1; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri -verify-machineinstrs < %s | FileCheck -check-prefix=CI -check-prefix=GCN -check-prefix=CIVI %s
2; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=VI -check-prefix=GFX89 -check-prefix=GCN -check-prefix=CIVI %s
3; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX89 -check-prefix=GFX9 -check-prefix=GCN %s
4
5; GCN-LABEL: {{^}}fneg_fabs_fadd_f16:
6; CI-DAG: v_cvt_f32_f16_e32
7; CI-DAG: v_cvt_f32_f16_e64 [[CVT_ABS_X:v[0-9]+]], |s{{[0-9]+}}|
8; CI: v_sub_f32_e32 v{{[0-9]+}}, v{{[0-9]+}}, [[CVT_ABS_X]]
9
10; GFX89-NOT: _and
11; GFX89: v_sub_f16_e64 {{v[0-9]+}}, {{s[0-9]+}}, |{{v[0-9]+}}|
12define amdgpu_kernel void @fneg_fabs_fadd_f16(half addrspace(1)* %out, half %x, half %y) {
13  %fabs = call half @llvm.fabs.f16(half %x)
14  %fsub = fsub half -0.0, %fabs
15  %fadd = fadd half %y, %fsub
16  store half %fadd, half addrspace(1)* %out, align 2
17  ret void
18}
19
20; GCN-LABEL: {{^}}fneg_fabs_fmul_f16:
21; CI-DAG: v_cvt_f32_f16_e32
22; CI-DAG: v_cvt_f32_f16_e64 [[CVT_NEG_ABS_X:v[0-9]+]], -|{{s[0-9]+}}|
23; CI: v_mul_f32_e32 {{v[0-9]+}},  {{v[0-9]+}}, [[CVT_NEG_ABS_X]]
24; CI: v_cvt_f16_f32_e32
25
26; GFX89-NOT: _and
27; GFX89: v_mul_f16_e64 [[MUL:v[0-9]+]], {{s[0-9]+}}, -|{{v[0-9]+}}|
28; GFX89-NOT: [[MUL]]
29; GFX89: {{flat|global}}_store_short v{{.+}}, [[MUL]]
30define amdgpu_kernel void @fneg_fabs_fmul_f16(half addrspace(1)* %out, half %x, half %y) {
31  %fabs = call half @llvm.fabs.f16(half %x)
32  %fsub = fsub half -0.0, %fabs
33  %fmul = fmul half %y, %fsub
34  store half %fmul, half addrspace(1)* %out, align 2
35  ret void
36}
37
38; DAGCombiner will transform:
39; (fabs (f16 bitcast (i16 a))) => (f16 bitcast (and (i16 a), 0x7FFFFFFF))
40; unless isFabsFree returns true
41
42; GCN-LABEL: {{^}}fneg_fabs_free_f16:
43; GCN: {{s_or_b32 s[0-9]+, s[0-9]+, 0x8000|s_bitset1_b32 s[0-9]+, 15}}
44define amdgpu_kernel void @fneg_fabs_free_f16(half addrspace(1)* %out, i16 %in) {
45  %bc = bitcast i16 %in to half
46  %fabs = call half @llvm.fabs.f16(half %bc)
47  %fsub = fsub half -0.0, %fabs
48  store half %fsub, half addrspace(1)* %out
49  ret void
50}
51
52; GCN-LABEL: {{^}}fneg_fabs_f16:
53; GCN: {{s_or_b32 s[0-9]+, s[0-9]+, 0x8000|s_bitset1_b32 s[0-9]+, 15}}
54define amdgpu_kernel void @fneg_fabs_f16(half addrspace(1)* %out, half %in) {
55  %fabs = call half @llvm.fabs.f16(half %in)
56  %fsub = fsub half -0.0, %fabs
57  store half %fsub, half addrspace(1)* %out, align 2
58  ret void
59}
60
61; GCN-LABEL: {{^}}v_fneg_fabs_f16:
62; GCN: v_or_b32_e32 v{{[0-9]+}}, 0x8000, v{{[0-9]+}}
63define amdgpu_kernel void @v_fneg_fabs_f16(half addrspace(1)* %out, half addrspace(1)* %in) {
64  %val = load half, half addrspace(1)* %in, align 2
65  %fabs = call half @llvm.fabs.f16(half %val)
66  %fsub = fsub half -0.0, %fabs
67  store half %fsub, half addrspace(1)* %out, align 2
68  ret void
69}
70
71; GCN-LABEL: {{^}}s_fneg_fabs_v2f16_non_bc_src:
72; GFX9-DAG: s_load_dword [[VAL:s[0-9]+]]
73; GFX9-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 0x40003c00
74; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[VAL]], [[K]]
75; GFX9: v_or_b32_e32 [[RESULT:v[0-9]+]], 0x80008000, [[ADD]]
76
77; VI: v_or_b32_e32 v{{[0-9]+}}, 0x80008000, v{{[0-9]+}}
78define amdgpu_kernel void @s_fneg_fabs_v2f16_non_bc_src(<2 x half> addrspace(1)* %out, <2 x half> %in) {
79  %add = fadd <2 x half> %in, <half 1.0, half 2.0>
80  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %add)
81  %fneg.fabs = fsub <2 x half> <half -0.0, half -0.0>, %fabs
82  store <2 x half> %fneg.fabs, <2 x half> addrspace(1)* %out
83  ret void
84}
85
86; FIXME: single bit op
87
88; Combine turns this into integer op when bitcast source (from load)
89
90; GCN-LABEL: {{^}}s_fneg_fabs_v2f16_bc_src:
91; GCN: s_or_b32 s{{[0-9]+}}, s{{[0-9]+}}, 0x80008000
92define amdgpu_kernel void @s_fneg_fabs_v2f16_bc_src(<2 x half> addrspace(1)* %out, <2 x half> %in) {
93  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %in)
94  %fneg.fabs = fsub <2 x half> <half -0.0, half -0.0>, %fabs
95  store <2 x half> %fneg.fabs, <2 x half> addrspace(1)* %out
96  ret void
97}
98
99; GCN-LABEL: {{^}}fneg_fabs_v4f16:
100; GCN: s_mov_b32 [[MASK:s[0-9]+]], 0x80008000
101; GCN: s_or_b32 s{{[0-9]+}}, s{{[0-9]+}}, [[MASK]]
102; GCN: s_or_b32 s{{[0-9]+}}, s{{[0-9]+}}, [[MASK]]
103; GCN: {{flat|global}}_store_dwordx2
104define amdgpu_kernel void @fneg_fabs_v4f16(<4 x half> addrspace(1)* %out, <4 x half> %in) {
105  %fabs = call <4 x half> @llvm.fabs.v4f16(<4 x half> %in)
106  %fsub = fsub <4 x half> <half -0.0, half -0.0, half -0.0, half -0.0>, %fabs
107  store <4 x half> %fsub, <4 x half> addrspace(1)* %out
108  ret void
109}
110
111; GCN-LABEL: {{^}}fold_user_fneg_fabs_v2f16:
112; CI: s_load_dword [[IN:s[0-9]+]]
113; CI: s_lshr_b32
114; CI: v_cvt_f32_f16_e64 v{{[0-9]+}}, |s{{[0-9]+}}|
115; CI: v_cvt_f32_f16_e64 v{{[0-9]+}}, |s{{[0-9]+}}|
116; CI: v_mul_f32_e32 v{{[0-9]+}}, -4.0, v{{[0-9]+}}
117; CI: v_mul_f32_e32 v{{[0-9]+}}, -4.0, v{{[0-9]+}}
118
119; VI: v_mul_f16_e64 v{{[0-9]+}}, |s{{[0-9]+}}|, -4.0
120; VI: v_mul_f16_sdwa v{{[0-9]+}}, |v{{[0-9]+}}|, v{{[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
121
122; GFX9: s_and_b32 [[ABS:s[0-9]+]], s{{[0-9]+}}, 0x7fff7fff
123; GFX9: v_pk_mul_f16 v{{[0-9]+}}, [[ABS]], -4.0 op_sel_hi:[1,0]
124define amdgpu_kernel void @fold_user_fneg_fabs_v2f16(<2 x half> addrspace(1)* %out, <2 x half> %in) #0 {
125  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %in)
126  %fneg.fabs = fsub <2 x half> <half -0.0, half -0.0>, %fabs
127  %mul = fmul <2 x half> %fneg.fabs, <half 4.0, half 4.0>
128  store <2 x half> %mul, <2 x half> addrspace(1)* %out
129  ret void
130}
131
132; GCN-LABEL: {{^}}s_fneg_multi_use_fabs_v2f16:
133; GFX9: s_and_b32 [[ABS:s[0-9]+]], s{{[0-9]+}}, 0x7fff7fff
134; GFX9: v_mov_b32_e32 [[V_ABS:v[0-9]+]], [[ABS]]
135; GFX9: s_xor_b32 [[NEG:s[0-9]+]], [[ABS]], 0x80008000
136; GFX9-DAG: v_mov_b32_e32 [[V_NEG:v[0-9]+]], [[NEG]]
137; GFX9-DAG: global_store_dword v{{[0-9]+}}, [[V_ABS]], s{{\[[0-9]+:[0-9]+\]}}
138; GFX9: global_store_dword v{{[0-9]+}}, [[V_NEG]], s{{\[[0-9]+:[0-9]+\]}}
139define amdgpu_kernel void @s_fneg_multi_use_fabs_v2f16(<2 x half> addrspace(1)* %out0, <2 x half> addrspace(1)* %out1, <2 x half> %in) {
140  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %in)
141  %fneg = fsub <2 x half> <half -0.0, half -0.0>, %fabs
142  store <2 x half> %fabs, <2 x half> addrspace(1)* %out0
143  store <2 x half> %fneg, <2 x half> addrspace(1)* %out1
144  ret void
145}
146
147; GCN-LABEL: {{^}}s_fneg_multi_use_fabs_foldable_neg_v2f16:
148; GFX9: s_and_b32 [[ABS:s[0-9]+]], s{{[0-9]+}}, 0x7fff7fff
149; GFX9: v_pk_mul_f16 v{{[0-9]+}}, [[ABS]], -4.0 op_sel_hi:[1,0]
150define amdgpu_kernel void @s_fneg_multi_use_fabs_foldable_neg_v2f16(<2 x half> addrspace(1)* %out0, <2 x half> addrspace(1)* %out1, <2 x half> %in) {
151  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %in)
152  %fneg = fsub <2 x half> <half -0.0, half -0.0>, %fabs
153  %mul = fmul <2 x half> %fneg, <half 4.0, half 4.0>
154  store <2 x half> %fabs, <2 x half> addrspace(1)* %out0
155  store <2 x half> %mul, <2 x half> addrspace(1)* %out1
156  ret void
157}
158
159declare half @llvm.fabs.f16(half) #1
160declare <2 x half> @llvm.fabs.v2f16(<2 x half>) #1
161declare <4 x half> @llvm.fabs.v4f16(<4 x half>) #1
162
163attributes #0 = { nounwind }
164attributes #1 = { nounwind readnone }
165