freedreno/ir3: add macro to declare variable length arrays
[mesa.git] / src / gallium / drivers / freedreno / ir3 / ir3.h
1 /*
2 * Copyright (c) 2013 Rob Clark <robdclark@gmail.com>
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 */
23
24 #ifndef IR3_H_
25 #define IR3_H_
26
27 #include <stdint.h>
28 #include <stdbool.h>
29
30 #include "util/u_debug.h"
31 #include "util/list.h"
32
33 #include "instr-a3xx.h"
34 #include "disasm.h" /* TODO move 'enum shader_t' somewhere else.. */
35
36 /* low level intermediate representation of an adreno shader program */
37
38 struct ir3_compiler;
39 struct ir3;
40 struct ir3_instruction;
41 struct ir3_block;
42
43 struct ir3_info {
44 uint32_t gpu_id;
45 uint16_t sizedwords;
46 uint16_t instrs_count; /* expanded to account for rpt's */
47 /* NOTE: max_reg, etc, does not include registers not touched
48 * by the shader (ie. vertex fetched via VFD_DECODE but not
49 * touched by shader)
50 */
51 int8_t max_reg; /* highest GPR # used by shader */
52 int8_t max_half_reg;
53 int16_t max_const;
54 };
55
56 struct ir3_register {
57 enum {
58 IR3_REG_CONST = 0x001,
59 IR3_REG_IMMED = 0x002,
60 IR3_REG_HALF = 0x004,
61 /* high registers are used for some things in compute shaders,
62 * for example. Seems to be for things that are global to all
63 * threads in a wave, so possibly these are global/shared by
64 * all the threads in the wave?
65 */
66 IR3_REG_HIGH = 0x008,
67 IR3_REG_RELATIV= 0x010,
68 IR3_REG_R = 0x020,
69 /* Most instructions, it seems, can do float abs/neg but not
70 * integer. The CP pass needs to know what is intended (int or
71 * float) in order to do the right thing. For this reason the
72 * abs/neg flags are split out into float and int variants. In
73 * addition, .b (bitwise) operations, the negate is actually a
74 * bitwise not, so split that out into a new flag to make it
75 * more clear.
76 */
77 IR3_REG_FNEG = 0x040,
78 IR3_REG_FABS = 0x080,
79 IR3_REG_SNEG = 0x100,
80 IR3_REG_SABS = 0x200,
81 IR3_REG_BNOT = 0x400,
82 IR3_REG_EVEN = 0x800,
83 IR3_REG_POS_INF= 0x1000,
84 /* (ei) flag, end-input? Set on last bary, presumably to signal
85 * that the shader needs no more input:
86 */
87 IR3_REG_EI = 0x2000,
88 /* meta-flags, for intermediate stages of IR, ie.
89 * before register assignment is done:
90 */
91 IR3_REG_SSA = 0x4000, /* 'instr' is ptr to assigning instr */
92 IR3_REG_ARRAY = 0x8000,
93 IR3_REG_PHI_SRC= 0x10000, /* phi src, regs[0]->instr points to phi */
94
95 } flags;
96 union {
97 /* normal registers:
98 * the component is in the low two bits of the reg #, so
99 * rN.x becomes: (N << 2) | x
100 */
101 int num;
102 /* immediate: */
103 int32_t iim_val;
104 uint32_t uim_val;
105 float fim_val;
106 /* relative: */
107 struct {
108 uint16_t id;
109 int16_t offset;
110 } array;
111 };
112
113 /* For IR3_REG_SSA, src registers contain ptr back to assigning
114 * instruction.
115 *
116 * For IR3_REG_ARRAY, the pointer is back to the last dependent
117 * array access (although the net effect is the same, it points
118 * back to a previous instruction that we depend on).
119 */
120 struct ir3_instruction *instr;
121
122 union {
123 /* used for cat5 instructions, but also for internal/IR level
124 * tracking of what registers are read/written by an instruction.
125 * wrmask may be a bad name since it is used to represent both
126 * src and dst that touch multiple adjacent registers.
127 */
128 unsigned wrmask;
129 /* for relative addressing, 32bits for array size is too small,
130 * but otoh we don't need to deal with disjoint sets, so instead
131 * use a simple size field (number of scalar components).
132 */
133 unsigned size;
134 };
135 };
136
137 struct ir3_instruction {
138 struct ir3_block *block;
139 opc_t opc;
140 enum {
141 /* (sy) flag is set on first instruction, and after sample
142 * instructions (probably just on RAW hazard).
143 */
144 IR3_INSTR_SY = 0x001,
145 /* (ss) flag is set on first instruction, and first instruction
146 * to depend on the result of "long" instructions (RAW hazard):
147 *
148 * rcp, rsq, log2, exp2, sin, cos, sqrt
149 *
150 * It seems to synchronize until all in-flight instructions are
151 * completed, for example:
152 *
153 * rsq hr1.w, hr1.w
154 * add.f hr2.z, (neg)hr2.z, hc0.y
155 * mul.f hr2.w, (neg)hr2.y, (neg)hr2.y
156 * rsq hr2.x, hr2.x
157 * (rpt1)nop
158 * mad.f16 hr2.w, hr2.z, hr2.z, hr2.w
159 * nop
160 * mad.f16 hr2.w, (neg)hr0.w, (neg)hr0.w, hr2.w
161 * (ss)(rpt2)mul.f hr1.x, (r)hr1.x, hr1.w
162 * (rpt2)mul.f hr0.x, (neg)(r)hr0.x, hr2.x
163 *
164 * The last mul.f does not have (ss) set, presumably because the
165 * (ss) on the previous instruction does the job.
