forked from JuliaLang/julia
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathaotcompile.cpp
2067 lines (1904 loc) · 86.6 KB
/
aotcompile.cpp
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
// This file is a part of Julia. License is MIT: https://julialang.org/license
#include "llvm-version.h"
#include "platform.h"
// target support
#if JL_LLVM_VERSION >= 170000
#include <llvm/TargetParser/Triple.h>
#else
#include <llvm/ADT/Triple.h>
#endif
#include "llvm/Support/CodeGen.h"
#include <llvm/ADT/Statistic.h>
#include <llvm/Analysis/TargetLibraryInfo.h>
#include <llvm/Analysis/TargetTransformInfo.h>
#include <llvm/IR/DataLayout.h>
#include <llvm/MC/TargetRegistry.h>
#include <llvm/Target/TargetMachine.h>
// analysis passes
#include <llvm/Analysis/Passes.h>
#include <llvm/IR/IRBuilder.h>
#include <llvm/IR/PassManager.h>
#include <llvm/IR/Verifier.h>
#include <llvm/Transforms/Utils/ModuleUtils.h>
#include <llvm/Passes/PassBuilder.h>
#include <llvm/Passes/PassPlugin.h>
#if defined(USE_POLLY)
#include <polly/RegisterPasses.h>
#include <polly/LinkAllPasses.h>
#include <polly/CodeGen/CodegenCleanup.h>
#if defined(USE_POLLY_ACC)
#include <polly/Support/LinkGPURuntime.h>
#endif
#endif
// for outputting code
#include <llvm/Bitcode/BitcodeWriter.h>
#include <llvm/Bitcode/BitcodeWriterPass.h>
#include <llvm/Bitcode/BitcodeReader.h>
#include "llvm/Object/ArchiveWriter.h"
#include <llvm/IR/IRPrintingPasses.h>
#include <llvm/IR/LegacyPassManagers.h>
#include <llvm/Transforms/Utils/Cloning.h>
#include <llvm/Support/FormatAdapters.h>
#include <llvm/Linker/Linker.h>
using namespace llvm;
#include "jitlayers.h"
#include "serialize.h"
#include "julia_assert.h"
#include "processor.h"
#define DEBUG_TYPE "julia_aotcompile"
STATISTIC(CICacheLookups, "Number of codeinst cache lookups");
STATISTIC(CreateNativeCalls, "Number of jl_create_native calls made");
STATISTIC(CreateNativeMethods, "Number of methods compiled for jl_create_native");
STATISTIC(CreateNativeMax, "Max number of methods compiled at once for jl_create_native");
STATISTIC(CreateNativeGlobals, "Number of globals compiled for jl_create_native");
static void addComdat(GlobalValue *G, Triple &T)
{
if (T.isOSBinFormatCOFF() && !G->isDeclaration()) {
// add __declspec(dllexport) to everything marked for export
assert(G->hasExternalLinkage() && "Cannot set DLLExport on non-external linkage!");
G->setDLLStorageClass(GlobalValue::DLLExportStorageClass);
}
}
typedef struct {
orc::ThreadSafeModule M;
SmallVector<GlobalValue*, 0> jl_sysimg_fvars;
SmallVector<GlobalValue*, 0> jl_sysimg_gvars;
std::map<jl_code_instance_t*, std::tuple<uint32_t, uint32_t>> jl_fvar_map;
SmallVector<void*, 0> jl_value_to_llvm;
SmallVector<jl_code_instance_t*, 0> jl_external_to_llvm;
} jl_native_code_desc_t;
extern "C" JL_DLLEXPORT_CODEGEN
void jl_get_function_id_impl(void *native_code, jl_code_instance_t *codeinst,
int32_t *func_idx, int32_t *specfunc_idx)
{
jl_native_code_desc_t *data = (jl_native_code_desc_t*)native_code;
if (data) {
// get the function index in the fvar lookup table
auto it = data->jl_fvar_map.find(codeinst);
if (it != data->jl_fvar_map.end()) {
std::tie(*func_idx, *specfunc_idx) = it->second;
}
}
}
extern "C" JL_DLLEXPORT_CODEGEN
void jl_get_llvm_gvs_impl(void *native_code, arraylist_t *gvs)
{
// map a memory location (jl_value_t or jl_binding_t) to a GlobalVariable
jl_native_code_desc_t *data = (jl_native_code_desc_t*)native_code;
arraylist_grow(gvs, data->jl_value_to_llvm.size());
memcpy(gvs->items, data->jl_value_to_llvm.data(), gvs->len * sizeof(void*));
}
extern "C" JL_DLLEXPORT_CODEGEN
void jl_get_llvm_external_fns_impl(void *native_code, arraylist_t *external_fns)
{
jl_native_code_desc_t *data = (jl_native_code_desc_t*)native_code;
arraylist_grow(external_fns, data->jl_external_to_llvm.size());
memcpy(external_fns->items, data->jl_external_to_llvm.data(),
external_fns->len * sizeof(jl_code_instance_t*));
}
extern "C" JL_DLLEXPORT_CODEGEN
LLVMOrcThreadSafeModuleRef jl_get_llvm_module_impl(void *native_code)
{
jl_native_code_desc_t *data = (jl_native_code_desc_t*)native_code;
if (data)
return wrap(&data->M);
else
return NULL;
}
extern "C" JL_DLLEXPORT_CODEGEN
GlobalValue* jl_get_llvm_function_impl(void *native_code, uint32_t idx)
{
jl_native_code_desc_t *data = (jl_native_code_desc_t*)native_code;
if (data)
return data->jl_sysimg_fvars[idx];
else
return NULL;
}
template<typename T>
static inline SmallVector<T*, 0> consume_gv(Module &M, const char *name, bool allow_bad_fvars)
{
// Get information about sysimg export functions from the two global variables.
