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; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 < %s | FileCheck -check-prefixes=GFX10 %s
6
7define half @v_constained_fma_f16_fpexcept_strict(half %x, half %y, half %z) #0 {
8; GCN-LABEL: v_constained_fma_f16_fpexcept_strict:
9; GCN:       ; %bb.0:
10; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
11; GCN-NEXT:    v_fma_f16 v0, v0, v1, v2
12; GCN-NEXT:    s_setpc_b64 s[30:31]
13;
14; GFX10-LABEL: v_constained_fma_f16_fpexcept_strict:
15; GFX10:       ; %bb.0:
16; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
17; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
18; GFX10-NEXT:    v_fma_f16 v0, v0, v1, v2
19; GFX10-NEXT:    s_setpc_b64 s[30:31]
20  %val = call half @llvm.experimental.constrained.fma.f16(half %x, half %y, half %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
21  ret half %val
22}
23
24define <2 x half> @v_constained_fma_v2f16_fpexcept_strict(<2 x half> %x, <2 x half> %y, <2 x half> %z) #0 {
25; GFX9-LABEL: v_constained_fma_v2f16_fpexcept_strict:
26; GFX9:       ; %bb.0:
27; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
28; GFX9-NEXT:    v_pk_fma_f16 v0, v0, v1, v2
29; GFX9-NEXT:    s_setpc_b64 s[30:31]
30;
31; GFX8-LABEL: v_constained_fma_v2f16_fpexcept_strict:
32; GFX8:       ; %bb.0:
33; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
34; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
35; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
36; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v0
37; GFX8-NEXT:    v_fma_f16 v3, v5, v4, v3
38; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
39; GFX8-NEXT:    v_fma_f16 v0, v0, v1, v2
40; GFX8-NEXT:    v_or_b32_e32 v0, v0, v3
41; GFX8-NEXT:    s_setpc_b64 s[30:31]
42;
43; GFX10-LABEL: v_constained_fma_v2f16_fpexcept_strict:
44; GFX10:       ; %bb.0:
45; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
46; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
47; GFX10-NEXT:    v_pk_fma_f16 v0, v0, v1, v2
48; GFX10-NEXT:    s_setpc_b64 s[30:31]
49  %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")
50  ret <2 x half> %val
51}
52
53define <3 x half> @v_constained_fma_v3f16_fpexcept_strict(<3 x half> %x, <3 x half> %y, <3 x half> %z) #0 {
54; GFX9-LABEL: v_constained_fma_v3f16_fpexcept_strict:
55; GFX9:       ; %bb.0:
56; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
57; GFX9-NEXT:    v_pk_fma_f16 v0, v0, v2, v4
58; GFX9-NEXT:    v_fma_f16 v1, v1, v3, v5
59; GFX9-NEXT:    s_setpc_b64 s[30:31]
60;
61; GFX8-LABEL: v_constained_fma_v3f16_fpexcept_strict:
62; GFX8:       ; %bb.0:
63; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
64; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
65; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v2
66; GFX8-NEXT:    v_lshrrev_b32_e32 v8, 16, v0
67; GFX8-NEXT:    v_fma_f16 v6, v8, v7, v6
68; GFX8-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
69; GFX8-NEXT:    v_fma_f16 v0, v0, v2, v4
70; GFX8-NEXT:    v_or_b32_e32 v0, v0, v6
71; GFX8-NEXT:    v_fma_f16 v1, v1, v3, v5
72; GFX8-NEXT:    s_setpc_b64 s[30:31]
73;
74; GFX10-LABEL: v_constained_fma_v3f16_fpexcept_strict:
75; GFX10:       ; %bb.0:
76; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
77; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
78; GFX10-NEXT:    v_pk_fma_f16 v0, v0, v2, v4
79; GFX10-NEXT:    v_fma_f16 v1, v1, v3, v5
80; GFX10-NEXT:    s_setpc_b64 s[30:31]
81  %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")
82  ret <3 x half> %val
83}
84
85define <4 x half> @v_constained_fma_v4f16_fpexcept_strict(<4 x half> %x, <4 x half> %y, <4 x half> %z) #0 {
86; GFX9-LABEL: v_constained_fma_v4f16_fpexcept_strict:
87; GFX9:       ; %bb.0:
88; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
89; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
90; GFX9-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
91; GFX9-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
92; GFX9-NEXT:    v_fma_f16 v6, v8, v7, v6
93; GFX9-NEXT:    v_lshrrev_b32_e32 v7, 16, v4
94; GFX9-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
95; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v0
96; GFX9-NEXT:    v_fma_f16 v1, v1, v3, v5
97; GFX9-NEXT:    v_fma_f16 v0, v0, v2, v4
98; GFX9-NEXT:    v_fma_f16 v7, v9, v8, v7
99; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff, v0
100; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff, 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_fmac_f16_e32 v5, v1, v3
136; GFX10-NEXT:    v_fmac_f16_e32 v6, v8, v7
137; GFX10-NEXT:    v_fmac_f16_e32 v9, v11, v10
138; GFX10-NEXT:    v_and_b32_e32 v0, 0xffff, v4
139; GFX10-NEXT:    v_and_b32_e32 v1, 0xffff, v5
140; GFX10-NEXT:    v_lshl_or_b32 v0, v9, 16, v0
141; GFX10-NEXT:    v_lshl_or_b32 v1, v6, 16, v1
