//===----------------------------------------------------------------------===//
//
// This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops
-// and generates target-independent LLVM-IR. Legalization of the IR is done
-// in the codegen. However, the vectorizes uses (will use) the codegen
-// interfaces to generate IR that is likely to result in an optimal binary.
+// and generates target-independent LLVM-IR.
+// The vectorizer uses the TargetTransformInfo analysis to estimate the costs
+// of instructions in order to estimate the profitability of vectorization.
//
-// The loop vectorizer combines consecutive loop iteration into a single
+// The loop vectorizer combines consecutive loop iterations into a single
// 'wide' iteration. After this transformation the index is incremented
// by the SIMD vector width, and not by one.
//
// D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.
//
// Variable uniformity checks are inspired by:
-// Karrenberg, R. and Hack, S. Whole Function Vectorization.
+// Karrenberg, R. and Hack, S. Whole Function Vectorization.
//
// Other ideas/concepts are from:
// A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.
#include "llvm/Pass.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
+#include "llvm/Support/PatternMatch.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/Support/ValueHandle.h"
+#include "llvm/Target/TargetLibraryInfo.h"
#include "llvm/Transforms/Scalar.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Local.h"
#include <map>
using namespace llvm;
+using namespace llvm::PatternMatch;
static cl::opt<unsigned>
VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
cl::desc("Enable if-conversion during vectorization."));
/// We don't vectorize loops with a known constant trip count below this number.
-static const unsigned TinyTripCountThreshold = 16;
+static cl::opt<unsigned>
+TinyTripCountVectorThreshold("vectorizer-min-trip-count", cl::init(16),
+ cl::Hidden,
+ cl::desc("Don't vectorize loops with a constant "
+ "trip count that is smaller than this "
+ "value."));
+
+/// We don't unroll loops with a known constant trip count below this number.
+static const unsigned TinyTripCountUnrollThreshold = 128;
-/// When performing a runtime memory check, do not check more than this
-/// number of pointers. Notice that the check is quadratic!
-static const unsigned RuntimeMemoryCheckThreshold = 4;
+/// When performing memory disambiguation checks at runtime do not make more
+/// than this number of comparisons.
+static const unsigned RuntimeMemoryCheckThreshold = 8;
-/// This is the highest vector width that we try to generate.
-static const unsigned MaxVectorSize = 8;
+/// Maximum simd width.
+static const unsigned MaxVectorWidth = 64;
-/// This is the highest Unroll Factor.
-static const unsigned MaxUnrollSize = 4;
+/// Maximum vectorization unroll count.
+static const unsigned MaxUnrollFactor = 16;
namespace {
class InnerLoopVectorizer {
public:
InnerLoopVectorizer(Loop *OrigLoop, ScalarEvolution *SE, LoopInfo *LI,
- DominatorTree *DT, DataLayout *DL, unsigned VecWidth,
+ DominatorTree *DT, DataLayout *DL,
+ const TargetLibraryInfo *TLI, unsigned VecWidth,
unsigned UnrollFactor)
- : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), VF(VecWidth),
- UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
+ : OrigLoop(OrigLoop), SE(SE), LI(LI), DT(DT), DL(DL), TLI(TLI),
+ VF(VecWidth), UF(UnrollFactor), Builder(SE->getContext()), Induction(0),
OldInduction(0), WidenMap(UnrollFactor) {}
// Perform the actual loop widening (vectorization).
/// Add code that checks at runtime if the accessed arrays overlap.
/// Returns the comparator value or NULL if no check is needed.
- Value *addRuntimeCheck(LoopVectorizationLegality *Legal,
- Instruction *Loc);
+ Instruction *addRuntimeCheck(LoopVectorizationLegality *Legal,
+ Instruction *Loc);
/// Create an empty loop, based on the loop ranges of the old loop.
void createEmptyLoop(LoopVectorizationLegality *Legal);
/// Copy and widen the instructions from the old loop.
/// of scalars.
void scalarizeInstruction(Instruction *Instr);
+ /// Vectorize Load and Store instructions,
+ void vectorizeMemoryInstruction(Instruction *Instr,
+ LoopVectorizationLegality *Legal);
+
/// Create a broadcast instruction. This method generates a broadcast
/// instruction (shuffle) for loop invariant values and for the induction
/// value. If this is the induction variable then we extend it to N, N+1, ...
/// This function adds 0, 1, 2 ... to each vector element, starting at zero.
/// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
/// The sequence starts at StartIndex.
- Value *getConsecutiveVector(Value* Val, unsigned StartIdx, bool Negate);
+ Value *getConsecutiveVector(Value* Val, int StartIdx, bool Negate);
/// When we go over instructions in the basic block we rely on previous
/// values within the current basic block or on loop invariant values.
/// broadcast them into a vector.
VectorParts &getVectorValue(Value *V);
- /// Get a uniform vector of constant integers. We use this to get
- /// vectors of ones and zeros for the reduction code.
- Constant* getUniformVector(unsigned Val, Type* ScalarTy);
-
/// Generate a shuffle sequence that will reverse the vector Vec.
Value *reverseVector(Value *Vec);
ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
/// \return True if 'Key' is saved in the Value Map.
- bool has(Value *Key) { return MapStoreage.count(Key); }
+ bool has(Value *Key) const { return MapStorage.count(Key); }
/// Initializes a new entry in the map. Sets all of the vector parts to the
/// save value in 'Val'.
/// \return A reference to a vector with splat values.
VectorParts &splat(Value *Key, Value *Val) {
- MapStoreage[Key].clear();
- MapStoreage[Key].append(UF, Val);
- return MapStoreage[Key];
+ VectorParts &Entry = MapStorage[Key];
+ Entry.assign(UF, Val);
+ return Entry;
}
///\return A reference to the value that is stored at 'Key'.
VectorParts &get(Value *Key) {
- if (!has(Key))
- MapStoreage[Key].resize(UF);
- return MapStoreage[Key];
+ VectorParts &Entry = MapStorage[Key];
+ if (Entry.empty())
+ Entry.resize(UF);
+ assert(Entry.size() == UF);
+ return Entry;
}
+ private:
/// The unroll factor. Each entry in the map stores this number of vector
/// elements.
unsigned UF;
/// Map storage. We use std::map and not DenseMap because insertions to a
/// dense map invalidates its iterators.
- std::map<Value*, VectorParts> MapStoreage;
+ std::map<Value *, VectorParts> MapStorage;
};
/// The original loop.
DominatorTree *DT;
/// Data Layout.
DataLayout *DL;
+ /// Target Library Info.
+ const TargetLibraryInfo *TLI;
+
/// The vectorization SIMD factor to use. Each vector will have this many
/// vector elements.
unsigned VF;
BasicBlock *LoopVectorBody;
///The scalar loop body.
BasicBlock *LoopScalarBody;
- ///The first bypass block.
- BasicBlock *LoopBypassBlock;
+ /// A list of all bypass blocks. The first block is the entry of the loop.
+ SmallVector<BasicBlock *, 4> LoopBypassBlocks;
/// The new Induction variable which was added to the new block.
PHINode *Induction;
/// The induction variable of the old basic block.
PHINode *OldInduction;
+ /// Holds the extended (to the widest induction type) start index.
+ Value *ExtendedIdx;
/// Maps scalars to widened vectors.
ValueMap WidenMap;
};
+/// \brief Check if conditionally executed loads are hoistable.
+///
+/// This class has two functions: isHoistableLoad and canHoistAllLoads.
+/// isHoistableLoad should be called on all load instructions that are executed
+/// conditionally. After all conditional loads are processed, the client should
+/// call canHoistAllLoads to determine if all of the conditional executed loads
+/// have an unconditional memory access to the same memory address in the loop.
+class LoadHoisting {
+ typedef SmallPtrSet<Value *, 8> MemorySet;
+
+ Loop *TheLoop;
+ DominatorTree *DT;
+ MemorySet CondLoadAddrSet;
+
+public:
+ LoadHoisting(Loop *L, DominatorTree *D) : TheLoop(L), DT(D) {}
+
+ /// \brief Check if the instruction is a load with a identifiable address.
+ bool isHoistableLoad(Instruction *L);
+
+ /// \brief Check if all of the conditional loads are hoistable because there
+ /// exists an unconditional memory access to the same address in the loop.
+ bool canHoistAllLoads();
+};
+
+bool LoadHoisting::isHoistableLoad(Instruction *L) {
+ LoadInst *LI = dyn_cast<LoadInst>(L);
+ if (!LI)
+ return false;
+
+ CondLoadAddrSet.insert(LI->getPointerOperand());
+ return true;
+}
+
+static void addMemAccesses(BasicBlock *BB, SmallPtrSet<Value *, 8> &Set) {
+ for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE; ++BI) {
+ if (LoadInst *LI = dyn_cast<LoadInst>(BI)) // Try a load.
+ Set.insert(LI->getPointerOperand());
+ else if (StoreInst *SI = dyn_cast<StoreInst>(BI)) // Try a store.
+ Set.insert(SI->getPointerOperand());
+ }
+}
+
+bool LoadHoisting::canHoistAllLoads() {
+ // No conditional loads.
+ if (CondLoadAddrSet.empty())
+ return true;
+
+ MemorySet UncondMemAccesses;
+ std::vector<BasicBlock*> &LoopBlocks = TheLoop->getBlocksVector();
+ BasicBlock *LoopLatch = TheLoop->getLoopLatch();
+
+ // Iterate over the unconditional blocks and collect memory access addresses.
+ for (unsigned i = 0, e = LoopBlocks.size(); i < e; ++i) {
+ BasicBlock *BB = LoopBlocks[i];
+
+ // Ignore conditional blocks.
+ if (BB != LoopLatch && !DT->dominates(BB, LoopLatch))
+ continue;
+
+ addMemAccesses(BB, UncondMemAccesses);
+ }
+
+ // And make sure there is a matching unconditional access for every
+ // conditional load.
+ for (MemorySet::iterator MI = CondLoadAddrSet.begin(),
+ ME = CondLoadAddrSet.end(); MI != ME; ++MI)
+ if (!UncondMemAccesses.count(*MI))
+ return false;
+
+ return true;
+}
+
/// LoopVectorizationLegality checks if it is legal to vectorize a loop, and
/// to what vectorization factor.
/// This class does not look at the profitability of vectorization, only the
class LoopVectorizationLegality {
public:
LoopVectorizationLegality(Loop *L, ScalarEvolution *SE, DataLayout *DL,
- DominatorTree *DT)
- : TheLoop(L), SE(SE), DL(DL), DT(DT), Induction(0) {}
+ DominatorTree *DT, TargetTransformInfo* TTI,
+ AliasAnalysis *AA, TargetLibraryInfo *TLI)
+ : TheLoop(L), SE(SE), DL(DL), DT(DT), TTI(TTI), AA(AA), TLI(TLI),
+ Induction(0), WidestIndTy(0), HasFunNoNaNAttr(false),
+ LoadSpeculation(L, DT) {}
/// This enum represents the kinds of reductions that we support.
enum ReductionKind {
- NoReduction, ///< Not a reduction.
- IntegerAdd, ///< Sum of numbers.
- IntegerMult, ///< Product of numbers.
- IntegerOr, ///< Bitwise or logical OR of numbers.
- IntegerAnd, ///< Bitwise or logical AND of numbers.
- IntegerXor ///< Bitwise or logical XOR of numbers.
+ RK_NoReduction, ///< Not a reduction.
+ RK_IntegerAdd, ///< Sum of integers.
+ RK_IntegerMult, ///< Product of integers.
+ RK_IntegerOr, ///< Bitwise or logical OR of numbers.
+ RK_IntegerAnd, ///< Bitwise or logical AND of numbers.
+ RK_IntegerXor, ///< Bitwise or logical XOR of numbers.
+ RK_IntegerMinMax, ///< Min/max implemented in terms of select(cmp()).
+ RK_FloatAdd, ///< Sum of floats.
+ RK_FloatMult, ///< Product of floats.
+ RK_FloatMinMax ///< Min/max implemented in terms of select(cmp()).
};
/// This enum represents the kinds of inductions that we support.
enum InductionKind {
- NoInduction, ///< Not an induction variable.
- IntInduction, ///< Integer induction variable. Step = 1.
- ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
- PtrInduction ///< Pointer induction variable. Step = sizeof(elem).
+ IK_NoInduction, ///< Not an induction variable.
+ IK_IntInduction, ///< Integer induction variable. Step = 1.
+ IK_ReverseIntInduction, ///< Reverse int induction variable. Step = -1.
+ IK_PtrInduction, ///< Pointer induction var. Step = sizeof(elem).
+ IK_ReversePtrInduction ///< Reverse ptr indvar. Step = - sizeof(elem).
+ };
+
+ // This enum represents the kind of minmax reduction.
+ enum MinMaxReductionKind {
+ MRK_Invalid,
+ MRK_UIntMin,
+ MRK_UIntMax,
+ MRK_SIntMin,
+ MRK_SIntMax,
+ MRK_FloatMin,
+ MRK_FloatMax
};
/// This POD struct holds information about reduction variables.
struct ReductionDescriptor {
- ReductionDescriptor() : StartValue(0), LoopExitInstr(0), Kind(NoReduction) {
- }
+ ReductionDescriptor() : StartValue(0), LoopExitInstr(0),
+ Kind(RK_NoReduction), MinMaxKind(MRK_Invalid) {}
- ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K)
- : StartValue(Start), LoopExitInstr(Exit), Kind(K) {}
+ ReductionDescriptor(Value *Start, Instruction *Exit, ReductionKind K,
+ MinMaxReductionKind MK)
+ : StartValue(Start), LoopExitInstr(Exit), Kind(K), MinMaxKind(MK) {}
// The starting value of the reduction.
// It does not have to be zero!
- Value *StartValue;
+ TrackingVH<Value> StartValue;
// The instruction who's value is used outside the loop.
Instruction *LoopExitInstr;
// The kind of the reduction.
ReductionKind Kind;
+ // If this a min/max reduction the kind of reduction.
+ MinMaxReductionKind MinMaxKind;
+ };
+
+ /// This POD struct holds information about a potential reduction operation.
+ struct ReductionInstDesc {
+ ReductionInstDesc(bool IsRedux, Instruction *I) :
+ IsReduction(IsRedux), PatternLastInst(I), MinMaxKind(MRK_Invalid) {}
+
+ ReductionInstDesc(Instruction *I, MinMaxReductionKind K) :
+ IsReduction(true), PatternLastInst(I), MinMaxKind(K) {}
+
+ // Is this instruction a reduction candidate.
+ bool IsReduction;
+ // The last instruction in a min/max pattern (select of the select(icmp())
+ // pattern), or the current reduction instruction otherwise.
+ Instruction *PatternLastInst;
+ // If this is a min/max pattern the comparison predicate.
+ MinMaxReductionKind MinMaxKind;
};
// This POD struct holds information about the memory runtime legality
}
/// Insert a pointer and calculate the start and end SCEVs.
- void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr);
+ void insert(ScalarEvolution *SE, Loop *Lp, Value *Ptr, bool WritePtr);
/// This flag indicates if we need to add the runtime check.
bool Need;
/// Holds the pointers that we need to check.
- SmallVector<Value*, 2> Pointers;
+ SmallVector<TrackingVH<Value>, 2> Pointers;
/// Holds the pointer value at the beginning of the loop.
SmallVector<const SCEV*, 2> Starts;
/// Holds the pointer value at the end of the loop.
SmallVector<const SCEV*, 2> Ends;
+ /// Holds the information if this pointer is used for writing to memory.
+ SmallVector<bool, 2> IsWritePtr;
};
/// A POD for saving information about induction variables.
struct InductionInfo {
InductionInfo(Value *Start, InductionKind K) : StartValue(Start), IK(K) {}
- InductionInfo() : StartValue(0), IK(NoInduction) {}
+ InductionInfo() : StartValue(0), IK(IK_NoInduction) {}
/// Start value.
- Value *StartValue;
+ TrackingVH<Value> StartValue;
/// Induction kind.
InductionKind IK;
};
/// induction descriptor.
typedef MapVector<PHINode*, InductionInfo> InductionList;
+ /// Alias(Multi)Map stores the values (GEPs or underlying objects and their
+ /// respective Store/Load instruction(s) to calculate aliasing.
+ typedef MapVector<Value*, Instruction* > AliasMap;
+ typedef DenseMap<Value*, std::vector<Instruction*> > AliasMultiMap;
+
/// Returns true if it is legal to vectorize this loop.
/// This does not mean that it is profitable to vectorize this
/// loop, only that it is legal to do so.
/// Returns the induction variables found in the loop.
InductionList *getInductionVars() { return &Inductions; }
+ /// Returns the widest induction type.
+ Type *getWidestInductionType() { return WidestIndTy; }
+
/// Returns True if V is an induction variable in this loop.
bool isInductionVariable(const Value *V);
/// Returns the information that we collected about runtime memory check.
