Coverage Report

Created: 2017-11-21 03:47

/Users/buildslave/jenkins/workspace/clang-stage2-coverage-R/llvm/tools/polly/lib/CodeGen/IslNodeBuilder.cpp
Line
Count
Source (jump to first uncovered line)
1
//===- IslNodeBuilder.cpp - Translate an isl AST into a LLVM-IR AST -------===//
2
//
3
//                     The LLVM Compiler Infrastructure
4
//
5
// This file is distributed under the University of Illinois Open Source
6
// License. See LICENSE.TXT for details.
7
//
8
//===----------------------------------------------------------------------===//
9
//
10
// This file contains the IslNodeBuilder, a class to translate an isl AST into
11
// a LLVM-IR AST.
12
//
13
//===----------------------------------------------------------------------===//
14
15
#include "polly/CodeGen/IslNodeBuilder.h"
16
#include "polly/CodeGen/BlockGenerators.h"
17
#include "polly/CodeGen/CodeGeneration.h"
18
#include "polly/CodeGen/IslAst.h"
19
#include "polly/CodeGen/IslExprBuilder.h"
20
#include "polly/CodeGen/LoopGenerators.h"
21
#include "polly/CodeGen/RuntimeDebugBuilder.h"
22
#include "polly/Config/config.h"
23
#include "polly/Options.h"
24
#include "polly/ScopInfo.h"
25
#include "polly/Support/GICHelper.h"
26
#include "polly/Support/SCEVValidator.h"
27
#include "polly/Support/ScopHelper.h"
28
#include "llvm/ADT/APInt.h"
29
#include "llvm/ADT/PostOrderIterator.h"
30
#include "llvm/ADT/SetVector.h"
31
#include "llvm/ADT/SmallPtrSet.h"
32
#include "llvm/ADT/Statistic.h"
33
#include "llvm/Analysis/LoopInfo.h"
34
#include "llvm/Analysis/RegionInfo.h"
35
#include "llvm/Analysis/ScalarEvolution.h"
36
#include "llvm/Analysis/ScalarEvolutionExpressions.h"
37
#include "llvm/IR/BasicBlock.h"
38
#include "llvm/IR/Constant.h"
39
#include "llvm/IR/Constants.h"
40
#include "llvm/IR/DataLayout.h"
41
#include "llvm/IR/DerivedTypes.h"
42
#include "llvm/IR/Dominators.h"
43
#include "llvm/IR/Function.h"
44
#include "llvm/IR/InstrTypes.h"
45
#include "llvm/IR/Instruction.h"
46
#include "llvm/IR/Instructions.h"
47
#include "llvm/IR/Type.h"
48
#include "llvm/IR/Value.h"
49
#include "llvm/Support/Casting.h"
50
#include "llvm/Support/CommandLine.h"
51
#include "llvm/Support/ErrorHandling.h"
52
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
53
#include "isl/aff.h"
54
#include "isl/aff_type.h"
55
#include "isl/ast.h"
56
#include "isl/ast_build.h"
57
#include "isl/isl-noexceptions.h"
58
#include "isl/map.h"
59
#include "isl/set.h"
60
#include "isl/union_map.h"
61
#include "isl/union_set.h"
62
#include "isl/val.h"
63
#include <algorithm>
64
#include <cassert>
65
#include <cstdint>
66
#include <cstring>
67
#include <string>
68
#include <utility>
69
#include <vector>
70
71
using namespace llvm;
72
using namespace polly;
73
74
#define DEBUG_TYPE "polly-codegen"
75
76
STATISTIC(VersionedScops, "Number of SCoPs that required versioning.");
77
78
STATISTIC(SequentialLoops, "Number of generated sequential for-loops");
79
STATISTIC(ParallelLoops, "Number of generated parallel for-loops");
80
STATISTIC(VectorLoops, "Number of generated vector for-loops");
81
STATISTIC(IfConditions, "Number of generated if-conditions");
82
83
static cl::opt<bool> PollyGenerateRTCPrint(
84
    "polly-codegen-emit-rtc-print",
85
    cl::desc("Emit code that prints the runtime check result dynamically."),
86
    cl::Hidden, cl::init(false), cl::ZeroOrMore, cl::cat(PollyCategory));
87
88
// If this option is set we always use the isl AST generator to regenerate
89
// memory accesses. Without this option set we regenerate expressions using the
90
// original SCEV expressions and only generate new expressions in case the
91
// access relation has been changed and consequently must be regenerated.
92
static cl::opt<bool> PollyGenerateExpressions(
93
    "polly-codegen-generate-expressions",
94
    cl::desc("Generate AST expressions for unmodified and modified accesses"),
95
    cl::Hidden, cl::init(false), cl::ZeroOrMore, cl::cat(PollyCategory));
96
97
static cl::opt<int> PollyTargetFirstLevelCacheLineSize(
98
    "polly-target-first-level-cache-line-size",
99
    cl::desc("The size of the first level cache line size specified in bytes."),
100
    cl::Hidden, cl::init(64), cl::ZeroOrMore, cl::cat(PollyCategory));
101
102
__isl_give isl_ast_expr *
103
IslNodeBuilder::getUpperBound(__isl_keep isl_ast_node *For,
104
298
                              ICmpInst::Predicate &Predicate) {
105
298
  isl_id *UBID, *IteratorID;
106
298
  isl_ast_expr *Cond, *Iterator, *UB, *Arg0;
107
298
  isl_ast_op_type Type;
108
298
109
298
  Cond = isl_ast_node_for_get_cond(For);
110
298
  Iterator = isl_ast_node_for_get_iterator(For);
111
298
  assert(isl_ast_expr_get_type(Cond) == isl_ast_expr_op &&
112
298
         "conditional expression is not an atomic upper bound");
113
298
114
298
  Type = isl_ast_expr_get_op_type(Cond);
115
298
116
298
  switch (Type) {
117
298
  case isl_ast_op_le:
118
217
    Predicate = ICmpInst::ICMP_SLE;
119
217
    break;
120
298
  case isl_ast_op_lt:
121
81
    Predicate = ICmpInst::ICMP_SLT;
122
81
    break;
123
298
  default:
124
0
    llvm_unreachable("Unexpected comparison type in loop condition");
125
298
  }
126
298
127
298
  Arg0 = isl_ast_expr_get_op_arg(Cond, 0);
128
298
129
298
  assert(isl_ast_expr_get_type(Arg0) == isl_ast_expr_id &&
130
298
         "conditional expression is not an atomic upper bound");
131
298
132
298
  UBID = isl_ast_expr_get_id(Arg0);
133
298
134
298
  assert(isl_ast_expr_get_type(Iterator) == isl_ast_expr_id &&
135
298
         "Could not get the iterator");
136
298
137
298
  IteratorID = isl_ast_expr_get_id(Iterator);
138
298
139
298
  assert(UBID == IteratorID &&
140
298
         "conditional expression is not an atomic upper bound");
141
298
142
298
  UB = isl_ast_expr_get_op_arg(Cond, 1);
143
298
144
298
  isl_ast_expr_free(Cond);
145
298
  isl_ast_expr_free(Iterator);
146
298
  isl_ast_expr_free(Arg0);
147
298
  isl_id_free(IteratorID);
148
298
  isl_id_free(UBID);
149
298
150
298
  return UB;
151
298
}
152
153
/// Return true if a return value of Predicate is true for the value represented
154
/// by passed isl_ast_expr_int.
155
static bool checkIslAstExprInt(__isl_take isl_ast_expr *Expr,
156
42
                               isl_bool (*Predicate)(__isl_keep isl_val *)) {
157
42
  if (isl_ast_expr_get_type(Expr) != isl_ast_expr_int) {
158
0
    isl_ast_expr_free(Expr);
159
0
    return false;
160
0
  }
161
42
  auto ExprVal = isl_ast_expr_get_val(Expr);
162
42
  isl_ast_expr_free(Expr);
163
42
  if (Predicate(ExprVal) != isl_bool_true) {
164
0
    isl_val_free(ExprVal);
165
0
    return false;
166
0
  }
167
42
  isl_val_free(ExprVal);
168
42
  return true;
169
42
}
170
171
22
int IslNodeBuilder::getNumberOfIterations(__isl_keep isl_ast_node *For) {
172
22
  assert(isl_ast_node_get_type(For) == isl_ast_node_for);
173
22
  auto Body = isl_ast_node_for_get_body(For);
174
22
175
22
  // First, check if we can actually handle this code.
176
22
  switch (isl_ast_node_get_type(Body)) {
177
22
  case isl_ast_node_user:
178
20
    break;
179
22
  case isl_ast_node_block: {
180
2
    isl_ast_node_list *List = isl_ast_node_block_get_children(Body);
181
4
    for (int i = 0; i < isl_ast_node_list_n_ast_node(List); 
++i2
) {
182
3
      isl_ast_node *Node = isl_ast_node_list_get_ast_node(List, i);
183
3
      int Type = isl_ast_node_get_type(Node);
184
3
      isl_ast_node_free(Node);
185
3
      if (Type != isl_ast_node_user) {
186
1
        isl_ast_node_list_free(List);
187
1
        isl_ast_node_free(Body);
188
1
        return -1;
189
1
      }
190
3
    }
191
2
    isl_ast_node_list_free(List);
192
1
    break;
193
2
  }
194
2
  default:
195
0
    isl_ast_node_free(Body);
196
0
    return -1;
197
21
  }
198
21
  isl_ast_node_free(Body);
199
21
200
21
  auto Init = isl_ast_node_for_get_init(For);
201
21
  if (!checkIslAstExprInt(Init, isl_val_is_zero))
202
0
    return -1;
203
21
  auto Inc = isl_ast_node_for_get_inc(For);
204
21
  if (!checkIslAstExprInt(Inc, isl_val_is_one))
205
0
    return -1;
206
21
  CmpInst::Predicate Predicate;
207
21
  auto UB = getUpperBound(For, Predicate);
208
21
  if (isl_ast_expr_get_type(UB) != isl_ast_expr_int) {
209
2
    isl_ast_expr_free(UB);
210
2
    return -1;
211
2
  }
212
19
  auto UpVal = isl_ast_expr_get_val(UB);
213
19
  isl_ast_expr_free(UB);
214
19
  int NumberIterations = isl_val_get_num_si(UpVal);
215
19
  isl_val_free(UpVal);
216
19
  if (NumberIterations < 0)
217
0
    return -1;
218
19
  if (Predicate == CmpInst::ICMP_SLT)
219
0
    return NumberIterations;
220
19
  else
221
19
    return NumberIterations + 1;
222
22
}
223
224
/// Extract the values and SCEVs needed to generate code for a block.
