[llvm-commits] [llvm] r129199 - /llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

[Chris Lattner]

Author: lattner
Date: Sat Apr  9 01:57:13 2011
New Revision: 129199

URL: http://llvm.org/viewvc/llvm-project?rev=129199&view=rev
Log:
Avoid excess precision issues that lead to generating host-compiler-specific code.
Switch lowering probably shouldn't be using FP for this.  This resolves PR9581.


Modified:
    llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

Modified: llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp?rev=129199&r1=129198&r2=129199&view=diff
==============================================================================
--- llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp (original)
+++ llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp Sat Apr  9 01:57:13 2011
@@ -2019,9 +2019,13 @@
     APInt Range = ComputeRange(LEnd, RBegin);
     assert((Range - 2ULL).isNonNegative() &&
            "Invalid case distance");
-    double LDensity = (double)LSize.roundToDouble() /
+    // Use volatile double here to avoid excess precision issues on some hosts,
+    // e.g. that use 80-bit X87 registers.
+    volatile double LDensity =
+       (double)LSize.roundToDouble() /
                            (LEnd - First + 1ULL).roundToDouble();
-    double RDensity = (double)RSize.roundToDouble() /
+    volatile double RDensity =
+      (double)RSize.roundToDouble() /
                            (Last - RBegin + 1ULL).roundToDouble();
     double Metric = Range.logBase2()*(LDensity+RDensity);
     // Should always split in some non-trivial place