Generalize IVUsers to track arbitrary expressions rather than expressions

explicitly split into stride-and-offset pairs. Also, add the
ability to track multiple post-increment loops on the same expression.

This refines the concept of "normalizing" SCEV expressions used for
to post-increment uses, and introduces a dedicated utility routine for
normalizing and denormalizing expressions.

This fixes the expansion of expressions which are post-increment users
of more than one loop at a time. More broadly, this takes LSR another
step closer to being able to reason about more than one loop at a time.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@100699 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 346 insertions, 224 deletions

diff --git a/lib/Analysis/IVUsers.cpp b/lib/Analysis/IVUsers.cpp
index 47b5d4a0f0..467f9dd840 100644
--- a/lib/Analysis/IVUsers.cpp
+++ b/lib/Analysis/IVUsers.cpp
@@ -62,120 +62,34 @@ static void CollectSubexprs(const SCEV *S,
   Ops.push_back(S);
 }
 
-/// getSCEVStartAndStride - Compute the start and stride of this expression,
-/// returning false if the expression is not a start/stride pair, or true if it
-/// is.  The stride must be a loop invariant expression, but the start may be
-/// a mix of loop invariant and loop variant expressions.  The start cannot,
-/// however, contain an AddRec from a different loop, unless that loop is an
-/// outer loop of the current loop.
-static bool getSCEVStartAndStride(const SCEV *&SH, Loop *L, Loop *UseLoop,
-                                  const SCEV *&Start, const SCEV *&Stride,
-                                  ScalarEvolution *SE, DominatorTree *DT) {
-  const SCEV *TheAddRec = Start;   // Initialize to zero.
-
-  // If the outer level is an AddExpr, the operands are all start values except
-  // for a nested AddRecExpr.
-  if (const SCEVAddExpr *AE = dyn_cast<SCEVAddExpr>(SH)) {
-    for (unsigned i = 0, e = AE->getNumOperands(); i != e; ++i)
-      if (const SCEVAddRecExpr *AddRec =
-             dyn_cast<SCEVAddRecExpr>(AE->getOperand(i)))
-        TheAddRec = SE->getAddExpr(AddRec, TheAddRec);
-      else
-        Start = SE->getAddExpr(Start, AE->getOperand(i));
-  } else if (isa<SCEVAddRecExpr>(SH)) {
-    TheAddRec = SH;
-  } else {
-    return false;  // not analyzable.
-  }
-
-  // Break down TheAddRec into its component parts.
-  SmallVector<const SCEV *, 4> Subexprs;
-  CollectSubexprs(TheAddRec, Subexprs, *SE);
-
-  // Look for an addrec on the current loop among the parts.
-  const SCEV *AddRecStride = 0;
-  for (SmallVectorImpl<const SCEV *>::iterator I = Subexprs.begin(),
-       E = Subexprs.end(); I != E; ++I) {
-    const SCEV *S = *I;
-    if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
-      if (AR->getLoop() == L) {
-        *I = AR->getStart();
-        AddRecStride = AR->getStepRecurrence(*SE);
-        break;
-      }
-  }
-  if (!AddRecStride)
-    return false;
-
-  // Add up everything else into a start value (which may not be
-  // loop-invariant).
-  const SCEV *AddRecStart = SE->getAddExpr(Subexprs);
-
-  // Use getSCEVAtScope to attempt to simplify other loops out of
-  // the picture.
-  AddRecStart = SE->getSCEVAtScope(AddRecStart, UseLoop);
-
-  Start = SE->getAddExpr(Start, AddRecStart);
-
-  // If stride is an instruction, make sure it properly dominates the header.
-  // Otherwise we could end up with a use before def situation.
-  if (!isa<SCEVConstant>(AddRecStride)) {
-    BasicBlock *Header = L->getHeader();
-    if (!AddRecStride->properlyDominates(Header, DT))
-      return false;
+/// isInteresting - Test whether the given expression is "interesting" when
+/// used by the given expression, within the context of analyzing the
+/// given loop.
+static bool isInteresting(const SCEV *S, const Instruction *I, const Loop *L) {
+  // Anything loop-invariant is interesting.
+  if (!isa<SCEVUnknown>(S) && S->isLoopInvariant(L))
+    return true;
 
-    DEBUG(dbgs() << "[";
-          WriteAsOperand(dbgs(), L->getHeader(), /*PrintType=*/false);
-          dbgs() << "] Variable stride: " << *AddRecStride << "\n");
+  // An addrec is interesting if it's affine or if it has an interesting start.
+  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
+    // Keep things simple. Don't touch loop-variant strides.
+    if (AR->getLoop() == L && (AR->isAffine() || !L->contains(I)))
+        return true;
+    // Otherwise recurse to see if the start value is interesting.
+    return isInteresting(AR->getStart(), I, L);
   }
 
