[llvm-commits] move CPU detection from lib/Target/$FOO to libSystem

[Daniel Dunbar]

On Sat, Nov 14, 2009 at 2:04 AM, Daniel Dunbar <daniel at zuster.org> wrote:
> Applied as r, with the extra StringMap include dropped. Thank you very

Err, r88768.

 - Daniel

> much, this has long been on my TODO list.
>
> Are you also working on the target features part of this work?
>
> On Sat, Nov 14, 2009 at 1:48 AM, Roman Divacky <rdivacky at freebsd.org> wrote:
>> On Sat, Nov 14, 2009 at 01:46:21AM -0800, Chandler Carruth wrote:
>>> On Sat, Nov 14, 2009 at 1:13 AM, Roman Divacky <rdivacky at freebsd.org> wrote:
>>> > On Fri, Nov 13, 2009 at 10:18:31PM -0800, Chandler Carruth wrote:
>>> >> On Fri, Nov 13, 2009 at 3:48 AM, Roman Divacky <rdivacky at freebsd.org> wrote:
>
>>> >> A meta-point: you're moving a lot more than just GetCurrentX86CPU into
>>> >> libSystem, can you expose the other functions and delete the copies in
>>> >> X86Subtarget? I'd rather avoid duplicating subtle code like that.
>>> >
>>> > I was told to not expose any target specific functionality, the
>>> > GetX86CpuIDAndInfo and DetectX86FamilyModel are fairly small and not
>>> > likely to change often. I dont like the duplication either butcant think of
>>> > any way to reach this X86 specific stuff from libSystem.
>>>
>>> I was thinking the other way around: expose this in libSystem so that
>>> the X86 target can leverage it. I agree libSystem shouldn't depend on
>>> libTarget.
>>
>> I know... I already did it like this but daniel preferred this way. I have
>> to agree that this looks better
>
>  http://llvm.org/bugs/show_bug.cgi?id=5389
> has the relevant context for those late to the party.
>
> Basically I want to eliminate any host specific code in the targets
> themselves -- the frontend and tools should have complete control over
> this. The idea is to add another routine to detect the set of
> "features", and then move tools like the JIT and llc to use this if
> they want to match the host, instead of relying on the backend to do
> it.
>
>  - Daniel
>
> On Fri, Nov 13, 2009 at 3:48 AM, Roman Divacky <rdivacky at freebsd.org> wrote:
>> hi
>>
>> the attached patch moves cpu detection code from lib/Target/$FOO
>> (for FOO=X86 :) ) to lib/System. This is necessary for clang
>> to be able to do -march=native. The patch was already reviewed
>> by Daniel Dunbar.
>>
>> Can you please review/commit this?
>>
>> thnx!
>>
>> roman
>>
>> p.s. CC: me as I am not subscribed to llvm-commits
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>
>>
>