pan/midgard: Use better heuristic for shader termination
[mesa.git] / src / panfrost / midgard / disassemble.c
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 *
8 * Permission is hereby granted, free of charge, to any person obtaining a copy
9 * of this software and associated documentation files (the "Software"), to deal
10 * in the Software without restriction, including without limitation the rights
11 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
12 * copies of the Software, and to permit persons to whom the Software is
13 * furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice shall be included in
16 * all copies or substantial portions of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
21 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
22 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
23 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
24 * THE SOFTWARE.
25 */
26
27 #include <stdio.h>
28 #include <stdint.h>
29 #include <stdlib.h>
30 #include <assert.h>
31 #include <inttypes.h>
32 #include <ctype.h>
33 #include <string.h>
34 #include "midgard.h"
35 #include "midgard-parse.h"
36 #include "midgard_ops.h"
37 #include "midgard_quirks.h"
38 #include "disassemble.h"
39 #include "helpers.h"
40 #include "util/half_float.h"
41 #include "util/u_math.h"
42
43 #define DEFINE_CASE(define, str) case define: { printf(str); break; }
44
45 static unsigned *midg_tags;
46 static bool is_instruction_int = false;
47
48 /* Stats */
49
50 static struct midgard_disasm_stats midg_stats;
51
52 /* Prints a short form of the tag for branching, the minimum needed to be
53 * legible and unambiguous */
54
55 static void
56 print_tag_short(unsigned tag)
57 {
58 switch (midgard_word_types[tag]) {
59 case midgard_word_type_texture:
60 printf("tex/%X", tag);
61 break;
62
63 case midgard_word_type_load_store:
64 printf("ldst");
65 break;
66
67 case midgard_word_type_alu:
68 printf("alu%u/%X", midgard_word_size[tag], tag);
69 break;
70
71 default:
72 printf("%s%X", (tag > 0) ? "" : "unk", tag);
73 break;
74 }
75 }
76
77 static void
78 print_alu_opcode(midgard_alu_op op)
79 {
80 bool int_op = false;
81
82 if (alu_opcode_props[op].name) {
83 printf("%s", alu_opcode_props[op].name);
84
85 int_op = midgard_is_integer_op(op);
86 } else
87 printf("alu_op_%02X", op);
88
89 /* For constant analysis */
90 is_instruction_int = int_op;
91 }
92
93 static void
94 print_ld_st_opcode(midgard_load_store_op op)
95 {
96 if (load_store_opcode_props[op].name)
97 printf("%s", load_store_opcode_props[op].name);
98 else
99 printf("ldst_op_%02X", op);
100 }
101
102 static bool is_embedded_constant_half = false;
103 static bool is_embedded_constant_int = false;
104
105 static char
106 prefix_for_bits(unsigned bits)
107 {
108 switch (bits) {
109 case 8:
110 return 'q';
111 case 16:
112 return 'h';
113 case 64:
114 return 'd';
115 default:
116 return 0;
117 }
118 }
119
120 /* For static analysis to ensure all registers are written at least once before
121 * use along the source code path (TODO: does this break done for complex CF?)
122 */
123
124 uint16_t midg_ever_written = 0;
125
126 static void
127 print_reg(unsigned reg, unsigned bits)
128 {
129 /* Perform basic static analysis for expanding constants correctly */
130
131 if (reg == 26) {
132 is_embedded_constant_int = is_instruction_int;
133 is_embedded_constant_half = (bits < 32);
134 }
135
136 unsigned uniform_reg = 23 - reg;
137 bool is_uniform = false;
138
139 /* For r8-r15, it could be a work or uniform. We distinguish based on
140 * the fact work registers are ALWAYS written before use, but uniform
141 * registers are NEVER written before use. */
142
143 if ((reg >= 8 && reg < 16) && !(midg_ever_written & (1 << reg)))
144 is_uniform = true;
145
146 /* r16-r23 are always uniform */
147
148 if (reg >= 16 && reg <= 23)
149 is_uniform = true;
150
151 /* Update the uniform count appropriately */
152
153 if (is_uniform)
154 midg_stats.uniform_count =
155 MAX2(uniform_reg + 1, midg_stats.uniform_count);
156
157 char prefix = prefix_for_bits(bits);
158
159 if (prefix)
160 putchar(prefix);
161
162 printf("r%u", reg);
163 }
164
165 static char *outmod_names_float[4] = {
166 "",
167 ".pos",
168 ".unk2",
169 ".sat"
170 };
171
172 static char *outmod_names_int[4] = {
173 ".isat",
174 ".usat",
175 "",
176 ".hi"
177 };
178
179 static char *srcmod_names_int[4] = {
180 "sext(",
181 "zext(",
182 "",
183 "("
184 };
185
186 static void
187 print_outmod(unsigned outmod, bool is_int)
188 {
189 printf("%s", is_int ? outmod_names_int[outmod] :
190 outmod_names_float[outmod]);
191 }
192
193 static void
194 print_quad_word(uint32_t *words, unsigned tabs)
195 {
196 unsigned i;
197
198 for (i = 0; i < 4; i++)
199 printf("0x%08X%s ", words[i], i == 3 ? "" : ",");
200
201 printf("\n");
202 }
203
204 static const char components[16] = "xyzwefghijklmnop";
205
206 /* Helper to print 4 chars of a swizzle */
207 static void
208 print_swizzle_helper(unsigned swizzle, bool upper)
209 {
