1; RUN: llc -march=hexagon -debug-only=pipeliner < %s -o - 2>&1 -pipeliner-experimental-cg=true | FileCheck %s
2; REQUIRES: asserts
3
4; Test that there is a chain edge between two dependent Phis.
5; The pipeliner tries to remove chains between unrelated Phis, but
6; was too aggressive in some cases. When this happens the two Phis may get
7; scheduled too far apart. In this case, the second Phi was scheduled in
8; the next stage.
9
10; CHECK: SU([[SU1:[0-9]+]]): %14:intregs = PHI %{{[0-9]+}}:intregs, %bb.0, %{{[0-9]+}}:intregs, %bb.1
11; CHECK: Successors:
12; CHECK: SU({{.*}}): Data Latency=0
13; CHECK: SU([[SU2:[0-9]+]]): Data Latency=0
14; CHECK: SU([[SU2]]):   %{{[0-9]+}}:intregs = PHI %{{[0-9]+}}:intregs, %bb.0, %14:intregs, %bb.1
15; CHECK: Predecessors:
16; CHECK: SU([[SU1]]): Data Latency=0
17
18%s.0 = type { i16, i8, i32, i8*, i8*, i8*, i8*, i8*, i8*, i32*, [2 x i32], i8*, i8*, i8*, %s.1, i8*, [8 x i8], i8 }
19%s.1 = type { i32, i16, i16 }
20%s.2 = type { i32, i32, i32, i32, i32, i32, i32, i32, i32 }
21
22@g0 = global i32 0, align 4
23@g1 = global i32 0, align 4
24@g2 = global i32 0, align 4
25@g3 = global i32 0, align 4
26@g4 = global i32 0, align 4
27@g5 = common global i32 0, align 4
28@g6 = external global %s.0
29@g7 = private unnamed_addr constant [4 x i8] c"%d\0A\00", align 1
30
31; Function Attrs: nounwind
32declare i32 @f0(%s.0* nocapture, i8* nocapture readonly, ...) #0
33
34; Function Attrs: nounwind
35define void @f1(%s.2* nocapture %a0, i32* nocapture readonly %a1, i32* nocapture readonly %a2, i16 signext %a3) #0 {
36b0:
37  %v0 = load i32, i32* %a2, align 4
38  %v1 = tail call i32 @llvm.hexagon.S2.asr.r.r.sat(i32 %v0, i32 2)
39  %v2 = tail call i32 @llvm.hexagon.A2.sath(i32 %v1)
40  store i32 0, i32* @g5, align 4
41  %v3 = load i32, i32* @g0, align 4
42  %v4 = load i32, i32* @g1, align 4
43  %v5 = load i32, i32* @g2, align 4
44  %v6 = load i32, i32* @g3, align 4
45  %v7 = load i32, i32* @g4, align 4
46  br label %b1
47
48b1:                                               ; preds = %b1, %b0
49  %v8 = phi i32 [ %v7, %b0 ], [ %v52, %b1 ]
50  %v9 = phi i32 [ %v6, %b0 ], [ %v50, %b1 ]
51  %v10 = phi i32 [ %v5, %b0 ], [ %v46, %b1 ]
52  %v11 = phi i32 [ %v4, %b0 ], [ %v44, %b1 ]
53  %v12 = phi i32 [ %v3, %b0 ], [ %v38, %b1 ]
54  %v13 = phi i32 [ 0, %b0 ], [ %v53, %b1 ]
55  %v14 = phi i32* [ %a2, %b0 ], [ %v26, %b1 ]
56  %v15 = phi i32* [ %a1, %b0 ], [ %v19, %b1 ]
57  %v16 = phi i32 [ %v2, %b0 ], [ %v32, %b1 ]
58  %v17 = phi i32 [ 0, %b0 ], [ %v25, %b1 ]
59  %v18 = phi i32 [ 0, %b0 ], [ %v16, %b1 ]
60  %v19 = getelementptr inbounds i32, i32* %v15, i32 1
61  %v20 = load i32, i32* %v15, align 4
62  %v21 = tail call i32 @llvm.hexagon.A2.asrh(i32 %v20)
63  %v22 = shl i32 %v21, 16
64  %v23 = ashr exact i32 %v22, 16
65  %v24 = tail call i32 @llvm.hexagon.S2.asr.r.r.sat(i32 %v23, i32 2)
66  %v25 = tail call i32 @llvm.hexagon.A2.sath(i32 %v24)
67  %v26 = getelementptr inbounds i32, i32* %v14, i32 1
68  %v27 = load i32, i32* %v14, align 4
69  %v28 = tail call i32 @llvm.hexagon.A2.asrh(i32 %v27)
