Broaden optimization of fcmp ([us]itofp x, constant) by instcombine.

The patch extends the optimization to cases where the constant's
magnitude is so small or large that the rounding of the conversion
is irrelevant.  The "so small" case includes negative zero.

Differential review: http://reviews.llvm.org/D11210

llvm-svn: 247708

1 files changed, 23 insertions, 14 deletions

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index b3366bb9888..f1a66816c4c 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -3734,15 +3734,7 @@ Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
 
   IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
 
-  // Check to see that the input is converted from an integer type that is small
-  // enough that preserves all bits.  TODO: check here for "known" sign bits.
-  // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
-  unsigned InputSize = IntTy->getScalarSizeInBits();
-
-  // If this is a uitofp instruction, we need an extra bit to hold the sign.
   bool LHSUnsigned = isa<UIToFPInst>(LHSI);
-  if (LHSUnsigned)
-    ++InputSize;
 
   if (I.isEquality()) {
     FCmpInst::Predicate P = I.getPredicate();
@@ -3769,13 +3761,30 @@ Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
     // equality compares as integer?
   }
 
-  // Comparisons with zero are a special case where we know we won't lose
-  // information.
-  bool IsCmpZero = RHS.isPosZero();
+  // Check to see that the input is converted from an integer type that is small
+  // enough that preserves all bits.  TODO: check here for "known" sign bits.
+  // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
+  unsigned InputSize = IntTy->getScalarSizeInBits();
 
-  // If the conversion would lose info, don't hack on this.
-  if ((int)InputSize > MantissaWidth && !IsCmpZero)
-    return nullptr;
+  // Following test does NOT adjust InputSize downwards for signed inputs, 
+  // because the most negative value still requires all the mantissa bits 
+  // to distinguish it from one less than that value.
+  if ((int)InputSize > MantissaWidth) {
+    // Conversion would lose accuracy. Check if loss can impact comparison.
+    int Exp = ilogb(RHS);
+    if (Exp == APFloat::IEK_Inf) {
+      int MaxExponent = ilogb(APFloat::getLargest(RHS.getSemantics()));
+      if (MaxExponent < (int)InputSize - !LHSUnsigned) 
+        // Conversion could create infinity.
+        return nullptr;
+    } else {
+      // Note that if RHS is zero or NaN, then Exp is negative 
+      // and first condition is trivially false.
+      if (MantissaWidth <= Exp && Exp <= (int)InputSize - !LHSUnsigned) 
+        // Conversion could affect comparison.
+        return nullptr;
+    }
+  }
 
   // Otherwise, we can potentially simplify the comparison.  We know that it
   // will always come through as an integer value and we know the constant is