[PATCH] D19654: PR27132: Proper mangling for __unaligned qualifier (now with PR27367 fixed)

[Andrey Bokhanko via cfe-commits]

andreybokhanko marked an inline comment as done.

================
Comment at: lib/AST/MicrosoftMangle.cpp:1583-1584
@@ -1579,2 +1582,4 @@
   case QMM_Result:
+    // Presence of __unaligned qualifier shouldn't affect mangling here.
+    Quals.removeUnaligned();
     if ((!IsPointer && Quals) || isa<TagType>(T)) {
----------------
majnemer wrote:
> majnemer wrote:
> > rnk wrote:
> > > majnemer wrote:
> > > > andreybokhanko wrote:
> > > > > majnemer wrote:
> > > > > > andreybokhanko wrote:
> > > > > > > Done. Test added.
> > > > > > Hmm, can you give a concrete example why we need this line?
> > > > > Sure. An example is:
> > > > > 
> > > > > __unaligned int unaligned_foo3() { return 0; }
> > > > > 
> > > > > MS mangles it as
> > > > > 
> > > > > ?unaligned_foo3@@YAHXZ
> > > > > 
> > > > > However, if __unaligned is taken into account, "if ((!IsPointer && Quals) || isa<TagType>(T))" computes to true and clang adds "?A", resulting to
> > > > > 
> > > > > ?unaligned_foo3@@YA?AHXZ
> > > > > 
> > > > > Yours,
> > > > > Andrey
> > > > > 
> > > > Wait, I thought __unaligned can only apply to pointer types.  Is this not so?!
> > > > Does `__unaligned int x;` really keep it's `__unaligned` qualifier?
> > > Yeah it does:
> > >   $ cat t.cpp
> > >   __unaligned int x;
> > >   $ cl -nologo -c t.cpp && dumpbin /symbols t.obj  | grep ?x
> > >   t.cpp
> > >   008 00000000 SECT3  notype       External     | ?x@@3HFA (int __unaligned x)
> > Woah.  So if you do:
> > 
> > > __unaligned int unaligned_foo3() { return 0; }
> > > auto z = foo3();
> > 
> > How is `z` mangled?
> `z` is mangled without the qualifiers.  In fact:
> 
> ```
> __unaligned int unaligned_foo3() { return 0; }
> __unaligned int z;
> auto z = unaligned_foo3();
> ```
> 
> Is an error:
> 
> 
> > x.cpp(3): error C2373: 'z': redefinition; different type modifiers
> > x.cpp(2): note: see declaration of 'z'
> 
> Do we have comparable behavior?
Not exactly the same, but comparable, indeed:

```
$ clang -cc1 ~/test.cpp -std=c++11 -triple i686-pc-win32 -fms-extensions
/nfs/ims/home/asbokhan/test.cpp:3:6: error: redefinition of 'z'
auto z = unaligned_foo3();
     ^
/nfs/ims/home/asbokhan/test.cpp:2:17: note: previous definition is here
__unaligned int z;
                ^
1 error generated.
```

This has nothing to do with __unaligned, though, as both MS compiler and us don't allow redefinitions. If a redefinition has a different __unaligned modifier, MS also notes this (with "; different type modifiers" suffix in the message), but it doesn't serve any practical purpose, as changing modifier won't make the code compilable.

Yours,
Andrey


http://reviews.llvm.org/D19654