-
Notifications
You must be signed in to change notification settings - Fork 3.4k
/
codegen_cpu.cc
1542 lines (1425 loc) · 65.4 KB
/
codegen_cpu.cc
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/*
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing,
* software distributed under the License is distributed on an
* "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
* KIND, either express or implied. See the License for the
* specific language governing permissions and limitations
* under the License.
*/
/*!
* \file codegen_cpu.cc
*/
#ifdef TVM_LLVM_VERSION
#include "codegen_cpu.h"
#include <llvm/ADT/SmallVector.h>
#include <llvm/ADT/StringRef.h>
#include <llvm/IR/Argument.h>
#include <llvm/IR/Attributes.h>
#include <llvm/IR/BasicBlock.h>
#include <llvm/IR/CallingConv.h>
#include <llvm/IR/Comdat.h>
#include <llvm/IR/Constants.h>
#include <llvm/IR/DIBuilder.h>
#include <llvm/IR/DebugInfoMetadata.h>
#include <llvm/IR/DebugLoc.h>
#include <llvm/IR/DerivedTypes.h>
#include <llvm/IR/Function.h>
#include <llvm/IR/GlobalVariable.h>
#include <llvm/IR/Instructions.h>
#include <llvm/IR/LLVMContext.h>
#include <llvm/IR/MDBuilder.h>
#include <llvm/IR/Metadata.h>
#include <llvm/IR/Module.h>
#if TVM_LLVM_VERSION >= 100
#include <llvm/Support/Alignment.h>
#endif
#include <llvm/Support/raw_ostream.h>
#include <llvm/Target/TargetMachine.h>
#include <llvm/Transforms/Utils/ModuleUtils.h>
#include <tvm/runtime/c_runtime_api.h>
#include <tvm/runtime/module.h>
#include <tvm/tir/analysis.h>
#include <algorithm>
#include <memory>
#include <tuple>
#include <unordered_map>
#include <unordered_set>
#include "../func_registry_generator.h"
#include "../metadata_utils.h"
#include "llvm_instance.h"
namespace tvm {
namespace codegen {
// Make these non-inline because of std::unique_ptr. See comment in
// codegen_llvm.cc for more information.
CodeGenCPU::CodeGenCPU() = default;
CodeGenCPU::~CodeGenCPU() = default;
void CodeGenCPU::Init(const std::string& module_name, LLVMTarget* llvm_target,
Optional<String> system_lib_prefix, bool dynamic_lookup,
bool target_c_runtime) {
CodeGenLLVM::Init(module_name, llvm_target, system_lib_prefix, dynamic_lookup, target_c_runtime);
system_lib_prefix_ = system_lib_prefix;
dbg_info_ = CreateDebugInfo(module_.get());
static_assert(sizeof(TVMValue) == sizeof(double), "invariant");
func_handle_map_.clear();
export_system_symbols_.clear();
// Runtime types.
t_tvm_shape_index_ =
llvm::Type::getIntNTy(*llvm_target_->GetContext(), DataType::ShapeIndex().bits());
// Defined in 3rdparty/dlpack/include/dlpack/dlpack.h:
// typedef struct { DLDeviceType device_type; int device_id; } DLDevice;
t_tvm_device_ = llvm::StructType::create({t_int_, t_int_});
// Defined in 3rdparty/dlpack/include/dlpack/dlpack.h:
// typedef struct { uint8_t code; uint8_t bits; uint16_t lanes; } DLDataType;
t_tvm_type_ = llvm::StructType::create({t_int8_, t_int8_, t_int16_});
// Defined in include/tvm/runtime/c_runtime_api.h:
// typedef void* TVMFunctionHandle;
t_tvm_func_handle_ = t_void_p_;
// Defined in 3rdparty/dlpack/include/dlpack/dlpack.h:
// typedef struct { ... } DLTensor;
t_tvm_array_ = llvm::StructType::create({t_void_p_, t_tvm_device_, t_int_, t_tvm_type_,
t_tvm_shape_index_->getPointerTo(),
t_tvm_shape_index_->getPointerTo(), t_int64_});
// Defined in include/tvm/runtime/c_runtime_api.h:
// typedef union { ... } TVMValue;
t_tvm_value_ = llvm::StructType::create({t_float64_});
// Defined in include/tvm/runtime/c_backend_api.h:
// typedef struct { void* sync_handle; int32_t num_task; } TVMParallelGroupEnv;
t_tvm_parallel_group_env_ = llvm::StructType::create({t_int32_->getPointerTo(), t_int32_});
// Defined in include/tvm/runtime/c_backend_api.h:
// typedef int (*TVMBackendPackedCFunc)(TVMValue* args, int* type_codes, int num_args,
// TVMValue* out_ret_value, int* out_ret_tcode,
// void* resource_handle);
ftype_tvm_backend_packed_c_func_ = llvm::FunctionType::get(
t_int_,
{t_void_p_, t_int_->getPointerTo(), t_int_, t_void_p_, t_int_->getPointerTo(), t_void_p_},
false);
t_tvm_crt_func_registry_ = llvm::StructType::create(
{t_char_->getPointerTo(), ftype_tvm_backend_packed_c_func_->getPointerTo()});
t_tvm_crt_module_ = llvm::StructType::create({t_tvm_crt_func_registry_->getPointerTo()});
// Defined in include/tvm/runtime/c_backend_api.h:
// typedef int (*FTVMParallelLambda)(int task_id, TVMParallelGroupEnv* penv, void* cdata);
ftype_tvm_parallel_lambda_ = llvm::FunctionType::get(
t_int_, {t_int_, t_tvm_parallel_group_env_->getPointerTo(), t_void_p_}, false);
md_tbaa_ctx_ptr_ = md_builder_->createTBAAScalarTypeNode("ctx_ptr", md_tbaa_root_);
// Runtime functions.
// Defined in include/tvm/runtime/c_runtime_api.h:
// int TVMFuncCall(TVMFunctionHandle func, TVMValue* arg_values, int* type_codes, int num_args,
// TVMValue* ret_val, int* ret_type_code);
ftype_tvm_func_call_ = llvm::FunctionType::get(
t_int_,
{t_tvm_func_handle_, t_tvm_value_->getPointerTo(), t_int_->getPointerTo(), t_int_,
t_tvm_value_->getPointerTo(), t_int_->getPointerTo()},
false);
// Defined in include/tvm/runtime/c_backend_api.h:
// int TVMBackendGetFuncFromEnv(void* mod_node, const char* func_name, TVMFunctionHandle* out);
ftype_tvm_get_func_from_env_ = llvm::FunctionType::get(
t_int_, {t_void_p_, t_char_->getPointerTo(), t_tvm_func_handle_->getPointerTo()}, false);
// Defined in include/tvm/runtime/c_runtime_api.h:
// void TVMAPISetLastError(const char* msg);
ftype_tvm_api_set_last_error_ =
llvm::FunctionType::get(t_void_, {t_char_->getPointerTo()}, false);
// Defined in include/tvm/runtime/c_backend_api.h:
// int TVMBackendParallelLaunch(FTVMParallelLambda flambda, void* cdata, int num_task);
ftype_tvm_parallel_launch_ = llvm::FunctionType::get(
t_int_, {ftype_tvm_parallel_lambda_->getPointerTo(), t_void_p_, t_int_}, false);
// Defined in include/tvm/runtime/c_backend_api.h:
// int TVMBackendParallelBarrier(int task_id, TVMParallelGroupEnv* penv);
ftype_tvm_parallel_barrier_ =
llvm::FunctionType::get(t_int_, {t_int_, t_tvm_parallel_group_env_->getPointerTo()}, false);
ftype_tvm_static_init_callback_ = llvm::FunctionType::get(t_int_, {t_void_p_}, false);
ftype_tvm_static_init_ =
llvm::FunctionType::get(t_int_,
{t_void_p_->getPointerTo(),
ftype_tvm_static_init_callback_->getPointerTo(), t_void_p_, t_int_},
false);
// initialize TVM runtime API
if (system_lib_prefix_.defined() && !target_c_runtime) {
// We will need this in environment for backward registration.
// Defined in include/tvm/runtime/c_backend_api.h:
// int TVMBackendRegisterSystemLibSymbol(const char* name, void* ptr);
f_tvm_register_system_symbol_ = llvm::Function::Create(
llvm::FunctionType::get(t_int_, {t_char_->getPointerTo(), t_void_p_}, false),
llvm::Function::ExternalLinkage, "TVMBackendRegisterSystemLibSymbol", module_.get());
} else {
f_tvm_register_system_symbol_ = nullptr;
}
if (dynamic_lookup || system_lib_prefix_.defined()) {
f_tvm_func_call_ = llvm::Function::Create(ftype_tvm_func_call_, llvm::Function::ExternalLinkage,
"TVMFuncCall", module_.get());
f_tvm_get_func_from_env_ =
llvm::Function::Create(ftype_tvm_get_func_from_env_, llvm::Function::ExternalLinkage,
"TVMBackendGetFuncFromEnv", module_.get());
f_tvm_api_set_last_error_ =
llvm::Function::Create(ftype_tvm_api_set_last_error_, llvm::Function::ExternalLinkage,
"TVMAPISetLastError", module_.get());
f_tvm_parallel_launch_ =
llvm::Function::Create(ftype_tvm_parallel_launch_, llvm::Function::ExternalLinkage,
"TVMBackendParallelLaunch", module_.get());
f_tvm_parallel_barrier_ =
llvm::Function::Create(ftype_tvm_parallel_barrier_, llvm::Function::ExternalLinkage,
"TVMBackendParallelBarrier", module_.get());
}
target_c_runtime_ = target_c_runtime;
InitGlobalContext(dynamic_lookup);
}
llvm::DISubprogram* CodeGenCPU::CreateDebugFunction(llvm::StringRef name,
const Array<Type>& param_types,
const Type& return_type) {
#if TVM_LLVM_VERSION < 50
return nullptr;
#else
llvm::SmallVector<llvm::Metadata*, 4> paramTys;
paramTys.push_back(GetDebugType(return_type));
for (const auto& param_type : param_types) {
paramTys.push_back(GetDebugType(param_type));
}
auto* DIFunctionTy = dbg_info_->di_builder_->createSubroutineType(
dbg_info_->di_builder_->getOrCreateTypeArray(paramTys));
bool local_to_unit = llvm::GlobalVariable::isLocalLinkage(llvm::GlobalValue::InternalLinkage);
#if TVM_LLVM_VERSION >= 80
auto SPFlags = llvm::DISubprogram::toSPFlags(local_to_unit, /*IsDefinition=*/true,
/*IsOptimized=*/true);
#else
bool SPFlags = /*IsOptimized=*/true;
#endif
auto* DIFunction = dbg_info_->di_builder_->createFunction(
/*Scope=*/dbg_info_->file_, /*Name=*/name, /*LinkageName=*/"",
/*File=*/dbg_info_->file_, /*LineNo=*/0, /*Ty=*/DIFunctionTy,
/*ScopeLine=*/0, /*Flags=*/llvm::DINode::FlagPrototyped, /*SPFlags=*/SPFlags);
return DIFunction;
#endif
}
llvm::DISubprogram* CodeGenCPU::CreateDebugFunction(const GlobalVar& gvar, const PrimFunc& func) {
std::string name = func->GetAttr<String>(tvm::attr::kGlobalSymbol).value_or(gvar->name_hint);
return CreateDebugFunction(name, func->params.Map(GetType), func->ret_type);
}
void CodeGenCPU::AddFunction(const GlobalVar& gvar, const PrimFunc& func) {
di_subprogram_ = CreateDebugFunction(gvar, func);
EmitDebugLocation(func->span);
CodeGenLLVM::AddFunction(gvar, func);
if (f_tvm_register_system_symbol_ != nullptr) {
if (auto global_symbol = func->GetAttr<String>(tvm::attr::kGlobalSymbol)) {
export_system_symbols_.emplace_back(
std::make_pair(global_symbol.value().operator std::string(), function_));
}
}
AddDebugInformation(function_, func->params.Map(GetType));
}
void CodeGenCPU::AddMainFunction(const std::string& entry_func_name) {
llvm::Function* f = module_->getFunction(entry_func_name);
ICHECK(f) << "Function " << entry_func_name << "does not in module";
llvm::Type* type = llvm::ArrayType::get(t_char_, entry_func_name.length() + 1);
llvm::GlobalVariable* global =
new llvm::GlobalVariable(*module_, type, true, llvm::GlobalValue::WeakAnyLinkage, nullptr,
runtime::symbol::tvm_module_main);
#if TVM_LLVM_VERSION >= 100
global->setAlignment(llvm::Align(1));
#else
global->setAlignment(1);
#endif
// comdat is needed for windows select any linking to work
// set comdat to Any(weak linking)
if (llvm_target_->GetOrCreateTargetMachine()->getTargetTriple().isOSWindows()) {
llvm::Comdat* comdat = module_->getOrInsertComdat(runtime::symbol::tvm_module_main);
comdat->setSelectionKind(llvm::Comdat::Any);
global->setComdat(comdat);
}
global->setInitializer(
llvm::ConstantDataArray::getString(*llvm_target_->GetContext(), entry_func_name));
global->setDLLStorageClass(llvm::GlobalVariable::DLLExportStorageClass);
}
std::unique_ptr<llvm::Module> CodeGenCPU::Finish() {
// link modules
if (dbg_info_ != nullptr) {
dbg_info_->di_builder_->finalize();
}
return CodeGenLLVM::Finish();
}
CodeGenLLVM::TypedPointer CodeGenCPU::CreateStructRefPtr(DataType t, llvm::Value* buf,
llvm::Value* index, int kind) {
if (kind < builtin::kArrKindBound_) {
if (buf->getType() == t_void_p_) {
buf = builder_->CreatePointerCast(buf, t_tvm_array_->getPointerTo());
} else {
ICHECK_EQ(buf->getType(), t_tvm_array_->getPointerTo());
}
}
switch (kind) {
case builtin::kArrAddr: {
return TypedPointer(t_tvm_array_, builder_->CreateInBoundsGEP(t_tvm_array_, buf, index));
}
case builtin::kArrData: {
llvm::Type* member_type = t_tvm_array_->getStructElementType(0);
llvm::Value* member_addr =
builder_->CreateInBoundsGEP(t_tvm_array_, buf, {index, ConstInt32(0)});
return TypedPointer(member_type, member_addr);
}
case builtin::kArrShape: {
llvm::Type* member_type = t_tvm_array_->getStructElementType(4);
llvm::Value* member_addr =
builder_->CreateInBoundsGEP(t_tvm_array_, buf, {index, ConstInt32(4)});
return TypedPointer(member_type, member_addr);
}
case builtin::kArrStrides: {
llvm::Type* member_type = t_tvm_array_->getStructElementType(5);
llvm::Value* member_addr =
builder_->CreateInBoundsGEP(t_tvm_array_, buf, {index, ConstInt32(5)});
return TypedPointer(member_type, member_addr);
}
case builtin::kArrNDim: {
llvm::Type* member_type = t_tvm_array_->getStructElementType(2);
llvm::Value* member_addr =
builder_->CreateInBoundsGEP(t_tvm_array_, buf, {index, ConstInt32(2)});
return TypedPointer(member_type, member_addr);
}
case builtin::kArrTypeCode: {
llvm::Type* member_type = t_tvm_array_->getStructElementType(3)->getStructElementType(0);
llvm::Value* member_addr =
builder_->CreateInBoundsGEP(t_tvm_array_, buf, {index, ConstInt32(3), ConstInt32(0)});
return TypedPointer(member_type, member_addr);
}
case builtin::kArrTypeBits: {
llvm::Type* member_type = t_tvm_array_->getStructElementType(3)->getStructElementType(1);
llvm::Value* member_addr =
builder_->CreateInBoundsGEP(t_tvm_array_, buf, {index, ConstInt32(3), ConstInt32(1)});
return TypedPointer(member_type, member_addr);
}
case builtin::kArrTypeLanes: {
llvm::Type* member_type = t_tvm_array_->getStructElementType(3)->getStructElementType(2);
llvm::Value* member_addr =
builder_->CreateInBoundsGEP(t_tvm_array_, buf, {index, ConstInt32(3), ConstInt32(2)});
return TypedPointer(member_type, member_addr);
}
case builtin::kArrByteOffset: {
llvm::Type* member_type = t_tvm_array_->getStructElementType(6);
llvm::Value* member_addr =
builder_->CreateInBoundsGEP(t_tvm_array_, buf, {index, ConstInt32(6)});
return TypedPointer(member_type, member_addr);
}
case builtin::kArrDeviceId: {
llvm::Type* member_type = t_tvm_array_->getStructElementType(1)->getStructElementType(1);
llvm::Value* member_addr =
builder_->CreateInBoundsGEP(t_tvm_array_, buf, {index, ConstInt32(1), ConstInt32(1)});
return TypedPointer(member_type, member_addr);
}
case builtin::kArrDeviceType: {
llvm::Type* member_type = t_tvm_array_->getStructElementType(1)->getStructElementType(0);
llvm::Value* member_addr =
builder_->CreateInBoundsGEP(t_tvm_array_, buf, {index, ConstInt32(1), ConstInt32(0)});
return TypedPointer(member_type, member_addr);
}
case builtin::kTVMValueContent: {
ICHECK_EQ(t.lanes(), 1);
if (t.is_bool()) {
// The stride between adjacent entries is still
// `sizeof(TVMValue)==64`, even if the enum currently holds a
// boolean.
buf = builder_->CreatePointerCast(buf, t_int64_->getPointerTo());
buf = builder_->CreateInBoundsGEP(t_int64_, buf, index);
buf = builder_->CreatePointerCast(buf, DTypeToLLVMType(t)->getPointerTo());
return TypedPointer(t_int8_, buf);
} else if (t.is_int() && t.bits() == 64) {
buf = builder_->CreatePointerCast(buf, t_int64_->getPointerTo());
return TypedPointer(t_int64_, builder_->CreateInBoundsGEP(t_int64_, buf, index));
} else if (t.is_float() && t.bits() == 64) {
buf = builder_->CreatePointerCast(buf, t_float64_->getPointerTo());
return TypedPointer(t_float64_, builder_->CreateInBoundsGEP(t_float64_, buf, index));
} else if (t.is_handle()) {
buf = builder_->CreatePointerCast(buf, t_tvm_value_->getPointerTo());
buf = builder_->CreateInBoundsGEP(t_tvm_value_, buf, index);
return TypedPointer(t_void_p_, builder_->CreatePointerCast(buf, t_void_p_->getPointerTo()));
} else {
LOG(DEBUG) << "DataType " << t << " cannot be stored into a TVMValue";
}
}
default:
LOG(FATAL) << "unknown field code";
}
}
llvm::Value* CodeGenCPU::CreateCallExtern(Type ret_type, String global_symbol,
const Array<PrimExpr>& args, bool skip_first_arg) {
std::vector<llvm::Value*> arg_values;
for (size_t i = static_cast<size_t>(skip_first_arg); i < args.size(); ++i) {
arg_values.push_back(MakeValue(args[i]));
}
std::vector<llvm::Type*> arg_types;
for (llvm::Value* v : arg_values) {
arg_types.push_back(v->getType());
}
llvm::FunctionType* ftype = llvm::FunctionType::get(GetLLVMType(ret_type), arg_types, false);
// Check if it is available in global function table as injected function.
auto callee = [&]() -> llvm::Value* {
if (auto it = gv_func_map_.find(global_symbol); it != gv_func_map_.end()) {
if (it->second == nullptr) {
it->second = InitContextPtr(ftype->getPointerTo(), "__" + global_symbol);
}
return GetContextPtr(it->second);
} else if (llvm::Function* f = module_->getFunction(MakeStringRef(global_symbol))) {
return f;
} else {
return llvm::Function::Create(ftype, llvm::Function::ExternalLinkage,
MakeStringRef(global_symbol), module_.get());
}
}();
if (callee->getType() != ftype->getPointerTo()) {
callee = builder_->CreatePointerCast(callee, ftype->getPointerTo());
}
return builder_->CreateCall(ftype, callee, arg_values);
}
llvm::GlobalVariable* CodeGenCPU::InitContextPtr(llvm::Type* p_type, std::string name) {
llvm::GlobalVariable* gv = new llvm::GlobalVariable(
*module_, p_type, false, llvm::GlobalValue::LinkOnceAnyLinkage, nullptr, name);
#if TVM_LLVM_VERSION >= 100
gv->setAlignment(llvm::Align(data_layout_->getTypeAllocSize(p_type)));
#else
gv->setAlignment(data_layout_->getTypeAllocSize(p_type));
#endif
gv->setInitializer(llvm::Constant::getNullValue(p_type));
gv->setDLLStorageClass(llvm::GlobalValue::DLLStorageClassTypes::DLLExportStorageClass);
// comdat is needed for windows select any linking to work
// set comdat to Any(weak linking)
if (llvm_target_->GetOrCreateTargetMachine()->getTargetTriple().isOSWindows()) {
llvm::Comdat* comdat = module_->getOrInsertComdat(name);
comdat->setSelectionKind(llvm::Comdat::Any);
gv->setComdat(comdat);
}
return gv;
}
llvm::Value* CodeGenCPU::GetContextPtr(llvm::GlobalVariable* gv) {
ICHECK(gv != nullptr);
#if TVM_LLVM_VERSION >= 110
llvm::LoadInst* faddr =
builder_->CreateAlignedLoad(gv->getValueType(), gv, llvm::Align(gv->getAlignment()));
#elif TVM_LLVM_VERSION >= 80
llvm::LoadInst* faddr = builder_->CreateAlignedLoad(gv->getValueType(), gv, gv->getAlignment());
#else
llvm::LoadInst* faddr = builder_->CreateAlignedLoad(gv, gv->getAlignment());
#endif
faddr->setMetadata("tbaa",
md_builder_->createTBAAStructTagNode(md_tbaa_ctx_ptr_, md_tbaa_ctx_ptr_, 0));
return faddr;
}
void CodeGenCPU::InitGlobalContext(bool dynamic_lookup) {
std::string ctx_symbol = system_lib_prefix_.value_or("") + tvm::runtime::symbol::tvm_module_ctx;
// Module context
gv_mod_ctx_ = InitContextPtr(t_void_p_, ctx_symbol);
// Register back the locations.
if (f_tvm_register_system_symbol_ != nullptr && !target_c_runtime_) {
export_system_symbols_.emplace_back(std::make_pair(ctx_symbol, gv_mod_ctx_));
} else {
if (!dynamic_lookup) {
gv_tvm_func_call_ = InitContextPtr(ftype_tvm_func_call_->getPointerTo(), "__TVMFuncCall");
gv_tvm_get_func_from_env_ = InitContextPtr(ftype_tvm_get_func_from_env_->getPointerTo(),
"__TVMBackendGetFuncFromEnv");
gv_tvm_api_set_last_error_ =
InitContextPtr(ftype_tvm_api_set_last_error_->getPointerTo(), "__TVMAPISetLastError");
gv_tvm_parallel_launch_ =
InitContextPtr(ftype_tvm_parallel_launch_->getPointerTo(), "__TVMBackendParallelLaunch");
gv_tvm_parallel_barrier_ = InitContextPtr(ftype_tvm_parallel_barrier_->getPointerTo(),
"__TVMBackendParallelBarrier");
// Mark as context functions
gv_func_map_["TVMBackendAllocWorkspace"] = nullptr;
gv_func_map_["TVMBackendFreeWorkspace"] = nullptr;
}
}
}
llvm::BasicBlock* CodeGenCPU::CheckCallSuccess(llvm::Value* retcode) {
// create emit codes that checks and load the function.
llvm::LLVMContext* ctx = llvm_target_->GetContext();
auto* fail_block = llvm::BasicBlock::Create(*ctx, "call_fail", function_);
auto* end_block = llvm::BasicBlock::Create(*ctx, "call_end", function_);
auto* succ = builder_->CreateICmpEQ(retcode, llvm::ConstantInt::get(t_int_, 0));
builder_->CreateCondBr(succ, end_block, fail_block, md_very_likely_branch_);
builder_->SetInsertPoint(fail_block);
// return the code.
builder_->CreateRet(retcode);
// otherwise set it to be new end.
