Let me show you a snippet of a lldb debug session in progress in my ObjC
variant:
-100000,100000,v,18.48
Process 45774 stopped
* thread #1, queue = 'com.apple.main-thread', stop reason = step in
frame #0: 0x0000000100000e2a multiple.debug`+[Foo
long:int:char:float:](self=Foo, _cmd=<no value available>,
_param=0x00007fff5fbff948) at multiple.m:15
12 char:(char) c
13 float:(float) d
14 {
15 printf( "%ld,%d,%c,%.2f\n", a, b, c, d);
-> 16 }
17
18 @end
(lldb) p *_param
(p.long:int:char:float:) $2 = (a = -100000, b = 100000, c = 'v', d =
18.4799995)
You can see that the parameter values `a,b,c,d` are actually fields of
a struct parameter `_param`. `_param` uniformly appears as the third
parameter after `self` and `_cmd`. `p _param->a` works of course, but it
would be nice to be able to say 'p a', since in the source code one sees
only `a`. `_param` is more or less an implementation detail.
A clue how to achieve this, would be very much appreciated.
Ciao
Nat!
[*] except, if it's a picture of thousand words 
Because we try as much as possible to let the compiler figure this sort of thing out for us, we implement the transparent lookups for this & self by compiling our expression in a context that poses as a method of the class whose method we are stopped in. For instance, for ObjC, we construct a category on the class, and put the expression text in a method in that category, and then compile all that and call the method.
That's done in ExpressionSourceCode::GetText.
The one hitch to this is that this works because the compiler that is linked into lldb knows about the ObjC model we are emulating. An unmodified clang won't do transparent lookup into some random argument that happens to be called _param. So you would have to build lldb with a version of clang that understands your lookup rules for this to work.
If that's not possible then you can try to do this by monkeying with the lookup rules implemented by lldb's ClangASTSource to return "_param.a" when it is just looking up "a". That's how we inject variables and types that are present in the local context into the expression as it is getting parsed. But at that point you are getting your hands into some fairly deep magic, and clang's Aslan is not as forgiving as Narnia's...
Jim
I think the cleanest solution would be to actually modify the compiler
to emit "correct" dwarf (as in, dwarf representing the code as it
actually looks like to user, and not some internal intermediate
representation). The dwarf expression in DW_AT_location can easily
handle the lookup of some field in a struct. Of course, this is
assuming you are actually able to modify the compiler.
(disclaimer: I know next to nothing, when it comes to objc)
Pavel Labath schrieb:
I think the cleanest solution would be to actually modify the compiler
to emit "correct" dwarf (as in, dwarf representing the code as it
actually looks like to user, and not some internal intermediate
representation). The dwarf expression in DW_AT_location can easily
handle the lookup of some field in a struct. Of course, this is
assuming you are actually able to modify the compiler.
(disclaimer: I know next to nothing, when it comes to objc)
That would probably be ideal. I have to see if I can do this. But my
metaABI is somewhat "hacked" into clang. So I am not saying it wouldn't
work, but I suspect it won't, as I am relying on existing clang code.
Ciao
Nat!
Pavel, can you say more about your idea?
In both ObjC and C++ methods, you can refer to an ivar either as "this->ivar" or just "ivar". But the DWARF for C++ doesn't tell lldb that a particular language supports referring to ivars transparently this way. It will say that the "this" parameter is artificial, and that it is the "object pointer". But it doesn't so far as I can tell record the fact that elements of that parameter can be transparently accessed.
I think it would be confusing for the debug information to record the transparently accessed ivars as pseudo-locals, the duplication would confuse folks, and that isn't how they are understood by the person writing the code. It might be good to propose a "DW_AT_transparent" attribute, and mark the ivars or maybe the parameter that way. I guessing that wasn't done because it was assumed that the debugger would know this sort of rule from the language in question.
As I understand it, problem here is that Nat's runtime has extra rules for transparent access that lldb doesn't know about.
Jim
I understood the problem differently, although I can't really say
whether it was correct or not. My impression was that the members of
the "_param" struct were **parameters** to the method, which have
become (due to some source code transformations, or ABI
considerations) members of a single struct parameter. So a programmer
would write (i'm not going to try to write this in objc)
int foo(int x, int y) { return x+y; }
and the some frontend would transform this into:
struct foo_params {int x,y;};
int foo(foo_params *_param) { return _param->x + _param->y; }
before passing it into the real compiler.
If this is the case then I think it makes sense to have the dwarf
describe the original source code, and not the intermediate form.
However, if it's something else, then that's likely not a good idea,
and I should probably shut up, as I obviously don't know enough about
objc to be of help here.
cheers,
pl
Pavel, this is pretty much exactly what happens. (Except that there is no frontend. The substitution is done during CodeGen. But that is just a detail)
I think a more general solution, that doesn't tie mulle-objc implicitly to the dwarf format would be better in the long term though.
Ciao
Nat!