r192312 - Initialize vtorDisp in class constructors and destructors

[Warren Hunt]

In all of the tests I have run the vtorDisp is at -4.  RecordlayoutBuilder
puts it there (there are weird padding rules when vtordisps are present,
but the location of a vtordisp with respect to its virtual base has always
been constant to my testing).  If you have any counterexamples I would love
to see them!

-Warren


On Thu, Oct 10, 2013 at 7:52 AM, Timur Iskhodzhanov <timurrrr at google.com>wrote:

> 2013/10/9 Timur Iskhodzhanov <timurrrr at google.com>:
> > Author: timurrrr
> > Date: Wed Oct  9 13:16:58 2013
> > New Revision: 192312
> >
> > URL: http://llvm.org/viewvc/llvm-project?rev=192312&view=rev
> > Log:
> > Initialize vtorDisp in class constructors and destructors
> >
> > Reviewed at http://llvm-reviews.chandlerc.com/D1867
> >
> > ...
> > Modified: cfe/trunk/lib/CodeGen/MicrosoftCXXABI.cpp
> > URL:
> http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/CodeGen/MicrosoftCXXABI.cpp?rev=192312&r1=192311&r2=192312&view=diff
> >
> ==============================================================================
> > --- cfe/trunk/lib/CodeGen/MicrosoftCXXABI.cpp (original)
> > +++ cfe/trunk/lib/CodeGen/MicrosoftCXXABI.cpp Wed Oct  9 13:16:58 2013
> > @@ -456,6 +459,61 @@ MicrosoftCXXABI::EmitCtorCompleteObjectH
> >    return SkipVbaseCtorsBB;
> >  }
> >
> > +void MicrosoftCXXABI::initializeHiddenVirtualInheritanceMembers(
> > +    CodeGenFunction &CGF, const CXXRecordDecl *RD) {
> > +  // In most cases, an override for a vbase virtual method can adjust
> > +  // the "this" parameter by applying a constant offset.
> > +  // However, this is not enough while a constructor or a destructor of
> some
> > +  // class X is being executed if all the following conditions are met:
> > +  //  - X has virtual bases, (1)
> > +  //  - X overrides a virtual method M of a vbase Y, (2)
> > +  //  - X itself is a vbase of the most derived class.
> > +  //
> > +  // If (1) and (2) are true, the vtorDisp for vbase Y is a hidden
> member of X
> > +  // which holds the extra amount of "this" adjustment we must do when
> we use
> > +  // the X vftables (i.e. during X ctor or dtor).
> > +  // Outside the ctors and dtors, the values of vtorDisps are zero.
> > +
> > +  const ASTRecordLayout &Layout = getContext().getASTRecordLayout(RD);
> > +  typedef ASTRecordLayout::VBaseOffsetsMapTy VBOffsets;
> > +  const VBOffsets &VBaseMap = Layout.getVBaseOffsetsMap();
> > +  CGBuilderTy &Builder = CGF.Builder;
> > +
> > +  unsigned AS =
> > +
>  cast<llvm::PointerType>(getThisValue(CGF)->getType())->getAddressSpace();
> > +  llvm::Value *Int8This = 0;  // Initialize lazily.
> > +
> > +  for (VBOffsets::const_iterator I = VBaseMap.begin(), E =
> VBaseMap.end();
> > +        I != E; ++I) {
> > +    if (!I->second.hasVtorDisp())
> > +      continue;
> > +
> > +    llvm::Value *VBaseOffset =
> CGM.getCXXABI().GetVirtualBaseClassOffset(
> > +        CGF, getThisValue(CGF), RD, I->first);
> > +    // FIXME: it doesn't look right that we SExt in
> GetVirtualBaseClassOffset()
> > +    // just to Trunc back immediately.
> > +    VBaseOffset = Builder.CreateTruncOrBitCast(VBaseOffset,
> CGF.Int32Ty);
> > +    uint64_t ConstantVBaseOffset =
> > +        Layout.getVBaseClassOffset(I->first).getQuantity();
> > +
> > +    // vtorDisp_for_vbase = vbptr[vbase_idx] - offsetof(RD, vbase).
> > +    llvm::Value *VtorDispValue = Builder.CreateSub(
> > +        VBaseOffset, llvm::ConstantInt::get(CGM.Int32Ty,
> ConstantVBaseOffset),
> > +        "vtordisp.value");
> > +
> > +    if (!Int8This)
> > +      Int8This = Builder.CreateBitCast(getThisValue(CGF),
> > +                                       CGF.Int8Ty->getPointerTo(AS));
> > +    llvm::Value *VtorDispPtr = Builder.CreateInBoundsGEP(Int8This,
> VBaseOffset);
> > +    // vtorDisp is always the 32-bits before the vbase in the class
> layout.
> > +    VtorDispPtr = Builder.CreateConstGEP1_32(VtorDispPtr, -4);
>
> After investigating some other tests, I'm pretty sure the "-4"
> vtordisp offset is not constant but rather might have a different
> value for more complex record layouts.
>
> Warren, do you have any insight to share on the vtordisp location in a
> record layout?
>
> > +    VtorDispPtr = Builder.CreateBitCast(
> > +        VtorDispPtr, CGF.Int32Ty->getPointerTo(AS), "vtordisp.ptr");
> > +
> > +    Builder.CreateStore(VtorDispValue, VtorDispPtr);
> > +  }
> > +}
> > +
> >  void MicrosoftCXXABI::EmitCXXConstructors(const CXXConstructorDecl *D) {
> >    // There's only one constructor type in this ABI.
> >    CGM.EmitGlobal(GlobalDecl(D, Ctor_Complete));
> >
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20131010/4e96fbd2/attachment.html>