[Bug target/86891] [9 Regression] __builtin_sub_overflow incorrect for unsigned types

[wilco at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86891

--- Comment #5 from Wilco <wilco at gcc dot gnu.org> ---
(In reply to Richard Earnshaw from comment #4)
> Yes, the extension should be zero-extend, not sign extend.  The plus
> operation is correct, however, since decrementing the first operand could
> lead to underflow if it was zero.  So the correct rtl would be 
> 
>   (compare ((zero_x(a)) (plus (zero_x(b) (ltu(cc, 0)))))
>   (minus (...))

Agreed. The issue is more widespread though, signed underflow doesn't work
either,
and subv<mode>4 simply uses gen_sub<mode>3_compare1 which does do a normal
compare, so the RTL does not compute the overflow flag eventhough the actual
compare does.

It seems like a bug when it does constant fold this RTL given it should have
been folded before expand. Using UNSPEC for these complex flag uses may be best
to be safe.