir_tex_src_sampler_offset:
1642 return nir_type_int;
1643
1644 default:
1645 unreachable("Invalid texture source type");
1646 }
1647 }
1648
1649 static inline unsigned
1650 nir_tex_instr_src_size(const nir_tex_instr *instr, unsigned src)
1651 {
1652 if (instr->src[src].src_type == nir_tex_src_coord)
1653 return instr->coord_components;
1654
1655 /* The MCS value is expected to be a vec4 returned by a txf_ms_mcs */
1656 if (instr->src[src].src_type == nir_tex_src_ms_mcs)
1657 return 4;
1658
1659 if (instr->src[src].src_type == nir_tex_src_ddx ||
1660 instr->src[src].src_type == nir_tex_src_ddy) {
1661 if (instr->is_array)
1662 return instr->coord_components - 1;
1663 else
1664 return instr->coord_components;
1665 }
1666
1667 /* Usual APIs don't allow cube + offset, but we allow it, with 2 coords for
1668 * the offset, since a cube maps to a single face.
1669 */
1670 if (instr->src[src].src_type == nir_tex_src_offset) {
1671 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE)
1672 return 2;
1673 else if (instr->is_array)
1674 return instr->coord_components - 1;
1675 else
1676 return instr->coord_components;
1677 }
1678
1679 return 1;
1680 }
1681
1682 static inline int
1683 nir_tex_instr_src_index(const nir_tex_instr *instr, nir_tex_src_type type)
1684 {
1685 for (unsigned i = 0; i < instr->num_srcs; i++)
1686 if (instr->src[i].src_type == type)
1687 return (int) i;
1688
1689 return -1;
1690 }
1691
1692 void nir_tex_instr_add_src(nir_tex_instr *tex,
1693 nir_tex_src_type src_type,
1694 nir_src src);
1695
1696 void nir_tex_instr_remove_src(nir_tex_instr *tex, unsigned src_idx);
1697
1698 typedef struct {
1699 nir_instr instr;
1700
1701 nir_const_value value;
1702
1703 nir_ssa_def def;
1704 } nir_load_const_instr;
1705
1706 typedef enum {
1707 nir_jump_return,
1708 nir_jump_break,
1709 nir_jump_continue,
1710 } nir_jump_type;
1711
1712 typedef struct {
1713 nir_instr instr;
1714 nir_jump_type type;
1715 } nir_jump_instr;
1716
1717 /* creates a new SSA variable in an undefined state */
1718
1719 typedef struct {
1720 nir_instr instr;
1721 nir_ssa_def def;
1722 } nir_ssa_undef_instr;
1723
1724 typedef struct {
1725 struct exec_node node;
1726
1727 /* The predecessor block corresponding to this source */
1728 struct nir_block *pred;
1729
1730 nir_src src;
1731 } nir_phi_src;
1732
1733 #define nir_foreach_phi_src(phi_src, phi) \
1734 foreach_list_typed(nir_phi_src, phi_src, node, &(phi)->srcs)
1735 #define nir_foreach_phi_src_safe(phi_src, phi) \
1736 foreach_list_typed_safe(nir_phi_src, phi_src, node, &(phi)->srcs)
1737
1738 typedef struct {
1739 nir_instr instr;
1740
1741 struct exec_list srcs; /** < list of nir_phi_src */
1742
1743 nir_dest dest;
1744 } nir_phi_instr;
1745
1746 typedef struct {
1747 struct exec_node node;
1748 nir_src src;
1749 nir_dest dest;
1750 } nir_parallel_copy_entry;
1751
1752 #define nir_foreach_parallel_copy_entry(entry, pcopy) \
1753 foreach_list_typed(nir_parallel_copy_entry, entry, node, &(pcopy)->entries)
1754
1755 typedef struct {
1756 nir_instr instr;
1757
1758 /* A list of nir_parallel_copy_entrys. The sources of all of the
1759 * entries are copied to the corresponding destinations "in parallel".
1760 * In other words, if we have two entries: a -> b and b -> a, the values
1761 * get swapped.
1762 */
1763 struct exec_list entries;
1764 } nir_parallel_copy_instr;
1765
1766 NIR_DEFINE_CAST(nir_instr_as_alu, nir_instr, nir_alu_instr, instr,
1767 type, nir_instr_type_alu)
1768 NIR_DEFINE_CAST(nir_instr_as_call, nir_instr, nir_call_instr, instr,
1769 type, nir_instr_type_call)
1770 NIR_DEFINE_CAST(nir_instr_as_jump, nir_instr, nir_jump_instr, instr,
1771 type, nir_instr_type_jump)
1772 NIR_DEFINE_CAST(nir_instr_as_tex, nir_instr, nir_tex_instr, instr,
1773 type, nir_instr_type_tex)
1774 NIR_DEFINE_CAST(nir_instr_as_intrinsic, nir_instr, nir_intrinsic_instr, instr,
1775 type, nir_instr_type_intrinsic)
1776 NIR_DEFINE_CAST(nir_instr_as_load_const, nir_instr, nir_load_const_instr, instr,
1777 type, nir_instr_type_load_const)
1778 NIR_DEFINE_CAST(nir_instr_as_ssa_undef, nir_instr, nir_ssa_undef_instr, instr,
1779 type, nir_instr_type_ssa_undef)
1780 NIR_DEFINE_CAST(nir_instr_as_phi, nir_instr, nir_phi_instr, instr,
1781 type, nir_instr_type_phi)
1782 NIR_DEFINE_CAST(nir_instr_as_parallel_copy, nir_instr,
1783 nir_parallel_copy_instr, instr,
1784 type, nir_instr_type_parallel_copy)
1785
1786 /*
1787 * Control flow
1788 *
1789 * Control flow consists of a tree of control flow nodes, which include
1790 * if-statements and loops. The leaves of the tree are basic blocks, lists of
1791 * instructions that always run start-to-finish. Each basic block also keeps
1792 * track of its successors (blocks which may run immediately after the current
1793 * block) and predecessors (blocks which could have run immediately before the
1794 * current block). Each function also has a start block and an end block which
1795 * all return statements point to (which is always empty). Together, all the
1796 * blocks with their predecessors and successors make up the control flow
1797 * graph (CFG) of the function. There are helpers that modify the tree of
1798 * control flow nodes while modifying the CFG appropriately; these should be
1799 * used instead of modifying the tree directly.
1800 */
1801
1802 typedef enum {
1803 nir_cf_node_block,
1804 nir_cf_node_if,
1805 nir_cf_node_loop,
1806 nir_cf_node_function
1807 } nir_cf_node_type;
1808
1809 typedef struct nir_cf_node {
1810 struct exec_node node;
1811 nir_cf_node_type type;
1812 struct nir_cf_node *parent;
1813 } nir_cf_node;
1814
1815 typedef struct nir_block {
1816 nir_cf_node cf_node;
1817
1818 struct exec_list instr_list; /** < list of nir_instr */
1819
1820 /** generic block index; generated by nir_index_blocks */
1821 unsigned index;
1822
1823 /*
1824 * Each block can only have up to 2 successors, so we put them in a simple
1825 * array - no need for anything more complicated.
1826 */
1827 struct nir_block *successors[2];
1828
1829 /* Set of nir_block predecessors in the CFG */
1830 struct set *predecessors;
1831
1832 /*
1833 * this node's immediate dominator in the dominance tree - set to NULL for
1834 * the start block.
1835 */
1836 struct nir_block *imm_dom;
1837
1838 /* This node's children in the dominance tree */
1839 unsigned num_dom_children;
1840 struct nir_block **dom_children;
1841
1842 /* Set of nir_blocks on the dominance frontier of this block */
1843 struct set *dom_frontier;
1844
1845 /*
1846 * These two indices have the property that dom_{pre,post}_index for each
1847 * child of this block in the dominance tree will always be between
1848 * dom_pre_index and dom_post_index for this block, which makes testing if
1849 * a given block is dominated by another block an O(1) operation.
