1; RUN: llc -march=amdgcn -mcpu=gfx901 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GFX9 -check-prefix=GCN %s
2; RUN: llc -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=VI -check-prefix=CIVI -check-prefix=GCN %s
3; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -check-prefix=CI -check-prefix=CIVI -check-prefix=GCN %s
4
5; GCN-LABEL: {{^}}s_abs_v2i16:
6; GFX9: s_load_dword [[VAL:s[0-9]+]]
7; GFX9: v_pk_sub_i16 [[SUB:v[0-9]+]], 0, [[VAL]]
8; GFX9: v_pk_max_i16 [[MAX:v[0-9]+]], [[VAL]], [[SUB]]
9; GFX9: v_pk_add_u16 [[ADD:v[0-9]+]], [[MAX]], 2
10
11; CIVI: v_sub_i32_e32
12; CIVI-DAG: v_sub_i32_e32
13; CIVI: v_bfe_i32
14; CIVI-DAG: v_bfe_i32
15; CIVI-DAG: v_add_i32_e32
16; CIVI-DAG: v_lshlrev_b32_e32 v{{[0-9]+}}, 16
17; CIVI: v_add_i32_e32
18; CIVI: v_and_b32_e32 v{{[0-9]+}}, 0xffff,
19; CIVI: v_or_b32_e32
20define void @s_abs_v2i16(<2 x i16> addrspace(1)* %out, <2 x i16> %val) #0 {
21  %neg = sub <2 x i16> zeroinitializer, %val
22  %cond = icmp sgt <2 x i16> %val, %neg
23  %res = select <2 x i1> %cond, <2 x i16> %val, <2 x i16> %neg
24  %res2 = add <2 x i16> %res, <i16 2, i16 2>
25  store <2 x i16> %res2, <2 x i16> addrspace(1)* %out, align 4
26  ret void
27}
28
29; GCN-LABEL: {{^}}v_abs_v2i16:
30; GFX9: flat_load_dword [[VAL:v[0-9]+]]
31; GFX9: v_pk_sub_i16 [[SUB:v[0-9]+]], 0, [[VAL]]
32; GFX9: v_pk_max_i16 [[MAX:v[0-9]+]], [[VAL]], [[SUB]]
33; GFX9: v_pk_add_u16 [[ADD:v[0-9]+]], [[MAX]], 2
34
35; VI: v_lshrrev_b32_e32 v{{[0-9]+}}, 16,
36; VI: v_sub_u16_e32 v{{[0-9]+}}, 0, v{{[0-9]+}}
37; VI: v_sub_u16_e32 v{{[0-9]+}}, 0, v{{[0-9]+}}
38; VI: v_max_i16_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
39; VI: v_max_i16_e32 v{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}
40; VI: v_add_u16_e32 v{{[0-9]+}}, 2, v{{[0-9]+}}
41; VI: v_add_u16_e32 v{{[0-9]+}}, 2, v{{[0-9]+}}
42; VI-NOT: v_and_b32
43; VI: v_or_b32_e32
44define void @v_abs_v2i16(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(1)* %src) #0 {
45  %tid = call i32 @llvm.amdgcn.workitem.id.x()
46  %gep.in = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %src, i32 %tid
47  %gep.out = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %out, i32 %tid
48  %val = load <2 x i16>, <2 x i16> addrspace(1)* %gep.in, align 4
49  %neg = sub <2 x i16> zeroinitializer, %val
50  %cond = icmp sgt <2 x i16> %val, %neg
51  %res = select <2 x i1> %cond, <2 x i16> %val, <2 x i16> %neg
52  %res2 = add <2 x i16> %res, <i16 2, i16 2>
53  store <2 x i16> %res2, <2 x i16> addrspace(1)* %gep.out, align 4
54  ret void
55}