RuntimePointerCheck *getRuntimePointerCheck() { return &PtrRtCheck; }
+
+ /// This function returns the identity element (or neutral element) for
+ /// the operation K.
+ static Constant *getReductionIdentity(ReductionKind K, Type *Tp);
private:
/// Check if a single basic block loop is vectorizable.
/// At this point we know that this is a loop with a constant trip count
/// Returns True, if 'Phi' is the kind of reduction variable for type
/// 'Kind'. If this is a reduction variable, it adds it to ReductionList.
bool AddReductionVar(PHINode *Phi, ReductionKind Kind);
- /// Returns true if the instruction I can be a reduction variable of type
- /// 'Kind'.
- bool isReductionInstr(Instruction *I, ReductionKind Kind);
+ /// Returns a struct describing if the instruction 'I' can be a reduction
+ /// variable of type 'Kind'. If the reduction is a min/max pattern of
+ /// select(icmp()) this function advances the instruction pointer 'I' from the
+ /// compare instruction to the select instruction and stores this pointer in
+ /// 'PatternLastInst' member of the returned struct.
+ ReductionInstDesc isReductionInstr(Instruction *I, ReductionKind Kind,
+ ReductionInstDesc &Desc);
+ /// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
+ /// pattern corresponding to a min(X, Y) or max(X, Y).
+ static ReductionInstDesc isMinMaxSelectCmpPattern(Instruction *I,
+ ReductionInstDesc &Prev);
/// Returns the induction kind of Phi. This function may return NoInduction
/// if the PHI is not an induction variable.
InductionKind isInductionVariable(PHINode *Phi);
/// Return true if can compute the address bounds of Ptr within the loop.
bool hasComputableBounds(Value *Ptr);
+ /// Return true if there is the chance of write reorder.
+ bool hasPossibleGlobalWriteReorder(Value *Object,
+ Instruction *Inst,
+ AliasMultiMap &WriteObjects,
+ unsigned MaxByteWidth);
+ /// Return the AA location for a load or a store.
+ AliasAnalysis::Location getLoadStoreLocation(Instruction *Inst);
+
/// The loop that we evaluate.
Loop *TheLoop;
ScalarEvolution *SE;
/// DataLayout analysis.
DataLayout *DL;
- // Dominators.
+ /// Dominators.
DominatorTree *DT;
+ /// Target Info.
+ TargetTransformInfo *TTI;
+ /// Alias Analysis.
+ AliasAnalysis *AA;
+ /// Target Library Info.
+ TargetLibraryInfo *TLI;
// --- vectorization state --- //
/// Notice that inductions don't need to start at zero and that induction
/// variables can be pointers.
InductionList Inductions;
+ /// Holds the widest induction type encountered.
+ Type *WidestIndTy;
/// Allowed outside users. This holds the reduction
/// vars which can be accessed from outside the loop.
/// We need to check that all of the pointers in this list are disjoint
/// at runtime.
RuntimePointerCheck PtrRtCheck;
+ /// Can we assume the absence of NaNs.
+ bool HasFunNoNaNAttr;
+
+ /// Utility to determine whether loads can be speculated.
+ LoadHoisting LoadSpeculation;
};
/// LoopVectorizationCostModel - estimates the expected speedups due to
public:
LoopVectorizationCostModel(Loop *L, ScalarEvolution *SE, LoopInfo *LI,
LoopVectorizationLegality *Legal,
- const TargetTransformInfo *TTI)
- : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI) {}
-
- /// \return The most profitable vectorization factor.
+ const TargetTransformInfo &TTI,
+ DataLayout *DL, const TargetLibraryInfo *TLI)
+ : TheLoop(L), SE(SE), LI(LI), Legal(Legal), TTI(TTI), DL(DL), TLI(TLI) {}
+
+ /// Information about vectorization costs
+ struct VectorizationFactor {
+ unsigned Width; // Vector width with best cost
+ unsigned Cost; // Cost of the loop with that width
+ };
+ /// \return The most profitable vectorization factor and the cost of that VF.
/// This method checks every power of two up to VF. If UserVF is not ZERO
/// then this vectorization factor will be selected if vectorization is
/// possible.
- unsigned selectVectorizationFactor(bool OptForSize, unsigned UserVF);
+ VectorizationFactor selectVectorizationFactor(bool OptForSize,
+ unsigned UserVF);
+ /// \return The size (in bits) of the widest type in the code that
+ /// needs to be vectorized. We ignore values that remain scalar such as
+ /// 64 bit loop indices.
+ unsigned getWidestType();
/// \return The most profitable unroll factor.
/// If UserUF is non-zero then this method finds the best unroll-factor
/// based on register pressure and other parameters.
- unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF);
+ /// VF and LoopCost are the selected vectorization factor and the cost of the
+ /// selected VF.
+ unsigned selectUnrollFactor(bool OptForSize, unsigned UserUF, unsigned VF,
+ unsigned LoopCost);
/// \brief A struct that represents some properties of the register usage
/// of a loop.
/// the scalar type.
static Type* ToVectorTy(Type *Scalar, unsigned VF);
+ /// Returns whether the instruction is a load or store and will be a emitted
+ /// as a vector operation.
+ bool isConsecutiveLoadOrStore(Instruction *I);
+
/// The loop that we evaluate.
Loop *TheLoop;
/// Scev analysis.
/// Vectorization legality.
LoopVectorizationLegality *Legal;
/// Vector target information.
- const TargetTransformInfo *TTI;
+ const TargetTransformInfo &TTI;
+ /// Target data layout information.
+ DataLayout *DL;
+ /// Target Library Info.
+ const TargetLibraryInfo *TLI;
+};
+
+/// Utility class for getting and setting loop vectorizer hints in the form
+/// of loop metadata.
+struct LoopVectorizeHints {
+ /// Vectorization width.
+ unsigned Width;
+ /// Vectorization unroll factor.
+ unsigned Unroll;
+
+ LoopVectorizeHints(const Loop *L)
+ : Width(VectorizationFactor)
+ , Unroll(VectorizationUnroll)
+ , LoopID(L->getLoopID()) {
+ getHints(L);
+ // The command line options override any loop metadata except for when
+ // width == 1 which is used to indicate the loop is already vectorized.
+ if (VectorizationFactor.getNumOccurrences() > 0 && Width != 1)
+ Width = VectorizationFactor;
+ if (VectorizationUnroll.getNumOccurrences() > 0)
+ Unroll = VectorizationUnroll;
+ }
+
+ /// Return the loop vectorizer metadata prefix.
+ static StringRef Prefix() { return "llvm.vectorizer."; }
+
+ MDNode *createHint(LLVMContext &Context, StringRef Name, unsigned V) {
+ SmallVector<Value*, 2> Vals;
+ Vals.push_back(MDString::get(Context, Name));
+ Vals.push_back(ConstantInt::get(Type::getInt32Ty(Context), V));
+ return MDNode::get(Context, Vals);
+ }
+
+ /// Mark the loop L as already vectorized by setting the width to 1.
+ void setAlreadyVectorized(Loop *L) {
+ LLVMContext &Context = L->getHeader()->getContext();
+
+ Width = 1;
+
+ // Create a new loop id with one more operand for the already_vectorized
+ // hint. If the loop already has a loop id then copy the existing operands.
+ SmallVector<Value*, 4> Vals(1);
+ if (LoopID)
+ for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i)
+ Vals.push_back(LoopID->getOperand(i));
+
+ Twine Name = Prefix() + "width";
+ Vals.push_back(createHint(Context, Name.str(), Width));
+
+ MDNode *NewLoopID = MDNode::get(Context, Vals);
+ // Set operand 0 to refer to the loop id itself.
+ NewLoopID->replaceOperandWith(0, NewLoopID);
+
+ L->setLoopID(NewLoopID);
+ if (LoopID)
+ LoopID->replaceAllUsesWith(NewLoopID);
+
+ LoopID = NewLoopID;
+ }
+
+private:
+ MDNode *LoopID;
+
+ /// Find hints specified in the loop metadata.
+ void getHints(const Loop *L) {
+ if (!LoopID)
+ return;
+
+ // First operand should refer to the loop id itself.
+ assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
+ assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
+
+ for (unsigned i = 1, ie = LoopID->getNumOperands(); i < ie; ++i) {
+ const MDString *S = 0;
+ SmallVector<Value*, 4> Args;
+
+ // The expected hint is either a MDString or a MDNode with the first
+ // operand a MDString.
+ if (const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i))) {
+ if (!MD || MD->getNumOperands() == 0)
+ continue;
+ S = dyn_cast<MDString>(MD->getOperand(0));
+ for (unsigned i = 1, ie = MD->getNumOperands(); i < ie; ++i)
+ Args.push_back(MD->getOperand(i));
+ } else {
+ S = dyn_cast<MDString>(LoopID->getOperand(i));
+ assert(Args.size() == 0 && "too many arguments for MDString");
+ }
+
+ if (!S)
+ continue;
+
+ // Check if the hint starts with the vectorizer prefix.
+ StringRef Hint = S->getString();
+ if (!Hint.startswith(Prefix()))
+ continue;
+ // Remove the prefix.
+ Hint = Hint.substr(Prefix().size(), StringRef::npos);
+
+ if (Args.size() == 1)
+ getHint(Hint, Args[0]);
+ }
+ }
+
+ // Check string hint with one operand.
+ void getHint(StringRef Hint, Value *Arg) {
+ const ConstantInt *C = dyn_cast<ConstantInt>(Arg);
+ if (!C) return;
+ unsigned Val = C->getZExtValue();
+
+ if (Hint == "width") {
+ assert(isPowerOf2_32(Val) && Val <= MaxVectorWidth &&
+ "Invalid width metadata");
+ Width = Val;
+ } else if (Hint == "unroll") {
+ assert(isPowerOf2_32(Val) && Val <= MaxUnrollFactor &&
+ "Invalid unroll metadata");
+ Unroll = Val;
+ } else
+ DEBUG(dbgs() << "LV: ignoring unknown hint " << Hint);
+ }
};
/// The LoopVectorize Pass.
LoopInfo *LI;
TargetTransformInfo *TTI;
DominatorTree *DT;
+ AliasAnalysis *AA;
+ TargetLibraryInfo *TLI;
virtual bool runOnLoop(Loop *L, LPPassManager &LPM) {
// We only vectorize innermost loops.
SE = &getAnalysis<ScalarEvolution>();
DL = getAnalysisIfAvailable<DataLayout>();
LI = &getAnalysis<LoopInfo>();
- TTI = getAnalysisIfAvailable<TargetTransformInfo>();
+ TTI = &getAnalysis<TargetTransformInfo>();
DT = &getAnalysis<DominatorTree>();
+ AA = getAnalysisIfAvailable<AliasAnalysis>();
+ TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
+
+ if (DL == NULL) {
+ DEBUG(dbgs() << "LV: Not vectorizing because of missing data layout");
+ return false;
+ }
DEBUG(dbgs() << "LV: Checking a loop in \"" <<
L->getHeader()->getParent()->getName() << "\"\n");
+ LoopVectorizeHints Hints(L);
+
+ if (Hints.Width == 1) {
+ DEBUG(dbgs() << "LV: Not vectorizing.\n");
+ return false;
+ }
+
// Check if it is legal to vectorize the loop.
- LoopVectorizationLegality LVL(L, SE, DL, DT);
+ LoopVectorizationLegality LVL(L, SE, DL, DT, TTI, AA, TLI);
if (!LVL.canVectorize()) {
DEBUG(dbgs() << "LV: Not vectorizing.\n");
return false;
}
// Use the cost model.
- LoopVectorizationCostModel CM(L, SE, LI, &LVL, TTI);
+ LoopVectorizationCostModel CM(L, SE, LI, &LVL, *TTI, DL, TLI);
- // Check the function attribues to find out if this function should be
+ // Check the function attributes to find out if this function should be
// optimized for size.
Function *F = L->getHeader()->getParent();
Attribute::AttrKind SzAttr = Attribute::OptimizeForSize;
return false;
}
- unsigned VF = CM.selectVectorizationFactor(OptForSize, VectorizationFactor);
- unsigned UF = CM.selectUnrollFactor(OptForSize, VectorizationUnroll);
+ // Select the optimal vectorization factor.
+ LoopVectorizationCostModel::VectorizationFactor VF;
+ VF = CM.selectVectorizationFactor(OptForSize, Hints.Width);
+ // Select the unroll factor.
+ unsigned UF = CM.selectUnrollFactor(OptForSize, Hints.Unroll, VF.Width,
+ VF.Cost);
- if (VF == 1) {
+ if (VF.Width == 1) {
DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");
return false;
}
- DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF << ") in "<<
+ DEBUG(dbgs() << "LV: Found a vectorizable loop ("<< VF.Width << ") in "<<
F->getParent()->getModuleIdentifier()<<"\n");
DEBUG(dbgs() << "LV: Unroll Factor is " << UF << "\n");
- // If we decided that it is *legal* to vectorizer the loop then do it.
- InnerLoopVectorizer LB(L, SE, LI, DT, DL, VF, UF);
+ // If we decided that it is *legal* to vectorize the loop then do it.
+ InnerLoopVectorizer LB(L, SE, LI, DT, DL, TLI, VF.Width, UF);
LB.vectorize(&LVL);
+ // Mark the loop as already vectorized to avoid vectorizing again.
+ Hints.setAlreadyVectorized(L);
+
DEBUG(verifyFunction(*L->getHeader()->getParent()));
return true;
}
LoopPass::getAnalysisUsage(AU);
AU.addRequiredID(LoopSimplifyID);
AU.addRequiredID(LCSSAID);
+ AU.addRequired<DominatorTree>();
AU.addRequired<LoopInfo>();
AU.addRequired<ScalarEvolution>();
- AU.addRequired<DominatorTree>();
+ AU.addRequired<TargetTransformInfo>();
AU.addPreserved<LoopInfo>();
AU.addPreserved<DominatorTree>();
}
void
LoopVectorizationLegality::RuntimePointerCheck::insert(ScalarEvolution *SE,
- Loop *Lp, Value *Ptr) {
+ Loop *Lp, Value *Ptr,
+ bool WritePtr) {
const SCEV *Sc = SE->getSCEV(Ptr);
const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
assert(AR && "Invalid addrec expression");
Pointers.push_back(Ptr);
Starts.push_back(AR->getStart());
Ends.push_back(ScEnd);
+ IsWritePtr.push_back(WritePtr);
}
Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
return Shuf;
}
-Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, unsigned StartIdx,
+Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, int StartIdx,
bool Negate) {
assert(Val->getType()->isVectorTy() && "Must be a vector");
assert(Val->getType()->getScalarType()->isIntegerTy() &&
// Create a vector of consecutive numbers from zero to VF.
for (int i = 0; i < VLen; ++i) {
- int Idx = Negate ? (-i): i;
- Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx));
+ int64_t Idx = Negate ? (-i) : i;
+ Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx, Negate));
}
// Add the consecutive indices to the vector value.
int LoopVectorizationLegality::isConsecutivePtr(Value *Ptr) {
assert(Ptr->getType()->isPointerTy() && "Unexpected non ptr");
+ // Make sure that the pointer does not point to structs.
+ if (cast<PointerType>(Ptr->getType())->getElementType()->isAggregateType())
+ return 0;
// If this value is a pointer induction variable we know it is consecutive.
PHINode *Phi = dyn_cast_or_null<PHINode>(Ptr);
if (Phi && Inductions.count(Phi)) {
InductionInfo II = Inductions[Phi];
- if (PtrInduction == II.IK)
+ if (IK_PtrInduction == II.IK)
return 1;
+ else if (IK_ReversePtrInduction == II.IK)
+ return -1;
}
GetElementPtrInst *Gep = dyn_cast_or_null<GetElementPtrInst>(Ptr);
unsigned NumOperands = Gep->getNumOperands();
Value *LastIndex = Gep->getOperand(NumOperands - 1);
+ Value *GpPtr = Gep->getPointerOperand();
+ // If this GEP value is a consecutive pointer induction variable and all of
+ // the indices are constant then we know it is consecutive. We can
+ Phi = dyn_cast<PHINode>(GpPtr);
+ if (Phi && Inductions.count(Phi)) {
+
+ // Make sure that the pointer does not point to structs.
+ PointerType *GepPtrType = cast<PointerType>(GpPtr->getType());
+ if (GepPtrType->getElementType()->isAggregateType())
+ return 0;
+
+ // Make sure that all of the index operands are loop invariant.
+ for (unsigned i = 1; i < NumOperands; ++i)
+ if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
+ return 0;
+
+ InductionInfo II = Inductions[Phi];
+ if (IK_PtrInduction == II.IK)
+ return 1;
+ else if (IK_ReversePtrInduction == II.IK)
+ return -1;
+ }
+
// Check that all of the gep indices are uniform except for the last.
for (unsigned i = 0; i < NumOperands - 1; ++i)
if (!SE->isLoopInvariant(SE->getSCEV(Gep->getOperand(i)), TheLoop))
// If this scalar is unknown, assume that it is a constant or that it is
// loop invariant. Broadcast V and save the value for future uses.