225
static int findReferencesInBlock(struct SubtreeReferences &References,
226
29
                                 const ScopStmt *Stmt, BasicBlock *BB) {
227
192
  for (Instruction &Inst : *BB) {
228
192
    // Include invariant loads
229
192
    if (isa<LoadInst>(Inst))
230
21
      if (Value *InvariantLoad = References.GlobalMap.lookup(&Inst))
231
7
        References.Values.insert(InvariantLoad);
232
192
233
365
    for (Value *SrcVal : Inst.operands()) {
234
365
      auto *Scope = References.LI.getLoopFor(BB);
235
365
      if (canSynthesize(SrcVal, References.S, &References.SE, Scope)) {
236
264
        References.SCEVs.insert(References.SE.getSCEVAtScope(SrcVal, Scope));
237
264
        continue;
238
264
      } else 
if (Value *101
NewVal101
= References.GlobalMap.lookup(SrcVal))
239
6
        References.Values.insert(NewVal);
240
365
    }
241
192
  }
242
29
  return 0;
243
29
}
244
245
isl_stat addReferencesFromStmt(const ScopStmt *Stmt, void *UserPtr,
246
29
                               bool CreateScalarRefs) {
247
29
  auto &References = *static_cast<struct SubtreeReferences *>(UserPtr);
248
29
249
29
  if (Stmt->isBlockStmt())
250
29
    findReferencesInBlock(References, Stmt, Stmt->getBasicBlock());
251
0
  else {
252
0
    assert(Stmt->isRegionStmt() &&
253
0
           "Stmt was neither block nor region statement");
254
0
    for (BasicBlock *BB : Stmt->getRegion()->blocks())
255
0
      findReferencesInBlock(References, Stmt, BB);
256
0
  }
257
29
258
46
  for (auto &Access : *Stmt) {
259
46
    if (References.ParamSpace) {
260
0
      isl::space ParamSpace = Access->getLatestAccessRelation().get_space();
261
0
      (*References.ParamSpace) =
262
0
          References.ParamSpace->align_params(ParamSpace);
263
0
    }
264
46
265
46
    if (Access->isLatestArrayKind()) {
266
44
      auto *BasePtr = Access->getLatestScopArrayInfo()->getBasePtr();
267
44
      if (Instruction *OpInst = dyn_cast<Instruction>(BasePtr))
268
11
        if (Stmt->getParent()->contains(OpInst))
269
4
          continue;
270
40
271
40
      References.Values.insert(BasePtr);
272
40
      continue;
273
40
    }
274
2
275
2
    if (CreateScalarRefs)
276
2
      References.Values.insert(References.BlockGen.getOrCreateAlloca(*Access));
277
46
  }
278
29
279
29
  return isl_stat_ok;
280
29
}
281
282
/// Extract the out-of-scop values and SCEVs referenced from a set describing
283
/// a ScopStmt.
284
///
285
/// This includes the SCEVUnknowns referenced by the SCEVs used in the
286
/// statement and the base pointers of the memory accesses. For scalar
287
/// statements we force the generation of alloca memory locations and list
288
/// these locations in the set of out-of-scop values as well.
289
///
290
/// @param Set     A set which references the ScopStmt we are interested in.
291
/// @param UserPtr A void pointer that can be casted to a SubtreeReferences
292
///                structure.
293
static isl_stat addReferencesFromStmtSet(__isl_take isl_set *Set,
294
29
                                         void *UserPtr) {
295
29
  isl_id *Id = isl_set_get_tuple_id(Set);
296
29
  auto *Stmt = static_cast<const ScopStmt *>(isl_id_get_user(Id));
297
29
  isl_id_free(Id);
298
29
  isl_set_free(Set);
299
29
  return addReferencesFromStmt(Stmt, UserPtr);
300
29
}
301
302
/// Extract the out-of-scop values and SCEVs referenced from a union set
303
/// referencing multiple ScopStmts.
304
///
305
/// This includes the SCEVUnknowns referenced by the SCEVs used in the
306
/// statement and the base pointers of the memory accesses. For scalar
307
/// statements we force the generation of alloca memory locations and list
308
/// these locations in the set of out-of-scop values as well.
309
///
310
/// @param USet       A union set referencing the ScopStmts we are interested
311
///                   in.
312
/// @param References The SubtreeReferences data structure through which
313
///                   results are returned and further information is
314
///                   provided.
315
static void
316
addReferencesFromStmtUnionSet(isl_union_set *USet,
317
28
                              struct SubtreeReferences &References) {
318
28
  isl_union_set_foreach_set(USet, addReferencesFromStmtSet, &References);
319
28
  isl_union_set_free(USet);
320
28
}
321
322
__isl_give isl_union_map *
323
47
IslNodeBuilder::getScheduleForAstNode(__isl_keep isl_ast_node *For) {
324
47
  return IslAstInfo::getSchedule(For);
325
47
}
326
327
void IslNodeBuilder::getReferencesInSubtree(__isl_keep isl_ast_node *For,
328
                                            SetVector<Value *> &Values,
329
28
                                            SetVector<const Loop *> &Loops) {
330
28
  SetVector<const SCEV *> SCEVs;
331
28
  struct SubtreeReferences References = {
332
28
      LI, SE, S, ValueMap, Values, SCEVs, getBlockGenerator(), nullptr};
333
28
334
28
  for (const auto &I : IDToValue)
335
28
    
Values.insert(I.second)21
;
336
28
337
28
  // NOTE: this is populated in IslNodeBuilder::addParameters
338
28
  for (const auto &I : OutsideLoopIterations)
339
28
    
Values.insert(cast<SCEVUnknown>(I.second)->getValue())2
;
340
28
341
28
  isl_union_set *Schedule = isl_union_map_domain(getScheduleForAstNode(For));
342
28
  addReferencesFromStmtUnionSet(Schedule, References);
343
28
344
197
  for (const SCEV *Expr : SCEVs) {
345
197
    findValues(Expr, SE, Values);
346
197
    findLoops(Expr, Loops);
347
197
  }
348
28
349
73
  Values.remove_if([](const Value *V) { return isa<GlobalValue>(V); });
350
28
351
28
  /// Note: Code generation of induction variables of loops outside Scops
352
28
  ///
353
28
  /// Remove loops that contain the scop or that are part of the scop, as they
354
28
  /// are considered local. This leaves only loops that are before the scop, but
355
28
  /// do not contain the scop itself.
356
28
  /// We ignore loops perfectly contained in the Scop because these are already
357
28
  /// generated at `IslNodeBuilder::addParameters`. These `Loops` are loops
358
28
  /// whose induction variables are referred to by the Scop, but the Scop is not
359
28
  /// fully contained in these Loops. Since there can be many of these,
360
28
  /// we choose to codegen these on-demand.
361
28
  /// @see IslNodeBuilder::materializeNonScopLoopInductionVariable.
362
42
  Loops.remove_if([this](const Loop *L) {
363
42
    return S.contains(L) || 
L->contains(S.getEntry())3
;
364
42
  });
365
28
366
28
  // Contains Values that may need to be replaced with other values
367
28
  // due to replacements from the ValueMap. We should make sure
368
28
  // that we return correctly remapped values.
369
28
  // NOTE: this code path is tested by:
370
28
  //     1.  test/Isl/CodeGen/OpenMP/single_loop_with_loop_invariant_baseptr.ll
371
28
  //     2.  test/Isl/CodeGen/OpenMP/loop-body-references-outer-values-3.ll
372
28
  SetVector<Value *> ReplacedValues;
373
66
  for (Value *V : Values) {
374
66
    ReplacedValues.insert(getLatestValue(V));
375
66
  }
376
28
  Values = ReplacedValues;
377
28
}
378
379
28
void IslNodeBuilder::updateValues(ValueMapT &NewValues) {
380
28
  SmallPtrSet<Value *, 5> Inserted;
381
28
382
28
  for (const auto &I : IDToValue) {
383
21
    IDToValue[I.first] = NewValues[I.second];
384
21
    Inserted.insert(I.second);
385
21
  }
386
28
387
64
  for (const auto &I : NewValues) {
388
64
    if (Inserted.count(I.first))
389
0
      continue;
390
64
391
64
    ValueMap[I.first] = I.second;
392
64
  }
393
28
}
394
395
66
Value *IslNodeBuilder::getLatestValue(Value *Original) const {
396
66
  auto It = ValueMap.find(Original);
397
66
  if (It == ValueMap.end())
398
63
    return Original;
399
3
  return It->second;
400
3
}
401
402
void IslNodeBuilder::createUserVector(__isl_take isl_ast_node *User,
403
                                      std::vector<Value *> &IVS,
404
                                      __isl_take isl_id *IteratorID,
405
20
                                      __isl_take isl_union_map *Schedule) {
406
20
  isl_ast_expr *Expr = isl_ast_node_user_get_expr(User);
407
20
  isl_ast_expr *StmtExpr = isl_ast_expr_get_op_arg(Expr, 0);
408
20
  isl_id *Id = isl_ast_expr_get_id(StmtExpr);
409
20
  isl_ast_expr_free(StmtExpr);
410
20
  ScopStmt *Stmt = (ScopStmt *)isl_id_get_user(Id);
411
20
  std::vector<LoopToScevMapT> VLTS(IVS.size());
412
20
413
20
  isl_union_set *Domain = isl_union_set_from_set(Stmt->getDomain().release());
414
20
  Schedule = isl_union_map_intersect_domain(Schedule, Domain);
415
20
  isl_map *S = isl_map_from_union_map(Schedule);
416
20
417
20
  auto *NewAccesses = createNewAccesses(Stmt, User);
418
20
  createSubstitutionsVector(Expr, Stmt, VLTS, IVS, IteratorID);
419
20
  VectorBlockGenerator::generate(BlockGen, *Stmt, VLTS, S, NewAccesses);
420
20
  isl_id_to_ast_expr_free(NewAccesses);
421
20
  isl_map_free(S);
422
20
  isl_id_free(Id);
423
20
  isl_ast_node_free(User);
424
20
}
425
426
22
void IslNodeBuilder::createMark(__isl_take isl_ast_node *Node) {
427
22
  auto *Id = isl_ast_node_mark_get_id(Node);
428
22
  auto Child = isl_ast_node_mark_get_node(Node);
429
22
  isl_ast_node_free(Node);
430
22
  // If a child node of a 'SIMD mark' is a loop that has a single iteration,
431
22
  // it will be optimized away and we should skip it.
