1; RUN: opt -S -march=r600 -mcpu=cayman -basic-aa -loop-vectorize -force-vector-interleave=1 -force-vector-width=4 -dce -instcombine < %s | FileCheck %s
2
3; Check vectorization that would ordinarily require a runtime bounds
4; check on the pointers when mixing address spaces. For now we cannot
5; assume address spaces do not alias, and we can't assume that
6; different pointers are directly comparable.
7;
8; These all test this basic loop for different combinations of address
9; spaces, and swapping in globals or adding noalias.
10;
11;void foo(int addrspace(N)* [noalias] a, int addrspace(M)* [noalias] b, int n)
12;{
13;    for (int i = 0; i < n; ++i)
14;    {
15;        a[i] = 3 * b[i];
16;    }
17;}
18
19; Artificial datalayout
20target datalayout = "e-p:32:32:32-p1:16:16:16-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-v16:16:16-v24:32:32-v32:32:32-v48:64:64-v64:64:64-v96:128:128-v128:128:128-v192:256:256-v256:256:256-v512:512:512-v1024:1024:1024-v2048:2048:2048-n32:64"
21
22
23@g_as1 = common addrspace(1) global [1024 x i32] zeroinitializer, align 16
24@q_as2 = common addrspace(2) global [1024 x i32] zeroinitializer, align 16
25
26; Both parameters are unidentified objects with the same address
27; space, so this should vectorize normally.
28define void @foo(i32 addrspace(1)* %a, i32 addrspace(1)* %b, i32 %n) #0 {
29; CHECK-LABEL: @foo(
30; CHECK: <4 x i32>
31; CHECK: ret
32
33entry:
34  %cmp1 = icmp slt i32 0, %n
35  br i1 %cmp1, label %for.body, label %for.end
36
37for.body:                                         ; preds = %entry, %for.body
38  %i.02 = phi i32 [ %inc, %for.body ], [ 0, %entry ]
39  %idxprom = sext i32 %i.02 to i64
40  %arrayidx = getelementptr inbounds i32, i32 addrspace(1)* %b, i64 %idxprom
41  %0 = load i32, i32 addrspace(1)* %arrayidx, align 4
42  %mul = mul nsw i32 %0, 3
43  %idxprom1 = sext i32 %i.02 to i64
44  %arrayidx2 = getelementptr inbounds i32, i32 addrspace(1)* %a, i64 %idxprom1
45  store i32 %mul, i32 addrspace(1)* %arrayidx2, align 4
46  %inc = add nsw i32 %i.02, 1
47  %cmp = icmp slt i32 %inc, %n
48  br i1 %cmp, label %for.body, label %for.end
49
50for.end:                                          ; preds = %for.body, %entry
51  ret void
52}
53
54; Parameters are unidentified and different address spaces, so cannot vectorize.
55define void @bar0(i32* %a, i32 addrspace(1)* %b, i32 %n) #0 {
56; CHECK-LABEL: @bar0(
57; CHECK-NOT: <4 x i32>
58; CHECK: ret
59
60entry:
61  %cmp1 = icmp slt i32 0, %n
62  br i1 %cmp1, label %for.body, label %for.end
63
64for.body:                                         ; preds = %entry, %for.body
65  %i.02 = phi i32 [ %inc, %for.body ], [ 0, %entry ]
66  %idxprom = sext i32 %i.02 to i64
67  %arrayidx = getelementptr inbounds i32, i32 addrspace(1)* %b, i64 %idxprom
68  %0 = load i32, i32 addrspace(1)* %arrayidx, align 4
69  %mul = mul nsw i32 %0, 3
70  %idxprom1 = sext i32 %i.02 to i64
71  %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 %idxprom1
72  store i32 %mul, i32* %arrayidx2, align 4
73  %inc = add nsw i32 %i.02, 1
74  %cmp = icmp slt i32 %inc, %n
75  br i1 %cmp, label %for.body, label %for.end
76
77for.end:                                          ; preds = %for.body, %entry
78  ret void
79}
80
81; Swapped arguments should be the same
82define void @bar1(i32 addrspace(1)* %a, i32* %b, i32 %n) #0 {
83; CHECK-LABEL: @bar1(
84; CHECK-NOT: <4 x i32>
85; CHECK: ret
86
87entry:
88  %cmp1 = icmp slt i32 0, %n
89  br i1 %cmp1, label %for.body, label %for.end
90
91for.body:                                         ; preds = %entry, %for.body
92  %i.02 = phi i32 [ %inc, %for.body ], [ 0, %entry ]
93  %idxprom = sext i32 %i.02 to i64
94  %arrayidx = getelementptr inbounds i32, i32* %b, i64 %idxprom
95  %0 = load i32, i32* %arrayidx, align 4
96  %mul = mul nsw i32 %0, 3
97  %idxprom1 = sext i32 %i.02 to i64
98  %arrayidx2 = getelementptr inbounds i32, i32 addrspace(1)* %a, i64 %idxprom1
99  store i32 %mul, i32 addrspace(1)* %arrayidx2, align 4
100  %inc = add nsw i32 %i.02, 1
101  %cmp = icmp slt i32 %inc, %n
102  br i1 %cmp, label %for.body, label %for.end
103
104for.end:                                          ; preds = %for.body, %entry
105  ret void
106}
107
108; We should still be able to vectorize with noalias even if the
109; address spaces are different.
