Fix SCEVExpander: don't try to expand quadratic recurrences outside a loop.

Partial fix for PR17459: wrong code at -O3 on x86_64-linux-gnu
(affecting trunk and 3.3)

When SCEV expands a recurrence outside of a loop it attempts to scale
by the stride of the recurrence. Chained recurrences don't work that
way. We could compute binomial coefficients, but would hve to
guarantee that the chained AddRec's are in a perfectly reduced form.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@193438 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 19 insertions, 2 deletions

diff --git a/lib/Transforms/Scalar/LoopStrengthReduce.cpp b/lib/Transforms/Scalar/LoopStrengthReduce.cpp
index 14cb979392..eff5268c44 100644
--- a/lib/Transforms/Scalar/LoopStrengthReduce.cpp
+++ b/lib/Transforms/Scalar/LoopStrengthReduce.cpp
@@ -1170,6 +1170,13 @@ public:
   /// may be used.
   bool AllFixupsOutsideLoop;
 
+  /// RigidFormula is set to true to guarantee that this use will be associated
+  /// with a single formula--the one that initially matched. Some SCEV
+  /// expressions cannot be expanded. This allows LSR to consider the registers
+  /// used by those expressions without the need to expand them later after
+  /// changing the formula.
+  bool RigidFormula;
+
   /// WidestFixupType - This records the widest use type for any fixup using
   /// this LSRUse. FindUseWithSimilarFormula can't consider uses with different
   /// max fixup widths to be equivalent, because the narrower one may be relying
@@ -1188,6 +1195,7 @@ public:
                                       MinOffset(INT64_MAX),
                                       MaxOffset(INT64_MIN),
                                       AllFixupsOutsideLoop(true),
+                                      RigidFormula(false),
                                       WidestFixupType(0) {}
 
   bool HasFormulaWithSameRegs(const Formula &F) const;
@@ -1214,6 +1222,9 @@ bool LSRUse::HasFormulaWithSameRegs(const Formula &F) const {
 /// InsertFormula - If the given formula has not yet been inserted, add it to
 /// the list, and return true. Return false otherwise.
 bool LSRUse::InsertFormula(const Formula &F) {
+  if (!Formulae.empty() && RigidFormula)
+    return false;
+
   SmallVector<const SCEV *, 4> Key = F.BaseRegs;
   if (F.ScaledReg) Key.push_back(F.ScaledReg);
   // Unstable sort by host order ok, because this is only used for uniquifying.
@@ -1433,7 +1444,7 @@ static unsigned getScalingFactorCost(const TargetTransformInfo &TTI,
   }
   case LSRUse::ICmpZero:
     // ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg.
-    // Therefore, return 0 in case F.Scale == -1. 
+    // Therefore, return 0 in case F.Scale == -1.
     return F.Scale != -1;
 
   case LSRUse::Basic:
@@ -2943,7 +2954,7 @@ void LSRInstance::CollectFixupsAndInitialFormulae() {
 
         // x == y  -->  x - y == 0
         const SCEV *N = SE.getSCEV(NV);
-        if (SE.isLoopInvariant(N, L) && isSafeToExpand(N)) {
+        if (SE.isLoopInvariant(N, L) && isSafeToExpand(N, SE)) {
           // S is normalized, so normalize N before folding it into S
           // to keep the result normalized.
           N = TransformForPostIncUse(Normalize, N, CI, 0,
@@ -2986,6 +2997,10 @@ void LSRInstance::CollectFixupsAndInitialFormulae() {
 /// and loop-computable portions.
 void
 LSRInstance::InsertInitialFormula(const SCEV *S, LSRUse &LU, size_t LUIdx) {
+  // Mark uses whose expressions cannot be expanded.
+  if (!isSafeToExpand(S, SE))
+    LU.RigidFormula = true;
+
   Formula F;
   F.InitialMatch(S, L, SE);
   bool Inserted = InsertFormula(LU, LUIdx, F);
@@ -4353,6 +4368,8 @@ Value *LSRInstance::Expand(const LSRFixup &LF,
                            SCEVExpander &Rewriter,
                            SmallVectorImpl<WeakVH> &DeadInsts) const {
   const LSRUse &LU = Uses[LF.LUIdx];
+  if (LU.RigidFormula)
+    return LF.OperandValToReplace;
 
   // Determine an input position which will be dominated by the operands and
   // which will dominate the result.