210 for (unsigned i = 0; i < 4; ++i) {
211 unsigned c = (swizzle >> (i * 2)) & 3;
212 c += upper*4;
213 printf("%c", components[c]);
214 }
215 }
216
217 /* Helper to print 8 chars of a swizzle, duplicating over */
218 static void
219 print_swizzle_helper_8(unsigned swizzle, bool upper)
220 {
221 for (unsigned i = 0; i < 4; ++i) {
222 unsigned c = (swizzle >> (i * 2)) & 3;
223 c *= 2;
224 c += upper*8;
225 printf("%c%c", components[c], components[c+1]);
226 }
227 }
228
229 static void
230 print_swizzle_vec16(unsigned swizzle, bool rep_high, bool rep_low,
231 midgard_dest_override override)
232 {
233 printf(".");
234
235 if (override == midgard_dest_override_upper) {
236 if (rep_high)
237 printf(" /* rep_high */ ");
238 if (rep_low)
239 printf(" /* rep_low */ ");
240
241 if (!rep_high && rep_low)
242 print_swizzle_helper_8(swizzle, true);
243 else
244 print_swizzle_helper_8(swizzle, false);
245 } else {
246 print_swizzle_helper_8(swizzle, rep_high & 1);
247 print_swizzle_helper_8(swizzle, !(rep_low & 1));
248 }
249 }
250
251 static void
252 print_swizzle_vec8(unsigned swizzle, bool rep_high, bool rep_low)
253 {
254 printf(".");
255
256 print_swizzle_helper(swizzle, rep_high & 1);
257 print_swizzle_helper(swizzle, !(rep_low & 1));
258 }
259
260 static void
261 print_swizzle_vec4(unsigned swizzle, bool rep_high, bool rep_low)
262 {
263 if (rep_high)
264 printf(" /* rep_high */ ");
265 if (rep_low)
266 printf(" /* rep_low */ ");
267
268 if (swizzle == 0xE4) return; /* xyzw */
269
270 printf(".");
271 print_swizzle_helper(swizzle, 0);
272 }
273 static void
274 print_swizzle_vec2(unsigned swizzle, bool rep_high, bool rep_low)
275 {
276 if (rep_high)
277 printf(" /* rep_high */ ");
278 if (rep_low)
279 printf(" /* rep_low */ ");
280
281 if (swizzle == 0xE4) return; /* XY */
282
283 printf(".");
284
285 for (unsigned i = 0; i < 4; i += 2) {
286 unsigned a = (swizzle >> (i * 2)) & 3;
287 unsigned b = (swizzle >> ((i+1) * 2)) & 3;
288
289 /* Normally we're adjacent, but if there's an issue, don't make
290 * it ambiguous */
291
292 if (a & 0x1)
293 printf("[%c%c]", components[a], components[b]);
294 else if (a == b)
295 printf("%c", components[a >> 1]);
296 else if (b == (a + 1))
297 printf("%c", "XY"[a >> 1]);
298 else
299 printf("[%c%c]", components[a], components[b]);
300 }
301 }
302
303 static int
304 bits_for_mode(midgard_reg_mode mode)
305 {
306 switch (mode) {
307 case midgard_reg_mode_8:
308 return 8;
309 case midgard_reg_mode_16:
310 return 16;
311 case midgard_reg_mode_32:
312 return 32;
313 case midgard_reg_mode_64:
314 return 64;
315 default:
316 unreachable("Invalid reg mode");
317 return 0;
318 }
319 }
320
321 static int
322 bits_for_mode_halved(midgard_reg_mode mode, bool half)
323 {
324 unsigned bits = bits_for_mode(mode);
325
326 if (half)
327 bits >>= 1;
328
329 return bits;
330 }
331
332 static void
333 print_vector_src(unsigned src_binary,
334 midgard_reg_mode mode, unsigned reg,
335 midgard_dest_override override, bool is_int)
336 {
337 midgard_vector_alu_src *src = (midgard_vector_alu_src *)&src_binary;
338
339 /* Modifiers change meaning depending on the op's context */
340
341 midgard_int_mod int_mod = src->mod;
342
343 if (is_int) {
344 printf("%s", srcmod_names_int[int_mod]);
345 } else {
346 if (src->mod & MIDGARD_FLOAT_MOD_NEG)
347 printf("-");
348
349 if (src->mod & MIDGARD_FLOAT_MOD_ABS)
350 printf("abs(");
351 }
352
353 //register
354 unsigned bits = bits_for_mode_halved(mode, src->half);
355 print_reg(reg, bits);
356
357 //swizzle
358 if (bits == 16) {
359 /* When the mode of the instruction is itself 16-bit,
360 * rep_low/high work more or less as expected. But if the mode
361 * is 32-bit and we're stepping down, you only have vec4 and
362 * the meaning shifts to rep_low as higher-half and rep_high is
363 * never seen. TODO: are other modes similar? */
364
365 if (mode == midgard_reg_mode_32) {
366 printf(".");
367 print_swizzle_helper(src->swizzle, src->rep_low);
368 assert(!src->rep_high);
369 } else {
370 print_swizzle_vec8(src->swizzle, src->rep_high, src->rep_low);
371 }
372 } else if (bits == 8)
373 print_swizzle_vec16(src->swizzle, src->rep_high, src->rep_low, override);
374 else if (bits == 32)
375 print_swizzle_vec4(src->swizzle, src->rep_high, src->rep_low);
376 else if (bits == 64)
377 print_swizzle_vec2(src->swizzle, src->rep_high, src->rep_low);
378
379 /* Since we wrapped with a function-looking thing */
380
381 if (is_int && int_mod == midgard_int_shift)
382 printf(") << %u", bits);
383 else if ((is_int && (int_mod != midgard_int_normal))
384 || (!is_int && src->mod & MIDGARD_FLOAT_MOD_ABS))
385 printf(")");
386 }
387
388 static uint16_t
389 decode_vector_imm(unsigned src2_reg, unsigned imm)
390 {
391 uint16_t ret;
392 ret = src2_reg << 11;
393 ret |= (imm & 0x7) << 8;
394 ret |= (imm >> 3) & 0xFF;
395 return ret;
396 }
397
398 static void
399 print_immediate(uint16_t imm)
400 {
401 if (is_instruction_int)
402 printf("#%u", imm);
403 else
404 printf("#%g", _mesa_half_to_float(imm));
