• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,VI,GFX89 %s
2; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,GFX9,GFX89 %s
3; RUN: llc -march=amdgcn -mcpu=kaveri -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,CI %s
4
5declare half @llvm.fabs.f16(half) #0
6declare half @llvm.canonicalize.f16(half) #0
7declare <2 x half> @llvm.fabs.v2f16(<2 x half>) #0
8declare <2 x half> @llvm.canonicalize.v2f16(<2 x half>) #0
9declare <3 x half> @llvm.canonicalize.v3f16(<3 x half>) #0
10declare <4 x half> @llvm.canonicalize.v4f16(<4 x half>) #0
11declare i32 @llvm.amdgcn.workitem.id.x() #0
12
13; GCN-LABEL: {{^}}test_fold_canonicalize_undef_value_f16:
14; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}}
15; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
16define amdgpu_kernel void @test_fold_canonicalize_undef_value_f16(half addrspace(1)* %out) #1 {
17  %canonicalized = call half @llvm.canonicalize.f16(half undef)
18  store half %canonicalized, half addrspace(1)* %out
19  ret void
20}
21
22; GCN-LABEL: {{^}}v_test_canonicalize_var_f16:
23; GFX89: v_max_f16_e32 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
24; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
25
26; CI: v_cvt_f32_f16_e32
27; CI: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}}
28define amdgpu_kernel void @v_test_canonicalize_var_f16(half addrspace(1)* %out) #1 {
29  %val = load half, half addrspace(1)* %out
30  %canonicalized = call half @llvm.canonicalize.f16(half %val)
31  store half %canonicalized, half addrspace(1)* undef
32  ret void
33}
34
35; GCN-LABEL: {{^}}s_test_canonicalize_var_f16:
36; GFX89: v_max_f16_e64 [[REG:v[0-9]+]], {{s[0-9]+}}, {{s[0-9]+}}
37; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
38define amdgpu_kernel void @s_test_canonicalize_var_f16(half addrspace(1)* %out, i16 zeroext %val.arg) #1 {
39  %val = bitcast i16 %val.arg to half
40  %canonicalized = call half @llvm.canonicalize.f16(half %val)
41  store half %canonicalized, half addrspace(1)* %out
42  ret void
43}
44
45; GCN-LABEL: {{^}}v_test_canonicalize_build_vector_v2f16:
46; GFX9: v_and_b32_e32 v0, 0xffff, v0
47; GFX9-NEXT: v_lshl_or_b32 v0, v1, 16, v0
48; GFX9-NEXT: v_pk_max_f16 v0, v0, v0
49
50; VI: v_max_f16_sdwa v1, v1, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
51; VI: v_max_f16_e32 v0, v0, v0
52; VI: v_or_b32_e32 v0, v0, v1
53define <2 x half> @v_test_canonicalize_build_vector_v2f16(half %lo, half %hi) #1 {
54  %ins0 = insertelement <2 x half> undef, half %lo, i32 0
55  %ins1 = insertelement <2 x half> %ins0, half %hi, i32 1
56  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %ins1)
57  ret <2 x half> %canonicalized
58}
59
60; GCN-LABEL: {{^}}v_test_canonicalize_fabs_var_f16:
61; GFX89: v_max_f16_e64 [[REG:v[0-9]+]], |{{v[0-9]+}}|, |{{v[0-9]+}}|
62; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
63define amdgpu_kernel void @v_test_canonicalize_fabs_var_f16(half addrspace(1)* %out) #1 {
64  %val = load half, half addrspace(1)* %out
65  %val.fabs = call half @llvm.fabs.f16(half %val)
66  %canonicalized = call half @llvm.canonicalize.f16(half %val.fabs)
67  store half %canonicalized, half addrspace(1)* %out
68  ret void
69}
70
71; GCN-LABEL: {{^}}v_test_canonicalize_fneg_fabs_var_f16:
72; GFX89: v_max_f16_e64 [[REG:v[0-9]+]], -|{{v[0-9]+}}|, -|{{v[0-9]+}}|
73; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
74
75; CI: v_cvt_f32_f16_e64 v{{[0-9]+}}, -|v{{[0-9]+}}|
76; CI: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}}
77define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f16(half addrspace(1)* %out) #1 {
78  %val = load half, half addrspace(1)* %out
79  %val.fabs = call half @llvm.fabs.f16(half %val)
80  %val.fabs.fneg = fneg half %val.fabs
81  %canonicalized = call half @llvm.canonicalize.f16(half %val.fabs.fneg)
82  store half %canonicalized, half addrspace(1)* %out
83  ret void
84}
85
86; GCN-LABEL: {{^}}v_test_canonicalize_fneg_var_f16:
87; GFX89: v_max_f16_e64 [[REG:v[0-9]+]], -{{v[0-9]+}}, -{{v[0-9]+}}
88; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
89
90; CI: v_cvt_f32_f16_e64 {{v[0-9]+}}, -{{v[0-9]+}}
91; CI: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}}
92define amdgpu_kernel void @v_test_canonicalize_fneg_var_f16(half addrspace(1)* %out) #1 {
93  %val = load half, half addrspace(1)* %out
94  %val.fneg = fneg half %val
95  %canonicalized = call half @llvm.canonicalize.f16(half %val.fneg)
96  store half %canonicalized, half addrspace(1)* %out
97  ret void
98}
99
