<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
</head>
<body>
<div class="moz-cite-prefix">On 7/11/20 12:58 AM, Richard Smith
wrote:<br>
</div>
<blockquote type="cite"
cite="mid:CAOfiQq=r3uL8rJisi4QwG-RZ+RmTF4p4yZnD41Oc8fuM=hV9Gg@mail.gmail.com">
<meta http-equiv="content-type" content="text/html; charset=UTF-8">
<div dir="ltr">
<div dir="ltr">On Fri, 10 Jul 2020 at 13:59, Vassil Vassilev via
cfe-dev <<a href="mailto:cfe-dev@lists.llvm.org"
moz-do-not-send="true">cfe-dev@lists.llvm.org</a>> wrote:<br>
</div>
<div class="gmail_quote">
<blockquote class="gmail_quote" style="margin:0px 0px 0px
0.8ex;border-left:1px solid
rgb(204,204,204);padding-left:1ex">
<div>
<div>Hi Richard,</div>
<div><br>
</div>
<div>On 7/10/20 11:10 PM, Richard Smith wrote:<br>
</div>
<blockquote type="cite">
<div dir="ltr">Hi Vassil,
<div><br>
</div>
<div>This is a very exciting proposal that I can
imagine bringing important benefits to the existing
cling users and also to the clang user and developer
community. Thank you for all the work you and your
team have done on cling so far and for offering to
bring that work under the LLVM umbrella!</div>
<div><br>
</div>
<div>Are you imagining cling being part of the clang
repository, or a separate LLVM subproject (with only
the changes necessary to support cling-style uses of
the clang libraries added to the clang tree)?</div>
</div>
</blockquote>
<p><br>
</p>
<p> Good question. In principle cling was developed with
the idea to become a separate LLVM subproject. Although
I'd easily see it fit in clang/tools/.<br>
</p>
<p><br>
</p>
<p> Nominally, cling has "high-energy physics"-specific
features such as the so called 'meta commands'. For
example, `[cling] .L some_file` would try to load a
library called some_file.so and if it does not exist,
try #include-ing a header with that name; `[cling] .x
script.C` includes script.C and calls a function named
`script`. I can imagine that broader community may not
like/use that. If we start trimming down features like
that then it won't really be cling anymore. Here is what
I would imagine as a way forward:</p>
<p> 1. Land as many cling/"incremental
compilation"-related patches as we can in clang.<br>
2. Build a simple tool, let's use a strawman name --
clang-repl, which only does the basics. For example, one
can feed it incremental C++ and execute it.<br>
3. Rework cling to use that infrastructure -- ideally,
implementing it's specific meta commands and other
domain-specific features such as dynamic scopes.</p>
<p> We could move any of the cling features which the
broader community finds useful closer to clang. For the
moment I am being conservative as this will also give us
the opportunity to rethink some of the features.</p>
<p> The hard part is what lives where. First bullet point
is clear. The second -- not so much. Clang has a
clang-interpreter in its examples folder and it looks a
little unmaintained. Maybe we can start repurposing that
to match 2.</p>
<p> As for cling itself there are some challenges we
should try to solve. Our community lives downstream
(currently llvm-5) and a straight-forward llvm upgrade +
bugfixing takes around 3 months due to the nature of our
software stacks. It would be a non-trivial task to move
the cling-based development in llvm upstream. My worry
is that HEP-cling will soon depart from LLVM-cling if we
don't get both communities on the same codebase (we have
experienced such a problem with the getFullyQualified*
interfaces). I am hoping that a middleman, such as
clang-repl, can help. When we move parts of cling in
clang we will develop and test the required
functionality using clang-repl. This way users will
enjoy cling-like experience and when cling upgrades its
llvm its codebase will become smaller in size.</p>
<p> Am I making sense?</p>
</div>
</blockquote>
<div>Yes, the above all makes sense to me. I agree that there
should be only one thing named 'cling', and that it should
broadly have the feature set that current 'cling' has. I
think there are a couple of ways we can get there while
still providing the a minimalist interpreter to a broader
audience: either we can build a simpler clang-interpreter
and a more advanced cling binary from a common set of
libraries, or we could produce a configurable binary that's
able to serve both rules depending on configuration or a
plugin or scripting system.</div>
</div>
</div>
</blockquote>
<p><br>
</p>
<p> Good point. We could make it extendable, and actually that
should be a design goal. The question how exactly is not very
clear to me. Can you elaborate on what you had in mind as
configuration or scripting system (plugin system I think I know
what you meant). I will give an example with 3 distinct features
in cling which we have implemented over the years and had
different requirements:</p>
<p> * <a moz-do-not-send="true"
href="https://llvm.org/devmtg/2013-11/slides/Vassilev-Poster.pdf">AST-based
automatic differentiation</a> with the <a
moz-do-not-send="true" href="https://github.com/vgvassilev/clad">clad
library</a> -- here we essentially extend cling's runtime by
