1 ; RUN: llc -mtriple=i686-linux -pre-RA-sched=source < %s | FileCheck %s
3 declare void @error(i32 %i, i32 %a, i32 %b)
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-LABEL: test_ifchains:
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
33 call void @error(i32 %i, i32 1, i32 %b)
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
43 call void @error(i32 %i, i32 1, i32 %b)
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
53 call void @error(i32 %i, i32 1, i32 %b)
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
63 call void @error(i32 %i, i32 1, i32 %b)
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
73 call void @error(i32 %i, i32 1, i32 %b)
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-LABEL: test_loop_cold_blocks:
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
104 call void @error(i32 %i, i32 1, i32 %base)
108 %unlikelycond2 = icmp sgt i32 %base, 21
109 br i1 %unlikelycond2, label %unlikely2, label %body3, !prof !0
112 call void @error(i32 %i, i32 2, i32 %base)
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
127 !0 = !{!"branch_weights", i32 4, i32 64}
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-LABEL: test_loop_early_exits:
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
155 %bailcond2 = icmp eq i32 %base, 43
156 br i1 %bailcond2, label %bail2, label %body3
162 %bailcond3 = icmp eq i32 %base, 44
163 br i1 %bailcond3, label %bail3, label %body4
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
180 ; Tail duplication during layout can entirely remove body0 by duplicating it
181 ; into the entry block and into body1. This is a good thing but it isn't what
182 ; this test is looking for. So to make the blocks longer so they don't get
183 ; duplicated, we add some calls to dummy.
184 declare void @dummy()
186 define i32 @test_loop_rotate(i32 %i, i32* %a) {
187 ; Check that we rotate conditional exits from the loop to the bottom of the
188 ; loop, eliminating unconditional branches to the top.
189 ; CHECK-LABEL: test_loop_rotate:
199 %iv = phi i32 [ 0, %entry ], [ %next, %body1 ]
200 %base = phi i32 [ 0, %entry ], [ %sum, %body1 ]
201 %next = add i32 %iv, 1
202 %exitcond = icmp eq i32 %next, %i
205 br i1 %exitcond, label %exit, label %body1
208 %arrayidx = getelementptr inbounds i32, i32* %a, i32 %iv
209 %0 = load i32, i32* %arrayidx
210 %sum = add nsw i32 %0, %base
211 %bailcond1 = icmp eq i32 %sum, 42
218 define i32 @test_no_loop_rotate(i32 %i, i32* %a) {
219 ; Check that we don't try to rotate a loop which is already laid out with
220 ; fallthrough opportunities into the top and out of the bottom.
221 ; CHECK-LABEL: test_no_loop_rotate:
231 %iv = phi i32 [ 0, %entry ], [ %next, %body1 ]
232 %base = phi i32 [ 0, %entry ], [ %sum, %body1 ]
233 %arrayidx = getelementptr inbounds i32, i32* %a, i32 %iv
234 %0 = load i32, i32* %arrayidx
235 %sum = add nsw i32 %0, %base
236 %bailcond1 = icmp eq i32 %sum, 42
237 br i1 %bailcond1, label %exit, label %body1
240 %next = add i32 %iv, 1
241 %exitcond = icmp eq i32 %next, %i
242 br i1 %exitcond, label %exit, label %body0
248 define i32 @test_loop_align(i32 %i, i32* %a) {
249 ; Check that we provide basic loop body alignment with the block placement
251 ; CHECK-LABEL: test_loop_align:
253 ; CHECK: .p2align [[ALIGN:[0-9]+]],
261 %iv = phi i32 [ 0, %entry ], [ %next, %body ]
262 %base = phi i32 [ 0, %entry ], [ %sum, %body ]
263 %arrayidx = getelementptr inbounds i32, i32* %a, i32 %iv
264 %0 = load i32, i32* %arrayidx
265 %sum = add nsw i32 %0, %base
266 %next = add i32 %iv, 1
267 %exitcond = icmp eq i32 %next, %i
268 br i1 %exitcond, label %exit, label %body
274 define i32 @test_nested_loop_align(i32 %i, i32* %a, i32* %b) {
275 ; Check that we provide nested loop body alignment.
276 ; CHECK-LABEL: test_nested_loop_align:
278 ; CHECK: .p2align [[ALIGN]],
279 ; CHECK-NEXT: %loop.body.1
280 ; CHECK: .p2align [[ALIGN]],
281 ; CHECK-NEXT: %inner.loop.body
282 ; CHECK-NOT: .p2align
286 br label %loop.body.1
289 %iv = phi i32 [ 0, %entry ], [ %next, %loop.body.2 ]
290 %arrayidx = getelementptr inbounds i32, i32* %a, i32 %iv
291 %bidx = load i32, i32* %arrayidx
292 br label %inner.loop.body
295 %inner.iv = phi i32 [ 0, %loop.body.1 ], [ %inner.next, %inner.loop.body ]
296 %base = phi i32 [ 0, %loop.body.1 ], [ %sum, %inner.loop.body ]
297 %scaled_idx = mul i32 %bidx, %iv
298 %inner.arrayidx = getelementptr inbounds i32, i32* %b, i32 %scaled_idx
299 %0 = load i32, i32* %inner.arrayidx
300 %sum = add nsw i32 %0, %base
301 %inner.next = add i32 %iv, 1
302 %inner.exitcond = icmp eq i32 %inner.next, %i
303 br i1 %inner.exitcond, label %loop.body.2, label %inner.loop.body
306 %next = add i32 %iv, 1
307 %exitcond = icmp eq i32 %next, %i
308 br i1 %exitcond, label %exit, label %loop.body.1
314 define void @unnatural_cfg1() {
315 ; Test that we can handle a loop with an inner unnatural loop at the end of
316 ; a function. This is a gross CFG reduced out of the single source GCC.
