disassemble.c revision 7ec681f3
1/* Author(s):
2 *   Connor Abbott
3 *   Alyssa Rosenzweig
4 *
5 * Copyright (c) 2013 Connor Abbott (connor@abbott.cx)
6 * Copyright (c) 2018 Alyssa Rosenzweig (alyssa@rosenzweig.io)
7 * Copyright (C) 2019-2020 Collabora, Ltd.
8 *
9 * Permission is hereby granted, free of charge, to any person obtaining a copy
10 * of this software and associated documentation files (the "Software"), to deal
11 * in the Software without restriction, including without limitation the rights
12 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
13 * copies of the Software, and to permit persons to whom the Software is
14 * furnished to do so, subject to the following conditions:
15 *
16 * The above copyright notice and this permission notice shall be included in
17 * all copies or substantial portions of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
22 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
24 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
25 * THE SOFTWARE.
26 */
27
28#include <stdio.h>
29#include <stdint.h>
30#include <stdlib.h>
31#include <assert.h>
32#include <inttypes.h>
33#include <ctype.h>
34#include <string.h>
35#include "midgard.h"
36#include "midgard_ops.h"
37#include "midgard_quirks.h"
38#include "disassemble.h"
39#include "helpers.h"
40#include "util/bitscan.h"
41#include "util/half_float.h"
42#include "util/u_math.h"
43
44#define DEFINE_CASE(define, str) case define: { fprintf(fp, str); break; }
45
46/* These are not mapped to hardware values, they just represent the possible
47 * implicit arg modifiers that some midgard opcodes have, which can be decoded
48 * from the opcodes via midgard_{alu,ldst,tex}_special_arg_mod() */
49typedef enum {
50        midgard_arg_mod_none = 0,
51        midgard_arg_mod_inv,
52        midgard_arg_mod_x2,
53} midgard_special_arg_mod;
54
55typedef struct {
56        unsigned *midg_tags;
57        struct midgard_disasm_stats midg_stats;
58
59        /* For static analysis to ensure all registers are written at least once before
60         * use along the source code path (TODO: does this break done for complex CF?)
61         */
62
63        uint16_t midg_ever_written;
64} disassemble_context;
65
66/* Transform an expanded writemask (duplicated 8-bit format) into its condensed
67 * form (one bit per component) */
68
69static inline unsigned
70condense_writemask(unsigned expanded_mask,
71                   unsigned bits_per_component)
72{
73        if (bits_per_component == 8) {
74                /* Duplicate every bit to go from 8 to 16-channel wrmask */
75                unsigned omask = 0;
76
77                for (unsigned i = 0; i < 8; ++i) {
78                        if (expanded_mask & (1 << i))
79                                omask |= (3 << (2 * i));
80                }
81
82                return omask;
83        }
84
85        unsigned slots_per_component = bits_per_component / 16;
86        unsigned max_comp = (16 * 8) / bits_per_component;
87        unsigned condensed_mask = 0;
88
89        for (unsigned i = 0; i < max_comp; i++) {
90                if (expanded_mask & (1 << (i * slots_per_component)))
91                        condensed_mask |= (1 << i);
92        }
93
94        return condensed_mask;
95}
96
97static bool
98print_alu_opcode(FILE *fp, midgard_alu_op op)
99{
100        if (alu_opcode_props[op].name)
101                fprintf(fp, "%s", alu_opcode_props[op].name);
102        else
103                fprintf(fp, "alu_op_%02X", op);
104
105        /* For constant analysis */
106        return midgard_is_integer_op(op);
107}
108
109static void
110print_ld_st_opcode(FILE *fp, midgard_load_store_op op)
111{
112        if (load_store_opcode_props[op].name)
113                fprintf(fp, "%s", load_store_opcode_props[op].name);
114        else
115                fprintf(fp, "ldst_op_%02X", op);
116}
117
118static void
119validate_sampler_type(enum mali_texture_op op, enum mali_sampler_type sampler_type)
120{
121        if (op == midgard_tex_op_mov || op == midgard_tex_op_barrier)
122                assert(sampler_type == 0);
123        else
124                assert(sampler_type > 0);
125}
126
127static void
128validate_expand_mode(midgard_src_expand_mode expand_mode,
129                     midgard_reg_mode reg_mode)
130{
131        switch (expand_mode) {
132        case midgard_src_passthrough:
133                break;
134
135        case midgard_src_rep_low:
136                assert(reg_mode == midgard_reg_mode_8 ||
137                       reg_mode == midgard_reg_mode_16);
138                break;
139
140        case midgard_src_rep_high:
141                assert(reg_mode == midgard_reg_mode_8 ||
142                       reg_mode == midgard_reg_mode_16);
143                break;
144
145        case midgard_src_swap:
146                assert(reg_mode == midgard_reg_mode_8 ||
147                       reg_mode == midgard_reg_mode_16);
148                break;
149
150        case midgard_src_expand_low:
151                assert(reg_mode != midgard_reg_mode_8);
152                break;
153
154        case midgard_src_expand_high:
155                assert(reg_mode != midgard_reg_mode_8);
156                break;
157
158        case midgard_src_expand_low_swap:
159                assert(reg_mode == midgard_reg_mode_16);
160                break;
161
162        case midgard_src_expand_high_swap:
163                assert(reg_mode == midgard_reg_mode_16);
164                break;
165
166        default:
167                unreachable("Invalid expand mode");
168                break;
169        }
170}
171
172static void
173print_alu_reg(disassemble_context *ctx, FILE *fp, unsigned reg, bool is_write)
174{
175        unsigned uniform_reg = 23 - reg;
176        bool is_uniform = false;
177
178        /* For r8-r15, it could be a work or uniform. We distinguish based on
179         * the fact work registers are ALWAYS written before use, but uniform
180         * registers are NEVER written before use. */
181
182        if ((reg >= 8 && reg < 16) && !(ctx->midg_ever_written & (1 << reg)))
183                is_uniform = true;
184
185        /* r16-r23 are always uniform */
186
187        if (reg >= 16 && reg <= 23)
188                is_uniform = true;
189
190        /* Update the uniform count appropriately */
191
192        if (is_uniform)
193                ctx->midg_stats.uniform_count =
194                        MAX2(uniform_reg + 1, ctx->midg_stats.uniform_count);
195
196        if (reg == REGISTER_UNUSED || reg == REGISTER_UNUSED + 1)
197                fprintf(fp, "TMP%u", reg - REGISTER_UNUSED);
198        else if (reg == REGISTER_TEXTURE_BASE || reg == REGISTER_TEXTURE_BASE + 1)
199                fprintf(fp, "%s%u", is_write ? "AT" : "TA",  reg - REGISTER_TEXTURE_BASE);
200        else if (reg == REGISTER_LDST_BASE || reg == REGISTER_LDST_BASE + 1)
201                fprintf(fp, "AL%u", reg - REGISTER_LDST_BASE);
202        else if (is_uniform)
203                fprintf(fp, "U%u", uniform_reg);
204        else if (reg == 31 && !is_write)
205                fprintf(fp, "PC_SP");
206        else
207                fprintf(fp, "R%u", reg);
208}
209
210static void
211print_ldst_write_reg(FILE *fp, unsigned reg)
212{
213        switch (reg) {
214        case 26:
215        case 27:
216                fprintf(fp, "AL%u", reg - REGISTER_LDST_BASE);
217                break;
218        case 28:
219        case 29:
220                fprintf(fp, "AT%u", reg - REGISTER_TEXTURE_BASE);
221                break;
222        case 31:
223                fprintf(fp, "PC_SP");
224                break;
225        default:
226                fprintf(fp, "R%d", reg);
227                break;
228        }
229}
230
231static void
232print_ldst_read_reg(FILE *fp, unsigned reg)
233{
234        switch (reg) {
235        case 0:
236        case 1:
237                fprintf(fp, "AL%u", reg);
238                break;
239        case 2:
240                fprintf(fp, "PC_SP");
241                break;
242        case 3:
243                fprintf(fp, "LOCAL_STORAGE_PTR");
244                break;
245        case 4:
246                fprintf(fp, "LOCAL_THREAD_ID");
247                break;
248        case 5:
249                fprintf(fp, "GROUP_ID");
250                break;
251        case 6:
252                fprintf(fp, "GLOBAL_THREAD_ID");
253                break;
254        case 7:
255                fprintf(fp, "0");
256                break;
257        default:
258                unreachable("Invalid load/store register read");
