[Bug rtl-optimization/123308] Recent ifcvt's change causes RISC-V regressions

[law at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123308

--- Comment #3 from Jeffrey A. Law <law at gcc dot gnu.org> ---
Unfortunately, those sub-word cases turn out to be semi-important; we cobbled
together some analysis code to look for cases where we were missing obviously
profitable opportunities to if-convert code and what popped out consistently
was 32-bit arithmetic on rv64.  And those usually have some mix of extenions on
the result and subregs on the inputs.  The shift count is obviously a pain
point as well.


We just want to be careful with the SUBREG and extension cases.  We started
handling them here and the result was a horrible mess from an implementation
standpoint.  We're still trying to figure out the right way to clean that mess
up for upstreaming.