Should we be testing more target triples? I'm not clear how much variety there is on 32-bit targets.
I think it's enough. Currently we only have darwin that has different f80 alignment on 32-bit:
Is there a bugzilla PR covering this?
Probability not. I guess few people concerns 32-bit darwin now. I didn't mean to fix the bug. We just have the same requirment in our downstream code.
I tried. Then I found the check doubled due to darwin and Linux have different inline asm comment. More importantly, the interesting info is buried by the large diff.
$ grep -r -E 'fldt|fstpt' llvm/test/CodeGen/X86/ | wc -l 994 $ grep -r -E 'fldt|fstpt' llvm/test/CodeGen/X86/ |cut -d ':' -f 1 |uniq | wc -l 58