Skip to content

Commit d749095

Browse files
fhahnfhahn
committed
[LAA] Add tests where we could derive NoDep due to no overlap.
Add additional tests where we can prove that the accesses are either completely before or after each other.
1 parent d367695 commit d749095

File tree

1 file changed

+319
-0
lines changed

1 file changed

+319
-0
lines changed

‎llvm/test/Analysis/LoopAccessAnalysis/accesses-completely-before-or-after.ll

Lines changed: 319 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,319 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --version 5
2+
; RUN: opt -passes="print<access-info>" -disable-output %s 2>&1 | FileCheck %s
3+
4+
define i32 @completely_before_or_after_true_dep_different_size(ptr %d) {
5+
; CHECK-LABEL: 'completely_before_or_after_true_dep_different_size'
6+
; CHECK-NEXT: loop:
7+
; CHECK-NEXT: Memory dependences are safe with run-time checks
8+
; CHECK-NEXT: Dependences:
9+
; CHECK-NEXT: Run-time memory checks:
10+
; CHECK-NEXT: Check 0:
11+
; CHECK-NEXT: Comparing group GRP0:
12+
; CHECK-NEXT: %gep.128.iv = getelementptr i64, ptr %gep.128, i64 %iv
13+
; CHECK-NEXT: Against group GRP1:
14+
; CHECK-NEXT: %gep.iv = getelementptr i32, ptr %d, i64 %iv
15+
; CHECK-NEXT: Grouped accesses:
16+
; CHECK-NEXT: Group GRP0:
17+
; CHECK-NEXT: (Low: (128 + %d) High: (384 + %d))
18+
; CHECK-NEXT: Member: {(128 + %d),+,8}<nw><%loop>
19+
; CHECK-NEXT: Group GRP1:
20+
; CHECK-NEXT: (Low: %d High: (128 + %d))
21+
; CHECK-NEXT: Member: {%d,+,4}<nw><%loop>
22+
; CHECK-EMPTY:
23+
; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
24+
; CHECK-NEXT: SCEV assumptions:
25+
; CHECK-EMPTY:
26+
; CHECK-NEXT: Expressions re-written:
27+
;
28+
entry:
29+
%gep.128 = getelementptr i8, ptr %d, i64 128
30+
br label %loop
31+
32+
loop:
33+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
34+
%gep.128.iv = getelementptr i64, ptr %gep.128, i64 %iv
35+
store i64 0, ptr %gep.128.iv, align 8
36+
%gep.iv = getelementptr i32, ptr %d, i64 %iv
37+
%l = load i32, ptr %gep.iv, align 4
38+
%iv.next = add i64 %iv, 1
39+
%ec = icmp eq i64 %iv.next, 32
40+
br i1 %ec, label %exit, label %loop
41+
42+
exit:
43+
ret i32 %l
44+
}
45+
46+
define i32 @may_overlap_true_dep_different_size(ptr %d) {
47+
; CHECK-LABEL: 'may_overlap_true_dep_different_size'
48+
; CHECK-NEXT: loop:
49+
; CHECK-NEXT: Memory dependences are safe with run-time checks
50+
; CHECK-NEXT: Dependences:
51+
; CHECK-NEXT: Run-time memory checks:
52+
; CHECK-NEXT: Check 0:
53+
; CHECK-NEXT: Comparing group GRP0:
54+
; CHECK-NEXT: %gep.128.iv = getelementptr i64, ptr %gep.128, i64 %iv
55+
; CHECK-NEXT: Against group GRP1:
56+
; CHECK-NEXT: %gep.iv = getelementptr i32, ptr %d, i64 %iv
57+
; CHECK-NEXT: Grouped accesses:
58+
; CHECK-NEXT: Group GRP0:
59+
; CHECK-NEXT: (Low: (127 + %d) High: (383 + %d))
60+
; CHECK-NEXT: Member: {(127 + %d),+,8}<nw><%loop>
61+
; CHECK-NEXT: Group GRP1:
62+
; CHECK-NEXT: (Low: %d High: (128 + %d))
63+
; CHECK-NEXT: Member: {%d,+,4}<nw><%loop>
64+
; CHECK-EMPTY:
65+
; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
66+
; CHECK-NEXT: SCEV assumptions:
67+
; CHECK-EMPTY:
68+
; CHECK-NEXT: Expressions re-written:
69+
;
70+
entry:
71+
%gep.128 = getelementptr i8, ptr %d, i64 127
72+
br label %loop
73+
74+
loop:
75+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
76+
%gep.128.iv = getelementptr i64, ptr %gep.128, i64 %iv
77+
store i64 0, ptr %gep.128.iv, align 8
78+
%gep.iv= getelementptr i32, ptr %d, i64 %iv
79+
%l = load i32, ptr %gep.iv, align 4
80+
%iv.next = add i64 %iv, 1
81+
%ec = icmp eq i64 %iv.next, 32
82+
br i1 %ec, label %exit, label %loop
83+
84+
exit:
85+
ret i32 %l
86+
}
87+
88+
define void @completely_after_stores_with_different_sizes(ptr %dst) {
89+
; CHECK-LABEL: 'completely_after_stores_with_different_sizes'
90+
; CHECK-NEXT: loop:
91+
; CHECK-NEXT: Memory dependences are safe with run-time checks
92+
; CHECK-NEXT: Dependences:
93+
; CHECK-NEXT: Run-time memory checks:
94+
; CHECK-NEXT: Check 0:
95+
; CHECK-NEXT: Comparing group GRP0:
96+
; CHECK-NEXT: %gep.iv = getelementptr i16, ptr %dst, i64 %iv
97+
; CHECK-NEXT: Against group GRP1:
98+
; CHECK-NEXT: %gep.dst.128.iv = getelementptr i8, ptr %gep.dst.128, i64 %iv
99+
; CHECK-NEXT: Grouped accesses:
100+
; CHECK-NEXT: Group GRP0:
101+
; CHECK-NEXT: (Low: %dst High: (128 + %dst)<nuw>)
102+
; CHECK-NEXT: Member: {%dst,+,2}<nw><%loop>
103+
; CHECK-NEXT: Group GRP1:
104+
; CHECK-NEXT: (Low: (128 + %dst)<nuw> High: (192 + %dst))
105+
; CHECK-NEXT: Member: {(128 + %dst)<nuw>,+,1}<nw><%loop>
106+
; CHECK-EMPTY:
107+
; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
108+
; CHECK-NEXT: SCEV assumptions:
109+
; CHECK-EMPTY:
110+
; CHECK-NEXT: Expressions re-written:
111+
;
112+
entry:
113+
%gep.dst.128 = getelementptr nuw i8, ptr %dst, i64 128
114+
br label %loop
115+
116+
loop:
117+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
118+
%gep.iv = getelementptr i16, ptr %dst, i64 %iv
119+
store i16 0, ptr %gep.iv, align 2
120+
%gep.dst.128.iv = getelementptr i8, ptr %gep.dst.128, i64 %iv
121+
store i8 0, ptr %gep.dst.128.iv, align 1
122+
%iv.next = add i64 %iv, 1
123+
%ec = icmp eq i64 %iv.next, 64
124+
br i1 %ec, label %exit, label %loop
125+
126+
exit:
127+
ret void
128+
}
129+
130+
131+
define void @may_overlap_stores_with_different_sizes(ptr %dst) {
132+
; CHECK-LABEL: 'may_overlap_stores_with_different_sizes'
133+
; CHECK-NEXT: loop:
134+
; CHECK-NEXT: Memory dependences are safe with run-time checks
135+
; CHECK-NEXT: Dependences:
136+
; CHECK-NEXT: Run-time memory checks:
137+
; CHECK-NEXT: Check 0:
138+
; CHECK-NEXT: Comparing group GRP0:
139+
; CHECK-NEXT: %gep.iv = getelementptr i16, ptr %dst, i64 %iv
140+
; CHECK-NEXT: Against group GRP1:
141+
; CHECK-NEXT: %gep.dst.128.iv = getelementptr i8, ptr %gep.dst.128, i64 %iv
142+
; CHECK-NEXT: Grouped accesses:
143+
; CHECK-NEXT: Group GRP0:
144+
; CHECK-NEXT: (Low: %dst High: (130 + %dst))
145+
; CHECK-NEXT: Member: {%dst,+,2}<nw><%loop>
146+
; CHECK-NEXT: Group GRP1:
147+
; CHECK-NEXT: (Low: (128 + %dst)<nuw> High: (193 + %dst))
148+
; CHECK-NEXT: Member: {(128 + %dst)<nuw>,+,1}<nw><%loop>
149+
; CHECK-EMPTY:
150+
; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
