1; RUN: llc -mtriple=i686-linux -pre-RA-sched=source < %s | FileCheck %s 2 3declare void @error(i32 %i, i32 %a, i32 %b) 4 5define i32 @test_ifchains(i32 %i, i32* %a, i32 %b) { 6; Test a chain of ifs, where the block guarded by the if is error handling code 7; that is not expected to run. 8; CHECK-LABEL: test_ifchains: 9; CHECK: %entry 10; CHECK-NOT: .align 11; CHECK: %else1 12; CHECK-NOT: .align 13; CHECK: %else2 14; CHECK-NOT: .align 15; CHECK: %else3 16; CHECK-NOT: .align 17; CHECK: %else4 18; CHECK-NOT: .align 19; CHECK: %exit 20; CHECK: %then1 21; CHECK: %then2 22; CHECK: %then3 23; CHECK: %then4 24; CHECK: %then5 25 26entry: 27 %gep1 = getelementptr i32, i32* %a, i32 1 28 %val1 = load i32, i32* %gep1 29 %cond1 = icmp ugt i32 %val1, 1 30 br i1 %cond1, label %then1, label %else1, !prof !0 31 32then1: 33 call void @error(i32 %i, i32 1, i32 %b) 34 br label %else1 35 36else1: 37 %gep2 = getelementptr i32, i32* %a, i32 2 38 %val2 = load i32, i32* %gep2 39 %cond2 = icmp ugt i32 %val2, 2 40 br i1 %cond2, label %then2, label %else2, !prof !0 41 42then2: 43 call void @error(i32 %i, i32 1, i32 %b) 44 br label %else2 45 46else2: 47 %gep3 = getelementptr i32, i32* %a, i32 3 48 %val3 = load i32, i32* %gep3 49 %cond3 = icmp ugt i32 %val3, 3 50 br i1 %cond3, label %then3, label %else3, !prof !0 51 52then3: 53 call void @error(i32 %i, i32 1, i32 %b) 54 br label %else3 55 56else3: 57 %gep4 = getelementptr i32, i32* %a, i32 4 58 %val4 = load i32, i32* %gep4 59 %cond4 = icmp ugt i32 %val4, 4 60 br i1 %cond4, label %then4, label %else4, !prof !0 61 62then4: 63 call void @error(i32 %i, i32 1, i32 %b) 64 br label %else4 65 66else4: 67 %gep5 = getelementptr i32, i32* %a, i32 3 68 %val5 = load i32, i32* %gep5 69 %cond5 = icmp ugt i32 %val5, 3 70 br i1 %cond5, label %then5, label %exit, !prof !0 71 72then5: 73 call void @error(i32 %i, i32 1, i32 %b) 74 br label %exit 75 76exit: 77 ret i32 %b 78} 79 80define i32 @test_loop_cold_blocks(i32 %i, i32* %a) { 81; Check that we sink cold loop blocks after the hot loop body. 82; CHECK-LABEL: test_loop_cold_blocks: 83; CHECK: %entry 84; CHECK-NOT: .align 85; CHECK: %unlikely1 86; CHECK-NOT: .align 87; CHECK: %unlikely2 88; CHECK: .align 89; CHECK: %body1 90; CHECK: %body2 91; CHECK: %body3 92; CHECK: %exit 93 94entry: 95 br label %body1 96 97body1: 98 %iv = phi i32 [ 0, %entry ], [ %next, %body3 ] 99 %base = phi i32 [ 0, %entry ], [ %sum, %body3 ] 100 %unlikelycond1 = icmp slt i32 %base, 42 101 br i1 %unlikelycond1, label %unlikely1, label %body2, !prof !0 102 103unlikely1: 104 call void @error(i32 %i, i32 1, i32 %base) 105 br label %body2 106 107body2: 108 %unlikelycond2 = icmp sgt i32 %base, 21 109 br i1 %unlikelycond2, label %unlikely2, label %body3, !prof !0 110 111unlikely2: 112 call void @error(i32 %i, i32 2, i32 %base) 113 br label %body3 114 115body3: 116 %arrayidx = getelementptr inbounds i32, i32* %a, i32 %iv 117 %0 = load i32, i32* %arrayidx 118 %sum = add nsw i32 %0, %base 119 %next = add i32 %iv, 1 120 %exitcond = icmp eq i32 %next, %i 121 br i1 %exitcond, label %exit, label %body1 122 123exit: 124 ret i32 %sum 125} 126 127!0 = !{!"branch_weights", i32 4, i32 64} 128 129define i32 @test_loop_early_exits(i32 %i, i32* %a) { 130; Check that we sink early exit blocks out of loop bodies. 131; CHECK-LABEL: test_loop_early_exits: 132; CHECK: %entry 133; CHECK: %body1 134; CHECK: %body2 135; CHECK: %body3 136; CHECK: %body4 137; CHECK: %exit 138; CHECK: %bail1 139; CHECK: %bail2 140; CHECK: %bail3 141 142entry: 143 br label %body1 144 145body1: 146 %iv = phi i32 [ 0, %entry ], [ %next, %body4 ] 147 %base = phi i32 [ 0, %entry ], [ %sum, %body4 ] 148 %bailcond1 = icmp eq i32 %base, 42 149 br i1 %bailcond1, label %bail1, label %body2 150 151bail1: 152 ret i32 -1 153 154body2: 155 %bailcond2 = icmp eq i32 %base, 43 156 br i1 %bailcond2, label %bail2, label %body3 157 158bail2: 159 ret i32 -2 160 161body3: 162 %bailcond3 = icmp eq i32 %base, 44 163 br i1 %bailcond3, label %bail3, label %body4 164 165bail3: 166 ret i32 -3 167 168body4: 169 %arrayidx = getelementptr inbounds i32, i32* %a, i32 %iv 170 %0 = load i32, i32* %arrayidx 171 %sum = add nsw i32 %0, %base 172 %next = add