https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102178
--- Comment #30 from Vladimir Makarov <vmakarov at gcc dot gnu.org> --- (In reply to Richard Biener from comment #29) > (In reply to Vladimir Makarov from comment #28) > > Could somebody benchmark the following patch on zen2 470.lbm. > > Code generation changes quite a bit, with the patch the offending function > is 16 bytes larger. I see no large immediate moves to GPRs anymore but > there is still a lot of spilling of XMMs to GPRs. Performance is > unchanged by the patch: > > 470.lbm 13740 128 107 S 13740 128 107 S > 470.lbm 13740 128 107 * 13740 128 107 S > 470.lbm 13740 128 107 S 13740 128 107 * > > Thank you very much for testing the patch, Richard. The results mean no go for the patch to me. > Without knowing much of the code I wonder if we can check whether the move > will be to a reg in GENERAL_REGS? That is, do we know whether there are > (besides some special constants like zero), immediate moves to the > destination register class? > There are no such info from the target code. Ideally we need to have the cost of loading *particular* immediate value into register class on the same cost basis as load/store. Still to use this info efficiently choosing alternatives should be based on costs not on the hints and some machine independent general heuristics (as now). > That said, given the result on LBM I'd not change this at this point. > > Honza wanted to look at the move pattern to try to mitigate the > GPR spilling of XMMs. > > I do think that we need to take costs into account at some point and get > rid of the reload style hand-waving with !?* in the move patterns. In general I am agree with the direction but it will be quite hard to do. I know it well from my experience to change register class cost calculation algorithm in IRA (the experimental code can be found on the branch ira-select). I expect huge number of test failures and some benchmark performance degradation practically for any targets and a big involvement of target maintainers to fix them. Although it is possible to try to do this for one target at the time.
