1 /*
2 * Copyright (c) 2013 Rob Clark <robdclark@gmail.com>
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 */
23
24 #include <assert.h>
25 #include <stdbool.h>
26 #include <stdint.h>
27 #include <stdio.h>
28 #include <stdlib.h>
29 #include <string.h>
30
31 #include <util/log.h>
32 #include <util/u_debug.h>
33
34 #include "isa/isa.h"
35
36 #include "disasm.h"
37 #include "instr-a3xx.h"
38
39 static enum debug_t debug;
40
41 static const char *levels[] = {
42 "",
43 "\t",
44 "\t\t",
45 "\t\t\t",
46 "\t\t\t\t",
47 "\t\t\t\t\t",
48 "\t\t\t\t\t\t",
49 "\t\t\t\t\t\t\t",
50 "\t\t\t\t\t\t\t\t",
51 "\t\t\t\t\t\t\t\t\t",
52 "x",
53 "x",
54 "x",
55 "x",
56 "x",
57 "x",
58 };
59
60 struct disasm_ctx {
61 FILE *out;
62 struct isa_decode_options *options;
63 unsigned level;
64 unsigned extra_cycles;
65
66 /**
67 * nop_count/has_end used to detect the real end of shader. Since
68 * in some cases there can be a epilogue following an `end` we look
69 * for a sequence of `nop`s following the `end`
70 */
71 int nop_count; /* number of nop's since non-nop instruction: */
72 bool has_end; /* have we seen end instruction */
73
74 int cur_n; /* current instr # */
75 int cur_opc_cat; /* current opc_cat */
76
77 int sfu_delay;
78
79 /**
80 * State accumulated decoding fields of the current instruction,
81 * handled after decoding is complete (ie. at start of next instr)
82 */
83 struct {
84 bool ss;
85 uint8_t nop;
86 uint8_t repeat;
87 } last;
88
89 /**
90 * State accumulated decoding fields of src or dst register
91 */
92 struct {
93 bool half;
94 bool r;
95 enum {
96 FILE_GPR = 1,
97 FILE_CONST = 2,
98 } file;
99 unsigned num;
100 } reg;
101
102 struct shader_stats *stats;
103 };
104
105 static void
print_stats(struct disasm_ctx * ctx)106 print_stats(struct disasm_ctx *ctx)
107 {
108 if (ctx->options->gpu_id >= 600) {
109 /* handle MERGEREGS case.. this isn't *entirely* accurate, as
110 * you can have shader stages not using merged register file,
111 * but it is good enough for a guestimate:
112 */
113 unsigned n = (ctx->stats->halfreg + 1) / 2;
114
115 ctx->stats->halfreg = 0;
116 ctx->stats->fullreg = MAX2(ctx->stats->fullreg, n);
117 }
118
119 unsigned instructions = ctx->cur_n + ctx->extra_cycles + 1;
120
121 fprintf(ctx->out, "%sStats:\n", levels[ctx->level]);
122 fprintf(ctx->out,
123 "%s- shaderdb: %u instr, %u nops, %u non-nops, %u mov, %u cov\n",
124 levels[ctx->level], instructions, ctx->stats->nops,
125 instructions - ctx->stats->nops, ctx->stats->mov_count,
126 ctx->stats->cov_count);
127
128 fprintf(ctx->out,
129 "%s- shaderdb: %u last-baryf, %d half, %d full, %u constlen\n",
130 levels[ctx->level], ctx->stats->last_baryf,
131 DIV_ROUND_UP(ctx->stats->halfreg, 4),
132 DIV_ROUND_UP(ctx->stats->fullreg, 4),