405 }
406
407 static void
408 update_dest(unsigned reg)
409 {
410 /* We should record writes as marking this as a work register. Store
411 * the max register in work_count; we'll add one at the end */
412
413 if (reg < 16) {
414 midg_stats.work_count = MAX2(reg, midg_stats.work_count);
415 midg_ever_written |= (1 << reg);
416 }
417 }
418
419 static void
420 print_dest(unsigned reg, midgard_reg_mode mode, midgard_dest_override override)
421 {
422 /* Depending on the mode and override, we determine the type of
423 * destination addressed. Absent an override, we address just the
424 * type of the operation itself */
425
426 unsigned bits = bits_for_mode(mode);
427
428 if (override != midgard_dest_override_none)
429 bits /= 2;
430
431 update_dest(reg);
432 print_reg(reg, bits);
433 }
434
435 static void
436 print_mask_vec16(uint8_t mask, midgard_dest_override override)
437 {
438 printf(".");
439
440 for (unsigned i = 0; i < 8; i++) {
441 if (mask & (1 << i))
442 printf("%c%c",
443 components[i*2 + 0],
444 components[i*2 + 1]);
445 }
446 }
447
448 /* For 16-bit+ masks, we read off from the 8-bit mask field. For 16-bit (vec8),
449 * it's just one bit per channel, easy peasy. For 32-bit (vec4), it's one bit
450 * per channel with one duplicate bit in the middle. For 64-bit (vec2), it's
451 * one-bit per channel with _3_ duplicate bits in the middle. Basically, just
452 * subdividing the 128-bit word in 16-bit increments. For 64-bit, we uppercase
453 * the mask to make it obvious what happened */
454
455 static void
456 print_mask(uint8_t mask, unsigned bits, midgard_dest_override override)
457 {
458 if (bits == 8) {
459 print_mask_vec16(mask, override);
460 return;
461 }
462
463 /* Skip 'complete' masks */
464
465 if (override == midgard_dest_override_none) {
466 if (bits >= 32 && mask == 0xFF) return;
467
468 if (bits == 16) {
469 if (mask == 0x0F)
470 return;
471 else if (mask == 0xF0) {
472 printf("'");
473 return;
474 }
475 }
476 }
477
478 printf(".");
479
480 unsigned skip = (bits / 16);
481 bool uppercase = bits > 32;
482 bool tripped = false;
483
484 /* To apply an upper destination override, we "shift" the alphabet.
485 * E.g. with an upper override on 32-bit, instead of xyzw, print efgh.
486 * For upper 16-bit, instead of xyzwefgh, print ijklmnop */
487
488 const char *alphabet = components;
489
490 if (override == midgard_dest_override_upper)
491 alphabet += (128 / bits);
492
493 for (unsigned i = 0; i < 8; i += skip) {
494 bool a = (mask & (1 << i)) != 0;
495
496 for (unsigned j = 1; j < skip; ++j) {
497 bool dupe = (mask & (1 << (i + j))) != 0;
498 tripped |= (dupe != a);
499 }
500
501 if (a) {
502 char c = alphabet[i / skip];
503
504 if (uppercase)
505 c = toupper(c);
506
507 printf("%c", c);
508 }
509 }
510
511 if (tripped)
512 printf(" /* %X */", mask);
513 }
514
515 /* Prints the 4-bit masks found in texture and load/store ops, as opposed to
516 * the 8-bit masks found in (vector) ALU ops. Supports texture-style 16-bit
517 * mode as well, but not load/store-style 16-bit mode. */
518
519 static void
520 print_mask_4(unsigned mask, bool upper)
521 {
522 if (mask == 0xF) {
523 if (upper)
524 printf("'");
525
526 return;
527 }
528
529 printf(".");
530
531 for (unsigned i = 0; i < 4; ++i) {
532 bool a = (mask & (1 << i)) != 0;
533 if (a)
534 printf("%c", components[i + (upper ? 4 : 0)]);
535 }
536 }
537
538 static void
539 print_vector_field(const char *name, uint16_t *words, uint16_t reg_word,
540 unsigned tabs)
541 {
542 midgard_reg_info *reg_info = (midgard_reg_info *)&reg_word;
543 midgard_vector_alu *alu_field = (midgard_vector_alu *) words;
544 midgard_reg_mode mode = alu_field->reg_mode;
545 unsigned override = alu_field->dest_override;
546
547 /* For now, prefix instruction names with their unit, until we
548 * understand how this works on a deeper level */
549 printf("%s.", name);
550
551 print_alu_opcode(alu_field->op);
552
553 /* Postfix with the size to disambiguate if necessary */
554 char postfix = prefix_for_bits(bits_for_mode(mode));
555 bool size_ambiguous = override != midgard_dest_override_none;
556
557 if (size_ambiguous)
558 printf("%c", postfix ? postfix : 'r');
559
560 /* Print the outmod, if there is one */
561 print_outmod(alu_field->outmod,
562 midgard_is_integer_out_op(alu_field->op));
563
564 printf(" ");
565
566 /* Mask denoting status of 8-lanes */
567 uint8_t mask = alu_field->mask;
568
569 /* First, print the destination */
570 print_dest(reg_info->out_reg, mode, alu_field->dest_override);
571
572 if (override != midgard_dest_override_none) {
573 bool modeable = (mode != midgard_reg_mode_8);
574 bool known = override != 0x3; /* Unused value */
575
576 if (!(modeable && known))
577 printf("/* do%u */ ", override);
578 }
579
580 print_mask(mask, bits_for_mode(mode), override);
581
582 printf(", ");
583
584 bool is_int = midgard_is_integer_op(alu_field->op);
585 print_vector_src(alu_field->src1, mode, reg_info->src1_reg, override, is_int);
586
587 printf(", ");
