On Fri, Mar 1, 2013 at 12:38 PM, Eric Botcazou <ebotca...@adacore.com> wrote:
> Hi,
>
> this is the ICE during the build of the Ada runtime on x86-64/Windows: the
> compiler stops on the assertion in declare_return_variable:
>
> var = return_slot;
> gcc_assert (TREE_CODE (var) != SSA_NAME);
>
> The problem stems for the fnsplit pass: it turns
>
> P.Sin (const long_long_float x)
> {
> long_long_float _3;
> boolean _4;
> q__double _9;
>
> <bb 2>:
> _3 = ABS_EXPR <x_2(D)>;
> _4 = _3 < 1.0e+0;
> if (_4 != 0)
> goto <bb 3>;
> else
> goto <bb 4>;
>
> <bb 3>:
> <retval> = x_2(D);
> goto <bb 5>;
>
> <bb 4>:
> _9 = q.sin (x_2(D));
> <retval> = _9;
>
> <bb 5>:
> return <retval>;
>
> }
>
> into
>
> P.Sin (const long_long_float x)
> {
> long_long_float _3;
> boolean _4;
>
> <bb 2>:
> _3 = ABS_EXPR <x_2(D)>;
> _4 = _3 < 1.0e+0;
> if (_4 != 0)
> goto <bb 3>;
> else
> goto <bb 4>;
>
> <bb 3>:
> <retval> = x_2(D);
> goto <bb 5>;
>
> <bb 4>:
> <retval> = p.sin.part (x_2(D)); [return slot optimization]
>
> <bb 5>:
> return <retval>;
>
> }
>
> Note the RSO flag on the call to the split part: it's bogus since the return
> type is scalar:
>
> <real_type 0x7ffff6da02a0 long_long_float sizes-gimplified visited XF>
>
> Fixed by checking that either the type of the result is not is_gimple_reg_type
> or the result is passed by reference before setting the RSO flag.
>
> Tested on x86_64-suse-linux, OK for the mainline?
But in theory this pessimizes code as aggregate_value_p returned true
for this? That is, isn't the bug that we rewrite a possible return-slot-decl
into SSA? (do we do that here?)
Thanks,
Richard.
>
> 2013-01-01 Eric Botcazou <ebotca...@adacore.com>
>
> PR tree-optimization/56424
> * ipa-split.c (split_function): Do not set the RSO flag if result is
> not by reference and its type is not a register type.
>
>
> 2013-01-01 Eric Botcazou <ebotca...@adacore.com>
>
> * gcc.dg/pr56424.c: New test.
>
>
> --
> Eric Botcazou
