Drop support for LLVM < 15, legacy pass manager and non-opaque IR pointers #4657
Conversation
LLVM 14 is the first version supporting both new pass manager and opaque IR pointers.
kinke
force-pushed
the
rm_llvm_11_12_13
branch
3 times, most recently
from
a326e47 to
9a611d5
Compare
The new min LLVM version isn't provided by Ubuntu 20 officially, so resort to the LLVM apt repo for now, as we are stuck with Ubuntu 20 due to the gdb regressions.
kinke
force-pushed
the
rm_llvm_11_12_13
branch
3 times, most recently
from
0ec399b to
8df40b4
Compare
|
Dropping typed IR pointers seems not really feasible yet:
So I guess that'll have to wait until we raise min LLVM to v17. |
|
Any opinions? Or concerns wrt. dropping the legacy pass manager support? As another option, we could drop LLVM 14 too, drop typed IR pointers already, and live with that one test failure for LLVM < 17 (a bit of a PITA, as the same failure causes a druntime integration test to fail too). I don't think that LLVM problem would be a huge deal in practice, for non-official packages using such an older LLVM (or our official FreeBSD package, which uses LLVM 15 IIRC). It'd be interesting to see when clang switched to opaque IR pointers by default. Edit: According to godbolt, clang switched in v15 already. |
|
No concerns from my side. Weka is using LLVM14 at the moment, but that's because of using an older LDC for now. The plan is to upgrade to 1.38 first, then upgrade LLVM, and then to later LDC. So no problem if 1.39 does not support LLVM14. |
Mainly because opaque IR pointers appear buggy in LLVM 14 (LLVM crashes when compiling druntime).
kinke
changed the title
|
This is ready; it's just that the 'required' CI jobs need to be adapted after landing this. I'll try to follow up with some pointer-bitcast cleanup; I have something locally already, but it appears to be misbehaving badly at the moment. ;) |
Conflicts: CMakeLists.txt tools/CMakeLists.txt
LLVM 14 is the first version supporting both new pass manager and opaque IR pointers.