[Bug target/102154] [12 Regression] ICE in extract_insn, at recog.c:2769 since r12-3277-gd2874d905647a1d146dafa60199d440e837adc4d

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102154

--- Comment #17 from Segher Boessenkool <segher at gcc dot gnu.org> ---
(In reply to Hongtao.liu from comment #15)
> as discussed in
> https://gcc.gnu.org/pipermail/gcc-patches/2021-August/578437.html, allow
> specific float-int subreg seems weird.

Indiscriminately allowing all of them is a bad idea.  No machine (that has
more than just integer hardware) can actually use those, resulting in ICEs
(like here) or inefficient code generated by reload, or a mix, or worse.

> And when i look the pattern, it also seems to be strange to disallow subreg
> in operands[1], IMHO things like TARGET_ALLOW_SF_SUBREG should be done in
> TARGET_CAN_CHANGE_MODE_CLASS, so reload/lra will do the right thing.

No.  We usually have this in the predicates.  But there simply is no even
remotely efficient way to do such moves, so we should not pretend there is.

It would be nicer to disallow such subregs elsewhere, sure.

(And rs6000_can_change_mode_class already disallows this move fwiw).