1; RUN: opt < %s -basicaa -gvn -S -die | FileCheck %s 2 3; 32-bit little endian target. 4target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128-n8:16:32" 5 6;; Trivial RLE test. 7define i32 @test0(i32 %V, i32* %P) { 8 store i32 %V, i32* %P 9 10 %A = load i32* %P 11 ret i32 %A 12; CHECK: @test0 13; CHECK: ret i32 %V 14} 15 16 17;;===----------------------------------------------------------------------===;; 18;; Tests for crashers 19;;===----------------------------------------------------------------------===;; 20 21;; PR5016 22define i8 @crash0({i32, i32} %A, {i32, i32}* %P) { 23 store {i32, i32} %A, {i32, i32}* %P 24 %X = bitcast {i32, i32}* %P to i8* 25 %Y = load i8* %X 26 ret i8 %Y 27} 28 29;; No PR filed, crashed in CaptureTracker. 30declare void @helper() 31define void @crash1() { 32 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* undef, i8* undef, i64 undef, i32 1, i1 false) nounwind 33 %tmp = load i8* bitcast (void ()* @helper to i8*) 34 %x = icmp eq i8 %tmp, 15 35 ret void 36} 37 38 39;;===----------------------------------------------------------------------===;; 40;; Store -> Load and Load -> Load forwarding where src and dst are different 41;; types, but where the base pointer is a must alias. 42;;===----------------------------------------------------------------------===;; 43 44;; i32 -> f32 forwarding. 45define float @coerce_mustalias1(i32 %V, i32* %P) { 46 store i32 %V, i32* %P 47 48 %P2 = bitcast i32* %P to float* 49 50 %A = load float* %P2 51 ret float %A 52; CHECK: @coerce_mustalias1 53; CHECK-NOT: load 54; CHECK: ret float 55} 56 57;; i32* -> float forwarding. 58define float @coerce_mustalias2(i32* %V, i32** %P) { 59 store i32* %V, i32** %P 60 61 %P2 = bitcast i32** %P to float* 62 63 %A = load float* %P2 64 ret float %A 65; CHECK: @coerce_mustalias2 66; CHECK-NOT: load 67; CHECK: ret float 68} 69 70;; float -> i32* forwarding. 71define i32* @coerce_mustalias3(float %V, float* %P) { 72 store float %V, float* %P 73 74 %P2 = bitcast float* %P to i32** 75 76 %A = load i32** %P2 77 ret i32* %A 78; CHECK: @coerce_mustalias3 79; CHECK-NOT: load 80; CHECK: ret i32* 81} 82 83;; i32 -> f32 load forwarding. 84define float @coerce_mustalias4(i32* %P, i1 %cond) { 85 %A = load i32* %P 86 87 %P2 = bitcast i32* %P to float* 88 %B = load float* %P2 89 br i1 %cond, label %T, label %F 90T: 91 ret float %B 92 93F: 94 %X = bitcast i32 %A to float 95 ret float %X 96 97; CHECK: @coerce_mustalias4 98; CHECK: %A = load i32* %P 99; CHECK-NOT: load 100; CHECK: ret float 101; CHECK: F: 102} 103 104;; i32 -> i8 forwarding 105define i8 @coerce_mustalias5(i32 %V, i32* %P) { 106 store i32 %V, i32* %P 107 108 %P2 = bitcast i32* %P to i8* 109 110 %A = load i8* %P2 111 ret i8 %A 112; CHECK: @coerce_mustalias5 113; CHECK-NOT: load 114; CHECK: ret i8 115} 116 117;; i64 -> float forwarding 118define float @coerce_mustalias6(i64 %V, i64* %P) { 119 store i64 %V, i64* %P 120 121 %P2 = bitcast i64* %P to float* 122 123 %A = load float* %P2 124 ret float %A 125; CHECK: @coerce_mustalias6 126; CHECK-NOT: load 127; CHECK: ret float 128} 129 130;; i64 -> i8* (32-bit) forwarding 131define i8* @coerce_mustalias7(i64 %V, i64* %P) { 132 store i64 %V, i64* %P 133 134 %P2 = bitcast i64* %P to i8** 135 136 %A = load i8** %P2 137 ret i8* %A 138; CHECK: @coerce_mustalias7 139; CHECK-NOT: load 140; CHECK: ret i8* 141} 142 143; memset -> i16 forwarding. 144define signext i16 @memset_to_i16_local(i16* %A) nounwind ssp { 145entry: 146 %conv = bitcast i16* %A to i8* 147 tail call void @llvm.memset.p0i8.i64(i8* %conv, i8 1, i64 200, i32 1, i1 false) 148 %arrayidx = getelementptr inbounds i16* %A, i64 42 149 %tmp2 = load i16* %arrayidx 150 ret i16 %tmp2 151; CHECK: @memset_to_i16_local 152; CHECK-NOT: load 153; CHECK: ret i16 257 154} 155 156; memset -> float forwarding. 157define float @memset_to_float_local(float* %A, i8 %Val) nounwind ssp { 158entry: 159 %conv = bitcast float* %A to i8* ; <i8*> [#uses=1] 160 tail call void @llvm.memset.p0i8.i64(i8* %conv, i8 %Val, i64 400, i32 1, i1 false) 161 %arrayidx = getelementptr inbounds float* %A, i64 42 ; <float*> [#uses=1] 162 %tmp2 = load float* %arrayidx ; <float> [#uses=1] 163 ret float %tmp2 164; CHECK: @memset_to_float_local 165; CHECK-NOT: load 166; CHECK: zext 167; CHECK-NEXT: shl 168; CHECK-NEXT: or 169; CHECK-NEXT: shl 170; CHECK-NEXT: or 171; CHECK-NEXT: bitcast 172; CHECK-NEXT: ret float 173} 174 175;; non-local memset -> i16 load forwarding. 176define i16 @memset_to_i16_nonlocal0(i16* %P, i1 %cond) { 177 %P3 = bitcast i16* %P to i8* 178 br i1 %cond, label %T, label %F 179T: 180 tail call void @llvm.memset.p0i8.i64(i8* %P3, i8 1, i64 400, i32 1, i1 false) 181 br label %Cont 182 183F: 184 tail call void @llvm.memset.p0i8.i64(i8* %P3, i8 2, i64 400, i32 1, i1 false) 185 br label %Cont 186 187Cont: 188 %P2 = getelementptr i16* %P, i32 4 189 %A = load i16* %P2 190 ret i16 %A 191 192; CHECK: @memset_to_i16_nonlocal0 193; CHECK: Cont: 194; CHECK-NEXT: %A = phi i16 [ 514, %F ], [ 257, %T ] 195; CHECK-NOT: load 196; CHECK: ret i16 %A 197} 198 199@GCst = constant {i32, float, i32 } { i32 42, float 14., i32 97 } 200 201; memset -> float forwarding. 202define float @memcpy_to_float_local(float* %A) nounwind ssp { 203entry: 204 %conv = bitcast float* %A to i8* ; <i8*> [#uses=1] 205 tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %conv, i8* bitcast ({i32, float, i32 }* @GCst to i8*), i64 12, i32 1, i1 false) 206 %arrayidx = getelementptr inbounds float* %A, i64 1 ; <float*> [#uses=1] 207 %tmp2 = load float* %arrayidx ; <float> [#uses=1] 208 ret float %tmp2 209; CHECK: @memcpy_to_float_local 210; CHECK-NOT: load 211; CHECK: ret float 1.400000e+01 212} 213 214 215 216;; non-local i32/float -> i8 load forwarding. 