1850 */
1851 unsigned dom_pre_index, dom_post_index;
1852
1853 /* live in and out for this block; used for liveness analysis */
1854 BITSET_WORD *live_in;
1855 BITSET_WORD *live_out;
1856 } nir_block;
1857
1858 static inline nir_instr *
1859 nir_block_first_instr(nir_block *block)
1860 {
1861 struct exec_node *head = exec_list_get_head(&block->instr_list);
1862 return exec_node_data(nir_instr, head, node);
1863 }
1864
1865 static inline nir_instr *
1866 nir_block_last_instr(nir_block *block)
1867 {
1868 struct exec_node *tail = exec_list_get_tail(&block->instr_list);
1869 return exec_node_data(nir_instr, tail, node);
1870 }
1871
1872 static inline bool
1873 nir_block_ends_in_jump(nir_block *block)
1874 {
1875 return !exec_list_is_empty(&block->instr_list) &&
1876 nir_block_last_instr(block)->type == nir_instr_type_jump;
1877 }
1878
1879 #define nir_foreach_instr(instr, block) \
1880 foreach_list_typed(nir_instr, instr, node, &(block)->instr_list)
1881 #define nir_foreach_instr_reverse(instr, block) \
1882 foreach_list_typed_reverse(nir_instr, instr, node, &(block)->instr_list)
1883 #define nir_foreach_instr_safe(instr, block) \
1884 foreach_list_typed_safe(nir_instr, instr, node, &(block)->instr_list)
1885 #define nir_foreach_instr_reverse_safe(instr, block) \
1886 foreach_list_typed_reverse_safe(nir_instr, instr, node, &(block)->instr_list)
1887
1888 typedef struct nir_if {
1889 nir_cf_node cf_node;
1890 nir_src condition;
1891
1892 struct exec_list then_list; /** < list of nir_cf_node */
1893 struct exec_list else_list; /** < list of nir_cf_node */
1894 } nir_if;
1895
1896 typedef struct {
1897 nir_if *nif;
1898
1899 /** Instruction that generates nif::condition. */
1900 nir_instr *conditional_instr;
1901
1902 /** Block within ::nif that has the break instruction. */
1903 nir_block *break_block;
1904
1905 /** Last block for the then- or else-path that does not contain the break. */
1906 nir_block *continue_from_block;
1907
1908 /** True when ::break_block is in the else-path of ::nif. */
1909 bool continue_from_then;
1910
1911 struct list_head loop_terminator_link;
1912 } nir_loop_terminator;
1913
1914 typedef struct {
1915 /* Number of instructions in the loop */
1916 unsigned num_instructions;
1917
1918 /* Maximum number of times the loop is run (if known) */
1919 unsigned max_trip_count;
1920
1921 /* Do we know the exact number of times the loop will be run */
1922 bool exact_trip_count_known;
1923
1924 /* Unroll the loop regardless of its size */
1925 bool force_unroll;
1926
1927 /* Does the loop contain complex loop terminators, continues or other
1928 * complex behaviours? If this is true we can't rely on
1929 * loop_terminator_list to be complete or accurate.
1930 */
1931 bool complex_loop;
1932
1933 nir_loop_terminator *limiting_terminator;
1934
1935 /* A list of loop_terminators terminating this loop. */
1936 struct list_head loop_terminator_list;
1937 } nir_loop_info;
1938
1939 typedef struct {
1940 nir_cf_node cf_node;
1941
1942 struct exec_list body; /** < list of nir_cf_node */
1943
1944 nir_loop_info *info;
1945 } nir_loop;
1946
1947 /**
1948 * Various bits of metadata that can may be created or required by
1949 * optimization and analysis passes
1950 */
1951 typedef enum {
1952 nir_metadata_none = 0x0,
1953 nir_metadata_block_index = 0x1,
1954 nir_metadata_dominance = 0x2,
1955 nir_metadata_live_ssa_defs = 0x4,
1956 nir_metadata_not_properly_reset = 0x8,
1957 nir_metadata_loop_analysis = 0x10,
1958 } nir_metadata;
1959
1960 typedef struct {
1961 nir_cf_node cf_node;
1962
1963 /** pointer to the function of which this is an implementation */
1964 struct nir_function *function;
1965
1966 struct exec_list body; /** < list of nir_cf_node */
1967
1968 nir_block *end_block;
1969
1970 /** list for all local variables in the function */
1971 struct exec_list locals;
1972
1973 /** list of local registers in the function */
1974 struct exec_list registers;
1975
1976 /** next available local register index */
1977 unsigned reg_alloc;
1978
1979 /** next available SSA value index */
1980 unsigned ssa_alloc;
1981
1982 /* total number of basic blocks, only valid when block_index_dirty = false */
1983 unsigned num_blocks;
1984
1985 nir_metadata valid_metadata;
1986 } nir_function_impl;
1987
1988 ATTRIBUTE_RETURNS_NONNULL static inline nir_block *
1989 nir_start_block(nir_function_impl *impl)
1990 {
1991 return (nir_block *) impl->body.head_sentinel.next;
1992 }
1993
1994 ATTRIBUTE_RETURNS_NONNULL static inline nir_block *
1995 nir_impl_last_block(nir_function_impl *impl)
1996 {
1997 return (nir_block *) impl->body.tail_sentinel.prev;
1998 }
1999
2000 static inline nir_cf_node *
2001 nir_cf_node_next(nir_cf_node *node)
2002 {
2003 struct exec_node *next = exec_node_get_next(&node->node);
2004 if (exec_node_is_tail_sentinel(next))
2005 return NULL;
2006 else
2007 return exec_node_data(nir_cf_node, next, node);
2008 }
2009
2010 static inline nir_cf_node *
2011 nir_cf_node_prev(nir_cf_node *node)
2012 {
2013 struct exec_node *prev = exec_node_get_prev(&node->node);
2014 if (exec_node_is_head_sentinel(prev))
2015 return NULL;
2016 else
2017 return exec_node_data(nir_cf_node, prev, node);
2018 }
2019
2020 static inline bool
2021 nir_cf_node_is_first(const nir_cf_node *node)
2022 {
2023 return exec_node_is_head_sentinel(node->node.prev);
2024 }
2025
2026 static inline bool
2027 nir_cf_node_is_last(const nir_cf_node *node)
2028 {
2029 return exec_node_is_tail_sentinel(node->node.next);
2030 }
2031
2032 NIR_DEFINE_CAST(nir_cf_node_as_block, nir_cf_node, nir_block, cf_node,
2033 type, nir_cf_node_block)
2034 NIR_DEFINE_CAST(nir_cf_node_as_if, nir_cf_node, nir_if, cf_node,
2035 type, nir_cf_node_if)
2036 NIR_DEFINE_CAST(nir_cf_node_as_loop, nir_cf_node, nir_loop, cf_node,
2037 type, nir_cf_node_loop)
2038 NIR_DEFINE_CAST(nir_cf_node_as_function, nir_cf_node,
2039 nir_function_impl, cf_node, type, nir_cf_node_function)
2040
2041 static inline nir_block *
2042 nir_if_first_then_block(nir_if *if_stmt)
2043 {
2044 struct exec_node *head = exec_list_get_head(&if_stmt->then_list);
2045 return nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
2046 }
2047
2048 static inline nir_block *
2049 nir_if_last_then_block(nir_if *if_stmt)
2050 {
2051 struct exec_node *tail = exec_list_get_tail(&if_stmt->then_list);
2052 return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
2053 }
2054
2055 static inline nir_block *
2056 nir_if_first_else_block(nir_if *if_stmt)
2057 {
2058 struct exec_node *head = exec_list_get_head(&if_stmt->else_list);
2059 return nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
2060 }
2061
2062 static inline nir_block *
2063 nir_if_last_else_block(nir_if *if_stmt)
2064 {
2065 struct exec_node *tail = exec_list_get_tail(&if_stmt->else_list);
2066 return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
2067 }
2068
2069 static inline nir_block *
2070 nir_loop_first_block(nir_loop *loop)
2071 {
2072 struct exec_node *head = exec_list_get_head(&loop->body);
2073 return nir_cf_node_as_block(exec_node_data(nir_cf_node, head, node));
2074 }
2075
2076 static inline nir_block *
2077 nir_loop_last_block(nir_loop *loop)
2078 {
2079 struct exec_node *tail = exec_list_get_tail(&loop->body);
2080 return nir_cf_node_as_block(exec_node_data(nir_cf_node, tail, node));
2081 }
2082
2083 typedef struct {
2084 uint8_t num_components;
2085 uint8_t bit_size;
2086 } nir_parameter;
2087
2088 typedef struct nir_function {
2089 struct exec_node node;
2090
2091 const char *name;
2092 struct nir_shader *shader;
2093
2094 unsigned num_params;
2095 nir_parameter *params;
2096
2097 /** The implementation of this function.
2098 *
2099 * If the function is only declared and not implemented, this is NULL.
