| blob | d798166a67839663788a1a8108aa17d34ba91ecc |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -global-isel=0 -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9,GFX9-SDAG %s
3 ; RUN: llc -global-isel=1 -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9,GFX9-GISEL %s
5 ; RUN: llc -global-isel=0 -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji < %s | FileCheck -check-prefixes=GCN,GFX8,GFX8-SDAG %s
6 ; RUN: llc -global-isel=1 -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji < %s | FileCheck -check-prefixes=GCN,GFX8,GFX8-GISEL %s
8 ; RUN: llc -global-isel=0 -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX10,GFX10-SDAG %s
9 ; RUN: llc -global-isel=1 -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX10,GFX10-GISEL %s
11 ; RUN: llc -global-isel=0 -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX11,GFX11-SDAG %s
12 ; RUN: llc -global-isel=1 -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 < %s | FileCheck -check-prefixes=GFX10PLUS,GFX11,GFX1-GISEL %s
15 ; FIXME: promotion not handled without f16 insts
17 define half @v_constained_fmul_f16_fpexcept_strict(half %x, half %y) #0 {
18 ; GCN-LABEL: v_constained_fmul_f16_fpexcept_strict:
19 ; GCN: ; %bb.0:
20 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
21 ; GCN-NEXT: v_mul_f16_e32 v0, v0, v1
22 ; GCN-NEXT: s_setpc_b64 s[30:31]
23 ;
24 ; GFX10PLUS-LABEL: v_constained_fmul_f16_fpexcept_strict:
25 ; GFX10PLUS: ; %bb.0:
26 ; GFX10PLUS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
27 ; GFX10PLUS-NEXT: v_mul_f16_e32 v0, v0, v1
28 ; GFX10PLUS-NEXT: s_setpc_b64 s[30:31]
29 %val = call half @llvm.experimental.constrained.fmul.f16(half %x, half %y, metadata !"round.tonearest", metadata !"fpexcept.strict")
30 ret half %val
31 }
33 define half @v_constained_fmul_f16_fpexcept_ignore(half %x, half %y) #0 {
34 ; GCN-LABEL: v_constained_fmul_f16_fpexcept_ignore:
35 ; GCN: ; %bb.0:
36 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
37 ; GCN-NEXT: v_mul_f16_e32 v0, v0, v1
38 ; GCN-NEXT: s_setpc_b64 s[30:31]
39 ;
40 ; GFX10PLUS-LABEL: v_constained_fmul_f16_fpexcept_ignore:
41 ; GFX10PLUS: ; %bb.0:
42 ; GFX10PLUS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
43 ; GFX10PLUS-NEXT: v_mul_f16_e32 v0, v0, v1
44 ; GFX10PLUS-NEXT: s_setpc_b64 s[30:31]
45 %val = call half @llvm.experimental.constrained.fmul.f16(half %x, half %y, metadata !"round.tonearest", metadata !"fpexcept.ignore")
46 ret half %val
47 }
49 define half @v_constained_fmul_f16_fpexcept_maytrap(half %x, half %y) #0 {
50 ; GCN-LABEL: v_constained_fmul_f16_fpexcept_maytrap:
51 ; GCN: ; %bb.0:
52 ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
53 ; GCN-NEXT: v_mul_f16_e32 v0, v0, v1
54 ; GCN-NEXT: s_setpc_b64 s[30:31]
55 ;
56 ; GFX10PLUS-LABEL: v_constained_fmul_f16_fpexcept_maytrap:
57 ; GFX10PLUS: ; %bb.0:
58 ; GFX10PLUS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
59 ; GFX10PLUS-NEXT: v_mul_f16_e32 v0, v0, v1
60 ; GFX10PLUS-NEXT: s_setpc_b64 s[30:31]
61 %val = call half @llvm.experimental.constrained.fmul.f16(half %x, half %y, metadata !"round.tonearest", metadata !"fpexcept.maytrap")
