1; RUN: opt < %s -analyze -scalar-evolution | FileCheck %s 2 3; ScalarEvolution should be able to understand the loop and eliminate the casts. 4 5; CHECK: {%d,+,4} 6 7define void @foo(i32* nocapture %d, i32 %n) nounwind { 8entry: 9 %0 = icmp sgt i32 %n, 0 ; <i1> [#uses=1] 10 br i1 %0, label %bb.nph, label %return 11 12bb.nph: ; preds = %entry 13 br label %bb 14 15bb: ; preds = %bb1, %bb.nph 16 %i.02 = phi i32 [ %5, %bb1 ], [ 0, %bb.nph ] ; <i32> [#uses=2] 17 %p.01 = phi i8 [ %4, %bb1 ], [ -1, %bb.nph ] ; <i8> [#uses=2] 18 %1 = sext i8 %p.01 to i32 ; <i32> [#uses=1] 19 %2 = sext i32 %i.02 to i64 ; <i64> [#uses=1] 20 %3 = getelementptr i32, i32* %d, i64 %2 ; <i32*> [#uses=1] 21 store i32 %1, i32* %3, align 4 22 %4 = add i8 %p.01, 1 ; <i8> [#uses=1] 23 %5 = add i32 %i.02, 1 ; <i32> [#uses=2] 24 br label %bb1 25 26bb1: ; preds = %bb 27 %6 = icmp slt i32 %5, %n ; <i1> [#uses=1] 28 br i1 %6, label %bb, label %bb1.return_crit_edge 29 30bb1.return_crit_edge: ; preds = %bb1 31 br label %return 32 33return: ; preds = %bb1.return_crit_edge, %entry 34 ret void 35} 36 37; ScalarEvolution should be able to find the maximum tripcount 38; of this multiple-exit loop, and if it doesn't know the exact 39; count, it should say so. 40 41; PR7845 42; CHECK: Loop %for.cond: <multiple exits> Unpredictable backedge-taken count. 43; CHECK: Loop %for.cond: max backedge-taken count is 5 44 45@.str = private constant [4 x i8] c"%d\0A\00" ; <[4 x i8]*> [#uses=2] 46 47define i32 @main() nounwind { 48entry: 49 br label %for.cond 50 51for.cond: ; preds = %for.inc, %entry 52 %g_4.0 = phi i32 [ 0, %entry ], [ %add, %for.inc ] ; <i32> [#uses=5] 53 %cmp = icmp slt i32 %g_4.0, 5 ; <i1> [#uses=1] 54 br i1 %cmp, label %for.body, label %for.end 55 56for.body: ; preds = %for.cond 57 %conv = trunc i32 %g_4.0 to i16 ; <i16> [#uses=1] 58 %tobool.not = icmp eq i16 %conv, 0 ; <i1> [#uses=1] 59 %tobool3 = icmp ne i32 %g_4.0, 0 ; <i1> [#uses=1] 60 %or.cond = and i1 %tobool.not, %tobool3 ; <i1> [#uses=1] 61 br i1 %or.cond, label %for.end, label %for.inc 62 63for.inc: ; preds = %for.body 64 %add = add nsw i32 %g_4.0, 1 ; <i32> [#uses=1] 65 br label %for.cond 66 67for.end: ; preds = %for.body, %for.cond 68 %call = call i32 (i8*, ...) @printf(i8* getelementptr inbounds ([4 x i8], [4 x i8]* @.str, i64 0, i64 0), i32 %g_4.0) nounwind ; <i32> [#uses=0] 69 ret i32 0 70} 71 72declare i32 @printf(i8*, ...) 73 74define void @test(i8* %a, i32 %n) nounwind { 75entry: 76 %cmp1 = icmp sgt i32 %n, 0 77 br i1 %cmp1, label %for.body.lr.ph, label %for.end 78 79for.body.lr.ph: ; preds = %entry 80 %tmp = zext i32 %n to i64 81 br label %for.body 82 83for.body: ; preds = %for.body, %for.body.lr.ph 84 %indvar = phi i64 [ %indvar.next, %for.body ], [ 0, %for.body.lr.ph ] 85 %arrayidx = getelementptr i8, i8* %a, i64 %indvar 86 store i8 0, i8* %arrayidx, align 1 87 %indvar.next = add i64 %indvar, 1 88 %exitcond = icmp ne i64 %indvar.next, %tmp 89 br i1 %exitcond, label %for.body, label %for.cond.for.end_crit_edge 90 91for.cond.for.end_crit_edge: ; preds = %for.body 92 br label %for.end 93 94for.end: ; preds = %for.cond.for.end_crit_edge, %entry 95 ret void 96} 97 98; CHECK: Determining loop execution counts for: @test 99; CHECK-NEXT: backedge-taken count is 100; CHECK-NEXT: max backedge-taken count is -1 101 102; PR19799: Indvars miscompile due to an incorrect max backedge taken count from SCEV. 103; CHECK-LABEL: @pr19799 104; CHECK: Loop %for.body.i: <multiple exits> Unpredictable backedge-taken count. 