[cfe-dev] -fspell-checking causing modifications to AST?
richard at metafoo.co.uk
Mon Jul 14 17:52:42 PDT 2014
On Mon, Jul 14, 2014 at 4:53 PM, Kevin Funk <kfunk at kde.org> wrote:
> I'm a bit puzzled by the following behavior of clang when inspecting the
> for the following code snippet:
> char foo;
> char bar = foo1;
> $ clang++ -cc1 -ast-dump test.cpp:
> main.cpp:2:12: error: use of undeclared identifier 'foo1'; did you mean
> `-VarDecl 0xa19cb0 <line:2:1, col:12> col:6 bar 'char' cinit
> `-ImplicitCastExpr 0xa19d60 <col:12> 'char' <LValueToRValue>
> `-DeclRefExpr 0xa19d38 <col:12> 'char' lvalue Var 0xa19c40 'foo' 'char'
> So, Clang seems to interpret 'foo1' as a typo 'foo'. This is still fine.
> However, Clang also "fixes up the code" and pretends that 'foo1' *is* 'foo'
> which -ast-dump shows (see last line of the dump). This is odd.
> Obviously, this is only the case if -fno-spell-checking is *not* passed to
> clang++. With -fno-spell-checking I get this instead:
> $ clang++ -cc1 -ast-dump test.cpp -fno-spell-checking
> `-VarDecl 0x1a79ca0 <line:2:1, col:6> col:6 bar 'char'
> This looks fine to me.
> So, my question: Is this intended behavior? Is this for error recovery in
Yes, this is intended. It is our policy that whenever we emit an error
message with a fix-it hint, we recover as if that fix-it hint had been
applied to the AST. This is not limited to typo-correction; it happens
whenever we emit an error with a fix-it.
If true, is there a way to both enable spell-checking but to *disable* it
> touching the AST?
Not currently, no. I think what you're really asking for is for us to turn
off all error recovery, so the AST contains only things that were in the
source code, and omits erroneous things rather than trying to recover from
errors? We could try to support such a mode, but I suspect that it would
degrade the diagnostic experience enough that you may want to run Clang
twice: once in this mode, and once to produce user-facing diagnostics.
Also, I don't expect that you'll find volunteers in the Clang community to
do the work, but if the design and rationale are reasonable and you can
provide a convincing argument that the mode will have sufficient ongoing
support work, we'd accept patches to implement this.
Reminder: I'm working on integrating Clang (read: libclang) in KDevelop,
> I'm looking at this from an development tool perspective. Magic behavior
> inside Clang is somewhat undesirable there :)
Can you explain a bit more about your use cases? I assume you want to
provide some kind of AST introspection on not-necessarily-valid code, and
you don't want to expose our typo-correction results in that? Depending on
what you're trying to achieve, there might be lighter-weight approaches
(for instance, checking if the token at the given location actually refers
to the right name).
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the cfe-dev