1 ; RUN: llc -mtriple=i686-linux -pre-RA-sched=source < %s | FileCheck %s
2 ; RUN: opt -disable-output -passes=debugify < %s
4 declare void @error(i32 %i, i32 %a, i32 %b)
6 define i32 @test_ifchains(i32 %i, ptr %a, i32 %b) {
7 ; Test a chain of ifs, where the block guarded by the if is error handling code
8 ; that is not expected to run.
9 ; CHECK-LABEL: test_ifchains:
28 %gep1 = getelementptr i32, ptr %a, i32 1
29 %val1 = load i32, ptr %gep1
30 %cond1 = icmp ugt i32 %val1, 1
31 br i1 %cond1, label %then1, label %else1, !prof !0
34 call void @error(i32 %i, i32 1, i32 %b)
38 %gep2 = getelementptr i32, ptr %a, i32 2
39 %val2 = load i32, ptr %gep2
40 %cond2 = icmp ugt i32 %val2, 2
41 br i1 %cond2, label %then2, label %else2, !prof !0
44 call void @error(i32 %i, i32 1, i32 %b)
48 %gep3 = getelementptr i32, ptr %a, i32 3
49 %val3 = load i32, ptr %gep3
50 %cond3 = icmp ugt i32 %val3, 3
51 br i1 %cond3, label %then3, label %else3, !prof !0
54 call void @error(i32 %i, i32 1, i32 %b)
58 %gep4 = getelementptr i32, ptr %a, i32 4
59 %val4 = load i32, ptr %gep4
60 %cond4 = icmp ugt i32 %val4, 4
61 br i1 %cond4, label %then4, label %else4, !prof !0
64 call void @error(i32 %i, i32 1, i32 %b)
68 %gep5 = getelementptr i32, ptr %a, i32 3
69 %val5 = load i32, ptr %gep5
70 %cond5 = icmp ugt i32 %val5, 3
71 br i1 %cond5, label %then5, label %exit, !prof !0
74 call void @error(i32 %i, i32 1, i32 %b)
81 define i32 @test_loop_cold_blocks(i32 %i, ptr %a) {
82 ; Check that we sink cold loop blocks after the hot loop body.
83 ; CHECK-LABEL: test_loop_cold_blocks:
99 %iv = phi i32 [ 0, %entry ], [ %next, %body3 ]
100 %base = phi i32 [ 0, %entry ], [ %sum, %body3 ]
101 %unlikelycond1 = icmp slt i32 %base, 42
102 br i1 %unlikelycond1, label %unlikely1, label %body2, !prof !0
105 call void @error(i32 %i, i32 1, i32 %base)
109 %unlikelycond2 = icmp sgt i32 %base, 21
110 br i1 %unlikelycond2, label %unlikely2, label %body3, !prof !0
113 call void @error(i32 %i, i32 2, i32 %base)
117 %arrayidx = getelementptr inbounds i32, ptr %a, i32 %iv
118 %0 = load i32, ptr %arrayidx
119 %sum = add nsw i32 %0, %base
120 %next = add i32 %iv, 1
121 %exitcond = icmp eq i32 %next, %i
122 br i1 %exitcond, label %exit, label %body1
128 !0 = !{!"branch_weights", i32 1, i32 64}
130 define i32 @test_loop_early_exits(i32 %i, ptr %a) {
131 ; Check that we sink early exit blocks out of loop bodies.
132 ; CHECK-LABEL: test_loop_early_exits:
147 %iv = phi i32 [ 0, %entry ], [ %next, %body4 ]
148 %base = phi i32 [ 0, %entry ], [ %sum, %body4 ]
149 %bailcond1 = icmp eq i32 %base, 42
150 br i1 %bailcond1, label %bail1, label %body2
156 %bailcond2 = icmp eq i32 %base, 43
157 br i1 %bailcond2, label %bail2, label %body3
163 %bailcond3 = icmp eq i32 %base, 44
164 br i1 %bailcond3, label %bail3, label %body4
170 %arrayidx = getelementptr inbounds i32, ptr %a, i32 %iv
171 %0 = load i32, ptr %arrayidx
172 %sum = add nsw i32 %0, %base
173 %next = add i32 %iv, 1
174 %exitcond = icmp eq i32 %next, %i
175 br i1 %exitcond, label %exit, label %body1
181 ; Tail duplication during layout can entirely remove body0 by duplicating it
182 ; into the entry block and into body1. This is a good thing but it isn't what
183 ; this test is looking for. So to make the blocks longer so they don't get
184 ; duplicated, we add some calls to dummy.
185 declare void @dummy()
187 define i32 @test_loop_rotate(i32 %i, ptr %a) {
188 ; Check that we rotate conditional exits from the loop to the bottom of the
189 ; loop, eliminating unconditional branches to the top.
190 ; CHECK-LABEL: test_loop_rotate:
200 %iv = phi i32 [ 0, %entry ], [ %next, %body1 ]
201 %base = phi i32 [ 0, %entry ], [ %sum, %body1 ]
202 %next = add i32 %iv, 1
203 %exitcond = icmp eq i32 %next, %i
206 br i1 %exitcond, label %exit, label %body1
209 %arrayidx = getelementptr inbounds i32, ptr %a, i32 %iv
210 %0 = load i32, ptr %arrayidx
211 %sum = add nsw i32 %0, %base
212 %bailcond1 = icmp eq i32 %sum, 42
219 define i32 @test_no_loop_rotate(i32 %i, ptr %a) {
220 ; Check that we don't try to rotate a loop which is already laid out with
221 ; fallthrough opportunities into the top and out of the bottom.