100; GCN-LABEL: {{^}}v_test_no_denormals_canonicalize_fneg_var_f16:
101; VI: v_mul_f16_e32 [[REG:v[0-9]+]], -1.0, v{{[0-9]+}}
102; GFX9: v_max_f16_e64 [[REG:v[0-9]+]], -v{{[0-9]+}}, -v{{[0-9]+}}
103; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
104define amdgpu_kernel void @v_test_no_denormals_canonicalize_fneg_var_f16(half addrspace(1)* %out) #2 {
105  %val = load half, half addrspace(1)* %out
106  %val.fneg = fneg half %val
107  %canonicalized = call half @llvm.canonicalize.f16(half %val.fneg)
108  store half %canonicalized, half addrspace(1)* %out
109  ret void
110}
111
112; GCN-LABEL: {{^}}v_test_no_denormals_canonicalize_fneg_fabs_var_f16:
113; VI: v_mul_f16_e64 [[REG:v[0-9]+]], -1.0, |v{{[0-9]+}}|
114; GFX9: v_max_f16_e64 [[REG:v[0-9]+]], -|v{{[0-9]+}}|, -|v{{[0-9]+}}|
115
116; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
117
118; CI: v_cvt_f32_f16_e64 {{v[0-9]+}}, -|{{v[0-9]+}}|
119; CI: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}}
120define amdgpu_kernel void @v_test_no_denormals_canonicalize_fneg_fabs_var_f16(half addrspace(1)* %out) #2 {
121  %val = load half, half addrspace(1)* %out
122  %val.fabs = call half @llvm.fabs.f16(half %val)
123  %val.fabs.fneg = fneg half %val.fabs
124  %canonicalized = call half @llvm.canonicalize.f16(half %val.fabs.fneg)
125  store half %canonicalized, half addrspace(1)* %out
126  ret void
127}
128
129; GCN-LABEL: {{^}}test_fold_canonicalize_p0_f16:
130; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}}
131; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
132define amdgpu_kernel void @test_fold_canonicalize_p0_f16(half addrspace(1)* %out) #1 {
133  %canonicalized = call half @llvm.canonicalize.f16(half 0.0)
134  store half %canonicalized, half addrspace(1)* %out
135  ret void
136}
137
138; GCN-LABEL: {{^}}test_fold_canonicalize_n0_f16:
139; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffff8000{{$}}
140; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
141define amdgpu_kernel void @test_fold_canonicalize_n0_f16(half addrspace(1)* %out) #1 {
142  %canonicalized = call half @llvm.canonicalize.f16(half -0.0)
143  store half %canonicalized, half addrspace(1)* %out
144  ret void
145}
146
147; GCN-LABEL: {{^}}test_fold_canonicalize_p1_f16:
148; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3c00{{$}}
149; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
150define amdgpu_kernel void @test_fold_canonicalize_p1_f16(half addrspace(1)* %out) #1 {
151  %canonicalized = call half @llvm.canonicalize.f16(half 1.0)
152  store half %canonicalized, half addrspace(1)* %out
153  ret void
154}
155
156; GCN-LABEL: {{^}}test_fold_canonicalize_n1_f16:
157; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffffbc00{{$}}
158; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
159define amdgpu_kernel void @test_fold_canonicalize_n1_f16(half addrspace(1)* %out) #1 {
160  %canonicalized = call half @llvm.canonicalize.f16(half -1.0)
161  store half %canonicalized, half addrspace(1)* %out
162  ret void
163}
164
165; GCN-LABEL: {{^}}test_fold_canonicalize_literal_f16:
166; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x4c00{{$}}
167; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
168define amdgpu_kernel void @test_fold_canonicalize_literal_f16(half addrspace(1)* %out) #1 {
169  %canonicalized = call half @llvm.canonicalize.f16(half 16.0)
170  store half %canonicalized, half addrspace(1)* %out
171  ret void
172}
173
174; GCN-LABEL: {{^}}test_default_denormals_fold_canonicalize_denormal0_f16:
175; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3ff{{$}}
176; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
177define amdgpu_kernel void @test_default_denormals_fold_canonicalize_denormal0_f16(half addrspace(1)* %out) #1 {
178  %canonicalized = call half @llvm.canonicalize.f16(half 0xH03FF)
179  store half %canonicalized, half addrspace(1)* %out
180  ret void
181}
182
183; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal0_f16:
184; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3ff{{$}}
185; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
186define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal0_f16(half addrspace(1)* %out) #3 {
187  %canonicalized = call half @llvm.canonicalize.f16(half 0xH03FF)
188  store half %canonicalized, half addrspace(1)* %out
189  ret void
190}
191
192; GCN-LABEL: {{^}}test_default_denormals_fold_canonicalize_denormal1_f16:
193; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffff83ff{{$}}
194; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
195define amdgpu_kernel void @test_default_denormals_fold_canonicalize_denormal1_f16(half addrspace(1)* %out) #1 {
196  %canonicalized = call half @llvm.canonicalize.f16(half 0xH83FF)
197  store half %canonicalized, half addrspace(1)* %out
198  ret void
199}
200
201; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal1_f16:
202; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffff83ff{{$}}
203; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
204define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal1_f16(half addrspace(1)* %out) #3 {
205  %canonicalized = call half @llvm.canonicalize.f16(half 0xH83FF)
206  store half %canonicalized, half addrspace(1)* %out
207  ret void
208}
209
210; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_f16:
211; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7c00{{$}}
212; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
213define amdgpu_kernel void @test_fold_canonicalize_qnan_f16(half addrspace(1)* %out) #1 {
214  %canonicalized = call half @llvm.canonicalize.f16(half 0xH7C00)
215  store half %canonicalized, half addrspace(1)* %out
216  ret void
217}
218
219; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg1_f16:
220; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}}
221; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
222define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg1_f16(half addrspace(1)* %out) #1 {
223  %canonicalized = call half @llvm.canonicalize.f16(half bitcast (i16 -1 to half))
224  store half %canonicalized, half addrspace(1)* %out
225  ret void
226}
227
228; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg2_f16:
229; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}}
230; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
231define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg2_f16(half addrspace(1)* %out) #1 {
232  %canonicalized = call half @llvm.canonicalize.f16(half bitcast (i16 -2 to half))
233  store half %canonicalized, half addrspace(1)* %out
234  ret void
235}
236
237; GCN-LABEL: {{^}}test_fold_canonicalize_snan0_value_f16:
238; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}}
239; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
240define amdgpu_kernel void @test_fold_canonicalize_snan0_value_f16(half addrspace(1)* %out) #1 {
241  %canonicalized = call half @llvm.canonicalize.f16(half 0xH7C01)
242  store half %canonicalized, half addrspace(1)* %out
243  ret void
244}
245
246; GCN-LABEL: {{^}}test_fold_canonicalize_snan1_value_f16:
247; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}}
248; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
249define amdgpu_kernel void @test_fold_canonicalize_snan1_value_f16(half addrspace(1)* %out) #1 {
250  %canonicalized = call half @llvm.canonicalize.f16(half 0xH7DFF)
251  store half %canonicalized, half addrspace(1)* %out
252  ret void
253}
254
255; GCN-LABEL: {{^}}test_fold_canonicalize_snan2_value_f16:
256; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}}
257; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
258define amdgpu_kernel void @test_fold_canonicalize_snan2_value_f16(half addrspace(1)* %out) #1 {
259  %canonicalized = call half @llvm.canonicalize.f16(half 0xHFDFF)
260  store half %canonicalized, half addrspace(1)* %out
261  ret void
262}
263
264; GCN-LABEL: {{^}}test_fold_canonicalize_snan3_value_f16:
265; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e00{{$}}
266; GFX89: {{flat|global}}_store_short v{{.+}}, [[REG]]
267define amdgpu_kernel void @test_fold_canonicalize_snan3_value_f16(half addrspace(1)* %out) #1 {
268  %canonicalized = call half @llvm.canonicalize.f16(half 0xHFC01)
269  store half %canonicalized, half addrspace(1)* %out
270  ret void
271}
272
273; GCN-LABEL: {{^}}v_test_canonicalize_var_v2f16:
274; VI-DAG: v_max_f16_sdwa [[REG0:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
275; VI-DAG: v_max_f16_e32 [[REG1:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}}
276; VI-NOT: v_and_b32
277
278; GFX9: v_pk_max_f16 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+$}}
279; GFX9: global_store_dword v{{.+}}, [[REG]], s
280define amdgpu_kernel void @v_test_canonicalize_var_v2f16(<2 x half> addrspace(1)* %out) #1 {
281  %tid = call i32 @llvm.amdgcn.workitem.id.x()
282  %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
283  %val = load <2 x half>, <2 x half> addrspace(1)* %gep
284  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %val)
285  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
286  ret void
287}
288
289; GCN-LABEL: {{^}}v_test_canonicalize_fabs_var_v2f16:
290; VI: v_max_f16_sdwa [[REG0:v[0-9]+]], |v{{[0-9]+}}|, |v{{[0-9]+}}| dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
291; VI: v_max_f16_e64 [[REG1:v[0-9]+]], |v{{[0-9]+}}|, |v{{[0-9]+}}|
292; VI-NOT: 0xffff
293; VI: v_or_b32
294
295; GFX9: v_and_b32_e32 [[ABS:v[0-9]+]], 0x7fff7fff, v{{[0-9]+}}
296; GFX9: v_pk_max_f16 [[REG:v[0-9]+]], [[ABS]], [[ABS]]{{$}}
297; GFX89: {{flat|global}}_store_dword
298define amdgpu_kernel void @v_test_canonicalize_fabs_var_v2f16(<2 x half> addrspace(1)* %out) #1 {
299  %tid = call i32 @llvm.amdgcn.workitem.id.x()
300  %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
301  %val = load <2 x half>, <2 x half> addrspace(1)* %gep
302  %val.fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %val)
303  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %val.fabs)
304  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
305  ret void
306}
307
308; GCN-LABEL: {{^}}v_test_canonicalize_fneg_fabs_var_v2f16:
309; VI-DAG: v_max_f16_sdwa [[REG0:v[0-9]+]], -|v{{[0-9]+}}|, -|v{{[0-9]+}}| dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
310; VI-DAG: v_max_f16_e64 [[REG1:v[0-9]+]], -|v{{[0-9]+}}|, -|v{{[0-9]+}}|
311; VI: v_or_b32
312
313; GFX9: v_and_b32_e32 [[ABS:v[0-9]+]], 0x7fff7fff, v{{[0-9]+}}
314; GFX9: v_pk_max_f16 [[REG:v[0-9]+]], [[ABS]], [[ABS]] neg_lo:[1,1] neg_hi:[1,1]{{$}}
315; GFX89: {{flat|global}}_store_dword
316
317; CI: v_cvt_f32_f16
318; CI: v_cvt_f32_f16
319; CI: v_mul_f32_e32 v{{[0-9]+}}, 1.0
320; CI: v_mul_f32_e32 v{{[0-9]+}}, 1.0
321define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_v2f16(<2 x half> addrspace(1)* %out) #1 {
322  %tid = call i32 @llvm.amdgcn.workitem.id.x()
323  %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
324  %val = load <2 x half>, <2 x half> addrspace(1)* %gep
325  %val.fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %val)
326  %val.fabs.fneg = fneg <2 x half> %val.fabs
327  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %val.fabs.fneg)
328  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
329  ret void
330}
331
332; GCN-LABEL: {{^}}v_test_canonicalize_fneg_var_v2f16:
333; VI-DAG: v_max_f16_sdwa [[REG1:v[0-9]+]], -v{{[0-9]+}}, -v{{[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
334; VI-DAG: v_max_f16_e64 [[REG0:v[0-9]+]], -v{{[0-9]+}}, -v{{[0-9]+}}
335; VI-NOT: 0xffff
336
337; GFX9: v_pk_max_f16 [[REG:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}} neg_lo:[1,1] neg_hi:[1,1]{{$}}
338; GFX9: global_store_dword v{{[0-9]+}}, [[REG]], s
339define amdgpu_kernel void @v_test_canonicalize_fneg_var_v2f16(<2 x half> addrspace(1)* %out) #1 {
340  %tid = call i32 @llvm.amdgcn.workitem.id.x()
341  %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
342  %val = load <2 x half>, <2 x half> addrspace(1)* %gep
343  %fneg.val = fneg <2 x half> %val
344  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %fneg.val)
345  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
346  ret void
347}
348
349; GCN-LABEL: {{^}}s_test_canonicalize_var_v2f16:
350; VI-DAG: v_max_f16_sdwa [[REG0:v[0-9]+]], {{v[0-9]+}}, {{v[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
351; VI-DAG: v_max_f16_e64 [[REG1:v[0-9]+]], {{s[0-9]+}}, {{s[0-9]+}}
352; VI-NOT: v_and_b32
353
354; GFX9: v_pk_max_f16 [[REG:v[0-9]+]], {{s[0-9]+}}, {{s[0-9]+$}}
355; GFX9: global_store_dword v{{[0-9]+}}, [[REG]], s
356define amdgpu_kernel void @s_test_canonicalize_var_v2f16(<2 x half> addrspace(1)* %out, i32 zeroext %val.arg) #1 {
357  %val = bitcast i32 %val.arg to <2 x half>
358  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %val)
359  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
360  ret void
361}
362
363; GCN-LABEL: {{^}}test_fold_canonicalize_p0_v2f16:
364; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}}
365; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
366define amdgpu_kernel void @test_fold_canonicalize_p0_v2f16(<2 x half> addrspace(1)* %out) #1 {
367  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> zeroinitializer)
368  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
369  ret void
370}
371
372; GCN-LABEL: {{^}}test_fold_canonicalize_n0_v2f16:
373; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x80008000{{$}}
374; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
375define amdgpu_kernel void @test_fold_canonicalize_n0_v2f16(<2 x half> addrspace(1)* %out) #1 {
