1; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx900 -mattr=-flat-for-global -denormal-fp-math=preserve-sign -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GFX9 %s
2; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
3; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=CI %s
4
5
6; GCN-LABEL: {{^}}s_pack_v2f16:
7; GFX9: s_load_dword [[VAL0:s[0-9]+]]
8; GFX9: s_load_dword [[VAL1:s[0-9]+]]
9; GFX9: s_pack_ll_b32_b16 [[PACKED:s[0-9]+]], [[VAL0]], [[VAL1]]
10; GFX9: ; use [[PACKED]]
11define amdgpu_kernel void @s_pack_v2f16(i32 addrspace(4)* %in0, i32 addrspace(4)* %in1) #0 {
12  %val0 = load volatile i32, i32 addrspace(4)* %in0
13  %val1 = load volatile i32, i32 addrspace(4)* %in1
14  %lo.i = trunc i32 %val0 to i16
15  %hi.i = trunc i32 %val1 to i16
16  %lo = bitcast i16 %lo.i to half
17  %hi = bitcast i16 %hi.i to half
18  %vec.0 = insertelement <2 x half> undef, half %lo, i32 0
19  %vec.1 = insertelement <2 x half> %vec.0, half %hi, i32 1
20  %vec.i32 = bitcast <2 x half> %vec.1 to i32
21
22  call void asm sideeffect "; use $0", "s"(i32 %vec.i32) #0
23  ret void
24}
25
26; GCN-LABEL: {{^}}s_pack_v2f16_imm_lo:
27; GFX9: s_load_dword [[VAL1:s[0-9]+]]
28; GFX9: s_pack_ll_b32_b16 [[PACKED:s[0-9]+]], 0x1234, [[VAL1]]
29; GFX9: ; use [[PACKED]]
30define amdgpu_kernel void @s_pack_v2f16_imm_lo(i32 addrspace(4)* %in1) #0 {
31  %val1 = load i32, i32 addrspace(4)* %in1
32  %hi.i = trunc i32 %val1 to i16
33  %hi = bitcast i16 %hi.i to half
34  %vec.0 = insertelement <2 x half> undef, half 0xH1234, i32 0
35  %vec.1 = insertelement <2 x half> %vec.0, half %hi, i32 1
36  %vec.i32 = bitcast <2 x half> %vec.1 to i32
37
38  call void asm sideeffect "; use $0", "s"(i32 %vec.i32) #0
39  ret void
40}
41
42; GCN-LABEL: {{^}}s_pack_v2f16_imm_hi:
43; GFX9: s_load_dword [[VAL0:s[0-9]+]]
44; GFX9: s_pack_ll_b32_b16 [[PACKED:s[0-9]+]], [[VAL0]], 0x1234
45; GFX9: ; use [[PACKED]]
46define amdgpu_kernel void @s_pack_v2f16_imm_hi(i32 addrspace(4)* %in0) #0 {
47  %val0 = load i32, i32 addrspace(4)* %in0
48  %lo.i = trunc i32 %val0 to i16
49  %lo = bitcast i16 %lo.i to half
50  %vec.0 = insertelement <2 x half> undef, half %lo, i32 0
51  %vec.1 = insertelement <2 x half> %vec.0, half 0xH1234, i32 1
52  %vec.i32 = bitcast <2 x half> %vec.1 to i32
53
54  call void asm sideeffect "; use $0", "s"(i32 %vec.i32) #0
55  ret void
56}
57
58; GCN-LABEL: {{^}}v_pack_v2f16:
59; GFX9: global_load_dword [[VAL0:v[0-9]+]]
60; GFX9: global_load_dword [[VAL1:v[0-9]+]]
61
62; GFX9: v_and_b32_e32 [[ELT0:v[0-9]+]], 0xffff, [[VAL0]]
63; GFX9: v_lshl_or_b32 [[PACKED:v[0-9]+]], [[VAL1]], 16, [[ELT0]]
64; GFX9: ; use [[PACKED]]
65define amdgpu_kernel void @v_pack_v2f16(i32 addrspace(1)* %in0, i32 addrspace(1)* %in1) #0 {
66  %tid = call i32 @llvm.amdgcn.workitem.id.x()
67  %tid.ext = sext i32 %tid to i64
68  %in0.gep = getelementptr inbounds i32, i32 addrspace(1)* %in0, i64 %tid.ext
69  %in1.gep = getelementptr inbounds i32, i32 addrspace(1)* %in1, i64 %tid.ext
70  %val0 = load volatile i32, i32 addrspace(1)* %in0.gep
71  %val1 = load volatile i32, i32 addrspace(1)* %in1.gep
72  %lo.i = trunc i32 %val0 to i16
73  %hi.i = trunc i32 %val1 to i16
