1 ; RUN: opt -mtriple=amdgcn--amdhsa -data-layout=A5 -O3 -S -inline-threshold=1 < %s | FileCheck -check-prefixes=GCN,GCN-INL1,GCN-MAXBBDEF %s
2 ; RUN: opt -mtriple=amdgcn--amdhsa -data-layout=A5 -O3 -S < %s | FileCheck -check-prefixes=GCN,GCN-INLDEF,GCN-MAXBBDEF %s
3 ; RUN: opt -mtriple=amdgcn--amdhsa -data-layout=A5 -passes='default<O3>' -S -inline-threshold=1 < %s | FileCheck -check-prefixes=GCN,GCN-INL1,GCN-MAXBBDEF %s
4 ; RUN: opt -mtriple=amdgcn--amdhsa -data-layout=A5 -passes='default<O3>' -S < %s | FileCheck -check-prefixes=GCN,GCN-INLDEF,GCN-MAXBBDEF %s
5 ; RUN: opt -mtriple=amdgcn--amdhsa -data-layout=A5 -passes='default<O3>' -S -amdgpu-inline-max-bb=1 < %s | FileCheck -check-prefixes=GCN,GCN-MAXBB1 %s
7 define coldcc float @foo(float %x, float %y) {
9 %cmp = fcmp ogt float %x, 0.000000e+00
10 %div = fdiv float %y, %x
11 %mul = fmul float %x, %y
12 %cond = select i1 %cmp, float %div, float %mul
16 define coldcc void @foo_private_ptr(ptr addrspace(5) nocapture %p) {
18 %tmp1 = load float, ptr addrspace(5) %p, align 4
19 %cmp = fcmp ogt float %tmp1, 1.000000e+00
20 br i1 %cmp, label %if.then, label %if.end
22 if.then: ; preds = %entry
23 %div = fdiv float 1.000000e+00, %tmp1
24 store float %div, ptr addrspace(5) %p, align 4
27 if.end: ; preds = %if.then, %entry
31 define coldcc void @foo_private_ptr2(ptr addrspace(5) nocapture %p1, ptr addrspace(5) nocapture %p2) {
33 call void @forbid_sroa(ptr addrspace(5) %p1)
34 call void @forbid_sroa(ptr addrspace(5) %p2)
35 %tmp1 = load float, ptr addrspace(5) %p1, align 4
36 %cmp = fcmp ogt float %tmp1, 1.000000e+00
37 br i1 %cmp, label %if.then, label %if.end
40 %div = fdiv float 2.000000e+00, %tmp1
41 store float %div, ptr addrspace(5) %p2, align 4
48 define float @sin_wrapper(float %x) {
50 %call = tail call float @_Z3sinf(float %x)
54 define void @foo_noinline(ptr addrspace(5) nocapture %p) #0 {
56 %tmp1 = load float, ptr addrspace(5) %p, align 4
57 %mul = fmul float %tmp1, 2.000000e+00
58 store float %mul, ptr addrspace(5) %p, align 4
62 ; GCN: define amdgpu_kernel void @test_inliner(
63 ; GCN-INL1: %c1 = tail call coldcc float @foo(
64 ; GCN-INLDEF: %cmp.i = fcmp ogt float %tmp2, 0.000000e+00
65 ; GCN-MAXBBDEF: %div.i{{[0-9]*}} = fdiv float 1.000000e+00, %c
66 ; GCN-MAXBBDEF: %div.i{{[0-9]*}} = fdiv float 2.000000e+00, %tmp1.i
67 ; GCN-MAXBB1: call coldcc void @foo_private_ptr
68 ; GCN-MAXBB1: call coldcc void @foo_private_ptr2
69 ; GCN: call void @foo_noinline(
70 ; GCN: tail call float @_Z3sinf(
71 define amdgpu_kernel void @test_inliner(ptr addrspace(1) nocapture %a, i32 %n) {
73 %pvt_arr = alloca [64 x float], align 4, addrspace(5)
74 %tid = tail call i32 @llvm.amdgcn.workitem.id.x()
