1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s
3; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji < %s | FileCheck -check-prefixes=GCN,GFX8 %s
4; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10 %s
5
6define half @v_constained_fma_f16_fpexcept_strict(half %x, half %y, half %z) #0 {
7; GCN-LABEL: v_constained_fma_f16_fpexcept_strict:
8; GCN:       ; %bb.0:
9; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
10; GCN-NEXT:    v_fma_f16 v0, v0, v1, v2
11; GCN-NEXT:    s_setpc_b64 s[30:31]
12;
13; GFX10-LABEL: v_constained_fma_f16_fpexcept_strict:
14; GFX10:       ; %bb.0:
15; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
16; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
17; GFX10-NEXT:    v_fma_f16 v0, v0, v1, v2
18; GFX10-NEXT:    s_setpc_b64 s[30:31]
19  %val = call half @llvm.experimental.constrained.fma.f16(half %x, half %y, half %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
20  ret half %val
21}
22
23define <2 x half> @v_constained_fma_v2f16_fpexcept_strict(<2 x half> %x, <2 x half> %y, <2 x half> %z) #0 {
24; GFX9-LABEL: v_constained_fma_v2f16_fpexcept_strict:
25; GFX9:       ; %bb.0:
26; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
27; GFX9-NEXT:    v_pk_fma_f16 v0, v0, v1, v2
28; GFX9-NEXT:    s_setpc_b64 s[30:31]
29;
30; GFX8-LABEL: v_constained_fma_v2f16_fpexcept_strict:
31; GFX8:       ; %bb.0:
32; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
33; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
34; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
35; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v0
36; GFX8-NEXT:    v_fma_f16 v3, v5, v4, v3
37; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
38; GFX8-NEXT:    v_fma_f16 v0, v0, v1, v2
39; GFX8-NEXT:    v_or_b32_e32 v0, v0, v3
40; GFX8-NEXT:    s_setpc_b64 s[30:31]
41;
42; GFX10-LABEL: v_constained_fma_v2f16_fpexcept_strict:
43; GFX10:       ; %bb.0:
44; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
45; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
46; GFX10-NEXT:    v_pk_fma_f16 v0, v0, v1, v2
47; GFX10-NEXT:    s_setpc_b64 s[30:31]
48  %val = call <2 x half> @llvm.experimental.constrained.fma.v2f16(<2 x half> %x, <2 x half> %y, <2 x half> %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
49  ret <2 x half> %val
50}
51
52define <3 x half> @v_constained_fma_v3f16_fpexcept_strict(<3 x half> %x, <3 x half> %y, <3 x half> %z) #0 {
53; GFX9-LABEL: v_constained_fma_v3f16_fpexcept_strict:
54; GFX9:       ; %bb.0:
55; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
56; GFX9-NEXT:    v_pk_fma_f16 v0, v0, v2, v4
57; GFX9-NEXT:    v_fma_f16 v1, v1, v3, v5
58; GFX9-NEXT:    s_setpc_b64 s[30:31]
59;
60; GFX8-LABEL: v_constained_fma_v3f16_fpexcept_strict:
61; GFX8:       ; %bb.0:
62; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
63; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
64; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v2
65; GFX8-NEXT:    v_lshrrev_b32_e32 v8, 16, v0
66; GFX8-NEXT:    v_fma_f16 v6, v8, v7, v6
67; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
68; GFX8-NEXT:    v_fma_f16 v0, v0, v2, v4
69; GFX8-NEXT:    v_or_b32_e32 v0, v0, v6
70; GFX8-NEXT:    v_fma_f16 v1, v1, v3, v5
71; GFX8-NEXT:    s_setpc_b64 s[30:31]
72;
73; GFX10-LABEL: v_constained_fma_v3f16_fpexcept_strict:
74; GFX10:       ; %bb.0:
75; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
76; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
77; GFX10-NEXT:    v_pk_fma_f16 v0, v0, v2, v4
78; GFX10-NEXT:    v_fma_f16 v1, v1, v3, v5
79; GFX10-NEXT:    s_setpc_b64 s[30:31]
80  %val = call <3 x half> @llvm.experimental.constrained.fma.v3f16(<3 x half> %x, <3 x half> %y, <3 x half> %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
81  ret <3 x half> %val
82}
83
84define <4 x half> @v_constained_fma_v4f16_fpexcept_strict(<4 x half> %x, <4 x half> %y, <4 x half> %z) #0 {
