[LoopDataPrefetch + SystemZ] Let target decide on prefetching on a per loop basis

[jonpa]

Summary

This patch adds

• New arguments to getMinPrefetchStride() to let the target decide on a per-loop basis if software prefetching should be done even with a stride within the limit of the hw prefetcher.
• New TTI hook enableWritePrefetching() to let a target do write prefetching by default (defaults to false).

LoopDataPrefetch:

• A search through the whole loop to gather information before emitting any prefetches. This way the target can get information via new arguments to getMinPrefetchStride() and emit prefetches more selectively. Collected information includes: Does the loop have a call, how many memory accesses, how many of them are strided, how many prefetches will cover them. This is NFC to before as long as the target does not change its definition of getMinPrefetchStride().

• If a previous access to the same exact address was 'read', and the current one is 'write', make it a 'write' prefetch.

• If two accesses that are covered by the same prefetch do not dominate each other, put the prefetch in a block that dominates both of them.

• If a ConstantMaxTripCount is less than ItersAhead, then skip the loop.

SystemZ:

• increase the distance of prefetching (to meet the hot lbm loop with prefetching 9 iterations ahead).
• enable write prefetching by default.
• emit sw prefetching for any stride according to new heuristics in getMinPrefetchStride(), which includes lbm.
• Do we need a test to test getMinPrefetchStride()?

[jonpa]

llvm/lib/Transforms/Scalar/LoopDataPrefetch.cpp

233

The Prefetch struct MemI member is used solely for debug/ORE output. It is currently not guarded by '#ifndef NDEBUG', but maybe it should be, even though it's just one pointer? The debug output could perhaps be improved also, I merely tried to keep what there was while also printing the new values gathered. The MemI member is not needed for anything else, so if we could change the debug output instead perhaps it's not even needed, or?

251

I am trusting DT->findNearestCommonDominator() and DomBB->getTerminator() to do this, but I am not 100% sure that there is always a terminator in each block...?

[jonpa]

ping!

[jonpa]

Ping! Patch rebased.

We see that lbm gets a significant speedup with this patch on SystemZ.

[jonpa]

Ping!

Adding members of the Loop Optimization Working Group, as I do believe this patch deserves some attention.

[jonpa]

ping!

Did anyone try to enable write prefetching on lbm on another target than SystemZ?

[Meinersbur]

@greened is looking into generating the prefetcher properties using tblgen (D58736). Would this be compatible with that approach?

I like the clean test cases. Unfortunely, I don't have experience with prefetching in the lbm benchmark.

llvm/include/llvm/Analysis/TargetTransformInfo.h

1324–1327

Could you add explanations for the individual parameters to the doxygen comment?

llvm/lib/Transforms/Scalar/LoopDataPrefetch.cpp

233

I'd appreciate if it was guarded by #ifndef NDEBUG. It would make it more explicit and enforcing to be only used in debug builds.

238

[style] Please add empty lines between methods.

Also, add a doxygen comment about what this method and its parameters (and the class itself) are supposed to do.

244–245

[style] please clang-format the patch.

[jonpa]

Thanks for review! Patch updated.

@greened is looking into generating the prefetcher properties using tblgen (D58736). Would this be compatible with that approach?

I see that his patch is as well adding the option to specify write-prefetching per default, so that is a common goal. From what I can see that patch doesn't have any other functional changes to LoopDataPrefetch, so it seems complimentary. Personally, I do think that there is a need to do tuning in C++ for a given target, like my patch here for SystemZ - I am not sure if greened is aiming to generalize cache systems and in the future implement some kind of common heuristics for all? I would hope then that it would still be possible for each target to make its own decisions. I have a feeling that these systems are so complex that they are not quite predictable with generalized parameters.

Unfortunely, I don't have experience with prefetching in the lbm benchmark.

It goes over a huge memory array with both loads and stores. Enabling store prefetching on it seems like a good idea on any architecture...

llvm/lib/Transforms/Scalar/LoopDataPrefetch.cpp

233

I don't think that's possible given that it's also used by the ORE, or?

[Meinersbur]

llvm/include/llvm/Analysis/TargetTransformInfo.h

884

I am confused with the description of NumPrefetches. It looks like its passed the number of different memory accesses in the loop. "Number of prefetches needed for the strided accesses." sounds like hardware requirement, but its passed a property of the code. Can you clarify?

llvm/lib/Transforms/Scalar/LoopDataPrefetch.cpp

233

Well, then it's not only for debug output as the comment says. Remove the part in parens?

I'd appreciate if the other class members had doxygen comments as well.

237

[style] Start parameter names with capital letters (PascalCase)

[jonpa]

Patch updated per review.

Comments expanded. I hope multiple line doxygen comments are ok for a parameter?

[Meinersbur]

In D70228#1956686, @jonpa wrote:

Comments expanded. I hope multiple line doxygen comments are ok for a parameter?

Yes; I usually indent/align them, although not done automatically by clang format. For instance, the doxygen for getMinPrefetchStride could be improved:

/// Some HW prefetchers can handle accesses up to a certain constant stride.
/// Sometimes prefetching is beneficial even below the HW prefetcher limit,
/// and the arguments provided are meant to serve as a basis for deciding this
/// for a particular loop.
///
/// \param NumMemAccesses        Number of memory accesses in the loop.
/// \param NumStridedMemAccesses Number of the memory accesses that
///                              ScalarEvolution could find a known stride
///                              for.
/// \param NumPrefetches         Number of software prefetches that will be
///                              emitted as determined by the addresses
///                              involved and the cache line size.
/// \param HasCall               True if the loop contains a call.
///
/// \return This is the minimum stride in bytes where it makes sense to start
///         adding SW prefetches. The default is 1, i.e. prefetch with any
///         stride.

