On Mon, 25 Mar 2019, Richard Biener wrote: > > This PR and possibly quite some dups that have been accumulating lately > run into an artifact of DCEs edge removal code when generating debug > stmts. Here DCE removes edges from the CFG at the same time it removes > dead controlling stmts but before PHI nodes are removed. This causes > us to remove PHI arguments associated with removed edges and make > those PHI appearing as degenerate, causing wrong debug stmts to be > generated. Consider for example i_1 = PHI <i_2, 0> where the edge > leading to i_2 is removed - at the time we remove the PHI it looks > like i_1 = PHI <0> and a i = 0 debug-stmt is generated. > > The fix is to delay edge removal and move PHI removal back to the > place it had originally been, doing that in reverse dominator order > as intended for debug stmt creation (it jumped from where it'll be > after the patch to all-before to all-after). > > The testcase shows up as UNRESOLVED where it formerly was > wrong-debug because we change from i = 0 to i = NULL. I've seen > no way to mark "optimized-out" as expected and at -O1 we > retain the loop and the correct value of i. > > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress. > > OK (well, I think it is, just double-checking).
I have applied this now, it doesn't immediately solve any of the other issues. Richard.
