[C++ Patch] PR 89488 ("[9 Regression] ICE in merge_exception_specifiers, at cp/typeck2.c:2395")

[Jason Merrill]

On 2/26/19 11:02 AM, Paolo Carlini wrote:
> Hi,
> 
> On 26/02/19 15:28, Jason Merrill wrote:
>> On 2/25/19 10:27 AM, Paolo Carlini wrote:
>>> Hi,
>>>
>>> this error recovery regression has to do with the recent changes 
>>> committed by Jason for c++/88368. What happens is that 
>>> maybe_instantiate_noexcept fails the hard way, thus, toward the end 
>>> of the function, doesn't update TREE_TYPE (fn) and just returns 
>>> false. process_subob_fn doesn't notice and proceeds to call 
>>> merge_exception_specifiers anyway where of course the gcc_assert 
>>> (!DEFERRED_NOEXCEPT_SPEC_P (add)) triggers, because 
>>> maybe_instantiate_noexcept has not done its normal job. To improve 
>>> error-recovery I think we can simply leave *spec_p alone in such 
>>> cases, because we would merge the *spec_p with a 
>>> TYPE_RAISES_EXCEPTIONS (TREE_TYPE (fn)) where TREE_TYPE (fn) has not 
>>> been normally computed. I tried a few other things which prima facie 
>>> looked sensible but nothing else worked - eg, returning false from 
>>> maybe_instantiate_noexcept and also updating TREE_TYPE (fn) to a 
>>> noexcept_false_spec variant causes regressions exactly for the 
>>> testcases of c++/88368.
>>
>> If maybe_instantiate_noexcept returns false, I think we should set 
>> *spec_p to error_mark_node.
> 
> Sure, that certainly works, I tested it a couple of days ago and I'm 
> finishing testing the below now. The only difference is that during 
> error-recovery 'zl ()' is seen as seriously broken and we don't give the 
> second,  "cannot convert", error message, which we used to give.

OK if it passes.

Jason