1 ; RUN: llc -mtriple=i686-linux < %s | FileCheck %s 2 3 declare void @error(i32 %i, i32 %a, i32 %b) 4 5 define 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: 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 26 entry: 27 %gep1 = getelementptr i32* %a, i32 1 28 %val1 = load i32* %gep1 29 %cond1 = icmp ugt i32 %val1, 1 30 br i1 %cond1, label %then1, label %else1, !prof !0 31 32 then1: 33 call void @error(i32 %i, i32 1, i32 %b) 34 br label %else1 35 36 else1: 37 %gep2 = getelementptr i32* %a, i32 2 38 %val2 = load i32* %gep2 39 %cond2 = icmp ugt i32 %val2, 2 40 br i1 %cond2, label %then2, label %else2, !prof !0 41 42 then2: 43 call void @error(i32 %i, i32 1, i32 %b) 44 br label %else2 45 46 else2: 47 %gep3 = getelementptr i32* %a, i32 3 48 %val3 = load i32* %gep3 49 %cond3 = icmp ugt i32 %val3, 3 50 br i1 %cond3, label %then3, label %else3, !prof !0 51 52 then3: 53 call void @error(i32 %i, i32 1, i32 %b) 54 br label %else3 55 56 else3: 57 %gep4 = getelementptr i32* %a, i32 4 58 %val4 = load i32* %gep4 59 %cond4 = icmp ugt i32 %val4, 4 60 br i1 %cond4, label %then4, label %else4, !prof !0 61 62 then4: 63 call void @error(i32 %i, i32 1, i32 %b) 64 br label %else4 65 66 else4: 67 %gep5 = getelementptr i32* %a, i32 3 68 %val5 = load i32* %gep5 69 %cond5 = icmp ugt i32 %val5, 3 70 br i1 %cond5, label %then5, label %exit, !prof !0 71 72 then5: 73 call void @error(i32 %i, i32 1, i32 %b) 74 br label %exit 75 76 exit: 77 ret i32 %b 78 } 79 80 define i32 @test_loop_cold_blocks(i32 %i, i32* %a) { 81 ; Check that we sink cold loop blocks after the hot loop body. 82 ; CHECK: 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 94 entry: 95 br label %body1 96 97 body1: 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 !0 102 103 unlikely1: 104 call void @error(i32 %i, i32 1, i32 %base) 105 br label %body2 106 107 body2: 108 %unlikelycond2 = icmp sgt i32 %base, 21 109 br i1 %unlikelycond2, label %unlikely2, label %body3, !prof !0 110 111 unlikely2: 112 call void @error(i32 %i, i32 2, i32 %base) 113 br label %body3 114 115 body3: 116 %arrayidx = getelementptr inbounds i32* %a, i32 %iv 117 %0 = load 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 122 123 exit: 124 ret i32 %sum 125 } 126 127 !0 = metadata !{metadata !"branch_weights", i32 4, i32 64} 128 129 define i32 @test_loop_early_exits(i32 %i, i32* %a) { 130 ; Check that we sink early exit blocks out of loop bodies. 131 ; CHECK: 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 141 142 entry: 143 br label %body1 144 145 body1: 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 150 151 bail1: 152 ret i32 -1 153 154 body2: 155 %bailcond2 = icmp eq i32 %base, 43 156 br i1 %bailcond2, label %bail2, label %body3 157 158 bail2: 159 ret i32 -2 160 161 body3: 162 %bailcond3 = icmp eq i32 %base, 44 163 br i1 %bailcond3, label %bail3, label %body4 164 165 bail3: 166 ret i32 -3 167 168 body4: 169 %arrayidx = getelementptr inbounds i32* %a, i32 %iv 170 %0 = load 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 175 176 exit: 177 ret i32 %sum 178 } 179 180 define 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: test_loop_rotate: 184 ; CHECK: %entry 185 ; CHECK: %body1 186 ; CHECK: %body0 187 ; CHECK: %exit 188 189 entry: 190 br label %body0 191 192 body0: 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 198 199 body1: 200 %arrayidx = getelementptr inbounds i32* %a, i32 %iv 201 %0 = load i32* %arrayidx 202 %sum = add nsw i32 %0, %base 203 %bailcond1 = icmp eq i32 %sum, 42 204 br label %body0 205 206 exit: 207 ret i32 %base 208 } 209 210 define 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: test_no_loop_rotate: 214 ; CHECK: %entry 215 ; CHECK: %body0 216 ; CHECK: %body1 217 ; CHECK: %exit 218 219 entry: 220 br label %body0 221 222 body0: 223 %iv = phi i32 [ 0, %entry ], [ %next, %body1 ] 224 %base = phi i32 [ 0, %entry ], [ %sum, %body1 ] 225 %arrayidx = getelementptr inbounds i32* %a, i32 %iv 226 %0 = load i32* %arrayidx 227 %sum = add nsw i32 %0, %base 228 %bailcond1 = icmp eq i32 %sum, 42 229 br i1 %bailcond1, label %exit, label %body1 230 231 body1: 232 %next = add i32 %iv, 1 233 %exitcond = icmp eq i32 %next, %i 234 br i1 %exitcond, label %exit, label %body0 235 236 exit: 237 ret i32 %base 238 } 239 240 define void @test_loop_rotate_reversed_blocks() { 241 ; This test case (greatly reduced from an Olden bencmark) ensures that the loop 242 ; rotate implementation doesn't assume that loops are laid out in a particular 243 ; order. The first loop will get split into two basic blocks, with the loop 244 ; header coming after the loop latch. 245 ; 246 ; CHECK: test_loop_rotate_reversed_blocks 247 ; CHECK: %entry 248 ; Look for a jump into the middle of the loop, and no branches mid-way. 249 ; CHECK: jmp 250 ; CHECK: %loop1 251 ; CHECK-NOT: j{{\w*}} .LBB{{.*}} 252 ; CHECK: %loop1 253 ; CHECK: je 254 255 entry: 256 %cond1 = load volatile i1* undef 257 br i1 %cond1, label %loop2.preheader, label %loop1 258 259 loop1: 260 call i32 @f() 261 %cond2 = load volatile i1* undef 262 br i1 %cond2, label %loop2.preheader, label %loop1 263 264 loop2.preheader: 265 call i32 @f() 266 %cond3 = load volatile i1* undef 267 br i1 %cond3, label %exit, label %loop2 268 269 loop2: 270 call i32 @f() 271 %cond4 = load volatile i1* undef 272 br i1 %cond4, label %exit, label %loop2 273 274 exit: 275 ret void 276 } 277 278 define i32 @test_loop_align(i32 %i, i32* %a) { 279 ; Check that we provide basic loop body alignment with the block placement 280 ; pass. 281 ; CHECK: test_loop_align: 282 ; CHECK: %entry 283 ; CHECK: .align [[ALIGN:[0-9]+]], 284 ; CHECK-NEXT: %body 285 ; CHECK: %exit 286 287 entry: 288 br label %body 289 290 body: 291 %iv = phi i32 [ 0, %entry ], [ %next, %body ] 292 %base = phi i32 [ 0, %entry ], [ %sum, %body ] 293 %arrayidx = getelementptr inbounds i32* %a, i32 %iv 294 %0 = load i32* %arrayidx 295 %sum = add nsw i32 %0, %base 296 %next = add i32 %iv, 1 297 %exitcond = icmp eq i32 %next, %i 298 br i1 %exitcond, label %exit, label %body 299 300 exit: 301 ret i32 %sum 302 } 303 304 define i32 @test_nested_loop_align(i32 %i, i32* %a, i32* %b) { 305 ; Check that we provide nested loop body alignment. 306 ; CHECK: test_nested_loop_align: 307 ; CHECK: %entry 308 ; CHECK: .align [[ALIGN]], 309 ; CHECK-NEXT: %loop.body.1 310 ; CHECK: .align [[ALIGN]], 311 ; CHECK-NEXT: %inner.loop.body 312 ; CHECK-NOT: .align 313 ; CHECK: %exit 314 315 entry: 316 br label %loop.body.1 317 318 loop.body.1: 319 %iv = phi i32 [ 0, %entry ], [ %next, %loop.body.2 ] 320 %arrayidx = getelementptr inbounds i32* %a, i32 %iv 321 %bidx = load i32* %arrayidx 322 br label %inner.loop.body 323 324 inner.loop.body: 325 %inner.iv = phi i32 [ 0, %loop.body.1 ], [ %inner.next, %inner.loop.body ] 326 %base = phi i32 [ 0, %loop.body.1 ], [ %sum, %inner.loop.body ] 327 %scaled_idx = mul i32 %bidx, %iv 328 %inner.arrayidx = getelementptr inbounds i32* %b, i32 %scaled_idx 329 %0 = load i32* %inner.arrayidx 330 %sum = add nsw i32 %0, %base 331 %inner.next = add i32 %iv, 1 332 %inner.exitcond = icmp eq i32 %inner.next, %i 333 br i1 %inner.exitcond, label %loop.body.2, label %inner.loop.body 334 335 loop.body.2: 336 %next = add i32 %iv, 1 337 %exitcond = icmp eq i32 %next, %i 338 br i1 %exitcond, label %exit, label %loop.body.1 339 340 exit: 341 ret i32 %sum 342 } 343 344 define void @unnatural_cfg1() { 345 ; Test that we can handle a loop with an inner unnatural loop at the end of 346 ; a function. This is a gross CFG reduced out of the single source GCC. 347 ; CHECK: unnatural_cfg1 348 ; CHECK: %entry 349 ; CHECK: %loop.body1 350 ; CHECK: %loop.body2 351 ; CHECK: %loop.body3 352 353 entry: 354 br label %loop.header 355 356 loop.header: 357 br label %loop.body1 358 359 loop.body1: 360 br i1 undef, label %loop.body3, label %loop.body2 361 362 loop.body2: 363 %ptr = load i32** undef, align 4 364 br label %loop.body3 365 366 loop.body3: 367 %myptr = phi