317 ; CHECK-LABEL: unnatural_cfg1
324 br label %loop.header
330 br i1 undef, label %loop.body3, label %loop.body2
333 %ptr = load i32*, i32** undef, align 4
337 %myptr = phi i32* [ %ptr2, %loop.body5 ], [ %ptr, %loop.body2 ], [ undef, %loop.body1 ]
338 %bcmyptr = bitcast i32* %myptr to i32*
339 %val = load i32, i32* %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 i32*, i32** undef, align 4
351 define void @unnatural_cfg2() {
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
362 ; CHECK: %loop.inner2.begin
364 ; CHECK: %loop.inner1.begin
368 br label %loop.header
371 %comp0 = icmp eq i32* undef, null
372 br i1 %comp0, label %bail, label %loop.body1
375 %val0 = load i32*, i32** undef, align 4
376 br i1 undef, label %loop.body2, label %loop.inner1.begin
379 br i1 undef, label %loop.body4, label %loop.body3
382 %ptr1 = getelementptr inbounds i32, i32* %val0, i32 0
383 %castptr1 = bitcast i32* %ptr1 to i32**
384 %val1 = load i32*, i32** %castptr1, align 4
385 br label %loop.inner1.begin
388 %valphi = phi i32* [ %val2, %loop.inner1.end ], [ %val1, %loop.body3 ], [ %val0, %loop.body1 ]
389 %castval = bitcast i32* %valphi to i32*
390 %comp1 = icmp eq i32 undef, 48
391 br i1 %comp1, label %loop.inner1.end, label %loop.body4
394 %ptr2 = getelementptr inbounds i32, i32* %valphi, i32 0
395 %castptr2 = bitcast i32* %ptr2 to i32**
396 %val2 = load i32*, i32** %castptr2, align 4
397 br label %loop.inner1.begin
403 %comp2 = icmp ult i32 undef, 3
404 br i1 %comp2, label %loop.inner2.begin, label %loop.end
407 br i1 false, label %loop.end, label %loop.inner2.end
410 %comp3 = icmp eq i32 undef, 1769472
411 br i1 %comp3, label %loop.end, label %loop.inner2.begin
414 br label %loop.header
420 define i32 @problematic_switch() {
421 ; This function's CFG caused overlow in the machine branch probability
422 ; calculation, triggering asserts. Make sure we don't crash on it.
423 ; CHECK: problematic_switch
426 switch i32 undef, label %exit [
427 i32 879, label %bogus
467 %merge = phi i32 [ 3, %step ], [ 6, %entry ]
471 define void @fpcmp_unanalyzable_branch(i1 %cond) {
472 ; This function's CFG contains an once-unanalyzable branch (une on floating
473 ; points). As now it becomes analyzable, we should get best layout in which each
474 ; edge in 'entry' -> 'entry.if.then_crit_edge' -> 'if.then' -> 'if.end' is
476 ; CHECK-LABEL: fpcmp_unanalyzable_branch:
477 ; CHECK: # BB#0: # %entry
478 ; CHECK: # BB#1: # %entry.if.then_crit_edge
479 ; CHECK: .LBB10_5: # %if.then
480 ; CHECK: .LBB10_6: # %if.end
481 ; CHECK: # BB#3: # %exit
482 ; CHECK: jne .LBB10_4
483 ; CHECK-NEXT: jnp .LBB10_6
484 ; CHECK: jmp .LBB10_5
487 ; Note that this branch must be strongly biased toward
488 ; 'entry.if.then_crit_edge' to ensure that we would try to form a chain for
489 ; 'entry' -> 'entry.if.then_crit_edge' -> 'if.then' -> 'if.end'.
490 br i1 %cond, label %entry.if.then_crit_edge, label %lor.lhs.false, !prof !1
492 entry.if.then_crit_edge:
493 %.pre14 = load i8, i8* undef, align 1
497 br i1 undef, label %if.end, label %exit
500 %cmp.i = fcmp une double 0.000000e+00, undef
501 br i1 %cmp.i, label %if.then, label %if.end, !prof !3
504 %0 = phi i8 [ %.pre14, %entry.if.then_crit_edge ], [ undef, %exit ]
506 store i8 %1, i8* undef, align 4
513 !1 = !{!"branch_weights", i32 1000, i32 1}
514 !3 = !{!"branch_weights", i32 1, i32 1000}
518 declare i32 @h(i32 %x)
520 define i32 @test_global_cfg_break_profitability() {
521 ; Check that our metrics for the profitability of a CFG break are global rather
522 ; than local. A successor may be very hot, but if the current block isn't, it
523 ; doesn't matter. Within this test the 'then' block is slightly warmer than the
524 ; 'else' block, but not nearly enough to merit merging it with the exit block
525 ; even though the probability of 'then' branching to the 'exit' block is very
527 ; CHECK: test_global_cfg_break_profitability
528 ; CHECK: calll {{_?}}f
529 ; CHECK: calll {{_?}}g
530 ; CHECK: calll {{_?}}h
534 br i1 undef, label %then, label %else, !prof !2
537 %then.result = call i32 @f()
541 %else.result = call i32 @g()
545 %result = phi i32 [ %then.result, %then ], [ %else.result, %else ]
546 %result2 = call i32 @h(i32 %result)
550 !2 = !{!"branch_weights", i32 3, i32 1}
552 declare i32 @__gxx_personality_v0(...)