// Strip them from the Module so that it's easier to handle the uses.
GlobalVariable *gv = M.getGlobalVariable(name);
assert(gv && gv->hasInitializer());
ArrayType *Ty = cast<ArrayType>(gv->getInitializer()->getType());
unsigned nele = Ty->getArrayNumElements();
SmallVector<T*, 0> res(nele);
ConstantArray *ary = nullptr;
if (gv->getInitializer()->isNullValue()) {
for (unsigned i = 0; i < nele; ++i)
res[i] = cast<T>(Constant::getNullValue(Ty->getArrayElementType()));
}
else {
ary = cast<ConstantArray>(gv->getInitializer());
unsigned i = 0;
while (i < nele) {
llvm::Value *val = ary->getOperand(i)->stripPointerCasts();
if (allow_bad_fvars && (!isa<T>(val) || (isa<Function>(val) && cast<Function>(val)->isDeclaration()))) {
// Shouldn't happen in regular use, but can happen in bugpoint.
nele--;
continue;
}
res[i++] = cast<T>(val);
}
res.resize(nele);
}
assert(gv->use_empty());
gv->eraseFromParent();
if (ary && ary->use_empty())
ary->destroyConstant();
return res;
}
static Constant *get_ptrdiff32(Type *T_size, Constant *ptr, Constant *base)
{
if (ptr->getType()->isPointerTy())
ptr = ConstantExpr::getPtrToInt(ptr, T_size);
auto ptrdiff = ConstantExpr::getSub(ptr, base);
return T_size->getPrimitiveSizeInBits() > 32 ? ConstantExpr::getTrunc(ptrdiff, Type::getInt32Ty(ptr->getContext())) : ptrdiff;
}
static Constant *emit_offset_table(Module &M, Type *T_size, ArrayRef<Constant*> vars,
StringRef name, StringRef suffix)
{
auto T_int32 = Type::getInt32Ty(M.getContext());
uint32_t nvars = vars.size();
ArrayType *vars_type = ArrayType::get(T_int32, nvars + 1);
auto gv = new GlobalVariable(M, vars_type, true,
GlobalVariable::ExternalLinkage,
nullptr,
name + "_offsets" + suffix);
auto vbase = ConstantExpr::getPtrToInt(gv, T_size);
SmallVector<Constant*, 0> offsets(nvars + 1);
offsets[0] = ConstantInt::get(T_int32, nvars);
for (uint32_t i = 0; i < nvars; i++)
offsets[i + 1] = get_ptrdiff32(T_size, vars[i], vbase);
gv->setInitializer(ConstantArray::get(vars_type, offsets));
gv->setVisibility(GlobalValue::HiddenVisibility);
gv->setDSOLocal(true);
return vbase;
}
static void emit_table(Module &mod, ArrayRef<GlobalValue*> vars,
StringRef name, Type *T_psize)
{
// Emit a global variable with all the variable addresses.
size_t nvars = vars.size();
SmallVector<Constant*, 0> addrs(nvars);
for (size_t i = 0; i < nvars; i++) {
Constant *var = vars[i];
addrs[i] = ConstantExpr::getBitCast(var, T_psize);
}
ArrayType *vars_type = ArrayType::get(T_psize, nvars);
auto GV = new GlobalVariable(mod, vars_type, true,
GlobalVariable::ExternalLinkage,
ConstantArray::get(vars_type, addrs),
name);
GV->setVisibility(GlobalValue::HiddenVisibility);
GV->setDSOLocal(true);
}
static bool is_safe_char(unsigned char c)
{
return ('0' <= c && c <= '9') ||
('A' <= c && c <= 'Z') ||
('a' <= c && c <= 'z') ||
(c == '_' || c == '$') ||
(c >= 128 && c < 255);
}
static const char hexchars[16] = {
'0', '1', '2', '3', '4', '5', '6', '7',
'8', '9', 'A', 'B', 'C', 'D', 'E', 'F' };
static const char *const common_names[256] = {
// 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, a, b, c, d, e, f
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 0x00
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 0x10
"SP", "NOT", "DQT", "YY", 0, "REM", "AND", "SQT", // 0x20
"LPR", "RPR", "MUL", "SUM", 0, "SUB", "DOT", "DIV", // 0x28
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, "COL", 0, "LT", "EQ", "GT", "QQ", // 0x30
"AT", 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 0x40
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, "LBR", "RDV", "RBR", "POW", 0, // 0x50
"TIC", 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 0x60
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, "LCR", "OR", "RCR", "TLD", "DEL", // 0x70
0 }; // remainder is filled with zeros, though are also all safe characters
// reversibly removes special characters from the name of GlobalObjects,
// which might cause them to be treated special by LLVM or the system linker
// the only non-identifier characters we allow to appear are '.' and '$',
// and all of UTF-8 above code-point 128 (except 255)
// most are given "friendly" abbreviations
// the remaining few will print as hex
// e.g. mangles "llvm.a≠a$a!a##" as "llvmDOT.a≠a$aNOT.aYY.YY."
static void makeSafeName(GlobalObject &G)
{
StringRef Name = G.getName();
SmallVector<char, 32> SafeName;
for (unsigned char c : Name.bytes()) {
if (is_safe_char(c)) {
SafeName.push_back(c);
}
else {
if (common_names[c]) {
SafeName.push_back(common_names[c][0]);
SafeName.push_back(common_names[c][1]);
if (common_names[c][2])
SafeName.push_back(common_names[c][2]);
}
else {
SafeName.push_back(hexchars[(c >> 4) & 0xF]);
SafeName.push_back(hexchars[c & 0xF]);
}
SafeName.push_back('.');
}
}
if (SafeName.size() != Name.size())
G.setName(StringRef(SafeName.data(), SafeName.size()));
}
jl_code_instance_t *jl_ci_cache_lookup(const jl_cgparams_t &cgparams, jl_method_instance_t *mi, size_t world)
{
++CICacheLookups;
jl_value_t *ci = cgparams.lookup(mi, world, world);
JL_GC_PROMISE_ROOTED(ci);
jl_code_instance_t *codeinst = NULL;
if (ci != jl_nothing) {
codeinst = (jl_code_instance_t*)ci;
}
else {
if (cgparams.lookup != jl_rettype_inferred_addr) {
jl_error("Refusing to automatically run type inference with custom cache lookup.");
}
else {
codeinst = jl_type_infer(mi, world, SOURCE_MODE_ABI);
/* Even if this codeinst is ordinarily not cacheable, we need to force
* it into the cache here, since it was explicitly requested and is
* otherwise not reachable from anywhere in the system image.
*/
if (!jl_mi_cache_has_ci(mi, codeinst))
jl_mi_cache_insert(mi, codeinst);
}
}
return codeinst;
}
// takes the running content that has collected in the shadow module and dump it to disk
// this builds the object file portion of the sysimage files for fast startup, and can
// also be used be extern consumers like GPUCompiler.jl to obtain a module containing
// all reachable & inferrrable functions.
// The `policy` flag switches between the default mode `0` and the extern mode `1` used by GPUCompiler.
// `_imaging_mode` controls if raw pointers can be embedded (e.g. the code will be loaded into the same session).
// `_external_linkage` create linkages between pkgimages.
extern "C" JL_DLLEXPORT_CODEGEN
void *jl_create_native_impl(jl_array_t *methods, LLVMOrcThreadSafeModuleRef llvmmod, const jl_cgparams_t *cgparams, int _policy, int _imaging_mode, int _external_linkage, size_t _world)
{
JL_TIMING(NATIVE_AOT, NATIVE_Create);
++CreateNativeCalls;
CreateNativeMax.updateMax(jl_array_nrows(methods));
if (cgparams == NULL)
cgparams = &jl_default_cgparams;
jl_native_code_desc_t *data = new jl_native_code_desc_t;
CompilationPolicy policy = (CompilationPolicy) _policy;
bool imaging = imaging_default() || _imaging_mode == 1;
jl_method_instance_t *mi = NULL;
jl_code_info_t *src = NULL;
JL_GC_PUSH1(&src);
auto ct = jl_current_task;
bool timed = (ct->reentrant_timing & 1) == 0;
if (timed)
ct->reentrant_timing |= 1;
orc::ThreadSafeContext ctx;
orc::ThreadSafeModule backing;
if (!llvmmod) {
ctx = jl_ExecutionEngine->acquireContext();
backing = jl_create_ts_module("text", ctx);
}
orc::ThreadSafeModule &clone = llvmmod ? *unwrap(llvmmod) : backing;
auto ctxt = clone.getContext();
uint64_t compiler_start_time = 0;
uint8_t measure_compile_time_enabled = jl_atomic_load_relaxed(&jl_measure_compile_time_enabled);
if (measure_compile_time_enabled)
compiler_start_time = jl_hrtime();
// compile all methods for the current world and type-inference world
auto target_info = clone.withModuleDo([&](Module &M) {
return std::make_pair(M.getDataLayout(), Triple(M.getTargetTriple()));
});
jl_codegen_params_t params(ctxt, std::move(target_info.first), std::move(target_info.second));
params.params = cgparams;
params.imaging_mode = imaging;
params.debug_level = cgparams->debug_info_level;
params.external_linkage = _external_linkage;
size_t compile_for[] = { jl_typeinf_world, _world };
for (int worlds = 0; worlds < 2; worlds++) {
JL_TIMING(NATIVE_AOT, NATIVE_Codegen);
size_t this_world = compile_for[worlds];
if (!this_world)
continue;
// Don't emit methods for the typeinf_world with extern policy
if (policy != CompilationPolicy::Default && this_world == jl_typeinf_world)
continue;
size_t i, l;
for (i = 0, l = jl_array_nrows(methods); i < l; i++) {
// each item in this list is either a MethodInstance indicating something
// to compile, or an svec(rettype, sig) describing a C-callable alias to create.
