[llvm-commits] [llvm] r132676 - in /llvm/trunk: lib/CodeGen/SelectionDAG/FastISel.cpp test/CodeGen/X86/fast-isel-agg-constant.ll

[Jim Grosbach]

On Jun 9, 2011, at 12:48 PM, Roman Divacky wrote:

> On Thu, Jun 09, 2011 at 11:22:41AM -0700, Jim Grosbach wrote:
>> 
>> On Jun 9, 2011, at 1:08 AM, Roman Divacky wrote:
>> 
>>> On Wed, Jun 08, 2011 at 11:08:36PM -0700, Eric Christopher wrote:
>>>> 
>>>> On Jun 6, 2011, at 8:38 AM, Roman Divacky wrote:
>>>> 
>>>>> OK to commit? Or is there a better way to distinguish those two
>>>>> asm dialects?
>>>> 
>>>> Maybe check the Subtarget instead?
>>> 
>>> Where do I get it in MCExpr?
>> 
>> If at all possible, we want to avoid the Subtarget in MC layer operations. There are a few places it's done, but it's a layering violation that we're trying very hard to avoid. Long term, we want to refactor the target machine to allow us to decouple it from MC.
>> 
>> For purposes of something like this, you would theoretically just use the target triple. That is, if the triple specifies the target is Darwin (or not), behave accordingly. I don't think the triple is explicitly threaded through MC, though, so other than conceptually, I don't think that helps you much.
> 
> The thing is that in MCExpr there's absolutely nothing related to triples etc.
> It's just a pure container for an expression. Nothing from the outside world
> is there :)
> 
> Thats why I introduced the two new variant kinds, to distinguish the two
> dialects.

Yeah. It's a bit of an ugly solution, but it's pretty much necessary for now. IIRC, there are precedent examples of the same thing in there. We really need a target-hook that has access to the triple level information for this sort of thing, but there's no such beast right now.

-Jim