554 define void @test_eh_lpad_successor() personality i8* bitcast (i32 (...)* @__gxx_personality_v0 to i8*) {
555 ; Some times the landing pad ends up as the first successor of an invoke block.
556 ; When this happens, a strange result used to fall out of updateTerminators: we
557 ; didn't correctly locate the fallthrough successor, assuming blindly that the
558 ; first one was the fallthrough successor. As a result, we would add an
559 ; erroneous jump to the landing pad thinking *that* was the default successor.
560 ; CHECK-LABEL: test_eh_lpad_successor
566 invoke i32 @f() to label %preheader unwind label %lpad
572 %lpad.val = landingpad { i8*, i32 }
574 resume { i8*, i32 } %lpad.val
580 declare void @fake_throw() noreturn
582 define void @test_eh_throw() personality i8* bitcast (i32 (...)* @__gxx_personality_v0 to i8*) {
583 ; For blocks containing a 'throw' (or similar functionality), we have
584 ; a no-return invoke. In this case, only EH successors will exist, and
585 ; fallthrough simply won't occur. Make sure we don't crash trying to update
586 ; terminators for such constructs.
588 ; CHECK-LABEL: test_eh_throw
593 invoke void @fake_throw() to label %continue unwind label %cleanup
599 %0 = landingpad { i8*, i32 }
604 define void @test_unnatural_cfg_backwards_inner_loop() {
605 ; Test that when we encounter an unnatural CFG structure after having formed
606 ; a chain for an inner loop which happened to be laid out backwards we don't
607 ; attempt to merge onto the wrong end of the inner loop just because we find it
608 ; first. This was reduced from a crasher in GCC's single source.
610 ; CHECK-LABEL: test_unnatural_cfg_backwards_inner_loop
616 br i1 undef, label %loop2a, label %body
622 %next.load = load i32*, i32** undef
623 br i1 %comp.a, label %loop2a, label %loop2b
626 %var = phi i32* [ null, %entry ], [ null, %body ], [ %next.phi, %loop1 ]
627 %next.var = phi i32* [ null, %entry ], [ undef, %body ], [ %next.load, %loop1 ]
628 %comp.a = icmp eq i32* %var, null
632 %gep = getelementptr inbounds i32, i32* %var.phi, i32 0
633 %next.ptr = bitcast i32* %gep to i32**
634 store i32* %next.phi, i32** %next.ptr
638 %var.phi = phi i32* [ %next.phi, %loop2b ], [ %var, %loop2a ]
639 %next.phi = phi i32* [ %next.load, %loop2b ], [ %next.var, %loop2a ]
643 define void @unanalyzable_branch_to_loop_header() {
644 ; Ensure that we can handle unanalyzable branches into loop headers. We
645 ; pre-form chains for unanalyzable branches, and will find the tail end of that
646 ; at the start of the loop. This function uses floating point comparison
647 ; fallthrough because that happens to always produce unanalyzable branches on
650 ; CHECK-LABEL: unanalyzable_branch_to_loop_header
656 %cmp = fcmp une double 0.000000e+00, undef
657 br i1 %cmp, label %loop, label %exit
660 %cond = icmp eq i8 undef, 42
661 br i1 %cond, label %exit, label %loop
667 define void @unanalyzable_branch_to_best_succ(i1 %cond) {
668 ; Ensure that we can handle unanalyzable branches where the destination block
669 ; gets selected as the optimal successor to merge.
671 ; This branch is now analyzable and hence the destination block becomes the
672 ; hotter one. The right order is entry->bar->exit->foo.
674 ; CHECK-LABEL: unanalyzable_branch_to_best_succ
681 ; Bias this branch toward bar to ensure we form that chain.
682 br i1 %cond, label %bar, label %foo, !prof !1
685 %cmp = fcmp une double 0.000000e+00, undef
686 br i1 %cmp, label %bar, label %exit
696 define void @unanalyzable_branch_to_free_block(float %x) {
697 ; Ensure that we can handle unanalyzable branches where the destination block
698 ; gets selected as the best free block in the CFG.