jl_value_t *item = jl_array_ptr_ref(methods, i);
if (jl_is_simplevector(item)) {
if (worlds == 1)
jl_compile_extern_c(wrap(&clone), ¶ms, NULL, jl_svecref(item, 0), jl_svecref(item, 1));
continue;
}
mi = (jl_method_instance_t*)item;
src = NULL;
// if this method is generally visible to the current compilation world,
// and this is either the primary world, or not applicable in the primary world
// then we want to compile and emit this
if (jl_atomic_load_relaxed(&mi->def.method->primary_world) <= this_world && this_world <= jl_atomic_load_relaxed(&mi->def.method->deleted_world)) {
// find and prepare the source code to compile
jl_code_instance_t *codeinst = jl_ci_cache_lookup(*cgparams, mi, this_world);
if (codeinst && !params.compiled_functions.count(codeinst)) {
// now add it to our compilation results
JL_GC_PROMISE_ROOTED(codeinst->rettype);
orc::ThreadSafeModule result_m = jl_create_ts_module(name_from_method_instance(codeinst->def),
params.tsctx, clone.getModuleUnlocked()->getDataLayout(),
Triple(clone.getModuleUnlocked()->getTargetTriple()));
jl_llvm_functions_t decls = jl_emit_codeinst(result_m, codeinst, NULL, params);
if (result_m)
params.compiled_functions[codeinst] = {std::move(result_m), std::move(decls)};
}
}
}
// finally, make sure all referenced methods also get compiled or fixed up
jl_compile_workqueue(params, policy);
}
JL_GC_POP();
// process the globals array, before jl_merge_module destroys them
SmallVector<std::string, 0> gvars(params.global_targets.size());
data->jl_value_to_llvm.resize(params.global_targets.size());
StringSet<> gvars_names;
DenseSet<GlobalValue *> gvars_set;
size_t idx = 0;
for (auto &global : params.global_targets) {
gvars[idx] = global.second->getName().str();
global.second->setInitializer(literal_static_pointer_val(global.first, global.second->getValueType()));
assert(gvars_set.insert(global.second).second && "Duplicate gvar in params!");
assert(gvars_names.insert(gvars[idx]).second && "Duplicate gvar name in params!");
data->jl_value_to_llvm[idx] = global.first;
idx++;
}
CreateNativeMethods += params.compiled_functions.size();
size_t offset = gvars.size();
data->jl_external_to_llvm.resize(params.external_fns.size());
for (auto &extern_fn : params.external_fns) {
jl_code_instance_t *this_code = std::get<0>(extern_fn.first);
bool specsig = std::get<1>(extern_fn.first);
assert(specsig && "Error external_fns doesn't handle non-specsig yet");
(void) specsig;
GlobalVariable *F = extern_fn.second;
size_t idx = gvars.size() - offset;
assert(idx >= 0);
assert(idx < data->jl_external_to_llvm.size());
data->jl_external_to_llvm[idx] = this_code;
assert(gvars_set.insert(F).second && "Duplicate gvar in params!");
assert(gvars_names.insert(F->getName()).second && "Duplicate gvar name in params!");
gvars.push_back(std::string(F->getName()));
}
// clones the contents of the module `m` to the shadow_output collector
// while examining and recording what kind of function pointer we have
{
JL_TIMING(NATIVE_AOT, NATIVE_Merge);
Linker L(*clone.getModuleUnlocked());
for (auto &def : params.compiled_functions) {
jl_merge_module(clone, std::move(std::get<0>(def.second)));
jl_code_instance_t *this_code = def.first;
jl_llvm_functions_t decls = std::get<1>(def.second);
StringRef func = decls.functionObject;
StringRef cfunc = decls.specFunctionObject;
uint32_t func_id = 0;
uint32_t cfunc_id = 0;
if (func == "jl_fptr_args") {
func_id = -1;
}
else if (func == "jl_fptr_sparam") {
func_id = -2;
}
else {
//Safe b/c context is locked by params
data->jl_sysimg_fvars.push_back(cast<Function>(clone.getModuleUnlocked()->getNamedValue(func)));
func_id = data->jl_sysimg_fvars.size();
}
if (!cfunc.empty()) {
//Safe b/c context is locked by params
data->jl_sysimg_fvars.push_back(cast<Function>(clone.getModuleUnlocked()->getNamedValue(cfunc)));
cfunc_id = data->jl_sysimg_fvars.size();
}
data->jl_fvar_map[this_code] = std::make_tuple(func_id, cfunc_id);
}
if (params._shared_module) {
bool error = L.linkInModule(std::move(params._shared_module));
assert(!error && "Error linking in shared module");
(void)error;
}
}
// now get references to the globals in the merged module
// and set them to be internalized and initialized at startup
for (auto &global : gvars) {
//Safe b/c context is locked by params
GlobalVariable *G = cast<GlobalVariable>(clone.getModuleUnlocked()->getNamedValue(global));
assert(G->hasInitializer());
G->setLinkage(GlobalValue::InternalLinkage);
G->setDSOLocal(true);
data->jl_sysimg_gvars.push_back(G);