259        }
260}
261
262static void
263print_tex_reg(FILE *fp, unsigned reg, bool is_write)
264{
265        char *str = is_write ? "TA" : "AT";
266        int select = reg & 1;
267
268        switch (reg) {
269        case 0:
270        case 1:
271                fprintf(fp, "R%d", select);
272                break;
273        case 26:
274        case 27:
275                fprintf(fp, "AL%d", select);
276                break;
277        case 28:
278        case 29:
279                fprintf(fp, "%s%d", str, select);
280                break;
281        default:
282                unreachable("Invalid texture register");
283        }
284}
285
286
287static char *outmod_names_float[4] = {
288        "",
289        ".clamp_0_inf",
290        ".clamp_m1_1",
291        ".clamp_0_1"
292};
293
294static char *outmod_names_int[4] = {
295        ".ssat",
296        ".usat",
297        ".keeplo",
298        ".keephi"
299};
300
301static char *srcmod_names_int[4] = {
302        ".sext",
303        ".zext",
304        ".replicate",
305        ".lshift",
306};
307
308static char *argmod_names[3] = {
309        "",
310        ".inv",
311        ".x2",
312};
313
314static char *index_format_names[4] = {
315        "",
316        ".u64",
317        ".u32",
318        ".s32"
319};
320
321static void
322print_outmod(FILE *fp, unsigned outmod, bool is_int)
323{
324        fprintf(fp, "%s", is_int ? outmod_names_int[outmod] :
325                outmod_names_float[outmod]);
326}
327
328static void
329print_alu_outmod(FILE *fp, unsigned outmod, bool is_int, bool half)
330{
331        if (is_int && !half) {
332                assert(outmod == midgard_outmod_keeplo);
333                return;
334        }
335
336        if (!is_int && half)
337                fprintf(fp, ".shrink");
338
339        print_outmod(fp, outmod, is_int);
340}
341
342/* arg == 0 (dest), arg == 1 (src1), arg == 2 (src2) */
343static midgard_special_arg_mod
344midgard_alu_special_arg_mod(midgard_alu_op op, unsigned arg) {
345        midgard_special_arg_mod mod = midgard_arg_mod_none;
346
347        switch (op) {
348        case midgard_alu_op_ishladd:
349        case midgard_alu_op_ishlsub:
350                if (arg == 1) mod = midgard_arg_mod_x2;
351                break;
352
353        default:
354                break;
355        }
356
357        return mod;
358}
359
360static void
361print_quad_word(FILE *fp, uint32_t *words, unsigned tabs)
362{
363        unsigned i;
364
365        for (i = 0; i < 4; i++)
366                fprintf(fp, "0x%08X%s ", words[i], i == 3 ? "" : ",");
367
368        fprintf(fp, "\n");
369}
370
371static const char components[16] = "xyzwefghijklmnop";
372
373static int
374bits_for_mode(midgard_reg_mode mode)
375{
376        switch (mode) {
377        case midgard_reg_mode_8:
378                return 8;
379        case midgard_reg_mode_16:
380                return 16;
381        case midgard_reg_mode_32:
382                return 32;
383        case midgard_reg_mode_64:
384                return 64;
385        default:
386                unreachable("Invalid reg mode");
387                return 0;
388        }
389}
390
391static int
392bits_for_mode_halved(midgard_reg_mode mode, bool half)
393{
394        unsigned bits = bits_for_mode(mode);
395
396        if (half)
397                bits >>= 1;
398
399        return bits;
400}
401
402static void
403print_vec_selectors_64(FILE *fp, unsigned swizzle,
404                       midgard_reg_mode reg_mode,
405                       midgard_src_expand_mode expand_mode,
406                       unsigned selector_offset, uint8_t mask)
407{
408        bool expands = INPUT_EXPANDS(expand_mode);
409
410        unsigned comp_skip = expands ? 1 : 2;
411        unsigned mask_bit = 0;
412        for (unsigned i = selector_offset; i < 4; i += comp_skip, mask_bit += 4) {
413                if (!(mask & (1 << mask_bit))) continue;
414
415                unsigned a = (swizzle >> (i * 2)) & 3;
416
417                if (INPUT_EXPANDS(expand_mode)) {
418                        fprintf(fp, "%c", components[a]);
419                        continue;
420                }
421
422                unsigned b = (swizzle >> ((i+1) * 2)) & 3;
423
424                /* Normally we're adjacent, but if there's an issue,
425                 * don't make it ambiguous */
426
427                if (b == a + 1)
428                        fprintf(fp, "%c", a >> 1 ? 'Y' : 'X');
429                else
430                        fprintf(fp, "[%c%c]", components[a], components[b]);
431        }
432}
433
434static void
435print_vec_selectors(FILE *fp, unsigned swizzle,
436                    midgard_reg_mode reg_mode,
437                    unsigned selector_offset, uint8_t mask,
438                    unsigned *mask_offset)
439{
440        assert(reg_mode != midgard_reg_mode_64);
441
442        unsigned mask_skip = MAX2(bits_for_mode(reg_mode) / 16, 1);
443
444        bool is_vec16 = reg_mode == midgard_reg_mode_8;
445
446        for (unsigned i = 0; i < 4; i++, *mask_offset += mask_skip) {
447                if (!(mask & (1 << *mask_offset))) continue;
448
449                unsigned c = (swizzle >> (i * 2)) & 3;
450
451                /* Vec16 has two components per swizzle selector. */
452                if (is_vec16)
453                        c *= 2;
454
455                c += selector_offset;
456
457                fprintf(fp, "%c", components[c]);
458                if (is_vec16)
459                        fprintf(fp, "%c", components[c+1]);
460        }
461}
462
463static void
464print_vec_swizzle(FILE *fp, unsigned swizzle,
465                  midgard_src_expand_mode expand,
466                  midgard_reg_mode mode,
467                  uint8_t mask)
468{
469        unsigned bits = bits_for_mode_halved(mode, INPUT_EXPANDS(expand));
470
471        /* Swizzle selectors are divided in two halves that are always
472         * mirrored, the only difference is the starting component offset.
473         * The number represents an offset into the components[] array. */
474        unsigned first_half = 0;
475        unsigned second_half = (128 / bits) / 2; /* only used for 8 and 16-bit */
476
477        switch (expand) {
478        case midgard_src_passthrough:
479                if (swizzle == 0xE4) return; /* identity swizzle */
480                break;
481
482        case midgard_src_expand_low:
483                second_half /= 2;
484                break;
485
486        case midgard_src_expand_high:
487                first_half = second_half;
488                second_half += second_half / 2;
489                break;
490
491        /* The rest of the cases are only used for 8 and 16-bit */
492
493        case midgard_src_rep_low:
494                second_half = 0;
495                break;
496
497        case midgard_src_rep_high:
498                first_half = second_half;
499                break;
500
501        case midgard_src_swap:
502                first_half = second_half;
503                second_half = 0;
504                break;
505
506        case midgard_src_expand_low_swap:
507                first_half = second_half / 2;
508                second_half = 0;
509                break;
510
511        case midgard_src_expand_high_swap:
512                first_half = second_half + second_half / 2;
513                break;
514
515        default:
516                unreachable("Invalid expand mode");
517                break;
518        }
519
520        fprintf(fp, ".");
521
522        /* Vec2 are weird so we use a separate function to simplify things. */
523        if (mode == midgard_reg_mode_64) {
524                print_vec_selectors_64(fp, swizzle, mode, expand, first_half, mask);
525                return;
526        }
527
528        unsigned mask_offs = 0;
529        print_vec_selectors(fp, swizzle, mode, first_half, mask, &mask_offs);
530        if (mode == midgard_reg_mode_8 || mode == midgard_reg_mode_16)
531                print_vec_selectors(fp, swizzle, mode, second_half, mask, &mask_offs);
532}
533
534static void
535print_scalar_constant(FILE *fp, unsigned src_binary,
536                      const midgard_constants *consts,
537                      midgard_scalar_alu *alu)
538{
539        midgard_scalar_alu_src *src = (midgard_scalar_alu_src *)&src_binary;
540        assert(consts != NULL);
541
542        fprintf(fp, "#");
543        mir_print_constant_component(fp, consts, src->component,
544                                     src->full ?