588
589 if (reg_info->src2_imm) {
590 uint16_t imm = decode_vector_imm(reg_info->src2_reg, alu_field->src2 >> 2);
591 print_immediate(imm);
592 } else {
593 print_vector_src(alu_field->src2, mode,
594 reg_info->src2_reg, override, is_int);
595 }
596
597 midg_stats.instruction_count++;
598 printf("\n");
599 }
600
601 static void
602 print_scalar_src(unsigned src_binary, unsigned reg)
603 {
604 midgard_scalar_alu_src *src = (midgard_scalar_alu_src *)&src_binary;
605
606 if (src->negate)
607 printf("-");
608
609 if (src->abs)
610 printf("abs(");
611
612 print_reg(reg, src->full ? 32 : 16);
613
614 unsigned c = src->component;
615
616 if (src->full) {
617 assert((c & 1) == 0);
618 c >>= 1;
619 }
620
621 printf(".%c", components[c]);
622
623 if (src->abs)
624 printf(")");
625
626 }
627
628 static uint16_t
629 decode_scalar_imm(unsigned src2_reg, unsigned imm)
630 {
631 uint16_t ret;
632 ret = src2_reg << 11;
633 ret |= (imm & 3) << 9;
634 ret |= (imm & 4) << 6;
635 ret |= (imm & 0x38) << 2;
636 ret |= imm >> 6;
637 return ret;
638 }
639
640 static void
641 print_scalar_field(const char *name, uint16_t *words, uint16_t reg_word,
642 unsigned tabs)
643 {
644 midgard_reg_info *reg_info = (midgard_reg_info *)&reg_word;
645 midgard_scalar_alu *alu_field = (midgard_scalar_alu *) words;
646
647 if (alu_field->unknown)
648 printf("scalar ALU unknown bit set\n");
649
650 printf("%s.", name);
651 print_alu_opcode(alu_field->op);
652 print_outmod(alu_field->outmod,
653 midgard_is_integer_out_op(alu_field->op));
654 printf(" ");
655
656 bool full = alu_field->output_full;
657 update_dest(reg_info->out_reg);
658 print_reg(reg_info->out_reg, full ? 32 : 16);
659 unsigned c = alu_field->output_component;
660
661 if (full) {
662 assert((c & 1) == 0);
663 c >>= 1;
664 }
665
666 printf(".%c, ", components[c]);
667
668 print_scalar_src(alu_field->src1, reg_info->src1_reg);
669
670 printf(", ");
671
672 if (reg_info->src2_imm) {
673 uint16_t imm = decode_scalar_imm(reg_info->src2_reg,
674 alu_field->src2);
675 print_immediate(imm);
676 } else
677 print_scalar_src(alu_field->src2, reg_info->src2_reg);
678
679 midg_stats.instruction_count++;
680 printf("\n");
681 }
682
683 static void
684 print_branch_op(unsigned op)
685 {
686 switch (op) {
687 case midgard_jmp_writeout_op_branch_uncond:
688 printf("uncond.");
689 break;
690
691 case midgard_jmp_writeout_op_branch_cond:
692 printf("cond.");
693 break;
694
695 case midgard_jmp_writeout_op_writeout:
696 printf("write.");
697 break;
698
699 case midgard_jmp_writeout_op_tilebuffer_pending:
700 printf("tilebuffer.");
701 break;
702
703 case midgard_jmp_writeout_op_discard:
704 printf("discard.");
705 break;
706
707 default:
708 printf("unk%u.", op);
709 break;
710 }
711 }
712
713 static void
714 print_branch_cond(int cond)
715 {
716 switch (cond) {
717 case midgard_condition_write0:
718 printf("write0");
719 break;
720
721 case midgard_condition_false:
722 printf("false");
723 break;
724
725 case midgard_condition_true:
726 printf("true");
727 break;
728
729 case midgard_condition_always:
730 printf("always");
731 break;
732
733 default:
734 printf("unk%X", cond);
735 break;
736 }
737 }
738
739 static bool
740 print_compact_branch_writeout_field(uint16_t word)
741 {
742 midgard_jmp_writeout_op op = word & 0x7;
743 midg_stats.instruction_count++;
744
745 switch (op) {
746 case midgard_jmp_writeout_op_branch_uncond: {
747 midgard_branch_uncond br_uncond;
748 memcpy((char *) &br_uncond, (char *) &word, sizeof(br_uncond));
749 printf("br.uncond ");
750
751 if (br_uncond.unknown != 1)
752 printf("unknown:%u, ", br_uncond.unknown);
753
754 if (br_uncond.offset >= 0)
755 printf("+");
756
757 printf("%d -> ", br_uncond.offset);
758 print_tag_short(br_uncond.dest_tag);
759 printf("\n");
760
761 return br_uncond.offset >= 0;
762 }
763
764 case midgard_jmp_writeout_op_branch_cond:
765 case midgard_jmp_writeout_op_writeout:
766 case midgard_jmp_writeout_op_discard:
767 default: {
768 midgard_branch_cond br_cond;
769 memcpy((char *) &br_cond, (char *) &word, sizeof(br_cond));
770
771 printf("br.");
772
773 print_branch_op(br_cond.op);
774 print_branch_cond(br_cond.cond);
775
776 printf(" ");
777
778 if (br_cond.offset >= 0)
779 printf("+");
780
781 printf("%d -> ", br_cond.offset);
782 print_tag_short(br_cond.dest_tag);
783 printf("\n");
784
785 return br_cond.offset >= 0;
786 }
787 }
788
789 return false;
790 }
791
792 static bool
793 print_extended_branch_writeout_field(uint8_t *words, unsigned next)
794 {
795 midgard_branch_extended br;
796 memcpy((char *) &br, (char *) words, sizeof(br));
797
798 printf("brx.");
799
800 print_branch_op(br.op);
801
802 /* Condition codes are a LUT in the general case, but simply repeated 8 times for single-channel conditions.. Check this. */
803
804 bool single_channel = true;
805
806 for (unsigned i = 0; i < 16; i += 2) {
807 single_channel &= (((br.cond >> i) & 0x3) == (br.cond & 0x3));
808 }
809
810 if (single_channel)
811 print_branch_cond(br.cond & 0x3);
