[llvm-dev] [RFC] Generalize llvm.memcpy / llvm.memmove intrinsics.
Philip Reames via llvm-dev
llvm-dev at lists.llvm.org
Wed Aug 19 10:14:19 PDT 2015
On 08/19/2015 09:35 AM, Pete Cooper via llvm-dev wrote:
> Hey Lang
>> On Aug 18, 2015, at 6:04 PM, Lang Hames via llvm-dev <llvm-dev at lists.llvm.org> wrote:
>> Hi All,
>> I'd like to float two changes to the llvm.memcpy / llvm.memmove intrinsics.
>> (1) Add an i1 <mayPerfectlyAlias> argument to the llvm.memcpy intrinsic.
>> When set to '1' (the auto-upgrade default), this argument would indicate that the source and destination arguments may perfectly alias (otherwise they must not alias at all - memcpy prohibits partial overlap). While the C standard says that memcpy's arguments can't alias at all, perfect aliasing works in practice, and clang currently relies on this behavior: it emits llvm.memcpys for aggregate copies, despite the possibility of self-assignment.
>> Going forward, llvm.memcpy calls emitted for aggregate copies would have mayPerfectlyAlias set to '1'. Other uses of llvm.memcpy (including lowerings from memcpy calls) would have mapPerfectlyAlias set to '0'.
>> This change is motivated by poor optimization for small memcpys on targets with strict alignment requirements. When a user writes a small, unaligned memcpy we may transform it into an unaligned load/store pair in instcombine (See InstCombine::SimplifyMemTransfer), which is then broken up into an unwieldy series of smaller loads and stores during legalization. I have a fix for this issue which tags the pointers for unaligned load/store pairs with noalias metadata allowing CodeGen to produce better code during legalization, but it's not safe to apply while clang is emitting memcpys with pointers that may perfectly alias. If the 'mayPerfectlyAlias' flag were introduced, I could inspect that and add the noalias tag only if mayPerfectlyAlias is '0'.
>> Note: We could also achieve the desired effect by adding a new intrinsic (llvm.structcpy?) with semantics that match the current llvm.memcpy ones (i.e. perfect-aliasing or non-aliasing, but no partial), and then reclaim llvm.memcpy for non-aliasing pointers only. I floated this idea with David Majnemer on IRC and he suggested that adding a flag to llvm.memcpy might be less disruptive and easier to maintain - thanks for the suggestion David!
Given there's a semantically conservative interpretation and a more
optimistic one, this really sounds like a case for metadata not another
argument to the function. Our memcpy could keep it's current semantics,
and we could add a piece of metadata which says none of the arguments to
the call alias.
Actually, can't we already get this interpretation by marking both
argument points as noalias? Doesn't that require that they don't
overlap at all? I think we just need the ability to specify noalias at
the callsite for each argument. I don't know if that's been tried, but
it should work in theory. There are some issues with control dependence
of call site attributes though that we'd need to watch out for/fix.
>> (2) Allow different source and destination alignments on both llvm.memcpy / llvm.memmove.
>> Since I'm talking about changes to llvm.memcpy anyway, a few people asked me to float this one. Having separate alignments for the source and destination pointers may allow us to generate better code when one of the pointers has a higher alignment.
>> The auto-upgrade for this would be to set both source and destination alignment to the original 'align' value.
> FWIW, I have a patch for this lying around. I can dig it up. I use alignment attributes to do it as there’s no need for alignment to be its own argument any more.
This would be a nice cleanup in general. +1
>> Any thoughts?
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
More information about the llvm-dev