[DAGCombiner] restrict (float)((int) f) --> ftrunc with no-signed-zeros

As noted in the D44909 review, the transform from (fptosi+sitofp) to ftrunc 
can produce -0.0 where the original code does not:

#include <stdio.h>
  
int main(int argc) {
  float x;
  x = -0.8 * argc;
  printf("%f\n", (float)((int)x));
  return 0;
}

$ clang -O0 -mavx fp.c ; ./a.out 
0.000000
$ clang -O1 -mavx fp.c ; ./a.out 
-0.000000

Ideally, we'd use IR/node flags to predicate the transform, but the IR parser 
doesn't currently allow fast-math-flags on the cast instructions. So for now, 
just use the function attribute that corresponds to clang's "-fno-signed-zeros" 
option.

Differential Revision: https://reviews.llvm.org/D48085

llvm-svn: 335761

1 files changed, 7 insertions, 2 deletions

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 34a00733c39..1b88f7cf888 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -11123,9 +11123,14 @@ static SDValue foldFPToIntToFP(SDNode *N, SelectionDAG &DAG,
     return SDValue();
 
   // We only do this if the target has legal ftrunc. Otherwise, we'd likely be
-  // replacing casts with a libcall.
+  // replacing casts with a libcall. We also must be allowed to ignore -0.0
+  // because FTRUNC will return -0.0 for (-1.0, -0.0), but using integer
+  // conversions would return +0.0.
+  // FIXME: We should be able to use node-level FMF here.
+  // TODO: If strict math, should we use FABS (+ range check for signed cast)?
   EVT VT = N->getValueType(0);
-  if (!TLI.isOperationLegal(ISD::FTRUNC, VT))
+  if (!TLI.isOperationLegal(ISD::FTRUNC, VT) ||
+      !DAG.getTarget().Options.NoSignedZerosFPMath)
     return SDValue();
 
   // fptosi/fptoui round towards zero, so converting from FP to integer and