2
* Copyright © 2018 Valve Corporation
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:
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
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
20
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25
#include "aco_builder.h"
28
#include "common/ac_shader_util.h"
29
#include "common/sid.h"
35
const std::array<const char*, num_reduce_ops> reduce_ops = []()
37
std::array<const char*, num_reduce_ops> ret{};
39
ret[iadd16] = "iadd16";
40
ret[iadd32] = "iadd32";
41
ret[iadd64] = "iadd64";
43
ret[imul16] = "imul16";
44
ret[imul32] = "imul32";
45
ret[imul64] = "imul64";
46
ret[fadd16] = "fadd16";
47
ret[fadd32] = "fadd32";
48
ret[fadd64] = "fadd64";
49
ret[fmul16] = "fmul16";
50
ret[fmul32] = "fmul32";
51
ret[fmul64] = "fmul64";
53
ret[imin16] = "imin16";
54
ret[imin32] = "imin32";
55
ret[imin64] = "imin64";
57
ret[imax16] = "imax16";
58
ret[imax32] = "imax32";
59
ret[imax64] = "imax64";
61
ret[umin16] = "umin16";
62
ret[umin32] = "umin32";
63
ret[umin64] = "umin64";
65
ret[umax16] = "umax16";
66
ret[umax32] = "umax32";
67
ret[umax64] = "umax64";
68
ret[fmin16] = "fmin16";
69
ret[fmin32] = "fmin32";
70
ret[fmin64] = "fmin64";
71
ret[fmax16] = "fmax16";
72
ret[fmax32] = "fmax32";
73
ret[fmax64] = "fmax64";
75
ret[iand16] = "iand16";
76
ret[iand32] = "iand32";
77
ret[iand64] = "iand64";
83
ret[ixor16] = "ixor16";
84
ret[ixor32] = "ixor32";
85
ret[ixor64] = "ixor64";
90
print_reg_class(const RegClass rc, FILE* output)
92
if (rc.is_subdword()) {
93
fprintf(output, " v%ub: ", rc.bytes());
94
} else if (rc.type() == RegType::sgpr) {
95
fprintf(output, " s%u: ", rc.size());
96
} else if (rc.is_linear()) {
97
fprintf(output, " lv%u: ", rc.size());
99
fprintf(output, " v%u: ", rc.size());
104
print_physReg(PhysReg reg, unsigned bytes, FILE* output, unsigned flags)
107
fprintf(output, "m0");
108
} else if (reg == 106) {
109
fprintf(output, "vcc");
110
} else if (reg == 253) {
111
fprintf(output, "scc");
112
} else if (reg == 126) {
113
fprintf(output, "exec");
115
bool is_vgpr = reg / 256;
116
unsigned r = reg % 256;
117
unsigned size = DIV_ROUND_UP(bytes, 4);
118
if (size == 1 && (flags & print_no_ssa)) {
119
fprintf(output, "%c%d", is_vgpr ? 'v' : 's', r);
121
fprintf(output, "%c[%d", is_vgpr ? 'v' : 's', r);
123
fprintf(output, "-%d]", r + size - 1);
125
fprintf(output, "]");
127
if (reg.byte() || bytes % 4)
128
fprintf(output, "[%d:%d]", reg.byte() * 8, (reg.byte() + bytes) * 8);
133
print_constant(uint8_t reg, FILE* output)
135
if (reg >= 128 && reg <= 192) {
136
fprintf(output, "%d", reg - 128);
138
} else if (reg >= 192 && reg <= 208) {
139
fprintf(output, "%d", 192 - reg);
144
case 240: fprintf(output, "0.5"); break;
145
case 241: fprintf(output, "-0.5"); break;
146
case 242: fprintf(output, "1.0"); break;
147
case 243: fprintf(output, "-1.0"); break;
148
case 244: fprintf(output, "2.0"); break;
149
case 245: fprintf(output, "-2.0"); break;
150
case 246: fprintf(output, "4.0"); break;
151
case 247: fprintf(output, "-4.0"); break;
152
case 248: fprintf(output, "1/(2*PI)"); break;
157
aco_print_operand(const Operand* operand, FILE* output, unsigned flags)
159
if (operand->isLiteral() || (operand->isConstant() && operand->bytes() == 1)) {
160
if (operand->bytes() == 1)
161
fprintf(output, "0x%.2x", operand->constantValue());
162
else if (operand->bytes() == 2)
163
fprintf(output, "0x%.4x", operand->constantValue());
