I don't have an environment to test this, but I believe it should be enough to
get the JIT to work.
This came up in
https://github.com/google/jax/issues/5679
PiperOrigin-RevId: 356578451
Change-Id: I99a2aa0e87739b9edce81074fce1ca5c0bd25115
This solved the following compilation issue on windows:
bazel-out/x64_windows-opt/bin/external/llvm-project/llvm/lib/Target/AMDGPU\AMDGPUGenInstrInfo.inc(33991): fatal error C1091: compiler limit: string exceeds 65535 bytes in length
PiperOrigin-RevId: 293286375
Change-Id: Ibec6425c57d0e2748d9a76d187ce86a24e324696
This change to reflect MLIR's recent addition to LLVM's monorepo. Instead of
stripping llvm from the http archive, use build file map to refer to both MLIR
and LLVM within llvm-project and remove read-only copy that was temporarily
there. Also bump the commit version until after the integration of MLIR into
LLVM monorepo.
PiperOrigin-RevId: 287547194
Change-Id: I4459b96702dc5eb19b58c7ceaa6cabb16f530a43
TensorFlow now builds with:
--incompatible_disable_legacy_cc_provider
--incompatible_depset_is_not_iterable
--incompatible_depset_union
--incompatible_remap_main_repo
Fixes#26388
Working towards #25935
PiperOrigin-RevId: 241350192
I previously use hacky BUILD file to import prebuilt LLVM binaries from CMake to Bazel so that I can build `tfcompile` with Bazel on Windows, but since Tensorflow updates the LLVM revision used very frequently, this method is too troublesome.
I am trying to make XLA/AOT to work out-of-the-box on Windows with Bazel, then convince Tensorflow team to create a presubmit so that no one can break XLA/AOT on Windows anymore.
Summary:
- Remove unused `llvm_host_triple`.
- Rename `cmake_vars` to `posix_cmake_vars` and create a new `win32_cmake_vars` as Windows is sufficiently different from other platforms.
- Create `LLVM_DEFINES`, `LLVM_COPTS` and `LLVM_LINKOPTS` in `llvm.bzl` to set global compile/link flags for LLVM.
- Create macro `llvm_cc_library` that wrap around `cc_library` to ensure that all targets receive `LLVM_COPTS` consistently. Replace all `cc_library` with `llvm_cc_library`.
- For `support` library, include `lib/Support/Windows/*` for Windows and `lib/Support/Unix/*` for others.
If `llvm.BUILD` is not to be edited manually no matter what, can the script used to generate `llvm.BUILD` be open-sourced?
#15213
/cc @tatatodd
Imported from GitHub PR #18293
Copybara import of the project:
- a5e7b841d79878b12e0d0103e5b1813384a80aa8 [XLA/AOT] Build LLVM with Bazel on Windows by Loo Rong Jie <loorongjie@gmail.com>
- b3a20c7bbff95fd8ad9f44c653a3ecffc54de3b3 Add back LLVM_ENABLE_STATS by Loo Rong Jie <loorongjie@gmail.com>
- 787d451b495180f5c882d09d76222c157102f0fc Merge b3a20c7bbff95fd8ad9f44c653a3ecffc54de3b3 into 6a1ff... by Loo Rong Jie <loorongjie@gmail.com>
COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/tensorflow/pull/18293 from rongjiecomputer:llvm b3a20c7bbff95fd8ad9f44c653a3ecffc54de3b3
PiperOrigin-RevId: 208073601
- Remove LLVM_DEFINES from all targets but :config to avoid running into
Windows command line length limitations.
- Add LLVM_COPTS to cc_binary rules.
- Extract out a list of platform specific headers to a
LLVM_SUPPORT_PLATFORM_SPECIFIC_SRCS_GLOB variable.
PiperOrigin-RevId: 203806527
This lets external contributors add or remove compiler flags without having
access to the generator script (which isn't open source).
E.g. see #18293
PiperOrigin-RevId: 202322000
