hi, Pete, I understand InstCombine may simplify bitcast to nothing, but the order matters. AlwaysInliner happens in very early stage, and we never call it again. if InstCombine works, can we invoke it before Inliner? thanks, --lx On Mon, Jan 5, 2015 at 5:40 PM, Pete Cooper <peter_cooper at apple.com> wrote:> Hi lx, Philip > > I've seen an instcombine which helps with this situation. It fires when > the function types on both sides of the bitcast have the same number of > operands and compatible types. It then adds bitcasts on the arguments and > removes the one on the called function. > > I don't have IR to hand, but it would be worth passing your IR through > instcombine to see if that helps you. > > The idea of improving the inliner is also great, but you may find that > it's needed for cases other than this one if i'm right about the > instcombine. > > Thanks > Pete > > Sent from my iPhone > > On Jan 5, 2015, at 3:16 AM, Liu Xin <navy.xliu at gmail.com> wrote: > > Philip, > > I post here because I think AlwaysInliner should inline it. I want to > detect the indirect calls for Inliner, and I want to hear inputs. > > let me define indirect call first in my idea. In one single expression, > one function may be subject to bitcast more than one time. we can detect > this situation and treat it as a regular call of last function, is that > okay? > > thanks, > --lx > > > On Mon, Jan 5, 2015 at 7:32 AM, Philip Reames <listmail at philipreames.com> > wrote: > >> On 01/04/2015 12:04 AM, Liu Xin wrote: >> >>> >>> %294= callfloatbitcast (float(float, float*)* @__gpu_modff to >>> float(float, i64)*)(float%293, i64 %preg.212.addr.0) >>> >>> as you may know, some gpu backends don't support function call. we need >>> to make sure to inline all functions here. however, Inliner can not figure >>> out that this is a valid callsite in this form. actually, it is. in C >>> words, cast a function and then call should be treat as callsite, right? >>> >>> Generally, the inliner doesn't do much with indirect calls, but given >> there is no simpler canonical case here, I expect we'll have to. >> >> Its possible we might even want to define this as a direct call. I'm not >> sure what the expectations are with regards to the type of the function >> being called and the type of the callsite. I suspect a lot of code would >> get confused if getCalledFunction returned __gpu_modff with it's unbitcast >> type. That's possibly something we should fix though. >> >> We'll want to get other folks input here, but a small patch to the >> inliner to handle this case would seem reasonable to me. >> >> Philip >> > > _______________________________________________ > LLVM Developers mailing list > LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev > >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150105/bda1a63b/attachment.html>
Sent from my iPhone> On Jan 5, 2015, at 7:50 AM, Liu Xin <navy.xliu at gmail.com> wrote: > > hi, Pete, > > I understand InstCombine may simplify bitcast to nothing, but the order matters. AlwaysInliner happens in very early stage, and we never call it again. if InstCombine works, can we invoke it before Inliner?I don't know the order for sure, but I know instcombine runs a few times so probably one of those is before alwaysinline. Arguably, if the pass order is an issue, I'd be fine with making the code David found be a utility that the inliner can call. Not sure if anyone else thinks that's out of scope for the inliner, but I like the idea of it being able to make a call more inlinable. Thanks Pete> > thanks, > --lx > > >> On Mon, Jan 5, 2015 at 5:40 PM, Pete Cooper <peter_cooper at apple.com> wrote: >> Hi lx, Philip >> >> I've seen an instcombine which helps with this situation. It fires when the function types on both sides of the bitcast have the same number of operands and compatible types. It then adds bitcasts on the arguments and removes the one on the called function. >> >> I don't have IR to hand, but it would be worth passing your IR through instcombine to see if that helps you. >> >> The idea of improving the inliner is also great, but you may find that it's needed for cases other than this one if i'm right about the instcombine. >> >> Thanks >> Pete >> >> Sent from my iPhone >> >>> On Jan 5, 2015, at 3:16 AM, Liu Xin <navy.xliu at gmail.com> wrote: >>> >>> Philip, >>> >>> I post here because I think AlwaysInliner should inline it. I want to detect the indirect calls for Inliner, and I want to hear inputs. >>> >>> let me define indirect call first in my idea. In one single expression, one function may be subject to bitcast more than one time. we can detect this situation and treat it as a regular call of last function, is that okay? >>> >>> thanks, >>> --lx >>> >>> >>>> On Mon, Jan 5, 2015 at 7:32 AM, Philip Reames <listmail at philipreames.com> wrote: >>>>> On 01/04/2015 12:04 AM, Liu Xin wrote: >>>>> >>>>> %294= callfloatbitcast (float(float, float*)* @__gpu_modff to float(float, i64)*)(float%293, i64 %preg.212.addr.0) >>>>> >>>>> as you may know, some gpu backends don't support function call. we need to make sure to inline all functions here. however, Inliner can not figure out that this is a valid callsite in this form. actually, it is. in C words, cast a function and then call should be treat as callsite, right? >>>> Generally, the inliner doesn't do much with indirect calls, but given there is no simpler canonical case here, I expect we'll have to. >>>> >>>> Its possible we might even want to define this as a direct call. I'm not sure what the expectations are with regards to the type of the function being called and the type of the callsite. I suspect a lot of code would get confused if getCalledFunction returned __gpu_modff with it's unbitcast type. That's possibly something we should fix though. >>>> >>>> We'll want to get other folks input here, but a small patch to the inliner to handle this case would seem reasonable to me. >>>> >>>> Philip >>> >>> _______________________________________________ >>> LLVM Developers mailing list >>> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu >>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev >-------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150105/18d31c62/attachment.html>
On Mon, Jan 5, 2015 at 6:06 AM, Pete Cooper <peter_cooper at apple.com> wrote:> On Jan 5, 2015, at 7:50 AM, Liu Xin <navy.xliu at gmail.com> wrote: > > hi, Pete, > > I understand InstCombine may simplify bitcast to nothing, but the order > matters. AlwaysInliner happens in very early stage, and we never call it > again. if InstCombine works, can we invoke it before Inliner? > > I don't know the order for sure, but I know instcombine runs a few times > so probably one of those is before alwaysinline. > > Arguably, if the pass order is an issue, I'd be fine with making the code > David found be a utility that the inliner can call. Not sure if anyone else > thinks that's out of scope for the inliner, but I like the idea of it being > able to make a call more inlinable. >Honestly, I'm not sold on this one. always_inline is a really murky thing to begin with. I think the only reasonable stance is to insist that if the frontend *really* needs the inline to occur, it should arrange for the inline to be trivially do-able. Relying on the always inliner (much less other passes) having any kind of folding in order for the inlining to occur seems really risky. I'd actually like it if we could spec sufficiently restrictive rules, and make it a verify failure to violate them on a call marked always_inline so that frontends were required to diagnose this nicely to users or satisfy the constraints. Sadly, we are a very long way away from realizing a system like that. I think we should probably move in that direction though, even if it takes a long time to get there, as other patterns seem really hard to maintain and support long term. -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150106/57f8e4cc/attachment.html>