builder_->SetInsertPoint(end_block);
return end_block;
}
void CodeGenCPU::CreateComputeScope(const AttrStmtNode* op) {
EmitDebugLocation(op);
/*! \brief maintain states that should be guarded when step into compute scope */
struct ComputeScopeStates {
explicit ComputeScopeStates(CodeGenCPU* parent) : parent_(parent) {}
void EnterWithScope() {
std::swap(function_, parent_->function_);
std::swap(analyzer_, parent_->analyzer_);
std::swap(var_map_, parent_->var_map_);
std::swap(di_subprogram_, parent_->di_subprogram_);
}
void ExitWithScope() {
std::swap(function_, parent_->function_);
std::swap(analyzer_, parent_->analyzer_);
std::swap(var_map_, parent_->var_map_);
std::swap(di_subprogram_, parent_->di_subprogram_);
}
llvm::Function* function_{nullptr};
llvm::DISubprogram* di_subprogram_{nullptr};
std::unordered_map<const VarNode*, llvm::Value*> var_map_;
std::unique_ptr<arith::Analyzer> analyzer_{std::make_unique<arith::Analyzer>()};
CodeGenCPU* parent_;
};
// There are two reasons why we create another function for compute_scope
// - Make sure the generated compute function is clearly separately(though it can get inlined)
// - Set noalias on all the pointer arguments, some of them are loaded from TVMArgs.
// This is easier than set the alias scope manually.
Array<Var> vargs = tir::UndefinedVars(op->body, {});
std::vector<llvm::Value*> arg_values;
std::vector<llvm::Type*> arg_types;
for (Var v : vargs) {
llvm::Value* value = MakeValue(v);
value->setName(v->name_hint.c_str());
arg_values.push_back(value);
arg_types.push_back(value->getType());
}
llvm::FunctionType* ftype = llvm::FunctionType::get(t_int_, arg_types, false);
// $xxx_compute_ functions are not global. They should be marked as static (via InternalLinkage)
// to call them correctly on MIPS platform (CALL16 issue)
// Linkage ld Error: CALL16 reloc at 0x290 not against global symbol
const StringImmNode* value = op->value.as<StringImmNode>();
ICHECK(value != nullptr);
llvm::Function* fcompute = llvm::Function::Create(ftype, llvm::Function::InternalLinkage,
MakeStringRef(value->value), module_.get());
SetTargetAttributes(fcompute);
for (auto it = fcompute->arg_begin(); it != fcompute->arg_end(); it++) {
const Var& var = vargs[std::distance(fcompute->arg_begin(), it)];
it->setName(std::string(var->name_hint));
}
llvm::BasicBlock* compute_call_end = CheckCallSuccess(builder_->CreateCall(fcompute, arg_values));
llvm::LLVMContext* ctx = llvm_target_->GetContext();
// enter compute scope and setup compute function.
With<ComputeScopeStates> scope_states_guard(this);
size_t idx = 0;
for (auto it = fcompute->arg_begin(); it != fcompute->arg_end(); ++it, ++idx) {
llvm::Argument* v = &(*it);
const Var& var = vargs[idx];
var_map_[var.get()] = v;
if (var.dtype().is_handle() && !alias_var_set_.count(var.get())) {
// set non alias.
#if TVM_LLVM_VERSION >= 50
fcompute->addParamAttr(idx, llvm::Attribute::NoAlias);
// always not inline compute function to make the code structure clean
#else
fcompute->setDoesNotAlias(idx + 1);
#endif
fcompute->addFnAttr(llvm::Attribute::NoInline);
}
// Add alignment attribute if needed.
#if TVM_LLVM_VERSION >= 50
auto f = alloc_storage_info_.find(var.get());
if (f != alloc_storage_info_.end()) {
unsigned align = f->second.alignment;
if (align > 1) {
auto attr = llvm::Attribute::get(*ctx, llvm::Attribute::Alignment, align);
fcompute->addParamAttr(idx, attr);
}
}
#endif
}
function_ = fcompute;
di_subprogram_ = CreateDebugFunction(MakeStringRef(value->value), vargs.Map(GetType),
PrimType(DataType::Int(32)));
auto* compute_entry = llvm::BasicBlock::Create(*ctx, "entry", function_);
builder_->SetInsertPoint(compute_entry);
this->VisitStmt(op->body);
builder_->CreateRet(ConstInt32(0));
builder_->SetInsertPoint(compute_call_end);
AddDebugInformation(fcompute, vargs.Map(GetType));
}
CodeGenLLVM::TypedPointer CodeGenCPU::PackClosureData(const Array<Var>& vfields,
uint64_t* num_bytes,
std::string struct_name) {
if (vfields.size() == 0) {
*num_bytes = 0U;
return TypedPointer(t_void_p_, llvm::Constant::getNullValue(t_void_p_));
}
std::vector<llvm::Type*> fields;
for (Var v : vfields) {
auto it = var_map_.find(v.get());
ICHECK(it != var_map_.end());
fields.push_back(it->second->getType());
}
llvm::StructType* ctype = struct_name.size() ? llvm::StructType::create(fields, struct_name)
: llvm::StructType::create(fields);
llvm::AllocaInst* cvalue =
WithFunctionEntry([&]() { return builder_->CreateAlloca(ctype, ConstInt32(1)); });
llvm::Value* zero = ConstInt32(0);
for (size_t i = 0; i < vfields.size(); ++i) {
builder_->CreateStore(var_map_.at(vfields[i].get()),
builder_->CreateInBoundsGEP(ctype, cvalue, {zero, ConstInt32(i)}));
}
*num_bytes = data_layout_->getTypeAllocSize(ctype);
return TypedPointer(ctype, cvalue);
}
void CodeGenCPU::UnpackClosureData(TypedPointer cdata, const Array<Var>& vfields,
std::unordered_map<const VarNode*, llvm::Value*>* vmap) {
for (size_t i = 0; i < vfields.size(); ++i) {
llvm::Type* field_type = cdata.type->getStructElementType(i);
llvm::Value* field_addr =
builder_->CreateInBoundsGEP(cdata.type, cdata.addr, {ConstInt32(0), ConstInt32(i)});
llvm::Value* load =
builder_->CreateLoad(field_type, field_addr, std::string(vfields[i]->name_hint));
(*vmap)[vfields[i].get()] = load;
}
}
void CodeGenCPU::CreateParallelLaunch(const Stmt& body, int num_task, std::string name) {
// closure data
llvm::Function* f =
llvm::Function::Create(ftype_tvm_parallel_lambda_, llvm::Function::PrivateLinkage,
"__tvm_parallel_lambda", module_.get());
SetTargetAttributes(f);
// allocate and setup the closure, call the closure.
Array<Var> vfields = tir::UndefinedVars(body, {});
uint64_t nbytes;
TypedPointer cdata = PackClosureData(vfields, &nbytes, "closure_" + name);
#if TVM_LLVM_VERSION >= 90
auto launch_callee = llvm::FunctionCallee(ftype_tvm_parallel_launch_, RuntimeTVMParallelLaunch());
#else
auto launch_callee = RuntimeTVMParallelLaunch();
#endif
llvm::BasicBlock* par_launch_end = CheckCallSuccess(builder_->CreateCall(
launch_callee,
{f, builder_->CreatePointerCast(cdata.addr, t_void_p_), ConstInt32(num_task)}));
// Setup the closure function.
auto* lambda_entry =
llvm::BasicBlock::Create(*llvm_target_->GetContext(), "parallel_closure_entry", f);
builder_->SetInsertPoint(lambda_entry);
auto it = f->arg_begin();
llvm::Value* task_id = &(*it++);
task_id->setName("task_id");
llvm::Value* penv = &(*it++);
cdata.addr = builder_->CreatePointerCast(&(*it++), cdata.addr->getType());
// setup new variable map, swap it with current var context.
std::unordered_map<const VarNode*, llvm::Value*> new_vmap;
UnpackClosureData(cdata, vfields, &new_vmap);
// setup parallel env
ParallelEnv par_env;
par_env.task_id = Var("task_id", DataType::Int(32));
par_env.num_task = Var("num_task", DataType::Int(32));
new_vmap[par_env.task_id.get()] = task_id;
new_vmap[par_env.num_task.get()] = builder_->CreateLoad(
t_int32_,
builder_->CreateInBoundsGEP(t_tvm_parallel_group_env_, penv, {ConstInt32(0), ConstInt32(1)}),
"num_task");
par_env.penv = penv;
auto new_analyzer = std::make_unique<arith::Analyzer>();
std::swap(function_, f);
std::swap(parallel_env_, par_env);
std::swap(analyzer_, new_analyzer);
std::swap(var_map_, new_vmap);
this->VisitStmt(body);
builder_->CreateRet(ConstInt32(0));
// swap the var map back, now we are back on track.
std::swap(var_map_, new_vmap);
std::swap(analyzer_, new_analyzer);
std::swap(parallel_env_, par_env);
std::swap(function_, f);
ICHECK_NE(par_env.parallel_loop_count, 0) << "Cannot find parallel loop within parallel launch";
builder_->SetInsertPoint(par_launch_end);
}
llvm::Value* CodeGenCPU::CreateStaticHandle() {
llvm::GlobalVariable* gv =
new llvm::GlobalVariable(*module_, t_void_p_, false, llvm::GlobalValue::PrivateLinkage,
nullptr, "__tvm_static_handle");
#if TVM_LLVM_VERSION >= 100
gv->setAlignment(llvm::Align(data_layout_->getTypeAllocSize(t_void_p_)));
#else
gv->setAlignment(data_layout_->getTypeAllocSize(t_void_p_));
#endif
gv->setInitializer(llvm::Constant::getNullValue(t_void_p_));
return gv;
}
void CodeGenCPU::CreateStaticInit(const std::string& init_fname, const Stmt& body) {
// closure data
llvm::Function* f =
llvm::Function::Create(ftype_tvm_static_init_callback_, llvm::Function::PrivateLinkage,
"__tvm_static_init_lambda", module_.get());
SetTargetAttributes(f);
llvm::Value* gv = CreateStaticHandle();
llvm::Function* finit = module_->getFunction(init_fname);
if (finit == nullptr) {
finit = llvm::Function::Create(ftype_tvm_static_init_, llvm::Function::ExternalLinkage,
init_fname, module_.get());
}
// allocate and setup the closure, call the closure.
uint64_t nbytes;
Array<Var> vfields = tir::UndefinedVars(body, {});
TypedPointer cdata = PackClosureData(vfields, &nbytes);
llvm::BasicBlock* init_end = CheckCallSuccess(builder_->CreateCall(
finit, {gv, f, builder_->CreatePointerCast(cdata.addr, t_void_p_), ConstInt32(nbytes)}));
// Setup the closure function.
auto* lambda_entry = llvm::BasicBlock::Create(*llvm_target_->GetContext(), "entry", f);
builder_->SetInsertPoint(lambda_entry);
auto it = f->arg_begin();
cdata.addr = builder_->CreatePointerCast(&(*it++), cdata.addr->getType());
// setup new variable map, swap it with current var context.
std::unordered_map<const VarNode*, llvm::Value*> new_vmap;
UnpackClosureData(cdata, vfields, &new_vmap);
ICHECK(parallel_env_.penv == nullptr);
auto new_analyzer = std::make_unique<arith::Analyzer>();
std::swap(function_, f);
std::swap(analyzer_, new_analyzer);
std::swap(var_map_, new_vmap);
this->VisitStmt(body);
builder_->CreateRet(ConstInt32(0));
// swap the var map back, now we are back on track.
std::swap(var_map_, new_vmap);
std::swap(analyzer_, new_analyzer);
std::swap(function_, f);
builder_->SetInsertPoint(init_end);
}
llvm::Value* CodeGenCPU::GetPackedFuncHandle(const std::string& fname) {
// We will store the packed function handle in global space.
// Initialize it during the first call.
llvm::DataLayout layout(module_.get());
uint64_t align = layout.getTypeAllocSize(t_tvm_func_handle_);
auto it = func_handle_map_.find(fname);
llvm::GlobalVariable* hptr;
if (it == func_handle_map_.end()) {
// create global location for the handle
// create the function handle
hptr =
new llvm::GlobalVariable(*module_, t_tvm_func_handle_, false,
llvm::GlobalValue::InternalLinkage, nullptr, ".tvm_func." + fname);
#if TVM_LLVM_VERSION >= 100
hptr->setAlignment(llvm::Align(align));
#else
hptr->setAlignment(align);
#endif
hptr->setInitializer(llvm::Constant::getNullValue(t_tvm_func_handle_));
func_handle_map_[fname] = hptr;
} else {
hptr = it->second;
}
// create emit codes that checks and load the function.
