[LLVMdev] BlockFrequency spill weights

[Jakob Stoklund Olesen]

On Jun 17, 2013, at 10:48 AM, Benjamin Kramer <benny.kra at gmail.com> wrote:

> [Splitting this out from the original thread to reduce noise in it]
> 
> 
> On 17.06.2013, at 18:43, Jakob Stoklund Olesen <stoklund at 2pi.dk> wrote:
>> +LiveIntervals::getSpillWeight(bool isDef, bool isUse, BlockFrequency freq) {
>> +  return (isDef + isUse) * freq.getFrequency();
>> }
>> 
>> This computation can overflow.
> 
> Yep, I went down the easy route and converted it to floating point arithmetic. Is that OK here?

Yes, that should be fine.

+LiveIntervals::getSpillWeight(bool isDef, bool isUse, BlockFrequency freq) {
+  float EntryFreq = BlockFrequency::getEntryFrequency();
+  return (isDef + isUse) * (freq.getFrequency() / EntryFreq);

Nitpick: The float division can be constant folded and turned into a multiplication:

  const float Scale = 1.0f / getEntryFrequency();
  return (isDef + isUse) * (freq.getFrequency() * Scale);

I wouldn’t trust the compiler to do that without -fast-math.

> The attached patch also fixes the regression tests that failed. Mostly minor changes in register allocation. The SPARC one is a bit scary because an instruction now gets moved across inline assembly into a delay slot. Do you know if that is a safe transformation?

That sounds OK. As long as it is safe to move that instruction across the inline asm.

> <block-frequency-spilling.patch>

LGTM

Thanks,
/jakob