xref: /llvm-project/llvm/test/Transforms/SLPVectorizer/X86/vectorize-reorder-reuse.ll (revision b3480d5ede83720abf58dcadf4e79015e5a7e60a) 
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -passes=slp-vectorizer -S -mtriple=x86_64-unknown-linux-gnu -mcpu=bdver2 < %s | FileCheck %s
3
4define i32 @foo(ptr nocapture readonly %arr, i32 %a1, i32 %a2, i32 %a3, i32 %a4, i32 %a5, i32 %a6, i32 %a7, i32 %a8) {
5; CHECK-LABEL: @foo(
6; CHECK-NEXT:  entry:
7; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[ARR:%.*]], i64 1
8; CHECK-NEXT:    [[TMP0:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
9; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[TMP0]], [[A1:%.*]]
10; CHECK-NEXT:    [[ADD2:%.*]] = add i32 [[TMP0]], [[A2:%.*]]
11; CHECK-NEXT:    [[ADD4:%.*]] = add i32 [[TMP0]], [[A3:%.*]]
12; CHECK-NEXT:    [[ADD6:%.*]] = add i32 [[TMP0]], [[A4:%.*]]
13; CHECK-NEXT:    [[ADD8:%.*]] = add i32 [[TMP0]], [[A5:%.*]]
14; CHECK-NEXT:    [[ADD10:%.*]] = add i32 [[TMP0]], [[A6:%.*]]
15; CHECK-NEXT:    [[TMP1:%.*]] = load i32, ptr [[ARR]], align 4
16; CHECK-NEXT:    [[ADD12:%.*]] = add i32 [[TMP1]], [[A7:%.*]]
17; CHECK-NEXT:    [[ADD14:%.*]] = add i32 [[TMP1]], [[A8:%.*]]
18; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i32 [[ADD]], [[ADD2]]
19; CHECK-NEXT:    [[COND:%.*]] = select i1 [[CMP]], i32 [[ADD]], i32 [[ADD2]]
20; CHECK-NEXT:    [[CMP15:%.*]] = icmp ult i32 [[COND]], [[ADD4]]
21; CHECK-NEXT:    [[COND19:%.*]] = select i1 [[CMP15]], i32 [[COND]], i32 [[ADD4]]
22; CHECK-NEXT:    [[CMP20:%.*]] = icmp ult i32 [[COND19]], [[ADD6]]
23; CHECK-NEXT:    [[COND24:%.*]] = select i1 [[CMP20]], i32 [[COND19]], i32 [[ADD6]]
24; CHECK-NEXT:    [[CMP25:%.*]] = icmp ult i32 [[COND24]], [[ADD8]]
25; CHECK-NEXT:    [[COND29:%.*]] = select i1 [[CMP25]], i32 [[COND24]], i32 [[ADD8]]
26; CHECK-NEXT:    [[CMP30:%.*]] = icmp ult i32 [[COND29]], [[ADD10]]
27; CHECK-NEXT:    [[COND34:%.*]] = select i1 [[CMP30]], i32 [[COND29]], i32 [[ADD10]]
28; CHECK-NEXT:    [[CMP35:%.*]] = icmp ult i32 [[COND34]], [[ADD12]]
29; CHECK-NEXT:    [[COND39:%.*]] = select i1 [[CMP35]], i32 [[COND34]], i32 [[ADD12]]
30; CHECK-NEXT:    [[CMP40:%.*]] = icmp ult i32 [[COND39]], [[ADD14]]
31; CHECK-NEXT:    [[COND44:%.*]] = select i1 [[CMP40]], i32 [[COND39]], i32 [[ADD14]]
32; CHECK-NEXT:    ret i32 [[COND44]]
33;
34entry:
35  %arrayidx = getelementptr inbounds i32, ptr %arr, i64 1
36  %0 = load i32, ptr %arrayidx, align 4
37  %add = add i32 %0, %a1
38  %add2 = add i32 %0, %a2
39  %add4 = add i32 %0, %a3
40  %add6 = add i32 %0, %a4
41  %add8 = add i32 %0, %a5
42  %add10 = add i32 %0, %a6
43  %1 = load i32, ptr %arr, align 4
44  %add12 = add i32 %1, %a7
45  %add14 = add i32 %1, %a8
