llvm dev - Oct 2017 - Code coverage BoF - notes and updates

Hi Dean,

We didn't discuss using XRay instrumentation during the BoF but it is an
interesting idea (by the way, thanks for your talk about XRay internals!). XRay
provides the advantage of being able to turn profiling on and off, but I'm
not sure how the resulting data could be used.

The code coverage feature is highly dependent on the frontend's profile
counter placement. The mapping between counters and parts of the AST is used to
gather accurate information about regions within a line. For example, the
coverage tool can show you that the l.h.s of "true || false" is
evaluated once, and the r.h.s isn't evaluated. This works with arbitrarily
nested short-circuit operators.

It might be possible to use XRay instrumentation to gather profile data, but I
think it will be challenging to precisely map that data back to the AST nodes
the frontend knows about. The problem is similar to the one I've outlined in
section 3 ("Optimizing profile counter placement"). The idea there is
to map a minimal set of counters placed by IR PGO back to AST nodes: the one
sketch of a solution I have still depends on running the frontend counter
placement pass to achieve this.

What are your thoughts on this?

thanks,
vedant
> On Oct 24, 2017, at 1:53 PM, Dean Michael Berris <dean.berris at
gmail.com> wrote:
> 
> Thanks for the summary Vedant!
> 
> I'm sorry I missed this BoF session.
> 
> Others have mentioned the possibility of maybe using XRay for some of this
information (function-level coverage, maybe having more intrinsics for marking
branch/basic-block level instrumentation). Was this explored in the BoF? Is
there interest in potentially exploring this particular space?
> 
> Cheers
> 
>> On 25 Oct 2017, at 07:24, Vedant Kumar via llvm-dev <llvm-dev at
lists.llvm.org <mailto:llvm-dev at lists.llvm.org>> wrote:
>> 
>> Hello,
>> 
>> Our goals for the code coverage BoF (10/19) were to find areas where we
can improve the coverage tooling, and to learn more about how coverage is used.
I'd like to thank all of the attendees for their input and for making the
BoF productive. Special thanks to Mandeep Grang, who volunteered as a mic runner
at the last minute.
>> 
>> In this email I'll share my (rough) notes and outline some future
plans. Please feel free to ask for clarifications or to add your own notes.
>> 
>> Here are the slides from the BoF:
>>
https://docs.google.com/presentation/d/e/2PACX-1vS-rV02j1zhPq9Y6AtcUkbZW2c7Q5YYuQ6FPxN-aYiKwrw6c8DU3zW_RYeJlWPMZ5-S6hgz_CIcL8Gd/pub?start=false&loop=false&delayms=3000&slide=id.p
<https://docs.google.com/presentation/d/e/2PACX-1vS-rV02j1zhPq9Y6AtcUkbZW2c7Q5YYuQ6FPxN-aYiKwrw6c8DU3zW_RYeJlWPMZ5-S6hgz_CIcL8Gd/pub?start=false&loop=false&delayms=3000&slide=id.p>
>> 
>> 1. The header problem
>> 
>> Coverage instrumentation overhead is roughly quadratic in the number of
translation units in a project. The problem is that coverage mappings for
template instantiations and static inline functions from headers are pulled into
every TU. This bloats the profile metadata sections (which can slow down profile
I/O), results in large binaries, and causes long link times (or link failures).
>> 
>> We could solve this problem by maintaining an external coverage
database and discarding duplicate coverage mappings from the DB. Another idea is
to emit coverage mappings to a side file and unique them when generating
coverage reports. Both ideas require changes to the build workflow.
>> 
>> A third option is to emit named coverage mappings with linkonce_odr
linkage (for languages with an ODR). This would be a format-breaking change but
it wouldn't affect the build workflow. My plan is to try and evaluate this
idea in the coming week.
>> 
>> 2. HTML report quality
>> 
>> There seems to be widespread interest in improving the quality of
coverage reports. We need volunteers to work on this and would love your help!
Here are some desired features:
>> 
>> * Search and filtering for coverage summaries
>> * Collapsing parts of a coverage summary by subdirectory
>> * Automatically generating a top 10 list of code regions which need
better coverage
>> * Searching via complex queries (e.g: 'give me uncovered regions in
covered lines', or 'give me uncovered regions after a call')
>> * Generating coverage deltas between two profiles, and identifying
coverage regressions in a patch/commit
>> * Simplified tracking of coverage trends over time
>> 
>> There is some consensus that this functionality should not be built on
top of the existing llvm-cov C++ codebase. It might be better to develop these
features in a language more amenable to rapid prototyping and interoperation
with popular web application frameworks (perhaps Python). To facilitate this,
llvm-cov gained support for exporting all of its data to JSON (see
CoverageExporterJson.cpp
<https://github.com/llvm-mirror/llvm/blob/master/tools/llvm-cov/CoverageExporterJson.cpp>).
If you are interested in working on these features, I would be happy to work
with you on design issues and on code review.
>> 
>> 3. Optimizing profile counter placement
>> 
>> From Eli's notes:
>> 
>>> I remember we also spent some time discussing the counter
intrinsics, and whether we could produce a different set of intrinsics in the
frontend, and produce the counters later in the pipeline to avoid duplicate
counters.   I didn't completely follow that discussion; I haven't spent
much time looking at the counter intrinsics or how they're lowered.
>> 