222 ; CHECK-LABEL: test_no_loop_rotate:
232 %iv = phi i32 [ 0, %entry ], [ %next, %body1 ]
233 %base = phi i32 [ 0, %entry ], [ %sum, %body1 ]
234 %arrayidx = getelementptr inbounds i32, ptr %a, i32 %iv
235 %0 = load i32, ptr %arrayidx
236 %sum = add nsw i32 %0, %base
237 %bailcond1 = icmp eq i32 %sum, 42
238 br i1 %bailcond1, label %exit, label %body1
241 %next = add i32 %iv, 1
242 %exitcond = icmp eq i32 %next, %i
243 br i1 %exitcond, label %exit, label %body0
249 define i32 @test_loop_align(i32 %i, ptr %a) {
250 ; Check that we provide basic loop body alignment with the block placement
252 ; CHECK-LABEL: test_loop_align:
254 ; CHECK: .p2align [[ALIGN:[0-9]+]],
262 %iv = phi i32 [ 0, %entry ], [ %next, %body ]
263 %base = phi i32 [ 0, %entry ], [ %sum, %body ]
264 %arrayidx = getelementptr inbounds i32, ptr %a, i32 %iv
265 %0 = load i32, ptr %arrayidx
266 %sum = add nsw i32 %0, %base
267 %next = add i32 %iv, 1
268 %exitcond = icmp eq i32 %next, %i
269 br i1 %exitcond, label %exit, label %body
275 define i32 @test_nested_loop_align(i32 %i, ptr %a, ptr %b) {
276 ; Check that we provide nested loop body alignment.
277 ; CHECK-LABEL: test_nested_loop_align:
279 ; CHECK: .p2align [[ALIGN]],
280 ; CHECK-NEXT: %loop.body.1
281 ; CHECK: .p2align [[ALIGN]],
282 ; CHECK-NEXT: %inner.loop.body
283 ; CHECK-NOT: .p2align
287 br label %loop.body.1
290 %iv = phi i32 [ 0, %entry ], [ %next, %loop.body.2 ]
291 %arrayidx = getelementptr inbounds i32, ptr %a, i32 %iv
292 %bidx = load i32, ptr %arrayidx
293 br label %inner.loop.body
296 %inner.iv = phi i32 [ 0, %loop.body.1 ], [ %inner.next, %inner.loop.body ]
297 %base = phi i32 [ 0, %loop.body.1 ], [ %sum, %inner.loop.body ]
298 %scaled_idx = mul i32 %bidx, %iv
299 %inner.arrayidx = getelementptr inbounds i32, ptr %b, i32 %scaled_idx
300 %0 = load i32, ptr %inner.arrayidx
301 %sum = add nsw i32 %0, %base
302 %inner.next = add i32 %iv, 1
303 %inner.exitcond = icmp eq i32 %inner.next, %i
304 br i1 %inner.exitcond, label %loop.body.2, label %inner.loop.body
307 %next = add i32 %iv, 1
308 %exitcond = icmp eq i32 %next, %i
309 br i1 %exitcond, label %exit, label %loop.body.1
315 define void @unnatural_cfg1() {
316 ; Test that we can handle a loop with an inner unnatural loop at the end of
317 ; a function. This is a gross CFG reduced out of the single source GCC.
318 ; CHECK-LABEL: unnatural_cfg1
320 ; CHECK: %loop.header
324 br label %loop.header
330 br i1 undef, label %loop.body3, label %loop.body2
333 %ptr = load ptr, ptr undef, align 4
337 %myptr = phi ptr [ %ptr2, %loop.body5 ], [ %ptr, %loop.body2 ], [ undef, %loop.body1 ]
338 %bcmyptr = bitcast ptr %myptr to ptr
339 %val = load i32, ptr %bcmyptr, align 4
340 %comp = icmp eq i32 %val, 48
341 br i1 %comp, label %loop.body4, label %loop.body5
344 br i1 undef, label %loop.header, label %loop.body5
347 %ptr2 = load ptr, ptr undef, align 4
351 define void @unnatural_cfg2(ptr %p0, i32 %a0) {
352 ; Test that we can handle a loop with a nested natural loop *and* an unnatural
353 ; loop. This was reduced from a crash on block placement when run over
355 ; CHECK-LABEL: unnatural_cfg2
357 ; CHECK: %loop.header
361 ; CHECK: %loop.inner2.begin
363 ; CHECK: %loop.inner1.begin
367 br label %loop.header
370 %comp0 = icmp eq ptr %p0, null
371 br i1 %comp0, label %bail, label %loop.body1
374 %val0 = load ptr, ptr undef, align 4
375 br i1 undef, label %loop.body2, label %loop.inner1.begin
378 br i1 undef, label %loop.body4, label %loop.body3
381 %ptr1 = getelementptr inbounds i32, ptr %val0, i32 0
382 %castptr1 = bitcast ptr %ptr1 to ptr
383 %val1 = load ptr, ptr %castptr1, align 4
384 br label %loop.inner1.begin
387 %valphi = phi ptr [ %val2, %loop.inner1.end ], [ %val1, %loop.body3 ], [ %val0, %loop.body1 ]
388 %castval = bitcast ptr %valphi to ptr
389 %comp1 = icmp eq i32 %a0, 48
390 br i1 %comp1, label %loop.inner1.end, label %loop.body4
393 %ptr2 = getelementptr inbounds i32, ptr %valphi, i32 0
394 %castptr2 = bitcast ptr %ptr2 to ptr
395 %val2 = load ptr, ptr %castptr2, align 4
396 br label %loop.inner1.begin
402 %comp2 = icmp ult i32 %a0, 3
403 br i1 %comp2, label %loop.inner2.begin, label %loop.end
406 br i1 false, label %loop.end, label %loop.inner2.end
409 %comp3 = icmp eq i32 %a0, 1769472
410 br i1 %comp3, label %loop.end, label %loop.inner2.begin
413 br label %loop.header
419 define i32 @problematic_switch() {
420 ; This function's CFG caused overlow in the machine branch probability
421 ; calculation, triggering asserts. Make sure we don't crash on it.
