[llvm] [SLP][AMDGPU] Vectorize operands of non-trivially-vectorizable intrinsic calls (PR #189784)

[Alexey Bataev via llvm-commits]

================
@@ -29461,6 +29501,22 @@ bool SLPVectorizerPass::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
       PostProcessCmps.insert(cast<CmpInst>(&*It));
   }
 
+  DenseMap<Intrinsic::ID, SmallSetVector<Value *, 4>> IntrinsicSeedOps;
+  for (Instruction &I : *BB) {
+    if (R.isDeleted(&I))
+      continue;
+    // Collect operands of non-trivially vectorizable intrinsic calls (e.g.,
+    // llvm.amdgcn.exp2) and group by intrinsic ID, so their operands can be
+    // vectorized independently.
+    if (Value *Op = getNonTriviallyVectorizableIntrinsicCallOperand(&I))
+      IntrinsicSeedOps[cast<CallInst>(&I)->getIntrinsicID()].insert(Op);
+  }
+  // Try to vectorize per intrinsic call ID.
+  for (auto &[ID, Ops] : IntrinsicSeedOps) {
+    if (Ops.size() >= 2)
+      Changed |= tryToVectorizeList(Ops.getArrayRef(), R);
----------------
alexey-bataev wrote:

I think we already handle this in tryToVectorizeList

https://github.com/llvm/llvm-project/pull/189784