[llvm] [IA][RISCV] Support VP loads/stores in InterleavedAccessPass (PR #120490)

[Craig Topper via llvm-commits]

================
@@ -643,43 +666,92 @@ bool InterleavedAccessImpl::lowerDeinterleaveIntrinsic(
                                    DeinterleaveDeadInsts))
     return false;
 
-  LLVM_DEBUG(dbgs() << "IA: Found a deinterleave intrinsic: " << *DI
-                    << " with factor = " << DeinterleaveValues.size() << "\n");
+  const unsigned Factor = DeinterleaveValues.size();
 
-  // Try and match this with target specific intrinsics.
-  if (!TLI->lowerDeinterleaveIntrinsicToLoad(LI, DeinterleaveValues))
-    return false;
+  if (auto *VPLoad = dyn_cast<VPIntrinsic>(LoadedVal)) {
+    if (VPLoad->getIntrinsicID() != Intrinsic::vp_load)
+      return false;
+    // Check mask operand. Handle both all-true and interleaved mask.
+    Value *WideMask = VPLoad->getOperand(1);
+    std::optional<Value *> Mask = getMask(WideMask, Factor);
+    if (!Mask)
+      return false;
+
+    LLVM_DEBUG(dbgs() << "IA: Found a vp.load with deinterleave intrinsic "
+                      << *DI << " and factor = " << Factor << "\n");
+
+    // Since lowerInterleaveLoad expects Shuffles and LoadInst, use special
+    // TLI function to emit target-specific interleaved instruction.
+    if (!TLI->lowerInterleavedScalableLoad(VPLoad, *Mask, DI,
+                                           DeinterleaveValues))
----------------
topperc wrote:

The vectorizer shouldn't generate deinterleave/interleave intrinsics for fixed vectors, but maybe we should guard against handwritten IR?

https://github.com/llvm/llvm-project/pull/120490