i32 %iv, 1 173 %exitcond = icmp eq i32 %next, %i 174 br i1 %exitcond, label %exit, label %body1 175 176exit: 177 ret i32 %sum 178} 179 180define i32 @test_loop_rotate(i32 %i, i32* %a) { 181; Check that we rotate conditional exits from the loop to the bottom of the 182; loop, eliminating unconditional branches to the top. 183; CHECK-LABEL: test_loop_rotate: 184; CHECK: %entry 185; CHECK: %body1 186; CHECK: %body0 187; CHECK: %exit 188 189entry: 190 br label %body0 191 192body0: 193 %iv = phi i32 [ 0, %entry ], [ %next, %body1 ] 194 %base = phi i32 [ 0, %entry ], [ %sum, %body1 ] 195 %next = add i32 %iv, 1 196 %exitcond = icmp eq i32 %next, %i 197 br i1 %exitcond, label %exit, label %body1 198 199body1: 200 %arrayidx = getelementptr inbounds i32, i32* %a, i32 %iv 201 %0 = load i32, i32* %arrayidx 202 %sum = add nsw i32 %0, %base 203 %bailcond1 = icmp eq i32 %sum, 42 204 br label %body0 205 206exit: 207 ret i32 %base 208} 209 210define i32 @test_no_loop_rotate(i32 %i, i32* %a) { 211; Check that we don't try to rotate a loop which is already laid out with 212; fallthrough opportunities into the top and out of the bottom. 213; CHECK-LABEL: test_no_loop_rotate: 214; CHECK: %entry 215; CHECK: %body0 216; CHECK: %body1 217; CHECK: %exit 218 219entry: 220 br label %body0 221 222body0: 223 %iv = phi i32 [ 0, %entry ], [ %next, %body1 ] 224 %base = phi i32 [ 0, %entry ], [ %sum, %body1 ] 225 %arrayidx = getelementptr inbounds i32, i32* %a, i32 %iv 226 %0 = load i32, i32* %arrayidx 227 %sum = add nsw i32 %0, %base 228 %bailcond1 = icmp eq i32 %sum, 42 229 br i1 %bailcond1, label %exit, label %body1 230 231body1: 232 %next = add i32 %iv, 1 233 %exitcond = icmp eq i32 %next, %i 234 br i1 %exitcond, label %exit, label %body0 235 236exit: 237 ret i32 %base 238} 239 240define i32 @test_loop_align(i32 %i, i32* %a) { 241; Check that we provide basic loop body alignment with the block placement 242; pass. 243; CHECK-LABEL: test_loop_align: 244; CHECK: %entry 245; CHECK: .align [[ALIGN:[0-9]+]], 246; CHECK-NEXT: %body 247; CHECK: %exit 248 249entry: 250 br label %body 251 252body: 253 %iv = phi i32 [ 0, %entry ], [ %next, %body ] 254 %base = phi i32 [ 0, %entry ], [ %sum, %body ] 255 %arrayidx = getelementptr inbounds i32, i32* %a, i32 %iv 256 %0 = load i32, i32* %arrayidx 257 %sum = add nsw i32 %0, %base 258 %next = add i32 %iv, 1 259 %exitcond = icmp eq i32 %next, %i 260 br i1 %exitcond, label %exit, label %body 261 262exit: 263 ret i32 %sum 264} 265 266define i32 @test_nested_loop_align(i32 %i, i32* %a, i32* %b) { 267; Check that we provide nested loop body alignment. 268; CHECK-LABEL: test_nested_loop_align: 269; CHECK: %entry 270; CHECK: .align [[ALIGN]], 271; CHECK-NEXT: %loop.body.1 272; CHECK: .align [[ALIGN]], 273; CHECK-NEXT: %inner.loop.body 274; CHECK-NOT: .align 275; CHECK: %exit 276 277entry: 278 br label %loop.body.1 279 280loop.body.1: 281 %iv = phi i32 [ 0, %entry ], [ %next, %loop.body.2 ] 282 %arrayidx = getelementptr inbounds i32, i32* %a, i32 %iv 283 %bidx = load i32, i32* %arrayidx 284 br label %inner.loop.body 285 286inner.loop.body: 287 %inner.iv = phi i32 [ 0, %loop.body.1 ], [ %inner.next, %inner.loop.body ] 288 %base = phi i32 [ 0, %loop.body.1 ], [ %sum, %inner.loop.body ] 289 %scaled_idx = mul i32 %bidx, %iv 290 %inner.arrayidx = getelementptr inbounds i32, i32* %b, i32 %scaled_idx 291 %0 = load i32, i32* %inner.arrayidx 292 %sum = add nsw i32 %0, %base 293 %inner.next = add i32 %iv, 1 294 %inner.exitcond = icmp eq i32 %inner.next, %i 295 br i1 %inner.exitcond, label %loop.body.2, label %inner.loop.body 296 297loop.body.2: 298 %next = add i32 %iv, 1 299 %exitcond = icmp eq i32 %next, %i 300 br i1 %exitcond, label %exit, label %loop.body.1 301 302exit: 303 ret i32 %sum 304} 305 306define void @unnatural_cfg1() { 307; Test that we can handle a loop with an inner unnatural loop at the end of 308; a function. This is a gross CFG reduced out of the single source GCC. 309; CHECK: unnatural_cfg1 310; CHECK: %entry 311; CHECK: %loop.body1 312; CHECK: %loop.body2 313; CHECK: %loop.body3 314 315entry: 316 br label %loop.header 317 318loop.header: 319 br label %loop.body1 320 321loop.body1: 322 