133 DIV_ROUND_UP(ctx->stats->constlen, 4));
134
135 fprintf(
136 ctx->out,
137 "%s- shaderdb: %u cat0, %u cat1, %u cat2, %u cat3, %u cat4, %u cat5, %u cat6, %u cat7\n",
138 levels[ctx->level], ctx->stats->instrs_per_cat[0],
139 ctx->stats->instrs_per_cat[1], ctx->stats->instrs_per_cat[2],
140 ctx->stats->instrs_per_cat[3], ctx->stats->instrs_per_cat[4],
141 ctx->stats->instrs_per_cat[5], ctx->stats->instrs_per_cat[6],
142 ctx->stats->instrs_per_cat[7]);
143
144 fprintf(ctx->out, "%s- shaderdb: %u sstall, %u (ss), %u (sy)\n",
145 levels[ctx->level], ctx->stats->sstall, ctx->stats->ss,
146 ctx->stats->sy);
147 }
148
149 static const struct opc_info {
150 const char *name;
151 } opcs[1 << (3 + NOPC_BITS)] = {
152 #define OPC(cat, opc, name) [(opc)] = {#name}
153 /* clang-format off */
154 /* category 0: */
155 OPC(0, OPC_NOP, nop),
156 OPC(0, OPC_B, b),
157 OPC(0, OPC_JUMP, jump),
158 OPC(0, OPC_CALL, call),
159 OPC(0, OPC_RET, ret),
160 OPC(0, OPC_KILL, kill),
161 OPC(0, OPC_DEMOTE, demote),
162 OPC(0, OPC_END, end),
163 OPC(0, OPC_EMIT, emit),
164 OPC(0, OPC_CUT, cut),
165 OPC(0, OPC_CHMASK, chmask),
166 OPC(0, OPC_CHSH, chsh),
167 OPC(0, OPC_FLOW_REV, flow_rev),
168 OPC(0, OPC_PREDT, predt),
169 OPC(0, OPC_PREDF, predf),
170 OPC(0, OPC_PREDE, prede),
171 OPC(0, OPC_BKT, bkt),
172 OPC(0, OPC_STKS, stks),
173 OPC(0, OPC_STKR, stkr),
174 OPC(0, OPC_XSET, xset),
175 OPC(0, OPC_XCLR, xclr),
176 OPC(0, OPC_GETLAST, getlast),
177 OPC(0, OPC_GETONE, getone),
178 OPC(0, OPC_DBG, dbg),
179 OPC(0, OPC_SHPS, shps),
180 OPC(0, OPC_SHPE, shpe),
181
182 /* category 1: */
183 OPC(1, OPC_MOV, ),
184 OPC(1, OPC_MOVMSK, movmsk),
185 OPC(1, OPC_SWZ, swz),
186 OPC(1, OPC_SCT, sct),
187 OPC(1, OPC_GAT, gat),
188 OPC(1, OPC_BALLOT_MACRO, ballot.macro),
189 OPC(1, OPC_ANY_MACRO, any.macro),
190 OPC(1, OPC_ALL_MACRO, all.macro),
191 OPC(1, OPC_ELECT_MACRO, elect.macro),
192 OPC(1, OPC_READ_COND_MACRO, read_cond.macro),
193 OPC(1, OPC_READ_FIRST_MACRO, read_first.macro),
194 OPC(1, OPC_SWZ_SHARED_MACRO, swz_shared.macro),
195 OPC(1, OPC_SCAN_MACRO, scan.macro),
196 OPC(1, OPC_SCAN_CLUSTERS_MACRO, scan_clusters.macro),
197 OPC(1, OPC_SHPS_MACRO, shps.macro),
198 OPC(1, OPC_PUSH_CONSTS_LOAD_MACRO, push_consts_load.macro),
199
200 /* category 2: */
201 OPC(2, OPC_ADD_F, add.f),
202 OPC(2, OPC_MIN_F, min.f),
203 OPC(2, OPC_MAX_F, max.f),
204 OPC(2, OPC_MUL_F, mul.f),
205 OPC(2, OPC_SIGN_F, sign.f),
206 OPC(2, OPC_CMPS_F, cmps.f),
207 OPC(2, OPC_ABSNEG_F, absneg.f),
208 OPC(2, OPC_CMPV_F, cmpv.f),
209 OPC(2, OPC_FLOOR_F, floor.f),
210 OPC(2, OPC_CEIL_F, ceil.f),
211 OPC(2, OPC_RNDNE_F, rndne.f),
212 OPC(2, OPC_RNDAZ_F, rndaz.f),
213 OPC(2, OPC_TRUNC_F, trunc.f),
214 OPC(2, OPC_ADD_U, add.u),
215 OPC(2, OPC_ADD_S, add.s),