545                                     midgard_reg_mode_32 : midgard_reg_mode_16,
546                                     false, src->mod, alu->op);
547}
548
549static void
550print_vector_constants(FILE *fp, unsigned src_binary,
551                       const midgard_constants *consts,
552                       midgard_vector_alu *alu)
553{
554        midgard_vector_alu_src *src = (midgard_vector_alu_src *)&src_binary;
555        bool expands = INPUT_EXPANDS(src->expand_mode);
556        unsigned bits = bits_for_mode_halved(alu->reg_mode, expands);
557        unsigned max_comp = (sizeof(*consts) * 8) / bits;
558        unsigned comp_mask, num_comp = 0;
559
560        assert(consts);
561        assert(max_comp <= 16);
562
563        comp_mask = effective_writemask(alu->op, condense_writemask(alu->mask, bits));
564        num_comp = util_bitcount(comp_mask);
565
566        fprintf(fp, "<");
567        bool first = true;
568
569	for (unsigned i = 0; i < max_comp; ++i) {
570                if (!(comp_mask & (1 << i))) continue;
571
572                unsigned c = (src->swizzle >> (i * 2)) & 3;
573
574                if (bits == 16 && !expands) {
575                        bool upper = i >= 4;
576
577                        switch (src->expand_mode) {
578                        case midgard_src_passthrough:
579                                c += upper * 4;
580                                break;
581                        case midgard_src_rep_low:
582                                break;
583                        case midgard_src_rep_high:
584                                c += 4;
585                                break;
586                        case midgard_src_swap:
587                                c += !upper * 4;
588                                break;
589                        default:
590                                unreachable("invalid expand mode");
591                                break;
592                        }
593                } else if (bits == 32 && !expands) {
594                        /* Implicitly ok */
595                } else if (bits == 64 && !expands) {
596                        /* Implicitly ok */
597                } else if (bits == 8 && !expands) {
598                        bool upper = i >= 8;
599
600                        unsigned index = (i >> 1) & 3;
601                        unsigned base = (src->swizzle >> (index * 2)) & 3;
602                        c = base * 2;
603
604                        switch (src->expand_mode) {
605                        case midgard_src_passthrough:
606                                c += upper * 8;
607                                break;
608                        case midgard_src_rep_low:
609                                break;
610                        case midgard_src_rep_high:
611                                c += 8;
612                                break;
613                        case midgard_src_swap:
614                                c += !upper * 8;
615                                break;
616                        default:
617                                unreachable("invalid expand mode");
618                                break;
619                        }
620
621                        /* We work on twos, actually */
622                        if (i & 1)
623                                c++;
624                } else {
625                        printf(" (%u)", src->expand_mode);
626                }
627
628                if (first)
629                        first = false;
630                else
631                        fprintf(fp, ", ");
632
633                mir_print_constant_component(fp, consts, c, alu->reg_mode,
634                                             expands, src->mod, alu->op);
635        }
636
637        if (num_comp > 1)
638                fprintf(fp, ">");
639}
640
641static void
642print_srcmod(FILE *fp, bool is_int, bool expands, unsigned mod, bool scalar)
643{
644        /* Modifiers change meaning depending on the op's context */
645
646        if (is_int) {
647                if (expands)
648                        fprintf(fp, "%s", srcmod_names_int[mod]);
649        } else {
650                if (mod & MIDGARD_FLOAT_MOD_ABS)
651                        fprintf(fp, ".abs");
652                if (mod & MIDGARD_FLOAT_MOD_NEG)
653                        fprintf(fp, ".neg");
654                if (expands)
655                        fprintf(fp, ".widen");
656        }
657}
658
659static void
660print_vector_src(disassemble_context *ctx, FILE *fp, unsigned src_binary,
661                 midgard_reg_mode mode, unsigned reg,
662                 midgard_shrink_mode shrink_mode,
663                 uint8_t src_mask, bool is_int,
664                 midgard_special_arg_mod arg_mod)
665{
666        midgard_vector_alu_src *src = (midgard_vector_alu_src *)&src_binary;
667
668        validate_expand_mode(src->expand_mode, mode);
669
670        print_alu_reg(ctx, fp, reg, false);
671
672        print_vec_swizzle(fp, src->swizzle, src->expand_mode, mode, src_mask);
673
674        fprintf(fp, "%s", argmod_names[arg_mod]);
675
676        print_srcmod(fp, is_int, INPUT_EXPANDS(src->expand_mode), src->mod, false);
677}
678
679static uint16_t
680decode_vector_imm(unsigned src2_reg, unsigned imm)
681{
682        uint16_t ret;
683        ret = src2_reg << 11;
684        ret |= (imm & 0x7) << 8;
685        ret |= (imm >> 3) & 0xFF;
686        return ret;
687}
688
689static void
690print_immediate(FILE *fp, uint16_t imm, bool is_instruction_int)
691{
692        if (is_instruction_int)
693                fprintf(fp, "#%u", imm);
694        else
695                fprintf(fp, "#%g", _mesa_half_to_float(imm));
696}
697
698static void
699update_dest(disassemble_context *ctx, unsigned reg)
700{
701        /* We should record writes as marking this as a work register. Store
702         * the max register in work_count; we'll add one at the end */
703
704        if (reg < 16) {
705                ctx->midg_stats.work_count = MAX2(reg, ctx->midg_stats.work_count);
706                ctx->midg_ever_written |= (1 << reg);
707        }
708}
709
710static void
711print_dest(disassemble_context *ctx, FILE *fp, unsigned reg)
712{
713        update_dest(ctx, reg);
714        print_alu_reg(ctx, fp, reg, true);
715}
716
717/* For 16-bit+ masks, we read off from the 8-bit mask field. For 16-bit (vec8),
718 * it's just one bit per channel, easy peasy. For 32-bit (vec4), it's one bit
719 * per channel with one duplicate bit in the middle. For 64-bit (vec2), it's
720 * one-bit per channel with _3_ duplicate bits in the middle. Basically, just
721 * subdividing the 128-bit word in 16-bit increments. For 64-bit, we uppercase
722 * the mask to make it obvious what happened */
723
724static void
725print_alu_mask(FILE *fp, uint8_t mask, unsigned bits, midgard_shrink_mode shrink_mode)
726{
727        /* Skip 'complete' masks */
728
729        if (shrink_mode == midgard_shrink_mode_none && mask == 0xFF)
730                return;
731
732        fprintf(fp, ".");
733
734        unsigned skip = MAX2(bits / 16, 1);
735        bool uppercase = bits > 32;
736        bool tripped = false;
737
738        /* To apply an upper destination shrink_mode, we "shift" the alphabet.
739         * E.g. with an upper shrink_mode on 32-bit, instead of xyzw, print efgh.
740         * For upper 16-bit, instead of xyzwefgh, print ijklmnop */
741
742        const char *alphabet = components;
743
744        if (shrink_mode == midgard_shrink_mode_upper) {
745                assert(bits != 8);
746                alphabet += (128 / bits);
747        }
748
749        for (unsigned i = 0; i < 8; i += skip) {
750                bool a = (mask & (1 << i)) != 0;
751
752                for (unsigned j = 1; j < skip; ++j) {
753                        bool dupe = (mask & (1 << (i + j))) != 0;
754                        tripped |= (dupe != a);
755                }
756
757                if (a) {
758                        /* TODO: handle shrinking from 16-bit */
759                        unsigned comp_idx = bits == 8 ? i * 2 : i;
760                        char c = alphabet[comp_idx / skip];
761
762                        if (uppercase) {
763                                c = toupper(c);
764                                assert(c == 'X' || c == 'Y');
765                        }
766
767                        fprintf(fp, "%c", c);
768                        if (bits == 8)
769                                fprintf(fp, "%c", alphabet[comp_idx+1]);
770                }
771        }
772
773        if (tripped)
774                fprintf(fp, " /* %X */", mask);
775}
776
777/* TODO: 16-bit mode */
778static void
779print_ldst_mask(FILE *fp, unsigned mask, unsigned swizzle) {
780        fprintf(fp, ".");
781
782        for (unsigned i = 0; i < 4; ++i) {
783                bool write = (mask & (1 << i)) != 0;
784                unsigned c = (swizzle >> (i * 2)) & 3;
785                /* We can't omit the swizzle here since many ldst ops have a
786                 * combined swizzle/writemask, and it would be ambiguous to not
787                 * print the masked-out components. */
788                fprintf(fp, "%c", write ? components[c] : '~');
789        }
790}
791
792static void
793print_tex_mask(FILE *fp, unsigned mask, bool upper)
794{
795        if (mask == 0xF) {
796                if (upper)
797                        fprintf(fp, "'");
798
799                return;
800        }
801
802        fprintf(fp, ".");
803
804        for (unsigned i = 0; i < 4; ++i) {
805                bool a = (mask & (1 << i)) != 0;
806                if (a)
807                        fprintf(fp, "%c", components[i + (upper ? 4 : 0)]);
808        }
809}
810
811static void
812print_vector_field(disassemble_context *ctx, FILE *fp, const char *name,
813                   uint16_t *words, uint16_t reg_word,
814                   const midgard_constants *consts, unsigned tabs, bool verbose)
815{
816        midgard_reg_info *reg_info = (midgard_reg_info *)&reg_word;
817        midgard_vector_alu *alu_field = (midgard_vector_alu *) words;
818        midgard_reg_mode mode = alu_field->reg_mode;
819        midgard_alu_op op = alu_field->op;
820        unsigned shrink_mode = alu_field->shrink_mode;
821        bool is_int = midgard_is_integer_op(op);
822        bool is_int_out = midgard_is_integer_out_op(op);
823
824        if (verbose)
825                fprintf(fp, "%s.", name);
826
827        bool is_instruction_int = print_alu_opcode(fp, alu_field->op);
828
829        /* Print lane width */
830        fprintf(fp, ".%c%d", is_int_out ? 'i' : 'f', bits_for_mode(mode));
831
832        fprintf(fp, " ");
833
834        /* Mask denoting status of 8-lanes */
835        uint8_t mask = alu_field->mask;
836
837        /* First, print the destination */
838        print_dest(ctx, fp, reg_info->out_reg);
839
840        if (shrink_mode != midgard_shrink_mode_none) {
841                bool shrinkable = (mode != midgard_reg_mode_8);
842                bool known = shrink_mode != 0x3; /* Unused value */
843
844                if (!(shrinkable && known))
845                        fprintf(fp, "/* do%u */ ", shrink_mode);
846        }
847
848        /* Instructions like fdot4 do *not* replicate, ensure the
849         * mask is of only a single component */
850
851        unsigned rep = GET_CHANNEL_COUNT(alu_opcode_props[op].props);
852
853        if (rep) {
854                unsigned comp_mask = condense_writemask(mask, bits_for_mode(mode));
855                unsigned num_comp = util_bitcount(comp_mask);
856                if (num_comp != 1)
857                        fprintf(fp, "/* err too many components */");
858        }
859        print_alu_mask(fp, mask, bits_for_mode(mode), shrink_mode);
860
861        /* Print output modifiers */
862
863        print_alu_outmod(fp, alu_field->outmod, is_int_out, shrink_mode != midgard_shrink_mode_none);
864
865        /* Mask out unused components based on the writemask, but don't mask out
866         * components that are used for interlane instructions like fdot3. */
867        uint8_t src_mask =
868                rep ? expand_writemask(mask_of(rep), log2(128 / bits_for_mode(mode))) : mask;
869
870        fprintf(fp, ", ");
871
872        if (reg_info->src1_reg == REGISTER_CONSTANT)
873                print_vector_constants(fp, alu_field->src1, consts, alu_field);
874        else {
875                midgard_special_arg_mod argmod = midgard_alu_special_arg_mod(op, 1);
876                print_vector_src(ctx, fp, alu_field->src1, mode, reg_info->src1_reg,
877                                 shrink_mode, src_mask, is_int, argmod);
878        }
879
880        fprintf(fp, ", ");
881
882        if (reg_info->src2_imm) {
883                uint16_t imm = decode_vector_imm(reg_info->src2_reg, alu_field->src2 >> 2);
884                print_immediate(fp, imm, is_instruction_int);
885        } else if (reg_info->src2_reg == REGISTER_CONSTANT) {
886                print_vector_constants(fp, alu_field->src2, consts, alu_field);
887        } else {
888                midgard_special_arg_mod argmod = midgard_alu_special_arg_mod(op, 2);
889                print_vector_src(ctx, fp, alu_field->src2, mode, reg_info->src2_reg,
890                                 shrink_mode, src_mask, is_int, argmod);
891        }
892
893        ctx->midg_stats.instruction_count++;
894        fprintf(fp, "\n");
895}
896
897static void
898print_scalar_src(disassemble_context *ctx, FILE *fp, bool is_int, unsigned src_binary, unsigned reg)
899{
900        midgard_scalar_alu_src *src = (midgard_scalar_alu_src *)&src_binary;
901
902        print_alu_reg(ctx, fp, reg, false);
903
904        unsigned c = src->component;
905
906        if (src->full) {
907                assert((c & 1) == 0);
908                c >>= 1;
909        }
910
911        fprintf(fp, ".%c", components[c]);
912
913        print_srcmod(fp, is_int, !src->full, src->mod, true);
914}
915
916static uint16_t
917decode_scalar_imm(unsigned src2_reg, unsigned imm)
918{
919        uint16_t ret;
920        ret = src2_reg << 11;
921        ret |= (imm & 3) << 9;
922        ret |= (imm & 4) << 6;
923        ret |= (imm & 0x38) << 2;
924        ret |= imm >> 6;
925        return ret;
926}
927
928static void
929print_scalar_field(disassemble_context *ctx, FILE *fp, const char *name,
930                   uint16_t *words, uint16_t reg_word,
931                   const midgard_constants *consts, unsigned tabs, bool verbose)
932{
933        midgard_reg_info *reg_info = (midgard_reg_info *)&reg_word;
934        midgard_scalar_alu *alu_field = (midgard_scalar_alu *) words;
935        bool is_int = midgard_is_integer_op(alu_field->op);
936        bool is_int_out = midgard_is_integer_out_op(alu_field->op);
937        bool full = alu_field->output_full;
938
939        if (alu_field->unknown)
940                fprintf(fp, "scalar ALU unknown bit set\n");
941
942        if (verbose)
943                fprintf(fp, "%s.", name);
944
945        bool is_instruction_int = print_alu_opcode(fp, alu_field->op);
946
947        /* Print lane width, in this case the lane width is always 32-bit, but
948         * we print it anyway to make it consistent with the other instructions. */
949        fprintf(fp, ".%c32", is_int_out ? 'i' : 'f');
950
951        fprintf(fp, " ");
952
953        print_dest(ctx, fp, reg_info->out_reg);
954        unsigned c = alu_field->output_component;
955
956        if (full) {
957                assert((c & 1) == 0);
958                c >>= 1;
959        }
960
961        fprintf(fp, ".%c", components[c]);
962
963        print_alu_outmod(fp, alu_field->outmod, is_int_out, !full);
964
965        fprintf(fp, ", ");
966
967        if (reg_info->src1_reg == REGISTER_CONSTANT)
968                print_scalar_constant(fp, alu_field->src1, consts, alu_field);
969        else
970                print_scalar_src(ctx, fp, is_int, alu_field->src1, reg_info->src1_reg);
971
972        fprintf(fp, ", ");
973
974        if (reg_info->src2_imm) {
975                uint16_t imm = decode_scalar_imm(reg_info->src2_reg,
976                                                 alu_field->src2);
977                print_immediate(fp, imm, is_instruction_int);
978	} else if (reg_info->src2_reg == REGISTER_CONSTANT) {
979                print_scalar_constant(fp, alu_field->src2, consts, alu_field);
980        } else
981                print_scalar_src(ctx, fp, is_int, alu_field->src2, reg_info->src2_reg);
982
983        ctx->midg_stats.instruction_count++;
984        fprintf(fp, "\n");
985}
986
987static void
988print_branch_op(FILE *fp, unsigned op)
989{
990        switch (op) {
991        case midgard_jmp_writeout_op_branch_uncond:
992                fprintf(fp, "uncond.");
993                break;
994
995        case midgard_jmp_writeout_op_branch_cond:
996                fprintf(fp, "cond.");
997                break;
998
999        case midgard_jmp_writeout_op_writeout:
1000                fprintf(fp, "write.");
1001                break;
1002
1003        case midgard_jmp_writeout_op_tilebuffer_pending:
1004                fprintf(fp, "tilebuffer.");
1005                break;
1006
1007        case midgard_jmp_writeout_op_discard:
1008                fprintf(fp, "discard.");
1009                break;
1010
1011        default:
1012                fprintf(fp, "unk%u.", op);
1013                break;
1014        }
1015}
1016
1017static void
1018print_branch_cond(FILE *fp, int cond)
1019{
1020        switch (cond) {
1021        case midgard_condition_write0:
1022                fprintf(fp, "write0");
1023                break;
1024
1025        case midgard_condition_false:
1026                fprintf(fp, "false");
1027                break;
1028
1029        case midgard_condition_true:
1030                fprintf(fp, "true");
1031                break;
1032
1033        case midgard_condition_always:
1034                fprintf(fp, "always");
1035                break;
1036
1037        default:
1038                fprintf(fp, "unk%X", cond);
1039                break;
1040        }
1041}
1042
1043static bool
1044print_compact_branch_writeout_field(disassemble_context *ctx, FILE *fp, uint16_t word)
1045{
1046        midgard_jmp_writeout_op op = word & 0x7;
1047        ctx->midg_stats.instruction_count++;
1048
1049        switch (op) {
1050        case midgard_jmp_writeout_op_branch_uncond: {
1051                midgard_branch_uncond br_uncond;
1052                memcpy((char *) &br_uncond, (char *) &word, sizeof(br_uncond));
1053                fprintf(fp, "br.uncond ");
1054
1055                if (br_uncond.unknown != 1)
1056                        fprintf(fp, "unknown:%u, ", br_uncond.unknown);
1057
1058                if (br_uncond.offset >= 0)
1059                        fprintf(fp, "+");
1060
1061                fprintf(fp, "%d -> %s", br_uncond.offset,
1062                                midgard_tag_props[br_uncond.dest_tag].name);
1063                fprintf(fp, "\n");
1064
1065                return br_uncond.offset >= 0;
1066        }
1067
1068        case midgard_jmp_writeout_op_branch_cond:
1069        case midgard_jmp_writeout_op_writeout:
1070        case midgard_jmp_writeout_op_discard:
1071        default: {
1072                midgard_branch_cond br_cond;
1073                memcpy((char *) &br_cond, (char *) &word, sizeof(br_cond));
1074
1075                fprintf(fp, "br.");
1076
1077                print_branch_op(fp, br_cond.op);
1078                print_branch_cond(fp, br_cond.cond);
1079
1080                fprintf(fp, " ");
1081
1082                if (br_cond.offset >= 0)
1083                        fprintf(fp, "+");
1084
1085                fprintf(fp, "%d -> %s", br_cond.offset,
1086                                midgard_tag_props[br_cond.dest_tag].name);
1087                fprintf(fp, "\n");
1088
1089                return br_cond.offset >= 0;
1090        }
1091        }
1092
1093        return false;
1094}
1095
1096static bool
1097print_extended_branch_writeout_field(disassemble_context *ctx, FILE *fp, uint8_t *words,
1098                                     unsigned next)
1099{
1100        midgard_branch_extended br;
1101        memcpy((char *) &br, (char *) words, sizeof(br));
1102
1103        fprintf(fp, "brx.");
1104
1105        print_branch_op(fp, br.op);
1106
1107        /* Condition codes are a LUT in the general case, but simply repeated 8 times for single-channel conditions.. Check this. */
1108
1109        bool single_channel = true;
1110
1111        for (unsigned i = 0; i < 16; i += 2) {
1112                single_channel &= (((br.cond >> i) & 0x3) == (br.cond & 0x3));
1113        }
1114
1115        if (single_channel)
1116                print_branch_cond(fp, br.cond & 0x3);
1117        else
1118                fprintf(fp, "lut%X", br.cond);
1119
1120        if (br.unknown)
1121                fprintf(fp, ".unknown%u", br.unknown);
1122
1123        fprintf(fp, " ");
1124
1125        if (br.offset >= 0)
1126                fprintf(fp, "+");
1127
1128        fprintf(fp, "%d -> %s\n", br.offset,
1129                        midgard_tag_props[br.dest_tag].name);
1130
1131        unsigned I = next + br.offset * 4;
1132
1133        if (ctx->midg_tags[I] && ctx->midg_tags[I] != br.dest_tag) {
1134                fprintf(fp, "\t/* XXX TAG ERROR: jumping to %s but tagged %s \n",
1135                        midgard_tag_props[br.dest_tag].name,
1136                        midgard_tag_props[ctx->midg_tags[I]].name);
1137        }
1138
1139        ctx->midg_tags[I] = br.dest_tag;
1140
1141        ctx->midg_stats.instruction_count++;
1142        return br.offset >= 0;
1143}
1144
1145static unsigned
1146num_alu_fields_enabled(uint32_t control_word)
1147{
1148        unsigned ret = 0;
1149
1150        if ((control_word >> 17) & 1)
1151                ret++;
1152
1153        if ((control_word >> 19) & 1)
1154                ret++;
1155
1156        if ((control_word >> 21) & 1)
1157                ret++;
1158
1159        if ((control_word >> 23) & 1)
1160                ret++;
1161
1162        if ((control_word >> 25) & 1)
1163                ret++;
1164
1165        return ret;
1166}
1167
1168static bool
1169print_alu_word(disassemble_context *ctx, FILE *fp, uint32_t *words,
1170               unsigned num_quad_words, unsigned tabs, unsigned next,
1171               bool verbose)
1172{
1173        uint32_t control_word = words[0];
1174        uint16_t *beginning_ptr = (uint16_t *)(words + 1);
1175        unsigned num_fields = num_alu_fields_enabled(control_word);
1176        uint16_t *word_ptr = beginning_ptr + num_fields;
1177        unsigned num_words = 2 + num_fields;
1178        const midgard_constants *consts = NULL;
1179        bool branch_forward = false;
1180
1181        if ((control_word >> 17) & 1)
1182                num_words += 3;
1183
1184        if ((control_word >> 19) & 1)
1185                num_words += 2;
1186
1187        if ((control_word >> 21) & 1)
1188                num_words += 3;
1189
1190        if ((control_word >> 23) & 1)
1191                num_words += 2;
1192
1193        if ((control_word >> 25) & 1)
1194                num_words += 3;
1195
1196        if ((control_word >> 26) & 1)
1197                num_words += 1;
1198
1199        if ((control_word >> 27) & 1)
1200                num_words += 3;
1201
1202        if (num_quad_words > (num_words + 7) / 8) {
1203                assert(num_quad_words == (num_words + 15) / 8);
1204                //Assume that the extra quadword is constants
1205                consts = (midgard_constants *)(words + (4 * num_quad_words - 4));
1206        }
1207
1208        if ((control_word >> 16) & 1)
1209                fprintf(fp, "unknown bit 16 enabled\n");
1210
1211        if ((control_word >> 17) & 1) {
1212                print_vector_field(ctx, fp, "vmul", word_ptr, *beginning_ptr, consts, tabs, verbose);
1213                beginning_ptr += 1;
1214                word_ptr += 3;
1215        }
1216
1217        if ((control_word >> 18) & 1)
1218                fprintf(fp, "unknown bit 18 enabled\n");
1219
1220        if ((control_word >> 19) & 1) {
1221                print_scalar_field(ctx, fp, "sadd", word_ptr, *beginning_ptr, consts, tabs, verbose);
1222                beginning_ptr += 1;
1223                word_ptr += 2;
1224        }
1225
1226        if ((control_word >> 20) & 1)
1227                fprintf(fp, "unknown bit 20 enabled\n");
1228
1229        if ((control_word >> 21) & 1) {
1230                print_vector_field(ctx, fp, "vadd", word_ptr, *beginning_ptr, consts, tabs, verbose);
1231                beginning_ptr += 1;
1232                word_ptr += 3;
1233        }
1234
1235        if ((control_word >> 22) & 1)
1236                fprintf(fp, "unknown bit 22 enabled\n");
1237
1238        if ((control_word >> 23) & 1) {
1239                print_scalar_field(ctx, fp, "smul", word_ptr, *beginning_ptr, consts, tabs, verbose);
1240                beginning_ptr += 1;
1241                word_ptr += 2;
1242        }
1243
1244        if ((control_word >> 24) & 1)
1245                fprintf(fp, "unknown bit 24 enabled\n");
1246
1247        if ((control_word >> 25) & 1) {
1248                print_vector_field(ctx, fp, "lut", word_ptr, *beginning_ptr, consts, tabs, verbose);
1249                word_ptr += 3;
1250        }
1251
1252        if ((control_word >> 26) & 1) {
1253                branch_forward |= print_compact_branch_writeout_field(ctx, fp, *word_ptr);
1254                word_ptr += 1;
1255        }
1256
1257        if ((control_word >> 27) & 1) {
1258                branch_forward |= print_extended_branch_writeout_field(ctx, fp, (uint8_t *) word_ptr, next);
1259                word_ptr += 3;
1260        }
1261
1262        if (consts)
1263                fprintf(fp, "uconstants 0x%X, 0x%X, 0x%X, 0x%X\n",
1264                        consts->u32[0], consts->u32[1],
1265                        consts->u32[2], consts->u32[3]);
1266
1267        return branch_forward;
1268}
1269
1270/* TODO: how can we use this now that we know that these params can't be known
1271 * before run time in every single case? Maybe just use it in the cases we can? */
1272UNUSED static void
1273print_varying_parameters(FILE *fp, midgard_load_store_word *word)
1274{
1275        midgard_varying_params p = midgard_unpack_varying_params(*word);
1276
1277        /* If a varying, there are qualifiers */
1278        if (p.flat_shading)
1279                fprintf(fp, ".flat");
1280
1281        if (p.perspective_correction)
1282                fprintf(fp, ".correction");
1283
1284        if (p.centroid_mapping)
1285                fprintf(fp, ".centroid");
1286
1287        if (p.interpolate_sample)
1288                fprintf(fp, ".sample");
1289
1290        switch (p.modifier) {
1291                case midgard_varying_mod_perspective_y:
1292                        fprintf(fp, ".perspectivey");
1293                        break;
1294                case midgard_varying_mod_perspective_z:
1295                        fprintf(fp, ".perspectivez");
1296                        break;
1297                case midgard_varying_mod_perspective_w:
1298                        fprintf(fp, ".perspectivew");
1299                        break;
1300                default:
1301                        unreachable("invalid varying modifier");
1302                        break;
1303        }
1304}
1305
1306static bool
1307is_op_varying(unsigned op)
1308{
1309        switch (op) {
1310        case midgard_op_st_vary_16:
1311        case midgard_op_st_vary_32:
1312        case midgard_op_st_vary_32i:
1313        case midgard_op_st_vary_32u:
1314        case midgard_op_ld_vary_16:
1315        case midgard_op_ld_vary_32:
1316        case midgard_op_ld_vary_32i:
1317        case midgard_op_ld_vary_32u:
1318                return true;
1319        }
1320
1321        return false;
1322}
1323
1324static bool
1325is_op_attribute(unsigned op)
1326{
1327        switch (op) {
1328        case midgard_op_ld_attr_16:
1329        case midgard_op_ld_attr_32:
1330        case midgard_op_ld_attr_32i:
1331        case midgard_op_ld_attr_32u:
1332                return true;
1333        }
1334
1335        return false;
1336}
1337
1338/* Helper to print integer well-formatted, but only when non-zero. */
1339static void
1340midgard_print_sint(FILE *fp, int n)
1341{
1342        if (n > 0)
1343                fprintf(fp, " + 0x%X", n);
1344        else if (n < 0)
1345                fprintf(fp, " - 0x%X", -n);
1346}
1347
1348static void
1349update_stats(signed *stat, unsigned address)
1350{
1351        if (*stat >= 0)
1352                *stat = MAX2(*stat, address + 1);
1353}
1354
1355static void
1356print_load_store_instr(disassemble_context *ctx, FILE *fp, uint64_t data, bool verbose)
1357{
