Allow the construction of SCEVs with SCEVCouldNotCompute operands, by

implementing folding. Fixes PR2857.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@57049 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Analysis/ScalarEvolution.cpp b/lib/Analysis/ScalarEvolution.cpp
index 5125804..473057c 100644 (file)
--- a/lib/Analysis/ScalarEvolution.cpp
+++ b/lib/Analysis/ScalarEvolution.cpp
@@ -676,6 +676,9 @@ SCEVHandle ScalarEvolution::getTruncateExpr(const SCEVHandle &Op, const Type *Ty
       return getAddRecExpr(Operands, AddRec->getLoop());
   }
 
+  if (isa<SCEVCouldNotCompute>(Op))
+    return new SCEVCouldNotCompute();
+
   SCEVTruncateExpr *&Result = (*SCEVTruncates)[std::make_pair(Op, Ty)];
   if (Result == 0) Result = new SCEVTruncateExpr(Op, Ty);
   return Result;
@@ -691,6 +694,9 @@ SCEVHandle ScalarEvolution::getZeroExtendExpr(const SCEVHandle &Op, const Type *
   // operands (often constants).  This would allow analysis of something like
   // this:  for (unsigned char X = 0; X < 100; ++X) { int Y = X; }
 
+  if (isa<SCEVCouldNotCompute>(Op))
+    return new SCEVCouldNotCompute();
+
   SCEVZeroExtendExpr *&Result = (*SCEVZeroExtends)[std::make_pair(Op, Ty)];
   if (Result == 0) Result = new SCEVZeroExtendExpr(Op, Ty);
   return Result;
@@ -706,6 +712,9 @@ SCEVHandle ScalarEvolution::getSignExtendExpr(const SCEVHandle &Op, const Type *
   // operands (often constants).  This would allow analysis of something like
   // this:  for (signed char X = 0; X < 100; ++X) { int Y = X; }
 
+  if (isa<SCEVCouldNotCompute>(Op))
+    return new SCEVCouldNotCompute();
+
   SCEVSignExtendExpr *&Result = (*SCEVSignExtends)[std::make_pair(Op, Ty)];
   if (Result == 0) Result = new SCEVSignExtendExpr(Op, Ty);
   return Result;
@@ -734,6 +743,10 @@ SCEVHandle ScalarEvolution::getAddExpr(std::vector<SCEVHandle> &Ops) {
   // Sort by complexity, this groups all similar expression types together.
   GroupByComplexity(Ops);
 
+  // Could not compute plus anything equals could not compute.
+  if (isa<SCEVCouldNotCompute>(Ops.back()))
+    return new SCEVCouldNotCompute();
+
   // If there are any constants, fold them together.
   unsigned Idx = 0;
   if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
@@ -959,6 +972,21 @@ SCEVHandle ScalarEvolution::getMulExpr(std::vector<SCEVHandle> &Ops) {
   // Sort by complexity, this groups all similar expression types together.
   GroupByComplexity(Ops);
 
+  if (isa<SCEVCouldNotCompute>(Ops.back())) {
+    // CNC * 0 = 0
+    for (unsigned i = 0, e = Ops.size() - 1; i != e; ++i) {
+      if (Ops[i]->getSCEVType() != scConstant)
+        break;
+
+      SCEVConstant *SC = cast<SCEVConstant>(Ops[i]);
+      if (SC->getValue()->isMinValue(false))
+        return SC;
+    }
+
+    // Otherwise, we can't compute it.
+    return new SCEVCouldNotCompute();
+  }
+
   // If there are any constants, fold them together.
   unsigned Idx = 0;
   if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
@@ -1124,6 +1152,9 @@ SCEVHandle ScalarEvolution::getUDivExpr(const SCEVHandle &LHS, const SCEVHandle
 
   // FIXME: implement folding of (X*4)/4 when we know X*4 doesn't overflow.
 