br i1 undef, label %loop.body3, label %loop.body2 323 324loop.body2: 325 %ptr = load i32*, i32** undef, align 4 326 br label %loop.body3 327 328loop.body3: 329 %myptr = phi i32* [ %ptr2, %loop.body5 ], [ %ptr, %loop.body2 ], [ undef, %loop.body1 ] 330 %bcmyptr = bitcast i32* %myptr to i32* 331 %val = load i32, i32* %bcmyptr, align 4 332 %comp = icmp eq i32 %val, 48 333 br i1 %comp, label %loop.body4, label %loop.body5 334 335loop.body4: 336 br i1 undef, label %loop.header, label %loop.body5 337 338loop.body5: 339 %ptr2 = load i32*, i32** undef, align 4 340 br label %loop.body3 341} 342 343define void @unnatural_cfg2() { 344; Test that we can handle a loop with a nested natural loop *and* an unnatural 345; loop. This was reduced from a crash on block placement when run over 346; single-source GCC. 347; CHECK: unnatural_cfg2 348; CHECK: %entry 349; CHECK: %loop.body1 350; CHECK: %loop.body2 351; CHECK: %loop.body3 352; CHECK: %loop.inner1.begin 353; The end block is folded with %loop.body3... 354; CHECK-NOT: %loop.inner1.end 355; CHECK: %loop.body4 356; CHECK: %loop.inner2.begin 357; The loop.inner2.end block is folded 358; CHECK: %loop.header 359; CHECK: %bail 360 361entry: 362 br label %loop.header 363 364loop.header: 365 %comp0 = icmp eq i32* undef, null 366 br i1 %comp0, label %bail, label %loop.body1 367 368loop.body1: 369 %val0 = load i32*, i32** undef, align 4 370 br i1 undef, label %loop.body2, label %loop.inner1.begin 371 372loop.body2: 373 br i1 undef, label %loop.body4, label %loop.body3 374 375loop.body3: 376 %ptr1 = getelementptr inbounds i32, i32* %val0, i32 0 377 %castptr1 = bitcast i32* %ptr1 to i32** 378 %val1 = load i32*, i32** %castptr1, align 4 379 br label %loop.inner1.begin 380 381loop.inner1.begin: 382 %valphi = phi i32* [ %val2, %loop.inner1.end ], [ %val1, %loop.body3 ], [ %val0, %loop.body1 ] 383 %castval = bitcast i32* %valphi to i32* 384 %comp1 = icmp eq i32 undef, 48 385 br i1 %comp1, label %loop.inner1.end, label %loop.body4 386 387loop.inner1.end: 388 %ptr2 = getelementptr inbounds i32, i32* %valphi, i32 0 389 %castptr2 = bitcast i32* %ptr2 to i32** 390 %val2 = load i32*, i32** %castptr2, align 4 391 br label %loop.inner1.begin 392 393loop.body4.dead: 394 br label %loop.body4 395 396loop.body4: 397 %comp2 = icmp ult i32 undef, 3 398 br i1 %comp2, label %loop.inner2.begin, label %loop.end 399 400loop.inner2.begin: 401 br i1 false, label %loop.end, label %loop.inner2.end 402 403loop.inner2.end: 404 %comp3 = icmp eq i32 undef, 1769472 405 br i1 %comp3, label %loop.end, label %loop.inner2.begin 406 407loop.end: 408 br label %loop.header 409 410bail: 411 unreachable 412} 413 414define i32 @problematic_switch() { 415; This function's CFG caused overlow in the machine branch probability 416; calculation, triggering asserts. Make sure we don't crash on it. 417; CHECK: problematic_switch 418 419entry: 420 switch i32 undef, label %exit [ 421 i32 879, label %bogus 422 i32 877, label %step 423 i32 876, label %step 424 i32 875, label %step 425 i32 874, label %step 426 i32 873, label %step 427 i32 872, label %step 428 i32 868, label %step 429 i32 867, label %step 430 i32 866, label %step 431 i32 861, label %step 432 i32 860, label %step 433 i32 856, label %step 434 i32 855, label %step 435 i32 854, label %step 436 i32 831, label %step 437 i32 830, label %step 438 i32 829, label %step 439 i32 828, label %step 440 i32 815, label %step 441 i32 814, label %step 442 i32 811, label %step 443 i32 806, label %step 444 i32 805, label %step 445 i32 804, label %step 446 i32 803, label %step 447 i32 802, label %step 448 i32 801, label %step 449 i32 800, label %step 450 i32 799, label %step 451 i32 798, label %step 452 i32 797, label %step 453 i32 796, label %step 454 i32 795, label %step 455 ] 456bogus: 457 unreachable 458step: 459 br label %exit 460exit: 461 %merge = phi i32 [ 3, %step ], [ 6, %entry ] 462 ret i32 %merge 463} 464 465define void @fpcmp_unanalyzable_branch(i1 %cond) { 466; This function's CFG contains an unanalyzable branch that is likely to be 467; split due to having a different high-probability predecessor. 