1358        midgard_load_store_word *word = (midgard_load_store_word *) &data;
1359
1360        print_ld_st_opcode(fp, word->op);
1361
1362        if (word->op == midgard_op_trap) {
1363                fprintf(fp, " 0x%X\n", word->signed_offset);
1364                return;
1365        }
1366
1367        /* Print opcode modifiers */
1368
1369        if (OP_USES_ATTRIB(word->op)) {
1370                /* Print non-default attribute tables */
1371                bool default_secondary =
1372                        (word->op == midgard_op_st_vary_32) ||
1373                        (word->op == midgard_op_st_vary_16) ||
1374                        (word->op == midgard_op_st_vary_32u) ||
1375                        (word->op == midgard_op_st_vary_32i) ||
1376                        (word->op == midgard_op_ld_vary_32) ||
1377                        (word->op == midgard_op_ld_vary_16) ||
1378                        (word->op == midgard_op_ld_vary_32u) ||
1379                        (word->op == midgard_op_ld_vary_32i);
1380
1381                bool default_primary =
1382                        (word->op == midgard_op_ld_attr_32) ||
1383                        (word->op == midgard_op_ld_attr_16) ||
1384                        (word->op == midgard_op_ld_attr_32u) ||
1385                        (word->op == midgard_op_ld_attr_32i);
1386
1387                bool has_default = (default_secondary || default_primary);
1388                bool is_secondary = (word->index_format >> 1);
1389
1390                if (has_default && (is_secondary != default_secondary))
1391                        fprintf(fp, ".%s", is_secondary ? "secondary" : "primary");
1392        } else if (word->op == midgard_op_ld_cubemap_coords || OP_IS_PROJECTION(word->op))
1393                fprintf(fp, ".%s", word->bitsize_toggle ? "f32" : "f16");
1394
1395        fprintf(fp, " ");
1396
1397        /* src/dest register */
1398
1399        if (!OP_IS_STORE(word->op)) {
1400                print_ldst_write_reg(fp, word->reg);
1401
1402                /* Some opcodes don't have a swizzable src register, and
1403                 * instead the swizzle is applied before the result is written
1404                 * to the dest reg. For these ops, we combine the writemask
1405                 * with the swizzle to display them in the disasm compactly. */
1406                unsigned swizzle = word->swizzle;
1407                if ((OP_IS_REG2REG_LDST(word->op) &&
1408                        word->op != midgard_op_lea &&
1409                        word->op != midgard_op_lea_image) || OP_IS_ATOMIC(word->op))
1410                        swizzle = 0xE4;
1411                print_ldst_mask(fp, word->mask, swizzle);
1412        } else {
1413                print_ldst_read_reg(fp, word->reg);
1414                print_vec_swizzle(fp, word->swizzle, midgard_src_passthrough,
1415                                  midgard_reg_mode_32, 0xFF);
1416        }
1417
1418        /* ld_ubo args */
1419        if (OP_IS_UBO_READ(word->op)) {
1420                if (word->signed_offset & 1) { /* buffer index imm */
1421                        unsigned imm = midgard_unpack_ubo_index_imm(*word);
1422                        fprintf(fp, ", %u", imm);
1423                } else { /* buffer index from reg */
1424                        fprintf(fp, ", ");
1425                        print_ldst_read_reg(fp, word->arg_reg);
1426                        fprintf(fp, ".%c", components[word->arg_comp]);
1427                }
1428
1429                fprintf(fp, ", ");
1430                print_ldst_read_reg(fp, word->index_reg);
1431                fprintf(fp, ".%c", components[word->index_comp]);
1432                if (word->index_shift)
1433                        fprintf(fp, " lsl %u",  word->index_shift);
1434                midgard_print_sint(fp, UNPACK_LDST_UBO_OFS(word->signed_offset));
1435        }
1436
1437        /* mem addr expression */
1438        if (OP_HAS_ADDRESS(word->op)) {
1439                fprintf(fp, ", ");
1440                bool first = true;
1441
1442                /* Skip printing zero */
1443                if (word->arg_reg != 7 || verbose) {
1444                        print_ldst_read_reg(fp, word->arg_reg);
1445                        fprintf(fp, ".u%d.%c",
1446                                word->bitsize_toggle ? 64 : 32, components[word->arg_comp]);
1447                        first = false;
1448                }
1449
1450                if ((word->op < midgard_op_atomic_cmpxchg ||
1451                     word->op > midgard_op_atomic_cmpxchg64_be) &&
1452                     word->index_reg != 0x7) {
1453                        if (!first)
1454                                fprintf(fp, " + ");
1455
1456                        print_ldst_read_reg(fp, word->index_reg);
1457                        fprintf(fp, "%s.%c",
1458                                index_format_names[word->index_format],
1459                                components[word->index_comp]);
1460                        if (word->index_shift)
1461                                fprintf(fp, " lsl %u",  word->index_shift);
1462                }
1463
1464                midgard_print_sint(fp, word->signed_offset);
1465        }
1466
1467        /* src reg for reg2reg ldst opcodes */
1468        if (OP_IS_REG2REG_LDST(word->op)) {
1469                fprintf(fp, ", ");
1470                print_ldst_read_reg(fp, word->arg_reg);
1471                print_vec_swizzle(fp, word->swizzle, midgard_src_passthrough,
1472                                  midgard_reg_mode_32, 0xFF);
1473        }
1474
1475        /* atomic ops encode the source arg where the ldst swizzle would be. */
1476        if (OP_IS_ATOMIC(word->op)) {
1477                unsigned src = (word->swizzle >> 2) & 0x7;
1478                unsigned src_comp = word->swizzle & 0x3;
1479                fprintf(fp, ", ");
1480                print_ldst_read_reg(fp, src);
1481                fprintf(fp, ".%c", components[src_comp]);
1482        }
1483
1484        /* CMPXCHG encodes the extra comparison arg where the index reg would be. */
1485        if (word->op >= midgard_op_atomic_cmpxchg &&
1486            word->op <= midgard_op_atomic_cmpxchg64_be) {
1487                fprintf(fp, ", ");
1488                print_ldst_read_reg(fp, word->index_reg);
1489                fprintf(fp, ".%c", components[word->index_comp]);
1490        }
1491
1492        /* index reg for attr/vary/images, selector for ld/st_special */
1493        if (OP_IS_SPECIAL(word->op) || OP_USES_ATTRIB(word->op)) {
1494                fprintf(fp, ", ");
1495                print_ldst_read_reg(fp, word->index_reg);
1496                fprintf(fp, ".%c", components[word->index_comp]);
1497                if (word->index_shift)
1498                        fprintf(fp, " lsl %u",  word->index_shift);
1499                midgard_print_sint(fp, UNPACK_LDST_ATTRIB_OFS(word->signed_offset));
1500        }
1501
1502        /* vertex reg for attrib/varying ops, coord reg for image ops */
1503        if (OP_USES_ATTRIB(word->op)) {
1504                fprintf(fp, ", ");
1505                print_ldst_read_reg(fp, word->arg_reg);
1506
1507                if (OP_IS_IMAGE(word->op))
1508                        fprintf(fp, ".u%d", word->bitsize_toggle ? 64 : 32);
1509
1510                fprintf(fp, ".%c", components[word->arg_comp]);
1511
1512                if (word->bitsize_toggle && !OP_IS_IMAGE(word->op))
1513                        midgard_print_sint(fp, UNPACK_LDST_VERTEX_OFS(word->signed_offset));
1514        }
1515
1516        /* TODO: properly decode format specifier for PACK/UNPACK ops */
1517        if (OP_IS_PACK_COLOUR(word->op) || OP_IS_UNPACK_COLOUR(word->op)) {
1518                fprintf(fp, ", ");
1519                unsigned format_specifier = (word->signed_offset << 4) | word->index_shift;
1520                fprintf(fp, "0x%X", format_specifier);
1521        }
1522
1523        fprintf(fp, "\n");
1524
1525        /* Debugging stuff */
1526
1527        if (is_op_varying(word->op)) {
1528                /* Do some analysis: check if direct access */
1529
1530                if (word->index_reg == 0x7 && ctx->midg_stats.varying_count >= 0)
1531                        update_stats(&ctx->midg_stats.varying_count,
1532                                     UNPACK_LDST_ATTRIB_OFS(word->signed_offset));
1533                else
1534                        ctx->midg_stats.varying_count = -16;
1535        } else if (is_op_attribute(word->op)) {
1536                if (word->index_reg == 0x7 && ctx->midg_stats.attribute_count >= 0)
1537                        update_stats(&ctx->midg_stats.attribute_count,
1538                                     UNPACK_LDST_ATTRIB_OFS(word->signed_offset));
1539                else
1540                        ctx->midg_stats.attribute_count = -16;
1541        }
1542
1543        if (!OP_IS_STORE(word->op))
1544                update_dest(ctx, word->reg);
1545
1546        if (OP_IS_UBO_READ(word->op))
1547                update_stats(&ctx->midg_stats.uniform_buffer_count,
1548                             UNPACK_LDST_UBO_OFS(word->signed_offset));
1549
1550        ctx->midg_stats.instruction_count++;
1551}
1552
1553static void
1554print_load_store_word(disassemble_context *ctx, FILE *fp, uint32_t *word, bool verbose)
1555{
1556        midgard_load_store *load_store = (midgard_load_store *) word;
1557
1558        if (load_store->word1 != 3) {
1559                print_load_store_instr(ctx, fp, load_store->word1, verbose);
1560        }
1561
1562        if (load_store->word2 != 3) {
1563                print_load_store_instr(ctx, fp, load_store->word2, verbose);