2100 */
2101 nir_function_impl *impl;
2102
2103 bool is_entrypoint;
2104 } nir_function;
2105
2106 typedef enum {
2107 nir_lower_imul64 = (1 << 0),
2108 nir_lower_isign64 = (1 << 1),
2109 /** Lower all int64 modulus and division opcodes */
2110 nir_lower_divmod64 = (1 << 2),
2111 /** Lower all 64-bit umul_high and imul_high opcodes */
2112 nir_lower_imul_high64 = (1 << 3),
2113 nir_lower_mov64 = (1 << 4),
2114 nir_lower_icmp64 = (1 << 5),
2115 nir_lower_iadd64 = (1 << 6),
2116 nir_lower_iabs64 = (1 << 7),
2117 nir_lower_ineg64 = (1 << 8),
2118 nir_lower_logic64 = (1 << 9),
2119 nir_lower_minmax64 = (1 << 10),
2120 nir_lower_shift64 = (1 << 11),
2121 nir_lower_imul_2x32_64 = (1 << 12),
2122 } nir_lower_int64_options;
2123
2124 typedef enum {
2125 nir_lower_drcp = (1 << 0),
2126 nir_lower_dsqrt = (1 << 1),
2127 nir_lower_drsq = (1 << 2),
2128 nir_lower_dtrunc = (1 << 3),
2129 nir_lower_dfloor = (1 << 4),
2130 nir_lower_dceil = (1 << 5),
2131 nir_lower_dfract = (1 << 6),
2132 nir_lower_dround_even = (1 << 7),
2133 nir_lower_dmod = (1 << 8),
2134 nir_lower_fp64_full_software = (1 << 9),
2135 } nir_lower_doubles_options;
2136
2137 typedef struct nir_shader_compiler_options {
2138 bool lower_fdiv;
2139 bool lower_ffma;
2140 bool fuse_ffma;
2141 bool lower_flrp32;
2142 /** Lowers flrp when it does not support doubles */
2143 bool lower_flrp64;
2144 bool lower_fpow;
2145 bool lower_fsat;
2146 bool lower_fsqrt;
2147 bool lower_fmod32;
2148 bool lower_fmod64;
2149 /** Lowers ibitfield_extract/ubitfield_extract to ibfe/ubfe. */
2150 bool lower_bitfield_extract;
2151 /** Lowers ibitfield_extract/ubitfield_extract to bfm, compares, shifts. */
2152 bool lower_bitfield_extract_to_shifts;
2153 /** Lowers bitfield_insert to bfi/bfm */
2154 bool lower_bitfield_insert;
2155 /** Lowers bitfield_insert to bfm, compares, and shifts. */
2156 bool lower_bitfield_insert_to_shifts;
2157 /** Lowers bitfield_reverse to shifts. */
2158 bool lower_bitfield_reverse;
2159 /** Lowers bit_count to shifts. */
2160 bool lower_bit_count;
2161 /** Lowers bfm to shifts and subtracts. */
2162 bool lower_bfm;
2163 /** Lowers ifind_msb to compare and ufind_msb */
2164 bool lower_ifind_msb;
2165 /** Lowers find_lsb to ufind_msb and logic ops */
2166 bool lower_find_lsb;
2167 bool lower_uadd_carry;
2168 bool lower_usub_borrow;
2169 /** Lowers imul_high/umul_high to 16-bit multiplies and carry operations. */
2170 bool lower_mul_high;
2171 /** lowers fneg and ineg to fsub and isub. */
2172 bool lower_negate;
2173 /** lowers fsub and isub to fadd+fneg and iadd+ineg. */
2174 bool lower_sub;
2175
2176 /* lower {slt,sge,seq,sne} to {flt,fge,feq,fne} + b2f: */
2177 bool lower_scmp;
2178
2179 /** enables rules to lower idiv by power-of-two: */
2180 bool lower_idiv;
2181
2182 /** enables rules to lower isign to imin+imax */
2183 bool lower_isign;
2184
2185 /* Does the native fdot instruction replicate its result for four
2186 * components? If so, then opt_algebraic_late will turn all fdotN
2187 * instructions into fdot_replicatedN instructions.
2188 */
2189 bool fdot_replicates;
2190
2191 /** lowers ffloor to fsub+ffract: */
2192 bool lower_ffloor;
2193
2194 /** lowers ffract to fsub+ffloor: */
2195 bool lower_ffract;
2196
2197 /** lowers fceil to fneg+ffloor+fneg: */
2198 bool lower_fceil;
2199
2200 bool lower_ldexp;
2201
2202 bool lower_pack_half_2x16;
2203 bool lower_pack_unorm_2x16;
2204 bool lower_pack_snorm_2x16;
2205 bool lower_pack_unorm_4x8;
2206 bool lower_pack_snorm_4x8;
2207 bool lower_unpack_half_2x16;
2208 bool lower_unpack_unorm_2x16;
2209 bool lower_unpack_snorm_2x16;
2210 bool lower_unpack_unorm_4x8;
2211 bool lower_unpack_snorm_4x8;
2212
2213 bool lower_extract_byte;
2214 bool lower_extract_word;
2215
2216 bool lower_all_io_to_temps;
2217
2218 /**
2219 * Does the driver support real 32-bit integers? (Otherwise, integers
2220 * are simulated by floats.)
2221 */
2222 bool native_integers;
2223
2224 /* Indicates that the driver only has zero-based vertex id */
2225 bool vertex_id_zero_based;
2226
2227 /**
2228 * If enabled, gl_BaseVertex will be lowered as:
2229 * is_indexed_draw (~0/0) & firstvertex
2230 */
2231 bool lower_base_vertex;
2232
2233 /**
2234 * If enabled, gl_HelperInvocation will be lowered as:
2235 *
2236 * !((1 << sample_id) & sample_mask_in))
2237 *
2238 * This depends on some possibly hw implementation details, which may
2239 * not be true for all hw. In particular that the FS is only executed
2240 * for covered samples or for helper invocations. So, do not blindly
2241 * enable this option.
2242 *
2243 * Note: See also issue #22 in ARB_shader_image_load_store
2244 */
2245 bool lower_helper_invocation;
2246
2247 bool lower_cs_local_index_from_id;
2248 bool lower_cs_local_id_from_index;
2249
2250 bool lower_device_index_to_zero;
2251
2252 /* Set if nir_lower_wpos_ytransform() should also invert gl_PointCoord. */
2253 bool lower_wpos_pntc;
2254
2255 /**
2256 * Should nir_lower_io() create load_interpolated_input intrinsics?
2257 *
2258 * If not, it generates regular load_input intrinsics and interpolation
2259 * information must be inferred from the list of input nir_variables.
2260 */
2261 bool use_interpolated_input_intrinsics;
2262
2263 /* Lowers when 32x32->64 bit multiplication is not supported */
2264 bool lower_mul_2x32_64;
2265
2266 unsigned max_unroll_iterations;
2267
2268 nir_lower_int64_options lower_int64_options;
2269 nir_lower_doubles_options lower_doubles_options;
2270 } nir_shader_compiler_options;
2271
2272 typedef struct nir_shader {
2273 /** list of uniforms (nir_variable) */
2274 struct exec_list uniforms;
2275
2276 /** list of inputs (nir_variable) */
2277 struct exec_list inputs;
2278
2279 /** list of outputs (nir_variable) */
2280 struct exec_list outputs;
2281
2282 /** list of shared compute variables (nir_variable) */
2283 struct exec_list shared;
2284
2285 /** Set of driver-specific options for the shader.
2286 *
2287 * The memory for the options is expected to be kept in a single static
2288 * copy by the driver.
2289 */
2290 const struct nir_shader_compiler_options *options;
2291
2292 /** Various bits of compile-time information about a given shader */
2293 struct shader_info info;
2294
2295 /** list of global variables in the shader (nir_variable) */
2296 struct exec_list globals;
2297
2298 /** list of system value variables in the shader (nir_variable) */
2299 struct exec_list system_values;
2300
2301 struct exec_list functions; /** < list of nir_function */
2302
2303 /** list of global register in the shader */
2304 struct exec_list registers;
2305
2306 /** next available global register index */
2307 unsigned reg_alloc;
2308
2309 /**
2310 * the highest index a load_input_*, load_uniform_*, etc. intrinsic can
2311 * access plus one
2312 */
2313 unsigned num_inputs, num_uniforms, num_outputs, num_shared;
2314
2315 /** Constant data associated with this shader.
2316 *
2317 * Constant data is loaded through load_constant intrinsics. See also
2318 * nir_opt_large_constants.