165
fprintf(output, "0x%x", operand->constantValue());
166
} else if (operand->isConstant()) {
167
print_constant(operand->physReg().reg(), output);
168
} else if (operand->isUndefined()) {
169
print_reg_class(operand->regClass(), output);
170
fprintf(output, "undef");
172
if (operand->isLateKill())
173
fprintf(output, "(latekill)");
174
if (operand->is16bit())
175
fprintf(output, "(is16bit)");
176
if (operand->is24bit())
177
fprintf(output, "(is24bit)");
178
if ((flags & print_kill) && operand->isKill())
179
fprintf(output, "(kill)");
181
if (!(flags & print_no_ssa))
182
fprintf(output, "%%%d%s", operand->tempId(), operand->isFixed() ? ":" : "");
184
if (operand->isFixed())
185
print_physReg(operand->physReg(), operand->bytes(), output, flags);
190
print_definition(const Definition* definition, FILE* output, unsigned flags)
192
if (!(flags & print_no_ssa))
193
print_reg_class(definition->regClass(), output);
194
if (definition->isPrecise())
195
fprintf(output, "(precise)");
196
if (definition->isNUW())
197
fprintf(output, "(nuw)");
198
if (definition->isNoCSE())
199
fprintf(output, "(noCSE)");
200
if ((flags & print_kill) && definition->isKill())
201
fprintf(output, "(kill)");
202
if (!(flags & print_no_ssa))
203
fprintf(output, "%%%d%s", definition->tempId(), definition->isFixed() ? ":" : "");
205
if (definition->isFixed())
206
print_physReg(definition->physReg(), definition->bytes(), output, flags);
210
print_storage(storage_class storage, FILE* output)
212
fprintf(output, " storage:");
214
if (storage & storage_buffer)
215
printed += fprintf(output, "%sbuffer", printed ? "," : "");
216
if (storage & storage_atomic_counter)
217
printed += fprintf(output, "%satomic_counter", printed ? "," : "");
218
if (storage & storage_image)
219
printed += fprintf(output, "%simage", printed ? "," : "");
220
if (storage & storage_shared)
221
printed += fprintf(output, "%sshared", printed ? "," : "");
222
if (storage & storage_task_payload)
223
printed += fprintf(output, "%stask_payload", printed ? "," : "");
224
if (storage & storage_vmem_output)
225
printed += fprintf(output, "%svmem_output", printed ? "," : "");
226
if (storage & storage_scratch)
227
printed += fprintf(output, "%sscratch", printed ? "," : "");
228
if (storage & storage_vgpr_spill)
229
printed += fprintf(output, "%svgpr_spill", printed ? "," : "");
233
print_semantics(memory_semantics sem, FILE* output)
235
fprintf(output, " semantics:");
237
if (sem & semantic_acquire)
238
printed += fprintf(output, "%sacquire", printed ? "," : "");
239
if (sem & semantic_release)
240
printed += fprintf(output, "%srelease", printed ? "," : "");
241
if (sem & semantic_volatile)
242
printed += fprintf(output, "%svolatile", printed ? "," : "");
243
if (sem & semantic_private)
244
printed += fprintf(output, "%sprivate", printed ? "," : "");
245
if (sem & semantic_can_reorder)
246
printed += fprintf(output, "%sreorder", printed ? "," : "");
247
if (sem & semantic_atomic)
248
printed += fprintf(output, "%satomic", printed ? "," : "");
249
if (sem & semantic_rmw)
250
printed += fprintf(output, "%srmw", printed ? "," : "");
254
print_scope(sync_scope scope, FILE* output, const char* prefix = "scope")
256
fprintf(output, " %s:", prefix);
258
case scope_invocation: fprintf(output, "invocation"); break;
259
case scope_subgroup: fprintf(output, "subgroup"); break;
260
case scope_workgroup: fprintf(output, "workgroup"); break;
261
case scope_queuefamily: fprintf(output, "queuefamily"); break;
262
case scope_device: fprintf(output, "device"); break;
267
print_sync(memory_sync_info sync, FILE* output)
269
print_storage(sync.storage, output);
270
print_semantics(sync.semantics, output);
271
print_scope(sync.scope, output);
275
print_instr_format_specific(const Instruction* instr, FILE* output)
277
switch (instr->format) {
279
const SOPK_instruction& sopk = instr->sopk();
280
fprintf(output, " imm:%d", sopk.imm & 0x8000 ? (sopk.imm - 65536) : sopk.imm);
284
uint16_t imm = instr->sopp().imm;
285
switch (instr->opcode) {
286
case aco_opcode::s_waitcnt: {
287
/* we usually should check the chip class for vmcnt/lgkm, but
288
* insert_waitcnt() should fill it in regardless. */
289
unsigned vmcnt = (imm & 0xF) | ((imm & (0x3 << 14)) >> 10);
291
fprintf(output, " vmcnt(%d)", vmcnt);
292
if (((imm >> 4) & 0x7) < 0x7)
293
fprintf(output, " expcnt(%d)", (imm >> 4) & 0x7);
294
if (((imm >> 8) & 0x3F) < 0x3F)
295
fprintf(output, " lgkmcnt(%d)", (imm >> 8) & 0x3F);
298
case aco_opcode::s_endpgm:
299
case aco_opcode::s_endpgm_saved:
300
case aco_opcode::s_endpgm_ordered_ps_done:
301
case aco_opcode::s_wakeup:
302
case aco_opcode::s_barrier:
303
case aco_opcode::s_icache_inv:
304
case aco_opcode::s_ttracedata:
305
case aco_opcode::s_set_gpr_idx_off: {
308
case aco_opcode::s_sendmsg: {
309
unsigned id = imm & sendmsg_id_mask;
311
case sendmsg_none: fprintf(output, " sendmsg(MSG_NONE)"); break;
313
fprintf(output, " sendmsg(gs%s%s, %u)", imm & 0x10 ? ", cut" : "",
314
imm & 0x20 ? ", emit" : "", imm >> 8);
316
case _sendmsg_gs_done:
317
fprintf(output, " sendmsg(gs_done%s%s, %u)", imm & 0x10 ? ", cut" : "",
318
imm & 0x20 ? ", emit" : "", imm >> 8);
320
case sendmsg_save_wave: fprintf(output, " sendmsg(save_wave)"); break;
321
case sendmsg_stall_wave_gen: fprintf(output, " sendmsg(stall_wave_gen)"); break;
322
case sendmsg_halt_waves: fprintf(output, " sendmsg(halt_waves)"); break;
323
case sendmsg_ordered_ps_done: fprintf(output, " sendmsg(ordered_ps_done)"); break;
324
case sendmsg_early_prim_dealloc: fprintf(output, " sendmsg(early_prim_dealloc)"); break;
325
case sendmsg_gs_alloc_req: fprintf(output, " sendmsg(gs_alloc_req)"); break;
331
fprintf(output, " imm:%u", imm);
335
if (instr->sopp().block != -1)
336
fprintf(output, " block:BB%d", instr->sopp().block);
340
const SMEM_instruction& smem = instr->smem();
342
fprintf(output, " glc");
344
fprintf(output, " dlc");
346
fprintf(output, " nv");
347
print_sync(smem.sync, output);
350
case Format::VINTRP: {
351
const Interp_instruction& vintrp = instr->vintrp();
352
fprintf(output, " attr%d.%c", vintrp.attribute, "xyzw"[vintrp.component]);
356
const DS_instruction& ds = instr->ds();
358
fprintf(output, " offset0:%u", ds.offset0);
360
fprintf(output, " offset1:%u", ds.offset1);
362
fprintf(output, " gds");
363
print_sync(ds.sync, output);
366
case Format::MUBUF: {
367
const MUBUF_instruction& mubuf = instr->mubuf();
369
fprintf(output, " offset:%u", mubuf.offset);
371
fprintf(output, " offen");
373
fprintf(output, " idxen");
375
fprintf(output, " addr64");
377
fprintf(output, " glc");
379
fprintf(output, " dlc");
381
fprintf(output, " slc");
383
fprintf(output, " tfe");
385
fprintf(output, " lds");
386
if (mubuf.disable_wqm)
387
fprintf(output, " disable_wqm");
388
print_sync(mubuf.sync, output);
392