105; CHECK: Loop %for.body.i: max backedge-taken count is 1 106@a = common global i32 0, align 4 107 108define i32 @pr19799() { 109entry: 110 store i32 -1, i32* @a, align 4 111 br label %for.body.i 112 113for.body.i: ; preds = %for.cond.i, %entry 114 %storemerge1.i = phi i32 [ -1, %entry ], [ %add.i.i, %for.cond.i ] 115 %tobool.i = icmp eq i32 %storemerge1.i, 0 116 %add.i.i = add nsw i32 %storemerge1.i, 2 117 br i1 %tobool.i, label %bar.exit, label %for.cond.i 118 119for.cond.i: ; preds = %for.body.i 120 store i32 %add.i.i, i32* @a, align 4 121 %cmp.i = icmp slt i32 %storemerge1.i, 0 122 br i1 %cmp.i, label %for.body.i, label %bar.exit 123 124bar.exit: ; preds = %for.cond.i, %for.body.i 125 ret i32 0 126} 127 128; PR18886: Indvars miscompile due to an incorrect max backedge taken count from SCEV. 129; CHECK-LABEL: @pr18886 130; CHECK: Loop %for.body: <multiple exits> Unpredictable backedge-taken count. 131; CHECK: Loop %for.body: max backedge-taken count is 3 132@aa = global i64 0, align 8 133 134define i32 @pr18886() { 135entry: 136 store i64 -21, i64* @aa, align 8 137 br label %for.body 138 139for.body: 140 %storemerge1 = phi i64 [ -21, %entry ], [ %add, %for.cond ] 141 %tobool = icmp eq i64 %storemerge1, 0 142 %add = add nsw i64 %storemerge1, 8 143 br i1 %tobool, label %return, label %for.cond 144 145for.cond: 146 store i64 %add, i64* @aa, align 8 147 %cmp = icmp slt i64 %add, 9 148 br i1 %cmp, label %for.body, label %return 149 150return: 151 %retval.0 = phi i32 [ 1, %for.body ], [ 0, %for.cond ] 152 ret i32 %retval.0 153} 154 155; Here we have a must-exit loop latch that is not computable and a 156; may-exit early exit that can only have one non-exiting iteration 157; before the check is forever skipped. 158; 159; CHECK-LABEL: @cannot_compute_mustexit 160; CHECK: Loop %for.body.i: <multiple exits> Unpredictable backedge-taken count. 161; CHECK: Loop %for.body.i: Unpredictable max backedge-taken count. 162@b = common global i32 0, align 4 163 164define i32 @cannot_compute_mustexit() { 165entry: 166 store i32 -1, i32* @a, align 4 167 br label %for.body.i 168 169for.body.i: ; preds = %for.cond.i, %entry 170 %storemerge1.i = phi i32 [ -1, %entry ], [ %add.i.i, %for.cond.i ] 171 %tobool.i = icmp eq i32 %storemerge1.i, 0 172 %add.i.i = add nsw i32 %storemerge1.i, 2 173 br i1 %tobool.i, label %bar.exit, label %for.cond.i 174 175for.cond.i: ; preds = %for.body.i 176 store i32 %add.i.i, i32* @a, align 4 177 %ld = load volatile i32, i32* @b 178 %cmp.i = icmp ne i32 %ld, 0 179 br i1 %cmp.i, label %for.body.i, label %bar.exit 180 181bar.exit: ; preds = %for.cond.i, %for.body.i 182 ret i32 0 183} 184 185; This loop has two must-exits, both of which dominate the latch. The 186; MaxBECount should be the minimum of them. 187; 188; CHECK-LABEL: @two_mustexit 189; CHECK: Loop %for.body.i: <multiple exits> backedge-taken count is 1 190; CHECK: Loop %for.body.i: max backedge-taken count is 1 191define i32 @two_mustexit() { 192entry: 193 store i32 -1, i32* @a, align 4 194 br label %for.body.i 195 196for.body.i: ; preds = %for.cond.i, %entry 197 %storemerge1.i = phi i32 [ -1, %entry ], [ %add.i.i, %for.cond.i ] 198 %tobool.i = icmp sgt i32 %storemerge1.i, 0 199 %add.i.i = add nsw i32 %storemerge1.i, 2 200 br i1 %tobool.i, label %bar.exit, label %for.cond.i 201 202for.cond.i: ; preds = %for.body.i 203 store i32 %add.i.i, i32* @a, align 4 204 %cmp.i = icmp slt i32 %storemerge1.i, 3 205 br i1 %cmp.i, label %for.body.i, label %bar.exit 206 207bar.exit: ; preds = %for.cond.i, %for.body.i 208 ret i32 0 209} 210 211; CHECK-LABEL: @ne_max_trip_count_1 