376  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half -0.0, half -0.0>)
377  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
378  ret void
379}
380
381; GCN-LABEL: {{^}}test_fold_canonicalize_p1_v2f16:
382; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3c003c00{{$}}
383; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
384define amdgpu_kernel void @test_fold_canonicalize_p1_v2f16(<2 x half> addrspace(1)* %out) #1 {
385  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 1.0, half 1.0>)
386  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
387  ret void
388}
389
390; GCN-LABEL: {{^}}test_fold_canonicalize_n1_v2f16:
391; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0xbc00bc00{{$}}
392; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
393define amdgpu_kernel void @test_fold_canonicalize_n1_v2f16(<2 x half> addrspace(1)* %out) #1 {
394  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half -1.0, half -1.0>)
395  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
396  ret void
397}
398
399; GCN-LABEL: {{^}}test_fold_canonicalize_literal_v2f16:
400; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x4c004c00{{$}}
401; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
402define amdgpu_kernel void @test_fold_canonicalize_literal_v2f16(<2 x half> addrspace(1)* %out) #1 {
403  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 16.0, half 16.0>)
404  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
405  ret void
406}
407
408; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal0_v2f16:
409; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3ff03ff{{$}}
410; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
411define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal0_v2f16(<2 x half> addrspace(1)* %out) #1 {
412  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH03FF, half 0xH03FF>)
413  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
414  ret void
415}
416
417; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal0_v2f16:
418; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3ff03ff{{$}}
419; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
420define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal0_v2f16(<2 x half> addrspace(1)* %out) #3 {
421  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH03FF, half 0xH03FF>)
422  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
423  ret void
424}
425
426; GCN-LABEL: {{^}}test_no_denormals_fold_canonicalize_denormal1_v2f16:
427; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x83ff83ff{{$}}
428; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
429define amdgpu_kernel void @test_no_denormals_fold_canonicalize_denormal1_v2f16(<2 x half> addrspace(1)* %out) #1 {
430  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH83FF, half 0xH83FF>)
431  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
432  ret void
433}
434
435; GCN-LABEL: {{^}}test_denormals_fold_canonicalize_denormal1_v2f16:
436; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x83ff83ff{{$}}
437; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
438define amdgpu_kernel void @test_denormals_fold_canonicalize_denormal1_v2f16(<2 x half> addrspace(1)* %out) #3 {
439  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH83FF, half 0xH83FF>)
440  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
441  ret void
442}
443
444; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_v2f16:
445; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7c007c00{{$}}
446; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
447define amdgpu_kernel void @test_fold_canonicalize_qnan_v2f16(<2 x half> addrspace(1)* %out) #1 {
448  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH7C00, half 0xH7C00>)
449  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
450  ret void
451}
452
453; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg1_v2f16:
454; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00{{$}}
455; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
456define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg1_v2f16(<2 x half> addrspace(1)* %out) #1 {
457  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> bitcast (i32 -1 to <2 x half>))
458  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
459  ret void
460}
461
462; GCN-LABEL: {{^}}test_fold_canonicalize_qnan_value_neg2_v2f16:
463; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00{{$}}
464; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
465define amdgpu_kernel void @test_fold_canonicalize_qnan_value_neg2_v2f16(<2 x half> addrspace(1)* %out) #1 {
466  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half bitcast (i16 -2 to half), half bitcast (i16 -2 to half)>)
467  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
468  ret void
469}
470
471; GCN-LABEL: {{^}}test_fold_canonicalize_snan0_value_v2f16:
472; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00{{$}}
473; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
474define amdgpu_kernel void @test_fold_canonicalize_snan0_value_v2f16(<2 x half> addrspace(1)* %out) #1 {
475  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH7C01, half 0xH7C01>)
476  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
477  ret void
478}
479
480; GCN-LABEL: {{^}}test_fold_canonicalize_snan1_value_v2f16:
481; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00{{$}}
482; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
483define amdgpu_kernel void @test_fold_canonicalize_snan1_value_v2f16(<2 x half> addrspace(1)* %out) #1 {
484  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xH7DFF, half 0xH7DFF>)
485  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
486  ret void
487}
488
489; GCN-LABEL: {{^}}test_fold_canonicalize_snan2_value_v2f16:
490; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00{{$}}
491; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
492define amdgpu_kernel void @test_fold_canonicalize_snan2_value_v2f16(<2 x half> addrspace(1)* %out) #1 {
493  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xHFDFF, half 0xHFDFF>)
494  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
495  ret void
496}
497
498; GCN-LABEL: {{^}}test_fold_canonicalize_snan3_value_v2f16:
499; GFX89: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00{{$}}
500; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
501define amdgpu_kernel void @test_fold_canonicalize_snan3_value_v2f16(<2 x half> addrspace(1)* %out) #1 {
502  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> <half 0xHFC01, half 0xHFC01>)
503  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
504  ret void
505}
506
507; FIXME: Extra 4th component handled
508; GCN-LABEL: {{^}}v_test_canonicalize_var_v3f16:
509; GFX9: s_waitcnt
510; GFX9-NEXT: v_pk_max_f16 v0, v0, v0
511; GFX9-NEXT: v_pk_max_f16 v1, v1, v1
512; GFX9-NEXT: s_setpc_b64
513
514; VI-DAG: v_max_f16_sdwa [[CANON_ELT1:v[0-9]+]], v0, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
515; VI-DAG: v_max_f16_e32 [[CANON_ELT0:v[0-9]+]], v0, v0
516; VI-DAG: v_max_f16_e32 v1, v1, v1
517; VI-DAG: v_or_b32_e32 v0, [[CANON_ELT0]], [[CANON_ELT1]]
518
519; VI: s_setpc_b64
520define <3 x half> @v_test_canonicalize_var_v3f16(<3 x half> %val) #1 {
521  %canonicalized = call <3 x half> @llvm.canonicalize.v3f16(<3 x half> %val)
522  ret <3 x half> %canonicalized
523}
524
525; GCN-LABEL: {{^}}v_test_canonicalize_var_v4f16:
526; GFX9: s_waitcnt
527; GFX9-NEXT: v_pk_max_f16 v0, v0, v0
528; GFX9-NEXT: v_pk_max_f16 v1, v1, v1
529; GFX9-NEXT: s_setpc_b64
530
531; VI-DAG: v_max_f16_sdwa [[CANON_ELT3:v[0-9]+]], v1, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
532; VI-DAG: v_max_f16_e32 [[CANON_ELT2:v[0-9]+]], v1, v1
533; VI-DAG: v_max_f16_sdwa [[CANON_ELT1:v[0-9]+]], v0, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
534; VI-DAG: v_max_f16_e32 [[CANON_ELT0:v[0-9]+]], v0, v0
535; VI-DAG: v_or_b32_e32 v0, [[CANON_ELT0]], [[CANON_ELT1]]
536; VI-DAG: v_or_b32_e32 v1, [[CANON_ELT2]], [[CANON_ELT3]]
537; VI: s_setpc_b64
538define <4 x half> @v_test_canonicalize_var_v4f16(<4 x half> %val) #1 {
539  %canonicalized = call <4 x half> @llvm.canonicalize.v4f16(<4 x half> %val)
540  ret <4 x half> %canonicalized
541}
542
543; GCN-LABEL: {{^}}s_test_canonicalize_undef_v2f16:
544; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x7e007e00
545; GFX89: {{flat|global}}_store_dword v{{.+}}, [[REG]]
546define amdgpu_kernel void @s_test_canonicalize_undef_v2f16(<2 x half> addrspace(1)* %out) #1 {
547  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> undef)
548  store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
549  ret void
550}
551
552; GCN-LABEL: {{^}}v_test_canonicalize_reg_undef_v2f16:
553; GFX9: s_waitcnt
554; GFX9-NEXT: v_max_f16_e32 v0, v0, v0
555; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
556; GFX9-NEXT: s_setpc_b64
557
558; High bits known zero
559; FIXME: Should also be true on gfx9 by default?
