r254251 - ARM v8.1a adds Advanced SIMD instructions for Rounding Double Multiply
David Blaikie via cfe-commits
cfe-commits at lists.llvm.org
Tue Dec 1 11:09:07 PST 2015
On Tue, Dec 1, 2015 at 10:43 AM, Renato Golin <renato.golin at linaro.org>
> On 1 December 2015 at 18:29, David Blaikie <dblaikie at gmail.com> wrote:
> > Are they things the test-suite couldn't (either technically or
> > philosophically) cover, or only that it doesn't cover it at the moment,
> > could do so?
> IMO, it's a philosophical issue. The test-suite is a whole-program
> execution, and all that matters is the end result, the output of the
> program, not the program itself.
> Those tests are mostly large (> 3 lines of code) so, if we were trying
> to objdump and match for instruction patterns, any optimisation
> variance could destroy them, as well as you may find the pattern
> elsewhere in the same object, not belonging to the function you want
> (inlining, etc).
Not sure I follow - I'm not suggesting adding objdump/instruction checking
to existing large programs in the test-suite, but adding other tests to the
test-suite that do this and have appropriate input for it to be a
reliable/useful form of testing.
> All in all, all the problems with reducing C code
> would be pertinent here, so a big hammer to kill small issues. On the
> other hand, the Clang based asm/obj examples are all small, and
> hand-crafted to suit the problem being tested.
> We could, however, add a new directory in the test-suite for that kind
> of tests, like we already have for regressions in C/C++.
If a separate directory makes it more clear what the purpose is, sure,
seems fine to me, but I don't have strong or informed opinions on exactly
how the test-suite is laid out.
> That would
> dissociate the front-end from the back-end, and essentially any
> compiler used would have to abide by the rules of the test (which may
> not be true for some compilers).
This is perhaps an interesting part/point: In theory I would think that
these tests should be able to be phrased in a compiler-agnostic way, since
they're testing the compiler as a whole and testing behavior that the
source code demands, no?
> I don't mind where they end up, really, but we need that kind of test
> somewhere. Clang is an easy place because we know we have it built
> when the Clang repo is checked out.
> > That would sort of defeat the point of having the testing and projects
> > separated though - it would tie the tests together and produce most of
> > undesirable outcomes of having single end-to-end tests.
> Right, I agree, but since the relationship between Clang and LLVM is
> non-trivial, and there are lots of changes that need to be done on
> both sides, I can't see why we shouldn't have tests that span across
> Clearly, the sanitizer tests need Clang and Compiler-RT. Some Libc++
> tests need Compiler-RT, others need libunwind, on ARM I need Clang to
> build it. There are a lot of dependencies that are not there by
> accident, but by design.
> The more you move tests away from the big targets (LLVM, Clang), the
> less they're ran by people committing patches, and the harder will be
> to pick up the failures. Having said that, a decent buildbot coverage
> would account for most of those problems, but it would also put a
> stress on slow/expensive/experimental hardware support. It's a
> balance, I think.
Why would it require slow/expensive/experimental hardware? (& if it does
require that, how do you expect average developers to run them on a regular
> > (it would mean that at least the pure (or at least non-Clang) LLVM
> > would have a test to run where they would not if the test remained in
> > only)
> That is a good point. Right now, RT/libc++ tests in that category are
> controlled by enabling/disabling support via CMake flags. We may be
> able to do the same with Clang (if LLVM is not built with it, disable
> asm/obj tests). Moving them all to a regression package in the
> test-suite would be another option, but one that would take a lot
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the cfe-commits