[llvm] r218492 - Add the first backend support for on demand subtarget creation

[Eric Christopher]

On Wed, Oct 1, 2014 at 7:16 AM, Daniel Sanders
<Daniel.Sanders at imgtec.com> wrote:
>> -----Original Message-----
>> From: Eric Christopher [mailto:echristo at gmail.com]
>> Sent: 30 September 2014 17:49
>> To: Daniel Sanders
>> Cc: llvm-commits at cs.uiuc.edu
>> Subject: Re: [llvm] r218492 - Add the first backend support for on demand
>> subtarget creation
>>
>> On Tue, Sep 30, 2014 at 2:32 AM, Daniel Sanders
>> <Daniel.Sanders at imgtec.com> wrote:
>> > Hi Eric,
>> >
>> >>  if (hasMips16Attr)
>> >>    FS += FS.empty() ? "+mips16" : ",+mips16";
>> >>  else if (hasNoMips16Attr)
>> >>    FS += FS.empty() ? "-mips16" : ",-mips16";
>> >
>> > It looks like targets are going to end up using this kind of code a lot. Would
>> > it make sense to switch the StringRef for a vector-like object at some
>> point?
>>
>> Possibly? For mips16 it's particularly annoying (it should never have
>> been a bare attribute), but only if I want to keep backwards
>> compatibility with existing IR. Everything else has a tendency to put
>> stuff in the feature string. I'll get the clang patch committed today
>> emitting the cpu and feature string for every function so you can see
>> what it looks like.
>>
>> -eric
>
> In that case, it sounds like it's fairly unique to us. I guess it makes more sense for us to migrate the attribute towards the more conventional implementation. If I understand the IR backwards compatibility promises correctly, we should be able to completely remove the attribute by LLVM 3.7.

It is, and that would be good, thanks.

-eric