summaryrefslogtreecommitdiff
path: root/src/panfrost/midgard/midgard_ra.c
blob: 1e511ad92c255c9d92c690d3e612e92b62077326 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
/*
 * Copyright (C) 2018-2019 Alyssa Rosenzweig <alyssa@rosenzweig.io>
 * Copyright (C) 2019 Collabora, Ltd.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */

#include "compiler.h"
#include "midgard_ops.h"
#include "util/u_math.h"
#include "util/u_memory.h"
#include "midgard_quirks.h"

struct phys_reg {
        /* Physical register: 0-31 */
        unsigned reg;

        /* Byte offset into the physical register: 0-15 */
        unsigned offset;

        /* log2(bytes per component) for fast mul/div */
        unsigned shift;
};

/* Shift up by reg_offset and horizontally by dst_offset. */

static void
offset_swizzle(unsigned *swizzle, unsigned reg_offset, unsigned srcshift, unsigned dstshift, unsigned dst_offset)
{
        unsigned out[MIR_VEC_COMPONENTS];

        signed reg_comp = reg_offset >> srcshift;
        signed dst_comp = dst_offset >> dstshift;

        unsigned max_component = (16 >> srcshift) - 1;

        assert(reg_comp << srcshift == reg_offset);
        assert(dst_comp << dstshift == dst_offset);

        for (signed c = 0; c < MIR_VEC_COMPONENTS; ++c) {
                signed comp = MAX2(c - dst_comp, 0);
                out[c] = MIN2(swizzle[comp] + reg_comp, max_component);
        }

        memcpy(swizzle, out, sizeof(out));
}

/* Helper to return the default phys_reg for a given register */

static struct phys_reg
default_phys_reg(int reg, unsigned shift)
{
        struct phys_reg r = {
                .reg = reg,
                .offset = 0,
                .shift = shift
        };

        return r;
}

/* Determine which physical register, swizzle, and mask a virtual
 * register corresponds to */

static struct phys_reg
index_to_reg(compiler_context *ctx, struct lcra_state *l, unsigned reg, unsigned shift)
{
        /* Check for special cases */
        if (reg == ~0)
                return default_phys_reg(REGISTER_UNUSED, shift);
        else if (reg >= SSA_FIXED_MINIMUM)
                return default_phys_reg(SSA_REG_FROM_FIXED(reg), shift);
        else if (!l)
                return default_phys_reg(REGISTER_UNUSED, shift);

        struct phys_reg r = {
                .reg = l->solutions[reg] / 16,
                .offset = l->solutions[reg] & 0xF,
                .shift = shift
        };

        /* Report that we actually use this register, and return it */

        if (r.reg < 16)
                ctx->work_registers = MAX2(ctx->work_registers, r.reg);

        return r;
}

static void
set_class(unsigned *classes, unsigned node, unsigned class)
{
        if (node < SSA_FIXED_MINIMUM && class != classes[node]) {
                assert(classes[node] == REG_CLASS_WORK);
                classes[node] = class;
        }
}

/* Special register classes impose special constraints on who can read their
 * values, so check that */

static bool ASSERTED
check_read_class(unsigned *classes, unsigned tag, unsigned node)
{
        /* Non-nodes are implicitly ok */
        if (node >= SSA_FIXED_MINIMUM)
                return true;

        switch (classes[node]) {
        case REG_CLASS_LDST:
                return (tag == TAG_LOAD_STORE_4);
        case REG_CLASS_TEXR:
                return (tag == TAG_TEXTURE_4);
        case REG_CLASS_TEXW:
                return (tag != TAG_LOAD_STORE_4);
        case REG_CLASS_WORK:
                return IS_ALU(tag);
        default:
                unreachable("Invalid class");
        }
}

static bool ASSERTED
check_write_class(unsigned *classes, unsigned tag, unsigned node)
{
        /* Non-nodes are implicitly ok */
        if (node >= SSA_FIXED_MINIMUM)
                return true;

        switch (classes[node]) {
        case REG_CLASS_TEXR:
                return true;
        case REG_CLASS_TEXW:
                return (tag == TAG_TEXTURE_4);
        case REG_CLASS_LDST:
        case REG_CLASS_WORK:
                return IS_ALU(tag) || (tag == TAG_LOAD_STORE_4);
        default:
                unreachable("Invalid class");
        }
}