However, there is a tradeoff involved in number off lines changed in this diff, consistency with the other docygen comments in the same file, number of likes for the entire doxygen comment.

[jonpa]

In D70228#1956859, @Meinersbur wrote:

In D70228#1956686, @jonpa wrote:

Comments expanded. I hope multiple line doxygen comments are ok for a parameter?

Yes; I usually indent/align them, although not done automatically by clang format. For instance, the doxygen for getMinPrefetchStride could be improved:

/// Some HW prefetchers can handle accesses up to a certain constant stride.
/// Sometimes prefetching is beneficial even below the HW prefetcher limit,
/// and the arguments provided are meant to serve as a basis for deciding this
/// for a particular loop.
///
/// \param NumMemAccesses        Number of memory accesses in the loop.
/// \param NumStridedMemAccesses Number of the memory accesses that
///                              ScalarEvolution could find a known stride
///                              for.
/// \param NumPrefetches         Number of software prefetches that will be
///                              emitted as determined by the addresses
///                              involved and the cache line size.
/// \param HasCall               True if the loop contains a call.
///
/// \return This is the minimum stride in bytes where it makes sense to start
///         adding SW prefetches. The default is 1, i.e. prefetch with any
///         stride.

However, there is a tradeoff involved in number off lines changed in this diff, consistency with the other docygen comments in the same file, number of likes for the entire doxygen comment.

I see :-) Thanks for the review.

llvm/include/llvm/Analysis/TargetTransformInfo.h

public:
   /// \return How much before a load we should place the prefetch
   /// instruction.  This is currently measured in number of
   /// instructions.
   unsigned getPrefetchDistance() const;
 
   /// \return Some HW prefetchers can handle accesses up to a certain
   /// constant stride.  This is the minimum stride in bytes where it
   /// makes sense to start adding SW prefetches.  The default is 1,
-  /// i.e. prefetch with any stride.
-  unsigned getMinPrefetchStride() const;
+  /// i.e. prefetch with any stride.  Sometimes prefetching is beneficial
+  /// even below the HW prefetcher limit, and the arguments provided are
+  /// meant to serve as a basis for deciding this for a particular loop.
+  unsigned getMinPrefetchStride(unsigned NumMemAccesses,
+                                unsigned NumStridedMemAccesses,
+                                unsigned NumPrefetches,
+                                bool HasCall) const;
 
   /// \return The maximum number of iterations to prefetch ahead.  If
   /// the required number of iterations is more than this number, no
   /// prefetching is performed.
   unsigned getMaxPrefetchIterationsAhead() const;
 
+  /// \return True if prefetching should also be done for writes.
+  bool enableWritePrefetching() const;
+
   /// \return The maximum interleave factor that any transform should try to
   /// perform for this target. This number depends on the level of parallelism
   /// and the number of execution units in the CPU.
   unsigned getMaxInterleaveFactor(unsigned VF) const;
 
   /// Collect properties of V used in cost analysis, e.g. OP_PowerOf2.
   static OperandValueKind getOperandInfo(Value *V,
                                          OperandValueProperties &OpProps);
public:
   /// \return How much before a load we should place the prefetch
   /// instruction.  This is currently measured in number of
   /// instructions.
   virtual unsigned getPrefetchDistance() const = 0;
 
   /// \return Some HW prefetchers can handle accesses up to a certain
   /// constant stride.  This is the minimum stride in bytes where it
   /// makes sense to start adding SW prefetches.  The default is 1,
-  /// i.e. prefetch with any stride.
-  virtual unsigned getMinPrefetchStride() const = 0;
+  /// i.e. prefetch with any stride.  Sometimes prefetching is beneficial
+  /// even below the HW prefetcher limit, and the arguments provided are
+  /// meant to serve as a basis for deciding this for a particular loop.
+  virtual unsigned getMinPrefetchStride(unsigned NumMemAccesses,
+                                        unsigned NumStridedMemAccesses,
+                                        unsigned NumPrefetches,
+                                        bool HasCall) const = 0;
 
   /// \return The maximum number of iterations to prefetch ahead.  If
   /// the required number of iterations is more than this number, no
   /// prefetching is performed.
   virtual unsigned getMaxPrefetchIterationsAhead() const = 0;
 
+  /// \return True if prefetching should also be done for writes.
+  virtual bool enableWritePrefetching() const = 0;
+
   virtual unsigned getMaxInterleaveFactor(unsigned VF) = 0;
   virtual unsigned getArithmeticInstrCost(
       unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
       OperandValueKind Opd2Info, OperandValueProperties Opd1PropInfo,
       OperandValueProperties Opd2PropInfo, ArrayRef<const Value *> Args,
       const Instruction *CxtI = nullptr) = 0;
   virtual int getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
                              Type *SubTp) = 0;
public:
   ///
   unsigned getPrefetchDistance() const override {
     return Impl.getPrefetchDistance();
   }
 
   /// Return the minimum stride necessary to trigger software
   /// prefetching.
   ///
-  unsigned getMinPrefetchStride() const override {
-    return Impl.getMinPrefetchStride();
+  unsigned getMinPrefetchStride(unsigned NumMemAccesses,
+                                unsigned NumStridedMemAccesses,
+                                unsigned NumPrefetches,
+                                bool HasCall) const override {
+    return Impl.getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
+                                     NumPrefetches, HasCall);
   }
 
