commit 7d736b5b9adb055bf39341659985345680f67d09
author Benoit Jacob <benoitjacob@google.com> Thu Jan 18 11:54:39 2024 -0500
committer GitHub <noreply@github.com> Thu Jan 18 16:54:39 2024 +0000
tree 7386b6adedc6b7cfb700606fb33b73f4e021a562
parent 6ab1ed8e94d1e4a2ba07abbcab4f80302631ab8b

Ukernels: simplify the architecture-specific bitcode build. (#16126)

This PR basically re-does how we build and link architecture-specific
and generic bitcode, for CPU ukernels.

The main change is that we now build one separate ukernel bitcode file
for each CPU architecture, and it's the only thing that `iree-compile`
needs to load and link to. Before, it had to load and link together 3
different bitcode modules, two of which were actually compiled as
WebAssembly for genericity, and it involved weak symbols. Full inlining
is really important for performance here, and in this system, it
depended on two fragile mechanisms:
1. Reinterpreting a WebAssembly module as a native module.
2. Weak symbols.

That was done out of a concern for sharing generic code across
same-bitness architectures, but that turned out be substantially more
complex in both the ukernels and the compiler, and to suffer from
problems described in #16000, which this PR fixes. Namely, the inlining
across these modules and down to the calling dispatch function was
brittle, never worked outside of x86, and to make even that work, we had
to do some illegal target-attribute stripping that was effectively UB
and caused issues described in
https://github.com/llvm/llvm-project/issues/78206 (originally filed as a
LLVM bug but then understood to be caused by our own illegal
attribute-stripping).

So, we now build completely separate bitcode for each target CPU
architecture that we are supporting in `iree-compile`. At the moment,
that's roughly 20k of code for each additional architecture for which we
don't have dedicated optimized code paths. For comparison, each
architecture that does have dedicated code weighs roughly 100k. I found
that surprisingly large, until I remembered that IR is far more verbose
than object code. There are ways to control that if and when that
becomes a major concern:
* We could further trim the supported CPU architectures, but note that
we gain simplicity from not doing so at the moment.
* We could leave bitcode files separate rather than embedding them in
`iree-compile`, allowing each user to obtain their own for their target
architecture.
* We could just embed the ukernels *source* code, as pure C is so cheap
to compile - the main cost would be now having to essentially build/ship
`clang` as part of iree-compile, but supporting only C and not C++
shouldn't be that bad.
* One problem that that would also solve is that at the moment, for CPU
architectures for which `iree-compile` allows the user to pick from
multiple ABIs (looking at you, RISC-V --- you need to grow up, pick an
ABI and settle in life), compiling ukernels from source at that time
would make it trivial to have ukernels built for the right ABI. For now
in this PR we just pick whichever ABI our tests are currently using.

Note: this PR is broken into a few thematic commits to help reviewing.

Fixes #16000