[PATCH] D26053: [CodeGen] Break critical edge from RTC to original loop.

[Eli Friedman via llvm-commits]

efriedma added inline comments.


================
Comment at: lib/Support/ScopHelper.cpp:232
       : Expander(SCEVExpander(SE, DL, Name)), SE(SE), Name(Name), R(R),
-        VMap(VMap) {}
+        VMap(VMap), RTCBB(RTCBB) {}
 
----------------
grosser wrote:
> I like that this patch makes the insert position of new instructions explicit, as the old code was clearly very fragile.
> 
> This change exposes really an unnecessary strong connection between the insert location and the run-time block where we currently insert code to. I wonder if we not even want to just name this InsertBlock instead of RTCBB. 
I'm currently not completely sure what the intent here actually is... the code in question is exercised by precisely one regression test, and there aren't any comments explaining the logic.

I think it makes sense to make the variable name reflect the actual value; if someone does come through and clean this up at some point, they can rename it to something appropriate, or make the function take an actual insertion point as a parameter.


Repository:
  rL LLVM

https://reviews.llvm.org/D26053