560; VI: s_waitcnt
561; VI-NEXT: v_max_f16_e32 v0, v0, v0
562; VI-NEXT: s_setpc_b64
563define <2 x half> @v_test_canonicalize_reg_undef_v2f16(half %val) #1 {
564  %vec = insertelement <2 x half> undef, half %val, i32 0
565  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec)
566  ret <2 x half> %canonicalized
567}
568
569; GCN-LABEL: {{^}}v_test_canonicalize_undef_reg_v2f16:
570; GFX89: s_waitcnt
571; GFX89-NEXT: v_max_f16_sdwa v0, v0, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
572; GFX89-NEXT: s_setpc_b64
573define <2 x half> @v_test_canonicalize_undef_reg_v2f16(half %val) #1 {
574  %vec = insertelement <2 x half> undef, half %val, i32 1
575  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec)
576  ret <2 x half> %canonicalized
577}
578
579; GCN-LABEL: {{^}}v_test_canonicalize_undef_lo_imm_hi_v2f16:
580; GCN: s_waitcnt
581; GFX89-NEXT: v_mov_b32_e32 v0, 0x3c003c00
582; GFX89-NEXT: s_setpc_b64
583
584; CI-NEXT: v_mov_b32_e32 v0, 0x7fc00000
585; CI-NEXT: v_mov_b32_e32 v1, 1.0
586; CI-NEXT: s_setpc_b64
587define <2 x half> @v_test_canonicalize_undef_lo_imm_hi_v2f16() #1 {
588  %vec = insertelement <2 x half> undef, half 1.0, i32 1
589  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec)
590  ret <2 x half> %canonicalized
591}
592
593; GCN-LABEL: {{^}}v_test_canonicalize_imm_lo_undef_hi_v2f16:
594; GCN: s_waitcnt
595; GFX89-NEXT: v_mov_b32_e32 v0, 0x3c003c00
596; GFX89-NEXT: s_setpc_b64
597
598; CI-NEXT: v_mov_b32_e32 v0, 1.0
599; CI-NEXT: v_mov_b32_e32 v1, 0x7fc00000
600; CI-NEXT: s_setpc_b64
601define <2 x half> @v_test_canonicalize_imm_lo_undef_hi_v2f16() #1 {
602  %vec = insertelement <2 x half> undef, half 1.0, i32 0
603  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec)
604  ret <2 x half> %canonicalized
605}
606
607; GCN-LABEL: {{^}}v_test_canonicalize_undef_lo_k_hi_v2f16:
608; GCN: s_waitcnt
609; GFX89-NEXT: v_mov_b32_e32 v0, 0x4c004c00
610; GFX89-NEXT: s_setpc_b64
611
612; CI-NEXT: v_mov_b32_e32 v0, 0x7fc00000
613; CI-NEXT: v_mov_b32_e32 v1, 0x41800000
614; CI-NEXT: s_setpc_b64
615define <2 x half> @v_test_canonicalize_undef_lo_k_hi_v2f16() #1 {
616  %vec = insertelement <2 x half> undef, half 16.0, i32 1
617  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec)
618  ret <2 x half> %canonicalized
619}
620
621; GCN-LABEL: {{^}}v_test_canonicalize_k_lo_undef_hi_v2f16:
622; GCN: s_waitcnt
623; GFX89-NEXT: v_mov_b32_e32 v0, 0x4c004c00
624; GFX89-NEXT: s_setpc_b64
625
626; CI-NEXT: v_mov_b32_e32 v0, 0x41800000
627; CI-NEXT: v_mov_b32_e32 v1, 0x7fc00000
628; CI-NEXT: s_setpc_b64
629define <2 x half> @v_test_canonicalize_k_lo_undef_hi_v2f16() #1 {
630  %vec = insertelement <2 x half> undef, half 16.0, i32 0
631  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec)
632  ret <2 x half> %canonicalized
633}
634
635; GCN-LABEL: {{^}}v_test_canonicalize_reg_k_v2f16:
636; GFX9: s_waitcnt
637; GFX9-DAG: v_max_f16_e32 v0, v0, v0
638; GFX9-DAG: s_movk_i32 [[K:s[0-9]+]], 0x4000
639; GFX9: v_and_b32_e32 v0, 0xffff, v0
640; GFX9: v_lshl_or_b32 v0, [[K]], 16, v0
641; GFX9: s_setpc_b64
642
643; VI: s_waitcnt
644; VI-NEXT: v_max_f16_e32 v0, v0, v0
645; VI-NEXT: v_or_b32_e32 v0, 2.0, v0
646; VI-NEXT: s_setpc_b64
647define <2 x half> @v_test_canonicalize_reg_k_v2f16(half %val) #1 {
648  %vec0 = insertelement <2 x half> undef, half %val, i32 0
649  %vec1 = insertelement <2 x half> %vec0, half 2.0, i32 1
650  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec1)
651  ret <2 x half> %canonicalized
652}
653
654; GCN-LABEL: {{^}}v_test_canonicalize_k_reg_v2f16:
655; GFX9: v_max_f16_e32 v0, v0, v0
656; GFX9: v_mov_b32_e32 [[K:v[0-9]+]], 0x4000
657; GFX9: v_lshl_or_b32 v0, v0, 16, [[K]]
658; GFX9: s_setpc_b64
659
660; VI: s_waitcnt