llvm::LLVMContext* ctx = llvm_target_->GetContext();
llvm::BasicBlock* pre_block = builder_->GetInsertBlock();
auto* init_block = llvm::BasicBlock::Create(*ctx, "handle_init", function_);
auto* end_block = llvm::BasicBlock::Create(*ctx, "handle_init_end", function_);
#if TVM_LLVM_VERSION >= 110
llvm::Value* handle = builder_->CreateAlignedLoad(hptr->getValueType(), hptr, llvm::Align(align));
#elif TVM_LLVM_VERSION >= 80
llvm::Value* handle = builder_->CreateAlignedLoad(hptr->getValueType(), hptr, align);
#else
llvm::Value* handle = builder_->CreateAlignedLoad(hptr, align);
#endif
llvm::Value* handle_not_null =
builder_->CreateICmpNE(handle, llvm::Constant::getNullValue(t_tvm_func_handle_));
builder_->CreateCondBr(handle_not_null, end_block, init_block, md_very_likely_branch_);
// Initialize the handle if needed.
builder_->SetInsertPoint(init_block);
llvm::Value* out =
WithFunctionEntry([&]() { return builder_->CreateAlloca(t_tvm_func_handle_); });
#if TVM_LLVM_VERSION >= 110
llvm::LoadInst* ctx_load = builder_->CreateAlignedLoad(gv_mod_ctx_->getValueType(), gv_mod_ctx_,
llvm::Align(gv_mod_ctx_->getAlignment()));
#elif TVM_LLVM_VERSION >= 80
llvm::LoadInst* ctx_load = builder_->CreateAlignedLoad(gv_mod_ctx_->getValueType(), gv_mod_ctx_,
gv_mod_ctx_->getAlignment());
#else
llvm::LoadInst* ctx_load = builder_->CreateAlignedLoad(gv_mod_ctx_, gv_mod_ctx_->getAlignment());
#endif
ctx_load->setMetadata(
"tbaa", md_builder_->createTBAAStructTagNode(md_tbaa_ctx_ptr_, md_tbaa_ctx_ptr_, 0));
#if TVM_LLVM_VERSION >= 90
auto env_callee = llvm::FunctionCallee(ftype_tvm_get_func_from_env_, RuntimeTVMGetFuncFromEnv());
#else
auto env_callee = RuntimeTVMGetFuncFromEnv();
#endif
llvm::Value* retcode = builder_->CreateCall(env_callee, {ctx_load, GetConstString(fname), out});
init_block = CheckCallSuccess(retcode);
#if TVM_LLVM_VERSION >= 110
llvm::Value* loaded_handle =
builder_->CreateAlignedLoad(t_tvm_func_handle_, out, llvm::Align(align));
#elif TVM_LLVM_VERSION >= 80
llvm::Value* loaded_handle = builder_->CreateAlignedLoad(t_tvm_func_handle_, out, align);
#else
llvm::Value* loaded_handle = builder_->CreateAlignedLoad(out, align);
#endif
// Store the handle
builder_->CreateStore(loaded_handle, hptr);
builder_->CreateBr(end_block);
// end block
builder_->SetInsertPoint(end_block);
llvm::PHINode* phi = builder_->CreatePHI(t_tvm_func_handle_, 2);
phi->addIncoming(handle, pre_block);
phi->addIncoming(loaded_handle, init_block);
return phi;
}
CodeGenCPU::PackedCall CodeGenCPU::MakeCallPackedLowered(const Array<PrimExpr>& args,
const DataType& r_type,
const int64_t begin, const int64_t end,
bool use_string_lookup) {
std::string func_name = [&]() {
auto ptr = args[0].as<StringImmNode>();
ICHECK(ptr) << "Expected first argument of tir::Call to be "
<< "a string containing the callee's name, "
<< "but instead contained " << args[0];
return ptr->value;
}();
// call the function
int64_t nargs = end - begin;
ICHECK_GE(nargs, 0);
llvm::Value* stack_value = MakeValue(args[1]);
llvm::Value* stack_tcode = MakeValue(args[2]);
llvm::Value* arg_value = builder_->CreateInBoundsGEP(
t_tvm_value_, builder_->CreatePointerCast(stack_value, t_tvm_value_->getPointerTo()),
ConstInt32(begin));
TypedPointer arg_tcode =
CreateBufferPtr(stack_tcode, DataType::Int(32), {ConstInt32(begin)}, DataType::Int(32));
llvm::Value* ret_value = builder_->CreateInBoundsGEP(
t_tvm_value_, builder_->CreatePointerCast(stack_value, t_tvm_value_->getPointerTo()),
ConstInt32(end));
TypedPointer ret_tcode =
CreateBufferPtr(stack_tcode, DataType::Int(32), {ConstInt32(end)}, DataType::Int(32));
llvm::FunctionType* callee_ftype = nullptr;
llvm::Value* callee_value = nullptr;
std::vector<llvm::Value*> call_args;
if (use_string_lookup) {
callee_ftype = ftype_tvm_func_call_;
callee_value = RuntimeTVMFuncCall();
call_args.push_back(GetPackedFuncHandle(func_name));
call_args.insert(call_args.end(),
{arg_value, arg_tcode.addr, ConstInt32(nargs), ret_value, ret_tcode.addr});
} else {
callee_ftype = ftype_tvm_backend_packed_c_func_;
callee_value = module_->getFunction(func_name);
if (callee_value == nullptr) {
callee_value =
llvm::Function::Create(ftype_tvm_backend_packed_c_func_, llvm::Function::ExternalLinkage,
func_name, module_.get());
}
// NOTE: This is a bugfix to a previous coupled convention(in lower_tvm_builtin)
// The begin, end should correspond to the right location in cpacked excluding resource handle.
// TODO(tqchen): upstream the fix.
// nargs -= 1;
call_args.insert(call_args.end(), {
builder_->CreateBitCast(arg_value, t_void_p_),
arg_tcode.addr,
ConstInt32(nargs),
builder_->CreateBitCast(ret_value, t_void_p_),
ret_tcode.addr,
});
call_args.push_back(llvm::ConstantPointerNull::get(t_void_p_));
}
#if TVM_LLVM_VERSION >= 90
auto call_callee = llvm::FunctionCallee(callee_ftype, callee_value);
#else
(void)callee_ftype; // use callee_ftype to avoid unused variable warning when using older LLVM.
auto call_callee = callee_value;
#endif
llvm::Value* call = builder_->CreateCall(call_callee, call_args);
llvm::BasicBlock* end_block = CheckCallSuccess(call);
PackedCall pc = {nullptr};
if (!r_type.is_void()) {
// Load the return value and cast it to the designated type (r_type).