2319 */
2320 void *constant_data;
2321 unsigned constant_data_size;
2322 } nir_shader;
2323
2324 #define nir_foreach_function(func, shader) \
2325 foreach_list_typed(nir_function, func, node, &(shader)->functions)
2326
2327 static inline nir_function_impl *
2328 nir_shader_get_entrypoint(nir_shader *shader)
2329 {
2330 nir_function *func = NULL;
2331
2332 nir_foreach_function(function, shader) {
2333 assert(func == NULL);
2334 if (function->is_entrypoint) {
2335 func = function;
2336 #ifndef NDEBUG
2337 break;
2338 #endif
2339 }
2340 }
2341
2342 if (!func)
2343 return NULL;
2344
2345 assert(func->num_params == 0);
2346 assert(func->impl);
2347 return func->impl;
2348 }
2349
2350 nir_shader *nir_shader_create(void *mem_ctx,
2351 gl_shader_stage stage,
2352 const nir_shader_compiler_options *options,
2353 shader_info *si);
2354
2355 /** creates a register, including assigning it an index and adding it to the list */
2356 nir_register *nir_global_reg_create(nir_shader *shader);
2357
2358 nir_register *nir_local_reg_create(nir_function_impl *impl);
2359
2360 void nir_reg_remove(nir_register *reg);
2361
2362 /** Adds a variable to the appropriate list in nir_shader */
2363 void nir_shader_add_variable(nir_shader *shader, nir_variable *var);
2364
2365 static inline void
2366 nir_function_impl_add_variable(nir_function_impl *impl, nir_variable *var)
2367 {
2368 assert(var->data.mode == nir_var_function_temp);
2369 exec_list_push_tail(&impl->locals, &var->node);
2370 }
2371
2372 /** creates a variable, sets a few defaults, and adds it to the list */
2373 nir_variable *nir_variable_create(nir_shader *shader,
2374 nir_variable_mode mode,
2375 const struct glsl_type *type,
2376 const char *name);
2377 /** creates a local variable and adds it to the list */
2378 nir_variable *nir_local_variable_create(nir_function_impl *impl,
2379 const struct glsl_type *type,
2380 const char *name);
2381
2382 /** creates a function and adds it to the shader's list of functions */
2383 nir_function *nir_function_create(nir_shader *shader, const char *name);
2384
2385 nir_function_impl *nir_function_impl_create(nir_function *func);
2386 /** creates a function_impl that isn't tied to any particular function */
2387 nir_function_impl *nir_function_impl_create_bare(nir_shader *shader);
2388
2389 nir_block *nir_block_create(nir_shader *shader);
2390 nir_if *nir_if_create(nir_shader *shader);
2391 nir_loop *nir_loop_create(nir_shader *shader);
2392
2393 nir_function_impl *nir_cf_node_get_function(nir_cf_node *node);
2394
2395 /** requests that the given pieces of metadata be generated */
2396 void nir_metadata_require(nir_function_impl *impl, nir_metadata required, ...);
2397 /** dirties all but the preserved metadata */
2398 void nir_metadata_preserve(nir_function_impl *impl, nir_metadata preserved);
2399
2400 /** creates an instruction with default swizzle/writemask/etc. with NULL registers */
2401 nir_alu_instr *nir_alu_instr_create(nir_shader *shader, nir_op op);
2402
2403 nir_deref_instr *nir_deref_instr_create(nir_shader *shader,
2404 nir_deref_type deref_type);
2405
2406 nir_jump_instr *nir_jump_instr_create(nir_shader *shader, nir_jump_type type);
2407
2408 nir_load_const_instr *nir_load_const_instr_create(nir_shader *shader,
2409 unsigned num_components,
2410 unsigned bit_size);
2411
2412 nir_intrinsic_instr *nir_intrinsic_instr_create(nir_shader *shader,
2413 nir_intrinsic_op op);
2414
2415 nir_call_instr *nir_call_instr_create(nir_shader *shader,
2416 nir_function *callee);
2417
2418 nir_tex_instr *nir_tex_instr_create(nir_shader *shader, unsigned num_srcs);
2419
2420 nir_phi_instr *nir_phi_instr_create(nir_shader *shader);
2421
2422 nir_parallel_copy_instr *nir_parallel_copy_instr_create(nir_shader *shader);
2423
2424 nir_ssa_undef_instr *nir_ssa_undef_instr_create(nir_shader *shader,
2425 unsigned num_components,
2426 unsigned bit_size);
2427
2428 nir_const_value nir_alu_binop_identity(nir_op binop, unsigned bit_size);
2429
2430 /**
2431 * NIR Cursors and Instruction Insertion API
2432 * @{
2433 *
2434 * A tiny struct representing a point to insert/extract instructions or
2435 * control flow nodes. Helps reduce the combinatorial explosion of possible
2436 * points to insert/extract.
2437 *
2438 * \sa nir_control_flow.h
2439 */
2440 typedef enum {
2441 nir_cursor_before_block,
2442 nir_cursor_after_block,
2443 nir_cursor_before_instr,
2444 nir_cursor_after_instr,
2445 } nir_cursor_option;
2446
2447 typedef struct {
2448 nir_cursor_option option;
2449 union {
2450 nir_block *block;
2451 nir_instr *instr;
2452 };
2453 } nir_cursor;
2454
2455 static inline nir_block *
2456 nir_cursor_current_block(nir_cursor cursor)
2457 {
2458 if (cursor.option == nir_cursor_before_instr ||
2459 cursor.option == nir_cursor_after_instr) {
2460 return cursor.instr->block;
2461 } else {
2462 return cursor.block;
2463 }
2464 }
2465
2466 bool nir_cursors_equal(nir_cursor a, nir_cursor b);
2467
2468 static inline nir_cursor
2469 nir_before_block(nir_block *block)
2470 {
2471 nir_cursor cursor;
2472 cursor.option = nir_cursor_before_block;
2473 cursor.block = block;
2474 return cursor;
2475 }
2476
2477 static inline nir_cursor
2478 nir_after_block(nir_block *block)
2479 {
2480 nir_cursor cursor;
2481 cursor.option = nir_cursor_after_block;
2482 cursor.block = block;
2483 return cursor;
2484 }
2485
2486 static inline nir_cursor
2487 nir_before_instr(nir_instr *instr)
2488 {
2489 nir_cursor cursor;
2490 cursor.option = nir_cursor_before_instr;
2491 cursor.instr = instr;
2492 return cursor;
2493 }
2494
2495 static inline nir_cursor
2496 nir_after_instr(nir_instr *instr)
2497 {
2498 nir_cursor cursor;
2499 cursor.option = nir_cursor_after_instr;
2500 cursor.instr = instr;
2501 return cursor;
2502 }
2503
2504 static inline nir_cursor
2505 nir_after_block_before_jump(nir_block *block)
2506 {
2507 nir_instr *last_instr = nir_block_last_instr(block);
2508 if (last_instr && last_instr->type == nir_instr_type_jump) {
2509 return nir_before_instr(last_instr);
2510 } else {
2511 return nir_after_block(block);
2512 }
2513 }
2514
2515 static inline nir_cursor
2516 nir_before_src(nir_src *src, bool is_if_condition)
2517 {
2518 if (is_if_condition) {
2519 nir_block *prev_block =
2520 nir_cf_node_as_block(nir_cf_node_prev(&src->parent_if->cf_node));
2521 assert(!nir_block_ends_in_jump(prev_block));
2522 return nir_after_block(prev_block);
2523 } else if (src->parent_instr->type == nir_instr_type_phi) {
2524 #ifndef NDEBUG
2525 nir_phi_instr *cond_phi = nir_instr_as_phi(src->parent_instr);
2526 bool found = false;
2527 nir_foreach_phi_src(phi_src, cond_phi) {
2528 if (phi_src->src.ssa == src->ssa) {
2529 found = true;
2530 break;
2531 }
2532 }
2533 assert(found);
2534 #endif
2535 /* The LIST_ENTRY macro is a generic container-of macro, it just happens
2536 * to have a more specific name.