432
22
  if (strcmp(isl_id_get_name(Id), "SIMD") == 0 &&
433
22
      
isl_ast_node_get_type(Child) == isl_ast_node_for2
) {
434
2
    bool Vector = PollyVectorizerChoice == VECTORIZER_POLLY;
435
2
    int VectorWidth = getNumberOfIterations(Child);
436
2
    if (Vector && 1 < VectorWidth && VectorWidth <= 16)
437
2
      createForVector(Child, VectorWidth);
438
0
    else
439
0
      createForSequential(Child, true);
440
2
    isl_id_free(Id);
441
2
    return;
442
2
  }
443
20
  if (strcmp(isl_id_get_name(Id), "Inter iteration alias-free") == 0) {
444
2
    auto *BasePtr = static_cast<Value *>(isl_id_get_user(Id));
445
2
    Annotator.addInterIterationAliasFreeBasePtr(BasePtr);
446
2
  }
447
22
  create(Child);
448
22
  isl_id_free(Id);
449
22
}
450
451
void IslNodeBuilder::createForVector(__isl_take isl_ast_node *For,
452
19
                                     int VectorWidth) {
453
19
  isl_ast_node *Body = isl_ast_node_for_get_body(For);
454
19
  isl_ast_expr *Init = isl_ast_node_for_get_init(For);
455
19
  isl_ast_expr *Inc = isl_ast_node_for_get_inc(For);
456
19
  isl_ast_expr *Iterator = isl_ast_node_for_get_iterator(For);
457
19
  isl_id *IteratorID = isl_ast_expr_get_id(Iterator);
458
19
459
19
  Value *ValueLB = ExprBuilder.create(Init);
460
19
  Value *ValueInc = ExprBuilder.create(Inc);
461
19
462
19
  Type *MaxType = ExprBuilder.getType(Iterator);
463
19
  MaxType = ExprBuilder.getWidestType(MaxType, ValueLB->getType());
464
19
  MaxType = ExprBuilder.getWidestType(MaxType, ValueInc->getType());
465
19
466
19
  if (MaxType != ValueLB->getType())
467
0
    ValueLB = Builder.CreateSExt(ValueLB, MaxType);
468
19
  if (MaxType != ValueInc->getType())
469
0
    ValueInc = Builder.CreateSExt(ValueInc, MaxType);
470
19
471
19
  std::vector<Value *> IVS(VectorWidth);
472
19
  IVS[0] = ValueLB;
473
19
474
119
  for (int i = 1; i < VectorWidth; 
i++100
)
475
100
    IVS[i] = Builder.CreateAdd(IVS[i - 1], ValueInc, "p_vector_iv");
476
19
477
19
  isl_union_map *Schedule = getScheduleForAstNode(For);
478
19
  assert(Schedule && "For statement annotation does not contain its schedule");
479
19
480
19
  IDToValue[IteratorID] = ValueLB;
481
19
482
19
  switch (isl_ast_node_get_type(Body)) {
483
19
  case isl_ast_node_user:
484
18
    createUserVector(Body, IVS, isl_id_copy(IteratorID),
485
18
                     isl_union_map_copy(Schedule));
486
18
    break;
487
19
  case isl_ast_node_block: {
488
1
    isl_ast_node_list *List = isl_ast_node_block_get_children(Body);
489
1
490
3
    for (int i = 0; i < isl_ast_node_list_n_ast_node(List); 
++i2
)
491
2
      createUserVector(isl_ast_node_list_get_ast_node(List, i), IVS,
492
2
                       isl_id_copy(IteratorID), isl_union_map_copy(Schedule));
493
1
494
1
    isl_ast_node_free(Body);
495
1
    isl_ast_node_list_free(List);
496
1
    break;
497
19
  }
498
19
  default:
499
0
    isl_ast_node_dump(Body);
500
0
    llvm_unreachable("Unhandled isl_ast_node in vectorizer");
501
19
  }
502
19
503
19
  IDToValue.erase(IDToValue.find(IteratorID));
504
19
  isl_id_free(IteratorID);
505
19
  isl_union_map_free(Schedule);
506
19
507
19
  isl_ast_node_free(For);
508
19
  isl_ast_expr_free(Iterator);
509
19
510
19
  VectorLoops++;
511
19
}
512
513
/// Restore the initial ordering of dimensions of the band node
514
///
515
/// In case the band node represents all the dimensions of the iteration
516
/// domain, recreate the band node to restore the initial ordering of the
517
/// dimensions.
518
///
519
/// @param Node The band node to be modified.
520
/// @return The modified schedule node.
521
249
static bool IsLoopVectorizerDisabled(isl::ast_node Node) {
522
249
  assert(isl_ast_node_get_type(Node.keep()) == isl_ast_node_for);
523
249
  auto Body = Node.for_get_body();
524
249
  if (isl_ast_node_get_type(Body.keep()) != isl_ast_node_mark)
525
242
    return false;
526
7
  auto Id = Body.mark_get_id();
527
7
  if (strcmp(Id.get_name().c_str(), "Loop Vectorizer Disabled") == 0)
528
2
    return true;
529
5
  return false;
530
5
}
531
532
void IslNodeBuilder::createForSequential(__isl_take isl_ast_node *For,
533
249
                                         bool MarkParallel) {
534
249
  isl_ast_node *Body;
535
249
  isl_ast_expr *Init, *Inc, *Iterator, *UB;
536
249
  isl_id *IteratorID;
537
249
  Value *ValueLB, *ValueUB, *ValueInc;
538
249
  Type *MaxType;
539
249
  BasicBlock *ExitBlock;
540
249
  Value *IV;
541
249
  CmpInst::Predicate Predicate;
542
249
543
249
  bool LoopVectorizerDisabled =
544
249
      IsLoopVectorizerDisabled(isl::manage(isl_ast_node_copy(For)));
545
249
546
249
  Body = isl_ast_node_for_get_body(For);
547
249
548
249
  // isl_ast_node_for_is_degenerate(For)
549
249
  //
550
249
  // TODO: For degenerated loops we could generate a plain assignment.
551
249
  //       However, for now we just reuse the logic for normal loops, which will
552
249
  //       create a loop with a single iteration.
553
249
554
249
  Init = isl_ast_node_for_get_init(For);
555
249
  Inc = isl_ast_node_for_get_inc(For);
556
249
  Iterator = isl_ast_node_for_get_iterator(For);
557
249
  IteratorID = isl_ast_expr_get_id(Iterator);
558
249
  UB = getUpperBound(For, Predicate);
559
249
560
249
  ValueLB = ExprBuilder.create(Init);
561
249
  ValueUB = ExprBuilder.create(UB);
562
249
  ValueInc = ExprBuilder.create(Inc);
563
249
564
249
  MaxType = ExprBuilder.getType(Iterator);
565
249
  MaxType = ExprBuilder.getWidestType(MaxType, ValueLB->getType());
566
249
  MaxType = ExprBuilder.getWidestType(MaxType, ValueUB->getType());
567
249
  MaxType = ExprBuilder.getWidestType(MaxType, ValueInc->getType());
568
249
569
249
  if (MaxType != ValueLB->getType())
570
0
    ValueLB = Builder.CreateSExt(ValueLB, MaxType);
571
249
  if (MaxType != ValueUB->getType())
572
35
    ValueUB = Builder.CreateSExt(ValueUB, MaxType);
573
249
  if (MaxType != ValueInc->getType())
574
0
    ValueInc = Builder.CreateSExt(ValueInc, MaxType);
575
249
576
249
  // If we can show that LB <Predicate> UB holds at least once, we can
577
249
  // omit the GuardBB in front of the loop.
578
249
  bool UseGuardBB =
579
249
      !SE.isKnownPredicate(Predicate, SE.getSCEV(ValueLB), SE.getSCEV(ValueUB));
580
249
  IV = createLoop(ValueLB, ValueUB, ValueInc, Builder, LI, DT, ExitBlock,
581
249
                  Predicate, &Annotator, MarkParallel, UseGuardBB,
582
249
                  LoopVectorizerDisabled);
583
249
  IDToValue[IteratorID] = IV;
584
249
585
249
  create(Body);
586
249
587
249
  Annotator.popLoop(MarkParallel);
588
249
589
249
  IDToValue.erase(IDToValue.find(IteratorID));
590
249
591
249
  Builder.SetInsertPoint(&ExitBlock->front());
592
249
593
249
  isl_ast_node_free(For);
594
249
  isl_ast_expr_free(Iterator);
595
249
  isl_id_free(IteratorID);
596
249
597
249
  SequentialLoops++;
598
249
}
599
600
/// Remove the BBs contained in a (sub)function from the dominator tree.
601
///
602
/// This function removes the basic blocks that are part of a subfunction from
603
/// the dominator tree. Specifically, when generating code it may happen that at
604
/// some point the code generation continues in a new sub-function (e.g., when
605
/// generating OpenMP code). The basic blocks that are created in this
606
/// sub-function are then still part of the dominator tree of the original
607
/// function, such that the dominator tree reaches over function boundaries.
608
/// This is not only incorrect, but also causes crashes. This function now
609
/// removes from the dominator tree all basic blocks that are dominated (and
610
/// consequently reachable) from the entry block of this (sub)function.
611
///
612
/// FIXME: A LLVM (function or region) pass should not touch anything outside of
613
/// the function/region it runs on. Hence, the pure need for this function shows
614
/// that we do not comply to this rule. At the moment, this does not cause any
615
/// issues, but we should be aware that such issues may appear. Unfortunately
616
/// the current LLVM pass infrastructure does not allow to make Polly a module
617
/// or call-graph pass to solve this issue, as such a pass would not have access
618
/// to the per-function analyses passes needed by Polly. A future pass manager
619
/// infrastructure is supposed to enable such kind of access possibly allowing
620
/// us to create a cleaner solution here.
621
///
622
/// FIXME: Instead of adding the dominance information and then dropping it
623
/// later on, we should try to just not add it in the first place. This requires
624
/// some careful testing to make sure this does not break in interaction with
625
/// the SCEVBuilder and SplitBlock which may rely on the dominator tree or
626
/// which may try to update it.
627
///
628
/// @param F The function which contains the BBs to removed.
629
/// @param DT The dominator tree from which to remove the BBs.