Value *B = getBroadcastInstrs(V);
- WidenMap.splat(V, B);
- return WidenMap.get(V);
-}
-
-Constant*
-InnerLoopVectorizer::getUniformVector(unsigned Val, Type* ScalarTy) {
- return ConstantVector::getSplat(VF, ConstantInt::get(ScalarTy, Val, true));
+ return WidenMap.splat(V, B);
}
Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
"reverse");
}
+
+void InnerLoopVectorizer::vectorizeMemoryInstruction(Instruction *Instr,
+ LoopVectorizationLegality *Legal) {
+ // Attempt to issue a wide load.
+ LoadInst *LI = dyn_cast<LoadInst>(Instr);
+ StoreInst *SI = dyn_cast<StoreInst>(Instr);
+
+ assert((LI || SI) && "Invalid Load/Store instruction");
+
+ Type *ScalarDataTy = LI ? LI->getType() : SI->getValueOperand()->getType();
+ Type *DataTy = VectorType::get(ScalarDataTy, VF);
+ Value *Ptr = LI ? LI->getPointerOperand() : SI->getPointerOperand();
+ unsigned Alignment = LI ? LI->getAlignment() : SI->getAlignment();
+
+ unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ScalarDataTy);
+ unsigned VectorElementSize = DL->getTypeStoreSize(DataTy)/VF;
+
+ if (ScalarAllocatedSize != VectorElementSize)
+ return scalarizeInstruction(Instr);
+
+ // If the pointer is loop invariant or if it is non consecutive,
+ // scalarize the load.
+ int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
+ bool Reverse = ConsecutiveStride < 0;
+ bool UniformLoad = LI && Legal->isUniform(Ptr);
+ if (!ConsecutiveStride || UniformLoad)
+ return scalarizeInstruction(Instr);
+
+ Constant *Zero = Builder.getInt32(0);
+ VectorParts &Entry = WidenMap.get(Instr);
+
+ // Handle consecutive loads/stores.
+ GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
+ if (Gep && Legal->isInductionVariable(Gep->getPointerOperand())) {
+ Value *PtrOperand = Gep->getPointerOperand();
+ Value *FirstBasePtr = getVectorValue(PtrOperand)[0];
+ FirstBasePtr = Builder.CreateExtractElement(FirstBasePtr, Zero);
+
+ // Create the new GEP with the new induction variable.
+ GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
+ Gep2->setOperand(0, FirstBasePtr);
+ Gep2->setName("gep.indvar.base");
+ Ptr = Builder.Insert(Gep2);
+ } else if (Gep) {
+ assert(SE->isLoopInvariant(SE->getSCEV(Gep->getPointerOperand()),
+ OrigLoop) && "Base ptr must be invariant");
+
+ // The last index does not have to be the induction. It can be
+ // consecutive and be a function of the index. For example A[I+1];
+ unsigned NumOperands = Gep->getNumOperands();
+
+ Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
+ VectorParts &GEPParts = getVectorValue(LastGepOperand);
+ Value *LastIndex = GEPParts[0];
+ LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
+
+ // Create the new GEP with the new induction variable.
+ GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
+ Gep2->setOperand(NumOperands - 1, LastIndex);
+ Gep2->setName("gep.indvar.idx");
+ Ptr = Builder.Insert(Gep2);
+ } else {
+ // Use the induction element ptr.
+ assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
+ VectorParts &PtrVal = getVectorValue(Ptr);
+ Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
+ }
+
+ // Handle Stores:
+ if (SI) {
+ assert(!Legal->isUniform(SI->getPointerOperand()) &&
+ "We do not allow storing to uniform addresses");
+
+ VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
+ for (unsigned Part = 0; Part < UF; ++Part) {
+ // Calculate the pointer for the specific unroll-part.
+ Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
+
+ if (Reverse) {
+ // If we store to reverse consecutive memory locations then we need
+ // to reverse the order of elements in the stored value.
+ StoredVal[Part] = reverseVector(StoredVal[Part]);
+ // If the address is consecutive but reversed, then the
+ // wide store needs to start at the last vector element.
+ PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
+ PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
+ }
+
+ Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
+ Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
+ }
+ }
+
+ for (unsigned Part = 0; Part < UF; ++Part) {
+ // Calculate the pointer for the specific unroll-part.
+ Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
+
+ if (Reverse) {
+ // If the address is consecutive but reversed, then the
+ // wide store needs to start at the last vector element.
+ PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
+ PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
+ }
+
+ Value *VecPtr = Builder.CreateBitCast(PartPtr, DataTy->getPointerTo());
+ Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
+ cast<LoadInst>(LI)->setAlignment(Alignment);
+ Entry[Part] = Reverse ? reverseVector(LI) : LI;
+ }
+}
+
void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
// Holds vector parameters or scalars, in case of uniform vals.
// Create a new entry in the WidenMap and initialize it to Undef or Null.
VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
- // For each scalar that we create:
- for (unsigned Width = 0; Width < VF; ++Width) {
- // For each vector unroll 'part':
- for (unsigned Part = 0; Part < UF; ++Part) {
+ // For each vector unroll 'part':
+ for (unsigned Part = 0; Part < UF; ++Part) {
+ // For each scalar that we create:
+ for (unsigned Width = 0; Width < VF; ++Width) {
Instruction *Cloned = Instr->clone();
if (!IsVoidRetTy)
Cloned->setName(Instr->getName() + ".cloned");
}
}
-Value*
+Instruction *
InnerLoopVectorizer::addRuntimeCheck(LoopVectorizationLegality *Legal,
Instruction *Loc) {
LoopVectorizationLegality::RuntimePointerCheck *PtrRtCheck =
if (!PtrRtCheck->Need)
return NULL;
- Value *MemoryRuntimeCheck = 0;
+ Instruction *MemoryRuntimeCheck = 0;
unsigned NumPointers = PtrRtCheck->Pointers.size();
SmallVector<Value* , 2> Starts;
SmallVector<Value* , 2> Ends;
}
}
+ IRBuilder<> ChkBuilder(Loc);
+
for (unsigned i = 0; i < NumPointers; ++i) {
for (unsigned j = i+1; j < NumPointers; ++j) {
- Instruction::CastOps Op = Instruction::BitCast;
- Value *Start0 = CastInst::Create(Op, Starts[i], PtrArithTy, "bc", Loc);
- Value *Start1 = CastInst::Create(Op, Starts[j], PtrArithTy, "bc", Loc);
- Value *End0 = CastInst::Create(Op, Ends[i], PtrArithTy, "bc", Loc);
- Value *End1 = CastInst::Create(Op, Ends[j], PtrArithTy, "bc", Loc);
-
- Value *Cmp0 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
- Start0, End1, "bound0", Loc);
- Value *Cmp1 = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_ULE,
- Start1, End0, "bound1", Loc);
- Value *IsConflict = BinaryOperator::Create(Instruction::And, Cmp0, Cmp1,
- "found.conflict", Loc);
+ // No need to check if two readonly pointers intersect.
+ if (!PtrRtCheck->IsWritePtr[i] && !PtrRtCheck->IsWritePtr[j])
+ continue;
+
+ Value *Start0 = ChkBuilder.CreateBitCast(Starts[i], PtrArithTy, "bc");
+ Value *Start1 = ChkBuilder.CreateBitCast(Starts[j], PtrArithTy, "bc");
+ Value *End0 = ChkBuilder.CreateBitCast(Ends[i], PtrArithTy, "bc");
+ Value *End1 = ChkBuilder.CreateBitCast(Ends[j], PtrArithTy, "bc");
+
+ Value *Cmp0 = ChkBuilder.CreateICmpULE(Start0, End1, "bound0");
+ Value *Cmp1 = ChkBuilder.CreateICmpULE(Start1, End0, "bound1");
+ Value *IsConflict = ChkBuilder.CreateAnd(Cmp0, Cmp1, "found.conflict");
if (MemoryRuntimeCheck)
- MemoryRuntimeCheck = BinaryOperator::Create(Instruction::Or,
- MemoryRuntimeCheck,
- IsConflict,
- "conflict.rdx", Loc);
- else
- MemoryRuntimeCheck = IsConflict;
+ IsConflict = ChkBuilder.CreateOr(MemoryRuntimeCheck, IsConflict,
+ "conflict.rdx");
+ MemoryRuntimeCheck = cast<Instruction>(IsConflict);
}
}
the vectorized instructions while the old loop will continue to run the
scalar remainder.
- [ ] <-- vector loop bypass.
+ [ ] <-- vector loop bypass (may consist of multiple blocks).
/ |
/ v
| [ ] <-- vector pre header.
// induction variables. In the code below we also support a case where we
// don't have a single induction variable.
OldInduction = Legal->getInduction();
- Type *IdxTy = OldInduction ? OldInduction->getType() :
- DL->getIntPtrType(SE->getContext());
+ Type *IdxTy = Legal->getWidestInductionType();
// Find the loop boundaries.
const SCEV *ExitCount = SE->getExitCount(OrigLoop, OrigLoop->getLoopLatch());
// The loop index does not have to start at Zero. Find the original start
// value from the induction PHI node. If we don't have an induction variable
// then we know that it starts at zero.
- Value *StartIdx = OldInduction ?
- OldInduction->getIncomingValueForBlock(BypassBlock):
- ConstantInt::get(IdxTy, 0);
+ Builder.SetInsertPoint(BypassBlock->getTerminator());
+ Value *StartIdx = ExtendedIdx = OldInduction ?
+ Builder.CreateZExt(OldInduction->getIncomingValueForBlock(BypassBlock),
+ IdxTy):
+ ConstantInt::get(IdxTy, 0);
assert(BypassBlock && "Invalid loop structure");
-
- // Generate the code that checks in runtime if arrays overlap.
- Value *MemoryRuntimeCheck = addRuntimeCheck(Legal,
- BypassBlock->getTerminator());
+ LoopBypassBlocks.push_back(BypassBlock);
// Split the single block loop into the two loop structure described above.
BasicBlock *VectorPH =
BasicBlock *ScalarPH =
MiddleBlock->splitBasicBlock(MiddleBlock->getTerminator(), "scalar.ph");
- // This is the location in which we add all of the logic for bypassing
- // the new vector loop.
- Instruction *Loc = BypassBlock->getTerminator();
-
// Use this IR builder to create the loop instructions (Phi, Br, Cmp)
// inside the loop.
Builder.SetInsertPoint(VecBody->getFirstInsertionPt());
// times the unroll factor (num of SIMD instructions).
Constant *Step = ConstantInt::get(IdxTy, VF * UF);
+ // This is the IR builder that we use to add all of the logic for bypassing
+ // the new vector loop.
+ IRBuilder<> BypassBuilder(BypassBlock->getTerminator());
+
// We may need to extend the index in case there is a type mismatch.
// We know that the count starts at zero and does not overflow.
if (Count->getType() != IdxTy) {
// The exit count can be of pointer type. Convert it to the correct
// integer type.
if (ExitCount->getType()->isPointerTy())
- Count = CastInst::CreatePointerCast(Count, IdxTy, "ptrcnt.to.int", Loc);
+ Count = BypassBuilder.CreatePointerCast(Count, IdxTy, "ptrcnt.to.int");
else
- Count = CastInst::CreateZExtOrBitCast(Count, IdxTy, "zext.cnt", Loc);
+ Count = BypassBuilder.CreateZExtOrTrunc(Count, IdxTy, "cnt.cast");
}
// Add the start index to the loop count to get the new end index.
- Value *IdxEnd = BinaryOperator::CreateAdd(Count, StartIdx, "end.idx", Loc);
+ Value *IdxEnd = BypassBuilder.CreateAdd(Count, StartIdx, "end.idx");
// Now we need to generate the expression for N - (N % VF), which is
// the part that the vectorized body will execute.
- Value *R = BinaryOperator::CreateURem(Count, Step, "n.mod.vf", Loc);
- Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
- Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx,
- "end.idx.rnd.down", Loc);
+ Value *R = BypassBuilder.CreateURem(Count, Step, "n.mod.vf");
+ Value *CountRoundDown = BypassBuilder.CreateSub(Count, R, "n.vec");
+ Value *IdxEndRoundDown = BypassBuilder.CreateAdd(CountRoundDown, StartIdx,
+ "end.idx.rnd.down");
// Now, compare the new count to zero. If it is zero skip the vector loop and
// jump to the scalar loop.
- Value *Cmp = CmpInst::Create(Instruction::ICmp, CmpInst::ICMP_EQ,
- IdxEndRoundDown,
- StartIdx,
- "cmp.zero", Loc);
+ Value *Cmp = BypassBuilder.CreateICmpEQ(IdxEndRoundDown, StartIdx,
+ "cmp.zero");
+
+ BasicBlock *LastBypassBlock = BypassBlock;
+
+ // Generate the code that checks in runtime if arrays overlap. We put the
+ // checks into a separate block to make the more common case of few elements
+ // faster.
+ Instruction *MemRuntimeCheck = addRuntimeCheck(Legal,
+ BypassBlock->getTerminator());
+ if (MemRuntimeCheck) {
+ // Create a new block containing the memory check.
+ BasicBlock *CheckBlock = BypassBlock->splitBasicBlock(MemRuntimeCheck,
+ "vector.memcheck");
+ LoopBypassBlocks.push_back(CheckBlock);
+
+ // Replace the branch into the memory check block with a conditional branch
+ // for the "few elements case".
+ Instruction *OldTerm = BypassBlock->getTerminator();
+ BranchInst::Create(MiddleBlock, CheckBlock, Cmp, OldTerm);
+ OldTerm->eraseFromParent();
+
+ Cmp = MemRuntimeCheck;
+ LastBypassBlock = CheckBlock;
+ }
- // If we are using memory runtime checks, include them in.
- if (MemoryRuntimeCheck)
- Cmp = BinaryOperator::Create(Instruction::Or, Cmp, MemoryRuntimeCheck,
- "CntOrMem", Loc);
-
- BranchInst::Create(MiddleBlock, VectorPH, Cmp, Loc);
- // Remove the old terminator.
- Loc->eraseFromParent();
+ LastBypassBlock->getTerminator()->eraseFromParent();
+ BranchInst::Create(MiddleBlock, VectorPH, Cmp,
+ LastBypassBlock);
// We are going to resume the execution of the scalar loop.
// Go over all of the induction variables that we found and fix the
PHINode *ResumeIndex = 0;
LoopVectorizationLegality::InductionList::iterator I, E;
LoopVectorizationLegality::InductionList *List = Legal->getInductionVars();
+ // Set builder to point to last bypass block.
+ BypassBuilder.SetInsertPoint(LoopBypassBlocks.back()->getTerminator());
for (I = List->begin(), E = List->end(); I != E; ++I) {
PHINode *OrigPhi = I->first;
LoopVectorizationLegality::InductionInfo II = I->second;
- PHINode *ResumeVal = PHINode::Create(OrigPhi->getType(), 2, "resume.val",
+
+ Type *ResumeValTy = (OrigPhi == OldInduction) ? IdxTy : OrigPhi->getType();
+ PHINode *ResumeVal = PHINode::Create(ResumeValTy, 2, "resume.val",
MiddleBlock->getTerminator());
+ // We might have extended the type of the induction variable but we need a
+ // truncated version for the scalar loop.
+ PHINode *TruncResumeVal = (OrigPhi == OldInduction) ?
+ PHINode::Create(OrigPhi->getType(), 2, "trunc.resume.val",
+ MiddleBlock->getTerminator()) : 0;
+
Value *EndValue = 0;
switch (II.IK) {
- case LoopVectorizationLegality::NoInduction:
+ case LoopVectorizationLegality::IK_NoInduction:
llvm_unreachable("Unknown induction");
- case LoopVectorizationLegality::IntInduction: {
- // Handle the integer induction counter:
+ case LoopVectorizationLegality::IK_IntInduction: {
+ // Handle the integer induction counter.
assert(OrigPhi->getType()->isIntegerTy() && "Invalid type");
- assert(OrigPhi == OldInduction && "Unknown integer PHI");
- // We know what the end value is.
- EndValue = IdxEndRoundDown;
- // We also know which PHI node holds it.
- ResumeIndex = ResumeVal;
+
+ // We have the canonical induction variable.
+ if (OrigPhi == OldInduction) {
+ // Create a truncated version of the resume value for the scalar loop,
+ // we might have promoted the type to a larger width.
+ EndValue =
+ BypassBuilder.CreateTrunc(IdxEndRoundDown, OrigPhi->getType());
+ // The new PHI merges the original incoming value, in case of a bypass,
+ // or the value at the end of the vectorized loop.
+ for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
+ TruncResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
+ TruncResumeVal->addIncoming(EndValue, VecBody);
+
+ // We know what the end value is.
+ EndValue = IdxEndRoundDown;
+ // We also know which PHI node holds it.