700 ; CHECK-LABEL: unanalyzable_branch_to_free_block
708 br i1 undef, label %a, label %b
715 %cmp = fcmp une float %x, undef
716 br i1 %cmp, label %c, label %exit
726 define void @many_unanalyzable_branches() {
727 ; Ensure that we don't crash as we're building up many unanalyzable branches,
730 ; CHECK-LABEL: many_unanalyzable_branches
737 %val0 = load volatile float, float* undef
738 %cmp0 = fcmp une float %val0, undef
739 br i1 %cmp0, label %1, label %0
740 %val1 = load volatile float, float* undef
741 %cmp1 = fcmp une float %val1, undef
742 br i1 %cmp1, label %2, label %1
743 %val2 = load volatile float, float* undef
744 %cmp2 = fcmp une float %val2, undef
745 br i1 %cmp2, label %3, label %2
746 %val3 = load volatile float, float* undef
747 %cmp3 = fcmp une float %val3, undef
748 br i1 %cmp3, label %4, label %3
749 %val4 = load volatile float, float* undef
750 %cmp4 = fcmp une float %val4, undef
751 br i1 %cmp4, label %5, label %4
752 %val5 = load volatile float, float* undef
753 %cmp5 = fcmp une float %val5, undef
754 br i1 %cmp5, label %6, label %5
755 %val6 = load volatile float, float* undef
756 %cmp6 = fcmp une float %val6, undef
757 br i1 %cmp6, label %7, label %6
758 %val7 = load volatile float, float* undef
759 %cmp7 = fcmp une float %val7, undef
760 br i1 %cmp7, label %8, label %7
761 %val8 = load volatile float, float* undef
762 %cmp8 = fcmp une float %val8, undef
763 br i1 %cmp8, label %9, label %8
764 %val9 = load volatile float, float* undef
765 %cmp9 = fcmp une float %val9, undef
766 br i1 %cmp9, label %10, label %9
767 %val10 = load volatile float, float* undef
768 %cmp10 = fcmp une float %val10, undef
769 br i1 %cmp10, label %11, label %10
770 %val11 = load volatile float, float* undef
771 %cmp11 = fcmp une float %val11, undef
772 br i1 %cmp11, label %12, label %11
773 %val12 = load volatile float, float* undef
774 %cmp12 = fcmp une float %val12, undef
775 br i1 %cmp12, label %13, label %12
776 %val13 = load volatile float, float* undef
777 %cmp13 = fcmp une float %val13, undef
778 br i1 %cmp13, label %14, label %13
779 %val14 = load volatile float, float* undef
780 %cmp14 = fcmp une float %val14, undef
781 br i1 %cmp14, label %15, label %14
782 %val15 = load volatile float, float* undef
783 %cmp15 = fcmp une float %val15, undef
784 br i1 %cmp15, label %16, label %15
785 %val16 = load volatile float, float* undef
786 %cmp16 = fcmp une float %val16, undef
787 br i1 %cmp16, label %17, label %16
788 %val17 = load volatile float, float* undef
789 %cmp17 = fcmp une float %val17, undef
790 br i1 %cmp17, label %18, label %17
791 %val18 = load volatile float, float* undef
792 %cmp18 = fcmp une float %val18, undef
793 br i1 %cmp18, label %19, label %18
794 %val19 = load volatile float, float* undef
795 %cmp19 = fcmp une float %val19, undef
796 br i1 %cmp19, label %20, label %19
797 %val20 = load volatile float, float* undef
798 %cmp20 = fcmp une float %val20, undef
799 br i1 %cmp20, label %21, label %20
800 %val21 = load volatile float, float* undef
801 %cmp21 = fcmp une float %val21, undef
802 br i1 %cmp21, label %22, label %21
803 %val22 = load volatile float, float* undef
804 %cmp22 = fcmp une float %val22, undef
805 br i1 %cmp22, label %23, label %22
806 %val23 = load volatile float, float* undef
807 %cmp23 = fcmp une float %val23, undef
808 br i1 %cmp23, label %24, label %23
809 %val24 = load volatile float, float* undef
810 %cmp24 = fcmp une float %val24, undef
811 br i1 %cmp24, label %25, label %24
812 %val25 = load volatile float, float* undef
813 %cmp25 = fcmp une float %val25, undef
814 br i1 %cmp25, label %26, label %25
815 %val26 = load volatile float, float* undef
816 %cmp26 = fcmp une float %val26, undef
817 br i1 %cmp26, label %27, label %26
818 %val27 = load volatile float, float* undef
819 %cmp27 = fcmp une float %val27, undef
820 br i1 %cmp27, label %28, label %27
821 %val28 = load volatile float, float* undef
822 %cmp28 = fcmp une float %val28, undef
823 br i1 %cmp28, label %29, label %28
824 %val29 = load volatile float, float* undef
825 %cmp29 = fcmp une float %val29, undef
826 br i1 %cmp29, label %30, label %29
827 %val30 = load volatile float, float* undef
828 %cmp30 = fcmp une float %val30, undef
829 br i1 %cmp30, label %31, label %30
830 %val31 = load volatile float, float* undef
831 %cmp31 = fcmp une float %val31, undef
832 br i1 %cmp31, label %32, label %31
833 %val32 = load volatile float, float* undef
834 %cmp32 = fcmp une float %val32, undef
835 br i1 %cmp32, label %33, label %32
836 %val33 = load volatile float, float* undef
837 %cmp33 = fcmp une float %val33, undef
838 br i1 %cmp33, label %34, label %33
839 %val34 = load volatile float, float* undef