166 *
167 * The blob driver also seems to set it on WAR hazards, although
168 * not really clear if this is needed or just blob compiler being
169 * sloppy. So far I haven't found a case where removing the (ss)
170 * causes problems for WAR hazard, but I could just be getting
171 * lucky:
172 *
173 * rcp r1.y, r3.y
174 * (ss)(rpt2)mad.f32 r3.y, (r)c9.x, r1.x, (r)r3.z
175 *
176 */
177 IR3_INSTR_SS = 0x002,
178 /* (jp) flag is set on jump targets:
179 */
180 IR3_INSTR_JP = 0x004,
181 IR3_INSTR_UL = 0x008,
182 IR3_INSTR_3D = 0x010,
183 IR3_INSTR_A = 0x020,
184 IR3_INSTR_O = 0x040,
185 IR3_INSTR_P = 0x080,
186 IR3_INSTR_S = 0x100,
187 IR3_INSTR_S2EN = 0x200,
188 IR3_INSTR_G = 0x400,
189 /* meta-flags, for intermediate stages of IR, ie.
190 * before register assignment is done:
191 */
192 IR3_INSTR_MARK = 0x1000,
193 IR3_INSTR_UNUSED= 0x2000,
194 } flags;
195 int repeat;
196 #ifdef DEBUG
197 unsigned regs_max;
198 #endif
199 unsigned regs_count;
200 struct ir3_register **regs;
201 union {
202 struct {
203 char inv;
204 char comp;
205 int immed;
206 struct ir3_block *target;
207 } cat0;
208 struct {
209 type_t src_type, dst_type;
210 } cat1;
211 struct {
212 enum {
213 IR3_COND_LT = 0,
214 IR3_COND_LE = 1,
215 IR3_COND_GT = 2,
216 IR3_COND_GE = 3,
217 IR3_COND_EQ = 4,
218 IR3_COND_NE = 5,
219 } condition;
220 } cat2;
221 struct {
222 unsigned samp, tex;
223 type_t type;
224 } cat5;
225 struct {
226 type_t type;
227 int src_offset;
228 int dst_offset;
229 int iim_val; /* for ldgb/stgb, # of components */
230 } cat6;
231 /* for meta-instructions, just used to hold extra data
232 * before instruction scheduling, etc
233 */
234 struct {
235 int off; /* component/offset */
236 } fo;
237 struct {
238 /* used to temporarily hold reference to nir_phi_instr
239 * until we resolve the phi srcs
240 */
241 void *nphi;
242 } phi;
243 struct {
244 struct ir3_block *block;
245 } inout;
246 };
247
248 /* transient values used during various algorithms: */
249 union {
250 /* The instruction depth is the max dependency distance to output.
251 *
252 * You can also think of it as the "cost", if we did any sort of
253 * optimization for register footprint. Ie. a value that is just
254 * result of moving a const to a reg would have a low cost, so to
255 * it could make sense to duplicate the instruction at various
256 * points where the result is needed to reduce register footprint.
257 */
258 unsigned depth;
259 /* When we get to the RA stage, we no longer need depth, but
260 * we do need instruction's position/name:
261 */
262 struct {
263 uint16_t ip;
264 uint16_t name;
265 };
266 };
267
268 /* used for per-pass extra instruction data.
269 */
270 void *data;
271
272 /* Used during CP and RA stages. For fanin and shader inputs/
273 * outputs where we need a sequence of consecutive registers,
274 * keep track of each src instructions left (ie 'n-1') and right
275 * (ie 'n+1') neighbor. The front-end must insert enough mov's
276 * to ensure that each instruction has at most one left and at
277 * most one right neighbor. During the copy-propagation pass,
278 * we only remove mov's when we can preserve this constraint.
279 * And during the RA stage, we use the neighbor information to
280 * allocate a block of registers in one shot.
281 *
282 * TODO: maybe just add something like:
283 * struct ir3_instruction_ref {
284 * struct ir3_instruction *instr;
285 * unsigned cnt;
286 * }
287 *
288 * Or can we get away without the refcnt stuff? It seems like
289 * it should be overkill.. the problem is if, potentially after
290 * already eliminating some mov's, if you have a single mov that
291 * needs to be grouped with it's neighbors in two different
292 * places (ex. shader output and a fanin).
293 */
294 struct {
295 struct ir3_instruction *left, *right;
296 uint16_t left_cnt, right_cnt;
297 } cp;
298
299 /* an instruction can reference at most one address register amongst
300 * it's src/dst registers. Beyond that, you need to insert mov's.
