[PATCH] D37899: [SystemZ] Implement shouldCoalesce() to help regalloc to avoid running out of registers with GR128 regs

[Ulrich Weigand via Phabricator via llvm-commits]

uweigand added a comment.

I looked into making i128 legal when I recently added the 128-bit atomics, but in the end decided against it.   Not only would it be a lot of work (since you basically would have to reimplement all the operations on i128 that are currently handled by common code), but it is not clear that we actually always want it.

i128 lives in a even/odd register pair, but this is really only needed for the few special instructions that operate on those pairs.  If you had just a normal load followed by a normal store of a i128 value, you *want* to break it into two independent i64 copies -- there is no reason at all to constrain the register allocator to use a even/odd pair here.

I've been thinking about alternative solutions to this problem.  One way might be to prevent coalescing (as this patch does) and then try to *encourage* (but not enforce) use of the register pair earlier, e.g. by providing register allocator hints for those registers that would have been coalesced with the pair if we had not prevented this - sort of like the ARM register pair hints.  But it turns out this is no so easy, because the allocator will assume those hints conflict with the original register pair (unless we switch on enableSubRegLiveness, but this seems to have other problems, it looks like it causes a number of internal compiler errors in the test suite).

Another option might be to not *completely* prevent coalescing a 128-bit pair with its component registers, but allow it in cases where there is no excessive register pressure.  But I'm not sure exactly how to get at this information at the time we make the coalescing decision ...


https://reviews.llvm.org/D37899