i32* [ %ptr2, %loop.body5 ], [ %ptr, %loop.body2 ], [ undef, %loop.body1 ] 368 %bcmyptr = bitcast i32* %myptr to i32* 369 %val = load i32* %bcmyptr, align 4 370 %comp = icmp eq i32 %val, 48 371 br i1 %comp, label %loop.body4, label %loop.body5 372 373 loop.body4: 374 br i1 undef, label %loop.header, label %loop.body5 375 376 loop.body5: 377 %ptr2 = load i32** undef, align 4 378 br label %loop.body3 379 } 380 381 define void @unnatural_cfg2() { 382 ; Test that we can handle a loop with a nested natural loop *and* an unnatural 383 ; loop. This was reduced from a crash on block placement when run over 384 ; single-source GCC. 385 ; CHECK: unnatural_cfg2 386 ; CHECK: %entry 387 ; CHECK: %loop.body1 388 ; CHECK: %loop.body2 389 ; CHECK: %loop.body3 390 ; CHECK: %loop.inner1.begin 391 ; The end block is folded with %loop.body3... 392 ; CHECK-NOT: %loop.inner1.end 393 ; CHECK: %loop.body4 394 ; CHECK: %loop.inner2.begin 395 ; The loop.inner2.end block is folded 396 ; CHECK: %loop.header 397 ; CHECK: %bail 398 399 entry: 400 br label %loop.header 401 402 loop.header: 403 %comp0 = icmp eq i32* undef, null 404 br i1 %comp0, label %bail, label %loop.body1 405 406 loop.body1: 407 %val0 = load i32** undef, align 4 408 br i1 undef, label %loop.body2, label %loop.inner1.begin 409 410 loop.body2: 411 br i1 undef, label %loop.body4, label %loop.body3 412 413 loop.body3: 414 %ptr1 = getelementptr inbounds i32* %val0, i32 0 415 %castptr1 = bitcast i32* %ptr1 to i32** 416 %val1 = load i32** %castptr1, align 4 417 br label %loop.inner1.begin 418 419 loop.inner1.begin: 420 %valphi = phi i32* [ %val2, %loop.inner1.end ], [ %val1, %loop.body3 ], [ %val0, %loop.body1 ] 421 %castval = bitcast i32* %valphi to i32* 422 %comp1 = icmp eq i32 undef, 48 423 br i1 %comp1, label %loop.inner1.end, label %loop.body4 424 425 loop.inner1.end: 426 %ptr2 = getelementptr inbounds i32* %valphi, i32 0 427 %castptr2 = bitcast i32* %ptr2 to i32** 428 %val2 = load i32** %castptr2, align 4 429 br label %loop.inner1.begin 430 431 loop.body4.dead: 432 br label %loop.body4 433 434 loop.body4: 435 %comp2 = icmp ult i32 undef, 3 436 br i1 %comp2, label %loop.inner2.begin, label %loop.end 437 438 loop.inner2.begin: 439 br i1 false, label %loop.end, label %loop.inner2.end 440 441 loop.inner2.end: 442 %comp3 = icmp eq i32 undef, 1769472 443 br i1 %comp3, label %loop.end, label %loop.inner2.begin 444 445 loop.end: 446 br label %loop.header 447 448 bail: 449 unreachable 450 } 451 452 define i32 @problematic_switch() { 453 ; This function's CFG caused overlow in the machine branch probability 454 ; calculation, triggering asserts. Make sure we don't crash on it. 455 ; CHECK: problematic_switch 456 457 entry: 458 switch i32 undef, label %exit [ 459 i32 879, label %bogus 460 i32 877, label %step 461 i32 876, label %step 462 i32 875, label %step 463 i32 874, label %step 464 i32 873, label %step 465 i32 872, label %step 466 i32 868, label %step 467 i32 867, label %step 468 i32 866, label %step 469 i32 861, label %step 470 i32 860, label %step 471 i32 856, label %step 472 i32 855, label %step 473 i32 854, label %step 474 i32 831, label %step 475 i32 830, label %step 476 i32 829, label %step 477 i32 828, label %step 478 i32 815, label %step 479 i32 814, label %step 480 i32 811, label %step 481 i32 806, label %step 482 i32 805, label %step 483 i32 804, label %step 484 i32 803, label %step 485 i32 802, label %step 486 i32 801, label %step 487 i32 800, label %step 488 i32 799, label %step 489 i32 798, label %step 490 i32 797, label %step 491 i32 796, label %step 492 i32 795, label %step 493 ] 494 bogus: 495 unreachable 496 step: 497 br label %exit 498 exit: 499 %merge = phi i32 [ 3, %step ], [ 6, %entry ] 500 ret i32 %merge 501 } 502 503 define void @fpcmp_unanalyzable_branch(i1 %cond) { 504 ; This function's CFG contains an unanalyzable branch that is likely to be 505 ; split due to having a different high-probability predecessor. 