62 ret half %val
63 }
65 define <2 x half> @v_constained_fmul_v2f16_fpexcept_strict(<2 x half> %x, <2 x half> %y) #0 {
66 ; GFX9-LABEL: v_constained_fmul_v2f16_fpexcept_strict:
67 ; GFX9: ; %bb.0:
68 ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
69 ; GFX9-NEXT: v_pk_mul_f16 v0, v0, v1
70 ; GFX9-NEXT: s_setpc_b64 s[30:31]
71 ;
72 ; GFX8-SDAG-LABEL: v_constained_fmul_v2f16_fpexcept_strict:
73 ; GFX8-SDAG: ; %bb.0:
74 ; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
75 ; GFX8-SDAG-NEXT: v_mul_f16_sdwa v2, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
76 ; GFX8-SDAG-NEXT: v_mul_f16_e32 v0, v0, v1
77 ; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v0, v2
78 ; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
79 ;
80 ; GFX8-GISEL-LABEL: v_constained_fmul_v2f16_fpexcept_strict:
81 ; GFX8-GISEL: ; %bb.0:
82 ; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
83 ; GFX8-GISEL-NEXT: v_mul_f16_e32 v2, v0, v1
84 ; GFX8-GISEL-NEXT: v_mul_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
85 ; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v2, v0
86 ; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
87 ;
88 ; GFX10PLUS-LABEL: v_constained_fmul_v2f16_fpexcept_strict:
89 ; GFX10PLUS: ; %bb.0:
90 ; GFX10PLUS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
91 ; GFX10PLUS-NEXT: v_pk_mul_f16 v0, v0, v1
92 ; GFX10PLUS-NEXT: s_setpc_b64 s[30:31]
93 %val = call <2 x half> @llvm.experimental.constrained.fmul.v2f16(<2 x half> %x, <2 x half> %y, metadata !"round.tonearest", metadata !"fpexcept.strict")
94 ret <2 x half> %val
95 }
97 define <2 x half> @v_constained_fmul_v2f16_fpexcept_ignore(<2 x half> %x, <2 x half> %y) #0 {
98 ; GFX9-LABEL: v_constained_fmul_v2f16_fpexcept_ignore:
99 ; GFX9: ; %bb.0:
100 ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
101 ; GFX9-NEXT: v_pk_mul_f16 v0, v0, v1
102 ; GFX9-NEXT: s_setpc_b64 s[30:31]
103 ;
104 ; GFX8-SDAG-LABEL: v_constained_fmul_v2f16_fpexcept_ignore:
105 ; GFX8-SDAG: ; %bb.0:
106 ; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
107 ; GFX8-SDAG-NEXT: v_mul_f16_sdwa v2, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
108 ; GFX8-SDAG-NEXT: v_mul_f16_e32 v0, v0, v1
109 ; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v0, v2
110 ; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
111 ;
112 ; GFX8-GISEL-LABEL: v_constained_fmul_v2f16_fpexcept_ignore:
113 ; GFX8-GISEL: ; %bb.0:
114 ; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
115 ; GFX8-GISEL-NEXT: v_mul_f16_e32 v2, v0, v1
116 ; GFX8-GISEL-NEXT: v_mul_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
117 ; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v2, v0
118 ; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
119 ;
120 ; GFX10PLUS-LABEL: v_constained_fmul_v2f16_fpexcept_ignore:
121 ; GFX10PLUS: ; %bb.0:
122 ; GFX10PLUS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
123 ; GFX10PLUS-NEXT: v_pk_mul_f16 v0, v0, v1
124 ; GFX10PLUS-NEXT: s_setpc_b64 s[30:31]
