GkvJwa wrote:
> > This case(new and delete object) I think we should probably fix the EH
> > numbering algorithm, This does not conflict with This PR
>
> I argued throughout this PR that "fixing this in EH numbering" is not optimal
> since mixing SEH with C++ unwinding is generally not allowed in MS C++ and
> only seems to be supported by Borland C++. In the latter case I do not know
> how unwinding is generated (e.g. which personality is chosen).
>
> What i noticed though: even if we mark the dtor as virtual for the delete
> sample there is still a scope begin+end. That is also true if we delete the
> delete call and just have the ctor present. Since this is legal in MS C++ I
> guess one can argue that we should indeed respect c++ scope begin/end
> intrinsics in the `calculateSEHStateForAsynchEH` numbering algo.
I checked the IR for this case.
```
invoke.cont2: ; preds = %invoke.cont1
invoke void @llvm.seh.scope.begin()
to label %invoke.cont3 unwind label %ehcleanup
invoke.cont3: ; preds = %invoke.cont2
store volatile ptr %call, ptr %a, align 8
%0 = load volatile ptr, ptr %a, align 8
%isnull = icmp eq ptr %0, null
br i1 %isnull, label %delete.end, label %delete.notnull
delete.notnull: ; preds = %invoke.cont3
invoke void @llvm.seh.scope.begin()
to label %invoke.cont6 unwind label %ehcleanup10
invoke.cont6: ; preds = %delete.notnull
invoke void @"??1A@@QEAA@XZ"(ptr noundef nonnull align 1 dereferenceable(1)
%0) #9
to label %invoke.cont7 unwind label %ehcleanup10
invoke.cont7: ; preds = %invoke.cont6
invoke void @llvm.seh.try.end()
to label %invoke.cont8 unwind label %ehcleanup10
invoke.cont8: ; preds = %invoke.cont7
invoke void @"??3@YAXPEAX_K@Z"(ptr noundef %0, i64 noundef 1) #10
to label %invoke.cont9 unwind label %ehcleanup14
invoke.cont9: ; preds = %invoke.cont8
br label %delete.end
delete.end: ; preds = %invoke.cont9,
%invoke.cont3
invoke void @llvm.seh.try.end()
to label %invoke.cont13 unwind label %ehcleanup14
invoke.cont13: ; preds = %delete.end
%1 = call ptr @llvm.localaddress()
call void @"?fin$0@0@f@@"(i8 noundef 0, ptr noundef %1)
ret void
```
`delete.notnull` include `llvm.seh.scope.begin`, but `delete.end`
not(`delete.notnull` also br `delete.end` too
This prevents `new` and `delete` from being compiled into the `__try` block
simultaneously, unlike in MSVC.
https://github.com/llvm/llvm-project/pull/172287
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
