1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt < %s -gvn -S | FileCheck %s
3
4target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128-ni:1"
5target triple = "x86_64-unknown-linux-gnu"
6
7declare void @llvm.memset.p0i8.i8(i8*, i8, i32, i1)
8declare void @foo(i8*)
9
10define i8 @test(i1 %cmp) {
11; CHECK-LABEL: @test(
12; CHECK-NEXT:  entry:
13; CHECK-NEXT:    [[P:%.*]] = alloca i8
14; CHECK-NEXT:    store i8 5, i8* [[P]]
15; CHECK-NEXT:    br label [[HEADER:%.*]]
16; CHECK:       header:
17; CHECK-NEXT:    [[V:%.*]] = phi i8 [ 5, [[ENTRY:%.*]] ], [ -5, [[ALIVE:%.*]] ]
18; CHECK-NEXT:    [[I:%.*]] = phi i8 [ 0, [[ENTRY]] ], [ [[I_INC:%.*]], [[ALIVE]] ]
19; CHECK-NEXT:    br i1 [[CMP:%.*]], label [[ALIVE]], label [[DEAD:%.*]]
20; CHECK:       dead:
21; CHECK-NEXT:    call void @foo(i8* [[P]])
22; CHECK-NEXT:    [[I_1:%.*]] = add i8 [[I]], [[V]]
23; CHECK-NEXT:    br label [[ALIVE]]
24; CHECK:       alive:
25; CHECK-NEXT:    [[I_2:%.*]] = phi i8 [ [[I]], [[HEADER]] ], [ [[I_1]], [[DEAD]] ]
26; CHECK-NEXT:    store i8 -5, i8* [[P]]
27; CHECK-NEXT:    call void @llvm.memset.p0i8.i32(i8* align 1 [[P]], i8 0, i32 1, i1 false)
28; CHECK-NEXT:    [[I_INC]] = add i8 [[I_2]], 1
29; CHECK-NEXT:    [[CMP_LOOP:%.*]] = icmp ugt i8 [[I_INC]], 100
30; CHECK-NEXT:    br i1 [[CMP_LOOP]], label [[EXIT:%.*]], label [[HEADER]]
31; CHECK:       exit:
32; CHECK-NEXT:    ret i8 0
33;
34
35entry:
36  %p = alloca i8
37  %addr = getelementptr inbounds i8, i8* %p, i64 0
38  store i8 5, i8* %addr
39  br label %header
40header:
41  %i = phi i8 [0, %entry], [%i.inc, %backedge]
42  br i1 %cmp, label %alive, label %dead
43dead:
44  call void @foo(i8* %p)
45  %v = load i8, i8* %addr, !invariant.load !1
46  %i.1 = add i8 %i, %v
47  br label %alive
48alive:
49  %i.2 = phi i8 [%i, %header], [%i.1, %dead]
50  store i8 -5, i8* %addr
51  br label %backedge
52backedge:
53  call void @llvm.memset.p0i8.i8(i8 * align 1 %p, i8 0, i32 1, i1 false)
54  %i.inc = add i8 %i.2, 1
55  %cmp.loop = icmp ugt i8 %i.inc, 100
56  br i1 %cmp.loop, label %exit, label %header
57exit:
58  %res = load i8, i8* %addr
59  ret i8 %res
60}
61
62; Check that first two loads are not optimized out while the one marked with
63; invariant.load reuses %res1
64define i8 @test2(i1 %cmp, i8 *%p) {
65; CHECK-LABEL: @test2(
66; CHECK-NEXT:  entry:
67; CHECK-NEXT:    [[RES1:%.*]] = load i8, i8* [[P:%.*]], align 1
68; CHECK-NEXT:    call void @foo(i8* [[P]])
69; CHECK-NEXT:    br i1 [[CMP:%.*]], label [[B2:%.*]], label [[B1:%.*]]
70; CHECK:       b1:
71; CHECK-NEXT:    [[RES2:%.*]] = load i8, i8* [[P]]
72; CHECK-NEXT:    [[RES3:%.*]] = add i8 [[RES1]], [[RES2]]
73; CHECK-NEXT:    br label [[ALIVE:%.*]]
74; CHECK:       b2:
75; CHECK-NEXT:    [[RES_DEAD:%.*]] = add i8 [[RES1]], [[RES1]]
76; CHECK-NEXT:    br label [[ALIVE]]
77; CHECK:       alive:
78; CHECK-NEXT:    [[RES_PHI:%.*]] = phi i8 [ [[RES3]], [[B1]] ], [ [[RES_DEAD]], [[B2]] ]
79; CHECK-NEXT:    ret i8 [[RES_PHI]]
80;
81
82entry:
83  %res1 = load i8, i8* %p
84  call void @foo(i8 *%p)
85  br i1 %cmp, label %b2, label %b1
86b1:
87  %res2 = load i8, i8* %p
88  %res3 = add i8 %res1, %res2
89  br label %alive
90b2:
91  %v = load i8, i8* %p, !invariant.load !1
92  %res.dead = add i8 %v, %res1
93  br label %alive
94alive:
95  %res.phi = phi i8 [%res3, %b1], [%res.dead, %b2]
96  ret i8 %res.phi
97}
98
99; This is essentially the same test case as the above one but with %b1 and %b2
100; swapped in "br i1 %cmp, label %b1, label %b2" instruction. That helps us to
101; ensure that results doesn't depend on visiting order.