-  Stride = AddRecStride;
-  return true;
-}
-
-/// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
-/// and now we need to decide whether the user should use the preinc or post-inc
-/// value.  If this user should use the post-inc version of the IV, return true.
-///
-/// Choosing wrong here can break dominance properties (if we choose to use the
-/// post-inc value when we cannot) or it can end up adding extra live-ranges to
-/// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
-/// should use the post-inc value).
-static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
-                                       const Loop *L, DominatorTree *DT) {
-  // If the user is in the loop, use the preinc value.
-  if (L->contains(User)) return false;
-
-  BasicBlock *LatchBlock = L->getLoopLatch();
-  if (!LatchBlock)
+  // An add is interesting if any of its operands is.
+  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
+    for (SCEVAddExpr::op_iterator OI = Add->op_begin(), OE = Add->op_end();
+         OI != OE; ++OI)
+      if (isInteresting(*OI, I, L))
+        return true;
     return false;
+  }
 
-  // Ok, the user is outside of the loop.  If it is dominated by the latch
-  // block, use the post-inc value.
-  if (DT->dominates(LatchBlock, User->getParent()))
-    return true;
-
-  // There is one case we have to be careful of: PHI nodes.  These little guys
-  // can live in blocks that are not dominated by the latch block, but (since
-  // their uses occur in the predecessor block, not the block the PHI lives in)
-  // should still use the post-inc value.  Check for this case now.
-  PHINode *PN = dyn_cast<PHINode>(User);
-  if (!PN) return false;  // not a phi, not dominated by latch block.
-
-  // Look at all of the uses of IV by the PHI node.  If any use corresponds to
-  // a block that is not dominated by the latch block, give up and use the
-  // preincremented value.
-  unsigned NumUses = 0;
-  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
-    if (PN->getIncomingValue(i) == IV) {
-      ++NumUses;
-      if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
-        return false;
-    }
-
-  // Okay, all uses of IV by PN are in predecessor blocks that really are
-  // dominated by the latch block.  Use the post-incremented value.
-  return true;
+  // Nothing else is interesting here.
+  return false;
 }
 
 /// AddUsersIfInteresting - Inspect the specified instruction.  If it is a
@@ -196,16 +110,9 @@ bool IVUsers::AddUsersIfInteresting(Instruction *I) {
   const SCEV *ISE = SE->getSCEV(I);
   if (isa<SCEVCouldNotCompute>(ISE)) return false;
 
-  // Get the start and stride for this expression.
-  Loop *UseLoop = LI->getLoopFor(I->getParent());
-  const SCEV *Start = SE->getIntegerSCEV(0, ISE->getType());
-  const SCEV *Stride = Start;
-
-  if (!getSCEVStartAndStride(ISE, L, UseLoop, Start, Stride, SE, DT))
-    return false;  // Non-reducible symbolic expression, bail out.
-
-  // Keep things simple. Don't touch loop-variant strides.
-  if (!Stride->isLoopInvariant(L) && L->contains(I))
+  // If we've come to an uninteresting expression, stop the traversal and
+  // call this a user.
+  if (!isInteresting(ISE, I, L))
     return false;
 
   SmallPtrSet<Instruction *, 4> UniqueUsers;
@@ -241,27 +148,24 @@ bool IVUsers::AddUsersIfInteresting(Instruction *I) {
     }
 
     if (AddUserToIVUsers) {
-      // Okay, we found a user that we cannot reduce.  Analyze the instruction
-      // and decide what to do with it.  If we are a use inside of the loop, use
-      // the value before incrementation, otherwise use it after incrementation.
-      if (IVUseShouldUsePostIncValue(User, I, L, DT)) {
-        // The value used will be incremented by the stride more than we are
-        // expecting, so subtract this off.
-        const SCEV *NewStart = SE->getMinusSCEV(Start, Stride);
-        IVUses.push_back(new IVStrideUse(this, Stride, NewStart, User, I));
-        IVUses.back().setIsUseOfPostIncrementedValue(true);
-        DEBUG(dbgs() << "   USING POSTINC SCEV, START=" << *NewStart<< "\n");
-      } else {
-        IVUses.push_back(new IVStrideUse(this, Stride, Start, User, I));
-      }
+      // Okay, we found a user that we cannot reduce.
+      IVUses.push_back(new IVStrideUse(this, ISE, User, I));
+      IVStrideUse &NewUse = IVUses.back();
+      // Transform the expression into a normalized form.
+      NewUse.Expr =
+        TransformForPostIncUse(NormalizeAutodetect, NewUse.Expr,
+                               User, I,
+                               NewUse.PostIncLoops,
+                               *SE, *DT);
+      DEBUG(dbgs() << "   NORMALIZED TO: " << *NewUse.Expr << '\n');
     }
   }
   return true;
 }
 
-IVStrideUse &IVUsers::AddUser(const SCEV *Stride, const SCEV *Offset,
+IVStrideUse &IVUsers::AddUser(const SCEV *Expr,
                               Instruction *User, Value *Operand) {
-  IVUses.push_back(new IVStrideUse(this, Stride, Offset, User, Operand));
+  IVUses.push_back(new IVStrideUse(this, Expr, User, Operand));
   return IVUses.back();
 }
 
