1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2 ; RUN: opt -loop-versioning -S < %s | FileCheck %s -check-prefix=LV
4 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
8 ; for (int i = 0; i < n; i++) {
9 ; A[2 * index] = A[2 * index] + B[i];
13 ; SCEV is unable to prove that A[2 * i] does not overflow.
15 ; Analyzing the IR does not help us because the GEPs are not
16 ; affine AddRecExprs. However, we can turn them into AddRecExprs
17 ; using SCEV Predicates.
19 ; Once we have an affine expression we need to add an additional NUSW
20 ; to check that the pointers don't wrap since the GEPs are not
23 ; The expression for %mul_ext as analyzed by SCEV is
24 ; (zext i32 {0,+,2}<%for.body> to i64)
25 ; We have added the nusw flag to turn this expression into the SCEV expression:
26 ; i64 {0,+,2}<%for.body>
28 define void @f1(i16* noalias %a,
30 ; LV-NEXT: for.body.lver.check:
31 ; LV-NEXT: [[A2:%.*]] = ptrtoint i16* [[A:%.*]] to i64
32 ; LV-NEXT: [[TMP0:%.*]] = add i64 [[N:%.*]], -1
33 ; LV-NEXT: [[TMP1:%.*]] = trunc i64 [[TMP0]] to i32
34 ; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP1]])
35 ; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0
36 ; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1
37 ; LV-NEXT: [[TMP2:%.*]] = add i32 0, [[MUL_RESULT]]
38 ; LV-NEXT: [[TMP3:%.*]] = sub i32 0, [[MUL_RESULT]]
39 ; LV-NEXT: [[TMP4:%.*]] = icmp ugt i32 [[TMP3]], 0
40 ; LV-NEXT: [[TMP5:%.*]] = icmp ult i32 [[TMP2]], 0
41 ; LV-NEXT: [[TMP6:%.*]] = select i1 false, i1 [[TMP4]], i1 [[TMP5]]
42 ; LV-NEXT: [[TMP7:%.*]] = icmp ugt i64 [[TMP0]], 4294967295
43 ; LV-NEXT: [[TMP8:%.*]] = or i1 [[TMP6]], [[TMP7]]
44 ; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP8]], [[MUL_OVERFLOW]]
45 ; LV-NEXT: [[TMP10:%.*]] = or i1 false, [[TMP9]]
46 ; LV-NEXT: [[MUL3:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]])
47 ; LV-NEXT: [[MUL_RESULT4:%.*]] = extractvalue { i64, i1 } [[MUL3]], 0
48 ; LV-NEXT: [[MUL_OVERFLOW5:%.*]] = extractvalue { i64, i1 } [[MUL3]], 1
49 ; LV-NEXT: [[TMP11:%.*]] = add i64 [[A2]], [[MUL_RESULT4]]
50 ; LV-NEXT: [[TMP12:%.*]] = sub i64 [[A2]], [[MUL_RESULT4]]
51 ; LV-NEXT: [[TMP13:%.*]] = icmp ugt i64 [[TMP12]], [[A2]]
52 ; LV-NEXT: [[TMP14:%.*]] = icmp ult i64 [[TMP11]], [[A2]]
53 ; LV-NEXT: [[TMP15:%.*]] = select i1 false, i1 [[TMP13]], i1 [[TMP14]]
54 ; LV-NEXT: [[TMP16:%.*]] = or i1 [[TMP15]], [[MUL_OVERFLOW5]]
55 ; LV-NEXT: [[TMP17:%.*]] = or i1 [[TMP10]], [[TMP16]]
56 ; LV-NEXT: br i1 [[TMP17]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]]
57 ; LV: for.body.ph.lver.orig:
58 ; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]]
59 ; LV: for.body.lver.orig:
60 ; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
61 ; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC1_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
62 ; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2
63 ; LV-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = zext i32 [[MUL_LVER_ORIG]] to i64
64 ; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT_LVER_ORIG]]
65 ; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2
66 ; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]]
67 ; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2
68 ; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]]
69 ; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2
70 ; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1
71 ; LV-NEXT: [[INC1_LVER_ORIG]] = add i32 [[IND1_LVER_ORIG]], 1
72 ; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]]
73 ; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]]
75 ; LV-NEXT: br label [[FOR_BODY:%.*]]
77 ; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
78 ; LV-NEXT: [[IND1:%.*]] = phi i32 [ 0, [[FOR_BODY_PH]] ], [ [[INC1:%.*]], [[FOR_BODY]] ]
