1; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -mcpu=hawaii -passes=loop-unroll -unroll-threshold=57 -unroll-peel-count=0 -unroll-allow-partial=false -unroll-max-iteration-count-to-analyze=16 < %s | FileCheck %s 2 3@indices = external global [16 x i32] 4 5; CHECK-LABEL: @test_func_addrspacecast_cost_noop( 6; CHECK-NOT: br i1 7define amdgpu_kernel void @test_func_addrspacecast_cost_noop(ptr addrspace(1) noalias nocapture %out, ptr addrspace(1) noalias nocapture %in) #0 { 8entry: 9 br label %for.body 10 11for.body: 12 %indvars.iv = phi i32 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 13 %sum.02 = phi float [ %fmul, %for.body ], [ 0.0, %entry ] 14 %idx.ptr = getelementptr inbounds [16 x i32], ptr @indices, i32 0, i32 %indvars.iv 15 %index = load i32, ptr %idx.ptr 16 %arrayidx.in = getelementptr inbounds float, ptr addrspace(1) %in, i32 %index 17 %arrayidx.out = getelementptr inbounds float, ptr addrspace(1) %out, i32 %index 18 %cast.in = addrspacecast ptr addrspace(1) %arrayidx.in to ptr 19 %cast.out = addrspacecast ptr addrspace(1) %arrayidx.out to ptr 20 %load = load float, ptr %cast.in 21 %fmul = fmul float %load, %sum.02 22 store float %fmul, ptr %cast.out 23 %indvars.iv.next = add i32 %indvars.iv, 1 24 %exitcond = icmp eq i32 %indvars.iv.next, 16 25 br i1 %exitcond, label %for.end, label %for.body 26 27for.end: 28 ret void 29} 30 31; Free, but not a no-op 32; CHECK-LABEL: @test_func_addrspacecast_cost_free( 33; CHECK-NOT: br i1 34define amdgpu_kernel void @test_func_addrspacecast_cost_free(ptr noalias nocapture %out, ptr noalias nocapture %in) #0 { 35entry: 36 br label %for.body 37 38for.body: 39 %indvars.iv = phi i32 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 40 %sum.02 = phi float [ %fmul, %for.body ], [ 0.0, %entry ] 41 %idx.ptr = getelementptr inbounds [16 x i32], ptr @indices, i32 0, i32 %indvars.iv 42 %index = load i32, ptr %idx.ptr 43 %arrayidx.in = getelementptr inbounds float, ptr %in, i32 %index 44 %arrayidx.out = getelementptr inbounds float, ptr %out, i32 %index 45 %cast.in = addrspacecast ptr %arrayidx.in to ptr addrspace(3) 46 %cast.out = addrspacecast ptr %arrayidx.out to ptr addrspace(3) 47 %load = load float, ptr addrspace(3) %cast.in 48 %fmul = fmul float %load, %sum.02 49 store float %fmul, ptr addrspace(3) %cast.out 50 %indvars.iv.next = add i32 %indvars.iv, 1 51 %exitcond = icmp eq i32 %indvars.iv.next, 16 52 br i1 %exitcond, label %for.end, label %for.body 53 54for.end: 55 ret void 56} 57 58; CHECK-LABEL: @test_func_addrspacecast_cost_nonfree( 59; CHECK: br i1 %exitcond 60define amdgpu_kernel void @test_func_addrspacecast_cost_nonfree(ptr addrspace(3) noalias nocapture %out, ptr addrspace(3) noalias nocapture %in) #0 { 61entry: 62 br label %for.body 63 64for.body: 65 %indvars.iv = phi i32 [ %indvars.iv.next, %for.body ], [ 0, %entry ] 66 %sum.02 = phi float [ %fmul, %for.body ], [ 0.0, %entry ] 67 %idx.ptr = getelementptr inbounds [16 x i32], ptr @indices, i32 0, i32 %indvars.iv 68 %index = load i32, ptr %idx.ptr 69 %arrayidx.in = getelementptr inbounds float, ptr addrspace(3) %in, i32 %index 70 %arrayidx.out = getelementptr inbounds float, ptr addrspace(3) %out, i32 %index 71 %cast.in = addrspacecast ptr addrspace(3) %arrayidx.in to ptr 72 %cast.out = addrspacecast ptr addrspace(3) %arrayidx.out to ptr 73 %load = load float, ptr %cast.in 74 %fmul = fmul float %load, %sum.02 75 store float %fmul, ptr %cast.out 76 %indvars.iv.next = add i32 %indvars.iv, 1 77 %exitcond = icmp eq i32 %indvars.iv.next, 16 78 br i1 %exitcond, label %for.end, label %for.body 79 80for.end: 81 ret void 82} 83 84attributes #0 = { nounwind } 85attributes #1 = { nounwind readnone speculatable } 86