@@ -295,30 +199,10 @@ bool IVUsers::runOnLoop(Loop *l, LPPassManager &LPM) {
 /// getReplacementExpr - Return a SCEV expression which computes the
 /// value of the OperandValToReplace of the given IVStrideUse.
 const SCEV *IVUsers::getReplacementExpr(const IVStrideUse &U) const {
-  // Start with zero.
-  const SCEV *RetVal = SE->getIntegerSCEV(0, U.getStride()->getType());
-  // Create the basic add recurrence.
-  RetVal = SE->getAddRecExpr(RetVal, U.getStride(), L);
-  // Add the offset in a separate step, because it may be loop-variant.
-  RetVal = SE->getAddExpr(RetVal, U.getOffset());
-  // For uses of post-incremented values, add an extra stride to compute
-  // the actual replacement value.
-  if (U.isUseOfPostIncrementedValue())
-    RetVal = SE->getAddExpr(RetVal, U.getStride());
-  return RetVal;
-}
-
-/// getCanonicalExpr - Return a SCEV expression which computes the
-/// value of the SCEV of the given IVStrideUse, ignoring the 
-/// isUseOfPostIncrementedValue flag.
-const SCEV *IVUsers::getCanonicalExpr(const IVStrideUse &U) const {
-  // Start with zero.
-  const SCEV *RetVal = SE->getIntegerSCEV(0, U.getStride()->getType());
-  // Create the basic add recurrence.
-  RetVal = SE->getAddRecExpr(RetVal, U.getStride(), L);
-  // Add the offset in a separate step, because it may be loop-variant.
-  RetVal = SE->getAddExpr(RetVal, U.getOffset());
-  return RetVal;
+  PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(U.PostIncLoops);
+  return TransformForPostIncUse(Denormalize, U.getExpr(),
+                                U.getUser(), U.getOperandValToReplace(),
+                                Loops, *SE, *DT);
 }
 
 void IVUsers::print(raw_ostream &OS, const Module *M) const {
@@ -339,8 +223,13 @@ void IVUsers::print(raw_ostream &OS, const Module *M) const {
     WriteAsOperand(OS, UI->getOperandValToReplace(), false);
     OS << " = "
        << *getReplacementExpr(*UI);
-    if (UI->isUseOfPostIncrementedValue())
-      OS << " (post-inc)";
+    for (PostIncLoopSet::const_iterator
+         I = UI->PostIncLoops.begin(),
+         E = UI->PostIncLoops.end(); I != E; ++I) {
+      OS << " (post-inc with loop ";
+      WriteAsOperand(OS, (*I)->getHeader(), false);
+      OS << ")";
+    }
     OS << " in  ";
     UI->getUser()->print(OS, &Annotator);
     OS << '\n';
@@ -356,6 +245,39 @@ void IVUsers::releaseMemory() {
   IVUses.clear();
 }
 
+static const SCEVAddRecExpr *findAddRecForLoop(const SCEV *S, const Loop *L) {
+  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
+    if (AR->getLoop() == L)
+      return AR;
+    return findAddRecForLoop(AR->getStart(), L);
+  }
+
+  if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
+    for (SCEVAddExpr::op_iterator I = Add->op_begin(), E = Add->op_end();
+         I != E; ++I)
+      if (const SCEVAddRecExpr *AR = findAddRecForLoop(*I, L))
+        return AR;
+    return 0;
+  }
+
+  return 0;
+}
+
+const SCEV *IVStrideUse::getStride(const Loop *L) const {
+  if (const SCEVAddRecExpr *AR = findAddRecForLoop(getExpr(), L))
+    return AR->getStepRecurrence(*Parent->SE);
+  return 0;
+}
+
+void IVStrideUse::transformToPostInc(const Loop *L) {
+  PostIncLoopSet Loops;
+  Loops.insert(L);
+  Expr = TransformForPostIncUse(Normalize, Expr,
+                                getUser(), getOperandValToReplace(),
+                                Loops, *Parent->SE, *Parent->DT);
+  PostIncLoops.insert(L);
+}
+
 void IVStrideUse::deleted() {
   // Remove this user from the list.
   Parent->IVUses.erase(this);
diff --git a/lib/Analysis/ScalarEvolutionExpander.cpp b/lib/Analysis/ScalarEvolutionExpander.cpp
index 2e18ceac52..dd8ab431f3 100644
--- a/lib/Analysis/ScalarEvolutionExpander.cpp
+++ b/lib/Analysis/ScalarEvolutionExpander.cpp
@@ -966,9 +966,12 @@ Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
   // Determine a normalized form of this expression, which is the expression
   // before any post-inc adjustment is made.
   const SCEVAddRecExpr *Normalized = S;
-  if (L == PostIncLoop) {
-    const SCEV *Step = S->getStepRecurrence(SE);
-    Normalized = cast<SCEVAddRecExpr>(SE.getMinusSCEV(S, Step));
+  if (PostIncLoops.count(L)) {
+    PostIncLoopSet Loops;
+    Loops.insert(L);
+    Normalized =
+      cast<SCEVAddRecExpr>(TransformForPostIncUse(Normalize, S, 0, 0,
+                                                  Loops, SE, *SE.DT));
   }
 
