Matthias Braun via llvm-dev
2018-Jan-16 17:31 UTC
[llvm-dev] Checking when Register Allocation has been performed
Please don't rely on this for checking whether regalloc was run: You can have functions without vregs pre-RA[1]. We don't need or should track state such as pre/post-RA as part of the function. Instead it really is a property of where a pass was scheduled, so the pass should know and not the function. I'd recommend simply creating a pre-RA and a post-RA pass instead of scheduling the same pass twice. (Of course you can share most of the pass implementation if they turn out to be similar). - Matthias [1]: While currently we do not set the NoVRrgs flag pre-ra even if there are no vregs used, when loading a .mir file for example the flag is computed from scratch and will be set.> On Jan 15, 2018, at 12:26 PM, Craig Topper via llvm-dev <llvm-dev at lists.llvm.org> wrote: > > Maybe MF.getProperties().hasProperty(MachineFunctionProperties::Property::NoVRegs))? > > ~Craig > > On Mon, Jan 15, 2018 at 12:07 PM, Martin J. O'Riordan via llvm-dev <llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>> wrote: > Hi LLVM Devs, > > > > I have some shared code that performs lowering operations that can occur before or after register allocation. When it is pre-RA I want to only use virtual registers for intermediate results, but post-RA I have to use only a very restricted set of physical registers. > > > > Code generation using the restricted set is not as efficient as it is when I can use virtual registers. At the moment I have a clunky implementation for checking whether or not the register allocator has been run, and I am wondering if there is a “correct way” of checking whether the RA pass has been run? Some part of the register info API that I have missed perhaps? > > > > Thanks, > > > > MartinO > > > > > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org> > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev <http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev> > > > _______________________________________________ > LLVM Developers mailing list > llvm-dev at lists.llvm.org > http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180116/e3bc9d3e/attachment.html>
Martin J. O'Riordan via llvm-dev
2018-Jan-17 07:31 UTC
[llvm-dev] Checking when Register Allocation has been performed
Thanks Matthias,
I have both a pre-RA and a post-RA scheduler, and I had thought that I could
track “has RA happened?” by setting a flag in my pre-RA scheduler as it
completes - my suspicion (which you have confirmed) was that “#vregs == 0” was
not a safe assumption. What I cannot be sure of, is what passes execute after
my pre-RA scheduler but before RA, and what passes execute after RA but before
my post-RA scheduler since this is largely target-independent. If any of those
passes trigger a custom lowering action (frame setup for example), and it
requires additional scratch registers, then my code generation strategy needs to
change. As it happens, I know that frame lowering always occurs post-RA so that
is not a real example, but is it possible that another pass between pre-RA and
post-RA scheduling could trigger lowering actions? If not, then setting a flag
in my pre-RA scheduler will solve the problem perfectly.
All the best,
MartinO
From: mbraun at apple.com [mailto:mbraun at apple.com]
Sent: 16 January 2018 17:31
To: Craig Topper <craig.topper at gmail.com>
Cc: Martin J. O'Riordan <MartinO at theheart.ie>; LLVM Developers
<llvm-dev at lists.llvm.org>
Subject: Re: [llvm-dev] Checking when Register Allocation has been performed
Please don't rely on this for checking whether regalloc was run: You can
have functions without vregs pre-RA[1].
We don't need or should track state such as pre/post-RA as part of the
function. Instead it really is a property of where a pass was scheduled, so the
pass should know and not the function.
I'd recommend simply creating a pre-RA and a post-RA pass instead of
scheduling the same pass twice. (Of course you can share most of the pass
implementation if they turn out to be similar).
- Matthias
[1]: While currently we do not set the NoVRrgs flag pre-ra even if there are no
vregs used, when loading a .mir file for example the flag is computed from
scratch and will be set.
On Jan 15, 2018, at 12:26 PM, Craig Topper via llvm-dev <llvm-dev at
lists.llvm.org <mailto:llvm-dev at lists.llvm.org> > wrote:
Maybe
MF.getProperties().hasProperty(MachineFunctionProperties::Property::NoVRegs))?
~Craig
On Mon, Jan 15, 2018 at 12:07 PM, Martin J. O'Riordan via llvm-dev
<llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org> >
wrote:
Hi LLVM Devs,
I have some shared code that performs lowering operations that can occur before
or after register allocation. When it is pre-RA I want to only use virtual
registers for intermediate results, but post-RA I have to use only a very
restricted set of physical registers.
Code generation using the restricted set is not as efficient as it is when I can
use virtual registers. At the moment I have a clunky implementation for
checking whether or not the register allocator has been run, and I am wondering
if there is a “correct way” of checking whether the RA pass has been run? Some
part of the register info API that I have missed perhaps?
Thanks,
MartinO
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20180117/597b928e/attachment-0001.html>
Matthias Braun via llvm-dev
2018-Jan-17 18:04 UTC
[llvm-dev] Checking when Register Allocation has been performed
> On Jan 16, 2018, at 11:31 PM, Martin J. O'Riordan <MartinO at theheart.ie> wrote: > > Thanks Matthias, > > I have both a pre-RA and a post-RA scheduler, and I had thought that I could track “has RA happened?” by setting a flag in my pre-RA scheduler as it completes - my suspicion (which you have confirmed) was that “#vregs == 0” was not a safe assumption. What I cannot be sure of, is what passes execute after my pre-RA scheduler but before RA, and what passes execute after RA but before my post-RA scheduler since this is largely target-independent. If any of those passes trigger a custom lowering action (frame setup for example), and it requires additional scratch registers, then my code generation strategy needs to change. As it happens, I know that frame lowering always occurs post-RA so that is not a real example, but is it possible that another pass between pre-RA and post-RA scheduling could trigger lowering actions? If not, then setting a flag in my pre-RA scheduler will solve the problem perfectly.Do you have a specific callback in TargetInstrInfo or similar in mind? I would expect them to either be used pre-ra or post-ra but not both... - Matthias -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20180117/a40ea010/attachment.html>