>> Just to recap: the frontend emits calls to the llvm.instrprof_increment
intrinsic to implement counter updates. Each increment intrinsic is passed a
function name and a counter index (there's a mapping between AST nodes and
counter indices). The intrinsics are lowered in the InstrProfiling pass. During
lowering, an array of uint64_t counters is created for each function, and the
intrinsic calls are replaced by a load-add-store pattern.
>> 
>> Frontend counter updates can look highly redundant because of inlining.
It's common to see single basic blocks with tens of distinct counter
updates, most of which are redundant. One potential solution is to create a
minimal set of profile counter updates after the inliner runs, and to map these
counters back to AST nodes (https://bugs.llvm.org/show_bug.cgi?id=33500
<https://bugs.llvm.org/show_bug.cgi?id=33500>). This is the most promising
approach we know of to cut down on counter updates, but I don't have a
precise idea of how it would work. Here's a rough sketch of a solution:
>> 
>> * Have the frontend emit 'virtual' llvm.instrprof_increment
intrinsics. These will eventually be discarded during lowering.
>> * Run an early inlining step, then run the IR PGO pass.
>> * In the lowering step, emit a section into the object which describes
how to map the real counter updates to the virtual ones. I don't have a
clear idea of how to build or encode this mapping.
>> * Teach llvm-profdata how to reconstruct an indexed profile which the
frontend can understand (i.e map the real counters back to the virtual ones).
llvm-profdata would need to inspect the mapping section in the binary to
accomplish this.
>> 
>> 4. Optimizing profile counter updates
>> 
>> We had a few different suggestions to speed up profile counter updates:
>> 
>> * Make function counter arrays linkonce_odr when possible. This is
similar to the solution from the first section ("The header problem").
I'll try to evaluate this idea in the coming week.
>> * Enable register promotion for counter updates which occur within
loops. David Li has already done the work to enable this for IR PGO.
>> * Investigate the # of relocations emitted for counter updates. It
might be cheaper to load the address of the function counter array once and
index into it, instead of indexing into the global on each update.
>> * Use 32-bit counters. This would cut the size of the counters section
in half and speed up profile I/O.
>> * Use 1-bit counters. This could be useful for those who are only
interested in binary coverage. IMO there are other ideas we should try before
compromising on report accuracy.
>> * Use saturating counters. IMO this isn't likely to be a win in
common cases, but could increase compile time and code size.
>> 
>> 5. Using coverage interactively while hacking on llvm
>> 
>> During the BoF I mentioned that it can be really useful to see coverage
reporting interactively, as you're working on a patch. Here's a hacky
way to do this:
>> 
>> * Build your code as you normally would (say, "ninja opt")
>> * Change the files you're interested in
>> * cd to your build directory and export
CCC_OVERRIDE_OPTIONS="+-fcoverage-mapping
+-fprofile-instr-generate=/tmp/opt_%m.profraw"
>> * Rebuild ("ninja opt" again). This will enable coverage
instrumentation, but only for the files you've affected with your changes.
>> * Run a one-liner to generate a coverage report
(http://clang.llvm.org/docs/SourceBasedCodeCoverage.html#creating-coverage-reports
<http://clang.llvm.org/docs/SourceBasedCodeCoverage.html#creating-coverage-reports>)
>> 
>> I like this approach because it means I don't have to maintain a
separate, coverage-enabled build tree. It's an easy way to check that your
patches have decent test coverage. If I want to disable coverage reporting I
just need to unset CCC_OVERRIDE_OPTIONS and recompile.
>> 
>> 6. C APIs for libCoverage
>> 
>> We didn't get a chance to discuss this in detail during the BoF,
but I would like to upstream some C APIs to surface functionality from
libCoverage. This will make it easier for IDEs and editors to display coverage
information "in-line", right next to source code. Here's what that
might look like:
>> 
>>
https://developer.apple.com/library/content/documentation/DeveloperTools/Conceptual/testing_with_xcode/chapters/07-code_coverage.html
<https://developer.apple.com/library/content/documentation/DeveloperTools/Conceptual/testing_with_xcode/chapters/07-code_coverage.html>
>> 
>> If anyone has concerns about adding in these APIs, please let me know!
>> 
>> 7. Making use of debug info
>> 
>> From Eli's notes:
>> 
>>> It seemed like we got a lot of questions related to why we
aren't using debug info. :) It might be possible to come up with some sort
of hybrid which trades off runtime overhead for lower resolution, without
completely throwing away regions like gcov does.  But it would be a big project,
and the end result would still have a lot of the same problems as actual gcov in
terms of the optimizer destroying necessary info.
>> 
>> To add to this: I think there are a lot of unanswered questions here.
It's unclear how clang would decide to use debug info instead of regions, or
how the different types of coverage counters would interact. I'm not very
optimistic about this.
>> 
>> thanks,
>> vedant
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>
>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
> 
> -- Dean
> 
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20171024/d31ad9ec/attachment.html>

