[PATCH] D28073: Preserve domtree and loop-simplify for runtime unrolling.

[Michael Kuperstein via Phabricator via llvm-commits]

mkuper added inline comments.


================
Comment at: lib/Transforms/Utils/LoopUnroll.cpp:55
+static cl::opt<bool>
+UnrollVerifyDomtree("unroll-verify-domtree", cl::init(false), cl::Hidden,
+                    cl::desc("Verify domtree after unrolling"));
----------------
efriedma wrote:
> mkuper wrote:
> > We want to enforce this being on in new tests too, I assume?
> We probably want this on in new tests, yes, but I have no idea how we would enforce it.  
In code review, I meant. :-)


================
Comment at: lib/Transforms/Utils/LoopUnroll.cpp:619
+      if (BB == LatchBlock) {
+        // The latch is special because we emit unconditional branches in
+        // some cases.  Since the latch is always at the bottom of the loop,
----------------
efriedma wrote:
> mkuper wrote:
> > Isn't it the case that when the latch ends with an unconditional branch, that branch is towards the header block?
> > If so, we should not have any children outside the loop.
> Maybe this is more clear?  "The latch is special because we can emit an unconditional branch in the unrolled loop even if the original latch block ends in a conditional branch."
Ah, ok, this makes sense.
This still looks a bit weird to me, though. Any chance you can give an example of when this ends up different than the nearest common dominator with Latches[0]?

(Feel free to ignore me - I'm not sure I understand this enough to LGTM it anyway. :-) )



Repository:
  rL LLVM

https://reviews.llvm.org/D28073