[llvm] [RISCV] Add optimization for memset inline (PR #146673)

[Craig Topper via llvm-commits]

================
@@ -25011,8 +25015,23 @@ EVT RISCVTargetLowering::getOptimalMemOpType(
   // a large scalar constant and instead use vmv.v.x/i to do the
   // broadcast.  For everything else, prefer ELenVT to minimize VL and thus
   // maximize the chance we can encode the size in the vsetvli.
-  MVT ELenVT = MVT::getIntegerVT(Subtarget.getELen());
-  MVT PreferredVT = (Op.isMemset() && !Op.isZeroMemset()) ? MVT::i8 : ELenVT;
+  // If Op size is greater than LMUL8 memory operation, we don't support inline
+  // of memset. Return EVT based on Op size to avoid redundant splitting and
+  // merging operations if Op size is no greater than LMUL8 memory operation.
+  if (Op.isMemset()) {
+    if (!Op.isZeroMemset())
+      return EVT::getVectorVT(Context, MVT::i8, Op.size());
+    if (Op.size() >
+        Subtarget.getMaxLMULForFixedLengthVectors() * MinVLenInBytes)
+      return MVT::Other;
+    if (Subtarget.hasVInstructionsI64() && Op.size() % 8 == 0)
+      return EVT::getVectorVT(Context, MVT::i64, Op.size() / 8);
----------------
topperc wrote:

Do we need to check alignment for any of these types?

https://github.com/llvm/llvm-project/pull/146673