1 ; RUN: opt -disable-output < %s -aa-pipeline=scev-aa -passes=aa-eval -print-all-alias-modref-info \
2 ; RUN: 2>&1 | FileCheck %s
4 ; At the time of this writing, misses the example of the form
5 ; A[i+(j+1)] != A[i+j], which can arise from multi-dimensional array references,
6 ; and the example of the form A[0] != A[i+1], where i+1 is known to be positive.
8 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64"
10 ; p[i] and p[i+1] don't alias.
12 ; CHECK-LABEL: Function: loop
13 ; CHECK: NoAlias: double* %pi, double* %pi.next
15 define void @loop(ptr nocapture %p, i64 %n) nounwind {
17 %j = icmp sgt i64 %n, 0
18 br i1 %j, label %bb, label %return
21 %i = phi i64 [ 0, %entry ], [ %i.next, %bb ]
22 %pi = getelementptr double, ptr %p, i64 %i
23 %i.next = add i64 %i, 1
24 %pi.next = getelementptr double, ptr %p, i64 %i.next
25 %x = load double, ptr %pi
26 %y = load double, ptr %pi.next
27 %z = fmul double %x, %y
28 store double %z, ptr %pi
29 %exitcond = icmp eq i64 %i.next, %n
30 br i1 %exitcond, label %return, label %bb
36 ; Slightly more involved: p[j][i], p[j][i+1], and p[j+1][i] don't alias.
38 ; CHECK-LABEL: Function: nestedloop
39 ; CHECK: NoAlias: double* %pi.j, double* %pi.next.j
40 ; CHECK: NoAlias: double* %pi.j, double* %pi.j.next
41 ; CHECK: NoAlias: double* %pi.j.next, double* %pi.next.j
43 define void @nestedloop(ptr nocapture %p, i64 %m) nounwind {
45 %k = icmp sgt i64 %m, 0
46 br i1 %k, label %guard, label %return
49 %l = icmp sgt i64 91, 0
50 br i1 %l, label %outer.loop, label %return
53 %j = phi i64 [ 0, %guard ], [ %j.next, %outer.latch ]
57 %i = phi i64 [ 0, %outer.loop ], [ %i.next, %bb ]
58 %i.next = add i64 %i, 1
61 %pi.j = getelementptr double, ptr %p, i64 %e
62 %f = add i64 %i.next, %j
63 %pi.next.j = getelementptr double, ptr %p, i64 %f
64 %x = load double, ptr %pi.j
65 %y = load double, ptr %pi.next.j
66 %z = fmul double %x, %y
67 store double %z, ptr %pi.j
71 %pi.j.next = getelementptr double, ptr %p, i64 %g
72 %a = load double, ptr %pi.j.next
73 %b = fmul double %x, %a
74 store double %b, ptr %pi.j.next
76 %exitcond = icmp eq i64 %i.next, 91
77 br i1 %exitcond, label %outer.latch, label %bb
80 %j.next = add i64 %j, 91
81 %h = icmp eq i64 %j.next, %m
82 br i1 %h, label %return, label %outer.loop
88 ; Even more involved: same as nestedloop, but with a variable extent.
89 ; When n is 1, p[j+1][i] does alias p[j][i+1], and there's no way to
90 ; prove whether n will be greater than 1, so that relation will always
91 ; by MayAlias. The loop is guarded by a n > 0 test though, so
92 ; p[j+1][i] and p[j][i] can theoretically be determined to be NoAlias,
93 ; however the analysis currently doesn't do that.
94 ; TODO: Make the analysis smarter and turn that MayAlias into a NoAlias.
96 ; CHECK-LABEL: Function: nestedloop_more
97 ; CHECK: NoAlias: double* %pi.j, double* %pi.next.j
98 ; CHECK: MayAlias: double* %pi.j, double* %pi.j.next
100 define void @nestedloop_more(ptr nocapture %p, i64 %n, i64 %m) nounwind {
102 %k = icmp sgt i64 %m, 0
103 br i1 %k, label %guard, label %return
106 %l = icmp sgt i64 %n, 0
107 br i1 %l, label %outer.loop, label %return
110 %j = phi i64 [ 0, %guard ], [ %j.next, %outer.latch ]
114 %i = phi i64 [ 0, %outer.loop ], [ %i.next, %bb ]
115 %i.next = add i64 %i, 1
118 %pi.j = getelementptr double, ptr %p, i64 %e
119 %f = add i64 %i.next, %j
120 %pi.next.j = getelementptr double, ptr %p, i64 %f
121 %x = load double, ptr %pi.j
122 %y = load double, ptr %pi.next.j
123 %z = fmul double %x, %y
124 store double %z, ptr %pi.j
128 %pi.j.next = getelementptr double, ptr %p, i64 %g
129 %a = load double, ptr %pi.j.next
130 %b = fmul double %x, %a
131 store double %b, ptr %pi.j.next
133 %exitcond = icmp eq i64 %i.next, %n
134 br i1 %exitcond, label %outer.latch, label %bb
137 %j.next = add i64 %j, %n
138 %h = icmp eq i64 %j.next, %m
139 br i1 %h, label %return, label %outer.loop
145 ; ScalarEvolution expands field offsets into constants, which allows it to
146 ; do aggressive analysis. Contrast this with BasicAA, which works by
147 ; recognizing GEP idioms.