/* Prepass before RA to ensure special class restrictions are met. The idea is
 * to create a bit field of types of instructions that read a particular index.
 * Later, we'll add moves as appropriate and rewrite to specialize by type. */

static void
mark_node_class (unsigned *bitfield, unsigned node)
{
        if (node < SSA_FIXED_MINIMUM)
                BITSET_SET(bitfield, node);
}

void
mir_lower_special_reads(compiler_context *ctx)
{
        size_t sz = BITSET_WORDS(ctx->temp_count) * sizeof(BITSET_WORD);

        /* Bitfields for the various types of registers we could have. aluw can
         * be written by either ALU or load/store */

        unsigned *alur = calloc(sz, 1);
        unsigned *aluw = calloc(sz, 1);
        unsigned *brar = calloc(sz, 1);
        unsigned *ldst = calloc(sz, 1);
        unsigned *texr = calloc(sz, 1);
        unsigned *texw = calloc(sz, 1);

        /* Pass #1 is analysis, a linear scan to fill out the bitfields */

        mir_foreach_instr_global(ctx, ins) {
                switch (ins->type) {
                case TAG_ALU_4:
                        mark_node_class(aluw, ins->dest);
                        mark_node_class(alur, ins->src[0]);
                        mark_node_class(alur, ins->src[1]);
                        mark_node_class(alur, ins->src[2]);

                        if (ins->compact_branch && ins->writeout)
                                mark_node_class(brar, ins->src[0]);

                        break;

                case TAG_LOAD_STORE_4:
                        mark_node_class(aluw, ins->dest);
                        mark_node_class(ldst, ins->src[0]);
                        mark_node_class(ldst, ins->src[1]);
                        mark_node_class(ldst, ins->src[2]);
                        mark_node_class(ldst, ins->src[3]);
                        break;

                case TAG_TEXTURE_4:
                        mark_node_class(texr, ins->src[0]);
                        mark_node_class(texr, ins->src[1]);
                        mark_node_class(texr, ins->src[2]);
                        mark_node_class(texw, ins->dest);
                        break;
                }
        }

        /* Pass #2 is lowering now that we've analyzed all the classes.
         * Conceptually, if an index is only marked for a single type of use,
         * there is nothing to lower. If it is marked for different uses, we
         * split up based on the number of types of uses. To do so, we divide
         * into N distinct classes of use (where N>1 by definition), emit N-1
         * moves from the index to copies of the index, and finally rewrite N-1
         * of the types of uses to use the corresponding move */

        unsigned spill_idx = ctx->temp_count;

        for (unsigned i = 0; i < ctx->temp_count; ++i) {
                bool is_alur = BITSET_TEST(alur, i);
                bool is_aluw = BITSET_TEST(aluw, i);
                bool is_brar = BITSET_TEST(brar, i);
                bool is_ldst = BITSET_TEST(ldst, i);
                bool is_texr = BITSET_TEST(texr, i);
                bool is_texw = BITSET_TEST(texw, i);

                /* Analyse to check how many distinct uses there are. ALU ops
                 * (alur) can read the results of the texture pipeline (texw)
                 * but not ldst or texr. Load/store ops (ldst) cannot read
                 * anything but load/store inputs. Texture pipeline cannot read
                 * anything but texture inputs. TODO: Simplify.  */

                bool collision =
                        (is_alur && (is_ldst || is_texr)) ||
                        (is_ldst && (is_alur || is_texr || is_texw)) ||
                        (is_texr && (is_alur || is_ldst || is_texw)) ||
                        (is_texw && (is_aluw || is_ldst || is_texr)) ||
                        (is_brar && is_texw);
        
                if (!collision)
                        continue;

                /* Use the index as-is as the work copy. Emit copies for
                 * special uses */

                unsigned classes[] = { TAG_LOAD_STORE_4, TAG_TEXTURE_4, TAG_TEXTURE_4, TAG_ALU_4};
                bool collisions[] = { is_ldst, is_texr, is_texw && is_aluw, is_brar };

                for (unsigned j = 0; j < ARRAY_SIZE(collisions); ++j) {
                        if (!collisions[j]) continue;

                        /* When the hazard is from reading, we move and rewrite
                         * sources (typical case). When it's from writing, we
                         * flip the move and rewrite destinations (obscure,
                         * only from control flow -- impossible in SSA) */

                        bool hazard_write = (j == 2);

