Benchmarks on CI show that there are some performance regressions after
LLVM update due to worse inliner results. Explicit specification of
target-features fixes the problem. Interestingly, it seems that this is
not required for Darwin AArch64 target.
* Updated Windows requirements for building from source
* Removed Linux requirements on ncurses because LLVM is built without it
* Explicitly stated glibc version
Clang-produced and GCC-produced binaries might be ABI-incompatible on
MinGW. Explanation on GitHub: msys2/MINGW-packages/issues/6855#issuecomment-680859662.
TL;DR: GCC-generated sections are 16-byte-padded, while Clang ones are
not. It causes problems during merge of COMDAT sections.
I observed the problem during compilation of runtime tests, but it is
possible that the problem could affect main compilation pipeline as well.
https://reviews.llvm.org/D86659 (which landed in LLVM 12) fixes
the problem. So we have another motivation for switching to LLD besides
https://youtrack.jetbrains.com/issue/KT-47605.
The only known downside is unsupported defsym which causes slight binary
size increase. I think it is doable.
The right way is to add something like KonanTarget.MSVC_X64, but doing
so requires changes throughout whole compiler. It would be especially
painful in HostManager, where we would need to deprecate
KonanTarget.MINGW_X64 as host. Instead we "hack" ClangArgs to compile
for x86_64-pc-windows-msvc instead of x86_64-pc-windows-gnu in JNI case.
CI may contain custom MSVC and Windows Kit installation path, so we
should support it. Things might break when machine has several MSVC
installed (at custom and default path), but it sounds more like
incorrect environment setup problem than ours.
Since LLVM for Windows is now native instead on MinGW, we have to
compile code in a different environment in case of JNI. This commits
just separates ClangArgs into two subclasses without actual behavior
changes.
--no-threads flag was used, because LLD had deadlock long time ago.
This flag is removed from LLD (in favor of --threads=N) and it looks
like that the bug is gone now. So we can just drop --no-threads flag!
Recent versions of LLVM complain when caller and callee have
different vector attributes: https://reviews.llvm.org/D82562.
To mitigate this problem this commit changes how we include
polyHashUnroll* and polyHashTail into Traits: we duplicate them
inside each trait with correct attributes by using
```
#pragma clang attribute push(ATTRIBUTE, apply_to = function)
```
This solution is not pretty for sure, but Clang is not complaining
anymore.
This commit is important for several reasons:
1. Finally, LLVM update from 8.0.0 to 11.1.0
At the time of writing, LLVM 12.0.0 already came out, but we need to be
compatible with LLVM from Xcode 12.5, so that's why we stick to a bit
older version of LLVM.
2. These LLVM distributions are built with tools/llvm_builder/package.py
We finally managed to introduce an explicit process of building our LLVM
distributions, so we will be able to updated them more frequently and
adapt to our needs.
3. Native Windows LLVM instead of MinGW
Last but not least, we now use native Windows LLVM distribution. While
it might be harder to use (no more posix everywhere), simpler build
process (building msys2 is not that easy) and no need for MinGW-specific
patches makes such distribution way more convenient.
The next LLVM version will be 11.1.0, and we need to be sure that
Xcode supports bitcode of this version. So we bump minimal Xcode version
to 12.5, which should be OK, because by the time Kotlin 1.6.0 comes out,
Xcode 12.5 will be around 6 month old.
Also this commit simplifies tarball naming.
Implemented `test multiple cinterops with test source sets and compilations`
This test will assert source set dependencies
and shared native compilations.
^KT-47775
^KT-47053
^KT-47775 Verification Pending
This implementation will be lenient toward the common
*Test.dependsOn(*Main) mistake.
^KT-47053 Verification Pending
Support source sets that do not have a dedicated
shared native compilation. Also support additional visible source sets
coming from associate compilations.
Dmitriy Novozhilov