[LLVMdev] Target-specific defaults for LLVM tools?

[Sean Silva]

On Tue, Mar 17, 2015 at 11:13 AM, Dario Domizioli <dario.domizioli at gmail.com
> wrote:

> Hello, LLVM.
>
> We'd like to start a discussion about what would be the best way of
> supporting target-specific defaults for options in the LLVM tools.
>
> Problem description
> -------------------
>
> LLVM strives to be a generic compiler and it has a lot of code generation
> options which are (rightly) not linked to the target triple.
> However, some target triples do impose hard requirements for such options.
> For
> example, PS4 requires the relocation model to be PIC. Cross-compilers for
> specific target systems are likely to have similar restrictions.
>
> One very effective way to ensure that users of Clang/LLVM don't misuse the
> tools by passing the wrong options is to set the default values for those
> options in the cross-compiler once the target triple is known. This has the
> following rationale:
>
> - Correctness: the user must be able to generate correct code for the
> target
>   in the default workflow (with or without optimizations).
> - Performance: the user should be able to generate good code for the target
>   in the default workflow with optimizations enabled.
> - Usability: although the default workflow could potentially be defined as
>   "sorry, but you must pass options X, Y and Z to the tools", obviously it
> would
>   be easier if it did not require such esoteric options.
>
> And of course, from a company's point of view:
>
> - Commercial reasons: if a tool is released commercially, the user expects
> such
>   commercial product to work "out of the box" for the target system.
>
> This is very easy to do in Clang, as the driver does the job of
> "dispatching"
> the correct command line options to subsequent invocations of tools.
> Because of
> the driver being the "middle man", it can parse the options, understand
> what the
> target triple is, and invoke other tools accordingly (i.e. actively adding
> an
> option, or erroring out upon encountering an invalid set of options).
>
> A vendor can set the default target triple for Clang even at build time,
> and
> this approach seems to not cause any trouble in the Clang tests (which
> don't get
> as far as checking the output assembly). So for Clang the problem is
> solved to a
> certain degree.
>
> However, issues start to pop up when we consider all the other tools (opt,
> llc,
> and so on) and the complications of having for example a "cross-llc".
> Usage of
> such tools is becoming increasingly likely thanks to the existence of other
> tools and frontends which can generate LLVM IR. Some crazy people might
> even try
> to write IR instead of C++! :-)
>
> First of all, overriding the defaults based on the triple is harder in LLVM
> tools, because of how the cl::opt options work. If you consider llc, it
> determines the triple in this sequence:
>
> - Parse the IR and extract the triple if present.
> - If there's a triple on the command line, override the IR triple.
> - If no triple was found by either method, use the default triple.
>
> This process is performed in the compileModule() function, which happens
> way
> after the cl::init initializers for the cl::opt option defaults, and
> shortly
> after the cl::ParseCommandLineOptions() call. If the value of some option
> needs
> to be changed _after_ the triple has been determined, additional work has
> to be
> performed to tweak the cl::opt options on the fly, but only if they had not
> already been set on the command line (which would mean the user explicitly
> wanted a non-default value). It would be quite ugly, although doable.
>
> The real problem, however, is that tools are used not only by the end
> user, but
> also extensively by the LLVM tests that check the output assembly.
>

You have this completely backwards, and then some. The use case for these
tools for testing is not "also". That is the only current use case that
they are developed for. I'm not sure where you got the impression that they
are meant to be used by end users, but that is just completely, flat-out
incorrect.

