Re: [PATCH 4/N] Recover GOTO predictor.

["Martin Liška" <mliska@suse.cz>]

Richi?

Thanks

On 06/30/2017 03:48 PM, Martin Liška wrote:
> On 06/22/2017 12:27 PM, Richard Biener wrote:
>> On Wed, Jun 21, 2017 at 3:06 PM, Martin Liška <mliska@suse.cz> wrote:
>>> Hello.
>>>
>>> There's one additional predictor enhancement that is GOTO predict that
>>> used to working. Following patch adds expect statement for C and C++ family
>>> languages.
>>>
>>> There's one fallout which is vrp24.c test-case, where only 'Simplified relational'
>>> appears just once. Adding Richi and Patrick who can probably help how to fix the
>>> test-case.
>>
>> Happens to be optimized better now, there's only one predicate to simplify
>> left in the IL input to VRP1.  I suggest to change it to match 1 times and add
>> -fdump-tree-optimized to dg-options and
>>
>> /* { dg-final { scan-tree-dump-times "if " 3 "optimized" } } */
>>
>> to verify we have 3 conditions left.
> 
> One small note, I see 4 conditions in optimized dump:
> 
> sss (struct rtx_def * insn, int code1, int code2, int code3)
> {
>   int D1544;
>   struct rtx_def * body;
>   _Bool D1562;
> 
>   <bb 2> [100.00%] [count: INV]:
>   body_5 = insn_4(D)->u.fld[5].rt_rtx;
>   D1544_6 = body_5->code;
>   if (D1544_6 == 55)
>     goto <bb 7> (L7); [34.00%] [count: INV]
>   else
>     goto <bb 3>; [66.00%] [count: INV]
> 
>   <bb 3> [66.00%] [count: INV]:
>   if (code3_7(D) == 99)
>     goto <bb 4>; [34.00%] [count: INV]
>   else
>     goto <bb 8> (L16); [66.00%] [count: INV]
> 
>   <bb 4> [22.44%] [count: INV]:
>   D1562_9 = code1_8(D) == 10;
>   if (D1562_9 == 1)
>     goto <bb 7> (L7); [64.00%] [count: INV]
>   else
>     goto <bb 8> (L16); [36.00%] [count: INV]
> 
>   <bb 5> [9.82%] [count: INV]:
>   arf ();
> 
>   <bb 6> [46.68%] [count: INV]:
>   frob (); [tail call]
>   goto <bb 8> (L16); [100.00%] [count: INV]
> 
> L7 [48.36%] [count: INV]:
>   if (code2_12(D) == 42)
>     goto <bb 5>; [20.24%] [count: INV]
>   else
>     goto <bb 6>; [79.76%] [count: INV]
> 
> L16 [100.00%] [count: INV]:
>   return;
> 
> }
> 
> Is it a problem or adjusting to 4 is fine?
> 
> Martin
> 
>>
>>> Patch can bootstrap on ppc64le-redhat-linux and survives regression tests.
>>>
>>> Ready to be installed?
>>> Martin
>