1564        }
1565}
1566
1567static void
1568print_texture_reg_select(FILE *fp, uint8_t u, unsigned base)
1569{
1570        midgard_tex_register_select sel;
1571        memcpy(&sel, &u, sizeof(u));
1572
1573        print_tex_reg(fp, base + sel.select, false);
1574
1575        unsigned component = sel.component;
1576
1577        /* Use the upper half in half-reg mode */
1578        if (sel.upper) {
1579                assert(!sel.full);
1580                component += 4;
1581        }
1582
1583        fprintf(fp, ".%c.%d", components[component], sel.full ? 32 : 16);
1584
1585        assert(sel.zero == 0);
1586}
1587
1588static void
1589print_texture_format(FILE *fp, int format)
1590{
1591        /* Act like a modifier */
1592        fprintf(fp, ".");
1593
1594        switch (format) {
1595                DEFINE_CASE(1, "1d");
1596                DEFINE_CASE(2, "2d");
1597                DEFINE_CASE(3, "3d");
1598                DEFINE_CASE(0, "cube");
1599
1600        default:
1601                unreachable("Bad format");
1602        }
1603}
1604
1605static bool
1606midgard_op_has_helpers(unsigned op)
1607{
1608        switch (op) {
1609        case midgard_tex_op_normal:
1610        case midgard_tex_op_derivative:
1611                return true;
1612        default:
1613                return false;
1614        }
1615}
1616
1617static void
1618print_texture_op(FILE *fp, unsigned op)
1619{
1620        if (tex_opcode_props[op].name)
1621                fprintf(fp, "%s", tex_opcode_props[op].name);
1622        else
1623                fprintf(fp, "tex_op_%02X", op);
1624}
1625
1626static bool
1627texture_op_takes_bias(unsigned op)
1628{
1629        return op == midgard_tex_op_normal;
1630}
1631
1632static char
1633sampler_type_name(enum mali_sampler_type t)
1634{
1635        switch (t) {
1636        case MALI_SAMPLER_FLOAT:
1637                return 'f';
1638        case MALI_SAMPLER_UNSIGNED:
1639                return 'u';
1640        case MALI_SAMPLER_SIGNED:
1641                return 'i';
1642        default:
1643                return '?';
1644        }
1645
1646}
1647
1648static void
1649print_texture_barrier(FILE *fp, uint32_t *word)
1650{
1651        midgard_texture_barrier_word *barrier = (midgard_texture_barrier_word *) word;
1652
1653        if (barrier->type != TAG_TEXTURE_4_BARRIER)
1654                fprintf(fp, "/* barrier tag %X != tex/bar */ ", barrier->type);
1655
1656        if (!barrier->cont)
1657                fprintf(fp, "/* cont missing? */");
1658
1659        if (!barrier->last)
1660                fprintf(fp, "/* last missing? */");
1661
1662        if (barrier->zero1)
1663                fprintf(fp, "/* zero1 = 0x%X */ ", barrier->zero1);
1664
1665        if (barrier->zero2)
1666                fprintf(fp, "/* zero2 = 0x%X */ ", barrier->zero2);
1667
1668        if (barrier->zero3)
1669                fprintf(fp, "/* zero3 = 0x%X */ ", barrier->zero3);
1670
1671        if (barrier->zero4)
1672                fprintf(fp, "/* zero4 = 0x%X */ ", barrier->zero4);
1673
1674        if (barrier->zero5)
1675                fprintf(fp, "/* zero4 = 0x%" PRIx64 " */ ", barrier->zero5);
1676
1677        if (barrier->out_of_order)
1678                fprintf(fp, ".ooo%u", barrier->out_of_order);
1679
1680        fprintf(fp, "\n");
1681}
1682
1683#undef DEFINE_CASE
1684
1685static const char *
1686texture_mode(enum mali_texture_mode mode)
1687{
1688        switch (mode) {
1689        case TEXTURE_NORMAL: return "";
1690        case TEXTURE_SHADOW: return ".shadow";
1691        case TEXTURE_GATHER_SHADOW: return ".gather.shadow";
1692        case TEXTURE_GATHER_X: return ".gatherX";
1693        case TEXTURE_GATHER_Y: return ".gatherY";
1694        case TEXTURE_GATHER_Z: return ".gatherZ";
1695        case TEXTURE_GATHER_W: return ".gatherW";
1696        default: return "unk";
1697        }
1698}
1699
1700static const char *
1701derivative_mode(enum mali_derivative_mode mode)
1702{
1703        switch (mode) {
1704        case TEXTURE_DFDX: return ".x";
1705        case TEXTURE_DFDY: return ".y";
1706        default: return "unk";
1707        }
1708}
1709
1710static void
1711print_texture_word(disassemble_context *ctx, FILE *fp, uint32_t *word,
1712                   unsigned tabs, unsigned in_reg_base, unsigned out_reg_base)
1713{
1714        midgard_texture_word *texture = (midgard_texture_word *) word;
1715        ctx->midg_stats.helper_invocations |= midgard_op_has_helpers(texture->op);
1716        validate_sampler_type(texture->op, texture->sampler_type);
1717
1718        /* Broad category of texture operation in question */
1719        print_texture_op(fp, texture->op);
1720
1721        /* Barriers use a dramatically different code path */
1722        if (texture->op == midgard_tex_op_barrier) {
1723                print_texture_barrier(fp, word);
1724                return;
1725        } else if (texture->type == TAG_TEXTURE_4_BARRIER)
1726                fprintf (fp, "/* nonbarrier had tex/bar tag */ ");
1727        else if (texture->type == TAG_TEXTURE_4_VTX)
1728                fprintf (fp, ".vtx");
1729
1730        if (texture->op == midgard_tex_op_derivative)
1731                fprintf(fp, "%s", derivative_mode(texture->mode));
1732        else
1733                fprintf(fp, "%s", texture_mode(texture->mode));
1734
1735        /* Specific format in question */
1736        print_texture_format(fp, texture->format);
1737
1738        /* Instruction "modifiers" parallel the ALU instructions. */
1739
1740        if (texture->cont)
1741                fprintf(fp, ".cont");
1742
1743        if (texture->last)
1744                fprintf(fp, ".last");
1745
1746        if (texture->out_of_order)
1747                fprintf(fp, ".ooo%u", texture->out_of_order);
1748
1749        fprintf(fp, " ");
1750        print_tex_reg(fp, out_reg_base + texture->out_reg_select, true);
1751        print_tex_mask(fp, texture->mask, texture->out_upper);
1752        fprintf(fp, ".%c%d", texture->sampler_type == MALI_SAMPLER_FLOAT ? 'f' : 'i',
1753                             texture->out_full ? 32 : 16);
1754        assert(!(texture->out_full && texture->out_upper));
1755
1756        /* Output modifiers are only valid for float texture operations */
1757        if (texture->sampler_type == MALI_SAMPLER_FLOAT)
1758                print_outmod(fp, texture->outmod, false);
1759
1760        fprintf(fp, ", ");
1761
1762        /* Depending on whether we read from textures directly or indirectly,
1763         * we may be able to update our analysis */
1764
1765        if (texture->texture_register) {
1766                fprintf(fp, "texture[");
1767                print_texture_reg_select(fp, texture->texture_handle, in_reg_base);
1768                fprintf(fp, "], ");
1769
1770                /* Indirect, tut tut */
1771                ctx->midg_stats.texture_count = -16;
1772        } else {
1773                fprintf(fp, "texture%u, ", texture->texture_handle);
1774                update_stats(&ctx->midg_stats.texture_count, texture->texture_handle);
1775        }
1776
1777        /* Print the type, GL style */
1778        fprintf(fp, "%csampler", sampler_type_name(texture->sampler_type));
1779
1780        if (texture->sampler_register) {
1781                fprintf(fp, "[");
1782                print_texture_reg_select(fp, texture->sampler_handle, in_reg_base);
1783                fprintf(fp, "]");
1784
1785                ctx->midg_stats.sampler_count = -16;
1786        } else {
1787                fprintf(fp, "%u", texture->sampler_handle);
1788                update_stats(&ctx->midg_stats.sampler_count, texture->sampler_handle);
1789        }
1790
1791        print_vec_swizzle(fp, texture->swizzle, midgard_src_passthrough, midgard_reg_mode_32, 0xFF);
1792
1793        fprintf(fp, ", ");
1794
1795        midgard_src_expand_mode exp =
1796                texture->in_reg_upper ? midgard_src_expand_high : midgard_src_passthrough;
1797        print_tex_reg(fp, in_reg_base + texture->in_reg_select, false);
1798        print_vec_swizzle(fp, texture->in_reg_swizzle, exp, midgard_reg_mode_32, 0xFF);
1799        fprintf(fp, ".%d", texture->in_reg_full ? 32 : 16);
1800        assert(!(texture->in_reg_full && texture->in_reg_upper));
1801
1802        /* There is *always* an offset attached. Of
1803         * course, that offset is just immediate #0 for a
1804         * GLES call that doesn't take an offset. If there
1805         * is a non-negative non-zero offset, this is
1806         * specified in immediate offset mode, with the
1807         * values in the offset_* fields as immediates. If
1808         * this is a negative offset, we instead switch to
1809         * a register offset mode, where the offset_*
1810         * fields become register triplets */
1811
1812        if (texture->offset_register) {
1813                fprintf(fp, " + ");
1814
1815                bool full = texture->offset & 1;
1816                bool select = texture->offset & 2;
1817                bool upper = texture->offset & 4;
1818                unsigned swizzle = texture->offset >> 3;
1819                midgard_src_expand_mode exp =
1820                        upper ? midgard_src_expand_high : midgard_src_passthrough;
1821
1822                print_tex_reg(fp, in_reg_base + select, false);
1823                print_vec_swizzle(fp, swizzle, exp, midgard_reg_mode_32, 0xFF);
1824                fprintf(fp, ".%d", full ? 32 : 16);
1825                assert(!(texture->out_full && texture->out_upper));
1826
1827                fprintf(fp, ", ");
1828        } else if (texture->offset) {
1829                /* Only select ops allow negative immediate offsets, verify */
1830
1831                signed offset_x = (texture->offset & 0xF);
1832                signed offset_y = ((texture->offset >> 4) & 0xF);
1833                signed offset_z = ((texture->offset >> 8) & 0xF);
1834
1835                bool neg_x = offset_x < 0;
1836                bool neg_y = offset_y < 0;
1837                bool neg_z = offset_z < 0;
1838                bool any_neg = neg_x || neg_y || neg_z;
1839
1840                if (any_neg && texture->op != midgard_tex_op_fetch)
1841                        fprintf(fp, "/* invalid negative */ ");
1842
1843                /* Regardless, just print the immediate offset */
1844
1845                fprintf(fp, " + <%d, %d, %d>, ", offset_x, offset_y, offset_z);
1846        } else {
1847                fprintf(fp, ", ");
1848        }
1849
1850        char lod_operand = texture_op_takes_bias(texture->op) ? '+' : '=';
1851
1852        if (texture->lod_register) {
1853                fprintf(fp, "lod %c ", lod_operand);
1854                print_texture_reg_select(fp, texture->bias, in_reg_base);
1855                fprintf(fp, ", ");
1856
1857                if (texture->bias_int)
1858                        fprintf(fp, " /* bias_int = 0x%X */", texture->bias_int);
1859        } else if (texture->op == midgard_tex_op_fetch) {
1860                /* For texel fetch, the int LOD is in the fractional place and
1861                 * there is no fraction. We *always* have an explicit LOD, even
1862                 * if it's zero. */
1863
1864                if (texture->bias_int)
1865                        fprintf(fp, " /* bias_int = 0x%X */ ", texture->bias_int);
1866
1867                fprintf(fp, "lod = %u, ", texture->bias);
1868        } else if (texture->bias || texture->bias_int) {
1869                signed bias_int = texture->bias_int;
1870                float bias_frac = texture->bias / 256.0f;
1871                float bias = bias_int + bias_frac;
1872
1873                bool is_bias = texture_op_takes_bias(texture->op);
1874                char sign = (bias >= 0.0) ? '+' : '-';
1875                char operand = is_bias ? sign : '=';
1876
1877                fprintf(fp, "lod %c %f, ", operand, fabsf(bias));
1878        }
1879
1880        fprintf(fp, "\n");
1881
1882        /* While not zero in general, for these simple instructions the
1883         * following unknowns are zero, so we don't include them */
1884
1885        if (texture->unknown4 ||
1886            texture->unknown8) {
1887                fprintf(fp, "// unknown4 = 0x%x\n", texture->unknown4);
1888                fprintf(fp, "// unknown8 = 0x%x\n", texture->unknown8);
1889        }
1890
1891        ctx->midg_stats.instruction_count++;
1892}
1893
1894struct midgard_disasm_stats
1895disassemble_midgard(FILE *fp, uint8_t *code, size_t size, unsigned gpu_id, bool verbose)
1896{
1897        uint32_t *words = (uint32_t *) code;
1898        unsigned num_words = size / 4;
1899        int tabs = 0;
1900
1901        bool branch_forward = false;
1902
1903        int last_next_tag = -1;
1904
1905        unsigned i = 0;
1906
1907        disassemble_context ctx = {
1908                .midg_tags = calloc(sizeof(ctx.midg_tags[0]), num_words),
1909                .midg_stats = {0},
1910                .midg_ever_written = 0,
1911        };
1912
1913        while (i < num_words) {
1914                unsigned tag = words[i] & 0xF;
1915                unsigned next_tag = (words[i] >> 4) & 0xF;
1916                unsigned num_quad_words = midgard_tag_props[tag].size;
1917
1918                if (ctx.midg_tags[i] && ctx.midg_tags[i] != tag) {
1919                        fprintf(fp, "\t/* XXX: TAG ERROR branch, got %s expected %s */\n",
1920                                        midgard_tag_props[tag].name,
1921                                        midgard_tag_props[ctx.midg_tags[i]].name);
1922                }
1923
1924                ctx.midg_tags[i] = tag;
1925
1926                /* Check the tag. The idea is to ensure that next_tag is
1927                 * *always* recoverable from the disassembly, such that we may
1928                 * safely omit printing next_tag. To show this, we first
1929                 * consider that next tags are semantically off-byone -- we end
1930                 * up parsing tag n during step n+1. So, we ensure after we're
1931                 * done disassembling the next tag of the final bundle is BREAK
1932                 * and warn otherwise. We also ensure that the next tag is
1933                 * never INVALID. Beyond that, since the last tag is checked
1934                 * outside the loop, we can check one tag prior. If equal to
1935                 * the current tag (which is unique), we're done. Otherwise, we
1936                 * print if that tag was > TAG_BREAK, which implies the tag was
1937                 * not TAG_BREAK or TAG_INVALID. But we already checked for
1938                 * TAG_INVALID, so it's just if the last tag was TAG_BREAK that
1939                 * we're silent. So we throw in a print for break-next on at
1940                 * the end of the bundle (if it's not the final bundle, which
1941                 * we already check for above), disambiguating this case as
1942                 * well.  Hence in all cases we are unambiguous, QED. */
1943
1944                if (next_tag == TAG_INVALID)
1945                        fprintf(fp, "\t/* XXX: invalid next tag */\n");
1946
1947                if (last_next_tag > TAG_BREAK && last_next_tag != tag) {
1948                        fprintf(fp, "\t/* XXX: TAG ERROR sequence, got %s expexted %s */\n",
1949                                        midgard_tag_props[tag].name,
1950                                        midgard_tag_props[last_next_tag].name);
1951                }
1952
1953                last_next_tag = next_tag;
1954
1955                /* Tags are unique in the following way:
1956                 *
1957                 * INVALID, BREAK, UNKNOWN_*: verbosely printed
1958                 * TEXTURE_4_BARRIER: verified by barrier/!barrier op
1959                 * TEXTURE_4_VTX: .vtx tag printed
1960                 * TEXTURE_4: tetxure lack of barriers or .vtx
1961                 * TAG_LOAD_STORE_4: only load/store
1962                 * TAG_ALU_4/8/12/16: by number of instructions/constants
1963                 * TAG_ALU_4_8/12/16_WRITEOUT: ^^ with .writeout tag
1964                 */
1965
1966                switch (tag) {
1967                case TAG_TEXTURE_4_VTX ... TAG_TEXTURE_4_BARRIER: {
1968                        bool interpipe_aliasing =
1969                                midgard_get_quirks(gpu_id) & MIDGARD_INTERPIPE_REG_ALIASING;
1970
1971                        print_texture_word(&ctx, fp, &words[i], tabs,
1972                                        interpipe_aliasing ? 0 : REG_TEX_BASE,
1973                                        interpipe_aliasing ? REGISTER_LDST_BASE : REG_TEX_BASE);
1974                        break;
1975                }
1976
1977                case TAG_LOAD_STORE_4:
1978                        print_load_store_word(&ctx, fp, &words[i], verbose);
1979                        break;
1980
1981                case TAG_ALU_4 ... TAG_ALU_16_WRITEOUT:
1982                        branch_forward = print_alu_word(&ctx, fp, &words[i], num_quad_words, tabs, i + 4*num_quad_words, verbose);
1983
1984                        /* TODO: infer/verify me */
1985                        if (tag >= TAG_ALU_4_WRITEOUT)
1986                                fprintf(fp, "writeout\n");
1987
1988                        break;
1989
1990                default:
1991                        fprintf(fp, "Unknown word type %u:\n", words[i] & 0xF);
1992                        num_quad_words = 1;
1993                        print_quad_word(fp, &words[i], tabs);
1994                        fprintf(fp, "\n");
1995                        break;
1996                }
1997
1998                /* We are parsing per bundle anyway. Add before we start
1999                 * breaking out so we don't miss the final bundle. */
2000
2001                ctx.midg_stats.bundle_count++;
2002                ctx.midg_stats.quadword_count += num_quad_words;
2003
2004                /* Include a synthetic "break" instruction at the end of the
2005                 * bundle to signify that if, absent a branch, the shader
2006                 * execution will stop here. Stop disassembly at such a break
2007                 * based on a heuristic */
2008
2009                if (next_tag == TAG_BREAK) {
2010                        if (branch_forward) {
2011                                fprintf(fp, "break\n");
2012                        } else {
2013                                fprintf(fp, "\n");
2014                                break;
2015                        }
2016                }
2017
2018                fprintf(fp, "\n");
2019
2020                i += 4 * num_quad_words;
2021        }
2022
2023        if (last_next_tag != TAG_BREAK) {
2024                fprintf(fp, "/* XXX: shader ended with tag %s */\n",
2025                                midgard_tag_props[last_next_tag].name);
2026        }
2027
2028        free(ctx.midg_tags);
2029
2030        /* We computed work_count as max_work_registers, so add one to get the
2031         * count. If no work registers are written, you still have one work
2032         * reported, which is exactly what the hardware expects */
2033
2034        ctx.midg_stats.work_count++;
2035
2036        return ctx.midg_stats;
2037}
2038