1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -march=hexagon < %s | FileCheck %s
4 ; In this testcase, when loads were moved close to users, they were actualy
5 ; moved right before the consuming stores. This was after the store group
6 ; was moved, so the loads and stores ended up being interleaved. This violated
7 ; the assumption in store realigning that all loads were available before the
8 ; first store, causing some code depending on the loads being inserted before
10 ; Just make sure that this compiles ok.
12 ; Function Attrs: nounwind
13 define void @f0(ptr noalias nocapture readonly %a0, ptr noalias nocapture %a1, i32 %a2) #0 {
15 ; CHECK: // %bb.0: // %b0
17 ; CHECK-NEXT: p0 = cmp.eq(r2,#0)
18 ; CHECK-NEXT: if (p0.new) jumpr:nt r31
20 ; CHECK-NEXT: .p2align 4
21 ; CHECK-NEXT: .LBB0_1: // %b2
22 ; CHECK-NEXT: // =>This Inner Loop Header: Depth=1
24 ; CHECK-NEXT: v0.cur = vmem(r0+#0)
25 ; CHECK-NEXT: vmem(r1+#0) = v0
28 ; CHECK-NEXT: v29.cur = vmem(r0+#1)
29 ; CHECK-NEXT: vmem(r1+#1) = v29
32 ; CHECK-NEXT: v30.cur = vmem(r0+#2)
33 ; CHECK-NEXT: vmem(r1+#2) = v30
36 ; CHECK-NEXT: r0 = add(r0,#256)
37 ; CHECK-NEXT: r1 = add(r1,#256)
38 ; CHECK-NEXT: v31.cur = vmem(r0+#3)
39 ; CHECK-NEXT: vmem(r1+#3) = v31
42 ; CHECK-NEXT: jump .LBB0_1
45 %v0 = icmp eq i32 %a2, 0
46 br i1 %v0, label %b3, label %b1
51 b2: ; preds = %b2, %b1
52 %v3 = phi ptr [ %v16, %b2 ], [ %a1, %b1 ]
53 %v4 = phi ptr [ %v11, %b2 ], [ %a0, %b1 ]
54 %v5 = getelementptr inbounds <16 x i32>, ptr %v4, i32 1
55 %v6 = load <16 x i32>, ptr %v4, align 64
56 %v7 = getelementptr inbounds <16 x i32>, ptr %v4, i32 2
57 %v8 = load <16 x i32>, ptr %v5, align 64
58 %v9 = getelementptr inbounds <16 x i32>, ptr %v4, i32 3
59 %v10 = load <16 x i32>, ptr %v7, align 64
60 %v11 = getelementptr inbounds <16 x i32>, ptr %v4, i32 4
61 %v12 = load <16 x i32>, ptr %v9, align 64
62 %v13 = getelementptr inbounds <16 x i32>, ptr %v3, i32 1
63 store <16 x i32> %v6, ptr %v3, align 64
64 %v14 = getelementptr inbounds <16 x i32>, ptr %v3, i32 2
65 store <16 x i32> %v8, ptr %v13, align 64
66 %v15 = getelementptr inbounds <16 x i32>, ptr %v3, i32 3
67 store <16 x i32> %v10, ptr %v14, align 64
68 %v16 = getelementptr inbounds <16 x i32>, ptr %v3, i32 4
69 store <16 x i32> %v12, ptr %v15, align 64
76 attributes #0 = { nounwind "target-features"="+hvxv65,+hvx-length64b" }