56
57; GCN-LABEL: {{^}}s_abs_v2i16_2:
58; GFX9: s_load_dword [[VAL:s[0-9]+]]
59; GFX9: v_pk_sub_i16 [[SUB:v[0-9]+]], 0, [[VAL]]
60; GFX9: v_pk_max_i16 [[MAX:v[0-9]+]], [[VAL]], [[SUB]]
61; GFX9: v_pk_add_u16 [[ADD:v[0-9]+]], [[MAX]], 2
62define void @s_abs_v2i16_2(<2 x i16> addrspace(1)* %out, <2 x i16> %val) #0 {
63  %z0 = insertelement <2 x i16> undef, i16 0, i16 0
64  %z1 = insertelement <2 x i16> %z0, i16 0, i16 1
65  %t0 = insertelement <2 x i16> undef, i16 2, i16 0
66  %t1 = insertelement <2 x i16> %t0, i16 2, i16 1
67  %neg = sub <2 x i16> %z1, %val
68  %cond = icmp sgt <2 x i16> %val, %neg
69  %res = select <2 x i1> %cond, <2 x i16> %val, <2 x i16> %neg
70  %res2 = add <2 x i16> %res, %t1
71  store <2 x i16> %res2, <2 x i16> addrspace(1)* %out, align 4
72  ret void
73}
74
75; GCN-LABEL: {{^}}v_abs_v2i16_2:
76; GFX9: buffer_load_dword [[VAL:v[0-9]+]]
77; GFX9: v_pk_sub_i16 [[SUB:v[0-9]+]], 0, [[VAL]]
78; GFX9: v_pk_max_i16 [[MAX:v[0-9]+]], [[VAL]], [[SUB]]
79; GFX9: v_pk_add_u16 [[ADD:v[0-9]+]], [[MAX]], 2
80define void @v_abs_v2i16_2(<2 x i16> addrspace(1)* %out, <2 x i16> addrspace(1)* %src) #0 {
81  %z0 = insertelement <2 x i16> undef, i16 0, i16 0
82  %z1 = insertelement <2 x i16> %z0, i16 0, i16 1
83  %t0 = insertelement <2 x i16> undef, i16 2, i16 0
84  %t1 = insertelement <2 x i16> %t0, i16 2, i16 1
85  %val = load <2 x i16>, <2 x i16> addrspace(1)* %src, align 4
86  %neg = sub <2 x i16> %z1, %val
87  %cond = icmp sgt <2 x i16> %val, %neg
88  %res = select <2 x i1> %cond, <2 x i16> %val, <2 x i16> %neg
89  %res2 = add <2 x i16> %res, %t1
90  store <2 x i16> %res2, <2 x i16> addrspace(1)* %out, align 4
91  ret void
92}
93
94; GCN-LABEL: {{^}}s_abs_v4i16:
95; GFX9: s_load_dword [[VAL0:s[0-9]+]]
96; GFX9: s_load_dword [[VAL1:s[0-9]+]]
97; GFX9-DAG: v_pk_sub_i16 [[SUB0:v[0-9]+]], 0, [[VAL0]]
98; GFX9-DAG: v_pk_max_i16 [[MAX0:v[0-9]+]], [[VAL0]], [[SUB0]]
99; GFX9-DAG: v_pk_add_u16 [[ADD0:v[0-9]+]], [[MAX0]], 2
100
101; GFX9-DAG: v_pk_sub_i16 [[SUB1:v[0-9]+]], 0, [[VAL1]]
102; GFX9-DAG: v_pk_max_i16 [[MAX1:v[0-9]+]], [[VAL1]], [[SUB1]]
103; GFX9-DAG: v_pk_add_u16 [[ADD1:v[0-9]+]], [[MAX1]], 2
104define void @s_abs_v4i16(<4 x i16> addrspace(1)* %out, <4 x i16> %val) #0 {
105  %z0 = insertelement <4 x i16> undef, i16 0, i16 0
106  %z1 = insertelement <4 x i16> %z0, i16 0, i16 1
107  %z2 = insertelement <4 x i16> %z1, i16 0, i16 2
108  %z3 = insertelement <4 x i16> %z2, i16 0, i16 3
109  %t0 = insertelement <4 x i16> undef, i16 2, i16 0
110  %t1 = insertelement <4 x i16> %t0, i16 2, i16 1
