1; RUN: opt -S -march=r600 -mcpu=cayman -loop-vectorize -force-vector-unroll=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  br label %for.cond
35
36for.cond:                                         ; preds = %for.body, %entry
37  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
38  %cmp = icmp slt i32 %i.0, %n
39  br i1 %cmp, label %for.body, label %for.end
40
41for.body:                                         ; preds = %for.cond
42  %idxprom = sext i32 %i.0 to i64
43  %arrayidx = getelementptr inbounds i32 addrspace(1)* %b, i64 %idxprom
44  %0 = load i32 addrspace(1)* %arrayidx, align 4
45  %mul = mul nsw i32 %0, 3
46  %idxprom1 = sext i32 %i.0 to i64
47  %arrayidx2 = getelementptr inbounds i32 addrspace(1)* %a, i64 %idxprom1
48  store i32 %mul, i32 addrspace(1)* %arrayidx2, align 4
49  %inc = add nsw i32 %i.0, 1
50  br label %for.cond
51
52for.end:                                          ; preds = %for.cond
53  ret void
54}
55
56; Parameters are unidentified and different address spaces, so cannot vectorize.
57define void @bar0(i32* %a, i32 addrspace(1)* %b, i32 %n) #0 {
58; CHECK-LABEL: @bar0(
59; CHECK-NOT: <4 x i32>
60; CHECK: ret
61
62entry:
63  br label %for.cond
64
65for.cond:                                         ; preds = %for.body, %entry
66  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
67  %cmp = icmp slt i32 %i.0, %n
68  br i1 %cmp, label %for.body, label %for.end
69
70for.body:                                         ; preds = %for.cond
71  %idxprom = sext i32 %i.0 to i64
72  %arrayidx = getelementptr inbounds i32 addrspace(1)* %b, i64 %idxprom
73  %0 = load i32 addrspace(1)* %arrayidx, align 4
74  %mul = mul nsw i32 %0, 3
75  %idxprom1 = sext i32 %i.0 to i64
76  %arrayidx2 = getelementptr inbounds i32* %a, i64 %idxprom1
77  store i32 %mul, i32* %arrayidx2, align 4
78  %inc = add nsw i32 %i.0, 1
79  br label %for.cond
80
81for.end:                                          ; preds = %for.cond
82  ret void
83}
84
85; Swapped arguments should be the same
86define void @bar1(i32 addrspace(1)* %a, i32* %b, i32 %n) #0 {
87; CHECK-LABEL: @bar1(
88; CHECK-NOT: <4 x i32>
89; CHECK: ret
90
91entry:
92  br label %for.cond
93
94for.cond:                                         ; preds = %for.body, %entry
95  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
96  %cmp = icmp slt i32 %i.0, %n
97  br i1 %cmp, label %for.body, label %for.end
98
99for.body:                                         ; preds = %for.cond
100  %idxprom = sext i32 %i.0 to i64
101  %arrayidx = getelementptr inbounds i32* %b, i64 %idxprom
102  %0 = load i32* %arrayidx, align 4
103  %mul = mul nsw i32 %0, 3
104  %idxprom1 = sext i32 %i.0 to i64
105  %arrayidx2 = getelementptr inbounds i32 addrspace(1)* %a, i64 %idxprom1
106  store i32 %mul, i32 addrspace(1)* %arrayidx2, align 4
107  %inc = add nsw i32 %i.0, 1
108  br label %for.cond
109
110for.end:                                          ; preds = %for.cond
111  ret void
112}
113
114; We should still be able to vectorize with noalias even if the
115; address spaces are different.
116define void @bar2(i32* noalias %a, i32 addrspace(1)* noalias %b, i32 %n) #0 {
117; CHECK-LABEL: @bar2(
118; CHECK: <4 x i32>
119; CHECK: ret
120
121entry:
122  br label %for.cond
123
124for.cond:                                         ; preds = %for.body, %entry
125  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
126  %cmp = icmp slt i32 %i.0, %n
127  br i1 %cmp, label %for.body, label %for.end
128
129for.body:                                         ; preds = %for.cond
130  %idxprom = sext i32 %i.0 to i64
131  %arrayidx = getelementptr inbounds i32 addrspace(1)* %b, i64 %idxprom
132  %0 = load i32 addrspace(1)* %arrayidx, align 4
133  %mul = mul nsw i32 %0, 3
134  %idxprom1 = sext i32 %i.0 to i64
135  %arrayidx2 = getelementptr inbounds i32* %a, i64 %idxprom1
136  store i32 %mul, i32* %arrayidx2, align 4
137  %inc = add nsw i32 %i.0, 1
138  br label %for.cond
139
140for.end:                                          ; preds = %for.cond
141  ret void
142}
143
144; Store to identified global with different address space. This isn't
145; generally safe and shouldn't be vectorized.
