[PATCH] D96976: [analyzer] Fix reinterpret_cast handling for pointer-to-member

Wed Mar 3 01:44:31 PST 2021

RedDocMD added inline comments.

================
Comment at: clang/test/Analysis/reinterpret-cast-pointer-to-member.cpp:52-62
+void testMultiple() {
+  int F::*f = &F::field;
+  int A::*a = reinterpret_cast<int A::*>(f);
+  int C::*c = reinterpret_cast<int C::*>(f);
+  A aobj;
+  C cobj;
+  aobj.*a = 13;
----------------
steakhal wrote:
> RedDocMD wrote:
> > steakhal wrote:
> > > RedDocMD wrote:
> > > > steakhal wrote:
> > > > > Wait a minute. It's not how it works.
> > > > > How I imagine member pointers, they are just offsets.
> > > > > `&F::field` is notionally equivalent with `offsetof(F, field)`. That being said, You can not apply this member pointer to any object besides `F`.
> > > > > Imagine if the classes of the inheritance tree would have other fields as well.
> > > > > Then the `offsetof(T, field)` would be different for `F`, and `C`.
> > > > > 
> > > > > This example demonstrates that both of these member pointer dereferences are UB.
> > > > > https://godbolt.org/z/15sMEP
> > > > > It returns different values depending on the optimization level, which is a clear sign of UB.
> > > > > BTW this issue is closely related to strict aliasing.
> > > > The member access on A is definitely UB, I guess I will do what I proposed in the `Some` case.
> > > > I don't think the other one is.  Consider the following:
> > > > ```
> > > > struct A {};
> > > > struct B : public A {};
> > > > struct C {
> > > >   int field;
> > > > };
> > > > struct D : public C {};
> > > > struct E : public B, public D {};
> > > > struct F : public E {};
> > > > 
> > > > int main() {
> > > >   int F::* ff = &F::field;
> > > >   int C::* cf1 = static_cast<int C::*>(ff);
> > > >   int C::* cf2 = reinterpret_cast<int C::*>(ff);
> > > >   C c;
> > > >   c.*cf1 = 10;
> > > >   c.*cf2 = 10;
> > > >   return 0;
> > > > }
> > > > ```
> > > > `cf1` and `cf2` are the same thing, except that they are declared differently (one via `static_cast`, other via `reinterpret_cast`). If we look at the AST (truncated to the first three lines of main):
> > > > ```
> > > > CompoundStmt 0x1a4fe18 <col:12, line:18:1>
> > > >     |-DeclStmt 0x1a4f3a8 <line:11:3, col:26>
> > > >     | `-VarDecl 0x1a21078 <col:3, col:21> col:12 used ff 'int F::*' cinit
> > > >     |   `-ImplicitCastExpr 0x1a4f378 <col:17, col:21> 'int F::*' <BaseToDerivedMemberPointer (E -> D -> C)>
> > > >     |     `-UnaryOperator 0x1a4f360 <col:17, col:21> 'int C::*' prefix '&' cannot overflow
> > > >     |       `-DeclRefExpr 0x1a4f2f8 <col:18, col:21> 'int' lvalue Field 0x1a207d0 'field' 'int'
> > > >     |-DeclStmt 0x1a4f560 <line:12:3, col:43>
> > > >     | `-VarDecl 0x1a4f428 <col:3, col:42> col:12 used cf1 'int C::*' cinit
> > > >     |   `-CXXStaticCastExpr 0x1a4f518 <col:18, col:42> 'int C::*' static_cast<int struct C::*> <DerivedToBaseMemberPointer (E -> D -> C)>
> > > >     |     `-ImplicitCastExpr 0x1a4f500 <col:40> 'int F::*' <LValueToRValue> part_of_explicit_cast
> > > >     |       `-DeclRefExpr 0x1a4f498 <col:40> 'int F::*' lvalue Var 0x1a21078 'ff' 'int F::*'
> > > >     |-DeclStmt 0x1a4f6e8 <line:13:3, col:48>
> > > >     | `-VarDecl 0x1a4f5c8 <col:3, col:47> col:12 used cf2 'int C::*' cinit
> > > >     |   `-CXXReinterpretCastExpr 0x1a4f6b8 <col:18, col:47> 'int C::*' reinterpret_cast<int struct C::*> <ReinterpretMemberPointer>
> > > >     |     `-ImplicitCastExpr 0x1a4f6a0 <col:45> 'int F::*' <LValueToRValue> part_of_explicit_cast
> > > >     |       `-DeclRefExpr 0x1a4f638 <col:45> 'int F::*' lvalue Var 0x1a21078 'ff' 'int F::*'
> > > > ```
> > > > Notice how the `static_cast` figures out the path to the correct subobject. This is how member-pointers are handled as far as I can tell.
> > > > Unfortunately, Stroustrup is surprisingly scant when it comes to this topic. I am trying to dig through the Standard.
> > > I tried to highlight that the classes could have non-static data members which could cause your assumption about `cf1 == cf2` not to hold anymore. See my previously attached godbolt link.
> > > 
> > > `static_cast<T>(p)` makes sure that the //offset// that the member pointer `p` represents is updated accordingly to match the //offset// of the `field` member within the new type `T`. Usually, it involves some addition/subtraction to accommodate this.
> > > While `static_cast<T>(p)` does not update the value, just reinterprets it in a different way - which usually results in an invalid pointer though and a bunch of subtle rules kick in such as the type similarity, pointer interchangeability, which in general known as //strict-aliasing//.
> > > While static_cast<T>(p) does not update the value, just reinterprets it in a different way - which usually results in an invalid pointer   though and a bunch of subtle rules kick in such as the type similarity, pointer interchangeability, which in general known as strict-aliasing.
> > Presumably you mean `reinterpret_cast`.
> > 
> > Ah, I see. This makes it glaringly obvious: [[ https://godbolt.org/z/f6vs77 | Another Godbolt link ]]
> > So I guess handling `reinterpret_cast` for pointer-to-members pointless. I guess the idea will be to simply not analyze pointer-to-members which have been obtained through `reinterpret_cast.`
> > Presumably you mean reinterpret_cast.
> Yes, sorry for copy-pasting too much xD
> 
> > So I guess handling reinterpret_cast for pointer-to-members pointless.
> I don't know if it's really pointless, but hard to model precisely for sure.
> 
> > I guess the idea will be to simply not analyze pointer-to-members which have been obtained through `reinterpret_cast`.
> I guess, you could just return unknown? I don't know. This area is not my expertise, I just wanted to get you going with my observations.
> I guess, you could just return unknown? 
If I return `nullptr` as `PointerToMemberData`, then the analyzer will mark the pointer-to-member as unknown [this is done by `handleLVectorSplat`]
>I don't know if it's really pointless, but hard to model precisely for sure.
Really hard to model and has too many cases and ifs and buts. Like although the member-pointer obtained by `reinterpret_cast` probably has the wrong offset, we still don't get seg-fault. Curious, but seems to happen for both GCC and Clang.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D96976/new/

https://reviews.llvm.org/D96976