LoongArch: Set 4 * (issue rate) as the default for -falign-functions and -falign-loops

  According to Xuerui's LLVM changeset [1], doing so can make a
significant performace gain.

Bootstrapped and regtested on loongarch64-linux-gnu.  Ok for GCC 14?

[1]:https://reviews.llvm.org/D148622

gcc/ChangeLog:

	* config/loongarch/loongarch.cc
	(loongarch_option_override_internal): If -falign-functions is
	used but the alignment is not explicitly specified, set it to
	4 * loongarch_issue_rate ().  Likewise for -falign-loops.
---
 gcc/config/loongarch/loongarch.cc | 11 +++++++++++
 1 file changed, 11 insertions(+)
  

Hi,

Thanks for helping confirming on GCC and porting this! I'd never know 
even GCC lacked this adaptation without someone actually checking... Too 
many things are taken for granted these days.

On 2023/4/18 20:17, Xi Ruoyao wrote:
> According to Xuerui's LLVM changeset [1], doing so can make a
> significant performace gain.
>
> Bootstrapped and regtested on loongarch64-linux-gnu.  Ok for GCC 14?
>
> [1]:https://reviews.llvm.org/D148622
>
> gcc/ChangeLog:
>
> 	* config/loongarch/loongarch.cc
> 	(loongarch_option_override_internal): If -falign-functions is
> 	used but the alignment is not explicitly specified, set it to
> 	4 * loongarch_issue_rate ().  Likewise for -falign-loops.
> ---
>   gcc/config/loongarch/loongarch.cc | 11 +++++++++++
>   1 file changed, 11 insertions(+)
>
> diff --git a/gcc/config/loongarch/loongarch.cc b/gcc/config/loongarch/loongarch.cc
> index 06fc1cd0604..6552484de7c 100644
> --- a/gcc/config/loongarch/loongarch.cc
> +++ b/gcc/config/loongarch/loongarch.cc
> @@ -6236,6 +6236,17 @@ loongarch_option_override_internal (struct gcc_options *opts)
>         && !opts->x_optimize_size)
>       opts->x_flag_prefetch_loop_arrays = 1;
>   
> +  /* Align functions and loops to (issue rate) * (insn size) to improve
> +     the throughput of the fetching units.  */
What about gating all of these on !opts->x_optimize_size, similar to 
what aarch64 does?
> +  char *align = XNEWVEC (char, 16);
> +  sprintf (align, "%d", loongarch_issue_rate () * 4);
> +
> +  if (opts->x_flag_align_functions && !opts->x_str_align_functions)
> +    opts->x_str_align_functions = align;
> +
> +  if (opts->x_flag_align_loops && !opts->x_str_align_loops)
> +    opts->x_str_align_loops = align;
> +
>     if (TARGET_DIRECT_EXTERN_ACCESS && flag_shlib)
>       error ("%qs cannot be used for compiling a shared library",
>   	   "-mdirect-extern-access");
Otherwise LGTM, thanks!
  

On Tue, 2023-04-18 at 20:39 +0800, WANG Xuerui wrote:
> Hi,
> 
> Thanks for helping confirming on GCC and porting this! I'd never know 
> even GCC lacked this adaptation without someone actually checking... Too 
> many things are taken for granted these days.
> 
> On 2023/4/18 20:17, Xi Ruoyao wrote:
> > According to Xuerui's LLVM changeset [1], doing so can make a
> > significant performace gain.
> > 
> > Bootstrapped and regtested on loongarch64-linux-gnu.  Ok for GCC 14?
> > 
> > [1]:https://reviews.llvm.org/D148622
> > 
> > gcc/ChangeLog:
> > 
> >         * config/loongarch/loongarch.cc
> >         (loongarch_option_override_internal): If -falign-functions is
> >         used but the alignment is not explicitly specified, set it to
> >         4 * loongarch_issue_rate ().  Likewise for -falign-loops.
> > ---
> >   gcc/config/loongarch/loongarch.cc | 11 +++++++++++
> >   1 file changed, 11 insertions(+)
> > 
> > diff --git a/gcc/config/loongarch/loongarch.cc b/gcc/config/loongarch/loongarch.cc
> > index 06fc1cd0604..6552484de7c 100644
> > --- a/gcc/config/loongarch/loongarch.cc
> > +++ b/gcc/config/loongarch/loongarch.cc
> > @@ -6236,6 +6236,17 @@ loongarch_option_override_internal (struct gcc_options *opts)
> >         && !opts->x_optimize_size)
> >       opts->x_flag_prefetch_loop_arrays = 1;
> >   
> > +  /* Align functions and loops to (issue rate) * (insn size) to improve
> > +     the throughput of the fetching units.  */