                        unsigned idx = spill_idx++;

                        /* Insert move before each read/write, depending on the
                         * hazard we're trying to account for */

                        mir_foreach_instr_global_safe(ctx, pre_use) {
                                if (pre_use->type != classes[j])
                                        continue;

                                if (hazard_write) {
                                        if (pre_use->dest != i)
                                                continue;

                                        midgard_instruction m = v_mov(idx, i);
                                        m.dest_type = pre_use->dest_type;
                                        m.src_types[1] = m.dest_type;
                                        m.mask = pre_use->mask;

                                        midgard_instruction *use = mir_next_op(pre_use);
                                        assert(use);
                                        mir_insert_instruction_before(ctx, use, m);
                                        mir_rewrite_index_dst_single(pre_use, i, idx);
                                } else {
                                        if (!mir_has_arg(pre_use, i))
                                                continue;

                                        idx = spill_idx++;

                                        midgard_instruction m = v_mov(i, idx);
                                        m.mask = mir_from_bytemask(mir_round_bytemask_up(
                                                                mir_bytemask_of_read_components(pre_use, i), 32), 32);
                                        mir_insert_instruction_before(ctx, pre_use, m);
                                        mir_rewrite_index_src_single(pre_use, i, idx);
                                }
                        }
                }
        }

        free(alur);
        free(aluw);
        free(brar);
        free(ldst);
        free(texr);
        free(texw);
}

static void
mir_compute_interference(
                compiler_context *ctx,
                struct lcra_state *l)
{
        /* First, we need liveness information to be computed per block */
        mir_compute_liveness(ctx);

        /* We need to force r1.w live throughout a blend shader */

        if (ctx->is_blend) {
                unsigned r1w = ~0;

                mir_foreach_block(ctx, _block) {
                        midgard_block *block = (midgard_block *) _block;
                        mir_foreach_instr_in_block_rev(block, ins) {
                                if (ins->writeout)
                                        r1w = ins->dest;
                        }

                        if (r1w != ~0)
                                break;
                }

                mir_foreach_instr_global(ctx, ins) {
                        if (ins->dest < ctx->temp_count)
                                lcra_add_node_interference(l, ins->dest, mir_bytemask(ins), r1w, 0xF);
                }
        }

        /* Now that every block has live_in/live_out computed, we can determine
         * interference by walking each block linearly. Take live_out at the
         * end of each block and walk the block backwards. */

        mir_foreach_block(ctx, _blk) {
                midgard_block *blk = (midgard_block *) _blk;
                uint16_t *live = mem_dup(_blk->live_out, ctx->temp_count * sizeof(uint16_t));

                mir_foreach_instr_in_block_rev(blk, ins) {
                        /* Mark all registers live after the instruction as
                         * interfering with the destination */

                        unsigned dest = ins->dest;

                        if (dest < ctx->temp_count) {
                                for (unsigned i = 0; i < ctx->temp_count; ++i)
                                        if (live[i]) {
                                                unsigned mask = mir_bytemask(ins);
                                                lcra_add_node_interference(l, dest, mask, i, live[i]);
                                        }
                        }

                        /* Update live_in */
                        mir_liveness_ins_update(live, ins, ctx->temp_count);
                }

                free(live);
        }
}

static bool
mir_is_64(midgard_instruction *ins)
{
        if (nir_alu_type_get_type_size(ins->dest_type) == 64)
                return true;

        mir_foreach_src(ins, v) {
                if (nir_alu_type_get_type_size(ins->src_types[v]) == 64)
                        return true;
        }

        return false;
}

/* This routine performs the actual register allocation. It should be succeeded
 * by install_registers */

static struct lcra_state *
allocate_registers(compiler_context *ctx, bool *spilled)
{
        /* The number of vec4 work registers available depends on when the
         * uniforms start and the shader stage. By ABI we limit blend shaders
         * to 8 registers, should be lower XXX */
        int work_count = ctx->is_blend ? 8 :
                16 - MAX2((ctx->uniform_cutoff - 8), 0);

       /* No register allocation to do with no SSA */

        if (!ctx->temp_count)
                return NULL;

        /* Initialize LCRA. Allocate an extra node at the end for a precoloured
         * r1 for interference */

        struct lcra_state *l = lcra_alloc_equations(ctx->temp_count + 1, 5);
        unsigned node_r1 = ctx->temp_count;