840 %cmp34 = fcmp une float %val34, undef
841 br i1 %cmp34, label %35, label %34
842 %val35 = load volatile float, float* undef
843 %cmp35 = fcmp une float %val35, undef
844 br i1 %cmp35, label %36, label %35
845 %val36 = load volatile float, float* undef
846 %cmp36 = fcmp une float %val36, undef
847 br i1 %cmp36, label %37, label %36
848 %val37 = load volatile float, float* undef
849 %cmp37 = fcmp une float %val37, undef
850 br i1 %cmp37, label %38, label %37
851 %val38 = load volatile float, float* undef
852 %cmp38 = fcmp une float %val38, undef
853 br i1 %cmp38, label %39, label %38
854 %val39 = load volatile float, float* undef
855 %cmp39 = fcmp une float %val39, undef
856 br i1 %cmp39, label %40, label %39
857 %val40 = load volatile float, float* undef
858 %cmp40 = fcmp une float %val40, undef
859 br i1 %cmp40, label %41, label %40
860 %val41 = load volatile float, float* undef
861 %cmp41 = fcmp une float %val41, undef
862 br i1 %cmp41, label %42, label %41
863 %val42 = load volatile float, float* undef
864 %cmp42 = fcmp une float %val42, undef
865 br i1 %cmp42, label %43, label %42
866 %val43 = load volatile float, float* undef
867 %cmp43 = fcmp une float %val43, undef
868 br i1 %cmp43, label %44, label %43
869 %val44 = load volatile float, float* undef
870 %cmp44 = fcmp une float %val44, undef
871 br i1 %cmp44, label %45, label %44
872 %val45 = load volatile float, float* undef
873 %cmp45 = fcmp une float %val45, undef
874 br i1 %cmp45, label %46, label %45
875 %val46 = load volatile float, float* undef
876 %cmp46 = fcmp une float %val46, undef
877 br i1 %cmp46, label %47, label %46
878 %val47 = load volatile float, float* undef
879 %cmp47 = fcmp une float %val47, undef
880 br i1 %cmp47, label %48, label %47
881 %val48 = load volatile float, float* undef
882 %cmp48 = fcmp une float %val48, undef
883 br i1 %cmp48, label %49, label %48
884 %val49 = load volatile float, float* undef
885 %cmp49 = fcmp une float %val49, undef
886 br i1 %cmp49, label %50, label %49
887 %val50 = load volatile float, float* undef
888 %cmp50 = fcmp une float %val50, undef
889 br i1 %cmp50, label %51, label %50
890 %val51 = load volatile float, float* undef
891 %cmp51 = fcmp une float %val51, undef
892 br i1 %cmp51, label %52, label %51
893 %val52 = load volatile float, float* undef
894 %cmp52 = fcmp une float %val52, undef
895 br i1 %cmp52, label %53, label %52
896 %val53 = load volatile float, float* undef
897 %cmp53 = fcmp une float %val53, undef
898 br i1 %cmp53, label %54, label %53
899 %val54 = load volatile float, float* undef
900 %cmp54 = fcmp une float %val54, undef
901 br i1 %cmp54, label %55, label %54
902 %val55 = load volatile float, float* undef
903 %cmp55 = fcmp une float %val55, undef
904 br i1 %cmp55, label %56, label %55
905 %val56 = load volatile float, float* undef
906 %cmp56 = fcmp une float %val56, undef
907 br i1 %cmp56, label %57, label %56
908 %val57 = load volatile float, float* undef
909 %cmp57 = fcmp une float %val57, undef
910 br i1 %cmp57, label %58, label %57
911 %val58 = load volatile float, float* undef
912 %cmp58 = fcmp une float %val58, undef
913 br i1 %cmp58, label %59, label %58
914 %val59 = load volatile float, float* undef
915 %cmp59 = fcmp une float %val59, undef
916 br i1 %cmp59, label %60, label %59
917 %val60 = load volatile float, float* undef
918 %cmp60 = fcmp une float %val60, undef
919 br i1 %cmp60, label %61, label %60
920 %val61 = load volatile float, float* undef
921 %cmp61 = fcmp une float %val61, undef
922 br i1 %cmp61, label %62, label %61
923 %val62 = load volatile float, float* undef
924 %cmp62 = fcmp une float %val62, undef
925 br i1 %cmp62, label %63, label %62
926 %val63 = load volatile float, float* undef
927 %cmp63 = fcmp une float %val63, undef
928 br i1 %cmp63, label %64, label %63
929 %val64 = load volatile float, float* undef
930 %cmp64 = fcmp une float %val64, undef
931 br i1 %cmp64, label %65, label %64
938 define void @benchmark_heapsort(i32 %n, double* nocapture %ra) {
939 ; This test case comes from the heapsort benchmark, and exemplifies several
940 ; important aspects to block placement in the presence of loops:
941 ; 1) Loop rotation needs to *ensure* that the desired exiting edge can be
943 ; 2) The exiting edge from the loop which is rotated to be laid out at the
944 ; bottom of the loop needs to be exiting into the nearest enclosing loop (to
945 ; which there is an exit). Otherwise, we force that enclosing loop into
946 ; strange layouts that are siginificantly less efficient, often times making
949 ; CHECK-LABEL: @benchmark_heapsort
951 ; First rotated loop top.
954 ; %for.cond gets completely tail-duplicated away.
