xref: /llvm-project/llvm/test/CodeGen/Hexagon/remove_lsr.ll (revision 2208c97c1bec2512d4e47b6223db6d95a7037956)
1; Test fix for PR-13709.
2; RUN: llc -mtriple=hexagon < %s | FileCheck %s
3; CHECK: f0
4; CHECK-NOT: lsr(r{{[0-9]+}}:{{[0-9]+}}, #32)
5; CHECK-NOT: lsr(r{{[0-9]+}}:{{[0-9]+}}, #32)
6
7; Convert the sequence
8; r17:16 = lsr(r11:10, #32)
9; .. = r16
10; into
11; r17:16 = lsr(r11:10, #32)
12; .. = r11
13; This makes the lsr instruction dead and it gets removed subsequently
14; by a dead code removal pass.
15
16
17%s.0 = type { i64 }
18%s.1 = type { i32 }
19
20define void @f0(ptr nocapture %a0, ptr nocapture %a1, ptr nocapture %a2, ptr nocapture %a3, ptr nocapture %a4) #0 {
21b0:
22  %v0 = getelementptr %s.0, ptr %a0, i32 1
23  %v1 = getelementptr %s.1, ptr %a2, i32 1
24  %v2 = getelementptr %s.1, ptr %a1, i32 1
25  %v3 = getelementptr i8, ptr %a4, i32 1
26  %v4 = getelementptr i8, ptr %a3, i32 1
27  br label %b1
28
29b1:                                               ; preds = %b1, %b0
30  %v5 = phi i32 [ %v38, %b1 ], [ 2, %b0 ]
31  %v6 = phi ptr [ %v37, %b1 ], [ %v4, %b0 ]
32  %v7 = phi ptr [ %v36, %b1 ], [ %v3, %b0 ]
33  %v8 = phi ptr [ %v35, %b1 ], [ %v2, %b0 ]
34  %v9 = phi ptr [ %v34, %b1 ], [ %v1, %b0 ]
35  %v10 = phi ptr [ %v33, %b1 ], [ %v0, %b0 ]
36  %v11 = phi i8 [ undef, %b0 ], [ %v30, %b1 ]
37  %v12 = phi i8 [ undef, %b0 ], [ %v29, %b1 ]
38  %v13 = phi i64 [ undef, %b0 ], [ %v28, %b1 ]
39  %v17 = tail call i64 @llvm.hexagon.A2.vsubhs(i64 0, i64 %v13)
40  %v18 = sext i8 %v12 to i32
41  %v19 = trunc i64 %v13 to i32
42  %v20 = trunc i64 %v17 to i32
43  %v21 = tail call i32 @llvm.hexagon.C2.mux(i32 %v18, i32 %v19, i32 %v20)
44  store i32 %v21, ptr %v8, align 4
45  %v22 = sext i8 %v11 to i32
46  %v23 = lshr i64 %v13, 32
47  %v24 = trunc i64 %v23 to i32
48  %v25 = lshr i64 %v17, 32
49  %v26 = trunc i64 %v25 to i32
50  %v27 = tail call i32 @llvm.hexagon.C2.mux(i32 %v22, i32 %v24, i32 %v26)
51  store i32 %v27, ptr %v9, align 4
52  %v28 = load i64, ptr %v10, align 8
53  %v29 = load i8, ptr %v6, align 1
54  %v30 = load i8, ptr %v7, align 1
55  %v31 = trunc i32 %v5 to i8
56  %v32 = icmp eq i8 %v31, 32
57  %v33 = getelementptr %s.0, ptr %v10, i32 1
58  %v34 = getelementptr %s.1, ptr %v9, i32 1
59  %v35 = getelementptr %s.1, ptr %v8, i32 1
60  %v36 = getelementptr i8, ptr %v7, i32 1
61  %v37 = getelementptr i8, ptr %v6, i32 1
62  %v38 = add i32 %v5, 1
63  br i1 %v32, label %b2, label %b1
64
65b2:                                               ; preds = %b1
66  ret void
67}
68
69declare i64 @llvm.hexagon.A2.vsubhs(i64, i64) #1
70declare i32 @llvm.hexagon.C2.mux(i32, i32, i32) #1
71
72attributes #0 = { nounwind "target-cpu"="hexagonv5" }
73attributes #1 = { nounwind readnone "target-cpu"="hexagonv5" }
74