Don't inline on Windows on "[PyTorch] Make tls_local_dispatch_key_set…

… inlineable (reapply)"

FLAGS_disable_variable_dispatch had to go, but it looks like the only user was some benchmarks anyway.

Differential Revision: [D25547962](https://our.internmc.facebook.com/intern/diff/D25547962/)

**NOTE FOR REVIEWERS**: This PR has internal Facebook specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D25547962/)!

[ghstack-poisoned]

aten/src/ATen/native/cpu/CatKernel.cpp

@@ -20,6 +20,8 @@ struct InputMeta {
 
 template <typename scalar_t>
 void cat_serial_kernel_impl(Tensor& result, TensorList tensors, int64_t dim) {
+  TORCH_INTERNAL_ASSERT_DEBUG_ONLY(
+      dim >= 0 && dim < result.dim(), "dim out of range in cat_serial_kernel_impl");
   int64_t outer = result.numel() / (result.sizes()[dim] * result.strides()[dim]);
   scalar_t* result_data = result.data_ptr<scalar_t>();
   int64_t ninputs = tensors.size();

c10/core/impl/LocalDispatchKeySet.cpp

@@ -18,6 +18,12 @@ PODLocalDispatchKeySet raw_local_dispatch_key_set;
 
 #endif
 
+#ifdef _MSC_VER
+LocalDispatchKeySet tls_local_dispatch_key_set() {
+  return raw_local_dispatch_key_set;
+}
+#endif // _MSC_VER
+
 void _force_tls_local_dispatch_key_set(LocalDispatchKeySet key_set) {
   raw_local_dispatch_key_set = PODLocalDispatchKeySet {
     key_set.included_.raw_repr(),

c10/core/impl/LocalDispatchKeySet.h

@@ -52,6 +52,10 @@ struct C10_API LocalDispatchKeySet {
   DispatchKeySet excluded_;
 };
 
+// thread_local variables cannot be C10_API on Windows.
+#ifdef _MSC_VER
+C10_API LocalDispatchKeySet tls_local_dispatch_key_set();
+#else // _MSC_VER
 /// In the CAFFE2_FB_LIMITED_MOBILE_CAPABILITY build setting,
 /// thread_local is not supported.
 #ifndef CAFFE2_FB_LIMITED_MOBILE_CAPABILITY
@@ -65,6 +69,7 @@ inline C10_API LocalDispatchKeySet tls_local_dispatch_key_set() {
   // because they include this header.
   return raw_local_dispatch_key_set;
 }
+#endif // _MSC_VER
 
 // Internal, use ThreadLocalStateGuard
 C10_API void _force_tls_local_dispatch_key_set(LocalDispatchKeySet key_set);