[Mlir-commits] [mlir] [ROCDL] Added LDS barrier ops to ROCDL (gfx1250) (PR #171810)

[Jakub Kuderski]

================
@@ -1177,25 +1177,82 @@ def ROCDL_RawBufferAtomicCmpSwap :
 // Memory prefetch intrinsics
 
 def ROCDL_GlobalPrefetchOp :
-  ROCDL_IntrOp<"global.prefetch", [], [], [], 0, 0, 0, 0, [1], ["scope"]>,
-  Arguments<(ins Arg<LLVM_PointerInAddressSpace<1>, "", []>:$ptr, I32Attr:$scope)> {
+  ROCDL_IntrOp<"global.prefetch", [], [], [], 0, 0, 1, 0, [1], ["scope"]> {
+  dag args = (ins Arg<LLVM_PointerInAddressSpace<1>, "", [MemRead]>:$ptr,
+                  I32Attr:$scope);
+  let arguments = !con(args, baseArgs);
   let description = [{
     Prefetches 1 byte of data per lane from global memory into the WGP-cache or L2-cache.
     Available on gfx1250+.
   }];
   let results = (outs);
   let assemblyFormat = "$ptr `,` `scope` $scope attr-dict `:` qualified(type($ptr))";
+  let extraClassDefinition = [{
+    SmallVector<Value> $cppClass::getAccessedOperands() {
----------------
kuhar wrote:

What determined this choice of a return type? Is it for some interface / templated code? I think ideally we'd want to return a single value or `SmallVector<Value, 1>`

https://github.com/llvm/llvm-project/pull/171810