1; RUN: opt -mtriple=amdgcn-- -O1 -S < %s | FileCheck %s --check-prefixes=FUNC,LOOP
2; RUN: opt -mtriple=amdgcn-- -passes='default<O1>' -S < %s | FileCheck %s --check-prefixes=FUNC,LOOP
3; RUN: opt -mtriple=amdgcn-- -O1 -S -disable-promote-alloca-to-vector < %s | FileCheck %s --check-prefixes=FUNC,FULL-UNROLL
4; RUN: opt -mtriple=amdgcn-- -passes='default<O1>' -S -disable-promote-alloca-to-vector < %s | FileCheck %s --check-prefixes=FUNC,FULL-UNROLL
5
6target datalayout = "A5"
7
8; This test contains a simple loop that initializes an array declared in
9; private memory. This loop would be fully unrolled if we could not SROA
10; the alloca. Check that we successfully eliminate it before the unroll,
11; so that we do not need to fully unroll it.
12
13; FUNC-LABEL: @private_memory
14; LOOP-NOT: alloca
15; LOOP: loop.header:
16; LOOP: br i1 %{{[^,]+}}, label %exit, label %loop.header
17
18; FULL-UNROLL: alloca
19; FULL-UNROLL-COUNT-256: store i32 {{[0-9]+}}, i32 addrspace(5)*
20; FULL-UNROLL-NOT: br
21
22; FUNC: store i32 %{{[^,]+}}, i32 addrspace(1)* %out
23define amdgpu_kernel void @private_memory(i32 addrspace(1)* %out, i32 %n) {
24entry:
25  %alloca = alloca [16 x i32], addrspace(5)
26  br label %loop.header
27
28loop.header:
29  %counter = phi i32 [0, %entry], [%inc, %loop.inc]
30  br label %loop.body
31
32loop.body:
33  %salt = xor i32 %counter, %n
34  %idx = and i32 %salt, 15
35  %ptr = getelementptr [16 x i32], [16 x i32] addrspace(5)* %alloca, i32 0, i32 %idx
36  store i32 %counter, i32 addrspace(5)* %ptr
37  br label %loop.inc
38
39loop.inc:
40  %inc = add i32 %counter, 1
41  %cmp = icmp sge i32 %counter, 255
42  br i1 %cmp, label  %exit, label %loop.header
43
44exit:
45  %gep = getelementptr [16 x i32], [16 x i32] addrspace(5)* %alloca, i32 0, i32 %n
46  %load = load i32, i32 addrspace(5)* %gep
47  store i32 %load, i32 addrspace(1)* %out
48  ret void
49}
50