[Polly] [IslCodeGenerator] Add OpenMP support

[grosser]

Summary

This hooks the existing OpenMP support, as it was already available in
the CLooG backend, to our isl code generator backend.

This is the last feature, that was only available in the CLooG backend and that
blocked the by-default use of the isl backend. As the isl backend is with this
commit not only at feature parity, but has several advantages (run-time
alias checks, delinerization, ...), we aim to switch the default code generation
backend to isl. This switch will take place after a brief in-tree testing period
for the new OpenMP support in our isl backend.

TODO: - Still needs to be documented/style-checked

• We do not use -enable-polly-openmp yet
• We may want to reduce the use of OpenMP specific variable names

Comments are already appreciated.

[jdoerfert]

Reviewed half of it but have to go now

include/polly/CodeGen/IslExprBuilder.h

84

This change can/should be commited earlier, LGTM btw. but is there a particular reason for MapVector instead of SmallDenseMap or just DenseMap?

lib/CodeGen/IslCodeGeneration.cpp

106

I don't like the comment (style + content). It doesn't give any more information than the type of the map.

140

Please, do not use this awfull OMP any more but talk about parallel function arguments or sth. similar.

(also documentation)

141

When I see the type here I remember my patch to generalize the code generation (LoopGenerators). I hoped you would review that and work on that as a base. Any reasons not to?

178

isn't createForParallel much nicer?

299

I usually do not like this Object producing methodes, but prefere the object to be passed as reference instead, especially for Vectors/Sets/Maps/etc.

307

Any reason not to make this a static function? I'm not sure about such lambda's myself and I think we always used static functions so far.

309

I don't really see the benefit of static_cast here as User is a black box void*, however it might be good to start using static_cast instead of C-style casts all the time.

317

That is probably not going to work. The instruction might be in the SCoP but outside the parallel loop, thus still needed as an argument for the parallel subfunction.

If you agree, then I suggest to use the union_set_foreach_set only to collect all Stmts/BBs inside of For and then iterate again over them. This time we do the instruction operand checks but we check if the parent of an instruction operand is in the set of basic blocks we collected earlier. This way we should be able to identify all inner SCoP instructions we need to pass as arguments.

320

I would like the negated check better:

if (isa<Constant> || isa<Global>)

or

canSynthesise(...)

then continue, otherwise assume we need it.

335

Why back to std::set?

343

We could remove the need for the extra set by looking up I.first in IDToValue, if found we continue otherwise we go on.
Or even better we merge the two loops and use this lookup to decide what action to take. But now it looks to good to be true, what did I miss?

542

Back to stdlib...

543

This is hard to understand. If I got it correctly (but please tell me if I'm wrong), you collect all nodes of a function in post order in a set and erase them afterwards from the dominator tree of the function, right?

549

I stoped reviewing here, will continue tonight.

[jdoerfert]

I still haven't looked at the test cases but I commented the rest of the patch. We should add more test cases and review all of them later.

lib/CodeGen/IslCodeGeneration.cpp

560

remove the newline

563

I would call it It and ItId but that's not important.

608

I still do not know why we need this, shouldn't we just create the domtree on the fly (when we create the subfunc) correctly?

685

You already mentioned --enable-polly-openmp but I would like to rename that option when we are at it, e.g.,

--polly-enable-parallel

763

This needs some docu, I don't know why we need this here? Whats in which map?

953

Where is that one used? Was it missing before?

[grosser]

This new patch has the following changes:

• Addressed most of the review comments of Johannes
• Removed a couple of smaller changes that have been committed individually
• Extracted the changes to the IslAst into http://reviews.llvm.org/D6142
• Full support for SCEV based code generation
• Tested on LNT with and without SCEV based code generation
  • 1 failure with SCEV based code generation due to llvm.org/PR21204
  • 2 failures due to alias information being invalidated by openmp code generation

[grosser]

Sorry for the delay. I just pushed out a new patch which aims to address
your comments.

[jdoerfert]

Some "final" comments. Maybe you want to address some of them. If LNT looks good I would be fine with this patch.

lib/CodeGen/IslCodeGeneration.cpp

111

one more / please

151

What loop are left?

178

docu?

314

The name is not really descriptive.

362

I'm still not really convinced this is better/nicer than a static function.

366

I don't understand this. Can you add a comment, otherwise this is a magic condition that cannot be maintained or debuged.

372

I usually like the SmallPtrSet but that's taste I guess.

564

What about a flag to the loopgenerator such that it will not add these blocks to the DT at all (or just omit the DT argument)?

860

I don't understand the comment? e.g., What do you mean by ahead of time? When do we generate the references then?

lib/Support/SCEVValidator.cpp

365

There is another type of "SCEVVisitor", Sebastian pointed it out to me once. IT only has one function something like:

visit(const SCEV *S);

and you cast the SCEV yourself. I would suggest using that for visitors that are only interested in one or two SCEV types.

431

Same as above

[simbuerg]

lib/CodeGen/IslCodeGeneration.cpp

91

out-of-scop

307

Extract the values and SCEVs needed to generate code for a ScopStmt.

311

... when generating code for this statement.

362

Most likely the static function would get a descriptive name, such as: isGlobalValue, right? That is quite redundant given the only statement this function executes is: return isa<GlobalValue>(V);

Furthermore, I do not think that this static function would get called from somewhere else in here, so whats the point in having it around on the top-level.

624

... generate code for SCEVs...

lib/Support/SCEVValidator.cpp

365

Nice find, that simplifies so many things

[grosser]

Adresses the latest review comments from Andreas and Johannes

[grosser]

In the previous patch, I addressed most reviewer comments.

Two points where I did not follow reviewer suggestions:

• I kept the small lambda function.
• I clean up the dominance tree at the end to stay in sync with CLooG (but added a FIXME to possibly improve on this later)

lib/CodeGen/IslCodeGeneration.cpp

91

Fixed.

111

Fixed.

151

I added:

"Loops that are before the scop, but do not contain the scop itself are considered not locally defined."

178

I added:

+ / Create LLVM-IR that executes a for node thread parallel.
+ /
+ /// @param For The FOR isl_ast_node for which code is generated.

307

Fixed.

311

Fixed.

314

I forgot to add this to the patch, but I now renamed this to

void *UserPtr and struct FindValuesUser *User

362

I left the lambda for now.

Similar to Andreas I have the impression that this as good as we can find a use case for lambdas. Are you proposing to generally avoid lambdas?

366

This was already indirectly documented in the header of the surrounding function. For clarity I added a brief comment right above the stmt:

+ / Remove loops that contain the scop or that are part of the scop, as they
+ / are considered local. This leaves only loops that are before the scop, but
+ /// do not contain the scop itself.

372

Changed.

564

Interesting idea. This requires some work to understand if this is possible. I added a FIXME. For now I propose to use the same solution as we had in the CLooG backend.

624

Fixed.

860

Changed to:

+ We may also reference loops outside of the scop which do not contain the
+ scop itself, but as the number of such scops may be arbitrarily large we do
+ not generate code for them here, but only at the point of code generation
+ where these values are needed.

lib/Support/SCEVValidator.cpp

365

Changed. This really reduced the code needed here significantly.

431

Changed.

[jdoerfert]

You can get a looks good to me now, if LNT looks fine too I'd suggest you commit it.
After formating of course ;)

[grosser]

Committed in r222088

include/polly/CodeGen/IslExprBuilder.h

 //===-IslExprBuilder.h - Helper to generate code for isl AST expressions --===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef POLLY_ISL_EXPR_BUILDER_H
 #define POLLY_ISL_EXPR_BUILDER_H
 
 #include "polly/CodeGen/IRBuilder.h"
 
-#include "isl/ast.h"
+#include "llvm/ADT/MapVector.h"
 
-#include <map>
+#include "isl/ast.h"
 
 namespace llvm {
 class SCEVExpander;
 }
 
 namespace polly {
 
 /// @brief LLVM-IR generator for isl_ast_expr[essions]

 /// 2) We always flag computations with 'nsw'
 ///
 /// As isl_ast_expr[essions] assume arbitrary precision, no wrapping should
 /// ever occur in the generated LLVM-IR (assuming the data type chosen is large
 /// enough).
 class IslExprBuilder {
 public:
   /// @brief A map from isl_ids to llvm::Values.
-  typedef std::map<isl_id *, llvm::Value *> IDToValueTy;
+  typedef llvm::MapVector<isl_id *, llvm::Value *> IDToValueTy;
 
