Incorrect transformation of (sub,neg) to (sub) for fp

A pair (sub,neg) should not be transformed to (sub) for
floating point operations, otherwise we can lose the sign of
zero for instructions like this:
  - (A - B) != B - A if B == A

Change-Id: I4d612612d4dc0a067fac5721ad206f74168bcd36
Signed-off-by: Serguei Katkov <serguei.i.katkov@intel.com>

1 files changed, 3 insertions, 1 deletions

diff --git a/compiler/optimizing/instruction_simplifier.cc b/compiler/optimizing/instruction_simplifier.cc
index afbc490..180bdc9 100644
--- a/compiler/optimizing/instruction_simplifier.cc
+++ b/compiler/optimizing/instruction_simplifier.cc
@@ -424,7 +424,8 @@ void InstructionSimplifierVisitor::VisitNeg(HNeg* instruction) {
     return;
   }
 
-  if (input->IsSub() && input->HasOnlyOneNonEnvironmentUse()) {
+  if (input->IsSub() && input->HasOnlyOneNonEnvironmentUse() &&
+      !Primitive::IsFloatingPointType(input->GetType())) {
     // Replace code looking like
     //    SUB tmp, a, b
     //    NEG dst, tmp
@@ -435,6 +436,7 @@ void InstructionSimplifierVisitor::VisitNeg(HNeg* instruction) {
     // worse code. In particular, we do not want the live ranges of `a` and `b`
     // to be extended if we are not sure the initial 'SUB' instruction can be
     // removed.
+    // We do not perform optimization for fp because we could lose the sign of zero.
     HSub* sub = input->AsSub();
     HSub* new_sub =
         new (GetGraph()->GetArena()) HSub(instruction->GetType(), sub->GetRight(), sub->GetLeft());