On 03/24/2017 12:49 PM, Richard Biener wrote:
> On Fri, Mar 24, 2017 at 10:25 AM, Martin Liška <mli...@suse.cz> wrote:
>> Hello.
>>
>> Briefly described in the PR, running ICF (without any optimization level)
>> can create a thunk call
>> that does not use an SSA_NAME which is a default def of an argument of the
>> caller:
>>
>> c (complex float b)
>> {
>> complex float arg.1;
>> float retval.0;
>>
>> <bb 2> [100.00%]:
>> arg.1_2 = b;
>> retval.0_4 = a (arg.1_2); [tail call]
>>
>> return retval.0_4;
>>
>> }
>>
>> The "arg" variable creation was introduced by Jason in r207301.
>> As complex type is passed as invisible reference, we need to find address of
>> the argument.
>> Thus I'm suggesting to find it alternatively via SSA_NAME_DEF_STMT.
>>
>> Patch can bootstrap on ppc64le-redhat-linux and survives regression tests.
>> And fixes the ICE
>> on s390x cross compiler.
>>
>> Ready to be installed?
>
> Ick, that looks like fragile code ... looks like the
> call_from_thunk_p case expects
> the default def or the param-decl. So why not create it that way? Or
Not creating the "arg" and directly passing the PARM_DECL will break verifier:
/home/marxin/Programming/testcases/pr80104.c:10:1: error: invalid argument to
gimple call
}
^
b
# .MEM_2 = VDEF <.MEM_1(D)>
retval.0_3 = a (b); [tail call]
/home/marxin/Programming/testcases/pr80104.c:10:1: internal compiler error:
verify_gimple failed
> simply not do
> this "optimization"?
You mean understanding when an argument of a wrapper (thunk) will need
invisible reference and
if so, then ICF should not do the optimization?
Martin
>
> Richard.
>
>> Martin