661; VI-NEXT: v_max_f16_sdwa v0, v0, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
662; VI-NEXT: v_or_b32_e32 v0, 0x4000, v0
663; VI-NEXT: s_setpc_b64
664define <2 x half> @v_test_canonicalize_k_reg_v2f16(half %val) #1 {
665  %vec0 = insertelement <2 x half> undef, half 2.0, i32 0
666  %vec1 = insertelement <2 x half> %vec0, half %val, i32 1
667  %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %vec1)
668  ret <2 x half> %canonicalized
669}
670
671; GCN-LABEL: {{^}}s_test_canonicalize_undef_v4f16:
672; GCN: v_mov_b32_e32 v0, 0x7e007e00
673; GCN: v_mov_b32_e32 v1, v0
674define amdgpu_kernel void @s_test_canonicalize_undef_v4f16(<4 x half> addrspace(1)* %out) #1 {
675  %canonicalized = call <4 x half> @llvm.canonicalize.v4f16(<4 x half> undef)
676  store <4 x half> %canonicalized, <4 x half> addrspace(1)* %out
677  ret void
678}
679
680; GCN-LABEL: {{^}}v_test_canonicalize_reg_undef_undef_undef_v4f16:
681; GFX9: s_waitcnt
682; GFX9-NEXT: v_max_f16_e32 v0, v0, v0
683; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
684; GFX9-NEXT: v_mov_b32_e32 v1, 0
685; GFX9-NEXT: s_setpc_b64
686
687; VI: s_waitcnt
688; VI-NEXT: v_max_f16_e32 v0, v0, v0
689; VI-NEXT: v_or_b32_e32 v0, 0x7e000000, v0
690; VI-NEXT: v_mov_b32_e32 v1, 0x7e007e00
691; VI-NEXT: s_setpc_b64
692define <4 x half> @v_test_canonicalize_reg_undef_undef_undef_v4f16(half %val) #1 {
693  %vec = insertelement <4 x half> undef, half %val, i32 0
694  %canonicalized = call <4 x half> @llvm.canonicalize.v4f16(<4 x half> %vec)
695  ret <4 x half> %canonicalized
696}
697
698; GCN-LABEL: {{^}}v_test_canonicalize_reg_reg_undef_undef_v4f16:
699; GFX9: s_waitcnt
700; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
701; GFX9-NEXT: v_lshl_or_b32 v0, v1, 16, v0
702; GFX9-NEXT: v_pk_max_f16 v0, v0, v0
703; GFX9-NEXT: v_mov_b32_e32 v1, 0
704; GFX9-NEXT: s_setpc_b64
705
706; VI: s_waitcnt
707; VI-DAG: v_max_f16_e32 v0, v0, v0
708; VI-DAG: v_max_f16_sdwa v1, v1, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
709; VI: v_or_b32_e32 v0, v0, v1
710; VI-NEXT: v_mov_b32_e32 v1, 0x7e007e00
711; VI-NEXT: s_setpc_b64
712define <4 x half> @v_test_canonicalize_reg_reg_undef_undef_v4f16(half %val0, half %val1) #1 {
713  %vec0 = insertelement <4 x half> undef, half %val0, i32 0
714  %vec1 = insertelement <4 x half> %vec0, half %val1, i32 1
715  %canonicalized = call <4 x half> @llvm.canonicalize.v4f16(<4 x half> %vec1)
716  ret <4 x half> %canonicalized
717}
718
719; GCN-LABEL: {{^}}v_test_canonicalize_reg_undef_reg_reg_v4f16:
720; GFX9: s_waitcnt
721; GFX9-NEXT: v_and_b32_e32 v1, 0xffff, v1
722; GFX9-NEXT: v_max_f16_e32 v0, v0, v0
723; GFX9-NEXT: v_lshl_or_b32 v1, v2, 16, v1
724; GFX9-NEXT: v_and_b32_e32 v0, 0xffff, v0
725; GFX9-NEXT: v_pk_max_f16 v1, v1, v1
726; GFX9-NEXT: s_setpc_b64
727
728; VI: s_waitcnt
729; VI-NEXT: v_max_f16_e32 v0, v0, v0
730; VI-NEXT: v_max_f16_e32 v1, v1, v1
731; VI-NEXT: v_max_f16_sdwa v2, v2, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
732; VI-NEXT: v_or_b32_e32 v0, 0x7e000000, v0
733; VI-NEXT: v_or_b32_e32 v1, v1, v2
734; VI-NEXT: s_setpc_b64
735define <4 x half> @v_test_canonicalize_reg_undef_reg_reg_v4f16(half %val0, half %val1, half %val2) #1 {
736  %vec0 = insertelement <4 x half> undef, half %val0, i32 0
737  %vec1 = insertelement <4 x half> %vec0, half %val1, i32 2
738  %vec2 = insertelement <4 x half> %vec1, half %val2, i32 3
739  %canonicalized = call <4 x half> @llvm.canonicalize.v4f16(<4 x half> %vec2)
740  ret <4 x half> %canonicalized
741}
742
743attributes #0 = { nounwind readnone }
744attributes #1 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
745attributes #2 = { nounwind "denormal-fp-math"="preserve-sign,preserve-sign" }
746attributes #3 = { nounwind "denormal-fp-math-f32"="preserve-sign,preserve-sign" }
747