468; CHECK: fpcmp_unanalyzable_branch 469; CHECK: %entry 470; CHECK: %exit 471; CHECK-NOT: %if.then 472; CHECK-NOT: %if.end 473; CHECK-NOT: jne 474; CHECK-NOT: jnp 475; CHECK: jne 476; CHECK-NEXT: jnp 477; CHECK-NEXT: %if.then 478 479entry: 480; Note that this branch must be strongly biased toward 481; 'entry.if.then_crit_edge' to ensure that we would try to form a chain for 482; 'entry' -> 'entry.if.then_crit_edge' -> 'if.then'. It is the last edge in that 483; chain which would violate the unanalyzable branch in 'exit', but we won't even 484; try this trick unless 'if.then' is believed to almost always be reached from 485; 'entry.if.then_crit_edge'. 486 br i1 %cond, label %entry.if.then_crit_edge, label %lor.lhs.false, !prof !1 487 488entry.if.then_crit_edge: 489 %.pre14 = load i8, i8* undef, align 1 490 br label %if.then 491 492lor.lhs.false: 493 br i1 undef, label %if.end, label %exit 494 495exit: 496 %cmp.i = fcmp une double 0.000000e+00, undef 497 br i1 %cmp.i, label %if.then, label %if.end 498 499if.then: 500 %0 = phi i8 [ %.pre14, %entry.if.then_crit_edge ], [ undef, %exit ] 501 %1 = and i8 %0, 1 502 store i8 %1, i8* undef, align 4 503 br label %if.end 504 505if.end: 506 ret void 507} 508 509!1 = !{!"branch_weights", i32 1000, i32 1} 510 511declare i32 @f() 512declare i32 @g() 513declare i32 @h(i32 %x) 514 515define i32 @test_global_cfg_break_profitability() { 516; Check that our metrics for the profitability of a CFG break are global rather 517; than local. A successor may be very hot, but if the current block isn't, it 518; doesn't matter. Within this test the 'then' block is slightly warmer than the 519; 'else' block, but not nearly enough to merit merging it with the exit block 520; even though the probability of 'then' branching to the 'exit' block is very 521; high. 522; CHECK: test_global_cfg_break_profitability 523; CHECK: calll {{_?}}f 524; CHECK: calll {{_?}}g 525; CHECK: calll {{_?}}h 526; CHECK: ret 527 528entry: 529 br i1 undef, label %then, label %else, !prof !2 530 531then: 532 %then.result = call i32 @f() 533 br label %exit 534 535else: 536 %else.result = call i32 @g() 537 br label %exit 538 539exit: 540 %result = phi i32 [ %then.result, %then ], [ %else.result, %else ] 541 %result2 = call i32 @h(i32 %result) 542 ret i32 %result 543} 544 545!2 = !{!"branch_weights", i32 3, i32 1} 546 547declare i32 @__gxx_personality_v0(...) 548 549define void @test_eh_lpad_successor() personality i8* bitcast (i32 (...)* @__gxx_personality_v0 to i8*) { 550; Some times the landing pad ends up as the first successor of an invoke block. 551; When this happens, a strange result used to fall out of updateTerminators: we 552; didn't correctly locate the fallthrough successor, assuming blindly that the 553; first one was the fallthrough successor. As a result, we would add an 554; erroneous jump to the landing pad thinking *that* was the default successor. 555; CHECK: test_eh_lpad_successor 556; CHECK: %entry 557; CHECK-NOT: jmp 558; CHECK: %loop 559 560entry: 561 invoke i32 @f() to label %preheader unwind label %lpad 562 563preheader: 564 br label %loop 565 566lpad: 567 %lpad.val = landingpad { i8*, i32 } 568 cleanup 569 resume { i8*, i32 } %lpad.val 570 571loop: 572 br label %loop 573} 574 575declare void @fake_throw() noreturn 576 577define void @test_eh_throw() personality i8* bitcast (i32 (...)* @__gxx_personality_v0 to i8*) { 578; For blocks containing a 'throw' (or similar functionality), we have 579; a no-return invoke. In this case, only EH successors will exist, and 580; fallthrough simply won't occur. Make sure we don't crash trying to update 581; terminators for such constructs. 582; 583; CHECK: test_eh_throw 584; CHECK: %entry 585; CHECK: %cleanup 586 587entry: 588 invoke void @fake_throw() to label %continue unwind label %cleanup 589 590continue: 591 unreachable 592 593cleanup: 594 %0 = landingpad { i8*, i32 } 595 cleanup 596 unreachable 597} 598 599define void @test_unnatural_cfg_backwards_inner_loop() { 600; Test that when we encounter an unnatural CFG structure after having formed 601; a chain for an inner loop which happened to be laid out backwards we don't 602; attempt to merge onto the wrong end of the inner loop just because we find it 603; first. This was reduced from a crasher in GCC's single source. 