2537 */
2538 nir_phi_src *phi_src = LIST_ENTRY(nir_phi_src, src, src);
2539 return nir_after_block_before_jump(phi_src->pred);
2540 } else {
2541 return nir_before_instr(src->parent_instr);
2542 }
2543 }
2544
2545 static inline nir_cursor
2546 nir_before_cf_node(nir_cf_node *node)
2547 {
2548 if (node->type == nir_cf_node_block)
2549 return nir_before_block(nir_cf_node_as_block(node));
2550
2551 return nir_after_block(nir_cf_node_as_block(nir_cf_node_prev(node)));
2552 }
2553
2554 static inline nir_cursor
2555 nir_after_cf_node(nir_cf_node *node)
2556 {
2557 if (node->type == nir_cf_node_block)
2558 return nir_after_block(nir_cf_node_as_block(node));
2559
2560 return nir_before_block(nir_cf_node_as_block(nir_cf_node_next(node)));
2561 }
2562
2563 static inline nir_cursor
2564 nir_after_phis(nir_block *block)
2565 {
2566 nir_foreach_instr(instr, block) {
2567 if (instr->type != nir_instr_type_phi)
2568 return nir_before_instr(instr);
2569 }
2570 return nir_after_block(block);
2571 }
2572
2573 static inline nir_cursor
2574 nir_after_cf_node_and_phis(nir_cf_node *node)
2575 {
2576 if (node->type == nir_cf_node_block)
2577 return nir_after_block(nir_cf_node_as_block(node));
2578
2579 nir_block *block = nir_cf_node_as_block(nir_cf_node_next(node));
2580
2581 return nir_after_phis(block);
2582 }
2583
2584 static inline nir_cursor
2585 nir_before_cf_list(struct exec_list *cf_list)
2586 {
2587 nir_cf_node *first_node = exec_node_data(nir_cf_node,
2588 exec_list_get_head(cf_list), node);
2589 return nir_before_cf_node(first_node);
2590 }
2591
2592 static inline nir_cursor
2593 nir_after_cf_list(struct exec_list *cf_list)
2594 {
2595 nir_cf_node *last_node = exec_node_data(nir_cf_node,
2596 exec_list_get_tail(cf_list), node);
2597 return nir_after_cf_node(last_node);
2598 }
2599
2600 /**
2601 * Insert a NIR instruction at the given cursor.
2602 *
2603 * Note: This does not update the cursor.
2604 */
2605 void nir_instr_insert(nir_cursor cursor, nir_instr *instr);
2606
2607 static inline void
2608 nir_instr_insert_before(nir_instr *instr, nir_instr *before)
2609 {
2610 nir_instr_insert(nir_before_instr(instr), before);
2611 }
2612
2613 static inline void
2614 nir_instr_insert_after(nir_instr *instr, nir_instr *after)
2615 {
2616 nir_instr_insert(nir_after_instr(instr), after);
2617 }
2618
2619 static inline void
2620 nir_instr_insert_before_block(nir_block *block, nir_instr *before)
2621 {
2622 nir_instr_insert(nir_before_block(block), before);
2623 }
2624
2625 static inline void
2626 nir_instr_insert_after_block(nir_block *block, nir_instr *after)
2627 {
2628 nir_instr_insert(nir_after_block(block), after);
2629 }
2630
2631 static inline void
2632 nir_instr_insert_before_cf(nir_cf_node *node, nir_instr *before)
2633 {
2634 nir_instr_insert(nir_before_cf_node(node), before);
2635 }
2636
2637 static inline void
2638 nir_instr_insert_after_cf(nir_cf_node *node, nir_instr *after)
2639 {
2640 nir_instr_insert(nir_after_cf_node(node), after);
2641 }
2642
2643 static inline void
2644 nir_instr_insert_before_cf_list(struct exec_list *list, nir_instr *before)
2645 {
2646 nir_instr_insert(nir_before_cf_list(list), before);
2647 }
2648
2649 static inline void
2650 nir_instr_insert_after_cf_list(struct exec_list *list, nir_instr *after)
2651 {
2652 nir_instr_insert(nir_after_cf_list(list), after);
2653 }
2654
2655 void nir_instr_remove_v(nir_instr *instr);
2656
2657 static inline nir_cursor
2658 nir_instr_remove(nir_instr *instr)
2659 {
2660 nir_cursor cursor;
2661 nir_instr *prev = nir_instr_prev(instr);
2662 if (prev) {
2663 cursor = nir_after_instr(prev);
2664 } else {
2665 cursor = nir_before_block(instr->block);
2666 }
2667 nir_instr_remove_v(instr);
2668 return cursor;
2669 }
2670
2671 /** @} */
2672
2673 typedef bool (*nir_foreach_ssa_def_cb)(nir_ssa_def *def, void *state);
2674 typedef bool (*nir_foreach_dest_cb)(nir_dest *dest, void *state);
2675 typedef bool (*nir_foreach_src_cb)(nir_src *src, void *state);
2676 bool nir_foreach_ssa_def(nir_instr *instr, nir_foreach_ssa_def_cb cb,
2677 void *state);
2678 bool nir_foreach_dest(nir_instr *instr, nir_foreach_dest_cb cb, void *state);
2679 bool nir_foreach_src(nir_instr *instr, nir_foreach_src_cb cb, void *state);
2680
2681 nir_const_value *nir_src_as_const_value(nir_src src);
2682
2683 static inline struct nir_instr *
2684 nir_src_instr(const struct nir_src *src)
2685 {
2686 return src->is_ssa ? src->ssa->parent_instr : NULL;
2687 }
2688
2689 #define NIR_SRC_AS_(name, c_type, type_enum, cast_macro) \
2690 static inline c_type * \
2691 nir_src_as_ ## name (struct nir_src *src) \
2692 { \
2693 return src->is_ssa && src->ssa->parent_instr->type == type_enum \
2694 ? cast_macro(src->ssa->parent_instr) : NULL; \
2695 } \
2696 static inline const c_type * \
2697 nir_src_as_ ## name ## _const(const struct nir_src *src) \
2698 { \
2699 return src->is_ssa && src->ssa->parent_instr->type == type_enum \
2700 ? cast_macro(src->ssa->parent_instr) : NULL; \
2701 }
2702
2703 NIR_SRC_AS_(alu_instr, nir_alu_instr, nir_instr_type_alu, nir_instr_as_alu)
2704
2705 bool nir_src_is_dynamically_uniform(nir_src src);
2706 bool nir_srcs_equal(nir_src src1, nir_src src2);
2707 void nir_instr_rewrite_src(nir_instr *instr, nir_src *src, nir_src new_src);
2708 void nir_instr_move_src(nir_instr *dest_instr, nir_src *dest, nir_src *src);
2709 void nir_if_rewrite_condition(nir_if *if_stmt, nir_src new_src);
2710 void nir_instr_rewrite_dest(nir_instr *instr, nir_dest *dest,
2711 nir_dest new_dest);
2712
2713 void nir_ssa_dest_init(nir_instr *instr, nir_dest *dest,
2714 unsigned num_components, unsigned bit_size,
2715 const char *name);
2716 void nir_ssa_def_init(nir_instr *instr, nir_ssa_def *def,
2717 unsigned num_components, unsigned bit_size,
2718 const char *name);
2719 static inline void
2720 nir_ssa_dest_init_for_type(nir_instr *instr, nir_dest *dest,
2721 const struct glsl_type *type,
2722 const char *name)
2723 {
2724 assert(glsl_type_is_vector_or_scalar(type));
2725 nir_ssa_dest_init(instr, dest, glsl_get_components(type),
2726 glsl_get_bit_size(type), name);
2727 }
2728 void nir_ssa_def_rewrite_uses(nir_ssa_def *def, nir_src new_src);
2729 void nir_ssa_def_rewrite_uses_after(nir_ssa_def *def, nir_src new_src,
2730 nir_instr *after_me);
2731
2732 nir_component_mask_t nir_ssa_def_components_read(const nir_ssa_def *def);
2733
2734 /*
2735 * finds the next basic block in source-code order, returns NULL if there is
2736 * none
2737 */
2738
2739 nir_block *nir_block_cf_tree_next(nir_block *block);
2740
2741 /* Performs the opposite of nir_block_cf_tree_next() */
2742
2743 nir_block *nir_block_cf_tree_prev(nir_block *block);
2744
2745 /* Gets the first block in a CF node in source-code order */
2746
2747 nir_block *nir_cf_node_cf_tree_first(nir_cf_node *node);
2748
2749 /* Gets the last block in a CF node in source-code order */
2750
2751 nir_block *nir_cf_node_cf_tree_last(nir_cf_node *node);
2752
2753 /* Gets the next block after a CF node in source-code order */
2754
2755 nir_block *nir_cf_node_cf_tree_next(nir_cf_node *node);
2756
2757 /* Macros for loops that visit blocks in source-code order */
2758
2759 #define nir_foreach_block(block, impl) \
2760 for (nir_block *block = nir_start_block(impl); block != NULL; \
2761 block = nir_block_cf_tree_next(block))
2762
2763 #define nir_foreach_block_safe(block, impl) \
2764 for (nir_block *block = nir_start_block(impl), \
2765 *next = nir_block_cf_tree_next(block); \
2766 block != NULL; \
2767 block = next, next = nir_block_cf_tree_next(block))
2768
2769 #define nir_foreach_block_reverse(block, impl) \
2770 for (nir_block *block = nir_impl_last_block(impl); block != NULL; \
2771 block = nir_block_cf_tree_prev(block))
2772
2773 #define nir_foreach_block_reverse_safe(block, impl) \
2774 for (nir_block *block = nir_impl_last_block(impl), \
2775 *prev = nir_block_cf_tree_prev(block); \
2776 block != NULL; \
2777 block = prev, prev = nir_block_cf_tree_prev(block))
2778
2779 #define nir_foreach_block_in_cf_node(block, node) \
2780 for (nir_block *block = nir_cf_node_cf_tree_first(node); \
2781 block != nir_cf_node_cf_tree_next(node); \
2782 block = nir_block_cf_tree_next(block))
2783
2784 /* If the following CF node is an if, this function returns that if.