74  %lo = bitcast i16 %lo.i to half
75  %hi = bitcast i16 %hi.i to half
76  %vec.0 = insertelement <2 x half> undef, half %lo, i32 0
77  %vec.1 = insertelement <2 x half> %vec.0, half %hi, i32 1
78  %vec.i32 = bitcast <2 x half> %vec.1 to i32
79  call void asm sideeffect "; use $0", "v"(i32 %vec.i32) #0
80  ret void
81}
82
83; GCN-LABEL: {{^}}v_pack_v2f16_user:
84; GFX9: global_load_dword [[VAL0:v[0-9]+]]
85; GFX9: global_load_dword [[VAL1:v[0-9]+]]
86
87; GFX9: v_and_b32_e32 [[ELT0:v[0-9]+]], 0xffff, [[VAL0]]
88; GFX9: v_lshl_or_b32 [[PACKED:v[0-9]+]], [[VAL1]], 16, [[ELT0]]
89
90; GFX9: v_add_u32_e32 v{{[0-9]+}}, 9, [[PACKED]]
91define amdgpu_kernel void @v_pack_v2f16_user(i32 addrspace(1)* %in0, i32 addrspace(1)* %in1) #0 {
92  %tid = call i32 @llvm.amdgcn.workitem.id.x()
93  %tid.ext = sext i32 %tid to i64
94  %in0.gep = getelementptr inbounds i32, i32 addrspace(1)* %in0, i64 %tid.ext
95  %in1.gep = getelementptr inbounds i32, i32 addrspace(1)* %in1, i64 %tid.ext
96  %val0 = load volatile i32, i32 addrspace(1)* %in0.gep
97  %val1 = load volatile i32, i32 addrspace(1)* %in1.gep
98  %lo.i = trunc i32 %val0 to i16
99  %hi.i = trunc i32 %val1 to i16
100  %lo = bitcast i16 %lo.i to half
101  %hi = bitcast i16 %hi.i to half
102  %vec.0 = insertelement <2 x half> undef, half %lo, i32 0
103  %vec.1 = insertelement <2 x half> %vec.0, half %hi, i32 1
104  %vec.i32 = bitcast <2 x half> %vec.1 to i32
105  %foo = add i32 %vec.i32, 9
106  store volatile i32 %foo, i32 addrspace(1)* undef
107  ret void
108}
109
110; GCN-LABEL: {{^}}v_pack_v2f16_imm_lo:
111; GFX9-DAG: global_load_dword [[VAL1:v[0-9]+]]
112
113; GFX9-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 0x1234{{$}}
114; GFX9: v_lshl_or_b32 [[PACKED:v[0-9]+]], [[VAL1]], 16, [[K]]
115; GFX9: ; use [[PACKED]]
116define amdgpu_kernel void @v_pack_v2f16_imm_lo(i32 addrspace(1)* %in1) #0 {
117  %tid = call i32 @llvm.amdgcn.workitem.id.x()
118  %tid.ext = sext i32 %tid to i64
119  %in1.gep = getelementptr inbounds i32, i32 addrspace(1)* %in1, i64 %tid.ext
120  %val1 = load volatile i32, i32 addrspace(1)* %in1.gep
121  %hi.i = trunc i32 %val1 to i16
122  %hi = bitcast i16 %hi.i to half
123  %vec.0 = insertelement <2 x half> undef, half 0xH1234, i32 0
124  %vec.1 = insertelement <2 x half> %vec.0, half %hi, i32 1
125  %vec.i32 = bitcast <2 x half> %vec.1 to i32
126  call void asm sideeffect "; use $0", "v"(i32 %vec.i32) #0
127  ret void
128}
129
130; GCN-LABEL: {{^}}v_pack_v2f16_inline_imm_lo:
131; GFX9-DAG: global_load_dword [[VAL1:v[0-9]+]]
132
133; GFX9-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 0x4400{{$}}
134; GFX9: v_lshl_or_b32 [[PACKED:v[0-9]+]], [[VAL1]], 16, [[K]]
135
136; GFX9: ; use [[PACKED]]
137define amdgpu_kernel void @v_pack_v2f16_inline_imm_lo(i32 addrspace(1)* %in1) #0 {
138  %tid = call i32 @llvm.amdgcn.workitem.id.x()
139  %tid.ext = sext i32 %tid to i64
140  %in1.gep = getelementptr inbounds i32, i32 addrspace(1)* %in1, i64 %tid.ext
141  %val1 = load volatile i32, i32 addrspace(1)* %in1.gep
142  %hi.i = trunc i32 %val1 to i16
143  %hi = bitcast i16 %hi.i to half
144  %vec.0 = insertelement <2 x half> undef, half 4.0, i32 0