79 ; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2
80 ; LV-NEXT: [[MUL_EXT:%.*]] = zext i32 [[MUL]] to i64
81 ; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT]]
82 ; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2
83 ; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr i16, i16* [[B]], i64 [[IND]]
84 ; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2
85 ; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]]
86 ; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2
87 ; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1
88 ; LV-NEXT: [[INC1]] = add i32 [[IND1]], 1
89 ; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]]
90 ; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]]
91 ; LV: for.end.loopexit:
92 ; LV-NEXT: br label [[FOR_END:%.*]]
93 ; LV: for.end.loopexit6:
94 ; LV-NEXT: br label [[FOR_END]]
98 i16* noalias %b, i64 %N) {
102 for.body: ; preds = %for.body, %entry
103 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ]
104 %ind1 = phi i32 [ 0, %entry ], [ %inc1, %for.body ]
106 %mul = mul i32 %ind1, 2
107 %mul_ext = zext i32 %mul to i64
109 %arrayidxA = getelementptr i16, i16* %a, i64 %mul_ext
110 %loadA = load i16, i16* %arrayidxA, align 2
112 %arrayidxB = getelementptr i16, i16* %b, i64 %ind
113 %loadB = load i16, i16* %arrayidxB, align 2
115 %add = mul i16 %loadA, %loadB
117 store i16 %add, i16* %arrayidxA, align 2
119 %inc = add nuw nsw i64 %ind, 1
120 %inc1 = add i32 %ind1, 1
122 %exitcond = icmp eq i64 %inc, %N
123 br i1 %exitcond, label %for.end, label %for.body
125 for.end: ; preds = %for.body
130 ; unsigned index = n;
131 ; for (int i = 0; i < n; i++) {
132 ; A[2 * index] = A[2 * index] + B[i];
136 ; the SCEV expression for 2 * index is not an AddRecExpr
137 ; (and implictly not affine). However, we are able to make assumptions
138 ; that will turn the expression into an affine one and continue the
141 ; Once we have an affine expression we need to add an additional NUSW
142 ; to check that the pointers don't wrap since the GEPs are not
145 ; This loop has a negative stride for A, and the nusw flag is required in
146 ; order to properly extend the increment from i32 -4 to i64 -4.
148 ; The expression for %mul_ext as analyzed by SCEV is
149 ; (zext i32 {(2 * (trunc i64 %N to i32)),+,-2}<%for.body> to i64)
150 ; We have added the nusw flag to turn this expression into the following SCEV:
151 ; i64 {zext i32 (2 * (trunc i64 %N to i32)) to i64,+,-2}<%for.body>
153 define void @f2(i16* noalias %a,
155 ; LV-NEXT: for.body.lver.check:
156 ; LV-NEXT: [[A2:%.*]] = ptrtoint i16* [[A:%.*]] to i64
157 ; LV-NEXT: [[TRUNCN:%.*]] = trunc i64 [[N:%.*]] to i32
158 ; LV-NEXT: [[TMP0:%.*]] = add i64 [[N]], -1
159 ; LV-NEXT: [[TMP1:%.*]] = shl i32 [[TRUNCN]], 1
160 ; LV-NEXT: [[TMP2:%.*]] = trunc i64 [[TMP0]] to i32
161 ; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP2]])
162 ; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0
163 ; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1
164 ; LV-NEXT: [[TMP3:%.*]] = add i32 [[TMP1]], [[MUL_RESULT]]
165 ; LV-NEXT: [[TMP4:%.*]] = sub i32 [[TMP1]], [[MUL_RESULT]]
166 ; LV-NEXT: [[TMP5:%.*]] = icmp ugt i32 [[TMP4]], [[TMP1]]
167 ; LV-NEXT: [[TMP6:%.*]] = icmp ult i32 [[TMP3]], [[TMP1]]
168 ; LV-NEXT: [[TMP7:%.*]] = select i1 true, i1 [[TMP5]], i1 [[TMP6]]
169 ; LV-NEXT: [[TMP8:%.*]] = icmp ugt i64 [[TMP0]], 4294967295
170 ; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP7]], [[TMP8]]
171 ; LV-NEXT: [[TMP10:%.*]] = or i1 [[TMP9]], [[MUL_OVERFLOW]]
172 ; LV-NEXT: [[TMP11:%.*]] = or i1 false, [[TMP10]]
173 ; LV-NEXT: [[TMP12:%.*]] = trunc i64 [[N]] to i31
174 ; LV-NEXT: [[TMP13:%.*]] = zext i31 [[TMP12]] to i64
175 ; LV-NEXT: [[TMP14:%.*]] = shl nuw nsw i64 [[TMP13]], 2
176 ; LV-NEXT: [[TMP15:%.*]] = add i64 [[A2]], [[TMP14]]