812 else
813 printf("lut%X", br.cond);
814
815 if (br.unknown)
816 printf(".unknown%u", br.unknown);
817
818 printf(" ");
819
820 if (br.offset >= 0)
821 printf("+");
822
823 printf("%d -> ", br.offset);
824 print_tag_short(br.dest_tag);
825 printf("\n");
826
827 unsigned I = next + br.offset * 4;
828
829 if (midg_tags[I] && midg_tags[I] != br.dest_tag) {
830 printf("\t/* XXX TAG ERROR: jumping to ");
831 print_tag_short(br.dest_tag);
832 printf(" but tagged ");
833 print_tag_short(midg_tags[I]);
834 printf(" */\n");
835 }
836
837 midg_tags[I] = br.dest_tag;
838
839 midg_stats.instruction_count++;
840 return br.offset >= 0;
841 }
842
843 static unsigned
844 num_alu_fields_enabled(uint32_t control_word)
845 {
846 unsigned ret = 0;
847
848 if ((control_word >> 17) & 1)
849 ret++;
850
851 if ((control_word >> 19) & 1)
852 ret++;
853
854 if ((control_word >> 21) & 1)
855 ret++;
856
857 if ((control_word >> 23) & 1)
858 ret++;
859
860 if ((control_word >> 25) & 1)
861 ret++;
862
863 return ret;
864 }
865
866 static float
867 float_bitcast(uint32_t integer)
868 {
869 union {
870 uint32_t i;
871 float f;
872 } v;
873
874 v.i = integer;
875 return v.f;
876 }
877
878 static bool
879 print_alu_word(uint32_t *words, unsigned num_quad_words,
880 unsigned tabs, unsigned next)
881 {
882 uint32_t control_word = words[0];
883 uint16_t *beginning_ptr = (uint16_t *)(words + 1);
884 unsigned num_fields = num_alu_fields_enabled(control_word);
885 uint16_t *word_ptr = beginning_ptr + num_fields;
886 unsigned num_words = 2 + num_fields;
887 bool branch_forward = false;
888
889 if ((control_word >> 16) & 1)
890 printf("unknown bit 16 enabled\n");
891
892 if ((control_word >> 17) & 1) {
893 print_vector_field("vmul", word_ptr, *beginning_ptr, tabs);
894 beginning_ptr += 1;
895 word_ptr += 3;
896 num_words += 3;
897 }
898
899 if ((control_word >> 18) & 1)
900 printf("unknown bit 18 enabled\n");
901
902 if ((control_word >> 19) & 1) {
903 print_scalar_field("sadd", word_ptr, *beginning_ptr, tabs);
904 beginning_ptr += 1;
905 word_ptr += 2;
906 num_words += 2;
907 }
908
909 if ((control_word >> 20) & 1)
910 printf("unknown bit 20 enabled\n");
911
912 if ((control_word >> 21) & 1) {
913 print_vector_field("vadd", word_ptr, *beginning_ptr, tabs);
914 beginning_ptr += 1;
915 word_ptr += 3;
916 num_words += 3;
917 }
918
919 if ((control_word >> 22) & 1)
920 printf("unknown bit 22 enabled\n");
921
922 if ((control_word >> 23) & 1) {
923 print_scalar_field("smul", word_ptr, *beginning_ptr, tabs);
924 beginning_ptr += 1;
925 word_ptr += 2;
926 num_words += 2;
927 }
928
929 if ((control_word >> 24) & 1)
930 printf("unknown bit 24 enabled\n");
931
932 if ((control_word >> 25) & 1) {
933 print_vector_field("lut", word_ptr, *beginning_ptr, tabs);
934 word_ptr += 3;
935 num_words += 3;
936 }
937
938 if ((control_word >> 26) & 1) {
939 branch_forward |= print_compact_branch_writeout_field(*word_ptr);
940 word_ptr += 1;
941 num_words += 1;
942 }
943
944 if ((control_word >> 27) & 1) {
945 branch_forward |= print_extended_branch_writeout_field((uint8_t *) word_ptr, next);
946 word_ptr += 3;
947 num_words += 3;
948 }
949
950 if (num_quad_words > (num_words + 7) / 8) {
951 assert(num_quad_words == (num_words + 15) / 8);
952 //Assume that the extra quadword is constants
953 void *consts = words + (4 * num_quad_words - 4);
954
955 if (is_embedded_constant_int) {
956 if (is_embedded_constant_half) {
957 int16_t *sconsts = (int16_t *) consts;
958 printf("sconstants %d, %d, %d, %d\n",
959 sconsts[0],
960 sconsts[1],
961 sconsts[2],
962 sconsts[3]);
963 } else {
964 uint32_t *iconsts = (uint32_t *) consts;
965 printf("iconstants 0x%X, 0x%X, 0x%X, 0x%X\n",
966 iconsts[0],
967 iconsts[1],
968 iconsts[2],
969 iconsts[3]);
970 }
971 } else {
972 if (is_embedded_constant_half) {
973 uint16_t *hconsts = (uint16_t *) consts;
974 printf("hconstants %g, %g, %g, %g\n",
975 _mesa_half_to_float(hconsts[0]),
976 _mesa_half_to_float(hconsts[1]),
977 _mesa_half_to_float(hconsts[2]),
978 _mesa_half_to_float(hconsts[3]));
979 } else {
980 uint32_t *fconsts = (uint32_t *) consts;
981 printf("fconstants %g, %g, %g, %g\n",
982 float_bitcast(fconsts[0]),
983 float_bitcast(fconsts[1]),
984 float_bitcast(fconsts[2]),
985 float_bitcast(fconsts[3]));
986 }
987
988 }
989 }
990
991 return branch_forward;
992 }
993
994 static void
995 print_varying_parameters(midgard_load_store_word *word)
996 {
997 midgard_varying_parameter param;
998 unsigned v = word->varying_parameters;
999 memcpy(&param, &v, sizeof(param));
1000
1001 if (param.is_varying) {
1002 /* If a varying, there are qualifiers */
1003 if (param.flat)
1004 printf(".flat");
1005
1006 if (param.interpolation != midgard_interp_default) {
1007 if (param.interpolation == midgard_interp_centroid)
1008 printf(".centroid");
1009 else
1010 printf(".interp%d", param.interpolation);
1011 }
1012
1013 if (param.modifier != midgard_varying_mod_none) {
1014 if (param.modifier == midgard_varying_mod_perspective_w)
1015 printf(".perspectivew");
