It used to be necessary on Apple platforms to ship with the App Store,
but XCode 15 has stopped embedding LLVM bitcode and the App Store no
longer accepts apps with bitcode embedded.
Rollup of 7 pull requests
Successful merges:
- rust-lang/rust#144072 (update `Atomic*::from_ptr` and `Atomic*::as_ptr` docs)
- rust-lang/rust#144151 (`tests/ui/issues/`: The Issues Strike Back [1/N])
- rust-lang/rust#144300 (Clippy fixes for miropt-test-tools)
- rust-lang/rust#144399 (Add a ratchet for moving all standard library tests to separate packages)
- rust-lang/rust#144472 (str: Mark unstable `round_char_boundary` feature functions as const)
- rust-lang/rust#144503 (Various refactors to the codegen coordinator code (part 3))
- rust-lang/rust#144530 (coverage: Infer `instances_used` from `pgo_func_name_var_map`)
r? `@ghost`
`@rustbot` modify labels: rollup
Nobody seems to actually use this, while still adding some extra
complexity to the already rather complex codegen coordinator code.
It is also not supported by any backend other than the LLVM backend.
Various refactors to the LTO handling code (part 2)
Continuing from https://github.com/rust-lang/rust/pull/143388 this removes a bit of dead code and moves the LTO symbol export calculation from individual backends to cg_ssa.
gpu offload host code generation
r? ghost
This will generate most of the host side code to use llvm's offload feature.
The first PR will only handle automatic mem-transfers to and from the device.
So if a user calls a kernel, we will copy inputs back and forth, but we won't do the actual kernel launch.
Before merging, we will use LLVM's Info infrastructure to verify that the memcopies match what openmp offloa generates in C++. `LIBOMPTARGET_INFO=-1 ./my_rust_binary` should print that a memcpy to and later from the device is happening.
A follow-up PR will generate the actual device-side kernel which will then do computations on the GPU.
A third PR will implement manual host2device and device2host functionality, but the goal is to minimize cases where a user has to overwrite our default handling due to performance issues.
I'm trying to get a full MVP out first, so this just recognizes GPU functions based on magic names. The final frontend will obviously move this over to use proper macros, like I'm already doing it for the autodiff work.
This work will also be compatible with std::autodiff, so one can differentiate GPU kernels.
Tracking:
- https://github.com/rust-lang/rust/issues/131513
Various refactors to the LTO handling code
In particular reducing the sharing of code paths between fat and thin-LTO and making the fat LTO implementation more self-contained. This also moves some autodiff handling out of cg_ssa into cg_llvm given that Enzyme only works with LLVM anyway and an implementation for another backend may do things entirely differently. This will also make it a bit easier to split LTO handling out of the coordinator thread main loop into a separate loop, which should reduce the complexity of the coordinator thread.
Most uses of it either contain a fat or thin lto module. Only
WorkItem::LTO could contain both, but splitting that enum variant
doesn't complicate things much.
There is no safety contract and I don't think any of them can actually
cause UB in more ways than passing malicious source code to rustc can.
While LtoModuleCodegen::optimize says that the returned ModuleCodegen
points into the LTO module, the LTO module has already been dropped by
the time this function returns, so if the returned ModuleCodegen indeed
points into the LTO module, we would have seen crashes on every LTO
compilation, which we don't. As such the comment is outdated.
Autodiff flags
Interestingly, it seems that some other projects have conflicts with exactly the same LLVM optimization passes as autodiff.
At least `LLVMRustOptimize` has exactly the flags that we need to disable problematic opt passes.
This PR enables us to compile code where users differentiate two identical functions in the same module. This has been especially common in test cases, but it's not impossible to encounter in the wild.
It also enables two new flags for testing/debugging. I consider writing an MCP to upgrade PrintPasses to be a standalone -Z flag, since it is *not* the same as `-Z print-llvm-passes`, which IMHO gives less useful output. A discussion can be found here: [#t-compiler/llvm > Print llvm passes. @ 💬](https://rust-lang.zulipchat.com/#narrow/channel/187780-t-compiler.2Fllvm/topic/Print.20llvm.20passes.2E/near/511533038)
Finally, it improves `PrintModBefore` and `PrintModAfter`. They used to work reliable, but now we just schedule enzyme as part of an existing ModulePassManager (MPM). Since Enzyme is last in the MPM scheduling, PrintModBefore became very inaccurate. It used to print the input module, which we gave to the Enzyme and was great to create llvm-ir reproducer. However, lately the MPM would run the whole `default<O3>` pipeline, which heavily modifies the llvm module, before we pass it to Enzyme. That made it impossible to use the flag to create llvm-ir reproducers for Enzyme bugs. We now schedule a PrintModule pass just before Enzyme, solving this problem.
