[clang] [Clang] Allow VDBPSADBW intrinsics in constexpr (PR #188887)

[Simon Pilgrim via cfe-commits]

================
@@ -2816,6 +2816,75 @@ static bool interp__builtin_ia32_pmul(
   return true;
 }
 
+static bool interp__builtin_ia32_dbpsadbw(InterpState &S, CodePtr OpPC,
+                                          const CallExpr *Call) {
+  assert(Call->getNumArgs() == 3);
+  unsigned Imm = popToUInt64(S, Call->getArg(2));
+
+  const Pointer &Src2 = S.Stk.pop<Pointer>();
+  const Pointer &Src1 = S.Stk.pop<Pointer>();
+  const Pointer &Dst = S.Stk.peek<Pointer>();
+
+  const auto *SrcVT = Call->getArg(0)->getType()->castAs<VectorType>();
+  PrimType SrcElemT = *S.getContext().classify(SrcVT->getElementType());
+  unsigned SourceLen = SrcVT->getNumElements();
+
+  const auto *DestVT = Call->getType()->castAs<VectorType>();
+  PrimType DestElemT = *S.getContext().classify(DestVT->getElementType());
+  bool DestUnsigned = Call->getType()->isUnsignedIntegerOrEnumerationType();
+
+  constexpr unsigned LaneSize = 16; // 128-bit lane = 16 bytes
+
+  // Phase 1: Shuffle Src2 using all four 2-bit fields of imm8.
+  // Within each 128-bit lane, for group j (0..3), select a 4-byte block
+  // from Src2 based on bits [2*j+1:2*j] of imm8.
+  uint8_t Shuffled[64]; // max 512-bit = 64 bytes
----------------
RKSimon wrote:

Make it work with any width?
```suggestion
  SmallVector<uint8_t, 64> Shuffled(SourceLen);
```

https://github.com/llvm/llvm-project/pull/188887