422 ; CHECK: problematic_switch
425 switch i32 undef, label %exit [
426 i32 879, label %bogus
466 %merge = phi i32 [ 3, %step ], [ 6, %entry ]
470 define void @fpcmp_unanalyzable_branch(i1 %cond, double %a0) {
471 ; This function's CFG contains an once-unanalyzable branch (une on floating
472 ; points). As now it becomes analyzable, we should get best layout in which each
473 ; edge in 'entry' -> 'entry.if.then_crit_edge' -> 'if.then' -> 'if.end' is
475 ; CHECK-LABEL: fpcmp_unanalyzable_branch:
476 ; CHECK: # %bb.0: # %entry
477 ; CHECK: # %bb.1: # %entry.if.then_crit_edge
478 ; CHECK: .LBB10_5: # %if.then
479 ; CHECK: .LBB10_6: # %if.end
480 ; CHECK: # %bb.3: # %exit
481 ; CHECK: jne .LBB10_4
482 ; CHECK-NEXT: jnp .LBB10_6
483 ; CHECK: jmp .LBB10_5
486 ; Note that this branch must be strongly biased toward
487 ; 'entry.if.then_crit_edge' to ensure that we would try to form a chain for
488 ; 'entry' -> 'entry.if.then_crit_edge' -> 'if.then' -> 'if.end'.
489 br i1 %cond, label %entry.if.then_crit_edge, label %lor.lhs.false, !prof !1
491 entry.if.then_crit_edge:
492 %.pre14 = load i8, ptr undef, align 1
496 br i1 undef, label %if.end, label %exit
499 %cmp.i = fcmp une double 0.000000e+00, %a0
500 br i1 %cmp.i, label %if.then, label %if.end, !prof !3
503 %0 = phi i8 [ %.pre14, %entry.if.then_crit_edge ], [ undef, %exit ]
505 store i8 %1, ptr undef, align 4
512 !1 = !{!"branch_weights", i32 1000, i32 1}
513 !3 = !{!"branch_weights", i32 1, i32 1000}
517 declare i32 @h(i32 %x)
519 define i32 @test_global_cfg_break_profitability() {
520 ; Check that our metrics for the profitability of a CFG break are global rather
521 ; than local. A successor may be very hot, but if the current block isn't, it
522 ; doesn't matter. Within this test the 'then' block is slightly warmer than the
523 ; 'else' block, but not nearly enough to merit merging it with the exit block
524 ; even though the probability of 'then' branching to the 'exit' block is very
526 ; CHECK: test_global_cfg_break_profitability
527 ; CHECK: calll {{_?}}f
528 ; CHECK: calll {{_?}}g
529 ; CHECK: calll {{_?}}h
533 br i1 undef, label %then, label %else, !prof !2
536 %then.result = call i32 @f()
540 %else.result = call i32 @g()
544 %result = phi i32 [ %then.result, %then ], [ %else.result, %else ]
545 %result2 = call i32 @h(i32 %result)
549 !2 = !{!"branch_weights", i32 3, i32 1}
551 declare i32 @__gxx_personality_v0(...)
553 define void @test_eh_lpad_successor() personality ptr @__gxx_personality_v0 {
554 ; Some times the landing pad ends up as the first successor of an invoke block.
555 ; When this happens, a strange result used to fall out of updateTerminators: we
556 ; didn't correctly locate the fallthrough successor, assuming blindly that the
557 ; first one was the fallthrough successor. As a result, we would add an
558 ; erroneous jump to the landing pad thinking *that* was the default successor.
559 ; CHECK-LABEL: test_eh_lpad_successor
565 invoke i32 @f() to label %preheader unwind label %lpad
571 %lpad.val = landingpad { ptr, i32 }
573 resume { ptr, i32 } %lpad.val
579 declare void @fake_throw() noreturn
581 define void @test_eh_throw() personality ptr @__gxx_personality_v0 {
582 ; For blocks containing a 'throw' (or similar functionality), we have
583 ; a no-return invoke. In this case, only EH successors will exist, and
584 ; fallthrough simply won't occur. Make sure we don't crash trying to update
585 ; terminators for such constructs.
587 ; CHECK-LABEL: test_eh_throw
592 invoke void @fake_throw() to label %continue unwind label %cleanup
598 %0 = landingpad { ptr, i32 }
603 define void @test_unnatural_cfg_backwards_inner_loop() {
604 ; Test that when we encounter an unnatural CFG structure after having formed
605 ; a chain for an inner loop which happened to be laid out backwards we don't
606 ; attempt to merge onto the wrong end of the inner loop just because we find it
607 ; first. This was reduced from a crasher in GCC's single source.
609 ; CHECK-LABEL: test_unnatural_cfg_backwards_inner_loop
615 br i1 undef, label %loop2a, label %body
621 %next.load = load ptr, ptr undef
622 br i1 %comp.a, label %loop2a, label %loop2b
625 %var = phi ptr [ null, %entry ], [ null, %body ], [ %next.phi, %loop1 ]
626 %next.var = phi ptr [ null, %entry ], [ undef, %body ], [ %next.load, %loop1 ]
627 %comp.a = icmp eq ptr %var, null
631 %gep = getelementptr inbounds i32, ptr %var.phi, i32 0
632 %next.ptr = bitcast ptr %gep to ptr
633 store ptr %next.phi, ptr %next.ptr
637 %var.phi = phi ptr [ %next.phi, %loop2b ], [ %var, %loop2a ]
638 %next.phi = phi ptr [ %next.load, %loop2b ], [ %next.var, %loop2a ]
642 define void @unanalyzable_branch_to_loop_header(double %a0) {
643 ; Ensure that we can handle unanalyzable branches into loop headers. We
644 ; pre-form chains for unanalyzable branches, and will find the tail end of that
645 ; at the start of the loop. This function uses floating point comparison
646 ; fallthrough because that happens to always produce unanalyzable branches on
649 ; CHECK-LABEL: unanalyzable_branch_to_loop_header
655 %cmp = fcmp une double 0.000000e+00, %a0
656 br i1 %cmp, label %loop, label %exit
659 %cond = icmp eq i8 undef, 42
660 br i1 %cond, label %exit, label %loop
666 define void @unanalyzable_branch_to_best_succ(i1 %cond, double %a0) {
667 ; Ensure that we can handle unanalyzable branches where the destination block
668 ; gets selected as the optimal successor to merge.