   /// @brief Construct an IslExprBuilder.
   ///
   /// @param Builder The IRBuilder used to construct the isl_ast_expr[ession].
   ///                The insert location of this IRBuilder defines WHERE the
   ///                corresponding LLVM-IR is generated.
   ///
   /// @param IDToValue The isl_ast_expr[ession] may reference parameters or
public:
   /// possible output values.
   ///
   /// @param Expr The expression for which to find the type.
   /// @return The type with which the expression should be computed.
   llvm::IntegerType *getType(__isl_keep isl_ast_expr *Expr);
 
 private:
   PollyIRBuilder &Builder;
-  std::map<isl_id *, llvm::Value *> &IDToValue;
+  IDToValueTy &IDToValue;
 
   /// @brief A SCEVExpander to translate dimension sizes to llvm values.
   llvm::SCEVExpander &Expander;
 
   llvm::Value *createOp(__isl_take isl_ast_expr *Expr);
   llvm::Value *createOpUnary(__isl_take isl_ast_expr *Expr);
   llvm::Value *createOpAccess(__isl_take isl_ast_expr *Expr);
   llvm::Value *createOpBin(__isl_take isl_ast_expr *Expr);

include/polly/Support/SCEVValidator.h

 //===--- SCEVValidator.h - Detect Scops -------------------------*- C++ -*-===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 // Checks if a SCEV expression represents a valid affine expression.
 //===----------------------------------------------------------------------===//
 
 #ifndef POLLY_SCEV_VALIDATOR_H
 #define POLLY_SCEV_VALIDATOR_H
 
+#include "llvm/ADT/SetVector.h"
 #include <vector>
 
 namespace llvm {
 class Region;
 class SCEV;
 class ScalarEvolution;
 class Value;
+class Loop;
 }
 
 namespace polly {
+/// @brief Find the loops referenced from a SCEV expression.
+///
+/// @param Expr The SCEV expression to scan for loops.
+/// @param Loops A vector into which the found loops are inserted.
+void findLoops(const llvm::SCEV *Expr,
+               llvm::SetVector<const llvm::Loop *> &Loops);
+
+/// @brief Find the values referenced by SCEVUnknowns in a given SCEV expression.
+///
+/// @param Expr The SCEV expression to scan for SCEVUnknowns.
+/// @param Expr A vector into which the found values are inserted.
+void findValues(const llvm::SCEV *Expr, llvm::SetVector<llvm::Value *> &Values);
+
 /// Returns true when the SCEV contains references to instructions within the
 /// region.
 ///
 /// @param S The SCEV to analyze.
 /// @param R The region in which we look for dependences.
 bool hasScalarDepsInsideRegion(const llvm::SCEV *S, const llvm::Region *R);
 bool isAffineExpr(const llvm::Region *R, const llvm::SCEV *Expression,
                   llvm::ScalarEvolution &SE,
                   const llvm::Value *BaseAddress = 0);
 std::vector<const llvm::SCEV *>
 getParamsInAffineExpr(const llvm::Region *R, const llvm::SCEV *Expression,
                       llvm::ScalarEvolution &SE,
                       const llvm::Value *BaseAddress = 0);
 }
 
 #endif

lib/CodeGen/IslCodeGeneration.cpp

 #include "polly/CodeGen/IslAst.h"
 #include "polly/CodeGen/LoopGenerators.h"
 #include "polly/CodeGen/Utils.h"
 #include "polly/Dependences.h"
 #include "polly/LinkAllPasses.h"
 #include "polly/ScopInfo.h"
 #include "polly/Support/GICHelper.h"
 #include "polly/Support/ScopHelper.h"
+#include "polly/Support/SCEVValidator.h"
 #include "polly/TempScopInfo.h"
+
+#include "llvm/ADT/PostOrderIterator.h"
+#include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/Analysis/LoopInfo.h"
 #include "llvm/Analysis/PostDominators.h"
 #include "llvm/Analysis/ScalarEvolutionExpander.h"
 #include "llvm/IR/Module.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/IR/DataLayout.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"

 using namespace polly;
 using namespace llvm;
 
 #define DEBUG_TYPE "polly-codegen-isl"
 
 class IslNodeBuilder {
 public:
   IslNodeBuilder(PollyIRBuilder &Builder, ScopAnnotator &Annotator, Pass *P,
-                 LoopInfo &LI, ScalarEvolution &SE, DominatorTree &DT)
-      : Builder(Builder), Annotator(Annotator),
+                 const DataLayout &DL, LoopInfo &LI, ScalarEvolution &SE,
+                 DominatorTree &DT, Scop &S)
+      : S(S), Builder(Builder), Annotator(Annotator),
         Rewriter(new SCEVExpander(SE, "polly")),
-        ExprBuilder(Builder, IDToValue, *Rewriter), P(P), LI(LI), SE(SE),
-        DT(DT) {}
+        ExprBuilder(Builder, IDToValue, *Rewriter), P(P), DL(DL), LI(LI),
+        SE(SE), DT(DT) {}
 
   ~IslNodeBuilder() { delete Rewriter; }
 
   void addParameters(__isl_take isl_set *Context);
   void create(__isl_take isl_ast_node *Node);
   IslExprBuilder &getExprBuilder() { return ExprBuilder; }
 
 private:
+  Scop &S;
   PollyIRBuilder &Builder;
   ScopAnnotator &Annotator;
 
   /// @brief A SCEVExpander to create llvm values from SCEVs.
   SCEVExpander *Rewriter;
 
   IslExprBuilder ExprBuilder;
   Pass *P;
+  const DataLayout &DL;
   LoopInfo &LI;
   ScalarEvolution &SE;
   DominatorTree &DT;
 
+  /// @brief The current iteration of out-of-scop loops
+  ///
+  /// This map provides for a given loop a llvm::Value that contains the current
+  /// loop iteration.
+  LoopToScevMapT OutsideLoopIterations;
+
   // This maps an isl_id* to the Value* it has in the generated program. For now
   // on, the only isl_ids that are stored here are the newly calculated loop
   // ivs.
   IslExprBuilder::IDToValueTy IDToValue;
 
   /// Generate code for a given SCEV*
   ///
   /// This function generates code for a given SCEV expression. It generated
   /// code is emmitted at the end of the basic block our Builder currently
   /// points to and the resulting value is returned.
   ///
   /// @param Expr The expression to code generate.
   Value *generateSCEV(const SCEV *Expr);
 
+  /// A set of Value -> Value remappings to apply when generating new code.
+  ///
+  /// When generating new code for a ScopStmt this map is used to map certain
+  /// llvm::Values to new llvm::Values.
+  ValueMapT ValueMap;
+
   // Extract the upper bound of this loop
   //
   // The isl code generation can generate arbitrary expressions to check if the
   // upper bound of a loop is reached, but it provides an option to enforce
   // 'atomic' upper bounds. An 'atomic upper bound is always of the form
   // iv <= expr, where expr is an (arbitrary) expression not containing iv.
   //
   // This function extracts 'atomic' upper bounds. Polly, in general, requires
   // atomic upper bounds for the following reasons:
   //
   // 1. An atomic upper bound is loop invariant
   //
   //    It must not be calculated at each loop iteration and can often even be
   //    hoisted out further by the loop invariant code motion.
   //
   // 2. OpenMP needs a loop invarient upper bound to calculate the number
   //    of loop iterations.
   //
   // 3. With the existing code, upper bounds have been easier to implement.
   __isl_give isl_ast_expr *getUpperBound(__isl_keep isl_ast_node *For,
                                          CmpInst::Predicate &Predicate);
 
   unsigned getNumberOfIterations(__isl_keep isl_ast_node *For);
 
+  /// Compute the values and loops referenced in this subtree.
+  ///
+  /// This function looks at all ScopStmts scheduled below the provided For node
+  /// and finds the llvm::Value[s] and llvm::Loops[s] which are referenced but
+  /// not locally defined.
+  ///
+  /// Values that can be synthesized or that are available as globals are
+  /// considered locally defined.
+  ///
+  /// Loops that contain the scop or that are part of the scop are considered
+  /// locally defined. Loops that are before the scop, but do not contain the
+  /// scop itself are considered not locally defined.
+  ///
+  /// @param For    The node defining the subtree.
+  /// @param Values A vector that will be filled with the Values referenced in
+  ///               this subtree.
+  /// @param Loops  A vector that will be filled with the Loops referenced in
+  ///               this subtree.
+  void getReferencesInSubtree(__isl_keep isl_ast_node *For,
+                              SetVector<Value *> &Values,
+                              SetVector<const Loop *> &Loops);
+
+  /// Change the llvm::Value(s) used for code generation.
+  ///
+  /// When generating code certain values (e.g., references to induction
+  /// variables or array base pointers) in the original code may be replaced by
+  /// new values. This function allows to (partially) update the set of values
+  /// used. A typical use case for this function is the case when we continue
+  /// code generation in a subfunction/kernel function and need to explicitly
+  /// pass down certain values.
+  ///
+  /// @param NewValues A map that maps certain llvm::Values to new llvm::Values.
+  void updateValues(ParallelLoopGenerator::ValueToValueMapTy &NewValues);
+
   void createFor(__isl_take isl_ast_node *For);
   void createForVector(__isl_take isl_ast_node *For, int VectorWidth);
   void createForSequential(__isl_take isl_ast_node *For);
 
