blob: 59629140044e7ee6d95bbf0568b14d626baacbcb [file] [log] [blame]
Krzysztof Parzyszek046090d2018-03-12 14:01:28 +00001; RUN: llc -march=hexagon -O2 -debug-only=pipeliner < %s -o - 2>&1 > /dev/null | FileCheck %s
2
3; Test that the phi in the first epilog block is getter the correct
4; value from the kernel block. In this bug, the phi was using the value
5; defined in the loop instead of the Phi valued defined in the kernel.
6; We need to use the kernel's phi value (if the Phi in the kernel is the
7; last definition).
8
9; CHECK: New block
10; CHECK: %[[REG:([0-9]+)]]:intregs = PHI %{{.*}}, %[[REG1:([0-9]+)]]
11; CHECK: %[[REG1]]:intregs = A2_addi
12; CHECK: epilog:
13; CHECK: %{{[0-9]+}}:intregs = PHI %{{.*}}, %[[REG]]
14
15define void @f0(i32 %a0, i32 %a1) #0 {
16b0:
17 %v0 = icmp sgt i32 %a0, 64
18 br i1 %v0, label %b1, label %b3
19
20b1: ; preds = %b0
21 br label %b2
22
23b2: ; preds = %b2, %b1
24 %v1 = phi i32 [ %a0, %b1 ], [ %v13, %b2 ]
25 %v2 = phi <16 x i32>* [ null, %b1 ], [ %v3, %b2 ]
26 %v3 = getelementptr inbounds <16 x i32>, <16 x i32>* %v2, i32 1
27 %v4 = load <16 x i32>, <16 x i32>* %v2, align 64
28 %v5 = load <16 x i32>, <16 x i32>* undef, align 64
29 %v6 = tail call <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32> %v5, <16 x i32> undef, i32 1)
30 %v7 = tail call <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32> undef, <16 x i32> %v6)
31 %v8 = tail call <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32> undef, <32 x i32> %v7, i32 undef)
32 %v9 = tail call <32 x i32> @llvm.hexagon.V6.vdmpybus.dv.acc(<32 x i32> %v8, <32 x i32> undef, i32 undef)
33 %v10 = tail call <32 x i32> @llvm.hexagon.V6.vmpybus.acc(<32 x i32> %v9, <16 x i32> zeroinitializer, i32 undef)
34 %v11 = tail call <16 x i32> @llvm.hexagon.V6.lo(<32 x i32> %v10)
35 %v12 = tail call <16 x i32> @llvm.hexagon.V6.vasrhubsat(<16 x i32> undef, <16 x i32> %v11, i32 %a1)
36 store <16 x i32> %v12, <16 x i32>* null, align 64
37 %v13 = add nsw i32 %v1, -64
38 %v14 = icmp sgt i32 %v13, 64
39 br i1 %v14, label %b2, label %b3
40
41b3: ; preds = %b2, %b0
42 ret void
43}
44
45; Function Attrs: nounwind readnone
46declare <16 x i32> @llvm.hexagon.V6.valignbi(<16 x i32>, <16 x i32>, i32) #0
47
48; Function Attrs: nounwind readnone
49declare <32 x i32> @llvm.hexagon.V6.vcombine(<16 x i32>, <16 x i32>) #0
50
51; Function Attrs: nounwind readnone
52declare <32 x i32> @llvm.hexagon.V6.vdmpybus.dv.acc(<32 x i32>, <32 x i32>, i32) #0
53
54; Function Attrs: nounwind readnone
55declare <32 x i32> @llvm.hexagon.V6.vmpabus.acc(<32 x i32>, <32 x i32>, i32) #0
56
57; Function Attrs: nounwind readnone
58declare <32 x i32> @llvm.hexagon.V6.vmpybus.acc(<32 x i32>, <16 x i32>, i32) #0
59
60; Function Attrs: nounwind readnone
61declare <16 x i32> @llvm.hexagon.V6.vasrhubsat(<16 x i32>, <16 x i32>, i32) #0
62
63; Function Attrs: nounwind readnone
64declare <16 x i32> @llvm.hexagon.V6.lo(<32 x i32>) #0
65
66attributes #0 = { nounwind readnone "target-cpu"="hexagonv60" "target-features"="+hvxv60,+hvx-length64b" }