Re: [RFC] introduce --param max-lto-partition for having an upper bound on partition size

[Prathamesh Kulkarni <prathamesh dot kulkarni at linaro dot org>]

On 5 April 2016 at 16:58, Richard Biener <rguenther@suse.de> wrote:
> On Tue, 5 Apr 2016, Prathamesh Kulkarni wrote:
>
>> On 4 April 2016 at 19:44, Jan Hubicka <hubicka@ucw.cz> wrote:
>> >
>> >> diff --git a/gcc/lto/lto-partition.c b/gcc/lto/lto-partition.c
>> >> index 9eb63c2..bc0c612 100644
>> >> --- a/gcc/lto/lto-partition.c
>> >> +++ b/gcc/lto/lto-partition.c
>> >> @@ -511,9 +511,20 @@ lto_balanced_map (int n_lto_partitions)
>> >>    varpool_order.qsort (varpool_node_cmp);
>> >>
>> >>    /* Compute partition size and create the first partition.  */
>> >> +  if (PARAM_VALUE (MIN_PARTITION_SIZE) > PARAM_VALUE (MAX_PARTITION_SIZE))
>> >> +    fatal_error (input_location, "min partition size cannot be greater than max partition size");
>> >> +
>> >>    partition_size = total_size / n_lto_partitions;
>> >>    if (partition_size < PARAM_VALUE (MIN_PARTITION_SIZE))
>> >>      partition_size = PARAM_VALUE (MIN_PARTITION_SIZE);
>> >> +  else if (partition_size > PARAM_VALUE (MAX_PARTITION_SIZE))
>> >> +    {
>> >> +      n_lto_partitions = total_size / PARAM_VALUE (MAX_PARTITION_SIZE);
>> >> +      if (total_size % PARAM_VALUE (MAX_PARTITION_SIZE))
>> >> +     n_lto_partitions++;
>> >> +      partition_size = total_size / n_lto_partitions;
>> >> +    }
>> >
>> > lto_balanced_map actually works in a way that looks for cheapest cutpoint in range
>> > 3/4*parittion_size to 2*partition_size and picks the cheapest range.
>> > Setting partition_size to this value will thus not cause partitioner to produce smaller
>> > partitions only.  I suppose modify the conditional:
>> >
>> >       /* Partition is too large, unwind into step when best cost was reached and
>> >          start new partition.  */
>> >       if (partition->insns > 2 * partition_size)
>> >
>> > and/or in the code above set the partition_size to half of total_size/max_size.
>> >
>> > I know this is somewhat sloppy.  This was really just first cut implementation
>> > many years ago. I expected to reimplement it marter soon, but then there was
>> > never really a need for it (I am trying to avoid late IPA optimizations so the
>> > partitioning decisions should mostly affect compile time performance only).
>> > If ARM is more sensitive for partitining, perhaps it would make sense to try to
>> > look for something smarter.
>> >
>> >> +
>> >>    npartitions = 1;
>> >>    partition = new_partition ("");
>> >>    if (symtab->dump_file)
>> >> diff --git a/gcc/lto/lto.c b/gcc/lto/lto.c
>> >> index 9dd513f..294b8a4 100644
>> >> --- a/gcc/lto/lto.c
>> >> +++ b/gcc/lto/lto.c
>> >> @@ -3112,6 +3112,12 @@ do_whole_program_analysis (void)
>> >>    timevar_pop (TV_WHOPR_WPA);
>> >>
>> >>    timevar_push (TV_WHOPR_PARTITIONING);
>> >> +
>> >> +  if (flag_lto_partition != LTO_PARTITION_BALANCED
>> >> +      && PARAM_VALUE (MAX_PARTITION_SIZE) != INT_MAX)
>> >> +    fatal_error (input_location, "--param max-lto-partition should only"
>> >> +              " be used with balanced partitioning\n");
>> >> +
>> >
>> > I think we should wire in resonable MAX_PARTITION_SIZE default.  THe value you
>> > found experimentally may be a good start. For that reason we can't really
>> > refuse a value when !LTO_PARTITION_BALANCED.  Just document it as parameter for
>> > balanced partitioning only and add a parameter to lto_balanced_map specifying whether
>> > this param should be honored (because the same path is used for partitioning to one partition)
>> >
>> > Otherwise the patch looks good to me modulo missing documentation.
>> Thanks for the review. I have updated the patch.
>> Does this version look OK ?
>> I had randomly chosen 10000, not sure if that's an appropriate value
>> for default.
>
> I think it's way too small.  This is roughly the number of GIMPLE stmts
> (thus roughly the number of instructions).  So with say a 8 byte
> instruction format it is on the order of 80kB.  You'd want to have a
> default of at least several ten times of large-unit-insns (also 10000).
> I'd choose sth like 1000000 (one million).  I find the lto-min-partition
> number quite small as well (and up it by a factor of 10).
Done in this version.
Is it OK after bootstrap+test ?

Thanks,
Prathamesh
>
> Richard.
>
>> I have a silly question about partitioning: Does it hamper
>> transformations on ipa optimizations if caller and
>> callee get placed in separate partitions ? For instance if callee is
>> supposed to be inlined
>> into caller, would inlining still take place if callee and caller get
>> placed in separate partitions ?
>> I tried with a trivial example with -flto-partition=max
>> which created 3 partitions for 3 functions (bar, foo and main), and it was
>> able to inline bar into foo and foo into main.  I am not sure how that happens.
>> I thought ltrans can perform transformations on functions only within
>> a single partition
>> and not across partitions ?
>>
>> Thanks,
>> Prathamesh
>> >
>> > Honza
>>
>
> --
> Richard Biener <rguenther@suse.de>
> SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Graham Norton, HRB 21284 (AG Nuernberg)

Attachment: patch-4.diff
Description: Text document

Attachment: ChangeLog
Description: Binary data