[llvm] [RISCV] Lower unmasked zero-stride vp.stride to a splat of one scalar load. (PR #97394)

[Luke Lau via llvm-commits]

================
@@ -11666,31 +11666,56 @@ SDValue RISCVTargetLowering::lowerVPStridedLoad(SDValue Op,
   auto *VPNode = cast<VPStridedLoadSDNode>(Op);
   // Check if the mask is known to be all ones
   SDValue Mask = VPNode->getMask();
+  SDValue VL = VPNode->getVectorLength();
+  SDValue Stride = VPNode->getStride();
   bool IsUnmasked = ISD::isConstantSplatVectorAllOnes(Mask.getNode());
-
-  SDValue IntID = DAG.getTargetConstant(IsUnmasked ? Intrinsic::riscv_vlse
-                                                   : Intrinsic::riscv_vlse_mask,
-                                        DL, XLenVT);
-  SmallVector<SDValue, 8> Ops{VPNode->getChain(), IntID,
-                              DAG.getUNDEF(ContainerVT), VPNode->getBasePtr(),
-                              VPNode->getStride()};
-  if (!IsUnmasked) {
-    if (VT.isFixedLengthVector()) {
-      MVT MaskVT = ContainerVT.changeVectorElementType(MVT::i1);
-      Mask = convertToScalableVector(MaskVT, Mask, DAG, Subtarget);
+  SDValue Result, Chain;
+
+  // TODO: We restrict this to unmasked loads currently in consideration of
+  // the complexity of handling all falses masks.
+  MVT ScalarVT = ContainerVT.getVectorElementType();
+  if (IsUnmasked && isNullConstant(Stride) && ContainerVT.isInteger() &&
+      !Subtarget.hasOptimizedZeroStrideLoad()) {
+    SDValue ScalarLoad =
+        DAG.getExtLoad(ISD::EXTLOAD, DL, XLenVT, VPNode->getChain(),
+                       VPNode->getBasePtr(), ScalarVT, VPNode->getMemOperand());
+    Chain = ScalarLoad.getValue(1);
----------------
lukel97 wrote:

I just noticed this is the same code as how `Intrinsic::riscv_masked_strided_load` is lowered, but I'm wondering in both riscv_masked_strided_load and vp_load should we not be also checking the AVL/EVL is non-zero?

I see the requirement for it being unmasked was [discussed here](https://reviews.llvm.org/D137931) but I can't see the AVL mentioned. Maybe I'm missing something.

https://github.com/llvm/llvm-project/pull/97394