Jingyue Wu | 42f1d67 | 2015-07-28 18:22:40 +0000 | [diff] [blame] | 1 | ; RUN: opt < %s -S -analyze -scalar-evolution | FileCheck %s |
| 2 | |
| 3 | ; Positive and negative tests for inferring flags like nsw from |
| 4 | ; reasoning about how a poison value from overflow would trigger |
| 5 | ; undefined behavior. |
| 6 | |
| 7 | define void @foo() { |
| 8 | ret void |
| 9 | } |
| 10 | |
| 11 | ; Example where an add should get the nsw flag, so that a sext can be |
| 12 | ; distributed over the add. |
| 13 | define void @test-add-nsw(float* %input, i32 %offset, i32 %numIterations) { |
| 14 | ; CHECK-LABEL: @test-add-nsw |
| 15 | entry: |
| 16 | br label %loop |
| 17 | loop: |
| 18 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 19 | |
| 20 | ; CHECK: %index32 = |
| 21 | ; CHECK: --> {%offset,+,1}<nsw> |
| 22 | %index32 = add nsw i32 %i, %offset |
| 23 | |
| 24 | ; CHECK: %index64 = |
| 25 | ; CHECK: --> {(sext i32 %offset to i64),+,1}<nsw> |
| 26 | %index64 = sext i32 %index32 to i64 |
| 27 | |
| 28 | %ptr = getelementptr inbounds float, float* %input, i64 %index64 |
| 29 | %nexti = add nsw i32 %i, 1 |
| 30 | %f = load float, float* %ptr, align 4 |
| 31 | call void @foo() |
| 32 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 33 | br i1 %exitcond, label %exit, label %loop |
| 34 | exit: |
| 35 | ret void |
| 36 | } |
| 37 | |
| 38 | ; Example where an add should get the nuw flag. |
| 39 | define void @test-add-nuw(float* %input, i32 %offset, i32 %numIterations) { |
| 40 | ; CHECK-LABEL: @test-add-nuw |
| 41 | entry: |
| 42 | br label %loop |
| 43 | loop: |
| 44 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 45 | |
| 46 | ; CHECK: %index32 = |
| 47 | ; CHECK: --> {%offset,+,1}<nuw> |
| 48 | %index32 = add nuw i32 %i, %offset |
| 49 | |
| 50 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 51 | %nexti = add nuw i32 %i, 1 |
| 52 | %f = load float, float* %ptr, align 4 |
| 53 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 54 | br i1 %exitcond, label %exit, label %loop |
| 55 | |
| 56 | exit: |
| 57 | ret void |
| 58 | } |
| 59 | |
Sanjoy Das | 70c2bbd | 2016-05-29 00:31:18 +0000 | [diff] [blame] | 60 | define void @test-add-nuw-from-icmp(float* %input, i32 %offset, |
| 61 | i32 %numIterations) { |
| 62 | ; CHECK-LABEL: @test-add-nuw-from-icmp |
| 63 | entry: |
| 64 | br label %loop |
| 65 | loop: |
| 66 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 67 | |
| 68 | ; CHECK: %index32 = |
| 69 | ; CHECK: --> {%offset,+,1}<nuw> |
| 70 | %index32 = add nuw i32 %i, %offset |
| 71 | %cmp = icmp sgt i32 %index32, 0 |
| 72 | %cmp.idx = sext i1 %cmp to i32 |
| 73 | |
| 74 | %ptr = getelementptr inbounds float, float* %input, i32 %cmp.idx |
| 75 | %nexti = add nuw i32 %i, 1 |
| 76 | %f = load float, float* %ptr, align 4 |
| 77 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 78 | br i1 %exitcond, label %exit, label %loop |
| 79 | |
| 80 | exit: |
| 81 | ret void |
| 82 | } |
| 83 | |
Jingyue Wu | 42f1d67 | 2015-07-28 18:22:40 +0000 | [diff] [blame] | 84 | ; With no load to trigger UB from poison, we cannot infer nsw. |
| 85 | define void @test-add-no-load(float* %input, i32 %offset, i32 %numIterations) { |
| 86 | ; CHECK-LABEL: @test-add-no-load |
| 87 | entry: |
| 88 | br label %loop |
| 89 | loop: |
| 90 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 91 | |
| 92 | ; CHECK: %index32 = |
| 93 | ; CHECK: --> {%offset,+,1}<nw> |
| 94 | %index32 = add nsw i32 %i, %offset |
| 95 | |
| 96 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 97 | %nexti = add nuw i32 %i, 1 |
| 98 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 99 | br i1 %exitcond, label %exit, label %loop |
| 100 | |
| 101 | exit: |
| 102 | ret void |
| 103 | } |
| 104 | |
| 105 | ; The current code is only supposed to look at the loop header, so |
