32 lines · plain
1; RUN: llc < %s | FileCheck %s2; RUN: %if ptxas %{ llc < %s | %ptxas-verify %}3 4target triple = "nvptx64-nvidia-cuda"5 6; CHECK-LABEL: test17; CHECK: ld.global.v2.b328; CHECK: ld.global.v2.b329; CHECK: st.global.v2.b3210; CHECK: st.global.v2.b3211define void @test1(ptr addrspace(1) noalias align 8 %in, ptr addrspace(1) noalias align 8 %out) {12 %in.1 = getelementptr float, ptr addrspace(1) %in, i32 113 %in.2 = getelementptr float, ptr addrspace(1) %in, i32 214 %in.3 = getelementptr float, ptr addrspace(1) %in, i32 315 %v0 = load float, ptr addrspace(1) %in, align 816 %v1 = load float, ptr addrspace(1) %in.1, align 417 %v2 = load float, ptr addrspace(1) %in.2, align 818 %v3 = load float, ptr addrspace(1) %in.3, align 419 %sum0 = fadd float %v0, %v120 %sum1 = fadd float %v1, %v221 %sum2 = fadd float %v3, %v122 %sum3 = fadd float %v2, %v323 %out.1 = getelementptr float, ptr addrspace(1) %out, i32 124 %out.2 = getelementptr float, ptr addrspace(1) %out, i32 225 %out.3 = getelementptr float, ptr addrspace(1) %out, i32 326 store float %sum0, ptr addrspace(1) %out, align 827 store float %sum1, ptr addrspace(1) %out.1, align 428 store float %sum2, ptr addrspace(1) %out.2, align 829 store float %sum3, ptr addrspace(1) %out.3, align 430 ret void31}32