506 ; CHECK: fpcmp_unanalyzable_branch 507 ; CHECK: %entry 508 ; CHECK: %exit 509 ; CHECK-NOT: %if.then 510 ; CHECK-NOT: %if.end 511 ; CHECK-NOT: jne 512 ; CHECK-NOT: jnp 513 ; CHECK: jne 514 ; CHECK-NEXT: jnp 515 ; CHECK-NEXT: %if.then 516 517 entry: 518 ; Note that this branch must be strongly biased toward 519 ; 'entry.if.then_crit_edge' to ensure that we would try to form a chain for 520 ; 'entry' -> 'entry.if.then_crit_edge' -> 'if.then'. It is the last edge in that 521 ; chain which would violate the unanalyzable branch in 'exit', but we won't even 522 ; try this trick unless 'if.then' is believed to almost always be reached from 523 ; 'entry.if.then_crit_edge'. 524 br i1 %cond, label %entry.if.then_crit_edge, label %lor.lhs.false, !prof !1 525 526 entry.if.then_crit_edge: 527 %.pre14 = load i8* undef, align 1, !tbaa !0 528 br label %if.then 529 530 lor.lhs.false: 531 br i1 undef, label %if.end, label %exit 532 533 exit: 534 %cmp.i = fcmp une double 0.000000e+00, undef 535 br i1 %cmp.i, label %if.then, label %if.end 536 537 if.then: 538 %0 = phi i8 [ %.pre14, %entry.if.then_crit_edge ], [ undef, %exit ] 539 %1 = and i8 %0, 1 540 store i8 %1, i8* undef, align 4, !tbaa !0 541 br label %if.end 542 543 if.end: 544 ret void 545 } 546 547 !1 = metadata !{metadata !"branch_weights", i32 1000, i32 1} 548 549 declare i32 @f() 550 declare i32 @g() 551 declare i32 @h(i32 %x) 552 553 define i32 @test_global_cfg_break_profitability() { 554 ; Check that our metrics for the profitability of a CFG break are global rather 555 ; than local. A successor may be very hot, but if the current block isn't, it 556 ; doesn't matter. Within this test the 'then' block is slightly warmer than the 557 ; 'else' block, but not nearly enough to merit merging it with the exit block 558 ; even though the probability of 'then' branching to the 'exit' block is very 559 ; high. 560 ; CHECK: test_global_cfg_break_profitability 561 ; CHECK: calll {{_?}}f 562 ; CHECK: calll {{_?}}g 563 ; CHECK: calll {{_?}}h 564 ; CHECK: ret 565 566 entry: 567 br i1 undef, label %then, label %else, !prof !2 568 569 then: 570 %then.result = call i32 @f() 571 br label %exit 572 573 else: 574 %else.result = call i32 @g() 575 br label %exit 576 577 exit: 578 %result = phi i32 [ %then.result, %then ], [ %else.result, %else ] 579 %result2 = call i32 @h(i32 %result) 580 ret i32 %result 581 } 582 583 !2 = metadata !{metadata !"branch_weights", i32 3, i32 1} 584 585 declare i32 @__gxx_personality_v0(...) 586 587 define void @test_eh_lpad_successor() { 588 ; Some times the landing pad ends up as the first successor of an invoke block. 589 ; When this happens, a strange result used to fall out of updateTerminators: we 590 ; didn't correctly locate the fallthrough successor, assuming blindly that the 591 ; first one was the fallthrough successor. As a result, we would add an 592 ; erroneous jump to the landing pad thinking *that* was the default successor. 593 ; CHECK: test_eh_lpad_successor 594 ; CHECK: %entry 595 ; CHECK-NOT: jmp 596 ; CHECK: %loop 597 598 entry: 599 invoke i32 @f() to label %preheader unwind label %lpad 600 601 preheader: 602 br label %loop 603 604 lpad: 605 %lpad.val = landingpad { i8*, i32 } personality i8* bitcast (i32 (...)* @__gxx_personality_v0 to i8*) 606 cleanup 607 resume { i8*, i32 } %lpad.val 608 609 loop: 610 br label %loop 611 } 612 613 declare void @fake_throw() noreturn 614 615 define void @test_eh_throw() { 616 ; For blocks containing a 'throw' (or similar functionality), we have 617 ; a no-return invoke. In this case, only EH successors will exist, and 618 ; fallthrough simply won't occur. Make sure we don't crash trying to update 619 ; terminators for such constructs. 620 ; 621 ; CHECK: test_eh_throw 622 ; CHECK: %entry 623 ; CHECK: %cleanup 624 625 entry: 626 invoke void @fake_throw() to label %continue unwind label %cleanup 627 628 continue: 629 unreachable 630 631 cleanup: 632 %0 = landingpad { i8*, i32 } personality i8* bitcast (i32 (...)* @__gxx_personality_v0 to i8*) 633 cleanup 634 unreachable 635 } 636 637 define void @test_unnatural_cfg_backwards_inner_loop() { 638 ; Test that when we encounter an unnatural CFG structure after having formed 639 ; a chain for an inner loop which happened to be laid out backwards we don't 640 ; attempt to merge onto the wrong end of the inner loop just because we find it 641 ; first. This was reduced from a crasher in GCC's single source. 642 ; 643 ; CHECK: test_unnatural_cfg_backwards_inner_loop 644 ; CHECK: %entry 645 ; CHECK: [[BODY:# BB#[0-9]+]]: 646 ; CHECK: %loop2b 647 ; CHECK: %loop1 648 ; CHECK: %loop2a 649 650 entry: 651 br i1 undef, label %loop2a, label %body 652 653 body: 654 br label %loop2a 655 656 loop1: 657 %next.load = load i32** undef 658 br i1 %comp.a, label %loop2a, label %loop2b 659 660 loop2a: 661 %var = phi i32* [ null, %entry ], [ null, %body ], [ %next.phi, %loop1 ] 662 %next.var = phi i32* [ null, %entry ], [ undef, %body ], [ %next.load, %loop1 ] 663 %comp.a = icmp eq i32* %var, null 664 br label %loop3 665 666 loop2b: 667 %gep = getelementptr inbounds i32* %var.phi, i32 0 668 %next.ptr = bitcast i32* %gep to i32** 669 store i32* %next.phi, i32** %next.ptr 670 br label %loop3 671 672 loop3: 673 %var.phi = phi i32* [ %next.phi, %loop2b ], [ %var, %loop2a ] 674 %next.phi = phi i32* [ %next.load, %loop2b ], [ %next.var, %loop2a ] 675 br label %loop1 676 } 677 678 define void @unanalyzable_branch_to_loop_header() { 679 ; Ensure that we can handle unanalyzable branches into loop headers. We 680 ; pre-form chains for unanalyzable branches, and will find the tail end of that 681 ; at the start of the loop. This function uses floating point comparison 682 ; fallthrough because that happens to always produce unanalyzable branches on 683 ; x86. 684 ; 685 ; CHECK: unanalyzable_branch_to_loop_header 686 ; CHECK: %entry 687 ; CHECK: %loop 688 ; CHECK: %exit 689 690 entry: 691 %cmp = fcmp une double 0.000000e+00, undef 692 br i1 %cmp, label %loop, label %exit 693 694 loop: 695 %cond = icmp eq i8 undef, 42 696 br i1 %cond, label %exit, label %loop 697 698 exit: 699 ret void 700 } 701 702 define void @unanalyzable_branch_to_best_succ(i1 %cond) { 703 ; Ensure that we can handle unanalyzable branches where the destination block 704 ; gets selected as the optimal sucessor to merge. 705 ; 706 ; CHECK: unanalyzable_branch_to_best_succ 707 ; CHECK: %entry 708 ; CHECK: %foo 709 ; CHECK: %bar 710 ; CHECK: %exit 711 712 entry: 713 ; Bias this branch toward bar to ensure we form that chain. 714 br i1 %cond, label %bar, label %foo, !prof !1 715 716 foo: 717 %cmp = fcmp une double 0.000000e+00, undef 718 br i1 %cmp, label %bar, label %exit 719 720 bar: 721 call i32 @f() 722 br label %exit 723 724 exit: 725 ret void 726 } 727 728 define void @unanalyzable_branch_to_free_block(float %x) { 729 ; Ensure that we can handle unanalyzable branches where the destination block 730 ; gets selected as the best free block in the CFG. 731 ; 732 ; CHECK: unanalyzable_branch_to_free_block 733 ; CHECK: %entry 734 ; CHECK: %a 735 ; CHECK: %b 736 ; CHECK: %c 737 ; CHECK: %exit 738 739 entry: 740 br i1 undef, label %a, label %b 741 742 a: 743 call i32 @f() 744 br label %c 745 746 b: 747 %cmp = fcmp une float %x, undef 748 br i1 %cmp, label %c, label %exit 749 750 c: 751 call i32 @g() 752 br label %exit 753 754 exit: 755 ret void 756 } 757 758 define void @many_unanalyzable_branches() { 759 ; Ensure that we don't crash as we're building up many unanalyzable branches, 760 ; blocks, and loops. 