1016 else if (param.modifier == midgard_varying_mod_perspective_z)
1017 printf(".perspectivez");
1018 else
1019 printf(".mod%d", param.modifier);
1020 }
1021 } else if (param.flat || param.interpolation || param.modifier) {
1022 printf(" /* is_varying not set but varying metadata attached */");
1023 }
1024
1025 if (param.zero0 || param.zero1 || param.zero2)
1026 printf(" /* zero tripped, %u %u %u */ ", param.zero0, param.zero1, param.zero2);
1027 }
1028
1029 static bool
1030 is_op_varying(unsigned op)
1031 {
1032 switch (op) {
1033 case midgard_op_st_vary_16:
1034 case midgard_op_st_vary_32:
1035 case midgard_op_st_vary_32i:
1036 case midgard_op_st_vary_32u:
1037 case midgard_op_ld_vary_16:
1038 case midgard_op_ld_vary_32:
1039 case midgard_op_ld_vary_32i:
1040 case midgard_op_ld_vary_32u:
1041 return true;
1042 }
1043
1044 return false;
1045 }
1046
1047 static bool
1048 is_op_attribute(unsigned op)
1049 {
1050 switch (op) {
1051 case midgard_op_ld_attr_16:
1052 case midgard_op_ld_attr_32:
1053 case midgard_op_ld_attr_32i:
1054 case midgard_op_ld_attr_32u:
1055 return true;
1056 }
1057
1058 return false;
1059 }
1060
1061 static void
1062 print_load_store_arg(uint8_t arg, unsigned index)
1063 {
1064 /* Try to interpret as a register */
1065 midgard_ldst_register_select sel;
1066 memcpy(&sel, &arg, sizeof(arg));
1067
1068 /* If unknown is set, we're not sure what this is or how to
1069 * interpret it. But if it's zero, we get it. */
1070
1071 if (sel.unknown) {
1072 printf("0x%02X", arg);
1073 return;
1074 }
1075
1076 unsigned reg = REGISTER_LDST_BASE + sel.select;
1077 char comp = components[sel.component];
1078
1079 printf("r%u.%c", reg, comp);
1080
1081 /* Only print a shift if it's non-zero. Shifts only make sense for the
1082 * second index. For the first, we're not sure what it means yet */
1083
1084 if (index == 1) {
1085 if (sel.shift)
1086 printf(" << %u", sel.shift);
1087 } else {
1088 printf(" /* %X */", sel.shift);
1089 }
1090 }
1091
1092 static void
1093 update_stats(signed *stat, unsigned address)
1094 {
1095 if (*stat >= 0)
1096 *stat = MAX2(*stat, address + 1);
1097 }
1098
1099 static void
1100 print_load_store_instr(uint64_t data,
1101 unsigned tabs)
1102 {
1103 midgard_load_store_word *word = (midgard_load_store_word *) &data;
1104
1105 print_ld_st_opcode(word->op);
1106
1107 unsigned address = word->address;
1108
1109 if (is_op_varying(word->op)) {
1110 print_varying_parameters(word);
1111
1112 /* Do some analysis: check if direct cacess */
1113
1114 if ((word->arg_2 == 0x1E) && midg_stats.varying_count >= 0)
1115 update_stats(&midg_stats.varying_count, address);
1116 else
1117 midg_stats.varying_count = -16;
1118 } else if (is_op_attribute(word->op)) {
1119 if ((word->arg_2 == 0x1E) && midg_stats.attribute_count >= 0)
1120 update_stats(&midg_stats.attribute_count, address);
1121 else
1122 midg_stats.attribute_count = -16;
1123 }
1124
1125 printf(" r%u", word->reg);
1126 print_mask_4(word->mask, false);
1127
1128 if (!OP_IS_STORE(word->op))
1129 update_dest(word->reg);
1130
1131 bool is_ubo = OP_IS_UBO_READ(word->op);
1132
1133 if (is_ubo) {
1134 /* UBOs use their own addressing scheme */
1135
1136 int lo = word->varying_parameters >> 7;
1137 int hi = word->address;
1138
1139 /* TODO: Combine fields logically */
1140 address = (hi << 3) | lo;
1141 }
1142
1143 printf(", %u", address);
1144
1145 print_swizzle_vec4(word->swizzle, false, false);
1146
1147 printf(", ");
1148
1149 if (is_ubo) {
1150 printf("ubo%u", word->arg_1);
1151 update_stats(&midg_stats.uniform_buffer_count, word->arg_1);
1152 } else
1153 print_load_store_arg(word->arg_1, 0);
1154
1155 printf(", ");
1156 print_load_store_arg(word->arg_2, 1);
1157 printf(" /* %X */\n", word->varying_parameters);
1158
1159 midg_stats.instruction_count++;
1160 }
1161
1162 static void
1163 print_load_store_word(uint32_t *word, unsigned tabs)
1164 {
1165 midgard_load_store *load_store = (midgard_load_store *) word;
1166
1167 if (load_store->word1 != 3) {
1168 print_load_store_instr(load_store->word1, tabs);
1169 }
1170
1171 if (load_store->word2 != 3) {
1172 print_load_store_instr(load_store->word2, tabs);
1173 }
1174 }
1175
1176 static void
1177 print_texture_reg_select(uint8_t u, unsigned base)
1178 {
1179 midgard_tex_register_select sel;
1180 memcpy(&sel, &u, sizeof(u));
1181
1182 if (!sel.full)
1183 printf("h");
1184
1185 printf("r%u", base + sel.select);
1186
1187 unsigned component = sel.component;
1188
1189 /* Use the upper half in half-reg mode */
1190 if (sel.upper) {
1191 assert(!sel.full);
1192 component += 4;
1193 }
1194
1195 printf(".%c", components[component]);
1196
1197 assert(sel.zero == 0);
1198 }
1199
1200 static void
1201 print_texture_format(int format)
1202 {
1203 /* Act like a modifier */
1204 printf(".");
1205
1206 switch (format) {
1207 DEFINE_CASE(MALI_TEX_1D, "1d");
1208 DEFINE_CASE(MALI_TEX_2D, "2d");
1209 DEFINE_CASE(MALI_TEX_3D, "3d");
1210 DEFINE_CASE(MALI_TEX_CUBE, "cube");
1211
1212 default:
1213 unreachable("Bad format");
1214 }
1215 }
1216
1217 static bool
1218 midgard_op_has_helpers(unsigned op, bool gather)
1219 {
1220 if (gather)
1221 return true;
1222
1223 switch (op) {
1224 case TEXTURE_OP_NORMAL:
1225 case TEXTURE_OP_DFDX:
1226 case TEXTURE_OP_DFDY:
1227 return true;
1228 default:
1229 return false;
1230 }
1231 }
1232
1233 static void
1234 print_texture_op(unsigned op, bool gather)
1235 {
1236 /* Act like a bare name, like ESSL functions */
1237
1238 if (gather) {
1239 printf("textureGather");
1240
1241 unsigned component = op >> 4;
1242 unsigned bottom = op & 0xF;
1243
1244 if (bottom != 0x2)
1245 printf("_unk%u", bottom);
1246
1247 printf(".%c", components[component]);
1248 return;
1249 }
1250
1251 switch (op) {
1252 DEFINE_CASE(TEXTURE_OP_NORMAL, "texture");
1253 DEFINE_CASE(TEXTURE_OP_LOD, "textureLod");
1254 DEFINE_CASE(TEXTURE_OP_TEXEL_FETCH, "texelFetch");
1255 DEFINE_CASE(TEXTURE_OP_DFDX, "dFdx");
1256 DEFINE_CASE(TEXTURE_OP_DFDY, "dFdy");
1257
1258 default:
1259 printf("tex_%X", op);
1260 break;
1261 }
1262 }
1263
1264 static bool
1265 texture_op_takes_bias(unsigned op)
1266 {
1267 return op == TEXTURE_OP_NORMAL;
1268 }
1269
1270 static char
1271 sampler_type_name(enum mali_sampler_type t)
1272 {
1273 switch (t) {
1274 case MALI_SAMPLER_FLOAT:
1275 return 'f';
1276 case MALI_SAMPLER_UNSIGNED:
1277 return 'u';
1278 case MALI_SAMPLER_SIGNED:
1279 return 'i';
1280 default:
1281 return '?';
1282 }
1283
1284 }
1285
1286 #undef DEFINE_CASE
1287
1288 static void
1289 print_texture_word(uint32_t *word, unsigned tabs, unsigned in_reg_base, unsigned out_reg_base)
1290 {
1291 midgard_texture_word *texture = (midgard_texture_word *) word;
1292
1293 midg_stats.helper_invocations |=
1294 midgard_op_has_helpers(texture->op, texture->is_gather);
1295
1296 /* Broad category of texture operation in question */
1297 print_texture_op(texture->op, texture->is_gather);
1298
1299 /* Specific format in question */
1300 print_texture_format(texture->format);
1301
1302 /* Instruction "modifiers" parallel the ALU instructions. */
1303
1304 if (texture->shadow)
1305 printf(".shadow");
1306
1307 if (texture->cont)
1308 printf(".cont");
1309
1310 if (texture->last)
1311 printf(".last");
1312
1313 /* Output modifiers are always interpreted floatly */
1314 print_outmod(texture->outmod, false);
1315
1316 printf(" %sr%u", texture->out_full ? "" : "h",
1317 out_reg_base + texture->out_reg_select);
1318 print_mask_4(texture->mask, texture->out_upper);
1319 assert(!(texture->out_full && texture->out_upper));
1320 printf(", ");
1321
1322 /* Depending on whether we read from textures directly or indirectly,
1323 * we may be able to update our analysis */
1324
1325 if (texture->texture_register) {
1326 printf("texture[");
1327 print_texture_reg_select(texture->texture_handle, in_reg_base);
1328 printf("], ");
1329
1330 /* Indirect, tut tut */
1331 midg_stats.texture_count = -16;
1332 } else {
1333 printf("texture%u, ", texture->texture_handle);
1334 update_stats(&midg_stats.texture_count, texture->texture_handle);
1335 }
1336
1337 /* Print the type, GL style */
1338 printf("%csampler", sampler_type_name(texture->sampler_type));
1339
1340 if (texture->sampler_register) {
1341 printf("[");
1342 print_texture_reg_select(texture->sampler_handle, in_reg_base);
1343 printf("]");
1344
1345 midg_stats.sampler_count = -16;
1346 } else {
1347 printf("%u", texture->sampler_handle);
1348 update_stats(&midg_stats.sampler_count, texture->sampler_handle);
1349 }
1350
1351 print_swizzle_vec4(texture->swizzle, false, false);
1352 printf(", %sr%u", texture->in_reg_full ? "" : "h", in_reg_base + texture->in_reg_select);
1353 assert(!(texture->in_reg_full && texture->in_reg_upper));
1354
1355 /* TODO: integrate with swizzle */
1356 if (texture->in_reg_upper)
1357 printf("'");
1358
1359 print_swizzle_vec4(texture->in_reg_swizzle, false, false);
1360
1361 /* There is *always* an offset attached. Of
1362 * course, that offset is just immediate #0 for a
1363 * GLES call that doesn't take an offset. If there
1364 * is a non-negative non-zero offset, this is
1365 * specified in immediate offset mode, with the
1366 * values in the offset_* fields as immediates. If
1367 * this is a negative offset, we instead switch to
1368 * a register offset mode, where the offset_*
1369 * fields become register triplets */
1370
1371 if (texture->offset_register) {
1372 printf(" + ");
1373
1374 bool full = texture->offset & 1;
1375 bool select = texture->offset & 2;
1376 bool upper = texture->offset & 4;
1377
1378 printf("%sr%u", full ? "" : "h", in_reg_base + select);
1379 assert(!(texture->out_full && texture->out_upper));
1380
1381 /* TODO: integrate with swizzle */
1382 if (upper)
1383 printf("'");
1384
1385 print_swizzle_vec4(texture->offset >> 3, false, false);
1386
1387 printf(", ");
1388 } else if (texture->offset) {
1389 /* Only select ops allow negative immediate offsets, verify */
1390
1391 signed offset_x = (texture->offset & 0xF);
