MemCpyOptimizer.cpp

//===- MemCpyOptimizer.cpp - Optimize use of memcpy and friends -----------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This pass performs various transformations related to eliminating memcpy
// calls, or transforming sets of stores into memset's.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "memcpyopt"
#include "llvm/Transforms/Scalar.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/Dominators.h"
#include "llvm/Analysis/MemoryDependenceAnalysis.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/GlobalVariable.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetLibraryInfo.h"
#include "llvm/Transforms/Utils/Local.h"
#include <list>
using namespace llvm;

STATISTIC(NumMemCpyInstr, "Number of memcpy instructions deleted");
STATISTIC(NumMemSetInfer, "Number of memsets inferred");
STATISTIC(NumMoveToCpy,   "Number of memmoves converted to memcpy");
STATISTIC(NumCpyToSet,    "Number of memcpys converted to memset");

static int64_t GetOffsetFromIndex(const GEPOperator *GEP, unsigned Idx,
                                  bool &VariableIdxFound, const DataLayout &TD){
  // Skip over the first indices.
  gep_type_iterator GTI = gep_type_begin(GEP);
  for (unsigned i = 1; i != Idx; ++i, ++GTI)
    /*skip along*/;

  // Compute the offset implied by the rest of the indices.
  int64_t Offset = 0;
  for (unsigned i = Idx, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
    ConstantInt *OpC = dyn_cast<ConstantInt>(GEP->getOperand(i));
    if (OpC == 0)
      return VariableIdxFound = true;
    if (OpC->isZero()) continue;  // No offset.

    // Handle struct indices, which add their field offset to the pointer.
    if (StructType *STy = dyn_cast<StructType>(*GTI)) {
      Offset += TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
      continue;
    }

    // Otherwise, we have a sequential type like an array or vector.  Multiply
    // the index by the ElementSize.
    uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
    Offset += Size*OpC->getSExtValue();
  }

  return Offset;
}

/// IsPointerOffset - Return true if Ptr1 is provably equal to Ptr2 plus a
/// constant offset, and return that constant offset.  For example, Ptr1 might
/// be &A[42], and Ptr2 might be &A[40].  In this case offset would be -8.
static bool IsPointerOffset(Value *Ptr1, Value *Ptr2, int64_t &Offset,
                            const DataLayout &TD) {
  Ptr1 = Ptr1->stripPointerCasts();
  Ptr2 = Ptr2->stripPointerCasts();
  GEPOperator *GEP1 = dyn_cast<GEPOperator>(Ptr1);
  GEPOperator *GEP2 = dyn_cast<GEPOperator>(Ptr2);

  bool VariableIdxFound = false;

  // If one pointer is a GEP and the other isn't, then see if the GEP is a
  // constant offset from the base, as in "P" and "gep P, 1".
  if (GEP1 && GEP2 == 0 && GEP1->getOperand(0)->stripPointerCasts() == Ptr2) {
    Offset = -GetOffsetFromIndex(GEP1, 1, VariableIdxFound, TD);
    return !VariableIdxFound;
  }

  if (GEP2 && GEP1 == 0 && GEP2->getOperand(0)->stripPointerCasts() == Ptr1) {
    Offset = GetOffsetFromIndex(GEP2, 1, VariableIdxFound, TD);
    return !VariableIdxFound;
  }

  // Right now we handle the case when Ptr1/Ptr2 are both GEPs with an identical
  // base.  After that base, they may have some number of common (and
  // potentially variable) indices.  After that they handle some constant
  // offset, which determines their offset from each other.  At this point, we
  // handle no other case.
  if (!GEP1 || !GEP2 || GEP1->getOperand(0) != GEP2->getOperand(0))
    return false;

  // Skip any common indices and track the GEP types.
  unsigned Idx = 1;
  for (; Idx != GEP1->getNumOperands() && Idx != GEP2->getNumOperands(); ++Idx)
    if (GEP1->getOperand(Idx) != GEP2->getOperand(Idx))
      break;

  int64_t Offset1 = GetOffsetFromIndex(GEP1, Idx, VariableIdxFound, TD);
  int64_t Offset2 = GetOffsetFromIndex(GEP2, Idx, VariableIdxFound, TD);
  if (VariableIdxFound) return false;

  Offset = Offset2-Offset1;
  return true;
}


/// MemsetRange - Represents a range of memset'd bytes with the ByteVal value.
/// This allows us to analyze stores like:
///   store 0 -> P+1
///   store 0 -> P+0
///   store 0 -> P+3
///   store 0 -> P+2
/// which sometimes happens with stores to arrays of structs etc.  When we see
/// the first store, we make a range [1, 2).  The second store extends the range
/// to [0, 2).  The third makes a new range [2, 3).  The fourth store joins the
/// two ranges into [0, 3) which is memset'able.
namespace {
struct MemsetRange {
  // Start/End - A semi range that describes the span that this range covers.
  // The range is closed at the start and open at the end: [Start, End).
  int64_t Start, End;

  /// StartPtr - The getelementptr instruction that points to the start of the
  /// range.
  Value *StartPtr;

  /// Alignment - The known alignment of the first store.
  unsigned Alignment;

  /// TheStores - The actual stores that make up this range.
  SmallVector<Instruction*, 16> TheStores;

  bool isProfitableToUseMemset(const DataLayout &TD) const;

};
} // end anon namespace

bool MemsetRange::isProfitableToUseMemset(const DataLayout &TD) const {
  // If we found more than 4 stores to merge or 16 bytes, use memset.
  if (TheStores.size() >= 4 || End-Start >= 16) return true;

  // If there is nothing to merge, don't do anything.
  if (TheStores.size() < 2) return false;

  // If any of the stores are a memset, then it is always good to extend the
  // memset.
  for (unsigned i = 0, e = TheStores.size(); i != e; ++i)
    if (!isa<StoreInst>(TheStores[i]))
      return true;

  // Assume that the code generator is capable of merging pairs of stores
  // together if it wants to.
  if (TheStores.size() == 2) return false;

  // If we have fewer than 8 stores, it can still be worthwhile to do this.
  // For example, merging 4 i8 stores into an i32 store is useful almost always.
  // However, merging 2 32-bit stores isn't useful on a 32-bit architecture (the
  // memset will be split into 2 32-bit stores anyway) and doing so can
  // pessimize the llvm optimizer.
  //
  // Since we don't have perfect knowledge here, make some assumptions: assume
  // the maximum GPR width is the same size as the largest legal integer
  // size. If so, check to see whether we will end up actually reducing the
  // number of stores used.
  unsigned Bytes = unsigned(End-Start);
  unsigned MaxIntSize = TD.getLargestLegalIntTypeSize();
  if (MaxIntSize == 0)
    MaxIntSize = 1;
  unsigned NumPointerStores = Bytes / MaxIntSize;

  // Assume the remaining bytes if any are done a byte at a time.
  unsigned NumByteStores = Bytes - NumPointerStores * MaxIntSize;

  // If we will reduce the # stores (according to this heuristic), do the
  // transformation.  This encourages merging 4 x i8 -> i32 and 2 x i16 -> i32
  // etc.
  return TheStores.size() > NumPointerStores+NumByteStores;
}


namespace {
class MemsetRanges {
  /// Ranges - A sorted list of the memset ranges.  We use std::list here
  /// because each element is relatively large and expensive to copy.
  std::list<MemsetRange> Ranges;
  typedef std::list<MemsetRange>::iterator range_iterator;
  const DataLayout &TD;
public:
  MemsetRanges(const DataLayout &td) : TD(td) {}