761 ; 762 ; CHECK: many_unanalyzable_branches 763 ; CHECK: %entry 764 ; CHECK: %exit 765 766 entry: 767 br label %0 768 769 %val0 = load volatile float* undef 770 %cmp0 = fcmp une float %val0, undef 771 br i1 %cmp0, label %1, label %0 772 %val1 = load volatile float* undef 773 %cmp1 = fcmp une float %val1, undef 774 br i1 %cmp1, label %2, label %1 775 %val2 = load volatile float* undef 776 %cmp2 = fcmp une float %val2, undef 777 br i1 %cmp2, label %3, label %2 778 %val3 = load volatile float* undef 779 %cmp3 = fcmp une float %val3, undef 780 br i1 %cmp3, label %4, label %3 781 %val4 = load volatile float* undef 782 %cmp4 = fcmp une float %val4, undef 783 br i1 %cmp4, label %5, label %4 784 %val5 = load volatile float* undef 785 %cmp5 = fcmp une float %val5, undef 786 br i1 %cmp5, label %6, label %5 787 %val6 = load volatile float* undef 788 %cmp6 = fcmp une float %val6, undef 789 br i1 %cmp6, label %7, label %6 790 %val7 = load volatile float* undef 791 %cmp7 = fcmp une float %val7, undef 792 br i1 %cmp7, label %8, label %7 793 %val8 = load volatile float* undef 794 %cmp8 = fcmp une float %val8, undef 795 br i1 %cmp8, label %9, label %8 796 %val9 = load volatile float* undef 797 %cmp9 = fcmp une float %val9, undef 798 br i1 %cmp9, label %10, label %9 799 %val10 = load volatile float* undef 800 %cmp10 = fcmp une float %val10, undef 801 br i1 %cmp10, label %11, label %10 802 %val11 = load volatile float* undef 803 %cmp11 = fcmp une float %val11, undef 804 br i1 %cmp11, label %12, label %11 805 %val12 = load volatile float* undef 806 %cmp12 = fcmp une float %val12, undef 807 br i1 %cmp12, label %13, label %12 808 %val13 = load volatile float* undef 809 %cmp13 = fcmp une float %val13, undef 810 br i1 %cmp13, label %14, label %13 811 %val14 = load volatile float* undef 812 %cmp14 = fcmp une float %val14, undef 813 br i1 %cmp14, label %15, label %14 814 %val15 = load volatile float* undef 815 %cmp15 = fcmp une float %val15, undef 816 br i1 %cmp15, label %16, label %15 817 %val16 = load volatile float* undef 818 %cmp16 = fcmp une float %val16, undef 819 br i1 %cmp16, label %17, label %16 820 %val17 = load volatile float* undef 821 %cmp17 = fcmp une float %val17, undef 822 br i1 %cmp17, label %18, label %17 823 %val18 = load volatile float* undef 824 %cmp18 = fcmp une float %val18, undef 825 br i1 %cmp18, label %19, label %18 826 %val19 = load volatile float* undef 827 %cmp19 = fcmp une float %val19, undef 828 br i1 %cmp19, label %20, label %19 829 %val20 = load volatile float* undef 830 %cmp20 = fcmp une float %val20, undef 831 br i1 %cmp20, label %21, label %20 832 %val21 = load volatile float* undef 833 %cmp21 = fcmp une float %val21, undef 834 br i1 %cmp21, label %22, label %21 835 %val22 = load volatile float* undef 836 %cmp22 = fcmp une float %val22, undef 837 br i1 %cmp22, label %23, label %22 838 %val23 = load volatile float* undef 839 %cmp23 = fcmp une float %val23, undef 840 br i1 %cmp23, label %24, label %23 841 %val24 = load volatile float* undef 842 %cmp24 = fcmp une float %val24, undef 843 br i1 %cmp24, label %25, label %24 844 %val25 = load volatile float* undef 845 %cmp25 = fcmp une float %val25, undef 846 br i1 %cmp25, label %26, label %25 847 %val26 = load volatile float* undef 848 %cmp26 = fcmp une float %val26, undef 849 br i1 %cmp26, label %27, label %26 850 %val27 = load volatile float* undef 851 %cmp27 = fcmp une float %val27, undef 852 br i1 %cmp27, label %28, label %27 853 %val28 = load volatile float* undef 854 %cmp28 = fcmp une float %val28, undef 855 br i1 %cmp28, label %29, label %28 856 %val29 = load volatile float* undef 857 %cmp29 = fcmp une float %val29, undef 858 br i1 %cmp29, label %30, label %29 859 %val30 = load volatile float* undef 860 %cmp30 = fcmp une float %val30, undef 861 br i1 %cmp30, label %31, label %30 862 %val31 = load volatile float* undef 863 %cmp31 = fcmp une float %val31, undef 864 br i1 %cmp31, label %32, label %31 865 %val32 = load volatile float* undef 866 %cmp32 = fcmp une float %val32, undef 867 br i1 %cmp32, label %33, label %32 868 %val33 = load volatile float* undef 869 %cmp33 = fcmp une float %val33, undef 870 br i1 %cmp33, label %34, label %33 871 %val34 = load volatile float* undef 872 %cmp34 = fcmp une float %val34, undef 873 br i1 %cmp34, label %35, label %34 874 %val35 = load volatile float* undef 875 %cmp35 = fcmp une float %val35, undef 