+ ResumeIndex = ResumeVal;
+ break;
+ }
+
+ // Not the canonical induction variable - add the vector loop count to the
+ // start value.
+ Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
+ II.StartValue->getType(),
+ "cast.crd");
+ EndValue = BypassBuilder.CreateAdd(CRD, II.StartValue , "ind.end");
break;
}
- case LoopVectorizationLegality::ReverseIntInduction: {
+ case LoopVectorizationLegality::IK_ReverseIntInduction: {
// Convert the CountRoundDown variable to the PHI size.
- unsigned CRDSize = CountRoundDown->getType()->getScalarSizeInBits();
- unsigned IISize = II.StartValue->getType()->getScalarSizeInBits();
- Value *CRD = CountRoundDown;
- if (CRDSize > IISize)
- CRD = CastInst::Create(Instruction::Trunc, CountRoundDown,
- II.StartValue->getType(),
- "tr.crd", BypassBlock->getTerminator());
- else if (CRDSize < IISize)
- CRD = CastInst::Create(Instruction::SExt, CountRoundDown,
- II.StartValue->getType(),
- "sext.crd", BypassBlock->getTerminator());
- // Handle reverse integer induction counter:
- EndValue = BinaryOperator::CreateSub(II.StartValue, CRD, "rev.ind.end",
- BypassBlock->getTerminator());
+ Value *CRD = BypassBuilder.CreateSExtOrTrunc(CountRoundDown,
+ II.StartValue->getType(),
+ "cast.crd");
+ // Handle reverse integer induction counter.
+ EndValue = BypassBuilder.CreateSub(II.StartValue, CRD, "rev.ind.end");
break;
}
- case LoopVectorizationLegality::PtrInduction: {
+ case LoopVectorizationLegality::IK_PtrInduction: {
// For pointer induction variables, calculate the offset using
// the end index.
- EndValue = GetElementPtrInst::Create(II.StartValue, CountRoundDown,
- "ptr.ind.end",
- BypassBlock->getTerminator());
+ EndValue = BypassBuilder.CreateGEP(II.StartValue, CountRoundDown,
+ "ptr.ind.end");
+ break;
+ }
+ case LoopVectorizationLegality::IK_ReversePtrInduction: {
+ // The value at the end of the loop for the reverse pointer is calculated
+ // by creating a GEP with a negative index starting from the start value.
+ Value *Zero = ConstantInt::get(CountRoundDown->getType(), 0);
+ Value *NegIdx = BypassBuilder.CreateSub(Zero, CountRoundDown,
+ "rev.ind.end");
+ EndValue = BypassBuilder.CreateGEP(II.StartValue, NegIdx,
+ "rev.ptr.ind.end");
break;
}
}// end of case
// The new PHI merges the original incoming value, in case of a bypass,
// or the value at the end of the vectorized loop.
- ResumeVal->addIncoming(II.StartValue, BypassBlock);
+ for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I) {
+ if (OrigPhi == OldInduction)
+ ResumeVal->addIncoming(StartIdx, LoopBypassBlocks[I]);
+ else
+ ResumeVal->addIncoming(II.StartValue, LoopBypassBlocks[I]);
+ }
ResumeVal->addIncoming(EndValue, VecBody);
// Fix the scalar body counter (PHI node).
unsigned BlockIdx = OrigPhi->getBasicBlockIndex(ScalarPH);
- OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
+ // The old inductions phi node in the scalar body needs the truncated value.
+ if (OrigPhi == OldInduction)
+ OrigPhi->setIncomingValue(BlockIdx, TruncResumeVal);
+ else
+ OrigPhi->setIncomingValue(BlockIdx, ResumeVal);
}
// If we are generating a new induction variable then we also need to
assert(!ResumeIndex && "Unexpected resume value found");
ResumeIndex = PHINode::Create(IdxTy, 2, "new.indc.resume.val",
MiddleBlock->getTerminator());
- ResumeIndex->addIncoming(StartIdx, BypassBlock);
+ for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
+ ResumeIndex->addIncoming(StartIdx, LoopBypassBlocks[I]);
ResumeIndex->addIncoming(IdxEndRoundDown, VecBody);
}
// Insert the new loop into the loop nest and register the new basic blocks.
if (ParentLoop) {
ParentLoop->addChildLoop(Lp);
+ for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
+ ParentLoop->addBasicBlockToLoop(LoopBypassBlocks[I], LI->getBase());
ParentLoop->addBasicBlockToLoop(ScalarPH, LI->getBase());
ParentLoop->addBasicBlockToLoop(VectorPH, LI->getBase());
ParentLoop->addBasicBlockToLoop(MiddleBlock, LI->getBase());
LoopExitBlock = ExitBlock;
LoopVectorBody = VecBody;
LoopScalarBody = OldBasicBlock;
- LoopBypassBlock = BypassBlock;
}
/// This function returns the identity element (or neutral element) for
/// the operation K.
-static unsigned
-getReductionIdentity(LoopVectorizationLegality::ReductionKind K) {
+Constant*
+LoopVectorizationLegality::getReductionIdentity(ReductionKind K, Type *Tp) {
switch (K) {
- case LoopVectorizationLegality::IntegerXor:
- case LoopVectorizationLegality::IntegerAdd:
- case LoopVectorizationLegality::IntegerOr:
+ case RK_IntegerXor:
+ case RK_IntegerAdd:
+ case RK_IntegerOr:
// Adding, Xoring, Oring zero to a number does not change it.
- return 0;
- case LoopVectorizationLegality::IntegerMult:
+ return ConstantInt::get(Tp, 0);
+ case RK_IntegerMult:
// Multiplying a number by 1 does not change it.
- return 1;
- case LoopVectorizationLegality::IntegerAnd:
+ return ConstantInt::get(Tp, 1);
+ case RK_IntegerAnd:
// AND-ing a number with an all-1 value does not change it.
- return -1;
+ return ConstantInt::get(Tp, -1, true);
+ case RK_FloatMult:
+ // Multiplying a number by 1 does not change it.
+ return ConstantFP::get(Tp, 1.0L);
+ case RK_FloatAdd:
+ // Adding zero to a number does not change it.
+ return ConstantFP::get(Tp, 0.0L);
default:
llvm_unreachable("Unknown reduction kind");
}
}
-static bool
-isTriviallyVectorizableIntrinsic(Instruction *Inst) {
- IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst);
- if (!II)
- return false;
- switch (II->getIntrinsicID()) {
- case Intrinsic::sqrt:
- case Intrinsic::sin:
- case Intrinsic::cos:
- case Intrinsic::exp:
- case Intrinsic::exp2:
- case Intrinsic::log:
- case Intrinsic::log10:
- case Intrinsic::log2:
- case Intrinsic::fabs:
- case Intrinsic::floor:
- case Intrinsic::ceil:
- case Intrinsic::trunc:
- case Intrinsic::rint:
- case Intrinsic::nearbyint:
- case Intrinsic::pow:
- case Intrinsic::fma:
- case Intrinsic::fmuladd:
- return true;
+static Intrinsic::ID
+getIntrinsicIDForCall(CallInst *CI, const TargetLibraryInfo *TLI) {
+ // If we have an intrinsic call, check if it is trivially vectorizable.
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI)) {
+ switch (II->getIntrinsicID()) {
+ case Intrinsic::sqrt:
+ case Intrinsic::sin:
+ case Intrinsic::cos:
+ case Intrinsic::exp:
+ case Intrinsic::exp2:
+ case Intrinsic::log:
+ case Intrinsic::log10:
+ case Intrinsic::log2:
+ case Intrinsic::fabs:
+ case Intrinsic::floor:
+ case Intrinsic::ceil:
+ case Intrinsic::trunc:
+ case Intrinsic::rint:
+ case Intrinsic::nearbyint:
+ case Intrinsic::pow:
+ case Intrinsic::fma:
+ case Intrinsic::fmuladd:
+ return II->getIntrinsicID();
+ default:
+ return Intrinsic::not_intrinsic;
+ }
+ }
+
+ if (!TLI)
+ return Intrinsic::not_intrinsic;
+
+ LibFunc::Func Func;
+ Function *F = CI->getCalledFunction();
+ // We're going to make assumptions on the semantics of the functions, check
+ // that the target knows that it's available in this environment.
+ if (!F || !TLI->getLibFunc(F->getName(), Func))
+ return Intrinsic::not_intrinsic;
+
+ // Otherwise check if we have a call to a function that can be turned into a
+ // vector intrinsic.
+ switch (Func) {
default:
- return false;
+ break;
+ case LibFunc::sin:
+ case LibFunc::sinf:
+ case LibFunc::sinl:
+ return Intrinsic::sin;
+ case LibFunc::cos:
+ case LibFunc::cosf:
+ case LibFunc::cosl:
+ return Intrinsic::cos;
+ case LibFunc::exp:
+ case LibFunc::expf:
+ case LibFunc::expl:
+ return Intrinsic::exp;
+ case LibFunc::exp2:
+ case LibFunc::exp2f:
+ case LibFunc::exp2l:
+ return Intrinsic::exp2;
+ case LibFunc::log:
+ case LibFunc::logf:
+ case LibFunc::logl:
+ return Intrinsic::log;
+ case LibFunc::log10:
+ case LibFunc::log10f:
+ case LibFunc::log10l:
+ return Intrinsic::log10;
+ case LibFunc::log2:
+ case LibFunc::log2f:
+ case LibFunc::log2l:
+ return Intrinsic::log2;
+ case LibFunc::fabs:
+ case LibFunc::fabsf:
+ case LibFunc::fabsl:
+ return Intrinsic::fabs;
+ case LibFunc::floor:
+ case LibFunc::floorf:
+ case LibFunc::floorl:
+ return Intrinsic::floor;
+ case LibFunc::ceil:
+ case LibFunc::ceilf:
+ case LibFunc::ceill:
+ return Intrinsic::ceil;
+ case LibFunc::trunc:
+ case LibFunc::truncf:
+ case LibFunc::truncl:
+ return Intrinsic::trunc;
+ case LibFunc::rint:
+ case LibFunc::rintf:
+ case LibFunc::rintl:
+ return Intrinsic::rint;
+ case LibFunc::nearbyint:
+ case LibFunc::nearbyintf:
+ case LibFunc::nearbyintl:
+ return Intrinsic::nearbyint;
+ case LibFunc::pow:
+ case LibFunc::powf:
+ case LibFunc::powl:
+ return Intrinsic::pow;
}
- return false;
+
+ return Intrinsic::not_intrinsic;
+}
+
+/// This function translates the reduction kind to an LLVM binary operator.
+static unsigned
+getReductionBinOp(LoopVectorizationLegality::ReductionKind Kind) {
+ switch (Kind) {
+ case LoopVectorizationLegality::RK_IntegerAdd:
+ return Instruction::Add;
+ case LoopVectorizationLegality::RK_IntegerMult:
+ return Instruction::Mul;
+ case LoopVectorizationLegality::RK_IntegerOr:
+ return Instruction::Or;
+ case LoopVectorizationLegality::RK_IntegerAnd:
+ return Instruction::And;
+ case LoopVectorizationLegality::RK_IntegerXor:
+ return Instruction::Xor;
+ case LoopVectorizationLegality::RK_FloatMult:
+ return Instruction::FMul;
+ case LoopVectorizationLegality::RK_FloatAdd:
+ return Instruction::FAdd;
+ case LoopVectorizationLegality::RK_IntegerMinMax:
+ return Instruction::ICmp;
+ case LoopVectorizationLegality::RK_FloatMinMax:
+ return Instruction::FCmp;
+ default:
+ llvm_unreachable("Unknown reduction operation");
+ }
+}
+
+Value *createMinMaxOp(IRBuilder<> &Builder,
+ LoopVectorizationLegality::MinMaxReductionKind RK,
+ Value *Left,
+ Value *Right) {
+ CmpInst::Predicate P = CmpInst::ICMP_NE;
+ switch (RK) {
+ default:
+ llvm_unreachable("Unknown min/max reduction kind");
+ case LoopVectorizationLegality::MRK_UIntMin:
+ P = CmpInst::ICMP_ULT;
+ break;
+ case LoopVectorizationLegality::MRK_UIntMax:
+ P = CmpInst::ICMP_UGT;
+ break;
+ case LoopVectorizationLegality::MRK_SIntMin:
+ P = CmpInst::ICMP_SLT;
+ break;
+ case LoopVectorizationLegality::MRK_SIntMax:
+ P = CmpInst::ICMP_SGT;
+ break;
+ case LoopVectorizationLegality::MRK_FloatMin:
+ P = CmpInst::FCMP_OLT;
+ break;
+ case LoopVectorizationLegality::MRK_FloatMax:
+ P = CmpInst::FCMP_OGT;
+ break;
+ }
+
+ Value *Cmp;
+ if (RK == LoopVectorizationLegality::MRK_FloatMin || RK == LoopVectorizationLegality::MRK_FloatMax)
+ Cmp = Builder.CreateFCmp(P, Left, Right, "rdx.minmax.cmp");
+ else
+ Cmp = Builder.CreateICmp(P, Left, Right, "rdx.minmax.cmp");
+
+ Value *Select = Builder.CreateSelect(Cmp, Left, Right, "rdx.minmax.select");
+ return Select;
}
void
// the cost-model.
//
//===------------------------------------------------===//
- BasicBlock &BB = *OrigLoop->getHeader();
- Constant *Zero =
- ConstantInt::get(IntegerType::getInt32Ty(BB.getContext()), 0);
+ Constant *Zero = Builder.getInt32(0);
// In order to support reduction variables we need to be able to vectorize
// Phi nodes. Phi nodes have cycles, so we need to vectorize them in two
// To do so, we need to generate the 'identity' vector and overide
// one of the elements with the incoming scalar reduction. We need
// to do it in the vector-loop preheader.
- Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
+ Builder.SetInsertPoint(LoopBypassBlocks.front()->getTerminator());
// This is the vector-clone of the value that leaves the loop.
VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
// Find the reduction identity variable. Zero for addition, or, xor,
// one for multiplication, -1 for And.
- Constant *Identity = getUniformVector(getReductionIdentity(RdxDesc.Kind),
- VecTy->getScalarType());
-
- // This vector is the Identity vector where the first element is the
- // incoming scalar reduction.
- Value *VectorStart = Builder.CreateInsertElement(Identity,
- RdxDesc.StartValue, Zero);
+ Value *Identity;
+ Value *VectorStart;
+ if (RdxDesc.Kind == LoopVectorizationLegality::RK_IntegerMinMax ||
+ RdxDesc.Kind == LoopVectorizationLegality::RK_FloatMinMax) {
+ // MinMax reduction have the start value as their identify.
+ VectorStart = Identity = Builder.CreateVectorSplat(VF, RdxDesc.StartValue,
+ "minmax.ident");
+ } else {
+ Constant *Iden =
+ LoopVectorizationLegality::getReductionIdentity(RdxDesc.Kind,
+ VecTy->getScalarType());
+ Identity = ConstantVector::getSplat(VF, Iden);
+
+ // This vector is the Identity vector where the first element is the
+ // incoming scalar reduction.
+ VectorStart = Builder.CreateInsertElement(Identity,
+ RdxDesc.StartValue, Zero);
+ }
// Fix the vector-loop phi.
// We created the induction variable so we know that the
Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
VectorParts &Val = getVectorValue(LoopVal);
for (unsigned part = 0; part < UF; ++part) {
- // Make sure to add the reduction stat value only to the
+ // Make sure to add the reduction stat value only to the
// first unroll part.
Value *StartVal = (part == 0) ? VectorStart : Identity;
cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
Value *StartVal = (part == 0) ? VectorStart : Identity;
- NewPhi->addIncoming(StartVal, LoopBypassBlock);
+ for (unsigned I = 0, E = LoopBypassBlocks.size(); I != E; ++I)
+ NewPhi->addIncoming(StartVal, LoopBypassBlocks[I]);
NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
RdxParts.push_back(NewPhi);
}
// Reduce all of the unrolled parts into a single vector.
Value *ReducedPartRdx = RdxParts[0];
+ unsigned Op = getReductionBinOp(RdxDesc.Kind);
for (unsigned part = 1; part < UF; ++part) {
- switch (RdxDesc.Kind) {
- case LoopVectorizationLegality::IntegerAdd:
- ReducedPartRdx =
- Builder.CreateAdd(RdxParts[part], ReducedPartRdx, "add.rdx");
- break;
- case LoopVectorizationLegality::IntegerMult:
- ReducedPartRdx =
- Builder.CreateMul(RdxParts[part], ReducedPartRdx, "mul.rdx");
- break;
- case LoopVectorizationLegality::IntegerOr:
- ReducedPartRdx =
- Builder.CreateOr(RdxParts[part], ReducedPartRdx, "or.rdx");
- break;
- case LoopVectorizationLegality::IntegerAnd:
- ReducedPartRdx =
- Builder.CreateAnd(RdxParts[part], ReducedPartRdx, "and.rdx");
- break;
- case LoopVectorizationLegality::IntegerXor:
- ReducedPartRdx =
- Builder.CreateXor(RdxParts[part], ReducedPartRdx, "xor.rdx");
- break;
- default:
- llvm_unreachable("Unknown reduction operation");
- }
+ if (Op != Instruction::ICmp && Op != Instruction::FCmp)
+ ReducedPartRdx = Builder.CreateBinOp((Instruction::BinaryOps)Op,
+ RdxParts[part], ReducedPartRdx,
+ "bin.rdx");
+ else
+ ReducedPartRdx = createMinMaxOp(Builder, RdxDesc.MinMaxKind,
+ ReducedPartRdx, RdxParts[part]);
}
-
// VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
// and vector ops, reducing the set of values being computed by half each
ConstantVector::get(ShuffleMask),
"rdx.shuf");
- // Emit the operation on the shuffled value.