630
28
static void removeSubFuncFromDomTree(Function *F, DominatorTree &DT) {
631
28
  DomTreeNode *N = DT.getNode(&F->getEntryBlock());
632
28
  std::vector<BasicBlock *> Nodes;
633
28
634
28
  // We can only remove an element from the dominator tree, if all its children
635
28
  // have been removed. To ensure this we obtain the list of nodes to remove
636
28
  // using a post-order tree traversal.
637
297
  for (po_iterator<DomTreeNode *> I = po_begin(N), E = po_end(N); I != E; 
++I269
)
638
269
    Nodes.push_back(I->getBlock());
639
28
640
28
  for (BasicBlock *BB : Nodes)
641
269
    DT.eraseNode(BB);
642
28
}
643
644
28
void IslNodeBuilder::createForParallel(__isl_take isl_ast_node *For) {
645
28
  isl_ast_node *Body;
646
28
  isl_ast_expr *Init, *Inc, *Iterator, *UB;
647
28
  isl_id *IteratorID;
648
28
  Value *ValueLB, *ValueUB, *ValueInc;
649
28
  Type *MaxType;
650
28
  Value *IV;
651
28
  CmpInst::Predicate Predicate;
652
28
653
28
  // The preamble of parallel code interacts different than normal code with
654
28
  // e.g., scalar initialization. Therefore, we ensure the parallel code is
655
28
  // separated from the last basic block.
656
28
  BasicBlock *ParBB = SplitBlock(Builder.GetInsertBlock(),
657
28
                                 &*Builder.GetInsertPoint(), &DT, &LI);
658
28
  ParBB->setName("polly.parallel.for");
659
28
  Builder.SetInsertPoint(&ParBB->front());
660
28
661
28
  Body = isl_ast_node_for_get_body(For);
662
28
  Init = isl_ast_node_for_get_init(For);
663
28
  Inc = isl_ast_node_for_get_inc(For);
664
28
  Iterator = isl_ast_node_for_get_iterator(For);
665
28
  IteratorID = isl_ast_expr_get_id(Iterator);
666
28
  UB = getUpperBound(For, Predicate);
667
28
668
28
  ValueLB = ExprBuilder.create(Init);
669
28
  ValueUB = ExprBuilder.create(UB);
670
28
  ValueInc = ExprBuilder.create(Inc);
671
28
672
28
  // OpenMP always uses SLE. In case the isl generated AST uses a SLT
673
28
  // expression, we need to adjust the loop bound by one.
674
28
  if (Predicate == CmpInst::ICMP_SLT)
675
8
    ValueUB = Builder.CreateAdd(
676
8
        ValueUB, Builder.CreateSExt(Builder.getTrue(), ValueUB->getType()));
677
28
678
28
  MaxType = ExprBuilder.getType(Iterator);
679
28
  MaxType = ExprBuilder.getWidestType(MaxType, ValueLB->getType());
680
28
  MaxType = ExprBuilder.getWidestType(MaxType, ValueUB->getType());
681
28
  MaxType = ExprBuilder.getWidestType(MaxType, ValueInc->getType());
682
28
683
28
  if (MaxType != ValueLB->getType())
684
0
    ValueLB = Builder.CreateSExt(ValueLB, MaxType);
685
28
  if (MaxType != ValueUB->getType())
686
4
    ValueUB = Builder.CreateSExt(ValueUB, MaxType);
687
28
  if (MaxType != ValueInc->getType())
688
0
    ValueInc = Builder.CreateSExt(ValueInc, MaxType);
689
28
690
28
  BasicBlock::iterator LoopBody;
691
28
692
28
  SetVector<Value *> SubtreeValues;
693
28
  SetVector<const Loop *> Loops;
694
28
695
28
  getReferencesInSubtree(For, SubtreeValues, Loops);
696
28
697
28
  // Create for all loops we depend on values that contain the current loop
698
28
  // iteration. These values are necessary to generate code for SCEVs that
699
28
  // depend on such loops. As a result we need to pass them to the subfunction.
700
28
  // See [Code generation of induction variables of loops outside Scops]
701
28
  for (const Loop *L : Loops) {
702
1
    Value *LoopInductionVar = materializeNonScopLoopInductionVariable(L);
703
1
    SubtreeValues.insert(LoopInductionVar);
704
1
  }
705
28
706
28
  ValueMapT NewValues;
707
28
  ParallelLoopGenerator ParallelLoopGen(Builder, LI, DT, DL);
708
28
709
28
  IV = ParallelLoopGen.createParallelLoop(ValueLB, ValueUB, ValueInc,
710
28
                                          SubtreeValues, NewValues, &LoopBody);
711
28
  BasicBlock::iterator AfterLoop = Builder.GetInsertPoint();
712
28
  Builder.SetInsertPoint(&*LoopBody);
713
28
714
28
  // Remember the parallel subfunction
715
28
  ParallelSubfunctions.push_back(LoopBody->getFunction());
716
28
717
28
  // Save the current values.
718
28
  auto ValueMapCopy = ValueMap;
719
28
  IslExprBuilder::IDToValueTy IDToValueCopy = IDToValue;
720
28
721
28
  updateValues(NewValues);
722
28
  IDToValue[IteratorID] = IV;
723
28
724
28
  ValueMapT NewValuesReverse;
725
28
726
28
  for (auto P : NewValues)
727
64
    NewValuesReverse[P.second] = P.first;
728
28
729
28
  Annotator.addAlternativeAliasBases(NewValuesReverse);
730
28
731
28
  create(Body);
732
28
733
28
  Annotator.resetAlternativeAliasBases();
734
28
  // Restore the original values.
735
28
  ValueMap = ValueMapCopy;
736
28
  IDToValue = IDToValueCopy;
737
28
738
28
  Builder.SetInsertPoint(&*AfterLoop);
739
28
  removeSubFuncFromDomTree((*LoopBody).getParent()->getParent(), DT);
740
28
741
28
  for (const Loop *L : Loops)
742
28
    
OutsideLoopIterations.erase(L)1
;
743
28
744
28
  isl_ast_node_free(For);
745
28
  isl_ast_expr_free(Iterator);
746
28
  isl_id_free(IteratorID);
747
28
748
28
  ParallelLoops++;
749
28
}
750
751
/// Return whether any of @p Node's statements contain partial accesses.
752
///
753
/// Partial accesses are not supported by Polly's vector code generator.
754
17
static bool hasPartialAccesses(__isl_take isl_ast_node *Node) {
755
17
  return isl_ast_node_foreach_descendant_top_down(
756
17
             Node,
757
36
             [](isl_ast_node *Node, void *User) -> isl_bool {
758
36
               if (isl_ast_node_get_type(Node) != isl_ast_node_user)
759
18
                 return isl_bool_true;
760
18
761
18
               isl::ast_expr Expr = give(isl_ast_node_user_get_expr(Node));
762
18
               isl::ast_expr StmtExpr =
763
18
                   give(isl_ast_expr_get_op_arg(Expr.keep(), 0));
764
18
               isl::id Id = give(isl_ast_expr_get_id(StmtExpr.keep()));
765
18
766
18
               ScopStmt *Stmt =
767
18
                   static_cast<ScopStmt *>(isl_id_get_user(Id.keep()));
768
18
               isl::set StmtDom = Stmt->getDomain();
769
27
               for (auto *MA : *Stmt) {
770
27
                 if (MA->isLatestPartialAccess())
771
0
                   return isl_bool_error;
772
18
               }
773
18
               return isl_bool_true;
774
18
             },
775
17
             nullptr) == isl_stat_error;
776
17
}
777
778
294
void IslNodeBuilder::createFor(__isl_take isl_ast_node *For) {
779
294
  bool Vector = PollyVectorizerChoice == VECTORIZER_POLLY;
780
294
781
294
  if (Vector && 
IslAstInfo::isInnermostParallel(For)27
&&
782
294
      
!IslAstInfo::isReductionParallel(For)20
) {
783
20
    int VectorWidth = getNumberOfIterations(For);
784
20
    if (1 < VectorWidth && 
VectorWidth <= 1617
&&
!hasPartialAccesses(For)17
) {
785
17
      createForVector(For, VectorWidth);
786
17
      return;
787
17
    }
788
277
  }
789
277
790
277
  if (IslAstInfo::isExecutedInParallel(For)) {
791
28
    createForParallel(For);
792
28
    return;
793
28
  }
794
249
  bool Parallel =
795
249
      (IslAstInfo::isParallel(For) && 
!IslAstInfo::isReductionParallel(For)16
);
796
294
  createForSequential(For, Parallel);
797
294
}
798
799
68
void IslNodeBuilder::createIf(__isl_take isl_ast_node *If) {
800
68
  isl_ast_expr *Cond = isl_ast_node_if_get_cond(If);
801
68
802
68
  Function *F = Builder.GetInsertBlock()->getParent();
803
68
  LLVMContext &Context = F->getContext();
804
68
805
68
  BasicBlock *CondBB = SplitBlock(Builder.GetInsertBlock(),
806
68
                                  &*Builder.GetInsertPoint(), &DT, &LI);
807
68
  CondBB->setName("polly.cond");
808
68
  BasicBlock *MergeBB = SplitBlock(CondBB, &CondBB->front(), &DT, &LI);
809
68
  MergeBB->setName("polly.merge");
810
68
  BasicBlock *ThenBB = BasicBlock::Create(Context, "polly.then", F);
811
68
  BasicBlock *ElseBB = BasicBlock::Create(Context, "polly.else", F);
812
68
813
68
  DT.addNewBlock(ThenBB, CondBB);
814
68
  DT.addNewBlock(ElseBB, CondBB);
815
68
  DT.changeImmediateDominator(MergeBB, CondBB);
816
68
817
68
  Loop *L = LI.getLoopFor(CondBB);
818
68
  if (L) {
819
31
    L->addBasicBlockToLoop(ThenBB, LI);
820
31
    L->addBasicBlockToLoop(ElseBB, LI);
821
31
  }
822
68
823
68
  CondBB->getTerminator()->eraseFromParent();
824
68
825
68
  Builder.SetInsertPoint(CondBB);
826
68
  Value *Predicate = ExprBuilder.create(Cond);
827
68
  Builder.CreateCondBr(Predicate, ThenBB, ElseBB);
828
68
  Builder.SetInsertPoint(ThenBB);
829
68
  Builder.CreateBr(MergeBB);
830
68
  Builder.SetInsertPoint(ElseBB);
831
68
  Builder.CreateBr(MergeBB);
832
68
  Builder.SetInsertPoint(&ThenBB->front());
833
68
834
68
  create(isl_ast_node_if_get_then(If));
