| blob | 441eec00d5e9cdc3e50a74c09b8b54cd10258ca9 |
1 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
2 # RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX6 %s
3 # RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX7 %s
4 # RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX10 %s
6 ---
7 name: test_fmad_s16_flush
8 machineFunctionInfo:
9 mode:
10 fp64-fp16-input-denormals: false
11 fp64-fp16-output-denormals: false
13 body: |
14 bb.0:
15 liveins: $vgpr0, $vgpr1, $vgpr2
17 ; GFX6-LABEL: name: test_fmad_s16_flush
18 ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
19 ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
20 ; GFX6: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr1
21 ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
22 ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[COPY1]](s32)
23 ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[COPY2]](s32)
24 ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
25 ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
26 ; GFX6: [[FMUL:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT]], [[FPEXT1]]
27 ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL]](s32)
28 ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
29 ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
30 ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT2]], [[FPEXT3]]
31 ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32)
32 ; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FPTRUNC1]](s16)
33 ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
34 ; GFX7-LABEL: name: test_fmad_s16_flush
35 ; GFX7: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
36 ; GFX7: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
37 ; GFX7: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr1
38 ; GFX7: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
39 ; GFX7: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[COPY1]](s32)
40 ; GFX7: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[COPY2]](s32)
41 ; GFX7: [[FMAD:%[0-9]+]]:_(s16) = G_FMAD [[TRUNC]], [[TRUNC1]], [[TRUNC2]]
42 ; GFX7: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FMAD]](s16)
43 ; GFX7: $vgpr0 = COPY [[ANYEXT]](s32)
44 ; GFX10-LABEL: name: test_fmad_s16_flush
45 ; GFX10: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
46 ; GFX10: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
47 ; GFX10: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr1
48 ; GFX10: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
49 ; GFX10: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[COPY1]](s32)
50 ; GFX10: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[COPY2]](s32)
51 ; GFX10: [[FMUL:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC]], [[TRUNC1]]
52 ; GFX10: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[FMUL]], [[TRUNC2]]
53 ; GFX10: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16)
54 ; GFX10: $vgpr0 = COPY [[ANYEXT]](s32)
55 %0:_(s32) = COPY $vgpr0
56 %1:_(s32) = COPY $vgpr1
57 %2:_(s32) = COPY $vgpr1
58 %3:_(s16) = G_TRUNC %0
59 %4:_(s16) = G_TRUNC %1
60 %5:_(s16) = G_TRUNC %2
61 %6:_(s16) = G_FMAD %3, %4, %5
62 %7:_(s32) = G_ANYEXT %6
63 $vgpr0 = COPY %7
64 ...
66 ---
67 name: test_fmad_v2s16_flush
68 machineFunctionInfo:
69 mode:
70 fp64-fp16-input-denormals: false
71 fp64-fp16-output-denormals: false
73 body: |
74 bb.0:
75 liveins: $vgpr0, $vgpr1, $vgpr2
77 ; GFX6-LABEL: name: test_fmad_v2s16_flush
78 ; GFX6: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
79 ; GFX6: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
80 ; GFX6: [[COPY2:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr2
81 ; GFX6: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>)
82 ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
83 ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
84 ; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
85 ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
86 ; GFX6: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[COPY1]](<2 x s16>)
87 ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
88 ; GFX6: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
89 ; GFX6: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
90 ; GFX6: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[COPY2]](<2 x s16>)
91 ; GFX6: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
92 ; GFX6: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
93 ; GFX6: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
94 ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
95 ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
96 ; GFX6: [[FMUL:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT]], [[FPEXT1]]
97 ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL]](s32)
98 ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
99 ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC4]](s16)
100 ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT2]], [[FPEXT3]]
101 ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32)
102 ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
103 ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC3]](s16)
104 ; GFX6: [[FMUL1:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT4]], [[FPEXT5]]
105 ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL1]](s32)
106 ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16)
107 ; GFX6: [[FPEXT7:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC5]](s16)
108 ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT6]], [[FPEXT7]]
109 ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32)
110 ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC1]](s16)
111 ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC3]](s16)
112 ; GFX6: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32)
113 ; GFX6: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]]
114 ; GFX6: [[BITCAST3:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
115 ; GFX6: $vgpr0 = COPY [[BITCAST3]](<2 x s16>)
116 ; GFX7-LABEL: name: test_fmad_v2s16_flush
117 ; GFX7: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
118 ; GFX7: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
119 ; GFX7: [[COPY2:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr2
120 ; GFX7: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>)
121 ; GFX7: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
122 ; GFX7: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
123 ; GFX7: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
124 ; GFX7: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
125 ; GFX7: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[COPY1]](<2 x s16>)
126 ; GFX7: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
127 ; GFX7: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
128 ; GFX7: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
129 ; GFX7: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[COPY2]](<2 x s16>)
130 ; GFX7: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
131 ; GFX7: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
132 ; GFX7: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
133 ; GFX7: [[FMAD:%[0-9]+]]:_(s16) = G_FMAD [[TRUNC]], [[TRUNC2]], [[TRUNC4]]
134 ; GFX7: [[FMAD1:%[0-9]+]]:_(s16) = G_FMAD [[TRUNC1]], [[TRUNC3]], [[TRUNC5]]
135 ; GFX7: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FMAD]](s16)
136 ; GFX7: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FMAD1]](s16)
137 ; GFX7: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32)
138 ; GFX7: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]]
139 ; GFX7: [[BITCAST3:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
140 ; GFX7: $vgpr0 = COPY [[BITCAST3]](<2 x s16>)
141 ; GFX10-LABEL: name: test_fmad_v2s16_flush
142 ; GFX10: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
143 ; GFX10: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
144 ; GFX10: [[COPY2:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr2
145 ; GFX10: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>)
146 ; GFX10: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
147 ; GFX10: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
148 ; GFX10: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
149 ; GFX10: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
150 ; GFX10: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[COPY1]](<2 x s16>)
151 ; GFX10: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
152 ; GFX10: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
153 ; GFX10: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
154 ; GFX10: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[COPY2]](<2 x s16>)
155 ; GFX10: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
156 ; GFX10: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
157 ; GFX10: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
158 ; GFX10: [[FMUL:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC]], [[TRUNC2]]
159 ; GFX10: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[FMUL]], [[TRUNC4]]
160 ; GFX10: [[FMUL1:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC1]], [[TRUNC3]]
161 ; GFX10: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[FMUL1]], [[TRUNC5]]
162 ; GFX10: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16)
163 ; GFX10: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16)
164 ; GFX10: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32)
165 ; GFX10: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>)
166 %0:_(<2 x s16>) = COPY $vgpr0
167 %1:_(<2 x s16>) = COPY $vgpr1
168 %2:_(<2 x s16>) = COPY $vgpr2
169 %3:_(<2 x s16>) = G_FMAD %0, %1, %2
170 $vgpr0 = COPY %3
171 ...