const MIMG_instruction& mimg = instr->mimg();
393
unsigned identity_dmask =
394
!instr->definitions.empty() ? (1 << instr->definitions[0].size()) - 1 : 0xf;
395
if ((mimg.dmask & identity_dmask) != identity_dmask)
396
fprintf(output, " dmask:%s%s%s%s", mimg.dmask & 0x1 ? "x" : "",
397
mimg.dmask & 0x2 ? "y" : "", mimg.dmask & 0x4 ? "z" : "",
398
mimg.dmask & 0x8 ? "w" : "");
400
case ac_image_1d: fprintf(output, " 1d"); break;
401
case ac_image_2d: fprintf(output, " 2d"); break;
402
case ac_image_3d: fprintf(output, " 3d"); break;
403
case ac_image_cube: fprintf(output, " cube"); break;
404
case ac_image_1darray: fprintf(output, " 1darray"); break;
405
case ac_image_2darray: fprintf(output, " 2darray"); break;
406
case ac_image_2dmsaa: fprintf(output, " 2dmsaa"); break;
407
case ac_image_2darraymsaa: fprintf(output, " 2darraymsaa"); break;
410
fprintf(output, " unrm");
412
fprintf(output, " glc");
414
fprintf(output, " dlc");
416
fprintf(output, " slc");
418
fprintf(output, " tfe");
420
fprintf(output, " da");
422
fprintf(output, " lwe");
423
if (mimg.r128 || mimg.a16)
424
fprintf(output, " r128/a16");
426
fprintf(output, " d16");
427
if (mimg.disable_wqm)
428
fprintf(output, " disable_wqm");
429
print_sync(mimg.sync, output);
433
const Export_instruction& exp = instr->exp();
434
unsigned identity_mask = exp.compressed ? 0x5 : 0xf;
435
if ((exp.enabled_mask & identity_mask) != identity_mask)
436
fprintf(output, " en:%c%c%c%c", exp.enabled_mask & 0x1 ? 'r' : '*',
437
exp.enabled_mask & 0x2 ? 'g' : '*', exp.enabled_mask & 0x4 ? 'b' : '*',
438
exp.enabled_mask & 0x8 ? 'a' : '*');
440
fprintf(output, " compr");
442
fprintf(output, " done");
444
fprintf(output, " vm");
446
if (exp.dest <= V_008DFC_SQ_EXP_MRT + 7)
447
fprintf(output, " mrt%d", exp.dest - V_008DFC_SQ_EXP_MRT);
448
else if (exp.dest == V_008DFC_SQ_EXP_MRTZ)
449
fprintf(output, " mrtz");
450
else if (exp.dest == V_008DFC_SQ_EXP_NULL)
451
fprintf(output, " null");
452
else if (exp.dest >= V_008DFC_SQ_EXP_POS && exp.dest <= V_008DFC_SQ_EXP_POS + 3)
453
fprintf(output, " pos%d", exp.dest - V_008DFC_SQ_EXP_POS);
454
else if (exp.dest >= V_008DFC_SQ_EXP_PARAM && exp.dest <= V_008DFC_SQ_EXP_PARAM + 31)
455
fprintf(output, " param%d", exp.dest - V_008DFC_SQ_EXP_PARAM);
458
case Format::PSEUDO_BRANCH: {
459
const Pseudo_branch_instruction& branch = instr->branch();
460
/* Note: BB0 cannot be a branch target */
461
if (branch.target[0] != 0)
462
fprintf(output, " BB%d", branch.target[0]);
463
if (branch.target[1] != 0)
464
fprintf(output, ", BB%d", branch.target[1]);
467
case Format::PSEUDO_REDUCTION: {
468
const Pseudo_reduction_instruction& reduce = instr->reduction();
469
fprintf(output, " op:%s", reduce_ops[reduce.reduce_op]);
470
if (reduce.cluster_size)
471
fprintf(output, " cluster_size:%u", reduce.cluster_size);
474
case Format::PSEUDO_BARRIER: {
475
const Pseudo_barrier_instruction& barrier = instr->barrier();
476
print_sync(barrier.sync, output);
477
print_scope(barrier.exec_scope, output, "exec_scope");
482
case Format::SCRATCH: {
483
const FLAT_instruction& flat = instr->flatlike();
485
fprintf(output, " offset:%u", flat.offset);
487
fprintf(output, " glc");
489
fprintf(output, " dlc");
491
fprintf(output, " slc");
493
fprintf(output, " lds");
495
fprintf(output, " nv");
496
if (flat.disable_wqm)
497
fprintf(output, " disable_wqm");
498
print_sync(flat.sync, output);
501
case Format::MTBUF: {
502
