[llvm-commits] patch: partial DSE

[Peter Cooper]

Yeah, i had been considering calling Lang's function to get the list of instructions the memset would decompose into then use that to work out whether trimming the size of some of them would be ok, but thats done in the backend so it would be very nasty to call it from DSE.

On Nov 1, 2011, at 3:15 PM, Eric Christopher wrote:

>>> 
>> 
>> There aren't actually any vector types involved here; the issue is
>> that, for example, a 32 byte memset is cheaper than a 31-byte memset
>> under the default settings on x86-64.  I'm not sure what the right
>> approach is here.
> 
> Good point, I'd taken the 16 as a 16-byte vector issue and not the sizes.
> This is probably related to the patches that Lang did recently trying to figure
> out what the best number of "loads" is for a memset/memcpy as well.
> 
> Maybe a TLI.shouldCallMemcpy(Ty) type of call?
> 
> Not sure.
> 
> -eric