+  /// Create LLVM-IR that executes a for node thread parallel.
+  ///
+  /// @param For The FOR isl_ast_node for which code is generated.
+  void createForParallel(__isl_take isl_ast_node *For);
+
   /// Generate LLVM-IR that computes the values of the original induction
   /// variables in function of the newly generated loop induction variables.
   ///
   /// Example:
   ///
   ///   // Original
   ///   for i
   ///     for j
unsigned IslNodeBuilder::getNumberOfIterations(__isl_keep isl_ast_node *For) {
   isl_union_map *Schedule = IslAstInfo::getSchedule(For);
   isl_set *LoopDomain = isl_set_from_union_set(isl_union_map_range(Schedule));
   int NumberOfIterations = polly::getNumberOfIterations(LoopDomain);
   if (NumberOfIterations == -1)
     return -1;
   return NumberOfIterations + 1;
 }
 
+struct FindValuesUser {
+  LoopInfo &LI;
+  ScalarEvolution &SE;
+  Region &R;
+  SetVector<Value *> &Values;
+  SetVector<const SCEV *> &SCEVs;
+};
+
+/// Extract the values and SCEVs needed to generate code for a ScopStmt.
+///
+/// This function extracts a ScopStmt from a given isl_set and computes the
+/// Values this statement depends on as well as a set of SCEV expressions that
+/// need to be synthesized when generating code for this statment.
+static int findValuesInStmt(isl_set *Set, void *User) {
+  isl_id *Id = isl_set_get_tuple_id(Set);
+  struct FindValuesUser *U = static_cast<struct FindValuesUser *>(User);
+  const ScopStmt *Stmt = static_cast<const ScopStmt *>(isl_id_get_user(Id));
+  const BasicBlock *BB = Stmt->getBasicBlock();
+
+  // Check all the operands of instructions in the basic block.
+  for (const Instruction &Inst : *BB) {
+    for (Value *SrcVal : Inst.operands()) {
+      if (Instruction *OpInst = dyn_cast<Instruction>(SrcVal))
+        if (canSynthesize(OpInst, &U->LI, &U->SE, &U->R)) {
+          U->SCEVs.insert(U->SE.getSCEVAtScope(OpInst, U->LI.getLoopFor(BB)));
+          continue;
+        }
+      if (Instruction *OpInst = dyn_cast<Instruction>(SrcVal))
+        if (Stmt->getParent()->getRegion().contains(OpInst))
+          continue;
+
+      if (isa<Instruction>(SrcVal) || isa<Argument>(SrcVal))
+        U->Values.insert(SrcVal);
+    }
+  }
+  isl_id_free(Id);
+  isl_set_free(Set);
+  return 0;
+}
+
+void IslNodeBuilder::getReferencesInSubtree(__isl_keep isl_ast_node *For,
+                                            SetVector<Value *> &Values,
+                                            SetVector<const Loop *> &Loops) {
+
+  SetVector<const SCEV *> SCEVs;
+  struct FindValuesUser FindValues = {LI, SE, S.getRegion(), Values, SCEVs};
+
+  for (const auto &I : IDToValue)
+    Values.insert(I.second);
+
+  for (const auto &I : OutsideLoopIterations)
+    Values.insert(cast<SCEVUnknown>(I.second)->getValue());
+
+  isl_union_set *Schedule = isl_union_map_domain(IslAstInfo::getSchedule(For));
+
+  isl_union_set_foreach_set(Schedule, findValuesInStmt, &FindValues);
+  isl_union_set_free(Schedule);
+
+  for (const SCEV *Expr : SCEVs) {
+    findValues(Expr, Values);
+    findLoops(Expr, Loops);
+  }
+
+  Values.remove_if([](const Value *V) { return isa<GlobalValue>(V); });
+
+  /// Remove loops that contain the scop or that are part of the scop, as they
+  /// are considered local. This leaves only loops that are before the scop, but
+  /// do not contain the scop itself.
+  Loops.remove_if([this](const Loop *L) {
+    return this->S.getRegion().contains(L) ||
+           L->contains(S.getRegion().getEntry());
+  });
+}
+
+void IslNodeBuilder::updateValues(
+    ParallelLoopGenerator::ValueToValueMapTy &NewValues) {
+  SmallPtrSet<Value *, 5> Inserted;
+
+  for (const auto &I : IDToValue) {
+    IDToValue[I.first] = NewValues[I.second];
+    Inserted.insert(I.second);
+  }
+
+  for (const auto &I : NewValues) {
+    if (Inserted.count(I.first))
+      continue;
+
+    ValueMap[I.first] = I.second;
+  }
+}
+
 void IslNodeBuilder::createUserVector(__isl_take isl_ast_node *User,
                                       std::vector<Value *> &IVS,
                                       __isl_take isl_id *IteratorID,
                                       __isl_take isl_union_map *Schedule) {
   isl_ast_expr *Expr = isl_ast_node_user_get_expr(User);
   isl_ast_expr *StmtExpr = isl_ast_expr_get_op_arg(Expr, 0);
   isl_id *Id = isl_ast_expr_get_id(StmtExpr);
   isl_ast_expr_free(StmtExpr);
case isl_ast_node_block: {
     isl_ast_node_list_free(List);
     break;
   }
   default:
     isl_ast_node_dump(Body);
     llvm_unreachable("Unhandled isl_ast_node in vectorizer");
   }
 
-  IDToValue.erase(IteratorID);
+  IDToValue.erase(IDToValue.find(IteratorID));
   isl_id_free(IteratorID);
   isl_union_map_free(Schedule);
 
   isl_ast_node_free(For);
   isl_ast_expr_free(Iterator);
 }
 
 void IslNodeBuilder::createForSequential(__isl_take isl_ast_node *For) {
void IslNodeBuilder::createForSequential(__isl_take isl_ast_node *For) {
   IV = createLoop(ValueLB, ValueUB, ValueInc, Builder, P, LI, DT, ExitBlock,
                   Predicate, &Annotator, Parallel, UseGuardBB);
   IDToValue[IteratorID] = IV;
 
   create(Body);
 
   Annotator.popLoop(Parallel);
 
-  IDToValue.erase(IteratorID);
+  IDToValue.erase(IDToValue.find(IteratorID));
 
   Builder.SetInsertPoint(ExitBlock->begin());
 
   isl_ast_node_free(For);
   isl_ast_expr_free(Iterator);
   isl_id_free(IteratorID);
 }
 
