summaryrefslogtreecommitdiff
path: root/src/panfrost/midgard/midgard_ra.c
blob: 39ced65e272621141edfb6d04754f334b2ea7c47 (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
/*
 * 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 "lcra.h"

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

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

        /* Number of bytes in a component of this register */
        unsigned size;
};

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

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

        signed reg_comp = reg_offset / srcsize;
        signed dst_comp = dst_offset / srcsize;

        unsigned max_component = (16 / srcsize) - 1;

        assert(reg_comp * srcsize == reg_offset);
        assert(dst_comp * srcsize == 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, midgard_reg_mode size)
{
        struct phys_reg r = {
                .reg = reg,
                .offset = 0,
                .size = mir_bytes_for_mode(size)
        };

        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, midgard_reg_mode size)
{
        /* Check for special cases */
        if (reg == ~0)
                return default_phys_reg(REGISTER_UNUSED, size);
        else if (reg >= SSA_FIXED_MINIMUM)
                return default_phys_reg(SSA_REG_FROM_FIXED(reg), size);
        else if (!l)
                return default_phys_reg(REGISTER_UNUSED, size);

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

        /* 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
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
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]);
                        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++;

                        midgard_instruction m = hazard_write ?
                                v_mov(idx, i) : v_mov(i, 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;
                                } else {
                                        if (!mir_has_arg(pre_use, i))
                                                continue;
                                }

                                if (hazard_write) {
                                        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 {
                                        idx = spill_idx++;
                                        m = v_mov(i, idx);
                                        m.mask = mir_from_bytemask(mir_bytemask_of_read_components(pre_use, i), midgard_reg_mode_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);
}

/* We register allocate after scheduling, so we need to ensure instructions
 * executing in parallel within a segment of a bundle don't clobber each
 * other's registers. This is mostly a non-issue thanks to scheduling, but
 * there are edge cases. In particular, after a register is written in a
 * segment, it interferes with anything reading. */

static void
mir_compute_segment_interference(
                compiler_context *ctx,
                struct lcra_state *l,
                midgard_bundle *bun,
                unsigned pivot,
                unsigned i)
{
        for (unsigned j = pivot; j < i; ++j) {
                mir_foreach_src(bun->instructions[j], s) {
                        if (bun->instructions[j]->src[s] >= ctx->temp_count)
                                continue;

                        for (unsigned q = pivot; q < i; ++q) {
                                if (bun->instructions[q]->dest >= ctx->temp_count)
                                        continue;

                                /* See dEQP-GLES2.functional.shaders.return.output_write_in_func_dynamic_fragment */

                                if (q >= j) {
                                        if (!(bun->instructions[j]->unit == UNIT_SMUL && bun->instructions[q]->unit == UNIT_VLUT))
                                                continue;
                                }

                                unsigned mask = mir_bytemask(bun->instructions[q]);
                                unsigned rmask = mir_bytemask_of_read_components(bun->instructions[j], bun->instructions[j]->src[s]);
                                lcra_add_node_interference(l, bun->instructions[q]->dest, mask, bun->instructions[j]->src[s], rmask);
                        }
                }
        }
}

static void
mir_compute_bundle_interference(
                compiler_context *ctx,
                struct lcra_state *l,
                midgard_bundle *bun)
{
        if (!IS_ALU(bun->tag))
                return;

        bool old = bun->instructions[0]->unit >= UNIT_VADD;
        unsigned pivot = 0;

        for (unsigned i = 1; i < bun->instruction_count; ++i) {
                bool new = bun->instructions[i]->unit >= UNIT_VADD;

                if (old != new) {
                        mir_compute_segment_interference(ctx, l, bun, 0, i);
                        pivot = i;
                        break;
                }
        }

        mir_compute_segment_interference(ctx, l, bun, pivot, bun->instruction_count);
}

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);

        /* 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) {
                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);
                }

                mir_foreach_bundle_in_block(blk, bun)
                        mir_compute_bundle_interference(ctx, l, bun);

                free(live);
        }
}

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

struct lcra_state *
allocate_registers(compiler_context *ctx, bool *spilled)
{
        /* The number of vec4 work registers available depends on when the
         * uniforms start, so compute that first */
        int work_count = 16 - MAX2((ctx->uniform_cutoff - 8), 0);

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

        if (!ctx->temp_count)
                return NULL;

        struct lcra_state *l = lcra_alloc_equations(ctx->temp_count, 1, 8, 16, 5);

        /* 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);

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

        mir_foreach_instr_global(ctx, ins) {
                if (ins->dest >= SSA_FIXED_MINIMUM) continue;