Based on the PrintPass output, it also _seems_ like changing `registerEnzymeAndPassPipeline(PB, true);` to `registerEnzymeAndPassPipeline(PB, false);` has no effect. In theory, the bool should tell Enzyme to schedule some helpful passes in the PassBuilder. However, since it doesn't do anything and I'm not 100% sure anymore on whether we really need it, I'll just disable it for now and postpone investigations.
r? ``@oli-obk``
closes#139471
Tracking:
- https://github.com/rust-lang/rust/issues/124509
Autodiff batching
Enzyme supports batching, which is especially known from the ML side when training neural networks.
There we would normally have a training loop, where in each iteration we would pass in some data (e.g. an image), and a target vector. Based on how close we are with our prediction we compute our loss, and then use backpropagation to compute the gradients and update our weights.
That's quite inefficient, so what you normally do is passing in a batch of 8/16/.. images and targets, and compute the gradients for those all at once, allowing better optimizations.
Enzyme supports batching in two ways, the first one (which I implemented here) just accepts a Batch size,
and then each Dual/Duplicated argument has not one, but N shadow arguments. So instead of
```rs
for i in 0..100 {
df(x[i], y[i], 1234);
}
```
You can now do
```rs
for i in 0..100.step_by(4) {
df(x[i+0],x[i+1],x[i+2],x[i+3], y[i+0], y[i+1], y[i+2], y[i+3], 1234);
}
```
which will give the same results, but allows better compiler optimizations. See the testcase for details.
There is a second variant, where we can mark certain arguments and instead of having to pass in N shadow arguments, Enzyme assumes that the argument is N times longer. I.e. instead of accepting 4 slices with 12 floats each, we would accept one slice with 48 floats. I'll implement this over the next days.
I will also add more tests for both modes.
For any one preferring some more interactive explanation, here's a video of Tim's llvm dev talk, where he presents his work. https://www.youtube.com/watch?v=edvaLAL5RqU
I'll also add some other docs to the dev guide and user docs in another PR.
r? ghost
Tracking:
- https://github.com/rust-lang/rust/issues/124509
- https://github.com/rust-lang/rust/issues/135283
Rename `is_like_osx` to `is_like_darwin`
Replace `is_like_osx` with `is_like_darwin`, which more closely describes reality (OS X is the pre-2016 name for macOS, and is by now quite outdated; Darwin is the overall name for the OS underlying Apple's macOS, iOS, etc.).
``@rustbot`` label O-apple
r? compiler
The embedded bitcode should always be prepared for LTO/ThinLTO
Fixes#115344. Fixes#117220.
There are currently two methods for generating bitcode that used for LTO. One method involves using `-C linker-plugin-lto` to emit object files as bitcode, which is the typical setting used by cargo. The other method is through `-C embed-bitcode=yes`.
When using with `-C embed-bitcode=yes -C lto=no`, we run a complete non-LTO LLVM pipeline to obtain bitcode, then the bitcode is used for LTO. We run the Call Graph Profile Pass twice on the same module.
This PR is doing something similar to LLVM's `buildFatLTODefaultPipeline`, obtaining the bitcode for embedding after running `buildThinLTOPreLinkDefaultPipeline`.
r? nikic
`rustc_codegen_llvm` relied on `Deref` impls where `Deref::Target` was
or contained an extern type - in my experimental implementation of
rust-lang/rfcs#3729, this isn't possible as the `Target` associated
type's `?Sized` bound cannot be relaxed backwards compatibly (unless we
come up with some way of doing this).
In later pull requests with the rust-lang/rfcs#3729 implementation,
breakage like this could only occur for nightly users relying on the
`extern_types` feature.
Upstreaming this to avoid needing to keep carrying this patch locally,
and I think it'll necessarily need to change eventually.
