Fix handling of flag_rename_registers.

Message ID 12fd0c18-79e0-d20d-2973-92639071c050@suse.cz
State New
Headers show
Series
  • Fix handling of flag_rename_registers.
Related show

Commit Message

Martin Liška Oct. 12, 2021, 12:17 p.m.
Hello.

The option is disabled in rs6000 target with:

     { OPT_LEVELS_ALL, OPT_frename_registers, NULL, 0 },

Thus, we have to do an auto-detection only if it's really unset and also
equal to the Init value.

Patch can bootstrap on x86_64-linux-gnu and survives regression tests.
And the problematic test-case works on ppc64le.

Ready to be installed?
Thanks,
Martin

	PR target/102688

gcc/ChangeLog:

	* common.opt: Enable flag_rename_registers by default.
	* toplev.c (process_options): Auto-detect flag_rename_registers
	only if it is not turned off in a target.
---
  gcc/common.opt | 2 +-
  gcc/toplev.c   | 3 ++-
  2 files changed, 3 insertions(+), 2 deletions(-)

-- 
2.33.0

Comments

Jeff Law via Gcc-patches Oct. 12, 2021, 1:37 p.m. | #1
On Tue, Oct 12, 2021 at 2:18 PM Martin Liška <mliska@suse.cz> wrote:
>

> Hello.

>

> The option is disabled in rs6000 target with:

>

>      { OPT_LEVELS_ALL, OPT_frename_registers, NULL, 0 },

>

> Thus, we have to do an auto-detection only if it's really unset and also

> equal to the Init value.

>

> Patch can bootstrap on x86_64-linux-gnu and survives regression tests.

> And the problematic test-case works on ppc64le.

>

> Ready to be installed?


Hmm, I can see how it fixes the reported problem but I think the
thing is fragile.  I wonder if we can express things like

+  if (!OPTION_SET_P (flag_web))
     flag_web = flag_unroll_loops;

or

+  if (!OPTION_SET_P (flag_rename_registers))
     flag_rename_registers = flag_unroll_loops;

by adding EnabledBy(funroll-loops) to the respective options instead
(and funroll-loops EnabledBy(funroll-all-loops))

All SET_OPTION_IF_UNSET are fragile with respect to target overrides
(-fprofile-use does a lot of those for example).

I suppose opts_set could also record whether the target overrided
sth with its option_optimization_table.

Richard.

> Thanks,

> Martin

>

>         PR target/102688

>

> gcc/ChangeLog:

>

>         * common.opt: Enable flag_rename_registers by default.

>         * toplev.c (process_options): Auto-detect flag_rename_registers

>         only if it is not turned off in a target.

> ---

>   gcc/common.opt | 2 +-

>   gcc/toplev.c   | 3 ++-

>   2 files changed, 3 insertions(+), 2 deletions(-)

>

> diff --git a/gcc/common.opt b/gcc/common.opt

> index 4099effcc80..2c6be1bdd36 100644

> --- a/gcc/common.opt

> +++ b/gcc/common.opt

> @@ -2399,7 +2399,7 @@ Common Var(flag_live_range_shrinkage) Init(0) Optimization

>   Relief of register pressure through live range shrinkage.

>

>   frename-registers

> -Common Var(flag_rename_registers) Optimization

> +Common Var(flag_rename_registers) Init(1) Optimization

>   Perform a register renaming optimization pass.

>

>   fschedule-fusion

> diff --git a/gcc/toplev.c b/gcc/toplev.c

> index 167feac2583..ee7d8854f90 100644

> --- a/gcc/toplev.c

> +++ b/gcc/toplev.c

> @@ -1335,7 +1335,8 @@ process_options (bool no_backend)

>     if (!OPTION_SET_P (flag_web))

>       flag_web = flag_unroll_loops;

>

> -  if (!OPTION_SET_P (flag_rename_registers))

> +  /* The option can be turned off in a target.  */

> +  if (!OPTION_SET_P (flag_rename_registers) && flag_rename_registers)

>       flag_rename_registers = flag_unroll_loops;

>

>     if (flag_non_call_exceptions)

> --

> 2.33.0

>
Martin Liška Oct. 12, 2021, 2:03 p.m. | #2
On 10/12/21 15:37, Richard Biener wrote:
> On Tue, Oct 12, 2021 at 2:18 PM Martin Liška <mliska@suse.cz> wrote:

>>

>> Hello.

