[llvm-dev] DCE in the presence of control flow.

Sat Jan 30 08:02:49 PST 2016

(BTW hal, there is a nice web version of the code from keith's presentation
here:
https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=3&ved=0ahUKEwibwpvI8dHKAhVC0WMKHbdkDGYQFggtMAI&url=https%3A%2F%2Fwww.cs.princeton.edu%2F~ras%2Fdead.ps&usg=AFQjCNGxo6X07t5-9gFtzCbWpjz25afMBg&sig2=E6qUdZS8-x-88uJaIRZ8-A

The rest of code to the massively scalar compiler from rice, which had all
these nice SSA passes written in noweb, doesn't appear to be available
generally anymore, sadly. It was a nice reference.)

"I had assumed you would treat phi nodes differently from other operations
in that they don’t need to keep the block alive just to retain the data
flow facts but it would be simplest to do that."

Actually, they do.
At least in dead code elimination. This is true even if the incoming edge
is an empty block.

Phi nodes have data and control dependencies.  If the output of a phi is
necessary, both the data and control ependencies of the phi are necessary.

Now, in practice, yes, a phi node argument is only necessary *if* removing
it will not cause the phi node output to produce a different value in the
necessary uses.

But this is not DCE, it's value numbering :-)

Indeed, all the good value numbering/etc algorithms will determine if they
can replace/remove phi node arguments and detect unreachable blocks and
edges.
But that is a significantly more complicated ballgame.

In DCE, phi nodes will end up either entirely dead, or entirely alive.

On Sat, Jan 30, 2016 at 7:02 AM, David Callahan <dcallahan at fb.com> wrote:

> I had assumed you would treat phi nodes differently from other operations
> in that they don’t need to keep the block alive just to retain the data
> flow facts but it would be simplest to do that.
> Thanks Daniel
>
> From: Daniel Berlin <dberlin at dberlin.org>
> Date: Friday, January 29, 2016 at 10:26 PM
> To: David Callahan <dcallahan at fb.com>, Hal Finkel <hfinkel at anl.gov>
> Cc: LLVM Dev Mailing list <llvm-dev at lists.llvm.org>
> Subject: Re: [llvm-dev] DCE in the presence of control flow.
>
> In practice, APT is not faster to build than rdf.
> The df calculator we use is linear time and quite fast.
>
> Updating is also pretty trivial since it's only deletes of dead and
> unreachable code.  So anything it reached can be replaced with undef in
> most cases.
>
> Cd-dce is not slower in GCC than dce
>
> On Fri, Jan 29, 2016, 8:31 PM David Callahan <dcallahan at fb.com> wrote:
>
>> I think you can also avoid the RDF computation using a more directed form
>> of control dependence testing such as described in
>>
>> Keshav Pingali and Gianfranco Bilardi. 1997. Optimal control dependence
>> computation and the Roman chariots problem. ACM Trans. Program. Lang. Syst.
>> 19, 3 (May 1997), 462-491. DOI=http://dx.doi.org/10.1145/256167.256217
>> <https://urldefense.proofpoint.com/v2/url?u=http-3A__dx.doi.org_10.1145_256167.256217&d=CwMFaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=lFyiPUrFdOHdaobP7i4hoA&m=85DfLWatM6urqzYLUvrWBgn5LKUxCbQ3YheI3kh5XdU&s=qJbIjqiOKcfzbhZ33-9K0SGsouUl_rvtRgEareBdqGY&e=>
>>
>>
>> However one challenge seems to be fixing the SSA graph after deleting
>> essentially arbitrary connected regions of the control flow graph. It may
>> be that the common case where deleted control flow has a single entry
>> and a common post-dominator which seems straight forward but in the general
>> case it seems much harder. Any prior experience on that problem?
>> david
>>
>> From: Hal Finkel <hfinkel at anl.gov>
>> Date: Friday, January 29, 2016 at 4:50 PM
>> To: Daniel Berlin <dberlin at dberlin.org>
>>
>> Cc: LLVM Dev Mailing list <llvm-dev at lists.llvm.org>, David Callahan <
>> dcallahan at fb.com>
>> Subject: Re: [llvm-dev] DCE in the presence of control flow.
>>
>>
>>
>> ------------------------------
>>
>> *From: *"Daniel Berlin" <dberlin at dberlin.org>
>> *To: *"Hal Finkel" <hfinkel at anl.gov>
>> *Cc: *"LLVM Dev Mailing list" <llvm-dev at lists.llvm.org>, "David
>> Callahan" <dcallahan at fb.com>
>> *Sent: *Friday, January 29, 2016 12:48:37 AM
>> *Subject: *Re: [llvm-dev] DCE in the presence of control flow.
>>
>>
>>
>> On Thu, Jan 28, 2016 at 10:09 PM, Hal Finkel <hfinkel at anl.gov> wrote:
>>
>>>
>>> ------------------------------
>>>
>>> *From: *"David Callahan via llvm-dev" <llvm-dev at lists.llvm.org>
>>> *To: *"Daniel Berlin" <dberlin at dberlin.org>, "LLVM Dev Mailing list" <
>>> llvm-dev at lists.llvm.org>
>>> *Sent: *Thursday, January 28, 2016 11:35:49 PM
>>> *Subject: *Re: [llvm-dev] DCE in the presence of control flow.
>>>
>>> Thanks
>>> Also I found that some cases are also caught by a specialized routine to
>>> remove dead loops which is missing the case I noticed.
>>> odavd
>>>
>>> From: Daniel Berlin <dberlin at dberlin.org>
>>> Date: Thursday, January 28, 2016 at 8:45 PM
>>> To: David Callahan <dcallahan at fb.com>, LLVM Dev Mailing list <
>>> llvm-dev at lists.llvm.org>
>>> Subject: Re: [llvm-dev] DCE in the presence of control flow.
>>>
>>> The post dominators computation costs more on llvm than GCC because of
>>> how the respective cfgs work under the covers.  Even for GCC, when we
>>> implemented cd-dce, it only catches 1-2% more cases.  It's not really worth
>>> the cost in llvm unless postdom comes free
>>>
>>>
>>> A 1-2% reduction in code size seems like it might well be worth a
>>> post-dom calculation.
>>>
>>
>> 1-2% more cases != 1-2% reduction in code size. In particular, it assumes
>> nothing else will catch those cases :)
>>
>> Fair enough.
>>
>>
>> The cases are mostly caught by SimplifyCFG/etc anyway
>>
>> In any case, here are the numbers from when it was turned on by default:
>>
>> https://gcc.gnu.org/ml/gcc-patches/2004-01/msg00675.html
>>
>> Note: GCC is at least 3x faster at computing post-dom than LLVM
>>
>> Why?
>>
>>
>> Also, what exactly is the algorithm using post-dom info?
>>>
>>>
>> So, to review for a second, right now, the algorithm answers the question
>> when is a branch necessary with "always" :)
>>
>> The real answer is "when an already-necessary operation depends on its
>> existence".  This is of course, requires control-dependence to answer.
>> So if you take our current DCE algorithm, and instead of marking
>> terminators always-live, it simply marks control dependent edges of those
>> operands as necessary, and branches that generate those edges.
>>
>> (IE
>>
>> for each block in RDF(useful block):
>>   mark terminator of block as useful
>>
>>
>> FWIW, Keith Cooper's slide deck on this has a nice explanation:
>> https://www.cs.rice.edu/~keith/512/2011/Lectures/L04Dead-1up.pdf
>> <https://urldefense.proofpoint.com/v2/url?u=https-3A__www.cs.rice.edu_-7Ekeith_512_2011_Lectures_L04Dead-2D1up.pdf&d=CwMFaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=lFyiPUrFdOHdaobP7i4hoA&m=85DfLWatM6urqzYLUvrWBgn5LKUxCbQ3YheI3kh5XdU&s=C3xP_URm9hGCDdAsg-pKQsJ6sWhEVYgyYAGKbhC6ENU&e=>
>>
>> We might be able to do this without precomputing an RDF, however. For
>> example, you could solve a data-flow problem on the reverse CFG, where for
>> each block you solve for the "next" live instruction. Then a branch is
>> alive only if the next live instruction for each successor is different.
>> You'd need to have "next-live-instruction phi nodes" for cases where
>> there's not a unique answer, etc.
>>
>> Thanks again,
>> Hal
>>
>>
>> I suppose that we're looking for cases where we have a CFG diamond with
>>> one having only dead instructions, and loops with all dead instructions,
>>> etc.
>>>
>>
>> Yes.  Loops with all dead instructions includes "loops with no
>> side-effects outside of the loop"
>>
>>
>>>
>>>
>> Thanks again,
>>> Hal
>>>
>>>
>>> On Wed, Jan 27, 2016, 1:56 PM David Callahan via llvm-dev <
>>> llvm-dev at lists.llvm.org> wrote:
>>>
>>>> I have been looking at some internal codes looking for differences
>>>> between Clang (specifically 3.7.1) and gcc (typically 4.8.1 but sometimes
>>>> later).
>>>>
>>>>
>>>>
>>>> One area where I bumped into was dead code elimination in the presence
>>>> of complex control flow. I note that the “aggressive dead code elimination”
>>>> (ADCE.cpp) treats all branch operations as live (isa<TerminatorInst>(I)).
>>>> Doing more requires some approximation to control dependence.
>>>>
>>>>
>>>>
>>>> I note SimplifyCFG indirectly handles some simple cases. It will
>>>> speculate the contents of a basic block into a predecessor but this is
>>>> insufficient for more complex structures. This probably cherry-picks the
>>>> most common cases by frequency.
>>>>
>>>>
>>>>
>>>> Have their been prior attempts strengthen dead code elimination w.r.t.
>>>> control flow? If so, any guidance on what went wrong?
>>>>
>>>>
>>>>
>>>> Thanks
>>>>
>>>> David
>>>>
>>>>
>>>> _______________________________________________
>>>> LLVM Developers mailing list
>>>> llvm-dev at lists.llvm.org
>>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>> <https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.llvm.org_cgi-2Dbin_mailman_listinfo_llvm-2Ddev&d=CwMFaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=lFyiPUrFdOHdaobP7i4hoA&m=pCcZikgFQttaHaETuHc6G00dgArj_Spf58imKkXlTqk&s=NTh5Q1gE2ANS1rQYN9XFok_t8wvWCu1dzzzvHfv3hlI&e=>
>>>>
>>>
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> llvm-dev at lists.llvm.org
>>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>> <https://urldefense.proofpoint.com/v2/url?u=http-3A__lists.llvm.org_cgi-2Dbin_mailman_listinfo_llvm-2Ddev&d=CwMFaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=lFyiPUrFdOHdaobP7i4hoA&m=7nHK35pXeFc5plcmZyxIPJwB8O8ZyHCJ4P89mN8BoZQ&s=BIA3X52dktMwTkOdClno4EdVyZHbON1Yd_OB0ZehhB0&e=>
>>>
>>>
>>>
>>>
>>> --
>>> Hal Finkel
>>> Assistant Computational Scientist
>>> Leadership Computing Facility
>>> Argonne National Laboratory
>>>
>>
>>
>>
>>
>> --
>> Hal Finkel
>> Assistant Computational Scientist
>> Leadership Computing Facility
>> Argonne National Laboratory
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20160130/8b9039ce/attachment.html>