301 *
302 * NOTE: do not write this directly, use ir3_instr_set_address()
303 */
304 struct ir3_instruction *address;
305
306 /* Entry in ir3_block's instruction list: */
307 struct list_head node;
308
309 #ifdef DEBUG
310 uint32_t serialno;
311 #endif
312 };
313
314 static inline struct ir3_instruction *
315 ir3_neighbor_first(struct ir3_instruction *instr)
316 {
317 int cnt = 0;
318 while (instr->cp.left) {
319 instr = instr->cp.left;
320 if (++cnt > 0xffff) {
321 debug_assert(0);
322 break;
323 }
324 }
325 return instr;
326 }
327
328 static inline int ir3_neighbor_count(struct ir3_instruction *instr)
329 {
330 int num = 1;
331
332 debug_assert(!instr->cp.left);
333
334 while (instr->cp.right) {
335 num++;
336 instr = instr->cp.right;
337 if (num > 0xffff) {
338 debug_assert(0);
339 break;
340 }
341 }
342
343 return num;
344 }
345
346 /*
347 * Stupid/simple growable array implementation:
348 */
349 #define DECLARE_ARRAY(type, name) \
350 unsigned name ## _count, name ## _sz; \
351 type * name;
352
353 #define array_insert(ctx, arr, val) do { \
354 if (arr ## _count == arr ## _sz) { \
355 arr ## _sz = MAX2(2 * arr ## _sz, 16); \
356 arr = reralloc_size(ctx, arr, arr ## _sz * sizeof(arr[0])); \
357 } \
358 arr[arr ##_count++] = val; \
359 } while (0)
360
361 struct ir3 {
362 struct ir3_compiler *compiler;
363
364 unsigned ninputs, noutputs;
365 struct ir3_instruction **inputs;
366 struct ir3_instruction **outputs;
367
368 /* Track bary.f (and ldlv) instructions.. this is needed in
369 * scheduling to ensure that all varying fetches happen before
370 * any potential kill instructions. The hw gets grumpy if all
371 * threads in a group are killed before the last bary.f gets
372 * a chance to signal end of input (ei).
373 */
374 DECLARE_ARRAY(struct ir3_instruction *, baryfs);
375
376 /* Track all indirect instructions (read and write). To avoid
377 * deadlock scenario where an address register gets scheduled,
378 * but other dependent src instructions cannot be scheduled due
379 * to dependency on a *different* address register value, the
380 * scheduler needs to ensure that all dependencies other than
381 * the instruction other than the address register are scheduled
382 * before the one that writes the address register. Having a
383 * convenient list of instructions that reference some address
384 * register simplifies this.
385 */
386 DECLARE_ARRAY(struct ir3_instruction *, indirects);
387
388 /* and same for instructions that consume predicate register: */
389 DECLARE_ARRAY(struct ir3_instruction *, predicates);
390
391 /* Track texture sample instructions which need texture state
392 * patched in (for astc-srgb workaround):
393 */
394 DECLARE_ARRAY(struct ir3_instruction *, astc_srgb);
395
396 /* List of blocks: */
397 struct list_head block_list;
398
399 /* List of ir3_array's: */
400 struct list_head array_list;
401 };
402
403 typedef struct nir_variable nir_variable;
404
405 struct ir3_array {
406 struct list_head node;
407 unsigned length;
408 unsigned id;
409
410 nir_variable *var;
411
412 /* We track the last write and last access (read or write) to
413 * setup dependencies on instructions that read or write the
414 * array. Reads can be re-ordered wrt. other reads, but should
415 * not be re-ordered wrt. to writes. Writes cannot be reordered
416 * wrt. any other access to the array.
417 *
418 * So array reads depend on last write, and array writes depend
419 * on the last access.
420 */
421 struct ir3_instruction *last_write, *last_access;
422
423 /* extra stuff used in RA pass: */
424 unsigned base; /* base vreg name */
425 unsigned reg; /* base physical reg */
426 uint16_t start_ip, end_ip;
427 };
428
429 struct ir3_array * ir3_lookup_array(struct ir3 *ir, unsigned id);
430
431 typedef struct nir_block nir_block;
432
433 struct ir3_block {
434 struct list_head node;
435 struct ir3 *shader;
436
437 nir_block *nblock;
438
439 struct list_head instr_list; /* list of ir3_instruction */
440
441 /* each block has either one or two successors.. in case of
442 * two successors, 'condition' decides which one to follow.
443 * A block preceding an if/else has two successors.
444 */
445 struct ir3_instruction *condition;
446 struct ir3_block *successors[2];
447
448 uint16_t start_ip, end_ip;
449
450 /* Track instructions which do not write a register but other-
451 * wise must not be discarded (such as kill, stg, etc)
452 */
453 DECLARE_ARRAY(struct ir3_instruction *, keeps);
454
455 /* used for per-pass extra block data. Mainly used right
456 * now in RA step to track livein/liveout.
