two passes of simplifycfg needed to simplify code

[llvmbot]

Bugzilla Link	1786
Resolution	FIXED
Resolved on	Apr 24, 2010 07:06
Version	2.0
OS	Linux
Attachments	bitcode for test case
Reporter	LLVM Bugzilla Contributor
CC	@jayfoad,@nlewycky

Extended Description

I run this on the attached bitcode file pre.bc:

$ opt -f -o post.bc pre.bc -simplifycfg

Then in post.bc, basic block bb17 is only reachable from cond_next27, and cond_next27 is only reachable from bb17. Neither is reachable from the entry block so they should have been removed.

Running "opt ... -simplifycfg -simplifycfg" still doesn't remove them, but I guess that's because the pass manager still only runs the pass once, or something.

If I run simplifycfg again on post.bc, they do get removed.

[llvmbot]

assigned to @lattner

[llvmbot]

Bug, I think.

[nlewycky]

simplifycfg proves the branch in %cond_false, removing the only branch to %cond_true15 and deleting that block. This orphans %bb17 which is the root of (%cond_true26, %cond_next27 and %bb32) but there's a backedge from %cond_next27 to %bb17 making it appear still in use.

In order to detect this, it needs to query the dominator information. Unfortunately, the dominator information is out of date. I'm marking this as blocked by the dominator update mechanism, even though I suppose it could be special-cased in simplifycfg.

[lattner]

simplifycfg only handles trivially dead blocks. For more aggressive dce please use the adce pass.

[llvmbot]

simplifycfg only handles trivially dead blocks

In that case my complaint is that it's not idempotent! I don't know if that's a requirement for llvm passes, but it seems like a nice property, especially since it seems difficult to persuade the pass manager to run it twice.

[lattner]

Ah, I didn't get it that "opt -simplifycfg -simplifycfg" didn't work but "opt -simplifycfg | opt -simplifycfg" did. Reopening.

[lattner]

Fixed, patch here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20071112/055567.html

Thanks for reporting this!

-Chris