https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118992
--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> --- (In reply to H.J. Lu from comment #6) > This works for x86-64: > > diff --git a/gcc/cse.cc b/gcc/cse.cc > index 70d5caac4ca..786624cd890 100644 > --- a/gcc/cse.cc > +++ b/gcc/cse.cc > @@ -2287,6 +2287,10 @@ hash_rtx (const_rtx x, machine_mode mode, > record = true; > else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC) > record = true; > + else if (REG_ATTRS (x) > + && REG_EXPR (x) > + && TREE_CODE (REG_EXPR (x)) == PARM_DECL) > + record = true; as a heuristic this doesn't make much sense > else if (targetm.small_register_classes_for_mode_p (GET_MODE (x))) > record = false; > else if (targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno))) so which one is true for x86? The 2nd? This is actually a quite bad heuristic given it applies 'locally' instead when performing a replacement which can take into account distance and register pressure.
