r180603 - C++1y: support simple variable assignments in constexpr functions.

[Richard Smith]

On Sun, May 5, 2013 at 8:45 AM, Abramo Bagnara
<abramo.bagnara at bugseng.com>wrote:

> Il 05/05/2013 17:05, Richard Smith ha scritto:
> > On Sun, May 5, 2013 at 2:53 AM, Abramo Bagnara
> > <abramo.bagnara at bugseng.com <mailto:abramo.bagnara at bugseng.com>> wrote:
> >
> >     Il 26/04/2013 16:36, Richard Smith ha scritto:
> >     > Modified: cfe/trunk/lib/Sema/SemaInit.cpp
> >     > URL:
> >
> http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/Sema/SemaInit.cpp?rev=180603&r1=180602&r2=180603&view=diff
> >     >
> >
> ==============================================================================
> >     > --- cfe/trunk/lib/Sema/SemaInit.cpp (original)
> >     > +++ cfe/trunk/lib/Sema/SemaInit.cpp Fri Apr 26 09:36:30 2013
> >     > @@ -97,6 +97,7 @@ static void CheckStringInit(Expr *Str, Q
> >     >      DeclT = S.Context.getConstantArrayType(IAT->getElementType(),
> >     >                                             ConstVal,
> >     >                                             ArrayType::Normal, 0);
> >     > +    Str->setType(DeclT);
> >     >      return;
> >     >    }
> >
> >     Is this change deliberate? It seems to introduce a regression:
> >
> >     $ cat z.c
> >
> >     void f() {
> >       signed char s[] = "a";
> >       unsigned char u[] = "a";
> >     }
> >     $ _clang -cc1 -ast-dump z.c
> >     TranslationUnitDecl 0x67d96e0 <<invalid sloc>>
> >     |-TypedefDecl 0x67d9bc0 <<invalid sloc>> __int128_t '__int128'
> >     |-TypedefDecl 0x67d9c20 <<invalid sloc>> __uint128_t 'unsigned
> __int128'
> >     |-TypedefDecl 0x67d9f70 <<invalid sloc>> __builtin_va_list
> >     '__va_list_tag [1]'
> >     `-FunctionDecl 0x67da010 <z.c:2:1, line:5:1> f 'void ()'
> >       `-CompoundStmt 0x67da350 <line:2:10, line:5:1>
> >         |-DeclStmt 0x67da208 <line:3:3, col:24>
> >         | `-VarDecl 0x67da100 <col:3, col:21> s 'signed char [2]'
> >         |   `-StringLiteral 0x67da198 <col:21> 'signed char [2]' lvalue
> "a"
> >         `-DeclStmt 0x67da338 <line:4:3, col:26>
> >           `-VarDecl 0x67da270 <col:3, col:23> u 'unsigned char [2]'
> >             `-StringLiteral 0x67da2c8 <col:23> 'unsigned char [2]'
> >     lvalue "a"
> >
> >     Type of string literal should be plain char.
> >
> >
> > Yes, this is deliberate; we intended to set the string literal's type to
> > the type of the initialized variable (otherwise we would be initializing
> > an array of 'unsigned char' from an array of 'char'), but accidentally
> > only updated it in either the array bound or the type, but not both.
> >
> >   unsigned char a[] = "foo", b[4] = "bar";
> >
> > ... used to produce ...
> >
> > |-VarDecl 0x6222ad0 <<stdin>:1:1, col:21> a 'unsigned char [4]'
> > | `-StringLiteral 0x6222ba8 <col:21> 'const char [4]' lvalue "foo"
> > `-VarDecl 0x6222c60 <col:1, col:35> b 'unsigned char [4]'
> >   `-StringLiteral 0x6222cb8 <col:35> 'unsigned char [4]' lvalue "bar"
>
> BTW: uniformity is still not there:
> $ cat z.c
> void f() {
>   unsigned char q[] = ("a");
> }
> $ _clang -cc1 -ast-dump z.c
> TranslationUnitDecl 0x6f296e0 <<invalid sloc>>
> |-TypedefDecl 0x6f29bc0 <<invalid sloc>> __int128_t '__int128'
> |-TypedefDecl 0x6f29c20 <<invalid sloc>> __uint128_t 'unsigned __int128'
> |-TypedefDecl 0x6f29f70 <<invalid sloc>> __builtin_va_list
> '__va_list_tag [1]'
> `-FunctionDecl 0x6f2a010 <z.c:2:1, line:4:1> f 'void ()'
>   `-CompoundStmt 0x6f2a240 <line:2:10, line:4:1>
>     `-DeclStmt 0x6f2a228 <line:3:3, col:28>
>       `-VarDecl 0x6f2a100 <col:3, col:27> q 'unsigned char [2]'
>         `-ParenExpr 0x6f2a1c8 <col:23, col:27> 'unsigned char [2]' lvalue
>           `-StringLiteral 0x6f2a198 <col:24> 'char [2]' lvalue "a"
>

Yes, this is definitely wrong. A ParenExpr shouldn't change the type of its
operand.

IMHO the type of StringLiteral and ParenExpr should be char[] and as the
> use of string literal for array initialization is definitely a special
> case, nothing is wrong if its type is the one mandated by the standard
> and different from initialized decl (i.e. "initialize this unsigned char
> array from this byte sequence that is plain char typed")


There are several problems with that approach (or one problem viewed in
several different ways):
 - it introduces a change in type with no AST node performing the conversion
 - the StringLiteral expression actually *is* creating an object of type
unsigned char[2] in this case
 - in C++, the standard-mandated type is *const* char[N], but such a
StringLiteral can still be the value of a non-const array

Analogous to how we handle InitListExprs, we could keep separate syntactic
and semantic forms for a StringLiteral used as an initializer (or maybe
just separate syntactic and semantic types).
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20130505/f04b6213/attachment.html>