   /// Return the maximum prefetch distance in terms of loop
   /// iterations.
   ///
   unsigned getMaxPrefetchIterationsAhead() const override {
     return Impl.getMaxPrefetchIterationsAhead();
   }
 
+  /// \return True if prefetching should also be done for writes.
+  bool enableWritePrefetching() const override {
+    return Impl.enableWritePrefetching();
+  }
+
   unsigned getMaxInterleaveFactor(unsigned VF) override {
     return Impl.getMaxInterleaveFactor(VF);
   }
   unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
                                             unsigned &JTSize,
                                             ProfileSummaryInfo *PSI,
                                             BlockFrequencyInfo *BFI) override {
     return Impl.getEstimatedNumberOfCaseClusters(SI, JTSize, PSI, BFI);

llvm/include/llvm/Analysis/TargetTransformInfoImpl.h

llvm::Optional<unsigned> getCacheAssociativity(
     case TargetTransformInfo::CacheLevel::L2D:
       return llvm::Optional<unsigned>();
     }
 
     llvm_unreachable("Unknown TargetTransformInfo::CacheLevel");
   }
 
   unsigned getPrefetchDistance() const { return 0; }
-  unsigned getMinPrefetchStride() const { return 1; }
+  unsigned getMinPrefetchStride(unsigned NumMemAccesses,
+                                unsigned NumStridedMemAccesses,
+                                unsigned NumPrefetches,
+                                bool HasCall) const { return 1; }
   unsigned getMaxPrefetchIterationsAhead() const { return UINT_MAX; }
+  bool enableWritePrefetching() const { return false; }
 
   unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
 
   unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
                                   TTI::OperandValueKind Opd1Info,
                                   TTI::OperandValueKind Opd2Info,
                                   TTI::OperandValueProperties Opd1PropInfo,
                                   TTI::OperandValueProperties Opd2PropInfo,

llvm/include/llvm/CodeGen/BasicTTIImpl.h

public:
   virtual unsigned getCacheLineSize() const {
     return getST()->getCacheLineSize();
   }
 
   virtual unsigned getPrefetchDistance() const {
     return getST()->getPrefetchDistance();
   }
 
-  virtual unsigned getMinPrefetchStride() const {
-    return getST()->getMinPrefetchStride();
+  virtual unsigned getMinPrefetchStride(unsigned NumMemAccesses,
+                                        unsigned NumStridedMemAccesses,
+                                        unsigned NumPrefetches,
+                                        bool HasCall) const {
+    return getST()->getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
+                                         NumPrefetches, HasCall);
   }
 
   virtual unsigned getMaxPrefetchIterationsAhead() const {
     return getST()->getMaxPrefetchIterationsAhead();
   }
 
+  virtual bool enableWritePrefetching() const {
+    return getST()->enableWritePrefetching();
+  }
+
   /// @}
 
   /// \name Vector TTI Implementations
   /// @{
 
   unsigned getRegisterBitWidth(bool Vector) const { return 32; }
 
   /// Estimate the overhead of scalarizing an instruction. Insert and Extract

llvm/include/llvm/MC/MCSubtargetInfo.h

public:
   ///
   virtual unsigned getPrefetchDistance() const;
 
   /// Return the maximum prefetch distance in terms of loop
   /// iterations.
   ///
   virtual unsigned getMaxPrefetchIterationsAhead() const;
 
+  /// \return True if prefetching should also be done for writes.
+  ///
+  virtual bool enableWritePrefetching() const;
+
   /// Return the minimum stride necessary to trigger software
   /// prefetching.
   ///
-  virtual unsigned getMinPrefetchStride() const;
+  virtual unsigned getMinPrefetchStride(unsigned NumMemAccesses,
+                                        unsigned NumStridedMemAccesses,
+                                        unsigned NumPrefetches,
+                                        bool HasCall) const;
 };
 
 } // end namespace llvm
 
 #endif // LLVM_MC_MCSUBTARGETINFO_H

llvm/lib/Analysis/TargetTransformInfo.cpp

llvm::Optional<unsigned> TargetTransformInfo::getCacheAssociativity(
   CacheLevel Level) const {
   return TTIImpl->getCacheAssociativity(Level);
 }
 
 unsigned TargetTransformInfo::getPrefetchDistance() const {
   return TTIImpl->getPrefetchDistance();
 }
 
-unsigned TargetTransformInfo::getMinPrefetchStride() const {
-  return TTIImpl->getMinPrefetchStride();
+unsigned TargetTransformInfo::getMinPrefetchStride(unsigned NumMemAccesses,
+                                                  unsigned NumStridedMemAccesses,
+                                                   unsigned NumPrefetches,
+                                                   bool HasCall) const {
+  return TTIImpl->getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
+                                       NumPrefetches, HasCall);
 }
 
 unsigned TargetTransformInfo::getMaxPrefetchIterationsAhead() const {
   return TTIImpl->getMaxPrefetchIterationsAhead();
 }
 
+bool TargetTransformInfo::enableWritePrefetching() const {
+  return TTIImpl->enableWritePrefetching();
+}
+
 unsigned TargetTransformInfo::getMaxInterleaveFactor(unsigned VF) const {
   return TTIImpl->getMaxInterleaveFactor(VF);
 }
 