85; GFX9-LABEL: v_constained_fma_v4f16_fpexcept_strict:
86; GFX9:       ; %bb.0:
87; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
88; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
89; GFX9-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
90; GFX9-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
91; GFX9-NEXT:    v_fma_f16 v6, v8, v7, v6
92; GFX9-NEXT:    v_lshrrev_b32_e32 v7, 16, v4
93; GFX9-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
94; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v0
95; GFX9-NEXT:    v_fma_f16 v1, v1, v3, v5
96; GFX9-NEXT:    v_fma_f16 v0, v0, v2, v4
97; GFX9-NEXT:    v_mov_b32_e32 v2, 0xffff
98; GFX9-NEXT:    v_fma_f16 v7, v9, v8, v7
99; GFX9-NEXT:    v_and_b32_e32 v0, v2, v0
100; GFX9-NEXT:    v_and_b32_e32 v1, v2, v1
101; GFX9-NEXT:    v_lshl_or_b32 v0, v7, 16, v0
102; GFX9-NEXT:    v_lshl_or_b32 v1, v6, 16, v1
103; GFX9-NEXT:    s_setpc_b64 s[30:31]
104;
105; GFX8-LABEL: v_constained_fma_v4f16_fpexcept_strict:
106; GFX8:       ; %bb.0:
107; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
108; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
109; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
110; GFX8-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
111; GFX8-NEXT:    v_fma_f16 v6, v8, v7, v6
112; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v4
113; GFX8-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
114; GFX8-NEXT:    v_lshrrev_b32_e32 v9, 16, v0
115; GFX8-NEXT:    v_fma_f16 v7, v9, v8, v7
116; GFX8-NEXT:    v_fma_f16 v0, v0, v2, v4
117; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v7
118; GFX8-NEXT:    v_fma_f16 v1, v1, v3, v5
119; GFX8-NEXT:    v_or_b32_e32 v0, v0, v2
120; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v6
121; GFX8-NEXT:    v_or_b32_e32 v1, v1, v2
122; GFX8-NEXT:    s_setpc_b64 s[30:31]
123;
124; GFX10-LABEL: v_constained_fma_v4f16_fpexcept_strict:
125; GFX10:       ; %bb.0:
126; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
127; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
128; GFX10-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
129; GFX10-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
130; GFX10-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
131; GFX10-NEXT:    v_lshrrev_b32_e32 v9, 16, v4
132; GFX10-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
133; GFX10-NEXT:    v_lshrrev_b32_e32 v11, 16, v0
134; GFX10-NEXT:    v_fmac_f16_e32 v4, v0, v2
135; GFX10-NEXT:    v_mov_b32_e32 v0, 0xffff
136; GFX10-NEXT:    v_fmac_f16_e32 v5, v1, v3
137; GFX10-NEXT:    v_fmac_f16_e32 v6, v8, v7
138; GFX10-NEXT:    v_fmac_f16_e32 v9, v11, v10
139; GFX10-NEXT:    v_and_b32_e32 v1, v0, v4
140; GFX10-NEXT:    v_and_b32_e32 v2, v0, v5
141; GFX10-NEXT:    v_lshl_or_b32 v0, v9, 16, v1
142; GFX10-NEXT:    v_lshl_or_b32 v1, v6, 16, v2
143; GFX10-NEXT:    s_setpc_b64 s[30:31]
144  %val = call <4 x half> @llvm.experimental.constrained.fma.v4f16(<4 x half> %x, <4 x half> %y, <4 x half> %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
145  ret <4 x half> %val
146}
147
148define half @v_constained_fma_f16_fpexcept_strict_fneg(half %x, half %y, half %z) #0 {
149; GCN-LABEL: v_constained_fma_f16_fpexcept_strict_fneg:
150; GCN:       ; %bb.0:
151; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
152; GCN-NEXT:    v_fma_f16 v0, v0, v1, -v2
153; GCN-NEXT:    s_setpc_b64 s[30:31]
154;
155; GFX10-LABEL: v_constained_fma_f16_fpexcept_strict_fneg:
156; GFX10:       ; %bb.0:
157; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
158; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
159; GFX10-NEXT:    v_fma_f16 v0, v0, v1, -v2
160; GFX10-NEXT:    s_setpc_b64 s[30:31]
161  %neg.z = fneg half %z
162  %val = call half @llvm.experimental.constrained.fma.f16(half %x, half %y, half %neg.z, metadata !"round.tonearest", metadata !"fpexcept.strict")
163  ret half %val
164}
165
