xref: /llvm-project/llvm/test/CodeGen/Hexagon/fp_latency.ll (revision 2208c97c1bec2512d4e47b6223db6d95a7037956)
1; RUN: llc -O2 -mtriple=hexagon -fp-contract=fast -pipeliner-prune-loop-carried=false < %s | FileCheck %s
2
3; Test that there is 1 packet between the FP result and its use.
4
5; CHECK: loop0([[LOOP0:.LBB[0-9_]+]],
6; CHECK: [[LOOP0]]
7; CHECK: [[REG0:(r[0-9]+)]] += sfmpy(r{{[0-9]+}},r{{[0-9]+}})
8; CHECK: }
9; CHECK: }
10; CHECK: r{{[0-9]+}} = {{.*}}[[REG0]]
11
12; Function Attrs: nounwind readnone
13define void @f0(i32 %a0, i32 %a1) #0 {
14b0:
15  %v0 = alloca [1000 x float], align 64
16  %v1 = alloca [1000 x float], align 64
17  %v2 = alloca [1000 x float], align 64
18  %v3 = alloca [1000 x float], align 64
19  call void @llvm.lifetime.start.p0(i64 4000, ptr %v0) #2
20  call void @llvm.lifetime.start.p0(i64 4000, ptr %v1) #2
21  call void @llvm.lifetime.start.p0(i64 4000, ptr %v2) #2
22  call void @llvm.lifetime.start.p0(i64 4000, ptr %v3) #2
23  %v8 = icmp sgt i32 %a1, 0
24  %v9 = add i32 %a1, -1
25  br label %b1
26
27b1:                                               ; preds = %b3, %b0
28  %v11 = phi i32 [ 0, %b0 ], [ %v34, %b3 ]
29  br i1 %v8, label %b2, label %b3
30
31b2:                                               ; preds = %b2, %b1
32  %v12 = phi ptr [ %v33, %b2 ], [ %v3, %b1 ]
33  %v13 = phi i32 [ %v31, %b2 ], [ 0, %b1 ]
34  %v14 = mul nsw i32 %v13, %a1
35  %v15 = add nsw i32 %v14, %v11
36  %v16 = getelementptr inbounds [1000 x float], ptr %v1, i32 0, i32 %v15
37  %v17 = load float, ptr %v16, align 4, !tbaa !0
38  %v18 = fmul float %v17, %v17
39  %v19 = mul nsw i32 %v13, 25
40  %v20 = add nsw i32 %v19, %v11
41  %v21 = getelementptr inbounds [1000 x float], ptr %v2, i32 0, i32 %v20
42  %v22 = load float, ptr %v21, align 4, !tbaa !0
43  %v23 = fmul float %v22, %v22
44  %v24 = fadd float %v18, %v23
45  %v25 = load float, ptr %v12, align 4, !tbaa !0
46  %v26 = fmul float %v25, %v25
47  %v27 = fadd float %v24, %v26
48  %v28 = getelementptr inbounds [1000 x float], ptr %v0, i32 0, i32 %v20
49  %v29 = load float, ptr %v28, align 4, !tbaa !0
50  %v30 = fadd float %v29, %v27
51  store float %v30, ptr %v28, align 4, !tbaa !0
52  %v31 = add nuw nsw i32 %v13, 1
53  %v32 = icmp eq i32 %v13, %v9
54  %v33 = getelementptr float, ptr %v12, i32 1
55  br i1 %v32, label %b3, label %b2
56
57b3:                                               ; preds = %b2, %b1
58  %v34 = add nuw nsw i32 %v11, 1
59  %v35 = icmp eq i32 %v34, 25
60  br i1 %v35, label %b4, label %b1
61
62b4:                                               ; preds = %b3
63  call void @llvm.lifetime.end.p0(i64 4000, ptr %v3) #2
64  call void @llvm.lifetime.end.p0(i64 4000, ptr %v2) #2
65  call void @llvm.lifetime.end.p0(i64 4000, ptr %v1) #2
66  call void @llvm.lifetime.end.p0(i64 4000, ptr %v0) #2
67  ret void
68}
69
70; Function Attrs: argmemonly nounwind
71declare void @llvm.lifetime.start.p0(i64, ptr nocapture) #1
72
73; Function Attrs: argmemonly nounwind
74declare void @llvm.lifetime.end.p0(i64, ptr nocapture) #1
75
76attributes #0 = { nounwind readnone "target-cpu"="hexagonv60" "target-features"="+hvxv60,+hvx-length64b" }
77attributes #1 = { argmemonly nounwind }
78attributes #2 = { nounwind }
79
80!0 = !{!1, !1, i64 0}
81!1 = !{!"float", !2, i64 0}
82!2 = !{!"omnipotent char", !3, i64 0}
83!3 = !{!"Simple C/C++ TBAA"}
84