[llvm-dev] Representing MIPS ABI information in the triple as ARM/X86 do for EABI/EABIHF/X32

[Daniel Sanders via llvm-dev]

Hi Eric,

It's the unsolved problems on the pass-MCTargetOptions-everywhere path that are my main concern with that approach rather than the amount of work. The first problem is that the result of IRObjectFile::CollectAsmUndefinedRefs() depends on the ABI but IRObjectFile doesn't know it. How would you deliver the ABI to IRObjectFile? The second problem is that IRLinker will link incompatible IR together because it doesn't know the ABI's involved in the link. It thinks it's sufficient to check that the triples are the same. How would you prevent it from linking IR intended for O32 with IR intended for N64?

I'm also worried about the risk of undiscovered problems on the pass-MCTargetOptions-everywhere path but it's difficult to discuss those for obvious reasons. However, I think it's safe to say that there could be some major problems waiting to be discovered given that we don't have any targets that implement a major ABI with this approach.

> There's already some ABI support in MCTargetOptions that I used when fixing up the ARM port a while back and this use can either be extended or changed depending on what's necessary

That's correct and r224492 is quite similar to what I'm trying to do in this patch series (see below). I should mention that even that commit relies on the ABI being encoded in the triple though. In computeTargetABI() from ARMTargetMachine.cpp, MCTargetOptions::ABIName specifies the ABI if it is a non-empty string but the code otherwise falls back on extracting the ABI from the triple. The clang driver has similar code, it determines the value for the –target-abi option from –mabi if it's given and falls back on calculating it from the triple using the same code as computeTargetABI(). The overall effect is that the ABI information is delivered to some parts of LLVM via both MCTargetOptions::ABIName and the triple, and other parts solely via the triple (e.g. ARMAsmPrinter and ARMISelLowering). Both copies of this information always agree when –mabi is not given but they may differ when –mabi is given.

As I mentioned above, I'm trying to do a very similar thing to r224492 but I have an additional problem to deal with. In my scenario, clang doesn't start off with a triple that clearly specifies the ABI (mips64-linux-gnu can be N32 or N64 depending on –mabi) which makes it impossible for the bits of LLVM that can't see MCTargetOptions to know what to do.  I therefore start by normalizing the triple in clang to one that does specify the ABI clearly and is consistent with the –mabi option. With that problem solved, the rest of the patch series is consistent with r224492 except that MCTargetOptions::ABIName and the triple can never disagree. Personally, I would stop the patch series at this point and not go on to duplicate the ABI information into MCTargetOptions::ABIName (because this permits inconsistency and it doesn't make sense to put N64 code inside an O32 ELF) but if you feel strongly about this then I can add another patch that causes clang to pass the information in both the triple and the –target-abi option and have the Mips backend accept this. With that additional patch, this patch series will be completely consistent with r224492.

> Even if it weren't an API problem I'm not sure it'd be a good idea. I think we could add a with-abi configuration option to match the current gcc one rather than tacking it onto triple parsing.

I thought you were strongly against configure-time options based on previous discussions. Has this changed?

> ps. It also doesn't appear that Matthew (who took over my gcc responsibilities, thanks!) has added support for this in gcc yet either.
> echristo at dzur ~/b/build-gcc-mips64abin64> .../configure --target=mips64-linux-gnuabi64 --enable-languages=c
> echristo at dzur ~/tmp> ~/builds/build-gcc-mips64abin64/gcc/cc1 foo.c -o -
>          .file      1 "foo.c"
>          .section .mdebug.abiN32

It's true that upstream GCC doesn't know about this triple yet but that's because downstream users of GCC don't need to push their changes upstream when they redefine triples. Fedora and Gentoo both use --with-abi while Debian uses source patches. Upstream GCC only contains the default-defaults which may be different from the actual-defaults that are compiled in to the toolchain. In contrast, LLVM requires all the (often conflicting) meanings of a triple to be upstream. This is part of a separate problem that I plan to address in a later discussion so I'll leave this for another thread.

From: Eric Christopher [mailto:echristo at gmail.com]
Sent: 01 July 2016 23:05
To: Daniel Sanders; llvm-dev at lists.llvm.org
Cc: rafael.espindola at gmail.com; Matthew Fortune
Subject: Re: Representing MIPS ABI information in the triple as ARM/X86 do for EABI/EABIHF/X32

Hi Daniel,

While I'm sympathetic to the amount of work associated this appears to be attempting to work around an API problem by changing triples needlessly. There's already some ABI support in MCTargetOptions that I used when fixing up the ARM port a while back and this use can either be extended or changed depending on what's necessary. It may require plumbing, but that's expected with an API being expanded for new uses. It appears that the command line syntax is sufficient for the needs of enumerating target plus abi and if we need to add more defaults then it shouldn't be too bad to do so.

Even if it weren't an API problem I'm not sure it'd be a good idea. I think we could add a with-abi configuration option to match the current gcc one rather than tacking it onto triple parsing.

That said, I'm more than happy to help you detangle the existing API and come up with a good strategy for the MC level API.

Thanks!

-eric

ps. It also doesn't appear that Matthew (who took over my gcc responsibilities, thanks!) has added support for this in gcc yet either.

echristo at dzur ~/b/build-gcc-mips64abin64> .../configure --target=mips64-linux-gnuabi64 --enable-languages=c
echristo at dzur ~/tmp> ~/builds/build-gcc-mips64abin64/gcc/cc1 foo.c -o -
          .file      1 "foo.c"
          .section .mdebug.abiN32
...