2785 * Otherwise, it returns NULL.
2786 */
2787 nir_if *nir_block_get_following_if(nir_block *block);
2788
2789 nir_loop *nir_block_get_following_loop(nir_block *block);
2790
2791 void nir_index_local_regs(nir_function_impl *impl);
2792 void nir_index_global_regs(nir_shader *shader);
2793 void nir_index_ssa_defs(nir_function_impl *impl);
2794 unsigned nir_index_instrs(nir_function_impl *impl);
2795
2796 void nir_index_blocks(nir_function_impl *impl);
2797
2798 void nir_print_shader(nir_shader *shader, FILE *fp);
2799 void nir_print_shader_annotated(nir_shader *shader, FILE *fp, struct hash_table *errors);
2800 void nir_print_instr(const nir_instr *instr, FILE *fp);
2801 void nir_print_deref(const nir_deref_instr *deref, FILE *fp);
2802
2803 nir_shader *nir_shader_clone(void *mem_ctx, const nir_shader *s);
2804 nir_function_impl *nir_function_impl_clone(const nir_function_impl *fi);
2805 nir_constant *nir_constant_clone(const nir_constant *c, nir_variable *var);
2806 nir_variable *nir_variable_clone(const nir_variable *c, nir_shader *shader);
2807
2808 nir_shader *nir_shader_serialize_deserialize(void *mem_ctx, nir_shader *s);
2809
2810 #ifndef NDEBUG
2811 void nir_validate_shader(nir_shader *shader, const char *when);
2812 void nir_metadata_set_validation_flag(nir_shader *shader);
2813 void nir_metadata_check_validation_flag(nir_shader *shader);
2814
2815 static inline bool
2816 should_skip_nir(const char *name)
2817 {
2818 static const char *list = NULL;
2819 if (!list) {
2820 /* Comma separated list of names to skip. */
2821 list = getenv("NIR_SKIP");
2822 if (!list)
2823 list = "";
2824 }
2825
2826 if (!list[0])
2827 return false;
2828
2829 return comma_separated_list_contains(list, name);
2830 }
2831
2832 static inline bool
2833 should_clone_nir(void)
2834 {
2835 static int should_clone = -1;
2836 if (should_clone < 0)
2837 should_clone = env_var_as_boolean("NIR_TEST_CLONE", false);
2838
2839 return should_clone;
2840 }
2841
2842 static inline bool
2843 should_serialize_deserialize_nir(void)
2844 {
2845 static int test_serialize = -1;
2846 if (test_serialize < 0)
2847 test_serialize = env_var_as_boolean("NIR_TEST_SERIALIZE", false);
2848
2849 return test_serialize;
2850 }
2851
2852 static inline bool
2853 should_print_nir(void)
2854 {
2855 static int should_print = -1;
2856 if (should_print < 0)
2857 should_print = env_var_as_boolean("NIR_PRINT", false);
2858
2859 return should_print;
2860 }
2861 #else
2862 static inline void nir_validate_shader(nir_shader *shader, const char *when) { (void) shader; (void)when; }
2863 static inline void nir_metadata_set_validation_flag(nir_shader *shader) { (void) shader; }
2864 static inline void nir_metadata_check_validation_flag(nir_shader *shader) { (void) shader; }
2865 static inline bool should_skip_nir(UNUSED const char *pass_name) { return false; }
2866 static inline bool should_clone_nir(void) { return false; }
2867 static inline bool should_serialize_deserialize_nir(void) { return false; }
2868 static inline bool should_print_nir(void) { return false; }
2869 #endif /* NDEBUG */
2870
2871 #define _PASS(pass, nir, do_pass) do { \
2872 if (should_skip_nir(#pass)) { \
2873 printf("skipping %s\n", #pass); \
2874 break; \
2875 } \
2876 do_pass \
2877 nir_validate_shader(nir, "after " #pass); \
2878 if (should_clone_nir()) { \
2879 nir_shader *clone = nir_shader_clone(ralloc_parent(nir), nir); \
2880 ralloc_free(nir); \
2881 nir = clone; \
2882 } \
2883 if (should_serialize_deserialize_nir()) { \
2884 void *mem_ctx = ralloc_parent(nir); \
2885 nir = nir_shader_serialize_deserialize(mem_ctx, nir); \
2886 } \
2887 } while (0)
2888
2889 #define NIR_PASS(progress, nir, pass, ...) _PASS(pass, nir, \
2890 nir_metadata_set_validation_flag(nir); \
2891 if (should_print_nir()) \
2892 printf("%s\n", #pass); \
2893 if (pass(nir, ##__VA_ARGS__)) { \
2894 progress = true; \
2895 if (should_print_nir()) \
2896 nir_print_shader(nir, stdout); \
2897 nir_metadata_check_validation_flag(nir); \
2898 } \
2899 )
2900
2901 #define NIR_PASS_V(nir, pass, ...) _PASS(pass, nir, \
2902 if (should_print_nir()) \
2903 printf("%s\n", #pass); \
2904 pass(nir, ##__VA_ARGS__); \
2905 if (should_print_nir()) \
2906 nir_print_shader(nir, stdout); \
2907 )
2908
2909 #define NIR_SKIP(name) should_skip_nir(#name)
2910
2911 void nir_calc_dominance_impl(nir_function_impl *impl);
2912 void nir_calc_dominance(nir_shader *shader);
2913
2914 nir_block *nir_dominance_lca(nir_block *b1, nir_block *b2);
2915 bool nir_block_dominates(nir_block *parent, nir_block *child);
2916
2917 void nir_dump_dom_tree_impl(nir_function_impl *impl, FILE *fp);
2918 void nir_dump_dom_tree(nir_shader *shader, FILE *fp);
2919
2920 void nir_dump_dom_frontier_impl(nir_function_impl *impl, FILE *fp);
2921 void nir_dump_dom_frontier(nir_shader *shader, FILE *fp);
2922
2923 void nir_dump_cfg_impl(nir_function_impl *impl, FILE *fp);
2924 void nir_dump_cfg(nir_shader *shader, FILE *fp);
2925
2926 int nir_gs_count_vertices(const nir_shader *shader);
2927
2928 bool nir_shrink_vec_array_vars(nir_shader *shader, nir_variable_mode modes);
2929 bool nir_split_array_vars(nir_shader *shader, nir_variable_mode modes);
2930 bool nir_split_var_copies(nir_shader *shader);
2931 bool nir_split_per_member_structs(nir_shader *shader);
2932 bool nir_split_struct_vars(nir_shader *shader, nir_variable_mode modes);
2933
2934 bool nir_lower_returns_impl(nir_function_impl *impl);
2935 bool nir_lower_returns(nir_shader *shader);
2936
2937 bool nir_inline_functions(nir_shader *shader);
2938
2939 bool nir_propagate_invariant(nir_shader *shader);
2940
2941 void nir_lower_var_copy_instr(nir_intrinsic_instr *copy, nir_shader *shader);
2942 void nir_lower_deref_copy_instr(struct nir_builder *b,
2943 nir_intrinsic_instr *copy);
2944 bool nir_lower_var_copies(nir_shader *shader);
2945
2946 void nir_fixup_deref_modes(nir_shader *shader);
2947
2948 bool nir_lower_global_vars_to_local(nir_shader *shader);
2949
2950 bool nir_lower_indirect_derefs(nir_shader *shader, nir_variable_mode modes);
2951
2952 bool nir_lower_locals_to_regs(nir_shader *shader);
2953
2954 void nir_lower_io_to_temporaries(nir_shader *shader,
2955 nir_function_impl *entrypoint,
2956 bool outputs, bool inputs);
2957
2958 void nir_shader_gather_info(nir_shader *shader, nir_function_impl *entrypoint);
2959
2960 void nir_assign_var_locations(struct exec_list *var_list, unsigned *size,
2961 int (*type_size)(const struct glsl_type *));
2962
2963 /* Some helpers to do very simple linking */
2964 bool nir_remove_unused_varyings(nir_shader *producer, nir_shader *consumer);
2965 bool nir_remove_unused_io_vars(nir_shader *shader, struct exec_list *var_list,
2966 uint64_t *used_by_other_stage,
2967 uint64_t *used_by_other_stage_patches);
2968 void nir_compact_varyings(nir_shader *producer, nir_shader *consumer,
2969 bool default_to_smooth_interp);
2970 void nir_link_xfb_varyings(nir_shader *producer, nir_shader *consumer);
2971 bool nir_link_opt_varyings(nir_shader *producer, nir_shader *consumer);
2972
2973 typedef enum {
2974 /* If set, this forces all non-flat fragment shader inputs to be
2975 * interpolated as if with the "sample" qualifier. This requires
2976 * nir_shader_compiler_options::use_interpolated_input_intrinsics.
