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_fmac_f16_e32 v2, v0, v1
18; GFX10-NEXT:    v_mov_b32_e32 v0, 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 v5, 16, v0
35; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
36; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
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_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
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 v8, 16, v0
65; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
66; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v2
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_sdwa v0, v0, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
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_mov_b32_e32 v6, v5
79; GFX10-NEXT:    v_pk_fma_f16 v0, v0, v2, v4
80; GFX10-NEXT:    v_fmac_f16_e32 v6, v1, v3
81; GFX10-NEXT:    v_mov_b32_e32 v1, v6
82; GFX10-NEXT:    s_setpc_b64 s[30:31]
83  %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")
84  ret <3 x half> %val
85}
86
87define <4 x half> @v_constained_fma_v4f16_fpexcept_strict(<4 x half> %x, <4 x half> %y, <4 x half> %z) #0 {
88; GFX9-LABEL: v_constained_fma_v4f16_fpexcept_strict:
89; GFX9:       ; %bb.0:
90; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
91; GFX9-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
92; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
93; GFX9-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
94; GFX9-NEXT:    v_fma_f16 v6, v8, v7, v6
95; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v0
96; GFX9-NEXT:    v_lshrrev_b32_e32 v7, 16, v4
97; GFX9-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
98; GFX9-NEXT:    v_fma_f16 v0, v0, v2, v4
99; GFX9-NEXT:    v_mov_b32_e32 v2, 0xffff
100; GFX9-NEXT:    v_fma_f16 v1, v1, v3, v5
101; GFX9-NEXT:    v_and_b32_e32 v1, v2, v1
102; GFX9-NEXT:    v_fma_f16 v7, v9, v8, v7
103; GFX9-NEXT:    v_and_b32_e32 v0, v2, v0
104; GFX9-NEXT:    v_lshl_or_b32 v0, v7, 16, v0
105; GFX9-NEXT:    v_lshl_or_b32 v1, v6, 16, v1
106; GFX9-NEXT:    s_setpc_b64 s[30:31]
107;
108; GFX8-LABEL: v_constained_fma_v4f16_fpexcept_strict:
109; GFX8:       ; %bb.0:
110; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
111; GFX8-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
112; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
113; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
114; GFX8-NEXT:    v_fma_f16 v6, v8, v7, v6
115; GFX8-NEXT:    v_lshrrev_b32_e32 v9, 16, v0
116; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v4
117; GFX8-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
118; GFX8-NEXT:    v_fma_f16 v7, v9, v8, v7
119; GFX8-NEXT:    v_fma_f16 v0, v0, v2, v4
120; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v7
121; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
122; GFX8-NEXT:    v_fma_f16 v1, v1, v3, v5
123; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v6
124; GFX8-NEXT:    v_or_b32_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
125; GFX8-NEXT:    s_setpc_b64 s[30:31]
126;
127; GFX10-LABEL: v_constained_fma_v4f16_fpexcept_strict:
128; GFX10:       ; %bb.0:
129; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
130; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
131; GFX10-NEXT:    v_mov_b32_e32 v14, v5
132; GFX10-NEXT:    v_mov_b32_e32 v15, v4
133; GFX10-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
134; GFX10-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
135; GFX10-NEXT:    v_lshrrev_b32_e32 v10, 16, v2
136; GFX10-NEXT:    v_lshrrev_b32_e32 v5, 16, v14
137; GFX10-NEXT:    v_lshrrev_b32_e32 v4, 16, v15
138; GFX10-NEXT:    v_lshrrev_b32_e32 v11, 16, v0
139; GFX10-NEXT:    v_fmac_f16_e32 v15, v0, v2
140; GFX10-NEXT:    v_mov_b32_e32 v0, 0xffff
141; GFX10-NEXT:    v_fmac_f16_e32 v14, v1, v3
142; GFX10-NEXT:    v_fmac_f16_e32 v5, v8, v7
143; GFX10-NEXT:    v_fmac_f16_e32 v4, v11, v10
144; GFX10-NEXT:    v_and_b32_e32 v1, v0, v15
145; GFX10-NEXT:    v_and_b32_e32 v2, v0, v14
146; GFX10-NEXT:    v_lshl_or_b32 v0, v4, 16, v1
147; GFX10-NEXT:    v_lshl_or_b32 v1, v5, 16, v2
148; GFX10-NEXT:    s_setpc_b64 s[30:31]
149  %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")
150  ret <4 x half> %val
151}
152
153define half @v_constained_fma_f16_fpexcept_strict_fneg(half %x, half %y, half %z) #0 {
154; GCN-LABEL: v_constained_fma_f16_fpexcept_strict_fneg:
155; GCN:       ; %bb.0:
156; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