DataType r_api_type = tir::APIType(r_type);
llvm::Type* llvm_r_api_type = DTypeToLLVMType(r_api_type);
llvm::Value* load_ptr = builder_->CreatePointerCast(ret_value, llvm_r_api_type->getPointerTo());
#if TVM_LLVM_VERSION >= 110
llvm::Value* rvalue = builder_->CreateAlignedLoad(llvm_r_api_type, load_ptr, llvm::Align(8));
#elif TVM_LLVM_VERSION >= 80
llvm::Value* rvalue = builder_->CreateAlignedLoad(llvm_r_api_type, load_ptr, 8);
#else
llvm::Value* rvalue = builder_->CreateAlignedLoad(load_ptr, 8);
#endif
pc.ret_value = CreateCast(r_api_type, r_type, rvalue);
// Load the return type code.
#if TVM_LLVM_VERSION >= 110
pc.ret_tcode = builder_->CreateAlignedLoad(ret_tcode.type, ret_tcode.addr, llvm::Align(8));
#elif TVM_LLVM_VERSION >= 80
pc.ret_tcode = builder_->CreateAlignedLoad(ret_tcode.type, ret_tcode.addr, 8);
#else
pc.ret_tcode = builder_->CreateAlignedLoad(ret_tcode.addr, 8);
#endif
}
pc.end_block = end_block;
return pc;
}
llvm::Value* CodeGenCPU::CreateCallPacked(const CallNode* op, bool use_string_lookup) {
auto expected_num_args = use_string_lookup ? 5U : 6U;
ICHECK_EQ(op->args.size(), expected_num_args);
PackedCall pc = MakeCallPackedLowered(op->args, op->dtype, op->args[3].as<IntImmNode>()->value,
op->args[4].as<IntImmNode>()->value, use_string_lookup);
return pc.ret_value;
}
llvm::Value* CodeGenCPU::CreateCallTracePacked(const CallNode* op) {
ICHECK_EQ(op->args.size(), 6U);
PackedCall pc = MakeCallPackedLowered(op->args, op->dtype, op->args[3].as<IntImmNode>()->value,
op->args[4].as<IntImmNode>()->value, true);
llvm::LLVMContext* ctx = llvm_target_->GetContext();
// Get traced value.
llvm::Value* traced_value = MakeValue(op->args[5]);
// The update_block handles case when we need to update the return value.
llvm::BasicBlock* update_block = llvm::BasicBlock::Create(*ctx, "update_block", function_);
// The continue_block handles case when we need to return original
// traced value.
llvm::BasicBlock* continue_block = llvm::BasicBlock::Create(*ctx, "continue_block", function_);
// Check the ret_type_code and create cmp instruction.
llvm::Value* cmp =
builder_->CreateICmpNE(pc.ret_tcode, llvm::ConstantInt::get(t_int_, kTVMNullptr));
builder_->CreateCondBr(cmp, update_block, continue_block);
builder_->SetInsertPoint(update_block);
builder_->CreateBr(continue_block);
builder_->SetInsertPoint(continue_block);
// The return value depends on from what bb we come from.
llvm::PHINode* phi_rvalue = builder_->CreatePHI(traced_value->getType(), 2);
phi_rvalue->addIncoming(pc.ret_value, update_block);
phi_rvalue->addIncoming(traced_value, pc.end_block);
return phi_rvalue;
}
llvm::Value* CodeGenCPU::RuntimeTVMFuncCall() {
if (f_tvm_func_call_ != nullptr) return f_tvm_func_call_;
return GetContextPtr(gv_tvm_func_call_);
}
llvm::Value* CodeGenCPU::RuntimeTVMGetFuncFromEnv() {
if (f_tvm_get_func_from_env_ != nullptr) return f_tvm_get_func_from_env_;
return GetContextPtr(gv_tvm_get_func_from_env_);
}
llvm::Value* CodeGenCPU::RuntimeTVMAPISetLastError() {
if (f_tvm_api_set_last_error_ != nullptr) return f_tvm_api_set_last_error_;
return GetContextPtr(gv_tvm_api_set_last_error_);
}
llvm::Value* CodeGenCPU::RuntimeTVMParallelLaunch() {
if (f_tvm_parallel_launch_ != nullptr) return f_tvm_parallel_launch_;
return GetContextPtr(gv_tvm_parallel_launch_);
}
llvm::Value* CodeGenCPU::RuntimeTVMParallelBarrier() {
if (f_tvm_parallel_barrier_ != nullptr) return f_tvm_parallel_barrier_;
return GetContextPtr(gv_tvm_parallel_barrier_);
}
/*! \brief Defines LLVM Types for each Metadata member type. */
struct MetadataLlvmTypes {
llvm::Type* t_float64;
llvm::Type* t_uint8;
llvm::Type* t_int64;
llvm::Type* t_bool;
llvm::Type* t_cstring;
llvm::Type* t_void_p;
llvm::StructType* t_data_type;
/*! \brief Maps a MetadataBase subclass' type_key to its corresponding LLVM StructType. */
::std::unordered_map<std::string, llvm::StructType*> structs_by_type_key;
};
class MetadataTypeDefiner : public AttrVisitor {
public:
MetadataTypeDefiner(llvm::LLVMContext* ctx, struct MetadataLlvmTypes* llvm_types)
: ctx_{ctx}, llvm_types_{llvm_types} {}
void Visit(const char* key, double* value) final {
elements_.emplace_back(llvm_types_->t_float64);
}
void Visit(const char* key, int64_t* value) final {
elements_.emplace_back(llvm_types_->t_int64);
}
void Visit(const char* key, uint64_t* value) final {
elements_.emplace_back(llvm_types_->t_int64);
}
void Visit(const char* key, int* value) final { elements_.emplace_back(llvm_types_->t_int64); }
void Visit(const char* key, bool* value) final { elements_.emplace_back(llvm_types_->t_bool); }
void Visit(const char* key, std::string* value) final {
elements_.emplace_back(llvm_types_->t_cstring);
}