[AArch64] Add ZIP and UZP shuffle costs. (#88150)

This adds some costs for the shuffle instructions that should be lowered
to zip1/zip2/uzp1/uzp2 instructions.

Author: davemgreen

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -11851,35 +11851,6 @@ static bool isREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
   return true;
 }
 
-static bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
-  unsigned NumElts = VT.getVectorNumElements();
-  if (NumElts % 2 != 0)
-    return false;
-  WhichResult = (M[0] == 0 ? 0 : 1);
-  unsigned Idx = WhichResult * NumElts / 2;
-  for (unsigned i = 0; i != NumElts; i += 2) {
-    if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
-        (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
-      return false;
-    Idx += 1;
-  }
-
-  return true;
-}
-
-static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
-  unsigned NumElts = VT.getVectorNumElements();
-  WhichResult = (M[0] == 0 ? 0 : 1);
-  for (unsigned i = 0; i != NumElts; ++i) {
-    if (M[i] < 0)
-      continue; // ignore UNDEF indices
-    if ((unsigned)M[i] != 2 * i + WhichResult)
-      return false;
-  }
-
-  return true;
-}
-
 static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
   unsigned NumElts = VT.getVectorNumElements();
   if (NumElts % 2 != 0)

llvm/lib/Target/AArch64/AArch64PerfectShuffle.h

@@ -16,6 +16,8 @@
 
 #include "llvm/ADT/ArrayRef.h"
 
+namespace llvm {
+
 // 31 entries have cost 0
 // 756 entries have cost 1
 // 3690 entries have cost 2
@@ -6618,4 +6620,35 @@ static unsigned getPerfectShuffleCost(llvm::ArrayRef<int> M) {
   return (PFEntry >> 30) + 1;
 }
 
+inline bool isZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
+  unsigned NumElts = VT.getVectorNumElements();
+  if (NumElts % 2 != 0)
+    return false;
+  WhichResult = (M[0] == 0 ? 0 : 1);
+  unsigned Idx = WhichResult * NumElts / 2;
+  for (unsigned i = 0; i != NumElts; i += 2) {
+    if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
+        (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx + NumElts))
+      return false;
+    Idx += 1;
+  }
+
+  return true;
+}
+
+inline bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
+  unsigned NumElts = VT.getVectorNumElements();
+  WhichResult = (M[0] == 0 ? 0 : 1);
+  for (unsigned i = 0; i != NumElts; ++i) {
+    if (M[i] < 0)
+      continue; // ignore UNDEF indices
+    if ((unsigned)M[i] != 2 * i + WhichResult)
+      return false;
+  }
+
+  return true;
+}
+
+} // namespace llvm
+
 #endif

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

@@ -3932,6 +3932,16 @@ InstructionCost AArch64TTIImpl::getShuffleCost(
       }))
     return 0;
 
+  // Check for other shuffles that are not SK_ kinds but we have native
+  // instructions for, for example ZIP and UZP.
+  unsigned Unused;
+  if (LT.second.isFixedLengthVector() &&
+      LT.second.getVectorNumElements() == Mask.size() &&
+      (Kind == TTI::SK_PermuteTwoSrc || Kind == TTI::SK_PermuteSingleSrc) &&
+      (isZIPMask(Mask, LT.second, Unused) ||
+       isUZPMask(Mask, LT.second, Unused)))
+    return 1;
+
   if (Kind == TTI::SK_Broadcast || Kind == TTI::SK_Transpose ||
       Kind == TTI::SK_Select || Kind == TTI::SK_PermuteSingleSrc ||
       Kind == TTI::SK_Reverse || Kind == TTI::SK_Splice) {