[llvm] r310510 - [SimplifyCFG] remove checks for crasher test from r310481

[Davide Italiano via llvm-commits]

On Thu, Aug 10, 2017 at 12:43 AM, Davide Italiano <davide at freebsd.org> wrote:
> On Wed, Aug 9, 2017 at 12:42 PM, Sanjay Patel <spatel at rotateright.com> wrote:
>> The IR difference is real - it seems to depend on whether you're building
>> debug or release.
>>
>> This didn't change if I specified triple+layout with:
>> target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
>> target triple = "x86_64-unknown-linux-gnu"
>>
>> Here's what I show using a debug build of opt on x86-64 on macOS (note:
>> different blocks, different phi, etc):
>
> While I think this is unrelated to the change, it's an annoying piece
> of non-determinism.
> I'll try to shed some light on it today.
>
> Thanks,
>

I'm finally able to reproduce this.
It reproduces only on MacOS (not on Linux nor Windows), when building
trunk (debug/release) with the following compiler:

$ clang --version
Apple LLVM version 7.3.0 (clang-703.0.31)
Target: x86_64-apple-darwin15.4.0
Thread model: posix
InstalledDir: /Applications/Xcode.app/Contents/Developer/Toolchains/XcodeDefault.xctoolchain/usr/bin

Does this match the version you have? I'll consider building with a
different compiler on Mac (e.g. GCC) because this could possibly be a
miscompile introduced by the LLVM clang shipped by Apple.

Thanks,

--
Davide