217define i8 @coerce_mustalias_nonlocal0(i32* %P, i1 %cond) { 218 %P2 = bitcast i32* %P to float* 219 %P3 = bitcast i32* %P to i8* 220 br i1 %cond, label %T, label %F 221T: 222 store i32 42, i32* %P 223 br label %Cont 224 225F: 226 store float 1.0, float* %P2 227 br label %Cont 228 229Cont: 230 %A = load i8* %P3 231 ret i8 %A 232 233; CHECK: @coerce_mustalias_nonlocal0 234; CHECK: Cont: 235; CHECK: %A = phi i8 [ 236; CHECK-NOT: load 237; CHECK: ret i8 %A 238} 239 240 241;; non-local i32/float -> i8 load forwarding. This also tests that the "P3" 242;; bitcast equivalence can be properly phi translated. 243define i8 @coerce_mustalias_nonlocal1(i32* %P, i1 %cond) { 244 %P2 = bitcast i32* %P to float* 245 br i1 %cond, label %T, label %F 246T: 247 store i32 42, i32* %P 248 br label %Cont 249 250F: 251 store float 1.0, float* %P2 252 br label %Cont 253 254Cont: 255 %P3 = bitcast i32* %P to i8* 256 %A = load i8* %P3 257 ret i8 %A 258 259;; FIXME: This is disabled because this caused a miscompile in the llvm-gcc 260;; bootstrap, see r82411 261; 262; HECK: @coerce_mustalias_nonlocal1 263; HECK: Cont: 264; HECK: %A = phi i8 [ 265; HECK-NOT: load 266; HECK: ret i8 %A 267} 268 269 270;; non-local i32 -> i8 partial redundancy load forwarding. 271define i8 @coerce_mustalias_pre0(i32* %P, i1 %cond) { 272 %P3 = bitcast i32* %P to i8* 273 br i1 %cond, label %T, label %F 274T: 275 store i32 42, i32* %P 276 br label %Cont 277 278F: 279 br label %Cont 280 281Cont: 282 %A = load i8* %P3 283 ret i8 %A 284 285; CHECK: @coerce_mustalias_pre0 286; CHECK: F: 287; CHECK: load i8* %P3 288; CHECK: Cont: 289; CHECK: %A = phi i8 [ 290; CHECK-NOT: load 291; CHECK: ret i8 %A 292} 293 294;;===----------------------------------------------------------------------===;; 295;; Store -> Load and Load -> Load forwarding where src and dst are different 296;; types, and the reload is an offset from the store pointer. 297;;===----------------------------------------------------------------------===;; 298 299;; i32 -> i8 forwarding. 300;; PR4216 301define i8 @coerce_offset0(i32 %V, i32* %P) { 302 store i32 %V, i32* %P 303 304 %P2 = bitcast i32* %P to i8* 305 %P3 = getelementptr i8* %P2, i32 2 306 307 %A = load i8* %P3 308 ret i8 %A 309; CHECK: @coerce_offset0 310; CHECK-NOT: load 311; CHECK: ret i8 312} 313 314;; non-local i32/float -> i8 load forwarding. 315define i8 @coerce_offset_nonlocal0(i32* %P, i1 %cond) { 316 %P2 = bitcast i32* %P to float* 317 %P3 = bitcast i32* %P to i8* 318 %P4 = getelementptr i8* %P3, i32 2 319 br i1 %cond, label %T, label %F 320T: 321 store i32 42, i32* %P 322 br label %Cont 323 324F: 325 store float 1.0, float* %P2 326 br label %Cont 327 328Cont: 329 %A = load i8* %P4 330 ret i8 %A 331 332; CHECK: @coerce_offset_nonlocal0 333; CHECK: Cont: 334; CHECK: %A = phi i8 [ 335; CHECK-NOT: load 336; CHECK: ret i8 %A 337} 338 339 340;; non-local i32 -> i8 partial redundancy load forwarding. 