[llvm] [AArch64][Codegen]Transform saturating smull to sqdmulh (PR #143671)

[David Green via llvm-commits]

================
@@ -20717,6 +20717,83 @@ static SDValue performBuildVectorCombine(SDNode *N,
   return SDValue();
 }
 
+// A special combine for the vqdmulh family of instructions.
+// truncate( smin( sra ( mul( sext v0, sext v1 ) ), SHIFT_AMOUNT ),
+// SATURATING_VAL ) can be reduced to sqdmulh(...)
+static SDValue trySQDMULHCombine(SDNode *N,
+                                 TargetLowering::DAGCombinerInfo &DCI,
+                                 SelectionDAG &DAG) {
+
+  if (N->getOpcode() != ISD::TRUNCATE)
+    return SDValue();
----------------
davemgreen wrote:

Can I make a strange suggestion - instead of matching `truncate( smin( sra(...)`, can we match `smin(sra(...` and produce `sext(sqdmulh(..))`. It has the chance of matching in more cases, and the trunc(sext()) should disappear where it is not necessary. (Providing it is equivalent of course).

https://github.com/llvm/llvm-project/pull/143671