[llvm-dev] Encode target-abi into LLVM bitcode for LTO.

Tue Jan 7 17:26:54 PST 2020

On Tue, Jan 7, 2020 at 3:18 PM Daniel Sanders via llvm-dev <
llvm-dev at lists.llvm.org> wrote:

>
>
> On Jan 7, 2020, at 13:57, David Blaikie <dblaikie at gmail.com> wrote:
>
>
>
> On Mon, Jan 6, 2020 at 6:05 PM Daniel Sanders <daniel_l_sanders at apple.com>
> wrote:
>
>>
>>
>> On Jan 6, 2020, at 14:29, David Blaikie via llvm-dev <
>> llvm-dev at lists.llvm.org> wrote:
>>
>>
>>
>> On Mon, Jan 6, 2020 at 5:58 AM Zakk <zakk0610 at gmail.com> wrote:
>>
>>>
>>>
>>> David Blaikie <dblaikie at gmail.com> 於 2020年1月6日 週一 下午2:23寫道：
>>>
>>>> If this is something that can vary per file in a compilation and
>>>> resolve correctly when one object file is built with one ABI and another
>>>> object file is built with a different ABI (that seems to be antithetical to
>>>> the concept of "ABI" Though) - then it should be a subtarget feature.
>>>>
>>>> ABI is generally something that has to be agreed upon across object
>>>> files - so it wouldn't make sense to link two object files with two
>>>> different ABIs. What's going on here that makes that valid in this case?
>>>>
>>>>
>>> Are you talking about that "[mips] Pass ABI name via -target-abi
>>> instead of target-features"?
>>>
>>
>> I'm not talking about that patch in particular (I have no specific
>> knowledge of mips or its implementation) - but speaking about the general
>> design of LLVM's subtarget features.
>>
>> Might be interesting to know why that change was made & may help explain
>> what's going on here.
>>
>>
>> It's been a while so I don't remember the detail but IIRC one of the
>> reasons was that mips had a feature bit per ABI and had a lot of duplicated
>> code sanity checking that only one bit was enabled and deriving the ABI
>> from the feature bits. The -target-abi option already existed and using
>> that prevented the possibility of having more than one ABI selected.
>>
>> There was a lot of code (some of which didn't have access to target
>> features) in the backend that tried to derive the ABI from the arch
>> component of the triple (e.g. mips64 => n64 ABI) even though there were
>> multiple possible ABI's for each arch (mips64 => o32, n32, or n64 ABI's)
>> and there isn't a canonical choice for any given triple (it varies between
>> linux distributions and toolchains in general). Settling on -target-abi
>> allowed us to sort out the inconsistencies in the backends opinion of what
>> the selected ABI was. It also allowed us to move the selection of the ABI
>> into the frontend where disagreements between distributions/toolchains on
>> what each triple means was easier to deal with.
>>
>
> Is this something that can vary per function in a program? (that seems
> confusing to me - ABI is usually, sort of by definition, the thing that all
> parts of the program have to agree with (at least on either side of any
> function call - I suppose different functions could have different ABIs as
> long as the function declarations carried ABI information so callers could
> cooperate, etc)) It sounds to me like that's what Zakk is
> suggesting/grappling with.
>
>
> No, it was a per-binary thing for mips and was stored in the ELF header.
> Ignoring a couple quirks*, every object in the program had to agree on the
> ABI in order to link.
>
> I'm not particularly familiar with LTO but going by the description of the
> problem it seems to me that the overall issue is that for 1, 2, and 5, each
> module fails to completely describe the contents. They each have a label
> saying it's riscv64, elf, etc. but it doesn't mention lp64d anywhere. As a
> result you can't check that you aren't trying to mix incompatible modules
> and can only trust (and require) the command line option. It's worth
> mentioning that DataLayout tends to change for different ABI's so the ABI
> is kind-of there but there isn't anything that really guarantees that
> there's a 1:1 relationship.
>
> 3 and 4 fix the problem of the missing labels but the snag with 4 is that
> target features are overridable at the function level too and that doesn't
> really make sense for ABI's (it's fine for calling conventions but that's
> only part of the ABI and calling conventions are described elsewhere in the
> IR anyway). Without changing the IR, 3 looks like the only one that solves
> the overall problem but then you have potential for problems where the
> official triple for a platform doesn't match what needs to be in the triple
> metadata in the IR. For example, mips64-linux-gnu can be N32 or N64 ABI (or
> more rarely O32) depending on the OS/distribution/toolchain/version. FWIW,
> back when I worked on it, we were generally moving towards the idea of
> canonical triples which contained the ABI and some lowering code on the
> user facing interfaces to disambiguate things like mips64-linux-gnu to
> mips64-linux-gnuabin32.
>
>
To reply here a bit:

I worry about target triple being used, but I think I do/did agree that
it's probably the best we can move to in the near term. My concern is that
we will have diverged from more "canonical" triples that are used in other
places just for our compilation model. I'd love to be able to encode ABI
into the module in some way and make it an error to link two modules that
have incompatible ABIs and am definitely up to ideas on how to encode
target specific module level data into the module. I'd like to avoid
metadata if possible. Any thoughts here on how you'd like to see it encoded
for the long term?