const MTBUF_instruction& mtbuf = instr->mtbuf();
503
fprintf(output, " dfmt:");
504
switch (mtbuf.dfmt) {
505
case V_008F0C_BUF_DATA_FORMAT_8: fprintf(output, "8"); break;
506
case V_008F0C_BUF_DATA_FORMAT_16: fprintf(output, "16"); break;
507
case V_008F0C_BUF_DATA_FORMAT_8_8: fprintf(output, "8_8"); break;
508
case V_008F0C_BUF_DATA_FORMAT_32: fprintf(output, "32"); break;
509
case V_008F0C_BUF_DATA_FORMAT_16_16: fprintf(output, "16_16"); break;
510
case V_008F0C_BUF_DATA_FORMAT_10_11_11: fprintf(output, "10_11_11"); break;
511
case V_008F0C_BUF_DATA_FORMAT_11_11_10: fprintf(output, "11_11_10"); break;
512
case V_008F0C_BUF_DATA_FORMAT_10_10_10_2: fprintf(output, "10_10_10_2"); break;
513
case V_008F0C_BUF_DATA_FORMAT_2_10_10_10: fprintf(output, "2_10_10_10"); break;
514
case V_008F0C_BUF_DATA_FORMAT_8_8_8_8: fprintf(output, "8_8_8_8"); break;
515
case V_008F0C_BUF_DATA_FORMAT_32_32: fprintf(output, "32_32"); break;
516
case V_008F0C_BUF_DATA_FORMAT_16_16_16_16: fprintf(output, "16_16_16_16"); break;
517
case V_008F0C_BUF_DATA_FORMAT_32_32_32: fprintf(output, "32_32_32"); break;
518
case V_008F0C_BUF_DATA_FORMAT_32_32_32_32: fprintf(output, "32_32_32_32"); break;
519
case V_008F0C_BUF_DATA_FORMAT_RESERVED_15: fprintf(output, "reserved15"); break;
521
fprintf(output, " nfmt:");
522
switch (mtbuf.nfmt) {
523
case V_008F0C_BUF_NUM_FORMAT_UNORM: fprintf(output, "unorm"); break;
524
case V_008F0C_BUF_NUM_FORMAT_SNORM: fprintf(output, "snorm"); break;
525
case V_008F0C_BUF_NUM_FORMAT_USCALED: fprintf(output, "uscaled"); break;
526
case V_008F0C_BUF_NUM_FORMAT_SSCALED: fprintf(output, "sscaled"); break;
527
case V_008F0C_BUF_NUM_FORMAT_UINT: fprintf(output, "uint"); break;
528
case V_008F0C_BUF_NUM_FORMAT_SINT: fprintf(output, "sint"); break;
529
case V_008F0C_BUF_NUM_FORMAT_SNORM_OGL: fprintf(output, "snorm"); break;
530
case V_008F0C_BUF_NUM_FORMAT_FLOAT: fprintf(output, "float"); break;
533
fprintf(output, " offset:%u", mtbuf.offset);
535
fprintf(output, " offen");
537
fprintf(output, " idxen");
539
fprintf(output, " glc");
541
fprintf(output, " dlc");
543
fprintf(output, " slc");
545
fprintf(output, " tfe");
546
if (mtbuf.disable_wqm)
547
fprintf(output, " disable_wqm");
548
print_sync(mtbuf.sync, output);
551
case Format::VOP3P: {
552
if (instr->vop3p().clamp)
553
fprintf(output, " clamp");
560
if (instr->isVOP3()) {
561
const VOP3_instruction& vop3 = instr->vop3();
563
case 1: fprintf(output, " *2"); break;
564
case 2: fprintf(output, " *4"); break;
565
case 3: fprintf(output, " *0.5"); break;
568
fprintf(output, " clamp");
569
if (vop3.opsel & (1 << 3))
570
fprintf(output, " opsel_hi");
571
} else if (instr->isDPP16()) {
572
const DPP16_instruction& dpp = instr->dpp16();
573
if (dpp.dpp_ctrl <= 0xff) {
574
fprintf(output, " quad_perm:[%d,%d,%d,%d]", dpp.dpp_ctrl & 0x3, (dpp.dpp_ctrl >> 2) & 0x3,
575
(dpp.dpp_ctrl >> 4) & 0x3, (dpp.dpp_ctrl >> 6) & 0x3);
576
} else if (dpp.dpp_ctrl >= 0x101 && dpp.dpp_ctrl <= 0x10f) {
577
fprintf(output, " row_shl:%d", dpp.dpp_ctrl & 0xf);
578
} else if (dpp.dpp_ctrl >= 0x111 && dpp.dpp_ctrl <= 0x11f) {
579
fprintf(output, " row_shr:%d", dpp.dpp_ctrl & 0xf);
580
} else if (dpp.dpp_ctrl >= 0x121 && dpp.dpp_ctrl <= 0x12f) {
581
fprintf(output, " row_ror:%d", dpp.dpp_ctrl & 0xf);
582
} else if (dpp.dpp_ctrl == dpp_wf_sl1) {
583
fprintf(output, " wave_shl:1");
584
} else if (dpp.dpp_ctrl == dpp_wf_rl1) {
585