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

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

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

                lcra_set_alignment(l, dest, 2); /* (1 << 2) = 4 */

                /* XXX: Ensure swizzles align the right way with more LCRA constraints? */
                if (ins->type == TAG_ALU_4 && ins->alu.reg_mode != midgard_reg_mode_32)
                        lcra_set_alignment(l, dest, 3); /* (1 << 3) = 8 */
        }

        for (unsigned i = 0; i < ctx->temp_count; ++i)
                lcra_restrict_range(l, i, (found_class[i] + 1) * 4);
        
        free(found_class);

        /* 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);

                        if (OP_IS_VEC4_ONLY(ins->load_store.op))
                                lcra_restrict_range(l, ins->dest, 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);
                }
        }

        /* 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]));
        }

        /* Mark writeout to r0 */
        mir_foreach_instr_global(ctx, ins) {
                if (ins->compact_branch && ins->writeout && ins->src[0] < ctx->temp_count)
                        l->solutions[ins->src[0]] = 0;
        }
        
        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)
{
        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], mir_srcsize(ins, 0));
                struct phys_reg src2 = index_to_reg(ctx, l, ins->src[1], mir_srcsize(ins, 1));
                struct phys_reg dest = index_to_reg(ctx, l, ins->dest, mir_typesize(ins));

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

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

                offset_swizzle(ins->swizzle[0], src1.offset, src1.size, dest_offset);

                ins->registers.src1_reg = src1.reg;

                ins->registers.src2_imm = ins->has_inline_constant;

                if (ins->has_inline_constant) {
                        /* Encode inline 16-bit constant. See disassembler for
                         * where the algorithm is from */

                        ins->registers.src2_reg = ins->inline_constant >> 11;

                        int lower_11 = ins->inline_constant & ((1 << 12) - 1);
                        uint16_t imm = ((lower_11 >> 8) & 0x7) |
                                       ((lower_11 & 0xFF) << 3);

                        ins->alu.src2 = imm << 2;
                } else {
                        midgard_vector_alu_src mod2 =
                                vector_alu_from_unsigned(ins->alu.src2);
                        offset_swizzle(ins->swizzle[1], src2.offset, src2.size, dest_offset);
                        ins->alu.src2 = vector_alu_srco_unsigned(mod2);

                        ins->registers.src2_reg = src2.reg;
                }

                ins->registers.out_reg = 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->load_store.op);

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

                        ins->load_store.reg = src.reg - 26;
                        offset_swizzle(ins->swizzle[0], src.offset, src.size, 0);
               } else {
                        struct phys_reg dst = index_to_reg(ctx, l, ins->dest, mir_typesize(ins));

                        ins->load_store.reg = dst.reg;
                        offset_swizzle(ins->swizzle[0], 0, 4, dst.offset);
                        mir_set_bytemask(ins, mir_bytemask(ins) << dst.offset);
                }

                /* We also follow up by actual arguments */

                unsigned src2 = ins->src[1];
                unsigned src3 = ins->src[2];

                if (src2 != ~0) {
                        struct phys_reg src = index_to_reg(ctx, l, src2, mir_srcsize(ins, 1));
                        unsigned component = src.offset / src.size;
                        assert(component * src.size == src.offset);
                        ins->load_store.arg_1 |= midgard_ldst_reg(src.reg, component);
                }

                if (src3 != ~0) {
                        struct phys_reg src = index_to_reg(ctx, l, src3, mir_srcsize(ins, 2));
                        unsigned component = src.offset / src.size;
                        assert(component * src.size == src.offset);
                        ins->load_store.arg_2 |= midgard_ldst_reg(src.reg, component);
                }
 
                break;
        }

        case TAG_TEXTURE_4: {
                /* Grab RA results */
                struct phys_reg dest = index_to_reg(ctx, l, ins->dest, mir_typesize(ins));
                struct phys_reg coord = index_to_reg(ctx, l, ins->src[1], mir_srcsize(ins, 1));
                struct phys_reg lod = index_to_reg(ctx, l, ins->src[2], mir_srcsize(ins, 2));

                assert(dest.reg == 28 || dest.reg == 29);
                assert(coord.reg == 28 || coord.reg == 29);

                /* First, install the texture coordinate */
                ins->texture.in_reg_full = 1;
                ins->texture.in_reg_upper = 0;
                ins->texture.in_reg_select = coord.reg - 28;
                offset_swizzle(ins->swizzle[1], coord.offset, coord.size, 0);

                /* Next, install the destination */
                ins->texture.out_full = 1;
                ins->texture.out_upper = 0;
                ins->texture.out_reg_select = dest.reg - 28;
                offset_swizzle(ins->swizzle[0], 0, 4, 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));
                        midgard_tex_register_select sel = {
                                .select = lod.reg,
                                .full = 1,
                                .component = lod.offset / 4
                        };

                        uint8_t packed;
                        memcpy(&packed, &sel, sizeof(packed));
                        ins->texture.bias = packed;
                }

                break;
        }

        default:
                break;
        }
}

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