216 OPC(2, OPC_SUB_U, sub.u),
217 OPC(2, OPC_SUB_S, sub.s),
218 OPC(2, OPC_CMPS_U, cmps.u),
219 OPC(2, OPC_CMPS_S, cmps.s),
220 OPC(2, OPC_MIN_U, min.u),
221 OPC(2, OPC_MIN_S, min.s),
222 OPC(2, OPC_MAX_U, max.u),
223 OPC(2, OPC_MAX_S, max.s),
224 OPC(2, OPC_ABSNEG_S, absneg.s),
225 OPC(2, OPC_AND_B, and.b),
226 OPC(2, OPC_OR_B, or.b),
227 OPC(2, OPC_NOT_B, not.b),
228 OPC(2, OPC_XOR_B, xor.b),
229 OPC(2, OPC_CMPV_U, cmpv.u),
230 OPC(2, OPC_CMPV_S, cmpv.s),
231 OPC(2, OPC_MUL_U24, mul.u24),
232 OPC(2, OPC_MUL_S24, mul.s24),
233 OPC(2, OPC_MULL_U, mull.u),
234 OPC(2, OPC_BFREV_B, bfrev.b),
235 OPC(2, OPC_CLZ_S, clz.s),
236 OPC(2, OPC_CLZ_B, clz.b),
237 OPC(2, OPC_SHL_B, shl.b),
238 OPC(2, OPC_SHR_B, shr.b),
239 OPC(2, OPC_ASHR_B, ashr.b),
240 OPC(2, OPC_BARY_F, bary.f),
241 OPC(2, OPC_MGEN_B, mgen.b),
242 OPC(2, OPC_GETBIT_B, getbit.b),
243 OPC(2, OPC_SETRM, setrm),
244 OPC(2, OPC_CBITS_B, cbits.b),
245 OPC(2, OPC_SHB, shb),
246 OPC(2, OPC_MSAD, msad),
247
248 /* category 3: */
249 OPC(3, OPC_MAD_U16, mad.u16),
250 OPC(3, OPC_MADSH_U16, madsh.u16),
251 OPC(3, OPC_MAD_S16, mad.s16),
252 OPC(3, OPC_MADSH_M16, madsh.m16),
253 OPC(3, OPC_MAD_U24, mad.u24),
254 OPC(3, OPC_MAD_S24, mad.s24),
255 OPC(3, OPC_MAD_F16, mad.f16),
256 OPC(3, OPC_MAD_F32, mad.f32),
257 OPC(3, OPC_SEL_B16, sel.b16),
258 OPC(3, OPC_SEL_B32, sel.b32),
259 OPC(3, OPC_SEL_S16, sel.s16),
260 OPC(3, OPC_SEL_S32, sel.s32),
261 OPC(3, OPC_SEL_F16, sel.f16),
262 OPC(3, OPC_SEL_F32, sel.f32),
263 OPC(3, OPC_SAD_S16, sad.s16),
264 OPC(3, OPC_SAD_S32, sad.s32),
265 OPC(3, OPC_SHRM, shrm),
266 OPC(3, OPC_SHLM, shlm),
267 OPC(3, OPC_SHRG, shrg),
268 OPC(3, OPC_SHLG, shlg),
269 OPC(3, OPC_ANDG, andg),
270 OPC(3, OPC_DP2ACC, dp2acc),
271 OPC(3, OPC_DP4ACC, dp4acc),
272 OPC(3, OPC_WMM, wmm),
273 OPC(3, OPC_WMM_ACCU, wmm.accu),
274
275 /* category 4: */
276 OPC(4, OPC_RCP, rcp),
277 OPC(4, OPC_RSQ, rsq),
278 OPC(4, OPC_LOG2, log2),
279 OPC(4, OPC_EXP2, exp2),
280 OPC(4, OPC_SIN, sin),
281 OPC(4, OPC_COS, cos),
282 OPC(4, OPC_SQRT, sqrt),
283 OPC(4, OPC_HRSQ, hrsq),
284 OPC(4, OPC_HLOG2, hlog2),
285 OPC(4, OPC_HEXP2, hexp2),
286
287 /* category 5: */
288 OPC(5, OPC_ISAM, isam),
289 OPC(5, OPC_ISAML, isaml),
290 OPC(5, OPC_ISAMM, isamm),
291 OPC(5, OPC_SAM, sam),
292 OPC(5, OPC_SAMB, samb),
293 OPC(5, OPC_SAML, saml),
294 OPC(5, OPC_SAMGQ, samgq),
295 OPC(5, OPC_GETLOD, getlod),
296 OPC(5, OPC_CONV, conv),
297 OPC(5, OPC_CONVM, convm),
298 OPC(5, OPC_GETSIZE, getsize),
299 OPC(5, OPC_GETBUF, getbuf),
300 OPC(5, OPC_GETPOS, getpos),
301 OPC(5, OPC_GETINFO, getinfo),
302 OPC(5, OPC_DSX, dsx),
303 OPC(5, OPC_DSY, dsy),
304 OPC(5, OPC_GATHER4R, gather4r),
305 OPC(5, OPC_GATHER4G, gather4g),
306 OPC(5, OPC_GATHER4B, gather4b),
307 OPC(5, OPC_GATHER4A, gather4a),