149 %struct.A = type { %struct.B, i32, i32 }
150 %struct.B = type { double }
152 ; CHECK-LABEL: Function: foo
153 ; CHECK-DAG: NoAlias: %struct.B* %A, i32* %Z
154 ; CHECK-DAG: NoAlias: %struct.B* %A, %struct.B* %C
155 ; CHECK-DAG: MustAlias: %struct.B* %C, i32* %Z
156 ; CHECK-DAG: NoAlias: %struct.B* %A, i32* %C
157 ; CHECK-DAG: MustAlias: i32* %C, i32* %Z
158 ; CHECK-DAG: MustAlias: %struct.B* %C, i32* %Y
159 ; CHECK-DAG: MustAlias: i32* %C, i32* %Y
163 %A = alloca %struct.A
164 %Z = getelementptr %struct.A, ptr %A, i32 0, i32 1
165 %C = getelementptr %struct.B, ptr %A, i32 1
166 %Y = getelementptr %struct.A, ptr %A, i32 0, i32 1
167 load %struct.B, ptr %A
168 load %struct.B, ptr %C
175 ; CHECK-LABEL: Function: bar
176 ; CHECK-DAG: NoAlias: %struct.B* %M, i32* %P
177 ; CHECK-DAG: NoAlias: %struct.B* %M, %struct.B* %R
178 ; CHECK-DAG: MustAlias: i32* %P, %struct.B* %R
179 ; CHECK-DAG: NoAlias: %struct.B* %M, i32* %R
180 ; CHECK-DAG: MustAlias: i32* %P, i32* %R
181 ; CHECK-DAG: MustAlias: %struct.B* %R, i32* %V
182 ; CHECK-DAG: MustAlias: i32* %R, i32* %V
185 %M = alloca %struct.A
186 %P = getelementptr %struct.A, ptr %M, i32 0, i32 1
187 %R = getelementptr %struct.B, ptr %M, i32 1
188 %V = getelementptr %struct.A, ptr %M, i32 0, i32 1
189 load %struct.B, ptr %M
190 load %struct.B, ptr %R
197 ; CHECK: Function: nonnegative: 2 pointers, 0 call sites
198 ; CHECK: NoAlias: i64* %arrayidx, i64* %p
200 define void @nonnegative(ptr %p) nounwind {
204 for.body: ; preds = %entry, %for.body
205 %i = phi i64 [ %inc, %for.body ], [ 0, %entry ] ; <i64> [#uses=2]
206 %inc = add nsw i64 %i, 1 ; <i64> [#uses=2]
207 %arrayidx = getelementptr inbounds i64, ptr %p, i64 %inc
208 store i64 0, ptr %arrayidx
209 %tmp6 = load i64, ptr %p ; <i64> [#uses=1]
210 %cmp = icmp slt i64 %inc, %tmp6 ; <i1> [#uses=1]
211 br i1 %cmp, label %for.body, label %for.end
213 for.end: ; preds = %for.body, %entry
217 ; CHECK-LABEL: Function: test_no_dom: 3 pointers, 0 call sites
218 ; CHECK: MayAlias: double* %addr1, double* %data
219 ; CHECK: NoAlias: double* %addr2, double* %data
220 ; CHECK: MayAlias: double* %addr1, double* %addr2
222 ; In this case, checking %addr1 and %add2 involves two addrecs in two
223 ; different loops where neither dominates the other. This used to crash
224 ; because we expected the arguments to an AddExpr to have a strict
226 define void @test_no_dom(ptr %data, i1 %arg) {
228 load double, ptr %data
232 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.latch ]
233 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
234 br i1 %arg, label %subloop1, label %subloop2
237 %iv1 = phi i32 [0, %for.body], [%iv1.next, %subloop1]
238 %iv1.next = add i32 %iv1, 1
239 %addr1 = getelementptr double, ptr %data, i32 %iv1
240 store double 0.0, ptr %addr1
241 %cmp1 = icmp slt i32 %iv1, 200
242 br i1 %cmp1, label %subloop1, label %for.latch
245 %iv2 = phi i32 [400, %for.body], [%iv2.next, %subloop2]
246 %iv2.next = add i32 %iv2, 1
247 %addr2 = getelementptr double, ptr %data, i32 %iv2
248 store double 0.0, ptr %addr2
249 %cmp2 = icmp slt i32 %iv2, 600
250 br i1 %cmp2, label %subloop2, label %for.latch
259 declare ptr @get_addr(i32 %i)
261 ; CHECK-LABEL: Function: test_no_dom2: 3 pointers, 2 call sites
262 ; CHECK: MayAlias: double* %addr1, double* %data
263 ; CHECK: MayAlias: double* %addr2, double* %data
264 ; CHECK: MayAlias: double* %addr1, double* %addr2
266 ; In this case, checking %addr1 and %add2 involves two addrecs in two
267 ; different loops where neither dominates the other. This is analogous
268 ; to test_no_dom, but involves SCEVUnknown as opposed to SCEVAddRecExpr.
