[PATCH] D14358: DWARF's forward decl of a template should have template parameters.

[Robinson, Paul via cfe-commits]

| Why is matching by name insufficient/not correct?
I'm told we look at the mangled names in the ELF symbol table, demangle them, and look in the DWARF for the corresponding types.  Now, the mangled name (for predefined types in particular) provides a type description, not the name-as-emitted-by-Clang, and in fact the same type can have more than one name ('const int' and 'int const' for a trivial example).  The name Clang provides in the DWARF does not necessarily match the name produced by the demangler; this makes name-matching way more trouble than you'd think.  We're not interested in teaching the debugger how to parse template instantiation names.
Having the template type parameter means we have an unambiguous description of the type, and can match it easily.

| including unreferenced entities fails source fidelity
I'll assume you meant to say _excluding_ unreferenced entities fails source fidelity, which is quite true, but there is a valid engineering tradeoff in that what the DWARF actually contains (or not, in the case of, say, unused function declarations or unreferenced class contents) represents one possible valid source that could have produced the same object.  (I'm curious why an unreferenced formal parameter of a function still gets described, if this is your argument for omitting template parameters.)
Omitting template parameters however is not the same as omitting unreferenced entities, because the template parameters *are* referenced—by the template instantiation itself; and, omitting them from the source does not produce a valid program.  Now, one of the 3 debuggers Clang explicitly supports (i.e. gdb) seems not to mind that they're missing, but the other two would benefit from having these things, and I would really like to have Clang produce these things.

Thanks,
--paulr

From: David Blaikie [mailto:dblaikie at gmail.com]
Sent: Monday, November 09, 2015 1:46 PM
To: Robinson, Paul
Cc: reviews+D14358+public+d3104135076f0a10 at reviews.llvm.org; cfe-commits (cfe-commits at lists.llvm.org)
Subject: Re: [PATCH] D14358: DWARF's forward decl of a template should have template parameters.



On Thu, Nov 5, 2015 at 11:05 AM, Robinson, Paul <Paul_Robinson at playstation.sony.com<mailto:Paul_Robinson at playstation.sony.com>> wrote:
| What was your primary motivation?
A similar concern to PR20455 from our own debugger.  It much helps matching up the forward declaration and definition to have the parameters properly specified.

Why is matching by name insufficient/not correct?


| maybe it's possible to remangle the template using just the string name
I have no idea what you're talking about here.

Looking at PR20455 you linked, LLDB isn't finding the right function because of mangling:


call to a function 'basic_string<char, char_traits<char> >::operator[](int) const' ('_ZNK12basic_stringIc17char_traits<char>EixEi') that is not present in the target
It hasn't created the correct mangled name of operator[] - what I was saying is it might be possible to parse the template parameter from the pretty name, and use that to produce the mangled name. It /looks/ like GDB can manage this. Maybe only because we also include the mangled name of the member function? Not sure.

| | Choosing to emit a forward/incomplete declaration in the first place fails source fidelity,
| How so?
When the source has a full definition but Clang chooses to emit only the declaration, per CGDebugInfo.cpp/shouldOmitDefinition().

Sure, in the same way that including unreferenced entities fails source fidelity - all tradeoffs to reduce debug info size.

Though the behavior is visible in a simpler example that doesn't have that failing (& if your change goes in, the test case should probably be simplified like this):

template<typename T> struct foo;
foo<int> *f;

--paulr

From: David Blaikie [mailto:dblaikie at gmail.com<mailto:dblaikie at gmail.com>]
Sent: Thursday, November 05, 2015 12:10 AM
To: Robinson, Paul
Cc: reviews+D14358+public+d3104135076f0a10 at reviews.llvm.org<mailto:reviews%2BD14358%2Bpublic%2Bd3104135076f0a10 at reviews.llvm.org>; cfe-commits (cfe-commits at lists.llvm.org<mailto:cfe-commits at lists.llvm.org>)

Subject: Re: [PATCH] D14358: DWARF's forward decl of a template should have template parameters.