}
CreateNativeGlobals += gvars.size();
//Safe b/c context is locked by params
auto TT = Triple(clone.getModuleUnlocked()->getTargetTriple());
Function *juliapersonality_func = nullptr;
if (TT.isOSWindows() && TT.getArch() == Triple::x86_64) {
// setting the function personality enables stack unwinding and catching exceptions
// so make sure everything has something set
Type *T_int32 = Type::getInt32Ty(clone.getModuleUnlocked()->getContext());
juliapersonality_func = Function::Create(FunctionType::get(T_int32, true),
Function::ExternalLinkage, "__julia_personality", clone.getModuleUnlocked());
juliapersonality_func->setDLLStorageClass(GlobalValue::DLLImportStorageClass);
}
// move everything inside, now that we've merged everything
// (before adding the exported headers)
if (policy == CompilationPolicy::Default) {
//Safe b/c context is locked by params
for (GlobalObject &G : clone.getModuleUnlocked()->global_objects()) {
if (!G.isDeclaration()) {
G.setLinkage(GlobalValue::InternalLinkage);
G.setDSOLocal(true);
makeSafeName(G);
if (Function *F = dyn_cast<Function>(&G)) {
if (TT.isOSWindows() && TT.getArch() == Triple::x86_64) {
// Add unwind exception personalities to functions to handle async exceptions
F->setPersonalityFn(juliapersonality_func);
}
}
}
}
}
data->M = std::move(clone);
if (timed) {
if (measure_compile_time_enabled) {
auto end = jl_hrtime();
jl_atomic_fetch_add_relaxed(&jl_cumulative_compile_time, end - compiler_start_time);
}
ct->reentrant_timing &= ~1ull;
}
if (ctx.getContext()) {
jl_ExecutionEngine->releaseContext(std::move(ctx));
}
return (void*)data;
}
static object::Archive::Kind getDefaultForHost(Triple &triple)
{
if (triple.isOSDarwin())
return object::Archive::K_DARWIN;
return object::Archive::K_GNU;
}
typedef Error ArchiveWriterError;
static void reportWriterError(const ErrorInfoBase &E)
{
std::string err = E.message();
jl_safe_printf("ERROR: failed to emit output file %s\n", err.c_str());
}
static void injectCRTAlias(Module &M, StringRef name, StringRef alias, FunctionType *FT)
{
Function *target = M.getFunction(alias);
if (!target) {
target = Function::Create(FT, Function::ExternalLinkage, alias, M);
}
Function *interposer = Function::Create(FT, Function::InternalLinkage, name, M);
appendToCompilerUsed(M, {interposer});
llvm::IRBuilder<> builder(BasicBlock::Create(M.getContext(), "top", interposer));
SmallVector<Value *, 4> CallArgs;
for (auto &arg : interposer->args())
CallArgs.push_back(&arg);
auto val = builder.CreateCall(target, CallArgs);
builder.CreateRet(val);
}
void multiversioning_preannotate(Module &M);
// See src/processor.h for documentation about this table. Corresponds to jl_image_shard_t.
static GlobalVariable *emit_shard_table(Module &M, Type *T_size, Type *T_psize, unsigned threads) {
SmallVector<Constant *, 0> tables(sizeof(jl_image_shard_t) / sizeof(void *) * threads);
for (unsigned i = 0; i < threads; i++) {
auto suffix = "_" + std::to_string(i);
auto create_gv = [&](StringRef name, bool constant) {
auto gv = new GlobalVariable(M, T_size, constant,
GlobalValue::ExternalLinkage, nullptr, name + suffix);
gv->setVisibility(GlobalValue::HiddenVisibility);
gv->setDSOLocal(true);
return gv;
};
auto table = tables.data() + i * sizeof(jl_image_shard_t) / sizeof(void *);
table[offsetof(jl_image_shard_t, fvar_count) / sizeof(void*)] = create_gv("jl_fvar_count", true);
table[offsetof(jl_image_shard_t, fvar_ptrs) / sizeof(void*)] = create_gv("jl_fvar_ptrs", true);
table[offsetof(jl_image_shard_t, fvar_idxs) / sizeof(void*)] = create_gv("jl_fvar_idxs", true);
table[offsetof(jl_image_shard_t, gvar_offsets) / sizeof(void*)] = create_gv("jl_gvar_offsets", true);
table[offsetof(jl_image_shard_t, gvar_idxs) / sizeof(void*)] = create_gv("jl_gvar_idxs", true);
table[offsetof(jl_image_shard_t, clone_slots) / sizeof(void*)] = create_gv("jl_clone_slots", true);
table[offsetof(jl_image_shard_t, clone_ptrs) / sizeof(void*)] = create_gv("jl_clone_ptrs", true);
table[offsetof(jl_image_shard_t, clone_idxs) / sizeof(void*)] = create_gv("jl_clone_idxs", true);
}
auto tables_arr = ConstantArray::get(ArrayType::get(T_psize, tables.size()), tables);
auto tables_gv = new GlobalVariable(M, tables_arr->getType(), false,
GlobalValue::ExternalLinkage, tables_arr, "jl_shard_tables");
tables_gv->setVisibility(GlobalValue::HiddenVisibility);
tables_gv->setDSOLocal(true);
return tables_gv;
}
// See src/processor.h for documentation about this table. Corresponds to jl_image_ptls_t.