+/// @brief Remove the BBs contained in a (sub)function from the dominator tree.
+///
+/// This function removes the basic blocks that are part of a subfunction from
+/// the dominator tree. Specifically, when generating code it may happen that at
+/// some point the code generation continues in a new sub-function (e.g., when
+/// generating OpenMP code). The basic blocks that are created in this
+/// sub-function are then still part of the dominator tree of the original
+/// function, such that the dominator tree reaches over function boundaries.
+/// This is not only incorrect, but also causes crashes. This function now
+/// removes from the dominator tree all basic blocks that are dominated (and
+/// consequently reachable) from the entry block of this (sub)function.
+///
+/// FIXME: A LLVM (function or region) pass should not touch anything outside of
+/// the function/region it runs on. Hence, the pure need for this function shows
+/// that we do not comply to this rule. At the moment, this does not cause any
+/// issues, but we should be aware that such issues may appear. Unfortunately
+/// the current LLVM pass infrastructure does not allow to make Polly a module
+/// or call-graph pass to solve this issue, as such a pass would not have access
+/// to the per-function analyses passes needed by Polly. A future pass manager
+/// infrastructure is supposed to enable such kind of access possibly allowing
+/// us to create a cleaner solution here.
+///
+/// FIXME: Instead of adding the dominance information and then dropping it
+/// later on, we should try to just not add it in the first place. This requires
+/// some careful testing to make sure this does not break in interaction with
+/// the SCEVBuilder and SplitBlock which may rely on the dominator tree or
+/// which may try to update it.
+///
+/// @param F The function which contains the BBs to removed.
+/// @param DT The dominator tree from which to remove the BBs.
+static void removeSubFuncFromDomTree(Function *F, DominatorTree &DT) {
+  DomTreeNode *N = DT.getNode(&F->getEntryBlock());
+  std::vector<BasicBlock *> Nodes;
+
+  // We can only remove an element from the dominator tree, if all its children
+  // have been removed. To ensure this we obtain the list of nodes to remove
+  // using a post-order tree traversal.
+  for (po_iterator<DomTreeNode *> I = po_begin(N), E = po_end(N); I != E; ++I)
+    Nodes.push_back(I->getBlock());
+
+  for (BasicBlock *BB : Nodes)
+    DT.eraseNode(BB);
+}
+
+void IslNodeBuilder::createForParallel(__isl_take isl_ast_node *For) {
+  isl_ast_node *Body;
+  isl_ast_expr *Init, *Inc, *Iterator, *UB;
+  isl_id *IteratorID;
+  Value *ValueLB, *ValueUB, *ValueInc;
+  Type *MaxType;
+  Value *IV;
+  CmpInst::Predicate Predicate;
+
+  Body = isl_ast_node_for_get_body(For);
+  Init = isl_ast_node_for_get_init(For);
+  Inc = isl_ast_node_for_get_inc(For);
+  Iterator = isl_ast_node_for_get_iterator(For);
+  IteratorID = isl_ast_expr_get_id(Iterator);
+  UB = getUpperBound(For, Predicate);
+
+  ValueLB = ExprBuilder.create(Init);
+  ValueUB = ExprBuilder.create(UB);
+  ValueInc = ExprBuilder.create(Inc);
+
+  // OpenMP always uses SLE. In case the isl generated AST uses a SLT
+  // expression, we need to adjust the loop blound by one.
+  if (Predicate == CmpInst::ICMP_SLT)
+    ValueUB = Builder.CreateAdd(
+        ValueUB, Builder.CreateSExt(Builder.getTrue(), ValueUB->getType()));
+
+  MaxType = ExprBuilder.getType(Iterator);
+  MaxType = ExprBuilder.getWidestType(MaxType, ValueLB->getType());
+  MaxType = ExprBuilder.getWidestType(MaxType, ValueUB->getType());
+  MaxType = ExprBuilder.getWidestType(MaxType, ValueInc->getType());
+
+  if (MaxType != ValueLB->getType())
+    ValueLB = Builder.CreateSExt(ValueLB, MaxType);
+  if (MaxType != ValueUB->getType())
+    ValueUB = Builder.CreateSExt(ValueUB, MaxType);
+  if (MaxType != ValueInc->getType())
+    ValueInc = Builder.CreateSExt(ValueInc, MaxType);
+
+  BasicBlock::iterator LoopBody;
+
+  SetVector<Value *> SubtreeValues;
+  SetVector<const Loop *> Loops;
+
+  getReferencesInSubtree(For, SubtreeValues, Loops);
+
+  // Create for all loops we depend on values that contain the current loop
+  // iteration. These values are necessary to generate code for SCEVs that
+  // depend on such loops. As a result we need to pass them to the subfunction.
+  for (const Loop *L : Loops) {
+    const SCEV *OuterLIV = SE.getAddRecExpr(SE.getUnknown(Builder.getInt64(0)),
+                                            SE.getUnknown(Builder.getInt64(1)),
+                                            L, SCEV::FlagAnyWrap);
+    Value *V = generateSCEV(OuterLIV);
+    OutsideLoopIterations[L] = SE.getUnknown(V);
+    SubtreeValues.insert(V);
+  }
+
+  ParallelLoopGenerator::ValueToValueMapTy NewValues;
+  ParallelLoopGenerator ParallelLoopGen(Builder, P, LI, DT, DL);
+
+  IV = ParallelLoopGen.createParallelLoop(ValueLB, ValueUB, ValueInc,
+                                          SubtreeValues, NewValues, &LoopBody);
+  BasicBlock::iterator AfterLoop = Builder.GetInsertPoint();
+  Builder.SetInsertPoint(LoopBody);
+
+  // Save the current values.
+  ValueMapT ValueMapCopy = ValueMap;
+  IslExprBuilder::IDToValueTy IDToValueCopy = IDToValue;
+
+  updateValues(NewValues);
+  IDToValue[IteratorID] = IV;
+
+  create(Body);
+
+  // Restore the original values.
+  ValueMap = ValueMapCopy;
+  IDToValue = IDToValueCopy;
+
+  Builder.SetInsertPoint(AfterLoop);
+  removeSubFuncFromDomTree((*LoopBody).getParent()->getParent(), DT);
+
+  for (const Loop *L : Loops)
+    OutsideLoopIterations.erase(L);
+
+  isl_ast_node_free(For);
+  isl_ast_expr_free(Iterator);
+  isl_id_free(IteratorID);
+}
+
 void IslNodeBuilder::createFor(__isl_take isl_ast_node *For) {
   bool Vector = PollyVectorizerChoice != VECTORIZER_NONE;
 
   if (Vector && IslAstInfo::isInnermostParallel(For) &&
       !IslAstInfo::isReductionParallel(For)) {
     int VectorWidth = getNumberOfIterations(For);
     if (1 < VectorWidth && VectorWidth <= 16) {
       createForVector(For, VectorWidth);
       return;
     }
   }
+
+  if (IslAstInfo::isExecutedInParallel(For)) {
+    createForParallel(For);
+    return;
+  }
   createForSequential(For);
 }
 
 void IslNodeBuilder::createIf(__isl_take isl_ast_node *If) {
   isl_ast_expr *Cond = isl_ast_node_if_get_cond(If);
 
   Function *F = Builder.GetInsertBlock()->getParent();
   LLVMContext &Context = F->getContext();
for (int i = 0; i < isl_ast_expr_get_op_n_arg(Expr) - 1; ++i) {
     // result will always fit into the type of the original induction variable
     // (because we calculate a value of the original induction variable).
     const Value *OldIV = Stmt->getInductionVariableForDimension(i);
     if (OldIV) {
       V = Builder.CreateIntCast(V, OldIV->getType(), true);
       VMap[OldIV] = V;
     }
   }
 
+  // Add the current ValueMap to our per-statement value map.
+  //
+  // This is needed e.g. to rewrite array base addresses when moving code
+  // into a parallely executed subfunction.
+  VMap.insert(ValueMap.begin(), ValueMap.end());
+
   isl_ast_expr_free(Expr);
 }
 
 void IslNodeBuilder::createSubstitutionsVector(
     __isl_take isl_ast_expr *Expr, ScopStmt *Stmt, VectorValueMapT &VMap,
     std::vector<LoopToScevMapT> &VLTS, std::vector<Value *> &IVS,
     __isl_take isl_id *IteratorID) {
   int i = 0;
void IslNodeBuilder::createUser(__isl_take isl_ast_node *User) {
   isl_id *Id;
   ScopStmt *Stmt;
 
   isl_ast_expr *Expr = isl_ast_node_user_get_expr(User);
   isl_ast_expr *StmtExpr = isl_ast_expr_get_op_arg(Expr, 0);
   Id = isl_ast_expr_get_id(StmtExpr);
   isl_ast_expr_free(StmtExpr);
 
+  LTS.insert(OutsideLoopIterations.begin(), OutsideLoopIterations.end());
+
   Stmt = (ScopStmt *)isl_id_get_user(Id);
 
   createSubstitutions(Expr, Stmt, VMap, LTS);
   BlockGenerator::generate(Builder, *Stmt, VMap, LTS, P, LI, SE,
                            IslAstInfo::getBuild(User), &ExprBuilder);
 
   isl_ast_node_free(User);
   isl_id_free(Id);
for (unsigned i = 0; i < isl_set_dim(Context, isl_dim_param); ++i) {
     isl_id *Id;
 
     Id = isl_set_get_dim_id(Context, isl_dim_param, i);
     IDToValue[Id] = generateSCEV((const SCEV *)isl_id_get_user(Id));
 
     isl_id_free(Id);
   }
 
+  // Generate values for the current loop iteration for all surrounding loops.
+  //
+  // We may also reference loops outside of the scop which do not contain the
+  // scop itself, but as the number of such scops may be arbitrarily large we do
+  // not generate code for them here, but only at the point of code generation
+  // where these values are needed.
+  Region &R = S.getRegion();
+  Loop *L = LI.getLoopFor(R.getEntry());
+
+  while (L != nullptr && R.contains(L))
+    L = L->getParentLoop();
+
+  while (L != nullptr) {
+    const SCEV *OuterLIV = SE.getAddRecExpr(SE.getUnknown(Builder.getInt64(0)),
+                                            SE.getUnknown(Builder.getInt64(1)),
+                                            L, SCEV::FlagAnyWrap);
+    Value *V = generateSCEV(OuterLIV);
+    OutsideLoopIterations[L] = SE.getUnknown(V);
+    L = L->getParentLoop();
+  }
+
   isl_set_free(Context);
 }
 
 Value *IslNodeBuilder::generateSCEV(const SCEV *Expr) {
   Instruction *InsertLocation = --(Builder.GetInsertBlock()->end());
   return Rewriter->expandCodeFor(Expr, cast<IntegerType>(Expr->getType()),
                                  InsertLocation);
 }
 
 namespace {
 class IslCodeGeneration : public ScopPass {
 public:
   static char ID;
 