70  %v29 = shl i32 %v28, 16
71  %v30 = ashr exact i32 %v29, 16
72  %v31 = tail call i32 @llvm.hexagon.S2.asr.r.r.sat(i32 %v30, i32 2)
73  %v32 = tail call i32 @llvm.hexagon.A2.sath(i32 %v31)
74  %v33 = shl i32 %v17, 16
75  %v34 = ashr exact i32 %v33, 16
76  %v35 = tail call i32 @llvm.hexagon.M2.mpy.nac.sat.ll.s1(i32 %v12, i32 %v34, i32 %v34)
77  %v36 = shl i32 %v16, 16
78  %v37 = ashr exact i32 %v36, 16
79  %v38 = tail call i32 @llvm.hexagon.M2.mpy.nac.sat.ll.s1(i32 %v35, i32 %v37, i32 %v37)
80  store i32 %v38, i32* @g0, align 4
81  %v39 = shl i32 %v25, 16
82  %v40 = ashr exact i32 %v39, 16
83  %v41 = tail call i32 @llvm.hexagon.M2.mpy.acc.sat.ll.s1(i32 %v11, i32 %v40, i32 %v34)
84  %v42 = shl i32 %v32, 16
85  %v43 = ashr exact i32 %v42, 16
86  %v44 = tail call i32 @llvm.hexagon.M2.mpy.acc.sat.ll.s1(i32 %v41, i32 %v43, i32 %v37)
87  store i32 %v44, i32* @g1, align 4
88  %v45 = tail call i32 @llvm.hexagon.M2.mpy.acc.sat.ll.s1(i32 %v10, i32 %v43, i32 %v34)
89  %v46 = tail call i32 @llvm.hexagon.M2.mpy.acc.sat.ll.s1(i32 %v45, i32 %v40, i32 %v37)
90  store i32 %v46, i32* @g2, align 4
91  %v47 = tail call i32 @llvm.hexagon.M2.mpy.acc.sat.ll.s1(i32 %v9, i32 %v40, i32 0)
92  %v48 = shl i32 %v18, 16
93  %v49 = ashr exact i32 %v48, 16
94  %v50 = tail call i32 @llvm.hexagon.M2.mpy.acc.sat.ll.s1(i32 %v47, i32 %v43, i32 %v49)
95  store i32 %v50, i32* @g3, align 4
96  %v51 = tail call i32 @llvm.hexagon.M2.mpy.acc.sat.ll.s1(i32 %v8, i32 %v43, i32 0)
97  %v52 = tail call i32 @llvm.hexagon.M2.mpy.acc.sat.ll.s1(i32 %v51, i32 %v40, i32 %v49)
98  store i32 %v52, i32* @g4, align 4
99  %v53 = add nsw i32 %v13, 1
100  %v54 = icmp slt i32 %v53, 4
101  store i32 %v53, i32* @g5, align 4
102  br i1 %v54, label %b1, label %b2
103
104b2:                                               ; preds = %b1
105  %v55 = tail call i32 (%s.0*, i8*, ...) @f0(%s.0* @g6, i8* getelementptr inbounds ([4 x i8], [4 x i8]* @g7, i32 0, i32 0), i32 %v46) #2
106  %v56 = load i32, i32* @g2, align 4
107  %v57 = getelementptr inbounds %s.2, %s.2* %a0, i32 0, i32 5
108  store i32 %v56, i32* %v57, align 4, !tbaa !0
109  ret void
110}
111
112; Function Attrs: nounwind readnone
113declare i32 @llvm.hexagon.A2.sath(i32) #1
114
115; Function Attrs: nounwind readnone
116declare i32 @llvm.hexagon.S2.asr.r.r.sat(i32, i32) #1
117
118; Function Attrs: nounwind readnone
119declare i32 @llvm.hexagon.A2.asrh(i32) #1
120
121; Function Attrs: nounwind readnone
122declare i32 @llvm.hexagon.M2.mpy.nac.sat.ll.s1(i32, i32, i32) #1
123
124; Function Attrs: nounwind readnone
125declare i32 @llvm.hexagon.M2.mpy.acc.sat.ll.s1(i32, i32, i32) #1
126
127attributes #0 = { nounwind "target-cpu"="hexagonv55" }
128attributes #1 = { nounwind readnone }
129attributes #2 = { nounwind }
130
131!0 = !{!1, !2, i64 20}
132!1 = !{!"", !2, i64 0, !2, i64 4, !2, i64 8, !2, i64 12, !2, i64 16, !2, i64 20, !2, i64 24, !2, i64 28, !2, i64 32}
133!2 = !{!"int", !3, i64 0}
134!3 = !{!"omnipotent char", !4, i64 0}
135!4 = !{!"Simple C/C++ TBAA"}
136