876 br i1 %cmp35, label %36, label %35 877 %val36 = load volatile float* undef 878 %cmp36 = fcmp une float %val36, undef 879 br i1 %cmp36, label %37, label %36 880 %val37 = load volatile float* undef 881 %cmp37 = fcmp une float %val37, undef 882 br i1 %cmp37, label %38, label %37 883 %val38 = load volatile float* undef 884 %cmp38 = fcmp une float %val38, undef 885 br i1 %cmp38, label %39, label %38 886 %val39 = load volatile float* undef 887 %cmp39 = fcmp une float %val39, undef 888 br i1 %cmp39, label %40, label %39 889 %val40 = load volatile float* undef 890 %cmp40 = fcmp une float %val40, undef 891 br i1 %cmp40, label %41, label %40 892 %val41 = load volatile float* undef 893 %cmp41 = fcmp une float %val41, undef 894 br i1 %cmp41, label %42, label %41 895 %val42 = load volatile float* undef 896 %cmp42 = fcmp une float %val42, undef 897 br i1 %cmp42, label %43, label %42 898 %val43 = load volatile float* undef 899 %cmp43 = fcmp une float %val43, undef 900 br i1 %cmp43, label %44, label %43 901 %val44 = load volatile float* undef 902 %cmp44 = fcmp une float %val44, undef 903 br i1 %cmp44, label %45, label %44 904 %val45 = load volatile float* undef 905 %cmp45 = fcmp une float %val45, undef 906 br i1 %cmp45, label %46, label %45 907 %val46 = load volatile float* undef 908 %cmp46 = fcmp une float %val46, undef 909 br i1 %cmp46, label %47, label %46 910 %val47 = load volatile float* undef 911 %cmp47 = fcmp une float %val47, undef 912 br i1 %cmp47, label %48, label %47 913 %val48 = load volatile float* undef 914 %cmp48 = fcmp une float %val48, undef 915 br i1 %cmp48, label %49, label %48 916 %val49 = load volatile float* undef 917 %cmp49 = fcmp une float %val49, undef 918 br i1 %cmp49, label %50, label %49 919 %val50 = load volatile float* undef 920 %cmp50 = fcmp une float %val50, undef 921 br i1 %cmp50, label %51, label %50 922 %val51 = load volatile float* undef 923 %cmp51 = fcmp une float %val51, undef 924 br i1 %cmp51, label %52, label %51 925 %val52 = load volatile float* undef 926 %cmp52 = fcmp une float %val52, undef 927 br i1 %cmp52, label %53, label %52 928 %val53 = load volatile float* undef 929 %cmp53 = fcmp une float %val53, undef 930 br i1 %cmp53, label %54, label %53 931 %val54 = load volatile float* undef 932 %cmp54 = fcmp une float %val54, undef 933 br i1 %cmp54, label %55, label %54 934 %val55 = load volatile float* undef 935 %cmp55 = fcmp une float %val55, undef 936 br i1 %cmp55, label %56, label %55 937 %val56 = load volatile float* undef 938 %cmp56 = fcmp une float %val56, undef 939 br i1 %cmp56, label %57, label %56 940 %val57 = load volatile float* undef 941 %cmp57 = fcmp une float %val57, undef 942 br i1 %cmp57, label %58, label %57 943 %val58 = load volatile float* undef 944 %cmp58 = fcmp une float %val58, undef 945 br i1 %cmp58, label %59, label %58 946 %val59 = load volatile float* undef 947 %cmp59 = fcmp une float %val59, undef 948 br i1 %cmp59, label %60, label %59 949 %val60 = load volatile float* undef 950 %cmp60 = fcmp une float %val60, undef 951 br i1 %cmp60, label %61, label %60 952 %val61 = load volatile float* undef 953 %cmp61 = fcmp une float %val61, undef 954 br i1 %cmp61, label %62, label %61 955 %val62 = load volatile float* undef 956 %cmp62 = fcmp une float %val62, undef 957 br i1 %cmp62, label %63, label %62 958 %val63 = load volatile float* undef 959 %cmp63 = fcmp une float %val63, undef 960 br i1 %cmp63, label %64, label %63 961 %val64 = load volatile float* undef 962 %cmp64 = fcmp une float %val64, undef 963 br i1 %cmp64, label %65, label %64 964 965 br label %exit 966 exit: 967 ret void 968 } 969 970 define void @benchmark_heapsort(i32 %n, double* nocapture %ra) { 971 ; This test case comes from the heapsort benchmark, and exemplifies several 972 ; important aspects to block placement in the presence of loops: 973 ; 1) Loop rotation needs to *ensure* that the desired exiting edge can be 974 ; a fallthrough. 