>>

>> The option is disabled in rs6000 target with:

>>

>>       { OPT_LEVELS_ALL, OPT_frename_registers, NULL, 0 },

>>

>> Thus, we have to do an auto-detection only if it's really unset and also

>> equal to the Init value.

>>

>> Patch can bootstrap on x86_64-linux-gnu and survives regression tests.

>> And the problematic test-case works on ppc64le.

>>

>> Ready to be installed?

> 

> Hmm, I can see how it fixes the reported problem but I think the

> thing is fragile.


You are fully right, it's quite fragile.

> I wonder if we can express things like

> 

> +  if (!OPTION_SET_P (flag_web))

>       flag_web = flag_unroll_loops;

> 

> or

> 

> +  if (!OPTION_SET_P (flag_rename_registers))

>       flag_rename_registers = flag_unroll_loops;

> 

> by adding EnabledBy(funroll-loops) to the respective options instead

> (and funroll-loops EnabledBy(funroll-all-loops))


Testing that approach, I like it.

Note that my fix:

if (!OPTION_SET_P (flag_rename_registers) && flag_rename_registers)

won't work if one target sets flag_rename_registers = 1 and another to flag_rename_registers = 0.
Then one can't use Init setting a proper default value.

> 

> All SET_OPTION_IF_UNSET are fragile with respect to target overrides

> (-fprofile-use does a lot of those for example).

> 

> I suppose opts_set could also record whether the target overrided

> sth with its option_optimization_table.


I can experiment with a patch where SET_OPTION_IF_UNSET modified opts_set.

Thanks for clever feedback.
Martin

> 

> Richard.

> 

>> Thanks,

>> Martin

>>

>>          PR target/102688

>>

>> gcc/ChangeLog:

>>

>>          * common.opt: Enable flag_rename_registers by default.

>>          * toplev.c (process_options): Auto-detect flag_rename_registers

>>          only if it is not turned off in a target.

>> ---

>>    gcc/common.opt | 2 +-

>>    gcc/toplev.c   | 3 ++-

>>    2 files changed, 3 insertions(+), 2 deletions(-)

>>

>> diff --git a/gcc/common.opt b/gcc/common.opt

>> index 4099effcc80..2c6be1bdd36 100644

>> --- a/gcc/common.opt

>> +++ b/gcc/common.opt

>> @@ -2399,7 +2399,7 @@ Common Var(flag_live_range_shrinkage) Init(0) Optimization

>>    Relief of register pressure through live range shrinkage.

>>

>>    frename-registers

>> -Common Var(flag_rename_registers) Optimization

>> +Common Var(flag_rename_registers) Init(1) Optimization

>>    Perform a register renaming optimization pass.

>>

>>    fschedule-fusion

>> diff --git a/gcc/toplev.c b/gcc/toplev.c

>> index 167feac2583..ee7d8854f90 100644

>> --- a/gcc/toplev.c

>> +++ b/gcc/toplev.c

>> @@ -1335,7 +1335,8 @@ process_options (bool no_backend)

>>      if (!OPTION_SET_P (flag_web))

>>        flag_web = flag_unroll_loops;

>>

>> -  if (!OPTION_SET_P (flag_rename_registers))

>> +  /* The option can be turned off in a target.  */

>> +  if (!OPTION_SET_P (flag_rename_registers) && flag_rename_registers)

>>        flag_rename_registers = flag_unroll_loops;

>>

>>      if (flag_non_call_exceptions)

>> --

>> 2.33.0

>>
Martin Liška Oct. 12, 2021, 3:11 p.m. | #3
On 10/12/21 15:37, Richard Biener wrote:
> by adding EnabledBy(funroll-loops) to the respective options instead

> (and funroll-loops EnabledBy(funroll-all-loops))


All right, so the suggested approach works correctly.

Patch can bootstrap on x86_64-linux-gnu and survives regression tests.

