AMDGPU: Mark ds append/consume intrinsics with align 4 (llvm#110533)

Manual says the low 2 bits of the pointer are ignored.

Author: arsenm

llvm/include/llvm/IR/IntrinsicsAMDGPU.td

@@ -558,7 +558,8 @@ class AMDGPUDSAppendConsumedIntrinsic : Intrinsic<
   [llvm_anyptr_ty, // LDS or GDS ptr
    llvm_i1_ty], // isVolatile
    [IntrConvergent, IntrWillReturn, IntrArgMemOnly,
-    NoCapture<ArgIndex<0>>, ImmArg<ArgIndex<1>>, IntrNoCallback, IntrNoFree],
+    Align<ArgIndex<0>, 4>, NoCapture<ArgIndex<0>>,
+    ImmArg<ArgIndex<1>>, IntrNoCallback, IntrNoFree],
    "",
    [SDNPMemOperand]
 >;

llvm/test/Assembler/amdgcn-intrinsic-attributes.ll

@@ -0,0 +1,21 @@
+; REQUIRES: amdgpu-registered-target
+
+; RUN: llvm-as < %s | llvm-dis | FileCheck %s
+
+; Test assumed alignment parameter
+
+; CHECK: declare i32 @llvm.amdgcn.ds.append.p3(ptr addrspace(3) nocapture align 4, i1 immarg) #0
+
+define i32 @ds_append(ptr addrspace(3) %ptr) {
+  %ret = call i32 @llvm.amdgcn.ds.append.p3(ptr addrspace(3) %ptr, i1 false)
+  ret i32 %ret
+}
+
+; Test assumed alignment parameter
+; CHECK: declare i32 @llvm.amdgcn.ds.consume.p3(ptr addrspace(3) nocapture align 4, i1 immarg) #0
+define i32 @ds_consume(ptr addrspace(3) %ptr) {
+  %ret = call i32 @llvm.amdgcn.ds.consume.p3(ptr addrspace(3) %ptr, i1 false)
+  ret i32 %ret
+}
+
+; CHECK: attributes #0 = { convergent nocallback nofree nounwind willreturn memory(argmem: readwrite) }