125 %val = call <2 x half> @llvm.experimental.constrained.fmul.v2f16(<2 x half> %x, <2 x half> %y, metadata !"round.tonearest", metadata !"fpexcept.ignore")
126 ret <2 x half> %val
127 }
129 define <2 x half> @v_constained_fmul_v2f16_fpexcept_maytrap(<2 x half> %x, <2 x half> %y) #0 {
130 ; GFX9-LABEL: v_constained_fmul_v2f16_fpexcept_maytrap:
131 ; GFX9: ; %bb.0:
132 ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
133 ; GFX9-NEXT: v_pk_mul_f16 v0, v0, v1
134 ; GFX9-NEXT: s_setpc_b64 s[30:31]
135 ;
136 ; GFX8-SDAG-LABEL: v_constained_fmul_v2f16_fpexcept_maytrap:
137 ; GFX8-SDAG: ; %bb.0:
138 ; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
139 ; GFX8-SDAG-NEXT: v_mul_f16_sdwa v2, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
140 ; GFX8-SDAG-NEXT: v_mul_f16_e32 v0, v0, v1
141 ; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v0, v2
142 ; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
143 ;
144 ; GFX8-GISEL-LABEL: v_constained_fmul_v2f16_fpexcept_maytrap:
145 ; GFX8-GISEL: ; %bb.0:
146 ; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
147 ; GFX8-GISEL-NEXT: v_mul_f16_e32 v2, v0, v1
148 ; GFX8-GISEL-NEXT: v_mul_f16_sdwa v0, v0, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
149 ; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v2, v0
150 ; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
151 ;
152 ; GFX10PLUS-LABEL: v_constained_fmul_v2f16_fpexcept_maytrap:
153 ; GFX10PLUS: ; %bb.0:
154 ; GFX10PLUS-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
155 ; GFX10PLUS-NEXT: v_pk_mul_f16 v0, v0, v1
156 ; GFX10PLUS-NEXT: s_setpc_b64 s[30:31]
157 %val = call <2 x half> @llvm.experimental.constrained.fmul.v2f16(<2 x half> %x, <2 x half> %y, metadata !"round.tonearest", metadata !"fpexcept.maytrap")
158 ret <2 x half> %val
159 }
161 define <3 x half> @v_constained_fmul_v3f16_fpexcept_strict(<3 x half> %x, <3 x half> %y) #0 {
162 ; GFX9-SDAG-LABEL: v_constained_fmul_v3f16_fpexcept_strict:
163 ; GFX9-SDAG: ; %bb.0:
164 ; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
165 ; GFX9-SDAG-NEXT: v_pk_mul_f16 v0, v0, v2
166 ; GFX9-SDAG-NEXT: v_mul_f16_e32 v1, v1, v3
167 ; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
168 ;
169 ; GFX9-GISEL-LABEL: v_constained_fmul_v3f16_fpexcept_strict:
170 ; GFX9-GISEL: ; %bb.0:
171 ; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
172 ; GFX9-GISEL-NEXT: v_pk_mul_f16 v0, v0, v2
173 ; GFX9-GISEL-NEXT: v_pk_mul_f16 v1, v1, v3
174 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
175 ;
176 ; GFX8-SDAG-LABEL: v_constained_fmul_v3f16_fpexcept_strict:
177 ; GFX8-SDAG: ; %bb.0:
178 ; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
179 ; GFX8-SDAG-NEXT: v_mul_f16_sdwa v4, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
180 ; GFX8-SDAG-NEXT: v_mul_f16_e32 v0, v0, v2
181 ; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v0, v4
182 ; GFX8-SDAG-NEXT: v_mul_f16_e32 v1, v1, v3
183 ; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
184 ;
185 ; GFX8-GISEL-LABEL: v_constained_fmul_v3f16_fpexcept_strict:
186 ; GFX8-GISEL: ; %bb.0:
187 ; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