> On 25 Oct 2017, at 10:06, Vedant Kumar <vsk at apple.com> wrote:
> 
> Hi Dean,
> 
> We didn't discuss using XRay instrumentation during the BoF but it is
an interesting idea (by the way, thanks for your talk about XRay internals!).
XRay provides the advantage of being able to turn profiling on and off, but
I'm not sure how the resulting data could be used.
The talk was my pleasure -- hopefully when the video comes out, more people can
get up to speed on the intricacies involved in making things work. :)

The dynamic control might not be as important as being able to patch pre-main
and get execution coverage.
> 
> The code coverage feature is highly dependent on the frontend's profile
counter placement. The mapping between counters and parts of the AST is used to
gather accurate information about regions within a line. For example, the
coverage tool can show you that the l.h.s of "true || false" is
evaluated once, and the r.h.s isn't evaluated. This works with arbitrarily
nested short-circuit operators.
> 
Yes, this is cool stuff. The way we try to do this in XRay is at least to do the
following:

- When we're lowering the functions into assembler, we're able to
maintain a side-table (separate section) that maps where the instrumentation
points are. The instrumentation points are currently only automatically placed
for function entry and exits. We use heuristics to figure out which functions
get instrumented, but can be flag-controlled.

- XRay also has support for inserting intrinsics at the IR level that mark
"custom events" which take a pointer and a size -- the handler in the
XRay runtime can then do what it deems useful to do at these custom event
pointer. The downside to this intrinsic is that it behaves as a function call,
and might cause some (re)ordering issues or prevent some important optimisations
for code motion. Maybe coverage mode runs don't necessarily care about the
optimisations that happen and this is not an issue.

It seems to me that the combination of emitting custom events along with a
handler implementation at a basic block level and still be able to correlate
information more precisely to the AST (source-level) might be good to do. DWARF
info might not be the best thing to use to determine whether a specific sub-line
path was taken, but it's certainly helpful. The best I can think of at least
with XRay as something to build upon is a way of also marking where at
source-level certain branches/blocks are. This might be doable in a compact
manner as a CFG path, and if there's a consistent numbering of the CFG paths
those could be used to track the counters/latches in some sort of table.

I don't have a very clear picture yet in my head how that table might look
like or how that's different compared to the way it's being done now, so
I'll have to sit on it a bit longer and maybe whiteboard/sketch it out a bit
more.
> It might be possible to use XRay instrumentation to gather profile data,
but I think it will be challenging to precisely map that data back to the AST
nodes the frontend knows about. The problem is similar to the one I've
outlined in section 3 ("Optimizing profile counter placement"). The
idea there is to map a minimal set of counters placed by IR PGO back to AST
nodes: the one sketch of a solution I have still depends on running the frontend
counter placement pass to achieve this.
> 
> What are your thoughts on this?
> 
There's probably a representation we can use that will indicate which paths
in a CFG for a function were visited, and associating which parts of the AST
have been covered already. Assuming for example that we can give all paths from
a function's entry to a function's exits a unique number (for that
function) then we can certainly re-create that path as we're going through
the execution path and collect that information for coverage/reconstruction
purposes.

The advantage of using XRay in the collection path is so you can process the
data either in the handler implementation(s) while the program is running and
collecting the pre-computed coverage paths somehow (through files, or in memory
collection) or even collecting a bit more information and processing it offline
(things like collecting frequency and some PMU data while you're at it).

Currently I don't have time to explore this further, but if you or someone
else might be interested in giving this a go, I'd certainly be happy to
further explore at least the design of a potential alternative solution that
involves using XRay for coverage/PGO data gathering.

