Rollup merge of #57268 - peterhj:peterhj-optmergefunc, r=nagisa

Add a target option "merge-functions", and a corresponding -Z flag (works around #57356)

This commit adds a target option "merge-functions", which takes values in ("disabled", "trampolines", or "aliases" (default is "aliases")), to allow targets to opt out of the MergeFunctions LLVM pass. Additionally, the latest commit also adds an optional -Z flag, "merge-functions", which takes the same values and has precedence over the target option when both are specified.

This works around https://github.com/rust-lang/rust/issues/57356.

cc @eddyb @japaric @oli-obk @nox @nagisa

Also thanks to @denzp and @gnzlbg for discussing this on rust-cuda!

### Motivation

Basically, the problem is that the MergeFunctions pass, which rustc currently enables by default at -O2 and -O3 [1], and `extern "ptx-kernel"` functions (specific to the NVPTX target) are currently not compatible with each other. If the MergeFunctions pass is allowed to run, rustc can generate invalid PTX assembly (i.e. a PTX file that is not accepted by the native PTX assembler `ptxas`). Therefore we would like a way to opt out of the MergeFunctions pass, which is what our target option does.

### Related work

The current behavior of rustc is to enable MergeFunctions at -O2 and -O3 [1], and also to enable the use of function aliases within MergeFunctions [2] [3]. MergeFunctions seems to have some benefits, such as reducing code size and fixing a crash [4], which is why it is enabled. However, MergeFunctions both with and without function aliases is incompatible with the NVPTX target; a more detailed example for both cases is given below.

clang's "solution" is to have a "-fmerge-functions" flag that opts in to the MergeFunctions pass, but it is not enabled by default.

### Examples/more details

Consider an example Rust lib using `extern "ptx-kernel"` functions: https://github.com/peterhj/nvptx-mergefunc-bug/blob/master/nocore.rs. If we try to compile this with nightly rustc, we get the following compiler error:

    LLVM ERROR: Module has aliases, which NVPTX does not support.

This error happens because: (1) functions `foo` and `bar` have the same body, so are candidates to be merged by MergeFunctions; and (2) rustc configures MergeFunctions to generate function aliases using the "mergefunc-use-aliases" LLVM option [2] [3], but the NVPTX backend does not support those aliases.

Okay, so we can try omitting "mergefunc-use-aliases", and then rustc will happily emit PTX assembly: https://github.com/peterhj/nvptx-mergefunc-bug/blob/master/nocore-mergefunc-nousealiases-bad.ptx. However, this PTX is invalid! When we try to assemble it with `ptxas` (I'm on the CUDA 9.2 toolchain), we get an assembler error:

    ptxas nocore-mergefunc-nousealiases-bad.ptx, line 38; error   : Illegal call target, device function expected
    ptxas fatal   : Ptx assembly aborted due to errors

What's happening is that MergeFunctions rewrites the `bar` function to call `foo`. However, directly calling an `extern "ptx-kernel"` function from another `extern "ptx-kernel"` is wrong.

If we disable the MergeFunctions pass from running at all, rustc generates correct PTX assembly: https://github.com/peterhj/nvptx-mergefunc-bug/blob/master/nocore-nomergefunc-ok.ptx

[1] https://github.com/rust-lang/rust/blob/a36b960df626cbb8bea74f01243318b73f0bd201/src/librustc_codegen_ssa/back/write.rs#L155
[2] https://github.com/rust-lang/rust/blob/a36b960df626cbb8bea74f01243318b73f0bd201/src/librustc_codegen_llvm/llvm_util.rs#L64
[3] https://github.com/rust-lang/rust/pull/56358
[4] https://github.com/rust-lang/rust/pull/49479

0 files changed, 0 insertions, 0 deletions