308 OPC(5, OPC_SAMGP0, samgp0),
309 OPC(5, OPC_SAMGP1, samgp1),
310 OPC(5, OPC_SAMGP2, samgp2),
311 OPC(5, OPC_SAMGP3, samgp3),
312 OPC(5, OPC_DSXPP_1, dsxpp.1),
313 OPC(5, OPC_DSYPP_1, dsypp.1),
314 OPC(5, OPC_RGETPOS, rgetpos),
315 OPC(5, OPC_RGETINFO, rgetinfo),
316 OPC(5, OPC_BRCST_ACTIVE, brcst.active),
317 OPC(5, OPC_QUAD_SHUFFLE_BRCST, quad_shuffle.brcst),
318 OPC(5, OPC_QUAD_SHUFFLE_HORIZ, quad_shuffle.horiz),
319 OPC(5, OPC_QUAD_SHUFFLE_VERT, quad_shuffle.vert),
320 OPC(5, OPC_QUAD_SHUFFLE_DIAG, quad_shuffle.diag),
321 OPC(5, OPC_TCINV, tcinv),
322 /* macros are needed here for ir3_print */
323 OPC(5, OPC_DSXPP_MACRO, dsxpp.macro),
324 OPC(5, OPC_DSYPP_MACRO, dsypp.macro),
325
326
327 /* category 6: */
328 OPC(6, OPC_LDG, ldg),
329 OPC(6, OPC_LDG_A, ldg.a),
330 OPC(6, OPC_LDL, ldl),
331 OPC(6, OPC_LDP, ldp),
332 OPC(6, OPC_STG, stg),
333 OPC(6, OPC_STG_A, stg.a),
334 OPC(6, OPC_STL, stl),
335 OPC(6, OPC_STP, stp),
336 OPC(6, OPC_LDIB, ldib),
337 OPC(6, OPC_G2L, g2l),
338 OPC(6, OPC_L2G, l2g),
339 OPC(6, OPC_PREFETCH, prefetch),
340 OPC(6, OPC_LDLW, ldlw),
341 OPC(6, OPC_STLW, stlw),
342 OPC(6, OPC_RESFMT, resfmt),
343 OPC(6, OPC_RESINFO, resinfo),
344 OPC(6, OPC_ATOMIC_ADD, atomic.add),
345 OPC(6, OPC_ATOMIC_SUB, atomic.sub),
346 OPC(6, OPC_ATOMIC_XCHG, atomic.xchg),
347 OPC(6, OPC_ATOMIC_INC, atomic.inc),
348 OPC(6, OPC_ATOMIC_DEC, atomic.dec),
349 OPC(6, OPC_ATOMIC_CMPXCHG, atomic.cmpxchg),
350 OPC(6, OPC_ATOMIC_MIN, atomic.min),
351 OPC(6, OPC_ATOMIC_MAX, atomic.max),
352 OPC(6, OPC_ATOMIC_AND, atomic.and),
353 OPC(6, OPC_ATOMIC_OR, atomic.or),
354 OPC(6, OPC_ATOMIC_XOR, atomic.xor),
355 OPC(6, OPC_ATOMIC_B_ADD, atomic.b.add),
356 OPC(6, OPC_ATOMIC_B_SUB, atomic.b.sub),
357 OPC(6, OPC_ATOMIC_B_XCHG, atomic.b.xchg),
358 OPC(6, OPC_ATOMIC_B_INC, atomic.b.inc),
359 OPC(6, OPC_ATOMIC_B_DEC, atomic.b.dec),
360 OPC(6, OPC_ATOMIC_B_CMPXCHG, atomic.b.cmpxchg),
361 OPC(6, OPC_ATOMIC_B_MIN, atomic.b.min),
362 OPC(6, OPC_ATOMIC_B_MAX, atomic.b.max),
363 OPC(6, OPC_ATOMIC_B_AND, atomic.b.and),
364 OPC(6, OPC_ATOMIC_B_OR, atomic.b.or),
365 OPC(6, OPC_ATOMIC_B_XOR, atomic.b.xor),
366 OPC(6, OPC_ATOMIC_S_ADD, atomic.s.add),
367 OPC(6, OPC_ATOMIC_S_SUB, atomic.s.sub),
368 OPC(6, OPC_ATOMIC_S_XCHG, atomic.s.xchg),
369 OPC(6, OPC_ATOMIC_S_INC, atomic.s.inc),
370 OPC(6, OPC_ATOMIC_S_DEC, atomic.s.dec),
371 OPC(6, OPC_ATOMIC_S_CMPXCHG, atomic.s.cmpxchg),
372 OPC(6, OPC_ATOMIC_S_MIN, atomic.s.min),
373 OPC(6, OPC_ATOMIC_S_MAX, atomic.s.max),
374 OPC(6, OPC_ATOMIC_S_AND, atomic.s.and),
375 OPC(6, OPC_ATOMIC_S_OR, atomic.s.or),
376 OPC(6, OPC_ATOMIC_S_XOR, atomic.s.xor),
377 OPC(6, OPC_ATOMIC_G_ADD, atomic.g.add),
378 OPC(6, OPC_ATOMIC_G_SUB, atomic.g.sub),
379 OPC(6, OPC_ATOMIC_G_XCHG, atomic.g.xchg),
