1; RUN: opt -mtriple=amdgcn-- -loop-unroll -simplifycfg -sroa %s -S -o - | FileCheck %s
2; RUN: opt -mtriple=r600-- -loop-unroll -simplifycfg -sroa %s -S -o - | FileCheck %s
3
4target datalayout = "e-p:64:64-p1:64:64-p2:32:32-p3:32:32-p4:64:64-p5:32:32-p6:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024-v2048:2048-n32:64-S32-A5"
5
6; This test contains a simple loop that initializes an array declared in
7; private memory.  We want to make sure these kinds of loops are always
8; unrolled, because private memory is slow.
9
10; CHECK-LABEL: @private_memory
11; CHECK-NOT: alloca
12; CHECK: store i32 5, i32 addrspace(1)* %out
13define amdgpu_kernel void @private_memory(i32 addrspace(1)* %out) {
14entry:
15  %0 = alloca [32 x i32], addrspace(5)
16  br label %loop.header
17
18loop.header:
19  %counter = phi i32 [0, %entry], [%inc, %loop.inc]
20  br label %loop.body
21
22loop.body:
23  %ptr = getelementptr [32 x i32], [32 x i32] addrspace(5)* %0, i32 0, i32 %counter
24  store i32 %counter, i32 addrspace(5)* %ptr
25  br label %loop.inc
26
27loop.inc:
28  %inc = add i32 %counter, 1
29  %1 = icmp sge i32 %counter, 32
30  br i1 %1, label  %exit, label %loop.header
31
32exit:
33  %2 = getelementptr [32 x i32], [32 x i32] addrspace(5)* %0, i32 0, i32 5
34  %3 = load i32, i32 addrspace(5)* %2
35  store i32 %3, i32 addrspace(1)* %out
36  ret void
37}
38
39; Check that loop is unrolled for local memory references
40
41; CHECK-LABEL: @local_memory
42; CHECK: getelementptr i32, i32 addrspace(1)* %out, i32 128
43; CHECK-NEXT: store
44; CHECK-NEXT: ret
45define amdgpu_kernel void @local_memory(i32 addrspace(1)* %out, i32 addrspace(3)* %lds) {
46entry:
47  br label %loop.header
48
49loop.header:
50  %counter = phi i32 [0, %entry], [%inc, %loop.inc]
51  br label %loop.body
52
53loop.body:
54  %ptr_lds = getelementptr i32, i32 addrspace(3)* %lds, i32 %counter
55  %val = load i32, i32 addrspace(3)* %ptr_lds
56  %ptr_out = getelementptr i32, i32 addrspace(1)* %out, i32 %counter
57  store i32 %val, i32 addrspace(1)* %ptr_out
58  br label %loop.inc
59
60loop.inc:
61  %inc = add i32 %counter, 1
62  %cond = icmp sge i32 %counter, 128
63  br i1 %cond, label  %exit, label %loop.header
64
65exit:
66  ret void
67}
68
69; Check that a loop with if inside completely unrolled to eliminate phi and if
70
71; CHECK-LABEL: @unroll_for_if
72; CHECK: entry:
73; CHECK-NEXT: getelementptr
74; CHECK-NEXT: store
75; CHECK-NEXT: getelementptr
76; CHECK-NEXT: store
77; CHECK-NOT: br
78define amdgpu_kernel void @unroll_for_if(i32 addrspace(5)* %a) {
79entry:
80  br label %for.body
81
82for.body:                                         ; preds = %entry, %for.inc
83  %i1 = phi i32 [ 0, %entry ], [ %inc, %for.inc ]
84  %and = and i32 %i1, 1
85  %tobool = icmp eq i32 %and, 0
86  br i1 %tobool, label %for.inc, label %if.then
87
88if.then:                                          ; preds = %for.body
89  %0 = sext i32 %i1 to i64
90  %arrayidx = getelementptr inbounds i32, i32 addrspace(5)* %a, i64 %0
91  store i32 0, i32 addrspace(5)* %arrayidx, align 4
92  br label %for.inc
93
94for.inc:                                          ; preds = %for.body, %if.then
95  %inc = add nuw nsw i32 %i1, 1
96  %cmp = icmp ult i32 %inc, 48
97  br i1 %cmp, label %for.body, label %for.end
98
99for.end:                                          ; preds = %for.cond
100  ret void
101}
102
103; Check that runtime unroll is enabled for local memory references
104
105; CHECK-LABEL: @local_memory_runtime
106; CHECK: loop.header:
107; CHECK: load i32, i32 addrspace(3)*
108; CHECK: load i32, i32 addrspace(3)*
109; CHECK: br i1
110; CHECK: loop.header.epil
111; CHECK: load i32, i32 addrspace(3)*
112; CHECK: ret
113define amdgpu_kernel void @local_memory_runtime(i32 addrspace(1)* %out, i32 addrspace(3)* %lds, i32 %n) {
114entry:
115  br label %loop.header
116
117loop.header:
118  %counter = phi i32 [0, %entry], [%inc, %loop.inc]
119  br label %loop.body
120
121loop.body:
122  %ptr_lds = getelementptr i32, i32 addrspace(3)* %lds, i32 %counter
123  %val = load i32, i32 addrspace(3)* %ptr_lds
124  %ptr_out = getelementptr i32, i32 addrspace(1)* %out, i32 %counter
125  store i32 %val, i32 addrspace(1)* %ptr_out
126  br label %loop.inc
127
128loop.inc:
129  %inc = add i32 %counter, 1
130  %cond = icmp sge i32 %counter, %n
131  br i1 %cond, label  %exit, label %loop.header
132
133exit:
134  ret void
135}
136