- switch (RdxDesc.Kind) {
- case LoopVectorizationLegality::IntegerAdd:
- TmpVec = Builder.CreateAdd(TmpVec, Shuf, "add.rdx");
- break;
- case LoopVectorizationLegality::IntegerMult:
- TmpVec = Builder.CreateMul(TmpVec, Shuf, "mul.rdx");
- break;
- case LoopVectorizationLegality::IntegerOr:
- TmpVec = Builder.CreateOr(TmpVec, Shuf, "or.rdx");
- break;
- case LoopVectorizationLegality::IntegerAnd:
- TmpVec = Builder.CreateAnd(TmpVec, Shuf, "and.rdx");
- break;
- case LoopVectorizationLegality::IntegerXor:
- TmpVec = Builder.CreateXor(TmpVec, Shuf, "xor.rdx");
- break;
- default:
- llvm_unreachable("Unknown reduction operation");
- }
+ if (Op != Instruction::ICmp && Op != Instruction::FCmp)
+ TmpVec = Builder.CreateBinOp((Instruction::BinaryOps)Op, TmpVec, Shuf,
+ "bin.rdx");
+ else
+ TmpVec = createMinMaxOp(Builder, RdxDesc.MinMaxKind, TmpVec, Shuf);
}
// The result is in the first element of the vector.
void
InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
BasicBlock *BB, PhiVector *PV) {
- Constant *Zero = Builder.getInt32(0);
-
// For each instruction in the old loop.
for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
VectorParts &Entry = WidenMap.get(it);
// We know that all PHIs in non header blocks are converted into
// selects, so we don't have to worry about the insertion order and we
// can just use the builder.
-
// At this point we generate the predication tree. There may be
// duplications since this is a simple recursive scan, but future
// optimizations will clean it up.
- VectorParts Cond = createEdgeMask(P->getIncomingBlock(0),
- P->getParent());
-
- for (unsigned part = 0; part < UF; ++part) {
- VectorParts &In0 = getVectorValue(P->getIncomingValue(0));
- VectorParts &In1 = getVectorValue(P->getIncomingValue(1));
- Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In1[part],
- "predphi");
+
+ unsigned NumIncoming = P->getNumIncomingValues();
+
+ // Generate a sequence of selects of the form:
+ // SELECT(Mask3, In3,
+ // SELECT(Mask2, In2,
+ // ( ...)))
+ for (unsigned In = 0; In < NumIncoming; In++) {
+ VectorParts Cond = createEdgeMask(P->getIncomingBlock(In),
+ P->getParent());
+ VectorParts &In0 = getVectorValue(P->getIncomingValue(In));
+
+ for (unsigned part = 0; part < UF; ++part) {
+ // We might have single edge PHIs (blocks) - use an identity
+ // 'select' for the first PHI operand.
+ if (In == 0)
+ Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
+ In0[part]);
+ else
+ // Select between the current value and the previous incoming edge
+ // based on the incoming mask.
+ Entry[part] = Builder.CreateSelect(Cond[part], In0[part],
+ Entry[part], "predphi");
+ }
}
continue;
}
Legal->getInductionVars()->lookup(P);
switch (II.IK) {
- case LoopVectorizationLegality::NoInduction:
+ case LoopVectorizationLegality::IK_NoInduction:
llvm_unreachable("Unknown induction");
- case LoopVectorizationLegality::IntInduction: {
- assert(P == OldInduction && "Unexpected PHI");
- Value *Broadcasted = getBroadcastInstrs(Induction);
- // After broadcasting the induction variable we need to make the
- // vector consecutive by adding 0, 1, 2 ...
+ case LoopVectorizationLegality::IK_IntInduction: {
+ assert(P->getType() == II.StartValue->getType() && "Types must match");
+ Type *PhiTy = P->getType();
+ Value *Broadcasted;
+ if (P == OldInduction) {
+ // Handle the canonical induction variable. We might have had to
+ // extend the type.
+ Broadcasted = Builder.CreateTrunc(Induction, PhiTy);
+ } else {
+ // Handle other induction variables that are now based on the
+ // canonical one.
+ Value *NormalizedIdx = Builder.CreateSub(Induction, ExtendedIdx,
+ "normalized.idx");
+ NormalizedIdx = Builder.CreateSExtOrTrunc(NormalizedIdx, PhiTy);
+ Broadcasted = Builder.CreateAdd(II.StartValue, NormalizedIdx,
+ "offset.idx");
+ }
+ Broadcasted = getBroadcastInstrs(Broadcasted);
+ // After broadcasting the induction variable we need to make the vector
+ // consecutive by adding 0, 1, 2, etc.
for (unsigned part = 0; part < UF; ++part)
Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
continue;
}
- case LoopVectorizationLegality::ReverseIntInduction:
- case LoopVectorizationLegality::PtrInduction:
+ case LoopVectorizationLegality::IK_ReverseIntInduction:
+ case LoopVectorizationLegality::IK_PtrInduction:
+ case LoopVectorizationLegality::IK_ReversePtrInduction:
// Handle reverse integer and pointer inductions.
- Value *StartIdx = 0;
- // If we have a single integer induction variable then use it.
- // Otherwise, start counting at zero.
- if (OldInduction) {
- LoopVectorizationLegality::InductionInfo OldII =
- Legal->getInductionVars()->lookup(OldInduction);
- StartIdx = OldII.StartValue;
- } else {
- StartIdx = ConstantInt::get(Induction->getType(), 0);
- }
+ Value *StartIdx = ExtendedIdx;
// This is the normalized GEP that starts counting at zero.
Value *NormalizedIdx = Builder.CreateSub(Induction, StartIdx,
"normalized.idx");
// Handle the reverse integer induction variable case.
- if (LoopVectorizationLegality::ReverseIntInduction == II.IK) {
+ if (LoopVectorizationLegality::IK_ReverseIntInduction == II.IK) {
IntegerType *DstTy = cast<IntegerType>(II.StartValue->getType());
Value *CNI = Builder.CreateSExtOrTrunc(NormalizedIdx, DstTy,
"resize.norm.idx");
// After broadcasting the induction variable we need to make the
// vector consecutive by adding ... -3, -2, -1, 0.
for (unsigned part = 0; part < UF; ++part)
- Entry[part] = getConsecutiveVector(Broadcasted, -VF * part, true);
+ Entry[part] = getConsecutiveVector(Broadcasted, -(int)VF * part,
+ true);
continue;
}
// Handle the pointer induction variable case.
assert(P->getType()->isPointerTy() && "Unexpected type.");
+ // Is this a reverse induction ptr or a consecutive induction ptr.
+ bool Reverse = (LoopVectorizationLegality::IK_ReversePtrInduction ==
+ II.IK);
+
// This is the vector of results. Notice that we don't generate
// vector geps because scalar geps result in better code.
for (unsigned part = 0; part < UF; ++part) {
Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
for (unsigned int i = 0; i < VF; ++i) {
- Constant *Idx = ConstantInt::get(Induction->getType(),
- i + part * VF);
- Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx,
- "gep.idx");
+ int EltIndex = (i + part * VF) * (Reverse ? -1 : 1);
+ Constant *Idx = ConstantInt::get(Induction->getType(), EltIndex);
+ Value *GlobalIdx;
+ if (!Reverse)
+ GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, "gep.idx");
+ else
+ GlobalIdx = Builder.CreateSub(Idx, NormalizedIdx, "gep.ridx");
+
Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
"next.gep");
VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
for (unsigned Part = 0; Part < UF; ++Part) {
Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
- // Update the NSW, NUW and Exact flags.
- BinaryOperator *VecOp = cast<BinaryOperator>(V);
- if (isa<OverflowingBinaryOperator>(BinOp)) {
+ // Update the NSW, NUW and Exact flags. Notice: V can be an Undef.
+ BinaryOperator *VecOp = dyn_cast<BinaryOperator>(V);
+ if (VecOp && isa<OverflowingBinaryOperator>(BinOp)) {
VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
}
- if (isa<PossiblyExactOperator>(VecOp))
+ if (
VecOp && isa<PossiblyExactOperator>(VecOp))
VecOp->setIsExact(BinOp->isExact());
Entry[Part] = V;
break;
}
- case Instruction::Store: {
- // Attempt to issue a wide store.
- StoreInst *SI = dyn_cast<StoreInst>(it);
- Type *StTy = VectorType::get(SI->getValueOperand()->getType(), VF);
- Value *Ptr = SI->getPointerOperand();
- unsigned Alignment = SI->getAlignment();
-
- assert(!Legal->isUniform(Ptr) &&
- "We do not allow storing to uniform addresses");
-
-
- int Stride = Legal->isConsecutivePtr(Ptr);
- bool Reverse = Stride < 0;
- if (Stride == 0) {
- scalarizeInstruction(it);
+ case Instruction::Store:
+ case Instruction::Load:
+ vectorizeMemoryInstruction(it, Legal);
break;
- }
-
- // Handle consecutive stores.
-
- GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
- if (Gep) {
- // The last index does not have to be the induction. It can be
- // consecutive and be a function of the index. For example A[I+1];
- unsigned NumOperands = Gep->getNumOperands();
-
- Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
- VectorParts &GEPParts = getVectorValue(LastGepOperand);
- Value *LastIndex = GEPParts[0];
- LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
-
- // Create the new GEP with the new induction variable.
- GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
- Gep2->setOperand(NumOperands - 1, LastIndex);
- Ptr = Builder.Insert(Gep2);
- } else {
- // Use the induction element ptr.
- assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
- VectorParts &PtrVal = getVectorValue(Ptr);
- Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
- }
-
- VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
- for (unsigned Part = 0; Part < UF; ++Part) {
- // Calculate the pointer for the specific unroll-part.
- Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
-
- if (Reverse) {
- // If we store to reverse consecutive memory locations then we need
- // to reverse the order of elements in the stored value.
- StoredVal[Part] = reverseVector(StoredVal[Part]);
- // If the address is consecutive but reversed, then the
- // wide store needs to start at the last vector element.
- PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
- PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
- }
-
- Value *VecPtr = Builder.CreateBitCast(PartPtr, StTy->getPointerTo());
- Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
- }
- break;
- }
- case Instruction::Load: {
- // Attempt to issue a wide load.
- LoadInst *LI = dyn_cast<LoadInst>(it);
- Type *RetTy = VectorType::get(LI->getType(), VF);
- Value *Ptr = LI->getPointerOperand();
- unsigned Alignment = LI->getAlignment();
-
- // If the pointer is loop invariant or if it is non consecutive,
- // scalarize the load.
- int Stride = Legal->isConsecutivePtr(Ptr);
- bool Reverse = Stride < 0;
- if (Legal->isUniform(Ptr) || Stride == 0) {
- scalarizeInstruction(it);
- break;
- }
-
- GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
- if (Gep) {
- // The last index does not have to be the induction. It can be
- // consecutive and be a function of the index. For example A[I+1];
- unsigned NumOperands = Gep->getNumOperands();
-
- Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
- VectorParts &GEPParts = getVectorValue(LastGepOperand);
- Value *LastIndex = GEPParts[0];
- LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
-
- // Create the new GEP with the new induction variable.
- GetElementPtrInst *Gep2 = cast<GetElementPtrInst>(Gep->clone());
- Gep2->setOperand(NumOperands - 1, LastIndex);
- Ptr = Builder.Insert(Gep2);
- } else {
- // Use the induction element ptr.
- assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
- VectorParts &PtrVal = getVectorValue(Ptr);
- Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
- }
-
- for (unsigned Part = 0; Part < UF; ++Part) {
- // Calculate the pointer for the specific unroll-part.
- Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
-
- if (Reverse) {
- // If the address is consecutive but reversed, then the
- // wide store needs to start at the last vector element.
- PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
- PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
- }
-
- Value *VecPtr = Builder.CreateBitCast(PartPtr, RetTy->getPointerTo());
- Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
- cast<LoadInst>(LI)->setAlignment(Alignment);
- Entry[Part] = Reverse ? reverseVector(LI) : LI;
- }
- break;
- }
case Instruction::ZExt:
case Instruction::SExt:
case Instruction::FPToUI:
}
case Instruction::Call: {
- assert(isTriviallyVectorizableIntrinsic(it));
+ // Ignore dbg intrinsics.
+ if (isa<DbgInfoIntrinsic>(it))
+ break;
+
Module *M = BB->getParent()->getParent();
- IntrinsicInst *II = cast<IntrinsicInst>(it);
- Intrinsic::ID ID = II->getIntrinsicID();
+ CallInst *CI = cast<CallInst>(it);
+ Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
+ assert(ID && "Not an intrinsic call!");
for (unsigned Part = 0; Part < UF; ++Part) {
SmallVector<Value*, 4> Args;
- for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i) {
- VectorParts &Arg = getVectorValue(II->getArgOperand(i));
+ for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i) {
+ VectorParts &Arg = getVectorValue(CI->getArgOperand(i));
Args.push_back(Arg[Part]);
}
- Type *Tys[] = { VectorType::get(II->getType()->getScalarType(), VF) };
+ Type *Tys[] = { VectorType::get(CI->getType()->getScalarType(), VF) };
Function *F = Intrinsic::getDeclaration(M, ID, Tys);
Entry[Part] = Builder.CreateCall(F, Args);
}
SE->forgetLoop(OrigLoop);
// Update the dominator tree information.
- assert(DT->properlyDominates(LoopBypassBlock, LoopExitBlock) &&
+ assert(DT->properlyDominates(LoopBypassBlocks.front(), LoopExitBlock) &&
"Entry does not dominate exit.");
- DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlock);
+ for (unsigned I = 1, E = LoopBypassBlocks.size(); I != E; ++I)
+ DT->addNewBlock(LoopBypassBlocks[I], LoopBypassBlocks[I-1]);
+ DT->addNewBlock(LoopVectorPreHeader, LoopBypassBlocks.back());
DT->addNewBlock(LoopVectorBody, LoopVectorPreHeader);
- DT->addNewBlock(LoopMiddleBlock, LoopBypassBlock);
+ DT->addNewBlock(LoopMiddleBlock, LoopBypassBlocks.front());
DT->addNewBlock(LoopScalarPreHeader, LoopMiddleBlock);
DT->changeImmediateDominator(LoopScalarBody, LoopScalarPreHeader);
DT->changeImmediateDominator(LoopExitBlock, LoopMiddleBlock);
if (!isa<BranchInst>(BB->getTerminator()))
return false;
- // We must have at most two predecessors because we need to convert
- // all PHIs to selects.
- unsigned Preds = std::distance(pred_begin(BB), pred_end(BB));
- if (Preds > 2)
- return false;
-
// We must be able to predicate all blocks that need to be predicated.
if (blockNeedsPredication(BB) && !blockCanBePredicated(BB))
return false;
}
+ // Check that we can actually speculate the hoistable loads.
+ if (!LoadSpeculation.canHoistAllLoads())
+ return false;
+
// We can if-convert this loop.
return true;
}
bool LoopVectorizationLegality::canVectorize() {
- assert(TheLoop->getLoopPreheader() && "No preheader!!");
+ // We must have a loop in canonical form. Loops with indirectbr in them cannot
+ // be canonicalized.
+ if (!TheLoop->getLoopPreheader())
+ return false;
// We can only vectorize innermost loops.
if (TheLoop->getSubLoopsVector().size())
// Do not loop-vectorize loops with a tiny trip count.
unsigned TC = SE->getSmallConstantTripCount(TheLoop, Latch);
- if (TC > 0u && TC < TinyTripCountThreshold) {
+ if (TC > 0u && TC < TinyTripCountVectorThreshold) {
DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " <<
"This loop is not worth vectorizing.\n");
return false;
return true;
}
+static Type *convertPointerToIntegerType(DataLayout &DL, Type *Ty) {
+ if (Ty->isPointerTy())
+ return DL.getIntPtrType(Ty->getContext());
+ return Ty;
+}
+
+static Type* getWiderType(DataLayout &DL, Type *Ty0, Type *Ty1) {
+ Ty0 = convertPointerToIntegerType(DL, Ty0);
+ Ty1 = convertPointerToIntegerType(DL, Ty1);
+ if (Ty0->getScalarSizeInBits() > Ty1->getScalarSizeInBits())
+ return Ty0;
+ return Ty1;
+}
+
bool LoopVectorizationLegality::canVectorizeInstrs() {
BasicBlock *PreHeader = TheLoop->getLoopPreheader();
BasicBlock *Header = TheLoop->getHeader();
+ // Look for the attribute signaling the absence of NaNs.