   // Strip off any non-loop-dominating component from the addrec start.
@@ -1002,7 +1005,7 @@ Value *SCEVExpander::expandAddRecExprLiterally(const SCEVAddRecExpr *S) {
 
   // Accommodate post-inc mode, if necessary.
   Value *Result;
-  if (L != PostIncLoop)
+  if (!PostIncLoops.count(L))
     Result = PN;
   else {
     // In PostInc mode, use the post-incremented value.
@@ -1274,7 +1277,7 @@ Value *SCEVExpander::expand(const SCEV *S) {
       // If the SCEV is computable at this level, insert it into the header
       // after the PHIs (and after any other instructions that we've inserted
       // there) so that it is guaranteed to dominate any user inside the loop.
-      if (L && S->hasComputableLoopEvolution(L) && L != PostIncLoop)
+      if (L && S->hasComputableLoopEvolution(L) && !PostIncLoops.count(L))
         InsertPt = L->getHeader()->getFirstNonPHI();
       while (isInsertedInstruction(InsertPt) || isa<DbgInfoIntrinsic>(InsertPt))
         InsertPt = llvm::next(BasicBlock::iterator(InsertPt));
@@ -1296,7 +1299,7 @@ Value *SCEVExpander::expand(const SCEV *S) {
   Value *V = visit(S);
 
   // Remember the expanded value for this SCEV at this location.
-  if (!PostIncLoop)
+  if (PostIncLoops.empty())
     InsertedExpressions[std::make_pair(S, InsertPt)] = V;
 
   restoreInsertPoint(SaveInsertBB, SaveInsertPt);
@@ -1304,7 +1307,7 @@ Value *SCEVExpander::expand(const SCEV *S) {
 }
 
 void SCEVExpander::rememberInstruction(Value *I) {
-  if (!PostIncLoop)
+  if (PostIncLoops.empty())
     InsertedValues.insert(I);
 