Cheers
> thanks,
> vedant
> 
>> On Oct 24, 2017, at 1:53 PM, Dean Michael Berris <dean.berris at
gmail.com <mailto:dean.berris at gmail.com>> wrote:
>> 
>> Thanks for the summary Vedant!
>> 
>> I'm sorry I missed this BoF session.
>> 
>> Others have mentioned the possibility of maybe using XRay for some of
this information (function-level coverage, maybe having more intrinsics for
marking branch/basic-block level instrumentation). Was this explored in the BoF?
Is there interest in potentially exploring this particular space?
>> 
>> Cheers
>> 
>>> On 25 Oct 2017, at 07:24, Vedant Kumar via llvm-dev <llvm-dev at
lists.llvm.org <mailto:llvm-dev at lists.llvm.org>> wrote:
>>> 
>>> Hello,
>>> 
>>> Our goals for the code coverage BoF (10/19) were to find areas
where we can improve the coverage tooling, and to learn more about how coverage
is used. I'd like to thank all of the attendees for their input and for
making the BoF productive. Special thanks to Mandeep Grang, who volunteered as a
mic runner at the last minute.
>>> 
>>> In this email I'll share my (rough) notes and outline some
future plans. Please feel free to ask for clarifications or to add your own
notes.
>>> 
>>> Here are the slides from the BoF:
>>>
https://docs.google.com/presentation/d/e/2PACX-1vS-rV02j1zhPq9Y6AtcUkbZW2c7Q5YYuQ6FPxN-aYiKwrw6c8DU3zW_RYeJlWPMZ5-S6hgz_CIcL8Gd/pub?start=false&loop=false&delayms=3000&slide=id.p
<https://docs.google.com/presentation/d/e/2PACX-1vS-rV02j1zhPq9Y6AtcUkbZW2c7Q5YYuQ6FPxN-aYiKwrw6c8DU3zW_RYeJlWPMZ5-S6hgz_CIcL8Gd/pub?start=false&loop=false&delayms=3000&slide=id.p>
>>> 
>>> 1. The header problem
>>> 
>>> Coverage instrumentation overhead is roughly quadratic in the
number of translation units in a project. The problem is that coverage mappings
for template instantiations and static inline functions from headers are pulled
into every TU. This bloats the profile metadata sections (which can slow down
profile I/O), results in large binaries, and causes long link times (or link
failures).
>>> 
>>> We could solve this problem by maintaining an external coverage
database and discarding duplicate coverage mappings from the DB. Another idea is
to emit coverage mappings to a side file and unique them when generating
coverage reports. Both ideas require changes to the build workflow.
>>> 
>>> A third option is to emit named coverage mappings with linkonce_odr
linkage (for languages with an ODR). This would be a format-breaking change but
it wouldn't affect the build workflow. My plan is to try and evaluate this
idea in the coming week.
>>> 
>>> 2. HTML report quality
>>> 
>>> There seems to be widespread interest in improving the quality of
coverage reports. We need volunteers to work on this and would love your help!
Here are some desired features:
>>> 
>>> * Search and filtering for coverage summaries
>>> * Collapsing parts of a coverage summary by subdirectory
>>> * Automatically generating a top 10 list of code regions which need
better coverage
>>> * Searching via complex queries (e.g: 'give me uncovered
regions in covered lines', or 'give me uncovered regions after a
call')
>>> * Generating coverage deltas between two profiles, and identifying
coverage regressions in a patch/commit
>>> * Simplified tracking of coverage trends over time
>>> 
>>> There is some consensus that this functionality should not be built
on top of the existing llvm-cov C++ codebase. It might be better to develop
these features in a language more amenable to rapid prototyping and
interoperation with popular web application frameworks (perhaps Python). To
facilitate this, llvm-cov gained support for exporting all of its data to JSON
(see CoverageExporterJson.cpp
<https://github.com/llvm-mirror/llvm/blob/master/tools/llvm-cov/CoverageExporterJson.cpp>).
If you are interested in working on these features, I would be happy to work
with you on design issues and on code review.
>>> 
>>> 3. Optimizing profile counter placement
>>> 
>>> From Eli's notes:
>>> 
>>>> I remember we also spent some time discussing the counter
intrinsics, and whether we could produce a different set of intrinsics in the
frontend, and produce the counters later in the pipeline to avoid duplicate
counters.   I didn't completely follow that discussion; I haven't spent
much time looking at the counter intrinsics or how they're lowered.
>>> 
>>> Just to recap: the frontend emits calls to the
llvm.instrprof_increment intrinsic to implement counter updates. Each increment
intrinsic is passed a function name and a counter index (there's a mapping
between AST nodes and counter indices). The intrinsics are lowered in the
InstrProfiling pass. During lowering, an array of uint64_t counters is created
for each function, and the intrinsic calls are replaced by a load-add-store
pattern.
>>> 
>>> Frontend counter updates can look highly redundant because of
inlining. It's common to see single basic blocks with tens of distinct
counter updates, most of which are redundant. One potential solution is to
