[Bug rtl-optimization/68194] [6 Regression] wrong code at -O2 and -O3 on x86_64-linux-gnu

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68194

--- Comment #2 from ktkachov at gcc dot gnu.org ---
Created attachment 36640
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=36640&action=edit
cprop dump before the suspected bad transformation

Confirmed the failure.
However I think the cited revision (RTL if-conversion change) exposes another
bug.

I think cse goes wrong and ends up CSEing a memory load from a symbol_ref to
'f':
(insn 67 66 68 13 (set (reg:HI 108 [ f ])
        (mem/c:HI (symbol_ref:DI ("f") [flags 0x2]  <var_decl 0x7f6a29585c60
f>) [2 f+0 S2 A16])) fail.c:28 88 {*movhi_internal}
     (nil))
(insn 68 67 69 13 (set (reg:SI 109 [ f ])
        (sign_extend:SI (subreg:QI (reg:HI 108 [ f ]) 0))) fail.c:28 146
{extendqisi2}
     (expr_list:REG_DEAD (reg:HI 108 [ f ])
        (nil)))

in basic block 13 gets transformed to:
(insn 68 66 69 13 (set (reg:SI 109 [ f ])
        (sign_extend:SI (subreg:QI (reg:HI 112 [ f ]) 0))) fail.c:28 146
{extendqisi2}
     (expr_list:REG_DEAD (reg:HI 108 [ f ])
        (nil)))

because (symbol_ref:DI ("f") gets loaded into reg 112 in another basic block
(bb 8).
This seems bogus to me.
Indeed, adding -fno-rerun-cse-after-loop (the flag controlling cse2) to the
flags makes the testcase pass