1; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
2
3declare half @llvm.canonicalize.f16(half) #0
4
5; GCN-LABEL: {{^}}v_test_canonicalize_var_f16:
6; GCN: v_mul_f16_e32 [[REG:v[0-9]+]], 1.0, {{v[0-9]+}}
7; GCN: buffer_store_short [[REG]]
8define void @v_test_canonicalize_var_f16(half addrspace(1)* %out) #1 {
9  %val = load half, half addrspace(1)* %out
10  %canonicalized = call half @llvm.canonicalize.f16(half %val)
11  store half %canonicalized, half addrspace(1)* %out
12  ret void
13}
14
15; GCN-LABEL: {{^}}s_test_canonicalize_var_f16:
16; GCN: v_mul_f16_e64 [[REG:v[0-9]+]], 1.0, {{s[0-9]+}}
17; GCN: buffer_store_short [[REG]]
18define void @s_test_canonicalize_var_f16(half addrspace(1)* %out, i16 zeroext %val.arg) #1 {
19  %val = bitcast i16 %val.arg to half
20  %canonicalized = call half @llvm.canonicalize.f16(half %val)
21  store half %canonicalized, half addrspace(1)* %out
22  ret void
23}
24
25; GCN-LABEL: {{^}}test_fold_canonicalize_p0_f16:
26; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}}
27; GCN: buffer_store_short [[REG]]
28define void @test_fold_canonicalize_p0_f16(half addrspace(1)* %out) #1 {
29  %canonicalized = call half @llvm.canonicalize.f16(half 0.0)
30  store half %canonicalized, half addrspace(1)* %out
31  ret void
32}
33
34; GCN-LABEL: {{^}}test_fold_canonicalize_n0_f16:
35; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffff8000{{$}}
36; GCN: buffer_store_short [[REG]]
37define void @test_fold_canonicalize_n0_f16(half addrspace(1)* %out) #1 {
38  %canonicalized = call half @llvm.canonicalize.f16(half -0.0)
39  store half %canonicalized, half addrspace(1)* %out
40  ret void
41}
42
43; GCN-LABEL: {{^}}test_fold_canonicalize_p1_f16:
44; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3c00{{$}}
45; GCN: buffer_store_short [[REG]]
46define void @test_fold_canonicalize_p1_f16(half addrspace(1)* %out) #1 {
47  %canonicalized = call half @llvm.canonicalize.f16(half 1.0)
48  store half %canonicalized, half addrspace(1)* %out
49  ret void
50}
51
52; GCN-LABEL: {{^}}test_fold_canonicalize_n1_f16:
53; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffffbc00{{$}}
54; GCN: buffer_store_short [[REG]]
55define void @test_fold_canonicalize_n1_f16(half addrspace(1)* %out) #1 {
56  %canonicalized = call half @llvm.canonicalize.f16(half -1.0)
57  store half %canonicalized, half addrspace(1)* %out
58  ret void
59}
60
61; GCN-LABEL: {{^}}test_fold_canonicalize_literal_f16:
62; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x4c00{{$}}
63; GCN: buffer_store_short [[REG]]
64define void @test_fold_canonicalize_literal_f16(half addrspace(1)* %out) #1 {
65  %canonicalized = call half @llvm.canonicalize.f16(half 16.0)
66  store half %canonicalized, half addrspace(1)* %out
67  ret void
68}
69
70; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal0_f16:
71; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}}
72; GCN: buffer_store_short [[REG]]
73define void @test_no_denormals_fold_canonicalize_denormal0_f16(half addrspace(1)* %out) #1 {
74  %canonicalized = call half @llvm.canonicalize.f16(half 0xH03FF)
75  store half %canonicalized, half addrspace(1)* %out
76  ret void
77}
78
79; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal0_f16:
80; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3ff{{$}}
81; GCN: buffer_store_short [[REG]]
82define void @test_denormals_fold_canonicalize_denormal0_f16(half addrspace(1)* %out) #3 {
83  %canonicalized = call half @llvm.canonicalize.f16(half 0xH03FF)
84  store half %canonicalized, half addrspace(1)* %out
85  ret void
86}
87
88; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal1_f16:
89; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}}