create a minimal set of profile counter updates after the inliner runs, and to
map these counters back to AST nodes
(https://bugs.llvm.org/show_bug.cgi?id=33500
<https://bugs.llvm.org/show_bug.cgi?id=33500>). This is the most promising
approach we know of to cut down on counter updates, but I don't have a
precise idea of how it would work. Here's a rough sketch of a solution:
>>> 
>>> * Have the frontend emit 'virtual' llvm.instrprof_increment
intrinsics. These will eventually be discarded during lowering.
>>> * Run an early inlining step, then run the IR PGO pass.
>>> * In the lowering step, emit a section into the object which
describes how to map the real counter updates to the virtual ones. I don't
have a clear idea of how to build or encode this mapping.
>>> * Teach llvm-profdata how to reconstruct an indexed profile which
the frontend can understand (i.e map the real counters back to the virtual
ones). llvm-profdata would need to inspect the mapping section in the binary to
accomplish this.
>>> 
>>> 4. Optimizing profile counter updates
>>> 
>>> We had a few different suggestions to speed up profile counter
updates:
>>> 
>>> * Make function counter arrays linkonce_odr when possible. This is
similar to the solution from the first section ("The header problem").
I'll try to evaluate this idea in the coming week.
>>> * Enable register promotion for counter updates which occur within
loops. David Li has already done the work to enable this for IR PGO.
>>> * Investigate the # of relocations emitted for counter updates. It
might be cheaper to load the address of the function counter array once and
index into it, instead of indexing into the global on each update.
>>> * Use 32-bit counters. This would cut the size of the counters
section in half and speed up profile I/O.
>>> * Use 1-bit counters. This could be useful for those who are only
interested in binary coverage. IMO there are other ideas we should try before
compromising on report accuracy.
>>> * Use saturating counters. IMO this isn't likely to be a win in
common cases, but could increase compile time and code size.
>>> 
>>> 5. Using coverage interactively while hacking on llvm
>>> 
>>> During the BoF I mentioned that it can be really useful to see
coverage reporting interactively, as you're working on a patch. Here's a
hacky way to do this:
>>> 
>>> * Build your code as you normally would (say, "ninja
opt")
>>> * Change the files you're interested in
>>> * cd to your build directory and export
CCC_OVERRIDE_OPTIONS="+-fcoverage-mapping
+-fprofile-instr-generate=/tmp/opt_%m.profraw"
>>> * Rebuild ("ninja opt" again). This will enable coverage
instrumentation, but only for the files you've affected with your changes.
>>> * Run a one-liner to generate a coverage report
(http://clang.llvm.org/docs/SourceBasedCodeCoverage.html#creating-coverage-reports
<http://clang.llvm.org/docs/SourceBasedCodeCoverage.html#creating-coverage-reports>)
>>> 
>>> I like this approach because it means I don't have to maintain
a separate, coverage-enabled build tree. It's an easy way to check that your
patches have decent test coverage. If I want to disable coverage reporting I
just need to unset CCC_OVERRIDE_OPTIONS and recompile.
>>> 
>>> 6. C APIs for libCoverage
>>> 
>>> We didn't get a chance to discuss this in detail during the
BoF, but I would like to upstream some C APIs to surface functionality from
libCoverage. This will make it easier for IDEs and editors to display coverage
information "in-line", right next to source code. Here's what that
might look like:
>>> 
>>>
https://developer.apple.com/library/content/documentation/DeveloperTools/Conceptual/testing_with_xcode/chapters/07-code_coverage.html
<https://developer.apple.com/library/content/documentation/DeveloperTools/Conceptual/testing_with_xcode/chapters/07-code_coverage.html>
>>> 
>>> If anyone has concerns about adding in these APIs, please let me
know!
>>> 
>>> 7. Making use of debug info
>>> 
>>> From Eli's notes:
>>> 
>>>> It seemed like we got a lot of questions related to why we
aren't using debug info. :) It might be possible to come up with some sort
of hybrid which trades off runtime overhead for lower resolution, without
completely throwing away regions like gcov does.  But it would be a big project,
and the end result would still have a lot of the same problems as actual gcov in
terms of the optimizer destroying necessary info.
>>> 
>>> To add to this: I think there are a lot of unanswered questions
here. It's unclear how clang would decide to use debug info instead of
regions, or how the different types of coverage counters would interact. I'm
not very optimistic about this.
>>> 
>>> thanks,
>>> vedant
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> llvm-dev at lists.llvm.org <mailto:llvm-dev at
lists.llvm.org>
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
<http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev>
>> 
>> -- Dean
>> 
> 
-- Dean