| 106 | ; it should not infer nsw in this case, as that would require looking |
| 107 | ; outside the loop header. |
| 108 | define void @test-add-not-header(float* %input, i32 %offset, i32 %numIterations) { |
| 109 | ; CHECK-LABEL: @test-add-not-header |
| 110 | entry: |
| 111 | br label %loop |
| 112 | loop: |
| 113 | %i = phi i32 [ %nexti, %loop2 ], [ 0, %entry ] |
| 114 | br label %loop2 |
| 115 | loop2: |
| 116 | |
| 117 | ; CHECK: %index32 = |
| 118 | ; CHECK: --> {%offset,+,1}<nw> |
| 119 | %index32 = add nsw i32 %i, %offset |
| 120 | |
| 121 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 122 | %nexti = add nsw i32 %i, 1 |
| 123 | %f = load float, float* %ptr, align 4 |
| 124 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 125 | br i1 %exitcond, label %exit, label %loop |
| 126 | exit: |
| 127 | ret void |
| 128 | } |
| 129 | |
| 130 | ; Same thing as test-add-not-header, but in this case only the load |
| 131 | ; instruction is outside the loop header. |
| 132 | define void @test-add-not-header2(float* %input, i32 %offset, i32 %numIterations) { |
| 133 | ; CHECK-LABEL: @test-add-not-header2 |
| 134 | entry: |
| 135 | br label %loop |
| 136 | loop: |
| 137 | %i = phi i32 [ %nexti, %loop2 ], [ 0, %entry ] |
| 138 | |
| 139 | ; CHECK: %index32 = |
Sanjoy Das | a6155b6 | 2016-04-22 17:41:06 +0000 | [diff] [blame] | 140 | ; CHECK: --> {%offset,+,1}<nsw> |
Jingyue Wu | 42f1d67 | 2015-07-28 18:22:40 +0000 | [diff] [blame] | 141 | %index32 = add nsw i32 %i, %offset |
| 142 | |
| 143 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 144 | %nexti = add nsw i32 %i, 1 |
| 145 | br label %loop2 |
| 146 | loop2: |
| 147 | %f = load float, float* %ptr, align 4 |
| 148 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 149 | br i1 %exitcond, label %exit, label %loop |
| 150 | exit: |
| 151 | ret void |
| 152 | } |
| 153 | |
Sanjoy Das | a6155b6 | 2016-04-22 17:41:06 +0000 | [diff] [blame] | 154 | ; Similar to test-add-not-header, but in this case the load |
| 155 | ; instruction may not be executed. |
| 156 | define void @test-add-not-header3(float* %input, i32 %offset, i32 %numIterations, |
| 157 | i1* %cond_buf) { |
| 158 | ; CHECK-LABEL: @test-add-not-header3 |
| 159 | entry: |
| 160 | br label %loop |
| 161 | loop: |
| 162 | %i = phi i32 [ %nexti, %loop2 ], [ 0, %entry ] |
| 163 | |
| 164 | ; CHECK: %index32 = |
| 165 | ; CHECK: --> {%offset,+,1}<nw> |
| 166 | %index32 = add nsw i32 %i, %offset |
| 167 | |
| 168 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 169 | %nexti = add nsw i32 %i, 1 |
| 170 | %cond = load volatile i1, i1* %cond_buf |
| 171 | br i1 %cond, label %loop2, label %exit |
| 172 | loop2: |
| 173 | %f = load float, float* %ptr, align 4 |
| 174 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 175 | br i1 %exitcond, label %exit, label %loop |
| 176 | exit: |
| 177 | ret void |
| 178 | } |
| 179 | |
| 180 | ; Same thing as test-add-not-header2, except we have a few extra |
| 181 | ; blocks. |
| 182 | define void @test-add-not-header4(float* %input, i32 %offset, i32 %numIterations) { |
| 183 | ; CHECK-LABEL: @test-add-not-header4 |
| 184 | entry: |
| 185 | br label %loop |
| 186 | loop: |
| 187 | %i = phi i32 [ %nexti, %loop2 ], [ 0, %entry ] |
| 188 | |
| 189 | ; CHECK: %index32 = |
| 190 | ; CHECK: --> {%offset,+,1}<nsw> |
| 191 | %index32 = add nsw i32 %i, %offset |
| 192 | |
| 193 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 194 | %nexti = add nsw i32 %i, 1 |
| 195 | br label %loop3 |
| 196 | loop3: |
| 197 | br label %loop4 |
| 198 | loop4: |
| 199 | br label %loop2 |
| 200 | loop2: |
| 201 | %f = load float, float* %ptr, align 4 |
| 202 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 203 | br i1 %exitcond, label %exit, label %loop |
| 204 | exit: |
| 205 | ret void |
| 206 | } |
| 207 | |