   IslCodeGeneration() : ScopPass(ID) {}
 
+  /// @brief The datalayout used
+  const DataLayout *DL;
+
   /// @name The analysis passes we need to generate code.
   ///
   ///{
   LoopInfo *LI;
   IslAstInfo *AI;
   DominatorTree *DT;
   ScalarEvolution *SE;
   ///}
Value *buildRTC(PollyIRBuilder &Builder, IslExprBuilder &ExprBuilder) {
     return RTC;
   }
 
   bool runOnScop(Scop &S) {
     LI = &getAnalysis<LoopInfo>();
     AI = &getAnalysis<IslAstInfo>();
     DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
     SE = &getAnalysis<ScalarEvolution>();
+    DL = &getAnalysis<DataLayoutPass>().getDataLayout();
 
     assert(!S.getRegion().isTopLevelRegion() &&
            "Top level regions are not supported");
 
     // Build the alias scopes for annotations first.
     if (PollyAnnotateAliasScopes)
       Annotator.buildAliasScopes(S);
 
     BasicBlock *EnteringBB = simplifyRegion(&S, this);
     PollyIRBuilder Builder = createPollyIRBuilder(EnteringBB, Annotator);
 
-    IslNodeBuilder NodeBuilder(Builder, Annotator, this, *LI, *SE, *DT);
+    IslNodeBuilder NodeBuilder(Builder, Annotator, this, *DL, *LI, *SE, *DT, S);
     NodeBuilder.addParameters(S.getContext());
 
     Value *RTC = buildRTC(Builder, NodeBuilder.getExprBuilder());
     BasicBlock *StartBlock = executeScopConditionally(S, this, RTC);
     Builder.SetInsertPoint(StartBlock->begin());
 
     NodeBuilder.create(AI->getAst());
     return true;
   }
 
   virtual void printScop(raw_ostream &OS) const {}
 
   virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+    AU.addRequired<DataLayoutPass>();
     AU.addRequired<DominatorTreeWrapperPass>();
     AU.addRequired<IslAstInfo>();
     AU.addRequired<RegionInfoPass>();
     AU.addRequired<ScalarEvolution>();
     AU.addRequired<ScopDetection>();
     AU.addRequired<ScopInfo>();
     AU.addRequired<LoopInfo>();
 

lib/Support/SCEVValidator.cpp

if (BaseAddress == V) {
       DEBUG(dbgs() << "INVALID: UnknownExpr references BaseAddress\n");
       return ValidatorResult(SCEVType::INVALID);
     }
 
     return ValidatorResult(SCEVType::PARAM, Expr);
   }
 };
 
 /// @brief Check whether a SCEV refers to an SSA name defined inside a region.
 ///
 struct SCEVInRegionDependences
     : public SCEVVisitor<SCEVInRegionDependences, bool> {
 public:
   /// Returns true when the SCEV has SSA names defined in region R.
   static bool hasDependences(const SCEV *S, const Region *R) {
     SCEVInRegionDependences Ignore(R);
     return Ignore.visit(S);
bool visitAddRecExpr(const SCEVAddRecExpr *Expr) {
     if (visit(Expr->getStart()))
       return true;
 
     for (size_t i = 0; i < Expr->getNumOperands(); ++i)
       if (visit(Expr->getOperand(i)))
         return true;
 
     return false;
   }
 
   bool visitSMaxExpr(const SCEVSMaxExpr *Expr) {
     for (size_t i = 0; i < Expr->getNumOperands(); ++i)
       if (visit(Expr->getOperand(i)))
         return true;
 
     return false;
   }
bool visitUnknown(const SCEVUnknown *Expr) {
     return false;
   }
 
 private:
   const Region *R;
 };
 
 namespace polly {
+/// Find all loops referenced in SCEVAddRecExprs.
+class SCEVFindLoops {
+  SetVector<const Loop *> &Loops;
+
+public:
+  SCEVFindLoops(SetVector<const Loop *> &Loops) : Loops(Loops) {}
+
+  bool follow(const SCEV *S) {
+    if (const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S))
+      Loops.insert(AddRec->getLoop());
+    return true;
+  }
+  bool isDone() { return false; }
+};
+
+void findLoops(const SCEV *Expr, SetVector<const Loop *> &Loops) {
+  SCEVFindLoops FindLoops(Loops);
+  SCEVTraversal<SCEVFindLoops> ST(FindLoops);
+  ST.visitAll(Expr);
+}
+
+/// Find all values referenced in SCEVUnknowns.
+class SCEVFindValues {
+  SetVector<Value *> &Values;
+
+public:
+  SCEVFindValues(SetVector<Value *> &Values) : Values(Values) {}
+
+  bool follow(const SCEV *S) {
+    if (const SCEVUnknown *Unknown = dyn_cast<SCEVUnknown>(S))
+      Values.insert(Unknown->getValue());
+    return true;
+  }
+  bool isDone() { return false; }
+};
+
+void findValues(const SCEV *Expr, SetVector<Value *> &Values) {
+  SCEVFindValues FindValues(Values);
+  SCEVTraversal<SCEVFindValues> ST(FindValues);
+  ST.visitAll(Expr);
+}
+
 bool hasScalarDepsInsideRegion(const SCEV *Expr, const Region *R) {
   return SCEVInRegionDependences::hasDependences(Expr, R);
 }
 
 bool isAffineExpr(const Region *R, const SCEV *Expr, ScalarEvolution &SE,
                   const Value *BaseAddress) {
   if (isa<SCEVCouldNotCompute>(Expr))
     return false;

test/Isl/CodeGen/OpenMP/loop-body-references-outer-iv.ll

+; RUN: opt %loadPolly -polly-parallel -polly-ast -analyze < %s | FileCheck %s -check-prefix=AST
+; RUN: opt %loadPolly -polly-parallel -polly-codegen-isl -S -polly-codegen-scev -verify-dom-info < %s | FileCheck %s -check-prefix=IR
+
+; This code has failed the scev based code generation as the scev in the scop
+; contains an AddRecExpr of an outer loop. When generating code, we did not
+; properly forward the value of this expression to the subfunction.
+
+; AST: #pragma omp parallel for
+; AST: for (int c1 = 0; c1 <= 1023; c1 += 1)
+; AST:  Stmt_for_j(c1);
+
+; IR: @single_parallel_loop.polly.subfn
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
+target triple = "x86_64-unknown-linux-gnu"
+
+@A = common global [1024 x float] zeroinitializer, align 16
+
+define void @single_parallel_loop() nounwind {
+entry:
+  br label %for.i
+
+for.i:
+  %indvar.i = phi i64 [ %indvar.i.next, %for.i.inc], [ 0, %entry ]
+  br label %for.j
+
+for.j:
+  %indvar.j = phi i64 [ %indvar.j.next, %for.j], [ 0, %for.i ]
+  %sum = add i64 %indvar.j, %indvar.i
+  %scevgep = getelementptr [1024 x float]* @A, i64 0, i64 %sum
+  store float 0.0, float *%scevgep
+  %indvar.j.next = add i64 %indvar.j, 1
+  %exitcond.j = icmp slt i64 %indvar.j.next, 1024
+  br i1 %exitcond.j, label %for.j, label %for.i.inc
+
+for.i.inc:
+  fence seq_cst
+  %indvar.i.next = add i64 %indvar.i, 1
+  %exitcond.i = icmp ne i64 %indvar.i.next, 1024
+  br i1 %exitcond.i, label %for.i, label %exit
+
+exit:
+  ret void
+}

test/Isl/CodeGen/OpenMP/loop-body-references-outer-values-2.ll

+; RUN: opt %loadPolly -polly-parallel -polly-ast -analyze < %s | FileCheck %s -check-prefix=AST
+; RUN: opt %loadPolly -polly-parallel -polly-codegen-isl -S -verify-dom-info < %s | FileCheck %s -check-prefix=IR
+; RUN: opt %loadPolly -polly-parallel -polly-codegen-isl -S -polly-codegen-scev -verify-dom-info < %s | FileCheck %s -check-prefix=IR-SCEV
+
+; AST: #pragma simd
+; AST: #pragma omp parallel for
+; AST: for (int c1 = 0; c1 <= 1023; c1 += 1)
+; AST:   Stmt_for_i(c1);
+
+; IR: %[[gep:[._a-zA-Z0-9]*]] = getelementptr inbounds { [1024 x double]*, i64 }* %polly.par.userContext, i32 0, i32 1
+; IR: store i64 %extern, i64* %[[gep]]
+
+; IR-SCEV: getelementptr inbounds { [1024 x double]* }* %polly.par.userContext, i32 0, i32 0
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @kernel_trmm([1024 x double]* %B) {
+entry:
+  br label %for.cond1.preheader
+
+for.cond1.preheader:
+  %extern = add i64 1, 0
+  br label %for.i
+
+for.i:
+  %indvar.i = phi i64 [ %indvar.i.next, %for.i ], [ 0, %for.cond1.preheader ]
+  %getelementptr = getelementptr [1024 x double]* %B, i64 %extern, i64 %indvar.i
+  store double 0.000000e+00, double* %getelementptr
+  %indvar.i.next = add i64 %indvar.i, 1
+  %exitcond.i = icmp ne i64 %indvar.i.next, 1024
+  br i1 %exitcond.i, label %for.i, label %end
+
+end:
+  ret void
+}