212; CHECK: Loop %for.body: max backedge-taken count is 7 213define i32 @ne_max_trip_count_1(i32 %n) { 214entry: 215 %masked = and i32 %n, 7 216 br label %for.body 217 218for.body: 219 %i = phi i32 [ 0, %entry ], [ %add, %for.body ] 220 %add = add nsw i32 %i, 1 221 %cmp = icmp ne i32 %i, %masked 222 br i1 %cmp, label %for.body, label %bar.exit 223 224bar.exit: 225 ret i32 0 226} 227 228; CHECK-LABEL: @ne_max_trip_count_2 229; CHECK: Loop %for.body: max backedge-taken count is -1 230define i32 @ne_max_trip_count_2(i32 %n) { 231entry: 232 %masked = and i32 %n, 7 233 br label %for.body 234 235for.body: 236 %i = phi i32 [ 0, %entry ], [ %add, %for.body ] 237 %add = add nsw i32 %i, 1 238 %cmp = icmp ne i32 %add, %masked 239 br i1 %cmp, label %for.body, label %bar.exit 240 241bar.exit: 242 ret i32 0 243} 244 245; CHECK-LABEL: @ne_max_trip_count_3 246; CHECK: Loop %for.body: max backedge-taken count is 6 247define i32 @ne_max_trip_count_3(i32 %n) { 248entry: 249 %masked = and i32 %n, 7 250 %guard = icmp eq i32 %masked, 0 251 br i1 %guard, label %exit, label %for.preheader 252 253for.preheader: 254 br label %for.body 255 256for.body: 257 %i = phi i32 [ 0, %for.preheader ], [ %add, %for.body ] 258 %add = add nsw i32 %i, 1 259 %cmp = icmp ne i32 %add, %masked 260 br i1 %cmp, label %for.body, label %loop.exit 261 262loop.exit: 263 br label %exit 264 265exit: 266 ret i32 0 267} 268 269; CHECK-LABEL: @ne_max_trip_count_4 270; CHECK: Loop %for.body: max backedge-taken count is -2 271define i32 @ne_max_trip_count_4(i32 %n) { 272entry: 273 %guard = icmp eq i32 %n, 0 274 br i1 %guard, label %exit, label %for.preheader 275 276for.preheader: 277 br label %for.body 278 279for.body: 280 %i = phi i32 [ 0, %for.preheader ], [ %add, %for.body ] 281 %add = add nsw i32 %i, 1 282 %cmp = icmp ne i32 %add, %n 283 br i1 %cmp, label %for.body, label %loop.exit 284 285loop.exit: 286 br label %exit 287 288exit: 289 ret i32 0 290} 291 292; The end bound of the loop can change between iterations, so the exact trip 293; count is unknown, but SCEV can calculate the max trip count. 294define void @changing_end_bound(i32* %n_addr, i32* %addr) { 295; CHECK-LABEL: Determining loop execution counts for: @changing_end_bound 296; CHECK: Loop %loop: Unpredictable backedge-taken count. 297; CHECK: Loop %loop: max backedge-taken count is 2147483646 298entry: 299 br label %loop 300 301loop: 302 %iv = phi i32 [ 0, %entry ], [ %iv.next, %loop ] 303 %acc = phi i32 [ 0, %entry ], [ %acc.next, %loop ] 304 %val = load atomic i32, i32* %addr unordered, align 4 305 fence acquire 306 %acc.next = add i32 %acc, %val 307 %iv.next = add nsw i32 %iv, 1 308 %n = load atomic i32, i32* %n_addr unordered, align 4 309 %cmp = icmp slt i32 %iv.next, %n 310 br i1 %cmp, label %loop, label %loop.exit 311 312loop.exit: 313 ret void 314} 315 316; Similar test as above, but unknown start value. 317; Also, there's no nsw on the iv.next, but SCEV knows 318; the termination condition is LT, so the IV cannot wrap. 319define void @changing_end_bound2(i32 %start, i32* %n_addr, i32* %addr) { 320; CHECK-LABEL: Determining loop execution counts for: @changing_end_bound2 321; CHECK: Loop %loop: Unpredictable backedge-taken count. 