46  %cmp = icmp ult i32 %add, %add2
47  %cond = select i1 %cmp, i32 %add, i32 %add2
48  %cmp15 = icmp ult i32 %cond, %add4
49  %cond19 = select i1 %cmp15, i32 %cond, i32 %add4
50  %cmp20 = icmp ult i32 %cond19, %add6
51  %cond24 = select i1 %cmp20, i32 %cond19, i32 %add6
52  %cmp25 = icmp ult i32 %cond24, %add8
53  %cond29 = select i1 %cmp25, i32 %cond24, i32 %add8
54  %cmp30 = icmp ult i32 %cond29, %add10
55  %cond34 = select i1 %cmp30, i32 %cond29, i32 %add10
56  %cmp35 = icmp ult i32 %cond34, %add12
57  %cond39 = select i1 %cmp35, i32 %cond34, i32 %add12
58  %cmp40 = icmp ult i32 %cond39, %add14
59  %cond44 = select i1 %cmp40, i32 %cond39, i32 %add14
60  ret i32 %cond44
61}
62
63define i32 @foo1(ptr nocapture readonly %arr, i32 %a1, i32 %a2, i32 %a3, i32 %a4, i32 %a5, i32 %a6, i32 %a7, i32 %a8) {
64; CHECK-LABEL: @foo1(
65; CHECK-NEXT:  entry:
66; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[ARR:%.*]], i64 1
67; CHECK-NEXT:    [[TMP0:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
68; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[TMP0]], [[A1:%.*]]
69; CHECK-NEXT:    [[ARRAYIDX1:%.*]] = getelementptr inbounds i32, ptr [[ARR]], i64 2
70; CHECK-NEXT:    [[TMP1:%.*]] = load i32, ptr [[ARRAYIDX1]], align 4
71; CHECK-NEXT:    [[ADD2:%.*]] = add i32 [[TMP1]], [[A2:%.*]]
72; CHECK-NEXT:    [[ARRAYIDX3:%.*]] = getelementptr inbounds i32, ptr [[ARR]], i64 3
73; CHECK-NEXT:    [[TMP2:%.*]] = load i32, ptr [[ARRAYIDX3]], align 4
74; CHECK-NEXT:    [[ADD4:%.*]] = add i32 [[TMP2]], [[A3:%.*]]
75; CHECK-NEXT:    [[ADD6:%.*]] = add i32 [[TMP0]], [[A4:%.*]]
76; CHECK-NEXT:    [[ADD8:%.*]] = add i32 [[TMP0]], [[A5:%.*]]
77; CHECK-NEXT:    [[TMP3:%.*]] = load i32, ptr [[ARR]], align 4
78; CHECK-NEXT:    [[ADD10:%.*]] = add i32 [[TMP3]], [[A6:%.*]]
79; CHECK-NEXT:    [[ADD12:%.*]] = add i32 [[TMP1]], [[A7:%.*]]
80; CHECK-NEXT:    [[ADD14:%.*]] = add i32 [[TMP0]], [[A8:%.*]]
81; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i32 [[ADD]], [[ADD2]]
82; CHECK-NEXT:    [[COND:%.*]] = select i1 [[CMP]], i32 [[ADD]], i32 [[ADD2]]
83; CHECK-NEXT:    [[CMP15:%.*]] = icmp ult i32 [[COND]], [[ADD4]]
84; CHECK-NEXT:    [[COND19:%.*]] = select i1 [[CMP15]], i32 [[COND]], i32 [[ADD4]]
85; CHECK-NEXT:    [[CMP20:%.*]] = icmp ult i32 [[COND19]], [[ADD6]]
86; CHECK-NEXT:    [[COND24:%.*]] = select i1 [[CMP20]], i32 [[COND19]], i32 [[ADD6]]
87; CHECK-NEXT:    [[CMP25:%.*]] = icmp ult i32 [[COND24]], [[ADD8]]
88; CHECK-NEXT:    [[COND29:%.*]] = select i1 [[CMP25]], i32 [[COND24]], i32 [[ADD8]]
89; CHECK-NEXT:    [[CMP30:%.*]] = icmp ult i32 [[COND29]], [[ADD10]]
90; CHECK-NEXT:    [[COND34:%.*]] = select i1 [[CMP30]], i32 [[COND29]], i32 [[ADD10]]
