On Saturday, 30 March 2019 at 15:30:16 UTC, Thomas Mader wrote:
The lowmem switch tests [1] fail for me because I use ltsmaster
as the bootstrap compiler.
1) I think those tests should not run in this szenario.
2) I know that I need to rebuild ldc with itself to get the
lowmem switch support in the long run but I wonder if that
would not be better integrated into your cmake build via an
option.
Otherwise the packagers out there need to do it for themselves.
Ideally, all packagers would by now have installed a 2-step
bootstrapping process (ltsmaster -> master/recent version ->
master) or an incremental process (previous version -> master).
That also prevents us from having to backport support for newer
LLVM versions to ltsmaster (e.g., ltsmaster doesn't support new
LLVM 8). 32-bit ltsmaster isn't able to compile current master,
and principal compilability of master with the ltsmaster/2.068
front-end and libs is also only possible due to some slight
front-end adaptations (DMD requires 2.072 or so), whose number is
only going to grow until it becomes totally infeasible.
In short, ltsmaster won't be able to bootstrap current LDC master
forever, and I guess that point is reached pretty soon.
As to integrating the build-with-itself option in CMake, I'm
definitely against increasing CMake complexity, as invoking CMake
+ ninja/make twice is trivial and *way* more flexible.
And as to the 3 new dmd-testsuite tests using -lowmem, they are
specifically there to check that the switch works, so if compiled
with ltsmaster, they should fail to clearly indicate that the
compiler isn't fully working due to an inadequate host compiler.
If anyone doesn't care about that, the 3 files can simply be
removed before running the tests.
