[llvm] [RISCV][DAG][TLI] Avoid scalarizing length decreasing shuffles (PR #115532)

[Philip Reames via llvm-commits]

================
@@ -545,6 +545,11 @@ class TargetLoweringBase {
     return DefinedValues < 3;
   }
 
+  // In SDAG construction, should length decreasing shuffles be expanded
+  // to a sequence of extracts and inserts if they can't be recognized
+  // via the two extracted operand form?
+  virtual bool shouldScalarizeLengthDecreasingShuffle() const { return true; }
----------------
preames wrote:

> Do something unconditionally and fix it up in a combiner if preferred, like anything else. 

I agree this is the goal.  My hope was that once this initial step is in, we'd move the scalarization into a DAG combine.  It's a bit tricky because of the difference between a one time and iterative decision.  The other direction we can approach this is to unconditional scalarize, but I would argue that the simpler IR should be the initial canonical form.  

> Failing that, at minimum a TLI hook needs some kind of context if it's not going to be a simple bool field

Are you saying you'd rather see this be a boolean field on TargetLoweringBase to avoid the virtual dispatch?  If so, I'm happy to make that change.  I'm just not clear on what you're asking for.

I'll also note that I was thinking ahead to the next piece I mentioned previously, and that a profitability hook for the scalarization in DAG will almost certainly require a type based hook.  



https://github.com/llvm/llvm-project/pull/115532