[llvm] [AMDGPU] Cost of i8 vector insert/extract is free in some cases (PR #194991)

[Matt Arsenault via llvm-commits]

================
@@ -1009,19 +1009,32 @@ InstructionCost GCNTTIImpl::getVectorInstrCost(
   case Instruction::InsertElement: {
     unsigned EltSize
       = DL.getTypeSizeInBits(cast<VectorType>(ValTy)->getElementType());
+    // Dynamic indexing isn't free and is best avoided.
+    if (Index == ~0u)
+      return 2;
     if (EltSize < 32) {
       if (EltSize == 16 && Index == 0 && ST->has16BitInsts())
         return 0;
+      // Some i8 inserts and extracts are free so we want to reduce the
+      // cost to avoid scalarization. We limit the zero cost cases to avoid
+      // adversely impacting all i8 vectorizing.
+      if (EltSize == 8) {
+        unsigned NumElts = cast<FixedVectorType>(ValTy)->getNumElements();
+        if (NumElts >= 4 && isPowerOf2_32(NumElts)) {
+          // Extracts at indices aligned to 32-bit boundaries (0, 4, 8, 12 for
+          // v16i8) are free as they access the low byte of each VGPR. Other
+          // indices require bit manipulation (shifts/byte selects) and cost 1.
----------------
arsenm wrote:

Theoretically any byte or short mask is free with SDWA depending on the user 

https://github.com/llvm/llvm-project/pull/194991