91; CHECK-NEXT:    [[CMP35:%.*]] = icmp ult i32 [[COND34]], [[ADD12]]
92; CHECK-NEXT:    [[COND39:%.*]] = select i1 [[CMP35]], i32 [[COND34]], i32 [[ADD12]]
93; CHECK-NEXT:    [[CMP40:%.*]] = icmp ult i32 [[COND39]], [[ADD14]]
94; CHECK-NEXT:    [[COND44:%.*]] = select i1 [[CMP40]], i32 [[COND39]], i32 [[ADD14]]
95; CHECK-NEXT:    ret i32 [[COND44]]
96;
97entry:
98  %arrayidx = getelementptr inbounds i32, ptr %arr, i64 1
99  %0 = load i32, ptr %arrayidx, align 4
100  %add = add i32 %0, %a1
101  %arrayidx1 = getelementptr inbounds i32, ptr %arr, i64 2
102  %1 = load i32, ptr %arrayidx1, align 4
103  %add2 = add i32 %1, %a2
104  %arrayidx3 = getelementptr inbounds i32, ptr %arr, i64 3
105  %2 = load i32, ptr %arrayidx3, align 4
106  %add4 = add i32 %2, %a3
107  %add6 = add i32 %0, %a4
108  %add8 = add i32 %0, %a5
109  %3 = load i32, ptr %arr, align 4
110  %add10 = add i32 %3, %a6
111  %add12 = add i32 %1, %a7
112  %add14 = add i32 %0, %a8
113  %cmp = icmp ult i32 %add, %add2
114  %cond = select i1 %cmp, i32 %add, i32 %add2
115  %cmp15 = icmp ult i32 %cond, %add4
116  %cond19 = select i1 %cmp15, i32 %cond, i32 %add4
117  %cmp20 = icmp ult i32 %cond19, %add6
118  %cond24 = select i1 %cmp20, i32 %cond19, i32 %add6
119  %cmp25 = icmp ult i32 %cond24, %add8
120  %cond29 = select i1 %cmp25, i32 %cond24, i32 %add8
121  %cmp30 = icmp ult i32 %cond29, %add10
122  %cond34 = select i1 %cmp30, i32 %cond29, i32 %add10
123  %cmp35 = icmp ult i32 %cond34, %add12
124  %cond39 = select i1 %cmp35, i32 %cond34, i32 %add12
125  %cmp40 = icmp ult i32 %cond39, %add14
126  %cond44 = select i1 %cmp40, i32 %cond39, i32 %add14
127  ret i32 %cond44
128}
129
130define i32 @foo2(ptr nocapture readonly %arr, i32 %a1, i32 %a2, i32 %a3, i32 %a4, i32 %a5, i32 %a6, i32 %a7, i32 %a8) {
131; CHECK-LABEL: @foo2(
132; CHECK-NEXT:  entry:
133; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds i32, ptr [[ARR:%.*]], i64 3
134; CHECK-NEXT:    [[TMP0:%.*]] = load i32, ptr [[ARRAYIDX]], align 4
135; CHECK-NEXT:    [[ADD:%.*]] = add i32 [[TMP0]], [[A1:%.*]]
136; CHECK-NEXT:    [[ARRAYIDX1:%.*]] = getelementptr inbounds i32, ptr [[ARR]], i64 2
137; CHECK-NEXT:    [[TMP1:%.*]] = load i32, ptr [[ARRAYIDX1]], align 4
138; CHECK-NEXT:    [[ADD2:%.*]] = add i32 [[TMP1]], [[A2:%.*]]
139; CHECK-NEXT:    [[ADD4:%.*]] = add i32 [[TMP0]], [[A3:%.*]]
140; CHECK-NEXT:    [[TMP2:%.*]] = load i32, ptr [[ARR]], align 4
141; CHECK-NEXT:    [[ADD6:%.*]] = add i32 [[TMP2]], [[A4:%.*]]
142; CHECK-NEXT:    [[ARRAYIDX7:%.*]] = getelementptr inbounds i32, ptr [[ARR]], i64 1
143; CHECK-NEXT:    [[TMP3:%.*]] = load i32, ptr [[ARRAYIDX7]], align 4