269 define void @test_no_dom2(ptr %data, i1 %arg) {
271 load double, ptr %data
275 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.latch ]
276 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
277 br i1 %arg, label %subloop1, label %subloop2
280 %iv1 = phi i32 [0, %for.body], [%iv1.next, %subloop1]
281 %iv1.next = add i32 %iv1, 1
282 %addr1 = call ptr @get_addr(i32 %iv1)
283 store double 0.0, ptr %addr1
284 %cmp1 = icmp slt i32 %iv1, 200
285 br i1 %cmp1, label %subloop1, label %for.latch
288 %iv2 = phi i32 [400, %for.body], [%iv2.next, %subloop2]
289 %iv2.next = add i32 %iv2, 1
290 %addr2 = call ptr @get_addr(i32 %iv2)
291 store double 0.0, ptr %addr2
292 %cmp2 = icmp slt i32 %iv2, 600
293 br i1 %cmp2, label %subloop2, label %for.latch
303 ; CHECK-LABEL: Function: test_dom: 3 pointers, 0 call sites
304 ; CHECK: MayAlias: double* %addr1, double* %data
305 ; CHECK: NoAlias: double* %addr2, double* %data
306 ; CHECK: NoAlias: double* %addr1, double* %addr2
308 ; This is a variant of test_non_dom where the second subloop is
309 ; dominated by the first. As a result of that, we can nest the
310 ; addrecs and cancel out the %data base pointer.
311 define void @test_dom(ptr %data) {
313 load double, ptr %data
317 %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.latch ]
318 %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
322 %iv1 = phi i32 [0, %for.body], [%iv1.next, %subloop1]
323 %iv1.next = add i32 %iv1, 1
324 %addr1 = getelementptr double, ptr %data, i32 %iv1
325 store double 0.0, ptr %addr1
326 %cmp1 = icmp slt i32 %iv1, 200
327 br i1 %cmp1, label %subloop1, label %subloop2
330 %iv2 = phi i32 [400, %subloop1], [%iv2.next, %subloop2]
331 %iv2.next = add i32 %iv2, 1
332 %addr2 = getelementptr double, ptr %data, i32 %iv2
333 store double 0.0, ptr %addr2
334 %cmp2 = icmp slt i32 %iv2, 600
335 br i1 %cmp2, label %subloop2, label %for.latch
344 ; CHECK-LABEL: Function: test_different_pointer_bases_of_inttoptr: 2 pointers, 0 call sites
345 ; CHECK: NoAlias: <16 x i8>* %tmp5, <16 x i8>* %tmp7
347 define void @test_different_pointer_bases_of_inttoptr() {
352 %tmp = phi i32 [ %next, %for.body ], [ 1, %entry ]
353 %tmp1 = shl nsw i32 %tmp, 1
354 %tmp2 = add nuw nsw i32 %tmp1, %tmp1
355 %tmp3 = mul nsw i32 %tmp2, 1408
356 %tmp4 = add nsw i32 %tmp3, 1408
357 %tmp5 = getelementptr inbounds i8, ptr inttoptr (i32 1024 to ptr), i32 %tmp1
358 %tmp6 = load <16 x i8>, ptr %tmp5, align 1
359 %tmp7 = getelementptr inbounds i8, ptr inttoptr (i32 4096 to ptr), i32 %tmp4
360 store <16 x i8> %tmp6, ptr %tmp7, align 1
362 %next = add i32 %tmp, 2
363 %exitcond = icmp slt i32 %next, 10
364 br i1 %exitcond, label %for.body, label %for.end