        /* Starts of classes, in bytes */
        l->class_start[REG_CLASS_WORK]  = 16 * 0;
        l->class_start[REG_CLASS_LDST]  = 16 * 26;
        l->class_start[REG_CLASS_TEXR]  = 16 * 28;
        l->class_start[REG_CLASS_TEXW]  = 16 * 28;

        l->class_size[REG_CLASS_WORK] = 16 * work_count;
        l->class_size[REG_CLASS_LDST]  = 16 * 2;
        l->class_size[REG_CLASS_TEXR]  = 16 * 2;
        l->class_size[REG_CLASS_TEXW]  = 16 * 2;

        lcra_set_disjoint_class(l, REG_CLASS_TEXR, REG_CLASS_TEXW);

        /* To save space on T*20, we don't have real texture registers.
         * Instead, tex inputs reuse the load/store pipeline registers, and
         * tex outputs use work r0/r1. Note we still use TEXR/TEXW classes,
         * noting that this handles interferences and sizes correctly. */

        if (ctx->quirks & MIDGARD_INTERPIPE_REG_ALIASING) {
                l->class_start[REG_CLASS_TEXR] = l->class_start[REG_CLASS_LDST];
                l->class_start[REG_CLASS_TEXW] = l->class_start[REG_CLASS_WORK];
        }

        unsigned *found_class = calloc(sizeof(unsigned), ctx->temp_count);
        unsigned *min_alignment = calloc(sizeof(unsigned), ctx->temp_count);
        unsigned *min_bound = calloc(sizeof(unsigned), ctx->temp_count);

        mir_foreach_instr_global(ctx, ins) {
                /* Swizzles of 32-bit sources on 64-bit instructions need to be
                 * aligned to either bottom (xy) or top (zw). More general
                 * swizzle lowering should happen prior to scheduling (TODO),
                 * but once we get RA we shouldn't disrupt this further. Align
                 * sources of 64-bit instructions. */

                if (ins->type == TAG_ALU_4 && mir_is_64(ins)) {
                        mir_foreach_src(ins, v) {
                                unsigned s = ins->src[v];

                                if (s < ctx->temp_count)
                                        min_alignment[s] = 3;
                        }
                }

                if (ins->type == TAG_LOAD_STORE_4 && OP_HAS_ADDRESS(ins->op)) {
                        mir_foreach_src(ins, v) {
                                unsigned s = ins->src[v];
                                unsigned size = nir_alu_type_get_type_size(ins->src_types[v]);

                                if (s < ctx->temp_count)
                                        min_alignment[s] = (size == 64) ? 3 : 2;
                        }
                }

                if (ins->dest >= SSA_FIXED_MINIMUM) continue;

                unsigned size = nir_alu_type_get_type_size(ins->dest_type);

                if (ins->is_pack)
                        size = 32;

                /* 0 for x, 1 for xy, 2 for xyz, 3 for xyzw */
                int comps1 = util_logbase2(ins->mask);

                int bytes = (comps1 + 1) * (size / 8);

                /* Use the largest class if there's ambiguity, this
                 * handles partial writes */

                int dest = ins->dest;
                found_class[dest] = MAX2(found_class[dest], bytes);

                min_alignment[dest] =
                        (size == 16) ? 1 : /* (1 << 1) = 2-byte */
                        (size == 32) ? 2 : /* (1 << 2) = 4-byte */
                        (size == 64) ? 3 : /* (1 << 3) = 8-byte */
                        3; /* 8-bit todo */

                /* We can't cross xy/zw boundaries. TODO: vec8 can */
                if (size == 16)
                        min_bound[dest] = 8;

                /* We don't have a swizzle for the conditional and we don't
                 * want to muck with the conditional itself, so just force
                 * alignment for now */

                if (ins->type == TAG_ALU_4 && OP_IS_CSEL_V(ins->op)) {
                        min_alignment[dest] = 4; /* 1 << 4= 16-byte = vec4 */

                        /* LCRA assumes bound >= alignment */
                        min_bound[dest] = 16;
                }

                /* Since ld/st swizzles and masks are 32-bit only, we need them
                 * aligned to enable final packing */
                if (ins->type == TAG_LOAD_STORE_4)
                        min_alignment[dest] = MAX2(min_alignment[dest], 2);
        }

        for (unsigned i = 0; i < ctx->temp_count; ++i) {
                lcra_set_alignment(l, i, min_alignment[i] ? min_alignment[i] : 2,
                                min_bound[i] ? min_bound[i] : 16);
                lcra_restrict_range(l, i, found_class[i]);
        }
        
        free(found_class);
        free(min_alignment);
        free(min_bound);

