|
| 1 | +; RUN: opt < %s -passes=loop-interchange -pass-remarks-output=%t -disable-output |
| 2 | +; RUN: FileCheck -input-file %t %s |
| 3 | + |
| 4 | +; This is a collection of cases where the loops were incorrectly interchanged |
| 5 | +; due to mishandling of scalar dependencies. |
| 6 | + |
| 7 | +;; The original code is as follows, with some simplifications from the one in |
| 8 | +;; the issue #46867. The interchange is illegal due to the dependency of `s`. |
| 9 | +;; |
| 10 | +;; void issue46867(int *s, int c, int ff[4][9]) { |
| 11 | +;; for (int d = 0; d <= 2; d++) { |
| 12 | +;; for (int e = 0; e <= 2; e++) { |
| 13 | +;; if ((long)(ff[e][d] && (*s = 3), c) % 4073709551606) |
| 14 | +;; ++*s; |
| 15 | +;; } |
| 16 | +;; } |
| 17 | +;; } |
| 18 | + |
| 19 | +; CHECK: --- !Missed |
| 20 | +; CHECK-NEXT: Pass: loop-interchange |
| 21 | +; CHECK-NEXT: Name: Dependence |
| 22 | +; CHECK-NEXT: Function: issue46867 |
| 23 | +; CHECK-NEXT: Args: |
| 24 | +; CHECK-NEXT: - String: Cannot interchange loops due to dependences. |
| 25 | +; CHECK: --- !Missed |
| 26 | +; CHECK-NEXT: Pass: loop-interchange |
| 27 | +; CHECK-NEXT: Name: Dependence |
| 28 | +; CHECK-NEXT: Function: issue46867 |
| 29 | +; CHECK-NEXT: Args: |
| 30 | +; CHECK-NEXT: - String: Cannot interchange loops due to dependences. |
| 31 | +define void @issue46867(ptr noundef captures(none) %s, i32 noundef %c, ptr noundef readonly captures(none) %ff) { |
| 32 | +entry: |
| 33 | + %tobool7.not = icmp eq i32 %c, 0 |
| 34 | + br i1 %tobool7.not, label %for.cond1.preheader.us.preheader, label %entry.split |
| 35 | + |
| 36 | +for.cond1.preheader.us.preheader: |
| 37 | + br label %for.cond1.preheader.us |
| 38 | + |
| 39 | +for.cond1.preheader.us: |
| 40 | + %indvars.iv31 = phi i64 [ 0, %for.cond1.preheader.us.preheader ], [ %indvars.iv.next32, %for.cond.cleanup3.split.us.us ] |
| 41 | + br label %for.body4.us.us |
| 42 | + |
| 43 | +for.body4.us.us: |
| 44 | + %indvars.iv27 = phi i64 [ %indvars.iv.next28, %land.end.us.us ], [ 0, %for.cond1.preheader.us ] |
| 45 | + %arrayidx6.us.us = getelementptr inbounds nuw [9 x i32], ptr %ff, i64 %indvars.iv27, i64 %indvars.iv31 |
| 46 | + %0 = load i32, ptr %arrayidx6.us.us, align 4 |
| 47 | + %tobool.not.us.us = icmp eq i32 %0, 0 |
| 48 | + br i1 %tobool.not.us.us, label %land.end.us.us, label %land.rhs.us.us |
| 49 | + |
| 50 | +land.rhs.us.us: |
| 51 | + store i32 3, ptr %s, align 4 |
| 52 | + br label %land.end.us.us |
| 53 | + |
| 54 | +land.end.us.us: |
| 55 | + %indvars.iv.next28 = add nuw nsw i64 %indvars.iv27, 1 |
| 56 | + %exitcond30 = icmp ne i64 %indvars.iv.next28, 3 |
| 57 | + br i1 %exitcond30, label %for.body4.us.us, label %for.cond.cleanup3.split.us.us |
| 58 | + |
| 59 | +for.cond.cleanup3.split.us.us: |
| 60 | + %indvars.iv.next32 = add nuw nsw i64 %indvars.iv31, 1 |
| 61 | + %exitcond34 = icmp ne i64 %indvars.iv.next32, 3 |
| 62 | + br i1 %exitcond34, label %for.cond1.preheader.us, label %for.cond.cleanup.loopexit |
| 63 | + |
| 64 | +entry.split: |
| 65 | + %s.promoted19 = load i32, ptr %s, align 4 |
| 66 | + br label %for.cond1.preheader |
| 67 | + |
| 68 | +for.cond1.preheader: |
| 69 | + %indvars.iv23 = phi i64 [ 0, %entry.split ], [ %indvars.iv.next24, %for.cond.cleanup3.split ] |
| 70 | + %s.promoted20 = phi i32 [ %s.promoted19, %entry.split ], [ %inc.lcssa, %for.cond.cleanup3.split ] |
| 71 | + br label %for.body4 |
| 72 | + |
| 73 | +for.cond.cleanup.loopexit: |
| 74 | + br label %for.cond.cleanup |
| 75 | + |
| 76 | +for.cond.cleanup.loopexit21: |
| 77 | + br label %for.cond.cleanup |
| 78 | + |
| 79 | +for.cond.cleanup: |
| 80 | + ret void |
| 81 | + |
| 82 | +for.cond.cleanup3.split: |
| 83 | + %inc.lcssa = phi i32 [ %inc, %land.end ] |
| 84 | + %indvars.iv.next24 = add nuw nsw i64 %indvars.iv23, 1 |
| 85 | + %exitcond26 = icmp ne i64 %indvars.iv.next24, 3 |
| 86 | + br i1 %exitcond26, label %for.cond1.preheader, label %for.cond.cleanup.loopexit21 |
| 87 | + |
| 88 | +for.body4: |
| 89 | + %indvars.iv = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next, %land.end ] |
| 90 | + %1 = phi i32 [ %s.promoted20, %for.cond1.preheader ], [ %inc, %land.end ] |
| 91 | + %arrayidx6 = getelementptr inbounds nuw [9 x i32], ptr %ff, i64 %indvars.iv, i64 %indvars.iv23 |
| 92 | + %2 = load i32, ptr %arrayidx6, align 4 |
| 93 | + %tobool.not = icmp eq i32 %2, 0 |
| 94 | + br i1 %tobool.not, label %land.end, label %land.rhs |
| 95 | + |
| 96 | +land.rhs: |
| 97 | + store i32 3, ptr %s, align 4 |
| 98 | + br label %land.end |
| 99 | + |
| 100 | +land.end: |
| 101 | + %3 = phi i32 [ 3, %land.rhs ], [ %1, %for.body4 ] |
| 102 | + %inc = add nsw i32 %3, 1 |
| 103 | + store i32 %inc, ptr %s, align 4 |
| 104 | + %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 |
| 105 | + %exitcond = icmp ne i64 %indvars.iv.next, 3 |
| 106 | + br i1 %exitcond, label %for.body4, label %for.cond.cleanup3.split |
| 107 | +} |
| 108 | + |
| 109 | + |
| 110 | +;; The original code is as follows, with some simplifications from the one in |
| 111 | +;; the issue #47401. The interchange is illegal due to the dependency of `e`. |
| 112 | +;; |
| 113 | +;; void issue47401(int *e, int bb[][8]) { |
| 114 | +;; for (int c = 0; c <= 7; c++) |
| 115 | +;; for (int d = 4; d; d--) |
| 116 | +;; bb[d + 2][c] && (*e = bb[d][0]); |
| 117 | +;; } |
| 118 | + |
| 119 | +; CHECK: --- !Missed |
| 120 | +; CHECK-NEXT: Pass: loop-interchange |
| 121 | +; CHECK-NEXT: Name: Dependence |
| 122 | +; CHECK-NEXT: Function: issue47401 |
| 123 | +; CHECK-NEXT: Args: |
| 124 | +; CHECK-NEXT: - String: Cannot interchange loops due to dependences. |
| 125 | +define void @issue47401(ptr noundef writeonly captures(none) %e, ptr noundef readonly captures(none) %bb) { |
| 126 | +entry: |
| 127 | + br label %for.cond1.preheader |
| 128 | + |
| 129 | +for.cond1.preheader: |
| 130 | + %indvars.iv22 = phi i64 [ 0, %entry ], [ %indvars.iv.next23, %for.cond.cleanup2 ] |
| 131 | + br label %for.body3 |
| 132 | + |
| 133 | +for.cond.cleanup: |
| 134 | + ret void |
| 135 | + |
| 136 | +for.cond.cleanup2: |
| 137 | + %indvars.iv.next23 = add nuw nsw i64 %indvars.iv22, 1 |
| 138 | + %exitcond = icmp ne i64 %indvars.iv.next23, 8 |
| 139 | + br i1 %exitcond, label %for.cond1.preheader, label %for.cond.cleanup |
| 140 | + |
| 141 | +for.body3: |
| 142 | + %indvars.iv = phi i64 [ 4, %for.cond1.preheader ], [ %indvars.iv.next, %land.end ] |
| 143 | + %0 = getelementptr [8 x i32], ptr %bb, i64 %indvars.iv |
| 144 | + %arrayidx = getelementptr i8, ptr %0, i64 64 |
| 145 | + %arrayidx5 = getelementptr inbounds nuw [8 x i32], ptr %arrayidx, i64 0, i64 %indvars.iv22 |
| 146 | + %1 = load i32, ptr %arrayidx5, align 4 |
| 147 | + %tobool6.not = icmp eq i32 %1, 0 |
| 148 | + br i1 %tobool6.not, label %land.end, label %land.rhs |
| 149 | + |
| 150 | +land.rhs: |
| 151 | + %2 = load i32, ptr %0, align 4 |
| 152 | + store i32 %2, ptr %e, align 4 |
| 153 | + br label %land.end |
| 154 | + |
| 155 | +land.end: |
| 156 | + %indvars.iv.next = add nsw i64 %indvars.iv, -1 |
| 157 | + %tobool.not = icmp eq i64 %indvars.iv.next, 0 |
| 158 | + br i1 %tobool.not, label %for.cond.cleanup2, label %for.body3 |
| 159 | +} |
| 160 | + |
| 161 | +;; The original code is as follows, with some simplifications from the one in |
| 162 | +;; the issue #47295. The interchange is illegal due to the dependency of `f`. |
| 163 | +;; |
| 164 | +;; void issue47295(int *f, int cc[4][4]) { |
| 165 | +;; for (int i = 0; i <= 3; i++) { |
| 166 | +;; for (int j = 0; j <= 3; j++) { |
| 167 | +;; *f ^= 0x1000; |
| 168 | +;; cc[j][i] = *f; |
| 169 | +;; } |
| 170 | +;; } |
| 171 | +;; } |
| 172 | + |
| 173 | +; CHECK: --- !Missed |
| 174 | +; CHECK-NEXT: Pass: loop-interchange |
| 175 | +; CHECK-NEXT: Name: Dependence |
| 176 | +; CHECK-NEXT: Function: issue47295 |
| 177 | +; CHECK-NEXT: Args: |
| 178 | +; CHECK-NEXT: - String: Cannot interchange loops due to dependences. |
| 179 | +define void @issue47295(ptr noundef captures(none) %f, ptr noundef writeonly captures(none) %cc) { |
| 180 | +entry: |
| 181 | + br label %for.cond1.preheader |
| 182 | + |
| 183 | +for.cond1.preheader: |
| 184 | + %indvars.iv18 = phi i64 [ 0, %entry ], [ %indvars.iv.next19, %for.cond.cleanup3 ] |
| 185 | + br label %for.body4 |
| 186 | + |
| 187 | +for.cond.cleanup: |
| 188 | + ret void |
| 189 | + |
| 190 | +for.cond.cleanup3: |
| 191 | + %indvars.iv.next19 = add nuw nsw i64 %indvars.iv18, 1 |
| 192 | + %exitcond21 = icmp ne i64 %indvars.iv.next19, 4 |
| 193 | + br i1 %exitcond21, label %for.cond1.preheader, label %for.cond.cleanup |
| 194 | + |
| 195 | +for.body4: |
| 196 | + %indvars.iv = phi i64 [ 0, %for.cond1.preheader ], [ %indvars.iv.next, %for.body4 ] |
| 197 | + %0 = load i32, ptr %f, align 4 |
| 198 | + %xor = xor i32 %0, 4096 |
| 199 | + store i32 %xor, ptr %f, align 4 |
| 200 | + %arrayidx6 = getelementptr inbounds nuw [4 x i32], ptr %cc, i64 %indvars.iv, i64 %indvars.iv18 |
| 201 | + store i32 %xor, ptr %arrayidx6, align 4 |
| 202 | + %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 |
| 203 | + %exitcond = icmp ne i64 %indvars.iv.next, 4 |
| 204 | + br i1 %exitcond, label %for.body4, label %for.cond.cleanup3 |
| 205 | +} |
| 206 | + |
| 207 | +;; The original code is as follows, with some simplifications from the one in |
| 208 | +;; the issue #54176. The interchange is illegal due to the dependency of `aa`. |
| 209 | +;; |
| 210 | +;; void issue54176(int n, int m, float aa[1024][128], float bb[1024][128], float cc[1024][128]) { |
| 211 | +;; for (int j = 1; j < 128; j++) { |
| 212 | +;; for (int i = 1; i < 1024; i++) { |
| 213 | +;; cc[i][j] = aa[1][j]; |
| 214 | +;; aa[1][j-1] += bb[i][j]; |
| 215 | +;; } |
| 216 | +;; } |
| 217 | +;; } |
| 218 | + |
| 219 | +; CHECK: --- !Missed |
| 220 | +; CHECK-NEXT: Pass: loop-interchange |
| 221 | +; CHECK-NEXT: Name: Dependence |
| 222 | +; CHECK-NEXT: Function: issue54176 |
| 223 | +; CHECK-NEXT: Args: |
| 224 | +; CHECK-NEXT: - String: Cannot interchange loops due to dependences. |
| 225 | +define void @issue54176(i32 noundef %n, i32 noundef %m, ptr noundef captures(none) %aa, ptr noundef readonly captures(none) %bb, ptr noundef writeonly captures(none) %cc) { |
| 226 | + |
| 227 | +entry: |
| 228 | + %arrayidx = getelementptr inbounds nuw i8, ptr %aa, i64 512 |
| 229 | + br label %for.cond1.preheader |
| 230 | + |
| 231 | +for.cond1.preheader: |
| 232 | + %indvars.iv32 = phi i64 [ 1, %entry ], [ %indvars.iv.next33, %for.cond.cleanup3 ] |
| 233 | + %arrayidx5 = getelementptr inbounds nuw [128 x float], ptr %arrayidx, i64 0, i64 %indvars.iv32 |
| 234 | + %0 = add nsw i64 %indvars.iv32, -1 |
| 235 | + %arrayidx16 = getelementptr inbounds [128 x float], ptr %arrayidx, i64 0, i64 %0 |
| 236 | + br label %for.body4 |
| 237 | + |
| 238 | +for.cond.cleanup3: |
| 239 | + %indvars.iv.next33 = add nuw nsw i64 %indvars.iv32, 1 |
| 240 | + %exitcond36 = icmp ne i64 %indvars.iv.next33, 128 |
| 241 | + br i1 %exitcond36, label %for.cond1.preheader, label %for.cond.cleanup |
| 242 | + |
| 243 | +for.body4: |
| 244 | + %indvars.iv = phi i64 [ 1, %for.cond1.preheader ], [ %indvars.iv.next, %for.body4 ] |
| 245 | + %1 = load float, ptr %arrayidx5, align 4 |
| 246 | + %arrayidx9 = getelementptr inbounds nuw [128 x float], ptr %cc, i64 %indvars.iv, i64 %indvars.iv32 |
| 247 | + store float %1, ptr %arrayidx9, align 4 |
| 248 | + %arrayidx13 = getelementptr inbounds nuw [128 x float], ptr %bb, i64 %indvars.iv, i64 %indvars.iv32 |
| 249 | + %2 = load float, ptr %arrayidx13, align 4 |
| 250 | + %3 = load float, ptr %arrayidx16, align 4 |
| 251 | + %add = fadd float %2, %3 |
| 252 | + store float %add, ptr %arrayidx16, align 4 |
| 253 | + %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 |
| 254 | + %exitcond = icmp ne i64 %indvars.iv.next, 1024 |
| 255 | + br i1 %exitcond, label %for.body4, label %for.cond.cleanup3 |
| 256 | + |
| 257 | +for.cond.cleanup: |
| 258 | + ret void |
| 259 | +} |
| 260 | + |
| 261 | +;; The original code is as follows, with some simplifications from the one in |
| 262 | +;; the issue #116114. The interchange is illegal due to the dependency of `A`. |
| 263 | +;; |
| 264 | +;; void issue116114(int *A, int x, unsigned N, unsigned M) { |
| 265 | +;; for (unsigned m = 0; m < M; ++m) |
| 266 | +;; for (unsigned i = 0U; i < N - 1; ++i) { |
| 267 | +;; A[i] = A[i + 1] + x; |
| 268 | +;; } |
| 269 | +;; } |
| 270 | + |
| 271 | +; CHECK: --- !Missed |
| 272 | +; CHECK-NEXT: Pass: loop-interchange |
| 273 | +; CHECK-NEXT: Name: Dependence |
| 274 | +; CHECK-NEXT: Function: issue116114 |
| 275 | +; CHECK-NEXT: Args: |
| 276 | +; CHECK-NEXT: - String: Cannot interchange loops due to dependences. |
| 277 | +define void @issue116114(ptr noundef captures(none) %A, i32 noundef %x, i32 noundef %N, i32 noundef %M) { |
| 278 | +entry: |
| 279 | + %cmp18.not = icmp eq i32 %M, 0 |
| 280 | + br i1 %cmp18.not, label %for.cond.cleanup, label %for.cond1.preheader.lr.ph |
| 281 | + |
| 282 | +for.cond1.preheader.lr.ph: |
| 283 | + %sub = add i32 %N, -1 |
| 284 | + %cmp216.not = icmp eq i32 %sub, 0 |
| 285 | + br i1 %cmp216.not, label %for.cond1.preheader.preheader, label %for.cond1.preheader.us.preheader |
| 286 | + |
| 287 | +for.cond1.preheader.us.preheader: |
| 288 | + br label %for.cond1.preheader.us |
| 289 | + |
| 290 | +for.cond1.preheader.preheader: |
| 291 | + br label %for.cond.cleanup.loopexit |
| 292 | + |
| 293 | +for.cond1.preheader.us: |
| 294 | + %m.019.us = phi i32 [ %inc9.us, %for.cond1.for.cond.cleanup3_crit_edge.us ], [ 0, %for.cond1.preheader.us.preheader ] |
| 295 | + %wide.trip.count = zext i32 %sub to i64 |
| 296 | + br label %for.body4.us |
| 297 | + |
| 298 | +for.body4.us: |
| 299 | + %indvars.iv = phi i64 [ 0, %for.cond1.preheader.us ], [ %indvars.iv.next, %for.body4.us ] |
| 300 | + %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1 |
| 301 | + %arrayidx.us = getelementptr inbounds nuw i32, ptr %A, i64 %indvars.iv.next |
| 302 | + %0 = load i32, ptr %arrayidx.us, align 4 |
| 303 | + %add5.us = add nsw i32 %0, %x |
| 304 | + %arrayidx7.us = getelementptr inbounds nuw i32, ptr %A, i64 %indvars.iv |
| 305 | + store i32 %add5.us, ptr %arrayidx7.us, align 4 |
| 306 | + %exitcond = icmp ne i64 %indvars.iv.next, %wide.trip.count |
| 307 | + br i1 %exitcond, label %for.body4.us, label %for.cond1.for.cond.cleanup3_crit_edge.us |
| 308 | + |
| 309 | +for.cond1.for.cond.cleanup3_crit_edge.us: |
| 310 | + %inc9.us = add nuw i32 %m.019.us, 1 |
| 311 | + %exitcond22 = icmp ne i32 %inc9.us, %M |
| 312 | + br i1 %exitcond22, label %for.cond1.preheader.us, label %for.cond.cleanup.loopexit20 |
| 313 | + |
| 314 | +for.cond.cleanup.loopexit: |
| 315 | + br label %for.cond.cleanup |
| 316 | + |
| 317 | +for.cond.cleanup.loopexit20: |
| 318 | + br label %for.cond.cleanup |
| 319 | + |
| 320 | +for.cond.cleanup: |
| 321 | + ret void |
| 322 | +} |
0 commit comments