On Fri, Jun 24, 2016 at 3:02 AM Daniel Sanders <Daniel.Sanders at imgtec.com<mailto:Daniel.Sanders at imgtec.com>> wrote:
Hi,

Having recently enabled IAS by default for the MIPS O32 ABI, I'm now trying to do the same thing for the MIPS N64 ABI. Unfortunately, it is not currently possible to enable IAS by default for the N64 ABI without also enabling it for the N32 ABI because this information is not reflected in the triple and that's the only information MipsMCAsmInfo has. This would be fine if it N32 was also in a good state but the current N32 ABI support for IAS is badly broken and will likely take considerable effort to fix (and fixing it also requires solving the same key problem as enabling IAS for just N64). I therefore want to separate the two ABI's so that I can finish off N64 and then fix N32 afterwards.

I've posted a series of patches that allow us to distinguish N32 and N64 so that we can enable IAS for only N64  (D21465, D21467, D21069 for LLVM along with D21070, D21072 for clang). During the review of D21467, Rafael asked me to explain my approach here. I'm hoping that people will agree that this is an acceptable approach to take and that it's no different from how other targets handle certain ABI's and how we already handle -m32/-m64/-EL/-EB/etc.

The approach I'm using is to encode the ABI information into the triple in the same way that ARM does for EABI and hardfloat, and the same way that X86 does for X32. These targets define variants of Triple::GNU for each of these ABI's in the form of the Triple::GNUEABI, Triple::GNUEABIHF, and Triple::GNUX32 values of the environment component of the triple and it's up to the frontend and/or API-user to pass the right thing to the backend.

For MIPS, I'm defining Triple::GNUABI32, Triple::GNUABIN32, and Triple::GNUABI64. All three of these are supported by tools like binutils (by virtue of a wildcard match '*linux-gnu*') and of these three, Triple::GNUABI32 is not in general use, Triple::GNUABIN32 will be used by Debian if/when we do an N32 port, and Triple::GNUABI64 is currently used by Debian for our N64 port. Once the ABI is in our backend triples, it can be used by MipsMCAsmInfo to enable IAS by default for N64 without also enabling it for N32. In later patches, such as D1292 this same information will also be used to fix the N32 support for IAS.

At this point, you may be wondering what happens to Triple::GNU for triples like mips-linux-gnu and mips64-linux-gnu. The answer to this is that the API-user (e.g. clang) is expected to normalize such triples down to one of Triple::GNUABI32/GNUABIN32/GNUABI64. This can be as simple as adding:
  Triple ABITriple = TT.getABIVariant(ABIName); // ABIName can be the empty string to get the default ABI.
  if (ABITriple.getArch() != Triple::UnknownArch) {
    TT = ABITriple;
    ABIName = ""; // <- Only needed if this would end up in MCTargetOptions::ABIName.
  }
to the appropriate place in the caller. This is the same way clang handles the -m32, -m64, -EL, and -EB options with those options using get32BitArchVariant(), get64BitArchVariant(), getLittleEndianArchVariant(), and getBigEndianArchVariant() respectively to transform the triple for the backend.

In summary, human end users of tools such as clang will continue using the same triples as they always have. These tools will then resolve them down to the effective triple (for example, in functions like clang's ComputeEffectiveClangTriple(), ComputeLLVMTriple(), and computeTargetTriple()). The MIPS backend will only see the triples that have a specific known ABI. I'd like to stress that this process is the same as the one we already use today for other options. For example 'x86_64-linux-gnu-clang -m32' transforms the triple with get32BitArchVariant() and passes 'i386-linux-gnu' to the backend.

So far I've only been talking about Linux/GNU but the patch series also covers Linux/Musl, Linux/Android, and FreeBSD. I haven't talked about those because they are handled in the same way as the Linux/GNU case described above except that they use Triple::ABI32/ABIN32/ABI64 (or Android32/Android64 in the case of Android) instead of Triple::GNUABI32/GNUABIN32/GNUABI64.

Why is the Triple approach better for MIPS than MCTargetOptions::ABIName?
=========================================================================

A major pragmatic reason is that it works today and the MCTargetOption::ABIName approach doesn't. It's also in keeping with other targets whereas until fairly recently MIPS was the sole user of MCTargetOptions::ABIName. There is currently a single PowerPC test and six ARM tests that use the -target-abi option that feeds MCTargetOptions::ABIName at the moment and both of these targets are only using it for relatively minor tweaks to the ABI.

As things are today, MCTargetOptions::ABIName is good at fairly small tweaks to the ABI such as calling convention changes but it's unable to deal with more fundamental changes such as selecting between ELFCLASS32/ELFCLASS64 or enabling IAS. Some of these limitations look fixable if given months-years of effort but I can see potential problems once the obvious plumbing problem is fixed. Some of these problems such as preventing the linking of incompatible IR files in the IRLinker would require me to represent the ABI in the LLVM-IR (it's currently handled solely by the triple) which has been strongly opposed in earlier discussions about triples.

I can provide more detail on some of the specific problems if required but this boils down to a choice between an easy paved road that other people use and that I know will end in a working state, or an unexplored path with known difficulties and an uncertain result.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160705/be5f59d3/attachment-0001.html>