llvm dev - Jun 2013 - [LLVMdev] MI-Sched temporarily enabled on x86.

I'm briefly enabling the MachineScheduler pass for x86 tonight to collect
information on any failures that may show up.

To properly enable the new scheduler, this hook is implemented in X86Subtarget:

  bool enableMachineScheduler() const LLVM_OVERRIDE { return true; }

The MachineScheduler pass itself can be enabled/disabled with
-enable-misched=true/false. But most of the codegen changes result rather from
moving to source-order SD scheduling and subsequent register coalescing
decisions. This makes it extremely challenging to keep the unit tests
functioning during the transition. So you'll see unit test churn when I flip
the switch.

Later, sometime in July, I'll plan to make the permanent change for X86,
with all unit tests switched over to source order scheduling. (I have some
compile-time improvements to make first). Changing the default for ARM will
follow shortly thereafter.

-Andy

On Mon, Jun 24, 2013 at 2:41 AM, Andrew Trick <atrick at apple.com> wrote:
> I'm briefly enabling the MachineScheduler pass for x86 tonight to
collect
> information on any failures that may show up.
>
> To properly enable the new scheduler, this hook is implemented in
> X86Subtarget:
>
>   bool enableMachineScheduler() const LLVM_OVERRIDE { return true; }
>
> The MachineScheduler pass itself can be enabled/disabled with
> -enable-misched=true/false. But most of the codegen changes result rather
> from moving to source-order SD scheduling and subsequent register
> coalescing decisions. This makes it extremely challenging to keep the unit
> tests functioning during the transition. So you'll see unit test churn
when
> I flip the switch.
>
This is really cool. Is there a way to provide a flag to enable the whole
thing so that folks can help you test stuff in advance of making it the
default? (Yes, I'm volunteering. =D)
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20130624/bd015f5d/attachment.html>

On Jun 24, 2013, at 2:47 AM, Chandler Carruth <chandlerc at google.com>
wrote:
> 
> On Mon, Jun 24, 2013 at 2:41 AM, Andrew Trick <atrick at apple.com>
wrote:
> I'm briefly enabling the MachineScheduler pass for x86 tonight to
collect information on any failures that may show up.
> 
> To properly enable the new scheduler, this hook is implemented in
X86Subtarget:
> 
>   bool enableMachineScheduler() const LLVM_OVERRIDE { return true; }
> 
> The MachineScheduler pass itself can be enabled/disabled with
-enable-misched=true/false. But most of the codegen changes result rather from
moving to source-order SD scheduling and subsequent register coalescing
decisions. This makes it extremely challenging to keep the unit tests
functioning during the transition. So you'll see unit test churn when I flip
the switch.
> 
> This is really cool. Is there a way to provide a flag to enable the whole
thing so that folks can help you test stuff in advance of making it the default?
(Yes, I'm volunteering. =D)
I'll probably turn it off soon ;)

If want the same behavior as the target hook you can use these flags:

-pre-RA-sched=source -join-globalcopies -enable-misched

Thanks for any testing! Bug reports would be great.

-Andy
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20130624/83ad28ec/attachment.html>