fprintf(output, " wave_rol:1");
586
} else if (dpp.dpp_ctrl == dpp_wf_sr1) {
587
fprintf(output, " wave_shr:1");
588
} else if (dpp.dpp_ctrl == dpp_wf_rr1) {
589
fprintf(output, " wave_ror:1");
590
} else if (dpp.dpp_ctrl == dpp_row_mirror) {
591
fprintf(output, " row_mirror");
592
} else if (dpp.dpp_ctrl == dpp_row_half_mirror) {
593
fprintf(output, " row_half_mirror");
594
} else if (dpp.dpp_ctrl == dpp_row_bcast15) {
595
fprintf(output, " row_bcast:15");
596
} else if (dpp.dpp_ctrl == dpp_row_bcast31) {
597
fprintf(output, " row_bcast:31");
599
fprintf(output, " dpp_ctrl:0x%.3x", dpp.dpp_ctrl);
601
if (dpp.row_mask != 0xf)
602
fprintf(output, " row_mask:0x%.1x", dpp.row_mask);
603
if (dpp.bank_mask != 0xf)
604
fprintf(output, " bank_mask:0x%.1x", dpp.bank_mask);
606
fprintf(output, " bound_ctrl:1");
607
} else if (instr->isDPP8()) {
608
const DPP8_instruction& dpp = instr->dpp8();
609
fprintf(output, " dpp8:[%d,%d,%d,%d,%d,%d,%d,%d]", dpp.lane_sel[0], dpp.lane_sel[1],
610
dpp.lane_sel[2], dpp.lane_sel[3], dpp.lane_sel[4], dpp.lane_sel[5], dpp.lane_sel[6],
612
} else if (instr->isSDWA()) {
613
const SDWA_instruction& sdwa = instr->sdwa();
615
case 1: fprintf(output, " *2"); break;
616
case 2: fprintf(output, " *4"); break;
617
case 3: fprintf(output, " *0.5"); break;
620
fprintf(output, " clamp");
621
if (!instr->isVOPC()) {
622
char sext = sdwa.dst_sel.sign_extend() ? 's' : 'u';
623
unsigned offset = sdwa.dst_sel.offset();
624
if (instr->definitions[0].isFixed())
625
offset += instr->definitions[0].physReg().byte();
626
switch (sdwa.dst_sel.size()) {
627
case 1: fprintf(output, " dst_sel:%cbyte%u", sext, offset); break;
628
case 2: fprintf(output, " dst_sel:%cword%u", sext, offset >> 1); break;
629
case 4: fprintf(output, " dst_sel:dword"); break;
632
if (instr->definitions[0].bytes() < 4)
633
fprintf(output, " dst_preserve");
635
for (unsigned i = 0; i < std::min<unsigned>(2, instr->operands.size()); i++) {
636
char sext = sdwa.sel[i].sign_extend() ? 's' : 'u';
637
unsigned offset = sdwa.sel[i].offset();
638
if (instr->operands[i].isFixed())
639
offset += instr->operands[i].physReg().byte();
640
switch (sdwa.sel[i].size()) {
641
case 1: fprintf(output, " src%d_sel:%cbyte%u", i, sext, offset); break;
642
case 2: fprintf(output, " src%d_sel:%cword%u", i, sext, offset >> 1); break;
643
case 4: fprintf(output, " src%d_sel:dword", i); break;
651
aco_print_instr(const Instruction* instr, FILE* output, unsigned flags)
653
if (!instr->definitions.empty()) {
654
for (unsigned i = 0; i < instr->definitions.size(); ++i) {
655
print_definition(&instr->definitions[i], output, flags);
656
if (i + 1 != instr->definitions.size())
657
fprintf(output, ", ");
659
fprintf(output, " = ");
661
fprintf(output, "%s", instr_info.name[(int)instr->opcode]);
662
if (instr->operands.size()) {
663
const unsigned num_operands = instr->operands.size();
664
bool* const abs = (bool*)alloca(num_operands * sizeof(bool));
665
bool* const neg = (bool*)alloca(num_operands * sizeof(bool));
666
bool* const opsel = (bool*)alloca(num_operands * sizeof(bool));
667
bool* const f2f32 = (bool*)alloca(num_operands * sizeof(bool));
668
for (unsigned i = 0; i < num_operands; ++i) {
674
bool is_mad_mix = instr->opcode == aco_opcode::v_fma_mix_f32 ||
675
instr->opcode == aco_opcode::v_fma_mixlo_f16 ||
676
instr->opcode == aco_opcode::v_fma_mixhi_f16;
677
if (instr->isVOP3()) {
678