        /* Next, we'll determine semantic class. We default to zero (work).
         * But, if we're used with a special operation, that will force us to a
         * particular class. Each node must be assigned to exactly one class; a
         * prepass before RA should have lowered what-would-have-been
         * multiclass nodes into a series of moves to break it up into multiple
         * nodes (TODO) */

        mir_foreach_instr_global(ctx, ins) {
                /* Check if this operation imposes any classes */

                if (ins->type == TAG_LOAD_STORE_4) {
                        set_class(l->class, ins->src[0], REG_CLASS_LDST);
                        set_class(l->class, ins->src[1], REG_CLASS_LDST);
                        set_class(l->class, ins->src[2], REG_CLASS_LDST);
                        set_class(l->class, ins->src[3], REG_CLASS_LDST);

                        if (OP_IS_VEC4_ONLY(ins->op)) {
                                lcra_restrict_range(l, ins->dest, 16);
                                lcra_restrict_range(l, ins->src[0], 16);
                                lcra_restrict_range(l, ins->src[1], 16);
                                lcra_restrict_range(l, ins->src[2], 16);
                                lcra_restrict_range(l, ins->src[3], 16);
                        }
                } else if (ins->type == TAG_TEXTURE_4) {
                        set_class(l->class, ins->dest, REG_CLASS_TEXW);
                        set_class(l->class, ins->src[0], REG_CLASS_TEXR);
                        set_class(l->class, ins->src[1], REG_CLASS_TEXR);
                        set_class(l->class, ins->src[2], REG_CLASS_TEXR);
                        set_class(l->class, ins->src[3], REG_CLASS_TEXR);
                }
        }

        /* Check that the semantics of the class are respected */
        mir_foreach_instr_global(ctx, ins) {
                assert(check_write_class(l->class, ins->type, ins->dest));
                assert(check_read_class(l->class, ins->type, ins->src[0]));
                assert(check_read_class(l->class, ins->type, ins->src[1]));
                assert(check_read_class(l->class, ins->type, ins->src[2]));
                assert(check_read_class(l->class, ins->type, ins->src[3]));
        }

        /* Mark writeout to r0, depth to r1.x, stencil to r1.y,
         * render target to r1.z, unknown to r1.w */
        mir_foreach_instr_global(ctx, ins) {
                if (!(ins->compact_branch && ins->writeout)) continue;

                if (ins->src[0] < ctx->temp_count)
                        l->solutions[ins->src[0]] = 0;

                if (ins->src[2] < ctx->temp_count)
                        l->solutions[ins->src[2]] = (16 * 1) + COMPONENT_X * 4;

                if (ins->src[3] < ctx->temp_count)
                        l->solutions[ins->src[3]] = (16 * 1) + COMPONENT_Y * 4;

                if (ins->src[1] < ctx->temp_count)
                        l->solutions[ins->src[1]] = (16 * 1) + COMPONENT_Z * 4;

                if (ins->dest < ctx->temp_count)
                        l->solutions[ins->dest] = (16 * 1) + COMPONENT_W * 4;
        }

        /* Destinations of instructions in a writeout block cannot be assigned
         * to r1 unless they are actually used as r1 from the writeout itself,
         * since the writes to r1 are special. A code sequence like:
         *
         *      sadd.fmov r1.x, [...]
         *      vadd.fadd r0, r1, r2
         *      [writeout branch]
         *
         * will misbehave since the r1.x write will be interpreted as a
         * gl_FragDepth write so it won't show up correctly when r1 is read in
         * the following segment. We model this as interference.
         */

        l->solutions[node_r1] = (16 * 1);

        mir_foreach_block(ctx, _blk) {
                midgard_block *blk = (midgard_block *) _blk;

                mir_foreach_bundle_in_block(blk, v) {
                        /* We need at least a writeout and nonwriteout instruction */
                        if (v->instruction_count < 2)
                                continue;

                        /* Branches always come at the end */
                        midgard_instruction *br = v->instructions[v->instruction_count - 1];

                        if (!br->writeout)
                                continue;

                        for (signed i = v->instruction_count - 2; i >= 0; --i) {
                                midgard_instruction *ins = v->instructions[i];