188 ; GFX8-GISEL-NEXT: v_mul_f16_e32 v4, v0, v2
189 ; GFX8-GISEL-NEXT: v_mul_f16_sdwa v0, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
190 ; GFX8-GISEL-NEXT: v_mul_f16_e32 v1, v1, v3
191 ; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v4, v0
192 ; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
193 ;
194 ; GFX10-SDAG-LABEL: v_constained_fmul_v3f16_fpexcept_strict:
195 ; GFX10-SDAG: ; %bb.0:
196 ; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
197 ; GFX10-SDAG-NEXT: v_pk_mul_f16 v0, v0, v2
198 ; GFX10-SDAG-NEXT: v_mul_f16_e32 v1, v1, v3
199 ; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
200 ;
201 ; GFX10-GISEL-LABEL: v_constained_fmul_v3f16_fpexcept_strict:
202 ; GFX10-GISEL: ; %bb.0:
203 ; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
204 ; GFX10-GISEL-NEXT: v_pk_mul_f16 v0, v0, v2
205 ; GFX10-GISEL-NEXT: v_pk_mul_f16 v1, v1, v3
206 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
207 ;
208 ; GFX11-SDAG-LABEL: v_constained_fmul_v3f16_fpexcept_strict:
209 ; GFX11-SDAG: ; %bb.0:
210 ; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
211 ; GFX11-SDAG-NEXT: v_pk_mul_f16 v0, v0, v2
212 ; GFX11-SDAG-NEXT: v_mul_f16_e32 v1, v1, v3
213 ; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
214 ;
215 ; GFX1-GISEL-LABEL: v_constained_fmul_v3f16_fpexcept_strict:
216 ; GFX1-GISEL: ; %bb.0:
217 ; GFX1-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
218 ; GFX1-GISEL-NEXT: v_pk_mul_f16 v0, v0, v2
219 ; GFX1-GISEL-NEXT: v_pk_mul_f16 v1, v1, v3
220 ; GFX1-GISEL-NEXT: s_setpc_b64 s[30:31]
221 %val = call <3 x half> @llvm.experimental.constrained.fmul.v3f16(<3 x half> %x, <3 x half> %y, metadata !"round.tonearest", metadata !"fpexcept.strict")
222 ret <3 x half> %val
223 }
225 ; FIXME: Scalarized
226 define <4 x half> @v_constained_fmul_v4f16_fpexcept_strict(<4 x half> %x, <4 x half> %y) #0 {
227 ; GFX9-SDAG-LABEL: v_constained_fmul_v4f16_fpexcept_strict:
228 ; GFX9-SDAG: ; %bb.0:
229 ; GFX9-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
230 ; GFX9-SDAG-NEXT: v_mul_f16_sdwa v4, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
231 ; GFX9-SDAG-NEXT: v_mul_f16_sdwa v5, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
232 ; GFX9-SDAG-NEXT: v_mul_f16_e32 v1, v1, v3
233 ; GFX9-SDAG-NEXT: v_mul_f16_e32 v0, v0, v2
234 ; GFX9-SDAG-NEXT: s_mov_b32 s4, 0x5040100
235 ; GFX9-SDAG-NEXT: v_perm_b32 v0, v5, v0, s4
236 ; GFX9-SDAG-NEXT: v_perm_b32 v1, v4, v1, s4
237 ; GFX9-SDAG-NEXT: s_setpc_b64 s[30:31]
238 ;
239 ; GFX9-GISEL-LABEL: v_constained_fmul_v4f16_fpexcept_strict:
240 ; GFX9-GISEL: ; %bb.0:
241 ; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
242 ; GFX9-GISEL-NEXT: v_pk_mul_f16 v0, v0, v2
243 ; GFX9-GISEL-NEXT: v_pk_mul_f16 v1, v1, v3
244 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31]
245 ;
246 ; GFX8-SDAG-LABEL: v_constained_fmul_v4f16_fpexcept_strict:
247 ; GFX8-SDAG: ; %bb.0:
248 ; GFX8-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
249 ; GFX8-SDAG-NEXT: v_mul_f16_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
250 ; GFX8-SDAG-NEXT: v_mul_f16_sdwa v5, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