145  %vec.1 = insertelement <2 x half> %vec.0, half %hi, i32 1
146  %vec.i32 = bitcast <2 x half> %vec.1 to i32
147  call void asm sideeffect "; use $0", "v"(i32 %vec.i32) #0
148  ret void
149}
150
151; GCN-LABEL: {{^}}v_pack_v2f16_imm_hi:
152; GFX9-DAG: global_load_dword [[VAL0:v[0-9]+]]
153
154; GFX9-DAG: s_movk_i32 [[K:s[0-9]+]], 0x1234
155; GFX9: v_and_b32_e32 [[MASKED:v[0-9]+]], 0xffff, [[VAL0]]
156; GFX9: v_lshl_or_b32 [[PACKED:v[0-9]+]], [[K]], 16, [[MASKED]]
157
158; GFX9: ; use [[PACKED]]
159define amdgpu_kernel void @v_pack_v2f16_imm_hi(i32 addrspace(1)* %in0) #0 {
160  %tid = call i32 @llvm.amdgcn.workitem.id.x()
161  %tid.ext = sext i32 %tid to i64
162  %in0.gep = getelementptr inbounds i32, i32 addrspace(1)* %in0, i64 %tid.ext
163  %val0 = load volatile i32, i32 addrspace(1)* %in0.gep
164  %lo.i = trunc i32 %val0 to i16
165  %lo = bitcast i16 %lo.i to half
166  %vec.0 = insertelement <2 x half> undef, half %lo, i32 0
167  %vec.1 = insertelement <2 x half> %vec.0, half 0xH1234, i32 1
168  %vec.i32 = bitcast <2 x half> %vec.1 to i32
169  call void asm sideeffect "; use $0", "v"(i32 %vec.i32) #0
170  ret void
171}
172
173; GCN-LABEL: {{^}}v_pack_v2f16_inline_f16imm_hi:
174; GFX9-DAG: global_load_dword [[VAL:v[0-9]+]]
175
176; GFX9-DAG: s_movk_i32 [[K:s[0-9]+]], 0x3c00
177; GFX9: v_and_b32_e32 [[MASKED:v[0-9]+]], 0xffff, [[VAL]]
178; GFX9: v_lshl_or_b32 [[PACKED:v[0-9]+]], [[K]], 16, [[MASKED]]
179
180; GFX9: ; use [[PACKED]]
181define amdgpu_kernel void @v_pack_v2f16_inline_f16imm_hi(i32 addrspace(1)* %in0) #0 {
182  %tid = call i32 @llvm.amdgcn.workitem.id.x()
183  %tid.ext = sext i32 %tid to i64
184  %in0.gep = getelementptr inbounds i32, i32 addrspace(1)* %in0, i64 %tid.ext
185  %val0 = load volatile i32, i32 addrspace(1)* %in0.gep
186  %lo.i = trunc i32 %val0 to i16
187  %lo = bitcast i16 %lo.i to half
188  %vec.0 = insertelement <2 x half> undef, half %lo, i32 0
189  %vec.1 = insertelement <2 x half> %vec.0, half 1.0, i32 1
190  %vec.i32 = bitcast <2 x half> %vec.1 to i32
191  call void asm sideeffect "; use $0", "v"(i32 %vec.i32) #0
192  ret void
193}
194
195; GCN-LABEL: {{^}}v_pack_v2f16_inline_imm_hi:
196; GFX9: global_load_dword [[VAL:v[0-9]+]]
197
198; GFX9: v_and_b32_e32 [[MASKED:v[0-9]+]], 0xffff, [[VAL]]
199; GFX9: v_lshl_or_b32 [[PACKED:v[0-9]+]], 64, 16, [[MASKED]]
200
201; GFX9: ; use [[PACKED]]
202define amdgpu_kernel void @v_pack_v2f16_inline_imm_hi(i32 addrspace(1)* %in0) #0 {
203  %tid = call i32 @llvm.amdgcn.workitem.id.x()
204  %tid.ext = sext i32 %tid to i64
205  %in0.gep = getelementptr inbounds i32, i32 addrspace(1)* %in0, i64 %tid.ext
206  %val0 = load volatile i32, i32 addrspace(1)* %in0.gep
207  %lo.i = trunc i32 %val0 to i16
208  %lo = bitcast i16 %lo.i to half
209  %vec.0 = insertelement <2 x half> undef, half %lo, i32 0
210  %vec.1 = insertelement <2 x half> %vec.0, half 0xH0040, i32 1
211  %vec.i32 = bitcast <2 x half> %vec.1 to i32
212  call void asm sideeffect "; use $0", "v"(i32 %vec.i32) #0
213  ret void
214}
215
216declare i32 @llvm.amdgcn.workitem.id.x() #1
217
218attributes #0 = { nounwind }
219attributes #1 = { nounwind readnone }
220