                                if (ins->dest >= ctx->temp_count)
                                        continue;

                                bool used_as_r1 = (br->dest == ins->dest);

                                mir_foreach_src(br, s)
                                        used_as_r1 |= (s > 0) && (br->src[s] == ins->dest);

                                if (!used_as_r1)
                                        lcra_add_node_interference(l, ins->dest, mir_bytemask(ins), node_r1, 0xFFFF);
                        }
                }
        }

        /* Precolour blend input to r0. Note writeout is necessarily at the end
         * and blend shaders are single-RT only so there is only a single
         * writeout block, so this cannot conflict with the writeout r0 (there
         * is no need to have an intermediate move) */

        if (ctx->blend_input != ~0) {
                assert(ctx->blend_input < ctx->temp_count);
                l->solutions[ctx->blend_input] = 0;
        }

        /* Same for the dual-source blend input/output, except here we use r2,
         * which is also set in the fragment shader. */

        if (ctx->blend_src1 != ~0) {
                assert(ctx->blend_src1 < ctx->temp_count);
                l->solutions[ctx->blend_src1] = (16 * 2);
                ctx->work_registers = MAX2(ctx->work_registers, 2);
        }

        mir_compute_interference(ctx, l);

        *spilled = !lcra_solve(l);
        return l;
}


/* Once registers have been decided via register allocation
 * (allocate_registers), we need to rewrite the MIR to use registers instead of
 * indices */

static void
install_registers_instr(
        compiler_context *ctx,
        struct lcra_state *l,
        midgard_instruction *ins)
{
        unsigned src_shift[MIR_SRC_COUNT];

        for (unsigned i = 0; i < MIR_SRC_COUNT; ++i) {
                src_shift[i] =
                        util_logbase2(nir_alu_type_get_type_size(ins->src_types[i]) / 8);
        }

        unsigned dest_shift =
                util_logbase2(nir_alu_type_get_type_size(ins->dest_type) / 8);

        switch (ins->type) {
        case TAG_ALU_4:
        case TAG_ALU_8:
        case TAG_ALU_12:
        case TAG_ALU_16: {
                 if (ins->compact_branch)
                         return;

                struct phys_reg src1 = index_to_reg(ctx, l, ins->src[0], src_shift[0]);
                struct phys_reg src2 = index_to_reg(ctx, l, ins->src[1], src_shift[1]);
                struct phys_reg dest = index_to_reg(ctx, l, ins->dest, dest_shift);

                mir_set_bytemask(ins, mir_bytemask(ins) << dest.offset);

                unsigned dest_offset =
                        GET_CHANNEL_COUNT(alu_opcode_props[ins->op].props) ? 0 :
                        dest.offset;

                offset_swizzle(ins->swizzle[0], src1.offset, src1.shift, dest.shift, dest_offset);
                if (!ins->has_inline_constant)
                        offset_swizzle(ins->swizzle[1], src2.offset, src2.shift, dest.shift, dest_offset);
                if (ins->src[0] != ~0)
                        ins->src[0] = SSA_FIXED_REGISTER(src1.reg);
                if (ins->src[1] != ~0)
                        ins->src[1] = SSA_FIXED_REGISTER(src2.reg);
                if (ins->dest != ~0)
                        ins->dest = SSA_FIXED_REGISTER(dest.reg);
                break;
        }

        case TAG_LOAD_STORE_4: {
                /* Which physical register we read off depends on
                 * whether we are loading or storing -- think about the
                 * logical dataflow */

                bool encodes_src = OP_IS_STORE(ins->op);

                if (encodes_src) {
                        struct phys_reg src = index_to_reg(ctx, l, ins->src[0], src_shift[0]);
                        assert(src.reg == 26 || src.reg == 27);

                        ins->src[0] = SSA_FIXED_REGISTER(src.reg);
                        offset_swizzle(ins->swizzle[0], src.offset, src.shift, 0, 0);
               } else {
                        struct phys_reg dst = index_to_reg(ctx, l, ins->dest, dest_shift);

                        ins->dest = SSA_FIXED_REGISTER(dst.reg);
                        offset_swizzle(ins->swizzle[0], 0, 2, 2, dst.offset);
                        mir_set_bytemask(ins, mir_bytemask(ins) << dst.offset);
                }