 TargetTransformInfo::OperandValueKind
 TargetTransformInfo::getOperandInfo(Value *V, OperandValueProperties &OpProps) {
   OperandValueKind OpInfo = OK_AnyValue;
   OpProps = OP_None;

llvm/lib/MC/MCSubtargetInfo.cpp

 unsigned MCSubtargetInfo::getPrefetchDistance() const {
   return 0;
 }
 
 unsigned MCSubtargetInfo::getMaxPrefetchIterationsAhead() const {
   return UINT_MAX;
 }
 
-unsigned MCSubtargetInfo::getMinPrefetchStride() const {
+bool MCSubtargetInfo::enableWritePrefetching() const {
+  return false;
+}
+
+unsigned MCSubtargetInfo::getMinPrefetchStride(unsigned NumMemAccesses,
+                                               unsigned NumStridedMemAccesses,
+                                               unsigned NumPrefetches,
+                                               bool HasCall) const {
   return 1;
 }

llvm/lib/Target/AArch64/AArch64Subtarget.h

public:
   bool useRSqrt() const { return UseRSqrt; }
   bool force32BitJumpTables() const { return Force32BitJumpTables; }
   unsigned getMaxInterleaveFactor() const { return MaxInterleaveFactor; }
   unsigned getVectorInsertExtractBaseCost() const {
     return VectorInsertExtractBaseCost;
   }
   unsigned getCacheLineSize() const override { return CacheLineSize; }
   unsigned getPrefetchDistance() const override { return PrefetchDistance; }
-  unsigned getMinPrefetchStride() const override { return MinPrefetchStride; }
+  unsigned getMinPrefetchStride(unsigned NumMemAccesses,
+                                unsigned NumStridedMemAccesses,
+                                unsigned NumPrefetches,
+                                bool HasCall) const override {
+    return MinPrefetchStride;
+  }
   unsigned getMaxPrefetchIterationsAhead() const override {
     return MaxPrefetchIterationsAhead;
   }
   unsigned getPrefFunctionLogAlignment() const {
     return PrefFunctionLogAlignment;
   }
   unsigned getPrefLoopLogAlignment() const { return PrefLoopLogAlignment; }
 

llvm/lib/Target/SystemZ/SystemZTargetTransformInfo.h

public:
 
   /// \name Vector TTI Implementations
   /// @{
 
   unsigned getNumberOfRegisters(unsigned ClassID) const;
   unsigned getRegisterBitWidth(bool Vector) const;
 
   unsigned getCacheLineSize() const override { return 256; }
-  unsigned getPrefetchDistance() const override { return 2000; }
-  unsigned getMinPrefetchStride() const override { return 2048; }
+  unsigned getPrefetchDistance() const override { return 4500; }
+  unsigned getMinPrefetchStride(unsigned NumMemAccesses,
+                                unsigned NumStridedMemAccesses,
+                                unsigned NumPrefetches,
+                                bool HasCall) const override;
+  bool enableWritePrefetching() const override { return true; }
 
   bool hasDivRemOp(Type *DataType, bool IsSigned);
   bool prefersVectorizedAddressing() { return false; }
   bool LSRWithInstrQueries() { return true; }
   bool supportsEfficientVectorElementLoadStore() { return true; }
   bool enableInterleavedAccessVectorization() { return true; }
 
   int getArithmeticInstrCost(

llvm/lib/Target/SystemZ/SystemZTargetTransformInfo.cpp

 unsigned SystemZTTIImpl::getRegisterBitWidth(bool Vector) const {
   if (!Vector)
     return 64;
   if (ST->hasVector())
     return 128;
   return 0;
 }
 
+unsigned SystemZTTIImpl::getMinPrefetchStride(unsigned NumMemAccesses,
+                                              unsigned NumStridedMemAccesses,
+                                              unsigned NumPrefetches,
+                                              bool HasCall) const {
+  // Don't prefetch a loop with many far apart accesses.
+  if (NumPrefetches > 16)
+    return UINT_MAX;
+
+  // Emit prefetch instructions for smaller strides in cases where we think
+  // the hardware prefetcher might not be able to keep up.
+  if (NumStridedMemAccesses > 32 &&
+      NumStridedMemAccesses == NumMemAccesses && !HasCall)
+    return 1;
+
+  return ST->hasMiscellaneousExtensions3() ? 8192 : 2048;
+}
+
 bool SystemZTTIImpl::hasDivRemOp(Type *DataType, bool IsSigned) {
   EVT VT = TLI->getValueType(DL, DataType);
   return (VT.isScalarInteger() && TLI->isTypeLegal(VT));
 }
 
 // Return the bit size for the scalar type or vector element
 // type. getScalarSizeInBits() returns 0 for a pointer type.
 static unsigned getScalarSizeInBits(Type *Ty) {

llvm/lib/Transforms/Scalar/LoopDataPrefetch.cpp

 #include "llvm/Analysis/AssumptionCache.h"
 #include "llvm/Analysis/CodeMetrics.h"
 #include "llvm/Analysis/LoopInfo.h"
 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
 #include "llvm/Analysis/ScalarEvolution.h"
 #include "llvm/Analysis/ScalarEvolutionExpander.h"
 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/CodeGen/TargetLowering.h"
 #include "llvm/IR/CFG.h"
 #include "llvm/IR/Dominators.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/Module.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Transforms/Scalar.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"