111  %t2 = insertelement <4 x i16> %t1, i16 2, i16 2
112  %t3 = insertelement <4 x i16> %t2, i16 2, i16 3
113  %neg = sub <4 x i16> %z3, %val
114  %cond = icmp sgt <4 x i16> %val, %neg
115  %res = select <4 x i1> %cond, <4 x i16> %val, <4 x i16> %neg
116  %res2 = add <4 x i16> %res, %t3
117  store <4 x i16> %res2, <4 x i16> addrspace(1)* %out, align 4
118  ret void
119}
120
121; GCN-LABEL: {{^}}v_abs_v4i16:
122; GFX9: buffer_load_dwordx2 v{{\[}}[[VAL0:[0-9]+]]:[[VAL1:[0-9]+]]{{\]}}
123
124; GFX9-DAG: v_pk_sub_i16 [[SUB0:v[0-9]+]], 0, v[[VAL0]]
125; GFX9-DAG: v_pk_max_i16 [[MAX0:v[0-9]+]], v[[VAL0]], [[SUB0]]
126; GFX9-DAG: v_pk_add_u16 [[ADD0:v[0-9]+]], [[MAX0]], 2
127
128; GFX9-DAG: v_pk_sub_i16 [[SUB1:v[0-9]+]], 0, v[[VAL1]]
129; GFX9-DAG: v_pk_max_i16 [[MAX1:v[0-9]+]], v[[VAL1]], [[SUB1]]
130; GFX9-DAG: v_pk_add_u16 [[ADD1:v[0-9]+]], [[MAX1]], 2
131define void @v_abs_v4i16(<4 x i16> addrspace(1)* %out, <4 x i16> addrspace(1)* %src) #0 {
132  %z0 = insertelement <4 x i16> undef, i16 0, i16 0
133  %z1 = insertelement <4 x i16> %z0, i16 0, i16 1
134  %z2 = insertelement <4 x i16> %z1, i16 0, i16 2
135  %z3 = insertelement <4 x i16> %z2, i16 0, i16 3
136  %t0 = insertelement <4 x i16> undef, i16 2, i16 0
137  %t1 = insertelement <4 x i16> %t0, i16 2, i16 1
138  %t2 = insertelement <4 x i16> %t1, i16 2, i16 2
139  %t3 = insertelement <4 x i16> %t2, i16 2, i16 3
140  %val = load <4 x i16>, <4 x i16> addrspace(1)* %src, align 4
141  %neg = sub <4 x i16> %z3, %val
142  %cond = icmp sgt <4 x i16> %val, %neg
143  %res = select <4 x i1> %cond, <4 x i16> %val, <4 x i16> %neg
144  %res2 = add <4 x i16> %res, %t3
145  store <4 x i16> %res2, <4 x i16> addrspace(1)* %out, align 4
146  ret void
147}
148
149; GCN-LABEL: {{^}}s_min_max_v2i16:
150define void @s_min_max_v2i16(<2 x i16> addrspace(1)* %out0, <2 x i16> addrspace(1)* %out1, <2 x i16> %val0, <2 x i16> %val1) #0 {
151  %cond0 = icmp sgt <2 x i16> %val0, %val1
152  %sel0 = select <2 x i1> %cond0, <2 x i16> %val0, <2 x i16> %val1
153  %sel1 = select <2 x i1> %cond0, <2 x i16> %val1, <2 x i16> %val0
154
155  store volatile <2 x i16> %sel0, <2 x i16> addrspace(1)* %out0, align 4
156  store volatile <2 x i16> %sel1, <2 x i16> addrspace(1)* %out1, align 4
157  ret void
158}
159
160; GCN-LABEL: {{^}}v_min_max_v2i16:
161define void @v_min_max_v2i16(<2 x i16> addrspace(1)* %out0, <2 x i16> addrspace(1)* %out1, <2 x i16> addrspace(1)* %ptr0, <2 x i16> addrspace(1)* %ptr1) #0 {
162  %val0 = load volatile <2 x i16>, <2 x i16> addrspace(1)* %ptr0