   // If we just claimed an existing instruction and that instruction had
diff --git a/lib/Analysis/ScalarEvolutionNormalization.cpp b/lib/Analysis/ScalarEvolutionNormalization.cpp
new file mode 100644
index 0000000000..75c381d5ef
--- /dev/null
+++ b/lib/Analysis/ScalarEvolutionNormalization.cpp
@@ -0,0 +1,150 @@
+//===- ScalarEvolutionNormalization.cpp - See below -------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements utilities for working with "normalized" expressions.
+// See the comments at the top of ScalarEvolutionNormalization.h for details.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Analysis/Dominators.h"
+#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/Analysis/ScalarEvolutionNormalization.h"
+using namespace llvm;
+
+/// IVUseShouldUsePostIncValue - We have discovered a "User" of an IV expression
+/// and now we need to decide whether the user should use the preinc or post-inc
+/// value.  If this user should use the post-inc version of the IV, return true.
+///
+/// Choosing wrong here can break dominance properties (if we choose to use the
+/// post-inc value when we cannot) or it can end up adding extra live-ranges to
+/// the loop, resulting in reg-reg copies (if we use the pre-inc value when we
+/// should use the post-inc value).
+static bool IVUseShouldUsePostIncValue(Instruction *User, Instruction *IV,
+                                       const Loop *L, DominatorTree *DT) {
+  // If the user is in the loop, use the preinc value.
+  if (L->contains(User)) return false;
+
+  BasicBlock *LatchBlock = L->getLoopLatch();
+  if (!LatchBlock)
+    return false;
+
+  // Ok, the user is outside of the loop.  If it is dominated by the latch
+  // block, use the post-inc value.
+  if (DT->dominates(LatchBlock, User->getParent()))
+    return true;
+
+  // There is one case we have to be careful of: PHI nodes.  These little guys
+  // can live in blocks that are not dominated by the latch block, but (since
+  // their uses occur in the predecessor block, not the block the PHI lives in)
+  // should still use the post-inc value.  Check for this case now.
+  PHINode *PN = dyn_cast<PHINode>(User);
+  if (!PN) return false;  // not a phi, not dominated by latch block.
+
+  // Look at all of the uses of IV by the PHI node.  If any use corresponds to
+  // a block that is not dominated by the latch block, give up and use the
+  // preincremented value.
+  unsigned NumUses = 0;
+  for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
+    if (PN->getIncomingValue(i) == IV) {
+      ++NumUses;
+      if (!DT->dominates(LatchBlock, PN->getIncomingBlock(i)))
+        return false;
+    }
+
+  // Okay, all uses of IV by PN are in predecessor blocks that really are
+  // dominated by the latch block.  Use the post-incremented value.
+  return true;
+}
+
+const SCEV *llvm::TransformForPostIncUse(TransformKind Kind,
+                                         const SCEV *S,
+                                         Instruction *User,
+                                         Value *OperandValToReplace,
+                                         PostIncLoopSet &Loops,
+                                         ScalarEvolution &SE,
+                                         DominatorTree &DT) {
+  if (isa<SCEVConstant>(S) || isa<SCEVUnknown>(S))
+    return S;
+  if (const SCEVCastExpr *X = dyn_cast<SCEVCastExpr>(S)) {
+    const SCEV *O = X->getOperand();
+    const SCEV *N = TransformForPostIncUse(Kind, O, User, OperandValToReplace,
+                                           Loops, SE, DT);
+    if (O != N)
+      switch (S->getSCEVType()) {
+      case scZeroExtend: return SE.getZeroExtendExpr(N, S->getType());
+      case scSignExtend: return SE.getSignExtendExpr(N, S->getType());
+      case scTruncate: return SE.getTruncateExpr(N, S->getType());
+      default: llvm_unreachable("Unexpected SCEVCastExpr kind!");
+      }
+    return S;
+  }
+  if (const SCEVNAryExpr *X = dyn_cast<SCEVNAryExpr>(S)) {
+    SmallVector<const SCEV *, 8> Operands;
+    bool Changed = false;
+    for (SCEVNAryExpr::op_iterator I = X->op_begin(), E = X->op_end();
+         I != E; ++I) {
+      const SCEV *O = *I;
+      const SCEV *N = TransformForPostIncUse(Kind, O, User, OperandValToReplace,
+                                             Loops, SE, DT);
+      Changed |= N != O;
+      Operands.push_back(N);
+    }
+    if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
+      // An addrec. This is the interesting part.
+      const Loop *L = AR->getLoop();
+      const SCEV *Result = SE.getAddRecExpr(Operands, L);
+      switch (Kind) {
+      default: llvm_unreachable("Unexpected transform name!");
+      case NormalizeAutodetect:
+        if (Instruction *OI = dyn_cast<Instruction>(OperandValToReplace))
+          if (IVUseShouldUsePostIncValue(User, OI, L, &DT)) {
+            Result = SE.getMinusSCEV(Result, AR->getStepRecurrence(SE));
+            Loops.insert(L);
+          }
+        break;
+      case Normalize:
+        if (Loops.count(L))
+          Result = SE.getMinusSCEV(Result, AR->getStepRecurrence(SE));
+        break;
+      case Denormalize:
+        if (Loops.count(L)) {
+          const SCEV *TransformedStep =
+            TransformForPostIncUse(Kind, AR->getStepRecurrence(SE),
+                                   User, OperandValToReplace, Loops, SE, DT);
+          Result = SE.getAddExpr(Result, TransformedStep);
+        }
+        break;
+      }
+      return Result;
+    }
+    if (Changed)
+      switch (S->getSCEVType()) {
+      case scAddExpr: return SE.getAddExpr(Operands);
+      case scMulExpr: return SE.getMulExpr(Operands);
+      case scSMaxExpr: return SE.getSMaxExpr(Operands);
+      case scUMaxExpr: return SE.getUMaxExpr(Operands);
+      default: llvm_unreachable("Unexpected SCEVNAryExpr kind!");
+      }
+    return S;
+  }
+  if (const SCEVUDivExpr *X = dyn_cast<SCEVUDivExpr>(S)) {
+    const SCEV *LO = X->getLHS();
+    const SCEV *RO = X->getRHS();
+    const SCEV *LN = TransformForPostIncUse(Kind, LO, User, OperandValToReplace,
+                                            Loops, SE, DT);
+    const SCEV *RN = TransformForPostIncUse(Kind, RO, User, OperandValToReplace,
+                                            Loops, SE, DT);
+    if (LO != LN || RO != RN)
+      return SE.getUDivExpr(LN, RN);
+    return S;
+  }
+  llvm_unreachable("Unexpected SCEV kind!");
+  return 0;
+}
diff --git a/lib/Transforms/Scalar/IndVarSimplify.cpp b/lib/Transforms/Scalar/IndVarSimplify.cpp
index 6605666e45..1a58b6644f 100644
--- a/lib/Transforms/Scalar/IndVarSimplify.cpp
+++ b/lib/Transforms/Scalar/IndVarSimplify.cpp
@@ -454,6 +454,46 @@ bool IndVarSimplify::runOnLoop(Loop *L, LPPassManager &LPM) {
   return Changed;
 }
 
