[llvm-dev] [RFC] [X86] Emit unaligned vector moves on avx machine with option control.
James Y Knight via llvm-dev
llvm-dev at lists.llvm.org
Wed Apr 14 11:57:52 PDT 2021
This is not a principled change -- it avoids a problem arising from *one* use
of alignment information, but there are other uses of alignment in LLVM,
and those will still cause problems, potentially less clearly. So, I think
that this will not be a useful option to provide to users, in this form.
What I suspect you *actually* want here is an option to tell Clang not to
infer load/store alignments based on object types or alignment attributes
-- instead treating everything as being potentially aligned to 1 unless the
allocation is seen (e.g. global/local variables). Clang would still need to
use the usual alignment computation for variable definitions and structure
layout, but not memory operations. If clang emits "load ... align 1"
instructions in LLVM IR, the right thing would then happen in the X86
My initial inclination is that this feature is also not
particularly worthwhile to implement, but I'm open to being convinced that
this is indeed valuable enough to be worthwhile. It should actually work
reliably, and is somewhat in line with other such "not-quite-C" flags we
provide (e.g. -fno-delete-null-pointer-checks). Of course, even with such
an implementation, you can still have a problem with user code depending on
alignof() returning a reliable answer (e.g., llvm::PointerUnion). Not much
can be done about that.
On Wed, Apr 14, 2021 at 2:07 PM Liu, Chen3 via llvm-dev <
llvm-dev at lists.llvm.org> wrote:
> Hi all.
> We want to make a patch to always emit unaligned vector move instructions
> on AVX machine with option control. We do this for the following reason:
> 1. With AVX the performance for aligned vector move and unaligned
> vector move on X86 are the same if the address is aligned. In this case we
> prefer to use unaligned move because it can avoid some run time exceptions;
> 2. This fixes an inconsistency in optimization: suppose a load
> operation was merged into another instruction (e.g., load and add becomes
> `add [memop]'). If a misaligned pointer is passed to the two-instruction
> sequence, it will
> raise an exception. If the same pointer is passed to the memop
> instruction, it will work. Thus, the behavior of misalignment depends upon
> what optimization levels and passes are applied, and small source changes
> could cause
> issues to appear and disappear. It's better for the user to consistently
> use unaligned load/store to improve the debug experience;
> 1. Makes good use of HW that is capable of handling misaligned data
> gracefully. It is not necessarily a bug in users code but a third-part
> library. For example it would allow using a library built in old ages where
> stack alignment was 4-byte only.
> 2. Compatible with ICC so that users can easily use llvm;
> Roman Lebedev is worried that this patch will hide UB. In our opinions, UB
> doesn't have to mean raise an exception. The example code(
> https://godbolt.org/z/43bYPraoa) does have UB behavior but it is still
> valid (and reasonable) to interpret that UB as `go slower',
> instead of `raise exception'. Besides, as default we still emit aligned
> instructions as before, but we provide an option for users with this need.
> We have two patches discussing this issue, one of which has been abandoned:
> https://reviews.llvm.org/D88396 (abandoned)
> https://reviews.llvm.org/D99565 (in review)
> Chen Liu.
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-dev