173 ---
174 name: test_fmad_v4s16_flush
175 machineFunctionInfo:
176 mode:
177 fp64-fp16-input-denormals: false
178 fp64-fp16-output-denormals: false
180 body: |
181 bb.0:
182 liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $vgpr4_vgpr5
184 ; GFX6-LABEL: name: test_fmad_v4s16_flush
185 ; GFX6: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
186 ; GFX6: [[COPY1:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr2_vgpr3
187 ; GFX6: [[COPY2:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr4_vgpr5
188 ; GFX6: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
189 ; GFX6: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
190 ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
191 ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
192 ; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
193 ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
194 ; GFX6: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
195 ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
196 ; GFX6: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
197 ; GFX6: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
198 ; GFX6: [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY1]](<4 x s16>)
199 ; GFX6: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
200 ; GFX6: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
201 ; GFX6: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
202 ; GFX6: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
203 ; GFX6: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
204 ; GFX6: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST3]](s32)
205 ; GFX6: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
206 ; GFX6: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR3]](s32)
207 ; GFX6: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
208 ; GFX6: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
209 ; GFX6: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST4]](s32)
210 ; GFX6: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
211 ; GFX6: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR4]](s32)
212 ; GFX6: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV5]](<2 x s16>)
213 ; GFX6: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST5]](s32)
214 ; GFX6: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
215 ; GFX6: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR5]](s32)
216 ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
217 ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC4]](s16)
218 ; GFX6: [[FMUL:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT]], [[FPEXT1]]
219 ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL]](s32)
220 ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
221 ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC8]](s16)
222 ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT2]], [[FPEXT3]]
223 ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32)
224 ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
225 ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC5]](s16)
226 ; GFX6: [[FMUL1:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT4]], [[FPEXT5]]
227 ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL1]](s32)
228 ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16)
229 ; GFX6: [[FPEXT7:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC9]](s16)
230 ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT6]], [[FPEXT7]]
231 ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32)
232 ; GFX6: [[FPEXT8:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
233 ; GFX6: [[FPEXT9:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC6]](s16)
234 ; GFX6: [[FMUL2:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT8]], [[FPEXT9]]
235 ; GFX6: [[FPTRUNC4:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL2]](s32)
236 ; GFX6: [[FPEXT10:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC4]](s16)
237 ; GFX6: [[FPEXT11:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC10]](s16)
238 ; GFX6: [[FADD2:%[0-9]+]]:_(s32) = G_FADD [[FPEXT10]], [[FPEXT11]]
239 ; GFX6: [[FPTRUNC5:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD2]](s32)
240 ; GFX6: [[FPEXT12:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC3]](s16)
241 ; GFX6: [[FPEXT13:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC7]](s16)
242 ; GFX6: [[FMUL3:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT12]], [[FPEXT13]]
243 ; GFX6: [[FPTRUNC6:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL3]](s32)
244 ; GFX6: [[FPEXT14:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC6]](s16)
245 ; GFX6: [[FPEXT15:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC11]](s16)
246 ; GFX6: [[FADD3:%[0-9]+]]:_(s32) = G_FADD [[FPEXT14]], [[FPEXT15]]
247 ; GFX6: [[FPTRUNC7:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD3]](s32)
248 ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC1]](s16)
249 ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC3]](s16)
250 ; GFX6: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32)
251 ; GFX6: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]]
252 ; GFX6: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
253 ; GFX6: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC5]](s16)
254 ; GFX6: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC7]](s16)
255 ; GFX6: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C]](s32)
256 ; GFX6: [[OR1:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL1]]
257 ; GFX6: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
258 ; GFX6: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST6]](<2 x s16>), [[BITCAST7]](<2 x s16>)
259 ; GFX6: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
260 ; GFX7-LABEL: name: test_fmad_v4s16_flush
261 ; GFX7: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
262 ; GFX7: [[COPY1:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr2_vgpr3
263 ; GFX7: [[COPY2:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr4_vgpr5
264 ; GFX7: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
265 ; GFX7: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
266 ; GFX7: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
267 ; GFX7: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
268 ; GFX7: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
269 ; GFX7: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
270 ; GFX7: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
271 ; GFX7: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
272 ; GFX7: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
273 ; GFX7: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
274 ; GFX7: [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY1]](<4 x s16>)
275 ; GFX7: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
276 ; GFX7: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
277 ; GFX7: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
278 ; GFX7: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
279 ; GFX7: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
280 ; GFX7: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST3]](s32)
281 ; GFX7: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
282 ; GFX7: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR3]](s32)
283 ; GFX7: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
284 ; GFX7: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
285 ; GFX7: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST4]](s32)
286 ; GFX7: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
287 ; GFX7: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR4]](s32)
288 ; GFX7: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV5]](<2 x s16>)
289 ; GFX7: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST5]](s32)
290 ; GFX7: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
291 ; GFX7: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR5]](s32)