2977 */
2978 nir_lower_io_force_sample_interpolation = (1 << 1),
2979 } nir_lower_io_options;
2980 bool nir_lower_io(nir_shader *shader,
2981 nir_variable_mode modes,
2982 int (*type_size)(const struct glsl_type *),
2983 nir_lower_io_options);
2984
2985 typedef enum {
2986 /**
2987 * An address format which is a simple 32-bit global GPU address.
2988 */
2989 nir_address_format_32bit_global,
2990
2991 /**
2992 * An address format which is a simple 64-bit global GPU address.
2993 */
2994 nir_address_format_64bit_global,
2995
2996 /**
2997 * An address format which is comprised of a vec2 where the first
2998 * component is a vulkan descriptor index and the second is an offset.
2999 */
3000 nir_address_format_vk_index_offset,
3001 } nir_address_format;
3002 bool nir_lower_explicit_io(nir_shader *shader,
3003 nir_variable_mode modes,
3004 nir_address_format);
3005
3006 nir_src *nir_get_io_offset_src(nir_intrinsic_instr *instr);
3007 nir_src *nir_get_io_vertex_index_src(nir_intrinsic_instr *instr);
3008
3009 bool nir_is_per_vertex_io(const nir_variable *var, gl_shader_stage stage);
3010
3011 bool nir_lower_regs_to_ssa_impl(nir_function_impl *impl);
3012 bool nir_lower_regs_to_ssa(nir_shader *shader);
3013 bool nir_lower_vars_to_ssa(nir_shader *shader);
3014
3015 bool nir_remove_dead_derefs(nir_shader *shader);
3016 bool nir_remove_dead_derefs_impl(nir_function_impl *impl);
3017 bool nir_remove_dead_variables(nir_shader *shader, nir_variable_mode modes);
3018 bool nir_lower_constant_initializers(nir_shader *shader,
3019 nir_variable_mode modes);
3020
3021 bool nir_move_load_const(nir_shader *shader);
3022 bool nir_move_vec_src_uses_to_dest(nir_shader *shader);
3023 bool nir_lower_vec_to_movs(nir_shader *shader);
3024 void nir_lower_alpha_test(nir_shader *shader, enum compare_func func,
3025 bool alpha_to_one);
3026 bool nir_lower_alu(nir_shader *shader);
3027 bool nir_lower_alu_to_scalar(nir_shader *shader);
3028 bool nir_lower_bool_to_float(nir_shader *shader);
3029 bool nir_lower_bool_to_int32(nir_shader *shader);
3030 bool nir_lower_load_const_to_scalar(nir_shader *shader);
3031 bool nir_lower_read_invocation_to_scalar(nir_shader *shader);
3032 bool nir_lower_phis_to_scalar(nir_shader *shader);
3033 void nir_lower_io_arrays_to_elements(nir_shader *producer, nir_shader *consumer);
3034 void nir_lower_io_arrays_to_elements_no_indirects(nir_shader *shader,
3035 bool outputs_only);
3036 void nir_lower_io_to_scalar(nir_shader *shader, nir_variable_mode mask);
3037 void nir_lower_io_to_scalar_early(nir_shader *shader, nir_variable_mode mask);
3038
3039 typedef struct nir_lower_subgroups_options {
3040 uint8_t subgroup_size;
3041 uint8_t ballot_bit_size;
3042 bool lower_to_scalar:1;
3043 bool lower_vote_trivial:1;
3044 bool lower_vote_eq_to_ballot:1;
3045 bool lower_subgroup_masks:1;
3046 bool lower_shuffle:1;
3047 bool lower_shuffle_to_32bit:1;
3048 bool lower_quad:1;
3049 } nir_lower_subgroups_options;
3050
3051 bool nir_lower_subgroups(nir_shader *shader,
3052 const nir_lower_subgroups_options *options);
3053
3054 bool nir_lower_system_values(nir_shader *shader);
3055
3056 enum PACKED nir_lower_tex_packing {
3057 nir_lower_tex_packing_none = 0,
3058 /* The sampler returns up to 2 32-bit words of half floats or 16-bit signed
3059 * or unsigned ints based on the sampler type
3060 */
3061 nir_lower_tex_packing_16,
3062 /* The sampler returns 1 32-bit word of 4x8 unorm */
3063 nir_lower_tex_packing_8,
3064 };
3065
3066 typedef struct nir_lower_tex_options {
3067 /**
3068 * bitmask of (1 << GLSL_SAMPLER_DIM_x) to control for which
3069 * sampler types a texture projector is lowered.
3070 */
3071 unsigned lower_txp;
3072
3073 /**
3074 * If true, lower away nir_tex_src_offset for all texelfetch instructions.
3075 */
3076 bool lower_txf_offset;
3077
3078 /**
3079 * If true, lower away nir_tex_src_offset for all rect textures.
3080 */
3081 bool lower_rect_offset;
3082
3083 /**
3084 * If true, lower rect textures to 2D, using txs to fetch the
3085 * texture dimensions and dividing the texture coords by the
3086 * texture dims to normalize.
3087 */
3088 bool lower_rect;
3089
3090 /**
3091 * If true, convert yuv to rgb.
3092 */
3093 unsigned lower_y_uv_external;
3094 unsigned lower_y_u_v_external;
3095 unsigned lower_yx_xuxv_external;
3096 unsigned lower_xy_uxvx_external;
3097 unsigned lower_ayuv_external;
3098 unsigned lower_xyuv_external;
3099
3100 /**
3101 * To emulate certain texture wrap modes, this can be used
3102 * to saturate the specified tex coord to [0.0, 1.0]. The
3103 * bits are according to sampler #, ie. if, for example:
3104 *
3105 * (conf->saturate_s & (1 << n))
3106 *
3107 * is true, then the s coord for sampler n is saturated.
3108 *
3109 * Note that clamping must happen *after* projector lowering
3110 * so any projected texture sample instruction with a clamped
3111 * coordinate gets automatically lowered, regardless of the
3112 * 'lower_txp' setting.
3113 */
3114 unsigned saturate_s;
3115 unsigned saturate_t;
3116 unsigned saturate_r;
3117
3118 /* Bitmask of textures that need swizzling.
3119 *
3120 * If (swizzle_result & (1 << texture_index)), then the swizzle in
3121 * swizzles[texture_index] is applied to the result of the texturing
3122 * operation.
3123 */
3124 unsigned swizzle_result;
3125
3126 /* A swizzle for each texture. Values 0-3 represent x, y, z, or w swizzles
3127 * while 4 and 5 represent 0 and 1 respectively.
3128 */
3129 uint8_t swizzles[32][4];
3130
3131 /* Can be used to scale sampled values in range required by the format. */
3132 float scale_factors[32];
3133
3134 /**
3135 * Bitmap of textures that need srgb to linear conversion. If
3136 * (lower_srgb & (1 << texture_index)) then the rgb (xyz) components
3137 * of the texture are lowered to linear.
3138 */
3139 unsigned lower_srgb;
3140
3141 /**
3142 * If true, lower nir_texop_txd on cube maps with nir_texop_txl.
3143 */
3144 bool lower_txd_cube_map;
3145
3146 /**
3147 * If true, lower nir_texop_txd on 3D surfaces with nir_texop_txl.
3148 */
3149 bool lower_txd_3d;
3150
3151 /**
3152 * If true, lower nir_texop_txd on shadow samplers (except cube maps)
3153 * with nir_texop_txl. Notice that cube map shadow samplers are lowered
3154 * with lower_txd_cube_map.
3155 */
3156 bool lower_txd_shadow;
3157
3158 /**
3159 * If true, lower nir_texop_txd on all samplers to a nir_texop_txl.
3160 * Implies lower_txd_cube_map and lower_txd_shadow.
3161 */
3162 bool lower_txd;
3163
3164 /**
3165 * If true, lower nir_texop_txb that try to use shadow compare and min_lod
3166 * at the same time to a nir_texop_lod, some math, and nir_texop_tex.
3167 */
3168 bool lower_txb_shadow_clamp;
3169
3170 /**
3171 * If true, lower nir_texop_txd on shadow samplers when it uses min_lod
3172 * with nir_texop_txl. This includes cube maps.
3173 */
3174 bool lower_txd_shadow_clamp;
3175
3176 /**
3177 * If true, lower nir_texop_txd on when it uses both offset and min_lod
3178 * with nir_texop_txl. This includes cube maps.
3179 */
3180 bool lower_txd_offset_clamp;
3181
3182 /**
3183 * If true, apply a .bagr swizzle on tg4 results to handle Broadcom's
3184 * mixed-up tg4 locations.
3185 */
3186 bool lower_tg4_broadcom_swizzle;
3187
3188 enum nir_lower_tex_packing lower_tex_packing[32];
3189 } nir_lower_tex_options;
3190
3191 bool nir_lower_tex(nir_shader *shader,
3192 const nir_lower_tex_options *options);
3193
3194 bool nir_lower_idiv(nir_shader *shader);
3195
3196 bool nir_lower_clip_vs(nir_shader *shader, unsigned ucp_enables, bool use_vars);
3197 bool nir_lower_clip_fs(nir_shader *shader, unsigned ucp_enables);
3198 bool nir_lower_clip_cull_distance_arrays(nir_shader *nir);
3199
3200 void nir_lower_two_sided_color(nir_shader *shader);
3201
3202 bool nir_lower_clamp_color_outputs(nir_shader *shader);
3203
3204 void nir_lower_passthrough_edgeflags(nir_shader *shader);
3205 bool nir_lower_patch_vertices(nir_shader *nir, unsigned static_count,
3206 const gl_state_index16 *uniform_state_tokens);
3207
3208 typedef struct nir_lower_wpos_ytransform_options {
3209 gl_state_index16 state_tokens[STATE_LENGTH];
3210 bool fs_coord_origin_upper_left :1;
3211 bool fs_coord_origin_lower_left :1;
3212 bool fs_coord_pixel_center_integer :1;
3213 bool fs_coord_pixel_center_half_integer :1;
3214 } nir_lower_wpos_ytransform_options;
3215
3216 bool nir_lower_wpos_ytransform(nir_shader *shader,
3217 const nir_lower_wpos_ytransform_options *options);
3218 bool nir_lower_wpos_center(nir_shader *shader, const bool for_sample_shading);
3219
3220 typedef struct nir_lower_drawpixels_options {
3221 gl_state_index16 texcoord_state_tokens[STATE_LENGTH];
3222 gl_state_index16 scale_state_tokens[STATE_LENGTH];
3223 gl_state_index16 bias_state_tokens[STATE_LENGTH];
3224 unsigned drawpix_sampler;
3225 unsigned pixelmap_sampler;
3226 bool pixel_maps :1;
3227 bool scale_and_bias :1;
3228 } nir_lower_drawpixels_options;
3229
3230 void nir_lower_drawpixels(nir_shader *shader,
3231 const nir_lower_drawpixels_options *options);
3232
3233 typedef struct nir_lower_bitmap_options {
3234 unsigned sampler;
3235 bool swizzle_xxxx;
3236 } nir_lower_bitmap_options;
3237
3238 void nir_lower_bitmap(nir_shader *shader, const nir_lower_bitmap_options *options);
3239
3240 bool nir_lower_atomics_to_ssbo(nir_shader *shader, unsigned ssbo_offset);
3241
3242 typedef enum {
3243 nir_lower_int_source_mods = 1 << 0,
3244 nir_lower_float_source_mods = 1 << 1,
3245 nir_lower_triop_abs = 1 << 2,
3246 nir_lower_all_source_mods = (1 << 3) - 1
3247 } nir_lower_to_source_mods_flags;
3248
3249
3250 bool nir_lower_to_source_mods(nir_shader *shader, nir_lower_to_source_mods_flags options);
3251
3252 bool nir_lower_gs_intrinsics(nir_shader *shader);
3253
3254 typedef unsigned (*nir_lower_bit_size_callback)(const nir_alu_instr *, void *);
3255
3256 bool nir_lower_bit_size(nir_shader *shader,
3257 nir_lower_bit_size_callback callback,
3258 void *callback_data);
3259
3260 bool nir_lower_int64(nir_shader *shader, nir_lower_int64_options options);
3261
3262 bool nir_lower_doubles(nir_shader *shader, nir_lower_doubles_options options);
3263 bool nir_lower_pack(nir_shader *shader);
3264
3265 bool nir_normalize_cubemap_coords(nir_shader *shader);
3266
3267 void nir_live_ssa_defs_impl(nir_function_impl *impl);
3268
3269 void nir_loop_analyze_impl(nir_function_impl *impl,
3270 nir_variable_mode indirect_mask);
3271
3272 bool nir_ssa_defs_interfere(nir_ssa_def *a, nir_ssa_def *b);
3273
3274 bool nir_repair_ssa_impl(nir_function_impl *impl);
3275 bool nir_repair_ssa(nir_shader *shader);
3276
3277 void nir_convert_loop_to_lcssa(nir_loop *loop);
3278
3279 /* If phi_webs_only is true, only convert SSA values involved in phi nodes to
3280 * registers. If false, convert all values (even those not involved in a phi
3281 * node) to registers.
3282 */
3283 bool nir_convert_from_ssa(nir_shader *shader, bool phi_webs_only);
3284
3285 bool nir_lower_phis_to_regs_block(nir_block *block);
3286 bool nir_lower_ssa_defs_to_regs_block(nir_block *block);
3287 bool nir_rematerialize_derefs_in_use_blocks_impl(nir_function_impl *impl);
3288
3289 bool nir_opt_algebraic(nir_shader *shader);
3290 bool nir_opt_algebraic_before_ffma(nir_shader *shader);
3291 bool nir_opt_algebraic_late(nir_shader *shader);
3292 bool nir_opt_constant_folding(nir_shader *shader);
3293
3294 bool nir_opt_global_to_local(nir_shader *shader);
3295
3296 bool nir_copy_prop(nir_shader *shader);
3297
3298 bool nir_opt_copy_prop_vars(nir_shader *shader);
3299
3300 bool nir_opt_cse(nir_shader *shader);
3301
3302 bool nir_opt_dce(nir_shader *shader);
3303
3304 bool nir_opt_dead_cf(nir_shader *shader);
3305
3306 bool nir_opt_dead_write_vars(nir_shader *shader);
3307
3308 bool nir_opt_deref(nir_shader *shader);
3309
3310 bool nir_opt_find_array_copies(nir_shader *shader);
3311
3312 bool nir_opt_gcm(nir_shader *shader, bool value_number);
3313
3314 bool nir_opt_idiv_const(nir_shader *shader, unsigned min_bit_size);
3315
3316 bool nir_opt_if(nir_shader *shader);
3317
3318 bool nir_opt_intrinsics(nir_shader *shader);
3319
3320 bool nir_opt_large_constants(nir_shader *shader,
3321 glsl_type_size_align_func size_align,
3322 unsigned threshold);
3323
3324 bool nir_opt_loop_unroll(nir_shader *shader, nir_variable_mode indirect_mask);
3325
3326 bool nir_opt_move_comparisons(nir_shader *shader);
3327
3328 bool nir_opt_move_load_ubo(nir_shader *shader);
3329
3330 bool nir_opt_peephole_select(nir_shader *shader, unsigned limit,
3331 bool indirect_load_ok, bool expensive_alu_ok);
3332
3333 bool nir_opt_remove_phis(nir_shader *shader);
3334
3335 bool nir_opt_shrink_load(nir_shader *shader);
3336
3337 bool nir_opt_trivial_continues(nir_shader *shader);
3338
3339 bool nir_opt_undef(nir_shader *shader);
3340
3341 bool nir_opt_conditional_discard(nir_shader *shader);
3342
3343 void nir_sweep(nir_shader *shader);
3344
3345 void nir_remap_dual_slot_attributes(nir_shader *shader,
3346 uint64_t *dual_slot_inputs);
3347 uint64_t nir_get_single_slot_attribs_mask(uint64_t attribs, uint64_t dual_slot);
3348
3349 nir_intrinsic_op nir_intrinsic_from_system_value(gl_system_value val);
3350 gl_system_value nir_system_value_from_intrinsic(nir_intrinsic_op intrin);
3351
3352 #ifdef __cplusplus
3353 } /* extern "C" */
3354 #endif
3355
3356 #endif /* NIR_H */