 
 STATISTIC(NumPrefetches, "Number of prefetches inserted");
 
 namespace {
 
 /// Loop prefetch implementation class.
 class LoopDataPrefetch {
 public:
-  LoopDataPrefetch(AssumptionCache *AC, LoopInfo *LI, ScalarEvolution *SE,
-                   const TargetTransformInfo *TTI,
+  LoopDataPrefetch(AssumptionCache *AC, DominatorTree *DT, LoopInfo *LI,
+                   ScalarEvolution *SE, const TargetTransformInfo *TTI,
                    OptimizationRemarkEmitter *ORE)
-      : AC(AC), LI(LI), SE(SE), TTI(TTI), ORE(ORE) {}
+      : AC(AC), DT(DT), LI(LI), SE(SE), TTI(TTI), ORE(ORE) {}
 
   bool run();
 
 private:
   bool runOnLoop(Loop *L);
 
   /// Check if the stride of the accesses is large enough to
   /// warrant a prefetch.
-  bool isStrideLargeEnough(const SCEVAddRecExpr *AR);
+  bool isStrideLargeEnough(const SCEVAddRecExpr *AR, unsigned TargetMinStride);
 
-  unsigned getMinPrefetchStride() {
+  unsigned getMinPrefetchStride(unsigned NumMemAccesses,
+                                unsigned NumStridedMemAccesses,
+                                unsigned NumPrefetches,
+                                bool HasCall) {
     if (MinPrefetchStride.getNumOccurrences() > 0)
       return MinPrefetchStride;
-    return TTI->getMinPrefetchStride();
+    return TTI->getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
+                                     NumPrefetches, HasCall);
   }
 
   unsigned getPrefetchDistance() {
     if (PrefetchDistance.getNumOccurrences() > 0)
       return PrefetchDistance;
     return TTI->getPrefetchDistance();
   }
 
   unsigned getMaxPrefetchIterationsAhead() {
     if (MaxPrefetchIterationsAhead.getNumOccurrences() > 0)
       return MaxPrefetchIterationsAhead;
     return TTI->getMaxPrefetchIterationsAhead();
   }
 
+  bool doPrefetchWrites() {
+    if (PrefetchWrites.getNumOccurrences() > 0)
+      return PrefetchWrites;
+    return TTI->enableWritePrefetching();
+  }
+
   AssumptionCache *AC;
+  DominatorTree *DT;
   LoopInfo *LI;
   ScalarEvolution *SE;
   const TargetTransformInfo *TTI;
   OptimizationRemarkEmitter *ORE;
 };
 
 /// Legacy class for inserting loop data prefetches.
 class LoopDataPrefetchLegacyPass : public FunctionPass {
 public:
   static char ID; // Pass ID, replacement for typeid
   LoopDataPrefetchLegacyPass() : FunctionPass(ID) {
     initializeLoopDataPrefetchLegacyPassPass(*PassRegistry::getPassRegistry());
   }
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.addRequired<AssumptionCacheTracker>();
+    AU.addRequired<DominatorTreeWrapperPass>();
     AU.addPreserved<DominatorTreeWrapperPass>();
     AU.addRequired<LoopInfoWrapperPass>();
     AU.addPreserved<LoopInfoWrapperPass>();
     AU.addRequired<OptimizationRemarkEmitterWrapperPass>();
     AU.addRequired<ScalarEvolutionWrapperPass>();
     AU.addPreserved<ScalarEvolutionWrapperPass>();
     AU.addRequired<TargetTransformInfoWrapperPass>();
   }

 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
 INITIALIZE_PASS_END(LoopDataPrefetchLegacyPass, "loop-data-prefetch",
                     "Loop Data Prefetch", false, false)
 
 FunctionPass *llvm::createLoopDataPrefetchPass() {
   return new LoopDataPrefetchLegacyPass();
 }
 
-bool LoopDataPrefetch::isStrideLargeEnough(const SCEVAddRecExpr *AR) {
-  unsigned TargetMinStride = getMinPrefetchStride();
+bool LoopDataPrefetch::isStrideLargeEnough(const SCEVAddRecExpr *AR,
+                                           unsigned TargetMinStride) {
   // No need to check if any stride goes.
   if (TargetMinStride <= 1)
     return true;
 
   const auto *ConstStride = dyn_cast<SCEVConstant>(AR->getStepRecurrence(*SE));
   // If MinStride is set, don't prefetch unless we can ensure that stride is
   // larger.
   if (!ConstStride)
     return false;
 
   unsigned AbsStride = std::abs(ConstStride->getAPInt().getSExtValue());
   return TargetMinStride <= AbsStride;
 }
 
 PreservedAnalyses LoopDataPrefetchPass::run(Function &F,
                                             FunctionAnalysisManager &AM) {
+  DominatorTree *DT = &AM.getResult<DominatorTreeAnalysis>(F);
   LoopInfo *LI = &AM.getResult<LoopAnalysis>(F);
   ScalarEvolution *SE = &AM.getResult<ScalarEvolutionAnalysis>(F);
   AssumptionCache *AC = &AM.getResult<AssumptionAnalysis>(F);
   OptimizationRemarkEmitter *ORE =
       &AM.getResult<OptimizationRemarkEmitterAnalysis>(F);
   const TargetTransformInfo *TTI = &AM.getResult<TargetIRAnalysis>(F);
 
-  LoopDataPrefetch LDP(AC, LI, SE, TTI, ORE);
+  LoopDataPrefetch LDP(AC, DT, LI, SE, TTI, ORE);
   bool Changed = LDP.run();
 
   if (Changed) {
     PreservedAnalyses PA;
     PA.preserve<DominatorTreeAnalysis>();
     PA.preserve<LoopAnalysis>();
     return PA;
   }
 
   return PreservedAnalyses::all();
 }
 
 bool LoopDataPrefetchLegacyPass::runOnFunction(Function &F) {
   if (skipFunction(F))
     return false;
 