835
68
836
68
  Builder.SetInsertPoint(&ElseBB->front());
837
68
838
68
  if (isl_ast_node_if_has_else(If))
839
10
    create(isl_ast_node_if_get_else(If));
840
68
841
68
  Builder.SetInsertPoint(&MergeBB->front());
842
68
843
68
  isl_ast_node_free(If);
844
68
845
68
  IfConditions++;
846
68
}
847
848
__isl_give isl_id_to_ast_expr *
849
IslNodeBuilder::createNewAccesses(ScopStmt *Stmt,
850
453
                                  __isl_keep isl_ast_node *Node) {
851
453
  isl_id_to_ast_expr *NewAccesses =
852
453
      isl_id_to_ast_expr_alloc(Stmt->getParent()->getIslCtx().get(), 0);
853
453
854
453
  auto *Build = IslAstInfo::getBuild(Node);
855
453
  assert(Build && "Could not obtain isl_ast_build from user node");
856
453
  Stmt->setAstBuild(isl::manage(isl_ast_build_copy(Build)));
857
453
858
995
  for (auto *MA : *Stmt) {
859
995
    if (!MA->hasNewAccessRelation()) {
860
779
      if (PollyGenerateExpressions) {
861
2
        if (!MA->isAffine())
862
0
          continue;
863
2
        if (MA->getLatestScopArrayInfo()->getBasePtrOriginSAI())
864
0
          continue;
865
2
866
2
        auto *BasePtr =
867
2
            dyn_cast<Instruction>(MA->getLatestScopArrayInfo()->getBasePtr());
868
2
        if (BasePtr && 
Stmt->getParent()->getRegion().contains(BasePtr)0
)
869
0
          continue;
870
777
      } else {
871
777
        continue;
872
777
      }
873
218
    }
874
218
    assert(MA->isAffine() &&
875
218
           "Only affine memory accesses can be code generated");
876
218
877
218
    auto Schedule = isl_ast_build_get_schedule(Build);
878
218
879
#ifndef NDEBUG
880
    if (MA->isRead()) {
881
      auto Dom = Stmt->getDomain().release();
882
      auto SchedDom = isl_set_from_union_set(
883
          isl_union_map_domain(isl_union_map_copy(Schedule)));
884
      auto AccDom = isl_map_domain(MA->getAccessRelation().release());
885
      Dom = isl_set_intersect_params(Dom,
886
                                     Stmt->getParent()->getContext().release());
887
      SchedDom = isl_set_intersect_params(
888
          SchedDom, Stmt->getParent()->getContext().release());
889
      assert(isl_set_is_subset(SchedDom, AccDom) &&
890
             "Access relation not defined on full schedule domain");
891
      assert(isl_set_is_subset(Dom, AccDom) &&
892
             "Access relation not defined on full domain");
893
      isl_set_free(AccDom);
894
      isl_set_free(SchedDom);
895
      isl_set_free(Dom);
896
    }
897
#endif
898
899
218
    auto PWAccRel =
900
218
        MA->applyScheduleToAccessRelation(isl::manage(Schedule)).release();
901
218
902
218
    // isl cannot generate an index expression for access-nothing accesses.
903
218
    isl::set AccDomain =
904
218
        give(isl_pw_multi_aff_domain(isl_pw_multi_aff_copy(PWAccRel)));
905
218
    if (isl_set_is_empty(AccDomain.keep()) == isl_bool_true) {
906
2
      isl_pw_multi_aff_free(PWAccRel);
907
2
      continue;
908
2
    }
909
216
910
216
    auto AccessExpr = isl_ast_build_access_from_pw_multi_aff(Build, PWAccRel);
911
216
    NewAccesses =
912
216
        isl_id_to_ast_expr_set(NewAccesses, MA->getId().release(), AccessExpr);
913
216
  }
914
453
915
453
  return NewAccesses;
916
453
}
917
918
void IslNodeBuilder::createSubstitutions(__isl_take isl_ast_expr *Expr,
919
552
                                         ScopStmt *Stmt, LoopToScevMapT &LTS) {
920
552
  assert(isl_ast_expr_get_type(Expr) == isl_ast_expr_op &&
921
552
         "Expression of type 'op' expected");
922
552
  assert(isl_ast_expr_get_op_type(Expr) == isl_ast_op_call &&
923
552
         "Operation of type 'call' expected");
924
1.16k
  for (int i = 0; i < isl_ast_expr_get_op_n_arg(Expr) - 1; 
++i611
) {
925
611
    isl_ast_expr *SubExpr;
926
611
    Value *V;
927
611
928
611
    SubExpr = isl_ast_expr_get_op_arg(Expr, i + 1);
929
611
    V = ExprBuilder.create(SubExpr);
930
611
    ScalarEvolution *SE = Stmt->getParent()->getSE();
931
611
    LTS[Stmt->getLoopForDimension(i)] = SE->getUnknown(V);
932
611
  }
933
552
934
552
  isl_ast_expr_free(Expr);
935
552
}
936
937
void IslNodeBuilder::createSubstitutionsVector(
938
    __isl_take isl_ast_expr *Expr, ScopStmt *Stmt,
939
    std::vector<LoopToScevMapT> &VLTS, std::vector<Value *> &IVS,
940
20
    __isl_take isl_id *IteratorID) {
941
20
  int i = 0;
942
20
943
20
  Value *OldValue = IDToValue[IteratorID];
944
123
  for (Value *IV : IVS) {
945
123
    IDToValue[IteratorID] = IV;
946
123
    createSubstitutions(isl_ast_expr_copy(Expr), Stmt, VLTS[i]);
947
123
    i++;
948
123
  }
949
20
950
20
  IDToValue[IteratorID] = OldValue;
951
20
  isl_id_free(IteratorID);
952
20
  isl_ast_expr_free(Expr);
953
20
}
954
955
void IslNodeBuilder::generateCopyStmt(
956
4
    ScopStmt *Stmt, __isl_keep isl_id_to_ast_expr *NewAccesses) {
957
4
  assert(Stmt->size() == 2);
958
4
  auto ReadAccess = Stmt->begin();
959
4
  auto WriteAccess = ReadAccess++;
960
4
  assert((*ReadAccess)->isRead() && (*WriteAccess)->isMustWrite());
961
4
  assert((*ReadAccess)->getElementType() == (*WriteAccess)->getElementType() &&
962
4
         "Accesses use the same data type");
963
4
  assert((*ReadAccess)->isArrayKind() && (*WriteAccess)->isArrayKind());
964
4
  auto *AccessExpr =
965
4
      isl_id_to_ast_expr_get(NewAccesses, (*ReadAccess)->getId().release());
966
4
  auto *LoadValue = ExprBuilder.create(AccessExpr);
967
4
  AccessExpr =
968
4
      isl_id_to_ast_expr_get(NewAccesses, (*WriteAccess)->getId().release());
969
4
  auto *StoreAddr = ExprBuilder.createAccessAddress(AccessExpr);
970
4
  Builder.CreateStore(LoadValue, StoreAddr);
971
4
}
972
973
23
Value *IslNodeBuilder::materializeNonScopLoopInductionVariable(const Loop *L) {
974
23
  assert(OutsideLoopIterations.find(L) == OutsideLoopIterations.end() &&
975
23
         "trying to materialize loop induction variable twice");
976
23
  const SCEV *OuterLIV = SE.getAddRecExpr(SE.getUnknown(Builder.getInt64(0)),
977
23
                                          SE.getUnknown(Builder.getInt64(1)), L,
978
23
                                          SCEV::FlagAnyWrap);
979
23
  Value *V = generateSCEV(OuterLIV);
980
23
  OutsideLoopIterations[L] = SE.getUnknown(V);
981
23
  return V;
982
23
}
983
984
433
void IslNodeBuilder::createUser(__isl_take isl_ast_node *User) {
985
433
  LoopToScevMapT LTS;
986
433
  isl_id *Id;
987
433
  ScopStmt *Stmt;
988
433
989
433
  isl_ast_expr *Expr = isl_ast_node_user_get_expr(User);
990
433
  isl_ast_expr *StmtExpr = isl_ast_expr_get_op_arg(Expr, 0);
991
433
  Id = isl_ast_expr_get_id(StmtExpr);
992
433
  isl_ast_expr_free(StmtExpr);
993
433
994
433
  LTS.insert(OutsideLoopIterations.begin(), OutsideLoopIterations.end());
995
433
996
433
  Stmt = (ScopStmt *)isl_id_get_user(Id);
997
433
  auto *NewAccesses = createNewAccesses(Stmt, User);
998
433
  if (Stmt->isCopyStmt()) {
999
4
    generateCopyStmt(Stmt, NewAccesses);
1000
4
    isl_ast_expr_free(Expr);
1001
429
  } else {
1002
429
    createSubstitutions(Expr, Stmt, LTS);
1003
429
1004
429
    if (Stmt->isBlockStmt())
1005
394
      BlockGen.copyStmt(*Stmt, LTS, NewAccesses);
1006
35
    else
1007
35
      RegionGen.copyStmt(*Stmt, LTS, NewAccesses);
1008
429
  }
1009
433
1010
433
  isl_id_to_ast_expr_free(NewAccesses);
1011
433
  isl_ast_node_free(User);
1012
433
  isl_id_free(Id);
1013
433
}
1014
1015
115
void IslNodeBuilder::createBlock(__isl_take isl_ast_node *Block) {
1016
115
  isl_ast_node_list *List = isl_ast_node_block_get_children(Block);
1017
115
1018
387
  for (int i = 0; i < isl_ast_node_list_n_ast_node(List); 
++i272
)
1019
272
    create(isl_ast_node_list_get_ast_node(List, i));
1020
115
1021
115
  isl_ast_node_free(Block);
1022
115
  isl_ast_node_list_free(List);
1023
115
}
1024
1025
932
void IslNodeBuilder::create(__isl_take isl_ast_node *Node) {
1026
932
  switch (isl_ast_node_get_type(Node)) {
1027
932
  case isl_ast_node_error:
1028
0
    llvm_unreachable("code generation error");
1029
932
  case isl_ast_node_mark:
1030
22
    createMark(Node);
1031
22
    return;
1032
932
  case isl_ast_node_for:
1033
294
    createFor(Node);
1034
294
    return;
1035
932
  case isl_ast_node_if:
1036
68
    createIf(Node);
1037
68
    return;
1038
932
  case isl_ast_node_user:
1039
433
    createUser(Node);
1040
433
    return;
1041
932
  case isl_ast_node_block:
1042
115
    createBlock(Node);
1043
115
    return;
1044
0
  }
1045
0
1046
0
  llvm_unreachable("Unknown isl_ast_node type");
1047
0
}
1048
1049
242
bool IslNodeBuilder::materializeValue(isl_id *Id) {
1050
242
  // If the Id is already mapped, skip it.