Ready to be installed?
Thanks,
Martin
From c42efec30d7cce36c92d9369791826c9120dd3d1 Mon Sep 17 00:00:00 2001
From: Martin Liska <mliska@suse.cz>
Date: Tue, 12 Oct 2021 16:05:49 +0200
Subject: [PATCH] Fix handling of flag_rename_registers by a target.

gcc/ChangeLog:

	* common.opt: Use EnabledBy instead of detection in
	finish_options and process_options.
	* opts.c (finish_options): Remove handling of
	x_flag_unroll_all_loops.
	* toplev.c (process_options): Likewise for flag_web and
	flag_rename_registers.

diff --git a/gcc/common.opt b/gcc/common.opt
index 4099effcc80..bcbf95aada3 100644
--- a/gcc/common.opt
+++ b/gcc/common.opt
@@ -2399,7 +2399,7 @@ Common Var(flag_live_range_shrinkage) Init(0) Optimization
 Relief of register pressure through live range shrinkage.
 
 frename-registers
-Common Var(flag_rename_registers) Optimization
+Common Var(flag_rename_registers) Optimization EnabledBy(funroll-loops)
 Perform a register renaming optimization pass.
 
 fschedule-fusion
@@ -2939,7 +2939,7 @@ Common Var(flag_unroll_loops) Optimization
 Perform loop unrolling when iteration count is known.
 
 funroll-all-loops
-Common Var(flag_unroll_all_loops) Optimization
+Common Var(flag_unroll_all_loops) Optimization EnabledBy(funroll-all-loops)
 Perform loop unrolling for all loops.
 
 funroll-completely-grow-size
@@ -3158,7 +3158,7 @@ Common Var(flag_value_profile_transformations) Optimization
 Use expression value profiles in optimizations.
 
 fweb
-Common Var(flag_web) Optimization
+Common Var(flag_web) Optimization EnabledBy(funroll-loops)
 Construct webs and split unrelated uses of single variable.
 
 ftree-builtin-call-dce
diff --git a/gcc/opts.c b/gcc/opts.c
index 2116c2991dd..fc71b6e4242 100644
--- a/gcc/opts.c
+++ b/gcc/opts.c
@@ -1321,11 +1321,6 @@ finish_options (struct gcc_options *opts, struct gcc_options *opts_set,
 				       opts->x_flag_live_patching,
 				       loc);
 
-  /* Unrolling all loops implies that standard loop unrolling must also
-     be done.  */
-  if (opts->x_flag_unroll_all_loops)
-    opts->x_flag_unroll_loops = 1;
-
   /* Allow cunroll to grow size accordingly.  */
   if (!opts_set->x_flag_cunroll_grow_size)
     opts->x_flag_cunroll_grow_size
diff --git a/gcc/toplev.c b/gcc/toplev.c
index 167feac2583..81546b19e91 100644
--- a/gcc/toplev.c
+++ b/gcc/toplev.c
@@ -1331,13 +1331,6 @@ process_options (bool no_backend)
       flag_abi_version = 2;
     }
 
-  /* web and rename-registers help when run after loop unrolling.  */
-  if (!OPTION_SET_P (flag_web))
-    flag_web = flag_unroll_loops;
-
-  if (!OPTION_SET_P (flag_rename_registers))
-    flag_rename_registers = flag_unroll_loops;
-
   if (flag_non_call_exceptions)
     flag_asynchronous_unwind_tables = 1;
   if (flag_asynchronous_unwind_tables)
Jeff Law via Gcc-patches Oct. 13, 2021, 8:39 a.m. | #4
On Tue, Oct 12, 2021 at 5:11 PM Martin Liška <mliska@suse.cz> wrote:
>

> On 10/12/21 15:37, Richard Biener wrote:

> > by adding EnabledBy(funroll-loops) to the respective options instead

> > (and funroll-loops EnabledBy(funroll-all-loops))

>

> All right, so the suggested approach works correctly.

>

> Patch can bootstrap on x86_64-linux-gnu and survives regression tests.

>

> Ready to be installed?


 funroll-all-loops
-Common Var(flag_unroll_all_loops) Optimization
+Common Var(flag_unroll_all_loops) Optimization EnabledBy(funroll-all-loops)

that should be on funroll-loops?