75 %arrayidx = getelementptr inbounds float, ptr addrspace(1) %a, i32 %tid
76 %tmp2 = load float, ptr addrspace(1) %arrayidx, align 4
77 %add = add i32 %tid, 1
78 %arrayidx2 = getelementptr inbounds float, ptr addrspace(1) %a, i32 %add
79 %tmp5 = load float, ptr addrspace(1) %arrayidx2, align 4
80 %c1 = tail call coldcc float @foo(float %tmp2, float %tmp5)
82 %arrayidx5 = getelementptr inbounds [64 x float], ptr addrspace(5) %pvt_arr, i32 0, i32 %or
83 store float %c1, ptr addrspace(5) %arrayidx5, align 4
84 %arrayidx7 = getelementptr inbounds [64 x float], ptr addrspace(5) %pvt_arr, i32 0, i32 %or
85 call coldcc void @foo_private_ptr(ptr addrspace(5) %arrayidx7)
86 %arrayidx8 = getelementptr inbounds [64 x float], ptr addrspace(5) %pvt_arr, i32 0, i32 1
87 %arrayidx9 = getelementptr inbounds [64 x float], ptr addrspace(5) %pvt_arr, i32 0, i32 2
88 call coldcc void @foo_private_ptr2(ptr addrspace(5) %arrayidx8, ptr addrspace(5) %arrayidx9)
89 call void @foo_noinline(ptr addrspace(5) %arrayidx7)
90 %and = and i32 %tid, %n
91 %arrayidx11 = getelementptr inbounds [64 x float], ptr addrspace(5) %pvt_arr, i32 0, i32 %and
92 %tmp12 = load float, ptr addrspace(5) %arrayidx11, align 4
93 %c2 = call float @sin_wrapper(float %tmp12)
94 store float %c2, ptr addrspace(5) %arrayidx7, align 4
95 %xor = xor i32 %tid, %n
96 %arrayidx16 = getelementptr inbounds [64 x float], ptr addrspace(5) %pvt_arr, i32 0, i32 %xor
97 %tmp16 = load float, ptr addrspace(5) %arrayidx16, align 4
98 store float %tmp16, ptr addrspace(1) %arrayidx, align 4
102 ; GCN: define amdgpu_kernel void @test_inliner_multi_pvt_ptr(
103 ; GCN-MAXBBDEF: %div.i{{[0-9]*}} = fdiv float 2.000000e+00, %tmp1.i
104 ; GCN-MAXBB1: call coldcc void @foo_private_ptr2
105 define amdgpu_kernel void @test_inliner_multi_pvt_ptr(ptr addrspace(1) nocapture %a, i32 %n, float %v) {
107 %pvt_arr1 = alloca [32 x float], align 4, addrspace(5)
108 %pvt_arr2 = alloca [32 x float], align 4, addrspace(5)
109 %tid = tail call i32 @llvm.amdgcn.workitem.id.x()
110 %arrayidx = getelementptr inbounds float, ptr addrspace(1) %a, i32 %tid
111 %or = or i32 %tid, %n
112 %arrayidx4 = getelementptr inbounds [32 x float], ptr addrspace(5) %pvt_arr1, i32 0, i32 %or
113 %arrayidx5 = getelementptr inbounds [32 x float], ptr addrspace(5) %pvt_arr2, i32 0, i32 %or
114 store float %v, ptr addrspace(5) %arrayidx4, align 4
115 store float %v, ptr addrspace(5) %arrayidx5, align 4
116 %arrayidx8 = getelementptr inbounds [32 x float], ptr addrspace(5) %pvt_arr1, i32 0, i32 1
117 %arrayidx9 = getelementptr inbounds [32 x float], ptr addrspace(5) %pvt_arr2, i32 0, i32 2
118 call coldcc void @foo_private_ptr2(ptr addrspace(5) %arrayidx8, ptr addrspace(5) %arrayidx9)
