[llvm-dev] is_stmt and column numbers

[Robinson, Paul via llvm-dev]

> -----Original Message-----
> From: Roger Pack <rogerdpack2 at gmail.com>
> Sent: Friday, March 27, 2020 12:18 AM
> To: Robinson, Paul <paul.robinson at sony.com>
> Cc: llvm-dev at lists.llvm.org
> Subject: Re: [llvm-dev] is_stmt and column numbers
> 
> On Tue, Dec 17, 2019 at 8:35 AM Robinson, Paul <paul.robinson at sony.com>
> wrote:
> >
> >
> >
> > > -----Original Message-----
> > > From: llvm-dev <llvm-dev-bounces at lists.llvm.org> On Behalf Of Roger
> Pack
> > > via llvm-dev
> > > Sent: Tuesday, December 17, 2019 8:31 AM
> > > To: llvm-dev <llvm-dev at lists.llvm.org>
> > > Subject: [llvm-dev] is_stmt and column numbers
> > >
> > > It appears to me that from this:
> > >
> > > https://iis-git.ee.ethz.ch/H2020-
> > > Compiler/llvm/commit/1d68fc5021ca8c704e21f171d98bb18eb396a7fa
> > >
> > > That LLVM basically emits DWARF expressions where is_stmt is true only
> > > "once per line".  Is that still true?  Is there an option to emit more
> > > than one per line?
> > >
> > >
> > > Thanks all.
> > >
> > > -Roger Pack-
> > >
> > > refs:
> > > https://github.com/crystal-lang/crystal/pull/8499
> > > https://github.com/crystal-lang/crystal/issues/8319
> >
> > is_stmt is based fairly strictly on "the line number changed" (while
> > ignoring line 0).  This works reasonably well at -O0 and gets clunkier
> as
> > optimization level increases.  I've never been happy about it, but have
> > never found the time to come up with something better.  The optimizer
> has
> > been getting better at not introducing gratuitous line-number changes,
> if
> > that helps.
> >
> > There is no option to make it work differently.  We can (usually do)
> emit
> > column numbers, which gets you more line-table entries, although they
> are
> > not flagged with is_stmt.  I know at least one debugger that gave up on
> > is_stmt "because it's not reliable."
> > --paulr
> 
> OK that makes sense.  I've noticed that there'll be some DWARF
> statements that set the line number to "0".
> 
>   [0x0000048c]  Special opcode 72: advance Address by 5 to 0x1000015b0
> and Line by -3 to 0
>   [0x0000048d]  Special opcode 78: advance Address by 5 to 0x1000015b5
> and Line by 3 to 3
> 
> I assume this means some kind of internal processing?  Anyway the
> current work around crystal uses is to basically only accept DWARF
> statements that have a positive line number [not sure if that's right
> or not...], or if they're is_stmt then it accepts them :)
> Thanks!

The DWARF spec says line 0 means there is no relevant source location.
Sometimes it means an optimization combined operations from multiple 
source locations, and it would be misleading to pick one arbitrarily; 
sometimes it means the instruction came from an "artificial" bit of 
source, for example a default constructor in C++, that really has no 
worthwhile source location.

I can't imagine a correct producer ever setting is_stmt for line 0.
--paulr