                /* We also follow up by actual arguments */

                for (int i = 1; i <= 3; i++) {
                        unsigned src_index = ins->src[i];
                        if (src_index != ~0) {
                                struct phys_reg src = index_to_reg(ctx, l, src_index, src_shift[i]);
                                unsigned component = src.offset >> src.shift;
                                assert(component << src.shift == src.offset);
                                ins->src[i] = SSA_FIXED_REGISTER(src.reg);
                                ins->swizzle[i][0] += component;
                        }
                }

                break;
        }

        case TAG_TEXTURE_4: {
                if (ins->op == TEXTURE_OP_BARRIER)
                        break;

                /* Grab RA results */
                struct phys_reg dest = index_to_reg(ctx, l, ins->dest, dest_shift);
                struct phys_reg coord = index_to_reg(ctx, l, ins->src[1], src_shift[1]);
                struct phys_reg lod = index_to_reg(ctx, l, ins->src[2], src_shift[2]);
                struct phys_reg offset = index_to_reg(ctx, l, ins->src[3], src_shift[3]);

                /* First, install the texture coordinate */
                if (ins->src[1] != ~0)
                        ins->src[1] = SSA_FIXED_REGISTER(coord.reg);
                offset_swizzle(ins->swizzle[1], coord.offset, coord.shift, dest.shift, 0);

                /* Next, install the destination */
                if (ins->dest != ~0)
                        ins->dest = SSA_FIXED_REGISTER(dest.reg);
                offset_swizzle(ins->swizzle[0], 0, 2, dest.shift,
                                dest_shift == 1 ? dest.offset % 8 :
                                dest.offset);
                mir_set_bytemask(ins, mir_bytemask(ins) << dest.offset);

                /* If there is a register LOD/bias, use it */
                if (ins->src[2] != ~0) {
                        assert(!(lod.offset & 3));
                        ins->src[2] = SSA_FIXED_REGISTER(lod.reg);
                        ins->swizzle[2][0] = lod.offset / 4;
                }

                /* If there is an offset register, install it */
                if (ins->src[3] != ~0) {
                        ins->src[3] = SSA_FIXED_REGISTER(offset.reg);
                        ins->swizzle[3][0] = offset.offset / 4;
                }

                break;
        }

        default:
                break;
        }
}

static void
install_registers(compiler_context *ctx, struct lcra_state *l)
{
        mir_foreach_instr_global(ctx, ins)
                install_registers_instr(ctx, l, ins);
}


/* If register allocation fails, find the best spill node */

static signed
mir_choose_spill_node(
                compiler_context *ctx,
                struct lcra_state *l)
{
        /* We can't spill a previously spilled value or an unspill */

        mir_foreach_instr_global(ctx, ins) {
                if (ins->no_spill & (1 << l->spill_class)) {
                        lcra_set_node_spill_cost(l, ins->dest, -1);

                        if (l->spill_class != REG_CLASS_WORK) {
                                mir_foreach_src(ins, s)
                                        lcra_set_node_spill_cost(l, ins->src[s], -1);
                        }
                }
        }

        return lcra_get_best_spill_node(l);
}

/* Once we've chosen a spill node, spill it */

static void
mir_spill_register(
                compiler_context *ctx,
                unsigned spill_node,
                unsigned spill_class,
                unsigned *spill_count)
{
        if (spill_class == REG_CLASS_WORK && ctx->is_blend)
                unreachable("Blend shader spilling is currently unimplemented");

        unsigned spill_index = ctx->temp_count;

        /* We have a spill node, so check the class. Work registers
         * legitimately spill to TLS, but special registers just spill to work
         * registers */

        bool is_special = spill_class != REG_CLASS_WORK;
        bool is_special_w = spill_class == REG_CLASS_TEXW;

        /* Allocate TLS slot (maybe) */
        unsigned spill_slot = !is_special ? (*spill_count)++ : 0;

        /* For TLS, replace all stores to the spilled node. For
         * special reads, just keep as-is; the class will be demoted
         * implicitly. For special writes, spill to a work register */

        if (!is_special || is_special_w) {
                if (is_special_w)
                        spill_slot = spill_index++;

                mir_foreach_block(ctx, _block) {
                midgard_block *block = (midgard_block *) _block;
                mir_foreach_instr_in_block_safe(block, ins) {
                        if (ins->dest != spill_node) continue;

                        midgard_instruction st;

