38 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck %s3; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck %s4 5declare { i32, i32 } @llvm.amdgcn.ds.bvh.stack.rtn(i32, i32, <4 x i32>, i32 immarg)6 7define amdgpu_gs void @test_ds_bvh_stack(i32 %addr, i32 %data0, <4 x i32> %data1, ptr addrspace(1) %out) {8; CHECK-LABEL: test_ds_bvh_stack:9; CHECK: ; %bb.0:10; CHECK-NEXT: ds_bvh_stack_rtn_b32 v1, v0, v1, v[2:5]11; CHECK-NEXT: s_waitcnt lgkmcnt(0)12; CHECK-NEXT: v_add_nc_u32_e32 v0, v1, v013; CHECK-NEXT: global_store_b32 v[6:7], v0, off14; CHECK-NEXT: s_endpgm15 %pair = call { i32, i32 } @llvm.amdgcn.ds.bvh.stack.rtn(i32 %addr, i32 %data0, <4 x i32> %data1, i32 0)16 %vdst = extractvalue { i32, i32 } %pair, 017 %newaddr = extractvalue { i32, i32 } %pair, 118 %res = add i32 %vdst, %newaddr19 store i32 %res, ptr addrspace(1) %out, align 420 ret void21}22 23define amdgpu_gs void @test_ds_bvh_stack_1(i32 %addr, i32 %data0, <4 x i32> %data1, ptr addrspace(1) %out) {24; CHECK-LABEL: test_ds_bvh_stack_1:25; CHECK: ; %bb.0:26; CHECK-NEXT: ds_bvh_stack_rtn_b32 v1, v0, v1, v[2:5] offset:127; CHECK-NEXT: s_waitcnt lgkmcnt(0)28; CHECK-NEXT: v_add_nc_u32_e32 v0, v1, v029; CHECK-NEXT: global_store_b32 v[6:7], v0, off30; CHECK-NEXT: s_endpgm31 %pair = call { i32, i32 } @llvm.amdgcn.ds.bvh.stack.rtn(i32 %addr, i32 %data0, <4 x i32> %data1, i32 1)32 %vdst = extractvalue { i32, i32 } %pair, 033 %newaddr = extractvalue { i32, i32 } %pair, 134 %res = add i32 %vdst, %newaddr35 store i32 %res, ptr addrspace(1) %out, align 436 ret void37}38