[PATCH][RegAlloc] Make tryInstructionSplit less aggressive.

[Quentin Colombet]

Committed as r198369.

Thanks again Jakob!

-Quentin
On Dec 23, 2013, at 6:25 PM, Quentin Colombet <qcolombet at apple.com> wrote:

> Thanks Jakob for the quick review!
> 
> I'll integrate your suggestions and I'll commit after the break so that I am around in case something goes wrong.
> 
> Thanks again.
> 
> -Quentin
> On Dec 23, 2013, at 3:00 PM, Jakob Stoklund Olesen <stoklund at 2pi.dk> wrote:
> 
>> 
>> On Dec 23, 2013, at 2:35 PM, Quentin Colombet <qcolombet at apple.com> wrote:
>> 
>>> Hi Jakob,
>>> 
>>> Here is the reworked patch.
>>> 
>>> Now the MachineInstr API offers new methods to query the effect of the operand constraints on the register class of a virtual register or on a given register class.
>>> The new API is used in MachineRegisterInfo::recomputeRegClass as well as in RAGreedy::tryInstructionSplit.
>> 
>> Thanks, Quentin.
>> 
>> This is looking good.
>> 
>> I think I would prefer to omit the MI::getLargest* functions. You will almost always be looking at multiple instructions, and I’m afraid it can get expensive to call getLargestLegalSuperClass() redundantly for all of them.
>> 
>> The getRegClassConstraintEffect* functions look great.
>> 
>> +  const TargetRegisterClass *SuperRC = TRI->getLargestLegalSuperClass(CurRC);
>> +  // Split around every non-copy instruction if this split will relax
>> +  // the constraints on the virtual register.
>> +  // Otherwise, splitting just inserts uncoalescable copies that do not help
>> +  // the allocation.
>>  for (unsigned i = 0; i != Uses.size(); ++i) {
>>    if (const MachineInstr *MI = Indexes->getInstructionFromIndex(Uses[i]))
>> -      if (MI->isFullCopy()) {
>> +      if (MI->isFullCopy() ||
>> +          SuperRC == MI->getLargestRegClassForVReg(VirtReg.reg, TII, TRI, MRI,
>> +                                                   /* ExploreBundle */ true)) {
>> 
>> Here, I would just pass SuperRC to MI->getRegClassConstraintEffectForVReg() instead.
>> 
>> Also, don’t compare the register classes by equality, but check that the returned register class actually has fewer allocatable registers. See for example GPR/rGPR on ARM. RegisterClassInfo::getNumAllocatableRegs() is very fast.
>> 
>> +  if (ExploreBundle && (isBundle() || isBundled())) {
>> +    MachineBasicBlock::const_instr_iterator I = this;
>> +    MachineBasicBlock::const_instr_iterator E = getParent()->instr_end();
>> +    while (CurRC && ++I != E && I->isInsideBundle())
>> +      // Accumulate the constraints through the instructions.
>> +      CurRC = I->getRegClassConstraintEffectForVRegImpl(Reg, CurRC, TII, TRI);
>> +  }
>> 
>> See MIBundleOperands.
>> 
>> Go ahead and commit with those changes.
>> 
>> Thanks,
>> /jakob
>> 
> 
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits