[PATCH] D46039: Fix compile time hang in LSR

[Evgeny Stupachenko via Phabricator via llvm-commits]

evstupac added inline comments.


================
Comment at: lib/Transforms/Scalar/LoopStrengthReduce.cpp:3605-3606
+      // x to Depth.
+      GenerateReassociations(LU, LUIdx, LU.Formulae.back(),
+                             Depth + 1 + (Log2_32(AddOps.size()) >> 2));
   }
----------------
junbuml wrote:
> IIUC, this change perform like  if there are many reassociations at one level, then we limit the depth in the next levels. right? If we want to guarantee certain number of reassociations at each level, we may need to limit the number of function call here at each level.  I'm not sure which one is beneficial for performance in general between limiting the depth or limiting a certain number of reassociations at each level with a fixed depth. 
Yes. However the change do not limit linear case (1st level). If we want to limit 1st level, we'll need to decide which operand we use and which not, making algorithm more complicated with questionable compile time gain. Doing nothing in case we exceed some limit can hurt performance. I think that for linear cases we should leave number of reassociations as is. If you have example that shows we should limit this as well, please share.


Repository:
  rL LLVM

https://reviews.llvm.org/D46039