const VOP3_instruction& vop3 = instr->vop3();
679
for (unsigned i = 0; i < MIN2(num_operands, 3); ++i) {
680
abs[i] = vop3.abs[i];
681
neg[i] = vop3.neg[i];
682
opsel[i] = vop3.opsel & (1 << i);
684
} else if (instr->isDPP16()) {
685
const DPP16_instruction& dpp = instr->dpp16();
686
for (unsigned i = 0; i < MIN2(num_operands, 2); ++i) {
691
} else if (instr->isSDWA()) {
692
const SDWA_instruction& sdwa = instr->sdwa();
693
for (unsigned i = 0; i < MIN2(num_operands, 2); ++i) {
694
abs[i] = sdwa.abs[i];
695
neg[i] = sdwa.neg[i];
698
} else if (instr->isVOP3P() && is_mad_mix) {
699
const VOP3P_instruction& vop3p = instr->vop3p();
700
for (unsigned i = 0; i < MIN2(num_operands, 3); ++i) {
701
abs[i] = vop3p.neg_hi[i];
702
neg[i] = vop3p.neg_lo[i];
703
f2f32[i] = vop3p.opsel_hi & (1 << i);
704
opsel[i] = f2f32[i] && (vop3p.opsel_lo & (1 << i));
707
for (unsigned i = 0; i < num_operands; ++i) {
709
fprintf(output, ", ");
711
fprintf(output, " ");
714
fprintf(output, "-");
716
fprintf(output, "|");
718
fprintf(output, "hi(");
720
fprintf(output, "lo(");
721
aco_print_operand(&instr->operands[i], output, flags);
722
if (f2f32[i] || opsel[i])
723
fprintf(output, ")");
725
fprintf(output, "|");
727
if (instr->isVOP3P() && !is_mad_mix) {
728
const VOP3P_instruction& vop3 = instr->vop3p();
729
if ((vop3.opsel_lo & (1 << i)) || !(vop3.opsel_hi & (1 << i))) {
730
fprintf(output, ".%c%c", vop3.opsel_lo & (1 << i) ? 'y' : 'x',
731
vop3.opsel_hi & (1 << i) ? 'y' : 'x');
733
if (vop3.neg_lo[i] && vop3.neg_hi[i])
734
fprintf(output, "*[-1,-1]");
735
else if (vop3.neg_lo[i])
736
fprintf(output, "*[-1,1]");
737
else if (vop3.neg_hi[i])
738
fprintf(output, "*[1,-1]");
742
print_instr_format_specific(instr, output);
746
print_block_kind(uint16_t kind, FILE* output)
748
if (kind & block_kind_uniform)
749
fprintf(output, "uniform, ");
750
if (kind & block_kind_top_level)
751
fprintf(output, "top-level, ");
752
if (kind & block_kind_loop_preheader)
753
fprintf(output, "loop-preheader, ");
754
if (kind & block_kind_loop_header)
755
fprintf(output, "loop-header, ");
756
if (kind & block_kind_loop_exit)
757
fprintf(output, "loop-exit, ");
758
if (kind & block_kind_continue)
759
fprintf(output, "continue, ");
760
if (kind & block_kind_break)
761
fprintf(output, "break, ");
762
if (kind & block_kind_continue_or_break)
763
fprintf(output, "continue_or_break, ");
764
if (kind & block_kind_branch)
765
fprintf(output, "branch, ");
766
if (kind & block_kind_merge)
767
fprintf(output, "merge, ");
768
if (kind & block_kind_invert)
769
fprintf(output, "invert, ");
770
if (kind & block_kind_uses_discard)
771
fprintf(output, "discard, ");
772
if (kind & block_kind_needs_lowering)
773
fprintf(output, "needs_lowering, ");
774
if (kind & block_kind_export_end)
775
fprintf(output, "export_end, ");
779
print_stage(Stage stage, FILE* output)
781
fprintf(output, "ACO shader stage: ");
783
if (stage == compute_cs)
784
fprintf(output, "compute_cs");
785
else if (stage == fragment_fs)
786
fprintf(output, "fragment_fs");
787
else if (stage == gs_copy_vs)
788
fprintf(output, "gs_copy_vs");
789
else if (stage == vertex_ls)
790
fprintf(output, "vertex_ls");
791
else if (stage == vertex_es)
792
fprintf(output, "vertex_es");
793
else if (stage == vertex_vs)
794
fprintf(output, "vertex_vs");
795
else if (stage == tess_control_hs)
796
fprintf(output, "tess_control_hs");
797
else if (stage == vertex_tess_control_hs)
798
