llvm dev - Nov 2016 - BoF: Debug info for optimized code.

Debugging optimized code is a topic that generated a lot of
interest at previous dev meetings, but often fell somewhat short
since we've all been very busy with improving the compile-time
performance impact of debug info. This year we'd like to make up
for this by inviting to a dedicated session on just "-O" and
"-g"!

We will be giving a quick introduction to summarize the current
state of debug info handling in LLVM and highlight some of the
biggest problem areas.  To get a productive discussion going,
here are some points we'd like to invite you to think about ahead
of time:

* What are the major pain points that you or your customers
  experience? What works — what doesn't?

* Do you have out-of-tree patches that improve the optimized
  debugging experience? What are your experiences with it and
  what are the challenges in upstreaming them?

* We would really like to start tracking debug info quality the
  same way we track code size and performance numbers. What are
  useful metrics for your use-case?

See you all at the dev meeting!

Adrian Prantl
Fred Riss

I would love to be able to attend this BoF, but unfortunately I cannot.  I would
however like to add a couple of things for consideration during this discussion.

Our custom processor is a VLIW processor, and I think that there is a growing
trend in CPU design toward VLIW as transistor count increases, but clock speed
does not.  When we compile at '-O0' we do not use more than one
functional-unit at a time (except for the predication unit plus one other
because we have to).  This means that each "instruction" corresponds
to a particular line of code, and an assembly '.loc' directive can be
emitted for that instruction which produces a reasonably straight-forward
correspondence between the instruction and the source code location.

Optimisation for scalar processors introduces all sorts of fun - code-motion,
code-duplication, code-elimination, you name it - but even so, an instruction
generally corresponds to some line of code in the source even if it is not
obvious to the person debugging their code why the line-tracking jumps around
all the time.

But VLIW optimisation introduces a new twist in that each of the
functional-units can correspond to different lines of code, and this is really
very hard to understand in optimised code when debugging.  The more functional
units the VLIW architecture supports, the more difficult this becomes.  And to
complicate this further, predication units can control some functional-units and
not others, so a single instruction may contain elements which are conditional
and other elements which are not, and of course corresponding to several
different lines of code.

I do not even pretend to know much about Dwarf and the representation of debug
information, but it does appear that there is little or no support for the idea
that a single "instruction" can correspond to multiple diverse lines
in the source file.  It would be useful for an assembly '.loc' directive
to take a list of source locations, and for the debug meta-data to represent
this one-to-many mapping ('.debug_{[a]ranges|lines}' etc.).  At the
moment we are limited to picking one of the several possible locations, and
hoping that it is helpful to the programmer when debugging their optimised
programs.

Perhaps this is simply my misunderstanding of the degree of meta-data support in
Dwarf, but I don't see any obvious way of representing the one-to-many
mapping in LLVM for bundles, or the selective predication of the elements in the
bundle.  This kind of bundling happens very late in the MI level, and by then
the DebugLoc abstractions are getting quite difficult.  For instance, how do I
tell LLVM that some of the instructions in a bundle that contain a predicate are
predicated by that predicate, and while instructions are always active?  And how
does this pass on to the debug meta-data?  The Dwarf support in LLVM is
(thankfully) largely independent of the target, but it still needs to have
support for these abstractions and use-cases.

Another area that is really very painful for programmers, whether VLIW or
scalar, is C++ templates.  This is an enormously complex area to debug, in
particular because the optimised program usually involves extensive use of
inlining and multiple tiers of template specialisations.  Even reasoning about
the specialisation, whether at '-O0' or fully optimised is incredibly
hard.

I don't yet have out-of-tree contributions to make, because I have to admit
I struggle with this too, but I have some ideas which I would like to promote in
the future once I have ironed out the details.  But it would be good to have the
topic of VLIW and C++ templates added to the hot-topics for the BoF.

Thanks and I am looking forward to reading about the outcome from this BoF,

	MartinO

-----Original Message-----
From: llvm-dev [mailto:llvm-dev-bounces at lists.llvm.org] On Behalf Of Adrian
Prantl via llvm-dev
Sent: 30 October 2016 15:54
To: LLVM Dev <llvm-dev at lists.llvm.org>
Subject: [llvm-dev] BoF: Debug info for optimized code.

Debugging optimized code is a topic that generated a lot of interest at previous
dev meetings, but often fell somewhat short since we've all been very busy
with improving the compile-time performance impact of debug info. This year
we'd like to make up for this by inviting to a dedicated session on just
"-O" and "-g"!

We will be giving a quick introduction to summarize the current state of debug
info handling in LLVM and highlight some of the biggest problem areas.  To get a
productive discussion going, here are some points we'd like to invite you to
think about ahead of time:

* What are the major pain points that you or your customers
  experience? What works — what doesn't?

* Do you have out-of-tree patches that improve the optimized
  debugging experience? What are your experiences with it and
  what are the challenges in upstreaming them?

* We would really like to start tracking debug info quality the
  same way we track code size and performance numbers. What are
  useful metrics for your use-case?

See you all at the dev meeting!

Adrian Prantl
Fred Riss
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

On 11/1/2016 4:28 PM, Martin J. O'Riordan via llvm-dev
wrote:
> I do not even pretend to know much about Dwarf and the representation of
debug information, but it does appear that there is little or no support for the
idea that a single "instruction" can correspond to multiple diverse
lines in the source file.
There is.  There is even a patch for LLVM:
https://reviews.llvm.org/D16697

-Krzysztof

-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, 
hosted by The Linux Foundation