On Wed, Nov 4, 2015 at 11:32 PM, Robinson, Paul <Paul_Robinson at playstation.sony.com<mailto:Paul_Robinson at playstation.sony.com>> wrote:
Would citing PR20455 help?  It wasn't actually my primary motivation but it's not too far off.  Having the template parameters there lets you know what's going on in the DWARF, without having to fetch and parse the name string of every struct you come across.  Actually I'm not sure parsing the name string is unambiguous either; each parameter is either a typename, or an expression, but without the parameter DIEs you don't know which, a-priori.  (What does <foo> mean? Depends on whether you think it should be a type name or a value; you can't tell, syntactically, you have to do some lookups.  Ah, but if you had the parameter DIEs, you would Just Know.)

For LLDB's needs, I'm not sure it's sufficient either - but I wouldn't mind an answer before we use it as the basis for this change (it sounds like maybe it's possible to remangle the template using just the string name, rather than needing an explicit representation of the parameters)

What was your primary motivation?

 Choosing to emit a forward/incomplete declaration in the first place fails source fidelity,

How so? You might have only a template declaration (template<typename T> struct foo; foo<int> *f;) or you may've only instantiated the declaration (the C++ language requires you to instantiate or avoid instantiating certain things in certain places, so in some contexts you /only/ have an instantiated declaration, not a definition)

but it is a practical engineering tradeoff of compile/link performance against utility; and, after all, the source *could* have been written that way, with no semantic difference.  But, if we're going to emit a white-lie incomplete declaration, we should do so correctly.

Again, "correct" in DWARF is a fairly nebulous concept.

--paulr

P.S. We should talk about this forward-declaration tactic wrt LTO sometime.  I have a case where a nested class got forward-declared; it's entirely conceivable that the outer class with the inner forward-declared class would end up being picked by LTO, leaving the user with no debug info for the inner class contents.

I believe this Just Works(tm). The things that can vary per-insntance of a type (implicit special members, member template implicit specializations, and nested types*) are not added to the type's child list, but they reference the child as their parent. So they continue to apply no matter which instance of the type is picked for uniquing (because of the name-based referencing, so the nested type definition just says "my parent is _Zfoo" and whatever _Zfoo we end up picking in the LTO linking/metadata deduplication will serve that role just fine)

* we could just do a better job of modelling nested types (& other non-globally scoped types) in a way that more closely models the source by emitting a declaration where they were declared, and a definition where they are defined (with the usual DW_AT_specification to wire them up)


From: David Blaikie [mailto:dblaikie at gmail.com<mailto:dblaikie at gmail.com>]
Sent: Wednesday, November 04, 2015 8:30 PM
To: reviews+D14358+public+d3104135076f0a10 at reviews.llvm.org<mailto:reviews%2BD14358%2Bpublic%2Bd3104135076f0a10 at reviews.llvm.org>; Robinson, Paul
Subject: Re: [PATCH] D14358: DWARF's forward decl of a template should have template parameters.



On Wed, Nov 4, 2015 at 5:53 PM, Paul Robinson via cfe-commits <cfe-commits at lists.llvm.org<mailto:cfe-commits at lists.llvm.org>> wrote:
probinson added a comment.

GCC 4.8.4 on Linux doesn't produce these, but DWARF 4 section 5.5.8 says a class template instantiation is just like the equivalent non-template class entry, with the exception of the template parameter entries.  I read that as meaning an incomplete description (i.e. with DW_AT_declaration) lets you omit all the other children, but not the template parameters.

As usual, I think it's pretty hard to argue that DWARF /requires/ anything (permissive & all that). And I'm not sure that having these is particularly valuable/useful - what use do you have in mind for them?

Wouldn't hurt to have some size info about the cost here, though I don't imagine it's massive, it does open us up to emitting a whole slew of new types (the types the template is instantiated with, and anything that depends on - breaking/avoiding type edges can, in my experience, be quite beneficial (I described an example of this in my lightning talk last week)).


I don't think omitting the template DIEs was an intentional optimization, in the sense of being a decision separate from deciding to emit the incomplete/forward declaration in the first place.  They were just omitted because we were omitting everything, but everything turns out to be non-compliant.


http://reviews.llvm.org/D14358



_______________________________________________
cfe-commits mailing list
cfe-commits at lists.llvm.org<mailto:cfe-commits at lists.llvm.org>
http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20151109/015f02aa/attachment-0001.html>