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20171025/5603a99d/attachment-0001.html>

> On Oct 25, 2017, at 3:47 AM, Dean Michael Berris <dean.berris at
gmail.com> wrote:
> 
>> 
>> On 25 Oct 2017, at 10:06, Vedant Kumar <vsk at apple.com
<mailto:vsk at apple.com>> wrote:
>> 
>> Hi Dean,
>> 
>> We didn't discuss using XRay instrumentation during the BoF but it
is an interesting idea (by the way, thanks for your talk about XRay internals!).
XRay provides the advantage of being able to turn profiling on and off, but
I'm not sure how the resulting data could be used.
> 
> The talk was my pleasure -- hopefully when the video comes out, more people
can get up to speed on the intricacies involved in making things work. :)
> 
> The dynamic control might not be as important as being able to patch
pre-main and get execution coverage.
> 
>> 
>> The code coverage feature is highly dependent on the frontend's
profile counter placement. The mapping between counters and parts of the AST is
used to gather accurate information about regions within a line. For example,
the coverage tool can show you that the l.h.s of "true || false" is
evaluated once, and the r.h.s isn't evaluated. This works with arbitrarily
nested short-circuit operators.
>> 
> 
> Yes, this is cool stuff. The way we try to do this in XRay is at least to
do the following:
> 
> - When we're lowering the functions into assembler, we're able to
maintain a side-table (separate section) that maps where the instrumentation
points are. The instrumentation points are currently only automatically placed
for function entry and exits. We use heuristics to figure out which functions
get instrumented, but can be flag-controlled.
> 
> - XRay also has support for inserting intrinsics at the IR level that mark
"custom events" which take a pointer and a size -- the handler in the
XRay runtime can then do what it deems useful to do at these custom event
pointer.
Very cool!
> The downside to this intrinsic is that it behaves as a function call, and
might cause some (re)ordering issues or prevent some important optimisations for
code motion. Maybe coverage mode runs don't necessarily care about the
optimisations that happen and this is not an issue.
I'm not sure how much we benefit from the optimizer's handling of the
load-add-store counter update pattern. I have seen examples where it's able
to, e.g unroll a loop and perform a single counter update, or vectorize counter
updates. Would it ever make sense to inline the handler at custom event points,
or is this counter to XRay's design?
> It seems to me that the combination of emitting custom events along with a
handler implementation at a basic block level and still be able to correlate
information more precisely to the AST (source-level) might be good to do. DWARF
info might not be the best thing to use to determine whether a specific sub-line
path was taken, but it's certainly helpful. The best I can think of at least
with XRay as something to build upon is a way of also marking where at
source-level certain branches/blocks are. This might be doable in a compact
manner as a CFG path, and if there's a consistent numbering of the CFG paths
those could be used to track the counters/latches in some sort of table.
> 
> I don't have a very clear picture yet in my head how that table might
look like or how that's different compared to the way it's being done
now, so I'll have to sit on it a bit longer and maybe whiteboard/sketch it
out a bit more.
The profiling pass used for coverage isn't path sensitive, it just looks at
basic block counts (not edge counts). I think one way to get started with XRay
is to emit custom event points everywhere the llvm.instrprof_increment intrinsic
is currently used (see clang's CodeGenPGO). The next step would be to
massage the resulting profile into a form the coverage reader can understand.

That said, I think there are some people interested in working with
path-sensitive coverage. That's not on my radar right now, but if someone
does want to pick this project up, I think it would be useful to at least
consider 'standardizing' on one type of instrumentation (like XRay).
>> It might be possible to use XRay instrumentation to gather profile
data, but I think it will be challenging to precisely map that data back to the
AST nodes the frontend knows about. The problem is similar to the one I've
outlined in section 3 ("Optimizing profile counter placement"). The
idea there is to map a minimal set of counters placed by IR PGO back to AST
nodes: the one sketch of a solution I have still depends on running the frontend
counter placement pass to achieve this.
>> 
>> What are your thoughts on this?
>> 
> 
> There's probably a representation we can use that will indicate which
paths in a CFG for a function were visited, and associating which parts of the
AST have been covered already. Assuming for example that we can give all paths
from a function's entry to a function's exits a unique number (for that
function) then we can certainly re-create that path as we're going through
the execution path and collect that information for coverage/reconstruction
purposes.
> 
> The advantage of using XRay in the collection path is so you can process
the data either in the handler implementation(s) while the program is running
and collecting the pre-computed coverage paths somehow (through files, or in
memory collection) or even collecting a bit more information and processing it
offline (things like collecting frequency and some PMU data while you're at
it).
Reflecting on coverage while a program is running is useful enough that the
profiling runtime has basic support for it. But I agree with your points here.
The tooling and flexibility you can get with XRay makes it an interesting
choice.

