[cfe-dev] [RFC][OpenCL] Add builtin types and functions from the standard headers implicitly in the driver
James Y Knight via cfe-dev
cfe-dev at lists.llvm.org
Fri Feb 5 14:12:43 PST 2021
The whole tablegen thing seems like a sad path to have gone down, although
I can certainly see the practical benefits. Substantially the same problem
of compilation-speed exists for <x86intrin.h> (and friends), and I really
don't think we want to start defining all of *those* with a tablegen rule.
It would be really nice to instead somehow take advantage of the modules
infrastructure to fix this problem -- I'd really love it if we could
somehow start shipping a pre-built module artifact for our giant intrinsics
headers, included with the compiler distribution. And then use that by
default, regardless of whether users are otherwise enabling modules. If we
got that to work, we could use the same solution for both X86 and opencl.
However, independent of that discussion -- and more to the immediate issue
you're trying to raise -- your proposal seems like it's resulting in a very
complex set of options, and I'm not sure what the purpose is.
IIUC, the overall desire is to have, by default, these tens-of-thousands of
prototypes available to all OpenCL compilations. But, I don't see any
reason why users should care exactly HOW these are provided. I'd expect
that a given prototype should be provided either by the tablegen-based
builtin code, OR an auto-included header file -- not both, and not neither.
What difference does it make if the builtin-tablegen code doesn't provide
100% of the declarations, so long as the remainder are provided by an
automatically-included header? Why do you want to make users choose between
getting a half-baked set of function prototypes (tablegen version) and slow
compilation (auto-including giant header), when you can have the fully
correct set of functions AND nearly-as-fast compilation, by simply
supplementing tablegen with an auto-included header providing the remainder?
And then you need just a single user-visible option: to disable the
automatic declarations (via both tablegen + autoinclude).
On Wed, Feb 3, 2021 at 11:58 AM Anastasia Stulova via cfe-dev <
cfe-dev at lists.llvm.org> wrote:
> Hello,
>
> I would like to check if there is any feedback to the following proposal
> for
> improving the interface of standard type and function includes.
>
> Background
>
> Currently the default clang driver provides incomplete functionality for
> OpenCL
> because the headers with builtin function declarations are not included by
> default. The header can only be added using frontend flags requiring
> ‘-cc1’ or
> other frontend forwarding options
> (https://clang.llvm.org/docs/UsersManual.html#opencl-header). This means
> it
> is inaccessible to users in a conventional way.
>
> We propose to add the implicit header by default when a source is compiled
> in
> OpenCL mode. A review for this has been uploaded by Matt a few months ago:
> https://reviews.llvm.org/D78979. Note that the standard library
> functionality is
> added by default in OpenCL C without using include directives in the
> compiling
> sources. This means all builtin function declarations (there are 17000 of
> them) are to be parsed every time the source is compiled because which
> builtins
> are used by the kernels is not known beforehand. This impacts the
> compilation
> speed. For example, parsing a simple kernel with the builtin function
> declarations is 138 times slower in a Debug build and 13 times slower in a
> Release build than parsing the same kernel without those declarations.
>
> To mitigate the overhead of parsing the full header, a fast Tablegen based
> solution has been developed
> (https://llvm.org/devmtg/2019-10/talk-abstracts.html#lit5). The parsing
> speed with
> this mechanism for builtin function declarations is only 1.3 times slower
> in a Debug
> build and 1.05 times slower in a Release build compared to clang without
> the
> builtins. While this mechanism covers most of OpenCL standard
> functions it lacks two main classes of builtins: builtins defined by
> vendor extensions and builtins with enum arguments.
>
> Proposal
>
> We propose the following changes in the clang driver interface for OpenCL:
> - Enable the fast Tablegen based builtin function declaration mechanism by
> default in the clang driver. This makes the majority of OpenCL builtin
> functions available.
> - In addition, include the small header opencl-c-base.h by default in the
> clang
> driver. This header provides basic types and constants.
> No frontend or driver flags will be needed to allow using the majority of
> OpenCL
> types and functions from the standard, at a very low parsing speed
> increase.
>
> Since the Tablegen mechanism has some small overhead and it is not fully
> complete, we propose to add the following additional clang driver flags:
> 1. Add a new clang driver flag -cl-no-stdinc (*) that disables such extra
> includes to
> minimize further compilation speed or for the use cases that don’t require
> standard libraries. The majority of OpenCL clang tests will use this
> option.
> 2. [Optionally, if there is enough interest] Add a new clang driver flag
> -cl-all-stdinc (*) that will include the full header instead of using the
> Tablegen
> mechanism, at the cost of a significant increase in parsing time.
>
> At present we propose no change to the ‘cc1’ interface, but in the future
> it is
> expected that the functionality will be aligned between driver and
> frontend
> command line interfaces for the OpenCL headers.
>
> (*) The exact spelling of command line options is to be discussed.
>
> Summary
>
> This proposal enhances the clang driver with full functionality of OpenCL
> C by
> adding builtin function declarations implicitly without the need for any
> frontend flags to be given in the command line.
>
> The default clang behavior proposed is not expected to negatively impact
> users
> of clang as the parsing speed difference remains within the same order of
> magnitude. While the fast header mechanism matures, a fallback mechanism
> will
> be provided if needed that would allow switching to the slow header with
> the
> full functionality using a new driver flag. For backward compatibility,
> another
> flag is provided to disable all OpenCL declarations that are not builtin
> into
> the compiler.
>
> The solution proposed improves the driver interface and reduces risks of
> forcing the OpenCL community to update their use of clang due to
> significant
> regression of the compilation speed.
> _______________________________________________
> cfe-dev mailing list
> cfe-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20210205/038aeebd/attachment.html>
More information about the cfe-dev
mailing list