975 ; 2) The exiting edge from the loop which is rotated to be laid out at the 976 ; bottom of the loop needs to be exiting into the nearest enclosing loop (to 977 ; which there is an exit). Otherwise, we force that enclosing loop into 978 ; strange layouts that are siginificantly less efficient, often times maing 979 ; it discontiguous. 980 ; 981 ; CHECK: @benchmark_heapsort 982 ; CHECK: %entry 983 ; First rotated loop top. 984 ; CHECK: .align 985 ; CHECK: %while.end 986 ; CHECK: %for.cond 987 ; CHECK: %if.then 988 ; CHECK: %if.else 989 ; CHECK: %if.end10 990 ; Second rotated loop top 991 ; CHECK: .align 992 ; CHECK: %if.then24 993 ; CHECK: %while.cond.outer 994 ; Third rotated loop top 995 ; CHECK: .align 996 ; CHECK: %while.cond 997 ; CHECK: %while.body 998 ; CHECK: %land.lhs.true 999 ; CHECK: %if.then19 1000 ; CHECK: %if.then19 1001 ; CHECK: %if.then8 1002 ; CHECK: ret 1003 1004 entry: 1005 %shr = ashr i32 %n, 1 1006 %add = add nsw i32 %shr, 1 1007 %arrayidx3 = getelementptr inbounds double* %ra, i64 1 1008 br label %for.cond 1009 1010 for.cond: 1011 %ir.0 = phi i32 [ %n, %entry ], [ %ir.1, %while.end ] 1012 %l.0 = phi i32 [ %add, %entry ], [ %l.1, %while.end ] 1013 %cmp = icmp sgt i32 %l.0, 1 1014 br i1 %cmp, label %if.then, label %if.else 1015 1016 if.then: 1017 %dec = add nsw i32 %l.0, -1 1018 %idxprom = sext i32 %dec to i64 1019 %arrayidx = getelementptr inbounds double* %ra, i64 %idxprom 1020 %0 = load double* %arrayidx, align 8 1021 br label %if.end10 1022 1023 if.else: 1024 %idxprom1 = sext i32 %ir.0 to i64 1025 %arrayidx2 = getelementptr inbounds double* %ra, i64 %idxprom1 1026 %1 = load double* %arrayidx2, align 8 1027 %2 = load double* %arrayidx3, align 8 1028 store double %2, double* %arrayidx2, align 8 1029 %dec6 = add nsw i32 %ir.0, -1 1030 %cmp7 = icmp eq i32 %dec6, 1 1031 br i1 %cmp7, label %if.then8, label %if.end10 1032 1033 if.then8: 1034 store double %1, double* %arrayidx3, align 8 1035 ret void 1036 1037 if.end10: 1038 %ir.1 = phi i32 [ %ir.0, %if.then ], [ %dec6, %if.else ] 1039 %l.1 = phi i32 [ %dec, %if.then ], [ %l.0, %if.else ] 1040 %rra.0 = phi double [ %0, %if.then ], [ %1, %if.else ] 1041 %add31 = add nsw i32 %ir.1, 1 1042 br label %while.cond.outer 1043 1044 while.cond.outer: 1045 %j.0.ph.in = phi i32 [ %l.1, %if.end10 ], [ %j.1, %if.then24 ] 1046 %j.0.ph = shl i32 %j.0.ph.in, 1 1047 br label %while.cond 1048 1049 while.cond: 1050 %j.0 = phi i32 [ %add31, %if.end20 ], [ %j.0.ph, %while.cond.outer ] 1051 %cmp11 = icmp sgt i32 %j.0, %ir.1 1052 br i1 %cmp11, label %while.end, label %while.body 1053 1054 while.body: 1055 %cmp12 = icmp slt i32 %j.0, %ir.1 1056 br i1 %cmp12, label %land.lhs.true, label %if.end20 1057 1058 land.lhs.true: 1059 %idxprom13 = sext i32 %j.0 to i64 1060 %arrayidx14 = getelementptr inbounds double* %ra, i64 %idxprom13 1061 %3 = load double* %arrayidx14, align 8 1062 %add15 = add nsw i32 %j.0, 1 1063 %idxprom16 = sext i32 %add15 to i64 1064 %arrayidx17 = getelementptr inbounds double* %ra, i64 %idxprom16 1065 %4 = load double* %arrayidx17, align 8 1066 %cmp18 = fcmp olt double %3, %4 1067 br i1 %cmp18, label %if.then19, label %if.end20 1068 1069 if.then19: 1070 br label %if.end20 1071 1072 if.end20: 1073 %j.1 = phi i32 [ %add15, %if.then19 ], [ %j.0, %land.lhs.true ], [ %j.0, %while.body ] 1074 %idxprom21 = sext i32 %j.1 to i64 1075 %arrayidx22 = getelementptr inbounds double* %ra, i64 %idxprom21 1076 %5 = load double* %arrayidx22, align 8 1077 %cmp23 = fcmp olt double %rra.0, %5 1078 br i1 %cmp23, label %if.then24, label %while.cond 1079 1080 if.then24: 1081 %idxprom27 = sext i32 %j.0.ph.in to i64 1082 %arrayidx28 = getelementptr inbounds double* %ra, i64 %idxprom27 1083 store double %5, double* %arrayidx28, align 8 1084 br label %while.cond.outer 1085 1086 while.end: 1087 %idxprom33 = sext i32 %j.0.ph.in to i64 1088 %arrayidx34 = getelementptr inbounds double* %ra, i64 %idxprom33 1089 store double %rra.0, double* %arrayidx34, align 8 1090 br label %for.cond 1091 } 1092