177 ; LV-NEXT: [[MUL3:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]])
178 ; LV-NEXT: [[MUL_RESULT4:%.*]] = extractvalue { i64, i1 } [[MUL3]], 0
179 ; LV-NEXT: [[MUL_OVERFLOW5:%.*]] = extractvalue { i64, i1 } [[MUL3]], 1
180 ; LV-NEXT: [[TMP16:%.*]] = add i64 [[TMP15]], [[MUL_RESULT4]]
181 ; LV-NEXT: [[TMP17:%.*]] = sub i64 [[TMP15]], [[MUL_RESULT4]]
182 ; LV-NEXT: [[TMP18:%.*]] = icmp ugt i64 [[TMP17]], [[TMP15]]
183 ; LV-NEXT: [[TMP19:%.*]] = icmp ult i64 [[TMP16]], [[TMP15]]
184 ; LV-NEXT: [[TMP20:%.*]] = select i1 true, i1 [[TMP18]], i1 [[TMP19]]
185 ; LV-NEXT: [[TMP21:%.*]] = or i1 [[TMP20]], [[MUL_OVERFLOW5]]
186 ; LV-NEXT: [[TMP22:%.*]] = or i1 [[TMP11]], [[TMP21]]
187 ; LV-NEXT: br i1 [[TMP22]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]]
188 ; LV: for.body.ph.lver.orig:
189 ; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]]
190 ; LV: for.body.lver.orig:
191 ; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
192 ; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH_LVER_ORIG]] ], [ [[DEC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
193 ; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2
194 ; LV-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = zext i32 [[MUL_LVER_ORIG]] to i64
195 ; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT_LVER_ORIG]]
196 ; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2
197 ; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]]
198 ; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2
199 ; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]]
200 ; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2
201 ; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1
202 ; LV-NEXT: [[DEC_LVER_ORIG]] = sub i32 [[IND1_LVER_ORIG]], 1
203 ; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]]
204 ; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]]
206 ; LV-NEXT: br label [[FOR_BODY:%.*]]
208 ; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
209 ; LV-NEXT: [[IND1:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH]] ], [ [[DEC:%.*]], [[FOR_BODY]] ]
210 ; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2
211 ; LV-NEXT: [[MUL_EXT:%.*]] = zext i32 [[MUL]] to i64
212 ; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT]]
213 ; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2
214 ; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr i16, i16* [[B]], i64 [[IND]]
215 ; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2
216 ; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]]
217 ; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2
218 ; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1
219 ; LV-NEXT: [[DEC]] = sub i32 [[IND1]], 1
220 ; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]]
221 ; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]]
222 ; LV: for.end.loopexit:
223 ; LV-NEXT: br label [[FOR_END:%.*]]
224 ; LV: for.end.loopexit6:
225 ; LV-NEXT: br label [[FOR_END]]
229 i16* noalias %b, i64 %N) {
231 %TruncN = trunc i64 %N to i32
234 for.body: ; preds = %for.body, %entry
235 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ]
236 %ind1 = phi i32 [ %TruncN, %entry ], [ %dec, %for.body ]
238 %mul = mul i32 %ind1, 2
239 %mul_ext = zext i32 %mul to i64
241 %arrayidxA = getelementptr i16, i16* %a, i64 %mul_ext
242 %loadA = load i16, i16* %arrayidxA, align 2
244 %arrayidxB = getelementptr i16, i16* %b, i64 %ind
245 %loadB = load i16, i16* %arrayidxB, align 2
247 %add = mul i16 %loadA, %loadB
249 store i16 %add, i16* %arrayidxA, align 2
251 %inc = add nuw nsw i64 %ind, 1
252 %dec = sub i32 %ind1, 1
254 %exitcond = icmp eq i64 %inc, %N
255 br i1 %exitcond, label %for.end, label %for.body
257 for.end: ; preds = %for.body
261 ; We replicate the tests above, but this time sign extend 2 * index instead
262 ; of zero extending it.
264 ; The expression for %mul_ext as analyzed by SCEV is
265 ; i64 (sext i32 {0,+,2}<%for.body> to i64)
266 ; We have added the nssw flag to turn this expression into the following SCEV:
267 ; i64 {0,+,2}<%for.body>
269 define void @f3(i16* noalias %a,
271 ; LV-NEXT: for.body.lver.check:
272 ; LV-NEXT: [[A2:%.*]] = ptrtoint i16* [[A:%.*]] to i64
273 ; LV-NEXT: [[TMP0:%.*]] = add i64 [[N:%.*]], -1
274 ; LV-NEXT: [[TMP1:%.*]] = trunc i64 [[TMP0]] to i32
275 ; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP1]])
276 ; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0
277 ; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1
278 ; LV-NEXT: [[TMP2:%.*]] = add i32 0, [[MUL_RESULT]]
279 ; LV-NEXT: [[TMP3:%.*]] = sub i32 0, [[MUL_RESULT]]
280 ; LV-NEXT: [[TMP4:%.*]] = icmp sgt i32 [[TMP3]], 0
281 ; LV-NEXT: [[TMP5:%.*]] = icmp slt i32 [[TMP2]], 0
282 ; LV-NEXT: [[TMP6:%.*]] = select i1 false, i1 [[TMP4]], i1 [[TMP5]]
283 ; LV-NEXT: [[TMP7:%.*]] = icmp ugt i64 [[TMP0]], 4294967295
284 ; LV-NEXT: [[TMP8:%.*]] = or i1 [[TMP6]], [[TMP7]]
285 ; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP8]], [[MUL_OVERFLOW]]
286 ; LV-NEXT: [[TMP10:%.*]] = or i1 false, [[TMP9]]
287 ; LV-NEXT: [[MUL3:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]])
288 ; LV-NEXT: [[MUL_RESULT4:%.*]] = extractvalue { i64, i1 } [[MUL3]], 0
289 ; LV-NEXT: [[MUL_OVERFLOW5:%.*]] = extractvalue { i64, i1 } [[MUL3]], 1
290 ; LV-NEXT: [[TMP11:%.*]] = add i64 [[A2]], [[MUL_RESULT4]]
291 ; LV-NEXT: [[TMP12:%.*]] = sub i64 [[A2]], [[MUL_RESULT4]]
292 ; LV-NEXT: [[TMP13:%.*]] = icmp ugt i64 [[TMP12]], [[A2]]
293 ; LV-NEXT: [[TMP14:%.*]] = icmp ult i64 [[TMP11]], [[A2]]
294 ; LV-NEXT: [[TMP15:%.*]] = select i1 false, i1 [[TMP13]], i1 [[TMP14]]
295 ; LV-NEXT: [[TMP16:%.*]] = or i1 [[TMP15]], [[MUL_OVERFLOW5]]
296 ; LV-NEXT: [[TMP17:%.*]] = or i1 [[TMP10]], [[TMP16]]
297 ; LV-NEXT: br i1 [[TMP17]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]]
298 ; LV: for.body.ph.lver.orig:
299 ; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]]
300 ; LV: for.body.lver.orig:
301 ; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
302 ; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC1_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
303 ; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2
304 ; LV-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = sext i32 [[MUL_LVER_ORIG]] to i64
305 ; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT_LVER_ORIG]]
306 ; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2
307 ; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]]
308 ; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2
309 ; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]]
310 ; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2
311 ; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1
312 ; LV-NEXT: [[INC1_LVER_ORIG]] = add i32 [[IND1_LVER_ORIG]], 1
313 ; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]]
314 ; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]]
316 ; LV-NEXT: br label [[FOR_BODY:%.*]]
318 ; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
319 ; LV-NEXT: [[IND1:%.*]] = phi i32 [ 0, [[FOR_BODY_PH]] ], [ [[INC1:%.*]], [[FOR_BODY]] ]
320 ; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2
321 ; LV-NEXT: [[MUL_EXT:%.*]] = sext i32 [[MUL]] to i64
322 ; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT]]