providing a `clad::differentiate`, `clad::gradient`,
`clad::hessian` and `clad::jacobian` primitives. Each primitive is
a specially annotated wrapper over a function, say `double
pow2(double x) { return x*x; }; auto pow2dx =
clad::differentiate(pow2,/*wrt*/0);`. Here we let clang build a
valid AST and the plugin creates the first order derivative and
swaps the DeclRefExpr just before codegen so that we call the
derivative instead. This is achievable by the current clang plugin
system ( a bit problematic on windows as clang plugins do not work
there ).</p>
<p> * Language extensions which require Sema support -- we have a
legacy feature which should define a variable on the prompt if not
defined (something like implicit auto) `cling[] i = 13;` should be
translated into `cling[] auto i = 13;` if I is undefined. We solve
that by adding some last resort lookup callback which marks `i` of
dependent type so that we can produce an AST which we can later
'fix'.</p>
<p> * Language extensions which require delayed lookup rules (aka
dynamic scope) -- ROOT has an I/O system bound to cling people can
write:`if (TFile::Open("file_that_has_hist_cpp_obj.root"))
hist->Draw();`. Here we use the approach from the previous
bullet and synthesize `if
(TFile::Open("file_that_has_hist_cpp_obj.root"))
eval<void>("hist->Draw()", /*escape some context*/...);`.</p>
<p><br>
</p>
<p> The implementation of these three features can be considered as
possible with current clang. The issue is that it seems more like
hacking clang rather than extending it. If we can come up with a
sound way of implementing these features that would be awesome.<br>
</p>
<p><br>
</p>
<blockquote type="cite"
cite="mid:CAOfiQq=r3uL8rJisi4QwG-RZ+RmTF4p4yZnD41Oc8fuM=hV9Gg@mail.gmail.com">
<div dir="ltr">
<div class="gmail_quote">
<div><br>
</div>
<div>One other thing I think we should consider: there will be
substantial overlap between the incremental compilation,
code generation, REPL, etc. of cling and that of lldb.</div>
</div>
</div>
</blockquote>
<p><br>
</p>
<p> I would love to hear opinions from the lldb folks. We have
chatted number of times and I have looked at how they do it. I
think lldb spawns (used to spawn last time I looked) a compiler
instance per input line. That is not acceptable for cling due to
its high-performance requirements. Most of the issues that need
solving for lldb comes from materializing debug information to
AST. LLDB folks, correct me if I am wrong.</p>
<p> That being said doesn't mean that we should not aim for
centralizing the incremental compilation for both projects. We
should but may be challenging because of the different focus which
defines project priorities.<br>
</p>
<p><br>
</p>
<blockquote type="cite"
cite="mid:CAOfiQq=r3uL8rJisi4QwG-RZ+RmTF4p4yZnD41Oc8fuM=hV9Gg@mail.gmail.com">
<div dir="ltr">
<div class="gmail_quote">
<div> For the initial integration of cling into LLVM, there's
probably not much we can do about that, but it would seem
beneficial for both cling and lldb if common parts could be
shared where possible. As an extreme example, if we could
fully unify the projects to the point where a user could
switch into an 'lldb mode' in the middle of a cling session
to do step-by-step debugging of code entered into the REPL,
that would seem like an incredibly useful feature. Perhaps
there's some common set of base functionality that can be
factored out of lldb and cling and unified. It would likely
be a good idea to start talking to the lldb folks about that
early, in case it guides your work porting cling to trunk.</div>
</div>
</div>
</blockquote>
<p><br>
</p>
<p> Indeed. There have been user requests to be able to run
step-by-step in cling. That would be the ultimate long term goal!<br>
</p>
<p><br>
</p>
<blockquote type="cite"
cite="mid:CAOfiQq=r3uL8rJisi4QwG-RZ+RmTF4p4yZnD41Oc8fuM=hV9Gg@mail.gmail.com">
<div dir="ltr">
<div class="gmail_quote">
<blockquote class="gmail_quote" style="margin:0px 0px 0px
0.8ex;border-left:1px solid
rgb(204,204,204);padding-left:1ex">
<div>
<blockquote type="cite">
<div class="gmail_quote">
<div dir="ltr" class="gmail_attr">On Thu, 9 Jul 2020
at 13:46, Vassil Vassilev via cfe-dev <<a
href="mailto:cfe-dev@lists.llvm.org"
target="_blank" moz-do-not-send="true">cfe-dev@lists.llvm.org</a>>
wrote:<br>
</div>
<blockquote class="gmail_quote" style="margin:0px 0px
0px 0.8ex;border-left:1px solid
rgb(204,204,204);padding-left:1ex">Motivation<br>
===<br>
<br>
Over the last decade we have developed an
interactive, interpretative <br>
C++ (aka REPL) as part of the high-energy physics
(HEP) data analysis <br>
project -- ROOT [1-2]. We invested a significant
effort to replace the <br>
CINT C++ interpreter with a newly implemented REPL
based on llvm -- <br>
cling [3]. The cling infrastructure is a core
component of the data <br>
analysis framework of ROOT and runs in production
for approximately 5 <br>
years.<br>
<br>
Cling is also a standalone tool, which has a
growing community outside <br>
of our field. Cling’s user community includes users
in finance, biology <br>
and in a few companies with proprietary software.