+// FIXME: It is an extremely bad idea to indvar substitute anything more
+// complex than affine induction variables.  Doing so will put expensive
+// polynomial evaluations inside of the loop, and the str reduction pass
+// currently can only reduce affine polynomials.  For now just disable
+// indvar subst on anything more complex than an affine addrec, unless
+// it can be expanded to a trivial value.
+static bool isSafe(const SCEV *S, const Loop *L) {
+  // Loop-invariant values are safe.
+  if (S->isLoopInvariant(L)) return true;
+
+  // Affine addrecs are safe. Non-affine are not, because LSR doesn't know how
+  // to transform them into efficient code.
+  if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S))
+    return AR->isAffine();
+
+  // An add is safe it all its operands are safe.
+  if (const SCEVCommutativeExpr *Commutative = dyn_cast<SCEVCommutativeExpr>(S)) {
+    for (SCEVCommutativeExpr::op_iterator I = Commutative->op_begin(),
+         E = Commutative->op_end(); I != E; ++I)
+      if (!isSafe(*I, L)) return false;
+    return true;
+  }
+  
+  // A cast is safe if its operand is.
+  if (const SCEVCastExpr *C = dyn_cast<SCEVCastExpr>(S))
+    return isSafe(C->getOperand(), L);
+
+  // A udiv is safe if its operands are.
+  if (const SCEVUDivExpr *UD = dyn_cast<SCEVUDivExpr>(S))
+    return isSafe(UD->getLHS(), L) &&
+           isSafe(UD->getRHS(), L);
+
+  // SCEVUnknown is always safe.
+  if (isa<SCEVUnknown>(S))
+    return true;
+
+  // Nothing else is safe.
+  return false;
+}
+
 void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
   SmallVector<WeakVH, 16> DeadInsts;
 
@@ -465,7 +505,6 @@ void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
   // the need for the code evaluation methods to insert induction variables
   // of different sizes.
   for (IVUsers::iterator UI = IU->begin(), E = IU->end(); UI != E; ++UI) {
-    const SCEV *Stride = UI->getStride();
     Value *Op = UI->getOperandValToReplace();
     const Type *UseTy = Op->getType();
     Instruction *User = UI->getUser();
@@ -486,7 +525,7 @@ void IndVarSimplify::RewriteIVExpressions(Loop *L, SCEVExpander &Rewriter) {
     // currently can only reduce affine polynomials.  For now just disable
     // indvar subst on anything more complex than an affine addrec, unless
     // it can be expanded to a trivial value.
-    if (!AR->isLoopInvariant(L) && !Stride->isLoopInvariant(L))
+    if (!isSafe(AR, L))
       continue;
 
     // Determine the insertion point for this user. By default, insert
diff --git a/lib/Transforms/Scalar/LoopStrengthReduce.cpp b/lib/Transforms/Scalar/LoopStrengthReduce.cpp
index 625a75d6cc..631092b326 100644
--- a/lib/Transforms/Scalar/LoopStrengthReduce.cpp
+++ b/lib/Transforms/Scalar/LoopStrengthReduce.cpp
@@ -781,10 +781,10 @@ struct LSRFixup {
   /// will be replaced.
   Value *OperandValToReplace;
 
-  /// PostIncLoop - If this user is to use the post-incremented value of an
+  /// PostIncLoops - If this user is to use the post-incremented value of an
   /// induction variable, this variable is non-null and holds the loop
   /// associated with the induction variable.
-  const Loop *PostIncLoop;
+  PostIncLoopSet PostIncLoops;
 
   /// LUIdx - The index of the LSRUse describing the expression which
   /// this fixup needs, minus an offset (below).
@@ -795,6 +795,8 @@ struct LSRFixup {
   /// offsets, for example in an unrolled loop.
   int64_t Offset;
 
+  bool isUseFullyOutsideLoop(const Loop *L) const;
+
   LSRFixup();
 
   void print(raw_ostream &OS) const;
@@ -804,9 +806,24 @@ struct LSRFixup {
 }
 
 LSRFixup::LSRFixup()
-  : UserInst(0), OperandValToReplace(0), PostIncLoop(0),
+  : UserInst(0), OperandValToReplace(0),
     LUIdx(~size_t(0)), Offset(0) {}
 
+/// isUseFullyOutsideLoop - Test whether this fixup always uses its
+/// value outside of the given loop.
+bool LSRFixup::isUseFullyOutsideLoop(const Loop *L) const {
+  // PHI nodes use their value in their incoming blocks.
+  if (const PHINode *PN = dyn_cast<PHINode>(UserInst)) {
+    for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
+      if (PN->getIncomingValue(i) == OperandValToReplace &&
+          L->contains(PN->getIncomingBlock(i)))
+        return false;
+    return true;
+  }
+
+  return !L->contains(UserInst);
+}
+
 void LSRFixup::print(raw_ostream &OS) const {
   OS << "UserInst=";
   // Store is common and interesting enough to be worth special-casing.
@@ -821,9 +838,10 @@ void LSRFixup::print(raw_ostream &OS) const {
   OS << ", OperandValToReplace=";
   WriteAsOperand(OS, OperandValToReplace, /*PrintType=*/false);
 