test/Isl/CodeGen/OpenMP/loop-body-references-outer-values-3.ll

+; RUN: opt %loadPolly -basicaa -polly-parallel -polly-ast -analyze < %s | FileCheck %s -check-prefix=AST
+; RUN: opt %loadPolly -basicaa -polly-parallel -polly-codegen-isl -S -verify-dom-info < %s | FileCheck %s -check-prefix=IR
+; RUN: opt %loadPolly -basicaa -polly-parallel -polly-codegen-isl -S -polly-codegen-scev -verify-dom-info < %s | FileCheck %s -check-prefix=IR
+
+; The interesting part of this test case is the instruction:
+;   %tmp = bitcast i8* %call to i64**
+; which is not part of the scop. In the SCEV based code generation not '%tmp',
+; but %call is a parameter of the SCoP and we need to make sure its value is
+; properly forwarded to the subfunction.
+
+; AST: #pragma omp parallel for
+; AST: for (int c1 = 0; c1 < cols; c1 += 1)
+; AST:   Stmt_for_body(c1);
+
+; IR: @foo.polly.subfn
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @foo(i64 %cols, i8* noalias %call) {
+entry:
+  %tmp = bitcast i8* %call to i64**
+  br label %for.body
+
+for.body:
+  %indvar = phi i64 [ %indvar.next, %for.body ], [ 0, %entry ]
+  %arrayidx = getelementptr inbounds i64** %tmp, i64 0
+  %tmp1 = load i64** %arrayidx, align 8
+  %arrayidx.2 = getelementptr inbounds i64* %tmp1, i64 %indvar
+  store i64 1, i64* %arrayidx.2, align 4
+  %indvar.next = add nsw i64 %indvar, 1
+  %cmp = icmp slt i64 %indvar.next, %cols
+  br i1 %cmp, label %for.body, label %end
+
+end:
+  ret void
+}
+
+; Another variation of this test case, now with even more of the index
+; expression defined outside of the scop.
+
+; AST: #pragma omp parallel for
+; AST: for (int c1 = 0; c1 < cols; c1 += 1)
+; AST:   Stmt_for_body(c1);
+
+; IR: @bar.polly.subfn
+
+define void @bar(i64 %cols, i8* noalias %call) {
+entry:
+  %tmp = bitcast i8* %call to i64**
+  %arrayidx = getelementptr inbounds i64** %tmp, i64 0
+  br label %for.body
+
+for.body:
+  %indvar = phi i64 [ %indvar.next, %for.body ], [ 0, %entry ]
+  %tmp1 = load i64** %arrayidx, align 8
+  %arrayidx.2 = getelementptr inbounds i64* %tmp1, i64 %indvar
+  store i64 1, i64* %arrayidx.2, align 4
+  %indvar.next = add nsw i64 %indvar, 1
+  %cmp = icmp slt i64 %indvar.next, %cols
+  br i1 %cmp, label %for.body, label %end
+
+end:
+  ret void
+}

test/Isl/CodeGen/OpenMP/loop-body-references-outer-values.ll

+; RUN: opt %loadPolly -polly-parallel -polly-ast -analyze < %s | FileCheck %s -check-prefix=AST
+; RUN: opt %loadPolly -polly-parallel -polly-codegen-isl -S < %s | FileCheck %s -check-prefix=IR
+; RUN: opt %loadPolly -polly-parallel -polly-codegen-isl -S -polly-codegen-scev < %s | FileCheck %s -check-prefix=IR
+
+; Make sure we correctly forward the reference to 'A' to the OpenMP subfunction.
+;
+; void loop_references_outer_ids(float *A) {
+;   for (long i = 0; i < 100; i++)
+;     A[i] = i;
+; }
+
+
+; AST: #pragma simd
+; AST: #pragma omp parallel for
+; AST: for (int c1 = 0; c1 <= 99; c1 += 1)
+; AST:   Stmt_for_body(c1);
+
+; IR-LABEL: polly.start:
+; IR-NEXT:  %0 = bitcast { float* }* %polly.par.userContext to i8*
+; IR-NEXT:  call void @llvm.lifetime.start(i64 8, i8* %0)
+; IR-NEXT:  %1 = getelementptr inbounds { float* }* %polly.par.userContext, i32 0, i32 0
+; IR-NEXT:  store float* %A, float** %1
+; IR-NEXT:  %polly.par.userContext1 = bitcast { float* }* %polly.par.userContext to i8*
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+define void @loop_references_outer_ids(float* %A) {
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc, %entry
+  %i.0 = phi i64 [ 0, %entry ], [ %inc, %for.inc ]
+  %exitcond = icmp ne i64 %i.0, 100
+  br i1 %exitcond, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  %conv = sitofp i64 %i.0 to float
+  %arrayidx = getelementptr inbounds float* %A, i64 %i.0
+  store float %conv, float* %arrayidx, align 4
+  br label %for.inc
+
+for.inc:                                          ; preds = %for.body
+  %inc = add nsw i64 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  ret void
+}

test/Isl/CodeGen/OpenMP/loop-bounds-reference-outer-ids.ll

+; RUN: opt %loadPolly -polly-parallel -polly-ast -analyze < %s | FileCheck %s -check-prefix=AST
+; RUN: opt %loadPolly -polly-parallel -polly-codegen-isl -S < %s | FileCheck %s -check-prefix=IR
+; RUN: opt %loadPolly -polly-parallel -polly-codegen-isl -S -polly-codegen-scev < %s | FileCheck %s -check-prefix=IR
+;
+; float A[100];
+;
+; void loop_references_outer_ids(long n) {
+;   for (long i = 0; i < 100; i++)
+;     for (long j = 0; j < 100; j++)
+;       for (long k = 0; k < n + i; k++)
+;         A[j] += i + j + k;
+; }
+
+; In this test case we verify that the j-loop is generated as OpenMP parallel
+; loop and that the values of 'i' and 'n', needed in the loop bounds of the
+; k-loop, are correctly passed to the subfunction.
+
+; AST: #pragma minimal dependence distance: 1
+; AST: for (int c1 = max(0, -n + 1); c1 <= 99; c1 += 1)
+; AST:   #pragma omp parallel for
+; AST:   for (int c3 = 0; c3 <= 99; c3 += 1)
+; AST:     #pragma minimal dependence distance: 1
+; AST:     for (int c5 = 0; c5 < n + c1; c5 += 1)
+; AST:       Stmt_for_body6(c1, c3, c5);
+
+; IR: %polly.par.userContext = alloca { i64, i64 }
+; IR: %4 = bitcast { i64, i64 }* %polly.par.userContext to i8*
+; IR-NEXT: call void @llvm.lifetime.start(i64 16, i8* %4)
+; IR-NEXT: %5 = getelementptr inbounds { i64, i64 }* %polly.par.userContext, i32 0, i32 0
+; IR-NEXT: store i64 %n, i64* %5
+; IR-NEXT: %6 = getelementptr inbounds { i64, i64 }* %polly.par.userContext, i32 0, i32 1
+; IR-NEXT: store i64 %polly.indvar, i64* %6
+; IR-NEXT: %polly.par.userContext1 = bitcast { i64, i64 }* %polly.par.userContext to i8*
+
+; IR-LABEL: @loop_references_outer_ids.polly.subfn(i8* %polly.par.userContext)
+; IR: %polly.par.userContext1 = bitcast i8* %polly.par.userContext to { i64, i64 }*
+; IR-NEXT: %0 = getelementptr inbounds { i64, i64 }* %polly.par.userContext1, i32 0, i32 0
+; IR-NEXT: %1 = load i64* %0
+; IR-NEXT: %2 = getelementptr inbounds { i64, i64 }* %polly.par.userContext1, i32 0, i32 1
+; IR-NEXT: %3 = load i64* %2
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+
+@A = common global [100 x float] zeroinitializer, align 16
+
+define void @loop_references_outer_ids(i64 %n) {
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.inc13, %entry
+  %i.0 = phi i64 [ 0, %entry ], [ %inc14, %for.inc13 ]
+  %exitcond1 = icmp ne i64 %i.0, 100
+  br i1 %exitcond1, label %for.body, label %for.end15
+
+for.body:                                         ; preds = %for.cond
+  br label %for.cond1
+
+for.cond1:                                        ; preds = %for.inc10, %for.body
+  %j.0 = phi i64 [ 0, %for.body ], [ %inc11, %for.inc10 ]
+  %exitcond = icmp ne i64 %j.0, 100
+  br i1 %exitcond, label %for.body3, label %for.end12
+
+for.body3:                                        ; preds = %for.cond1
+  br label %for.cond4
+
+for.cond4:                                        ; preds = %for.inc, %for.body3
+  %k.0 = phi i64 [ 0, %for.body3 ], [ %inc, %for.inc ]
+  %add = add nsw i64 %i.0, %n
+  %cmp5 = icmp slt i64 %k.0, %add
+  br i1 %cmp5, label %for.body6, label %for.end
+
+for.body6:                                        ; preds = %for.cond4
+  %add7 = add nsw i64 %i.0, %j.0
+  %add8 = add nsw i64 %add7, %k.0
+  %conv = sitofp i64 %add8 to float
+  %arrayidx = getelementptr inbounds [100 x float]* @A, i64 0, i64 %j.0
+  %tmp = load float* %arrayidx, align 4
+  %add9 = fadd float %tmp, %conv
+  store float %add9, float* %arrayidx, align 4
+  br label %for.inc
+
+for.inc:                                          ; preds = %for.body6
+  %inc = add nsw i64 %k.0, 1
+  br label %for.cond4
+
+for.end:                                          ; preds = %for.cond4
+  br label %for.inc10
+
+for.inc10:                                        ; preds = %for.end
+  %inc11 = add nsw i64 %j.0, 1
+  br label %for.cond1
+
+for.end12:                                        ; preds = %for.cond1
+  br label %for.inc13
+
+for.inc13:                                        ; preds = %for.end12
+  %inc14 = add nsw i64 %i.0, 1
+  br label %for.cond
+
+for.end15:                                        ; preds = %for.cond
+  ret void
+}