static GlobalVariable *emit_ptls_table(Module &M, Type *T_size, Type *T_psize) {
std::array<Constant *, 3> ptls_table{
new GlobalVariable(M, T_size, false, GlobalValue::ExternalLinkage, Constant::getNullValue(T_size), "jl_pgcstack_func_slot"),
new GlobalVariable(M, T_size, false, GlobalValue::ExternalLinkage, Constant::getNullValue(T_size), "jl_pgcstack_key_slot"),
new GlobalVariable(M, T_size, false, GlobalValue::ExternalLinkage, Constant::getNullValue(T_size), "jl_tls_offset"),
};
for (auto &gv : ptls_table) {
cast<GlobalVariable>(gv)->setVisibility(GlobalValue::HiddenVisibility);
cast<GlobalVariable>(gv)->setDSOLocal(true);
}
auto ptls_table_arr = ConstantArray::get(ArrayType::get(T_psize, ptls_table.size()), ptls_table);
auto ptls_table_gv = new GlobalVariable(M, ptls_table_arr->getType(), false,
GlobalValue::ExternalLinkage, ptls_table_arr, "jl_ptls_table");
ptls_table_gv->setVisibility(GlobalValue::HiddenVisibility);
ptls_table_gv->setDSOLocal(true);
return ptls_table_gv;
}
// See src/processor.h for documentation about this table. Corresponds to jl_image_header_t.
static GlobalVariable *emit_image_header(Module &M, unsigned threads, unsigned nfvars, unsigned ngvars) {
constexpr uint32_t version = 1;
std::array<uint32_t, 4> header{
version,
threads,
nfvars,
ngvars,
};
auto header_arr = ConstantDataArray::get(M.getContext(), header);
auto header_gv = new GlobalVariable(M, header_arr->getType(), false,
GlobalValue::InternalLinkage, header_arr, "jl_image_header");
return header_gv;
}
// Grab fvars and gvars data from the module
static void get_fvars_gvars(Module &M, DenseMap<GlobalValue *, unsigned> &fvars, DenseMap<GlobalValue *, unsigned> &gvars) {
auto fvars_gv = M.getGlobalVariable("jl_fvars");
auto gvars_gv = M.getGlobalVariable("jl_gvars");
auto fvars_idxs = M.getGlobalVariable("jl_fvar_idxs");
auto gvars_idxs = M.getGlobalVariable("jl_gvar_idxs");
assert(fvars_gv);
assert(gvars_gv);
assert(fvars_idxs);
assert(gvars_idxs);
auto fvars_init = cast<ConstantArray>(fvars_gv->getInitializer());
auto gvars_init = cast<ConstantArray>(gvars_gv->getInitializer());
for (unsigned i = 0; i < fvars_init->getNumOperands(); ++i) {
auto gv = cast<GlobalValue>(fvars_init->getOperand(i)->stripPointerCasts());
assert(gv && gv->hasName() && "fvar must be a named global");
assert(!fvars.count(gv) && "Duplicate fvar");
fvars[gv] = i;
}
assert(fvars.size() == fvars_init->getNumOperands());
for (unsigned i = 0; i < gvars_init->getNumOperands(); ++i) {
auto gv = cast<GlobalValue>(gvars_init->getOperand(i)->stripPointerCasts());
assert(gv && gv->hasName() && "gvar must be a named global");
assert(!gvars.count(gv) && "Duplicate gvar");
gvars[gv] = i;
}
assert(gvars.size() == gvars_init->getNumOperands());
fvars_gv->eraseFromParent();
gvars_gv->eraseFromParent();
fvars_idxs->eraseFromParent();
gvars_idxs->eraseFromParent();
}
// Weight computation
// It is important for multithreaded image building to be able to split work up
// among the threads equally. The weight calculated here is an estimation of
// how expensive a particular function is going to be to compile.
