Emmiting new intrinsic that strips invariant.groups to make
devirtulization sound, as described in RFC: Devirtualization v2.
Details
Diff Detail
- Repository
- rC Clang
Event Timeline
clang/include/clang/AST/DeclCXX.h | ||
---|---|---|
779 ↗ | (On Diff #148312) | isCompleteDefinition checks whether this declaration of the class is a definition, not whether it has a definition anywhere; the latter is what you need here. You can use hasDefinition to check that. Please also check hasAnyDependentBases() (or add a comment to this function to indicate that it may return false for a templated class whose instantiations might be dynamic classes) -- if a class has dependent bases, we might not find out that it's a dynamic class until it's instantiated. |
clang/lib/CodeGen/CGExprScalar.cpp | ||
1626–1627 ↗ | (On Diff #148312) | Are there any cases where we need a barrier when the destination type is a dynamic type here? |
clang/lib/CodeGen/CGExprScalar.cpp | ||
---|---|---|
1626–1627 ↗ | (On Diff #148312) | No, I don't think so. I will also add test for that. |
Added launder when going from possiblyNotDynamic to possiblyDynamic emitting strip for pointer -> int only if poitner is possiblyDynamic
clang/include/clang/AST/DeclCXX.h | ||
---|---|---|
784 ↗ | (On Diff #148746) | Both of these new methods deserve doc comments explaining that they're conservative checks because the class might be incomplete or dependent. I think NonDynamic would read better than NotDynamic. |
clang/lib/CodeGen/CGExprScalar.cpp | ||
1623 ↗ | (On Diff #148746) | Unnecessary getTypePtr(). |
1624 ↗ | (On Diff #148746) | The opposite of Dst is Src. Alternatively, the opposite of Source is Destination (or Result). Please pick. |
1633 ↗ | (On Diff #148746) | If you made a couple of tiny helper functions here that you could invoke on either SourceClassDecl or DstClassDecl, you could avoid some redundant logic and also make the calls self-documenting enough to legibly inline into the if-conditions. ...in fact, since you start from a QualType in every case, maybe you should just define the helper as a method there. |
1647 ↗ | (On Diff #148746) | Incidentally, how do you protect against code like this? A *ptr; reinterpret_cast<B *&>(ptr) = new B(); ptr->foo(); Presumably there needs to be a launder/strip here, but I guess it would have to be introduced by the middle-end when forwarding the store? The way I've written this is an aliasing violation, but (1) I assume your pass isn't disabled whenever strict-aliasing is disabled and (2) you can do this with a memcpy and still pretty reliably expect that LLVM will be able to eventually forward the store. |
1791 ↗ | (On Diff #148746) | Another place you could definitely just use that helper function on QualType. |
1802 ↗ | (On Diff #148746) | Same. |
3305 ↗ | (On Diff #148746) | Yeah, helper function on QualType, please. :) |
clang/lib/CodeGen/CGExprScalar.cpp | ||
---|---|---|
1623 ↗ | (On Diff #148746) | getType returns QualType and it does not have getPointeeCXXRecordDecl. Am I missing something? |
1647 ↗ | (On Diff #148746) | Can you add more info on what is A and B so I can make sure I understand it correctly? I haven't thought about the strict aliasing. I think the only sane way would be to require strict aliasing for the strict vtable pointers. |
Refactor
clang/lib/CodeGen/CGExprScalar.cpp | ||
---|---|---|
1633 ↗ | (On Diff #148746) | oh yea, it is definitely better! |
clang/lib/CodeGen/CGExprScalar.cpp | ||
---|---|---|
1624 ↗ | (On Diff #148746) | Not sure if it still holds (if the DestTy should be DstTy, but I haven't declared it myself) |
clang/lib/CodeGen/CGExprScalar.cpp | ||
---|---|---|
1623 ↗ | (On Diff #148746) | operator-> on QualType drills to the Type*. |
1647 ↗ | (On Diff #148746) | It's whatever case you're worried about such that you have to launder member accesses and bitcasts. And like I mentioned, relying on strict aliasing isn't enough because you can do it legally with memcpy. Maybe it's okay to consider it UB? I'm not sure about that. |
clang/lib/CodeGen/CGExprScalar.cpp | ||
---|---|---|
1647 ↗ | (On Diff #148746) | AFAIK reinterpreting one class as another is UB if they are not in hierarchy (especially calling virtual function on reinterpreted class), not sure if strict aliasing should allow it anyway (if it would be a hand written custom vptr then it should be ok with strict aliasing turned off, but with vptr I don't think it is legal). |
clang/lib/CodeGen/CGExprScalar.cpp | ||
---|---|---|
1647 ↗ | (On Diff #148746) | OK, here's how I think about what we're doing here: We view the IR-level pointer value for a pointer to a dynamic class type as being a fat pointer, containing the actual pointer value and also a tag indicating the dynamic type of the object (only notionally, though -- the actual bit representation of the pointer doesn't include the extra information, but we don't ever emit IR that inspects the bit representation of the fat pointer to avoid exposing that fact). In that model, if you try to type pun between a pointer to a dynamic class type and a pointer to a non-dynamic-class type, that can't be expected to work because the (notional) value is different, much as type punning between a derived and base class pointer wouldn't work for a pointer to something other than a base class at offset zero. I think @rjmccall's example is OK, because both A and B would need to be dynamic class types in a hierarchy to work, and that means we'd be using the same notional pointer representation. A slight variation of that example: struct A {}; struct B : A { virtual void f(); }; struct C : B { void f(); } c; A *p = &c; B *q; memcpy(&q, &p, sizeof(B*)); // or q = std::bit_cast<B*>(p); q->f(); ... would be UB, because the representation of an A* and a B* are different (a B* contains a tag and an A* does not). |
clang/lib/CodeGen/CGExprScalar.cpp | ||
---|---|---|
1647 ↗ | (On Diff #148746) | Does this answer satisfy you John? Can I push it to trunk? |
clang/lib/CodeGen/CGExprScalar.cpp | ||
---|---|---|
1647 ↗ | (On Diff #148746) | Yeah, Richard's answer makes sense to me. |