Re: [llvm-dev] Creating sanity in LLVM/Clang's atomics impl

[James Y Knight via llvm-dev]

I've uploaded a few changes to review for implementing Step "A1" in the atomic cleanup plan:

http://reviews.llvm.org/D18200

  Add __atomic_* lowering to AtomicExpandPass.

http://reviews.llvm.org/D18201

  Switch over targets to use AtomicExpandPass, and clean up target atomics code.

There's also this, semi-related cleanup:

http://reviews.llvm.org/D17933

 Clang: Set MaxAtomicInlineWidth properly for i386, i486, and x86-64 cpus without cmpxchg16b. Also, remove all the manual definition of

GCC_HAVE_SYNC_COMPARE_AND_SWAP_* macros.

There's some kinda annoying issues in LLVM I've run into so far:

1) The required duplication of logic for which platforms have what-sized atomics between LLVM and Clang is unfortunate. This falls into the general problem of clang's lib/Basic/Targets.cpp file being mostly an awful repetition of information available in the LLVM backends.

We really ought to have Clang start depending on the LLVM backends.

2) The inability to access C ABI information in LLVM, whilst creating the libcalls is somewhat of a problem.

The atomic libcalls need the following information:

- What "size_t" and "int" mean? (DL.getIntPtrType() seems okayish for the former (malloc already does that), and can use getInt32() for the latter...except that MSP430 seems to use 16-bit "int")

- What sizes of integers actually exist? You need to know which functions of the form "T __atomic_load_N(T*, int)", with some integral type T, can actually exist in the atomics library. (e.g. if the platform has no __int128, there won't be an "__int128 __atomic_load_16(__int128*, int)" you can actually call.

All that information is available in Clang, but not in LLVM. I think such low-level C ABI details would be a good idea to have in LLVM.

I also ran into a couple of other issues whilst on this quest:

1) The C11 _Atomic ABI is incompatible between GCC and Clang . As is libc++'s std::atomic. (But not libstdc++: that uses the same ABI in both compilers).

Also, in Clang, _Atomic and libc++'s std::atomic use unfortunate choices of "maximum size to promote" that depend on the platform, and don't allow for easy future extension of the set of supported atomic sizes.

https://llvm.org/bugs/show_bug.cgi?id=26462

2) LLVM may not have a strong enough memory model to support GCC's __sync_* builtins properly (this may affect only ARMv8/AArch64 at the moment?). GCC has (recently) decided that __sync_fetch_and_add(&val, 4); needs *stronger* memory barriers than "__atomic_fetch_add(&val, 4, __ATOMIC_SEQ_CST);" . In particular, the former should not allow the processor to reorder a subsequent store before the store-release instruction at the end of the LL/SC sequence generated by atomic_fetch_add.

The C11, and LLVM, definition of a seq_cst memory operation only makes it ordered with other seq_cst operations, NOT with other non-seq_cst atomics. But, apparently GCC feels that the definition of its __sync_fetch_and_add builtin requires that it act like it act like a seq_cst FENCE, preventing non-seq_cst memory operations from moving across it. And thus, it now emits an explicit barrier instruction after the store-release.

https://gcc.gnu.org/PR65697

Should we do similar to GCC, and add a 6th kind of ordering, "seq_cst_fence" to model this?

On Wed, Jan 27, 2016 at 2:25 PM, James Knight <jykn...@google.com> wrote:

Right now, the atomics implementation in clang is a mess. It has three basically completely different code-paths:

1. There's the legacy __sync_* builtins, which clang lowers directly to atomic IR instructions. Then, the llvm atomic IR instructions themselves can sometimes emit libcalls to __sync_* library functions (which are basically undocumented, and users are typically responsible for implementing themselves if they want it).
2. There's the new __atomic_* builtins, which clang will, depending on size and alignment and target, lower either to a libcall to a "standardized-by-GCC" __atomic_* library function (implemented by libatomic), or, to the atomic IR instructions. (Also included here are the __c11_atomic_* builtins, which are almost-identical clang-specific aliases of the __atomic_* builtins. They share 99% of the same behavior and code.)
3. There's the C11 "_Atomic" type modifier and some OpenMP stuff, which clang lowers into atomic IR or libcalls to __atomic_* library functions, almost completely separately from #2. (Note: doesn't apply to C++ std::atomic<>, which gets implemented with the builtins). Beyond just a lot of duplicate code, the _Atomic impl is just broken: sometimes it emits llvm IR instructions directly (letting things fall back to __sync_* calls), and sometimes it checks if it should emit a libcall first (falling back to __atomic_* calls). This is certainly bad behavior.

I'd like to make a proposal for cleaning this mess up.