380 OPC(6, OPC_ATOMIC_G_INC, atomic.g.inc),
381 OPC(6, OPC_ATOMIC_G_DEC, atomic.g.dec),
382 OPC(6, OPC_ATOMIC_G_CMPXCHG, atomic.g.cmpxchg),
383 OPC(6, OPC_ATOMIC_G_MIN, atomic.g.min),
384 OPC(6, OPC_ATOMIC_G_MAX, atomic.g.max),
385 OPC(6, OPC_ATOMIC_G_AND, atomic.g.and),
386 OPC(6, OPC_ATOMIC_G_OR, atomic.g.or),
387 OPC(6, OPC_ATOMIC_G_XOR, atomic.g.xor),
388 OPC(6, OPC_LDGB, ldgb),
389 OPC(6, OPC_STGB, stgb),
390 OPC(6, OPC_STIB, stib),
391 OPC(6, OPC_LDC, ldc),
392 OPC(6, OPC_LDLV, ldlv),
393 OPC(6, OPC_PIPR, pipr),
394 OPC(6, OPC_PIPC, pipc),
395 OPC(6, OPC_EMIT2, emit),
396 OPC(6, OPC_ENDLS, endls),
397 OPC(6, OPC_GETSPID, getspid),
398 OPC(6, OPC_GETWID, getwid),
399 OPC(6, OPC_GETFIBERID, getfiberid),
400 OPC(6, OPC_STC, stc),
401 OPC(6, OPC_STSC, stsc),
402 OPC(6, OPC_LDC_K, ldc.k),
403 OPC(6, OPC_LDG_K, ldg.k),
404
405 OPC(6, OPC_SPILL_MACRO, spill.macro),
406 OPC(6, OPC_RELOAD_MACRO, reload.macro),
407
408 OPC(7, OPC_BAR, bar),
409 OPC(7, OPC_FENCE, fence),
410 OPC(7, OPC_LOCK, lock),
411 OPC(7, OPC_UNLOCK, unlock),
412 /* clang-format on */
413 #undef OPC
414 };
415
416 const char *
disasm_a3xx_instr_name(opc_t opc)417 disasm_a3xx_instr_name(opc_t opc)
418 {
419 if (opc_cat(opc) == OPC_META)
420 return "??meta??";
421 return opcs[opc].name;
422 }
423
424 static void
disasm_field_cb(void * d,const char * field_name,struct isa_decode_value * val)425 disasm_field_cb(void *d, const char *field_name, struct isa_decode_value *val)
426 {
427 struct disasm_ctx *ctx = d;
428
429 if (!strcmp(field_name, "NAME")) {
430 if (!strcmp("nop", val->str)) {
431 if (ctx->has_end) {
432 ctx->nop_count++;
433 if (ctx->nop_count > 3) {
434 ctx->options->stop = true;
435 }
436 }
437 ctx->stats->nops += 1 + ctx->last.repeat;
438 } else {
439 ctx->nop_count = 0;
440 }
441
442 if (!strcmp("end", val->str)) {
443 ctx->has_end = true;
444 ctx->nop_count = 0;
445 } else if (!strcmp("chsh", val->str)) {
446 ctx->options->stop = true;
447 } else if (!strcmp("bary.f", val->str)) {
448 ctx->stats->last_baryf = ctx->cur_n;
449 }
450 } else if (!strcmp(field_name, "REPEAT")) {
451 ctx->extra_cycles += val->num;
452 ctx->stats->instrs_per_cat[ctx->cur_opc_cat] += val->num;
453 ctx->last.repeat = val->num;
454 } else if (!strcmp(field_name, "NOP")) {
455 ctx->extra_cycles += val->num;
456 ctx->stats->instrs_per_cat[0] += val->num;
457 ctx->stats->nops += val->num;
458 ctx->last.nop = val->num;
459 } else if (!strcmp(field_name, "SY")) {
460 ctx->stats->sy += val->num;
461 } else if (!strcmp(field_name, "SS")) {
462 ctx->stats->ss += val->num;
463 ctx->last.ss = !!val->num;
464 } else if (!strcmp(field_name, "CONST")) {
465 ctx->reg.num = val->num;
466 ctx->reg.file = FILE_CONST;
467 } else if (!strcmp(field_name, "GPR")) {