670 ; This branch is now analyzable and hence the destination block becomes the
671 ; hotter one. The right order is entry->bar->exit->foo.
673 ; CHECK-LABEL: unanalyzable_branch_to_best_succ
680 ; Bias this branch toward bar to ensure we form that chain.
681 br i1 %cond, label %bar, label %foo, !prof !1
684 %cmp = fcmp une double 0.000000e+00, %a0
685 br i1 %cmp, label %bar, label %exit
695 define void @unanalyzable_branch_to_free_block(float %x) {
696 ; Ensure that we can handle unanalyzable branches where the destination block
697 ; gets selected as the best free block in the CFG.
699 ; CHECK-LABEL: unanalyzable_branch_to_free_block
707 br i1 undef, label %a, label %b
714 %cmp = fcmp une float %x, 0.0
715 br i1 %cmp, label %c, label %exit
725 define void @many_unanalyzable_branches() {
726 ; Ensure that we don't crash as we're building up many unanalyzable branches,
729 ; CHECK-LABEL: many_unanalyzable_branches
736 %val0 = load volatile float, ptr undef
737 %cmp0 = fcmp une float %val0, 0.0
738 br i1 %cmp0, label %1, label %0
739 %val1 = load volatile float, ptr undef
740 %cmp1 = fcmp une float %val1, 0.0
741 br i1 %cmp1, label %2, label %1
742 %val2 = load volatile float, ptr undef
743 %cmp2 = fcmp une float %val2, 0.0
744 br i1 %cmp2, label %3, label %2
745 %val3 = load volatile float, ptr undef
746 %cmp3 = fcmp une float %val3, 0.0
747 br i1 %cmp3, label %4, label %3
748 %val4 = load volatile float, ptr undef
749 %cmp4 = fcmp une float %val4, 0.0
750 br i1 %cmp4, label %5, label %4
751 %val5 = load volatile float, ptr undef
752 %cmp5 = fcmp une float %val5, 0.0
753 br i1 %cmp5, label %6, label %5
754 %val6 = load volatile float, ptr undef
755 %cmp6 = fcmp une float %val6, 0.0
756 br i1 %cmp6, label %7, label %6
757 %val7 = load volatile float, ptr undef
758 %cmp7 = fcmp une float %val7, 0.0
759 br i1 %cmp7, label %8, label %7
760 %val8 = load volatile float, ptr undef
761 %cmp8 = fcmp une float %val8, 0.0
762 br i1 %cmp8, label %9, label %8
763 %val9 = load volatile float, ptr undef
764 %cmp9 = fcmp une float %val9, 0.0
765 br i1 %cmp9, label %10, label %9
766 %val10 = load volatile float, ptr undef
767 %cmp10 = fcmp une float %val10, 0.0
768 br i1 %cmp10, label %11, label %10
769 %val11 = load volatile float, ptr undef
770 %cmp11 = fcmp une float %val11, 0.0
771 br i1 %cmp11, label %12, label %11
772 %val12 = load volatile float, ptr undef
773 %cmp12 = fcmp une float %val12, 0.0
774 br i1 %cmp12, label %13, label %12
775 %val13 = load volatile float, ptr undef
776 %cmp13 = fcmp une float %val13, 0.0
777 br i1 %cmp13, label %14, label %13
778 %val14 = load volatile float, ptr undef
779 %cmp14 = fcmp une float %val14, 0.0
780 br i1 %cmp14, label %15, label %14
781 %val15 = load volatile float, ptr undef
782 %cmp15 = fcmp une float %val15, 0.0
783 br i1 %cmp15, label %16, label %15
784 %val16 = load volatile float, ptr undef
785 %cmp16 = fcmp une float %val16, 0.0
786 br i1 %cmp16, label %17, label %16
787 %val17 = load volatile float, ptr undef
788 %cmp17 = fcmp une float %val17, 0.0
789 br i1 %cmp17, label %18, label %17
790 %val18 = load volatile float, ptr undef
791 %cmp18 = fcmp une float %val18, 0.0
792 br i1 %cmp18, label %19, label %18
793 %val19 = load volatile float, ptr undef
794 %cmp19 = fcmp une float %val19, 0.0
795 br i1 %cmp19, label %20, label %19
796 %val20 = load volatile float, ptr undef
797 %cmp20 = fcmp une float %val20, 0.0
798 br i1 %cmp20, label %21, label %20
799 %val21 = load volatile float, ptr undef
800 %cmp21 = fcmp une float %val21, 0.0
801 br i1 %cmp21, label %22, label %21
802 %val22 = load volatile float, ptr undef
803 %cmp22 = fcmp une float %val22, 0.0
804 br i1 %cmp22, label %23, label %22
805 %val23 = load volatile float, ptr undef
806 %cmp23 = fcmp une float %val23, 0.0
807 br i1 %cmp23, label %24, label %23
808 %val24 = load volatile float, ptr undef
809 %cmp24 = fcmp une float %val24, 0.0
810 br i1 %cmp24, label %25, label %24
811 %val25 = load volatile float, ptr undef
812 %cmp25 = fcmp une float %val25, 0.0
813 br i1 %cmp25, label %26, label %25
814 %val26 = load volatile float, ptr undef
815 %cmp26 = fcmp une float %val26, 0.0
816 br i1 %cmp26, label %27, label %26
817 %val27 = load volatile float, ptr undef
818 %cmp27 = fcmp une float %val27, 0.0
819 br i1 %cmp27, label %28, label %27
820 %val28 = load volatile float, ptr undef
821 %cmp28 = fcmp une float %val28, 0.0
822 br i1 %cmp28, label %29, label %28
823 %val29 = load volatile float, ptr undef
824 %cmp29 = fcmp une float %val29, 0.0
825 br i1 %cmp29, label %30, label %29
826 %val30 = load volatile float, ptr undef
827 %cmp30 = fcmp une float %val30, 0.0
828 br i1 %cmp30, label %31, label %30
829 %val31 = load volatile float, ptr undef
830 %cmp31 = fcmp une float %val31, 0.0
831 br i1 %cmp31, label %32, label %31
