1; RUN: llc -march=amdgcn -verify-machineinstrs -disable-block-placement < %s | FileCheck %s
2
3; Check that invariant compare is hoisted out of the loop.
4; At the same time condition shall not be serialized into a VGPR and deserialized later
5; using another v_cmp + v_cndmask, but used directly in s_and_saveexec_b64.
6
7; CHECK: v_cmp_{{..}}_u32_e{{32|64}} [[COND:s\[[0-9]+:[0-9]+\]|vcc]]
8; CHECK: BB0_1:
9; CHECK-NOT: v_cmp
10; CHECK_NOT: v_cndmask
11; CHECK: s_and_saveexec_b64 s[{{[[0-9]+:[0-9]+}}], [[COND]]
12; CHECK: ; %bb.2:
13
14define amdgpu_kernel void @hoist_cond(float addrspace(1)* nocapture %arg, float addrspace(1)* noalias nocapture readonly %arg1, i32 %arg3, i32 %arg4) {
15bb:
16  %tmp = tail call i32 @llvm.amdgcn.workitem.id.x() #0
17  %tmp5 = icmp ult i32 %tmp, %arg3
18  br label %bb1
19
20bb1:                                              ; preds = %bb3, %bb
21  %tmp7 = phi i32 [ %arg4, %bb ], [ %tmp16, %bb3 ]
22  %tmp8 = phi float [ 0.000000e+00, %bb ], [ %tmp15, %bb3 ]
23  br i1 %tmp5, label %bb2, label %bb3
24
25bb2:                                              ; preds = %bb1
26  %tmp10 = zext i32 %tmp7 to i64
27  %tmp11 = getelementptr inbounds float, float addrspace(1)* %arg1, i64 %tmp10
28  %tmp12 = load float, float addrspace(1)* %tmp11, align 4
29  br label %bb3
30
31bb3:                                             ; preds = %bb2, %bb1
32  %tmp14 = phi float [ %tmp12, %bb2 ], [ 0.000000e+00, %bb1 ]
33  %tmp15 = fadd float %tmp8, %tmp14
34  %tmp16 = add i32 %tmp7, -1
35  %tmp17 = icmp eq i32 %tmp16, 0
36  br i1 %tmp17, label %bb4, label %bb1
37
38bb4:                                             ; preds = %bb3
39  store float %tmp15, float addrspace(1)* %arg, align 4
40  ret void
41}
42
43; Function Attrs: nounwind readnone
44declare i32 @llvm.amdgcn.workitem.id.x() #0
45
46attributes #0 = { nounwind readnone }
47