163  %val1 = load volatile <2 x i16>, <2 x i16> addrspace(1)* %ptr1
164
165  %cond0 = icmp sgt <2 x i16> %val0, %val1
166  %sel0 = select <2 x i1> %cond0, <2 x i16> %val0, <2 x i16> %val1
167  %sel1 = select <2 x i1> %cond0, <2 x i16> %val1, <2 x i16> %val0
168
169  store volatile <2 x i16> %sel0, <2 x i16> addrspace(1)* %out0, align 4
170  store volatile <2 x i16> %sel1, <2 x i16> addrspace(1)* %out1, align 4
171  ret void
172}
173
174; GCN-LABEL: {{^}}s_min_max_v4i32:
175define void @s_min_max_v4i32(<4 x i16> addrspace(1)* %out0, <4 x i16> addrspace(1)* %out1, <4 x i16> %val0, <4 x i16> %val1) #0 {
176  %cond0 = icmp sgt <4 x i16> %val0, %val1
177  %sel0 = select <4 x i1> %cond0, <4 x i16> %val0, <4 x i16> %val1
178  %sel1 = select <4 x i1> %cond0, <4 x i16> %val1, <4 x i16> %val0
179
180  store volatile <4 x i16> %sel0, <4 x i16> addrspace(1)* %out0, align 4
181  store volatile <4 x i16> %sel1, <4 x i16> addrspace(1)* %out1, align 4
182  ret void
183}
184
185; GCN-LABEL: {{^}}v_min_max_v2i16_user:
186define void @v_min_max_v2i16_user(<2 x i16> addrspace(1)* %out0, <2 x i16> addrspace(1)* %out1, <2 x i16> addrspace(1)* %ptr0, <2 x i16> addrspace(1)* %ptr1) #0 {
187  %val0 = load volatile <2 x i16>, <2 x i16> addrspace(1)* %ptr0
188  %val1 = load volatile <2 x i16>, <2 x i16> addrspace(1)* %ptr1
189
190  %cond0 = icmp sgt <2 x i16> %val0, %val1
191  %sel0 = select <2 x i1> %cond0, <2 x i16> %val0, <2 x i16> %val1
192  %sel1 = select <2 x i1> %cond0, <2 x i16> %val1, <2 x i16> %val0
193
194  store volatile <2 x i16> %sel0, <2 x i16> addrspace(1)* %out0, align 4
195  store volatile <2 x i16> %sel1, <2 x i16> addrspace(1)* %out1, align 4
196  store volatile <2 x i1> %cond0, <2 x i1> addrspace(1)* undef
197  ret void
198}
199
200; GCN-LABEL: {{^}}u_min_max_v2i16:
201; GFX9: v_pk_max_u16 v{{[0-9]+}}, v{{[0-9]+}}, s{{[0-9]+}}
202; GFX9: v_pk_min_u16 v{{[0-9]+}}, s{{[0-9]+}}, v{{[0-9]+}}
203define void @u_min_max_v2i16(<2 x i16> addrspace(1)* %out0, <2 x i16> addrspace(1)* %out1, <2 x i16> %val0, <2 x i16> %val1) nounwind {
204  %cond0 = icmp ugt <2 x i16> %val0, %val1
205  %sel0 = select <2 x i1> %cond0, <2 x i16> %val0, <2 x i16> %val1
206  %sel1 = select <2 x i1> %cond0, <2 x i16> %val1, <2 x i16> %val0
207
208  store volatile <2 x i16> %sel0, <2 x i16> addrspace(1)* %out0, align 4
209  store volatile <2 x i16> %sel1, <2 x i16> addrspace(1)* %out1, align 4
210  ret void
211}
212
213declare i32 @llvm.amdgcn.workitem.id.x() #1
214
215attributes #0 = { nounwind }
216attributes #1 = { nounwind readnone }
217