On Tue, Oct 25, 2011 at 2:15 PM, Tom de Vries <tom_devr...@mentor.com> wrote:
> Richard,
>
> I have a patch for the i686 bootstrap problem reported in PR50763 comment 10.
>
> pr50763-2.c looks like this before tail_merge_optimize:
> ...
> std_canonical_va_list_type (union tree_node * typeD.1608)
> {
> _BoolD.1576 pretmp.6D.2739;
> union tree_node * pretmp.5D.2738;
> _BoolD.1576 pretmp.4D.2737;
> union tree_node * pretmp.3D.2736;
> intD.6 D.2734;
> _BoolD.1576 D.2733;
> union tree_node * D.2725;
> intD.6 D.2722;
> intD.6 D.2717;
>
> # BLOCK 2 freq:10000
> # PRED: ENTRY [100.0%] (fallthru,exec)
> # VUSE <.MEMD.2730_12(D)>
> D.2717_4 = typeD.1608_3(D)->baseD.1605.codeD.1597;
> if (D.2717_4 != 0)
> goto <bb 3>;
> else
> goto <bb 4>;
> # SUCC: 3 [50.0%] (true,exec) 4 [50.0%] (false,exec)
>
> # BLOCK 3 freq:5000
> # PRED: 2 [50.0%] (true,exec)
> # VUSE <.MEMD.2730_12(D)>
> # PT = anything
> typeD.1608_5 = typeD.1608_3(D)->typedD.1606.typeD.1600;
> goto <bb 6>;
> # SUCC: 6 [100.0%] (fallthru,exec)
>
> # BLOCK 4 freq:5000
> # PRED: 2 [50.0%] (false,exec)
> # VUSE <.MEMD.2730_12(D)>
> # PT = anything
> typeD.1608_6 = typeD.1608_3(D)->typedD.1606.typeD.1600;
> # VUSE <.MEMD.2730_12(D)>
> D.2722_7 = typeD.1608_6->baseD.1605.codeD.1597;
> if (D.2722_7 != 0)
> goto <bb 5>;
> else
> goto <bb 7>;
> # SUCC: 5 [50.0%] (true,exec) 7 [50.0%] (false,exec)
>
> # BLOCK 7 freq:2500
> # PRED: 4 [50.0%] (false,exec)
> goto <bb 6>;
> # SUCC: 6 [100.0%] (fallthru)
>
> # BLOCK 5 freq:2500
> # PRED: 4 [50.0%] (true,exec)
> # SUCC: 6 [100.0%] (fallthru,exec)
>
> # BLOCK 6 freq:10000
> # PRED: 3 [100.0%] (fallthru,exec) 7 [100.0%] (fallthru) 5 [100.0%]
> (fallthru,exec)
> # PT = anything
>
> # typeD.1608_1 = PHI <typeD.1608_5(3), typeD.1608_3(D)(7), typeD.1608_6(5)>
> # VUSE <.MEMD.2730_12(D)>
> # PT = anything
> D.2725_9 = typeD.1608_1->type_commonD.1607.main_variantD.1604;
> D.2733_15 = D.2725_9 != 0B;
> D.2734_14 = (intD.6) D.2733_15;
> # VUSE <.MEMD.2730_12(D)>
> return D.2734_14;
> # SUCC: EXIT [100.0%]
>
> }
> ...
>
> tail_merge_optimize discovers that block 3 and 5 are identical, and removes
> block 5.
> In replace_block_by, we end up with phi_vuse1 == NULL_TREE and
> phi_vuse2 == .MEMD.2730_12(D).
>
> After the original fix to PR50763, this forces us to update the vuses, while
> that is not necessary.
>
> Attached patch fixes this problem by determining more precisely when the vuses
> need to be updated.
>
> bootstrapped and reg-tested on x86_64 and i686.
>
> OK for trunk?
Ok.
Thanks,
Richard.
> Thanks,
> - Tom
>
> 2011-10-25 Tom de Vries <t...@codesourcery.com>
>
> PR tree-optimization/50763
> * tree-ssa-tail-merge.c (replace_block_by): Update vops if phi_vuse1 or
> phi_vuse2 is NULL_TREE only if bb1 dominates or is dominated by bb2.
>
> * gcc.dg/pr50763-2.c: New test.
>