958 ; Second rotated loop top
961 ; CHECK: %while.cond.outer
962 ; Third rotated loop top
966 ; CHECK: %land.lhs.true
973 %shr = ashr i32 %n, 1
974 %add = add nsw i32 %shr, 1
975 %arrayidx3 = getelementptr inbounds double, double* %ra, i64 1
979 %ir.0 = phi i32 [ %n, %entry ], [ %ir.1, %while.end ]
980 %l.0 = phi i32 [ %add, %entry ], [ %l.1, %while.end ]
981 %cmp = icmp sgt i32 %l.0, 1
982 br i1 %cmp, label %if.then, label %if.else
985 %dec = add nsw i32 %l.0, -1
986 %idxprom = sext i32 %dec to i64
987 %arrayidx = getelementptr inbounds double, double* %ra, i64 %idxprom
988 %0 = load double, double* %arrayidx, align 8
992 %idxprom1 = sext i32 %ir.0 to i64
993 %arrayidx2 = getelementptr inbounds double, double* %ra, i64 %idxprom1
994 %1 = load double, double* %arrayidx2, align 8
995 %2 = load double, double* %arrayidx3, align 8
996 store double %2, double* %arrayidx2, align 8
997 %dec6 = add nsw i32 %ir.0, -1
998 %cmp7 = icmp eq i32 %dec6, 1
999 br i1 %cmp7, label %if.then8, label %if.end10
1002 store double %1, double* %arrayidx3, align 8
1006 %ir.1 = phi i32 [ %ir.0, %if.then ], [ %dec6, %if.else ]
1007 %l.1 = phi i32 [ %dec, %if.then ], [ %l.0, %if.else ]
1008 %rra.0 = phi double [ %0, %if.then ], [ %1, %if.else ]
1009 %add31 = add nsw i32 %ir.1, 1
1010 br label %while.cond.outer
1013 %j.0.ph.in = phi i32 [ %l.1, %if.end10 ], [ %j.1, %if.then24 ]
1014 %j.0.ph = shl i32 %j.0.ph.in, 1
1015 br label %while.cond
1018 %j.0 = phi i32 [ %add31, %if.end20 ], [ %j.0.ph, %while.cond.outer ]
1019 %cmp11 = icmp sgt i32 %j.0, %ir.1
1020 br i1 %cmp11, label %while.end, label %while.body
1023 %cmp12 = icmp slt i32 %j.0, %ir.1
1024 br i1 %cmp12, label %land.lhs.true, label %if.end20
1027 %idxprom13 = sext i32 %j.0 to i64
1028 %arrayidx14 = getelementptr inbounds double, double* %ra, i64 %idxprom13
1029 %3 = load double, double* %arrayidx14, align 8
1030 %add15 = add nsw i32 %j.0, 1
1031 %idxprom16 = sext i32 %add15 to i64
1032 %arrayidx17 = getelementptr inbounds double, double* %ra, i64 %idxprom16
1033 %4 = load double, double* %arrayidx17, align 8
1034 %cmp18 = fcmp olt double %3, %4
1035 br i1 %cmp18, label %if.then19, label %if.end20
1041 %j.1 = phi i32 [ %add15, %if.then19 ], [ %j.0, %land.lhs.true ], [ %j.0, %while.body ]
1042 %idxprom21 = sext i32 %j.1 to i64
1043 %arrayidx22 = getelementptr inbounds double, double* %ra, i64 %idxprom21
1044 %5 = load double, double* %arrayidx22, align 8
1045 %cmp23 = fcmp olt double %rra.0, %5
1046 br i1 %cmp23, label %if.then24, label %while.cond
1049 %idxprom27 = sext i32 %j.0.ph.in to i64
1050 %arrayidx28 = getelementptr inbounds double, double* %ra, i64 %idxprom27
1051 store double %5, double* %arrayidx28, align 8
1052 br label %while.cond.outer
1055 %idxprom33 = sext i32 %j.0.ph.in to i64
1056 %arrayidx34 = getelementptr inbounds double, double* %ra, i64 %idxprom33
1057 store double %rra.0, double* %arrayidx34, align 8
1061 declare void @cold_function() cold
1063 define i32 @test_cold_calls(i32* %a) {
1064 ; Test that edges to blocks post-dominated by cold calls are
1065 ; marked as not expected to be taken. They should be laid out
1067 ; CHECK-LABEL: test_cold_calls:
1074 %gep1 = getelementptr i32, i32* %a, i32 1
1075 %val1 = load i32, i32* %gep1
1076 %cond1 = icmp ugt i32 %val1, 1
1077 br i1 %cond1, label %then, label %else
1080 call void @cold_function()
1084 %gep2 = getelementptr i32, i32* %a, i32 2
1085 %val2 = load i32, i32* %gep2
1089 %ret = phi i32 [ %val1, %then ], [ %val2, %else ]
1093 ; Make sure we put landingpads out of the way.
1094 declare i32 @pers(...)
1100 define i32 @test_lp(i32 %a) personality i32 (...)* @pers {
1101 ; CHECK-LABEL: test_lp:
1110 %0 = icmp sgt i32 %a, 1
1111 br i1 %0, label %hot, label %cold, !prof !4
1114 %1 = invoke i32 @foo()
1115 to label %then unwind label %hotlp
1118 %2 = invoke i32 @bar()
1119 to label %then unwind label %coldlp
1122 %3 = phi i32 [ %1, %hot ], [ %2, %cold ]
1126 %4 = landingpad { i8*, i32 }
1131 %5 = landingpad { i8*, i32 }
1136 %6 = phi i32 [-1, %hotlp], [-2, %coldlp]
1141 !4 = !{!"branch_weights", i32 65536, i32 0}
1143 ; Make sure that ehpad are scheduled from the least probable one
1144 ; to the most probable one. See selectBestCandidateBlock as to why.
1145 declare void @clean();
1147 define void @test_flow_unwind() personality i32 (...)* @pers {
1148 ; CHECK-LABEL: test_flow_unwind:
1154 ; CHECK: %outercleanup
1156 %0 = invoke i32 @foo()
1157 to label %then unwind label %outerlp
1160 %1 = invoke i32 @bar()
1161 to label %exit unwind label %innerlp
1167 %2 = landingpad { i8*, i32 }
1169 br label %innercleanup
1172 %3 = landingpad { i8*, i32 }
1174 br label %outercleanup
1177 %4 = phi { i8*, i32 } [%2, %innercleanup], [%3, %outerlp]
1179 resume { i8*, i32 } %4
1183 br label %outercleanup
1186 declare void @hot_function()
1188 define void @test_hot_branch(i32* %a) {
1189 ; Test that a hot branch that has a probability a little larger than 80% will
1190 ; break CFG constrains when doing block placement.
1191 ; CHECK-LABEL: test_hot_branch:
1198 %gep1 = getelementptr i32, i32* %a, i32 1
1199 %val1 = load i32, i32* %gep1
1200 %cond1 = icmp ugt i32 %val1, 1
1201 br i1 %cond1, label %then, label %else, !prof !5
1204 call void @hot_function()
1208 call void @cold_function()
1212 call void @hot_function()
1216 define void @test_hot_branch_profile(i32* %a) !prof !6 {
1217 ; Test that a hot branch that has a probability a little larger than 50% will
1218 ; break CFG constrains when doing block placement when profile is available.
