[PATCH] D52548: Stop instcombining propagating wider shufflevector arguments to predecessors.

[Sanjay Patel via llvm-commits]

On Sat, Sep 29, 2018 at 7:52 AM Jan Vesely <jan.vesely at rutgers.edu> wrote:

>
> it creates <3 x i8>, which is then shuffled into another <3 x i8>, which
> is then bitcast to <4 x i8>:
>
>   %vecinit.i = insertelement <3 x i8> undef, i8 %2, i32 0
>   %vecinit1.i = insertelement <3 x i8> %vecinit.i, i8 %3, i32 1
>   %vecinit4.i = insertelement <3 x i8> %vecinit1.i, i8 %4, i32 2
>
>   %extractVec = shufflevector <3 x i8> %vecinit4.i, <3 x i8> undef, <4 x
> i32> <i32 0, i32 1, i32 2, i32 undef>
>   ^^^ the shuffle result should probably be <4 x i8> ????
>

Thanks for reporting the bug. The result of that shuffle is <4 x i8>; the
mask has 4 elements:
http://llvm.org/docs/LangRef.html#shufflevector-instruction

So I think we've exposed a backend bug. Can you file a bug report and
attach the test files there? Given that this patch fixes a miscompile in
IR, I'd rather not revert it unless we can't correct the backend in some
reasonable amount of time.

The test case also suggests that we can limit the fix that this patch
imposed: if we're shuffling only to extend a series of insertelement ops,
then that's a good candidate for widening because we are not introducing
any wider math ops. That's a workaround for the backend bug in this
particular case, but it doesn't solve the root cause of the crashing.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180929/76dd4f38/attachment.html>