119 %xor = xor i32 %tid, %n
120 %arrayidx15 = getelementptr inbounds [32 x float], ptr addrspace(5) %pvt_arr1, i32 0, i32 %xor
121 %arrayidx16 = getelementptr inbounds [32 x float], ptr addrspace(5) %pvt_arr2, i32 0, i32 %xor
122 %tmp15 = load float, ptr addrspace(5) %arrayidx15, align 4
123 %tmp16 = load float, ptr addrspace(5) %arrayidx16, align 4
124 %tmp17 = fadd float %tmp15, %tmp16
125 store float %tmp17, ptr addrspace(1) %arrayidx, align 4
129 ; GCN: define amdgpu_kernel void @test_inliner_multi_pvt_ptr_cutoff(
130 ; GCN-INL1: call coldcc void @foo_private_ptr2
131 ; GCN-INLDEF: %div.i{{[0-9]*}} = fdiv float 2.000000e+00, %tmp1.i
132 define amdgpu_kernel void @test_inliner_multi_pvt_ptr_cutoff(ptr addrspace(1) nocapture %a, i32 %n, float %v) {
134 %pvt_arr1 = alloca [32 x float], align 4, addrspace(5)
135 %pvt_arr2 = alloca [33 x float], align 4, addrspace(5)
136 %tid = tail call i32 @llvm.amdgcn.workitem.id.x()
137 %arrayidx = getelementptr inbounds float, ptr addrspace(1) %a, i32 %tid
138 %or = or i32 %tid, %n
139 %arrayidx4 = getelementptr inbounds [32 x float], ptr addrspace(5) %pvt_arr1, i32 0, i32 %or
140 %arrayidx5 = getelementptr inbounds [33 x float], ptr addrspace(5) %pvt_arr2, i32 0, i32 %or
141 store float %v, ptr addrspace(5) %arrayidx4, align 4
142 store float %v, ptr addrspace(5) %arrayidx5, align 4
143 %arrayidx8 = getelementptr inbounds [32 x float], ptr addrspace(5) %pvt_arr1, i32 0, i32 1
144 %arrayidx9 = getelementptr inbounds [33 x float], ptr addrspace(5) %pvt_arr2, i32 0, i32 2
145 call coldcc void @foo_private_ptr2(ptr addrspace(5) %arrayidx8, ptr addrspace(5) %arrayidx9)
146 %xor = xor i32 %tid, %n
147 %arrayidx15 = getelementptr inbounds [32 x float], ptr addrspace(5) %pvt_arr1, i32 0, i32 %xor
148 %arrayidx16 = getelementptr inbounds [33 x float], ptr addrspace(5) %pvt_arr2, i32 0, i32 %xor
149 %tmp15 = load float, ptr addrspace(5) %arrayidx15, align 4
150 %tmp16 = load float, ptr addrspace(5) %arrayidx16, align 4
151 %tmp17 = fadd float %tmp15, %tmp16
152 store float %tmp17, ptr addrspace(1) %arrayidx, align 4
156 ; GCN: define amdgpu_kernel void @test_inliner_maxbb_singlebb(
157 ; GCN: tail call float @_Z3sinf
158 define amdgpu_kernel void @test_inliner_maxbb_singlebb(ptr addrspace(1) nocapture %a, i32 %n) {
160 %cmp = icmp eq i32 %n, 1
161 br i1 %cmp, label %bb.1, label %bb.2
165 store float 1.0, ptr undef
169 %c = call float @sin_wrapper(float 1.0)
170 store float %c, ptr addrspace(1) %a
174 declare i32 @llvm.amdgcn.workitem.id.x() #1
175 declare float @_Z3sinf(float) #1
176 declare void @forbid_sroa(ptr addrspace(5) nocapture %p)
178 attributes #0 = { noinline }
179 attributes #1 = { nounwind readnone }