struct FunctionInfo {
size_t weight;
size_t bbs;
size_t insts;
size_t clones;
};
static FunctionInfo getFunctionWeight(const Function &F)
{
FunctionInfo info;
info.weight = 1;
info.bbs = F.size();
info.insts = 0;
info.clones = 1;
for (const BasicBlock &BB : F) {
info.insts += BB.size();
}
if (F.hasFnAttribute("julia.mv.clones")) {
auto val = F.getFnAttribute("julia.mv.clones").getValueAsString();
// base16, so must be at most 4 * length bits long
// popcount gives number of clones
#if JL_LLVM_VERSION >= 170000
info.clones = APInt(val.size() * 4, val, 16).popcount() + 1;
#else
info.clones = APInt(val.size() * 4, val, 16).countPopulation() + 1;
#endif
}
info.weight += info.insts;
// more basic blocks = more complex than just sum of insts,
// add some weight to it
info.weight += info.bbs;
info.weight *= info.clones;
return info;
}
struct ModuleInfo {
Triple triple;
size_t globals;
size_t funcs;
size_t bbs;
size_t insts;
size_t clones;
size_t weight;
};
ModuleInfo compute_module_info(Module &M) {
ModuleInfo info;
info.triple = Triple(M.getTargetTriple());
info.globals = 0;
info.funcs = 0;
info.bbs = 0;
info.insts = 0;
info.clones = 0;
info.weight = 0;
for (auto &G : M.global_values()) {
if (G.isDeclaration()) {
continue;
}
info.globals++;
if (auto F = dyn_cast<Function>(&G)) {
info.funcs++;
auto func_info = getFunctionWeight(*F);
info.bbs += func_info.bbs;
info.insts += func_info.insts;
info.clones += func_info.clones;
info.weight += func_info.weight;
} else {
info.weight += 1;
}
}
return info;
}
struct Partition {
StringMap<bool> globals;
StringMap<unsigned> fvars;
StringMap<unsigned> gvars;
size_t weight;
};
static bool canPartition(const GlobalValue &G) {
if (auto F = dyn_cast<Function>(&G)) {
if (F->hasFnAttribute(Attribute::AlwaysInline))
return false;
}
return true;
}
static inline bool verify_partitioning(const SmallVectorImpl<Partition> &partitions, const Module &M, size_t fvars_size, size_t gvars_size) {
bool bad = false;
#ifndef JL_NDEBUG
SmallVector<uint32_t, 0> fvars(fvars_size);
SmallVector<uint32_t, 0> gvars(gvars_size);
StringMap<uint32_t> GVNames;
for (uint32_t i = 0; i < partitions.size(); i++) {
for (auto &name : partitions[i].globals) {
if (GVNames.count(name.getKey())) {
bad = true;
dbgs() << "Duplicate global name " << name.getKey() << " in partitions " << i << " and " << GVNames[name.getKey()] << "\n";
}
GVNames[name.getKey()] = i;
}
for (auto &fvar : partitions[i].fvars) {
if (fvars[fvar.second] != 0) {
bad = true;
dbgs() << "Duplicate fvar " << fvar.first() << " in partitions " << i << " and " << fvars[fvar.second] - 1 << "\n";
}
fvars[fvar.second] = i+1;
}
for (auto &gvar : partitions[i].gvars) {
if (gvars[gvar.second] != 0) {
bad = true;
dbgs() << "Duplicate gvar " << gvar.first() << " in partitions " << i << " and " << gvars[gvar.second] - 1 << "\n";
}
gvars[gvar.second] = i+1;
}
}
for (auto &GV : M.global_values()) {
if (GV.isDeclaration()) {
if (GVNames.count(GV.getName())) {
bad = true;
dbgs() << "Global " << GV.getName() << " is a declaration but is in partition " << GVNames[GV.getName()] << "\n";
}
} else {
// Local global values are not partitioned
if (!canPartition(GV)) {
if (GVNames.count(GV.getName())) {
bad = true;
dbgs() << "Shouldn't have partitioned " << GV.getName() << ", but is in partition " << GVNames[GV.getName()] << "\n";
}
continue;
}
if (!GVNames.count(GV.getName())) {
bad = true;
dbgs() << "Global " << GV << " not in any partition\n";
}
for (ConstantUses<GlobalValue> uses(const_cast<GlobalValue*>(&GV), const_cast<Module&>(M)); !uses.done(); uses.next()) {
auto val = uses.get_info().val;
if (!GVNames.count(val->getName())) {
bad = true;
dbgs() << "Global " << val->getName() << " used by " << GV.getName() << ", which is not in any partition\n";
continue;
}
if (GVNames[val->getName()] != GVNames[GV.getName()]) {
bad = true;
dbgs() << "Global " << val->getName() << " used by " << GV.getName() << ", which is in partition " << GVNames[GV.getName()] << " but " << val->getName() << " is in partition " << GVNames[val->getName()] << "\n";
}
}
}
}
for (uint32_t i = 0; i < fvars_size; i++) {
if (fvars[i] == 0) {
bad = true;
dbgs() << "fvar " << i << " not in any partition\n";
}
}
for (uint32_t i = 0; i < gvars_size; i++) {
if (gvars[i] == 0) {
bad = true;
dbgs() << "gvar " << i << " not in any partition\n";
}
}
#endif
return !bad;
}
// Chop a module up as equally as possible by weight into threads partitions
static SmallVector<Partition, 32> partitionModule(Module &M, unsigned threads) {