-  if (PostIncLoop) {
+  for (PostIncLoopSet::const_iterator I = PostIncLoops.begin(),
+       E = PostIncLoops.end(); I != E; ++I) {
     OS << ", PostIncLoop=";
-    WriteAsOperand(OS, PostIncLoop->getHeader(), /*PrintType=*/false);
+    WriteAsOperand(OS, (*I)->getHeader(), /*PrintType=*/false);
   }
 
   if (LUIdx != ~size_t(0))
@@ -1545,8 +1563,9 @@ LSRInstance::OptimizeLoopTermCond() {
             !DT.properlyDominates(UI->getUser()->getParent(), ExitingBlock)) {
           // Conservatively assume there may be reuse if the quotient of their
           // strides could be a legal scale.
-          const SCEV *A = CondUse->getStride();
-          const SCEV *B = UI->getStride();
+          const SCEV *A = CondUse->getStride(L);
+          const SCEV *B = UI->getStride(L);
+          if (!A || !B) continue;
           if (SE.getTypeSizeInBits(A->getType()) !=
               SE.getTypeSizeInBits(B->getType())) {
             if (SE.getTypeSizeInBits(A->getType()) >
@@ -1598,7 +1617,7 @@ LSRInstance::OptimizeLoopTermCond() {
         ExitingBlock->getInstList().insert(TermBr, Cond);
 
         // Clone the IVUse, as the old use still exists!
-        CondUse = &IU.AddUser(CondUse->getStride(), CondUse->getOffset(),
+        CondUse = &IU.AddUser(CondUse->getExpr(),
                               Cond, CondUse->getOperandValToReplace());
         TermBr->replaceUsesOfWith(OldCond, Cond);
       }
@@ -1607,9 +1626,7 @@ LSRInstance::OptimizeLoopTermCond() {
     // If we get to here, we know that we can transform the setcc instruction to
     // use the post-incremented version of the IV, allowing us to coalesce the
     // live ranges for the IV correctly.
-    CondUse->setOffset(SE.getMinusSCEV(CondUse->getOffset(),
-                                       CondUse->getStride()));
-    CondUse->setIsUseOfPostIncrementedValue(true);
+    CondUse->transformToPostInc(L);
     Changed = true;
 
     PostIncs.insert(Cond);
@@ -1717,19 +1734,24 @@ void LSRInstance::CollectInterestingTypesAndFactors() {
   SmallSetVector<const SCEV *, 4> Strides;
 
   // Collect interesting types and strides.
+  SmallVector<const SCEV *, 4> Worklist;
   for (IVUsers::const_iterator UI = IU.begin(), E = IU.end(); UI != E; ++UI) {
-    const SCEV *Stride = UI->getStride();
+    const SCEV *Expr = UI->getExpr();
 
     // Collect interesting types.
-    Types.insert(SE.getEffectiveSCEVType(Stride->getType()));
-
-    // Add the stride for this loop.
-    Strides.insert(Stride);
-
-    // Add strides for other mentioned loops.
-    for (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(UI->getOffset());
-         AR; AR = dyn_cast<SCEVAddRecExpr>(AR->getStart()))
-      Strides.insert(AR->getStepRecurrence(SE));
+    Types.insert(SE.getEffectiveSCEVType(Expr->getType()));
+
+    // Add strides for mentioned loops.
+    Worklist.push_back(Expr);
+    do {
+      const SCEV *S = Worklist.pop_back_val();
+      if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(S)) {
+        Strides.insert(AR->getStepRecurrence(SE));
+        Worklist.push_back(AR->getStart());
+      } else if (const SCEVAddExpr *Add = dyn_cast<SCEVAddExpr>(S)) {
+        Worklist.insert(Worklist.end(), Add->op_begin(), Add->op_end());
+      }
+    } while (!Worklist.empty());
   }
 
   // Compute interesting factors from the set of interesting strides.
@@ -1776,8 +1798,7 @@ void LSRInstance::CollectFixupsAndInitialFormulae() {
     LSRFixup &LF = getNewFixup();
     LF.UserInst = UI->getUser();
     LF.OperandValToReplace = UI->getOperandValToReplace();
-    if (UI->isUseOfPostIncrementedValue())
-      LF.PostIncLoop = L;
+    LF.PostIncLoops = UI->getPostIncLoops();
 
     LSRUse::KindType Kind = LSRUse::Basic;
     const Type *AccessTy = 0;
@@ -1786,7 +1807,7 @@ void LSRInstance::CollectFixupsAndInitialFormulae() {
       AccessTy = getAccessType(LF.UserInst);
     }
 
-    const SCEV *S = IU.getCanonicalExpr(*UI);
+    const SCEV *S = UI->getExpr();
 
     // Equality (== and !=) ICmps are special. We can rewrite (i == N) as
     // (N - i == 0), and this allows (N - i) to be the expression that we work
@@ -1824,7 +1845,7 @@ void LSRInstance::CollectFixupsAndInitialFormulae() {
     LF.LUIdx = P.first;
     LF.Offset = P.second;
     LSRUse &LU = Uses[LF.LUIdx];
-    LU.AllFixupsOutsideLoop &= !L->contains(LF.UserInst);
+    LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
 
