[polly] r246441 - Always use the branch instructions to model the PHI-node writes

[Tobias Grosser via llvm-commits]

On 08/31/2015 04:23 PM, Johannes Doerfert wrote:
> I was pretty sure this worked at some point. Why else do we have the
> code below? It will create a dependence from the Instruction that is
> incoming to the PHI to the predecessor block of the PHI.

Right, I also remember we discussed this before. Maybe the codegen code
you have written was indeed correct and I broke this when moving more
knowledge into the ScopInfo modeling?

At least now we have a test case that ensures we do not
regress here again. I am indeed surprised this was not covered. I
had the feeling we committed this with pretty good test coverage
originally.

Best,
Tobias