[PATCH] Correctly fill up cgraph_node::local.versionable flag.

[Martin Liška]

On 10/17/2015 09:49 PM, Jan Hubicka wrote:
>> Hello.
>>
>> I've been working on HSA branch, where we have a cloning pass running with all
>> optimization levels. The patch makes computation of cgraph_node::local.versionability
>> independent on IPA CP and uses the flag to verify that a function can be cloned.
>>
>> The patch can bootstrap on x86_64-linux-pc and survives test suite.
>>
>> Ready for trunk?
>> Thanks,
>> Martin
>
>> >From d17b51257d5e01ab6bd9a018b08f8ed6fd39c029 Mon Sep 17 00:00:00 2001
>> From: marxin <mliska@suse.cz>
>> Date: Thu, 8 Oct 2015 17:57:30 +0200
>> Subject: [PATCH 1/3] Correctly fill up cgraph_node::local.versionable flag.
>>
>> gcc/ChangeLog:
>>
>> 2015-10-15  Martin Liska  <mliska@suse.cz>
>>
>> 	* cgraphclones.c (cgraph_node::create_virtual_clone):
>> 	Verify cgraph_node.local.versionable instead of calling
>> 	tree_versionable_function_p.
>> 	* ipa-cp.c (determine_versionability): Save the information
>> 	to ipa_node_params summary.
>> 	(ipcp_versionable_function_p): Use it.
>> 	(ipcp_propagate_stage): Pass IPA_NODE_REF to a called function.
>> 	(ipcp_generate_summary): Do not compute cgraph_node
>> 	versionability.
>
> If you want to use the flag at WPA time, you need to stream it for LTO.
> I suppose this only passed testing because we have no testcases checking that
> ipa-cp happens with LTO.

Hi.

The flag is already streamed.

>
>> 	* ipa-inline-analysis.c (inline_generate_summary): Compute
>> 	visibility for all cgraph nodes.
>> 	* ipa-prop.c (ipa_node_params_t::duplicate): Duplicate
>> 	ipa_node_params::versionability.
>> 	* ipa-prop.h (struct ipa_node_params): Declare it.
>> ---
>>   gcc/cgraphclones.c        |  2 +-
>>   gcc/ipa-cp.c              | 15 ++++++---------
>>   gcc/ipa-inline-analysis.c |  4 ++++
>>   gcc/ipa-prop.c            |  1 +
>>   gcc/ipa-prop.h            |  2 ++
>>   5 files changed, 14 insertions(+), 10 deletions(-)
>>
>> diff --git a/gcc/cgraphclones.c b/gcc/cgraphclones.c
>> index e51431c..5c04dc4 100644
>> --- a/gcc/cgraphclones.c
>> +++ b/gcc/cgraphclones.c
>> @@ -570,7 +570,7 @@ cgraph_node::create_virtual_clone (vec<cgraph_edge *> redirect_callers,
>>     char *name;
>>
>>     if (!in_lto_p)
>> -    gcc_checking_assert (tree_versionable_function_p (old_decl));
>> +    gcc_checking_assert (local.versionable);
>
> Then you should be able to drop in_lto_p here.

Yes, it works.

I'm sending final version of the patch which I'm going to install to trunk.

Thanks for the review,
Martin

>
> OK with that change.
>
> Honza
>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-Correctly-fill-up-cgraph_node-local.versionable-flag.patch
Type: text/x-patch
Size: 4942 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20151019/e27d00cf/attachment.bin>