1; RUN: opt < %s -basicaa -slp-vectorizer -S | FileCheck %s
2target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
3target triple = "x86_64-apple-macosx10.9.0"
4
5@A = common global [2000 x double] zeroinitializer, align 16
6@B = common global [2000 x double] zeroinitializer, align 16
7@C = common global [2000 x float] zeroinitializer, align 16
8@D = common global [2000 x float] zeroinitializer, align 16
9
10; Currently SCEV isn't smart enough to figure out that accesses
11; A[3*i], A[3*i+1] and A[3*i+2] are consecutive, but in future
12; that would hopefully be fixed. For now, check that this isn't
13; vectorized.
14; CHECK-LABEL: foo_3double
15; CHECK-NOT: x double>
16; Function Attrs: nounwind ssp uwtable
17define void @foo_3double(i32 %u) #0 {
18entry:
19  %u.addr = alloca i32, align 4
20  store i32 %u, i32* %u.addr, align 4
21  %mul = mul nsw i32 %u, 3
22  %idxprom = sext i32 %mul to i64
23  %arrayidx = getelementptr inbounds [2000 x double]* @A, i32 0, i64 %idxprom
24  %0 = load double* %arrayidx, align 8
25  %arrayidx4 = getelementptr inbounds [2000 x double]* @B, i32 0, i64 %idxprom
26  %1 = load double* %arrayidx4, align 8
27  %add5 = fadd double %0, %1
28  store double %add5, double* %arrayidx, align 8
29  %add11 = add nsw i32 %mul, 1
30  %idxprom12 = sext i32 %add11 to i64
31  %arrayidx13 = getelementptr inbounds [2000 x double]* @A, i32 0, i64 %idxprom12
32  %2 = load double* %arrayidx13, align 8
33  %arrayidx17 = getelementptr inbounds [2000 x double]* @B, i32 0, i64 %idxprom12
34  %3 = load double* %arrayidx17, align 8
35  %add18 = fadd double %2, %3
36  store double %add18, double* %arrayidx13, align 8
37  %add24 = add nsw i32 %mul, 2
38  %idxprom25 = sext i32 %add24 to i64
39  %arrayidx26 = getelementptr inbounds [2000 x double]* @A, i32 0, i64 %idxprom25
40  %4 = load double* %arrayidx26, align 8
41  %arrayidx30 = getelementptr inbounds [2000 x double]* @B, i32 0, i64 %idxprom25
42  %5 = load double* %arrayidx30, align 8
43  %add31 = fadd double %4, %5
44  store double %add31, double* %arrayidx26, align 8
45  ret void
46}
47
48; SCEV should be able to tell that accesses A[C1 + C2*i], A[C1 + C2*i], ...
49; A[C1 + C2*i] are consecutive, if C2 is a power of 2, and C2 > C1 > 0.
50; Thus, the following code should be vectorized.
51; CHECK-LABEL: foo_2double
52; CHECK: x double>
53; Function Attrs: nounwind ssp uwtable
54define void @foo_2double(i32 %u) #0 {
55entry:
56  %u.addr = alloca i32, align 4
57  store i32 %u, i32* %u.addr, align 4
58  %mul = mul nsw i32 %u, 2
59  %idxprom = sext i32 %mul to i64
60  %arrayidx = getelementptr inbounds [2000 x double]* @A, i32 0, i64 %idxprom
61  %0 = load double* %arrayidx, align 8
62  %arrayidx4 = getelementptr inbounds [2000 x double]* @B, i32 0, i64 %idxprom
63  %1 = load double* %arrayidx4, align 8
64  %add5 = fadd double %0, %1
65  store double %add5, double* %arrayidx, align 8
66  %add11 = add nsw i32 %mul, 1
67  %idxprom12 = sext i32 %add11 to i64
68  %arrayidx13 = getelementptr inbounds [2000 x double]* @A, i32 0, i64 %idxprom12
69  %2 = load double* %arrayidx13, align 8
70  %arrayidx17 = getelementptr inbounds [2000 x double]* @B, i32 0, i64 %idxprom12
71  %3 = load double* %arrayidx17, align 8
72  %add18 = fadd double %2, %3
73  store double %add18, double* %arrayidx13, align 8
74  ret void
75}
76
77; Similar to the previous test, but with different datatype.