+  DominatorTree *DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
   LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
   ScalarEvolution *SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
   AssumptionCache *AC =
       &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
   OptimizationRemarkEmitter *ORE =
       &getAnalysis<OptimizationRemarkEmitterWrapperPass>().getORE();
   const TargetTransformInfo *TTI =
       &getAnalysis<TargetTransformInfoWrapperPass>().getTTI(F);
 
-  LoopDataPrefetch LDP(AC, LI, SE, TTI, ORE);
+  LoopDataPrefetch LDP(AC, DT, LI, SE, TTI, ORE);
   return LDP.run();
 }
 
 bool LoopDataPrefetch::run() {
   // If PrefetchDistance is not set, don't run the pass.  This gives an
   // opportunity for targets to run this pass for selected subtargets only
   // (whose TTI sets PrefetchDistance).
   if (getPrefetchDistance() == 0)
     return false;
   assert(TTI->getCacheLineSize() && "Cache line size is not set for target");
 
   bool MadeChange = false;
 
   for (Loop *I : *LI)
     for (auto L = df_begin(I), LE = df_end(I); L != LE; ++L)
       MadeChange |= runOnLoop(*L);
 
   return MadeChange;
 }
 
+struct Prefetch {
+  const SCEVAddRecExpr *LSCEVAddRec;
+  Instruction *InsertPt;
+  bool Writes;
+  /// The prefetched instruction (for debug output)
+  Instruction* MemI;
+  Prefetch(const SCEVAddRecExpr *lscev, Instruction *i) :
+    LSCEVAddRec(lscev), InsertPt(nullptr), Writes(false), MemI(nullptr) {
+    addInstruction(i);
+  };
+  void addInstruction(Instruction *I, DominatorTree *DT = nullptr,
+                      int64_t PtrDiff = 0) {
+    if (!InsertPt) {
+      MemI = I;
+      InsertPt = I;
+      Writes = isa<StoreInst>(I);
+    }
+    else {
+      BasicBlock *PrefBB = InsertPt->getParent();
+      BasicBlock *InsBB = I->getParent();
+      if (PrefBB != InsBB) {
+        BasicBlock *DomBB =DT->findNearestCommonDominator(PrefBB, InsBB);
+        if (DomBB != PrefBB)
+          InsertPt = DomBB->getTerminator();
+      }
+
+      if (isa<StoreInst>(I) && PtrDiff == 0)
+        Writes = true;
+    }
+  }
+};
+
 bool LoopDataPrefetch::runOnLoop(Loop *L) {
   bool MadeChange = false;
 
   // Only prefetch in the inner-most loop
   if (!L->empty())
     return MadeChange;
 
   SmallPtrSet<const Value *, 32> EphValues;
   CodeMetrics::collectEphemeralValues(L, AC, EphValues);
 
   // Calculate the number of iterations ahead to prefetch
   CodeMetrics Metrics;
+  bool HasCall = false;
   for (const auto BB : L->blocks()) {
     // If the loop already has prefetches, then assume that the user knows
     // what they are doing and don't add any more.
-    for (auto &I : *BB)
-      if (CallInst *CI = dyn_cast<CallInst>(&I))
-        if (Function *F = CI->getCalledFunction())
+    for (auto &I : *BB) {
+      if (isa<CallInst>(&I) || isa<InvokeInst>(&I)) {
+        ImmutableCallSite CS(&I);
+        if (const Function *F = CS.getCalledFunction()) {
           if (F->getIntrinsicID() == Intrinsic::prefetch)
             return MadeChange;
-
+          if (TTI->isLoweredToCall(F))
+            HasCall = true;
+        } else { // indirect call.
+          HasCall = true;
+        }
+      }
+    }
     Metrics.analyzeBasicBlock(BB, *TTI, EphValues);
   }
   unsigned LoopSize = Metrics.NumInsts;
   if (!LoopSize)
     LoopSize = 1;
 
   unsigned ItersAhead = getPrefetchDistance() / LoopSize;
   if (!ItersAhead)
     ItersAhead = 1;
 
   if (ItersAhead > getMaxPrefetchIterationsAhead())
     return MadeChange;
 
-  LLVM_DEBUG(dbgs() << "Prefetching " << ItersAhead
-                    << " iterations ahead (loop size: " << LoopSize << ") in "
-                    << L->getHeader()->getParent()->getName() << ": " << *L);
+  unsigned ConstantMaxTripCount = SE->getSmallConstantMaxTripCount(L);
+  if (ConstantMaxTripCount && ConstantMaxTripCount < ItersAhead + 1)
+    return MadeChange;
 
-  SmallVector<std::pair<Instruction *, const SCEVAddRecExpr *>, 16> PrefLoads;
-  for (const auto BB : L->blocks()) {
+  unsigned NumMemAccesses = 0;
+  unsigned NumStridedMemAccesses = 0;
+  SmallVector<Prefetch, 16> Prefetches;
+  for (const auto BB : L->blocks())
     for (auto &I : *BB) {
       Value *PtrValue;
       Instruction *MemI;
 
       if (LoadInst *LMemI = dyn_cast<LoadInst>(&I)) {
         MemI = LMemI;
         PtrValue = LMemI->getPointerOperand();
       } else if (StoreInst *SMemI = dyn_cast<StoreInst>(&I)) {
-        if (!PrefetchWrites) continue;
+        if (!doPrefetchWrites()) continue;
         MemI = SMemI;
         PtrValue = SMemI->getPointerOperand();
       } else continue;
 
       unsigned PtrAddrSpace = PtrValue->getType()->getPointerAddressSpace();
       if (PtrAddrSpace)
         continue;
-
+      NumMemAccesses++;
       if (L->isLoopInvariant(PtrValue))
         continue;
 
       const SCEV *LSCEV = SE->getSCEV(PtrValue);
       const SCEVAddRecExpr *LSCEVAddRec = dyn_cast<SCEVAddRecExpr>(LSCEV);
       if (!LSCEVAddRec)
         continue;
+      NumStridedMemAccesses++;
 