thanks,
vedant
> Currently I don't have time to explore this further, but if you or
someone else might be interested in giving this a go, I'd certainly be happy
to further explore at least the design of a potential alternative solution that
involves using XRay for coverage/PGO data gathering.
> 
> Cheers
> 
>> thanks,
>> vedant
>> 
>>> On Oct 24, 2017, at 1:53 PM, Dean Michael Berris <dean.berris at
gmail.com <mailto:dean.berris at gmail.com>> wrote:
>>> 
>>> Thanks for the summary Vedant!
>>> 
>>> I'm sorry I missed this BoF session.
>>> 
>>> Others have mentioned the possibility of maybe using XRay for some
of this information (function-level coverage, maybe having more intrinsics for
marking branch/basic-block level instrumentation). Was this explored in the BoF?
Is there interest in potentially exploring this particular space?
>>> 
>>> Cheers
>>> 
>>>> On 25 Oct 2017, at 07:24, Vedant Kumar via llvm-dev
<llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>>
wrote:
>>>> 
>>>> Hello,
>>>> 
>>>> Our goals for the code coverage BoF (10/19) were to find areas
where we can improve the coverage tooling, and to learn more about how coverage
is used. I'd like to thank all of the attendees for their input and for
making the BoF productive. Special thanks to Mandeep Grang, who volunteered as a
mic runner at the last minute.
>>>> 
>>>> In this email I'll share my (rough) notes and outline some
future plans. Please feel free to ask for clarifications or to add your own
notes.
>>>> 
>>>> Here are the slides from the BoF:
>>>>
https://docs.google.com/presentation/d/e/2PACX-1vS-rV02j1zhPq9Y6AtcUkbZW2c7Q5YYuQ6FPxN-aYiKwrw6c8DU3zW_RYeJlWPMZ5-S6hgz_CIcL8Gd/pub?start=false&loop=false&delayms=3000&slide=id.p
<https://docs.google.com/presentation/d/e/2PACX-1vS-rV02j1zhPq9Y6AtcUkbZW2c7Q5YYuQ6FPxN-aYiKwrw6c8DU3zW_RYeJlWPMZ5-S6hgz_CIcL8Gd/pub?start=false&loop=false&delayms=3000&slide=id.p>
>>>> 
>>>> 1. The header problem
>>>> 
>>>> Coverage instrumentation overhead is roughly quadratic in the
number of translation units in a project. The problem is that coverage mappings
for template instantiations and static inline functions from headers are pulled
into every TU. This bloats the profile metadata sections (which can slow down
profile I/O), results in large binaries, and causes long link times (or link
failures).
>>>> 
>>>> We could solve this problem by maintaining an external coverage
database and discarding duplicate coverage mappings from the DB. Another idea is
to emit coverage mappings to a side file and unique them when generating
coverage reports. Both ideas require changes to the build workflow.
>>>> 
>>>> A third option is to emit named coverage mappings with
linkonce_odr linkage (for languages with an ODR). This would be a
format-breaking change but it wouldn't affect the build workflow. My plan is
to try and evaluate this idea in the coming week.
>>>> 
>>>> 2. HTML report quality
>>>> 
>>>> There seems to be widespread interest in improving the quality
of coverage reports. We need volunteers to work on this and would love your
help! Here are some desired features:
>>>> 
>>>> * Search and filtering for coverage summaries
>>>> * Collapsing parts of a coverage summary by subdirectory
>>>> * Automatically generating a top 10 list of code regions which
need better coverage
>>>> * Searching via complex queries (e.g: 'give me uncovered
regions in covered lines', or 'give me uncovered regions after a
call')
>>>> * Generating coverage deltas between two profiles, and
identifying coverage regressions in a patch/commit
>>>> * Simplified tracking of coverage trends over time
>>>> 
>>>> There is some consensus that this functionality should not be
built on top of the existing llvm-cov C++ codebase. It might be better to
develop these features in a language more amenable to rapid prototyping and
interoperation with popular web application frameworks (perhaps Python). To
facilitate this, llvm-cov gained support for exporting all of its data to JSON
(see CoverageExporterJson.cpp
<https://github.com/llvm-mirror/llvm/blob/master/tools/llvm-cov/CoverageExporterJson.cpp>).
If you are interested in working on these features, I would be happy to work
with you on design issues and on code review.
>>>> 
>>>> 3. Optimizing profile counter placement
>>>> 
>>>> From Eli's notes:
>>>> 
>>>>> I remember we also spent some time discussing the counter
intrinsics, and whether we could produce a different set of intrinsics in the
frontend, and produce the counters later in the pipeline to avoid duplicate
counters.   I didn't completely follow that discussion; I haven't spent
much time looking at the counter intrinsics or how they're lowered.
>>>> 
>>>> Just to recap: the frontend emits calls to the
llvm.instrprof_increment intrinsic to implement counter updates. Each increment
intrinsic is passed a function name and a counter index (there's a mapping
between AST nodes and counter indices). The intrinsics are lowered in the
InstrProfiling pass. During lowering, an array of uint64_t counters is created
for each function, and the intrinsic calls are replaced by a load-add-store
pattern.
>>>> 
>>>> Frontend counter updates can look highly redundant because of
inlining. It's common to see single basic blocks with tens of distinct