90; GCN: buffer_store_short [[REG]]
91define void @test_no_denormals_fold_canonicalize_denormal1_f16(half addrspace(1)* %out) #1 {
92  %canonicalized = call half @llvm.canonicalize.f16(half 0xH83FF)
93  store half %canonicalized, half addrspace(1)* %out
94  ret void
95}
96
97; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal1_f16:
98; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffff83ff{{$}}
99; GCN: buffer_store_short [[REG]]
100define void @test_denormals_fold_canonicalize_denormal1_f16(half addrspace(1)* %out) #3 {
101  %canonicalized = call half @llvm.canonicalize.f16(half 0xH83FF)
102  store half %canonicalized, half addrspace(1)* %out
103  ret void
104}
105
106; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_f16:
107; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7c00{{$}}
108; GCN: buffer_store_short [[REG]]
109define void @test_fold_canonicalize_qnan_f16(half addrspace(1)* %out) #1 {
110  %canonicalized = call half @llvm.canonicalize.f16(half 0xH7C00)
111  store half %canonicalized, half addrspace(1)* %out
112  ret void
113}
114
115; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg1_f16:
116; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}}
117; GCN: buffer_store_short [[REG]]
118define void @test_fold_canonicalize_qnan_value_neg1_f16(half addrspace(1)* %out) #1 {
119  %canonicalized = call half @llvm.canonicalize.f16(half bitcast (i16 -1 to half))
120  store half %canonicalized, half addrspace(1)* %out
121  ret void
122}
123
124; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg2_f16:
125; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}}
126; GCN: buffer_store_short [[REG]]
127define void @test_fold_canonicalize_qnan_value_neg2_f16(half addrspace(1)* %out) #1 {
128  %canonicalized = call half @llvm.canonicalize.f16(half bitcast (i16 -2 to half))
129  store half %canonicalized, half addrspace(1)* %out
130  ret void
131}
132
133; GCN-LABEL: {{^}}test_fold_canonicalize_snan0_value_f16:
134; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}}
135; GCN: buffer_store_short [[REG]]
136define void @test_fold_canonicalize_snan0_value_f16(half addrspace(1)* %out) #1 {
137  %canonicalized = call half @llvm.canonicalize.f16(half 0xH7C01)
138  store half %canonicalized, half addrspace(1)* %out
139  ret void
140}
141
142; GCN-LABEL: {{^}}test_fold_canonicalize_snan1_value_f16:
143; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}}
144; GCN: buffer_store_short [[REG]]
145define void @test_fold_canonicalize_snan1_value_f16(half addrspace(1)* %out) #1 {
146  %canonicalized = call half @llvm.canonicalize.f16(half 0xH7DFF)
147  store half %canonicalized, half addrspace(1)* %out
148  ret void
149}
150
151; GCN-LABEL: {{^}}test_fold_canonicalize_snan2_value_f16:
152; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}}
153; GCN: buffer_store_short [[REG]]
154define void @test_fold_canonicalize_snan2_value_f16(half addrspace(1)* %out) #1 {
155  %canonicalized = call half @llvm.canonicalize.f16(half 0xHFDFF)
156  store half %canonicalized, half addrspace(1)* %out
157  ret void
158}
159
160; GCN-LABEL: {{^}}test_fold_canonicalize_snan3_value_f16:
161; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}}
162; GCN: buffer_store_short [[REG]]
163define void @test_fold_canonicalize_snan3_value_f16(half addrspace(1)* %out) #1 {
164  %canonicalized = call half @llvm.canonicalize.f16(half 0xHFC01)
165  store half %canonicalized, half addrspace(1)* %out
166  ret void
167}
168
169attributes #0 = { nounwind readnone }
170attributes #1 = { nounwind }
171attributes #2 = { nounwind "target-features"="-fp16-denormals,-fp16-denormals" }
172attributes #3 = { nounwind "target-features"="+fp16-denormals,+fp64-denormals" }
173