[clang] [llvm] [SPIRV] GPU intrinsics (PR #131190)

[Matt Arsenault via llvm-commits]

================
@@ -2861,6 +2861,69 @@ def int_experimental_convergence_anchor
 def int_experimental_convergence_loop
   : DefaultAttrsIntrinsic<[llvm_token_ty], [], [IntrNoMem, IntrConvergent]>;
 
+//===------- GPU Intrinsics -----------------------------------------------===//
+
+class GPUIntrinsic<LLVMType ret_type, string name>
+  : DefaultAttrsIntrinsic<[ret_type],
+                          [],
+                          [NoUndef<RetIndex>, IntrNoMem, IntrSpeculatable]>,
+                          ClangBuiltin<name>;
----------------
arsenm wrote:

It will break the AMDGPU attributor.

> We don't need an extra test showing that the amdgpu attributor works correctly when shown unknown intrinsics, because if it doesn't have that property already it's already severely broken.

Yes we do. You're creating something that isn't an unknown intrinsic. It becomes a known intrinsic, which is assumed does not call other intrinsics. e.g. https://godbolt.org/z/zjPdEbz8j. The current reasonable assumption is intrinsics are leaves that do not call other intrinsics. It may work as hoped if you strip out some set of attributes (probably need to lose nocallback at least), but we need this test to show it.

> unless at some point we decide the amdgpu attributor should start working in terms of target agnostic intrinsics at which point it'll work fine. 

We don't want this, particularly for the nontrivial cases that aren't just read argument register.




https://github.com/llvm/llvm-project/pull/131190