     // If this is the first use of this LSRUse, give it a formula.
     if (LU.Formulae.empty()) {
@@ -1936,7 +1957,7 @@ LSRInstance::CollectLoopInvariantFixupsAndFormulae() {
         LF.LUIdx = P.first;
         LF.Offset = P.second;
         LSRUse &LU = Uses[LF.LUIdx];
-        LU.AllFixupsOutsideLoop &= L->contains(LF.UserInst);
+        LU.AllFixupsOutsideLoop &= LF.isUseFullyOutsideLoop(L);
         InsertSupplementalFormula(U, LU, LF.LUIdx);
         CountRegisters(LU.Formulae.back(), Uses.size() - 1);
         break;
@@ -2783,8 +2804,8 @@ Value *LSRInstance::Expand(const LSRFixup &LF,
                            SmallVectorImpl<WeakVH> &DeadInsts) const {
   const LSRUse &LU = Uses[LF.LUIdx];
 
-  // Then, collect some instructions which we will remain dominated by when
-  // expanding the replacement. These must be dominated by any operands that
+  // Then, collect some instructions which must be dominated by the
+  // expanding replacement. These must be dominated by any operands that
   // will be required in the expansion.
   SmallVector<Instruction *, 4> Inputs;
   if (Instruction *I = dyn_cast<Instruction>(LF.OperandValToReplace))
@@ -2793,8 +2814,8 @@ Value *LSRInstance::Expand(const LSRFixup &LF,
     if (Instruction *I =
           dyn_cast<Instruction>(cast<ICmpInst>(LF.UserInst)->getOperand(1)))
       Inputs.push_back(I);
-  if (LF.PostIncLoop) {
-    if (!L->contains(LF.UserInst))
+  if (LF.PostIncLoops.count(L)) {
+    if (LF.isUseFullyOutsideLoop(L))
       Inputs.push_back(L->getLoopLatch()->getTerminator());
     else
       Inputs.push_back(IVIncInsertPos);
@@ -2831,7 +2852,7 @@ Value *LSRInstance::Expand(const LSRFixup &LF,
 
   // Inform the Rewriter if we have a post-increment use, so that it can
   // perform an advantageous expansion.
-  Rewriter.setPostInc(LF.PostIncLoop);
+  Rewriter.setPostInc(LF.PostIncLoops);
 
   // This is the type that the user actually needs.
   const Type *OpTy = LF.OperandValToReplace->getType();
@@ -2855,24 +2876,11 @@ Value *LSRInstance::Expand(const LSRFixup &LF,
     const SCEV *Reg = *I;
     assert(!Reg->isZero() && "Zero allocated in a base register!");
 
-    // If we're expanding for a post-inc user for the add-rec's loop, make the
-    // post-inc adjustment.
-    const SCEV *Start = Reg;
-    while (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Start)) {
-      if (AR->getLoop() == LF.PostIncLoop) {
-        Reg = SE.getAddExpr(Reg, AR->getStepRecurrence(SE));
-        // If the user is inside the loop, insert the code after the increment
-        // so that it is dominated by its operand. If the original insert point
-        // was already dominated by the increment, keep it, because there may
-        // be loop-variant operands that need to be respected also.
-        if (L->contains(LF.UserInst) && !DT.dominates(IVIncInsertPos, IP)) {
-          IP = IVIncInsertPos;
-          while (isa<DbgInfoIntrinsic>(IP)) ++IP;
-        }
-        break;
-      }
-      Start = AR->getStart();
-    }
+    // If we're expanding for a post-inc user, make the post-inc adjustment.
+    PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
+    Reg = TransformForPostIncUse(Denormalize, Reg,
+                                 LF.UserInst, LF.OperandValToReplace,
+                                 Loops, SE, DT);
 
     Ops.push_back(SE.getUnknown(Rewriter.expandCodeFor(Reg, 0, IP)));
   }
@@ -2889,11 +2897,11 @@ Value *LSRInstance::Expand(const LSRFixup &LF,
   if (F.AM.Scale != 0) {
     const SCEV *ScaledS = F.ScaledReg;
 
-    // If we're expanding for a post-inc user for the add-rec's loop, make the
-    // post-inc adjustment.
-    if (const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(ScaledS))
-      if (AR->getLoop() == LF.PostIncLoop)
-        ScaledS = SE.getAddExpr(ScaledS, AR->getStepRecurrence(SE));
+    // If we're expanding for a post-inc user, make the post-inc adjustment.
+    PostIncLoopSet &Loops = const_cast<PostIncLoopSet &>(LF.PostIncLoops);
+    ScaledS = TransformForPostIncUse(Denormalize, ScaledS,
+                                     LF.UserInst, LF.OperandValToReplace,
+                                     Loops, SE, DT);
 
     if (LU.Kind == LSRUse::ICmpZero) {
       // An interesting way of "folding" with an icmp is to use a negated
@@ -2954,7 +2962,7 @@ Value *LSRInstance::Expand(const LSRFixup &LF,
   Value *FullV = Rewriter.expandCodeFor(FullS, Ty, IP);
 
   // We're done expanding now, so reset the rewriter.
-  Rewriter.setPostInc(0);
+  Rewriter.clearPostInc();
 
   // An ICmpZero Formula represents an ICmp which we're handling as a
   // comparison against zero. Now that we've expanded an expression for that