251 ; GFX8-SDAG-NEXT: v_mul_f16_e32 v1, v1, v3
252 ; GFX8-SDAG-NEXT: v_mul_f16_e32 v0, v0, v2
253 ; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v0, v5
254 ; GFX8-SDAG-NEXT: v_or_b32_e32 v1, v1, v4
255 ; GFX8-SDAG-NEXT: s_setpc_b64 s[30:31]
256 ;
257 ; GFX8-GISEL-LABEL: v_constained_fmul_v4f16_fpexcept_strict:
258 ; GFX8-GISEL: ; %bb.0:
259 ; GFX8-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
260 ; GFX8-GISEL-NEXT: v_mul_f16_e32 v4, v0, v2
261 ; GFX8-GISEL-NEXT: v_mul_f16_sdwa v0, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
262 ; GFX8-GISEL-NEXT: v_mul_f16_e32 v2, v1, v3
263 ; GFX8-GISEL-NEXT: v_mul_f16_sdwa v1, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
264 ; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v4, v0
265 ; GFX8-GISEL-NEXT: v_or_b32_e32 v1, v2, v1
266 ; GFX8-GISEL-NEXT: s_setpc_b64 s[30:31]
267 ;
268 ; GFX10-SDAG-LABEL: v_constained_fmul_v4f16_fpexcept_strict:
269 ; GFX10-SDAG: ; %bb.0:
270 ; GFX10-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
271 ; GFX10-SDAG-NEXT: v_mul_f16_sdwa v4, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
272 ; GFX10-SDAG-NEXT: v_mul_f16_sdwa v5, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1
273 ; GFX10-SDAG-NEXT: v_mul_f16_e32 v0, v0, v2
274 ; GFX10-SDAG-NEXT: v_mul_f16_e32 v1, v1, v3
275 ; GFX10-SDAG-NEXT: v_perm_b32 v0, v5, v0, 0x5040100
276 ; GFX10-SDAG-NEXT: v_perm_b32 v1, v4, v1, 0x5040100
277 ; GFX10-SDAG-NEXT: s_setpc_b64 s[30:31]
278 ;
279 ; GFX10-GISEL-LABEL: v_constained_fmul_v4f16_fpexcept_strict:
280 ; GFX10-GISEL: ; %bb.0:
281 ; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
282 ; GFX10-GISEL-NEXT: v_pk_mul_f16 v0, v0, v2
283 ; GFX10-GISEL-NEXT: v_pk_mul_f16 v1, v1, v3
284 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31]
285 ;
286 ; GFX11-SDAG-LABEL: v_constained_fmul_v4f16_fpexcept_strict:
287 ; GFX11-SDAG: ; %bb.0:
288 ; GFX11-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
289 ; GFX11-SDAG-NEXT: v_lshrrev_b32_e32 v4, 16, v3
290 ; GFX11-SDAG-NEXT: v_lshrrev_b32_e32 v5, 16, v2
291 ; GFX11-SDAG-NEXT: v_lshrrev_b32_e32 v6, 16, v0
292 ; GFX11-SDAG-NEXT: v_lshrrev_b32_e32 v7, 16, v1
293 ; GFX11-SDAG-NEXT: v_mul_f16_e32 v1, v1, v3
294 ; GFX11-SDAG-NEXT: v_mul_f16_e32 v0, v0, v2
295 ; GFX11-SDAG-NEXT: v_mul_f16_e32 v2, v6, v5
296 ; GFX11-SDAG-NEXT: v_mul_f16_e32 v3, v7, v4
297 ; GFX11-SDAG-NEXT: v_perm_b32 v0, v2, v0, 0x5040100
298 ; GFX11-SDAG-NEXT: v_perm_b32 v1, v3, v1, 0x5040100
299 ; GFX11-SDAG-NEXT: s_setpc_b64 s[30:31]
300 ;
301 ; GFX1-GISEL-LABEL: v_constained_fmul_v4f16_fpexcept_strict:
302 ; GFX1-GISEL: ; %bb.0:
303 ; GFX1-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
304 ; GFX1-GISEL-NEXT: v_pk_mul_f16 v0, v0, v2
305 ; GFX1-GISEL-NEXT: v_pk_mul_f16 v1, v1, v3
306 ; GFX1-GISEL-NEXT: s_setpc_b64 s[30:31]
307 %val = call <4 x half> @llvm.experimental.constrained.fmul.v4f16(<4 x half> %x, <4 x half> %y, metadata !"round.tonearest", metadata !"fpexcept.strict")
308 ret <4 x half> %val
309 }
311 define amdgpu_ps half @s_constained_fmul_f16_fpexcept_strict(half inreg %x, half inreg %y) #0 {
