[llvm] [WebAssembly] Fold extended vector shifts by constant to extmul (PR #184007)

[Sam Parker via llvm-commits]

================
@@ -2797,9 +2797,81 @@ static SDValue unrollVectorShift(SDValue Op, SelectionDAG &DAG) {
   return DAG.getBuildVector(Op.getValueType(), DL, UnrolledOps);
 }
 
+/// Convert a vector shift of an extended value into a multiplication of
+/// extended values. By converting the shift amount to a multiplier (1 << C)
+/// and wrapping it in a matching extend node, we enable the instruction
+/// selector to match the pattern to WebAssembly extended multiplication
+/// instructions (e.g., i32x4.extmul_low_i16x8_s). Inactive lanes in the
+/// multiplier vector are populated with undefs.
+///
+/// Example transformation:
+/// Before:
+///   t1: v8i16 = ...
+///   t2: v4i32 = WebAssemblyISD::EXTEND_LOW_S t1
+///   t3: v4i32 = BUILD_VECTOR Constant:i32<12>, Constant:i32<0>, ...
+///   t4: v4i32 = shl t2, t3
+///
+/// After:
+///   t1: v8i16 = ...
+///   t2: v4i32 = WebAssemblyISD::EXTEND_LOW_S t1
+///   t3: v8i16 = BUILD_VECTOR Constant:i16<4096>, Constant:i16<1>, undef, ...
+///   t4: v4i32 = WebAssemblyISD::EXTEND_LOW_S t3 t5: v4i32 = mul t2, t4
+static SDValue foldShiftByConstantToExtMul(SDValue Op, SelectionDAG &DAG) {
+  if (Op.getOpcode() != ISD::SHL || !Op.getValueType().isVector())
+    return SDValue();
+
+  SDValue RHS = Op.getOperand(1);
+  if (RHS.getOpcode() != ISD::BUILD_VECTOR)
----------------
sparker-arm wrote:

Well, is the SHL always going to operate on a `BUILD_VECTOR`? Could we not ever see a `ConstantSDNode` used directly?

https://github.com/llvm/llvm-project/pull/184007