Sanjoy Das | 08989c7 | 2017-04-30 19:41:19 +0000 | [diff] [blame] | 208 | ; Demonstrate why we need a Visited set in llvm::programUndefinedIfFullPoison. |
Sanjoy Das | a6155b6 | 2016-04-22 17:41:06 +0000 | [diff] [blame] | 209 | define void @test-add-not-header5(float* %input, i32 %offset) { |
| 210 | ; CHECK-LABEL: @test-add-not-header5 |
| 211 | entry: |
| 212 | br label %loop |
| 213 | loop: |
| 214 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 215 | |
| 216 | ; CHECK: %index32 = |
| 217 | ; CHECK: --> {%offset,+,1}<nw> |
| 218 | %index32 = add nsw i32 %i, %offset |
| 219 | |
| 220 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 221 | %nexti = add nsw i32 %i, 1 |
| 222 | br label %loop |
| 223 | |
| 224 | exit: |
| 225 | ret void |
| 226 | } |
| 227 | |
Jingyue Wu | 42f1d67 | 2015-07-28 18:22:40 +0000 | [diff] [blame] | 228 | ; The call instruction makes it not guaranteed that the add will be |
| 229 | ; executed, since it could run forever or throw an exception, so we |
| 230 | ; cannot assume that the UB is realized. |
| 231 | define void @test-add-call(float* %input, i32 %offset, i32 %numIterations) { |
| 232 | ; CHECK-LABEL: @test-add-call |
| 233 | entry: |
| 234 | br label %loop |
| 235 | loop: |
| 236 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 237 | |
| 238 | ; CHECK: %index32 = |
| 239 | ; CHECK: --> {%offset,+,1}<nw> |
| 240 | call void @foo() |
| 241 | %index32 = add nsw i32 %i, %offset |
| 242 | |
| 243 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 244 | %nexti = add nsw i32 %i, 1 |
| 245 | %f = load float, float* %ptr, align 4 |
| 246 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 247 | br i1 %exitcond, label %exit, label %loop |
| 248 | exit: |
| 249 | ret void |
| 250 | } |
| 251 | |
| 252 | ; Same issue as test-add-call, but this time the call is between the |
| 253 | ; producer of poison and the load that consumes it. |
| 254 | define void @test-add-call2(float* %input, i32 %offset, i32 %numIterations) { |
| 255 | ; CHECK-LABEL: @test-add-call2 |
| 256 | entry: |
| 257 | br label %loop |
| 258 | loop: |
| 259 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 260 | |
| 261 | ; CHECK: %index32 = |
| 262 | ; CHECK: --> {%offset,+,1}<nw> |
| 263 | %index32 = add nsw i32 %i, %offset |
| 264 | |
| 265 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 266 | %nexti = add nsw i32 %i, 1 |
| 267 | call void @foo() |
| 268 | %f = load float, float* %ptr, align 4 |
| 269 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 270 | br i1 %exitcond, label %exit, label %loop |
| 271 | exit: |
| 272 | ret void |
| 273 | } |
| 274 | |
Sanjoy Das | 5cd6c5ca | 2017-02-22 06:52:32 +0000 | [diff] [blame] | 275 | ; Any poison input makes getelementptr produce poison |
| 276 | define void @test-gep-propagates-poison(float* %input, i32 %offset, i32 %numIterations) { |
| 277 | ; CHECK-LABEL: @test-gep-propagates-poison |
Jingyue Wu | 42f1d67 | 2015-07-28 18:22:40 +0000 | [diff] [blame] | 278 | entry: |
| 279 | br label %loop |
| 280 | loop: |
| 281 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 282 | |
| 283 | ; CHECK: %index32 = |
Sanjoy Das | 5cd6c5ca | 2017-02-22 06:52:32 +0000 | [diff] [blame] | 284 | ; CHECK: --> {%offset,+,1}<nsw> |
Jingyue Wu | 42f1d67 | 2015-07-28 18:22:40 +0000 | [diff] [blame] | 285 | %index32 = add nsw i32 %i, %offset |
| 286 | |
| 287 | %ptr = getelementptr float, float* %input, i32 %index32 |
| 288 | %nexti = add nsw i32 %i, 1 |
| 289 | %f = load float, float* %ptr, align 4 |
| 290 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 291 | br i1 %exitcond, label %exit, label %loop |
| 292 | exit: |
| 293 | ret void |
| 294 | } |
| 295 | |
| 296 | ; Multiplication by a non-zero constant propagates poison if there is |
| 297 | ; a nuw or nsw flag on the multiplication. |
| 298 | define void @test-add-mul-propagates(float* %input, i32 %offset, i32 %numIterations) { |
