You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
An unintuitive problem that @paili0628 ran into for the #959 pass implementation is that the default pipeline now needs to call dead-group-removal twice. However, the -d <pass> option disables all instances the pass because it is excluded from the pass pipeline. The tests/passes/tdcc tests do -d post-opt which is an alias to -d dead-group-removal -d dead-cell-removal ... which hides the fact that the dead-group-removal pass is being implicitly disabled make the cause of the bug non-obvious.
We probably need to rethink the -d flag a little more and see if it makes sense. Long term, we probably need to re-design the pass manager.
The text was updated successfully, but these errors were encountered:
ahhh interesting and tricky! Maybe aliases should be more of a first class scheduling concept so disabling them is disabling a particular block in the main compilation chain rather than disabling each of its constituent passes
An unintuitive problem that @paili0628 ran into for the #959 pass implementation is that the default pipeline now needs to call
dead-group-removal
twice. However, the-d <pass>
option disables all instances the pass because it is excluded from the pass pipeline. Thetests/passes/tdcc
tests do-d post-opt
which is an alias to-d dead-group-removal -d dead-cell-removal ...
which hides the fact that thedead-group-removal
pass is being implicitly disabled make the cause of the bug non-obvious.We probably need to rethink the
-d flag
a little more and see if it makes sense. Long term, we probably need to re-design the pass manager.The text was updated successfully, but these errors were encountered: