1; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=SI %s
2; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=GFX89 -check-prefix=GFX8 %s
3; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=GFX89 -check-prefix=GFX9 %s
4
5declare half @llvm.copysign.f16(half, half)
6declare float @llvm.copysign.f32(float, float)
7declare double @llvm.copysign.f64(double, double)
8declare <2 x half> @llvm.copysign.v2f16(<2 x half>, <2 x half>)
9declare <3 x half> @llvm.copysign.v3f16(<3 x half>, <3 x half>)
10declare <4 x half> @llvm.copysign.v4f16(<4 x half>, <4 x half>)
11
12declare i32 @llvm.amdgcn.workitem.id.x()
13
14; GCN-LABEL: {{^}}test_copysign_f16:
15; SI: {{buffer|flat}}_load_ushort v[[MAG:[0-9]+]]
16; SI: {{buffer|flat}}_load_ushort v[[SIGN:[0-9]+]]
17; SI: s_brev_b32 s[[CONST:[0-9]+]], -2
18; SI-DAG: v_cvt_f32_f16_e32 v[[MAG_F32:[0-9]+]], v[[MAG]]
19; SI-DAG: v_cvt_f32_f16_e32 v[[SIGN_F32:[0-9]+]], v[[SIGN]]
20; SI: v_bfi_b32 v[[OUT_F32:[0-9]+]], s[[CONST]], v[[MAG_F32]], v[[SIGN_F32]]
21; SI: v_cvt_f16_f32_e32 v[[OUT:[0-9]+]], v[[OUT_F32]]
22; GFX89: {{buffer|flat}}_load_ushort v[[MAG:[0-9]+]]
23; GFX89: {{buffer|flat}}_load_ushort v[[SIGN:[0-9]+]]
24; GFX89: s_movk_i32 s[[CONST:[0-9]+]], 0x7fff
25; GFX89: v_bfi_b32 v[[OUT:[0-9]+]], s[[CONST]], v[[MAG]], v[[SIGN]]
26; GCN: buffer_store_short v[[OUT]]
27; GCN: s_endpgm
28define amdgpu_kernel void @test_copysign_f16(
29  half addrspace(1)* %arg_out,
30  half addrspace(1)* %arg_mag,
31  half addrspace(1)* %arg_sign) {
32entry:
33  %mag = load volatile half, half addrspace(1)* %arg_mag
34  %sign = load volatile half, half addrspace(1)* %arg_sign
35  %out = call half @llvm.copysign.f16(half %mag, half %sign)
36  store half %out, half addrspace(1)* %arg_out
37  ret void
38}
39
40; GCN-LABEL: {{^}}test_copysign_out_f32_mag_f16_sign_f32:
41; GCN-DAG: {{buffer|flat|global}}_load_ushort v[[MAG:[0-9]+]]
42; GCN-DAG: {{buffer|flat|global}}_load_dword v[[SIGN:[0-9]+]]
43; GCN-DAG: s_brev_b32 s[[CONST:[0-9]+]], -2
44; GCN-DAG: v_cvt_f32_f16_e32 v[[MAG_EXT:[0-9]+]], v[[MAG]]
45; GCN: v_bfi_b32 v[[OUT:[0-9]+]], s[[CONST]], v[[MAG_EXT]], v[[SIGN]]
46; GCN: buffer_store_dword v[[OUT]]
47; GCN: s_endpgm
48define amdgpu_kernel void @test_copysign_out_f32_mag_f16_sign_f32(
49  float addrspace(1)* %arg_out,
50  half addrspace(1)* %arg_mag,
51  float addrspace(1)* %arg_sign) {
52entry:
53  %tid = call i32 @llvm.amdgcn.workitem.id.x()
54  %arg_mag_gep = getelementptr half, half addrspace(1)* %arg_mag, i32 %tid
55  %mag = load half, half addrspace(1)* %arg_mag_gep
56  %mag.ext = fpext half %mag to float
57  %arg_sign_gep = getelementptr float, float addrspace(1)* %arg_sign, i32 %tid
58  %sign = load float, float addrspace(1)* %arg_sign_gep
59  %out = call float @llvm.copysign.f32(float %mag.ext, float %sign)
60  store float %out, float addrspace(1)* %arg_out
61  ret void
62}
63
64; GCN-LABEL: {{^}}test_copysign_out_f64_mag_f16_sign_f64:
65; GCN-DAG: {{buffer|flat|global}}_load_ushort v[[MAG:[0-9]+]]
66; GCN-DAG: {{buffer|flat|global}}_load_dwordx2 v{{\[}}[[SIGN_LO:[0-9]+]]:[[SIGN_HI:[0-9]+]]{{\]}}
67; GCN-DAG: s_brev_b32 s[[CONST:[0-9]+]], -2
68; GCN-DAG: v_cvt_f32_f16_e32 v[[MAG_EXT:[0-9]+]], v[[MAG]]
69; GCN-DAG: v_cvt_f64_f32_e32 v{{\[}}[[MAG_EXT_LO:[0-9]+]]:[[MAG_EXT_HI:[0-9]+]]{{\]}}, v[[MAG_EXT]]
70; GCN: v_bfi_b32 v[[OUT_HI:[0-9]+]], s[[CONST]], v[[MAG_EXT_HI]], v[[SIGN_HI]]
71; GCN: buffer_store_dwordx2 v{{\[}}[[MAG_EXT_LO]]:[[OUT_HI]]{{\]}}
72; GCN: s_endpgm
73define amdgpu_kernel void @test_copysign_out_f64_mag_f16_sign_f64(
74  double addrspace(1)* %arg_out,
75  half addrspace(1)* %arg_mag,
76  double addrspace(1)* %arg_sign) {
77entry:
78  %tid = call i32 @llvm.amdgcn.workitem.id.x()
79  %arg_mag_gep = getelementptr half, half addrspace(1)* %arg_mag, i32 %tid
80  %mag = load half, half addrspace(1)* %arg_mag_gep
81  %mag.ext = fpext half %mag to double
82  %arg_sign_gep = getelementptr double, double addrspace(1)* %arg_sign, i32 %tid
83  %sign = load double, double addrspace(1)* %arg_sign_gep
84  %out = call double @llvm.copysign.f64(double %mag.ext, double %sign)
85  store double %out, double addrspace(1)* %arg_out
86  ret void
87}
88
89; GCN-LABEL: {{^}}test_copysign_out_f32_mag_f32_sign_f16:
90; GCN-DAG: {{buffer|flat|global}}_load_dword v[[MAG:[0-9]+]]
91; GCN-DAG: {{buffer|flat|global}}_load_ushort v[[SIGN:[0-9]+]]
92; GCN-DAG: s_brev_b32 s[[CONST:[0-9]+]], -2
93; SI-DAG: v_cvt_f32_f16_e32 v[[SIGN_F32:[0-9]+]], v[[SIGN]]
94; SI: v_bfi_b32 v[[OUT:[0-9]+]], s[[CONST]], v[[MAG]], v[[SIGN_F32]]
95; GFX89-DAG: v_lshlrev_b32_e32 v[[SIGN_SHIFT:[0-9]+]], 16, v[[SIGN]]
96; GFX89: v_bfi_b32 v[[OUT:[0-9]+]], s[[CONST]], v[[MAG]], v[[SIGN_SHIFT]]
97; GCN: buffer_store_dword v[[OUT]]
98; GCN: s_endpgm
99define amdgpu_kernel void @test_copysign_out_f32_mag_f32_sign_f16(
100  float addrspace(1)* %arg_out,
101  float addrspace(1)* %arg_mag,
102  half addrspace(1)* %arg_sign) {
103entry:
104  %tid = call i32 @llvm.amdgcn.workitem.id.x()
105  %arg_mag_gep = getelementptr float, float addrspace(1)* %arg_mag, i32 %tid
106  %mag = load float, float addrspace(1)* %arg_mag_gep
107  %arg_sign_gep = getelementptr half, half addrspace(1)* %arg_sign, i32 %tid
108  %sign = load half, half addrspace(1)* %arg_sign_gep
109  %sign.ext = fpext half %sign to float
110  %out = call float @llvm.copysign.f32(float %mag, float %sign.ext)
111  store float %out, float addrspace(1)* %arg_out
112  ret void
113}
114
115; GCN-LABEL: {{^}}test_copysign_out_f64_mag_f64_sign_f16:
116; GCN-DAG: {{buffer|flat|global}}_load_dwordx2 v{{\[}}[[MAG_LO:[0-9]+]]:[[MAG_HI:[0-9]+]]{{\]}}
117; GCN-DAG: {{buffer|flat|global}}_load_ushort v[[SIGN:[0-9]+]]
118; GCN-DAG: s_brev_b32 s[[CONST:[0-9]+]], -2
119; SI-DAG: v_cvt_f32_f16_e32 v[[SIGN_F32:[0-9]+]], v[[SIGN]]
120; SI: v_bfi_b32 v[[OUT_HI:[0-9]+]], s[[CONST]], v[[MAG_HI]], v[[SIGN_F32]]
121; GFX89-DAG: v_lshlrev_b32_e32 v[[SIGN_SHIFT:[0-9]+]], 16, v[[SIGN]]
122; GFX89: v_bfi_b32 v[[OUT_HI:[0-9]+]], s[[CONST]], v[[MAG_HI]], v[[SIGN_SHIFT]]
123; GCN: buffer_store_dwordx2 v{{\[}}[[MAG_LO]]:[[OUT_HI]]{{\]}}
124; GCN: s_endpgm
125define amdgpu_kernel void @test_copysign_out_f64_mag_f64_sign_f16(
126  double addrspace(1)* %arg_out,
127  double addrspace(1)* %arg_mag,
128  half addrspace(1)* %arg_sign) {
129entry:
130  %tid = call i32 @llvm.amdgcn.workitem.id.x()
131  %arg_mag_gep = getelementptr double, double addrspace(1)* %arg_mag, i32 %tid
132  %mag = load double, double addrspace(1)* %arg_mag_gep
133  %arg_sign_gep = getelementptr half, half addrspace(1)* %arg_sign, i32 %tid
134  %sign = load half, half addrspace(1)* %arg_sign_gep
135  %sign.ext = fpext half %sign to double
136  %out = call double @llvm.copysign.f64(double %mag, double %sign.ext)
137  store double %out, double addrspace(1)* %arg_out
138  ret void
139}
140
141; GCN-LABEL: {{^}}test_copysign_out_f16_mag_f16_sign_f32:
142; GCN-DAG: {{buffer|flat|global}}_load_ushort v[[MAG:[0-9]+]]
143; GCN-DAG: {{buffer|flat|global}}_load_dword v[[SIGN:[0-9]+]]
144; SI-DAG: s_brev_b32 s[[CONST:[0-9]+]], -2
145; SI-DAG: v_cvt_f32_f16_e32 v[[MAG_F32:[0-9]+]], v[[MAG]]
146; SI: v_bfi_b32 v[[OUT_F32:[0-9]+]], s[[CONST]], v[[MAG_F32]], v[[SIGN]]
147; SI: v_cvt_f16_f32_e32 v[[OUT:[0-9]+]], v[[OUT_F32]]
148; GFX89-DAG: s_movk_i32 s[[CONST:[0-9]+]], 0x7fff
149; GFX89-DAG: v_lshrrev_b32_e32 v[[SIGN_SHIFT:[0-9]+]], 16, v[[SIGN]]
150; GFX89: v_bfi_b32 v[[OUT:[0-9]+]], s[[CONST]], v[[MAG]], v[[SIGN_SHIFT]]
151; GCN: buffer_store_short v[[OUT]]
152; GCN: s_endpgm
153define amdgpu_kernel void @test_copysign_out_f16_mag_f16_sign_f32(
154  half addrspace(1)* %arg_out,
155  half addrspace(1)* %arg_mag,
156  float addrspace(1)* %arg_sign) {
157entry:
158  %tid = call i32 @llvm.amdgcn.workitem.id.x()
159  %arg_mag_gep = getelementptr half, half addrspace(1)* %arg_mag, i32 %tid
160  %mag = load half, half addrspace(1)* %arg_mag_gep
161  %arg_sign_gep = getelementptr float, float addrspace(1)* %arg_sign, i32 %tid
162  %sign = load float, float addrspace(1)* %arg_sign_gep
163  %sign.trunc = fptrunc float %sign to half
164  %out = call half @llvm.copysign.f16(half %mag, half %sign.trunc)
165  store half %out, half addrspace(1)* %arg_out
166  ret void
167}
168
169; GCN-LABEL: {{^}}test_copysign_out_f16_mag_f16_sign_f64:
170; GCN-DAG: {{buffer|flat|global}}_load_ushort v[[MAG:[0-9]+]]
171; GCN-DAG: {{buffer|flat|global}}_load_dwordx2 v{{\[}}[[SIGN_LO:[0-9]+]]:[[SIGN_HI:[0-9]+]]{{\]}}
172; SI-DAG: s_brev_b32 s[[CONST:[0-9]+]], -2
173; SI-DAG: v_cvt_f32_f16_e32 v[[MAG_F32:[0-9]+]], v[[MAG]]
174; SI: v_bfi_b32 v[[OUT_F32:[0-9]+]], s[[CONST]], v[[MAG_F32]], v[[SIGN_HI]]
175; SI: v_cvt_f16_f32_e32 v[[OUT:[0-9]+]], v[[OUT_F32]]
176; GFX89-DAG: s_movk_i32 s[[CONST:[0-9]+]], 0x7fff
177; GFX89-DAG: v_lshrrev_b32_e32 v[[SIGN_SHIFT:[0-9]+]], 16, v[[SIGN_HI]]
178; GFX89: v_bfi_b32 v[[OUT:[0-9]+]], s[[CONST]], v[[MAG]], v[[SIGN_SHIFT]]
179; GCN: buffer_store_short v[[OUT]]
180; GCN: s_endpgm
181define amdgpu_kernel void @test_copysign_out_f16_mag_f16_sign_f64(
182  half addrspace(1)* %arg_out,
183  half addrspace(1)* %arg_mag,
184  double addrspace(1)* %arg_sign) {
185entry:
186  %tid = call i32 @llvm.amdgcn.workitem.id.x()
187  %arg_mag_gep = getelementptr half, half addrspace(1)* %arg_mag, i32 %tid
188  %mag = load half, half addrspace(1)* %arg_mag
189  %arg_sign_gep = getelementptr double, double addrspace(1)* %arg_sign, i32 %tid
190  %sign = load double, double addrspace(1)* %arg_sign_gep
191  %sign.trunc = fptrunc double %sign to half
192  %out = call half @llvm.copysign.f16(half %mag, half %sign.trunc)
193  store half %out, half addrspace(1)* %arg_out
194  ret void
195}
196
197; GCN-LABEL: {{^}}test_copysign_out_f16_mag_f32_sign_f16:
198; GCN-DAG: {{buffer|flat|global}}_load_dword v[[MAG:[0-9]+]]
199; GCN-DAG: {{buffer|flat|global}}_load_ushort v[[SIGN:[0-9]+]]
200; SI-DAG: s_brev_b32 s[[CONST:[0-9]+]], -2
201; SI-DAG: v_cvt_f16_f32_e32 v[[MAG_TRUNC:[0-9]+]], v[[MAG]]
202; SI-DAG: v_cvt_f32_f16_e32 v[[SIGN_F32:[0-9]+]], v[[SIGN]]
203; SI-DAG: v_cvt_f32_f16_e32 v[[MAG_F32:[0-9]+]], v[[MAG_TRUNC]]
204; SI: v_bfi_b32 v[[OUT_F32:[0-9]+]], s[[CONST]], v[[MAG_F32]], v[[SIGN_F32]]
205; SI: v_cvt_f16_f32_e32 v[[OUT:[0-9]+]], v[[OUT_F32]]
206; GFX89-DAG: s_movk_i32 s[[CONST:[0-9]+]], 0x7fff
207; GFX89-DAG: v_cvt_f16_f32_e32 v[[MAG_TRUNC:[0-9]+]], v[[MAG]]
208; GFX89: v_bfi_b32 v[[OUT:[0-9]+]], s[[CONST]], v[[MAG_TRUNC]], v[[SIGN]]
209; GCN: buffer_store_short v[[OUT]]
210; GCN: s_endpgm
211define amdgpu_kernel void @test_copysign_out_f16_mag_f32_sign_f16(
212  half addrspace(1)* %arg_out,
213  float addrspace(1)* %arg_mag,
214  half addrspace(1)* %arg_sign) {
215entry:
216  %tid = call i32 @llvm.amdgcn.workitem.id.x()
217  %arg_mag_gep = getelementptr float, float addrspace(1)* %arg_mag, i32 %tid
218  %mag = load float, float addrspace(1)* %arg_mag_gep
219  %mag.trunc = fptrunc float %mag to half
220  %arg_sign_gep = getelementptr half, half addrspace(1)* %arg_sign, i32 %tid
221  %sign = load half, half addrspace(1)* %arg_sign_gep
222  %out = call half @llvm.copysign.f16(half %mag.trunc, half %sign)
223  store half %out, half addrspace(1)* %arg_out
224  ret void
225}
226
227; GCN-LABEL: {{^}}test_copysign_out_f16_mag_f64_sign_f16:
228; GCN: v_bfi_b32
229; GCN: s_endpgm
230define amdgpu_kernel void @test_copysign_out_f16_mag_f64_sign_f16(
231  half addrspace(1)* %arg_out,
232  double addrspace(1)* %arg_mag,
233  half addrspace(1)* %arg_sign) {
234entry:
235  %mag = load double, double addrspace(1)* %arg_mag
236  %mag.trunc = fptrunc double %mag to half
237  %sign = load half, half addrspace(1)* %arg_sign
238  %out = call half @llvm.copysign.f16(half %mag.trunc, half %sign)
239  store half %out, half addrspace(1)* %arg_out
240  ret void
241}
242
243; GCN-LABEL: {{^}}test_copysign_v2f16:
244; GCN: v_bfi_b32
245; GCN: v_bfi_b32
246; VI: v_or_b32_sdwa v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
247; GCN: s_endpgm
248define amdgpu_kernel void @test_copysign_v2f16(
249  <2 x half> addrspace(1)* %arg_out,
250  <2 x half> %arg_mag,
251  <2 x half> %arg_sign) {
252entry:
253  %out = call <2 x half> @llvm.copysign.v2f16(<2 x half> %arg_mag, <2 x half> %arg_sign)
254  store <2 x half> %out, <2 x half> addrspace(1)* %arg_out
255  ret void
256}
257
258; GCN-LABEL: {{^}}test_copysign_v3f16:
259; GCN: v_bfi_b32
260; GCN: v_bfi_b32
261; GCN: v_bfi_b32
262; GCN: s_endpgm
263define amdgpu_kernel void @test_copysign_v3f16(
264  <3 x half> addrspace(1)* %arg_out,
265  <3 x half> %arg_mag,
266  <3 x half> %arg_sign) {
267entry:
268  %out = call <3 x half> @llvm.copysign.v3f16(<3 x half> %arg_mag, <3 x half> %arg_sign)
269  store <3 x half> %out, <3 x half> addrspace(1)* %arg_out
270  ret void
271}
272
273; GCN-LABEL: {{^}}test_copysign_v4f16:
274; GCN: v_bfi_b32
275; GCN: v_bfi_b32
276; GCN: v_bfi_b32
277; GCN: v_bfi_b32
278; GCN: s_endpgm
279define amdgpu_kernel void @test_copysign_v4f16(
280  <4 x half> addrspace(1)* %arg_out,
281  <4 x half> %arg_mag,
282  <4 x half> %arg_sign) {
283entry:
284  %out = call <4 x half> @llvm.copysign.v4f16(<4 x half> %arg_mag, <4 x half> %arg_sign)
285  store <4 x half> %out, <4 x half> addrspace(1)* %arg_out
286  ret void
287}
288