322; CHECK: Loop %loop: max backedge-taken count is -1 323entry: 324 br label %loop 325 326loop: 327 %iv = phi i32 [ %start, %entry ], [ %iv.next, %loop ] 328 %acc = phi i32 [ 0, %entry ], [ %acc.next, %loop ] 329 %val = load atomic i32, i32* %addr unordered, align 4 330 fence acquire 331 %acc.next = add i32 %acc, %val 332 %iv.next = add i32 %iv, 1 333 %n = load atomic i32, i32* %n_addr unordered, align 4 334 %cmp = icmp slt i32 %iv.next, %n 335 br i1 %cmp, label %loop, label %loop.exit 336 337loop.exit: 338 ret void 339} 340 341; changing end bound and greater than one stride 342define void @changing_end_bound3(i32 %start, i32* %n_addr, i32* %addr) { 343; CHECK-LABEL: Determining loop execution counts for: @changing_end_bound3 344; CHECK: Loop %loop: Unpredictable backedge-taken count. 345; CHECK: Loop %loop: max backedge-taken count is 1073741823 346entry: 347 br label %loop 348 349loop: 350 %iv = phi i32 [ %start, %entry ], [ %iv.next, %loop ] 351 %acc = phi i32 [ 0, %entry ], [ %acc.next, %loop ] 352 %val = load atomic i32, i32* %addr unordered, align 4 353 fence acquire 354 %acc.next = add i32 %acc, %val 355 %iv.next = add nsw i32 %iv, 4 356 %n = load atomic i32, i32* %n_addr unordered, align 4 357 %cmp = icmp slt i32 %iv.next, %n 358 br i1 %cmp, label %loop, label %loop.exit 359 360loop.exit: 361 ret void 362} 363 364; same as above test, but the IV can wrap around. 365; so the max backedge taken count is unpredictable. 366define void @changing_end_bound4(i32 %start, i32* %n_addr, i32* %addr) { 367; CHECK-LABEL: Determining loop execution counts for: @changing_end_bound4 368; CHECK: Loop %loop: Unpredictable backedge-taken count. 369; CHECK: Loop %loop: Unpredictable max backedge-taken count. 370entry: 371 br label %loop 372 373loop: 374 %iv = phi i32 [ %start, %entry ], [ %iv.next, %loop ] 375 %acc = phi i32 [ 0, %entry ], [ %acc.next, %loop ] 376 %val = load atomic i32, i32* %addr unordered, align 4 377 fence acquire 378 %acc.next = add i32 %acc, %val 379 %iv.next = add i32 %iv, 4 380 %n = load atomic i32, i32* %n_addr unordered, align 4 381 %cmp = icmp slt i32 %iv.next, %n 382 br i1 %cmp, label %loop, label %loop.exit 383 384loop.exit: 385 ret void 386} 387 388; unknown stride. Since it's not knownPositive, we do not estimate the max 389; backedge taken count. 390define void @changing_end_bound5(i32 %stride, i32 %start, i32* %n_addr, i32* %addr) { 391; CHECK-LABEL: Determining loop execution counts for: @changing_end_bound5 392; CHECK: Loop %loop: Unpredictable backedge-taken count. 393; CHECK: Loop %loop: Unpredictable max backedge-taken count. 394entry: 395 br label %loop 396 397loop: 398 %iv = phi i32 [ %start, %entry ], [ %iv.next, %loop ] 399 %acc = phi i32 [ 0, %entry ], [ %acc.next, %loop ] 400 %val = load atomic i32, i32* %addr unordered, align 4 401 fence acquire 402 %acc.next = add i32 %acc, %val 403 %iv.next = add nsw i32 %iv, %stride 404 %n = load atomic i32, i32* %n_addr unordered, align 4 405 %cmp = icmp slt i32 %iv.next, %n 406 br i1 %cmp, label %loop, label %loop.exit 407 408loop.exit: 409 ret void 410} 411 412; negative stride value 413define void @changing_end_bound6(i32 %start, i32* %n_addr, i32* %addr) { 414; CHECK-LABEL: Determining loop execution counts for: @changing_end_bound6 415; CHECK: Loop %loop: Unpredictable backedge-taken count. 416; CHECK: Loop %loop: Unpredictable max backedge-taken count. 417entry: 418 br label %loop 419 420loop: 421 %iv = phi i32 [ %start, %entry ], [ %iv.next, %loop ] 422 %acc = phi i32 [ 0, %entry ], [ %acc.next, %loop ] 423 %val = load atomic i32, i32* %addr unordered, align 4 424 fence acquire 425 %acc.next = add i32 %acc, %val 426 %iv.next = add nsw i32 %iv, -1 427 %n = load atomic i32, i32* %n_addr unordered, align 4 428 %cmp = icmp slt i32 %iv.next, %n 429 br i1 %cmp, label %loop, label %loop.exit 430 431loop.exit: 432 ret void 433} 434