1219 ; CHECK-LABEL: test_hot_branch_profile:
1226 %gep1 = getelementptr i32, i32* %a, i32 1
1227 %val1 = load i32, i32* %gep1
1228 %cond1 = icmp ugt i32 %val1, 1
1229 br i1 %cond1, label %then, label %else, !prof !7
1232 call void @hot_function()
1236 call void @cold_function()
1240 call void @hot_function()
1244 define void @test_hot_branch_triangle_profile(i32* %a) !prof !6 {
1245 ; Test that a hot branch that has a probability a little larger than 80% will
1246 ; break triangle shaped CFG constrains when doing block placement if profile
1248 ; CHECK-LABEL: test_hot_branch_triangle_profile:
1254 %gep1 = getelementptr i32, i32* %a, i32 1
1255 %val1 = load i32, i32* %gep1
1256 %cond1 = icmp ugt i32 %val1, 1
1257 br i1 %cond1, label %exit, label %then, !prof !5
1260 call void @hot_function()
1264 call void @hot_function()
1268 define void @test_hot_branch_triangle_profile_topology(i32* %a) !prof !6 {
1269 ; Test that a hot branch that has a probability between 50% and 66% will not
1270 ; break triangle shaped CFG constrains when doing block placement if profile
1272 ; CHECK-LABEL: test_hot_branch_triangle_profile_topology:
1278 %gep1 = getelementptr i32, i32* %a, i32 1
1279 %val1 = load i32, i32* %gep1
1280 %cond1 = icmp ugt i32 %val1, 1
1281 br i1 %cond1, label %exit, label %then, !prof !7
1284 call void @hot_function()
1288 call void @hot_function()
1295 define void @test_forked_hot_diamond(i32* %a) {
1296 ; Test that a hot-branch with probability > 80% followed by a 50/50 branch
1297 ; will not place the cold predecessor if the probability for the fallthrough
1299 ; CHECK-LABEL: test_forked_hot_diamond
1307 %gep1 = getelementptr i32, i32* %a, i32 1
1308 %val1 = load i32, i32* %gep1
1309 %cond1 = icmp ugt i32 %val1, 1
1310 br i1 %cond1, label %then, label %else, !prof !5
1313 call void @hot_function()
1314 %gep2 = getelementptr i32, i32* %a, i32 2
1315 %val2 = load i32, i32* %gep2
1316 %cond2 = icmp ugt i32 %val2, 2
1317 br i1 %cond2, label %fork1, label %fork2, !prof !8
1320 call void @cold_function()
1321 %gep3 = getelementptr i32, i32* %a, i32 3
1322 %val3 = load i32, i32* %gep3
1323 %cond3 = icmp ugt i32 %val3, 3
1324 br i1 %cond3, label %fork1, label %fork2, !prof !8
1335 call void @hot_function()
1339 define void @test_forked_hot_diamond_gets_cold(i32* %a) {
1340 ; Test that a hot-branch with probability > 80% followed by a 50/50 branch
1341 ; will place the cold predecessor if the probability for the fallthrough
1343 ; The probability for both branches is 85%. For then2 vs else1
1344 ; this results in a compounded probability of 83%.
1345 ; Neither then2->fork1 nor then2->fork2 has a large enough relative
1346 ; probability to break the CFG.
1348 ; then2 -> fork1 vs else1 -> fork1 = 71%
1349 ; then2 -> fork2 vs else2 -> fork2 = 74%
1350 ; CHECK-LABEL: test_forked_hot_diamond_gets_cold
1360 %gep1 = getelementptr i32, i32* %a, i32 1
1361 %val1 = load i32, i32* %gep1
1362 %cond1 = icmp ugt i32 %val1, 1
1363 br i1 %cond1, label %then1, label %else1, !prof !9
1366 call void @hot_function()
1367 %gep2 = getelementptr i32, i32* %a, i32 2
1368 %val2 = load i32, i32* %gep2
1369 %cond2 = icmp ugt i32 %val2, 2
1370 br i1 %cond2, label %then2, label %else2, !prof !9
1373 call void @cold_function()
1377 call void @hot_function()
1378 %gep3 = getelementptr i32, i32* %a, i32 3
1379 %val3 = load i32, i32* %gep2
1380 %cond3 = icmp ugt i32 %val2, 3
1381 br i1 %cond3, label %fork1, label %fork2, !prof !8
1384 call void @cold_function()
1396 call void @hot_function()
1400 define void @test_forked_hot_diamond_stays_hot(i32* %a) {
1401 ; Test that a hot-branch with probability > 88.88% (1:8) followed by a 50/50
1402 ; branch will not place the cold predecessor as the probability for the
1403 ; fallthrough stays above 80%
1404 ; (1:8) followed by (1:1) is still (1:4)
1405 ; Here we use 90% probability because two in a row
1406 ; have a 89 % probability vs the original branch.
