23 lines · plain
1; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=SI %s2 3declare i32 @llvm.amdgcn.mbcnt.lo(i32, i32) #04declare i32 @llvm.amdgcn.mbcnt.hi(i32, i32) #05 6; SI-LABEL: {{^}}test_array_ptr_calc:7; SI-DAG: v_mul_u32_u248; SI-DAG: v_mul_hi_u32_u249; SI: s_endpgm10define amdgpu_kernel void @test_array_ptr_calc(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %inA, ptr addrspace(1) noalias %inB) {11 %mbcnt.lo = call i32 @llvm.amdgcn.mbcnt.lo(i32 -1, i32 0)12 %tid = call i32 @llvm.amdgcn.mbcnt.hi(i32 -1, i32 %mbcnt.lo)13 %a_ptr = getelementptr [1025 x i32], ptr addrspace(1) %inA, i32 %tid, i32 014 %b_ptr = getelementptr i32, ptr addrspace(1) %inB, i32 %tid15 %a = load i32, ptr addrspace(1) %a_ptr16 %b = load i32, ptr addrspace(1) %b_ptr17 %result = add i32 %a, %b18 store i32 %result, ptr addrspace(1) %out19 ret void20}21 22attributes #0 = { nounwind readnone }23