832 %val32 = load volatile float, ptr undef
833 %cmp32 = fcmp une float %val32, 0.0
834 br i1 %cmp32, label %33, label %32
835 %val33 = load volatile float, ptr undef
836 %cmp33 = fcmp une float %val33, 0.0
837 br i1 %cmp33, label %34, label %33
838 %val34 = load volatile float, ptr undef
839 %cmp34 = fcmp une float %val34, 0.0
840 br i1 %cmp34, label %35, label %34
841 %val35 = load volatile float, ptr undef
842 %cmp35 = fcmp une float %val35, 0.0
843 br i1 %cmp35, label %36, label %35
844 %val36 = load volatile float, ptr undef
845 %cmp36 = fcmp une float %val36, 0.0
846 br i1 %cmp36, label %37, label %36
847 %val37 = load volatile float, ptr undef
848 %cmp37 = fcmp une float %val37, 0.0
849 br i1 %cmp37, label %38, label %37
850 %val38 = load volatile float, ptr undef
851 %cmp38 = fcmp une float %val38, 0.0
852 br i1 %cmp38, label %39, label %38
853 %val39 = load volatile float, ptr undef
854 %cmp39 = fcmp une float %val39, 0.0
855 br i1 %cmp39, label %40, label %39
856 %val40 = load volatile float, ptr undef
857 %cmp40 = fcmp une float %val40, 0.0
858 br i1 %cmp40, label %41, label %40
859 %val41 = load volatile float, ptr undef
860 %cmp41 = fcmp une float %val41, undef
861 br i1 %cmp41, label %42, label %41
862 %val42 = load volatile float, ptr undef
863 %cmp42 = fcmp une float %val42, 0.0
864 br i1 %cmp42, label %43, label %42
865 %val43 = load volatile float, ptr undef
866 %cmp43 = fcmp une float %val43, 0.0
867 br i1 %cmp43, label %44, label %43
868 %val44 = load volatile float, ptr undef
869 %cmp44 = fcmp une float %val44, 0.0
870 br i1 %cmp44, label %45, label %44
871 %val45 = load volatile float, ptr undef
872 %cmp45 = fcmp une float %val45, 0.0
873 br i1 %cmp45, label %46, label %45
874 %val46 = load volatile float, ptr undef
875 %cmp46 = fcmp une float %val46, 0.0
876 br i1 %cmp46, label %47, label %46
877 %val47 = load volatile float, ptr undef
878 %cmp47 = fcmp une float %val47, 0.0
879 br i1 %cmp47, label %48, label %47
880 %val48 = load volatile float, ptr undef
881 %cmp48 = fcmp une float %val48, 0.0
882 br i1 %cmp48, label %49, label %48
883 %val49 = load volatile float, ptr undef
884 %cmp49 = fcmp une float %val49, 0.0
885 br i1 %cmp49, label %50, label %49
886 %val50 = load volatile float, ptr undef
887 %cmp50 = fcmp une float %val50, 0.0
888 br i1 %cmp50, label %51, label %50
889 %val51 = load volatile float, ptr undef
890 %cmp51 = fcmp une float %val51, 0.0
891 br i1 %cmp51, label %52, label %51
892 %val52 = load volatile float, ptr undef
893 %cmp52 = fcmp une float %val52, 0.0
894 br i1 %cmp52, label %53, label %52
895 %val53 = load volatile float, ptr undef
896 %cmp53 = fcmp une float %val53, 0.0
897 br i1 %cmp53, label %54, label %53
898 %val54 = load volatile float, ptr undef
899 %cmp54 = fcmp une float %val54, 0.0
900 br i1 %cmp54, label %55, label %54
901 %val55 = load volatile float, ptr undef
902 %cmp55 = fcmp une float %val55, 0.0
903 br i1 %cmp55, label %56, label %55
904 %val56 = load volatile float, ptr undef
905 %cmp56 = fcmp une float %val56, 0.0
906 br i1 %cmp56, label %57, label %56
907 %val57 = load volatile float, ptr undef
908 %cmp57 = fcmp une float %val57, 0.0
909 br i1 %cmp57, label %58, label %57
910 %val58 = load volatile float, ptr undef
911 %cmp58 = fcmp une float %val58, 0.0
912 br i1 %cmp58, label %59, label %58
913 %val59 = load volatile float, ptr undef
914 %cmp59 = fcmp une float %val59, 0.0
915 br i1 %cmp59, label %60, label %59
916 %val60 = load volatile float, ptr undef
917 %cmp60 = fcmp une float %val60, 0.0
918 br i1 %cmp60, label %61, label %60
919 %val61 = load volatile float, ptr undef
920 %cmp61 = fcmp une float %val61, 0.0
921 br i1 %cmp61, label %62, label %61
922 %val62 = load volatile float, ptr undef
923 %cmp62 = fcmp une float %val62, 0.0
924 br i1 %cmp62, label %63, label %62
925 %val63 = load volatile float, ptr undef
926 %cmp63 = fcmp une float %val63, 0.0
927 br i1 %cmp63, label %64, label %63
928 %val64 = load volatile float, ptr undef
929 %cmp64 = fcmp une float %val64, 0.0
930 br i1 %cmp64, label %65, label %64
937 define void @benchmark_heapsort(i32 %n, ptr nocapture %ra) {
938 ; This test case comes from the heapsort benchmark, and exemplifies several
939 ; important aspects to block placement in the presence of loops:
940 ; 1) Loop rotation needs to *ensure* that the desired exiting edge can be
942 ; 2) The exiting edge from the loop which is rotated to be laid out at the
943 ; bottom of the loop needs to be exiting into the nearest enclosing loop (to
944 ; which there is an exit). Otherwise, we force that enclosing loop into
945 ; strange layouts that are siginificantly less efficient, often times making
948 ; CHECK-LABEL: @benchmark_heapsort
950 ; First rotated loop top.
953 ; %for.cond gets completely tail-duplicated away.
957 ; Second rotated loop top
958 ; CHECK: %while.cond.outer
959 ; Third rotated loop top
964 ; CHECK: %land.lhs.true
971 %shr = ashr i32 %n, 1
972 %add = add nsw i32 %shr, 1
973 %arrayidx3 = getelementptr inbounds double, ptr %ra, i64 1
977 %ir.0 = phi i32 [ %n, %entry ], [ %ir.1, %while.end ]
978 %l.0 = phi i32 [ %add, %entry ], [ %l.1, %while.end ]
979 %cmp = icmp sgt i32 %l.0, 1
980 br i1 %cmp, label %if.then, label %if.else
983 %dec = add nsw i32 %l.0, -1
984 %idxprom = sext i32 %dec to i64
985 %arrayidx = getelementptr inbounds double, ptr %ra, i64 %idxprom
986 %0 = load double, ptr %arrayidx, align 8
990 %idxprom1 = sext i32 %ir.0 to i64
991 %arrayidx2 = getelementptr inbounds double, ptr %ra, i64 %idxprom1
992 %1 = load double, ptr %arrayidx2, align 8
993 %2 = load double, ptr %arrayidx3, align 8
994 store double %2, ptr %arrayidx2, align 8
995 %dec6 = add nsw i32 %ir.0, -1
996 %cmp7 = icmp eq i32 %dec6, 1
997 br i1 %cmp7, label %if.then8, label %if.end10
1000 store double %1, ptr %arrayidx3, align 8
1004 %ir.1 = phi i32 [ %ir.0, %if.then ], [ %dec6, %if.else ]
1005 %l.1 = phi i32 [ %dec, %if.then ], [ %l.0, %if.else ]
1006 %rra.0 = phi double [ %0, %if.then ], [ %1, %if.else ]
1007 %add31 = add nsw i32 %ir.1, 1
1008 br label %while.cond.outer
1011 %j.0.ph.in = phi i32 [ %l.1, %if.end10 ], [ %j.1, %if.then24 ]
1012 %j.0.ph = shl i32 %j.0.ph.in, 1
1013 br label %while.cond
1016 %j.0 = phi i32 [ %add31, %if.end20 ], [ %j.0.ph, %while.cond.outer ]
1017 %cmp11 = icmp sgt i32 %j.0, %ir.1
1018 br i1 %cmp11, label %while.end, label %while.body
1021 %cmp12 = icmp slt i32 %j.0, %ir.1
1022 br i1 %cmp12, label %land.lhs.true, label %if.end20
1025 %idxprom13 = sext i32 %j.0 to i64
1026 %arrayidx14 = getelementptr inbounds double, ptr %ra, i64 %idxprom13
1027 %3 = load double, ptr %arrayidx14, align 8
1028 %add15 = add nsw i32 %j.0, 1
1029 %idxprom16 = sext i32 %add15 to i64
1030 %arrayidx17 = getelementptr inbounds double, ptr %ra, i64 %idxprom16
1031 %4 = load double, ptr %arrayidx17, align 8
1032 %cmp18 = fcmp olt double %3, %4
1033 br i1 %cmp18, label %if.then19, label %if.end20
1039 %j.1 = phi i32 [ %add15, %if.then19 ], [ %j.0, %land.lhs.true ], [ %j.0, %while.body ]
1040 %idxprom21 = sext i32 %j.1 to i64
1041 %arrayidx22 = getelementptr inbounds double, ptr %ra, i64 %idxprom21
1042 %5 = load double, ptr %arrayidx22, align 8
1043 %cmp23 = fcmp olt double %rra.0, %5
1044 br i1 %cmp23, label %if.then24, label %while.cond
1047 %idxprom27 = sext i32 %j.0.ph.in to i64
1048 %arrayidx28 = getelementptr inbounds double, ptr %ra, i64 %idxprom27
1049 store double %5, ptr %arrayidx28, align 8
1050 br label %while.cond.outer
1053 %idxprom33 = sext i32 %j.0.ph.in to i64
1054 %arrayidx34 = getelementptr inbounds double, ptr %ra, i64 %idxprom33
1055 store double %rra.0, ptr %arrayidx34, align 8
1059 declare void @cold_function() cold
1061 define i32 @test_cold_calls(ptr %a) {
1062 ; Test that edges to blocks post-dominated by cold calls are
1063 ; marked as not expected to be taken. They should be laid out
1065 ; CHECK-LABEL: test_cold_calls:
1072 %gep1 = getelementptr i32, ptr %a, i32 1
1073 %val1 = load i32, ptr %gep1
1074 %cond1 = icmp ugt i32 %val1, 1
1075 br i1 %cond1, label %then, label %else
1078 call void @cold_function()
1082 %gep2 = getelementptr i32, ptr %a, i32 2
1083 %val2 = load i32, ptr %gep2
1087 %ret = phi i32 [ %val1, %then ], [ %val2, %else ]
1091 ; Make sure we put landingpads out of the way.
1092 declare i32 @pers(...)
1098 define i32 @test_lp(i32 %a) personality ptr @pers {
1099 ; CHECK-LABEL: test_lp:
1108 %0 = icmp sgt i32 %a, 1
1109 br i1 %0, label %hot, label %cold, !prof !4
1112 %1 = invoke i32 @foo()
1113 to label %then unwind label %hotlp
1116 %2 = invoke i32 @bar()
1117 to label %then unwind label %coldlp
1120 %3 = phi i32 [ %1, %hot ], [ %2, %cold ]
1124 %4 = landingpad { ptr, i32 }
1129 %5 = landingpad { ptr, i32 }
1134 %6 = phi i32 [-1, %hotlp], [-2, %coldlp]
1139 !4 = !{!"branch_weights", i32 65536, i32 0}
1141 ; Make sure that ehpad are scheduled from the least probable one
1142 ; to the most probable one. See selectBestCandidateBlock as to why.