468 /* don't count GPR regs r48.x (shared) or higher: */
469 if (val->num < 48) {
470 ctx->reg.num = val->num;
471 ctx->reg.file = FILE_GPR;
472 }
473 } else if (!strcmp(field_name, "SRC_R") || !strcmp(field_name, "SRC1_R") ||
474 !strcmp(field_name, "SRC2_R") || !strcmp(field_name, "SRC3_R")) {
475 ctx->reg.r = val->num;
476 } else if (!strcmp(field_name, "DST")) {
477 /* Dest register is always repeated
478 *
479 * Note that this doesn't really properly handle instructions
480 * that write multiple components.. the old disasm didn't handle
481 * that case either.
482 */
483 ctx->reg.r = true;
484 } else if (strstr(field_name, "HALF")) {
485 ctx->reg.half = val->num;
486 } else if (!strcmp(field_name, "SWIZ")) {
487 unsigned num = (ctx->reg.num << 2) | val->num;
488 if (ctx->reg.r)
489 num += ctx->last.repeat;
490
491 if (ctx->reg.file == FILE_CONST) {
492 ctx->stats->constlen = MAX2(ctx->stats->constlen, num);
493 } else if (ctx->reg.file == FILE_GPR) {
494 if (ctx->reg.half) {
495 ctx->stats->halfreg = MAX2(ctx->stats->halfreg, num);
496 } else {
497 ctx->stats->fullreg = MAX2(ctx->stats->fullreg, num);
498 }
499 }
500
501 memset(&ctx->reg, 0, sizeof(ctx->reg));
502 }
503 }
504
505 /**
506 * Handle stat updates dealt with at the end of instruction decoding,
507 * ie. before beginning of next instruction
508 */
509 static void
disasm_handle_last(struct disasm_ctx * ctx)510 disasm_handle_last(struct disasm_ctx *ctx)
511 {
512 if (ctx->last.ss) {
513 ctx->stats->sstall += ctx->sfu_delay;
514 ctx->sfu_delay = 0;
515 }
516
517 if (ctx->cur_opc_cat == 4) {
518 ctx->sfu_delay = 10;
519 } else {
520 int n = MIN2(ctx->sfu_delay, 1 + ctx->last.repeat + ctx->last.nop);
521 ctx->sfu_delay -= n;
522 }
523
524 memset(&ctx->last, 0, sizeof(ctx->last));
525 }
526
527 static void
disasm_instr_cb(void * d,unsigned n,void * instr)528 disasm_instr_cb(void *d, unsigned n, void *instr)
529 {
530 struct disasm_ctx *ctx = d;
531 uint32_t *dwords = (uint32_t *)instr;
532 uint64_t val = dwords[1];
533 val = val << 32;
534 val |= dwords[0];
535
536 unsigned opc_cat = val >> 61;
537
538 /* There are some cases where we can get instr_cb called multiple
539 * times per instruction (like when we need an extra line for branch
540 * target labels), don't update stats in these cases:
541 */
542 if (n != ctx->cur_n) {
543 if (n > 0) {
544 disasm_handle_last(ctx);
545 }
546 ctx->stats->instrs_per_cat[opc_cat]++;
547 ctx->cur_n = n;
548
549 /* mov vs cov stats are a bit harder to fish out of the field
550 * names, because current ir3-cat1.xml doesn't use {NAME} for
551 * this distinction. So for now just handle this case with
552 * some hand-coded parsing:
553 */
554 if (opc_cat == 1) {
555 unsigned opc = (val >> 57) & 0x3;