292 ; GFX7: [[FMAD:%[0-9]+]]:_(s16) = G_FMAD [[TRUNC]], [[TRUNC4]], [[TRUNC8]]
293 ; GFX7: [[FMAD1:%[0-9]+]]:_(s16) = G_FMAD [[TRUNC1]], [[TRUNC5]], [[TRUNC9]]
294 ; GFX7: [[FMAD2:%[0-9]+]]:_(s16) = G_FMAD [[TRUNC2]], [[TRUNC6]], [[TRUNC10]]
295 ; GFX7: [[FMAD3:%[0-9]+]]:_(s16) = G_FMAD [[TRUNC3]], [[TRUNC7]], [[TRUNC11]]
296 ; GFX7: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FMAD]](s16)
297 ; GFX7: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FMAD1]](s16)
298 ; GFX7: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32)
299 ; GFX7: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]]
300 ; GFX7: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
301 ; GFX7: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FMAD2]](s16)
302 ; GFX7: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[FMAD3]](s16)
303 ; GFX7: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C]](s32)
304 ; GFX7: [[OR1:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL1]]
305 ; GFX7: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
306 ; GFX7: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST6]](<2 x s16>), [[BITCAST7]](<2 x s16>)
307 ; GFX7: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
308 ; GFX10-LABEL: name: test_fmad_v4s16_flush
309 ; GFX10: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
310 ; GFX10: [[COPY1:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr2_vgpr3
311 ; GFX10: [[COPY2:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr4_vgpr5
312 ; GFX10: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
313 ; GFX10: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
314 ; GFX10: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
315 ; GFX10: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
316 ; GFX10: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
317 ; GFX10: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
318 ; GFX10: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
319 ; GFX10: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
320 ; GFX10: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
321 ; GFX10: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
322 ; GFX10: [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY1]](<4 x s16>)
323 ; GFX10: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
324 ; GFX10: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
325 ; GFX10: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
326 ; GFX10: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
327 ; GFX10: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
328 ; GFX10: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST3]](s32)
329 ; GFX10: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
330 ; GFX10: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR3]](s32)
331 ; GFX10: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
332 ; GFX10: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
333 ; GFX10: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST4]](s32)
334 ; GFX10: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
335 ; GFX10: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR4]](s32)
336 ; GFX10: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV5]](<2 x s16>)
337 ; GFX10: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST5]](s32)
338 ; GFX10: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
339 ; GFX10: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR5]](s32)
340 ; GFX10: [[FMUL:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC]], [[TRUNC4]]
341 ; GFX10: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[FMUL]], [[TRUNC8]]
342 ; GFX10: [[FMUL1:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC1]], [[TRUNC5]]
343 ; GFX10: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[FMUL1]], [[TRUNC9]]
344 ; GFX10: [[FMUL2:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC2]], [[TRUNC6]]
345 ; GFX10: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[FMUL2]], [[TRUNC10]]
346 ; GFX10: [[FMUL3:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC3]], [[TRUNC7]]
347 ; GFX10: [[FADD3:%[0-9]+]]:_(s16) = G_FADD [[FMUL3]], [[TRUNC11]]
348 ; GFX10: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16)
349 ; GFX10: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16)
350 ; GFX10: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32)
351 ; GFX10: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD2]](s16)
352 ; GFX10: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD3]](s16)
353 ; GFX10: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT2]](s32), [[ANYEXT3]](s32)
354 ; GFX10: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
355 ; GFX10: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
356 %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
357 %1:_(<4 x s16>) = COPY $vgpr2_vgpr3
358 %2:_(<4 x s16>) = COPY $vgpr4_vgpr5
359 %3:_(<4 x s16>) = G_FMAD %0, %1, %2
360 $vgpr0_vgpr1 = COPY %3
361 ...
364 ---
365 name: test_fmad_s16_denorm
366 machineFunctionInfo:
367 mode:
368 fp64-fp16-input-denormals: true
369 fp64-fp16-output-denormals: true
371 body: |
372 bb.0:
373 liveins: $vgpr0, $vgpr1, $vgpr2
375 ; GFX6-LABEL: name: test_fmad_s16_denorm
376 ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
377 ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
378 ; GFX6: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr1
379 ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
380 ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[COPY1]](s32)
381 ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[COPY2]](s32)
382 ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
383 ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
384 ; GFX6: [[FMUL:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT]], [[FPEXT1]]
385 ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL]](s32)
386 ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
387 ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
388 ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT2]], [[FPEXT3]]
389 ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32)
390 ; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FPTRUNC1]](s16)
391 ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
392 ; GFX7-LABEL: name: test_fmad_s16_denorm
393 ; GFX7: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
394 ; GFX7: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
395 ; GFX7: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr1
396 ; GFX7: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
397 ; GFX7: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[COPY1]](s32)
398 ; GFX7: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[COPY2]](s32)
399 ; GFX7: [[FMUL:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC]], [[TRUNC1]]
400 ; GFX7: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[FMUL]], [[TRUNC2]]
401 ; GFX7: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16)
402 ; GFX7: $vgpr0 = COPY [[ANYEXT]](s32)
403 ; GFX10-LABEL: name: test_fmad_s16_denorm
404 ; GFX10: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
405 ; GFX10: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
406 ; GFX10: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr1
407 ; GFX10: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
408 ; GFX10: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[COPY1]](s32)
409 ; GFX10: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[COPY2]](s32)
410 ; GFX10: [[FMUL:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC]], [[TRUNC1]]
411 ; GFX10: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[FMUL]], [[TRUNC2]]
412 ; GFX10: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16)
413 ; GFX10: $vgpr0 = COPY [[ANYEXT]](s32)
414 %0:_(s32) = COPY $vgpr0
415 %1:_(s32) = COPY $vgpr1
416 %2:_(s32) = COPY $vgpr1
417 %3:_(s16) = G_TRUNC %0
418 %4:_(s16) = G_TRUNC %1
419 %5:_(s16) = G_TRUNC %2
420 %6:_(s16) = G_FMAD %3, %4, %5
421 %7:_(s32) = G_ANYEXT %6
422 $vgpr0 = COPY %7
423 ...
425 ---
426 name: test_fmad_s16_denorm_flags
427 machineFunctionInfo:
428 mode:
429 fp64-fp16-input-denormals: true
430 fp64-fp16-output-denormals: true
432 body: |
433 bb.0:
434 liveins: $vgpr0, $vgpr1, $vgpr2
436 ; GFX6-LABEL: name: test_fmad_s16_denorm_flags
437 ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
438 ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
439 ; GFX6: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr1
440 ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
441 ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[COPY1]](s32)
442 ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[COPY2]](s32)
443 ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
444 ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
445 ; GFX6: [[FMUL:%[0-9]+]]:_(s32) = nnan G_FMUL [[FPEXT]], [[FPEXT1]]
446 ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL]](s32)
447 ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
448 ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
449 ; GFX6: [[FADD:%[0-9]+]]:_(s32) = nnan G_FADD [[FPEXT2]], [[FPEXT3]]
450 ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32)
451 ; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FPTRUNC1]](s16)
452 ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
453 ; GFX7-LABEL: name: test_fmad_s16_denorm_flags
454 ; GFX7: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
455 ; GFX7: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
456 ; GFX7: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr1
457 ; GFX7: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
458 ; GFX7: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[COPY1]](s32)
459 ; GFX7: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[COPY2]](s32)
460 ; GFX7: [[FMUL:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC]], [[TRUNC1]]
461 ; GFX7: [[FADD:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL]], [[TRUNC2]]
462 ; GFX7: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16)
463 ; GFX7: $vgpr0 = COPY [[ANYEXT]](s32)
464 ; GFX10-LABEL: name: test_fmad_s16_denorm_flags
465 ; GFX10: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
466 ; GFX10: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
467 ; GFX10: [[COPY2:%[0-9]+]]:_(s32) = COPY $vgpr1
468 ; GFX10: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
469 ; GFX10: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[COPY1]](s32)
470 ; GFX10: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[COPY2]](s32)
471 ; GFX10: [[FMUL:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC]], [[TRUNC1]]
472 ; GFX10: [[FADD:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL]], [[TRUNC2]]
473 ; GFX10: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16)
474 ; GFX10: $vgpr0 = COPY [[ANYEXT]](s32)
475 %0:_(s32) = COPY $vgpr0
476 %1:_(s32) = COPY $vgpr1
477 %2:_(s32) = COPY $vgpr1
478 %3:_(s16) = G_TRUNC %0
479 %4:_(s16) = G_TRUNC %1
480 %5:_(s16) = G_TRUNC %2
481 %6:_(s16) = nnan G_FMAD %3, %4, %5
482 %7:_(s32) = G_ANYEXT %6
483 $vgpr0 = COPY %7
484 ...
486 ---
487 name: test_fmad_v2s16_denorm
488 machineFunctionInfo:
489 mode:
490 fp64-fp16-input-denormals: true
491 fp64-fp16-output-denormals: true
493 body: |
494 bb.0:
495 liveins: $vgpr0, $vgpr1, $vgpr2
497 ; GFX6-LABEL: name: test_fmad_v2s16_denorm
498 ; GFX6: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
499 ; GFX6: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
500 ; GFX6: [[COPY2:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr2
501 ; GFX6: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>)
502 ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
503 ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
504 ; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
505 ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
506 ; GFX6: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[COPY1]](<2 x s16>)
507 ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
508 ; GFX6: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
509 ; GFX6: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
510 ; GFX6: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[COPY2]](<2 x s16>)
511 ; GFX6: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
512 ; GFX6: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
513 ; GFX6: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
514 ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
515 ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
516 ; GFX6: [[FMUL:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT]], [[FPEXT1]]
517 ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL]](s32)
518 ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
519 ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC4]](s16)
520 ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT2]], [[FPEXT3]]
521 ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32)
522 ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
523 ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC3]](s16)
524 ; GFX6: [[FMUL1:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT4]], [[FPEXT5]]
525 ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL1]](s32)
526 ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16)
527 ; GFX6: [[FPEXT7:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC5]](s16)
528 ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT6]], [[FPEXT7]]
529 ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32)
530 ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC1]](s16)
531 ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC3]](s16)
532 ; GFX6: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32)
533 ; GFX6: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]]
534 ; GFX6: [[BITCAST3:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
535 ; GFX6: $vgpr0 = COPY [[BITCAST3]](<2 x s16>)
536 ; GFX7-LABEL: name: test_fmad_v2s16_denorm
537 ; GFX7: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
538 ; GFX7: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
539 ; GFX7: [[COPY2:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr2
540 ; GFX7: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>)
541 ; GFX7: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
542 ; GFX7: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
543 ; GFX7: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
544 ; GFX7: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
545 ; GFX7: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[COPY1]](<2 x s16>)
546 ; GFX7: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
547 ; GFX7: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
548 ; GFX7: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
549 ; GFX7: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[COPY2]](<2 x s16>)
550 ; GFX7: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
551 ; GFX7: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
552 ; GFX7: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
553 ; GFX7: [[FMUL:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC]], [[TRUNC2]]
554 ; GFX7: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[FMUL]], [[TRUNC4]]
555 ; GFX7: [[FMUL1:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC1]], [[TRUNC3]]
556 ; GFX7: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[FMUL1]], [[TRUNC5]]
557 ; GFX7: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FADD]](s16)
558 ; GFX7: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FADD1]](s16)
559 ; GFX7: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32)
560 ; GFX7: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]]
561 ; GFX7: [[BITCAST3:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
562 ; GFX7: $vgpr0 = COPY [[BITCAST3]](<2 x s16>)
563 ; GFX10-LABEL: name: test_fmad_v2s16_denorm
564 ; GFX10: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
565 ; GFX10: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
566 ; GFX10: [[COPY2:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr2
567 ; GFX10: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>)
568 ; GFX10: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
569 ; GFX10: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
570 ; GFX10: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
571 ; GFX10: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
572 ; GFX10: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[COPY1]](<2 x s16>)
573 ; GFX10: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
574 ; GFX10: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
575 ; GFX10: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
576 ; GFX10: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[COPY2]](<2 x s16>)
577 ; GFX10: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
578 ; GFX10: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
579 ; GFX10: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
580 ; GFX10: [[FMUL:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC]], [[TRUNC2]]
581 ; GFX10: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[FMUL]], [[TRUNC4]]
582 ; GFX10: [[FMUL1:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC1]], [[TRUNC3]]
583 ; GFX10: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[FMUL1]], [[TRUNC5]]
584 ; GFX10: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16)
585 ; GFX10: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16)
586 ; GFX10: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32)
587 ; GFX10: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>)
588 %0:_(<2 x s16>) = COPY $vgpr0
589 %1:_(<2 x s16>) = COPY $vgpr1
590 %2:_(<2 x s16>) = COPY $vgpr2
591 %3:_(<2 x s16>) = G_FMAD %0, %1, %2
592 $vgpr0 = COPY %3
593 ...