1143 declare void @clean();
1145 define void @test_flow_unwind() personality ptr @pers {
1146 ; CHECK-LABEL: test_flow_unwind:
1152 ; CHECK: %outercleanup
1154 %0 = invoke i32 @foo()
1155 to label %then unwind label %outerlp
1158 %1 = invoke i32 @bar()
1159 to label %exit unwind label %innerlp
1165 %2 = landingpad { ptr, i32 }
1167 br label %innercleanup
1170 %3 = landingpad { ptr, i32 }
1172 br label %outercleanup
1175 %4 = phi { ptr, i32 } [%2, %innercleanup], [%3, %outerlp]
1177 resume { ptr, i32 } %4
1181 br label %outercleanup
1184 declare void @hot_function()
1186 define void @test_hot_branch(ptr %a) {
1187 ; Test that a hot branch that has a probability a little larger than 80% will
1188 ; break CFG constrains when doing block placement.
1189 ; CHECK-LABEL: test_hot_branch:
1196 %gep1 = getelementptr i32, ptr %a, i32 1
1197 %val1 = load i32, ptr %gep1
1198 %cond1 = icmp ugt i32 %val1, 1
1199 br i1 %cond1, label %then, label %else, !prof !5
1202 call void @hot_function()
1206 call void @cold_function()
1210 call void @hot_function()
1214 define void @test_hot_branch_profile(ptr %a) !prof !6 {
1215 ; Test that a hot branch that has a probability a little larger than 50% will
1216 ; break CFG constrains when doing block placement when profile is available.
1217 ; CHECK-LABEL: test_hot_branch_profile:
1224 %gep1 = getelementptr i32, ptr %a, i32 1
1225 %val1 = load i32, ptr %gep1
1226 %cond1 = icmp ugt i32 %val1, 1
1227 br i1 %cond1, label %then, label %else, !prof !7
1230 call void @hot_function()
1234 call void @cold_function()
1238 call void @hot_function()
1242 define void @test_hot_branch_triangle_profile(ptr %a) !prof !6 {
1243 ; Test that a hot branch that has a probability a little larger than 80% will
1244 ; break triangle shaped CFG constrains when doing block placement if profile
1246 ; CHECK-LABEL: test_hot_branch_triangle_profile:
1252 %gep1 = getelementptr i32, ptr %a, i32 1
1253 %val1 = load i32, ptr %gep1
1254 %cond1 = icmp ugt i32 %val1, 1
1255 br i1 %cond1, label %exit, label %then, !prof !5
1258 call void @hot_function()
1262 call void @hot_function()
1266 define void @test_hot_branch_triangle_profile_topology(ptr %a) !prof !6 {
1267 ; Test that a hot branch that has a probability between 50% and 66% will not
1268 ; break triangle shaped CFG constrains when doing block placement if profile
1270 ; CHECK-LABEL: test_hot_branch_triangle_profile_topology:
1276 %gep1 = getelementptr i32, ptr %a, i32 1
1277 %val1 = load i32, ptr %gep1
1278 %cond1 = icmp ugt i32 %val1, 1
1279 br i1 %cond1, label %exit, label %then, !prof !7
1282 call void @hot_function()
1286 call void @hot_function()
1293 define void @test_forked_hot_diamond(ptr %a) {
1294 ; Test that a hot-branch with probability > 80% followed by a 50/50 branch
1295 ; will not place the cold predecessor if the probability for the fallthrough
1297 ; CHECK-LABEL: test_forked_hot_diamond
1305 %gep1 = getelementptr i32, ptr %a, i32 1
1306 %val1 = load i32, ptr %gep1
1307 %cond1 = icmp ugt i32 %val1, 1
1308 br i1 %cond1, label %then, label %else, !prof !5
1311 call void @hot_function()
1312 %gep2 = getelementptr i32, ptr %a, i32 2
1313 %val2 = load i32, ptr %gep2
1314 %cond2 = icmp ugt i32 %val2, 2
1315 br i1 %cond2, label %fork1, label %fork2, !prof !8
1318 call void @cold_function()
1319 %gep3 = getelementptr i32, ptr %a, i32 3
1320 %val3 = load i32, ptr %gep3
1321 %cond3 = icmp ugt i32 %val3, 3
1322 br i1 %cond3, label %fork1, label %fork2, !prof !8
1333 call void @hot_function()
1337 define void @test_forked_hot_diamond_gets_cold(ptr %a) {
1338 ; Test that a hot-branch with probability > 80% followed by a 50/50 branch
1339 ; will place the cold predecessor if the probability for the fallthrough
1341 ; The probability for both branches is 85%. For then2 vs else1
1342 ; this results in a compounded probability of 83%.
1343 ; Neither then2->fork1 nor then2->fork2 has a large enough relative
1344 ; probability to break the CFG.
1346 ; then2 -> fork1 vs else1 -> fork1 = 71%
1347 ; then2 -> fork2 vs else2 -> fork2 = 74%
1348 ; CHECK-LABEL: test_forked_hot_diamond_gets_cold
1358 %gep1 = getelementptr i32, ptr %a, i32 1
1359 %val1 = load i32, ptr %gep1
1360 %cond1 = icmp ugt i32 %val1, 1
1361 br i1 %cond1, label %then1, label %else1, !prof !9
1364 call void @hot_function()
1365 %gep2 = getelementptr i32, ptr %a, i32 2
1366 %val2 = load i32, ptr %gep2
1367 %cond2 = icmp ugt i32 %val2, 2
1368 br i1 %cond2, label %then2, label %else2, !prof !9
1371 call void @cold_function()
1375 call void @hot_function()
1376 %gep3 = getelementptr i32, ptr %a, i32 3
1377 %val3 = load i32, ptr %gep2
1378 %cond3 = icmp ugt i32 %val2, 3
1379 br i1 %cond3, label %fork1, label %fork2, !prof !8
1382 call void @cold_function()
1394 call void @hot_function()
1398 define void @test_forked_hot_diamond_stays_hot(ptr %a) {
1399 ; Test that a hot-branch with probability > 88.88% (1:8) followed by a 50/50
1400 ; branch will not place the cold predecessor as the probability for the
1401 ; fallthrough stays above 80%
1402 ; (1:8) followed by (1:1) is still (1:4)
1403 ; Here we use 90% probability because two in a row
1404 ; have a 89 % probability vs the original branch.