157; GCN-NEXT:    v_fma_f16 v0, v0, v1, -v2
158; GCN-NEXT:    s_setpc_b64 s[30:31]
159;
160; GFX10-LABEL: v_constained_fma_f16_fpexcept_strict_fneg:
161; GFX10:       ; %bb.0:
162; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
163; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
164; GFX10-NEXT:    v_fma_f16 v0, v0, v1, -v2
165; GFX10-NEXT:    s_setpc_b64 s[30:31]
166  %neg.z = fneg half %z
167  %val = call half @llvm.experimental.constrained.fma.f16(half %x, half %y, half %neg.z, metadata !"round.tonearest", metadata !"fpexcept.strict")
168  ret half %val
169}
170
171define half @v_constained_fma_f16_fpexcept_strict_fneg_fneg(half %x, half %y, half %z) #0 {
172; GCN-LABEL: v_constained_fma_f16_fpexcept_strict_fneg_fneg:
173; GCN:       ; %bb.0:
174; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
175; GCN-NEXT:    v_fma_f16 v0, -v0, -v1, v2
176; GCN-NEXT:    s_setpc_b64 s[30:31]
177;
178; GFX10-LABEL: v_constained_fma_f16_fpexcept_strict_fneg_fneg:
179; GFX10:       ; %bb.0:
180; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
181; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
182; GFX10-NEXT:    v_fmac_f16_e64 v2, -v0, -v1
183; GFX10-NEXT:    v_mov_b32_e32 v0, v2
184; GFX10-NEXT:    s_setpc_b64 s[30:31]
185  %neg.x = fneg half %x
186  %neg.y = fneg half %y
187  %val = call half @llvm.experimental.constrained.fma.f16(half %neg.x, half %neg.y, half %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
188  ret half %val
189}
190
191define half @v_constained_fma_f16_fpexcept_strict_fabs_fabs(half %x, half %y, half %z) #0 {
192; GCN-LABEL: v_constained_fma_f16_fpexcept_strict_fabs_fabs:
193; GCN:       ; %bb.0:
194; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
195; GCN-NEXT:    v_fma_f16 v0, |v0|, |v1|, v2
196; GCN-NEXT:    s_setpc_b64 s[30:31]
197;
198; GFX10-LABEL: v_constained_fma_f16_fpexcept_strict_fabs_fabs:
199; GFX10:       ; %bb.0:
200; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
201; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
202; GFX10-NEXT:    v_fmac_f16_e64 v2, |v0|, |v1|
203; GFX10-NEXT:    v_mov_b32_e32 v0, v2
204; GFX10-NEXT:    s_setpc_b64 s[30:31]
205  %neg.x = call half @llvm.fabs.f16(half %x)
206  %neg.y = call half @llvm.fabs.f16(half %y)
207  %val = call half @llvm.experimental.constrained.fma.f16(half %neg.x, half %neg.y, half %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
208  ret half %val
209}
210
211define <2 x half> @v_constained_fma_v2f16_fpexcept_strict_fneg_fneg(<2 x half> %x, <2 x half> %y, <2 x half> %z) #0 {
212; GFX9-LABEL: v_constained_fma_v2f16_fpexcept_strict_fneg_fneg:
213; GFX9:       ; %bb.0:
214; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
215; GFX9-NEXT:    v_pk_fma_f16 v0, v0, v1, v2 neg_lo:[1,1,0] neg_hi:[1,1,0]
216; GFX9-NEXT:    s_setpc_b64 s[30:31]
217;
218; GFX8-LABEL: v_constained_fma_v2f16_fpexcept_strict_fneg_fneg:
219; GFX8:       ; %bb.0:
220; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
221; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v0
222; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
223; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
224; GFX8-NEXT:    v_fma_f16 v3, -v5, -v4, v3
225; GFX8-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
226; GFX8-NEXT:    v_fma_f16 v0, -v0, -v1, v2
227; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
228; GFX8-NEXT:    s_setpc_b64 s[30:31]
229;
230; GFX10-LABEL: v_constained_fma_v2f16_fpexcept_strict_fneg_fneg:
231; GFX10:       ; %bb.0:
232; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
233; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
234; GFX10-NEXT:    v_pk_fma_f16 v0, v0, v1, v2 neg_lo:[1,1,0] neg_hi:[1,1,0]
235; GFX10-NEXT:    s_setpc_b64 s[30:31]
236  %neg.x = fneg <2 x half> %x
237  %neg.y = fneg <2 x half> %y
238  %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")
239  ret <2 x half> %val
240}
241
242declare half @llvm.fabs.f16(half) #1
243declare half @llvm.experimental.constrained.fma.f16(half, half, half, metadata, metadata) #1
244declare <2 x half> @llvm.experimental.constrained.fma.v2f16(<2 x half>, <2 x half>, <2 x half>, metadata, metadata) #1
245declare <3 x half> @llvm.experimental.constrained.fma.v3f16(<3 x half>, <3 x half>, <3 x half>, metadata, metadata) #1
246declare <4 x half> @llvm.experimental.constrained.fma.v4f16(<4 x half>, <4 x half>, <4 x half>, metadata, metadata) #1
247
248attributes #0 = { strictfp }
249attributes #1 = { inaccessiblememonly nounwind willreturn }
250