[ROCDL] Added global/flag data prefetch ops (llvm#171449)

This PR brings data prefetch ops to ROCDL for gfx1250 architecture.
Extended all necessary rocdl tests

Author: ravil-mobile

mlir/include/mlir/Dialect/LLVMIR/ROCDLOps.td

@@ -1173,6 +1173,31 @@ 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)> {
+  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))";
+}
+
+def ROCDL_FlatPrefetchOp :
+  ROCDL_IntrOp<"flat.prefetch", [], [], [], 0, 0, 0, 0, [1], ["scope"]>,
+  Arguments<(ins Arg<LLVM_PointerInAddressSpace<0>, "", []>:$ptr, I32Attr:$scope)> {
+  let description = [{
+    Prefetches 1 byte of data per lane using flat-memory addresses into the WGP-cache or L2-cache.
+    Available on gfx1250+.
+  }];
+  let results = (outs);
+  let assemblyFormat = "$ptr `,` `scope` $scope attr-dict `:` qualified(type($ptr))";
+}
+
 //===---------------------------------------------------------------------===//
 // MI-100 and MI-200 buffer atomic floating point add intrinsic
 

mlir/test/Dialect/LLVMIR/rocdl.mlir

@@ -878,6 +878,20 @@ llvm.func @rocdl.raw.ptr.buffer.i32(%rsrc : !llvm.ptr<8>,
   llvm.return
 }
 
+llvm.func @rocdl.global.prefetch(%ptr : !llvm.ptr<1>) {
+  // CHECK-LABEL: rocdl.global.prefetch
+  // CHECK: rocdl.global.prefetch %{{.*}}, scope 0 : !llvm.ptr<1>
+  rocdl.global.prefetch %ptr, scope 0 : !llvm.ptr<1>
+  llvm.return
+}
+
+llvm.func @rocdl.flat.prefetch(%ptr : !llvm.ptr) {
+  // CHECK-LABEL: rocdl.flat.prefetch
+  // CHECK: rocdl.flat.prefetch %{{.*}}, scope 0 : !llvm.ptr 
+  rocdl.flat.prefetch %ptr, scope 0 : !llvm.ptr
+  llvm.return
+}
+
 // -----
 
 llvm.func @rocdl.raw.buffer.f32(%rsrc : vector<4xi32>,

mlir/test/Target/LLVMIR/rocdl.mlir

@@ -1341,6 +1341,20 @@ llvm.func @rocdl.raw.ptr.buffer.load.lds(%rsrc : !llvm.ptr<8>, %dstLds : !llvm.p
   llvm.return
 }
 
+llvm.func @rocdl.global.prefetch(%ptr : !llvm.ptr<1>) {
+  // CHECK-LABEL: rocdl.global.prefetch
+  // CHECK: call void @llvm.amdgcn.global.prefetch(ptr addrspace(1) %{{.*}}, i32 0)
+  rocdl.global.prefetch %ptr, scope 0 : !llvm.ptr<1>
+  llvm.return
+}
+
+llvm.func @rocdl.flat.prefetch(%ptr : !llvm.ptr) {
+  // CHECK-LABEL: rocdl.flat.prefetch
+  // CHECK: call void @llvm.amdgcn.flat.prefetch(ptr %{{.*}}, i32 0)
+  rocdl.flat.prefetch %ptr, scope 0 : !llvm.ptr
+  llvm.return
+}
+
 llvm.func @rocdl.wmma.scale(%arg0: i32, %arg1: vector<4xf32>, %arg2: vector<8xi32>,
                             %arg3: vector<12xi32>, %arg5: vector<16xi32>,
                             %arg8: i64, %arg9: vector<8xf32>) -> vector<4xf32> {