https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108826
--- Comment #20 from Jeffrey A. Law <law at gcc dot gnu.org> --- No idea why you stated "simplify-rtx is essentially never a good place to optimize the generated machine code" That's just patently false. There are all kinds of opportunities where simplify-rtx is capable to eliminating operators. And yes, there many cases where 4->3 combination support would be useful, particularly on RISC-V, but I strongly suspect it would be useful on other ports as well. I'd *love* to be able to write a define_split for those cases, but I suspect the book keeping necessary to make that work in combine is insanely complex. So we're sadly stuck using define_insn_and_split idioms which inherently break the cost modeling in combine. It's a lousy state of affairs.
