[Bug rtl-optimization/82692] [8 Regression] Ordered comparisons used for unordered built-ins

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82692

--- Comment #17 from Segher Boessenkool <segher at gcc dot gnu.org> ---
So we have a compare:CCFPU, the resulting flags is used in a GE
only, and ix86_cc_mode thinks the best mode to use for that is CCFP.

Which is fine, except compare:CCFPU is a different instruction, and GE
for the resulting insn means a different thing?!

How is this supposed to work?  How can generic code know this?

Everything worked fine, except the compare insn did not do the side
effect of setting a status flag.  Perhaps an unspec (or even an
unspec_volatile) should have been used for the compare?