[llvm] [AMDGPU] Vectorize more 16 bit shuffles (PR #90648)

[Matt Arsenault via llvm-commits]

================
@@ -1129,31 +1129,55 @@ InstructionCost GCNTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
                                            int Index, VectorType *SubTp,
                                            ArrayRef<const Value *> Args,
                                            const Instruction *CxtI) {
+  if (!isa<FixedVectorType>(VT))
+    return BaseT::getShuffleCost(Kind, VT, Mask, CostKind, Index, SubTp);
+
   Kind = improveShuffleKindFromMask(Kind, Mask, VT, Index, SubTp);
-  // Treat extractsubvector as single op permutation.
-  bool IsExtractSubvector = Kind == TTI::SK_ExtractSubvector;
-  if (IsExtractSubvector)
-    Kind = TTI::SK_PermuteSingleSrc;
-
-  if (ST->hasVOP3PInsts()) {
-    if (cast<FixedVectorType>(VT)->getNumElements() == 2 &&
-        DL.getTypeSizeInBits(VT->getElementType()) == 16) {
-      // With op_sel VOP3P instructions freely can access the low half or high
-      // half of a register, so any swizzle is free.
 
-      switch (Kind) {
-      case TTI::SK_Broadcast:
-      case TTI::SK_Reverse:
-      case TTI::SK_PermuteSingleSrc:
+  // Larger vector widths may require additional instructions, but are
+  // typically cheaper than scalarized versions.
+  unsigned NumVectorElts = cast<FixedVectorType>(VT)->getNumElements();
+  if (ST->getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS &&
+      DL.getTypeSizeInBits(VT->getElementType()) == 16) {
+    bool HasVOP3P = ST->hasVOP3PInsts();
+    unsigned RequestedElts =
+        count_if(Mask, [](int MaskElt) { return MaskElt != -1; });
+    if (RequestedElts == 0)
+      return 0;
+    switch (Kind) {
+    case TTI::SK_Broadcast:
+    case TTI::SK_Reverse:
+    case TTI::SK_PermuteSingleSrc: {
+      // With op_sel VOP3P instructions freely can access the low half or high
+      // half of a register, so any swizzle of two elements is free.
+      if (HasVOP3P && NumVectorElts == 2)
         return 0;
-      default:
-        break;
-      }
+      unsigned NumPerms = alignTo(RequestedElts, 2) / 2;
+      // SK_Broadcast just reuses the same mask
+      unsigned NumPermMasks = Kind == TTI::SK_Broadcast ? 1 : NumPerms;
+      return NumPerms + NumPermMasks;
+    }
+    case TTI::SK_ExtractSubvector:
+    case TTI::SK_InsertSubvector: {
+      if (HasVOP3P && NumVectorElts == 2)
+        return 0;
+      // Insert/extract subvectors require only shifts / extract code to get the
+      // relevant bits
+      return alignTo(RequestedElts, 2) / 2;
----------------
arsenm wrote:

There's only a shift if you need to access the high half of a register 

https://github.com/llvm/llvm-project/pull/90648