For example, there is <br>
a xeus-cling jupyter kernel [4]. One of the major
challenges we face to <br>
foster that community is our cling-related patches
in llvm and clang <br>
forks. The benefits of using the LLVM community
standards for code <br>
reviews, release cycles and integration has been
mentioned a number of <br>
times by our "external" users.<br>
<br>
Last year we were awarded an NSF grant to improve
cling's sustainability <br>
and make it a standalone tool. We thank the LLVM
Foundation Board for <br>
supporting us with a non-binding letter of
collaboration which was <br>
essential for getting this grant.<br>
<br>
<br>
Background<br>
===<br>
<br>
Cling is a C++ interpreter built on top of clang and
llvm. In a <br>
nutshell, it uses clang's incremental compilation
facilities to process <br>
code chunk-by-chunk by assuming an ever-growing
translation unit [5]. <br>
Then code is lowered into llvm IR and run by the
llvm jit. Cling has <br>
implemented some language "extensions" such as
execution statements on <br>
the global scope and error recovery. Cling is in the
core of HEP -- it <br>
is heavily used during data analysis of exabytes of
particle physics <br>
data coming from the Large Hadron Collider (LHC) and
other particle <br>
physics experiments.<br>
<br>
<br>
Plans<br>
===<br>
<br>
The project foresees three main directions -- move
parts of cling <br>
upstream along with the clang and llvm features that
enable them; extend <br>
and generalize the language interoperability layer
around cling; and <br>
extend and generalize the OpenCL/CUDA support in
cling. We are at the <br>
early stages of the project and this email intends
to be an RFC for the <br>
first part -- upstreaming parts of cling. Please do
share your thoughts <br>
on the rest, too.<br>
<br>
<br>
Moving Parts of Cling Upstream<br>
---<br>
<br>
Over the years we have slowly moved some patches
upstream. However we <br>
still have around 100 patches in the clang fork.
Most of them are in the <br>
context of extending the incremental compilation
support for clang. The <br>
incremental compilation poses some challenges in the
clang <br>
infrastructure. For example, we need to tune CodeGen
to work with <br>
multiple llvm::Module instances, and finalize per
each <br>
end-of-translation unit (we have multiple of them).
Other changes <br>
include small adjustments in the FileManager's
caching mechanism, and <br>
bug fixes in the SourceManager (code which can be
reached mostly from <br>
within our setup). One conclusion we can draw from
our research is that <br>
the clang infrastructure fits amazingly well to
something which was not <br>
its main use case. The grand total of our diffs
against clang-9 is: `62 <br>
files changed, 1294 insertions(+), 231
deletions(-)`. Cling is currently <br>
being upgraded from llvm-5 to llvm-9.<br>
<br>
A major weakness of cling's infrastructure is that
it does not work with <br>
the clang Action infrastructure due to the lack of
an <br>
IncrementalAction. A possible way forward would be
to implement a <br>
clang::IncrementalAction as a starting point. This
way we should be able <br>
to reduce the amount of setup necessary to use the
incremental <br>
infrastructure in clang. However, this will be a bit
of a testing <br>
challenge -- cling lives downstream and some of the
new code may be <br>
impossible to pick straight away and use. Building a
mainline example <br>
tool such as clang-repl which gives us a way to test
that incremental <br>
case or repurpose the already existing
clang-interpreter may be able to <br>
address the issue. The major risk of the task is
avoiding code in the <br>
clang mainline which is untested by its HEP
production environment.<br>
There are several other types of patches to the ROOT
fork of Clang, <br>
including ones in the context of
performance,towards C++ modules <br>
support (D41416), and storage (does not have a patch
yet but has an open <br>
projects entry and somebody working on it). These
patches can be <br>
considered in parallel independently on the rest.<br>
<br>
Extend and Generalize the Language Interoperability
Layer Around Cling<br>
---<br>
<br>
HEP has extensive experience with on-demand python
interoperability <br>
using cppyy[6], which is built around the type
information provided by <br>
cling. Unlike tools with custom parsers such as swig
and sip and tools <br>
built on top of C-APIs such as boost.python and
pybind11, cling can <br>
