1 ; RUN: llc -march=hexagon -hexagon-initial-cfg-cleanup=0 -pipeliner-experimental-cg=true -disable-cgp-delete-phis < %s | FileCheck %s
3 ; Test epilogue generation when reading loop-carried dependency from a previous
4 ; stage. The first epilogue should read value from iteration N-1 of the kernel.
7 ; CHECK: r{{[0-9]+}} = add([[REG0:r([0-9]+)]],#8)
8 ; CHECK: [[REG0]] = [[REG1:r([0-9]+)]]
10 ; CHECK: = add([[REG1]],#8)
12 ; Function Attrs: nounwind
13 define ptr @f0(ptr nocapture readonly %a0, i32 %a1) #0 {
15 %v0 = alloca [129 x i32], align 8
16 br i1 undef, label %b1, label %b3
21 b2: ; preds = %b2, %b1
22 %v1 = phi ptr [ %a0, %b1 ], [ %v2, %b2 ]
23 %v2 = phi ptr [ undef, %b1 ], [ %v15, %b2 ]
24 %v3 = phi ptr [ null, %b1 ], [ %v4, %b2 ]
25 %v4 = phi ptr [ null, %b1 ], [ %v14, %b2 ]
26 %v5 = phi i32 [ 0, %b1 ], [ %v13, %b2 ]
27 %v6 = phi ptr [ undef, %b1 ], [ %v12, %b2 ]
28 %v7 = load i16, ptr %v2, align 2
29 %v8 = sext i16 %v7 to i32
30 %v9 = call i32 @llvm.hexagon.M2.mpy.ll.s0(i32 %v8, i32 %v8) #2
31 %v10 = load i16, ptr %v6, align 2
32 %v11 = call i32 @llvm.hexagon.M2.mpy.acc.sat.ll.s0(i32 %v9, i32 undef, i32 undef) #2
33 store i32 %v11, ptr %v4, align 4
34 %v12 = getelementptr inbounds i16, ptr %v6, i32 -1
36 %v14 = getelementptr inbounds i32, ptr %v3, i32 2
37 %v15 = getelementptr inbounds i16, ptr %v1, i32 2
38 %v16 = icmp slt i32 %v13, %a1
39 br i1 %v16, label %b2, label %b3
41 b3: ; preds = %b2, %b0
42 %out = phi ptr [ null, %b0 ], [ %v14, %b2 ]
46 ; Function Attrs: nounwind readnone
47 declare i32 @llvm.hexagon.M2.mpy.ll.s0(i32, i32) #1
49 ; Function Attrs: nounwind readnone
50 declare i32 @llvm.hexagon.M2.mpy.acc.sat.ll.s0(i32, i32, i32) #1
52 attributes #0 = { nounwind "target-cpu"="hexagonv60" }
53 attributes #1 = { nounwind readnone }
54 attributes #2 = { nounwind }