Compiler: type inference bug.

The recent MIR-level select optimization failed to properly manage
the type information of the new kMirOpSelect node, resulting in a
type inference mismatch that trigered a reduction of optimization
level (should not have affected correctness).

Change-Id: I4e104050674c8667b9dff15a401ab230a46e7968
diff --git a/src/compiler/dex/mir_optimization.cc b/src/compiler/dex/mir_optimization.cc
index 5345501..74b13fe 100644
--- a/src/compiler/dex/mir_optimization.cc
+++ b/src/compiler/dex/mir_optimization.cc
@@ -418,6 +418,13 @@
                   static_cast<bool*>(arena_->NewMem(sizeof(bool) * 1, false,
                                                     ArenaAllocator::kAllocDFInfo));
               mir->ssa_rep->fp_def[0] = if_true->ssa_rep->fp_def[0];
+              // Match type of uses to def.
+              mir->ssa_rep->fp_use =
+                  static_cast<bool*>(arena_->NewMem(sizeof(bool) * mir->ssa_rep->num_uses, false,
+                                                    ArenaAllocator::kAllocDFInfo));
+              for (int i = 0; i < mir->ssa_rep->num_uses; i++) {
+                mir->ssa_rep->fp_use[i] = mir->ssa_rep->fp_def[0];
+              }
               /*
                * There is usually a Phi node in the join block for our two cases.  If the
                * Phi node only contains our two cases as input, we will use the result