test/Isl/CodeGen/OpenMP/reference-other-bb.ll

+; RUN: opt %loadPolly -polly-parallel -polly-codegen-isl -S -polly-codegen-scev -verify-dom-info < %s | FileCheck %s -check-prefix=IR
+
+; IR: @foo.polly.subfn
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @foo(i32 %sendcount, i8* %recvbuf) {
+entry:
+  br label %sw.bb3
+
+sw.bb3:
+  %tmp = bitcast i8* %recvbuf to double*
+  %cmp75 = icmp sgt i32 %sendcount, 0
+  br i1 %cmp75, label %for.body, label %end
+
+for.body:
+  %i.16 = phi i32 [ %inc14, %for.body ], [ 0, %sw.bb3 ]
+  %idxprom11 = sext i32 %i.16 to i64
+  %arrayidx12 = getelementptr inbounds double* %tmp, i64 %idxprom11
+  store double 1.0, double* %arrayidx12, align 8
+  %inc14 = add nsw i32 %i.16, 1
+  %cmp7 = icmp slt i32 %inc14, %sendcount
+  br i1 %cmp7, label %for.body, label %end
+
+end:
+  ret void
+}

test/Isl/CodeGen/OpenMP/reference-preceeding-loop.ll

+; RUN: opt %loadPolly -polly-parallel -polly-ast -analyze -polly-codegen-scev < %s | FileCheck %s -check-prefix=AST
+; RUN: opt %loadPolly -polly-parallel -polly-codegen-isl -S -polly-codegen-scev -verify-dom-info < %s | FileCheck %s -check-prefix=IR
+
+
+; - Test the case where scalar evolution references a loop that is outside
+;   of the scop, but does not contain the scop.
+; - Test the case where two parallel subfunctions are created.
+
+; AST: if (symbol >= p_2 + 1) {
+; AST-NEXT:   #pragma simd
+; AST-NEXT:   #pragma omp parallel for
+; AST-NEXT:   for (int c1 = 0; c1 < p_0 + symbol; c1 += 1)
+; AST-NEXT:     Stmt_while_body(c1);
+; AST-NEXT: } else
+; AST-NEXT:   #pragma simd
+; AST-NEXT:   #pragma omp parallel for
+; AST-NEXT:   for (int c1 = 0; c1 <= p_0 + p_2; c1 += 1)
+; AST-NEXT:     Stmt_while_body(c1);
+
+; IR: @update_model.polly.subfn
+; IR: @update_model.polly.subfn1
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+@cum_freq = external global [258 x i64], align 16
+
+define void @update_model(i64 %symbol) {
+entry:
+  br label %for.one
+
+for.one:
+  %i.1 = phi i64 [ %dec17, %for.one ], [ %symbol, %entry ]
+  %dec17 = add nsw i64 %i.1, -1
+  br i1 undef, label %for.one, label %while.body
+
+while.body:
+  %indvar = phi i64 [ %sub42, %while.body ], [ %i.1, %for.one ]
+  %sub42 = add nsw i64 %indvar, -1
+  %arrayidx44 = getelementptr inbounds [258 x i64]* @cum_freq, i64 0, i64 %sub42
+  store i64 1, i64* %arrayidx44, align 4
+  %cmp40 = icmp sgt i64 %sub42, 0
+  br i1 %cmp40, label %while.body, label %while.end
+
+while.end:
+  ret void
+}