312 ; GCN-LABEL: s_constained_fmul_f16_fpexcept_strict:
313 ; GCN: ; %bb.0:
314 ; GCN-NEXT: v_mov_b32_e32 v0, s3
315 ; GCN-NEXT: v_mul_f16_e32 v0, s2, v0
316 ; GCN-NEXT: ; return to shader part epilog
317 ;
318 ; GFX10PLUS-LABEL: s_constained_fmul_f16_fpexcept_strict:
319 ; GFX10PLUS: ; %bb.0:
320 ; GFX10PLUS-NEXT: v_mul_f16_e64 v0, s2, s3
321 ; GFX10PLUS-NEXT: ; return to shader part epilog
322 %val = call half @llvm.experimental.constrained.fmul.f16(half %x, half %y, metadata !"round.tonearest", metadata !"fpexcept.strict")
323 ret half %val
324 }
326 define amdgpu_ps <2 x half> @s_constained_fmul_v2f16_fpexcept_strict(<2 x half> inreg %x, <2 x half> inreg %y) #0 {
327 ; GFX9-LABEL: s_constained_fmul_v2f16_fpexcept_strict:
328 ; GFX9: ; %bb.0:
329 ; GFX9-NEXT: v_mov_b32_e32 v0, s3
330 ; GFX9-NEXT: v_pk_mul_f16 v0, s2, v0
331 ; GFX9-NEXT: ; return to shader part epilog
332 ;
333 ; GFX8-SDAG-LABEL: s_constained_fmul_v2f16_fpexcept_strict:
334 ; GFX8-SDAG: ; %bb.0:
335 ; GFX8-SDAG-NEXT: s_lshr_b32 s0, s3, 16
336 ; GFX8-SDAG-NEXT: s_lshr_b32 s1, s2, 16
337 ; GFX8-SDAG-NEXT: v_mov_b32_e32 v0, s0
338 ; GFX8-SDAG-NEXT: v_mov_b32_e32 v1, s1
339 ; GFX8-SDAG-NEXT: v_mul_f16_sdwa v0, v1, v0 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
340 ; GFX8-SDAG-NEXT: v_mov_b32_e32 v1, s3
341 ; GFX8-SDAG-NEXT: v_mul_f16_e32 v1, s2, v1
342 ; GFX8-SDAG-NEXT: v_or_b32_e32 v0, v1, v0
343 ; GFX8-SDAG-NEXT: ; return to shader part epilog
344 ;
345 ; GFX8-GISEL-LABEL: s_constained_fmul_v2f16_fpexcept_strict:
346 ; GFX8-GISEL: ; %bb.0:
347 ; GFX8-GISEL-NEXT: s_lshr_b32 s0, s2, 16
348 ; GFX8-GISEL-NEXT: s_lshr_b32 s1, s3, 16
349 ; GFX8-GISEL-NEXT: v_mov_b32_e32 v0, s3
350 ; GFX8-GISEL-NEXT: v_mov_b32_e32 v1, s1
351 ; GFX8-GISEL-NEXT: v_mov_b32_e32 v2, s0
352 ; GFX8-GISEL-NEXT: v_mul_f16_e32 v0, s2, v0
353 ; GFX8-GISEL-NEXT: v_mul_f16_sdwa v1, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
354 ; GFX8-GISEL-NEXT: v_or_b32_e32 v0, v0, v1
355 ; GFX8-GISEL-NEXT: ; return to shader part epilog
356 ;
357 ; GFX10PLUS-LABEL: s_constained_fmul_v2f16_fpexcept_strict:
358 ; GFX10PLUS: ; %bb.0:
359 ; GFX10PLUS-NEXT: v_pk_mul_f16 v0, s2, s3
360 ; GFX10PLUS-NEXT: ; return to shader part epilog
361 %val = call <2 x half> @llvm.experimental.constrained.fmul.v2f16(<2 x half> %x, <2 x half> %y, metadata !"round.tonearest", metadata !"fpexcept.strict")
362 ret <2 x half> %val
363 }
365 declare half @llvm.experimental.constrained.fmul.f16(half, half, metadata, metadata) #1
366 declare <2 x half> @llvm.experimental.constrained.fmul.v2f16(<2 x half>, <2 x half>, metadata, metadata) #1
367 declare <3 x half> @llvm.experimental.constrained.fmul.v3f16(<3 x half>, <3 x half>, metadata, metadata) #1
368 declare <4 x half> @llvm.experimental.constrained.fmul.v4f16(<4 x half>, <4 x half>, metadata, metadata) #1
370 attributes #0 = { strictfp }
371 attributes #1 = { inaccessiblememonly nounwind willreturn }
372 ;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
373 ; GFX10: {{.*}}
374 ; GFX11: {{.*}}
375 ; GFX8: {{.*}}