                        /* Note: it's important to match the mask of the spill
                         * with the mask of the instruction whose destination
                         * we're spilling, or otherwise we'll read invalid
                         * components and can fail RA in a subsequent iteration
                         */

                        if (is_special_w) {
                                st = v_mov(spill_node, spill_slot);
                                st.no_spill |= (1 << spill_class);
                                st.mask = ins->mask;
                                st.dest_type = st.src_types[1] = ins->dest_type;
                        } else {
                                ins->dest = spill_index++;
                                ins->no_spill |= (1 << spill_class);
                                st = v_load_store_scratch(ins->dest, spill_slot, true, ins->mask);
                        }

                        /* Hint: don't rewrite this node */
                        st.hint = true;

                        mir_insert_instruction_after_scheduled(ctx, block, ins, st);

                        if (!is_special)
                                ctx->spills++;
                }
                }
        }

        /* For special reads, figure out how many bytes we need */
        unsigned read_bytemask = 0;

        mir_foreach_instr_global_safe(ctx, ins) {
                read_bytemask |= mir_bytemask_of_read_components(ins, spill_node);
        }

        /* Insert a load from TLS before the first consecutive
         * use of the node, rewriting to use spilled indices to
         * break up the live range. Or, for special, insert a
         * move. Ironically the latter *increases* register
         * pressure, but the two uses of the spilling mechanism
         * are somewhat orthogonal. (special spilling is to use
         * work registers to back special registers; TLS
         * spilling is to use memory to back work registers) */

        mir_foreach_block(ctx, _block) {
                midgard_block *block = (midgard_block *) _block;
                mir_foreach_instr_in_block(block, ins) {
                        /* We can't rewrite the moves used to spill in the
                         * first place. These moves are hinted. */
                        if (ins->hint) continue;

                        /* If we don't use the spilled value, nothing to do */
                        if (!mir_has_arg(ins, spill_node)) continue;

                        unsigned index = 0;

                        if (!is_special_w) {
                                index = ++spill_index;

                                midgard_instruction *before = ins;
                                midgard_instruction st;

                                if (is_special) {
                                        /* Move */
                                        st = v_mov(spill_node, index);
                                        st.no_spill |= (1 << spill_class);
                                } else {
                                        /* TLS load */
                                        st = v_load_store_scratch(index, spill_slot, false, 0xF);
                                }

                                /* Mask the load based on the component count
                                 * actually needed to prevent RA loops */

                                st.mask = mir_from_bytemask(mir_round_bytemask_up(
                                                        read_bytemask, 32), 32);

                                mir_insert_instruction_before_scheduled(ctx, block, before, st);
                        } else {
                                /* Special writes already have their move spilled in */
                                index = spill_slot;
                        }


                        /* Rewrite to use */
                        mir_rewrite_index_src_single(ins, spill_node, index);

                        if (!is_special)
                                ctx->fills++;
                }
        }

        /* Reset hints */

        mir_foreach_instr_global(ctx, ins) {
                ins->hint = false;
        }
}

/* Run register allocation in a loop, spilling until we succeed */

void
mir_ra(compiler_context *ctx)
{
        struct lcra_state *l = NULL;
        bool spilled = false;
        int iter_count = 1000; /* max iterations */

        /* Number of 128-bit slots in memory we've spilled into */
        unsigned spill_count = 0;


        mir_create_pipeline_registers(ctx);

        do {
                if (spilled) {
                        signed spill_node = mir_choose_spill_node(ctx, l);

                        if (spill_node == -1) {
                                fprintf(stderr, "ERROR: Failed to choose spill node\n");
                                lcra_free(l);
                                return;
                        }

                        mir_spill_register(ctx, spill_node, l->spill_class, &spill_count);
                }

                mir_squeeze_index(ctx);
                mir_invalidate_liveness(ctx);

                if (l) {
                        lcra_free(l);
                        l = NULL;
                }

                l = allocate_registers(ctx, &spilled);
        } while(spilled && ((iter_count--) > 0));

        if (iter_count <= 0) {
                fprintf(stderr, "panfrost: Gave up allocating registers, rendering will be incomplete\n");
                assert(0);
        }

        /* Report spilling information. spill_count is in 128-bit slots (vec4 x
         * fp32), but tls_size is in bytes, so multiply by 16 */

        ctx->tls_size = spill_count * 16;

        install_registers(ctx, l);

        lcra_free(l);
}