595 ---
596 name: test_fmad_v2s16_denorm_flags
597 machineFunctionInfo:
598 mode:
599 fp64-fp16-input-denormals: true
600 fp64-fp16-output-denormals: true
602 body: |
603 bb.0:
604 liveins: $vgpr0, $vgpr1, $vgpr2
606 ; GFX6-LABEL: name: test_fmad_v2s16_denorm_flags
607 ; GFX6: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
608 ; GFX6: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
609 ; GFX6: [[COPY2:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr2
610 ; GFX6: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>)
611 ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
612 ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
613 ; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
614 ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
615 ; GFX6: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[COPY1]](<2 x s16>)
616 ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
617 ; GFX6: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
618 ; GFX6: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
619 ; GFX6: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[COPY2]](<2 x s16>)
620 ; GFX6: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
621 ; GFX6: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
622 ; GFX6: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
623 ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
624 ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
625 ; GFX6: [[FMUL:%[0-9]+]]:_(s32) = nnan G_FMUL [[FPEXT]], [[FPEXT1]]
626 ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL]](s32)
627 ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
628 ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC4]](s16)
629 ; GFX6: [[FADD:%[0-9]+]]:_(s32) = nnan G_FADD [[FPEXT2]], [[FPEXT3]]
630 ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32)
631 ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
632 ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC3]](s16)
633 ; GFX6: [[FMUL1:%[0-9]+]]:_(s32) = nnan G_FMUL [[FPEXT4]], [[FPEXT5]]
634 ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL1]](s32)
635 ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16)
636 ; GFX6: [[FPEXT7:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC5]](s16)
637 ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = nnan G_FADD [[FPEXT6]], [[FPEXT7]]
638 ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32)
639 ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC1]](s16)
640 ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC3]](s16)
641 ; GFX6: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32)
642 ; GFX6: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]]
643 ; GFX6: [[BITCAST3:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
644 ; GFX6: $vgpr0 = COPY [[BITCAST3]](<2 x s16>)
645 ; GFX7-LABEL: name: test_fmad_v2s16_denorm_flags
646 ; GFX7: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
647 ; GFX7: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
648 ; GFX7: [[COPY2:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr2
649 ; GFX7: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>)
650 ; GFX7: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
651 ; GFX7: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
652 ; GFX7: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
653 ; GFX7: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
654 ; GFX7: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[COPY1]](<2 x s16>)
655 ; GFX7: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
656 ; GFX7: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
657 ; GFX7: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
658 ; GFX7: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[COPY2]](<2 x s16>)
659 ; GFX7: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
660 ; GFX7: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
661 ; GFX7: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
662 ; GFX7: [[FMUL:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC]], [[TRUNC2]]
663 ; GFX7: [[FADD:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL]], [[TRUNC4]]
664 ; GFX7: [[FMUL1:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC1]], [[TRUNC3]]
665 ; GFX7: [[FADD1:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL1]], [[TRUNC5]]
666 ; GFX7: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FADD]](s16)
667 ; GFX7: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FADD1]](s16)
668 ; GFX7: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32)
669 ; GFX7: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]]
670 ; GFX7: [[BITCAST3:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
671 ; GFX7: $vgpr0 = COPY [[BITCAST3]](<2 x s16>)
672 ; GFX10-LABEL: name: test_fmad_v2s16_denorm_flags
673 ; GFX10: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr0
674 ; GFX10: [[COPY1:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr1
675 ; GFX10: [[COPY2:%[0-9]+]]:_(<2 x s16>) = COPY $vgpr2
676 ; GFX10: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[COPY]](<2 x s16>)
677 ; GFX10: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
678 ; GFX10: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
679 ; GFX10: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
680 ; GFX10: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
681 ; GFX10: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[COPY1]](<2 x s16>)
682 ; GFX10: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
683 ; GFX10: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
684 ; GFX10: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
685 ; GFX10: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[COPY2]](<2 x s16>)
686 ; GFX10: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
687 ; GFX10: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
688 ; GFX10: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
689 ; GFX10: [[FMUL:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC]], [[TRUNC2]]
690 ; GFX10: [[FADD:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL]], [[TRUNC4]]
691 ; GFX10: [[FMUL1:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC1]], [[TRUNC3]]
692 ; GFX10: [[FADD1:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL1]], [[TRUNC5]]
693 ; GFX10: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16)
694 ; GFX10: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16)
695 ; GFX10: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32)
696 ; GFX10: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>)
697 %0:_(<2 x s16>) = COPY $vgpr0
698 %1:_(<2 x s16>) = COPY $vgpr1
699 %2:_(<2 x s16>) = COPY $vgpr2
700 %3:_(<2 x s16>) = nnan G_FMAD %0, %1, %2
701 $vgpr0 = COPY %3
702 ...
704 ---
705 name: test_fmad_v4s16_denorm
706 machineFunctionInfo:
707 mode:
708 fp64-fp16-input-denormals: true
709 fp64-fp16-output-denormals: true
711 body: |
712 bb.0:
713 liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $vgpr4_vgpr5
715 ; GFX6-LABEL: name: test_fmad_v4s16_denorm
716 ; GFX6: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
717 ; GFX6: [[COPY1:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr2_vgpr3
718 ; GFX6: [[COPY2:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr4_vgpr5
719 ; GFX6: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
720 ; GFX6: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
721 ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
722 ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
723 ; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
724 ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
725 ; GFX6: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
726 ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
727 ; GFX6: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
728 ; GFX6: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
729 ; GFX6: [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY1]](<4 x s16>)
730 ; GFX6: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
731 ; GFX6: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
732 ; GFX6: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
733 ; GFX6: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
734 ; GFX6: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
735 ; GFX6: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST3]](s32)
736 ; GFX6: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
737 ; GFX6: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR3]](s32)
738 ; GFX6: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
739 ; GFX6: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
740 ; GFX6: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST4]](s32)
741 ; GFX6: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
742 ; GFX6: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR4]](s32)
743 ; GFX6: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV5]](<2 x s16>)
744 ; GFX6: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST5]](s32)
745 ; GFX6: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
746 ; GFX6: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR5]](s32)
747 ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
748 ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC4]](s16)
749 ; GFX6: [[FMUL:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT]], [[FPEXT1]]
750 ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL]](s32)
751 ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
752 ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC8]](s16)
753 ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT2]], [[FPEXT3]]
754 ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32)
755 ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
756 ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC5]](s16)
757 ; GFX6: [[FMUL1:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT4]], [[FPEXT5]]
758 ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL1]](s32)
759 ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16)
760 ; GFX6: [[FPEXT7:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC9]](s16)
761 ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT6]], [[FPEXT7]]
762 ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32)
763 ; GFX6: [[FPEXT8:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
764 ; GFX6: [[FPEXT9:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC6]](s16)
765 ; GFX6: [[FMUL2:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT8]], [[FPEXT9]]
766 ; GFX6: [[FPTRUNC4:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL2]](s32)
767 ; GFX6: [[FPEXT10:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC4]](s16)
768 ; GFX6: [[FPEXT11:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC10]](s16)
769 ; GFX6: [[FADD2:%[0-9]+]]:_(s32) = G_FADD [[FPEXT10]], [[FPEXT11]]
770 ; GFX6: [[FPTRUNC5:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD2]](s32)
771 ; GFX6: [[FPEXT12:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC3]](s16)
772 ; GFX6: [[FPEXT13:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC7]](s16)
773 ; GFX6: [[FMUL3:%[0-9]+]]:_(s32) = G_FMUL [[FPEXT12]], [[FPEXT13]]
774 ; GFX6: [[FPTRUNC6:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL3]](s32)
775 ; GFX6: [[FPEXT14:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC6]](s16)
776 ; GFX6: [[FPEXT15:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC11]](s16)
777 ; GFX6: [[FADD3:%[0-9]+]]:_(s32) = G_FADD [[FPEXT14]], [[FPEXT15]]
778 ; GFX6: [[FPTRUNC7:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD3]](s32)
779 ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC1]](s16)
780 ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC3]](s16)
781 ; GFX6: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32)
782 ; GFX6: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]]
783 ; GFX6: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
784 ; GFX6: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC5]](s16)
785 ; GFX6: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC7]](s16)
786 ; GFX6: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C]](s32)
787 ; GFX6: [[OR1:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL1]]
788 ; GFX6: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
789 ; GFX6: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST6]](<2 x s16>), [[BITCAST7]](<2 x s16>)
790 ; GFX6: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
791 ; GFX7-LABEL: name: test_fmad_v4s16_denorm
792 ; GFX7: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
793 ; GFX7: [[COPY1:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr2_vgpr3
794 ; GFX7: [[COPY2:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr4_vgpr5
795 ; GFX7: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
796 ; GFX7: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
797 ; GFX7: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
798 ; GFX7: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
799 ; GFX7: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
800 ; GFX7: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
801 ; GFX7: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
802 ; GFX7: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
803 ; GFX7: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
804 ; GFX7: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
805 ; GFX7: [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY1]](<4 x s16>)
806 ; GFX7: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
807 ; GFX7: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
808 ; GFX7: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
809 ; GFX7: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
810 ; GFX7: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
811 ; GFX7: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST3]](s32)
812 ; GFX7: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
813 ; GFX7: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR3]](s32)
814 ; GFX7: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
815 ; GFX7: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
816 ; GFX7: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST4]](s32)
817 ; GFX7: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
818 ; GFX7: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR4]](s32)
819 ; GFX7: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV5]](<2 x s16>)
820 ; GFX7: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST5]](s32)
821 ; GFX7: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
822 ; GFX7: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR5]](s32)
823 ; GFX7: [[FMUL:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC]], [[TRUNC4]]
824 ; GFX7: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[FMUL]], [[TRUNC8]]
825 ; GFX7: [[FMUL1:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC1]], [[TRUNC5]]
826 ; GFX7: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[FMUL1]], [[TRUNC9]]
827 ; GFX7: [[FMUL2:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC2]], [[TRUNC6]]
828 ; GFX7: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[FMUL2]], [[TRUNC10]]
829 ; GFX7: [[FMUL3:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC3]], [[TRUNC7]]
830 ; GFX7: [[FADD3:%[0-9]+]]:_(s16) = G_FADD [[FMUL3]], [[TRUNC11]]
831 ; GFX7: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FADD]](s16)
832 ; GFX7: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FADD1]](s16)
833 ; GFX7: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32)
834 ; GFX7: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]]
835 ; GFX7: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
836 ; GFX7: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FADD2]](s16)
837 ; GFX7: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[FADD3]](s16)
838 ; GFX7: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C]](s32)
839 ; GFX7: [[OR1:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL1]]
840 ; GFX7: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
841 ; GFX7: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST6]](<2 x s16>), [[BITCAST7]](<2 x s16>)
842 ; GFX7: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
843 ; GFX10-LABEL: name: test_fmad_v4s16_denorm
844 ; GFX10: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
845 ; GFX10: [[COPY1:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr2_vgpr3
846 ; GFX10: [[COPY2:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr4_vgpr5
847 ; GFX10: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
848 ; GFX10: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
849 ; GFX10: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
850 ; GFX10: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
851 ; GFX10: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
852 ; GFX10: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
853 ; GFX10: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
854 ; GFX10: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
855 ; GFX10: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
856 ; GFX10: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
857 ; GFX10: [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY1]](<4 x s16>)
858 ; GFX10: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
859 ; GFX10: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
860 ; GFX10: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
861 ; GFX10: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
862 ; GFX10: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
863 ; GFX10: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST3]](s32)
864 ; GFX10: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
865 ; GFX10: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR3]](s32)
866 ; GFX10: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
867 ; GFX10: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
868 ; GFX10: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST4]](s32)
869 ; GFX10: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
870 ; GFX10: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR4]](s32)
871 ; GFX10: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV5]](<2 x s16>)
872 ; GFX10: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST5]](s32)
873 ; GFX10: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
874 ; GFX10: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR5]](s32)
875 ; GFX10: [[FMUL:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC]], [[TRUNC4]]
876 ; GFX10: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[FMUL]], [[TRUNC8]]
877 ; GFX10: [[FMUL1:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC1]], [[TRUNC5]]
878 ; GFX10: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[FMUL1]], [[TRUNC9]]
879 ; GFX10: [[FMUL2:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC2]], [[TRUNC6]]
880 ; GFX10: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[FMUL2]], [[TRUNC10]]
881 ; GFX10: [[FMUL3:%[0-9]+]]:_(s16) = G_FMUL [[TRUNC3]], [[TRUNC7]]
882 ; GFX10: [[FADD3:%[0-9]+]]:_(s16) = G_FADD [[FMUL3]], [[TRUNC11]]
883 ; GFX10: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16)
884 ; GFX10: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16)
885 ; GFX10: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32)
886 ; GFX10: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD2]](s16)
887 ; GFX10: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD3]](s16)
888 ; GFX10: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT2]](s32), [[ANYEXT3]](s32)
889 ; GFX10: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
890 ; GFX10: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
891 %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
892 %1:_(<4 x s16>) = COPY $vgpr2_vgpr3
893 %2:_(<4 x s16>) = COPY $vgpr4_vgpr5
894 %3:_(<4 x s16>) = G_FMAD %0, %1, %2
895 $vgpr0_vgpr1 = COPY %3
896 ...
899 ---
900 name: test_fmad_v4s16_denorm_flags
901 machineFunctionInfo:
902 mode:
903 fp64-fp16-input-denormals: true
904 fp64-fp16-output-denormals: true
906 body: |
907 bb.0:
908 liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $vgpr4_vgpr5
910 ; GFX6-LABEL: name: test_fmad_v4s16_denorm_flags
911 ; GFX6: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
912 ; GFX6: [[COPY1:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr2_vgpr3
913 ; GFX6: [[COPY2:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr4_vgpr5
914 ; GFX6: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
915 ; GFX6: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
916 ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
917 ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
918 ; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
919 ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
920 ; GFX6: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
921 ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
922 ; GFX6: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
923 ; GFX6: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
924 ; GFX6: [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY1]](<4 x s16>)
925 ; GFX6: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
926 ; GFX6: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
927 ; GFX6: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
928 ; GFX6: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
929 ; GFX6: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
930 ; GFX6: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST3]](s32)
931 ; GFX6: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
932 ; GFX6: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR3]](s32)
933 ; GFX6: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
934 ; GFX6: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
935 ; GFX6: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST4]](s32)
936 ; GFX6: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
937 ; GFX6: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR4]](s32)
938 ; GFX6: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV5]](<2 x s16>)
939 ; GFX6: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST5]](s32)
940 ; GFX6: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
941 ; GFX6: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR5]](s32)
942 ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16)
943 ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC4]](s16)
944 ; GFX6: [[FMUL:%[0-9]+]]:_(s32) = nnan G_FMUL [[FPEXT]], [[FPEXT1]]
945 ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL]](s32)
946 ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16)
947 ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC8]](s16)
948 ; GFX6: [[FADD:%[0-9]+]]:_(s32) = nnan G_FADD [[FPEXT2]], [[FPEXT3]]
949 ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32)
950 ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16)
951 ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC5]](s16)
952 ; GFX6: [[FMUL1:%[0-9]+]]:_(s32) = nnan G_FMUL [[FPEXT4]], [[FPEXT5]]
953 ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL1]](s32)
954 ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16)
955 ; GFX6: [[FPEXT7:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC9]](s16)
956 ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = nnan G_FADD [[FPEXT6]], [[FPEXT7]]
957 ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32)
958 ; GFX6: [[FPEXT8:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16)
959 ; GFX6: [[FPEXT9:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC6]](s16)
960 ; GFX6: [[FMUL2:%[0-9]+]]:_(s32) = nnan G_FMUL [[FPEXT8]], [[FPEXT9]]
961 ; GFX6: [[FPTRUNC4:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL2]](s32)
962 ; GFX6: [[FPEXT10:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC4]](s16)
963 ; GFX6: [[FPEXT11:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC10]](s16)
964 ; GFX6: [[FADD2:%[0-9]+]]:_(s32) = nnan G_FADD [[FPEXT10]], [[FPEXT11]]
965 ; GFX6: [[FPTRUNC5:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD2]](s32)
966 ; GFX6: [[FPEXT12:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC3]](s16)
967 ; GFX6: [[FPEXT13:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC7]](s16)
968 ; GFX6: [[FMUL3:%[0-9]+]]:_(s32) = nnan G_FMUL [[FPEXT12]], [[FPEXT13]]
969 ; GFX6: [[FPTRUNC6:%[0-9]+]]:_(s16) = G_FPTRUNC [[FMUL3]](s32)
970 ; GFX6: [[FPEXT14:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC6]](s16)
971 ; GFX6: [[FPEXT15:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC11]](s16)
972 ; GFX6: [[FADD3:%[0-9]+]]:_(s32) = nnan G_FADD [[FPEXT14]], [[FPEXT15]]
973 ; GFX6: [[FPTRUNC7:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD3]](s32)
974 ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC1]](s16)
975 ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC3]](s16)
976 ; GFX6: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32)
977 ; GFX6: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]]
978 ; GFX6: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
979 ; GFX6: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC5]](s16)
980 ; GFX6: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC7]](s16)
981 ; GFX6: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C]](s32)
982 ; GFX6: [[OR1:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL1]]
983 ; GFX6: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
984 ; GFX6: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST6]](<2 x s16>), [[BITCAST7]](<2 x s16>)
985 ; GFX6: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
986 ; GFX7-LABEL: name: test_fmad_v4s16_denorm_flags
987 ; GFX7: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
988 ; GFX7: [[COPY1:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr2_vgpr3
989 ; GFX7: [[COPY2:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr4_vgpr5
990 ; GFX7: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
991 ; GFX7: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
992 ; GFX7: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
993 ; GFX7: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
994 ; GFX7: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
995 ; GFX7: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
996 ; GFX7: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
997 ; GFX7: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
998 ; GFX7: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
999 ; GFX7: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
1000 ; GFX7: [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY1]](<4 x s16>)
1001 ; GFX7: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
1002 ; GFX7: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
1003 ; GFX7: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
1004 ; GFX7: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
1005 ; GFX7: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
1006 ; GFX7: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST3]](s32)
1007 ; GFX7: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
1008 ; GFX7: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR3]](s32)
1009 ; GFX7: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
1010 ; GFX7: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
1011 ; GFX7: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST4]](s32)
1012 ; GFX7: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
1013 ; GFX7: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR4]](s32)
1014 ; GFX7: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV5]](<2 x s16>)
1015 ; GFX7: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST5]](s32)
1016 ; GFX7: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
1017 ; GFX7: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR5]](s32)
1018 ; GFX7: [[FMUL:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC]], [[TRUNC4]]
1019 ; GFX7: [[FADD:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL]], [[TRUNC8]]
1020 ; GFX7: [[FMUL1:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC1]], [[TRUNC5]]
1021 ; GFX7: [[FADD1:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL1]], [[TRUNC9]]
1022 ; GFX7: [[FMUL2:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC2]], [[TRUNC6]]
1023 ; GFX7: [[FADD2:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL2]], [[TRUNC10]]
1024 ; GFX7: [[FMUL3:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC3]], [[TRUNC7]]
1025 ; GFX7: [[FADD3:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL3]], [[TRUNC11]]
1026 ; GFX7: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FADD]](s16)
1027 ; GFX7: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FADD1]](s16)
1028 ; GFX7: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32)
1029 ; GFX7: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]]
1030 ; GFX7: [[BITCAST6:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32)
1031 ; GFX7: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FADD2]](s16)
1032 ; GFX7: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[FADD3]](s16)
1033 ; GFX7: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C]](s32)
1034 ; GFX7: [[OR1:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL1]]
1035 ; GFX7: [[BITCAST7:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32)
1036 ; GFX7: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST6]](<2 x s16>), [[BITCAST7]](<2 x s16>)
1037 ; GFX7: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
1038 ; GFX10-LABEL: name: test_fmad_v4s16_denorm_flags
1039 ; GFX10: [[COPY:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr0_vgpr1
1040 ; GFX10: [[COPY1:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr2_vgpr3
1041 ; GFX10: [[COPY2:%[0-9]+]]:_(<4 x s16>) = COPY $vgpr4_vgpr5
1042 ; GFX10: [[UV:%[0-9]+]]:_(<2 x s16>), [[UV1:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY]](<4 x s16>)
1043 ; GFX10: [[BITCAST:%[0-9]+]]:_(s32) = G_BITCAST [[UV]](<2 x s16>)
1044 ; GFX10: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST]](s32)
1045 ; GFX10: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
1046 ; GFX10: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32)
1047 ; GFX10: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
1048 ; GFX10: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>)
1049 ; GFX10: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32)
1050 ; GFX10: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32)
1051 ; GFX10: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32)
1052 ; GFX10: [[UV2:%[0-9]+]]:_(<2 x s16>), [[UV3:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY1]](<4 x s16>)
1053 ; GFX10: [[BITCAST2:%[0-9]+]]:_(s32) = G_BITCAST [[UV2]](<2 x s16>)
1054 ; GFX10: [[TRUNC4:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST2]](s32)
1055 ; GFX10: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST2]], [[C]](s32)
1056 ; GFX10: [[TRUNC5:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR2]](s32)
1057 ; GFX10: [[BITCAST3:%[0-9]+]]:_(s32) = G_BITCAST [[UV3]](<2 x s16>)
1058 ; GFX10: [[TRUNC6:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST3]](s32)
1059 ; GFX10: [[LSHR3:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST3]], [[C]](s32)
1060 ; GFX10: [[TRUNC7:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR3]](s32)
1061 ; GFX10: [[UV4:%[0-9]+]]:_(<2 x s16>), [[UV5:%[0-9]+]]:_(<2 x s16>) = G_UNMERGE_VALUES [[COPY2]](<4 x s16>)
1062 ; GFX10: [[BITCAST4:%[0-9]+]]:_(s32) = G_BITCAST [[UV4]](<2 x s16>)
1063 ; GFX10: [[TRUNC8:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST4]](s32)
1064 ; GFX10: [[LSHR4:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST4]], [[C]](s32)
1065 ; GFX10: [[TRUNC9:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR4]](s32)
1066 ; GFX10: [[BITCAST5:%[0-9]+]]:_(s32) = G_BITCAST [[UV5]](<2 x s16>)
1067 ; GFX10: [[TRUNC10:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST5]](s32)
1068 ; GFX10: [[LSHR5:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST5]], [[C]](s32)
1069 ; GFX10: [[TRUNC11:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR5]](s32)
1070 ; GFX10: [[FMUL:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC]], [[TRUNC4]]
1071 ; GFX10: [[FADD:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL]], [[TRUNC8]]
1072 ; GFX10: [[FMUL1:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC1]], [[TRUNC5]]
1073 ; GFX10: [[FADD1:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL1]], [[TRUNC9]]
1074 ; GFX10: [[FMUL2:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC2]], [[TRUNC6]]
1075 ; GFX10: [[FADD2:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL2]], [[TRUNC10]]
1076 ; GFX10: [[FMUL3:%[0-9]+]]:_(s16) = nnan G_FMUL [[TRUNC3]], [[TRUNC7]]
1077 ; GFX10: [[FADD3:%[0-9]+]]:_(s16) = nnan G_FADD [[FMUL3]], [[TRUNC11]]
1078 ; GFX10: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16)
1079 ; GFX10: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16)
1080 ; GFX10: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32)
1081 ; GFX10: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD2]](s16)
1082 ; GFX10: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD3]](s16)
1083 ; GFX10: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT2]](s32), [[ANYEXT3]](s32)
1084 ; GFX10: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>)
1085 ; GFX10: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>)
1086 %0:_(<4 x s16>) = COPY $vgpr0_vgpr1
1087 %1:_(<4 x s16>) = COPY $vgpr2_vgpr3
1088 %2:_(<4 x s16>) = COPY $vgpr4_vgpr5
1089 %3:_(<4 x s16>) = nnan G_FMAD %0, %1, %2
1090 $vgpr0_vgpr1 = COPY %3
1091 ...