[Bug tree-optimization/48063] [4.6 Regression] ICE: verify_stmts failed: conversion of register to a different size with -fno-early-inlining

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48063

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-03-10 18:50:19 UTC ---
Again caused by http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=161655
setup_one_parameter has:
  if (value
      && value != error_mark_node
      && !useless_type_conversion_p (TREE_TYPE (p), TREE_TYPE (value)))
    {
      if (fold_convertible_p (TREE_TYPE (p), value))
        rhs = fold_build1 (NOP_EXPR, TREE_TYPE (p), value);
      else
        /* ???  For valid (GIMPLE) programs we should not end up here.
           Still if something has gone wrong and we end up with truly
           mismatched types here, fall back to using a VIEW_CONVERT_EXPR
           to not leak invalid GIMPLE to the following passes.  */
        rhs = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (p), value);
    }
which obviously can't work when the VCE is invalid.  We really shouldn't be
trying to inline in that case.