166define half @v_constained_fma_f16_fpexcept_strict_fneg_fneg(half %x, half %y, half %z) #0 {
167; GCN-LABEL: v_constained_fma_f16_fpexcept_strict_fneg_fneg:
168; GCN:       ; %bb.0:
169; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
170; GCN-NEXT:    v_fma_f16 v0, -v0, -v1, v2
171; GCN-NEXT:    s_setpc_b64 s[30:31]
172;
173; GFX10-LABEL: v_constained_fma_f16_fpexcept_strict_fneg_fneg:
174; GFX10:       ; %bb.0:
175; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
176; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
177; GFX10-NEXT:    v_fma_f16 v0, -v0, -v1, v2
178; GFX10-NEXT:    s_setpc_b64 s[30:31]
179  %neg.x = fneg half %x
180  %neg.y = fneg half %y
181  %val = call half @llvm.experimental.constrained.fma.f16(half %neg.x, half %neg.y, half %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
182  ret half %val
183}
184
185define half @v_constained_fma_f16_fpexcept_strict_fabs_fabs(half %x, half %y, half %z) #0 {
186; GCN-LABEL: v_constained_fma_f16_fpexcept_strict_fabs_fabs:
187; GCN:       ; %bb.0:
188; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
189; GCN-NEXT:    v_fma_f16 v0, |v0|, |v1|, v2
190; GCN-NEXT:    s_setpc_b64 s[30:31]
191;
192; GFX10-LABEL: v_constained_fma_f16_fpexcept_strict_fabs_fabs:
193; GFX10:       ; %bb.0:
194; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
195; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
196; GFX10-NEXT:    v_fma_f16 v0, |v0|, |v1|, v2
197; GFX10-NEXT:    s_setpc_b64 s[30:31]
198  %neg.x = call half @llvm.fabs.f16(half %x)
199  %neg.y = call half @llvm.fabs.f16(half %y)
200  %val = call half @llvm.experimental.constrained.fma.f16(half %neg.x, half %neg.y, half %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
201  ret half %val
202}
203
204define <2 x half> @v_constained_fma_v2f16_fpexcept_strict_fneg_fneg(<2 x half> %x, <2 x half> %y, <2 x half> %z) #0 {
205; GFX9-LABEL: v_constained_fma_v2f16_fpexcept_strict_fneg_fneg:
206; GFX9:       ; %bb.0:
207; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
208; GFX9-NEXT:    v_pk_fma_f16 v0, v0, v1, v2 neg_lo:[1,1,0] neg_hi:[1,1,0]
209; GFX9-NEXT:    s_setpc_b64 s[30:31]
210;
211; GFX8-LABEL: v_constained_fma_v2f16_fpexcept_strict_fneg_fneg:
212; GFX8:       ; %bb.0:
213; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
214; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
215; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
216; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v0
217; GFX8-NEXT:    v_fma_f16 v3, -v5, -v4, v3
218; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
219; GFX8-NEXT:    v_fma_f16 v0, -v0, -v1, v2
220; GFX8-NEXT:    v_or_b32_e32 v0, v0, v3
221; GFX8-NEXT:    s_setpc_b64 s[30:31]
222;
223; GFX10-LABEL: v_constained_fma_v2f16_fpexcept_strict_fneg_fneg:
224; GFX10:       ; %bb.0:
225; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
226; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
227; GFX10-NEXT:    v_pk_fma_f16 v0, v0, v1, v2 neg_lo:[1,1,0] neg_hi:[1,1,0]
228; GFX10-NEXT:    s_setpc_b64 s[30:31]
229  %neg.x = fneg <2 x half> %x
230  %neg.y = fneg <2 x half> %y
231  %val = call <2 x half> @llvm.experimental.constrained.fma.v2f16(<2 x half> %neg.x, <2 x half> %neg.y, <2 x half> %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
232  ret <2 x half> %val
233}
234
235declare half @llvm.fabs.f16(half) #1
236declare half @llvm.experimental.constrained.fma.f16(half, half, half, metadata, metadata) #1
237declare <2 x half> @llvm.experimental.constrained.fma.v2f16(<2 x half>, <2 x half>, <2 x half>, metadata, metadata) #1
238declare <3 x half> @llvm.experimental.constrained.fma.v3f16(<3 x half>, <3 x half>, <3 x half>, metadata, metadata) #1
239declare <4 x half> @llvm.experimental.constrained.fma.v4f16(<4 x half>, <4 x half>, <4 x half>, metadata, metadata) #1
240
241attributes #0 = { strictfp }
242attributes #1 = { inaccessiblememonly nounwind willreturn }
243