604; 605; CHECK: test_unnatural_cfg_backwards_inner_loop 606; CHECK: %entry 607; CHECK: [[BODY:# BB#[0-9]+]]: 608; CHECK: %loop2b 609; CHECK: %loop1 610; CHECK: %loop2a 611 612entry: 613 br i1 undef, label %loop2a, label %body 614 615body: 616 br label %loop2a 617 618loop1: 619 %next.load = load i32*, i32** undef 620 br i1 %comp.a, label %loop2a, label %loop2b 621 622loop2a: 623 %var = phi i32* [ null, %entry ], [ null, %body ], [ %next.phi, %loop1 ] 624 %next.var = phi i32* [ null, %entry ], [ undef, %body ], [ %next.load, %loop1 ] 625 %comp.a = icmp eq i32* %var, null 626 br label %loop3 627 628loop2b: 629 %gep = getelementptr inbounds i32, i32* %var.phi, i32 0 630 %next.ptr = bitcast i32* %gep to i32** 631 store i32* %next.phi, i32** %next.ptr 632 br label %loop3 633 634loop3: 635 %var.phi = phi i32* [ %next.phi, %loop2b ], [ %var, %loop2a ] 636 %next.phi = phi i32* [ %next.load, %loop2b ], [ %next.var, %loop2a ] 637 br label %loop1 638} 639 640define void @unanalyzable_branch_to_loop_header() { 641; Ensure that we can handle unanalyzable branches into loop headers. We 642; pre-form chains for unanalyzable branches, and will find the tail end of that 643; at the start of the loop. This function uses floating point comparison 644; fallthrough because that happens to always produce unanalyzable branches on 645; x86. 646; 647; CHECK: unanalyzable_branch_to_loop_header 648; CHECK: %entry 649; CHECK: %loop 650; CHECK: %exit 651 652entry: 653 %cmp = fcmp une double 0.000000e+00, undef 654 br i1 %cmp, label %loop, label %exit 655 656loop: 657 %cond = icmp eq i8 undef, 42 658 br i1 %cond, label %exit, label %loop 659 660exit: 661 ret void 662} 663 664define void @unanalyzable_branch_to_best_succ(i1 %cond) { 665; Ensure that we can handle unanalyzable branches where the destination block 666; gets selected as the optimal successor to merge. 667; 668; CHECK: unanalyzable_branch_to_best_succ 669; CHECK: %entry 670; CHECK: %foo 671; CHECK: %bar 672; CHECK: %exit 673 674entry: 675 ; Bias this branch toward bar to ensure we form that chain. 676 br i1 %cond, label %bar, label %foo, !prof !1 677 678foo: 679 %cmp = fcmp une double 0.000000e+00, undef 680 br i1 %cmp, label %bar, label %exit 681 682bar: 683 call i32 @f() 684 br label %exit 685 686exit: 687 ret void 688} 689 690define void @unanalyzable_branch_to_free_block(float %x) { 691; Ensure that we can handle unanalyzable branches where the destination block 692; gets selected as the best free block in the CFG. 693; 694; CHECK: unanalyzable_branch_to_free_block 695; CHECK: %entry 696; CHECK: %a 697; CHECK: %b 698; CHECK: %c 699; CHECK: %exit 700 701entry: 702 br i1 undef, label %a, label %b 703 704a: 705 call i32 @f() 706 br label %c 707 708b: 709 %cmp = fcmp une float %x, undef 710 br i1 %cmp, label %c, label %exit 711 712c: 713 call i32 @g() 714 br label %exit 715 716exit: 717 ret void 718} 719 720define void @many_unanalyzable_branches() { 721; Ensure that we don't crash as we're building up many unanalyzable branches, 722; blocks, and loops. 723; 724; CHECK: many_unanalyzable_branches 725; CHECK: %entry 726; CHECK: %exit 727 728entry: 729 br label %0 730 731 %val0 = load volatile float, float* undef 732 %cmp0 = fcmp une float %val0, undef 733 br i1 %cmp0, label %1, label %0 734 %val1 = load volatile float, float* undef 735 %cmp1 = fcmp une float %val1, undef 736 br i1 %cmp1, label %2, label %1 737 %val2 = load volatile float, float* undef 738 %cmp2 = fcmp une float %val2, undef 739 br i1 %cmp2, label %3, label %2 740 %val3 = load volatile float, float* undef 741 %cmp3 = fcmp une float %val3, undef 742 br i1 %cmp3, label %4, label %3 743 %val4 = load volatile float, float* undef 744 %cmp4 = fcmp une float %val4, undef 745 br i1 %cmp4, label %5, label %4 746 %val5 = load volatile float, float* undef 747 %cmp5 = fcmp une float %val5, undef 748 br i1 %cmp5, label %6, label %5 749 %val6 = load volatile float, float* undef 750 %cmp6 = fcmp une float %val6, undef 751 br i1 %cmp6, label %7, label %6 752 %val7 = load volatile float, float* undef 753 %cmp7 = fcmp une float %val7, undef 754 br i1 %cmp7, label %8, label %7 755 %val8 = load volatile float, float* undef 756 %cmp8 = fcmp une float %val8, undef 757 br i1 %cmp8, label %9, label %8 758 %val9 = load volatile float, float* undef 759 %cmp9 = fcmp une float %val9, undef 760 br i1 %cmp9, label %10, label %9 761 %val10 = load volatile float, float* undef 762 %cmp10 = fcmp une float %val10, undef 763 br i1 %cmp10, label %11, label %10 764 %val11 = load volatile float, float* undef 765 %cmp11 = fcmp une float %val11, undef 766 br i1 %cmp11, label %12, label %11 767 %val12 = load volatile float, float* undef 768 %cmp12 = fcmp une float %val12, undef 769 br i1 %cmp12, label %13, label %12 770 %val13 = load volatile float, float* undef 771 %cmp13 = fcmp une float %val13, undef 772 br i1 %cmp13, label %14, label %13 773 %val14 = load volatile float, float* undef 774 %cmp14 = fcmp une float %val14, undef 775 br i1 %cmp14, label %15, label %14 776 %val15 = load volatile float, float* undef 777 %cmp15 = fcmp une float %val15, undef 778 br i1 %cmp15, label %16, label %15 779 %val16 = load volatile float, float* undef 780 %cmp16 = fcmp une float %val16, undef 781 br i1 %cmp16, label %17, label %16 782 %val17 = load volatile float, float* undef 783 %cmp17 = fcmp une float %val17, undef 784 br i1 %cmp17, label %18, label %17 785 %val18 = load volatile float, float* undef 786 %cmp18 = fcmp une float %val18, undef 787 br i1 %cmp18, label %19, label %18 788 %val19 = load volatile float, float* undef 789 %cmp19 = fcmp une float %val19, undef 790 br i1 %cmp19, label %20, label %19 791 %val20 = load volatile float, float* undef 792 %cmp20 = fcmp une float %val20, undef 793 br i1 %cmp20, label %21, label %20 794 %val21 = load volatile float, float* undef 795 %cmp21 = fcmp une float %val21, undef 796 br i1 %cmp21, label %22, label %21 797 %val22 = load volatile float, float* undef 798 %cmp22 = fcmp une float %val22, undef 799 br i1 %cmp22, label %23, label %22 800 %val23 = load volatile float, float* undef 801 %cmp23 = fcmp une float %val23, undef 802 br i1 %cmp23, label %24, label %23 803 %val24 = load volatile float, float* undef 804 %cmp24 = fcmp une float %val24, undef 805 br i1 %cmp24, label %25, label %24 806 %val25 = load volatile float, float* undef 807 %cmp25 = fcmp une float %val25, undef 808 br i1 %cmp25, label %26, label %25 809 %val26 = load volatile float, float* undef 810 %cmp26 = fcmp une float %val26, undef 811 br i1 %cmp26, label %27, label %26 812 %val27 = load volatile float, float* undef 813 %cmp27 = fcmp une float %val27, undef 814 br i1 %cmp27, label %28, label %27 815 %val28 = load volatile float, float* undef 816 %cmp28 = fcmp une float %val28, undef 817 br i1 %cmp28, label %29, label %28 818 %val29 = load volatile float, float* undef 819 %cmp29 = fcmp une float %val29, undef 820 br i1 %cmp29, label %30, label %29 821 %val30 = load volatile float, float* undef 822 %cmp30 = fcmp une float %val30, undef 823 br i1 %cmp30, label %31, label %30 824 %val31 = load volatile float, float* undef 825 %cmp31 = fcmp une float %val31, undef 826 br i1 %cmp31, label %32, label %31 827 %val32 = load volatile float, float* undef 828 %cmp32 = fcmp une float %val32, undef 829 br i1 %cmp32, label %33, label %32 830 %val33 = load volatile float, float* undef 831 %cmp33 = fcmp une float %val33, undef 832 br i1 %cmp33, label %34, label %33 833 %val34 = load volatile float, float* undef 834 %cmp34 = fcmp une float %val34, undef 835 br i1 %cmp34, label %35, label %34 836 %val35 = load volatile float, float* undef 837 %cmp35 = fcmp une float %val35, undef 838 br i1 %cmp35, label %36, label %35 839 %val36 = load volatile float, float* undef 840 %cmp36 = fcmp une float %val36, undef 841 br i1 %cmp36, label %37, label %36 842 %val37 = load volatile float, float* undef 843 %cmp37 = fcmp une float %val37, undef 844 br i1 %cmp37, label %38, label %37 845 %val38 = load volatile float, float* undef 846 %cmp38 = fcmp une float %val38, undef 847 br i1 %cmp38, label %39, label %38 848 %val39 = load volatile float, float* undef 849 %cmp39 = fcmp une float %val39, undef 850 br i1 %cmp39, label %40, label %39 851 %val40 = load volatile float, float* undef 852 %cmp40 = fcmp une float %val40, undef 853 br i1 %cmp40, label %41, label %40 854 %val41 = load volatile float, float* undef 855 %cmp41 = fcmp une float %val41, undef 856 br i1 %cmp41, label %42, label %41 857 %val42 = load volatile float, float* undef 858 %cmp42 = fcmp une float %val42, undef 859 br i1 %cmp42, label %43, label %42 860 %val43 = load volatile float, float* undef 861 %cmp43 = fcmp une float %val43, undef 862 br i1 %cmp43, label %44, label %43 863 %val44 = load volatile float, float* undef 864 %cmp44 = fcmp une float %val44, undef 865 br i1 %cmp44, label %45, label %44 866 %val45 = load volatile float, float* undef 867 %cmp45 = fcmp une float %val45, undef 868 br i1 %cmp45, label %46, label %45 869 %val46 = load volatile float, float* undef 870 %cmp46 = fcmp une float %val46, undef 871 br i1 %cmp46, label %47, label %46 872 %val47 = load volatile float, float* undef 873 %cmp47 = fcmp une float %val47, undef 874 br i1 %cmp47, label %48, label %47 875 %val48 = load volatile float, float* undef 876 %cmp48 = fcmp une float %val48, undef 877 br i1 %cmp48, label %49, label %48 878 %val49 = load volatile float, float* undef 879 %cmp49 = fcmp une float %val49, undef 880 br i1 %cmp49, label %50, label %49 881 %val50 = load volatile float, float* undef 882 %cmp50 = fcmp une float %val50, undef 883 br i1 %cmp50, label %51, label %50 884 %val51 = load volatile float, float* undef 885 %cmp51 = fcmp une float %val51, undef 886 br i1 %cmp51, label %52, label %51 887 %val52 = load volatile float, float* undef 888 %cmp52 = fcmp une float %val52, undef 889 br i1 %cmp52, label %53, label %52 890 %val53 = load volatile float, float* undef 891 %cmp53 = fcmp une float %val53, undef 892 br i1 %cmp53, label %54, label %53 893 %val54 = load volatile float, float* undef 894 %cmp54 = fcmp une float %val54, undef 895 br i1 %cmp54, label %55, label %54 896 %val55 = load volatile float, float* undef 897 %cmp55 = fcmp une float %val55, undef 898 br i1 %cmp55, label %56, label %55 899 %val56 = load volatile float, float* undef 900 %cmp56 = fcmp une float %val56, undef 901 br i1 %cmp56, label %57, label %56 902 %val57 = load volatile float, float* undef 903 %cmp57 = fcmp une float %val57, undef 904 br i1 %cmp57, label %58, label %57 905 %val58 = load volatile float, float* undef 906 %cmp58 = fcmp une float %val58, undef 907 br i1 %cmp58, label %59, label %58 908 %val59 = load volatile float, float* undef 909 %cmp59 = fcmp une float %val59, undef 910 br i1 %cmp59, label %60, label %59 911 %val60 = load volatile float, float* undef 912 %cmp60 = fcmp une float %val60, undef 913 br i1 %cmp60, label %61, label %60 914 %val61 = load volatile float, float* undef 915 %cmp61 = fcmp une float %val61, undef 916 br i1 %cmp61, label %62, label %61 917 %val62 = load volatile float, float* undef 918 %cmp62 = fcmp une float %val62, undef 919 br i1 %cmp62, label %63, label %62 920 %val63 = load volatile float, float* undef 921 %cmp63 = fcmp une float %val63, undef 922 br i1 %cmp63, label %64, label %63 923 %val64 = load volatile float, float* undef 924 %cmp64 = fcmp une float %val64, undef 925 br i1 %cmp64, label %65, label %64 926 927 br label %exit 928exit: 929 ret void 930} 931 932define void @benchmark_heapsort(i32 %n, double* nocapture %ra) { 933; This test case comes from the heapsort benchmark, and exemplifies several 934; important aspects to block placement in the presence of loops: 935; 1) Loop rotation needs to *ensure* that the desired exiting edge can be 936; a fallthrough. 937; 2) The exiting edge from the loop which is rotated to be laid out at the 938; bottom of the loop needs to be exiting into the nearest enclosing loop (to 939; which there is an exit). Otherwise, we force that enclosing loop into 940; strange layouts that are siginificantly less efficient, often times maing 941; it discontiguous. 