provide information about memory management patterns
(eg refcounting) <br>
and instantiate templates on the fly.We feel that
functionality may not <br>
be of general interest to the llvm community but we
will prepare another <br>
RFC and send it here later on to gather feedback.<br>
<br>
<br>
Extend and Generalize the OpenCL/CUDA Support in
Cling<br>
---<br>
<br>
Cling can incrementally compile CUDA code [7-8]
allowing easier set up <br>
and enabling some interesting use cases. There are a
number of planned <br>
improvements including talking to HIP [9] and SYCL
to support more <br>
hardware architectures.<br>
<br>
<br>
<br>
The primary focus of our work is to upstreaming
functionality required <br>
to build an incremental compiler and rework cling
build against vanilla <br>
clang and llvm. The last two points are to give the
scope of the work <br>
which we will be doing the next 2-3 years. We will
send here RFCs for <br>
both of them to trigger technical discussion if
there is interest in <br>
pursuing this direction.<br>
<br>
<br>
Collaboration<br>
===<br>
<br>
Open source development nowadays relies on
reviewers. LLVM is no <br>
different and we will probably disturb a good number
of people in the <br>
community ;)We would like to invite anybody
interested in joining our <br>
incremental C++ activities to our open every second
week calls. <br>
Announcements will be done via google group:
compiler-research-announce <br>
(<a
href="https://groups.google.com/g/compiler-research-announce"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://groups.google.com/g/compiler-research-announce</a>).<br>
<br>
<br>
<br>
Many thanks!<br>
<br>
<br>
David & Vassil<br>
<br>
References<br>
===<br>
[1] ROOT GitHub <a
href="https://github.com/root-project/root"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://github.com/root-project/root</a><br>
[2] ROOT <a href="https://root.cern"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://root.cern</a><br>
[3] Cling <a
href="https://github.com/root-project/cling"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://github.com/root-project/cling</a><br>
[4] Xeus-Cling <br>
<a
href="https://blog.jupyter.org/xeus-is-now-a-jupyter-subproject-c4ec5a1bf30b"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://blog.jupyter.org/xeus-is-now-a-jupyter-subproject-c4ec5a1bf30b</a><br>
[5] Cling – The New Interactive Interpreter for ROOT
6, <br>
<a
href="https://iopscience.iop.org/article/10.1088/1742-6596/396/5/052071"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://iopscience.iop.org/article/10.1088/1742-6596/396/5/052071</a><br>
[6] High-performance Python-C++ bindings with PyPy
and Cling, <br>
<a
href="https://dl.acm.org/doi/10.5555/3019083.3019087"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://dl.acm.org/doi/10.5555/3019083.3019087</a><br>
[7] <br>
<a
href="https://indico.cern.ch/event/697389/contributions/3085538/attachments/1712698/2761717/2018_09_10_cling_CUDA.pdf"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://indico.cern.ch/event/697389/contributions/3085538/attachments/1712698/2761717/2018_09_10_cling_CUDA.pdf</a><br>
[8] CUDA C++ in Jupyter: Adding CUDA Runtime Support
to Cling', <br>
<a
href="https://zenodo.org/record/3713753#.Xu8jqvJRXxU"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://zenodo.org/record/3713753#.Xu8jqvJRXxU</a><br>
[9] HIP Programming Guide <br>
<a
href="https://rocmdocs.amd.com/en/latest/Programming_Guides/HIP-GUIDE.html"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://rocmdocs.amd.com/en/latest/Programming_Guides/HIP-GUIDE.html</a><br>
<br>
_______________________________________________<br>
cfe-dev mailing list<br>
<a href="mailto:cfe-dev@lists.llvm.org"
target="_blank" moz-do-not-send="true">cfe-dev@lists.llvm.org</a><br>
<a
href="https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev"
rel="noreferrer" target="_blank"
moz-do-not-send="true">https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev</a><br>
</blockquote>
</div>
</blockquote>
<p><br>
</p>
</div>
_______________________________________________<br>
cfe-dev mailing list<br>
<a href="mailto:cfe-dev@lists.llvm.org" target="_blank"
moz-do-not-send="true">cfe-dev@lists.llvm.org</a><br>
<a
href="https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev"
rel="noreferrer" target="_blank" moz-do-not-send="true">https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev</a><br>
</blockquote>
</div>
</div>
</blockquote>
<p><br>
</p>
</body>
</html>