[PATCH] Fix flakiness in fp16-promote.ll
ahmed.bougacha at gmail.com
Mon Apr 20 15:57:45 PDT 2015
In http://reviews.llvm.org/D9112#158400, @pirama wrote:
> I can probably add regular expressions to match registers. It is still a problem that a different instruction schedule will cause the test to fail.
> Is it possible that an explicit target CPU or another option in the command line will remove the variations? If so, we should do that instead of changing the test. I am not familiar enough with the backend to know this.
If you set an explicit triple, IIRC there's a default CPU chosen (in this case probably cortex-a8), so I don't expect that to vary. Schedule variations are still possible because of other changes missing from trunk though, probably like the register allocation issue here. (+ Renato, who would have the advice you're looking for)
In any case, bots are still broken, so this should be fixed one way or the other, soon ;)
More information about the llvm-commits