78 lines · plain
1; RUN: opt -S -mtriple=amdgcn-unknown-amdhsa -mcpu=hawaii -passes=loop-unroll -unroll-threshold=100 -unroll-peel-count=0 -unroll-allow-partial=false -unroll-max-iteration-count-to-analyze=16 < %s | FileCheck %s2 3; CHECK-LABEL: @test_intrinsic_call_cost(4; CHECK-NOT: br i15define amdgpu_kernel void @test_intrinsic_call_cost(ptr addrspace(1) noalias nocapture %out, ptr addrspace(1) noalias nocapture %in) #0 {6entry:7 br label %for.body8 9for.body:10 %indvars.iv = phi i32 [ %indvars.iv.next, %for.body ], [ 0, %entry ]11 %sum.02 = phi float [ %fmul, %for.body ], [ 0.0, %entry ]12 %arrayidx.in = getelementptr inbounds float, ptr addrspace(1) %in, i32 %indvars.iv13 %arrayidx.out = getelementptr inbounds float, ptr addrspace(1) %out, i32 %indvars.iv14 %load = load float, ptr addrspace(1) %arrayidx.in15 %call = call float @llvm.minnum.f32(float %load, float 1.0);16 %fmul = fmul float %call, %sum.0217 store float %fmul, ptr addrspace(1) %arrayidx.out18 %indvars.iv.next = add i32 %indvars.iv, 119 %exitcond = icmp eq i32 %indvars.iv.next, 1620 br i1 %exitcond, label %for.end, label %for.body21 22for.end:23 ret void24}25 26; CHECK-LABEL: @test_func_call_cost(27; CHECK: br i1 %exitcond28define amdgpu_kernel void @test_func_call_cost(ptr addrspace(1) noalias nocapture %out, ptr addrspace(1) noalias nocapture %in) #0 {29entry:30 br label %for.body31 32for.body:33 %indvars.iv = phi i32 [ %indvars.iv.next, %for.body ], [ 0, %entry ]34 %sum.02 = phi float [ %fmul, %for.body ], [ 0.0, %entry ]35 %arrayidx.in = getelementptr inbounds float, ptr addrspace(1) %in, i32 %indvars.iv36 %arrayidx.out = getelementptr inbounds float, ptr addrspace(1) %out, i32 %indvars.iv37 %load = load float, ptr addrspace(1) %arrayidx.in38 %fptr = load ptr, ptr addrspace(4) null39 %call = tail call float %fptr(float %load, float 1.0)40 %fmul = fmul float %call, %sum.0241 store float %fmul, ptr addrspace(1) %arrayidx.out42 %indvars.iv.next = add i32 %indvars.iv, 143 %exitcond = icmp eq i32 %indvars.iv.next, 1644 br i1 %exitcond, label %for.end, label %for.body45 46for.end:47 ret void48}49 50; CHECK-LABEL: @test_indirect_call_cost(51; CHECK: br i1 %exitcond52define amdgpu_kernel void @test_indirect_call_cost(ptr addrspace(1) noalias nocapture %out, ptr addrspace(1) noalias nocapture %in) #0 {53entry:54 br label %for.body55 56for.body:57 %indvars.iv = phi i32 [ %indvars.iv.next, %for.body ], [ 0, %entry ]58 %sum.02 = phi float [ %fmul, %for.body ], [ 0.0, %entry ]59 %arrayidx.in = getelementptr inbounds float, ptr addrspace(1) %in, i32 %indvars.iv60 %arrayidx.out = getelementptr inbounds float, ptr addrspace(1) %out, i32 %indvars.iv61 %load = load float, ptr addrspace(1) %arrayidx.in62 %min = call float @func(float %load, float 1.0);63 %fmul = fmul float %min, %sum.0264 store float %fmul, ptr addrspace(1) %arrayidx.out65 %indvars.iv.next = add i32 %indvars.iv, 166 %exitcond = icmp eq i32 %indvars.iv.next, 1667 br i1 %exitcond, label %for.end, label %for.body68 69for.end:70 ret void71}72 73declare float @llvm.minnum.f32(float, float) #174declare float @func(float, float) #175 76attributes #0 = { nounwind }77attributes #1 = { nounwind readnone speculatable }78