457 */
458 void *data;
459
460 #ifdef DEBUG
461 uint32_t serialno;
462 #endif
463 };
464
465 static inline uint32_t
466 block_id(struct ir3_block *block)
467 {
468 #ifdef DEBUG
469 return block->serialno;
470 #else
471 return (uint32_t)(unsigned long)block;
472 #endif
473 }
474
475 struct ir3 * ir3_create(struct ir3_compiler *compiler,
476 unsigned nin, unsigned nout);
477 void ir3_destroy(struct ir3 *shader);
478 void * ir3_assemble(struct ir3 *shader,
479 struct ir3_info *info, uint32_t gpu_id);
480 void * ir3_alloc(struct ir3 *shader, int sz);
481
482 struct ir3_block * ir3_block_create(struct ir3 *shader);
483
484 struct ir3_instruction * ir3_instr_create(struct ir3_block *block, opc_t opc);
485 struct ir3_instruction * ir3_instr_create2(struct ir3_block *block,
486 opc_t opc, int nreg);
487 struct ir3_instruction * ir3_instr_clone(struct ir3_instruction *instr);
488 const char *ir3_instr_name(struct ir3_instruction *instr);
489
490 struct ir3_register * ir3_reg_create(struct ir3_instruction *instr,
491 int num, int flags);
492 struct ir3_register * ir3_reg_clone(struct ir3 *shader,
493 struct ir3_register *reg);
494
495 void ir3_instr_set_address(struct ir3_instruction *instr,
496 struct ir3_instruction *addr);
497
498 static inline bool ir3_instr_check_mark(struct ir3_instruction *instr)
499 {
500 if (instr->flags & IR3_INSTR_MARK)
501 return true; /* already visited */
502 instr->flags |= IR3_INSTR_MARK;
503 return false;
504 }
505
506 void ir3_block_clear_mark(struct ir3_block *block);
507 void ir3_clear_mark(struct ir3 *shader);
508
509 unsigned ir3_count_instructions(struct ir3 *ir);
510
511 static inline int ir3_instr_regno(struct ir3_instruction *instr,
512 struct ir3_register *reg)
513 {
514 unsigned i;
515 for (i = 0; i < instr->regs_count; i++)
516 if (reg == instr->regs[i])
517 return i;
518 return -1;
519 }
520
521
522 #define MAX_ARRAYS 16
523
524 /* comp:
525 * 0 - x
526 * 1 - y
527 * 2 - z
528 * 3 - w
529 */
530 static inline uint32_t regid(int num, int comp)
531 {
532 return (num << 2) | (comp & 0x3);
533 }
534
535 static inline uint32_t reg_num(struct ir3_register *reg)
536 {
537 return reg->num >> 2;
538 }
539
540 static inline uint32_t reg_comp(struct ir3_register *reg)
541 {
542 return reg->num & 0x3;
543 }
544
545 static inline bool is_flow(struct ir3_instruction *instr)
546 {
547 return (opc_cat(instr->opc) == 0);
548 }
549
550 static inline bool is_kill(struct ir3_instruction *instr)
551 {
552 return instr->opc == OPC_KILL;
553 }
554
555 static inline bool is_nop(struct ir3_instruction *instr)
556 {
557 return instr->opc == OPC_NOP;
558 }
559
560 /* Is it a non-transformative (ie. not type changing) mov? This can
561 * also include absneg.s/absneg.f, which for the most part can be
562 * treated as a mov (single src argument).
563 */
564 static inline bool is_same_type_mov(struct ir3_instruction *instr)
565 {
566 struct ir3_register *dst = instr->regs[0];
567
568 /* mov's that write to a0.x or p0.x are special: */
569 if (dst->num == regid(REG_P0, 0))
570 return false;
571 if (dst->num == regid(REG_A0, 0))
572 return false;
573
574 if (dst->flags & (IR3_REG_RELATIV | IR3_REG_ARRAY))
575 return false;
576
577 switch (instr->opc) {
578 case OPC_MOV:
579 return instr->cat1.src_type == instr->cat1.dst_type;
580 case OPC_ABSNEG_F:
581 case OPC_ABSNEG_S:
582 return true;
583 default:
584 return false;
585 }
586 }
587
588 static inline bool is_alu(struct ir3_instruction *instr)
589 {
590 return (1 <= opc_cat(instr->opc)) && (opc_cat(instr->opc) <= 3);
591 }
592
593 static inline bool is_sfu(struct ir3_instruction *instr)
594 {
595 return (opc_cat(instr->opc) == 4);
596 }
597
598 static inline bool is_tex(struct ir3_instruction *instr)
599 {
600 return (opc_cat(instr->opc) == 5);
601 }
602
603 static inline bool is_mem(struct ir3_instruction *instr)
604 {
605 return (opc_cat(instr->opc) == 6);
606 }
607
608 static inline bool
609 is_store(struct ir3_instruction *instr)
610 {
611 /* these instructions, the "destination" register is
612 * actually a source, the address to store to.
613 */
614 switch (instr->opc) {
615 case OPC_STG:
616 case OPC_STGB:
617 case OPC_STP:
618 case OPC_STL:
619 case OPC_STLW:
620 case OPC_L2G:
621 case OPC_G2L:
622 return true;
623 default:
624 return false;
625 }
626 }
627
628 static inline bool is_load(struct ir3_instruction *instr)
629 {
630 switch (instr->opc) {
631 case OPC_LDG:
632 case OPC_LDGB:
633 case OPC_LDL:
634 case OPC_LDP:
635 case OPC_L2G:
636 case OPC_LDLW:
637 case OPC_LDC:
638 case OPC_LDLV:
639 /* probably some others too.. */
640 return true;
641 default:
642 return false;
643 }
644 }
645
646 static inline bool is_input(struct ir3_instruction *instr)
647 {
648 /* in some cases, ldlv is used to fetch varying without
649 * interpolation.. fortunately inloc is the first src
650 * register in either case
651 */
652 switch (instr->opc) {
653 case OPC_LDLV:
654 case OPC_BARY_F:
655 return true;
656 default:
657 return false;
658 }
659 }
660
661 static inline bool is_bool(struct ir3_instruction *instr)
662 {
663 switch (instr->opc) {
664 case OPC_CMPS_F:
665 case OPC_CMPS_S:
666 case OPC_CMPS_U:
667 return true;
668 default:
669 return false;
670 }
671 }
672
673 static inline bool is_meta(struct ir3_instruction *instr)
674 {
675 /* TODO how should we count PHI (and maybe fan-in/out) which
676 * might actually contribute some instructions to the final
677 * result?