323 ; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2
324 ; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr i16, i16* [[B]], i64 [[IND]]
325 ; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2
326 ; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]]
327 ; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2
328 ; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1
329 ; LV-NEXT: [[INC1]] = add i32 [[IND1]], 1
330 ; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]]
331 ; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]]
332 ; LV: for.end.loopexit:
333 ; LV-NEXT: br label [[FOR_END:%.*]]
334 ; LV: for.end.loopexit6:
335 ; LV-NEXT: br label [[FOR_END]]
339 i16* noalias %b, i64 %N) {
343 for.body: ; preds = %for.body, %entry
344 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ]
345 %ind1 = phi i32 [ 0, %entry ], [ %inc1, %for.body ]
347 %mul = mul i32 %ind1, 2
348 %mul_ext = sext i32 %mul to i64
350 %arrayidxA = getelementptr i16, i16* %a, i64 %mul_ext
351 %loadA = load i16, i16* %arrayidxA, align 2
353 %arrayidxB = getelementptr i16, i16* %b, i64 %ind
354 %loadB = load i16, i16* %arrayidxB, align 2
356 %add = mul i16 %loadA, %loadB
358 store i16 %add, i16* %arrayidxA, align 2
360 %inc = add nuw nsw i64 %ind, 1
361 %inc1 = add i32 %ind1, 1
363 %exitcond = icmp eq i64 %inc, %N
364 br i1 %exitcond, label %for.end, label %for.body
366 for.end: ; preds = %for.body
370 define void @f4(i16* noalias %a,
372 ; LV-NEXT: for.body.lver.check:
373 ; LV-NEXT: [[A2:%.*]] = ptrtoint i16* [[A:%.*]] to i64
374 ; LV-NEXT: [[TRUNCN:%.*]] = trunc i64 [[N:%.*]] to i32
375 ; LV-NEXT: [[TMP0:%.*]] = add i64 [[N]], -1
376 ; LV-NEXT: [[TMP1:%.*]] = shl i32 [[TRUNCN]], 1
377 ; LV-NEXT: [[TMP2:%.*]] = trunc i64 [[TMP0]] to i32
378 ; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP2]])
379 ; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0
380 ; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1
381 ; LV-NEXT: [[TMP3:%.*]] = add i32 [[TMP1]], [[MUL_RESULT]]
382 ; LV-NEXT: [[TMP4:%.*]] = sub i32 [[TMP1]], [[MUL_RESULT]]
383 ; LV-NEXT: [[TMP5:%.*]] = icmp sgt i32 [[TMP4]], [[TMP1]]
384 ; LV-NEXT: [[TMP6:%.*]] = icmp slt i32 [[TMP3]], [[TMP1]]
385 ; LV-NEXT: [[TMP7:%.*]] = select i1 true, i1 [[TMP5]], i1 [[TMP6]]
386 ; LV-NEXT: [[TMP8:%.*]] = icmp ugt i64 [[TMP0]], 4294967295
387 ; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP7]], [[TMP8]]
388 ; LV-NEXT: [[TMP10:%.*]] = or i1 [[TMP9]], [[MUL_OVERFLOW]]
389 ; LV-NEXT: [[TMP11:%.*]] = or i1 false, [[TMP10]]
390 ; LV-NEXT: [[TMP12:%.*]] = sext i32 [[TMP1]] to i64
391 ; LV-NEXT: [[TMP13:%.*]] = shl nsw i64 [[TMP12]], 1
392 ; LV-NEXT: [[TMP14:%.*]] = add i64 [[A2]], [[TMP13]]
393 ; LV-NEXT: [[MUL3:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]])
394 ; LV-NEXT: [[MUL_RESULT4:%.*]] = extractvalue { i64, i1 } [[MUL3]], 0
395 ; LV-NEXT: [[MUL_OVERFLOW5:%.*]] = extractvalue { i64, i1 } [[MUL3]], 1
396 ; LV-NEXT: [[TMP15:%.*]] = add i64 [[TMP14]], [[MUL_RESULT4]]
397 ; LV-NEXT: [[TMP16:%.*]] = sub i64 [[TMP14]], [[MUL_RESULT4]]
398 ; LV-NEXT: [[TMP17:%.*]] = icmp ugt i64 [[TMP16]], [[TMP14]]
399 ; LV-NEXT: [[TMP18:%.*]] = icmp ult i64 [[TMP15]], [[TMP14]]
400 ; LV-NEXT: [[TMP19:%.*]] = select i1 true, i1 [[TMP17]], i1 [[TMP18]]
401 ; LV-NEXT: [[TMP20:%.*]] = or i1 [[TMP19]], [[MUL_OVERFLOW5]]
402 ; LV-NEXT: [[TMP21:%.*]] = or i1 [[TMP11]], [[TMP20]]