> What about gating all of these on !opts->x_optimize_size, similar to 
> what aarch64 does?

opts->x_flag_align_functions and opts->x_flag_align_loops are only set
with -O2 or above unless the user manually uses -falign-functions or -
falign-loops.  If the user uses "-Os -falign-functions" as CFLAGS I'd
assume s(he) wants to optimize for size but keep the optimized function
alignment.

> > +  char *align = XNEWVEC (char, 16);
> > +  sprintf (align, "%d", loongarch_issue_rate () * 4);
> > +
> > +  if (opts->x_flag_align_functions && !opts->x_str_align_functions)
> > +    opts->x_str_align_functions = align;
> > +
> > +  if (opts->x_flag_align_loops && !opts->x_str_align_loops)
> > +    opts->x_str_align_loops = align;
> > +
> >     if (TARGET_DIRECT_EXTERN_ACCESS && flag_shlib)
> >       error ("%qs cannot be used for compiling a shared library",
> >            "-mdirect-extern-access");
> Otherwise LGTM, thanks!
  

On 2023/4/18 20:45, Xi Ruoyao wrote:
> On Tue, 2023-04-18 at 20:39 +0800, WANG Xuerui wrote:
>> Hi,
>>
>> Thanks for helping confirming on GCC and porting this! I'd never know
>> even GCC lacked this adaptation without someone actually checking... Too
>> many things are taken for granted these days.
>>
>> On 2023/4/18 20:17, Xi Ruoyao wrote:
>>> According to Xuerui's LLVM changeset [1], doing so can make a
>>> significant performace gain.

"doing so can gain significant performance" or "significant performance 
can be gained by ..."?

Also the other important thing is, guaranteeing alignment also makes 
performance *more predictable* in addition to generally making things 
faster. You may want to mention this too somewhere.

>>>
>>> Bootstrapped and regtested on loongarch64-linux-gnu.  Ok for GCC 14?
>>>
>>> [1]:https://reviews.llvm.org/D148622
nit: one space after the colon.
>>>
>>> gcc/ChangeLog:
>>>
>>>          * config/loongarch/loongarch.cc
>>>          (loongarch_option_override_internal): If -falign-functions is
>>>          used but the alignment is not explicitly specified, set it to
>>>          4 * loongarch_issue_rate ().  Likewise for -falign-loops.
>>> ---
>>>    gcc/config/loongarch/loongarch.cc | 11 +++++++++++
>>>    1 file changed, 11 insertions(+)
>>>
>>> diff --git a/gcc/config/loongarch/loongarch.cc b/gcc/config/loongarch/loongarch.cc
>>> index 06fc1cd0604..6552484de7c 100644
>>> --- a/gcc/config/loongarch/loongarch.cc
>>> +++ b/gcc/config/loongarch/loongarch.cc
>>> @@ -6236,6 +6236,17 @@ loongarch_option_override_internal (struct gcc_options *opts)
>>>          && !opts->x_optimize_size)
>>>        opts->x_flag_prefetch_loop_arrays = 1;
>>>    
>>> +  /* Align functions and loops to (issue rate) * (insn size) to improve
>>> +     the throughput of the fetching units.  */
>> What about gating all of these on !opts->x_optimize_size, similar to
>> what aarch64 does?
> opts->x_flag_align_functions and opts->x_flag_align_loops are only set
> with -O2 or above unless the user manually uses -falign-functions or -
> falign-loops.  If the user uses "-Os -falign-functions" as CFLAGS I'd
> assume s(he) wants to optimize for size but keep the optimized function
> alignment.