test/Isl/CodeGen/OpenMP/single_loop.ll

+; RUN: opt %loadPolly -polly-parallel -polly-ast -analyze < %s | FileCheck %s -check-prefix=AST
+; RUN: opt %loadPolly -polly-parallel -polly-codegen-isl -S -verify-dom-info < %s | FileCheck %s -check-prefix=IR
+; RUN: opt %loadPolly -polly-parallel -polly-codegen-isl -S -polly-codegen-scev -verify-dom-info < %s | FileCheck %s -check-prefix=IR
+
+; RUN: opt %loadPolly -polly-parallel -polly-import-jscop -polly-import-jscop-dir=%S -polly-ast -analyze < %s | FileCheck %s -check-prefix=AST-STRIDE4
+; RUN: opt %loadPolly -polly-parallel -polly-import-jscop -polly-import-jscop-dir=%S -polly-codegen-isl -S < %s | FileCheck %s -check-prefix=IR-STRIDE4
+; RUN: opt %loadPolly -polly-parallel -polly-import-jscop -polly-import-jscop-dir=%S -polly-codegen-isl -polly-codegen-scev -S < %s | FileCheck %s -check-prefix=IR-STRIDE4
+
+; This extensive test case tests the creation of the full set of OpenMP calls
+; as well as the subfunction creation using a trivial loop as example.
+
+; #define N 1024
+; float A[N];
+;
+; void single_parallel_loop(void) {
+;   for (long i = 0; i < N; i++)
+;     A[i] = 1;
+; }
+
+; AST: #pragma simd
+; AST: #pragma omp parallel for
+; AST: for (int c1 = 0; c1 <= 1023; c1 += 1)
+; AST:   Stmt_S(c1);
+
+; AST-STRIDE4: #pragma omp parallel for
+; AST-STRIDE4: for (int c1 = 0; c1 <= 1023; c1 += 4)
+; AST-STRIDE4:   #pragma simd
+; AST-STRIDE4:   for (int c2 = c1; c2 <= c1 + 3; c2 += 1)
+; AST-STRIDE4:     Stmt_S(c2);
+
+; IR-LABEL: single_parallel_loop()
+; IR-NEXT: entry
+; IR-NEXT:   %polly.par.userContext = alloca
+
+; IR-LABEL: polly.start:
+; IR-NEXT:   %0 = bitcast {}* %polly.par.userContext to i8*
+; IR-NEXT:   call void @llvm.lifetime.start(i64 0, i8* %0)
+; IR-NEXT:   %polly.par.userContext1 = bitcast {}* %polly.par.userContext to i8*
+; IR-NEXT:   call void @GOMP_parallel_loop_runtime_start(void (i8*)* @single_parallel_loop.polly.subfn, i8* %polly.par.userContext1, i32 0, i64 0, i64 1024, i64 1)
+; IR-NEXT:   call void @single_parallel_loop.polly.subfn(i8* %polly.par.userContext1)
+; IR-NEXT:   call void @GOMP_parallel_end()
+; IR-NEXT:   %1 = bitcast {}* %polly.par.userContext to i8*
+; IR-NEXT:   call void @llvm.lifetime.end(i64 8, i8* %1)
+; IR-NEXT:   br label %polly.merge_new_and_old
+
+; IR: define internal void @single_parallel_loop.polly.subfn(i8* %polly.par.userContext) #1
+; IR-LABEL: polly.par.setup:
+; IR-NEXT:   %polly.par.LBPtr = alloca i64
+; IR-NEXT:   %polly.par.UBPtr = alloca i64
+; IR-NEXT:   %polly.par.userContext1 =
+; IR:   br label %polly.par.checkNext
+
+; IR-LABEL: polly.par.exit:
+; IR-NEXT:   call void @GOMP_loop_end_nowait()
+; IR-NEXT:   ret void
+
+; IR-LABEL: polly.par.checkNext:
+; IR-NEXT:   %[[parnext:[._a-zA-Z0-9]*]] = call i8 @GOMP_loop_runtime_next(i64* %polly.par.LBPtr, i64* %polly.par.UBPtr)
+; IR-NEXT:   %[[cmp:[._a-zA-Z0-9]*]] = icmp ne i8 %[[parnext]], 0
+; IR-NEXT:   br i1 %[[cmp]], label %polly.par.loadIVBounds, label %polly.par.exit
+
+; IR-LABEL: polly.par.loadIVBounds:
+; IR-NEXT:   %polly.par.LB = load i64* %polly.par.LBPtr
+; IR-NEXT:   %polly.par.UB = load i64* %polly.par.UBPtr
+; IR-NEXT:   %polly.par.UBAdjusted = sub i64 %polly.par.UB, 1
+; IR-NEXT:   br label %polly.loop_preheader
+
+; IR-LABEL: polly.loop_exit:
+; IR-NEXT:   br label %polly.par.checkNext
+
+; IR-LABEL: polly.loop_header:
+; IR-NEXT:   %polly.indvar = phi i64 [ %polly.par.LB, %polly.loop_preheader ], [ %polly.indvar_next, %polly.stmt.S ]
+; IR-NEXT:   br label %polly.stmt.S
+
+; IR-LABEL: polly.stmt.S:
+; IR-NEXT:   %[[gep:[._a-zA-Z0-9]*]] = getelementptr [1024 x float]* {{.*}}, i64 0, i64 %polly.indvar
+; IR-NEXT:   store float 1.000000e+00, float* %[[gep]]
+; IR-NEXT:   %polly.indvar_next = add nsw i64 %polly.indvar, 1
+; IR-NEXT:   %polly.adjust_ub = sub i64 %polly.par.UBAdjusted, 1
+; IR-NEXT:   %polly.loop_cond = icmp sle i64 %polly.indvar, %polly.adjust_ub
+; IR-NEXT:   br i1 %polly.loop_cond, label %polly.loop_header, label %polly.loop_exit
+
+; IR-LABEL: polly.loop_preheader:
+; IR-NEXT:   br label %polly.loop_header
+
+; IR: attributes #1 = { "polly.skip.fn" }
+
+; IR-STRIDE4:   call void @GOMP_parallel_loop_runtime_start(void (i8*)* @single_parallel_loop.polly.subfn, i8* %polly.par.userContext1, i32 0, i64 0, i64 1024, i64 4)
+; IR-STRIDE4:  add nsw i64 %polly.indvar, 3
+; IR-STRIDE4:  %polly.indvar_next = add nsw i64 %polly.indvar, 4
+; IR-STRIDE4   %polly.adjust_ub = sub i64 %polly.par.UBAdjusted, 4
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
+target triple = "x86_64-unknown-linux-gnu"
+
+@A = common global [1024 x float] zeroinitializer, align 16
+
+define void @single_parallel_loop() nounwind {
+entry:
+  br label %for.i
+
+for.i:
+  %indvar = phi i64 [ %indvar.next, %for.inc], [ 0, %entry ]
+  %scevgep = getelementptr [1024 x float]* @A, i64 0, i64 %indvar
+  %exitcond = icmp ne i64 %indvar, 1024
+  br i1 %exitcond, label %S, label %exit
+
+S:
+  store float 1.0, float* %scevgep
+  br label %for.inc
+
+for.inc:
+  %indvar.next = add i64 %indvar, 1
+  br label %for.i
+
+exit:
+  ret void
+}

test/Isl/CodeGen/OpenMP/single_loop_with_loop_invariant_baseptr.ll

+; RUN: opt %loadPolly -tbaa -polly-parallel -polly-ast -analyze < %s | FileCheck %s -check-prefix=AST
+; RUN: opt %loadPolly -tbaa -polly-parallel -polly-codegen-isl -S -verify-dom-info < %s | FileCheck %s -check-prefix=IR
+; RUN: opt %loadPolly -tbaa -polly-parallel -polly-codegen-isl -S -polly-codegen-scev -verify-dom-info < %s | FileCheck %s -check-prefix=IR
+
+; #define N 1024
+; float A[N];
+;
+; void single_parallel_loop(void) {
+;   for (long i = 0; i < N; i++)
+;     A[i] = 1;
+; }
+
+; AST: #pragma simd
+; AST: #pragma omp parallel for
+; AST: for (int c1 = 0; c1 <= 1023; c1 += 1)
+; AST:   Stmt_S(c1);
+
+; IR: @single_parallel_loop.polly.subfn
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @single_parallel_loop(float** %A) nounwind {
+entry:
+  br label %for.i
+
+for.i:
+  %indvar = phi i64 [ %indvar.next, %for.inc], [ 0, %entry ]
+  %exitcond = icmp ne i64 %indvar, 1024
+  br i1 %exitcond, label %S, label %exit
+
+S:
+  %ptr = load float** %A,  !tbaa !2
+  %scevgep = getelementptr float* %ptr, i64 %indvar
+  %val = load float* %scevgep, !tbaa !6
+  %sum = fadd float %val, 1.0
+  store float %sum, float* %scevgep, !tbaa !6
+  br label %for.inc
+
+for.inc:
+  %indvar.next = add i64 %indvar, 1
+  br label %for.i
+
+exit:
+  ret void
+}
+
+!2 = metadata !{metadata !"float", metadata !3, i64 0}
+!3 = metadata !{metadata !"omnipotent char", metadata !4, i64 0}
+!4 = metadata !{metadata !"Simple C/C++ TBAA"}
+!6 = metadata !{metadata !"float *ptr", metadata !3, i64 0}

test/Isl/CodeGen/OpenMP/single_parallel_loop___%for.i---%exit.jscop

+{
+   "context" : "{  :  }",
+   "name" : "for.i => exit",
+   "statements" : [
+      {
+         "accesses" : [
+            {
+               "kind" : "write",
+               "relation" : "{ Stmt_S[i0] -> MemRef_A[i0] }"
+            }
+         ],
+         "domain" : "{ Stmt_S[i0] : i0 >= 0 and i0 <= 1023 }",
+         "name" : "Stmt_S",
+         "schedule" : "{ Stmt_S[i0] -> scattering[0, floor(i0/4) * 4, i0] }"
+      }
+   ]
+}

test/Isl/CodeGen/OpenMP/two-parallel-loops-reference-outer-indvar.ll

+; RUN: opt %loadPolly -polly-parallel -polly-ast -analyze -polly-codegen-scev < %s | FileCheck %s -check-prefix=AST
+; RUN: opt %loadPolly -polly-parallel -polly-codegen-isl -S -polly-codegen-scev -verify-dom-info < %s | FileCheck %s -check-prefix=IR
+
+; This test case verifies that we create correct code even if two OpenMP loops
+; share common outer variables.
+
+; AST: if (nj >= p_1 + 3) {
+; AST:   #pragma simd
+; AST:   #pragma omp parallel for
+; AST:   for (int c1 = 0; c1 < p_0 + nj - 1; c1 += 1)
+; AST:     Stmt_for_body35(c1);
+; AST: } else
+; AST:   #pragma simd
+; AST:   #pragma omp parallel for
+; AST:   for (int c1 = 0; c1 <= p_0 + p_1; c1 += 1)
+; AST:     Stmt_for_body35(c1);
+
+; IR: @foo.polly.subfn
+; IR: @foo.polly.subfn1
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @foo(i64 %nj, [512 x double]* %R) {
+entry:
+  br label %for.cond1.preheader
+
+for.cond1.preheader:
+  %k.014 = phi i64 [ %inc87, %for.inc86 ], [ 0, %entry ]
+  %j.010 = add nsw i64 %k.014, 1
+  br i1 undef, label %for.body35, label %for.inc86
+
+for.body35:
+  %j.012 = phi i64 [ %j.0, %for.body35 ], [ %j.010, %for.cond1.preheader ]
+  %arrayidx39 = getelementptr inbounds [512 x double]* %R, i64 0, i64 %j.012
+  store double 0.000000e+00, double* %arrayidx39
+  %j.0 = add nsw i64 %j.012, 1
+  %cmp34 = icmp slt i64 %j.0, %nj
+  br i1 %cmp34, label %for.body35, label %for.inc86
+
+for.inc86:
+  %inc87 = add nsw i64 %k.014, 1
+  br i1 undef, label %for.cond1.preheader, label %for.end88
+
+for.end88:
+  ret void
+}