[PATCH] Avoid empty .debug_loc entries and lists

[David Blaikie]

On Thu, Jun 4, 2015 at 4:46 PM, Duncan P. N. Exon Smith <
dexonsmith at apple.com> wrote:

>
> > On 2015 Jun 4, at 16:36, David Blaikie <dblaikie at gmail.com> wrote:
> >
> >
> >
> > On Thu, Jun 4, 2015 at 4:28 PM, Duncan P. N. Exon Smith <
> dexonsmith at apple.com> wrote:
> >
> > > On 2015 Jun 4, at 16:09, David Blaikie <dblaikie at gmail.com> wrote:
> > >
> > > Judging by the code, does this cause us not to create a variable if it
> has no locations? That's probably not quite right - even if we have no
> locations, it's probably a good idea/correct to describe the variable so
> shadowing, etc, works correctly (if someone tries to print the variable the
> compiler still finds the right variable but just reports that it doesn't
> know the location)
> >
> > So the code here already has an early return if there are no ranges:
> >
> >     for (const auto &I : DbgValues) {
> >       InlinedVariable IV = I.first;
> >       if (Processed.count(IV))
> >         continue;
> >
> >       // Instruction ranges, specifying where IV is accessible.
> >       const auto &Ranges = I.second;
> >       if (Ranges.empty())
> >         continue;
> >
> > My patch effectively hits this `continue` more often.  The difference
> > is cases where we *thought* we knew how to emit ranges, but in fact
> > didn't know how.
> >
> > I suppose I assumed the intended behaviour was "don't emit the
> > variable if we don't have any ranges for it" since that's what was
> > going on, but are you saying this `continue` is a bug?  (Or am I
> > totally missing your point?)
> >
> > How do we end up with an empty range list here? That seems strange - I
> assume we're just lazily creating the rang lists in
> calculateDbgValueHistory... (I could sort of understand empty entries in a
> range list - when it turns out the dbg_value intrinsic describes no
> instructions due to things being optimized away, hoisted here or there,
> etc).
>
> No idea, I just assume it happens because there's code.
>

Not sure what you mean.

What I'm trying to understand is whether or not we should be fixing this
bug closer to the source (in calculateDbgValueHistory) that's producing
empty ranges or empty entries.


> What's the right thing to do?


I don't have all the context here to know that off-hand. Trying to discuss
it to try to wrap my head around enough of it to offer a useful opinion -
sorry it takes a while.


> Should we create a variable the same
> was as the for loop below for function arguments (that doesn't
> attach ranges)?  Should we not create a variable?


If we get the same output between creating it here or in the later loop, it
doesn't matter in the broad sense.


> Is it actually
> *useful* that we create a .debug_loc list with no entries and point
> to it?
>

No, there's no benefit to pointing to an empty loc list compared to just
having no DW_AT_location at all.

- David


>
> >
> > > On 2015 Jun 4, at 16:13, Adrian Prantl <aprantl at apple.com> wrote:
> > >
> > > Even worse, when it is a function argument, omitting the variable will
> effectively change the function signature. (Yes, there is also the
> subroutine type...).
> >
> > No, there's a separate loop that guarantees all the function arguments
> > get created.
> >
> >     // Collect info for variables that were optimized out.
> >     for (const DILocalVariable *DV : SP->getVariables()) {
> >       if (!Processed.insert(InlinedVariable(DV, nullptr)).second)
> >         continue;
> >       if (LexicalScope *Scope =
> LScopes.findLexicalScope(DV->getScope())) {
> >         ensureAbstractVariableIsCreatedIfScoped(InlinedVariable(DV,
> nullptr),
> >                                                 Scope->getScopeNode());
> >         ConcreteVariables.push_back(make_unique<DbgVariable>(
> >             DV, /* IA */ nullptr, /* Expr */ nullptr, this));
> >         InfoHolder.addScopeVariable(Scope,
> ConcreteVariables.back().get());
> >       }
> >     }
> >
> > I haven't touched this.
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150604/dfc4d28d/attachment.html>