151+
; CHECK-NEXT: SCEV assumptions:
152+
; CHECK-EMPTY:
153+
; CHECK-NEXT: Expressions re-written:
154+
;
155+
entry:
156+
%gep.dst.128 = getelementptr nuw i8, ptr %dst, i64 128
157+
br label %loop
158+
159+
loop:
160+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
161+
%gep.iv = getelementptr i16, ptr %dst, i64 %iv
162+
store i16 0, ptr %gep.iv, align 2
163+
%gep.dst.128.iv = getelementptr i8, ptr %gep.dst.128, i64 %iv
164+
store i8 0, ptr %gep.dst.128.iv, align 1
165+
%iv.next = add i64 %iv, 1
166+
%ec = icmp eq i64 %iv.next, 65
167+
br i1 %ec, label %exit, label %loop
168+
169+
exit: ; preds = %loop
170+
ret void
171+
}
172+
173+
define void @completely_before_or_after_non_const_distance(ptr %dst) {
174+
; CHECK-LABEL: 'completely_before_or_after_non_const_distance'
175+
; CHECK-NEXT: loop:
176+
; CHECK-NEXT: Report: unsafe dependent memory operations in loop. Use #pragma clang loop distribute(enable) to allow loop distribution to attempt to isolate the offending operations into a separate loop
177+
; CHECK-NEXT: Unknown data dependence.
178+
; CHECK-NEXT: Dependences:
179+
; CHECK-NEXT: Unknown:
180+
; CHECK-NEXT: store i32 0, ptr %gep.iv.mul, align 4 ->
181+
; CHECK-NEXT: store i32 0, ptr %gep.off.iv, align 4
182+
; CHECK-EMPTY:
183+
; CHECK-NEXT: Run-time memory checks:
184+
; CHECK-NEXT: Grouped accesses:
185+
; CHECK-EMPTY:
186+
; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
187+
; CHECK-NEXT: SCEV assumptions:
188+
; CHECK-EMPTY:
189+
; CHECK-NEXT: Expressions re-written:
190+
;
191+
entry:
192+
%gep.off = getelementptr i8, ptr %dst, i64 576
193+
br label %loop
194+
195+
loop:
196+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
197+
%iv.mul = mul i64 %iv, 84
198+
%gep.404 = getelementptr i8, ptr %dst, i64 404
199+
%gep.iv.mul = getelementptr i8, ptr %gep.404, i64 %iv.mul
200+
store i32 0, ptr %gep.iv.mul, align 4
201+
%gep.off.iv = getelementptr i32, ptr %gep.off, i64 %iv
202+
store i32 0, ptr %gep.off.iv, align 4
203+
%iv.next = add i64 %iv, 1
204+
%ec = icmp eq i64 %iv.next, 3
205+
br i1 %ec, label %exit, label %loop
206+
207+
exit:
208+
ret void
209+
}
210+
211+
define void @overlap_non_const_distance(ptr %dst) {
212+
; CHECK-LABEL: 'overlap_non_const_distance'
213+
; CHECK-NEXT: loop:
214+
; CHECK-NEXT: Report: unsafe dependent memory operations in loop. Use #pragma clang loop distribute(enable) to allow loop distribution to attempt to isolate the offending operations into a separate loop
215+
; CHECK-NEXT: Unknown data dependence.
216+
; CHECK-NEXT: Dependences:
217+
; CHECK-NEXT: Unknown:
218+
; CHECK-NEXT: store i32 0, ptr %gep.iv.mul, align 4 ->
219+
; CHECK-NEXT: store i32 0, ptr %gep.off.iv, align 4
220+
; CHECK-EMPTY:
221+
; CHECK-NEXT: Run-time memory checks:
222+
; CHECK-NEXT: Grouped accesses:
223+
; CHECK-EMPTY:
224+
; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
225+
; CHECK-NEXT: SCEV assumptions:
226+
; CHECK-EMPTY:
227+
; CHECK-NEXT: Expressions re-written:
228+
;
229+
entry:
230+
%gep.off = getelementptr i8, ptr %dst, i64 575
231+
br label %loop
232+
233+
loop:
234+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
235+
%iv.mul = mul i64 %iv, 84
236+
%gep.404 = getelementptr i8, ptr %dst, i64 404
237+
%gep.iv.mul = getelementptr i8, ptr %gep.404, i64 %iv.mul
238+
store i32 0, ptr %gep.iv.mul, align 4
239+
%gep.off.iv = getelementptr i32, ptr %gep.off, i64 %iv
240+
store i32 0, ptr %gep.off.iv, align 4
241+
%iv.next = add i64 %iv, 1
242+
%ec = icmp eq i64 %iv.next, 3
243+
br i1 %ec, label %exit, label %loop
244+
245+
exit:
246+
ret void
247+
}
248+
249+
define void @accesses_completely_before_or_after_instead_backwards_vectorizable(ptr dereferenceable(800) %dst) {
250+
; CHECK-LABEL: 'accesses_completely_before_or_after_instead_backwards_vectorizable'
251+
; CHECK-NEXT: loop:
252+
; CHECK-NEXT: Memory dependences are safe with a maximum safe vector width of 128 bits
253+
; CHECK-NEXT: Dependences:
254+
; CHECK-NEXT: BackwardVectorizable:
255+
; CHECK-NEXT: store i16 0, ptr %gep.mul.2, align 2 ->
256+
; CHECK-NEXT: store i16 0, ptr %gep.iv.32, align 2
257+
; CHECK-EMPTY:
258+
; CHECK-NEXT: Run-time memory checks:
259+
; CHECK-NEXT: Grouped accesses:
260+
; CHECK-EMPTY:
261+
; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
262+
; CHECK-NEXT: SCEV assumptions:
263+
; CHECK-EMPTY:
264+
; CHECK-NEXT: Expressions re-written:
265+
;
266+
entry:
267+
br label %loop
268+
269+
loop:
270+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
271+
%mul.2 = shl i64 %iv, 1
272+
%gep.mul.2 = getelementptr i16, ptr %dst, i64 %mul.2
273+
store i16 0, ptr %gep.mul.2, align 2
274+
%iv.32 = add i64 %iv, 32
275+
%gep.iv.32 = getelementptr i16, ptr %dst, i64 %iv.32
276+
store i16 0, ptr %gep.iv.32, align 2
277+
%iv.next = add i64 %iv, 1
278+
%ec = icmp eq i64 %iv, 15
279+
br i1 %ec, label %exit, label %loop
280+
281+
exit:
282+
ret void
283+
}
284+
285+
define void @accesses_may_overlap_backwards_vectorizable(ptr dereferenceable(800) %dst) {
286+
; CHECK-LABEL: 'accesses_may_overlap_backwards_vectorizable'
287+
; CHECK-NEXT: loop:
288+
; CHECK-NEXT: Memory dependences are safe with a maximum safe vector width of 128 bits
289+
; CHECK-NEXT: Dependences:
290+
; CHECK-NEXT: BackwardVectorizable:
291+
; CHECK-NEXT: store i16 0, ptr %gep.mul.2, align 2 ->
292+
; CHECK-NEXT: store i16 0, ptr %gep.iv.32, align 2
293+
; CHECK-EMPTY:
294+
; CHECK-NEXT: Run-time memory checks:
295+
; CHECK-NEXT: Grouped accesses:
296+
; CHECK-EMPTY:
297+
; CHECK-NEXT: Non vectorizable stores to invariant address were not found in loop.
298+
; CHECK-NEXT: SCEV assumptions:
299+
; CHECK-EMPTY:
300+
; CHECK-NEXT: Expressions re-written:
301+
;
302+
entry:
303+
br label %loop
304+
305+
loop:
306+
%iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
307+
%mul.2 = shl i64 %iv, 1
308+
%gep.mul.2 = getelementptr i16, ptr %dst, i64 %mul.2
309+
store i16 0, ptr %gep.mul.2, align 2
310+
%iv.32 = add i64 %iv, 32
311+
%gep.iv.32 = getelementptr i16, ptr %dst, i64 %iv.32
312+
store i16 0, ptr %gep.iv.32, align 2
313+
%iv.next = add i64 %iv, 1
314+
%ec = icmp eq i64 %iv, 16
315+
br i1 %ec, label %exit, label %loop
316+
317+
exit:
318+
ret void
319+
}

0 commit comments

Comments
 (0)