+ Function &F = *Header->getParent();
+ if (F.hasFnAttribute("no-nans-fp-math"))
+ HasFunNoNaNAttr = F.getAttributes().getAttribute(
+ AttributeSet::FunctionIndex,
+ "no-nans-fp-math").getValueAsString() == "true";
+
// For each block in the loop.
for (Loop::block_iterator bb = TheLoop->block_begin(),
be = TheLoop->block_end(); bb != be; ++bb) {
++it) {
if (PHINode *Phi = dyn_cast<PHINode>(it)) {
- // This should not happen because the loop should be normalized.
- if (Phi->getNumIncomingValues() != 2) {
- DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
- return false;
- }
-
+ Type *PhiTy = Phi->getType();
// Check that this PHI type is allowed.
- if (!Phi->getType()->isIntegerTy() &&
- !Phi->getType()->isPointerTy()) {
+ if (!PhiTy->isIntegerTy() &&
+ !PhiTy->isFloatingPointTy() &&
+ !PhiTy->isPointerTy()) {
DEBUG(dbgs() << "LV: Found an non-int non-pointer PHI.\n");
return false;
}
if (*bb != Header)
continue;
+ // We only allow if-converted PHIs with more than two incoming values.
+ if (Phi->getNumIncomingValues() != 2) {
+ DEBUG(dbgs() << "LV: Found an invalid PHI.\n");
+ return false;
+ }
+
// This is the value coming from the preheader.
Value *StartValue = Phi->getIncomingValueForBlock(PreHeader);
// Check if this is an induction variable.
InductionKind IK = isInductionVariable(Phi);
- if (NoInduction != IK) {
+ if (IK_NoInduction != IK) {
+ // Get the widest type.
+ if (!WidestIndTy)
+ WidestIndTy = convertPointerToIntegerType(*DL, PhiTy);
+ else
+ WidestIndTy = getWiderType(*DL, PhiTy, WidestIndTy);
+
// Int inductions are special because we only allow one IV.
- if (IK == IntInduction) {
- if (Induction) {
- DEBUG(dbgs() << "LV: Found too many inductions."<< *Phi <<"\n");
- return false;
- }
- Induction = Phi;
+ if (IK == IK_IntInduction) {
+ // Use the phi node with the widest type as induction. Use the last
+ // one if there are multiple (no good reason for doing this other
+ // than it is expedient).
+ if (!Induction || PhiTy == WidestIndTy)
+ Induction = Phi;
}
DEBUG(dbgs() << "LV: Found an induction variable.\n");
continue;
}
- if (AddReductionVar(Phi, IntegerAdd)) {
+ if (AddReductionVar(Phi, RK_IntegerAdd)) {
DEBUG(dbgs() << "LV: Found an ADD reduction PHI."<< *Phi <<"\n");
continue;
}
- if (AddReductionVar(Phi, IntegerMult)) {
+ if (AddReductionVar(Phi, RK_IntegerMult)) {
DEBUG(dbgs() << "LV: Found a MUL reduction PHI."<< *Phi <<"\n");
continue;
}
- if (AddReductionVar(Phi, IntegerOr)) {
+ if (AddReductionVar(Phi, RK_IntegerOr)) {
DEBUG(dbgs() << "LV: Found an OR reduction PHI."<< *Phi <<"\n");
continue;
}
- if (AddReductionVar(Phi, IntegerAnd)) {
+ if (AddReductionVar(Phi, RK_IntegerAnd)) {
DEBUG(dbgs() << "LV: Found an AND reduction PHI."<< *Phi <<"\n");
continue;
}
- if (AddReductionVar(Phi, IntegerXor)) {
+ if (AddReductionVar(Phi, RK_IntegerXor)) {
DEBUG(dbgs() << "LV: Found a XOR reduction PHI."<< *Phi <<"\n");
continue;
}
+ if (AddReductionVar(Phi, RK_IntegerMinMax)) {
+ DEBUG(dbgs() << "LV: Found a MINMAX reduction PHI."<< *Phi <<"\n");
+ continue;
+ }
+ if (AddReductionVar(Phi, RK_FloatMult)) {
+ DEBUG(dbgs() << "LV: Found an FMult reduction PHI."<< *Phi <<"\n");
+ continue;
+ }
+ if (AddReductionVar(Phi, RK_FloatAdd)) {
+ DEBUG(dbgs() << "LV: Found an FAdd reduction PHI."<< *Phi <<"\n");
+ continue;
+ }
+ if (AddReductionVar(Phi, RK_FloatMinMax)) {
+ DEBUG(dbgs() << "LV: Found an float MINMAX reduction PHI."<< *Phi <<"\n");
+ continue;
+ }
DEBUG(dbgs() << "LV: Found an unidentified PHI."<< *Phi <<"\n");
return false;
}// end of PHI handling
- // We still don't handle functions.
+ // We still don't handle functions. However, we can ignore dbg intrinsic
+ // calls and we do handle certain intrinsic and libm functions.
CallInst *CI = dyn_cast<CallInst>(it);
- if (CI && !isTriviallyVectorizableIntrinsic(it)) {
+ if (CI && !getIntrinsicIDForCall(CI, TLI) && !isa<DbgInfoIntrinsic>(CI)) {
DEBUG(dbgs() << "LV: Found a call site.\n");
return false;
}
if (!Induction) {
DEBUG(dbgs() << "LV: Did not find one integer induction var.\n");
- assert(getInductionVars()->size() && "No induction variables");
+ if (Inductions.empty())
+ return false;
}
return true;
Uniforms.insert(I);
// Insert all operands.
- for (int i = 0, Op = I->getNumOperands(); i < Op; ++i) {
- Worklist.push_back(I->getOperand(i));
- }
+ Worklist.insert(Worklist.end(), I->op_begin(), I->op_end());
}
}
+AliasAnalysis::Location
+LoopVectorizationLegality::getLoadStoreLocation(Instruction *Inst) {
+ if (StoreInst *Store = dyn_cast<StoreInst>(Inst))
+ return AA->getLocation(Store);
+ else if (LoadInst *Load = dyn_cast<LoadInst>(Inst))
+ return AA->getLocation(Load);
+
+ llvm_unreachable("Should be either load or store instruction");
+}
+
+bool
+LoopVectorizationLegality::hasPossibleGlobalWriteReorder(
+ Value *Object,
+ Instruction *Inst,
+ AliasMultiMap& WriteObjects,
+ unsigned MaxByteWidth) {
+
+ AliasAnalysis::Location ThisLoc = getLoadStoreLocation(Inst);
+
+ std::vector<Instruction*>::iterator
+ it = WriteObjects[Object].begin(),
+ end = WriteObjects[Object].end();
+
+ for (; it != end; ++it) {
+ Instruction* I = *it;
+ if (I == Inst)
+ continue;
+
+ AliasAnalysis::Location ThatLoc = getLoadStoreLocation(I);
+ if (AA->alias(ThisLoc.getWithNewSize(MaxByteWidth),
+ ThatLoc.getWithNewSize(MaxByteWidth)))
+ return true;
+ }
+ return false;
+}
+
bool LoopVectorizationLegality::canVectorizeMemory() {
+
typedef SmallVector<Value*, 16> ValueVector;
typedef SmallPtrSet<Value*, 16> ValueSet;
// Holds the Load and Store *instructions*.
PtrRtCheck.Pointers.clear();
PtrRtCheck.Need = false;
+ const bool IsAnnotatedParallel = TheLoop->isAnnotatedParallel();
+
// For each block.
for (Loop::block_iterator bb = TheLoop->block_begin(),
be = TheLoop->block_end(); bb != be; ++bb) {
if (it->mayReadFromMemory()) {
LoadInst *Ld = dyn_cast<LoadInst>(it);
if (!Ld) return false;
- if (!Ld->isSimple()) {
+ if (!Ld->isSimple() && !IsAnnotatedParallel) {
DEBUG(dbgs() << "LV: Found a non-simple load.\n");
return false;
}
if (it->mayWriteToMemory()) {
StoreInst *St = dyn_cast<StoreInst>(it);
if (!St) return false;
- if (!St->isSimple()) {
+ if (!St->isSimple() && !IsAnnotatedParallel) {
DEBUG(dbgs() << "LV: Found a non-simple store.\n");
return false;
}
return true;
}
- // Holds the read and read-write *pointers* that we find.
- ValueVector Reads;
- ValueVector ReadWrites;
+ // Holds the read and read-write *pointers* that we find. These maps hold
+ // unique values for pointers (so no need for multi-map).
+ AliasMap Reads;
+ AliasMap ReadWrites;
// Holds the analyzed pointers. We don't want to call GetUnderlyingObjects
// multiple times on the same object. If the ptr is accessed twice, once
// If we did *not* see this pointer before, insert it to
// the read-write list. At this phase it is only a 'write' list.
if (Seen.insert(Ptr))
- ReadWrites.push_back(Ptr);
+ ReadWrites.insert(std::make_pair(Ptr, ST));
+ }
+
+ if (IsAnnotatedParallel) {
+ DEBUG(dbgs()
+ << "LV: A loop annotated parallel, ignore memory dependency "
+ << "checks.\n");
+ return true;
}
for (I = Loads.begin(), IE = Loads.end(); I != IE; ++I) {
// read a few words, modify, and write a few words, and some of the
// words may be written to the same address.
if (Seen.insert(Ptr) || 0 == isConsecutivePtr(Ptr))
- Reads.push_back(Ptr);
+ Reads.insert(std::make_pair(Ptr, LD));
}
// If we write (or read-write) to a single destination and there are no
return true;
}
+ unsigned NumReadPtrs = 0;
+ unsigned NumWritePtrs = 0;
+
// Find pointers with computable bounds. We are going to use this information
// to place a runtime bound check.
bool CanDoRT = true;
- for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I)
- if (hasComputableBounds(*I)) {
- PtrRtCheck.insert(SE, TheLoop, *I);
- DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
+ AliasMap::iterator MI, ME;
+ for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
+ Value *V = (*MI).first;
+ if (hasComputableBounds(V)) {
+ PtrRtCheck.insert(SE, TheLoop, V, true);
+ NumWritePtrs++;
+ DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
} else {
CanDoRT = false;
break;
}
- for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I)
- if (hasComputableBounds(*I)) {
- PtrRtCheck.insert(SE, TheLoop, *I);
- DEBUG(dbgs() << "LV: Found a runtime check ptr:" << **I <<"\n");
+ }
+ for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
+ Value *V = (*MI).first;
+ if (hasComputableBounds(V)) {
+ PtrRtCheck.insert(SE, TheLoop, V, false);
+ NumReadPtrs++;
+ DEBUG(dbgs() << "LV: Found a runtime check ptr:" << *V <<"\n");
} else {
CanDoRT = false;
break;
}
+ }
// Check that we did not collect too many pointers or found a
// unsizeable pointer.
- if (!CanDoRT || PtrRtCheck.Pointers.size() > RuntimeMemoryCheckThreshold) {
+ unsigned NumComparisons = (NumWritePtrs * (NumReadPtrs + NumWritePtrs - 1));
+ DEBUG(dbgs() << "LV: We need to compare " << NumComparisons << " ptrs.\n");
+ if (!CanDoRT || NumComparisons > RuntimeMemoryCheckThreshold) {
PtrRtCheck.reset();
CanDoRT = false;
}
bool NeedRTCheck = false;
+ // Biggest vectorized access possible, vector width * unroll factor.
+ // TODO: We're being very pessimistic here, find a way to know the
+ // real access width before getting here.
+ unsigned MaxByteWidth = (TTI->getRegisterBitWidth(true) / 8) *
+ TTI->getMaximumUnrollFactor();
// Now that the pointers are in two lists (Reads and ReadWrites), we
// can check that there are no conflicts between each of the writes and
// between the writes to the reads.
- ValueSet WriteObjects;
+ // Note that WriteObjects duplicates the stores (indexed now by underlying
+ // objects) to avoid pointing to elements inside ReadWrites.
+ // TODO: Maybe create a new type where they can interact without duplication.
+ AliasMultiMap WriteObjects;
ValueVector TempObjects;
// Check that the read-writes do not conflict with other read-write
// pointers.
bool AllWritesIdentified = true;
- for (I = ReadWrites.begin(), IE = ReadWrites.end(); I != IE; ++I) {
- GetUnderlyingObjects(*I, TempObjects, DL);
- for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
- it != e; ++it) {
- if (!isIdentifiedObject(*it)) {
- DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **it <<"\n");
+ for (MI = ReadWrites.begin(), ME = ReadWrites.end(); MI != ME; ++MI) {
+ Value *Val = (*MI).first;
+ Instruction *Inst = (*MI).second;
+
+ GetUnderlyingObjects(Val, TempObjects, DL);
+ for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
+ UI != UE; ++UI) {
+ if (!isIdentifiedObject(*UI)) {
+ DEBUG(dbgs() << "LV: Found an unidentified write ptr:"<< **UI <<"\n");
NeedRTCheck = true;
AllWritesIdentified = false;
}
- if (!WriteObjects.insert(*it)) {
+
+ // Never seen it before, can't alias.
+ if (WriteObjects[*UI].empty()) {
+ DEBUG(dbgs() << "LV: Adding Underlying value:" << **UI <<"\n");
+ WriteObjects[*UI].push_back(Inst);
+ continue;
+ }
+ // Direct alias found.
+ if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
- << **it <<"\n");
+ << **UI <<"\n");
+ return false;
+ }
+ DEBUG(dbgs() << "LV: Found a conflicting global value:"
+ << **UI <<"\n");
+ DEBUG(dbgs() << "LV: While examining store:" << *Inst <<"\n");
+ DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
+
+ // If global alias, make sure they do alias.
+ if (hasPossibleGlobalWriteReorder(*UI,
+ Inst,
+ WriteObjects,
+ MaxByteWidth)) {
+ DEBUG(dbgs() << "LV: Found a possible write-write reorder:" << **UI
+ << "\n");
return false;
}
+
+ // Didn't alias, insert into map for further reference.
+ WriteObjects[*UI].push_back(Inst);
}
TempObjects.clear();
}
/// Check that the reads don't conflict with the read-writes.
- for (I = Reads.begin(), IE = Reads.end(); I != IE; ++I) {
- GetUnderlyingObjects(*I, TempObjects, DL);
- for (ValueVector::iterator it=TempObjects.begin(), e=TempObjects.end();
- it != e; ++it) {
+ for (MI = Reads.begin(), ME = Reads.end(); MI != ME; ++MI) {
+ Value *Val = (*MI).first;
+ GetUnderlyingObjects(Val, TempObjects, DL);
+ for (ValueVector::iterator UI=TempObjects.begin(), UE=TempObjects.end();
+ UI != UE; ++UI) {
// If all of the writes are identified then we don't care if the read
// pointer is identified or not.
- if (!AllWritesIdentified && !isIdentifiedObject(*it)) {
- DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **it <<"\n");
+ if (!AllWritesIdentified && !isIdentifiedObject(*UI)) {
+ DEBUG(dbgs() << "LV: Found an unidentified read ptr:"<< **UI <<"\n");
NeedRTCheck = true;
}
- if (WriteObjects.count(*it)) {
- DEBUG(dbgs() << "LV: Found a possible read/write reorder:"
- << **it <<"\n");
+
+ // Never seen it before, can't alias.
+ if (WriteObjects[*UI].empty())
+ continue;
+ // Direct alias found.
+ if (!AA || dyn_cast<GlobalValue>(*UI) == NULL) {
+ DEBUG(dbgs() << "LV: Found a possible write-write reorder:"
+ << **UI <<"\n");
+ return false;
+ }
+ DEBUG(dbgs() << "LV: Found a global value: "
+ << **UI <<"\n");
+ Instruction *Inst = (*MI).second;
+ DEBUG(dbgs() << "LV: While examining load:" << *Inst <<"\n");
+ DEBUG(dbgs() << "LV: On value:" << *Val <<"\n");
+
+ // If global alias, make sure they do alias.
