[PATCH] Fix PR middle-end/33088

Richard Guenther richard.guenther@gmail.com
Wed Dec 12 22:38:00 GMT 2007

On Dec 12, 2007 11:21 PM, Eric Botcazou <ebotcazou@libertysurf.fr> wrote:
> As diagnosed by Joseph, this is a problem introduced by RTH's patch
>   http://gcc.gnu.org/ml/gcc-patches/2005-06/msg00860.html>
> which promotes partial stores to the real or imaginary part of a complex
> variable to total stores.
> This is achieved by generating a load of the other, unmodified part of the
> complex variables just before the total store, thus introducing a load of
> an undefined value if the variable is uninitialized.
> This works, i.e. the uninitialized loads are cleaned up, if the fully-fledged
> complex operations lowering pass is run, followed by a DCE pass.  The problem
> here is that the lowering pass is not run, even at -O or above, because the
> associated predicate returns false for the function at hand.
> The proposed fix is two-pronged: (1) not promoting the partial stores at -O0,
> (2) teaching the predicate to detect the uninitialized loads at -O1 or above.
> (1) is achieved by just tweaking is_gimple_reg_type, i.e. DECL_GIMPLE_REG_P is
> still being set on the individual variables.  This seems to work, but...
> Tested on i586-suse-linux, OK for mainline?

I think the idea and implementation are sound.  But, did you verify
that the places
we set DECL_GIMPLE_REG_P in the gimplifier are properly guarded with the
changed predicate?  That is, shouldn't the occurances of

      if ((TREE_CODE (TREE_TYPE (parm)) == COMPLEX_TYPE
           || TREE_CODE (TREE_TYPE (parm)) == VECTOR_TYPE)
        DECL_GIMPLE_REG_P (parm) = 1;

changed to use the is_gimple_reg_type predicate?  (at least in the gimplifier)

Thanks for tackling this bug!

> 2007-12-12  Eric Botcazou  <ebotcazou@libertysurf.fr>
>         PR middle-end/33088
>         * gimplify.c (gimplify_modify_expr_complex_part): Add note to comment.
>         * tree-complex.c (init_dont_simulate_again): Return true if there are
>         uninitialized loads generated by gimplify_modify_expr_complex_part.
>         * tree-gimple.c (is_gimple_reg_type): Return false for complex types
>         if not optimizing.
>         * tree-ssa.c (ssa_undefined_value_p): New predicate extracted from...
>         (warn_uninit): ...here.  Use ssa_undefined_value_p.
>         * tree-ssa-pre.c (is_undefined_value): Delete.
>         (phi_translate_1): Use ssa_undefined_value_p.
>         (add_to_exp_gen): Likewise.
>         (make_values_for_stmt): Likewise.
>         * tree-flow.h (ssa_undefined_value_p): Declare.
> 2007-12-12  Eric Botcazou  <ebotcazou@libertysurf.fr>
>         * gcc.dg/uninit-13.c: UnXFAIL.
>         * gcc.dg/complex-5.c: New testcase.
> --
> Eric Botcazou

More information about the Gcc-patches mailing list