[cfe-dev] RFC: Clang-Misexpect Proposal

[Paul Kirth via cfe-dev]

We would like to propose adding a new clang based tool, clang-mispredict,
for identifying potentially incorrect uses of __builtin_expect().

--- TLDR ---
What are we proposing?
* Adding a new tool, clang-mispredict, to the LLVM project.

What is its purpose?
* Identifying potentially problematic uses of __builtin_expect().

Why is this important?
* __builtin_expect() has a high performance cost when it is used
incorrectly.

Where will the code live?
* clang-mispredict will live in clang-tools-extra, and there will be some
additions to clang.

Open questions
* What is the correct heuristic for determining a mismatch?

Potential shortcomings
* The approach is sensitive to both profiling input and configuration.


--- Details ---

We plant to build a new tool, clang-mispredict, on top of Clang’s PGO
infrastructure. This tool will issue warnings about the use of
_builtin_expect() when collected PGO profile data shows a mismatch with the
use of __builtin_expect(). Existing solutions for validating these
annotations generally follow this approach: leverage dynamic profiling to
validate existing annotations. This is the approach used in the Linux
kernel, for example. Unfortunately, existing solutions seem to be custom
efforts specific to a particular code base. Supporting this in the LLVM
ecosystem gives developers a general way to check the use of
_builtin_expect() without creating custom instrumentation.

Additionally, if _builtin_expect() is used incorrectly, then developers may
notice a performance regression when switching to the new PM. This happens
because the New PM follows performance annotations more closely than the
legacy PM, and is therefore more likely to try and aggressively optimize
these cases. Our proposed tool can help developers narrow their focus to
potentially problematic areas during the transition.

Finally, it may also be beneficial to suggest annotations when profiling
suggests it would be beneficial. We consider this a desirable property not
only from a performance standpoint, but also for identifying potential bugs
(both in the profiling corpus and within the codebase).

We think having clang-mispredict behave similar to clang-tidy is good
approach. We can use a compile commands database in conjunction with
profiling data to drive our validation tool. More concretely, we plan to
use Clang’s existing PGO infrastructure to emit warnings where branch
weights are assigned, but stop compilation before the IR is passed to the
backend. This largely amounts to checking if the profile counters are
outside of certain thresholds for 'hot' and 'cold' code. Similar to clang
tidy, we plan to support suppressing warnings within the codebase through
use of comment strings.

The most obvious question about this approach is how to quantify what is
‘hot’ or ‘cold’. Initially, we plan to use the thresholds already present
in LLVM; however, we are happy to use any suitable heuristic or empirically
derived threshold. Input from the community about the correct heuristic are
most welcome.

Our strategy does have some shortcomings. The usefulness of the warnings
are directly related to how representative the profiling input actually was
when compared to its normal use. If the profile collected is not
representative of typical use, then the warnings may not reflect the ground
truth. Program and build configuration can also dramatically change which
paths will be executed, thus affecting the quality of the profile data and
the generated warnings.

Ultimately, we are proposing adding a new standalone tool that takes a
compile commands database and an instrumentation profile to emit warnings
in a clang-tidy fashion.

Please let us know if you have comments or concerns about this proposal.

Thanks!

-- 
Paul Kirth
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20190723/c17aeb67/attachment.html>