68 lines · plain
1; RUN: opt -mtriple=amdgcn--amdhsa -S -passes=inline -inline-threshold=0 < %s | FileCheck %s2 3define void @use_flat_ptr_arg(ptr nocapture %p) {4entry:5 %tmp1 = load float, ptr %p, align 46 %div = fdiv float 1.000000e+00, %tmp17 %add0 = fadd float %div, 1.08 %add1 = fadd float %add0, 1.09 %add2 = fadd float %add1, 1.010 %add3 = fadd float %add2, 1.011 %add4 = fadd float %add3, 1.012 %add5 = fadd float %add4, 1.013 %add6 = fadd float %add5, 1.014 %add7 = fadd float %add6, 1.015 %add8 = fadd float %add7, 1.016 %add9 = fadd float %add8, 1.017 %add10 = fadd float %add9, 1.018 store float %add10, ptr %p, align 419 ret void20}21 22define void @use_private_ptr_arg(ptr addrspace(5) nocapture %p) {23entry:24 %tmp1 = load float, ptr addrspace(5) %p, align 425 %div = fdiv float 1.000000e+00, %tmp126 %add0 = fadd float %div, 1.027 %add1 = fadd float %add0, 1.028 %add2 = fadd float %add1, 1.029 %add3 = fadd float %add2, 1.030 %add4 = fadd float %add3, 1.031 %add5 = fadd float %add4, 1.032 %add6 = fadd float %add5, 1.033 %add7 = fadd float %add6, 1.034 %add8 = fadd float %add7, 1.035 %add9 = fadd float %add8, 1.036 %add10 = fadd float %add9, 1.037 store float %add10, ptr addrspace(5) %p, align 438 ret void39}40 41; Test that the inline threshold is boosted if called with an42; addrspacecasted' alloca.43; CHECK-LABEL: @test_inliner_flat_ptr(44; CHECK: call i32 @llvm.amdgcn.workitem.id.x()45; CHECK-NOT: call {{[.*]}}@46define amdgpu_kernel void @test_inliner_flat_ptr(ptr addrspace(1) nocapture %a, i32 %n) {47entry:48 %pvt_arr = alloca [64 x float], align 4, addrspace(5)49 %tid = tail call i32 @llvm.amdgcn.workitem.id.x()50 %arrayidx = getelementptr inbounds float, ptr addrspace(1) %a, i32 %tid51 %tmp2 = load float, ptr addrspace(1) %arrayidx, align 452 %add = add i32 %tid, 153 %arrayidx2 = getelementptr inbounds float, ptr addrspace(1) %a, i32 %add54 %tmp5 = load float, ptr addrspace(1) %arrayidx2, align 455 %or = or i32 %tid, %n56 %arrayidx5 = getelementptr inbounds [64 x float], ptr addrspace(5) %pvt_arr, i32 0, i32 %or57 %arrayidx7 = getelementptr inbounds [64 x float], ptr addrspace(5) %pvt_arr, i32 0, i32 %or58 %to.flat = addrspacecast ptr addrspace(5) %arrayidx7 to ptr59 call void @use_private_ptr_arg(ptr addrspace(5) %arrayidx7)60 call void @use_flat_ptr_arg(ptr %to.flat)61 ret void62}63 64declare i32 @llvm.amdgcn.workitem.id.x() #165 66attributes #0 = { noinline }67attributes #1 = { nounwind readnone }68