341define i8 @coerce_offset_pre0(i32* %P, i1 %cond) { 342 %P3 = bitcast i32* %P to i8* 343 %P4 = getelementptr i8* %P3, i32 2 344 br i1 %cond, label %T, label %F 345T: 346 store i32 42, i32* %P 347 br label %Cont 348 349F: 350 br label %Cont 351 352Cont: 353 %A = load i8* %P4 354 ret i8 %A 355 356; CHECK: @coerce_offset_pre0 357; CHECK: F: 358; CHECK: load i8* %P4 359; CHECK: Cont: 360; CHECK: %A = phi i8 [ 361; CHECK-NOT: load 362; CHECK: ret i8 %A 363} 364 365define i32 @chained_load(i32** %p) { 366block1: 367 %A = alloca i32* 368 369 %z = load i32** %p 370 store i32* %z, i32** %A 371 br i1 true, label %block2, label %block3 372 373block2: 374 %a = load i32** %p 375 br label %block4 376 377block3: 378 %b = load i32** %p 379 br label %block4 380 381block4: 382 %c = load i32** %p 383 %d = load i32* %c 384 ret i32 %d 385 386; CHECK: @chained_load 387; CHECK: %z = load i32** %p 388; CHECK-NOT: load 389; CHECK: %d = load i32* %z 390; CHECK-NEXT: ret i32 %d 391} 392 393 394declare i1 @cond() readonly 395declare i1 @cond2() readonly 396 397define i32 @phi_trans2() { 398; CHECK: @phi_trans2 399entry: 400 %P = alloca i32, i32 400 401 br label %F1 402 403F1: 404 %A = phi i32 [1, %entry], [2, %F] 405 %cond2 = call i1 @cond() 406 br i1 %cond2, label %T1, label %TY 407 408T1: 409 %P2 = getelementptr i32* %P, i32 %A 410 %x = load i32* %P2 411 %cond = call i1 @cond2() 412 br i1 %cond, label %TX, label %F 413 414F: 415 %P3 = getelementptr i32* %P, i32 2 416 store i32 17, i32* %P3 417 418 store i32 42, i32* %P2 ; Provides "P[A]". 419 br label %F1 420 421TX: 422 ; This load should not be compiled to 'ret i32 42'. An overly clever 423 ; implementation of GVN would see that we're returning 17 if the loop 424 ; executes once or 42 if it executes more than that, but we'd have to do 425 ; loop restructuring to expose this, and GVN shouldn't do this sort of CFG 426 ; transformation. 427 428; CHECK: TX: 429; CHECK: ret i32 %x 430 ret i32 %x 431TY: 432 ret i32 0 433} 434 435define i32 @phi_trans3(i32* %p) { 436; CHECK: @phi_trans3 437block1: 438 br i1 true, label %block2, label %block3 439 440block2: 441 store i32 87, i32* %p 442 br label %block4 443 444block3: 445 %p2 = getelementptr i32* %p, i32 43 446 store i32 97, i32* %p2 447 br label %block4 448 449block4: 450 %A = phi i32 [-1, %block2], [42, %block3] 451 br i1 true, label %block5, label %exit 452 453; CHECK: block4: 454; CHECK-NEXT: %D = phi i32 [ 87, %block2 ], [ 97, %block3 ] 455; CHECK-NOT: load 456 457block5: 458 %B = add i32 %A, 1 459 br i1 true, label %block6, label %exit 460 461block6: 462 %C = getelementptr i32* %p, i32 %B 463 br i1 true, label %block7, label %exit 464 465block7: 466 %D = load i32* %C 467 ret i32 %D 468 469; CHECK: block7: 470; CHECK-NEXT: ret i32 %D 471 472exit: 473 ret i32 -1 474} 475 476define i8 @phi_trans4(i8* %p) { 477; CHECK: @phi_trans4 478entry: 479 %X3 = getelementptr i8* %p, i32 192 480 store i8 192, i8* %X3 481 482 %X = getelementptr i8* %p, i32 4 483 %Y = load i8* %X 484 br label %loop 485 486loop: 487 %i = phi i32 [4, %entry], [192, %loop] 488 %X2 = getelementptr i8* %p, i32 %i 489 %Y2 = load i8* %X2 490 491; CHECK: loop: 492; CHECK-NEXT: %Y2 = phi i8 [ %Y, %entry ], [ 0, %loop ] 493; CHECK-NOT: load i8 494 495 %cond = call i1 @cond2() 496 497 %Z = bitcast i8 *%X3 to i32* 498 store i32 0, i32* %Z 499 br i1 %cond, label %loop, label %out 500 501out: 502 %R = add i8 %Y, %Y2 503 ret i8 %R 504} 505 506define i8 @phi_trans5(i8* %p) { 507; CHECK: @phi_trans5 508entry: 509 510 %X4 = getelementptr i8* %p, i32 2 511 store i8 19, i8* %X4 512 513 %X = getelementptr i8* %p, i32 4 514 %Y = load i8* %X 515 br label %loop 516 517loop: 518 %i = phi i32 [4, %entry], [3, %cont] 519 %X2 = getelementptr i8* %p, i32 %i 520 %Y2 = load i8* %X2 ; Ensure this load is not being incorrectly replaced. 521 %cond = call i1 @cond2() 522 br i1 %cond, label %cont, label %out 523 524cont: 525 %Z = getelementptr i8* %X2, i32 -1 526 %Z2 = bitcast i8 *%Z to i32* 527 store i32 50462976, i32* %Z2 ;; (1 << 8) | (2 << 16) | (3 << 24) 528 529 530; CHECK: store i32 531; CHECK-NEXT: getelementptr i8* %p, i32 3 532; CHECK-NEXT: load i8* 533 br label %loop 534 535out: 536 %R = add i8 %Y, %Y2 537 ret i8 %R 538} 539 540 541; PR6642 542define i32 @memset_to_load() nounwind readnone { 543entry: 544 %x = alloca [256 x i32], align 4 ; <[256 x i32]*> [#uses=2] 545 %tmp = bitcast [256 x i32]* %x to i8* ; <i8*> [#uses=1] 546 call void @llvm.memset.p0i8.i64(i8* %tmp, i8 0, i64 1024, i32 4, i1 false) 547 %arraydecay = getelementptr inbounds [256 x i32]* %x, i32 0, i32 0 ; <i32*> 548 %tmp1 = load i32* %arraydecay ; <i32> [#uses=1] 549 ret i32 %tmp1 550; CHECK: @memset_to_load 551; CHECK: ret i32 0 552} 553 554 555;;===----------------------------------------------------------------------===;; 556;; Load -> Load forwarding in partial alias case. 557;;===----------------------------------------------------------------------===;; 558 559define i32 @load_load_partial_alias(i8* %P) nounwind ssp { 560entry: 561 %0 = bitcast i8* %P to i32* 562 %tmp2 = load i32* %0 563 %add.ptr = getelementptr inbounds i8* %P, i64 1 564 %tmp5 = load i8* %add.ptr 565 %conv = zext i8 %tmp5 to i32 566 %add = add nsw i32 %tmp2, %conv 567 ret i32 %add 568 569; TEMPORARILYDISABLED: @load_load_partial_alias 570; TEMPORARILYDISABLED: load i32* 571; TEMPORARILYDISABLED-NOT: load 572; TEMPORARILYDISABLED: lshr i32 {{.*}}, 8 573; TEMPORARILYDISABLED-NOT: load 574; TEMPORARILYDISABLED: trunc i32 {{.*}} to i8 575; TEMPORARILYDISABLED-NOT: load 576; TEMPORARILYDISABLED: ret i32 577} 578 579 580; Cross block partial alias case. 581define i32 @load_load_partial_alias_cross_block(i8* %P) nounwind ssp { 582entry: 583 %xx = bitcast i8* %P to i32* 584 %x1 = load i32* %xx, align 4 585 %cmp = icmp eq i32 %x1, 127 586 br i1 %cmp, label %land.lhs.true, label %if.end 587 588land.lhs.true: ; preds = %entry 589 %arrayidx4 = getelementptr inbounds i8* %P, i64 1 590 %tmp5 = load i8* %arrayidx4, align 