-      // Check if the stride of the accesses is large enough to warrant a
-      // prefetch.
-      if (!isStrideLargeEnough(LSCEVAddRec))
-        continue;
-
-      // We don't want to double prefetch individual cache lines. If this load
-      // is known to be within one cache line of some other load that has
-      // already been prefetched, then don't prefetch this one as well.
+      // We don't want to double prefetch individual cache lines. If this
+      // access is known to be within one cache line of some other one that
+      // has already been prefetched, then don't prefetch this one as well.
       bool DupPref = false;
-      for (const auto &PrefLoad : PrefLoads) {
-        const SCEV *PtrDiff = SE->getMinusSCEV(LSCEVAddRec, PrefLoad.second);
+      for (auto &Pref : Prefetches) {
+        const SCEV *PtrDiff = SE->getMinusSCEV(LSCEVAddRec, Pref.LSCEVAddRec);
         if (const SCEVConstant *ConstPtrDiff =
             dyn_cast<SCEVConstant>(PtrDiff)) {
           int64_t PD = std::abs(ConstPtrDiff->getValue()->getSExtValue());
           if (PD < (int64_t) TTI->getCacheLineSize()) {
+            Pref.addInstruction(MemI, DT, PD);
             DupPref = true;
             break;
           }
         }
       }
-      if (DupPref)
-        continue;
+      if (!DupPref)
+        Prefetches.push_back(Prefetch(LSCEVAddRec, MemI));
+    }
+
+  unsigned TargetMinStride =
+    getMinPrefetchStride(NumMemAccesses, NumStridedMemAccesses,
+                         Prefetches.size(), HasCall);
+
+  LLVM_DEBUG(dbgs() << "Prefetching " << ItersAhead
+             << " iterations ahead (loop size: " << LoopSize << ") in "
+             << L->getHeader()->getParent()->getName() << ": " << *L);
+  LLVM_DEBUG(dbgs() << "Loop has: "
+             << NumMemAccesses << " memory accesses, "
+             << NumStridedMemAccesses << " strided memory accesses, "
+             << Prefetches.size() << " potential prefetch(es), "
+             << "a minimum stride of " << TargetMinStride << ", "
+             << (HasCall ? "calls" : "no calls") << ".\n");
+
+  for (auto &P : Prefetches) {
+    // Check if the stride of the accesses is large enough to warrant a
+    // prefetch.
+    if (!isStrideLargeEnough(P.LSCEVAddRec, TargetMinStride))
 
-      const SCEV *NextLSCEV = SE->getAddExpr(LSCEVAddRec, SE->getMulExpr(
-        SE->getConstant(LSCEVAddRec->getType(), ItersAhead),
-        LSCEVAddRec->getStepRecurrence(*SE)));
-      if (!isSafeToExpand(NextLSCEV, *SE))
-        continue;
+    const SCEV *NextLSCEV = SE->getAddExpr(P.LSCEVAddRec, SE->getMulExpr(
+      SE->getConstant(P.LSCEVAddRec->getType(), ItersAhead),
+      P.LSCEVAddRec->getStepRecurrence(*SE)));
 
-      PrefLoads.push_back(std::make_pair(MemI, LSCEVAddRec));
+    BasicBlock *BB = P.InsertPt->getParent();
+    Type *I8Ptr = Type::getInt8PtrTy(BB->getContext(), 0/*PtrAddrSpace*/);
+    SCEVExpander SCEVE(*SE, BB->getModule()->getDataLayout(), "prefaddr");
+    Value *PrefPtrValue = SCEVE.expandCodeFor(NextLSCEV, I8Ptr, P.InsertPt);
 
-      Type *I8Ptr = Type::getInt8PtrTy(BB->getContext(), PtrAddrSpace);
-      SCEVExpander SCEVE(*SE, I.getModule()->getDataLayout(), "prefaddr");
-      Value *PrefPtrValue = SCEVE.expandCodeFor(NextLSCEV, I8Ptr, MemI);
-
-      IRBuilder<> Builder(MemI);
-      Module *M = BB->getParent()->getParent();
-      Type *I32 = Type::getInt32Ty(BB->getContext());
-      Function *PrefetchFunc = Intrinsic::getDeclaration(
-          M, Intrinsic::prefetch, PrefPtrValue->getType());
-      Builder.CreateCall(
-          PrefetchFunc,
-          {PrefPtrValue,
-           ConstantInt::get(I32, MemI->mayReadFromMemory() ? 0 : 1),
-           ConstantInt::get(I32, 3), ConstantInt::get(I32, 1)});
-      ++NumPrefetches;
-      LLVM_DEBUG(dbgs() << "  Access: " << *PtrValue << ", SCEV: " << *LSCEV
-                        << "\n");
-      ORE->emit([&]() {
-        return OptimizationRemark(DEBUG_TYPE, "Prefetched", MemI)
-               << "prefetched memory access";
+    IRBuilder<> Builder(P.InsertPt);
+         ConstantInt::get(I32, P.Writes),
+    LLVM_DEBUG(dbgs() << "  Access: "
+               << *P.MemI->getOperand(isa<LoadInst>(P.MemI) ? 0 : 1)
+               << ", SCEV: " << *P.LSCEVAddRec << "\n");
+        return OptimizationRemark(DEBUG_TYPE, "Prefetched", P.MemI)
       });
 
