[clang] [HLSL] Add `InterlockedExchange` function and resource methods (PR #222161)

[Helena Kotas via cfe-commits]

================
@@ -0,0 +1,48 @@
+// RUN: %clang_cc1 -std=hlsl2021 -finclude-default-header -triple \
+// RUN:   dxil-pc-shadermodel6.6-library %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,DXCHECK
+
+// RUN: %clang_cc1 -std=hlsl2021 -finclude-default-header -triple \
+// RUN:   spirv-pc-vulkan-library %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,SPVCHECK
+
+// Test basic lowering of HLSL InterlockedExchange to `atomicrmw xchg
+// monotonic`. InterlockedExchange always reports the previous value, so it
+// only has a 3-argument form.
+
+groupshared int  gs_i32;
+groupshared uint gs_u32;
+groupshared int64_t  gs_i64;
+groupshared uint64_t gs_u64;
+
+// CHECK-LABEL: define {{.*}}void @{{.*}}test_int_3arg
+// DXCHECK:  %[[R:.*]] = atomicrmw xchg ptr addrspace(3) {{.*}}@gs_i32{{.*}}, i32 %{{.*}} syncscope("workgroup") monotonic
+// SPVCHECK: %[[R:.*]] = atomicrmw xchg ptr addrspace(3) {{.*}}@gs_i32{{.*}}, i32 %{{.*}} syncscope("workgroup") monotonic
----------------
hekota wrote:

Looks like the DXIL and SPIR-V expected results are identical and we don't need a separate checks (applies to the rest of the test too):

```suggestion
// CHECK:  %[[R:.*]] = atomicrmw xchg ptr addrspace(3) {{.*}}@gs_i32{{.*}}, i32 %{{.*}} syncscope("workgroup") monotonic
```

https://github.com/llvm/llvm-project/pull/222161