678 */
679 return (opc_cat(instr->opc) == -1);
680 }
681
682 static inline bool writes_addr(struct ir3_instruction *instr)
683 {
684 if (instr->regs_count > 0) {
685 struct ir3_register *dst = instr->regs[0];
686 return reg_num(dst) == REG_A0;
687 }
688 return false;
689 }
690
691 static inline bool writes_pred(struct ir3_instruction *instr)
692 {
693 if (instr->regs_count > 0) {
694 struct ir3_register *dst = instr->regs[0];
695 return reg_num(dst) == REG_P0;
696 }
697 return false;
698 }
699
700 /* returns defining instruction for reg */
701 /* TODO better name */
702 static inline struct ir3_instruction *ssa(struct ir3_register *reg)
703 {
704 if (reg->flags & (IR3_REG_SSA | IR3_REG_ARRAY)) {
705 debug_assert(!(reg->instr && (reg->instr->flags & IR3_INSTR_UNUSED)));
706 return reg->instr;
707 }
708 return NULL;
709 }
710
711 static inline bool conflicts(struct ir3_instruction *a,
712 struct ir3_instruction *b)
713 {
714 return (a && b) && (a != b);
715 }
716
717 static inline bool reg_gpr(struct ir3_register *r)
718 {
719 if (r->flags & (IR3_REG_CONST | IR3_REG_IMMED))
720 return false;
721 if ((reg_num(r) == REG_A0) || (reg_num(r) == REG_P0))
722 return false;
723 return true;
724 }
725
726 static inline type_t half_type(type_t type)
727 {
728 switch (type) {
729 case TYPE_F32: return TYPE_F16;
730 case TYPE_U32: return TYPE_U16;
731 case TYPE_S32: return TYPE_S16;
732 case TYPE_F16:
733 case TYPE_U16:
734 case TYPE_S16:
735 return type;
736 default:
737 assert(0);
738 return ~0;
739 }
740 }
741
742 /* some cat2 instructions (ie. those which are not float) can embed an
743 * immediate:
744 */
745 static inline bool ir3_cat2_int(opc_t opc)
746 {
747 switch (opc) {
748 case OPC_ADD_U:
749 case OPC_ADD_S:
750 case OPC_SUB_U:
751 case OPC_SUB_S:
752 case OPC_CMPS_U:
753 case OPC_CMPS_S:
754 case OPC_MIN_U:
755 case OPC_MIN_S:
756 case OPC_MAX_U:
757 case OPC_MAX_S:
758 case OPC_CMPV_U:
759 case OPC_CMPV_S:
760 case OPC_MUL_U:
761 case OPC_MUL_S:
762 case OPC_MULL_U:
763 case OPC_CLZ_S:
764 case OPC_ABSNEG_S:
765 case OPC_AND_B:
766 case OPC_OR_B:
767 case OPC_NOT_B:
768 case OPC_XOR_B:
769 case OPC_BFREV_B:
770 case OPC_CLZ_B:
771 case OPC_SHL_B:
772 case OPC_SHR_B:
773 case OPC_ASHR_B:
774 case OPC_MGEN_B:
775 case OPC_GETBIT_B:
776 case OPC_CBITS_B:
777 case OPC_BARY_F:
778 return true;
779
780 default:
781 return false;
782 }
783 }
784
785
786 /* map cat2 instruction to valid abs/neg flags: */
787 static inline unsigned ir3_cat2_absneg(opc_t opc)
788 {
789 switch (opc) {
790 case OPC_ADD_F:
791 case OPC_MIN_F:
792 case OPC_MAX_F:
793 case OPC_MUL_F:
794 case OPC_SIGN_F:
795 case OPC_CMPS_F:
796 case OPC_ABSNEG_F:
797 case OPC_CMPV_F:
798 case OPC_FLOOR_F:
799 case OPC_CEIL_F:
800 case OPC_RNDNE_F:
801 case OPC_RNDAZ_F:
802 case OPC_TRUNC_F:
803 case OPC_BARY_F:
804 return IR3_REG_FABS | IR3_REG_FNEG;
805
806 case OPC_ADD_U:
807 case OPC_ADD_S:
808 case OPC_SUB_U:
809 case OPC_SUB_S:
810 case OPC_CMPS_U:
811 case OPC_CMPS_S:
812 case OPC_MIN_U:
813 case OPC_MIN_S:
814 case OPC_MAX_U:
815 case OPC_MAX_S:
816 case OPC_CMPV_U:
817 case OPC_CMPV_S:
818 case OPC_MUL_U:
819 case OPC_MUL_S:
820 case OPC_MULL_U:
821 case OPC_CLZ_S:
822 return 0;
823
824 case OPC_ABSNEG_S:
825 return IR3_REG_SABS | IR3_REG_SNEG;
826
827 case OPC_AND_B:
828 case OPC_OR_B:
829 case OPC_NOT_B:
830 case OPC_XOR_B:
831 case OPC_BFREV_B:
832 case OPC_CLZ_B:
833 case OPC_SHL_B:
834 case OPC_SHR_B:
835 case OPC_ASHR_B:
836 case OPC_MGEN_B:
837 case OPC_GETBIT_B:
838 case OPC_CBITS_B:
839 return IR3_REG_BNOT;
840
841 default:
842 return 0;
843 }
844 }
845
846 /* map cat3 instructions to valid abs/neg flags: */
847 static inline unsigned ir3_cat3_absneg(opc_t opc)
848 {
849 switch (opc) {
850 case OPC_MAD_F16:
851 case OPC_MAD_F32:
852 case OPC_SEL_F16:
853 case OPC_SEL_F32:
854 return IR3_REG_FNEG;
855
856 case OPC_MAD_U16:
857 case OPC_MADSH_U16:
858 case OPC_MAD_S16:
859 case OPC_MADSH_M16:
860 case OPC_MAD_U24:
861 case OPC_MAD_S24:
862 case OPC_SEL_S16:
863 case OPC_SEL_S32:
864 case OPC_SAD_S16:
865 case OPC_SAD_S32:
866 /* neg *may* work on 3rd src.. */
867
868 case OPC_SEL_B16:
869 case OPC_SEL_B32:
870
871 default:
872 return 0;
873 }
874 }
875
876 /* iterator for an instructions's sources (reg), also returns src #: */
877 #define foreach_src_n(__srcreg, __n, __instr) \
878 if ((__instr)->regs_count) \
879 for (unsigned __cnt = (__instr)->regs_count - 1, __n = 0; __n < __cnt; __n++) \
880 if ((__srcreg = (__instr)->regs[__n + 1]))
881
882 /* iterator for an instructions's sources (reg): */
883 #define foreach_src(__srcreg, __instr) \
884 foreach_src_n(__srcreg, __i, __instr)
885
886 static inline unsigned __ssa_src_cnt(struct ir3_instruction *instr)
887 {
888 if (instr->address)
889 return instr->regs_count + 1;
890 return instr->regs_count;
891 }
892
893 static inline struct ir3_instruction * __ssa_src_n(struct ir3_instruction *instr, unsigned n)
894 {
895 if (n == (instr->regs_count + 0))
896 return instr->address;
897 return ssa(instr->regs[n]);
898 }
899
900 #define __src_cnt(__instr) ((__instr)->address ? (__instr)->regs_count : (__instr)->regs_count - 1)
901
902 /* iterator for an instruction's SSA sources (instr), also returns src #: */
903 #define foreach_ssa_src_n(__srcinst, __n, __instr) \
904 if ((__instr)->regs_count) \
905 for (unsigned __cnt = __ssa_src_cnt(__instr), __n = 0; __n < __cnt; __n++) \
906 if ((__srcinst = __ssa_src_n(__instr, __n)))
907
908 /* iterator for an instruction's SSA sources (instr): */
909 #define foreach_ssa_src(__srcinst, __instr) \
910 foreach_ssa_src_n(__srcinst, __i, __instr)
911
912
913 /* dump: */
914 void ir3_print(struct ir3 *ir);
915 void ir3_print_instr(struct ir3_instruction *instr);
916
917 /* depth calculation: */
918 int ir3_delayslots(struct ir3_instruction *assigner,
919 struct ir3_instruction *consumer, unsigned n);
920 void ir3_insert_by_depth(struct ir3_instruction *instr, struct list_head *list);
921 void ir3_depth(struct ir3 *ir);
922
923 /* copy-propagate: */
924 struct ir3_shader_variant;
925 void ir3_cp(struct ir3 *ir, struct ir3_shader_variant *so);
926
927 /* group neighbors and insert mov's to resolve conflicts: */
928 void ir3_group(struct ir3 *ir);
929
930 /* scheduling: */
931 int ir3_sched(struct ir3 *ir);
932
933 /* register assignment: */
934 struct ir3_ra_reg_set * ir3_ra_alloc_reg_set(void *memctx);
935 int ir3_ra(struct ir3 *ir3, enum shader_t type,
936 bool frag_coord, bool frag_face);
937
938 /* legalize: */
939 void ir3_legalize(struct ir3 *ir, bool *has_samp, bool *has_ssbo, int *max_bary);
940
941 /* ************************************************************************* */
942 /* instruction helpers */
943
944 static inline struct ir3_instruction *
945 ir3_MOV(struct ir3_block *block, struct ir3_instruction *src, type_t type)
946 {
947 struct ir3_instruction *instr = ir3_instr_create(block, OPC_MOV);
948 ir3_reg_create(instr, 0, 0); /* dst */
949 if (src->regs[0]->flags & IR3_REG_ARRAY) {
950 struct ir3_register *src_reg =
951 ir3_reg_create(instr, 0, IR3_REG_ARRAY);
952 src_reg->array = src->regs[0]->array;
953 src_reg->instr = src;
954 } else {
955 ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = src;
956 }
957 debug_assert(!(src->regs[0]->flags & IR3_REG_RELATIV));
958 instr->cat1.src_type = type;
959 instr->cat1.dst_type = type;
960 return instr;
961 }
962
963 static inline struct ir3_instruction *
964 ir3_COV(struct ir3_block *block, struct ir3_instruction *src,
965 type_t src_type, type_t dst_type)
966 {
967 struct ir3_instruction *instr = ir3_instr_create(block, OPC_MOV);
968 ir3_reg_create(instr, 0, 0); /* dst */
969 ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = src;
970 instr->cat1.src_type = src_type;
971 instr->cat1.dst_type = dst_type;
972 debug_assert(!(src->regs[0]->flags & IR3_REG_ARRAY));
973 return instr;
974 }
975
976 static inline struct ir3_instruction *
977 ir3_NOP(struct ir3_block *block)
978 {
979 return ir3_instr_create(block, OPC_NOP);
980 }
981