counter updates, most of which are redundant. One potential solution is to
create a minimal set of profile counter updates after the inliner runs, and to
map these counters back to AST nodes
(https://bugs.llvm.org/show_bug.cgi?id=33500
<https://bugs.llvm.org/show_bug.cgi?id=33500>). This is the most promising
approach we know of to cut down on counter updates, but I don't have a
precise idea of how it would work. Here's a rough sketch of a solution:
>>>> 
>>>> * Have the frontend emit 'virtual'
llvm.instrprof_increment intrinsics. These will eventually be discarded during
lowering.
>>>> * Run an early inlining step, then run the IR PGO pass.
>>>> * In the lowering step, emit a section into the object which
describes how to map the real counter updates to the virtual ones. I don't
have a clear idea of how to build or encode this mapping.
>>>> * Teach llvm-profdata how to reconstruct an indexed profile
which the frontend can understand (i.e map the real counters back to the virtual
ones). llvm-profdata would need to inspect the mapping section in the binary to
accomplish this.
>>>> 
>>>> 4. Optimizing profile counter updates
>>>> 
>>>> We had a few different suggestions to speed up profile counter
updates:
>>>> 
>>>> * Make function counter arrays linkonce_odr when possible. This
is similar to the solution from the first section ("The header
problem"). I'll try to evaluate this idea in the coming week.
>>>> * Enable register promotion for counter updates which occur
within loops. David Li has already done the work to enable this for IR PGO.
>>>> * Investigate the # of relocations emitted for counter updates.
It might be cheaper to load the address of the function counter array once and
index into it, instead of indexing into the global on each update.
>>>> * Use 32-bit counters. This would cut the size of the counters
section in half and speed up profile I/O.
>>>> * Use 1-bit counters. This could be useful for those who are
only interested in binary coverage. IMO there are other ideas we should try
before compromising on report accuracy.
>>>> * Use saturating counters. IMO this isn't likely to be a
win in common cases, but could increase compile time and code size.
>>>> 
>>>> 5. Using coverage interactively while hacking on llvm
>>>> 
>>>> During the BoF I mentioned that it can be really useful to see
coverage reporting interactively, as you're working on a patch. Here's a
hacky way to do this:
>>>> 
>>>> * Build your code as you normally would (say, "ninja
opt")
>>>> * Change the files you're interested in
>>>> * cd to your build directory and export
CCC_OVERRIDE_OPTIONS="+-fcoverage-mapping
+-fprofile-instr-generate=/tmp/opt_%m.profraw"
>>>> * Rebuild ("ninja opt" again). This will enable
coverage instrumentation, but only for the files you've affected with your
changes.
>>>> * Run a one-liner to generate a coverage report
(http://clang.llvm.org/docs/SourceBasedCodeCoverage.html#creating-coverage-reports
<http://clang.llvm.org/docs/SourceBasedCodeCoverage.html#creating-coverage-reports>)
>>>> 
>>>> I like this approach because it means I don't have to
maintain a separate, coverage-enabled build tree. It's an easy way to check
that your patches have decent test coverage. If I want to disable coverage
reporting I just need to unset CCC_OVERRIDE_OPTIONS and recompile.
>>>> 
>>>> 6. C APIs for libCoverage
>>>> 
>>>> We didn't get a chance to discuss this in detail during the
BoF, but I would like to upstream some C APIs to surface functionality from
libCoverage. This will make it easier for IDEs and editors to display coverage
information "in-line", right next to source code. Here's what that
might look like:
>>>> 
>>>>
https://developer.apple.com/library/content/documentation/DeveloperTools/Conceptual/testing_with_xcode/chapters/07-code_coverage.html
<https://developer.apple.com/library/content/documentation/DeveloperTools/Conceptual/testing_with_xcode/chapters/07-code_coverage.html>
>>>> 
>>>> If anyone has concerns about adding in these APIs, please let
me know!
>>>> 
>>>> 7. Making use of debug info
>>>> 
>>>> From Eli's notes:
>>>> 
>>>>> It seemed like we got a lot of questions related to why we
aren't using debug info. :) It might be possible to come up with some sort
of hybrid which trades off runtime overhead for lower resolution, without
completely throwing away regions like gcov does.  But it would be a big project,
and the end result would still have a lot of the same problems as actual gcov in
terms of the optimizer destroying necessary info.
>>>> 
>>>> To add to this: I think there are a lot of unanswered questions
here. It's unclear how clang would decide to use debug info instead of
regions, or how the different types of coverage counters would interact. I'm
not very optimistic about this.
>>>> 
>>>> thanks,
>>>> vedant
>>>> _______________________________________________
>>>> LLVM Developers mailing list
>>>> llvm-dev at lists.llvm.org <mailto:llvm-dev at
lists.llvm.org>
>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
<http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev>
>>> 
>>> -- Dean
>>> 
>> 
> 
> -- Dean
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20171025/81147904/attachment-0001.html>