Say that we have two blocks, B1 and B2. B1 doesn't do anything to variable X. B2 reads from variable X. Say that CFG simplification decides to merge B1 and B2. Currently, the new block, B1', will still say that the variable at head for X is Phi, which is wrong - it should be GetLocal, since that's the first thing that B1' does to X. This is asymptomatic except if we then do constant folding on X. I'm still trying to figure out a test case for this, but it's not looking so good.
OK found a reduced test case. Making a layout test.
Created attachment 145685 [details] the patch
Comment on attachment 145685 [details] the patch r=me
Landed in http://trac.webkit.org/changeset/119464