1392 signed offset_y = ((texture->offset >> 4) & 0xF);
1393 signed offset_z = ((texture->offset >> 8) & 0xF);
1394
1395 bool neg_x = offset_x < 0;
1396 bool neg_y = offset_y < 0;
1397 bool neg_z = offset_z < 0;
1398 bool any_neg = neg_x || neg_y || neg_z;
1399
1400 if (any_neg && texture->op != TEXTURE_OP_TEXEL_FETCH)
1401 printf("/* invalid negative */ ");
1402
1403 /* Regardless, just print the immediate offset */
1404
1405 printf(" + <%d, %d, %d>, ", offset_x, offset_y, offset_z);
1406 } else {
1407 printf(", ");
1408 }
1409
1410 char lod_operand = texture_op_takes_bias(texture->op) ? '+' : '=';
1411
1412 if (texture->lod_register) {
1413 printf("lod %c ", lod_operand);
1414 print_texture_reg_select(texture->bias, in_reg_base);
1415 printf(", ");
1416
1417 if (texture->bias_int)
1418 printf(" /* bias_int = 0x%X */", texture->bias_int);
1419 } else if (texture->op == TEXTURE_OP_TEXEL_FETCH) {
1420 /* For texel fetch, the int LOD is in the fractional place and
1421 * there is no fraction / possibility of bias. We *always* have
1422 * an explicit LOD, even if it's zero. */
1423
1424 if (texture->bias_int)
1425 printf(" /* bias_int = 0x%X */ ", texture->bias_int);
1426
1427 printf("lod = %u, ", texture->bias);
1428 } else if (texture->bias || texture->bias_int) {
1429 signed bias_int = texture->bias_int;
1430 float bias_frac = texture->bias / 256.0f;
1431 float bias = bias_int + bias_frac;
1432
1433 bool is_bias = texture_op_takes_bias(texture->op);
1434 char sign = (bias >= 0.0) ? '+' : '-';
1435 char operand = is_bias ? sign : '=';
1436
1437 printf("lod %c %f, ", operand, fabsf(bias));
1438 }
1439
1440 printf("\n");
1441
1442 /* While not zero in general, for these simple instructions the
1443 * following unknowns are zero, so we don't include them */
1444
1445 if (texture->unknown4 ||
1446 texture->unknownA ||
1447 texture->unknown8) {
1448 printf("// unknown4 = 0x%x\n", texture->unknown4);
1449 printf("// unknownA = 0x%x\n", texture->unknownA);
1450 printf("// unknown8 = 0x%x\n", texture->unknown8);
1451 }
1452
1453 midg_stats.instruction_count++;
1454 }
1455
1456 struct midgard_disasm_stats
1457 disassemble_midgard(uint8_t *code, size_t size, unsigned gpu_id, gl_shader_stage stage)
1458 {
1459 uint32_t *words = (uint32_t *) code;
1460 unsigned num_words = size / 4;
1461 int tabs = 0;
1462
1463 bool branch_forward = false;
1464
1465 int last_next_tag = -1;
1466
1467 unsigned i = 0;
1468
1469 midg_tags = calloc(sizeof(midg_tags[0]), num_words);
1470
1471 /* Stats for shader-db */
1472 memset(&midg_stats, 0, sizeof(midg_stats));
1473 midg_ever_written = 0;
1474
1475 while (i < num_words) {
1476 unsigned tag = words[i] & 0xF;
1477 unsigned next_tag = (words[i] >> 4) & 0xF;
1478 printf("\t%X -> %X\n", tag, next_tag);
1479 unsigned num_quad_words = midgard_word_size[tag];
1480
1481 if (midg_tags[i] && midg_tags[i] != tag) {
1482 printf("\t/* XXX: TAG ERROR branch, got ");
1483 print_tag_short(tag);
1484 printf(" expected ");
1485 print_tag_short(midg_tags[i]);
1486 printf(" */\n");
1487 }
1488
1489 midg_tags[i] = tag;
1490
1491 /* Check the tag */
1492 if (last_next_tag > 1) {
1493 if (last_next_tag != tag) {
1494 printf("\t/* XXX: TAG ERROR sequence, got ");
1495 print_tag_short(tag);
1496 printf(" expected ");
1497 print_tag_short(last_next_tag);
1498 printf(" */\n");
1499 }
1500 } else {
1501 /* TODO: Check ALU case */
1502 }
1503
1504 last_next_tag = next_tag;
1505
1506 switch (midgard_word_types[tag]) {
1507 case midgard_word_type_texture: {
1508 bool interpipe_aliasing =
1509 midgard_get_quirks(gpu_id) & MIDGARD_INTERPIPE_REG_ALIASING;
1510
1511 print_texture_word(&words[i], tabs,
1512 interpipe_aliasing ? 0 : REG_TEX_BASE,
1513 interpipe_aliasing ? REGISTER_LDST_BASE : REG_TEX_BASE);
1514 break;
1515 }
1516
1517 case midgard_word_type_load_store:
1518 print_load_store_word(&words[i], tabs);
1519 break;
1520
1521 case midgard_word_type_alu:
1522 branch_forward = print_alu_word(&words[i], num_quad_words, tabs, i + 4*num_quad_words);
1523
1524 /* Reset word static analysis state */
1525 is_embedded_constant_half = false;
1526 is_embedded_constant_int = false;
1527
1528 break;
1529
1530 default:
1531 printf("Unknown word type %u:\n", words[i] & 0xF);
1532 num_quad_words = 1;
1533 print_quad_word(&words[i], tabs);
1534 printf("\n");
1535 break;
1536 }
1537
1538 /* We are parsing per bundle anyway. Add before we start
1539 * breaking out so we don't miss the final bundle. */
1540
1541 midg_stats.bundle_count++;
1542 midg_stats.quadword_count += num_quad_words;
1543
1544 printf("\n");
1545
1546 unsigned next = (words[i] & 0xF0) >> 4;
1547
1548 if (i < num_words && next == 1 && !branch_forward)
1549 break;
1550
1551 i += 4 * num_quad_words;
1552 }
1553
1554 /* We computed work_count as max_work_registers, so add one to get the
1555 * count. If no work registers are written, you still have one work
1556 * reported, which is exactly what the hardware expects */
1557
1558 midg_stats.work_count++;
1559
1560 return midg_stats;
1561 }