Can you verify that the two-step -funroll-all-loops -> -funroll-loops
-> -frename-registers
works and that it is not somehow dependent on ordering?  Otherwise we have to
use EnabledBy(funroll-loops,funroll-all-loops) on frename-registers.
I guess the
EnabledBy doesn't work if the target decides to set flag_unroll_loop in one of
its hooks rather than via its option table override?  (as said, this
is all a mess...)

But grep should be your friend telling whether any target overrides
any of the flags...

I do hope we can eventually reduce the number of pre-/post-/lang/target/common
processing phases for options :/  Meh.

Richard.

> Thanks,

> Martin
Martin Liška Oct. 13, 2021, 10:02 a.m. | #5
On 10/13/21 10:39, Richard Biener wrote:
> On Tue, Oct 12, 2021 at 5:11 PM Martin Liška <mliska@suse.cz> wrote:

>>

>> On 10/12/21 15:37, Richard Biener wrote:

>>> by adding EnabledBy(funroll-loops) to the respective options instead

>>> (and funroll-loops EnabledBy(funroll-all-loops))

>>

>> All right, so the suggested approach works correctly.

>>

>> Patch can bootstrap on x86_64-linux-gnu and survives regression tests.

>>

>> Ready to be installed?

> 

>   funroll-all-loops

> -Common Var(flag_unroll_all_loops) Optimization

> +Common Var(flag_unroll_all_loops) Optimization EnabledBy(funroll-all-loops)

> 

> that should be on funroll-loops?


Yeah, what a stupid error.

> 

> Can you verify that the two-step -funroll-all-loops -> -funroll-loops

> -> -frename-registers


Yes, verified that in debugger, it's not dependent on an ordering.

> works and that it is not somehow dependent on ordering?  Otherwise we have to

> use EnabledBy(funroll-loops,funroll-all-loops) on frename-registers.

> I guess the

> EnabledBy doesn't work if the target decides to set flag_unroll_loop in one of

> its hooks rather than via its option table override?  (as said, this

> is all a mess...)


It's a complete mess. The only override happens in
rs6000_override_options_after_change. I think it can also utilize EnabledBy, but
I would like to do it in a different patch.

> 

> But grep should be your friend telling whether any target overrides

> any of the flags...

> 

> I do hope we can eventually reduce the number of pre-/post-/lang/target/common

> processing phases for options :/  Meh.


Huh.

May I install this fixed patch once it's tested?

Martin

> 

> Richard.

> 

>> Thanks,

>> Martin
Jeff Law via Gcc-patches Oct. 13, 2021, 11:49 a.m. | #6
On Wed, Oct 13, 2021 at 12:02 PM Martin Liška <mliska@suse.cz> wrote:
>

> On 10/13/21 10:39, Richard Biener wrote:

> > On Tue, Oct 12, 2021 at 5:11 PM Martin Liška <mliska@suse.cz> wrote:

> >>

> >> On 10/12/21 15:37, Richard Biener wrote:

> >>> by adding EnabledBy(funroll-loops) to the respective options instead

> >>> (and funroll-loops EnabledBy(funroll-all-loops))

> >>

> >> All right, so the suggested approach works correctly.

> >>

> >> Patch can bootstrap on x86_64-linux-gnu and survives regression tests.

> >>

> >> Ready to be installed?

> >

> >   funroll-all-loops

> > -Common Var(flag_unroll_all_loops) Optimization

> > +Common Var(flag_unroll_all_loops) Optimization EnabledBy(funroll-all-loops)

> >

> > that should be on funroll-loops?

>

> Yeah, what a stupid error.

>

> >

> > Can you verify that the two-step -funroll-all-loops -> -funroll-loops

> > -> -frename-registers

>

> Yes, verified that in debugger, it's not dependent on an ordering.

>

> > works and that it is not somehow dependent on ordering?  Otherwise we have to

> > use EnabledBy(funroll-loops,funroll-all-loops) on frename-registers.

> > I guess the

> > EnabledBy doesn't work if the target decides to set flag_unroll_loop in one of

> > its hooks rather than via its option table override?  (as said, this

> > is all a mess...)