Ah, okay. Fine then.

BTW I've also added some comments to the commit message that I forgot to 
review earlier.

>>> +  char *align = XNEWVEC (char, 16);
>>> +  sprintf (align, "%d", loongarch_issue_rate () * 4);
>>> +
>>> +  if (opts->x_flag_align_functions && !opts->x_str_align_functions)
>>> +    opts->x_str_align_functions = align;
>>> +
>>> +  if (opts->x_flag_align_loops && !opts->x_str_align_loops)
>>> +    opts->x_str_align_loops = align;
>>> +
>>>      if (TARGET_DIRECT_EXTERN_ACCESS && flag_shlib)
>>>        error ("%qs cannot be used for compiling a shared library",
>>>             "-mdirect-extern-access");
>> Otherwise LGTM, thanks!
  

On Tue, 2023-04-18 at 20:51 +0800, WANG Xuerui wrote:
> 
> On 2023/4/18 20:45, Xi Ruoyao wrote:
> > On Tue, 2023-04-18 at 20:39 +0800, WANG Xuerui wrote:
> > > Hi,
> > > 
> > > Thanks for helping confirming on GCC and porting this! I'd never know
> > > even GCC lacked this adaptation without someone actually checking... Too
> > > many things are taken for granted these days.
> > > 
> > > On 2023/4/18 20:17, Xi Ruoyao wrote:
> > > > According to Xuerui's LLVM changeset [1], doing so can make a
> > > > significant performace gain.
> 
> "doing so can gain significant performance" or "significant performance 
> can be gained by ..."?
> 
> Also the other important thing is, guaranteeing alignment also makes 
> performance *more predictable* in addition to generally making things 
> faster. You may want to mention this too somewhere.

Will include this in the final commit or the next revision, thanks!

/* snip */
  

Hi, ruoyao:

Thank you so much for making this submission. But we are testing the 
impact of these two alignment parameters

(also including -falign-jumps and -falign-lables ) on performance. So 
before the result comes out, this patch will

not be merged into the main branch for the time being.

Thanks!

在 2023/4/18 下午8:17, Xi Ruoyao 写道:
> According to Xuerui's LLVM changeset [1], doing so can make a
> significant performace gain.
>
> Bootstrapped and regtested on loongarch64-linux-gnu.  Ok for GCC 14?
>
> [1]:https://reviews.llvm.org/D148622
>
> gcc/ChangeLog:
>
> 	* config/loongarch/loongarch.cc
> 	(loongarch_option_override_internal): If -falign-functions is
> 	used but the alignment is not explicitly specified, set it to
> 	4 * loongarch_issue_rate ().  Likewise for -falign-loops.
> ---
>   gcc/config/loongarch/loongarch.cc | 11 +++++++++++
>   1 file changed, 11 insertions(+)
>
> diff --git a/gcc/config/loongarch/loongarch.cc b/gcc/config/loongarch/loongarch.cc
> index 06fc1cd0604..6552484de7c 100644
> --- a/gcc/config/loongarch/loongarch.cc
> +++ b/gcc/config/loongarch/loongarch.cc
> @@ -6236,6 +6236,17 @@ loongarch_option_override_internal (struct gcc_options *opts)
>         && !opts->x_optimize_size)
>       opts->x_flag_prefetch_loop_arrays = 1;
>   
> +  /* Align functions and loops to (issue rate) * (insn size) to improve
> +     the throughput of the fetching units.  */
> +  char *align = XNEWVEC (char, 16);
> +  sprintf (align, "%d", loongarch_issue_rate () * 4);
> +
> +  if (opts->x_flag_align_functions && !opts->x_str_align_functions)
> +    opts->x_str_align_functions = align;
> +
> +  if (opts->x_flag_align_loops && !opts->x_str_align_loops)
> +    opts->x_str_align_loops = align;
> +
>     if (TARGET_DIRECT_EXTERN_ACCESS && flag_shlib)
>       error ("%qs cannot be used for compiling a shared library",
>   	   "-mdirect-extern-access");
  