-      MadeChange = true;
-    }
-  }
 
   return MadeChange;
 }

llvm/test/CodeGen/SystemZ/prefetch-02.ll

+; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z14 -prefetch-distance=100 \
+; RUN:   -stop-after=loop-data-prefetch | FileCheck %s -check-prefix=FAR-PREFETCH
+; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z14 -prefetch-distance=20 \
+; RUN:   -stop-after=loop-data-prefetch | FileCheck %s -check-prefix=NEAR-PREFETCH
+;
+; Check that prefetches are not emitted when the known constant trip count of
+; the loop is smaller than the estimated "iterations ahead" of the prefetch.
+;
+; FAR-PREFETCH-LABEL: fun
+; FAR-PREFETCH-NOT: call void @llvm.prefetch
+
+; NEAR-PREFETCH-LABEL: fun
+; NEAR-PREFETCH: call void @llvm.prefetch
+
+
+define void @fun(i32* nocapture %Src, i32* nocapture readonly %Dst) {
+entry:
+  br label %for.body
+
+for.cond.cleanup:                                 ; preds = %for.body
+  ret void
+
+for.body:                                         ; preds = %for.body, %entry
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next.9, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %Dst, i64 %indvars.iv
+  %0 = load i32, i32* %arrayidx, align 4
+  %arrayidx2 = getelementptr inbounds i32, i32* %Src, i64 %indvars.iv
+  store i32 %0, i32* %arrayidx2, align 4
+  %indvars.iv.next.9 = add nuw nsw i64 %indvars.iv, 1600
+  %cmp.9 = icmp ult i64 %indvars.iv.next.9, 11200
+  br i1 %cmp.9, label %for.body, label %for.cond.cleanup
+}
+

llvm/test/CodeGen/SystemZ/prefetch-03.ll

+; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z14 -prefetch-distance=50 \
+; RUN:   -loop-prefetch-writes -stop-after=loop-data-prefetch | FileCheck %s
+;
+; Check that prefetches are emitted in a position that is executed each
+; iteration for each targeted memory instruction. The two stores in %true and
+; %false are within one cache line in memory, so they should get a single
+; prefetch in %for.body.
+;
+; CHECK-LABEL: for.body
+; CHECK: call void @llvm.prefetch.p0i8(i8* {{.*}}, i32 0
+; CHECK: call void @llvm.prefetch.p0i8(i8* {{.*}}, i32 1
+; CHECK-LABEL: true
+; CHECK-LABEL: false
+; CHECK-LABEL: latch
+
+define void @fun(i32* nocapture %Src, i32* nocapture readonly %Dst) {
+entry:
+  br label %for.body
+
+for.body:
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next.9, %latch ]
+  %arrayidx = getelementptr inbounds i32, i32* %Dst, i64 %indvars.iv
+  %0 = load i32, i32* %arrayidx, align 4
+  %cmp = icmp sgt i32 %0, 0
+  br i1 %cmp, label %true, label %false
+
+true:  
+  %arrayidx2 = getelementptr inbounds i32, i32* %Src, i64 %indvars.iv
+  store i32 %0, i32* %arrayidx2, align 4
+  br label %latch
+
+false:
+  %a = add i64 %indvars.iv, 8
+  %arrayidx3 = getelementptr inbounds i32, i32* %Src, i64 %a
+  store i32 %0, i32* %arrayidx3, align 4
+  br label %latch
+
+latch:
+  %indvars.iv.next.9 = add nuw nsw i64 %indvars.iv, 1600
+  %cmp.9 = icmp ult i64 %indvars.iv.next.9, 11200
+  br i1 %cmp.9, label %for.body, label %for.cond.cleanup
+
+for.cond.cleanup:
+  ret void
+}
+

llvm/test/CodeGen/SystemZ/prefetch-04.ll

+; RUN: llc < %s -mtriple=s390x-linux-gnu -mcpu=z14 -prefetch-distance=20 \
+; RUN:   -loop-prefetch-writes -stop-after=loop-data-prefetch | FileCheck %s
+;
+; Check that for a load followed by a store to the same address gets a single
+; write prefetch.
+;
+; CHECK-LABEL: for.body
+; CHECK: call void @llvm.prefetch.p0i8(i8* %scevgep{{.*}}, i32 1, i32 3, i32 1
+; CHECK-not: call void @llvm.prefetch
+
+define void @fun(i32* nocapture %Src, i32* nocapture readonly %Dst) {
+entry:
+  br label %for.body
+
+for.body:
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next.9, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %Dst, i64 %indvars.iv
+  %0 = load i32, i32* %arrayidx, align 4
+  %a = add i32 %0, 128
+  store i32 %a, i32* %arrayidx, align 4
+  %indvars.iv.next.9 = add nuw nsw i64 %indvars.iv, 1600
+  %cmp.9 = icmp ult i64 %indvars.iv.next.9, 11200
+  br i1 %cmp.9, label %for.body, label %for.cond.cleanup
+
+for.cond.cleanup:
+  ret void
+}
+