| 299 | ; CHECK-LABEL: @test-add-mul-propagates |
| 300 | entry: |
| 301 | br label %loop |
| 302 | loop: |
| 303 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 304 | |
| 305 | ; CHECK: %index32 = |
| 306 | ; CHECK: --> {%offset,+,1}<nsw> |
| 307 | %index32 = add nsw i32 %i, %offset |
| 308 | |
| 309 | %indexmul = mul nuw i32 %index32, 2 |
| 310 | %ptr = getelementptr inbounds float, float* %input, i32 %indexmul |
| 311 | %nexti = add nsw i32 %i, 1 |
| 312 | %f = load float, float* %ptr, align 4 |
| 313 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 314 | br i1 %exitcond, label %exit, label %loop |
| 315 | exit: |
| 316 | ret void |
| 317 | } |
| 318 | |
Sanjoy Das | 5cd6c5ca | 2017-02-22 06:52:32 +0000 | [diff] [blame] | 319 | ; Any poison input to multiplication propages poison. |
| 320 | define void @test-mul-propagates-poison(float* %input, i32 %offset, i32 %numIterations) { |
| 321 | ; CHECK-LABEL: @test-mul-propagates-poison |
Jingyue Wu | 42f1d67 | 2015-07-28 18:22:40 +0000 | [diff] [blame] | 322 | entry: |
| 323 | br label %loop |
| 324 | loop: |
| 325 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 326 | |
| 327 | ; CHECK: %index32 = |
Sanjoy Das | 5cd6c5ca | 2017-02-22 06:52:32 +0000 | [diff] [blame] | 328 | ; CHECK: --> {%offset,+,1}<nsw> |
Jingyue Wu | 42f1d67 | 2015-07-28 18:22:40 +0000 | [diff] [blame] | 329 | %index32 = add nsw i32 %i, %offset |
| 330 | |
| 331 | %indexmul = mul nsw i32 %index32, %offset |
| 332 | %ptr = getelementptr inbounds float, float* %input, i32 %indexmul |
| 333 | %nexti = add nsw i32 %i, 1 |
| 334 | %f = load float, float* %ptr, align 4 |
| 335 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 336 | br i1 %exitcond, label %exit, label %loop |
| 337 | exit: |
| 338 | ret void |
| 339 | } |
| 340 | |
Sanjoy Das | 5cd6c5ca | 2017-02-22 06:52:32 +0000 | [diff] [blame] | 341 | define void @test-mul-propagates-poison-2(float* %input, i32 %offset, i32 %numIterations) { |
| 342 | ; CHECK-LABEL: @test-mul-propagates-poison-2 |
Jingyue Wu | 42f1d67 | 2015-07-28 18:22:40 +0000 | [diff] [blame] | 343 | entry: |
| 344 | br label %loop |
| 345 | loop: |
| 346 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 347 | |
| 348 | ; CHECK: %index32 = |
Sanjoy Das | 5cd6c5ca | 2017-02-22 06:52:32 +0000 | [diff] [blame] | 349 | ; CHECK: --> {%offset,+,1}<nsw> |
Jingyue Wu | 42f1d67 | 2015-07-28 18:22:40 +0000 | [diff] [blame] | 350 | %index32 = add nsw i32 %i, %offset |
| 351 | |
| 352 | %indexmul = mul i32 %index32, 2 |
| 353 | %ptr = getelementptr inbounds float, float* %input, i32 %indexmul |
| 354 | %nexti = add nsw i32 %i, 1 |
| 355 | %f = load float, float* %ptr, align 4 |
| 356 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 357 | br i1 %exitcond, label %exit, label %loop |
| 358 | exit: |
| 359 | ret void |
| 360 | } |
| 361 | |
| 362 | ; Division by poison triggers UB. |
| 363 | define void @test-add-div(float* %input, i32 %offset, i32 %numIterations) { |
| 364 | ; CHECK-LABEL: @test-add-div |
| 365 | entry: |
| 366 | br label %loop |
| 367 | loop: |
| 368 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 369 | |
| 370 | ; CHECK: %j = |
| 371 | ; CHECK: --> {%offset,+,1}<nsw> |
| 372 | %j = add nsw i32 %i, %offset |
| 373 | |
| 374 | %q = sdiv i32 %numIterations, %j |
| 375 | %nexti = add nsw i32 %i, 1 |
| 376 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 377 | br i1 %exitcond, label %exit, label %loop |
| 378 | exit: |
| 379 | ret void |
| 380 | } |
| 381 | |
| 382 | ; Remainder of poison by non-poison divisor does not trigger UB. |
| 383 | define void @test-add-div2(float* %input, i32 %offset, i32 %numIterations) { |
| 384 | ; CHECK-LABEL: @test-add-div2 |
| 385 | entry: |
| 386 | br label %loop |
| 387 | loop: |
| 388 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 389 | |
| 390 | ; CHECK: %j = |
| 391 | ; CHECK: --> {%offset,+,1}<nw> |
| 392 | %j = add nsw i32 %i, %offset |
| 393 | |
| 394 | %q = sdiv i32 %j, %numIterations |
| 395 | %nexti = add nsw i32 %i, 1 |
| 396 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 397 | br i1 %exitcond, label %exit, label %loop |
| 398 | exit: |
| 399 | ret void |
| 400 | } |
| 401 | |
| 402 | ; Store to poison address triggers UB. |
| 403 | define void @test-add-store(float* %input, i32 %offset, i32 %numIterations) { |
| 404 | ; CHECK-LABEL: @test-add-store |
| 405 | entry: |
| 406 | br label %loop |
| 407 | loop: |
| 408 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 409 | |
| 410 | ; CHECK: %index32 = |
| 411 | ; CHECK: --> {%offset,+,1}<nsw> |
| 412 | %index32 = add nsw i32 %i, %offset |
| 413 | |
| 414 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 415 | %nexti = add nsw i32 %i, 1 |
| 416 | store float 1.0, float* %ptr, align 4 |
| 417 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 418 | br i1 %exitcond, label %exit, label %loop |
| 419 | exit: |
| 420 | ret void |
| 421 | } |
| 422 | |
| 423 | ; Three sequential adds where the middle add should have nsw. There is |
| 424 | ; a special case for sequential adds and this test covers that. We have to |
| 425 | ; put the final add first in the program since otherwise the special case |
| 426 | ; is not triggered, hence the strange basic block ordering. |
| 427 | define void @test-add-twice(float* %input, i32 %offset, i32 %numIterations) { |
| 428 | ; CHECK-LABEL: @test-add-twice |
| 429 | entry: |
| 430 | br label %loop |
| 431 | loop2: |
| 432 | ; CHECK: %seq = |
| 433 | ; CHECK: --> {(2 + %offset),+,1}<nw> |
| 434 | %seq = add nsw nuw i32 %index32, 1 |
| 435 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 436 | br i1 %exitcond, label %exit, label %loop |
| 437 | |
| 438 | loop: |
| 439 | %i = phi i32 [ %nexti, %loop2 ], [ 0, %entry ] |
| 440 | |
| 441 | %j = add nsw i32 %i, 1 |
| 442 | ; CHECK: %index32 = |
Oleg Ranevskyy | eb4ecca | 2016-05-25 13:01:33 +0000 | [diff] [blame] | 443 | ; CHECK: --> {(1 + %offset)<nsw>,+,1}<nsw> |
Jingyue Wu | 42f1d67 | 2015-07-28 18:22:40 +0000 | [diff] [blame] | 444 | %index32 = add nsw i32 %j, %offset |
| 445 | |
| 446 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 447 | %nexti = add nsw i32 %i, 1 |
| 448 | store float 1.0, float* %ptr, align 4 |
| 449 | br label %loop2 |
| 450 | exit: |
| 451 | ret void |
| 452 | } |
Bjarke Hammersholt Roune | 9791ed4 | 2015-08-14 22:45:26 +0000 | [diff] [blame] | 453 | |
| 454 | ; Example where a mul should get the nsw flag, so that a sext can be |
| 455 | ; distributed over the mul. |
| 456 | define void @test-mul-nsw(float* %input, i32 %stride, i32 %numIterations) { |
| 457 | ; CHECK-LABEL: @test-mul-nsw |
| 458 | entry: |
| 459 | br label %loop |
| 460 | loop: |
| 461 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 462 | |
| 463 | ; CHECK: %index32 = |
| 464 | ; CHECK: --> {0,+,%stride}<nsw> |
| 465 | %index32 = mul nsw i32 %i, %stride |
| 466 | |
| 467 | ; CHECK: %index64 = |
| 468 | ; CHECK: --> {0,+,(sext i32 %stride to i64)}<nsw> |
| 469 | %index64 = sext i32 %index32 to i64 |
| 470 | |
| 471 | %ptr = getelementptr inbounds float, float* %input, i64 %index64 |
| 472 | %nexti = add nsw i32 %i, 1 |
| 473 | %f = load float, float* %ptr, align 4 |
| 474 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 475 | br i1 %exitcond, label %exit, label %loop |
| 476 | exit: |
| 477 | ret void |
| 478 | } |
| 479 | |
| 480 | ; Example where a mul should get the nuw flag. |
| 481 | define void @test-mul-nuw(float* %input, i32 %stride, i32 %numIterations) { |
| 482 | ; CHECK-LABEL: @test-mul-nuw |
| 483 | entry: |
| 484 | br label %loop |
| 485 | loop: |
| 486 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 487 | |
| 488 | ; CHECK: %index32 = |
