1; RUN: llc -verify-machineinstrs -march=amdgcn < %s | FileCheck -check-prefix=GCN %s
2; RUN: llc -verify-machineinstrs -march=amdgcn -mcpu=tonga < %s | FileCheck -check-prefix=GCN %s
3
4; When a frame index offset is more than 12-bits, make sure we don't store
5; it in mubuf's offset field.
6
7; Also, make sure we use the same register for storing the scratch buffer addresss
8; for both stores. This register is allocated by the register scavenger, so we
9; should be able to reuse the same regiser for each scratch buffer access.
10
11; GCN-LABEL: {{^}}legal_offset_fi:
12; GCN: v_mov_b32_e32 [[OFFSET:v[0-9]+]], 0{{$}}
13; GCN: buffer_store_dword v{{[0-9]+}}, [[OFFSET]], s[{{[0-9]+}}:{{[0-9]+}}], s{{[0-9]+}} offen
14; GCN: v_mov_b32_e32 [[OFFSET]], 0x8000
15; GCN: buffer_store_dword v{{[0-9]+}}, [[OFFSET]], s[{{[0-9]+}}:{{[0-9]+}}], s{{[0-9]+}} offen{{$}}
16
17define void @legal_offset_fi(i32 addrspace(1)* %out, i32 %cond, i32 %if_offset, i32 %else_offset) {
18entry:
19  %scratch0 = alloca [8192 x i32]
20  %scratch1 = alloca [8192 x i32]
21
22  %scratchptr0 = getelementptr [8192 x i32], [8192 x i32]* %scratch0, i32 0, i32 0
23  store i32 1, i32* %scratchptr0
24
25  %scratchptr1 = getelementptr [8192 x i32], [8192 x i32]* %scratch1, i32 0, i32 0
26  store i32 2, i32* %scratchptr1
27
28  %cmp = icmp eq i32 %cond, 0
29  br i1 %cmp, label %if, label %else
30
31if:
32  %if_ptr = getelementptr [8192 x i32], [8192 x i32]* %scratch0, i32 0, i32 %if_offset
33  %if_value = load i32, i32* %if_ptr
34  br label %done
35
36else:
37  %else_ptr = getelementptr [8192 x i32], [8192 x i32]* %scratch1, i32 0, i32 %else_offset
38  %else_value = load i32, i32* %else_ptr
39  br label %done
40
41done:
42  %value = phi i32 [%if_value, %if], [%else_value, %else]
43  store i32 %value, i32 addrspace(1)* %out
44  ret void
45
46  ret void
47
48}
49
50; GCN-LABEL: {{^}}legal_offset_fi_offset
51; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+}}:{{[0-9]+}}], s{{[0-9]+}} offen
52; GCN: v_add_i32_e32 [[OFFSET:v[0-9]+]], vcc, 0x8000
53; GCN: buffer_store_dword v{{[0-9]+}}, [[OFFSET]], s[{{[0-9]+}}:{{[0-9]+}}], s{{[0-9]+}} offen{{$}}
54
55define void @legal_offset_fi_offset(i32 addrspace(1)* %out, i32 %cond, i32 addrspace(1)* %offsets, i32 %if_offset, i32 %else_offset) {
56entry:
57  %scratch0 = alloca [8192 x i32]
58  %scratch1 = alloca [8192 x i32]
59
60  %offset0 = load i32, i32 addrspace(1)* %offsets
61  %scratchptr0 = getelementptr [8192 x i32], [8192 x i32]* %scratch0, i32 0, i32 %offset0
62  store i32 %offset0, i32* %scratchptr0
63
64  %offsetptr1 = getelementptr i32, i32 addrspace(1)* %offsets, i32 1
65  %offset1 = load i32, i32 addrspace(1)* %offsetptr1
66  %scratchptr1 = getelementptr [8192 x i32], [8192 x i32]* %scratch1, i32 0, i32 %offset1
67  store i32 %offset1, i32* %scratchptr1
68
69  %cmp = icmp eq i32 %cond, 0
70  br i1 %cmp, label %if, label %else
71
72if:
73  %if_ptr = getelementptr [8192 x i32], [8192 x i32]* %scratch0, i32 0, i32 %if_offset
74  %if_value = load i32, i32* %if_ptr
75  br label %done
76
77else:
78  %else_ptr = getelementptr [8192 x i32], [8192 x i32]* %scratch1, i32 0, i32 %else_offset
79  %else_value = load i32, i32* %else_ptr
80  br label %done
81
82done:
83  %value = phi i32 [%if_value, %if], [%else_value, %else]
84  store i32 %value, i32 addrspace(1)* %out
85  ret void
86}
87
88; GCN-LABEL: @neg_vaddr_offset
89; We can't prove %offset is positive, so we must do the computation with the
90; immediate in an add instruction instead of folding offset and the immediate into
91; the store instruction.
92; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offen{{$}}
93define void @neg_vaddr_offset(i32 %offset) {
94entry:
95  %array = alloca [8192 x i32]
96  %ptr_offset = add i32 %offset, 4
97  %ptr = getelementptr [8192 x i32], [8192 x i32]* %array, i32 0, i32 %ptr_offset
98  store i32 0, i32* %ptr
99  ret void
100}
101
102; GCN-LABEL: @pos_vaddr_offse
103; GCN: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offen offset:16
104define void @pos_vaddr_offset(i32 addrspace(1)* %out, i32 %offset) {
105entry:
106  %array = alloca [8192 x i32]
107  %ptr = getelementptr [8192 x i32], [8192 x i32]* %array, i32 0, i32 4
108  store i32 0, i32* %ptr
109  %load_ptr = getelementptr [8192 x i32], [8192 x i32]* %array, i32 0, i32 %offset
110  %val = load i32, i32* %load_ptr
111  store i32 %val, i32 addrspace(1)* %out
112  ret void
113}
114