xref: /llvm-project/llvm/test/CodeGen/AMDGPU/GlobalISel/strict_fma.f16.ll (revision 76c22b18eafd2156568d72e9df2ff7bd3457888a) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GCN,GFX9 %s
3; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji < %s | FileCheck -check-prefixes=GCN,GFX8 %s
4
5define half @v_constained_fma_f16_fpexcept_strict(half %x, half %y, half %z) #0 {
6; GCN-LABEL: v_constained_fma_f16_fpexcept_strict:
7; GCN:       ; %bb.0:
8; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
9; GCN-NEXT:    v_fma_f16 v0, v0, v1, v2
10; GCN-NEXT:    s_setpc_b64 s[30:31]
11  %val = call half @llvm.experimental.constrained.fma.f16(half %x, half %y, half %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
12  ret half %val
13}
14
15define <2 x half> @v_constained_fma_v2f16_fpexcept_strict(<2 x half> %x, <2 x half> %y, <2 x half> %z) #0 {
16; GFX9-LABEL: v_constained_fma_v2f16_fpexcept_strict:
17; GFX9:       ; %bb.0:
18; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
19; GFX9-NEXT:    v_pk_fma_f16 v0, v0, v1, v2
20; GFX9-NEXT:    s_setpc_b64 s[30:31]
21;
22; GFX8-LABEL: v_constained_fma_v2f16_fpexcept_strict:
23; GFX8:       ; %bb.0:
24; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
25; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
26; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
27; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
28; GFX8-NEXT:    v_fma_f16 v0, v0, v1, v2
29; GFX8-NEXT:    v_fma_f16 v1, v3, v4, v5
30; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
31; GFX8-NEXT:    v_or_b32_e32 v0, v0, v1
32; GFX8-NEXT:    s_setpc_b64 s[30:31]
33  %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")
34  ret <2 x half> %val
35}
36
37define <3 x half> @v_constained_fma_v3f16_fpexcept_strict(<3 x half> %x, <3 x half> %y, <3 x half> %z) #0 {
38; GFX9-LABEL: v_constained_fma_v3f16_fpexcept_strict:
39; GFX9:       ; %bb.0:
40; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
41; GFX9-NEXT:    v_pk_fma_f16 v0, v0, v2, v4
42; GFX9-NEXT:    v_pk_fma_f16 v1, v1, v3, v5
43; GFX9-NEXT:    s_setpc_b64 s[30:31]
44;
45; GFX8-LABEL: v_constained_fma_v3f16_fpexcept_strict:
46; GFX8:       ; %bb.0:
47; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
48; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
49; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v2
50; GFX8-NEXT:    v_lshrrev_b32_e32 v8, 16, v4
51; GFX8-NEXT:    v_fma_f16 v0, v0, v2, v4
52; GFX8-NEXT:    v_fma_f16 v2, v6, v7, v8
53; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
54; GFX8-NEXT:    v_fma_f16 v1, v1, v3, v5
55; GFX8-NEXT:    v_or_b32_e32 v0, v0, v2
56; GFX8-NEXT:    s_setpc_b64 s[30:31]
57  %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")
58  ret <3 x half> %val
59}
60
61define <4 x half> @v_constained_fma_v4f16_fpexcept_strict(<4 x half> %x, <4 x half> %y, <4 x half> %z) #0 {
62; GFX9-LABEL: v_constained_fma_v4f16_fpexcept_strict:
63; GFX9:       ; %bb.0:
64; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
65; GFX9-NEXT:    v_pk_fma_f16 v0, v0, v2, v4
66; GFX9-NEXT:    v_pk_fma_f16 v1, v1, v3, v5
67; GFX9-NEXT:    s_setpc_b64 s[30:31]
68;
69; GFX8-LABEL: v_constained_fma_v4f16_fpexcept_strict:
70; GFX8:       ; %bb.0:
71; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
72; GFX8-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
73; GFX8-NEXT:    v_lshrrev_b32_e32 v8, 16, v2
74; GFX8-NEXT:    v_lshrrev_b32_e32 v10, 16, v4
75; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 16, v1
76; GFX8-NEXT:    v_lshrrev_b32_e32 v9, 16, v3
77; GFX8-NEXT:    v_lshrrev_b32_e32 v11, 16, v5
78; GFX8-NEXT:    v_fma_f16 v0, v0, v2, v4
79; GFX8-NEXT:    v_fma_f16 v2, v6, v8, v10
80; GFX8-NEXT:    v_fma_f16 v1, v1, v3, v5
81; GFX8-NEXT:    v_fma_f16 v3, v7, v9, v11
82; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
83; GFX8-NEXT:    v_or_b32_e32 v0, v0, v2
84; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
85; GFX8-NEXT:    v_or_b32_e32 v1, v1, v2
86; GFX8-NEXT:    s_setpc_b64 s[30:31]
87  %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")
88  ret <4 x half> %val
89}
90
91define half @v_constained_fma_f16_fpexcept_strict_fneg(half %x, half %y, half %z) #0 {
92; GCN-LABEL: v_constained_fma_f16_fpexcept_strict_fneg:
93; GCN:       ; %bb.0:
94; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
95; GCN-NEXT:    v_fma_f16 v0, v0, v1, -v2
96; GCN-NEXT:    s_setpc_b64 s[30:31]
97  %neg.z = fneg half %z
98  %val = call half @llvm.experimental.constrained.fma.f16(half %x, half %y, half %neg.z, metadata !"round.tonearest", metadata !"fpexcept.strict")
99  ret half %val
100}
101
102define half @v_constained_fma_f16_fpexcept_strict_fneg_fneg(half %x, half %y, half %z) #0 {
103; GCN-LABEL: v_constained_fma_f16_fpexcept_strict_fneg_fneg:
104; GCN:       ; %bb.0:
105; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
106; GCN-NEXT:    v_fma_f16 v0, -v0, -v1, v2
107; GCN-NEXT:    s_setpc_b64 s[30:31]
108  %neg.x = fneg half %x
109  %neg.y = fneg half %y
110  %val = call half @llvm.experimental.constrained.fma.f16(half %neg.x, half %neg.y, half %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
111  ret half %val
112}
113
114define half @v_constained_fma_f16_fpexcept_strict_fabs_fabs(half %x, half %y, half %z) #0 {
115; GCN-LABEL: v_constained_fma_f16_fpexcept_strict_fabs_fabs:
116; GCN:       ; %bb.0:
117; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
118; GCN-NEXT:    v_fma_f16 v0, |v0|, |v1|, v2
119; GCN-NEXT:    s_setpc_b64 s[30:31]
120  %neg.x = call half @llvm.fabs.f16(half %x) #0
121  %neg.y = call half @llvm.fabs.f16(half %y) #0
122  %val = call half @llvm.experimental.constrained.fma.f16(half %neg.x, half %neg.y, half %z, metadata !"round.tonearest", metadata !"fpexcept.strict")
123  ret half %val
124}
125
126define <2 x half> @v_constained_fma_v2f16_fpexcept_strict_fneg_fneg(<2 x half> %x, <2 x half> %y, <2 x half> %z) #0 {
127; GFX9-LABEL: v_constained_fma_v2f16_fpexcept_strict_fneg_fneg:
128; GFX9:       ; %bb.0:
129; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
130; GFX9-NEXT:    v_pk_fma_f16 v0, v0, v1, v2 neg_lo:[1,1,0] neg_hi:[1,1,0]
131; GFX9-NEXT:    s_setpc_b64 s[30:31]
132;
133; GFX8-LABEL: v_constained_fma_v2f16_fpexcept_strict_fneg_fneg:
134; GFX8:       ; %bb.0:
135; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
136; GFX8-NEXT:    v_xor_b32_e32 v0, 0x80008000, v0
137; GFX8-NEXT:    v_xor_b32_e32 v1, 0x80008000, v1
138; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
139; GFX8-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
140; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
141; GFX8-NEXT:    v_fma_f16 v0, v0, v1, v2
142; GFX8-NEXT:    v_fma_f16 v1, v3, v4, v5
143; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
144; GFX8-NEXT:    v_or_b32_e32 v0, v0, v1
145; GFX8-NEXT:    s_setpc_b64 s[30:31]
146  %neg.x = fneg <2 x half> %x
147  %neg.y = fneg <2 x half> %y
148  %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")
149  ret <2 x half> %val
150}
151
152declare half @llvm.fabs.f16(half)
153declare half @llvm.experimental.constrained.fma.f16(half, half, half, metadata, metadata)
154declare <2 x half> @llvm.experimental.constrained.fma.v2f16(<2 x half>, <2 x half>, <2 x half>, metadata, metadata)
155declare <3 x half> @llvm.experimental.constrained.fma.v3f16(<3 x half>, <3 x half>, <3 x half>, metadata, metadata)
156declare <4 x half> @llvm.experimental.constrained.fma.v4f16(<4 x half>, <4 x half>, <4 x half>, metadata, metadata)
157
158attributes #0 = { strictfp }
159