556 unsigned src_type = (val >> 50) & 0x7;
557 unsigned dst_type = (val >> 46) & 0x7;
558
559 if (opc == 0) {
560 if (src_type == dst_type) {
561 ctx->stats->mov_count++;
562 } else {
563 ctx->stats->cov_count++;
564 }
565 }
566 }
567 }
568
569 ctx->cur_opc_cat = opc_cat;
570
571 if (debug & PRINT_RAW) {
572 fprintf(ctx->out, "%s:%d:%04d:%04d[%08xx_%08xx] ", levels[ctx->level],
573 opc_cat, n, ctx->extra_cycles + n, dwords[1], dwords[0]);
574 }
575 }
576
577 int
disasm_a3xx_stat(uint32_t * dwords,int sizedwords,int level,FILE * out,unsigned gpu_id,struct shader_stats * stats)578 disasm_a3xx_stat(uint32_t *dwords, int sizedwords, int level, FILE *out,
579 unsigned gpu_id, struct shader_stats *stats)
580 {
581 struct isa_decode_options decode_options = {
582 .gpu_id = gpu_id,
583 .show_errors = true,
584 .max_errors = 5,
585 .branch_labels = true,
586 .field_cb = disasm_field_cb,
587 .pre_instr_cb = disasm_instr_cb,
588 };
589 struct disasm_ctx ctx = {
590 .out = out,
591 .level = level,
592 .options = &decode_options,
593 .stats = stats,
594 .cur_n = -1,
595 };
596
597 memset(stats, 0, sizeof(*stats));
598
599 decode_options.cbdata = &ctx;
600
601 isa_disasm(dwords, sizedwords * 4, out, &decode_options);
602
603 disasm_handle_last(&ctx);
604
605 if (debug & PRINT_STATS)
606 print_stats(&ctx);
607
608 return 0;
609 }
610
611 void
disasm_a3xx_set_debug(enum debug_t d)612 disasm_a3xx_set_debug(enum debug_t d)
613 {
614 debug = d;
615 }
616
617 #include <setjmp.h>
618
619 static bool jmp_env_valid;
620 static jmp_buf jmp_env;
621
622 void
ir3_assert_handler(const char * expr,const char * file,int line,const char * func)623 ir3_assert_handler(const char *expr, const char *file, int line,
624 const char *func)
625 {
626 mesa_loge("%s:%u: %s: Assertion `%s' failed.", file, line, func, expr);
627 if (jmp_env_valid)
628 longjmp(jmp_env, 1);
629 abort();
630 }
631
632 #define TRY(x) \
633 do { \
634 assert(!jmp_env_valid); \
635 if (setjmp(jmp_env) == 0) { \
636 jmp_env_valid = true; \
637 x; \
638 } \
639 jmp_env_valid = false; \
640 } while (0)
641
642 int
disasm_a3xx(uint32_t * dwords,int sizedwords,int level,FILE * out,unsigned gpu_id)643 disasm_a3xx(uint32_t *dwords, int sizedwords, int level, FILE *out,
644 unsigned gpu_id)
645 {
646 struct shader_stats stats;
647 return disasm_a3xx_stat(dwords, sizedwords, level, out, gpu_id, &stats);
648 }
649
650 int
try_disasm_a3xx(uint32_t * dwords,int sizedwords,int level,FILE * out,unsigned gpu_id)651 try_disasm_a3xx(uint32_t *dwords, int sizedwords, int level, FILE *out,
652 unsigned gpu_id)
653 {
654 struct shader_stats stats;
655 int ret = -1;
656 TRY(ret = disasm_a3xx_stat(dwords, sizedwords, level, out, gpu_id, &stats));
657 return ret;
658 }
659