Re: [tail-merge, PR84956] Don't merge bbs with bb_has_abnormal_pred

[Richard Biener <rguenther at suse dot de>]

On Thu, 22 Mar 2018, Tom de Vries wrote:

> Hi,
> 
> This fixes a 6/7/8 regression, a P3 ICE in the tail-merge pass.
> 
> 
> Consider the test-case from the patch.
> 
> When compiling at -O2, duplicates are found in tail-merge:
> ...
> find_duplicates: <bb 6> duplicate of <bb 7>
> find_duplicates: <bb 6> duplicate of <bb 9>
> ...
> So, tail-merge calls replace_block_by (bb7, bb6).
> 
> However, bb7 has an abnormal edge bb5 -> bb7 as predecessor (bb6 has an
> abnormal edge as predecessor as well, but that doesn't look necessary to
> trigger the ICE):
> ...
> ;;   basic block 9, loop depth 0
> ;;    prev block 3, next block 4, flags: (NEW, REACHABLE)
> ;;    pred:       3 [50.0% (guessed)] (FALSE_VALUE,EXECUTABLE)
>   goto <bb 8>; [100.00%]
> ;;    succ:       8 [always]  (FALLTHRU,EXECUTABLE)
> 
> ;;   basic block 6, loop depth 0
> ;;    prev block 5, next block 7, flags: (NEW, REACHABLE, VISITED)
> ;;    pred:       5 [33.3% (guessed)] (ABNORMAL,EXECUTABLE)
>   goto <bb 8>; [100.00%]
> ;;    succ:       8 [always] (FALLTHRU,EXECUTABLE)
> 
> ;;   basic block 7, loop depth 0
> ;;    prev block 6, next block 8, flags: (NEW, REACHABLE, VISITED)
> ;;    pred:       5 [33.3% (guessed)]  (ABNORMAL,EXECUTABLE)
> ;;    succ:       8 [always]  (FALLTHRU,EXECUTABLE)
> ...
> 
> So when replace_block_by calls redirect_edge_and_branch (bb5 -> bb7, bb6) it
> lands in gimple_redirect_edge_and_branch, and fails here:
> ...
> 6024	  if (e->flags & EDGE_ABNORMAL)
> 6025	    return NULL;
> ...
> 
> which causes this assert to trigger:
> ...
>       gcc_assert (pred_edge != NULL);
> ...
> 
> 
> The patch fixes the ICE conservatively by skipping bbs with
> bb_has_abnormal_preds in find_clusters_1.
> 
> [ A more optimal fix could be to detect in this example that there's an
> abnormal edge bb5 -> bb6, and skip redirecting bb5 -> bb7 to bb6. ]
> 
> Bootstrapped and reg-tested on x86_64.
> 
> OK for stage4, gcc-6-branch and gcc-7-branch?

OK everywhere.

Thanks,
Richard.