403 ; LV-NEXT: br i1 [[TMP21]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]]
404 ; LV: for.body.ph.lver.orig:
405 ; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]]
406 ; LV: for.body.lver.orig:
407 ; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
408 ; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH_LVER_ORIG]] ], [ [[DEC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
409 ; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2
410 ; LV-NEXT: [[MUL_EXT_LVER_ORIG:%.*]] = sext i32 [[MUL_LVER_ORIG]] to i64
411 ; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT_LVER_ORIG]]
412 ; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2
413 ; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]]
414 ; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2
415 ; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]]
416 ; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2
417 ; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1
418 ; LV-NEXT: [[DEC_LVER_ORIG]] = sub i32 [[IND1_LVER_ORIG]], 1
419 ; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]]
420 ; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]]
422 ; LV-NEXT: br label [[FOR_BODY:%.*]]
424 ; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
425 ; LV-NEXT: [[IND1:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH]] ], [ [[DEC:%.*]], [[FOR_BODY]] ]
426 ; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2
427 ; LV-NEXT: [[MUL_EXT:%.*]] = sext i32 [[MUL]] to i64
428 ; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr i16, i16* [[A]], i64 [[MUL_EXT]]
429 ; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2
430 ; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr i16, i16* [[B]], i64 [[IND]]
431 ; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2
432 ; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]]
433 ; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2
434 ; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1
435 ; LV-NEXT: [[DEC]] = sub i32 [[IND1]], 1
436 ; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]]
437 ; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]]
438 ; LV: for.end.loopexit:
439 ; LV-NEXT: br label [[FOR_END:%.*]]
440 ; LV: for.end.loopexit6:
441 ; LV-NEXT: br label [[FOR_END]]
445 i16* noalias %b, i64 %N) {
447 %TruncN = trunc i64 %N to i32
450 for.body: ; preds = %for.body, %entry
451 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ]
452 %ind1 = phi i32 [ %TruncN, %entry ], [ %dec, %for.body ]
454 %mul = mul i32 %ind1, 2
455 %mul_ext = sext i32 %mul to i64
457 %arrayidxA = getelementptr i16, i16* %a, i64 %mul_ext
458 %loadA = load i16, i16* %arrayidxA, align 2
460 %arrayidxB = getelementptr i16, i16* %b, i64 %ind
461 %loadB = load i16, i16* %arrayidxB, align 2
463 %add = mul i16 %loadA, %loadB
465 store i16 %add, i16* %arrayidxA, align 2
467 %inc = add nuw nsw i64 %ind, 1
468 %dec = sub i32 %ind1, 1
470 %exitcond = icmp eq i64 %inc, %N
471 br i1 %exitcond, label %for.end, label %for.body
473 for.end: ; preds = %for.body
477 ; The following function is similar to the one above, but has the GEP
478 ; to pointer %A inbounds. The index %mul doesn't have the nsw flag.
479 ; This means that the SCEV expression for %mul can wrap and we need
480 ; a SCEV predicate to continue analysis.
482 ; We can still analyze this by adding the required no wrap SCEV predicates.
