[PATCH] D15548: Remove the restriction that known and unknown probabilities cannot coexist when being normalized.

Xinliang David Li via llvm-commits llvm-commits at lists.llvm.org
Wed Dec 16 10:55:57 PST 2015


On Wed, Dec 16, 2015 at 10:53 AM, Cong Hou <congh at google.com> wrote:
> On Wed, Dec 16, 2015 at 10:45 AM, Xinliang David Li <davidxl at google.com>
> wrote:
>>
>> On Wed, Dec 16, 2015 at 10:41 AM, Cong Hou <congh at google.com> wrote:
>> > congh added a comment.
>> >
>> > In http://reviews.llvm.org/D15548#312148, @davidxl wrote:
>> >
>> >> While it makes sense to do this -- can we start to teach producers not
>> >> to pass unknown Probabilities in the first place?
>> >
>> >
>> > If you search addSuccessor in LLVM, you will find about 200 users for
>> > different platforms without probability passed explicitly. Some are
>> > conditionally used such as:
>> >
>> > MBB->addSuccessor(Succ1);
>> > if (...)
>> >
>> >   MBB->addSuccessor(Succ2);
>> >
>> > It seems that a default probability is more convenient but I don't
>> > object to add probabilities explicitly. There are just some tradeoffs.
>>
>> yes -- that is a longer term cleanup. After all the cases are cleaned
>> up, we can add more restrictions.
>
>
> Agree. So can we check in this patch first to recover related build breakage
> from outside?

Right except that I have not reviewed the patch in detail yet.

David
>
>>
>>
>> David
>> >
>> >
>> > http://reviews.llvm.org/D15548
>> >
>> >
>> >
>
>


More information about the llvm-commits mailing list