102define i8 @test3(i1 %cmp, i8 *%p) {
103; CHECK-LABEL: @test3(
104; CHECK-NEXT:  entry:
105; CHECK-NEXT:    [[RES1:%.*]] = load i8, i8* [[P:%.*]], align 1
106; CHECK-NEXT:    call void @foo(i8* [[P]])
107; CHECK-NEXT:    br i1 [[CMP:%.*]], label [[B1:%.*]], label [[B2:%.*]]
108; CHECK:       b1:
109; CHECK-NEXT:    [[RES2:%.*]] = load i8, i8* [[P]]
110; CHECK-NEXT:    [[RES3:%.*]] = add i8 [[RES1]], [[RES2]]
111; CHECK-NEXT:    br label [[ALIVE:%.*]]
112; CHECK:       b2:
113; CHECK-NEXT:    [[RES_DEAD:%.*]] = add i8 [[RES1]], [[RES1]]
114; CHECK-NEXT:    br label [[ALIVE]]
115; CHECK:       alive:
116; CHECK-NEXT:    [[RES_PHI:%.*]] = phi i8 [ [[RES3]], [[B1]] ], [ [[RES_DEAD]], [[B2]] ]
117; CHECK-NEXT:    ret i8 [[RES_PHI]]
118;
119entry:
120  %res1 = load i8, i8* %p
121  call void @foo(i8 *%p)
122  br i1 %cmp, label %b1, label %b2
123b1:
124  %res2 = load i8, i8* %p
125  %res3 = add i8 %res1, %res2
126  br label %alive
127b2:
128  %v = load i8, i8* %p, !invariant.load !1
129  %res.dead = add i8 %v, %res1
130  br label %alive
131alive:
132  %res.phi = phi i8 [%res3, %b1], [%res.dead, %b2]
133  ret i8 %res.phi
134}
135
136
137; This is reduced test case catching regression in the first version of the
138; fix for invariant loads (https://reviews.llvm.org/D64405).
139define void @test4() {
140; CHECK-LABEL: @test4(
141; CHECK-NEXT:  entry:
142; CHECK-NEXT:    [[TMP0:%.*]] = load float, float* inttoptr (i64 8 to float*), align 4
143; CHECK-NEXT:    [[TMP1:%.*]] = fmul float [[TMP0]], [[TMP0]]
144; CHECK-NEXT:    br label [[FUSION_LOOP_HEADER_DIM_1_PREHEADER:%.*]]
145; CHECK:       fusion.loop_header.dim.1.preheader:
146; CHECK-NEXT:    [[TMP2:%.*]] = phi float [ [[TMP0]], [[ENTRY:%.*]] ], [ [[DOTPRE:%.*]], [[FUSION_LOOP_HEADER_DIM_1_PREHEADER]] ]
147; CHECK-NEXT:    [[FUSION_INVAR_ADDRESS_DIM_0_03:%.*]] = phi i64 [ 0, [[ENTRY]] ], [ [[INVAR_INC3:%.*]], [[FUSION_LOOP_HEADER_DIM_1_PREHEADER]] ]
148; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* null, i64 0, i64 [[FUSION_INVAR_ADDRESS_DIM_0_03]], i64 0, i64 2
149; CHECK-NEXT:    [[TMP4:%.*]] = fmul float [[TMP2]], [[TMP2]]
150; CHECK-NEXT:    [[INVAR_INC3]] = add nuw nsw i64 [[FUSION_INVAR_ADDRESS_DIM_0_03]], 1
151; CHECK-NEXT:    [[DOTPHI_TRANS_INSERT:%.*]] = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* null, i64 0, i64 [[INVAR_INC3]], i64 0, i64 2
152; CHECK-NEXT:    [[DOTPRE]] = load float, float* [[DOTPHI_TRANS_INSERT]], align 4, !invariant.load !0
153; CHECK-NEXT:    br label [[FUSION_LOOP_HEADER_DIM_1_PREHEADER]]
154;
155entry:
156  %0 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* null, i64 0, i64 0, i64 0, i64 2
157  %1 = load float, float* %0, align 4
158  %2 = fmul float %1, %1
159  br label %fusion.loop_header.dim.1.preheader
160
161fusion.loop_header.dim.1.preheader:               ; preds = %fusion.loop_header.dim.1.preheader, %entry
162  %fusion.invar_address.dim.0.03 = phi i64 [ 0, %entry ], [ %invar.inc3, %fusion.loop_header.dim.1.preheader ]
163  %3 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* null, i64 0, i64 %fusion.invar_address.dim.0.03, i64 0, i64 2
164  %4 = load float, float* %3, align 4, !invariant.load !1
165  %5 = fmul float %4, %4
166  %6 = getelementptr inbounds [2 x [1 x [4 x float]]], [2 x [1 x [4 x float]]]* null, i64 0, i64 %fusion.invar_address.dim.0.03, i64 0, i64 2
167  %7 = load float, float* %6, align 4, !invariant.load !1
168  %8 = fmul float %7, %7
169  %invar.inc3 = add nuw nsw i64 %fusion.invar_address.dim.0.03, 1
170  br label %fusion.loop_header.dim.1.preheader
171}
172
173!1 = !{}
174