+ if (hasPossibleGlobalWriteReorder(*UI,
+ Inst,
+ WriteObjects,
+ MaxByteWidth)) {
+ DEBUG(dbgs() << "LV: Found a possible read-write reorder:" << **UI
+ << "\n");
return false;
}
}
return true;
}
+static bool hasMultipleUsesOf(Instruction *I,
+ SmallPtrSet<Instruction *, 8> &Insts) {
+ unsigned NumUses = 0;
+ for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use) {
+ if (Insts.count(dyn_cast<Instruction>(*Use)))
+ ++NumUses;
+ if (NumUses > 1)
+ return true;
+ }
+
+ return false;
+}
+
+static bool areAllUsesIn(Instruction *I, SmallPtrSet<Instruction *, 8> &Set) {
+ for(User::op_iterator Use = I->op_begin(), E = I->op_end(); Use != E; ++Use)
+ if (!Set.count(dyn_cast<Instruction>(*Use)))
+ return false;
+ return true;
+}
+
bool LoopVectorizationLegality::AddReductionVar(PHINode *Phi,
ReductionKind Kind) {
if (Phi->getNumIncomingValues() != 2)
// This includes users of the reduction, variables (which form a cycle
// which ends in the phi node).
Instruction *ExitInstruction = 0;
+ // Indicates that we found a reduction operation in our scan.
+ bool FoundReduxOp = false;
+
+ // We start with the PHI node and scan for all of the users of this
+ // instruction. All users must be instructions that can be used as reduction
+ // variables (such as ADD). We must have a single out-of-block user. The cycle
+ // must include the original PHI.
+ bool FoundStartPHI = false;
+
+ // To recognize min/max patterns formed by a icmp select sequence, we store
+ // the number of instruction we saw from the recognized min/max pattern,
+ // to make sure we only see exactly the two instructions.
+ unsigned NumCmpSelectPatternInst = 0;
+ ReductionInstDesc ReduxDesc(false, 0);
+
+ SmallPtrSet<Instruction *, 8> VisitedInsts;
+ SmallVector<Instruction *, 8> Worklist;
+ Worklist.push_back(Phi);
+ VisitedInsts.insert(Phi);
+
+ // A value in the reduction can be used:
+ // - By the reduction:
+ // - Reduction operation:
+ // - One use of reduction value (safe).
+ // - Multiple use of reduction value (not safe).
+ // - PHI:
+ // - All uses of the PHI must be the reduction (safe).
+ // - Otherwise, not safe.
+ // - By one instruction outside of the loop (safe).
+ // - By further instructions outside of the loop (not safe).
+ // - By an instruction that is not part of the reduction (not safe).
+ // This is either:
+ // * An instruction type other than PHI or the reduction operation.
+ // * A PHI in the header other than the initial PHI.
+ while (!Worklist.empty()) {
+ Instruction *Cur = Worklist.back();
+ Worklist.pop_back();
- // Iter is our iterator. We start with the PHI node and scan for all of the
- // users of this instruction. All users must be instructions that can be
- // used as reduction variables (such as ADD). We may have a single
- // out-of-block user. The cycle must end with the original PHI.
- Instruction *Iter = Phi;
- while (true) {
- // If the instruction has no users then this is a broken
- // chain and can't be a reduction variable.
- if (Iter->use_empty())
+ // No Users.
+ // If the instruction has no users then this is a broken chain and can't be
+ // a reduction variable.
+ if (Cur->use_empty())
return false;
- // Did we find a user inside this loop already ?
- bool FoundInBlockUser = false;
- // Did we reach the initial PHI node already ?
- bool FoundStartPHI = false;
-
- // For each of the *users* of iter.
- for (Value::use_iterator it = Iter->use_begin(), e = Iter->use_end();
- it != e; ++it) {
- Instruction *U = cast<Instruction>(*it);
- // We already know that the PHI is a user.
- if (U == Phi) {
- FoundStartPHI = true;
- continue;
- }
+ bool IsAPhi = isa<PHINode>(Cur);
+
+ // A header PHI use other than the original PHI.
+ if (Cur != Phi && IsAPhi && Cur->getParent() == Phi->getParent())
+ return false;
+
+ // Reductions of instructions such as Div, and Sub is only possible if the
+ // LHS is the reduction variable.
+ if (!Cur->isCommutative() && !IsAPhi && !isa<SelectInst>(Cur) &&
+ !isa<ICmpInst>(Cur) && !isa<FCmpInst>(Cur) &&
+ !VisitedInsts.count(dyn_cast<Instruction>(Cur->getOperand(0))))
+ return false;
+
+ // Any reduction instruction must be of one of the allowed kinds.
+ ReduxDesc = isReductionInstr(Cur, Kind, ReduxDesc);
+ if (!ReduxDesc.IsReduction)
+ return false;
+
+ // A reduction operation must only have one use of the reduction value.
+ if (!IsAPhi && Kind != RK_IntegerMinMax && Kind != RK_FloatMinMax &&
+ hasMultipleUsesOf(Cur, VisitedInsts))
+ return false;
+
+ // All inputs to a PHI node must be a reduction value.
+ if(IsAPhi && Cur != Phi && !areAllUsesIn(Cur, VisitedInsts))
+ return false;
+
+ if (Kind == RK_IntegerMinMax && (isa<ICmpInst>(Cur) ||
+ isa<SelectInst>(Cur)))
+ ++NumCmpSelectPatternInst;
+ if (Kind == RK_FloatMinMax && (isa<FCmpInst>(Cur) ||
+ isa<SelectInst>(Cur)))
+ ++NumCmpSelectPatternInst;
+
+ // Check whether we found a reduction operator.
+ FoundReduxOp |= !IsAPhi;
+
+ // Process users of current instruction. Push non PHI nodes after PHI nodes
+ // onto the stack. This way we are going to have seen all inputs to PHI
+ // nodes once we get to them.
+ SmallVector<Instruction *, 8> NonPHIs;
+ SmallVector<Instruction *, 8> PHIs;
+ for (Value::use_iterator UI = Cur->use_begin(), E = Cur->use_end(); UI != E;
+ ++UI) {
+ Instruction *Usr = cast<Instruction>(*UI);
// Check if we found the exit user.
- BasicBlock *Parent = U->getParent();
+ BasicBlock *Parent = Usr->getParent();
if (!TheLoop->contains(Parent)) {
// Exit if you find multiple outside users.
if (ExitInstruction != 0)
return false;
- ExitInstruction = Iter;
+ ExitInstruction = Cur;
+ continue;
}
- // We allow in-loop PHINodes which are not the original reduction PHI
- // node. If this PHI is the only user of Iter (happens in IF w/ no ELSE
- // structure) then don't skip this PHI.
- if (isa<PHINode>(Iter) && isa<PHINode>(U) &&
- U->getParent() != TheLoop->getHeader() &&
- TheLoop->contains(U) &&
- Iter->getNumUses() > 1)
- continue;
+ // Process instructions only once (termination).
+ if (VisitedInsts.insert(Usr)) {
+ if (isa<PHINode>(Usr))
+ PHIs.push_back(Usr);
+ else
+ NonPHIs.push_back(Usr);
+ }
+ // Remember that we completed the cycle.
+ if (Usr == Phi)
+ FoundStartPHI = true;
+ }
+ Worklist.append(PHIs.begin(), PHIs.end());
+ Worklist.append(NonPHIs.begin(), NonPHIs.end());
+ }
- // We can't have multiple inside users.
- if (FoundInBlockUser)
- return false;
- FoundInBlockUser = true;
+ // This means we have seen one but not the other instruction of the
+ // pattern or more than just a select and cmp.
+ if ((Kind == RK_IntegerMinMax || Kind == RK_FloatMinMax) &&
+ NumCmpSelectPatternInst != 2)
+ return false;
- // Any reduction instr must be of one of the allowed kinds.
- if (!isReductionInstr(U, Kind))
- return false;
+ if (!FoundStartPHI || !FoundReduxOp || !ExitInstruction)
+ return false;
- // Reductions of instructions such as Div, and Sub is only
- // possible if the LHS is the reduction variable.
- if (!U->isCommutative() && U->getOperand(0) != Iter)
- return false;
+ // We found a reduction var if we have reached the original phi node and we
+ // only have a single instruction with out-of-loop users.
- Iter = U;
- }
+ // This instruction is allowed to have out-of-loop users.
+ AllowedExit.insert(ExitInstruction);
- // We found a reduction var if we have reached the original
- // phi node and we only have a single instruction with out-of-loop
- // users.
- if (FoundStartPHI && ExitInstruction) {
- // This instruction is allowed to have out-of-loop users.
- AllowedExit.insert(ExitInstruction);
+ // Save the description of this reduction variable.
+ ReductionDescriptor RD(RdxStart, ExitInstruction, Kind,
+ ReduxDesc.MinMaxKind);
+ Reductions[Phi] = RD;
+ // We've ended the cycle. This is a reduction variable if we have an
+ // outside user and it has a binary op.
- // Save the description of this reduction variable.
- ReductionDescriptor RD(RdxStart, ExitInstruction, Kind);
- Reductions[Phi] = RD;
- return true;
- }
+ return true;
+}
- // If we've reached the start PHI but did not find an outside user then
- // this is dead code. Abort.
- if (FoundStartPHI)
- return false;
+/// Returns true if the instruction is a Select(ICmp(X, Y), X, Y) instruction
+/// pattern corresponding to a min(X, Y) or max(X, Y).
+LoopVectorizationLegality::ReductionInstDesc
+LoopVectorizationLegality::isMinMaxSelectCmpPattern(Instruction *I,
+ ReductionInstDesc &Prev) {
+
+ assert((isa<ICmpInst>(I) || isa<FCmpInst>(I) || isa<SelectInst>(I)) &&
+ "Expect a select instruction");
+ Instruction *Cmp = 0;
+ SelectInst *Select = 0;
+
+ // We must handle the select(cmp()) as a single instruction. Advance to the
+ // select.
+ if ((Cmp = dyn_cast<ICmpInst>(I)) || (Cmp = dyn_cast<FCmpInst>(I))) {
+ if (!Cmp->hasOneUse() || !(Select = dyn_cast<SelectInst>(*I->use_begin())))
+ return ReductionInstDesc(false, I);
+ return ReductionInstDesc(Select, Prev.MinMaxKind);
}
+
+ // Only handle single use cases for now.
+ if (!(Select = dyn_cast<SelectInst>(I)))
+ return ReductionInstDesc(false, I);
+ if (!(Cmp = dyn_cast<ICmpInst>(I->getOperand(0))) &&
+ !(Cmp = dyn_cast<FCmpInst>(I->getOperand(0))))
+ return ReductionInstDesc(false, I);
+ if (!Cmp->hasOneUse())
+ return ReductionInstDesc(false, I);
+
+ Value *CmpLeft;
+ Value *CmpRight;
+
+ // Look for a min/max pattern.
+ if (m_UMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
+ return ReductionInstDesc(Select, MRK_UIntMin);
+ else if (m_UMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
+ return ReductionInstDesc(Select, MRK_UIntMax);
+ else if (m_SMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
+ return ReductionInstDesc(Select, MRK_SIntMax);
+ else if (m_SMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
+ return ReductionInstDesc(Select, MRK_SIntMin);
+ else if (m_OrdFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
+ return ReductionInstDesc(Select, MRK_FloatMin);
+ else if (m_OrdFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
+ return ReductionInstDesc(Select, MRK_FloatMax);
+ else if (m_UnordFMin(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
+ return ReductionInstDesc(Select, MRK_FloatMin);
+ else if (m_UnordFMax(m_Value(CmpLeft), m_Value(CmpRight)).match(Select))
+ return ReductionInstDesc(Select, MRK_FloatMax);
+
+ return ReductionInstDesc(false, I);
}
-bool
+LoopVectorizationLegality::ReductionInstDesc
LoopVectorizationLegality::isReductionInstr(Instruction *I,
- ReductionKind Kind) {
+ ReductionKind Kind,
+ ReductionInstDesc &Prev) {
+ bool FP = I->getType()->isFloatingPointTy();
+ bool FastMath = (FP && I->isCommutative() && I->isAssociative());
switch (I->getOpcode()) {
default:
- return false;
+ return ReductionInstDesc(false, I);
case Instruction::PHI:
- // possibly.
- return true;
+ if (FP && (Kind != RK_FloatMult && Kind != RK_FloatAdd &&
+ Kind != RK_FloatMinMax))
+ return ReductionInstDesc(false, I);
+ return ReductionInstDesc(I, Prev.MinMaxKind);
case Instruction::Sub:
case Instruction::Add:
- return Kind == IntegerAdd;
- case Instruction::SDiv:
- case Instruction::UDiv:
+ return ReductionInstDesc(Kind == RK_IntegerAdd, I);
case Instruction::Mul:
- return Kind == IntegerMult;
+ return ReductionInstDesc(Kind == RK_IntegerMult, I);
case Instruction::And:
- return Kind == IntegerAnd;
+ return ReductionInstDesc(Kind == RK_IntegerAnd, I);
case Instruction::Or:
- return Kind == IntegerOr;
+ return ReductionInstDesc(Kind == RK_IntegerOr, I);
case Instruction::Xor:
- return Kind == IntegerXor;
+ return ReductionInstDesc(Kind == RK_IntegerXor, I);
+ case Instruction::FMul:
+ return ReductionInstDesc(Kind == RK_FloatMult && FastMath, I);
+ case Instruction::FAdd:
+ return ReductionInstDesc(Kind == RK_FloatAdd && FastMath, I);
+ case Instruction::FCmp:
+ case Instruction::ICmp:
+ case Instruction::Select:
+ if (Kind != RK_IntegerMinMax &&
+ (!HasFunNoNaNAttr || Kind != RK_FloatMinMax))
+ return ReductionInstDesc(false, I);
+ return isMinMaxSelectCmpPattern(I, Prev);
}
}
Type *PhiTy = Phi->getType();
// We only handle integer and pointer inductions variables.
if (!PhiTy->isIntegerTy() && !PhiTy->isPointerTy())
- return NoInduction;
+ return IK_NoInduction;
- // Check that the PHI is consecutive and starts at zero.
+ // Check that the PHI is consecutive.
const SCEV *PhiScev = SE->getSCEV(Phi);
const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(PhiScev);
if (!AR) {
DEBUG(dbgs() << "LV: PHI is not a poly recurrence.\n");
- return NoInduction;
+ return IK_NoInduction;
}
const SCEV *Step = AR->getStepRecurrence(*SE);
// Integer inductions need to have a stride of one.
if (PhiTy->isIntegerTy()) {
if (Step->isOne())
- return IntInduction;
+ return IK_IntInduction;
if (Step->isAllOnesValue())
- return ReverseIntInduction;
- return NoInduction;
+ return IK_ReverseIntInduction;
+ return IK_NoInduction;
}
// Calculate the pointer stride and check if it is consecutive.
const SCEVConstant *C = dyn_cast<SCEVConstant>(Step);
if (!C)
- return NoInduction;
+ return IK_NoInduction;
assert(PhiTy->isPointerTy() && "The PHI must be a pointer");
uint64_t Size = DL->getTypeAllocSize(PhiTy->getPointerElementType());
if (C->getValue()->equalsInt(Size))
- return PtrInduction;
+ return IK_PtrInduction;
+ else if (C->getValue()->equalsInt(0 - Size))
+ return IK_ReversePtrInduction;
- return NoInduction;
+ return IK_NoInduction;
}
bool LoopVectorizationLegality::isInductionVariable(const Value *V) {
bool LoopVectorizationLegality::blockCanBePredicated(BasicBlock *BB) {
for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
- // We don't predicate loads/stores at the moment.
- if (it->mayReadFromMemory() || it->mayWriteToMemory() || it->mayThrow())
+ // We might be able to hoist the load.
+ if (it->mayReadFromMemory() && !LoadSpeculation.isHoistableLoad(it))
+ return false;
+
+ // We don't predicate stores at the moment.
+ if (it->mayWriteToMemory() || it->mayThrow())
return false;
// The instructions below can trap.
return AR->isAffine();
}
-unsigned
+LoopVectorizationCostModel::VectorizationFactor
LoopVectorizationCostModel::selectVectorizationFactor(bool OptForSize,
unsigned UserVF) {
+ // Width 1 means no vectorize
+ VectorizationFactor Factor = { 1U, 0U };
if (OptForSize && Legal->getRuntimePointerCheck()->Need) {
DEBUG(dbgs() << "LV: Aborting. Runtime ptr check is required in Os.\n");
- return 1;
+ return Factor;
}
// Find the trip count.
unsigned TC = SE->getSmallConstantTripCount(TheLoop, TheLoop->getLoopLatch());
DEBUG(dbgs() << "LV: Found trip count:"<<TC<<"\n");
+ unsigned WidestType = getWidestType();
+ unsigned WidestRegister = TTI.getRegisterBitWidth(true);
+ unsigned MaxVectorSize = WidestRegister / WidestType;
+ DEBUG(dbgs() << "LV: The Widest type: " << WidestType << " bits.\n");
+ DEBUG(dbgs() << "LV: The Widest register is:" << WidestRegister << "bits.\n");
+
+ if (MaxVectorSize == 0) {
+ DEBUG(dbgs() << "LV: The target has no vector registers.\n");
+ MaxVectorSize = 1;
+ }
+
+ assert(MaxVectorSize <= 32 && "Did not expect to pack so many elements"
+ " into one vector!");
+
unsigned VF = MaxVectorSize;
// If we optimize the program for size, avoid creating the tail loop.