BTW, as a sidenote [[

One thing that's important to remember: at least on the C/C++ level, if you support lock-free atomic-ops for a given size/alignment, ALL the atomic ops for that size/alignment must be lock-free. This property is usually quite easy, because you'll have LL/SC or CAS instructions, with which everything else is implementable, and thus wouldn't ever emit a libcall, except for misaligned or too-large atomics, in any case.

On the other hand, many older CPU models are missing those instructions, such as ARMv5, Sparcv8, and Intel 80386. When your minimum CPU is set to such a CPU, all atomic operations must be done via libcall -- it's not acceptable for atomic_store to emit an atomic "store" instruction, but atomic_fetch_add to require a libcall which is implemented with a lock. If that were to happen, then atomic_fetch_add could not actually be atomic versus a simultaneous atomic_store.

]]

So anyhow, there's basically two paths I think we could take for cleanup. I like "Alternative A" below better, but would be interested to hear if others have reasons to think the other would be preferable for some reason.

Both alternatives I've suggested will have the effect that the __sync_* builtins in clang will now lower to __atomic_* function calls on platforms without inline atomics (or for unsupported sizes), and C11 atomics will stop being schizophrenic. Clang will cease to ever emit a call to a __sync_* function from any of its builtins or language support. This could theoretically cause an incompatibility on some target. 

However, I think it ought to be pretty safe, by now: I can't imagine anyone has __sync_* functions implemented for their platform but doesn't have the __atomic_* library functions implemented, as the __atomic_* builtins (which already use those library calls) are in widespread use. Notably, both libstdc++ and libc++ use them, as does a lot of third-party code. IMO it's worth having that potential incompatibility, in order to simplify the situation for users (only one set of atomic library calls to worry about), and to have a single code-path for atomics in the compiler.

Alternative A: Move all atomic libcall emission into LLVM

In this alternative, LLVM will be responsible for lowering all atomic operations, for all three code-paths listed at the beginning. Clang will emit no libcalls for atomic operations.

A1) In LLVM: when lowering "store atomic", "load atomic", "atomicrmw", and "cmpxchg" instructions that aren't supported by the target, emit libcalls to the new __atomic_* functions, (rather than the current behavior of calling the legacy __sync_* functions.)

Additionally, I'd like to move the decision to emit a libcall into AtomicExpandPass, and remove the ability to use Expand in ISel to create a libcall for ISD::ATOMIC_*. Putting the decision there allows querying a target attribute, up front, before any other lowering decisions (both in AtomicExpandPass and in ISel). When choosing whether to inline or libcall, the target will get to see only the size of the operation, and not the operation itself or the type. This will ensure that the "all or none" property is adhered to.

(Q: what about the implementation of __atomic_is_lock_free/__atomic_always_lock_free in clang? The clang frontend can't query target information from the llvm target backend, can it? Is there some other way to expose the information of what sizes are supported by a backend, such that those builtins -- the latter of which must be a C++ constant expression -- can use it?)

A2) In clang, start removing the code that knows how to do optimized library call lowering for atomics -- always emit llvm atomic ops. Except for one case: clang will still need to emit library calls itself for data not aligned naturally for its size. The LLVM atomic instructions currently will not handle unaligned data, but that is allowed for the four "slab of memory" builtins (__atomic_load, __atomic_store, __atomic_compare_exchange, and __atomic_exchange).

A3) Hopefully after that removal, the three codepaths in clang will be sufficiently the same that they can be merged or have been simplified so drastically that the remaining duplication is not worthwhile to merge.

A4) In LLVM, add "align" attributes to cmpxchg and atomicrmw, and allow specifying "align" values for "load atomic" and "store atomic". LLVM will lower them to the generic library calls. In clang, start lowering misaligned atomics to these llvm instructions as well.

Alternative B: Move all libcall emission for atomics into Clang.

In this alternative, LLVM will never emit atomic libcalls from the atomic IR instructions. If the operation requested is not possible on a given target, that is an error. So, the cleanups here are to get clang to stop emitting LLVM IR atomics that cannot be lowered without libcalls.

B1) In Clang: have the legacy __sync_* builtins become essentially aliases for the __atomic_* builtins (thus they will emit calls to __atomic_* library functions when applicable).

B2) In Clang: Have the C11 _Atomic type support and openmp stuff call into the __atomic builtins' code to do their dirty work, instead of having its own separate implementation.

B3) After those two changes, I believe clang will only ever emit atomic IR instructions when they can be lowered. So then, in LLVM: get rid of the fallback to libcalls to __sync_* from the atomic IR instructions. If an atomic IR operation is requested, and not implementable lock-free, it will be an error, with no fallback.