//Start by stripping fvars and gvars, which helpfully removes their uses as well
DenseMap<GlobalValue *, unsigned> fvars, gvars;
get_fvars_gvars(M, fvars, gvars);
// Partition by union-find, since we only have def->use traversal right now
struct Partitioner {
struct Node {
GlobalValue *GV;
unsigned parent;
unsigned size;
size_t weight;
};
SmallVector<Node, 0> nodes;
DenseMap<GlobalValue *, unsigned> node_map;
unsigned merged;
unsigned make(GlobalValue *GV, size_t weight) {
unsigned idx = nodes.size();
nodes.push_back({GV, idx, 1, weight});
node_map[GV] = idx;
return idx;
}
unsigned find(unsigned idx) {
while (nodes[idx].parent != idx) {
nodes[idx].parent = nodes[nodes[idx].parent].parent;
idx = nodes[idx].parent;
}
return idx;
}
unsigned merge(unsigned x, unsigned y) {
x = find(x);
y = find(y);
if (x == y)
return x;
if (nodes[x].size < nodes[y].size)
std::swap(x, y);
nodes[y].parent = x;
nodes[x].size += nodes[y].size;
nodes[x].weight += nodes[y].weight;
merged++;
return x;
}
};
Partitioner partitioner;
for (auto &G : M.global_values()) {
if (G.isDeclaration())
continue;
if (!canPartition(G))
continue;
// Currently ccallable global aliases have extern linkage, we only want to make the
// internally linked functions/global variables extern+hidden
if (G.hasLocalLinkage()) {
G.setLinkage(GlobalValue::ExternalLinkage);
G.setVisibility(GlobalValue::HiddenVisibility);
}
if (auto F = dyn_cast<Function>(&G)) {
partitioner.make(&G, getFunctionWeight(*F).weight);
} else {
partitioner.make(&G, 1);
}
}
// Merge all uses to go together into the same partition
for (unsigned i = 0; i < partitioner.nodes.size(); ++i) {
for (ConstantUses<GlobalValue> uses(partitioner.nodes[i].GV, M); !uses.done(); uses.next()) {
auto val = uses.get_info().val;
auto idx = partitioner.node_map.find(val);
// This can fail if we can't partition a global, but it uses something we can partition
// This should be fixed by altering canPartition to not permit partitioning this global
assert(idx != partitioner.node_map.end());
partitioner.merge(i, idx->second);
}
}
SmallVector<Partition, 32> partitions(threads);
// always get the smallest partition first
auto pcomp = [](const Partition *p1, const Partition *p2) {
return p1->weight > p2->weight;
};
std::priority_queue<Partition *, SmallVector<Partition *, 0>, decltype(pcomp)> pq(pcomp);
for (unsigned i = 0; i < threads; ++i) {
pq.push(&partitions[i]);
}
SmallVector<unsigned, 0> idxs(partitioner.nodes.size());
std::iota(idxs.begin(), idxs.end(), 0);
std::sort(idxs.begin(), idxs.end(), [&](unsigned a, unsigned b) {
//because roots have more weight than their children,
//we can sort by weight and get the roots first
return partitioner.nodes[a].weight > partitioner.nodes[b].weight;
});
// Assign the root of each partition to a partition, then assign its children to the same one
for (unsigned idx = 0; idx < idxs.size(); ++idx) {
auto i = idxs[idx];
auto root = partitioner.find(i);
assert(root == i || partitioner.nodes[root].weight == 0);
if (partitioner.nodes[root].weight) {
auto &node = partitioner.nodes[root];
auto &P = *pq.top();
pq.pop();
auto name = node.GV->getName();
P.globals.insert({name, true});
if (fvars.count(node.GV))
P.fvars[name] = fvars[node.GV];
if (gvars.count(node.GV))
P.gvars[name] = gvars[node.GV];
P.weight += node.weight;
node.weight = 0;
node.size = &P - partitions.data();
pq.push(&P);
}
if (root != i) {
auto &node = partitioner.nodes[i];
assert(node.weight != 0);
// we assigned its root already, so just add it to the root's partition
// don't touch the priority queue, since we're not changing the weight
auto &P = partitions[partitioner.nodes[root].size];
auto name = node.GV->getName();
P.globals.insert({name, true});
if (fvars.count(node.GV))
P.fvars[name] = fvars[node.GV];
if (gvars.count(node.GV))
P.gvars[name] = gvars[node.GV];
node.weight = 0;
node.size = partitioner.nodes[root].size;
}
}
bool verified = verify_partitioning(partitions, M, fvars.size(), gvars.size());
assert(verified && "Partitioning failed to partition globals correctly");
(void) verified;
return partitions;
}
struct ImageTimer {
uint64_t elapsed = 0;
std::string name;
std::string desc;
void startTimer() {
elapsed = jl_hrtime();
}
void stopTimer() {
elapsed = jl_hrtime() - elapsed;
}
void init(const Twine &name, const Twine &desc) {
this->name = name.str();
this->desc = desc.str();
}
operator bool() const {
return elapsed != 0;
}
void print(raw_ostream &out, bool clear=false) {
if (!*this)
return;
out << llvm::formatv("{0:F3} ", elapsed / 1e9) << name << " " << desc << "\n";
if (clear)
elapsed = 0;