// If we are unable to calculate the trip count then don't try to vectorize.
if (TC < 2) {
DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
- return 1;
+ return Factor;
}
// Find the maximum SIMD width that can fit within the trip count.
// zero then we require a tail.
if (VF < 2) {
DEBUG(dbgs() << "LV: Aborting. A tail loop is required in Os.\n");
- return 1;
+ return Factor;
}
}
assert(isPowerOf2_32(UserVF) && "VF needs to be a power of two");
DEBUG(dbgs() << "LV: Using user VF "<<UserVF<<".\n");
- return UserVF;
- }
-
- if (!TTI) {
- DEBUG(dbgs() << "LV: No vector target information. Not vectorizing. \n");
- return 1;
+ Factor.Width = UserVF;
+ return Factor;
}
float Cost = expectedCost(1);
}
DEBUG(dbgs() << "LV: Selecting VF = : "<< Width << ".\n");
- return Width;
+ Factor.Width = Width;
+ Factor.Cost = Width * Cost;
+ return Factor;
+}
+
+unsigned LoopVectorizationCostModel::getWidestType() {
+ unsigned MaxWidth = 8;
+
+ // For each block.
+ for (Loop::block_iterator bb = TheLoop->block_begin(),
+ be = TheLoop->block_end(); bb != be; ++bb) {
+ BasicBlock *BB = *bb;
+
+ // For each instruction in the loop.
+ for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
+ Type *T = it->getType();
+
+ // Only examine Loads, Stores and PHINodes.
+ if (!isa<LoadInst>(it) && !isa<StoreInst>(it) && !isa<PHINode>(it))
+ continue;
+
+ // Examine PHI nodes that are reduction variables.
+ if (PHINode *PN = dyn_cast<PHINode>(it))
+ if (!Legal->getReductionVars()->count(PN))
+ continue;
+
+ // Examine the stored values.
+ if (StoreInst *ST = dyn_cast<StoreInst>(it))
+ T = ST->getValueOperand()->getType();
+
+ // Ignore loaded pointer types and stored pointer types that are not
+ // consecutive. However, we do want to take consecutive stores/loads of
+ // pointer vectors into account.
+ if (T->isPointerTy() && !isConsecutiveLoadOrStore(it))
+ continue;
+
+ MaxWidth = std::max(MaxWidth,
+ (unsigned)DL->getTypeSizeInBits(T->getScalarType()));
+ }
+ }
+
+ return MaxWidth;
}
unsigned
LoopVectorizationCostModel::selectUnrollFactor(bool OptForSize,
- unsigned UserUF) {
+ unsigned UserUF,
+ unsigned VF,
+ unsigned LoopCost) {
+
+ // -- The unroll heuristics --
+ // We unroll the loop in order to expose ILP and reduce the loop overhead.
+ // There are many micro-architectural considerations that we can't predict
+ // at this level. For example frontend pressure (on decode or fetch) due to
+ // code size, or the number and capabilities of the execution ports.
+ //
+ // We use the following heuristics to select the unroll factor:
+ // 1. If the code has reductions the we unroll in order to break the cross
+ // iteration dependency.
+ // 2. If the loop is really small then we unroll in order to reduce the loop
+ // overhead.
+ // 3. We don't unroll if we think that we will spill registers to memory due
+ // to the increased register pressure.
+
// Use the user preference, unless 'auto' is selected.
if (UserUF != 0)
return UserUF;
if (OptForSize)
return 1;
- unsigned TargetVectorRegisters = TTI->getNumberOfRegisters(true);
+ // Do not unroll loops with a relatively small trip count.
+ unsigned TC = SE->getSmallConstantTripCount(TheLoop,
+ TheLoop->getLoopLatch());
+ if (TC > 1 && TC < TinyTripCountUnrollThreshold)
+ return 1;
+
+ unsigned TargetVectorRegisters = TTI.getNumberOfRegisters(true);
DEBUG(dbgs() << "LV: The target has " << TargetVectorRegisters <<
" vector registers\n");
// fit without causing spills.
unsigned UF = (TargetVectorRegisters - R.LoopInvariantRegs) / R.MaxLocalUsers;
- // We don't want to unroll the loops to the point where they do not fit into
- // the decoded cache. Assume that we only allow 32 IR instructions.
- UF = std::min(UF, (32 / R.NumInstructions));
-
// Clamp the unroll factor ranges to reasonable factors.
+ unsigned MaxUnrollSize = TTI.getMaximumUnrollFactor();
+
+ // If we did not calculate the cost for VF (because the user selected the VF)
+ // then we calculate the cost of VF here.
+ if (LoopCost == 0)
+ LoopCost = expectedCost(VF);
+
+ // Clamp the calculated UF to be between the 1 and the max unroll factor
+ // that the target allows.
if (UF > MaxUnrollSize)
UF = MaxUnrollSize;
else if (UF < 1)
UF = 1;
- return UF;
+ if (Legal->getReductionVars()->size()) {
+ DEBUG(dbgs() << "LV: Unrolling because of reductions. \n");
+ return UF;
+ }
+
+ // We want to unroll tiny loops in order to reduce the loop overhead.
+ // We assume that the cost overhead is 1 and we use the cost model
+ // to estimate the cost of the loop and unroll until the cost of the
+ // loop overhead is about 5% of the cost of the loop.
+ DEBUG(dbgs() << "LV: Loop cost is "<< LoopCost <<" \n");
+ if (LoopCost < 20) {
+ DEBUG(dbgs() << "LV: Unrolling to reduce branch cost. \n");
+ unsigned NewUF = 20/LoopCost + 1;
+ return std::min(NewUF, UF);
+ }
+
+ DEBUG(dbgs() << "LV: Not Unrolling. \n");
+ return 1;
}
LoopVectorizationCostModel::RegisterUsage
// For each instruction in the old loop.
for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
+ // Skip dbg intrinsics.
+ if (isa<DbgInfoIntrinsic>(it))
+ continue;
+
unsigned C = getInstructionCost(it, VF);
Cost += C;
DEBUG(dbgs() << "LV: Found an estimated cost of "<< C <<" for VF " <<
unsigned
LoopVectorizationCostModel::getInstructionCost(Instruction *I, unsigned VF) {
- assert(TTI && "Invalid vector target transformation info");
-
// If we know that this instruction will remain uniform, check the cost of
// the scalar version.
if (Legal->isUniformAfterVectorization(I))
// TODO: We need to estimate the cost of intrinsic calls.
switch (I->getOpcode()) {
case Instruction::GetElementPtr:
- // We mark this instruction as zero-cost because scalar GEPs are usually
- // lowered to the intruction addressing mode. At the moment we don't
- // generate vector geps.
+ // We mark this instruction as zero-cost because the cost of GEPs in
+ // vectorized code depends on whether the corresponding memory instruction
+ // is scalarized or not. Therefore, we handle GEPs with the memory
+ // instruction cost.
return 0;
case Instruction::Br: {
- return TTI->getCFInstrCost(I->getOpcode());
+ return TTI.getCFInstrCost(I->getOpcode());
}
case Instruction::PHI:
//TODO: IF-converted IFs become selects.
case Instruction::AShr:
case Instruction::And:
case Instruction::Or:
- case Instruction::Xor:
- return TTI->getArithmeticInstrCost(I->getOpcode(), VectorTy);
+ case Instruction::Xor: {
+ // Certain instructions can be cheaper to vectorize if they have a constant
+ // second vector operand. One example of this are shifts on x86.
+ TargetTransformInfo::OperandValueKind Op1VK =
+ TargetTransformInfo::OK_AnyValue;
+ TargetTransformInfo::OperandValueKind Op2VK =
+ TargetTransformInfo::OK_AnyValue;
+
+ if (isa<ConstantInt>(I->getOperand(1)))
+ Op2VK = TargetTransformInfo::OK_UniformConstantValue;
+
+ return TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, Op1VK, Op2VK);
+ }
case Instruction::Select: {
SelectInst *SI = cast<SelectInst>(I);
const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());
bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));
Type *CondTy = SI->getCondition()->getType();
- if (ScalarCond)
+ if (!ScalarCond)
CondTy = VectorType::get(CondTy, VF);
- return TTI->getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
+ return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy);
}
case Instruction::ICmp:
case Instruction::FCmp: {
Type *ValTy = I->getOperand(0)->getType();
VectorTy = ToVectorTy(ValTy, VF);
- return TTI->getCmpSelInstrCost(I->getOpcode(), VectorTy);
+ return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy);
}
- case Instruction::Store: {
- StoreInst *SI = cast<StoreInst>(I);
- Type *ValTy = SI->getValueOperand()->getType();
+ case Instruction::Store:
+ case Instruction::Load: {
+ StoreInst *SI = dyn_cast<StoreInst>(I);
+ LoadInst *LI = dyn_cast<LoadInst>(I);
+ Type *ValTy = (SI ? SI->getValueOperand()->getType() :
+ LI->getType());
VectorTy = ToVectorTy(ValTy, VF);
+ unsigned Alignment = SI ? SI->getAlignment() : LI->getAlignment();
+ unsigned AS = SI ? SI->getPointerAddressSpace() :
+ LI->getPointerAddressSpace();
+ Value *Ptr = SI ? SI->getPointerOperand() : LI->getPointerOperand();
+ // We add the cost of address computation here instead of with the gep
+ // instruction because only here we know whether the operation is
+ // scalarized.
if (VF == 1)
- return TTI->getMemoryOpCost(I->getOpcode(), VectorTy,
- SI->getAlignment(),
- SI->getPointerAddressSpace());
-
- // Scalarized stores.
- int Stride = Legal->isConsecutivePtr(SI->getPointerOperand());
- bool Reverse = Stride < 0;
- if (0 == Stride) {
+ return TTI.getAddressComputationCost(VectorTy) +
+ TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
+
+ // Scalarized loads/stores.
+ int ConsecutiveStride = Legal->isConsecutivePtr(Ptr);
+ bool Reverse = ConsecutiveStride < 0;
+ unsigned ScalarAllocatedSize = DL->getTypeAllocSize(ValTy);
+ unsigned VectorElementSize = DL->getTypeStoreSize(VectorTy)/VF;
+ if (!ConsecutiveStride || ScalarAllocatedSize != VectorElementSize) {
unsigned Cost = 0;
-
// The cost of extracting from the value vector and pointer vector.
- Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF);
+ Type *PtrTy = ToVectorTy(Ptr->getType(), VF);
for (unsigned i = 0; i < VF; ++i) {
- Cost += TTI->getVectorInstrCost(Instruction::ExtractElement,
- VectorTy, i);
- Cost += TTI->getVectorInstrCost(Instruction::ExtractElement,
- PtrTy, i);
+ // The cost of extracting the pointer operand.
+ Cost += TTI.getVectorInstrCost(Instruction::ExtractElement, PtrTy, i);
+ // In case of STORE, the cost of ExtractElement from the vector.
+ // In case of LOAD, the cost of InsertElement into the returned
+ // vector.
+ Cost += TTI.getVectorInstrCost(SI ? Instruction::ExtractElement :
+ Instruction::InsertElement,
+ VectorTy, i);
}
- // The cost of the scalar stores.
- Cost += VF * TTI->getMemoryOpCost(I->getOpcode(),
- ValTy->getScalarType(),
- SI->getAlignment(),
- SI->getPointerAddressSpace());
+ // The cost of the scalar loads/stores.
+ Cost += VF * TTI.getAddressComputationCost(ValTy->getScalarType());
+ Cost += VF * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(),
+ Alignment, AS);
return Cost;
}
- // Wide stores.
- unsigned Cost = TTI->getMemoryOpCost(I->getOpcode(), VectorTy,
- SI->getAlignment(),
- SI->getPointerAddressSpace());
- if (Reverse)
- Cost += TTI->getShuffleCost(TargetTransformInfo::SK_Reverse,
- VectorTy, 0);
- return Cost;
- }
- case Instruction::Load: {
- LoadInst *LI = cast<LoadInst>(I);
+ // Wide load/stores.
+ unsigned Cost = TTI.getAddressComputationCost(VectorTy);
+ Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS);
- if (VF == 1)
- return TTI->getMemoryOpCost(I->getOpcode(), VectorTy,
- LI->getAlignment(),
- LI->getPointerAddressSpace());
-
- // Scalarized loads.
- int Stride = Legal->isConsecutivePtr(LI->getPointerOperand());
- bool Reverse = Stride < 0;
- if (0 == Stride) {
- unsigned Cost = 0;
- Type *PtrTy = ToVectorTy(I->getOperand(0)->getType(), VF);
-
- // The cost of extracting from the pointer vector.
- for (unsigned i = 0; i < VF; ++i)
- Cost += TTI->getVectorInstrCost(Instruction::ExtractElement,
- PtrTy, i);
-
- // The cost of inserting data to the result vector.
- for (unsigned i = 0; i < VF; ++i)
- Cost += TTI->getVectorInstrCost(Instruction::InsertElement,
- VectorTy, i);
-
- // The cost of the scalar stores.
- Cost += VF * TTI->getMemoryOpCost(I->getOpcode(),
- RetTy->getScalarType(),
- LI->getAlignment(),
- LI->getPointerAddressSpace());
- return Cost;
- }
-
- // Wide loads.
- unsigned Cost = TTI->getMemoryOpCost(I->getOpcode(), VectorTy,
- LI->getAlignment(),
- LI->getPointerAddressSpace());
if (Reverse)
- Cost += TTI->getShuffleCost(TargetTransformInfo::SK_Reverse,
+ Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
VectorTy, 0);
return Cost;
}
// The cost of these is the same as the scalar operation.
if (I->getOpcode() == Instruction::Trunc &&
Legal->isInductionVariable(I->getOperand(0)))
- return TTI->getCastInstrCost(I->getOpcode(), I->getType(),
- I->getOperand(0)->getType());
+ return TTI.getCastInstrCost(I->getOpcode(), I->getType(),
+ I->getOperand(0)->getType());
Type *SrcVecTy = ToVectorTy(I->getOperand(0)->getType(), VF);
- return TTI->getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
+ return TTI.getCastInstrCost(I->getOpcode(), VectorTy, SrcVecTy);
}
case Instruction::Call: {
- assert(isTriviallyVectorizableIntrinsic(I));
- IntrinsicInst *II = cast<IntrinsicInst>(I);
- Type *RetTy = ToVectorTy(II->getType(), VF);
+ CallInst *CI = cast<CallInst>(I);
+ Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
+ assert(ID && "Not an intrinsic call!");
+ Type *RetTy = ToVectorTy(CI->getType(), VF);
SmallVector<Type*, 4> Tys;
- for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i)
- Tys.push_back(ToVectorTy(II->getArgOperand(i)->getType(), VF));
- return TTI->getIntrinsicInstrCost(II->getIntrinsicID(), RetTy, Tys);
+ for (unsigned i = 0, ie = CI->getNumArgOperands(); i != ie; ++i)
+ Tys.push_back(ToVectorTy(CI->getArgOperand(i)->getType(), VF));
+ return TTI.getIntrinsicInstrCost(ID, RetTy, Tys);
}
default: {
// We are scalarizing the instruction. Return the cost of the scalar
unsigned Cost = 0;
if (!RetTy->isVoidTy() && VF != 1) {
- unsigned InsCost = TTI->getVectorInstrCost(Instruction::InsertElement,
- VectorTy);
- unsigned ExtCost = TTI->getVectorInstrCost(Instruction::ExtractElement,
- VectorTy);
+ unsigned InsCost = TTI.getVectorInstrCost(Instruction::InsertElement,
+ VectorTy);
+ unsigned ExtCost = TTI.getVectorInstrCost(Instruction::ExtractElement,
+ VectorTy);
// The cost of inserting the results plus extracting each one of the
// operands.
// The cost of executing VF copies of the scalar instruction. This opcode
// is unknown. Assume that it is the same as 'mul'.
- Cost += VF * TTI->getArithmeticInstrCost(Instruction::Mul, VectorTy);
+ Cost += VF * TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy);
return Cost;
}
}// end of switch.
static const char lv_name[] = "Loop Vectorization";
INITIALIZE_PASS_BEGIN(LoopVectorize, LV_NAME, lv_name, false, false)
INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
+INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
INITIALIZE_PASS_END(LoopVectorize, LV_NAME, lv_name, false, false)
}
}
+bool LoopVectorizationCostModel::isConsecutiveLoadOrStore(Instruction *Inst) {
+ // Check for a store.
+ if (StoreInst *ST = dyn_cast<StoreInst>(Inst))
+ return Legal->isConsecutivePtr(ST->getPointerOperand()) != 0;
+ // Check for a load.
+ if (LoadInst *LI = dyn_cast<LoadInst>(Inst))
+ return Legal->isConsecutivePtr(LI->getPointerOperand()) != 0;
+
+ return false;
+}
projects / oota-llvm.git / blobdiff