[PATCH] D122148: [SLP] Peak into loads when hitting the RecursionMaxDepth

Thu Mar 24 05:51:08 PDT 2022

ABataev added inline comments.

================
Comment at: llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:3998
+  if (Depth >= RecursionMaxDepth &&
+      !(S.MainOp && match(S.MainOp, m_ZExtOrSExtOrSelf(m_Load(m_Value()))))) {
     LLVM_DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n");
----------------
dmgreen wrote:
> ABataev wrote:
> > It might be not profitable, e.g. if vector extension is not free, while scalar extension is free, and loads are not vectorizable. Also, what if you have an extension of a load and some other instructions, not loads/extractelements, etc.?
> Im not sure I understand what you mean by unprofitable? This just stops zext(load being forced to be gather if it hits the max depth. It should just mean that those node are either better (not gathers) or the same and shouldn't lead to regressions. It was previously hitting an arbitrary limit - you could say the same where any arbitrary limit causes arbitrary problems. Giving the loads the ability to order nicely should be a bigger win.
> 
> For the second part - do you mean a AltOp? If so then that makes sense, we can add a check for that, making sure it is the same as the MainOp.
1. The cost of vector sext/zext is larger than the cost of scalar sext/zext (which might be free in many cases).
2. If S.MainOp is zext/sext(load), it does not mean that all values are zext/sext(load), they might be sext/sext(load,extract,binop,etc.), since you're checking only the mainop.  

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D122148/new/

https://reviews.llvm.org/D122148