1051
242
  if (!IDToValue.count(Id)) {
1052
170
    auto *ParamSCEV = (const SCEV *)isl_id_get_user(Id);
1053
170
    Value *V = nullptr;
1054
170
1055
170
    // Parameters could refer to invariant loads that need to be
1056
170
    // preloaded before we can generate code for the parameter. Thus,
1057
170
    // check if any value referred to in ParamSCEV is an invariant load
1058
170
    // and if so make sure its equivalence class is preloaded.
1059
170
    SetVector<Value *> Values;
1060
170
    findValues(ParamSCEV, SE, Values);
1061
176
    for (auto *Val : Values) {
1062
176
      // Check if the value is an instruction in a dead block within the SCoP
1063
176
      // and if so do not code generate it.
1064
176
      if (auto *Inst = dyn_cast<Instruction>(Val)) {
1065
35
        if (S.contains(Inst)) {
1066
16
          bool IsDead = true;
1067
16
1068
16
          // Check for "undef" loads first, then if there is a statement for
1069
16
          // the parent of Inst and lastly if the parent of Inst has an empty
1070
16
          // domain. In the first and last case the instruction is dead but if
1071
16
          // there is a statement or the domain is not empty Inst is not dead.
1072
16
          auto MemInst = MemAccInst::dyn_cast(Inst);
1073
16
          auto Address = MemInst ? 
MemInst.getPointerOperand()15
:
nullptr1
;
1074
16
          if (Address && SE.getUnknown(UndefValue::get(Address->getType())) ==
1075
14
                  SE.getPointerBase(SE.getSCEV(Address))) {
1076
16
          } else if (S.getStmtFor(Inst)) {
1077
1
            IsDead = false;
1078
15
          } else {
1079
15
            auto *Domain = S.getDomainConditions(Inst->getParent()).release();
1080
15
            IsDead = isl_set_is_empty(Domain);
1081
15
            isl_set_free(Domain);
1082
15
          }
1083
16
1084
16
          if (IsDead) {
1085
2
            V = UndefValue::get(ParamSCEV->getType());
1086
2
            break;
1087
2
          }
1088
174
        }
1089
35
      }
1090
174
1091
174
      if (auto *IAClass = S.lookupInvariantEquivClass(Val)) {
1092
12
        // Check if this invariant access class is empty, hence if we never
1093
12
        // actually added a loads instruction to it. In that case it has no
1094
12
        // (meaningful) users and we should not try to code generate it.
1095
12
        if (IAClass->InvariantAccesses.empty())
1096
0
          V = UndefValue::get(ParamSCEV->getType());
1097
12
1098
12
        if (!preloadInvariantEquivClass(*IAClass)) {
1099
3
          isl_id_free(Id);
1100
3
          return false;
1101
3
        }
1102
167
      }
1103
176
    }
1104
167
1105
167
    V = V ? 
V2
:
generateSCEV(ParamSCEV)165
;
1106
170
    IDToValue[Id] = V;
1107
170
  }
1108
242
1109
242
  isl_id_free(Id);
1110
239
  return true;
1111
242
}
1112
1113
132
bool IslNodeBuilder::materializeParameters(isl_set *Set) {
1114
341
  for (unsigned i = 0, e = isl_set_dim(Set, isl_dim_param); i < e; 
++i209
) {
1115
212
    if (!isl_set_involves_dims(Set, isl_dim_param, i, 1))
1116
155
      continue;
1117
57
    isl_id *Id = isl_set_get_dim_id(Set, isl_dim_param, i);
1118
57
    if (!materializeValue(Id))
1119
3
      return false;
1120
212
  }
1121
132
  
return true129
;
1122
132
}
1123
1124
285
bool IslNodeBuilder::materializeParameters() {
1125
285
  for (const SCEV *Param : S.parameters()) {
1126
185
    isl_id *Id = S.getIdForParam(Param).release();
1127
185
    if (!materializeValue(Id))
1128
0
      return false;
1129
285
  }
1130
285
  return true;
1131
285
}
1132
1133
/// Generate the computation of the size of the outermost dimension from the
1134
/// Fortran array descriptor (in this case, `@g_arr`). The final `%size`
1135
/// contains the size of the array.
1136
///
1137
/// %arrty = type { i8*, i64, i64, [3 x %desc.dimensionty] }
1138
/// %desc.dimensionty = type { i64, i64, i64 }
1139
/// @g_arr = global %arrty zeroinitializer, align 32
1140
/// ...
1141
/// %0 = load i64, i64* getelementptr inbounds
1142
///                       (%arrty, %arrty* @g_arr, i64 0, i32 3, i64 0, i32 2)
1143
/// %1 = load i64, i64* getelementptr inbounds
1144
///                      (%arrty, %arrty* @g_arr, i64 0, i32 3, i64 0, i32 1)
1145
/// %2 = sub nsw i64 %0, %1
1146
/// %size = add nsw i64 %2, 1
1147
static Value *buildFADOutermostDimensionLoad(Value *GlobalDescriptor,
1148
                                             PollyIRBuilder &Builder,
1149
2
                                             std::string ArrayName) {
1150
2
  assert(GlobalDescriptor && "invalid global descriptor given");
1151
2
1152
2
  Value *endIdx[4] = {Builder.getInt64(0), Builder.getInt32(3),
1153
2
                      Builder.getInt64(0), Builder.getInt32(2)};
1154
2
  Value *endPtr = Builder.CreateInBoundsGEP(GlobalDescriptor, endIdx,
1155
2
                                            ArrayName + "_end_ptr");
1156
2
  Value *end = Builder.CreateLoad(endPtr, ArrayName + "_end");
1157
2
1158
2
  Value *beginIdx[4] = {Builder.getInt64(0), Builder.getInt32(3),
1159
2
                        Builder.getInt64(0), Builder.getInt32(1)};
1160
2
  Value *beginPtr = Builder.CreateInBoundsGEP(GlobalDescriptor, beginIdx,
1161
2
                                              ArrayName + "_begin_ptr");
1162
2
  Value *begin = Builder.CreateLoad(beginPtr, ArrayName + "_begin");
1163
2
1164
2
  Value *size =
1165
2
      Builder.CreateNSWSub(end, begin, ArrayName + "_end_begin_delta");
1166
2
  Type *endType = dyn_cast<IntegerType>(end->getType());
1167
2
  assert(endType && "expected type of end to be integral");
1168
2
1169
2
  size = Builder.CreateNSWAdd(end,
1170
2
                              ConstantInt::get(endType, 1, /* signed = */ true),
1171
2
                              ArrayName + "_size");
1172
2
1173
2
  return size;
1174
2
}
1175
1176
285
bool IslNodeBuilder::materializeFortranArrayOutermostDimension() {
1177
610
  for (ScopArrayInfo *Array : S.arrays()) {
1178
610
    if (Array->getNumberOfDimensions() == 0)
1179
143
      continue;
1180
467
1181
467
    Value *FAD = Array->getFortranArrayDescriptor();
1182
467
    if (!FAD)
1183
465
      continue;
1184
2
1185
2
    isl_pw_aff *ParametricPwAff = Array->getDimensionSizePw(0).release();
1186
2
    assert(ParametricPwAff && "parametric pw_aff corresponding "
1187
2
                              "to outermost dimension does not "
1188
2
                              "exist");
1189
2
1190
2
    isl_id *Id = isl_pw_aff_get_dim_id(ParametricPwAff, isl_dim_param, 0);
1191
2
    isl_pw_aff_free(ParametricPwAff);
1192
2
1193
2
    assert(Id && "pw_aff is not parametric");
1194
2
1195
2
    if (IDToValue.count(Id)) {
1196
0
      isl_id_free(Id);
1197
0
      continue;
1198
0
    }
1199
2
1200
2
    Value *FinalValue =
1201
2
        buildFADOutermostDimensionLoad(FAD, Builder, Array->getName());
1202
2
    assert(FinalValue && "unable to build Fortran array "
1203
2
                         "descriptor load of outermost dimension");
1204
2
    IDToValue[Id] = FinalValue;
1205
2
    isl_id_free(Id);
1206
2
  }
1207
285
  return true;
1208
285
}
1209
1210
Value *IslNodeBuilder::preloadUnconditionally(isl_set *AccessRange,
1211
                                              isl_ast_build *Build,
1212
105
                                              Instruction *AccInst) {
1213
105
  isl_pw_multi_aff *PWAccRel = isl_pw_multi_aff_from_set(AccessRange);
1214
105
  isl_ast_expr *Access =
1215
105
      isl_ast_build_access_from_pw_multi_aff(Build, PWAccRel);
1216
105
  auto *Address = isl_ast_expr_address_of(Access);
1217
105
  auto *AddressValue = ExprBuilder.create(Address);
1218
105
  Value *PreloadVal;
1219
105
1220
105
  // Correct the type as the SAI might have a different type than the user
1221
105
  // expects, especially if the base pointer is a struct.
1222
105
  Type *Ty = AccInst->getType();
1223
105
1224
105
  auto *Ptr = AddressValue;
1225
105
  auto Name = Ptr->getName();
1226
105
  auto AS = Ptr->getType()->getPointerAddressSpace();
1227
105
  Ptr = Builder.CreatePointerCast(Ptr, Ty->getPointerTo(AS), Name + ".cast");
1228
105
  PreloadVal = Builder.CreateLoad(Ptr, Name + ".load");
1229
105
  if (LoadInst *PreloadInst = dyn_cast<LoadInst>(PreloadVal))
1230
105
    PreloadInst->setAlignment(dyn_cast<LoadInst>(AccInst)->getAlignment());
1231
105
1232
105
  // TODO: This is only a hot fix for SCoP sequences that use the same load
1233
105
  //       instruction contained and hoisted by one of the SCoPs.