1405 ; CHECK-LABEL: test_forked_hot_diamond_stays_hot
1415 %gep1 = getelementptr i32, ptr %a, i32 1
1416 %val1 = load i32, ptr %gep1
1417 %cond1 = icmp ugt i32 %val1, 1
1418 br i1 %cond1, label %then1, label %else1, !prof !10
1421 call void @hot_function()
1422 %gep2 = getelementptr i32, ptr %a, i32 2
1423 %val2 = load i32, ptr %gep2
1424 %cond2 = icmp ugt i32 %val2, 2
1425 br i1 %cond2, label %then2, label %else2, !prof !10
1428 call void @cold_function()
1432 call void @hot_function()
1433 %gep3 = getelementptr i32, ptr %a, i32 3
1434 %val3 = load i32, ptr %gep2
1435 %cond3 = icmp ugt i32 %val2, 3
1436 br i1 %cond3, label %fork1, label %fork2, !prof !8
1439 call void @cold_function()
1451 call void @hot_function()
1455 ; Because %endif has a higher frequency than %if, the calculations show we
1456 ; shouldn't tail-duplicate %endif so that we can place it after %if. We were
1457 ; previously undercounting the cost by ignoring execution frequency that didn't
1458 ; come from the %if->%endif path.
1459 ; CHECK-LABEL: higher_frequency_succ_tail_dup
1466 define void @higher_frequency_succ_tail_dup(i1 %a, i1 %b, i1 %c) {
1469 if: ; preds = %entry
1470 call void @effect(i32 0)
1471 br i1 %a, label %elseif, label %endif, !prof !11 ; even
1473 elseif: ; preds = %if
1474 call void @effect(i32 1)
1475 br i1 %b, label %else, label %endif, !prof !11 ; even
1477 else: ; preds = %elseif
1478 call void @effect(i32 2)
1481 endif: ; preds = %if, %elseif, %else
1482 br i1 %c, label %then, label %ret, !prof !12 ; 5 to 3
1484 then: ; preds = %endif
1485 call void @effect(i32 3)
1488 ret: ; preds = %endif, %then
1492 define i32 @not_rotate_if_extra_branch(i32 %count) {
1493 ; Test checks that there is no loop rotation
1494 ; if it introduces extra branch.
1495 ; Specifically in this case because best exit is .header
1496 ; but it has fallthrough to .middle block and last block in
1497 ; loop chain .slow does not have afallthrough to .header.
1498 ; CHECK-LABEL: not_rotate_if_extra_branch
1507 %sum.0 = shl nsw i32 %count, 1
1511 %i = phi i32 [ %i.1, %.backedge ], [ 0, %.entry ]
1512 %sum = phi i32 [ %sum.1, %.backedge ], [ %sum.0, %.entry ]
1513 %is_exc = icmp sgt i32 %i, 9000000
1514 br i1 %is_exc, label %.bailout, label %.middle, !prof !13
1517 %sum.2 = add nsw i32 %count, 1
1521 %pr.1 = and i32 %i, 1023
1522 %pr.2 = icmp eq i32 %pr.1, 0
1523 br i1 %pr.2, label %.slow, label %.backedge, !prof !14
1526 tail call void @effect(i32 %sum)
1530 %sum.1 = add nsw i32 %i, %sum
1531 %i.1 = add nsw i32 %i, 1
1532 %end = icmp slt i32 %i.1, %count
1533 br i1 %end, label %.header, label %.stop, !prof !15
1536 %sum.phi = phi i32 [ %sum.1, %.backedge ], [ %sum.2, %.bailout ]
1540 define i32 @not_rotate_if_extra_branch_regression(i32 %count, i32 %init) {
1541 ; This is a regression test against patch avoid loop rotation if
1542 ; it introduce an extra btanch.
1543 ; CHECK-LABEL: not_rotate_if_extra_branch_regression
1545 ; CHECK: %.first_backedge
1546 ; CHECK: %.second_header
1549 %sum.0 = shl nsw i32 %count, 1
1550 br label %.first_header
1553 %i = phi i32 [ %i.1, %.first_backedge ], [ 0, %.entry ]
1554 %is_bo1 = icmp sgt i32 %i, 9000000
1555 br i1 %is_bo1, label %.bailout, label %.first_backedge, !prof !14
1558 %i.1 = add nsw i32 %i, 1
1559 %end = icmp slt i32 %i.1, %count
1560 br i1 %end, label %.first_header, label %.second_header, !prof !13
1563 %j = phi i32 [ %j.1, %.second_backedge ], [ %init, %.first_backedge ]
1564 %end.2 = icmp sgt i32 %j, %count
1565 br i1 %end.2, label %.stop, label %.second_middle, !prof !14
1568 %is_slow = icmp sgt i32 %j, 9000000
1569 br i1 %is_slow, label %.slow, label %.second_backedge, !prof !14
1572 tail call void @effect(i32 %j)
1573 br label %.second_backedge
1576 %j.1 = add nsw i32 %j, 1
1577 %end.3 = icmp slt i32 %j, 10000000
1578 br i1 %end.3, label %.second_header, label %.stop, !prof !13
1581 %res = add nsw i32 %j, %i.1
1588 declare void @effect(i32)
1590 !5 = !{!"branch_weights", i32 84, i32 16}
1591 !6 = !{!"function_entry_count", i32 10}
1592 !7 = !{!"branch_weights", i32 60, i32 40}
1593 !8 = !{!"branch_weights", i32 5001, i32 4999}
1594 !9 = !{!"branch_weights", i32 85, i32 15}
1595 !10 = !{!"branch_weights", i32 90, i32 10}
1596 !11 = !{!"branch_weights", i32 1, i32 1}
1597 !12 = !{!"branch_weights", i32 5, i32 3}
1598 !13 = !{!"branch_weights", i32 1, i32 1}
1599 !14 = !{!"branch_weights", i32 1, i32 1023}
1600 !15 = !{!"branch_weights", i32 4095, i32 1}