| 489 | ; CHECK: --> {0,+,%stride}<nuw> |
| 490 | %index32 = mul nuw i32 %i, %stride |
| 491 | |
| 492 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 493 | %nexti = add nuw i32 %i, 1 |
| 494 | %f = load float, float* %ptr, align 4 |
| 495 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 496 | br i1 %exitcond, label %exit, label %loop |
| 497 | |
| 498 | exit: |
| 499 | ret void |
| 500 | } |
| 501 | |
| 502 | ; Example where a shl should get the nsw flag, so that a sext can be |
| 503 | ; distributed over the shl. |
| 504 | define void @test-shl-nsw(float* %input, i32 %start, i32 %numIterations) { |
| 505 | ; CHECK-LABEL: @test-shl-nsw |
| 506 | entry: |
| 507 | br label %loop |
| 508 | loop: |
| 509 | %i = phi i32 [ %nexti, %loop ], [ %start, %entry ] |
| 510 | |
| 511 | ; CHECK: %index32 = |
| 512 | ; CHECK: --> {(256 * %start),+,256}<nsw> |
| 513 | %index32 = shl nsw i32 %i, 8 |
| 514 | |
| 515 | ; CHECK: %index64 = |
| 516 | ; CHECK: --> {(sext i32 (256 * %start) to i64),+,256}<nsw> |
| 517 | %index64 = sext i32 %index32 to i64 |
| 518 | |
| 519 | %ptr = getelementptr inbounds float, float* %input, i64 %index64 |
| 520 | %nexti = add nsw i32 %i, 1 |
| 521 | %f = load float, float* %ptr, align 4 |
| 522 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 523 | br i1 %exitcond, label %exit, label %loop |
| 524 | exit: |
| 525 | ret void |
| 526 | } |
| 527 | |
| 528 | ; Example where a shl should get the nuw flag. |
| 529 | define void @test-shl-nuw(float* %input, i32 %numIterations) { |
| 530 | ; CHECK-LABEL: @test-shl-nuw |
| 531 | entry: |
| 532 | br label %loop |
| 533 | loop: |
| 534 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 535 | |
| 536 | ; CHECK: %index32 = |
| 537 | ; CHECK: --> {0,+,512}<nuw> |
| 538 | %index32 = shl nuw i32 %i, 9 |
| 539 | |
| 540 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 541 | %nexti = add nuw i32 %i, 1 |
| 542 | %f = load float, float* %ptr, align 4 |
| 543 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 544 | br i1 %exitcond, label %exit, label %loop |
| 545 | |
| 546 | exit: |
| 547 | ret void |
| 548 | } |
| 549 | |
| 550 | ; Example where a sub should *not* get the nsw flag, because of how |
| 551 | ; scalar evolution represents A - B as A + (-B) and -B can wrap even |
| 552 | ; in cases where A - B does not. |
| 553 | define void @test-sub-no-nsw(float* %input, i32 %start, i32 %sub, i32 %numIterations) { |
| 554 | ; CHECK-LABEL: @test-sub-no-nsw |
| 555 | entry: |
| 556 | br label %loop |
| 557 | loop: |
| 558 | %i = phi i32 [ %nexti, %loop ], [ %start, %entry ] |
| 559 | |
| 560 | ; CHECK: %index32 = |
| 561 | ; CHECK: --> {((-1 * %sub) + %start),+,1}<nw> |
| 562 | %index32 = sub nsw i32 %i, %sub |
| 563 | %index64 = sext i32 %index32 to i64 |
| 564 | |
| 565 | %ptr = getelementptr inbounds float, float* %input, i64 %index64 |
| 566 | %nexti = add nsw i32 %i, 1 |
| 567 | %f = load float, float* %ptr, align 4 |
| 568 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 569 | br i1 %exitcond, label %exit, label %loop |
| 570 | exit: |
| 571 | ret void |
| 572 | } |
| 573 | |
| 574 | ; Example where a sub should get the nsw flag as the RHS cannot be the |
| 575 | ; minimal signed value. |
| 576 | define void @test-sub-nsw(float* %input, i32 %start, i32 %sub, i32 %numIterations) { |
| 577 | ; CHECK-LABEL: @test-sub-nsw |
| 578 | entry: |
| 579 | %halfsub = ashr i32 %sub, 1 |
| 580 | br label %loop |
| 581 | loop: |
| 582 | %i = phi i32 [ %nexti, %loop ], [ %start, %entry ] |
| 583 | |
| 584 | ; CHECK: %index32 = |
Oleg Ranevskyy | eb4ecca | 2016-05-25 13:01:33 +0000 | [diff] [blame] | 585 | ; CHECK: --> {((-1 * %halfsub)<nsw> + %start)<nsw>,+,1}<nsw> |
Bjarke Hammersholt Roune | 9791ed4 | 2015-08-14 22:45:26 +0000 | [diff] [blame] | 586 | %index32 = sub nsw i32 %i, %halfsub |
| 587 | %index64 = sext i32 %index32 to i64 |
| 588 | |
| 589 | %ptr = getelementptr inbounds float, float* %input, i64 %index64 |
| 590 | %nexti = add nsw i32 %i, 1 |
| 591 | %f = load float, float* %ptr, align 4 |
| 592 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 593 | br i1 %exitcond, label %exit, label %loop |
| 594 | exit: |
| 595 | ret void |
| 596 | } |
| 597 | |
| 598 | ; Example where a sub should get the nsw flag, since the LHS is non-negative, |
| 599 | ; which implies that the RHS cannot be the minimal signed value. |
| 600 | define void @test-sub-nsw-lhs-non-negative(float* %input, i32 %sub, i32 %numIterations) { |
| 601 | ; CHECK-LABEL: @test-sub-nsw-lhs-non-negative |
| 602 | entry: |
| 603 | br label %loop |
| 604 | loop: |
| 605 | %i = phi i32 [ %nexti, %loop ], [ 0, %entry ] |
| 606 | |
| 607 | ; CHECK: %index32 = |
| 608 | ; CHECK: --> {(-1 * %sub),+,1}<nsw> |
| 609 | %index32 = sub nsw i32 %i, %sub |
| 610 | |
| 611 | ; CHECK: %index64 = |
Amara Emerson | 56dca4e3 | 2017-08-04 20:19:46 +0000 | [diff] [blame] | 612 | ; CHECK: --> {(-1 * (sext i32 %sub to i64))<nsw>,+,1}<nsw |
Bjarke Hammersholt Roune | 9791ed4 | 2015-08-14 22:45:26 +0000 | [diff] [blame] | 613 | %index64 = sext i32 %index32 to i64 |
| 614 | |
| 615 | %ptr = getelementptr inbounds float, float* %input, i64 %index64 |
| 616 | %nexti = add nsw i32 %i, 1 |
| 617 | %f = load float, float* %ptr, align 4 |
| 618 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 619 | br i1 %exitcond, label %exit, label %loop |
| 620 | exit: |
| 621 | ret void |
| 622 | } |
| 623 | |
Amara Emerson | 56dca4e3 | 2017-08-04 20:19:46 +0000 | [diff] [blame] | 624 | ; Example checking that a sext is pushed onto a sub's operands if the sub is an |
| 625 | ; overflow intrinsic. |
| 626 | define void @test-sext-sub(float* %input, i32 %sub, i32 %numIterations) { |
| 627 | ; CHECK-LABEL: @test-sext-sub |
| 628 | entry: |
| 629 | br label %loop |
| 630 | loop: |
| 631 | %i = phi i32 [ %nexti, %cont ], [ 0, %entry ] |
| 632 | |
| 633 | ; CHECK: %val = extractvalue { i32, i1 } %ssub, 0 |
| 634 | ; CHECK: --> {(-1 * %sub),+,1}<nw> |
| 635 | %ssub = tail call { i32, i1 } @llvm.ssub.with.overflow.i32(i32 %i, i32 %sub) |
| 636 | %val = extractvalue { i32, i1 } %ssub, 0 |
| 637 | %ovfl = extractvalue { i32, i1 } %ssub, 1 |
| 638 | br i1 %ovfl, label %trap, label %cont |
| 639 | |
| 640 | trap: |
| 641 | tail call void @llvm.trap() |
| 642 | unreachable |
| 643 | |
| 644 | cont: |
| 645 | ; CHECK: %index64 = |
| 646 | ; CHECK: --> {(-1 * (sext i32 %sub to i64))<nsw>,+,1}<nsw |
| 647 | %index64 = sext i32 %val to i64 |
| 648 | |
| 649 | %ptr = getelementptr inbounds float, float* %input, i64 %index64 |
| 650 | %nexti = add nsw i32 %i, 1 |
| 651 | %f = load float, float* %ptr, align 4 |
| 652 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 653 | br i1 %exitcond, label %exit, label %loop |
| 654 | exit: |
| 655 | ret void |
| 656 | } |
| 657 | |
Bjarke Hammersholt Roune | 9791ed4 | 2015-08-14 22:45:26 +0000 | [diff] [blame] | 658 | ; Two adds with a sub in the middle and the sub should have nsw. There is |
| 659 | ; a special case for sequential adds/subs and this test covers that. We have to |
| 660 | ; put the final add first in the program since otherwise the special case |
| 661 | ; is not triggered, hence the strange basic block ordering. |
| 662 | define void @test-sub-with-add(float* %input, i32 %offset, i32 %numIterations) { |
| 663 | ; CHECK-LABEL: @test-sub-with-add |
| 664 | entry: |
| 665 | br label %loop |
| 666 | loop2: |
| 667 | ; CHECK: %seq = |
| 668 | ; CHECK: --> {(2 + (-1 * %offset)),+,1}<nw> |
| 669 | %seq = add nsw nuw i32 %index32, 1 |
| 670 | %exitcond = icmp eq i32 %nexti, %numIterations |
| 671 | br i1 %exitcond, label %exit, label %loop |
| 672 | |
| 673 | loop: |
| 674 | %i = phi i32 [ %nexti, %loop2 ], [ 0, %entry ] |
| 675 | |