On Tue, 2023-04-18 at 21:06 +0800, Lulu Cheng wrote:
> Hi, ruoyao:
> 
> Thank you so much for making this submission. But we are testing the 
> impact of these two alignment parameters
> 
> (also including -falign-jumps and -falign-lables ) on performance. So 
> before the result comes out, this patch will
> 
> not be merged into the main branch for the time being.

Ok, I'll wait for the result.

> 
> Thanks!
> 
> 在 2023/4/18 下午8:17, Xi Ruoyao 写道:
> > According to Xuerui's LLVM changeset [1], doing so can make a
> > significant performace gain.
> > 
> > Bootstrapped and regtested on loongarch64-linux-gnu.  Ok for GCC 14?
> > 
> > [1]:https://reviews.llvm.org/D148622
> > 
> > gcc/ChangeLog:
> > 
> >         * config/loongarch/loongarch.cc
> >         (loongarch_option_override_internal): If -falign-functions
> > is
> >         used but the alignment is not explicitly specified, set it
> > to
> >         4 * loongarch_issue_rate ().  Likewise for -falign-loops.
> > ---
> >   gcc/config/loongarch/loongarch.cc | 11 +++++++++++
> >   1 file changed, 11 insertions(+)
> > 
> > diff --git a/gcc/config/loongarch/loongarch.cc
> > b/gcc/config/loongarch/loongarch.cc
> > index 06fc1cd0604..6552484de7c 100644
> > --- a/gcc/config/loongarch/loongarch.cc
> > +++ b/gcc/config/loongarch/loongarch.cc
> > @@ -6236,6 +6236,17 @@ loongarch_option_override_internal (struct
> > gcc_options *opts)
> >         && !opts->x_optimize_size)
> >       opts->x_flag_prefetch_loop_arrays = 1;
> >   
> > +  /* Align functions and loops to (issue rate) * (insn size) to
> > improve
> > +     the throughput of the fetching units.  */
> > +  char *align = XNEWVEC (char, 16);
> > +  sprintf (align, "%d", loongarch_issue_rate () * 4);
> > +
> > +  if (opts->x_flag_align_functions && !opts->x_str_align_functions)
> > +    opts->x_str_align_functions = align;
> > +
> > +  if (opts->x_flag_align_loops && !opts->x_str_align_loops)
> > +    opts->x_str_align_loops = align;
> > +
> >     if (TARGET_DIRECT_EXTERN_ACCESS && flag_shlib)
> >       error ("%qs cannot be used for compiling a shared library",
> >            "-mdirect-extern-access");
>
  

On Tue, 2023-04-18 at 21:06 +0800, Lulu Cheng wrote:
> Hi, ruoyao:
> 
> Thank you so much for making this submission. But we are testing the 
> impact of these two alignment parameters
> 
> (also including -falign-jumps and -falign-lables ) on performance. So 
> before the result comes out, this patch will
> 
> not be merged into the main branch for the time being.

Hi!

Is there an estimate when the benchmark will be done?  If it will be
done soon I'll wait for the result before performing a full system
rebuild, otherwise I'll use my gut feeling to specify a -falign-
functions= value for the build :).
  

