[clang] [X86] Allow PSHUFD/PSHUFLW/PSHUFW intrinsics in constexpr. (PR #161210)

[Simon Pilgrim via cfe-commits]

================
@@ -2773,6 +2773,62 @@ static bool interp__builtin_blend(InterpState &S, CodePtr OpPC,
   return true;
 }
 
+enum class Half { None, Low, High };
+
+static bool interp__builtin_ia32_pshuf(InterpState &S, CodePtr OpPC,
+                                       const CallExpr *Call, Half whichHalf) {
+  assert(Call->getNumArgs() == 2 && "masked forms handled via select*");
+  APSInt ControlImm = popToAPSInt(S, Call->getArg(1));
+  const Pointer &Src = S.Stk.pop<Pointer>();
+  const Pointer &Dst = S.Stk.peek<Pointer>();
+
+  unsigned NumElems = Dst.getNumElems();
+  PrimType ElemT = Dst.getFieldDesc()->getPrimType();
+
+  unsigned ElemBits = static_cast<unsigned>(primSize(ElemT) * 8);
+  if (ElemBits != 16 && ElemBits != 32)
+    return false;
+
+  unsigned LaneElts = 128u / ElemBits;
+  assert(LaneElts && (NumElems % LaneElts == 0));
+
+  uint8_t ctl = static_cast<uint8_t>(ControlImm.getZExtValue());
+
+  for (unsigned idx = 0; idx != NumElems; idx++) {
+    unsigned LaneBase = (idx / LaneElts) * LaneElts;
+    unsigned LaneIdx = idx % LaneElts;
+
+    unsigned SrcIdx = idx;
+
+    if (ElemBits == 32) {
+      unsigned sel = (ctl >> (2 * LaneIdx)) & 0x3;
----------------
RKSimon wrote:

common expression - hoist this: `unsigned Sel = (Ctrl >> (2 * LaneIdx)) & 0x3;` and then adjust it for the ShufHW case

https://github.com/llvm/llvm-project/pull/161210