-- Sean Silva


> Currently, there is a non-trivial number of LLVM tests that do not set the
> full
> triple, but only parts of it, and then rely on specific default code
> generation
> options. In the past, some defaults have been locked down as opposed to
> being
> auto-detected (for example, -mcpu=generic) specifically for the purpose of
> making tests stable and correct for cross-compilers.
>
> Changing the code to override the defaults for opt, llc, and so on based
> on the
> target triple can make some tests fail, because it would reinstate the same
> problems we had before locking down such defaults. Of course, such a change
> could be kept private and out of the open-source tree, and the tests could
> be
> modified accordingly. However, it is clear that it could become a large
> burden
> to support these private changes when new tests are constantly added to the
> codebase.
>
> We would like to start a discussion with the community about how best to
> support
> the use case of "target-specific defaults for LLVM tools", especially in
> the
> case of cross-compiling tools. Maybe somebody else is running into our same
> use-case.
> And this could even prove to be tied to the cl::opt reworking that was
> going on
> a while ago, so new requirements / limitations / ideas might emerge.
>
> We have identified a few approaches that could serve as the starting point
> of a
> discussion. I will briefly outline them below. We don't want to push for
> any of
> them, we just want to discuss them. There could even be more effective
> ways of
> solving the problem, which is why we are interested in having this
> discussion
> here.
>
>
> A - Fixed option defaults, irrelevant default triple logic
> ----------------------------------------------------------
>
> I believe this is the current situation, possibly caused by how cl::opt
> works.
> This alternative assumes that:
>
> - There is no functionality in the open-source tree to change the defaults
> based
>   on the target triple.
> - Vendors may or may not change the default triple in their
> cross-compilers.
>
> If a vendor wants to both change a tool to have a specific default triple,
> and
> cause option defaults to depend on the triple, it is the vendor's
> responsibility
> to adapt failing tests, and changes will have to be kept private. If a
> vendor
> tried to push them onto the community ("yes, this -mcpu=generic in the
> test is
> totally unnecessary, but can we have it in tree?"), such a change would
> never
> pass a review.
>
> If this approach is chosen, then vendors will have to maintain private
> changes,
> but the state of the LLVM tests in the open-source tree will be consistent
> and
> there will be no target-specific code to override option defaults.
>
>
> B - Triple-based option defaults, variable default triple logic
> ---------------------------------------------------------------
>
> This alternative is at the other end of the spectrum and it assumes that:
>
> - There is some functionality (a new API in llvm::cl?) in the open-source
> tree
>   to override the defaults based on the target triple.
> - The default target triple can be freely changed in cross-compilers, with
>   impunity.
>
> As described above, this would cause problems to some existing tests. It
> could
> be argued though that, if a test is unstable with respect to code
> generation
> options, then it should specify all options that it actually requires in
> order
> to run properly.
> This would mean that tests must specify a target triple in the RUN lines
> or in
> the IR, otherwise they cannot assume that the code will be generated as
> expected.
>
> This puts a burden on the test implementors, but could it be argued that
> this is
> technically correct? Shouldn't tests be robust? On the other hand, if
> tests had
> to specify a target triple, wouldn't this reduce coverage (as they'd be
> testing
> a single target as opposed to many)?
>
> If this approach was chosen, then a good number of LLVM tests in the
> open-source
> tree would need to be adapted, but vendors could be free to implement
> their own
> defaults and set the default triple without significant consequences.
>
>
> C - Triple-based option defaults, fixed default triple logic
> ------------------------------------------------------------
>
> This alternative assumes that:
>
> - There is some functionality (a new API in llvm::cl?) in the open-source
> tree
>   to override the defaults based on the target triple.
> - Vendors are NOT free to change the logic determining the default triple.
> - The default triple is inferred by some fixed logic, which is
> test-friendly.
>
> There is code in the Clang driver that parses the program name to identify
> a
> prefix that is used to determine the target triple. For example, I seem to
> understand that "x86_64-scei-ps4-clang" would indicate a Clang which has a
> PS4
> target triple by default.
> Opt, llc, and similar tools do not support anything of the sort, so at the
> moment the default target triple cannot be selected with the same approach.
>
> Anyway, consider the case where option defaults are set based on the target
> triple, but the default triple can be inferred from the name prefix, just
> as
> Clang does; if no such prefix is found, the defaulting falls back to what
> we do
> currently. This allows tests with RUN lines running "llc" to just use the
> triple
> that we currently use, and therefore all existing open-source tests will
> continue to work.
>
> At the same time, a vendor could release a tool with a name prefix and
> reap the
> benefits of having the desired defaults; and to make it convenient for the
> user,
> a symlink (or something more complex on Windows) could easily be used.
>
> If this approach was chosen, vendor-specific tests which expect the vendor
> defaults would have to somehow rename the tool on the fly before running
> it; or
> maybe they could change the build process to produce both an "llc" and a
> "<triple>-llc" executables. Also vendors would need to release the renamed
> tools, but this often happens anyway with cross-compilers. This pushes some
> burden on the vendor, but it is much less than maintaining a lot of private
> changes to the open-source code and tests.
>
> What does the community think?
> Discuss. :-)
>
> Cheers,
>     Dario Domizioli
>     SN Systems - Sony Computer Entertainment Group
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150317/acb37c3f/attachment.html>