110define void @bar2(i32* noalias %a, i32 addrspace(1)* noalias %b, i32 %n) #0 {
111; CHECK-LABEL: @bar2(
112; CHECK: <4 x i32>
113; CHECK: ret
114
115entry:
116  %cmp1 = icmp slt i32 0, %n
117  br i1 %cmp1, label %for.body, label %for.end
118
119for.body:                                         ; preds = %entry, %for.body
120  %i.02 = phi i32 [ %inc, %for.body ], [ 0, %entry ]
121  %idxprom = sext i32 %i.02 to i64
122  %arrayidx = getelementptr inbounds i32, i32 addrspace(1)* %b, i64 %idxprom
123  %0 = load i32, i32 addrspace(1)* %arrayidx, align 4
124  %mul = mul nsw i32 %0, 3
125  %idxprom1 = sext i32 %i.02 to i64
126  %arrayidx2 = getelementptr inbounds i32, i32* %a, i64 %idxprom1
127  store i32 %mul, i32* %arrayidx2, align 4
128  %inc = add nsw i32 %i.02, 1
129  %cmp = icmp slt i32 %inc, %n
130  br i1 %cmp, label %for.body, label %for.end
131
132for.end:                                          ; preds = %for.body, %entry
133  ret void
134}
135
136; Store to identified global with different address space. This isn't
137; generally safe and shouldn't be vectorized.
138define void @arst0(i32* %b, i32 %n) #0 {
139; CHECK-LABEL: @arst0(
140; CHECK-NOT: <4 x i32>
141; CHECK: ret
142
143entry:
144  %cmp1 = icmp slt i32 0, %n
145  br i1 %cmp1, label %for.body, label %for.end
146
147for.body:                                         ; preds = %entry, %for.body
148  %i.02 = phi i32 [ %inc, %for.body ], [ 0, %entry ]
149  %idxprom = sext i32 %i.02 to i64
150  %arrayidx = getelementptr inbounds i32, i32* %b, i64 %idxprom
151  %0 = load i32, i32* %arrayidx, align 4
152  %mul = mul nsw i32 %0, 3
153  %idxprom1 = sext i32 %i.02 to i64
154  %arrayidx2 = getelementptr inbounds [1024 x i32], [1024 x i32] addrspace(1)* @g_as1, i64 0, i64 %idxprom1
155  store i32 %mul, i32 addrspace(1)* %arrayidx2, align 4
156  %inc = add nsw i32 %i.02, 1
157  %cmp = icmp slt i32 %inc, %n
158  br i1 %cmp, label %for.body, label %for.end
159
160for.end:                                          ; preds = %for.body, %entry
161  ret void
162}
163
164
165; Load from identified global with different address space.
166; This isn't generally safe and shouldn't be vectorized.
167define void @arst1(i32* %b, i32 %n) #0 {
168; CHECK-LABEL: @arst1(
169; CHECK-NOT: <4 x i32>
170; CHECK: ret
171
172entry:
173  %cmp1 = icmp slt i32 0, %n
174  br i1 %cmp1, label %for.body, label %for.end
175
176for.body:                                         ; preds = %entry, %for.body
177  %i.02 = phi i32 [ %inc, %for.body ], [ 0, %entry ]
178  %idxprom = sext i32 %i.02 to i64
179  %arrayidx = getelementptr inbounds [1024 x i32], [1024 x i32] addrspace(1)* @g_as1, i64 0, i64 %idxprom
180  %0 = load i32, i32 addrspace(1)* %arrayidx, align 4
181  %mul = mul nsw i32 %0, 3
182  %idxprom1 = sext i32 %i.02 to i64
183  %arrayidx2 = getelementptr inbounds i32, i32* %b, i64 %idxprom1
184  store i32 %mul, i32* %arrayidx2, align 4
185  %inc = add nsw i32 %i.02, 1
186  %cmp = icmp slt i32 %inc, %n
187  br i1 %cmp, label %for.body, label %for.end
188
189for.end:                                          ; preds = %for.body, %entry
190  ret void
191}
192
193; Read and write to 2 identified globals in different address
194; spaces. This should be vectorized.
195define void @aoeu(i32 %n) #0 {
196; CHECK-LABEL: @aoeu(
197; CHECK: <4 x i32>
198; CHECK: ret
199
200entry:
201  %cmp1 = icmp slt i32 0, %n
202  br i1 %cmp1, label %for.body, label %for.end
203
204for.body:                                         ; preds = %entry, %for.body
205  %i.02 = phi i32 [ %inc, %for.body ], [ 0, %entry ]
206  %idxprom = sext i32 %i.02 to i64
207  %arrayidx = getelementptr inbounds [1024 x i32], [1024 x i32] addrspace(2)* @q_as2, i64 0, i64 %idxprom
208  %0 = load i32, i32 addrspace(2)* %arrayidx, align 4
209  %mul = mul nsw i32 %0, 3
210  %idxprom1 = sext i32 %i.02 to i64
211  %arrayidx2 = getelementptr inbounds [1024 x i32], [1024 x i32] addrspace(1)* @g_as1, i64 0, i64 %idxprom1
212  store i32 %mul, i32 addrspace(1)* %arrayidx2, align 4
213  %inc = add nsw i32 %i.02, 1
214  %cmp = icmp slt i32 %inc, %n
215  br i1 %cmp, label %for.body, label %for.end
216
217for.end:                                          ; preds = %for.body, %entry
218  ret void
219}
220
221attributes #0 = { nounwind uwtable "less-precise-fpmad"="false" "frame-pointer"="all" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
222