LLVM 12–17 have been dropped for Nixpkgs 25.11. As discussed recently
on Matrix, this backports upstream changes to allow the use of
LLVM 20 for all GHC versions from 9.4.8 onward.
I looked over GHC commits mentioning LLVM since the release of 9.4.8,
and read the discussions and issues around the relevant bumps, and
attempted to be quite thorough, but I obviously cannot guarantee that
this is wholly comprehensive. It seems like upstream generally bumps
the upper bound on the basis of “it builds successfully for me”,
with specific adaptations for new versions being fairly uncommon and
only coming for obvious build blockers or reactively in response to
bug reports. I have backported both kinds of changes here.
For some commits, trivial conflict resolutions and adaptations
were required. It would be possible to pass the affected files to
`fetchpatch` as `excludes` and keep smaller fix‐up patches in tree in
some cases, but I opted to keep it simple and vendor complete backport
patches instead. I did not attempt to backport every single change to
the LLVM backend, only those that seemed directly relevant to support
for newer versions; if you’d get the same issue with the older LLVM,
that’s just a GHC bug.
These changes should actually make it easier to cross‐compile for
new architectures, as more recent LLVMs will have better support for
newer platforms, and it will be easier to backport GHC changes to
enable new platforms with less drift in the backend.
These patches do result in two breaking changes. Firstly, the minimum
LLVM version is bumped to 13 across the board. This is irrelevant for
Nixpkgs as we pin a specific LLVM version anyway, and versions below
LLVM 18 will be removed imminently. Secondly, support for the hidden
`-fno-llvm-tbaa` flag is dropped. This can be replaced with custom
`-optlo` flags to control the passes more directly, but the main
use of this undocumented flag appears to have been to [work around]
the lack of support for newer LLVM versions, anyway.
[work around]: <https://gitlab.haskell.org/ghc/ghc/-/issues/22220>
I successfully built the following on `aarch64-linux`:
* `pkgsCross.armv7l-hf-multiplatform.buildPackages.haskell.compiler.ghc948`
* `pkgsCross.armv7l-hf-multiplatform.buildPackages.haskell.compiler.ghc967`
* `pkgsCross.armv7l-hf-multiplatform.buildPackages.haskell.compiler.ghc984`
* `pkgsCross.armv7l-hf-multiplatform.buildPackages.haskell.compiler.ghc9102`
* `pkgsCross.armv7l-hf-multiplatform.buildPackages.haskell.compiler.ghc9121`
* `pkgsCross.armv7l-hf-multiplatform.buildPackages.haskell.compiler.ghc9122`
* `pkgsCross.riscv64.haskell.compiler.ghc948`
The GHC 9.4.8 with an ARMv7 host platform segfaults when I try to run
GHC, though e.g. `ghc-pkg --help` runs successfully. The GHC 9.10.3
build targeting ARMv7 crashed inside `llc(1)`, so I tried RISC‐V,
which has some platform mismatch issue relating to `libffi`, so I
tried z/Architecture, which failed with an invalid floating point
constant in the LLVM IR, so I tried 64‐bit MIPS, which failed with
a different `libffi` issue, so I tried 32‐bit MIPS, which failed
to compile `compiler-rt`, so I gave up. I confirmed that both of the
ARMv7 issues reproduce with 944e8fd4f4,
the revision before they were bumped from their old versions of LLVM,
so these are not regressions.
I built a test program with the ARMv7 cross‐compilers and
confirmed that they run on the AArch64 builder. I also confirmed
that the cross‐compiled RISC‐V GHC successfully runs under
`qemu-riscv64(1)`. It will only try to build programs via the C
backend, though, as that is the only option for unregisterised™
targets, so it’s not clear to me how useful LLVM support in 9.4.8
really is for bootstrapping new platforms; I guess even RISC‐V
would require more backporting work to produce a cross‐compiled
GHC that will use LLVM to compile its own input. I didn’t bother
setting up all the binfmt machinery to get it through compiling and
running a test program, but it at least makes the attempt.
ghc902Binary's rts.conf indeed prescribes linking against numa for
aarch64-linux, but this is not the case with the later bindists we have
been providing numa for.
(cherry picked from commit 347062bef5)
This bootstraps mostly from ghc984Binary, at least for the major
platforms. The bootstrap paths for darwin become much shorter.
Unfortunately, the bootstrap paths for i686-linux become a bit longer,
so it's a trade-off.
(cherry picked from commit 5df6c9a28f)
Latest 9.10.x minor release is 9.10.2, which is also in Stackage 24.8.
Thus, dropping according to the GHC Deprecation Policy.
(cherry picked from commit ac5228c6c0)
ghc902Binary's rts.conf indeed prescribes linking against numa for
aarch64-linux, but this is not the case with the later bindists we have
been providing numa for.
This bootstraps mostly from ghc984Binary, at least for the major
platforms. The bootstrap paths for darwin become much shorter.
Unfortunately, the bootstrap paths for i686-linux become a bit longer,
so it's a trade-off.