+  if (isa<SCEVCouldNotCompute>(LHS) || isa<SCEVCouldNotCompute>(RHS))
+    return new SCEVCouldNotCompute();
+
   SCEVUDivExpr *&Result = (*SCEVUDivs)[std::make_pair(LHS, RHS)];
   if (Result == 0) Result = new SCEVUDivExpr(LHS, RHS);
   return Result;
@@ -1171,6 +1202,12 @@ SCEVHandle ScalarEvolution::getAddRecExpr(std::vector<SCEVHandle> &Operands,
     }
   }
 
+  // Refuse to build an AddRec out of SCEVCouldNotCompute.
+  for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
+    if (isa<SCEVCouldNotCompute>(Operands[i]))
+      return new SCEVCouldNotCompute();
+  }
+
   SCEVAddRecExpr *&Result =
     (*SCEVAddRecExprs)[std::make_pair(L, std::vector<SCEV*>(Operands.begin(),
                                                             Operands.end()))];
@@ -1193,6 +1230,21 @@ SCEVHandle ScalarEvolution::getSMaxExpr(std::vector<SCEVHandle> Ops) {
   // Sort by complexity, this groups all similar expression types together.
   GroupByComplexity(Ops);
 
+  if (isa<SCEVCouldNotCompute>(Ops.back())) {
+    // CNC smax +inf = +inf.
+    for (unsigned i = 0, e = Ops.size() - 1; i != e; ++i) {
+      if (Ops[i]->getSCEVType() != scConstant)
+        break;
+
+      SCEVConstant *SC = cast<SCEVConstant>(Ops[i]);
+      if (SC->getValue()->isMaxValue(true))
+        return SC;
+    }
+
+    // Otherwise, we can't compute it.
+    return new SCEVCouldNotCompute();
+  }
+
   // If there are any constants, fold them together.
   unsigned Idx = 0;
   if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {
@@ -1273,6 +1325,21 @@ SCEVHandle ScalarEvolution::getUMaxExpr(std::vector<SCEVHandle> Ops) {
   // Sort by complexity, this groups all similar expression types together.
   GroupByComplexity(Ops);
 
+  if (isa<SCEVCouldNotCompute>(Ops[0])) {
+    // CNC umax inf = inf.
+    for (unsigned i = 0, e = Ops.size() - 1; i != e; ++i) {
+      if (Ops[i]->getSCEVType() != scConstant)
+        break;
+
+      SCEVConstant *SC = cast<SCEVConstant>(Ops[i]);
+      if (SC->getValue()->isMaxValue(false))
+        return SC;
+    }
+
+    // Otherwise, we can't compute it.
+    return new SCEVCouldNotCompute();
+  }
+
   // If there are any constants, fold them together.
   unsigned Idx = 0;
   if (SCEVConstant *LHSC = dyn_cast<SCEVConstant>(Ops[0])) {

diff --git a/test/Transforms/IndVarsSimplify/2008-10-03-CouldNotCompute.ll b/test/Transforms/IndVarsSimplify/2008-10-03-CouldNotCompute.ll
new file mode 100644 (file)
index 0000000..c78188d
--- /dev/null
+++ b/test/Transforms/IndVarsSimplify/2008-10-03-CouldNotCompute.ll
@@ -0,0 +1,32 @@
+; RUN: llvm-as < %s | opt -indvars
+; PR2857
+
+@foo = external global i32             ; <i32*> [#uses=1]
+
+define void @test(i32 %n, i32 %arg) {
+entry:
+       br i1 false, label %bb.nph, label %return
+
+bb.nph:                ; preds = %entry
+       %0 = load i32* @foo, align 4            ; <i32> [#uses=1]
+       %1 = sext i32 %0 to i64         ; <i64> [#uses=1]
+       br label %bb
+
+bb:            ; preds = %bb, %bb.nph
+       %.in = phi i32 [ %2, %bb ], [ %n, %bb.nph ]             ; <i32> [#uses=1]
+       %val.02 = phi i64 [ %5, %bb ], [ 0, %bb.nph ]           ; <i64> [#uses=2]
+       %result.01 = phi i64 [ %4, %bb ], [ 0, %bb.nph ]                ; <i64> [#uses=1]
+       %2 = add i32 %.in, -1           ; <i32> [#uses=2]
+       %3 = mul i64 %1, %val.02                ; <i64> [#uses=1]
+       %4 = add i64 %3, %result.01             ; <i64> [#uses=2]
+       %5 = add i64 %val.02, 1         ; <i64> [#uses=1]
+       %6 = icmp sgt i32 %2, 0         ; <i1> [#uses=1]
+       br i1 %6, label %bb, label %bb3.bb4_crit_edge
+
+bb3.bb4_crit_edge:             ; preds = %bb
+       %.lcssa = phi i64 [ %4, %bb ]           ; <i64> [#uses=0]
+       ret void
+
+return:                ; preds = %entry
+       ret void
+}