982 #define INSTR0(name) \
983 static inline struct ir3_instruction * \
984 ir3_##name(struct ir3_block *block) \
985 { \
986 struct ir3_instruction *instr = \
987 ir3_instr_create(block, OPC_##name); \
988 return instr; \
989 }
990
991 #define INSTR1(name) \
992 static inline struct ir3_instruction * \
993 ir3_##name(struct ir3_block *block, \
994 struct ir3_instruction *a, unsigned aflags) \
995 { \
996 struct ir3_instruction *instr = \
997 ir3_instr_create(block, OPC_##name); \
998 ir3_reg_create(instr, 0, 0); /* dst */ \
999 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1000 return instr; \
1001 }
1002
1003 #define INSTR2(name) \
1004 static inline struct ir3_instruction * \
1005 ir3_##name(struct ir3_block *block, \
1006 struct ir3_instruction *a, unsigned aflags, \
1007 struct ir3_instruction *b, unsigned bflags) \
1008 { \
1009 struct ir3_instruction *instr = \
1010 ir3_instr_create(block, OPC_##name); \
1011 ir3_reg_create(instr, 0, 0); /* dst */ \
1012 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1013 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1014 return instr; \
1015 }
1016
1017 #define INSTR3(name) \
1018 static inline struct ir3_instruction * \
1019 ir3_##name(struct ir3_block *block, \
1020 struct ir3_instruction *a, unsigned aflags, \
1021 struct ir3_instruction *b, unsigned bflags, \
1022 struct ir3_instruction *c, unsigned cflags) \
1023 { \
1024 struct ir3_instruction *instr = \
1025 ir3_instr_create(block, OPC_##name); \
1026 ir3_reg_create(instr, 0, 0); /* dst */ \
1027 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1028 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1029 ir3_reg_create(instr, 0, IR3_REG_SSA | cflags)->instr = c; \
1030 return instr; \
1031 }
1032
1033 #define INSTR4(name) \
1034 static inline struct ir3_instruction * \
1035 ir3_##name(struct ir3_block *block, \
1036 struct ir3_instruction *a, unsigned aflags, \
1037 struct ir3_instruction *b, unsigned bflags, \
1038 struct ir3_instruction *c, unsigned cflags, \
1039 struct ir3_instruction *d, unsigned dflags) \
1040 { \
1041 struct ir3_instruction *instr = \
1042 ir3_instr_create2(block, OPC_##name, 5); \
1043 ir3_reg_create(instr, 0, 0); /* dst */ \
1044 ir3_reg_create(instr, 0, IR3_REG_SSA | aflags)->instr = a; \
1045 ir3_reg_create(instr, 0, IR3_REG_SSA | bflags)->instr = b; \
1046 ir3_reg_create(instr, 0, IR3_REG_SSA | cflags)->instr = c; \
1047 ir3_reg_create(instr, 0, IR3_REG_SSA | dflags)->instr = d; \
1048 return instr; \
1049 }
1050
1051 /* cat0 instructions: */
1052 INSTR0(BR);
1053 INSTR0(JUMP);
1054 INSTR1(KILL);
1055 INSTR0(END);
1056
1057 /* cat2 instructions, most 2 src but some 1 src: */
1058 INSTR2(ADD_F)
1059 INSTR2(MIN_F)
1060 INSTR2(MAX_F)
1061 INSTR2(MUL_F)
1062 INSTR1(SIGN_F)
1063 INSTR2(CMPS_F)
1064 INSTR1(ABSNEG_F)
1065 INSTR2(CMPV_F)
1066 INSTR1(FLOOR_F)
1067 INSTR1(CEIL_F)
1068 INSTR1(RNDNE_F)
1069 INSTR1(RNDAZ_F)
1070 INSTR1(TRUNC_F)
1071 INSTR2(ADD_U)
1072 INSTR2(ADD_S)
1073 INSTR2(SUB_U)
1074 INSTR2(SUB_S)
1075 INSTR2(CMPS_U)
1076 INSTR2(CMPS_S)
1077 INSTR2(MIN_U)
1078 INSTR2(MIN_S)
1079 INSTR2(MAX_U)
1080 INSTR2(MAX_S)
1081 INSTR1(ABSNEG_S)
1082 INSTR2(AND_B)
1083 INSTR2(OR_B)
1084 INSTR1(NOT_B)
1085 INSTR2(XOR_B)
1086 INSTR2(CMPV_U)
1087 INSTR2(CMPV_S)
1088 INSTR2(MUL_U)
1089 INSTR2(MUL_S)
1090 INSTR2(MULL_U)
1091 INSTR1(BFREV_B)
1092 INSTR1(CLZ_S)
1093 INSTR1(CLZ_B)
1094 INSTR2(SHL_B)
1095 INSTR2(SHR_B)
1096 INSTR2(ASHR_B)
1097 INSTR2(BARY_F)
1098 INSTR2(MGEN_B)
1099 INSTR2(GETBIT_B)
1100 INSTR1(SETRM)
1101 INSTR1(CBITS_B)
1102 INSTR2(SHB)
1103 INSTR2(MSAD)
1104
1105 /* cat3 instructions: */
1106 INSTR3(MAD_U16)
1107 INSTR3(MADSH_U16)
1108 INSTR3(MAD_S16)
1109 INSTR3(MADSH_M16)
1110 INSTR3(MAD_U24)
1111 INSTR3(MAD_S24)
1112 INSTR3(MAD_F16)
1113 INSTR3(MAD_F32)
1114 INSTR3(SEL_B16)
1115 INSTR3(SEL_B32)
1116 INSTR3(SEL_S16)
1117 INSTR3(SEL_S32)
1118 INSTR3(SEL_F16)
1119 INSTR3(SEL_F32)
1120 INSTR3(SAD_S16)
1121 INSTR3(SAD_S32)
1122
1123 /* cat4 instructions: */
1124 INSTR1(RCP)