484 define void @f5(i16* noalias %a,
486 ; LV-NEXT: for.body.lver.check:
487 ; LV-NEXT: [[A2:%.*]] = ptrtoint i16* [[A:%.*]] to i64
488 ; LV-NEXT: [[TRUNCN:%.*]] = trunc i64 [[N:%.*]] to i32
489 ; LV-NEXT: [[TMP0:%.*]] = add i64 [[N]], -1
490 ; LV-NEXT: [[TMP1:%.*]] = shl i32 [[TRUNCN]], 1
491 ; LV-NEXT: [[TMP2:%.*]] = trunc i64 [[TMP0]] to i32
492 ; LV-NEXT: [[MUL1:%.*]] = call { i32, i1 } @llvm.umul.with.overflow.i32(i32 2, i32 [[TMP2]])
493 ; LV-NEXT: [[MUL_RESULT:%.*]] = extractvalue { i32, i1 } [[MUL1]], 0
494 ; LV-NEXT: [[MUL_OVERFLOW:%.*]] = extractvalue { i32, i1 } [[MUL1]], 1
495 ; LV-NEXT: [[TMP3:%.*]] = add i32 [[TMP1]], [[MUL_RESULT]]
496 ; LV-NEXT: [[TMP4:%.*]] = sub i32 [[TMP1]], [[MUL_RESULT]]
497 ; LV-NEXT: [[TMP5:%.*]] = icmp sgt i32 [[TMP4]], [[TMP1]]
498 ; LV-NEXT: [[TMP6:%.*]] = icmp slt i32 [[TMP3]], [[TMP1]]
499 ; LV-NEXT: [[TMP7:%.*]] = select i1 true, i1 [[TMP5]], i1 [[TMP6]]
500 ; LV-NEXT: [[TMP8:%.*]] = icmp ugt i64 [[TMP0]], 4294967295
501 ; LV-NEXT: [[TMP9:%.*]] = or i1 [[TMP7]], [[TMP8]]
502 ; LV-NEXT: [[TMP10:%.*]] = or i1 [[TMP9]], [[MUL_OVERFLOW]]
503 ; LV-NEXT: [[TMP11:%.*]] = or i1 false, [[TMP10]]
504 ; LV-NEXT: [[TMP12:%.*]] = sext i32 [[TMP1]] to i64
505 ; LV-NEXT: [[TMP13:%.*]] = shl nsw i64 [[TMP12]], 1
506 ; LV-NEXT: [[TMP14:%.*]] = add i64 [[A2]], [[TMP13]]
507 ; LV-NEXT: [[MUL3:%.*]] = call { i64, i1 } @llvm.umul.with.overflow.i64(i64 4, i64 [[TMP0]])
508 ; LV-NEXT: [[MUL_RESULT4:%.*]] = extractvalue { i64, i1 } [[MUL3]], 0
509 ; LV-NEXT: [[MUL_OVERFLOW5:%.*]] = extractvalue { i64, i1 } [[MUL3]], 1
510 ; LV-NEXT: [[TMP15:%.*]] = add i64 [[TMP14]], [[MUL_RESULT4]]
511 ; LV-NEXT: [[TMP16:%.*]] = sub i64 [[TMP14]], [[MUL_RESULT4]]
512 ; LV-NEXT: [[TMP17:%.*]] = icmp ugt i64 [[TMP16]], [[TMP14]]
513 ; LV-NEXT: [[TMP18:%.*]] = icmp ult i64 [[TMP15]], [[TMP14]]
514 ; LV-NEXT: [[TMP19:%.*]] = select i1 true, i1 [[TMP17]], i1 [[TMP18]]
515 ; LV-NEXT: [[TMP20:%.*]] = or i1 [[TMP19]], [[MUL_OVERFLOW5]]
516 ; LV-NEXT: [[TMP21:%.*]] = or i1 [[TMP11]], [[TMP20]]
517 ; LV-NEXT: br i1 [[TMP21]], label [[FOR_BODY_PH_LVER_ORIG:%.*]], label [[FOR_BODY_PH:%.*]]
518 ; LV: for.body.ph.lver.orig:
519 ; LV-NEXT: br label [[FOR_BODY_LVER_ORIG:%.*]]
520 ; LV: for.body.lver.orig:
521 ; LV-NEXT: [[IND_LVER_ORIG:%.*]] = phi i64 [ 0, [[FOR_BODY_PH_LVER_ORIG]] ], [ [[INC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
522 ; LV-NEXT: [[IND1_LVER_ORIG:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH_LVER_ORIG]] ], [ [[DEC_LVER_ORIG:%.*]], [[FOR_BODY_LVER_ORIG]] ]
523 ; LV-NEXT: [[MUL_LVER_ORIG:%.*]] = mul i32 [[IND1_LVER_ORIG]], 2
524 ; LV-NEXT: [[ARRAYIDXA_LVER_ORIG:%.*]] = getelementptr inbounds i16, i16* [[A]], i32 [[MUL_LVER_ORIG]]