在 2023/5/29 下午2:09, Xi Ruoyao 写道:
> On Tue, 2023-04-18 at 21:06 +0800, Lulu Cheng wrote:
>> Hi, ruoyao:
>>
>> Thank you so much for making this submission. But we are testing the
>> impact of these two alignment parameters
>>
>> (also including -falign-jumps and -falign-lables ) on performance. So
>> before the result comes out, this patch will
>>
>> not be merged into the main branch for the time being.
> Hi!
>
> Is there an estimate when the benchmark will be done?  If it will be
> done soon I'll wait for the result before performing a full system
> rebuild, otherwise I'll use my gut feeling to specify a -falign-
> functions= value for the build :).
>
Sorry for taking so long to reply to the email. From our current test 
results,

the performance of the SPEC is best when combined with -falign-loops=16,

-falign-jumps=16, -falign-functions=32 and -falign-lables=16.
  

On Tue, 2023-05-30 at 09:30 +0800, Lulu Cheng wrote:
> 
> 在 2023/5/29 下午2:09, Xi Ruoyao 写道:
> > On Tue, 2023-04-18 at 21:06 +0800, Lulu Cheng wrote:
> > > Hi, ruoyao:
> > > 
> > > Thank you so much for making this submission. But we are testing
> > > the
> > > impact of these two alignment parameters
> > > 
> > > (also including -falign-jumps and -falign-lables ) on performance.
> > > So
> > > before the result comes out, this patch will
> > > 
> > > not be merged into the main branch for the time being.
> > Hi!
> > 
> > Is there an estimate when the benchmark will be done?  If it will be
> > done soon I'll wait for the result before performing a full system
> > rebuild, otherwise I'll use my gut feeling to specify a -falign-
> > functions= value for the build :).
> > 
> Sorry for taking so long to reply to the email. From our current test 
> results,
> 
> the performance of the SPEC is best when combined with -falign-
> loops=16,
> 
> -falign-jumps=16, -falign-functions=32 and -falign-lables=16.

I've completed a system rebuild with -falign-
{jumps,functions,labels}=16.  I've missed -falign-loops=16 but the doc
says -falign-labels=16 implies -falign-jumps=16 and -falign-loops=16 (if
-falign-jumps or -falign-loops are not set explicitly with a larger
value).

I'll make a patch to set -falign-functions=32 and -falign-labels=16 with
-mtune={la464,loongarch64} after setting a basic develop environment on
the new system...  And I'm wondering if things will change with LA664
:).
  

在 2023/6/12 下午5:19, Xi Ruoyao 写道:
> On Tue, 2023-05-30 at 09:30 +0800, Lulu Cheng wrote:
>> 在 2023/5/29 下午2:09, Xi Ruoyao 写道:
>>> On Tue, 2023-04-18 at 21:06 +0800, Lulu Cheng wrote:
>>>> Hi, ruoyao:
>>>>
>>>> Thank you so much for making this submission. But we are testing
>>>> the
>>>> impact of these two alignment parameters
>>>>
>>>> (also including -falign-jumps and -falign-lables ) on performance.
>>>> So
>>>> before the result comes out, this patch will
>>>>
>>>> not be merged into the main branch for the time being.
>>> Hi!
>>>
>>> Is there an estimate when the benchmark will be done?  If it will be
>>> done soon I'll wait for the result before performing a full system
>>> rebuild, otherwise I'll use my gut feeling to specify a -falign-
>>> functions= value for the build :).
>>>
>> Sorry for taking so long to reply to the email. From our current test
>> results,
>>
>> the performance of the SPEC is best when combined with -falign-
>> loops=16,
>>
>> -falign-jumps=16, -falign-functions=32 and -falign-lables=16.
> I've completed a system rebuild with -falign-
> {jumps,functions,labels}=16.  I've missed -falign-loops=16 but the doc
> says -falign-labels=16 implies -falign-jumps=16 and -falign-loops=16 (if
> -falign-jumps or -falign-loops are not set explicitly with a larger
> value).
>
> I'll make a patch to set -falign-functions=32 and -falign-labels=16 with
> -mtune={la464,loongarch64} after setting a basic develop environment on
> the new system...  And I'm wondering if things will change with LA664
> :).
>
>
We haven't tested it on LA664 yet, so it's not clear whether this is 
consistent with LA464.

This test will not be done on LA664 anytime soon. But once I've done the 
test, I'll let you know by email.