144; CHECK-NEXT:    [[ADD8:%.*]] = add i32 [[TMP3]], [[A5:%.*]]
145; CHECK-NEXT:    [[ADD10:%.*]] = add i32 [[TMP2]], [[A6:%.*]]
146; CHECK-NEXT:    [[ADD12:%.*]] = add i32 [[TMP1]], [[A7:%.*]]
147; CHECK-NEXT:    [[ADD14:%.*]] = add i32 [[TMP3]], [[A8:%.*]]
148; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i32 [[ADD]], [[ADD2]]
149; CHECK-NEXT:    [[COND:%.*]] = select i1 [[CMP]], i32 [[ADD]], i32 [[ADD2]]
150; CHECK-NEXT:    [[CMP15:%.*]] = icmp ult i32 [[COND]], [[ADD4]]
151; CHECK-NEXT:    [[COND19:%.*]] = select i1 [[CMP15]], i32 [[COND]], i32 [[ADD4]]
152; CHECK-NEXT:    [[CMP20:%.*]] = icmp ult i32 [[COND19]], [[ADD6]]
153; CHECK-NEXT:    [[COND24:%.*]] = select i1 [[CMP20]], i32 [[COND19]], i32 [[ADD6]]
154; CHECK-NEXT:    [[CMP25:%.*]] = icmp ult i32 [[COND24]], [[ADD8]]
155; CHECK-NEXT:    [[COND29:%.*]] = select i1 [[CMP25]], i32 [[COND24]], i32 [[ADD8]]
156; CHECK-NEXT:    [[CMP30:%.*]] = icmp ult i32 [[COND29]], [[ADD10]]
157; CHECK-NEXT:    [[COND34:%.*]] = select i1 [[CMP30]], i32 [[COND29]], i32 [[ADD10]]
158; CHECK-NEXT:    [[CMP35:%.*]] = icmp ult i32 [[COND34]], [[ADD12]]
159; CHECK-NEXT:    [[COND39:%.*]] = select i1 [[CMP35]], i32 [[COND34]], i32 [[ADD12]]
160; CHECK-NEXT:    [[CMP40:%.*]] = icmp ult i32 [[COND39]], [[ADD14]]
161; CHECK-NEXT:    [[COND44:%.*]] = select i1 [[CMP40]], i32 [[COND39]], i32 [[ADD14]]
162; CHECK-NEXT:    ret i32 [[COND44]]
163;
164entry:
165  %arrayidx = getelementptr inbounds i32, ptr %arr, i64 3
166  %0 = load i32, ptr %arrayidx, align 4
167  %add = add i32 %0, %a1
168  %arrayidx1 = getelementptr inbounds i32, ptr %arr, i64 2
169  %1 = load i32, ptr %arrayidx1, align 4
170  %add2 = add i32 %1, %a2
171  %add4 = add i32 %0, %a3
172  %2 = load i32, ptr %arr, align 4
173  %add6 = add i32 %2, %a4
174  %arrayidx7 = getelementptr inbounds i32, ptr %arr, i64 1
175  %3 = load i32, ptr %arrayidx7, align 4
176  %add8 = add i32 %3, %a5
177  %add10 = add i32 %2, %a6
178  %add12 = add i32 %1, %a7
179  %add14 = add i32 %3, %a8
180  %cmp = icmp ult i32 %add, %add2
181  %cond = select i1 %cmp, i32 %add, i32 %add2
182  %cmp15 = icmp ult i32 %cond, %add4
183  %cond19 = select i1 %cmp15, i32 %cond, i32 %add4
184  %cmp20 = icmp ult i32 %cond19, %add6
185  %cond24 = select i1 %cmp20, i32 %cond19, i32 %add6
186  %cmp25 = icmp ult i32 %cond24, %add8
187  %cond29 = select i1 %cmp25, i32 %cond24, i32 %add8
188  %cmp30 = icmp ult i32 %cond29, %add10
189  %cond34 = select i1 %cmp30, i32 %cond29, i32 %add10
190  %cmp35 = icmp ult i32 %cond34, %add12
191  %cond39 = select i1 %cmp35, i32 %cond34, i32 %add12
192  %cmp40 = icmp ult i32 %cond39, %add14
193  %cond44 = select i1 %cmp40, i32 %cond39, i32 %add14
194  ret i32 %cond44
195}
196