1407 ; CHECK-LABEL: test_forked_hot_diamond_stays_hot
1417 %gep1 = getelementptr i32, i32* %a, i32 1
1418 %val1 = load i32, i32* %gep1
1419 %cond1 = icmp ugt i32 %val1, 1
1420 br i1 %cond1, label %then1, label %else1, !prof !10
1423 call void @hot_function()
1424 %gep2 = getelementptr i32, i32* %a, i32 2
1425 %val2 = load i32, i32* %gep2
1426 %cond2 = icmp ugt i32 %val2, 2
1427 br i1 %cond2, label %then2, label %else2, !prof !10
1430 call void @cold_function()
1434 call void @hot_function()
1435 %gep3 = getelementptr i32, i32* %a, i32 3
1436 %val3 = load i32, i32* %gep2
1437 %cond3 = icmp ugt i32 %val2, 3
1438 br i1 %cond3, label %fork1, label %fork2, !prof !8
1441 call void @cold_function()
1453 call void @hot_function()
1457 ; Because %endif has a higher frequency than %if, the calculations show we
1458 ; shouldn't tail-duplicate %endif so that we can place it after %if. We were
1459 ; previously undercounting the cost by ignoring execution frequency that didn't
1460 ; come from the %if->%endif path.
1461 ; CHECK-LABEL: higher_frequency_succ_tail_dup
1468 define void @higher_frequency_succ_tail_dup(i1 %a, i1 %b, i1 %c) {
1471 if: ; preds = %entry
1472 call void @effect(i32 0)
1473 br i1 %a, label %elseif, label %endif, !prof !11 ; even
1475 elseif: ; preds = %if
1476 call void @effect(i32 1)
1477 br i1 %b, label %else, label %endif, !prof !11 ; even
1479 else: ; preds = %elseif
1480 call void @effect(i32 2)
1483 endif: ; preds = %if, %elseif, %else
1484 br i1 %c, label %then, label %ret, !prof !12 ; 5 to 3
1486 then: ; preds = %endif
1487 call void @effect(i32 3)
1490 ret: ; preds = %endif, %then
1494 define i32 @not_rotate_if_extra_branch(i32 %count) {
1495 ; Test checks that there is no loop rotation
1496 ; if it introduces extra branch.
1497 ; Specifically in this case because best exit is .header
1498 ; but it has fallthrough to .middle block and last block in
1499 ; loop chain .slow does not have afallthrough to .header.
1500 ; CHECK-LABEL: not_rotate_if_extra_branch
1509 %sum.0 = shl nsw i32 %count, 1
1513 %i = phi i32 [ %i.1, %.backedge ], [ 0, %.entry ]
1514 %sum = phi i32 [ %sum.1, %.backedge ], [ %sum.0, %.entry ]
1515 %is_exc = icmp sgt i32 %i, 9000000
1516 br i1 %is_exc, label %.bailout, label %.middle, !prof !13
1519 %sum.2 = add nsw i32 %count, 1
1523 %pr.1 = and i32 %i, 1023
1524 %pr.2 = icmp eq i32 %pr.1, 0
1525 br i1 %pr.2, label %.slow, label %.backedge, !prof !14
1528 tail call void @effect(i32 %sum)
1532 %sum.1 = add nsw i32 %i, %sum
1533 %i.1 = add nsw i32 %i, 1
1534 %end = icmp slt i32 %i.1, %count
1535 br i1 %end, label %.header, label %.stop, !prof !15
1538 %sum.phi = phi i32 [ %sum.1, %.backedge ], [ %sum.2, %.bailout ]
1542 define i32 @not_rotate_if_extra_branch_regression(i32 %count, i32 %init) {
1543 ; This is a regression test against patch avoid loop rotation if
1544 ; it introduce an extra btanch.
1545 ; CHECK-LABEL: not_rotate_if_extra_branch_regression
1547 ; CHECK: %.first_backedge
1549 ; CHECK: %.second_header
1551 %sum.0 = shl nsw i32 %count, 1
1552 br label %.first_header
1555 %i = phi i32 [ %i.1, %.first_backedge ], [ 0, %.entry ]
1556 %is_bo1 = icmp sgt i32 %i, 9000000
1557 br i1 %is_bo1, label %.bailout, label %.first_backedge, !prof !14
1560 %i.1 = add nsw i32 %i, 1
1561 %end = icmp slt i32 %i.1, %count
1562 br i1 %end, label %.first_header, label %.second_header, !prof !13
1565 %j = phi i32 [ %j.1, %.second_backedge ], [ %init, %.first_backedge ]
1566 %end.2 = icmp sgt i32 %j, %count
1567 br i1 %end.2, label %.stop, label %.second_middle, !prof !14
1570 %is_slow = icmp sgt i32 %j, 9000000
1571 br i1 %is_slow, label %.slow, label %.second_backedge, !prof !14
1574 tail call void @effect(i32 %j)
1575 br label %.second_backedge
1578 %j.1 = add nsw i32 %j, 1
1579 %end.3 = icmp slt i32 %j, 10000000
1580 br i1 %end.3, label %.second_header, label %.stop, !prof !13
1583 %res = add nsw i32 %j, %i.1
1590 declare void @effect(i32)
1592 !5 = !{!"branch_weights", i32 84, i32 16}
1593 !6 = !{!"function_entry_count", i32 10}
1594 !7 = !{!"branch_weights", i32 60, i32 40}
1595 !8 = !{!"branch_weights", i32 5001, i32 4999}
1596 !9 = !{!"branch_weights", i32 85, i32 15}
1597 !10 = !{!"branch_weights", i32 90, i32 10}
1598 !11 = !{!"branch_weights", i32 1, i32 1}
1599 !12 = !{!"branch_weights", i32 5, i32 3}
1600 !13 = !{!"branch_weights", i32 1, i32 1}
1601 !14 = !{!"branch_weights", i32 1, i32 1023}
1602 !15 = !{!"branch_weights", i32 4095, i32 1}