[PATCH] D20017: Aggressive choosing best loop top

[Ehsan Amiri via llvm-commits]

amehsan added inline comments.

================
Comment at: lib/CodeGen/MachineBlockPlacement.cpp:726-742
@@ -719,2 +725,19 @@
 /// header is always profitable.
+///
+/// a latch block with conditional exit, and similar to following cfg:
+///
+///            entry               original    better
+///              |                  layout     layout
+///  ------> loop.header (body)    --------    ------
+///  |97%    /       \             entry       entry
+///  |      /50%      \50%         loop.header latch
+///  --- latch <--- if.then        if.then     loop.header
+///         \    97%  /            latch       if.then
+///          \3%     /3%           loop.end    loop.end
+///           loop.end
+///
+/// "original layout" cause latch needs a branch jumping back to loop.header
+/// when condition is true, but in "better layout", latch can fall through
+/// loop.header without this jump.
+
 MachineBasicBlock *
----------------
amehsan wrote:
> Could this hurt performance for the following example
> 
> ```
> 
>      --------->entry
>     |  |         |
>     |  |    loop header
>     |  |         |    \
>     |  |         |      \
>     |  ---- --Latch1<----- if.then
>     |           |           |
>     ----------Latch2 <------
> 
> Probabilities that matter: loop.heaer--->if.then 99%,  if then->Latch 1 99%, Latch1 ---> header 80%
> ```
> 
> The problem is that before your change the branch from if.then to Latch 1 is fall thru. After your change it is a jump back. If we have a single latch this shouldn't hurt the performance, but with multiple latches, the number of fall-thrus that your code, removes is more than those that it creates. I am not sure how important this is.
> 
> Checking for a single latch should be straightforward using LoopInfo API if this is really a problematic edge case.
> 
In the graph that I have drawn, arrows from Latch1 and Latch2 should go back to loop header. I hope it was not confusing.


http://reviews.llvm.org/D20017