142; GFX10-NEXT:    s_setpc_b64 s[30:31]
143  %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")
144  ret <4 x half> %val
145}
146
147define half @v_constained_fma_f16_fpexcept_strict_fneg(half %x, half %y, half %z) #0 {
148; GCN-LABEL: v_constained_fma_f16_fpexcept_strict_fneg:
149; GCN:       ; %bb.0:
150; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
151; GCN-NEXT:    v_fma_f16 v0, v0, v1, -v2
152; GCN-NEXT:    s_setpc_b64 s[30:31]
153;
154; GFX10-LABEL: v_constained_fma_f16_fpexcept_strict_fneg:
155; GFX10:       ; %bb.0:
156; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
157; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
158; GFX10-NEXT:    v_fma_f16 v0, v0, v1, -v2
159; GFX10-NEXT:    s_setpc_b64 s[30:31]
160  %neg.z = fneg half %z
161  %val = call half @llvm.experimental.constrained.fma.f16(half %x, half %y, half %neg.z, metadata !"round.tonearest", metadata !"fpexcept.strict")
162  ret half %val
163}
164
165define half @v_constained_fma_f16_fpexcept_strict_fneg_fneg(half %x, half %y, half %z) #0 {
166; GCN-LABEL: v_constained_fma_f16_fpexcept_strict_fneg_fneg:
167; GCN:       ; %bb.0:
168; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
169; GCN-NEXT:    v_fma_f16 v0, -v0, -v1, v2
170; GCN-NEXT:    s_setpc_b64 s[30:31]
171;
172; GFX10-LABEL: v_constained_fma_f16_fpexcept_strict_fneg_fneg:
173; GFX10:       ; %bb.0:
174; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
175; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
176; GFX10-NEXT:    v_fma_f16 v0, -v0, -v1, v2
177; GFX10-NEXT:    s_setpc_b64 s[30:31]
178  %neg.x = fneg half %x
179  %neg.y = fneg half %y
180  %val = call half @llvm.experimental.constrained.fma.f16(half %neg.x, half %neg.y, half %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
181  ret half %val
182}
183
184define half @v_constained_fma_f16_fpexcept_strict_fabs_fabs(half %x, half %y, half %z) #0 {
185; GCN-LABEL: v_constained_fma_f16_fpexcept_strict_fabs_fabs:
186; GCN:       ; %bb.0:
187; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
188; GCN-NEXT:    v_fma_f16 v0, |v0|, |v1|, v2
189; GCN-NEXT:    s_setpc_b64 s[30:31]
190;
191; GFX10-LABEL: v_constained_fma_f16_fpexcept_strict_fabs_fabs:
192; GFX10:       ; %bb.0:
193; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
194; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
195; GFX10-NEXT:    v_fma_f16 v0, |v0|, |v1|, v2
196; GFX10-NEXT:    s_setpc_b64 s[30:31]
197  %neg.x = call half @llvm.fabs.f16(half %x)
198  %neg.y = call half @llvm.fabs.f16(half %y)
199  %val = call half @llvm.experimental.constrained.fma.f16(half %neg.x, half %neg.y, half %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
200  ret half %val
201}
202
203define <2 x half> @v_constained_fma_v2f16_fpexcept_strict_fneg_fneg(<2 x half> %x, <2 x half> %y, <2 x half> %z) #0 {
204; GFX9-LABEL: v_constained_fma_v2f16_fpexcept_strict_fneg_fneg:
205; GFX9:       ; %bb.0:
206; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
207; GFX9-NEXT:    v_pk_fma_f16 v0, v0, v1, v2 neg_lo:[1,1,0] neg_hi:[1,1,0]
208; GFX9-NEXT:    s_setpc_b64 s[30:31]
209;
210; GFX8-LABEL: v_constained_fma_v2f16_fpexcept_strict_fneg_fneg:
211; GFX8:       ; %bb.0:
212; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
213; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
214; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
215; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v0
216; GFX8-NEXT:    v_fma_f16 v3, -v5, -v4, v3
217; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
218; GFX8-NEXT:    v_fma_f16 v0, -v0, -v1, v2
219; GFX8-NEXT:    v_or_b32_e32 v0, v0, v3
220; GFX8-NEXT:    s_setpc_b64 s[30:31]
221;
222; GFX10-LABEL: v_constained_fma_v2f16_fpexcept_strict_fneg_fneg:
223; GFX10:       ; %bb.0:
224; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
225; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
226; GFX10-NEXT:    v_pk_fma_f16 v0, v0, v1, v2 neg_lo:[1,1,0] neg_hi:[1,1,0]
227; GFX10-NEXT:    s_setpc_b64 s[30:31]
228  %neg.x = fneg <2 x half> %x
229  %neg.y = fneg <2 x half> %y
230  %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")
231  ret <2 x half> %val
232}
233
234declare half @llvm.fabs.f16(half) #1
235declare half @llvm.experimental.constrained.fma.f16(half, half, half, metadata, metadata) #1
236declare <2 x half> @llvm.experimental.constrained.fma.v2f16(<2 x half>, <2 x half>, <2 x half>, metadata, metadata) #1
237declare <3 x half> @llvm.experimental.constrained.fma.v3f16(<3 x half>, <3 x half>, <3 x half>, metadata, metadata) #1
238declare <4 x half> @llvm.experimental.constrained.fma.v4f16(<4 x half>, <4 x half>, <4 x half>, metadata, metadata) #1
239
240attributes #0 = { strictfp }
241attributes #1 = { inaccessiblememonly nounwind willreturn }
242