1 591 %conv6 = zext i8 %tmp5 to i32 592 ret i32 %conv6 593 594if.end: 595 ret i32 52 596; TEMPORARILY_DISABLED: @load_load_partial_alias_cross_block 597; TEMPORARILY_DISABLED: land.lhs.true: 598; TEMPORARILY_DISABLED-NOT: load i8 599; TEMPORARILY_DISABLED: ret i32 %conv6 600} 601 602 603;;===----------------------------------------------------------------------===;; 604;; Load Widening 605;;===----------------------------------------------------------------------===;; 606 607%widening1 = type { i32, i8, i8, i8, i8 } 608 609@f = global %widening1 zeroinitializer, align 4 610 611define i32 @test_widening1(i8* %P) nounwind ssp noredzone { 612entry: 613 %tmp = load i8* getelementptr inbounds (%widening1* @f, i64 0, i32 1), align 4 614 %conv = zext i8 %tmp to i32 615 %tmp1 = load i8* getelementptr inbounds (%widening1* @f, i64 0, i32 2), align 1 616 %conv2 = zext i8 %tmp1 to i32 617 %add = add nsw i32 %conv, %conv2 618 ret i32 %add 619; CHECK: @test_widening1 620; CHECK-NOT: load 621; CHECK: load i16* 622; CHECK-NOT: load 623; CHECK-ret i32 624} 625 626define i32 @test_widening2() nounwind ssp noredzone { 627entry: 628 %tmp = load i8* getelementptr inbounds (%widening1* @f, i64 0, i32 1), align 4 629 %conv = zext i8 %tmp to i32 630 %tmp1 = load i8* getelementptr inbounds (%widening1* @f, i64 0, i32 2), align 1 631 %conv2 = zext i8 %tmp1 to i32 632 %add = add nsw i32 %conv, %conv2 633 634 %tmp2 = load i8* getelementptr inbounds (%widening1* @f, i64 0, i32 3), align 2 635 %conv3 = zext i8 %tmp2 to i32 636 %add2 = add nsw i32 %add, %conv3 637 638 %tmp3 = load i8* getelementptr inbounds (%widening1* @f, i64 0, i32 4), align 1 639 %conv4 = zext i8 %tmp3 to i32 640 %add3 = add nsw i32 %add2, %conv3 641 642 ret i32 %add3 643; CHECK: @test_widening2 644; CHECK-NOT: load 645; CHECK: load i32* 646; CHECK-NOT: load 647; CHECK-ret i32 648} 649 650declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i32, i1) nounwind 651 652declare void @llvm.memcpy.p0i8.p0i8.i64(i8* nocapture, i8* nocapture, i64, i32, i1) nounwind 653 654;;===----------------------------------------------------------------------===;; 655;; Load -> Store dependency which isn't interfered with by a call that happens 656;; before the pointer was captured. 657;;===----------------------------------------------------------------------===;; 658 659%class.X = type { [8 x i8] } 660 661@_ZTV1X = weak_odr constant [5 x i8*] zeroinitializer 662@_ZTV1Y = weak_odr constant [5 x i8*] zeroinitializer 663 664declare void @use() 665declare void @use3(i8***, i8**) 666 667; PR8908 668define void @test_escape1() nounwind { 669 %x = alloca i8**, align 8 670 store i8** getelementptr inbounds ([5 x i8*]* @_ZTV1X, i64 0, i64 2), i8*** %x, align 8 671 call void @use() nounwind 672 %DEAD = load i8*** %x, align 8 673 call void @use3(i8*** %x, i8** %DEAD) nounwind 674 ret void 675; CHECK: test_escape1 676; CHECK-NOT: DEAD 677; CHECK: ret 678} 679