| 676 | %j = add nsw i32 %i, 1 |
| 677 | ; CHECK: %index32 = |
Oleg Ranevskyy | eb4ecca | 2016-05-25 13:01:33 +0000 | [diff] [blame] | 678 | ; CHECK: --> {(1 + (-1 * %offset))<nsw>,+,1}<nsw> |
Bjarke Hammersholt Roune | 9791ed4 | 2015-08-14 22:45:26 +0000 | [diff] [blame] | 679 | %index32 = sub nsw i32 %j, %offset |
| 680 | |
| 681 | %ptr = getelementptr inbounds float, float* %input, i32 %index32 |
| 682 | %nexti = add nsw i32 %i, 1 |
| 683 | store float 1.0, float* %ptr, align 4 |
| 684 | br label %loop2 |
| 685 | exit: |
| 686 | ret void |
| 687 | } |
| 688 | |
| 689 | |
| 690 | ; Subtraction of two recurrences. The addition in the SCEV that this |
| 691 | ; maps to is NSW, but the negation of the RHS does not since that |
| 692 | ; recurrence could be the most negative representable value. |
| 693 | define void @subrecurrences(i32 %outer_l, i32 %inner_l, i32 %val) { |
| 694 | ; CHECK-LABEL: @subrecurrences |
| 695 | entry: |
| 696 | br label %outer |
| 697 | |
| 698 | outer: |
| 699 | %o_idx = phi i32 [ 0, %entry ], [ %o_idx.inc, %outer.be ] |
| 700 | %o_idx.inc = add nsw i32 %o_idx, 1 |
| 701 | %cond = icmp eq i32 %o_idx, %val |
| 702 | br i1 %cond, label %inner, label %outer.be |
| 703 | |
| 704 | inner: |
| 705 | %i_idx = phi i32 [ 0, %outer ], [ %i_idx.inc, %inner ] |
| 706 | %i_idx.inc = add nsw i32 %i_idx, 1 |
| 707 | ; CHECK: %v = |
| 708 | ; CHECK-NEXT: --> {{[{][{]}}-1,+,-1}<nw><%outer>,+,1}<nsw><%inner> |
| 709 | %v = sub nsw i32 %i_idx, %o_idx.inc |
| 710 | %forub = udiv i32 1, %v |
| 711 | %cond2 = icmp eq i32 %i_idx, %inner_l |
| 712 | br i1 %cond2, label %outer.be, label %inner |
| 713 | |
| 714 | outer.be: |
| 715 | %cond3 = icmp eq i32 %o_idx, %outer_l |
| 716 | br i1 %cond3, label %exit, label %outer |
| 717 | |
| 718 | exit: |
| 719 | ret void |
| 720 | } |
Hans Wennborg | 3879035 | 2016-08-17 22:50:18 +0000 | [diff] [blame] | 721 | |
| 722 | |
| 723 | ; PR28932: Don't assert on non-SCEV-able value %2. |
| 724 | %struct.anon = type { i8* } |
| 725 | @a = common global %struct.anon* null, align 8 |
| 726 | @b = common global i32 0, align 4 |
| 727 | declare { i32, i1 } @llvm.ssub.with.overflow.i32(i32, i32) |
| 728 | declare void @llvm.trap() |
| 729 | define i32 @pr28932() { |
| 730 | entry: |
| 731 | %.pre = load %struct.anon*, %struct.anon** @a, align 8 |
| 732 | %.pre7 = load i32, i32* @b, align 4 |
| 733 | br label %for.cond |
| 734 | |
| 735 | for.cond: ; preds = %cont6, %entry |
| 736 | %0 = phi i32 [ %3, %cont6 ], [ %.pre7, %entry ] |
| 737 | %1 = phi %struct.anon* [ %.ph, %cont6 ], [ %.pre, %entry ] |
| 738 | %tobool = icmp eq %struct.anon* %1, null |
| 739 | %2 = tail call { i32, i1 } @llvm.ssub.with.overflow.i32(i32 %0, i32 1) |
| 740 | %3 = extractvalue { i32, i1 } %2, 0 |
| 741 | %4 = extractvalue { i32, i1 } %2, 1 |
| 742 | %idxprom = sext i32 %3 to i64 |
| 743 | %5 = getelementptr inbounds %struct.anon, %struct.anon* %1, i64 0, i32 0 |
| 744 | %6 = load i8*, i8** %5, align 8 |
| 745 | %7 = getelementptr inbounds i8, i8* %6, i64 %idxprom |
| 746 | %8 = load i8, i8* %7, align 1 |
| 747 | br i1 %tobool, label %if.else, label %if.then |
| 748 | |
| 749 | if.then: ; preds = %for.cond |
| 750 | br i1 %4, label %trap, label %cont6 |
| 751 | |
| 752 | trap: ; preds = %if.else, %if.then |
| 753 | tail call void @llvm.trap() |
| 754 | unreachable |
| 755 | |
| 756 | if.else: ; preds = %for.cond |
| 757 | br i1 %4, label %trap, label %cont1 |
| 758 | |
| 759 | cont1: ; preds = %if.else |
| 760 | %conv5 = sext i8 %8 to i64 |
| 761 | %9 = inttoptr i64 %conv5 to %struct.anon* |
| 762 | store %struct.anon* %9, %struct.anon** @a, align 8 |
| 763 | br label %cont6 |
| 764 | |
| 765 | cont6: ; preds = %cont1, %if.then |
| 766 | %.ph = phi %struct.anon* [ %9, %cont1 ], [ %1, %if.then ] |
| 767 | store i32 %3, i32* @b, align 4 |
| 768 | br label %for.cond |
| 769 | } |