146define void @arst0(i32* %b, i32 %n) #0 {
147; CHECK-LABEL: @arst0(
148; CHECK-NOT: <4 x i32>
149; CHECK: ret
150
151entry:
152  br label %for.cond
153
154for.cond:                                         ; preds = %for.body, %entry
155  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
156  %cmp = icmp slt i32 %i.0, %n
157  br i1 %cmp, label %for.body, label %for.end
158
159for.body:                                         ; preds = %for.cond
160  %idxprom = sext i32 %i.0 to i64
161  %arrayidx = getelementptr inbounds i32* %b, i64 %idxprom
162  %0 = load i32* %arrayidx, align 4
163  %mul = mul nsw i32 %0, 3
164  %idxprom1 = sext i32 %i.0 to i64
165  %arrayidx2 = getelementptr inbounds [1024 x i32] addrspace(1)* @g_as1, i64 0, i64 %idxprom1
166  store i32 %mul, i32 addrspace(1)* %arrayidx2, align 4
167  %inc = add nsw i32 %i.0, 1
168  br label %for.cond
169
170for.end:                                          ; preds = %for.cond
171  ret void
172}
173
174
175; Load from identified global with different address space.
176; This isn't generally safe and shouldn't be vectorized.
177define void @arst1(i32* %b, i32 %n) #0 {
178; CHECK-LABEL: @arst1(
179; CHECK-NOT: <4 x i32>
180; CHECK: ret
181
182entry:
183  br label %for.cond
184
185for.cond:                                         ; preds = %for.body, %entry
186  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
187  %cmp = icmp slt i32 %i.0, %n
188  br i1 %cmp, label %for.body, label %for.end
189
190for.body:                                         ; preds = %for.cond
191  %idxprom = sext i32 %i.0 to i64
192  %arrayidx = getelementptr inbounds [1024 x i32] addrspace(1)* @g_as1, i64 0, i64 %idxprom
193  %0 = load i32 addrspace(1)* %arrayidx, align 4
194  %mul = mul nsw i32 %0, 3
195  %idxprom1 = sext i32 %i.0 to i64
196  %arrayidx2 = getelementptr inbounds i32* %b, i64 %idxprom1
197  store i32 %mul, i32* %arrayidx2, align 4
198  %inc = add nsw i32 %i.0, 1
199  br label %for.cond
200
201for.end:                                          ; preds = %for.cond
202  ret void
203}
204
205; Read and write to 2 identified globals in different address
206; spaces. This should be vectorized.
207define void @aoeu(i32 %n) #0 {
208; CHECK-LABEL: @aoeu(
209; CHECK: <4 x i32>
210; CHECK: ret
211
212entry:
213  br label %for.cond
214
215for.cond:                                         ; preds = %for.body, %entry
216  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
217  %cmp = icmp slt i32 %i.0, %n
218  br i1 %cmp, label %for.body, label %for.end
219
220for.body:                                         ; preds = %for.cond
221  %idxprom = sext i32 %i.0 to i64
222  %arrayidx = getelementptr inbounds [1024 x i32] addrspace(2)* @q_as2, i64 0, i64 %idxprom
223  %0 = load i32 addrspace(2)* %arrayidx, align 4
224  %mul = mul nsw i32 %0, 3
225  %idxprom1 = sext i32 %i.0 to i64
226  %arrayidx2 = getelementptr inbounds [1024 x i32] addrspace(1)* @g_as1, i64 0, i64 %idxprom1
227  store i32 %mul, i32 addrspace(1)* %arrayidx2, align 4
228  %inc = add nsw i32 %i.0, 1
229  br label %for.cond
230
231for.end:                                          ; preds = %for.cond
232  ret void
233}
234
235attributes #0 = { nounwind 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" }
236