This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Fix computation of register limit for -fsched-pressure
On Fri, Oct 28, 2016 at 12:27 PM, Tamar Christina
<Tamar.Christina@arm.com> wrote:
> Looking at it again,
>
> it seems to be that the testcase should be adjusted.
> There's no actual spilling. It just uses more registers than before due to the scheduling.
Sorry I didn't look into the test, but using more registers sounds
like a regression too? At least we need to make sure it's reasonable.
Thanks,
bin
>
> I will update the testcase.
>
> Thanks.
>
> ________________________________________
> From: gcc-patches-owner@gcc.gnu.org <gcc-patches-owner@gcc.gnu.org> on behalf of Tamar Christina <Tamar.Christina@arm.com>
> Sent: Friday, October 28, 2016 10:53:20 AM
> To: Pat Haugen; Maxim Kuvyrkov
> Cc: GCC Patches; nd
> Subject: Re: [PATCH] Fix computation of register limit for -fsched-pressure
>
> Forwarding to list as well.
> ________________________________________
> From: Tamar Christina
> Sent: Friday, October 28, 2016 10:52:17 AM
> To: Pat Haugen; Maxim Kuvyrkov
> Cc: GCC Patches
> Subject: Re: [PATCH] Fix computation of register limit for -fsched-pressure
>
> Hi Pat,
>
> The commit seems to be causing some odd stack spills on aarch64.
>
> I've created a new ticket https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78142
>
> Thanks,
> Tamar
>
> ________________________________________
> From: gcc-patches-owner@gcc.gnu.org <gcc-patches-owner@gcc.gnu.org> on behalf of Pat Haugen <pthaugen@linux.vnet.ibm.com>
> Sent: Tuesday, October 18, 2016 4:07:55 PM
> To: Maxim Kuvyrkov
> Cc: GCC Patches
> Subject: Re: [PATCH] Fix computation of register limit for -fsched-pressure
>
> On 10/18/2016 05:31 AM, Maxim Kuvyrkov wrote:
>>> > I see your point and agree that current code isn't optimal. However, I don't think your patch is accurate either. Consider https://gcc.gnu.org/onlinedocs/gccint/Register-Basics.html and let's assume that FIXED_REGISTERS in class CL is set for a third of the registers, and CALL_USED_REGISTERS is set to "1" for another third of registers. So we have a third available for zero-cost allocation (CALL_USED_REGISTERS-FIXED_REGISTERS), a third available for spill-cost allocation (ALL_REGISTERS-CALL_USED_REGISTERS) and a third non-available (FIXED_REGISTERS).
>>> >
>>> > For a non-loop-single-basic-block function we should be targeting only the third of register available at zero-cost -- correct?
> Yes.
>
> This is what is done by the current code, but, apparently, by accident. It seems that the right register count can be obtained with:
>>> >
>>> > for (int i = 0; i < ira_class_hard_regs_num[cl]; ++i)
>>> > - if (call_used_regs[ira_class_hard_regs[cl][i]])
>>> > - ++call_used_regs_num[cl];
>>> > + if (!call_used_regs[ira_class_hard_regs[cl][i]]
>>> > + || fixed_regs[ira_class_hard_regs[cl][i]])
>>> > + ++call_saved_regs_num[cl];
>>> >
>>> > Does this look correct to you?
>> Thinking some more, it seems like fixed_regs should not be available to the scheduler no matter what. Therefore, the number of fixed registers should be subtracted from ira_class_hard_regs_num[cl] without any scaling (entry_freq / bb_freq).
>
> Ahh, yes, I forgot about FIXED_REGISTERS being included in CALL_USED_REGISTERS. I agree they should be totally removed from the register limit calculation. I'll rework the patch.
>
> Thanks,
> Pat
>