78; CHECK-LABEL: foo_4float
79; CHECK: x float>
80; Function Attrs: nounwind ssp uwtable
81define void @foo_4float(i32 %u) #0 {
82entry:
83  %u.addr = alloca i32, align 4
84  store i32 %u, i32* %u.addr, align 4
85  %mul = mul nsw i32 %u, 4
86  %idxprom = sext i32 %mul to i64
87  %arrayidx = getelementptr inbounds [2000 x float]* @C, i32 0, i64 %idxprom
88  %0 = load float* %arrayidx, align 4
89  %arrayidx4 = getelementptr inbounds [2000 x float]* @D, i32 0, i64 %idxprom
90  %1 = load float* %arrayidx4, align 4
91  %add5 = fadd float %0, %1
92  store float %add5, float* %arrayidx, align 4
93  %add11 = add nsw i32 %mul, 1
94  %idxprom12 = sext i32 %add11 to i64
95  %arrayidx13 = getelementptr inbounds [2000 x float]* @C, i32 0, i64 %idxprom12
96  %2 = load float* %arrayidx13, align 4
97  %arrayidx17 = getelementptr inbounds [2000 x float]* @D, i32 0, i64 %idxprom12
98  %3 = load float* %arrayidx17, align 4
99  %add18 = fadd float %2, %3
100  store float %add18, float* %arrayidx13, align 4
101  %add24 = add nsw i32 %mul, 2
102  %idxprom25 = sext i32 %add24 to i64
103  %arrayidx26 = getelementptr inbounds [2000 x float]* @C, i32 0, i64 %idxprom25
104  %4 = load float* %arrayidx26, align 4
105  %arrayidx30 = getelementptr inbounds [2000 x float]* @D, i32 0, i64 %idxprom25
106  %5 = load float* %arrayidx30, align 4
107  %add31 = fadd float %4, %5
108  store float %add31, float* %arrayidx26, align 4
109  %add37 = add nsw i32 %mul, 3
110  %idxprom38 = sext i32 %add37 to i64
111  %arrayidx39 = getelementptr inbounds [2000 x float]* @C, i32 0, i64 %idxprom38
112  %6 = load float* %arrayidx39, align 4
113  %arrayidx43 = getelementptr inbounds [2000 x float]* @D, i32 0, i64 %idxprom38
114  %7 = load float* %arrayidx43, align 4
115  %add44 = fadd float %6, %7
116  store float %add44, float* %arrayidx39, align 4
117  ret void
118}
119
120; Similar to the previous tests, but now we are dealing with AddRec SCEV.
121; CHECK-LABEL: foo_loop
122; CHECK: x double>
123; Function Attrs: nounwind ssp uwtable
124define i32 @foo_loop(double* %A, i32 %n) #0 {
125entry:
126  %A.addr = alloca double*, align 8
127  %n.addr = alloca i32, align 4
128  %sum = alloca double, align 8
129  %i = alloca i32, align 4
130  store double* %A, double** %A.addr, align 8
131  store i32 %n, i32* %n.addr, align 4
132  store double 0.000000e+00, double* %sum, align 8
133  store i32 0, i32* %i, align 4
134  %cmp1 = icmp slt i32 0, %n
135  br i1 %cmp1, label %for.body.lr.ph, label %for.end
136
137for.body.lr.ph:                                   ; preds = %entry
138  br label %for.body
139
140for.body:                                         ; preds = %for.body.lr.ph, %for.body
141  %0 = phi i32 [ 0, %for.body.lr.ph ], [ %inc, %for.body ]
142  %1 = phi double [ 0.000000e+00, %for.body.lr.ph ], [ %add7, %for.body ]
143  %mul = mul nsw i32 %0, 2
144  %idxprom = sext i32 %mul to i64
145  %arrayidx = getelementptr inbounds double* %A, i64 %idxprom
146  %2 = load double* %arrayidx, align 8
147  %mul1 = fmul double 7.000000e+00, %2
148  %add = add nsw i32 %mul, 1
149  %idxprom3 = sext i32 %add to i64
150  %arrayidx4 = getelementptr inbounds double* %A, i64 %idxprom3
151  %3 = load double* %arrayidx4, align 8
152  %mul5 = fmul double 7.000000e+00, %3
153  %add6 = fadd double %mul1, %mul5
154  %add7 = fadd double %1, %add6
155  store double %add7, double* %sum, align 8
156  %inc = add nsw i32 %0, 1
157  store i32 %inc, i32* %i, align 4
158  %cmp = icmp slt i32 %inc, %n
159  br i1 %cmp, label %for.body, label %for.cond.for.end_crit_edge
160
161for.cond.for.end_crit_edge:                       ; preds = %for.body
162  %split = phi double [ %add7, %for.body ]
163  br label %for.end
164
165for.end:                                          ; preds = %for.cond.for.end_crit_edge, %entry
166  %.lcssa = phi double [ %split, %for.cond.for.end_crit_edge ], [ 0.000000e+00, %entry ]
167  %conv = fptosi double %.lcssa to i32
168  ret i32 %conv
169}
170
171attributes #0 = { nounwind ssp uwtable "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
172
173!llvm.ident = !{!0}
174
175!0 = !{!"clang version 3.5.0 "}
176