[llvm-dev] RFC: Generate plain !tbaa tags in place of !tbaa.struct ones
Hal Finkel via llvm-dev
llvm-dev at lists.llvm.org
Wed Nov 1 20:54:23 PDT 2017
On 10/31/2017 05:02 AM, Ivan Kosarev wrote:
> To clarify further, what this paper proposes is to use !tbaa for all
> kinds of accesses, including aggregate ones, so we don't need to
> bother trying to convert them when an aggregate access becomes a
> series of scalar accesses or vice versa. As I said, in most cases such
> conversions are not possible anyway, because !tbaa.struct tags do not
> refer to the type of the aggregate itself and only enumerate its members.
That makes sense, but I don't believe you actually said *that*. I agree,
the fact that the tbaa.struct metadata refers to the scalar types, and
not the type itself, will prohibit the conversion in general.
I agree that we should fix this. Given that we're planning to make
changes, perhaps major ones, to the TBAA representation, we should
account for this requirement as part of that larger discussion. I see
no need for a separate thread on this particular topic in the context of
those larger changes.
That having been said, if we desire to fix this before those larger
changes to the TBAA representation (to handle aggregates, unions, etc.),
I think we should do that. In this case, however, we can/should make a
small enhancement that addresses this representational deficiency. We
can, for example, generate tbaa.struct metadata where the tbaa struct
type is the first in the list. This is easy to distinguish from the
existing format because in the existing format the first operand is an
integer. This will provide the necessary information to allow SROA to
create struct access tags for the scalarized accesses.
> As of today, there is no adequate representation that could be used to
> describe accesses to individual slices of scalarized aggregates. But
> at very least we could decorate accesses to slices with tags of their
> aggregates. That requires aggregate accesses to be described the same
> way we describe scalar accesses.
> To provide some background on importance of this matter, SROA and
> passing flattened parameters/returning values in clang together
> produce 75% of all undecorated loads and 90% if all undecorated stores
> on the LLVM code base under -O1.
> On 31/10/17 10:44, Ivan Kosarev wrote:
>> In short, the problem with !tbaa.struct is that in most cases it
>> cannot be converted to !tbaa. For transformations like SROA this
>> means they cannot propagate !tbaa.struct tags for aggregate-accessing
>> instructions like memcpy() calls to the resulting loads and stores.
>> On 31/10/17 05:11, Hal Finkel wrote:
>>> On 10/18/2017 05:49 AM, Ivan Kosarev via llvm-dev wrote:
>>>> The motivation behind this proposal is to make it possible to
>>>> propagate TBAA information through scalarizing transformations,
>>>> such as SROA, that rewrite accesses to aggregates, e.g., memcpy()
>>>> calls, into accesses to scalars, that is, load and store
>>>> Currently, we decorate instructions that initialize and copy
>>>> aggregates with !tbaa.struct tags that generally cannot be
>>>> transformed to !tbaa tags. For this reason every time we
>>>> scalarize an aggregate, we leave resulting loads and stores
>>>> undecorated, meaning optimization of such instructions cannot
>>>> benefit from TBAA.
>>> I understand that we'd like to improve TBAA in a number of different
>>> ways, and that probably involves replacing !tbaa.struct with an
>>> enhanced !tbaa representation, but I don't understand the point of
>>> this proposal. As you state, there are lots of places that could use
>>> !tbaa.struct information, such as SROA, but don't. Is the current
>>> !tbaa.struct missing some information that is essential for that
>>>> Furthermore, our analysis indicates that the only place where
>>>> !tbaa.struct tags may potentially impact code generation is
>>>> simplification of memcpy() and memmove() calls, see
>>>> SimplifyMemTransfer() in InstCombineCalls.cpp. Ironically, what
>>>> the code that makes that sole use of such tags is trying to do is
>>>> to construct a !tbaa tag from the information encoded in the
>>>> given !tbaa.struct tag. Note that it can only do that if the
>>>> !tbaa.struct tag describes a structure with a single member of a
>>>> scalar type.
>>>> Here's how we propose to resolve the issue in terms of specific
>>>> 1. Extend the TBAA facilities in clang to support aggregate types
>>>> as final access types. This patch:
>>>> [CodeGen] Propagate may-alias'ness of lvalues with TBAA info
>>>> implements this for the needs of fixing issues with
>>>> propagation of TBAA information, which in turn is necessary to
>>>> support TBAA for unions. So once this patch is committed, this
>>>> step is done.
>>>> 2. Generate !tbaa tags in addition to !tbaa.struct tags for
>>>> aggregate accesses.
>>>> 3. Fix the TBAA analysis to support aggregate accesses as
>>>> explained in this proposal:
>>>> RFC: Resolving TBAA issues
>>>> 4. Switch the code that simplifies memcpy() and memmove() calls
>>>> to !tbaa tags.
>>>> 5. Remove the support for !tbaa.struct tags.
>>>> I guess we would want to remove all code which only purpose is
>>>> to generate such tags, but I'm not sure what we should do with
>>>> the MD_tbaa_struct enumerator. Some possible options:
>>>> a) Leave it as is, just don't use it.
>>>> b) Rename to something like MD_unused.
>>>> c) Remove, but do not change the values of other MD_*
>>>> d) Remove and adjust values of other MD_* enumerators
>>>> Or, maybe we want some multi-stage plan here?
>>>> Further steps are supposed to include things like fixing SROA to
>>>> propagate TBAA information.
>>>> Any feedback is highly appreciated.
Lead, Compiler Technology and Programming Languages
Leadership Computing Facility
Argonne National Laboratory
More information about the llvm-dev