> *Just for completeness, the quirks I can remember off-hand were:
> - IEEE754 1985 and 2008 would successfully cross-link unless you used a
> flag indicating that it mattered. This was because we wanted to omit the
> 1985 standard from newer chips but there were many ecosystems using it due
> to historical reasons. In practice, very few programs care about the tiny
> details (does negation trap, etc.) so we essentially force-migrated whole
> ecosystems by relaxing the link requirements and changing the default.
> - Along the same lines, we also supported cross-linking specific variants
> of the O32 ABI. There was only supposed to be one O32 but an unfortunate
> mis-reading of the ABI spec coupled with a failure to catch it with
> conformance tests split it in two. Luckily, Matthew Fortune found a way to
> reunite them without breaking either one by adding a third that followed
> the original intent of the spec and was compatible with either one (but not
> both at once) and then migrating everyone to that.
>
>
*sigh* I remember that. :)

Thanks for chiming in Daniel!

-eric

> If it can vary per function, then the ABI information shouldn't be used
> outside the per-function context (ie: no global variables/other output
> could depend on the ABI because which function's ABI would it depend on?).
>
>
>>
>> I don't know WHY -target-abi is passing via different option, not via
>>> -mattr (subtarget feature)
>>> maybe usually subtarget feature is used to manages different specific
>>> ISA.
>>>
>>>
>>>
>>>> On Sun, Jan 5, 2020 at 10:04 PM Zakk via llvm-dev <
>>>> llvm-dev at lists.llvm.org> wrote:
>>>>
>>>>> Hi all.
>>>>>
>>>>> There are two steps in LTO codegen so the problem is how to pass ABI
>>>>> info into LTO code generator.
>>>>>
>>>>> The easier way is pass -target-abi via option to LTO codegen, but
>>>>> there is linking issue when linking two bitcodes generated by different
>>>>> -mabi option. (see https://reviews.llvm.org/D71387#1792169)
>>>>>
>>>>> Usually the ABI info for a file is derived from target triple, mcpu or
>>>>> -mabi, but in RISC-V, target-abi is only derived from -mabi and -mattr
>>>>> option, so the one of solutions is encoding target-abi in IR via LLVM
>>>>> module flags metadata.
>>>>>
>>>>> But there is an another issue in assembler. In current LLVM design,
>>>>> there is no mechanism to extract info from IR before AsmBackend
>>>>> construction, so I use some little weird approach to init target-abi option
>>>>> before construct AsmBackend[1] or reassign target-abi option in
>>>>> getSubtargetImpl and do some hack in backend[2].
>>>>>
>>>>> 1. https://reviews.llvm.org/D72245#change-sHyISc6hOqcy (see llc.cpp)
>>>>> 2. https://reviews.llvm.org/D72246 (see RISCVAsmBackend.h)
>>>>>
>>>>> I think [1] and [2] are not good enough, the other ideals like
>>>>>
>>>>> 3. encode target abi info in triple name. ex. riscv64-unknown-elf-lp64d
>>>>> 4. encode target-abi into in target-feature (maybe it's not a good
>>>>> ideal because mips reverted this approach
>>>>> before.
>>>>> http://llvm.org/viewvc/llvm-project?view=revision&revision=227583)
>>>>>
>>>>> 5. users should pass target-abi themselves. (append
>>>>> -Wl,-plugin-opt=-target-abi=ipl32f when compiling with -mabi=ilp32f)
>>>>>
>>>>> Is it a good idea to encode target-abi into bitcode?
>>>>> If yes, is there another good approach to fix AsmBackend issue?
>>>>> I’d appreciate any help or suggestions.
>>>>>
>>>>> Thanks.
>>>>>
>>>>> --
>>>>> Best regards,
>>>>> Kuan-Hsu
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> LLVM Developers mailing list
>>>>> llvm-dev at lists.llvm.org
>>>>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>>>
>>>>
>>>
>>> --
>>> Best regards,
>>> Kuan-Hsu
>>>
>>>
>>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>
>>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200107/2fdfe4c2/attachment.html>