98 lines · plain
1; RUN: opt -passes=amdgpu-attributor < %s | llc | FileCheck %s2 3target triple = "amdgcn-amd-amdhsa"4 5; The call to intrinsic implicitarg_ptr reaches a load through a phi. The6; offsets of the phi cannot be determined, and hence the attirbutor assumes that7; hostcall is in use.8 9; CHECK-LABEL: amdhsa.kernels:10; CHECK: .value_kind: hidden_hostcall_buffer11; CHECK: .value_kind: hidden_multigrid_sync_arg12; CHECK-LABEL: .name: kernel_113 14define amdgpu_kernel void @kernel_1(ptr addrspace(1) %a, i64 %index1, i64 %index2, i1 %cond) {15entry:16 %tmp7 = tail call ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()17 br i1 %cond, label %old, label %new18 19old: ; preds = %entry20 %tmp4 = getelementptr i8, ptr addrspace(4) %tmp7, i64 %index121 br label %join22 23new: ; preds = %entry24 %tmp12 = getelementptr inbounds i8, ptr addrspace(4) %tmp7, i64 %index225 br label %join26 27join: ; preds = %new, %old28 %.in.in.in = phi ptr addrspace(4) [ %tmp12, %new ], [ %tmp4, %old ]29 30 ;;; THIS USE is where the offset into implicitarg_ptr is unknown31 %.in = load i16, ptr addrspace(4) %.in.in.in, align 232 33 %idx.ext = sext i16 %.in to i6434 %add.ptr3 = getelementptr inbounds i32, ptr addrspace(1) %a, i64 %idx.ext35 %tmp16 = atomicrmw add ptr addrspace(1) %add.ptr3, i32 15 syncscope("agent-one-as") monotonic, align 436 ret void37}38 39; The call to intrinsic implicitarg_ptr is combined with an offset produced by40; select'ing between two constants, before it is eventually used in a GEP to41; form the address of a load. This test ensures that AAPointerInfo can look42; through the select to maintain a set of indices, so that it can precisely43; determine that hostcall and other expensive implicit args are not in use.44 45; CHECK-NOT: hidden_hostcall_buffer46; CHECK-NOT: hidden_multigrid_sync_arg47; CHECK-LABEL: .name: kernel_248 49define amdgpu_kernel void @kernel_2(ptr addrspace(1) %a, i1 %cond) {50entry:51 %tmp7 = tail call ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()52 %tmp5 = select i1 %cond, i64 12, i64 1853 %tmp6 = getelementptr inbounds i8, ptr addrspace(4) %tmp7, i64 %tmp554 55 ;;; THIS USE is where multiple offsets are possible, relative to implicitarg_ptr56 %tmp9 = load i16, ptr addrspace(4) %tmp6, align 257 58 %idx.ext = sext i16 %tmp9 to i6459 %add.ptr3 = getelementptr inbounds i32, ptr addrspace(1) %a, i64 %idx.ext60 %tmp16 = atomicrmw add ptr addrspace(1) %add.ptr3, i32 15 syncscope("agent-one-as") monotonic, align 461 ret void62}63 64; CHECK-NOT: hidden_hostcall_buffer65; CHECK-NOT: hidden_multigrid_sync_arg66; CHECK-LABEL: .name: kernel_367 68define amdgpu_kernel void @kernel_3(ptr addrspace(1) %a, i1 %cond) {69entry:70 %tmp7 = tail call ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()71 br i1 %cond, label %old, label %new72 73old: ; preds = %entry74 %tmp4 = getelementptr i8, ptr addrspace(4) %tmp7, i64 1275 br label %join76 77new: ; preds = %entry78 %tmp12 = getelementptr inbounds i8, ptr addrspace(4) %tmp7, i64 1879 br label %join80 81join: ; preds = %new, %old82 %.in.in.in = phi ptr addrspace(4) [ %tmp12, %new ], [ %tmp4, %old ]83 84 ;;; THIS USE of implicitarg_ptr should not produce hostcall metadata85 %.in = load i16, ptr addrspace(4) %.in.in.in, align 286 87 %idx.ext = sext i16 %.in to i6488 %add.ptr3 = getelementptr inbounds i32, ptr addrspace(1) %a, i64 %idx.ext89 %tmp16 = atomicrmw add ptr addrspace(1) %add.ptr3, i32 15 syncscope("agent-one-as") monotonic, align 490 ret void91}92 93declare i32 @llvm.amdgcn.workitem.id.x()94 95declare align 4 ptr addrspace(4) @llvm.amdgcn.implicitarg.ptr()96 97declare i32 @llvm.amdgcn.workgroup.id.x()98