[clang] [Headers][X86] Allow MMX/SSE/AVX MOVMSK intrinsics to be used in constexpr (PR #161914)

[Simon Pilgrim via cfe-commits]

================
@@ -2817,6 +2817,46 @@ static bool interp__builtin_ia32_pshuf(InterpState &S, CodePtr OpPC,
   return true;
 }
 
+static bool interp__builtin_ia32_movmsk_op(InterpState &S, CodePtr OpPC,
+                                           const CallExpr *Call) {
+  assert(Call->getNumArgs() == 1);
+
+  const Pointer &Source = S.Stk.pop<Pointer>();
+
+  unsigned SourceLen = Source.getNumElems();
+  const QualType ElemQT = getElemType(Source);
+  const OptPrimType ElemPT = S.getContext().classify(ElemQT);
+  unsigned LaneWidth = S.getASTContext().getTypeSize(ElemQT);
+
+  if (ElemQT->isIntegerType()) {
+    unsigned Byte = 8;
+    unsigned ResultLen = (LaneWidth * SourceLen) / Byte;
+    APInt Result(ResultLen, 0);
+    unsigned ResultIdx = 0;
+    for (unsigned I = 0; I != SourceLen; ++I) {
+      APInt Lane;
+      INT_TYPE_SWITCH_NO_BOOL(*ElemPT,
+                              { Lane = Source.elem<T>(I).toAPSInt(); });
+      for (unsigned J = 0; J != LaneWidth; J += Byte) {
+        Result.setBitVal(ResultIdx++, Lane[J + 7]);
+      }
----------------
RKSimon wrote:

MOVMSK needs to work per element, not per byte - PMOVMSKB has a vXi8 type so each element happens to be byte size, but all the ElemBitWidth/BitsInAByte code is unnecessary

https://github.com/llvm/llvm-project/pull/161914