fprintf(output, "vertex_tess_control_hs");
799
else if (stage == tess_eval_es)
800
fprintf(output, "tess_eval_es");
801
else if (stage == tess_eval_vs)
802
fprintf(output, "tess_eval_vs");
803
else if (stage == geometry_gs)
804
fprintf(output, "geometry_gs");
805
else if (stage == vertex_geometry_gs)
806
fprintf(output, "vertex_geometry_gs");
807
else if (stage == tess_eval_geometry_gs)
808
fprintf(output, "tess_eval_geometry_gs");
809
else if (stage == vertex_ngg)
810
fprintf(output, "vertex_ngg");
811
else if (stage == tess_eval_ngg)
812
fprintf(output, "tess_eval_ngg");
813
else if (stage == vertex_geometry_ngg)
814
fprintf(output, "vertex_geometry_ngg");
815
else if (stage == tess_eval_geometry_ngg)
816
fprintf(output, "tess_eval_geometry_ngg");
817
else if (stage == mesh_ngg)
818
fprintf(output, "mesh_ngg");
819
else if (stage == task_cs)
820
fprintf(output, "task_cs");
822
fprintf(output, "unknown");
824
fprintf(output, "\n");
828
aco_print_block(const Block* block, FILE* output, unsigned flags, const live& live_vars)
830
fprintf(output, "BB%d\n", block->index);
831
fprintf(output, "/* logical preds: ");
832
for (unsigned pred : block->logical_preds)
833
fprintf(output, "BB%d, ", pred);
834
fprintf(output, "/ linear preds: ");
835
for (unsigned pred : block->linear_preds)
836
fprintf(output, "BB%d, ", pred);
837
fprintf(output, "/ kind: ");
838
print_block_kind(block->kind, output);
839
fprintf(output, "*/\n");
841
if (flags & print_live_vars) {
842
fprintf(output, "\tlive out:");
843
for (unsigned id : live_vars.live_out[block->index])
844
fprintf(output, " %%%d", id);
845
fprintf(output, "\n");
847
RegisterDemand demand = block->register_demand;
848
fprintf(output, "\tdemand: %u vgpr, %u sgpr\n", demand.vgpr, demand.sgpr);
852
for (auto const& instr : block->instructions) {
853
fprintf(output, "\t");
854
if (flags & print_live_vars) {
855
RegisterDemand demand = live_vars.register_demand[block->index][index];
856
fprintf(output, "(%3u vgpr, %3u sgpr) ", demand.vgpr, demand.sgpr);
858
if (flags & print_perf_info)
859
fprintf(output, "(%3u clk) ", instr->pass_flags);
861
aco_print_instr(instr.get(), output, flags);
862
fprintf(output, "\n");
868
aco_print_program(const Program* program, FILE* output, const live& live_vars, unsigned flags)
870
switch (program->progress) {
871
case CompilationProgress::after_isel: fprintf(output, "After Instruction Selection:\n"); break;
872
case CompilationProgress::after_spilling:
873
fprintf(output, "After Spilling:\n");
876
case CompilationProgress::after_ra: fprintf(output, "After RA:\n"); break;
879
print_stage(program->stage, output);
881
for (Block const& block : program->blocks)
882
aco_print_block(&block, output, flags, live_vars);
884
if (program->constant_data.size()) {
885
fprintf(output, "\n/* constant data */\n");
886
for (unsigned i = 0; i < program->constant_data.size(); i += 32) {
887
fprintf(output, "[%06d] ", i);
888
unsigned line_size = std::min<size_t>(program->constant_data.size() - i, 32);
889
for (unsigned j = 0; j < line_size; j += 4) {
890
unsigned size = std::min<size_t>(program->constant_data.size() - (i + j), 4);
892
memcpy(&v, &program->constant_data[i + j], size);
893
fprintf(output, " %08x", v);
895
fprintf(output, "\n");
899
fprintf(output, "\n");
903
aco_print_program(const Program* program, FILE* output, unsigned flags)
905
aco_print_program(program, output, live(), flags);