942; 943; CHECK: @benchmark_heapsort 944; CHECK: %entry 945; First rotated loop top. 946; CHECK: .align 947; CHECK: %while.end 948; CHECK: %for.cond 949; CHECK: %if.then 950; CHECK: %if.else 951; CHECK: %if.end10 952; Second rotated loop top 953; CHECK: .align 954; CHECK: %if.then24 955; CHECK: %while.cond.outer 956; Third rotated loop top 957; CHECK: .align 958; CHECK: %while.cond 959; CHECK: %while.body 960; CHECK: %land.lhs.true 961; CHECK: %if.then19 962; CHECK: %if.end20 963; CHECK: %if.then8 964; CHECK: ret 965 966entry: 967 %shr = ashr i32 %n, 1 968 %add = add nsw i32 %shr, 1 969 %arrayidx3 = getelementptr inbounds double, double* %ra, i64 1 970 br label %for.cond 971 972for.cond: 973 %ir.0 = phi i32 [ %n, %entry ], [ %ir.1, %while.end ] 974 %l.0 = phi i32 [ %add, %entry ], [ %l.1, %while.end ] 975 %cmp = icmp sgt i32 %l.0, 1 976 br i1 %cmp, label %if.then, label %if.else 977 978if.then: 979 %dec = add nsw i32 %l.0, -1 980 %idxprom = sext i32 %dec to i64 981 %arrayidx = getelementptr inbounds double, double* %ra, i64 %idxprom 982 %0 = load double, double* %arrayidx, align 8 983 br label %if.end10 984 985if.else: 986 %idxprom1 = sext i32 %ir.0 to i64 987 %arrayidx2 = getelementptr inbounds double, double* %ra, i64 %idxprom1 988 %1 = load double, double* %arrayidx2, align 8 989 %2 = load double, double* %arrayidx3, align 8 990 store double %2, double* %arrayidx2, align 8 991 %dec6 = add nsw i32 %ir.0, -1 992 %cmp7 = icmp eq i32 %dec6, 1 993 br i1 %cmp7, label %if.then8, label %if.end10 994 995if.then8: 996 store double %1, double* %arrayidx3, align 8 997 ret void 998 999if.end10: 1000 %ir.1 = phi i32 [ %ir.0, %if.then ], [ %dec6, %if.else ] 1001 %l.1 = phi i32 [ %dec, %if.then ], [ %l.0, %if.else ] 1002 %rra.0 = phi double [ %0, %if.then ], [ %1, %if.else ] 1003 %add31 = add nsw i32 %ir.1, 1 1004 br label %while.cond.outer 1005 1006while.cond.outer: 1007 %j.0.ph.in = phi i32 [ %l.1, %if.end10 ], [ %j.1, %if.then24 ] 1008 %j.0.ph = shl i32 %j.0.ph.in, 1 1009 br label %while.cond 1010 1011while.cond: 1012 %j.0 = phi i32 [ %add31, %if.end20 ], [ %j.0.ph, %while.cond.outer ] 1013 %cmp11 = icmp sgt i32 %j.0, %ir.1 1014 br i1 %cmp11, label %while.end, label %while.body 1015 1016while.body: 1017 %cmp12 = icmp slt i32 %j.0, %ir.1 1018 br i1 %cmp12, label %land.lhs.true, label %if.end20 1019 1020land.lhs.true: 1021 %idxprom13 = sext i32 %j.0 to i64 1022 %arrayidx14 = getelementptr inbounds double, double* %ra, i64 %idxprom13 1023 %3 = load double, double* %arrayidx14, align 8 1024 %add15 = add nsw i32 %j.0, 1 1025 %idxprom16 = sext i32 %add15 to i64 1026 %arrayidx17 = getelementptr inbounds double, double* %ra, i64 %idxprom16 1027 %4 = load double, double* %arrayidx17, align 8 1028 %cmp18 = fcmp olt double %3, %4 1029 br i1 %cmp18, label %if.then19, label %if.end20 1030 1031if.then19: 1032 br label %if.end20 1033 1034if.end20: 1035 %j.1 = phi i32 [ %add15, %if.then19 ], [ %j.0, %land.lhs.true ], [ %j.0, %while.body ] 1036 %idxprom21 = sext i32 %j.1 to i64 1037 %arrayidx22 = getelementptr inbounds double, double* %ra, i64 %idxprom21 1038 %5 = load double, double* %arrayidx22, align 8 1039 %cmp23 = fcmp olt double %rra.0, %5 1040 br i1 %cmp23, label %if.then24, label %while.cond 1041 1042if.then24: 1043 %idxprom27 = sext i32 %j.0.ph.in to i64 1044 %arrayidx28 = getelementptr inbounds double, double* %ra, i64 %idxprom27 1045 store double %5, double* %arrayidx28, align 8 1046 br label %while.cond.outer 1047 1048while.end: 1049 %idxprom33 = sext i32 %j.0.ph.in to i64 1050 %arrayidx34 = getelementptr inbounds double, double* %ra, i64 %idxprom33 1051 store double %rra.0, double* %arrayidx34, align 8 1052 br label %for.cond 1053} 1054 1055declare void @cold_function() cold 1056 1057define i32 @test_cold_calls(i32* %a) { 1058; Test that edges to blocks post-dominated by cold calls are 1059; marked as not expected to be taken. They should be laid out 1060; at the bottom. 1061; CHECK-LABEL: test_cold_calls: 1062; CHECK: %entry 1063; CHECK: %else 1064; CHECK: %exit 1065; CHECK: %then 1066 1067entry: 1068 %gep1 = getelementptr i32, i32* %a, i32 1 1069 %val1 = load i32, i32* %gep1 1070 %cond1 = icmp ugt i32 %val1, 1 1071 br i1 %cond1, label %then, label %else 1072 1073then: 1074 call void @cold_function() 1075 br label %exit 1076 1077else: 1078 %gep2 = getelementptr i32, i32* %a, i32 2 1079 %val2 = load i32, i32* %gep2 1080 br label %exit 1081 1082exit: 1083 %ret = phi i32 [ %val1, %then ], [ %val2, %else ] 1084 ret i32 %ret 1085} 1086