Re: [PATCH 2/2] arm: lib: xor-neon: disable clang vectorization

[Nathan Chancellor]

Hi Adrian,

On Fri, Nov 06, 2020 at 01:50:13PM +0200, Adrian Ratiu wrote:
> I tested Arnd's kernel patch from the LLVM bugtracker [1], but with the
> Clang v10.0.1 I still get warnings like the following even though the
> __restrict workaround seems to affect the generated instructions:
> 
> ./include/asm-generic/xor.h:15:2: remark: the cost-model indicates that
> interleaving is not beneficial [-Rpass-missed=loop-vectorize]
> ./include/asm-generic/xor.h:11:1: remark: List vectorization was possible
> but not beneficial with cost 0 >= 0 [-Rpass-missed=slp-vectorizer]
> xor_8regs_2(unsigned long bytes, unsigned long *__restrict p1, unsigned long
> *__restrict p2)
> 
> [1] https://bugs.llvm.org/show_bug.cgi?id=40976#c6

Ack, thanks for double checking!

> In my opinion we have 3 ways to go regarding this:
> 
> 1. Leave it as is and try to notify the user of the breakage (eg add a new
> warning). You previously said this is not a good idea because the user can't
> do anything about it. I agree.
> 
> 2. Somehow work around the compiler bug in the kernel which is what the LLVM
> bugtracker patch tries to do. This is a slippery slope even if we somehow
> get it right, especially since multiple Clang versions might be supported in
> the future and we don't know when the bug will be properly fixed by the
> compiler. In addition we're enabling and "hiding" possibly undefined
> behaviour.
> 
> 3. Disable the broken feature and once the compiler bug is fixed enable it
> back warning users of old compilers that there is an action they can take:
> upgrade. This is exactly how this was handled for GCC previously, so there
> is a precedent.
> 
> This implements the 3'rd scenario which is also the first thing Arnd
> suggested in the original patch. :)

I agree that number three is definitely the most robust against the
future. I know that I periodically grep the tree for "bugs.llvm.org"
because we always file something on LLVM's bug tracker when we have to
work around something in the kernel. I think this patch is totally fine
as is, hopefully we can get it fixed in LLVM sooner rather than later.

Cheers,
Nathan