1125 INSTR1(RSQ)
1126 INSTR1(LOG2)
1127 INSTR1(EXP2)
1128 INSTR1(SIN)
1129 INSTR1(COS)
1130 INSTR1(SQRT)
1131
1132 /* cat5 instructions: */
1133 INSTR1(DSX)
1134 INSTR1(DSY)
1135
1136 static inline struct ir3_instruction *
1137 ir3_SAM(struct ir3_block *block, opc_t opc, type_t type,
1138 unsigned wrmask, unsigned flags, unsigned samp, unsigned tex,
1139 struct ir3_instruction *src0, struct ir3_instruction *src1)
1140 {
1141 struct ir3_instruction *sam;
1142 struct ir3_register *reg;
1143
1144 sam = ir3_instr_create(block, opc);
1145 sam->flags |= flags;
1146 ir3_reg_create(sam, 0, 0)->wrmask = wrmask;
1147 if (src0) {
1148 reg = ir3_reg_create(sam, 0, IR3_REG_SSA);
1149 reg->wrmask = (1 << (src0->regs_count - 1)) - 1;
1150 reg->instr = src0;
1151 }
1152 if (src1) {
1153 reg = ir3_reg_create(sam, 0, IR3_REG_SSA);
1154 reg->instr = src1;
1155 reg->wrmask = (1 << (src1->regs_count - 1)) - 1;
1156 }
1157 sam->cat5.samp = samp;
1158 sam->cat5.tex = tex;
1159 sam->cat5.type = type;
1160
1161 return sam;
1162 }
1163
1164 /* cat6 instructions: */
1165 INSTR2(LDLV)
1166 INSTR2(LDG)
1167 INSTR3(STG)
1168 INSTR3(LDGB);
1169 INSTR4(STGB);
1170 INSTR4(ATOMIC_ADD);
1171 INSTR4(ATOMIC_SUB);
1172 INSTR4(ATOMIC_XCHG);
1173 INSTR4(ATOMIC_INC);
1174 INSTR4(ATOMIC_DEC);
1175 INSTR4(ATOMIC_CMPXCHG);
1176 INSTR4(ATOMIC_MIN);
1177 INSTR4(ATOMIC_MAX);
1178 INSTR4(ATOMIC_AND);
1179 INSTR4(ATOMIC_OR);
1180 INSTR4(ATOMIC_XOR);
1181
1182 /* ************************************************************************* */
1183 /* split this out or find some helper to use.. like main/bitset.h.. */
1184
1185 #include <string.h>
1186
1187 #define MAX_REG 256
1188
1189 typedef uint8_t regmask_t[2 * MAX_REG / 8];
1190
1191 static inline unsigned regmask_idx(struct ir3_register *reg)
1192 {
1193 unsigned num = (reg->flags & IR3_REG_RELATIV) ? reg->array.offset : reg->num;
1194 debug_assert(num < MAX_REG);
1195 if (reg->flags & IR3_REG_HALF)
1196 num += MAX_REG;
1197 return num;
1198 }
1199
1200 static inline void regmask_init(regmask_t *regmask)
1201 {
1202 memset(regmask, 0, sizeof(*regmask));
1203 }
1204
1205 static inline void regmask_set(regmask_t *regmask, struct ir3_register *reg)
1206 {
1207 unsigned idx = regmask_idx(reg);
1208 if (reg->flags & IR3_REG_RELATIV) {
1209 unsigned i;
1210 for (i = 0; i < reg->size; i++, idx++)
1211 (*regmask)[idx / 8] |= 1 << (idx % 8);
1212 } else {
1213 unsigned mask;
1214 for (mask = reg->wrmask; mask; mask >>= 1, idx++)
1215 if (mask & 1)
1216 (*regmask)[idx / 8] |= 1 << (idx % 8);
1217 }
1218 }
1219
1220 static inline void regmask_or(regmask_t *dst, regmask_t *a, regmask_t *b)
1221 {
1222 unsigned i;
1223 for (i = 0; i < ARRAY_SIZE(*dst); i++)
1224 (*dst)[i] = (*a)[i] | (*b)[i];
1225 }
1226
1227 /* set bits in a if not set in b, conceptually:
1228 * a |= (reg & ~b)
1229 */
1230 static inline void regmask_set_if_not(regmask_t *a,
1231 struct ir3_register *reg, regmask_t *b)
1232 {
1233 unsigned idx = regmask_idx(reg);
1234 if (reg->flags & IR3_REG_RELATIV) {
1235 unsigned i;
1236 for (i = 0; i < reg->size; i++, idx++)
1237 if (!((*b)[idx / 8] & (1 << (idx % 8))))
1238 (*a)[idx / 8] |= 1 << (idx % 8);
1239 } else {
1240 unsigned mask;
1241 for (mask = reg->wrmask; mask; mask >>= 1, idx++)
1242 if (mask & 1)
1243 if (!((*b)[idx / 8] & (1 << (idx % 8))))
1244 (*a)[idx / 8] |= 1 << (idx % 8);
1245 }
1246 }
1247
1248 static inline bool regmask_get(regmask_t *regmask,
1249 struct ir3_register *reg)
1250 {
1251 unsigned idx = regmask_idx(reg);
1252 if (reg->flags & IR3_REG_RELATIV) {
1253 unsigned i;
1254 for (i = 0; i < reg->size; i++, idx++)
1255 if ((*regmask)[idx / 8] & (1 << (idx % 8)))
1256 return true;
1257 } else {
1258 unsigned mask;
1259 for (mask = reg->wrmask; mask; mask >>= 1, idx++)
1260 if (mask & 1)
1261 if ((*regmask)[idx / 8] & (1 << (idx % 8)))
1262 return true;
1263 }
1264 return false;
1265 }
1266
1267 /* ************************************************************************* */
1268
1269 #endif /* IR3_H_ */