525 ; LV-NEXT: [[LOADA_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXA_LVER_ORIG]], align 2
526 ; LV-NEXT: [[ARRAYIDXB_LVER_ORIG:%.*]] = getelementptr inbounds i16, i16* [[B:%.*]], i64 [[IND_LVER_ORIG]]
527 ; LV-NEXT: [[LOADB_LVER_ORIG:%.*]] = load i16, i16* [[ARRAYIDXB_LVER_ORIG]], align 2
528 ; LV-NEXT: [[ADD_LVER_ORIG:%.*]] = mul i16 [[LOADA_LVER_ORIG]], [[LOADB_LVER_ORIG]]
529 ; LV-NEXT: store i16 [[ADD_LVER_ORIG]], i16* [[ARRAYIDXA_LVER_ORIG]], align 2
530 ; LV-NEXT: [[INC_LVER_ORIG]] = add nuw nsw i64 [[IND_LVER_ORIG]], 1
531 ; LV-NEXT: [[DEC_LVER_ORIG]] = sub i32 [[IND1_LVER_ORIG]], 1
532 ; LV-NEXT: [[EXITCOND_LVER_ORIG:%.*]] = icmp eq i64 [[INC_LVER_ORIG]], [[N]]
533 ; LV-NEXT: br i1 [[EXITCOND_LVER_ORIG]], label [[FOR_END_LOOPEXIT:%.*]], label [[FOR_BODY_LVER_ORIG]]
535 ; LV-NEXT: br label [[FOR_BODY:%.*]]
537 ; LV-NEXT: [[IND:%.*]] = phi i64 [ 0, [[FOR_BODY_PH]] ], [ [[INC:%.*]], [[FOR_BODY]] ]
538 ; LV-NEXT: [[IND1:%.*]] = phi i32 [ [[TRUNCN]], [[FOR_BODY_PH]] ], [ [[DEC:%.*]], [[FOR_BODY]] ]
539 ; LV-NEXT: [[MUL:%.*]] = mul i32 [[IND1]], 2
540 ; LV-NEXT: [[ARRAYIDXA:%.*]] = getelementptr inbounds i16, i16* [[A]], i32 [[MUL]]
541 ; LV-NEXT: [[LOADA:%.*]] = load i16, i16* [[ARRAYIDXA]], align 2
542 ; LV-NEXT: [[ARRAYIDXB:%.*]] = getelementptr inbounds i16, i16* [[B]], i64 [[IND]]
543 ; LV-NEXT: [[LOADB:%.*]] = load i16, i16* [[ARRAYIDXB]], align 2
544 ; LV-NEXT: [[ADD:%.*]] = mul i16 [[LOADA]], [[LOADB]]
545 ; LV-NEXT: store i16 [[ADD]], i16* [[ARRAYIDXA]], align 2
546 ; LV-NEXT: [[INC]] = add nuw nsw i64 [[IND]], 1
547 ; LV-NEXT: [[DEC]] = sub i32 [[IND1]], 1
548 ; LV-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INC]], [[N]]
549 ; LV-NEXT: br i1 [[EXITCOND]], label [[FOR_END_LOOPEXIT6:%.*]], label [[FOR_BODY]]
550 ; LV: for.end.loopexit:
551 ; LV-NEXT: br label [[FOR_END:%.*]]
552 ; LV: for.end.loopexit6:
553 ; LV-NEXT: br label [[FOR_END]]
557 i16* noalias %b, i64 %N) {
559 %TruncN = trunc i64 %N to i32
562 for.body: ; preds = %for.body, %entry
563 %ind = phi i64 [ 0, %entry ], [ %inc, %for.body ]
564 %ind1 = phi i32 [ %TruncN, %entry ], [ %dec, %for.body ]
566 %mul = mul i32 %ind1, 2
568 %arrayidxA = getelementptr inbounds i16, i16* %a, i32 %mul
569 %loadA = load i16, i16* %arrayidxA, align 2
571 %arrayidxB = getelementptr inbounds i16, i16* %b, i64 %ind
572 %loadB = load i16, i16* %arrayidxB, align 2
574 %add = mul i16 %loadA, %loadB
576 store i16 %add, i16* %arrayidxA, align 2
578 %inc = add nuw nsw i64 %ind, 1
579 %dec = sub i32 %ind1, 1
581 %exitcond = icmp eq i64 %inc, %N
582 br i1 %exitcond, label %for.end, label %for.body
584 for.end: ; preds = %for.body