[llvm] r262440 - revert r262424 because there's a *clang test* for AArch64 that checks -O3 asm output

[Sanjay Patel via llvm-commits]

Hi Renato -

Thanks for replying. I was going to ping the AArch64 experts about this.
There were 7 tests that broke with my patch in:
clang/test/CodeGen/aarch64-neon-misc.c

The failures are as expected from what I can tell: my instcombine change
would transform vector compare+sext into shifts. Note that this is an
existing transform, and I'm trying to extend it to catch the cases seen in
PR26701.

This is at least the 3rd time I've had a clang test fail on me after making
an LLVM optimizer change. So I admit that I wanted to just XFAIL the whole
thing, but I wouldn't do that. :)

The last time I hit this kind of problem, I left a FIXME:
http://reviews.llvm.org/rL256395

Regarding the validity of the test file itself - there was strong
indication that we shouldn't have end-to-end regression tests (even with
-O0) here:
http://lists.llvm.org/pipermail/cfe-commits/Week-of-Mon-20151123/143974.html
(cc'd Eric and Simon in case I missed anything after that).

So unless something has changed, I think we need to split each test in
aarch64-neon-misc.c (and other files like it) into pieces:
1. A clang test for the C-level intrinsics -> LLVM intrinsics and/or native
IR.
2. A codegen test for LLVM intrinsic/IR -> asm.



On Wed, Mar 2, 2016 at 2:05 AM, Renato Golin <renato.golin at linaro.org>
wrote:

> On 2 March 2016 at 01:04, Sanjay Patel via llvm-commits
> <llvm-commits at lists.llvm.org> wrote:
> > Author: spatel
> > Date: Tue Mar  1 19:04:09 2016
> > New Revision: 262440
> >
> > URL: http://llvm.org/viewvc/llvm-project?rev=262440&view=rev
> > Log:
> > revert r262424 because there's a *clang test* for AArch64 that checks
> -O3 asm output
> > that is broken by this change
>
> The question is whether the test caught something worthy or not.
>
> I'm not discussing the validity of such tests, just that we may need
> to move it down to LLVM with care, rather than ignore or remove the
> test.
>
> Which test was it and how did it fail?
>
> cheers,
> --renato
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160302/da8de5ce/attachment.html>