[llvm] [RISCV] Sink vp.splat operands of VP intrinsic. (PR #133245)

[via llvm-commits]

================
@@ -2772,6 +2772,40 @@ bool RISCVTTIImpl::canSplatOperand(Instruction *I, int Operand) const {
   }
 }
 
+bool RISCVTTIImpl::tryToSinkVPSplat(VPIntrinsic *VPI,
+                                    SmallVectorImpl<Use *> &Ops) const {
+  Value *EVL = VPI->getVectorLengthParam();
+  if (!EVL)
+    return false;
+
+  for (auto &Op : VPI->operands()) {
+    auto *I = dyn_cast<Instruction>(Op.get());
+    if (!I || I->getParent() == VPI->getParent() ||
+        llvm::is_contained(Ops, &Op))
+      continue;
+
+    // We are looking for a vp.splat that can be sunk.
+    if (!match(I, m_Intrinsic<Intrinsic::experimental_vp_splat>(
+                      m_Value(), m_AllOnes(), m_Specific(EVL))))
----------------
NexMing wrote:

> I understand that vp.splat is most likely to be all ones, I was just a little concerned (and maybe playing devil's advocate here) that we might miss some optimizations opportunities. But please don't let this block the patch.

In general, after vectorization, different basic blocks have different mask values, so the likelihood of vp.splat's mask matching (user) VP intrinsic in different basic blocks is rare. Therefore, I think the necessity of this optimization is not significant.

https://github.com/llvm/llvm-project/pull/133245