[Bug tree-optimization/79389] [7 Regression] 30% performance regression in SciMark2 MonteCarlo

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79389

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Well, at least in the -O2 case, it clearly shows that (RTL) if-conversion does
happen.
If SETCC + ZEXT + ADD is always a win compared to LEAL x+1 + CMOV*, then we
should just expand the cmov in that case to that or combine or peephole it to
that.
Now, with -m32 -march=i586 -O2 (so that cmov can't be used), I see
noce_try_addcc (which is exactly that, SETCC + ZEXT + ADD) fails because the
condition is complex:
(ne (and:QI (subreg:QI (zero_extract:SI (reg:HI 100)
                (const_int 8 [0x8])
                (const_int 8 [0x8])) 0)
        (const_int 5 [0x5]))
    (const_int 0 [0]))
and in that case it attempts to emit:
(insn 53 0 0 (set (reg:SI 102)
        (eq:SI (reg:CCNO 17 flags)
            (const_int 0 [0]))) -1
     (nil))
but that isn't recognized by the backend.