On 10/1/20 5:26 AM, Richard Biener wrote: > On Wed, 30 Sep 2020, Jason Merrill wrote: > >> On 9/28/20 3:09 PM, Jason Merrill wrote: >>> On 9/28/20 3:56 AM, Richard Biener wrote: >>>> On Fri, 25 Sep 2020, Jason Merrill wrote: >>>> >>>>> On 9/25/20 2:30 AM, Richard Biener wrote: >>>>>> On Thu, 24 Sep 2020, Jason Merrill wrote: >>>>>> >>>>>>> On 9/24/20 3:43 AM, Richard Biener wrote: >>>>>>>> On Wed, 23 Sep 2020, Jason Merrill wrote: >>>>>>>> >>>>>>>>> On 9/23/20 2:42 PM, Richard Biener wrote: >>>>>>>>>> On September 23, 2020 7:53:18 PM GMT+02:00, Jason Merrill >>>>>>>>>> >>>>>>>>>> wrote: >>>>>>>>>>> On 9/23/20 4:14 AM, Richard Biener wrote: >>>>>>>>>>>> C++ operator delete, when DECL_IS_REPLACEABLE_OPERATOR_DELETE_P, >>>>>>>>>>>> does not cause the deleted object to be escaped.? It also has no >>>>>>>>>>>> other interesting side-effects for PTA so skip it like we do >>>>>>>>>>>> for BUILT_IN_FREE. >>>>>>>>>>> >>>>>>>>>>> Hmm, this is true of the default implementation, but since the >>>>>>>>>>> function >>>>>>>>>>> >>>>>>>>>>> is replaceable, we don't know what a user definition might do with >>>>>>>>>>> the >>>>>>>>>>> pointer. >>>>>>>>>> >>>>>>>>>> But can the object still be 'used' after delete? Can delete fail / >>>>>>>>>> throw? >>>>>>>>>> >>>>>>>>>> What guarantee does the predicate give us? >>>>>>>>> >>>>>>>>> The deallocation function is called as part of a delete expression in >>>>>>>>> order >>>>>>>>> to >>>>>>>>> release the storage for an object, ending its lifetime (if it was not >>>>>>>>> ended >>>>>>>>> by >>>>>>>>> a destructor), so no, the object can't be used afterward. >>>>>>>> >>>>>>>> OK, but the delete operator can access the object contents if there >>>>>>>> wasn't a destructor ... >>>>>>> >>>>>>>>> A deallocation function that throws has undefined behavior. >>>>>>>> >>>>>>>> OK, so it seems the 'replaceable' operators are the global ones >>>>>>>> (for user-defined/class-specific placement variants I see arbitrary >>>>>>>> extra arguments that we'd possibly need to handle). >>>>>>>> >>>>>>>> I'm happy to revert but I'd like to have a testcase that FAILs >>>>>>>> with the patch ;) >>>>>>>> >>>>>>>> Now, the following aborts: >>>>>>>> >>>>>>>> struct X { >>>>>>>> ???? static struct X saved; >>>>>>>> ???? int *p; >>>>>>>> ???? X() { __builtin_memcpy (this, &saved, sizeof (X)); } >>>>>>>> }; >>>>>>>> void operator delete (void *p) >>>>>>>> { >>>>>>>> ???? __builtin_memcpy (&X::saved, p, sizeof (X)); >>>>>>>> } >>>>>>>> int main() >>>>>>>> { >>>>>>>> ???? int y = 1; >>>>>>>> ???? X *p = new X; >>>>>>>> ???? p->p = &y; >>>>>>>> ???? delete p; >>>>>>>> ???? X *q = new X; >>>>>>>> ???? *(q->p) = 2; >>>>>>>> ???? if (y != 2) >>>>>>>> ?????? __builtin_abort (); >>>>>>>> } >>>>>>>> >>>>>>>> and I could fix this by not making *p but what *p points to escape. >>>>>>>> The testcase is of course maximally awkward, but hey ... ;) >>>>>>>> >>>>>>>> Now this would all be moot if operator delete may not access >>>>>>>> the object (or if the object contents are undefined at that point). >>>>>>>> >>>>>>>> Oh, and the testcase segfaults when compiled with GCC 10 because >>>>>>>> there we elide the new X / delete p pair ... which is invalid then? >>>>>>>> Hmm, we emit >>>>>>>> >>>>>>>> ???? MEM[(struct X *)_8] ={v} {CLOBBER}; >>>>>>>> ???? operator delete (_8, 8); >>>>>>>> >>>>>>>> so the object contents are undefined _before_ calling delete >>>>>>>> even when I do not have a DTOR?? That is, the above, >>>>>>>> w/o -fno-lifetime-dse, makes the PTA patch OK for the testcase. >>>>>>> >>>>>>> Yes, all classes have a destructor, even if it's trivial, so the >>>>>>> object's >>>>>>> lifetime definitely ends before the call to operator delete. This is >>>>>>> less >>>>>>> clear for scalar objects, but treating them similarly would be >>>>>>> consistent >>>>>>> with >>>>>>> other recent changes, so I think it's fine for us to assume that scalar >>>>>>> objects are also invalidated before the call to operator delete. But of >>>>>>> course this doesn't apply to explicit calls to operator delete outside >>>>>>> of a >>>>>>> delete expression. >>>>>> >>>>>> OK, so change the testcase main slightly to >>>>>> >>>>>> int main() >>>>>> { >>>>>> ??? int y = 1; >>>>>> ??? X *p = new X; >>>>>> ??? p->p = &y; >>>>>> ??? ::operator delete(p); >>>>>> ??? X *q = new X; >>>>>> ??? *(q->p) = 2; >>>>>> ??? if (y != 2) >>>>>> ????? __builtin_abort (); >>>>>> } >>>>>> >>>>>> in this case the lifetime of *p does not end before calling >>>>>> ::operator delete() and delete can stash the object contents >>>>>> somewhere before ending its lifetime.? For the very same reason >>>>>> we may not elide a new/delete pair like in >>>>>> >>>>>> int main() >>>>>> { >>>>>> ??? int *p = new int; >>>>>> ??? *p = 1; >>>>>> ??? ::operator delete (p); >>>>>> } >>>>> >>>>> Correct; the permission to elide new/delete pairs are for the expressions, >>>>> not >>>>> the functions. >>>>> >>>>>> which we before the change did not do only because calling >>>>>> operator delete made p escape.? Unfortunately points-to analysis >>>>>> cannot really reconstruct whether delete was called as part of >>>>>> a delete expression or directly (and thus whether object lifetime >>>>>> ended already), neither can DCE.? So I guess we need to mark >>>>>> the operator delete call in some way to make those transforms >>>>>> safe.? At least currently any operator delete call makes the >>>>>> alias guarantee of a operator new call moot by forcing the object >>>>>> to be aliased with all global and escaped memory ... >>>>>> >>>>>> Looks like there are some unallocated flags for CALL_EXPR we could >>>>>> pick but I wonder if we can recycle protected_flag which is >>>>>> >>>>>> ???????? CALL_FROM_THUNK_P and >>>>>> ???????? CALL_ALLOCA_FOR_VAR_P in >>>>>> ???????????? CALL_EXPR >>>>>> >>>>>> for calls to DECL_IS_OPERATOR_{NEW,DELETE}_P, thus whether >>>>>> we have CALL_FROM_THUNK_P for those operators.? Guess picking >>>>>> a new flag is safer. >>>>> >>>>> We won't ever call those operators from a thunk, so it should be OK to >>>>> reuse >>>>> it. >>>>> >>>>>> But, does it seem correct that we need to distinguish >>>>>> delete expressions from plain calls to operator delete? >>>>> >>>>> A reason for that distinction came up in the context of omitting >>>>> new/delete >>>>> pairs: we want to consider the operator first called by the new or delete >>>>> expression, not a call from that first operator to another operator >>>>> new/delete >>>>> and exposed by inlining. >>>>> >>>>> https://gcc.gnu.org/pipermail/gcc-patches/2020-April/543404.html >>>>> >>>>>> In this context I also wonder about non-replaceable operator delete, >>>>>> specifically operator delete in classes - are there any semantic >>>>>> differences between those or why did we choose to only mark >>>>>> the replaceable ones? >>>>> >>>>> The standard says that for omitting a 'new' allocation, the operator new >>>>> has >>>>> to be a replaceable one, but does not say the same about 'delete'; it just >>>>> says that if the allocation was omitted, the delete-expression does not >>>>> call a >>>>> deallocation function.? It may not be necessary to make this distinction >>>>> for >>>>> delete.? And this distinction could be local to the front end. >>>>> >>>>> In the front end, we currently have cxx_replaceable_global_alloc_fn that >>>>> already ignores the replaceability of operator delete.? And we have >>>>> CALL_FROM_NEW_OR_DELETE_P, that would just need to move into the middle >>>>> end. >>>>> And perhaps get renamed to CALL_OMITTABLE_NEW_OR_DELETE_P, and not get set >>>>> for >>>>> calls to non-replaceable operator new. >>>> >>>> CALL_FROM_NEW_OR_DELETE_P indeed looks like the best fit - it's >>>> only evaluated when cxx_replaceable_global_alloc_fn matches in the C++ >>>> FE so could be made to cover only replaceable variants. >>>> >>>> CALL_REPLACEABLE_NEW_OR_DELETE_P () maybe, since we already use >>>> REPLACEABLE for the fndecl flags?? OMITTABLE is too specific >>>> for the PTA case where it really matters whether the object >>>> lifetime ends before the delete call, not whether it can be >>>> omitted (hmm, guess that's not 100% overlap then either...). >>> >>> That seems like good overlap to me, if we agree that object lifetime ends >>> before any delete call from a delete-expression, whether or not the operator >>> delete is replaceable. >>> >>>> Mind doing the C++ side of things recycling protected_flag as suggested? >>> >>> OK. > > Find attached a patch series, your patch plus the GIMPLE side of the fix. > This shows that the C++ FE side is possibly incomplete with for > example g++.dg/pr94314-2.C now FAILing. There we have > > A *a = new A (argc); > delete a; > > being expanded to > > < (void) (a = TARGET_EXPR ;, try > { > A::A ((struct A *) D.2357, argc); > } > catch > { > operator delete (D.2357); > }, (struct A *) D.2357;) >>>>>; > < if (SAVE_EXPR != 0B) > { > try > { > *SAVE_EXPR = {CLOBBER}; > } > finally > { > operator delete ((void *) SAVE_EXPR ); > } > } > else > { > <<< Unknown tree: void_cst >>> > } >>>>>; > return = 0; > > where the operator delete call in the catch {} expression is > not marked as CALL_FROM_NEW_OR_DELETE_P, possibly because this > call is compiler generated. > > So it's probably technically true that CALL_FROM_NEW_OR_DELETE_P is > false but we still expect the same guarantees to hold here, in > particular we expect to elide the new/delete pair? That seems like an oversight in the standard. This first patch sets the flag. The second patch in the file removes consideration of whether an operator delete is replaceable, as I was discussing earlier.