>

> It's a complete mess. The only override happens in

> rs6000_override_options_after_change. I think it can also utilize EnabledBy, but

> I would like to do it in a different patch.

>

> >

> > But grep should be your friend telling whether any target overrides

> > any of the flags...

> >

> > I do hope we can eventually reduce the number of pre-/post-/lang/target/common

> > processing phases for options :/  Meh.

>

> Huh.

>

> May I install this fixed patch once it's tested?


Yes please.

Thanks,
Richard.

> Martin

>

> >

> > Richard.

> >

> >> Thanks,

> >> Martin

>
Jeff Law via Gcc-patches Oct. 14, 2021, 2:27 p.m. | #7
On 10/13/2021 4:02 AM, Martin Liška wrote:
>

>> works and that it is not somehow dependent on ordering?  Otherwise we 

>> have to

>> use EnabledBy(funroll-loops,funroll-all-loops) on frename-registers.

>> I guess the

>> EnabledBy doesn't work if the target decides to set flag_unroll_loop 

>> in one of

>> its hooks rather than via its option table override?  (as said, this

>> is all a mess...)

>

> It's a complete mess. The only override happens in

> rs6000_override_options_after_change. I think it can also utilize 

> EnabledBy, but

> I would like to do it in a different patch.

So what's the preferred way to handle this?  We're in the process of 
evaluating -frename-registers on our target right now and subject to 
verification of a couple issues, our inclination is to turn it on for 
our target at -O2.

Jeff
Martin Liška Oct. 15, 2021, 3:27 p.m. | #8
On 10/14/21 16:27, Jeff Law wrote:
> So what's the preferred way to handle this?  We're in the process of evaluating -frename-registers on our target right now and subject to verification of a couple issues, our inclination is to turn it on for our target at -O2.

> 

> Jeff


I think the best approach is doing that in TARGET_OPTION_OPTIMIZATION_TABLE like c6x does:

static const struct default_options c6x_option_optimization_table[] =
   {
     { OPT_LEVELS_1_PLUS, OPT_frename_registers, NULL, 1 },
...
}

Cheers,
Martin
Jeff Law Oct. 16, 2021, 7:51 p.m. | #9
On 10/15/2021 9:27 AM, Martin Liška wrote:
> On 10/14/21 16:27, Jeff Law wrote:

>> So what's the preferred way to handle this?  We're in the process of 

>> evaluating -frename-registers on our target right now and subject to 

>> verification of a couple issues, our inclination is to turn it on for 

>> our target at -O2.

>>

>> Jeff

>

> I think the best approach is doing that in 

> TARGET_OPTION_OPTIMIZATION_TABLE like c6x does:

>

> static const struct default_options c6x_option_optimization_table[] =

>   {

>     { OPT_LEVELS_1_PLUS, OPT_frename_registers, NULL, 1 },

> ...

> }

ACK.  So nothing particularly special :-)

jeff

Patch

diff --git a/gcc/common.opt b/gcc/common.opt
index 4099effcc80..2c6be1bdd36 100644
--- a/gcc/common.opt
+++ b/gcc/common.opt
@@ -2399,7 +2399,7 @@  Common Var(flag_live_range_shrinkage) Init(0) Optimization
  Relief of register pressure through live range shrinkage.
  
  frename-registers
-Common Var(flag_rename_registers) Optimization
+Common Var(flag_rename_registers) Init(1) Optimization
  Perform a register renaming optimization pass.
  
  fschedule-fusion
diff --git a/gcc/toplev.c b/gcc/toplev.c
index 167feac2583..ee7d8854f90 100644
--- a/gcc/toplev.c
+++ b/gcc/toplev.c
@@ -1335,7 +1335,8 @@  process_options (bool no_backend)
    if (!OPTION_SET_P (flag_web))
      flag_web = flag_unroll_loops;
  
-  if (!OPTION_SET_P (flag_rename_registers))
+  /* The option can be turned off in a target.  */
+  if (!OPTION_SET_P (flag_rename_registers) && flag_rename_registers)
      flag_rename_registers = flag_unroll_loops;
  
    if (flag_non_call_exceptions)