[clang] [Clang] Allow AVX/AVX512 subvector shuffles in constexpr (PR #168700)

[Simon Pilgrim via cfe-commits]

================
@@ -4847,6 +4847,39 @@ bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const CallExpr *Call,
     return interp__builtin_elementwise_triop(S, OpPC, Call,
                                              llvm::APIntOps::fshr);
 
+  case X86::BI__builtin_ia32_shuf_f32x4_256:
+  case X86::BI__builtin_ia32_shuf_i32x4_256:
+  case X86::BI__builtin_ia32_shuf_f64x2_256:
+  case X86::BI__builtin_ia32_shuf_i64x2_256:
+  case X86::BI__builtin_ia32_shuf_f32x4:
+  case X86::BI__builtin_ia32_shuf_i32x4:
+  case X86::BI__builtin_ia32_shuf_f64x2:
+  case X86::BI__builtin_ia32_shuf_i64x2: {
+    // Destination and sources A, B all have the same type.
+    QualType VecQT = Call->getArg(0)->getType();
+    const auto *VecT = VecQT->castAs<VectorType>();
+    unsigned NumElems = VecT->getNumElements();
+    unsigned ElemBits = S.getASTContext().getTypeSize(VecT->getElementType());
+    constexpr unsigned LaneBits = 128u;
----------------
RKSimon wrote:

```suggestion
    unsigned LaneBits = 128u;
```

https://github.com/llvm/llvm-project/pull/168700