[Bug target/92905] [10 Regression] Spills float-int union to memory

[vmakarov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92905

--- Comment #5 from Vladimir Makarov <vmakarov at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #3)
> Note, it isn't about <g>, using r<i>m in the first alternative of the
> reverted define_insn works well too, as well as swapping the alternatives
> (that is in that case basically what the trunk has, except in the second
> alternative the second input is r<i>m instead of m.  If the second
> alternative of second input is <i>m,
> it works the same as current trunk (unnecessary spill), if it is rm, it
> works well.
> Now, no idea if this isn't a bug in LRA or if there are some rules that the
> seemingly redundant constraints actually aren't redundant.

The culprit is 'g' vs 'm'.  When we have 'g' matching pseudo assigned to hard
register, LRA ignores memory in 'g' when considering preferred_reload_class.

Reload also treats such situation in analogous way.  And actually LRA adapted
the reload code.

So I can try to solve this PR.  But it will take some time.  Mostly any patch
in this sensitive area should be tested and benchmarked well.