1234
105
  if (SE.isSCEVable(Ty))
1235
90
    SE.forgetValue(AccInst);
1236
105
1237
105
  return PreloadVal;
1238
105
}
1239
1240
Value *IslNodeBuilder::preloadInvariantLoad(const MemoryAccess &MA,
1241
108
                                            isl_set *Domain) {
1242
108
  isl_set *AccessRange = isl_map_range(MA.getAddressFunction().release());
1243
108
  AccessRange = isl_set_gist_params(AccessRange, S.getContext().release());
1244
108
1245
108
  if (!materializeParameters(AccessRange)) {
1246
0
    isl_set_free(AccessRange);
1247
0
    isl_set_free(Domain);
1248
0
    return nullptr;
1249
0
  }
1250
108
1251
108
  auto *Build =
1252
108
      isl_ast_build_from_context(isl_set_universe(S.getParamSpace().release()));
1253
108
  isl_set *Universe = isl_set_universe(isl_set_get_space(Domain));
1254
108
  bool AlwaysExecuted = isl_set_is_equal(Domain, Universe);
1255
108
  isl_set_free(Universe);
1256
108
1257
108
  Instruction *AccInst = MA.getAccessInstruction();
1258
108
  Type *AccInstTy = AccInst->getType();
1259
108
1260
108
  Value *PreloadVal = nullptr;
1261
108
  if (AlwaysExecuted) {
1262
84
    PreloadVal = preloadUnconditionally(AccessRange, Build, AccInst);
1263
84
    isl_ast_build_free(Build);
1264
84
    isl_set_free(Domain);
1265
84
    return PreloadVal;
1266
84
  }
1267
24
1268
24
  if (!materializeParameters(Domain)) {
1269
3
    isl_ast_build_free(Build);
1270
3
    isl_set_free(AccessRange);
1271
3
    isl_set_free(Domain);
1272
3
    return nullptr;
1273
3
  }
1274
21
1275
21
  isl_ast_expr *DomainCond = isl_ast_build_expr_from_set(Build, Domain);
1276
21
  Domain = nullptr;
1277
21
1278
21
  ExprBuilder.setTrackOverflow(true);
1279
21
  Value *Cond = ExprBuilder.create(DomainCond);
1280
21
  Value *OverflowHappened = Builder.CreateNot(ExprBuilder.getOverflowState(),
1281
21
                                              "polly.preload.cond.overflown");
1282
21
  Cond = Builder.CreateAnd(Cond, OverflowHappened, "polly.preload.cond.result");
1283
21
  ExprBuilder.setTrackOverflow(false);
1284
21
1285
21
  if (!Cond->getType()->isIntegerTy(1))
1286
0
    Cond = Builder.CreateIsNotNull(Cond);
1287
21
1288
21
  BasicBlock *CondBB = SplitBlock(Builder.GetInsertBlock(),
1289
21
                                  &*Builder.GetInsertPoint(), &DT, &LI);
1290
21
  CondBB->setName("polly.preload.cond");
1291
21
1292
21
  BasicBlock *MergeBB = SplitBlock(CondBB, &CondBB->front(), &DT, &LI);
1293
21
  MergeBB->setName("polly.preload.merge");
1294
21
1295
21
  Function *F = Builder.GetInsertBlock()->getParent();
1296
21
  LLVMContext &Context = F->getContext();
1297
21
  BasicBlock *ExecBB = BasicBlock::Create(Context, "polly.preload.exec", F);
1298
21
1299
21
  DT.addNewBlock(ExecBB, CondBB);
1300
21
  if (Loop *L = LI.getLoopFor(CondBB))
1301
6
    L->addBasicBlockToLoop(ExecBB, LI);
1302
21
1303
21
  auto *CondBBTerminator = CondBB->getTerminator();
1304
21
  Builder.SetInsertPoint(CondBBTerminator);
1305
21
  Builder.CreateCondBr(Cond, ExecBB, MergeBB);
1306
21
  CondBBTerminator->eraseFromParent();
1307
21
1308
21
  Builder.SetInsertPoint(ExecBB);
1309
21
  Builder.CreateBr(MergeBB);
1310
21
1311
21
  Builder.SetInsertPoint(ExecBB->getTerminator());
1312
21
  Value *PreAccInst = preloadUnconditionally(AccessRange, Build, AccInst);
1313
21
  Builder.SetInsertPoint(MergeBB->getTerminator());
1314
21
  auto *MergePHI = Builder.CreatePHI(
1315
21
      AccInstTy, 2, "polly.preload." + AccInst->getName() + ".merge");
1316
21
  PreloadVal = MergePHI;
1317
21
1318
21
  if (!PreAccInst) {
1319
0
    PreloadVal = nullptr;
1320
0
    PreAccInst = UndefValue::get(AccInstTy);
1321
0
  }
1322
108
1323
108
  MergePHI->addIncoming(PreAccInst, ExecBB);
1324
108
  MergePHI->addIncoming(Constant::getNullValue(AccInstTy), CondBB);
1325
108
1326
108
  isl_ast_build_free(Build);
1327
108
  return PreloadVal;
1328
108
}
1329
1330
bool IslNodeBuilder::preloadInvariantEquivClass(
1331
147
    InvariantEquivClassTy &IAClass) {
1332
147
  // For an equivalence class of invariant loads we pre-load the representing
1333
147
  // element with the unified execution context. However, we have to map all
1334
147
  // elements of the class to the one preloaded load as they are referenced
1335
147
  // during the code generation and therefor need to be mapped.
1336
147
  const MemoryAccessList &MAs = IAClass.InvariantAccesses;
1337
147
  if (MAs.empty())
1338
2
    return true;
1339
145
1340
145
  MemoryAccess *MA = MAs.front();
1341
145
  assert(MA->isArrayKind() && MA->isRead());
1342
145
1343
145
  // If the access function was already mapped, the preload of this equivalence
1344
145
  // class was triggered earlier already and doesn't need to be done again.
1345
145
  if (ValueMap.count(MA->getAccessInstruction()))
1346
30
    return true;
1347
115
1348
115
  // Check for recursion which can be caused by additional constraints, e.g.,
1349
115
  // non-finite loop constraints. In such a case we have to bail out and insert
1350
115
  // a "false" runtime check that will cause the original code to be executed.
1351
115
  auto PtrId = std::make_pair(IAClass.IdentifyingPointer, IAClass.AccessType);
1352
115
  if (!PreloadedPtrs.insert(PtrId).second)
1353
4
    return false;
1354
111
1355
111
  // The execution context of the IAClass.
1356
111
  isl::set &ExecutionCtx = IAClass.ExecutionContext;
1357
111
1358
111
  // If the base pointer of this class is dependent on another one we have to
1359
111
  // make sure it was preloaded already.
1360
111
  auto *SAI = MA->getScopArrayInfo();
1361
111
  if (auto *BaseIAClass = S.lookupInvariantEquivClass(SAI->getBasePtr())) {
1362
22
    if (!preloadInvariantEquivClass(*BaseIAClass))
1363
1
      return false;
1364
21
1365
21
    // After we preloaded the BaseIAClass we adjusted the BaseExecutionCtx and
1366
21
    // we need to refine the ExecutionCtx.
1367
21
    isl::set BaseExecutionCtx = BaseIAClass->ExecutionContext;
1368
21
    ExecutionCtx = ExecutionCtx.intersect(BaseExecutionCtx);
1369
21
  }
1370
111
1371
111
  // If the size of a dimension is dependent on another class, make sure it is
1372
111
  // preloaded.
1373
111
  
for (unsigned i = 1, e = SAI->getNumberOfDimensions(); 110
i < e110
;
++i0
) {
1374
2
    const SCEV *Dim = SAI->getDimensionSize(i);
1375
2
    SetVector<Value *> Values;
1376
2
    findValues(Dim, SE, Values);
1377
2
    for (auto *Val : Values) {
1378
2
      if (auto *BaseIAClass = S.lookupInvariantEquivClass(Val)) {
1379
2
        if (!preloadInvariantEquivClass(*BaseIAClass))
1380
2
          return false;
1381
0
1382
0
        // After we preloaded the BaseIAClass we adjusted the BaseExecutionCtx
1383
0
        // and we need to refine the ExecutionCtx.
1384
0
        isl::set BaseExecutionCtx = BaseIAClass->ExecutionContext;
1385
0
        ExecutionCtx = ExecutionCtx.intersect(BaseExecutionCtx);
1386
0
      }
1387
2
    }
1388
2
  }
1389
110
1390
110
  Instruction *AccInst = MA->getAccessInstruction();
1391
108
  Type *AccInstTy = AccInst->getType();
1392
108
1393
108
  Value *PreloadVal = preloadInvariantLoad(*MA, ExecutionCtx.copy());
1394
108
  if (!PreloadVal)
1395
3
    return false;
1396
105
1397
115
  
for (const MemoryAccess *MA : MAs) 105
{
1398
115
    Instruction *MAAccInst = MA->getAccessInstruction();
1399
115
    assert(PreloadVal->getType() == MAAccInst->getType());
1400
115
    ValueMap[MAAccInst] = PreloadVal;
1401
115
  }
1402
105
1403
105
  if (SE.isSCEVable(AccInstTy)) {
1404
90
    isl_id *ParamId = S.getIdForParam(SE.getSCEV(AccInst)).release();
1405
90
    if (ParamId)
1406
21
      IDToValue[ParamId] = PreloadVal;
1407
90
    isl_id_free(ParamId);
1408
90
  }
1409
105
1410
105
  BasicBlock *EntryBB = &Builder.GetInsertBlock()->getParent()->getEntryBlock();
1411
105
  auto *Alloca = new AllocaInst(AccInstTy, DL.getAllocaAddrSpace(),
1412
105
                                AccInst->getName() + ".preload.s2a");
1413
105
  Alloca->insertBefore(&*EntryBB->getFirstInsertionPt());
1414
105
  Builder.CreateStore(PreloadVal, Alloca);
1415
105
  ValueMapT PreloadedPointer;
1416
105
  PreloadedPointer[PreloadVal] = AccInst;
1417
105
  Annotator.addAlternativeAliasBases(PreloadedPointer);
1418
105
1419
105
  for (auto *DerivedSAI : SAI->getDerivedSAIs()) {
1420
46
    Value *BasePtr = DerivedSAI->getBasePtr();
1421
46
1422
52
    for (const MemoryAccess *MA : MAs) {
1423
52
      // As the derived SAI information is quite coarse, any load from the
1424
52
      // current SAI could be the base pointer of the derived SAI, however we
1425
52
      // should only change the base pointer of the derived SAI if we actually
1426
52
      // preloaded it.
1427
52
      if (BasePtr == MA->getOriginalBaseAddr()) {
1428
0
        assert(BasePtr->getType() == PreloadVal->getType());
1429
0
        DerivedSAI->setBasePtr(PreloadVal);
1430
0
      }
1431
52
1432
52
      // For scalar derived SAIs we remap the alloca used for the derived value.
1433
52
      if (BasePtr == MA->getAccessInstruction())
1434
34
        ScalarMap[DerivedSAI] = Alloca;
1435
52
    }
1436
46
  }
1437
105
1438
115
  for (const MemoryAccess *MA : MAs) {
1439
115
    Instruction *MAAccInst = MA->getAccessInstruction();
1440
115
    // Use the escape system to get the correct value to users outside the SCoP.
1441
115
    BlockGenerator::EscapeUserVectorTy EscapeUsers;
1442
115
    for (auto *U : MAAccInst->users())
1443
120
      if (Instruction *UI = dyn_cast<Instruction>(U))
1444
120
        if (!S.contains(UI))
1445
7
          EscapeUsers.push_back(UI);
1446
115
1447
115
    if (EscapeUsers.empty())
1448
109
      continue;
1449
6
1450
6
    EscapeMap[MA->getAccessInstruction()] =
1451
6
        std::make_pair(Alloca, std::move(EscapeUsers));
1452
6
  }
1453
147
1454
147
  return true;
1455
147
}
1456
1457
289
void IslNodeBuilder::allocateNewArrays(BBPair StartExitBlocks) {
1458
623
  for (auto &SAI : S.arrays()) {
1459
623
    if (SAI->getBasePtr())
1460
613
      continue;
1461
10
1462
10
    assert(SAI->getNumberOfDimensions() > 0 && SAI->getDimensionSize(0) &&
1463
10
           "The size of the outermost dimension is used to declare newly "
1464
10
           "created arrays that require memory allocation.");
1465
10
1466
10
    Type *NewArrayType = nullptr;
1467
10
1468
10
    // Get the size of the array = size(dim_1)*...*size(dim_n)
1469
10
    uint64_t ArraySizeInt = 1;
1470
30
    for (int i = SAI->getNumberOfDimensions() - 1; i >= 0; 
i--20
) {
1471
20
      auto *DimSize = SAI->getDimensionSize(i);
1472
20
      unsigned UnsignedDimSize = static_cast<const SCEVConstant *>(DimSize)
1473
20
                                     ->getAPInt()
1474
20
                                     .getLimitedValue();
1475
20
1476
20
      if (!NewArrayType)
1477
10
        NewArrayType = SAI->getElementType();
1478
20
1479
20
      NewArrayType = ArrayType::get(NewArrayType, UnsignedDimSize);
1480
20
      ArraySizeInt *= UnsignedDimSize;
1481
20
    }
1482
10
1483
10
    if (SAI->isOnHeap()) {
1484
3
      LLVMContext &Ctx = NewArrayType->getContext();
1485
3
1486
3
      // Get the IntPtrTy from the Datalayout
1487
3
      auto IntPtrTy = DL.getIntPtrType(Ctx);
1488
3
1489
3
      // Get the size of the element type in bits
1490
3
      unsigned Size = SAI->getElemSizeInBytes();
1491
3
1492
3
      // Insert the malloc call at polly.start
1493
3
      auto InstIt = std::get<0>(StartExitBlocks)->getTerminator();
1494
3
      auto *CreatedArray = CallInst::CreateMalloc(
1495
3
          &*InstIt, IntPtrTy, SAI->getElementType(),
1496
3
          ConstantInt::get(Type::getInt64Ty(Ctx), Size),
1497
3
          ConstantInt::get(Type::getInt64Ty(Ctx), ArraySizeInt), nullptr,
1498
3
          SAI->getName());
1499
3
1500
3
      SAI->setBasePtr(CreatedArray);
1501
3
1502
3
      // Insert the free call at polly.exiting
1503
3
      CallInst::CreateFree(CreatedArray,
1504
3
                           std::get<1>(StartExitBlocks)->getTerminator());
1505
7
    } else {
1506
7
      auto InstIt = Builder.GetInsertBlock()
1507
7
                        ->getParent()
1508
7
                        ->getEntryBlock()
1509
7
                        .getTerminator();
1510
7
1511
7
      auto *CreatedArray = new AllocaInst(NewArrayType, DL.getAllocaAddrSpace(),
1512
7
                                          SAI->getName(), &*InstIt);
1513
7
      CreatedArray->setAlignment(PollyTargetFirstLevelCacheLineSize);
1514
7
      SAI->setBasePtr(CreatedArray);
1515
7
    }
1516
623
  }
1517
289
}
1518
1519
289
bool IslNodeBuilder::preloadInvariantLoads() {
1520
289
  auto &InvariantEquivClasses = S.getInvariantAccesses();
1521
289
  if (InvariantEquivClasses.empty())
1522
225
    return true;
1523
64
1524
64
  BasicBlock *PreLoadBB = SplitBlock(Builder.GetInsertBlock(),
1525
64
                                     &*Builder.GetInsertPoint(), &DT, &LI);
1526
64
  PreLoadBB->setName("polly.preload.begin");
1527
64
  Builder.SetInsertPoint(&PreLoadBB->front());
1528
64
1529
64
  for (auto &IAClass : InvariantEquivClasses)
1530
111
    if (!preloadInvariantEquivClass(IAClass))
1531
4
      return false;
1532
60
1533
60
  return true;
1534
60
}
1535
1536
285
void IslNodeBuilder::addParameters(__isl_take isl_set *Context) {
1537
285
  // Materialize values for the parameters of the SCoP.
1538
285
  materializeParameters();
1539
285
1540
285
  // materialize the outermost dimension parameters for a Fortran array.
1541
285
  // NOTE: materializeParameters() does not work since it looks through
1542
285
  // the SCEVs. We don't have a corresponding SCEV for the array size
1543
285
  // parameter
1544
285
  materializeFortranArrayOutermostDimension();
1545
285
1546
285
  // Generate values for the current loop iteration for all surrounding loops.
1547
285
  //
1548
285
  // We may also reference loops outside of the scop which do not contain the
1549
285
  // scop itself, but as the number of such scops may be arbitrarily large we do
1550
285
  // not generate code for them here, but only at the point of code generation
1551
285
  // where these values are needed.
1552
285
  Loop *L = LI.getLoopFor(S.getEntry());
1553
285
1554
501
  while (L != nullptr && 
S.contains(L)238
)
1555
285
    
L = L->getParentLoop()216
;
1556
285
1557
307
  while (L != nullptr) {
1558
22
    materializeNonScopLoopInductionVariable(L);
1559
22
    L = L->getParentLoop();
1560
22
  }
1561
285
1562
285
  isl_set_free(Context);
1563
285
}
1564
1565
188
Value *IslNodeBuilder::generateSCEV(const SCEV *Expr) {
1566
188
  /// We pass the insert location of our Builder, as Polly ensures during IR
1567
188
  /// generation that there is always a valid CFG into which instructions are
1568
188
  /// inserted. As a result, the insertpoint is known to be always followed by a
1569
188
  /// terminator instruction. This means the insert point may be specified by a
1570
188
  /// terminator instruction, but it can never point to an ->end() iterator
1571
188
  /// which does not have a corresponding instruction. Hence, dereferencing
1572
188
  /// the insertpoint to obtain an instruction is known to be save.
1573
188
  ///
1574
188
  /// We also do not need to update the Builder here, as new instructions are
1575
188
  /// always inserted _before_ the given InsertLocation. As a result, the
1576
188
  /// insert location remains valid.
1577
188
  assert(Builder.GetInsertBlock()->end() != Builder.GetInsertPoint() &&
1578
188
         "Insert location points after last valid instruction");
1579
188
  Instruction *InsertLocation = &*Builder.GetInsertPoint();
1580
188
  return expandCodeFor(S, SE, DL, "polly", Expr, Expr->getType(),
1581
188
                       InsertLocation, &ValueMap,
1582
188
                       StartBlock->getSinglePredecessor());
1583
188
}
1584
1585
/// The AST expression we generate to perform the run-time check assumes
1586
/// computations on integer types of infinite size. As we only use 64-bit
1587
/// arithmetic we check for overflows, in case of which we set the result
1588
/// of this run-time check to false to be conservatively correct,
1589
285
Value *IslNodeBuilder::createRTC(isl_ast_expr *Condition) {
1590
285
  auto ExprBuilder = getExprBuilder();
1591
285
1592
285
  // In case the AST expression has integers larger than 64 bit, bail out. The
1593
285
  // resulting LLVM-IR will contain operations on types that use more than 64
1594
285
  // bits. These are -- in case wrapping intrinsics are used -- translated to
1595
285
  // runtime library calls that are not available on all systems (e.g., Android)
1596
285
  // and consequently will result in linker errors.
1597
285
  if (ExprBuilder.hasLargeInts(isl::manage(isl_ast_expr_copy(Condition)))) {
1598
7
    isl_ast_expr_free(Condition);
1599
7
    return Builder.getFalse();
1600
7
  }
1601
278
1602
278
  ExprBuilder.setTrackOverflow(true);
1603
278
  Value *RTC = ExprBuilder.create(Condition);
1604
278
  if (!RTC->getType()->isIntegerTy(1))
1605
170
    RTC = Builder.CreateIsNotNull(RTC);
1606
278
  Value *OverflowHappened =
1607
278
      Builder.CreateNot(ExprBuilder.getOverflowState(), "polly.rtc.overflown");
1608
278
1609
278
  if (PollyGenerateRTCPrint) {
1610
0
    auto *F = Builder.GetInsertBlock()->getParent();
1611
0
    RuntimeDebugBuilder::createCPUPrinter(
1612
0
        Builder,
1613
0
        "F: " + F->getName().str() + " R: " + S.getRegion().getNameStr() +
1614
0
            "RTC: ",
1615
0
        RTC, " Overflow: ", OverflowHappened,
1616
0
        "\n"
1617
0
        "  (0 failed, -1 succeeded)\n"
1618
0
        "  (if one or both are 0 falling back to original code, if both are -1 "
1619
0
        "executing Polly code)\n");
1620
0
  }
1621
278
1622
278
  RTC = Builder.CreateAnd(RTC, OverflowHappened, "polly.rtc.result");
1623
278
  ExprBuilder.setTrackOverflow(false);
1624
278
1625
278
  if (!isa<ConstantInt>(RTC))
1626
107
    VersionedScops++;
1627
285
1628
285
  return RTC;
1629
285
}