Re: add INLINEABLE to maybe, either, bool

I read this discussion and thought that one small thing we could do is to clarify the documentation for INLINE a bit. Here's what I came up with; comments welcome. I'll commit this if there are no objections. <para> GHC (with <option>-O</option>, as always) tries to inline (or “unfold”) functions/values that are “small enough,” thus avoiding the call overhead and possibly exposing other more-wonderful optimisations. GHC has a set of heuristics, tuned over a long period of time using many benchmarks, that decide when it is beneficial to inline a function at its call site. The heuristics are designed to inline functions when it appears to be beneficial to do so, but without incurring excessive code bloat. If a function looks too big, it won't be inlined, and functions larger than a certain size will not even have their definition exported in the interface file. Some of the thresholds that govern these heuristic decisions can be changed using flags, see <xref linkend="options-f" />. </para> <para> Normally GHC will do a reasonable job of deciding by itself when it is a good idea to inline a function. However, sometimes you might want to override the default behaviour. For exmaple, if you have a key function that is important to inline because it leads to further optimisations, but GHC judges it to be too big to inline. </para> Cheers, Simon On 16/09/2013 19:03, Austin Seipp wrote:

In light of some recent conversations with others and self-review, I realize my prior messages may have been too strong, come off as hostile, or outright combative.

I'd like to publicly apologize for that: I'm sorry to ruffle feathers. (As a GHC developer, what seems 'obvious' to me is much different than most people, I realize.)

Now, in light of some discussions I had on IRC, there *are* things we can do here, and I'd like to lightly recap my position and some other points. To wit:

* I think it is bad to overuse things like INLINE, and I believe it encourages people to not understand the implications of what the compiler is doing (all programmers generally must have some intuition and control over their programs, and how they run.)

* Using the INLINE hammer everywhere makes it *incredibly* difficult to see where GHC deficiencies are, and that's not what we want - it hurts our ability to have informed decisions and examples. I also find it slightly disheartening that many people don't think GHC can handle cases like this.

* But Haskell is a language where inlining may not make a constant-factor difference, but *orders of magnitude difference*. vector-algorithms is a good example of this, and I'm not sure anyone knows how to 'fix it' so it doesn't have to INLINE literally everything. We're talking 10 orders of magnitude difference, if I remember my conversations with Dan/Edward correctly. lens is a lesser example: there are cases where GHC won't inline due to fear of work duplication or other unusual cases, but we can tackle these in GHC in some cases (and have.)

* We tend to be quite sensitive to performance matters as a community, I feel.

* And sometimes, things are hard. Even for people like Simon, 'fixing' bad inliner behavior can be a monstrous task, and INLINE is certainly a way to help the compiler when its hands are tied.

Ultimately, nobody is wrong here. But we have options, and two of them people brought up are good ones I think.

1.) Perhaps GHC should have a flag to warn you if you use INLINE/INLINEABLE on a definition that the inliner would have dealt with anyway. This should not be on by default with -Wall. But it would give us a useful tool to examine our assumptions more easily in a lot of cases.

2.) GHC does have a testsuite with many performance-related tests, and tests that check the Core. We could easily add a test that checked the core output of bool, maybe, and either (and other functions, as time may go on.) This is much easier and probably more robust than trying to contrive an example of what the performance difference might be.

Personally, I am way more interested in #1, as opposed to #2 (a failure to inline something so small would quickly be noticed - because lots of things probably won't inline at that point and our tests will fail!) However, I believe both of these are relatively easy, and quite feasible to implement.

Unfortunately, I have about 10,000,000 things on my plate with the upcoming release. So I'm afraid I don't have time to do these myself.

So, patches welcome! However, I am more than willing to help people get their feet wet in doing the work. You can email me (same email I'm using now,) or contact me on IRC (freenode, nick 'thoughtpolice') if you prefer more real time communication. I'll help you to the best of my abilities if you'd like to give it a go.

On Mon, Sep 16, 2013 at 4:56 PM, Austin Seipp wrote:

...

A serious question: if you don't even trust GHC to inline 'bool', 'maybe' or 'either', given their triviality, do you trust it to ever inline anything at all? I'm being completely honest here.

It still ignores the question of *why* the inliner is failing to do what you want. If the type inferencer fails to infer the type of an utterly trivial function - let's again say 'bool :: a -> a -> Bool -> a', as it's type is about as trivial as it's definition - it is almost certainly broken. By the same token, GHC not inlining 'bool' under -O would almost certainly be a bug too, in my eyes. The definition is trivial to the point where we should not ask "what if it doesn't inline" - we should figure out WHY it does not do so. Maybe INLINE would be a justified way of fixing it, but in this case it's just unnecessary and has been verified as such.

By the same token, we also don't encourage people to wildly put `seq` everywhere, or make everything on earth strict just because it makes them feel good.

A compiler must work on a broad range of programs for a broad range of use cases. There are certainly some cases that the compiler is *not* tuned for. In some of these cases, we work to make them more efficient. We patch the compiler to make it better where-ever possible. But this case? This is nothing but a premature optimization in my eyes - and one that even people like Edward or myself are guilty of, for sure.* And I am repenting by rejecting the "INLINE school of thought" (or INLINE school of hammers, as it were.)

If you want to make the argument that 'bool' - or something else even - should be INLINE, by all means do so. But if you're going to do so without any empirical cases, or examples of why it should be so (especially when we have already checked the interface files,) and just say it lets you sleep better at night? I simply do not buy it.

On Mon, Sep 16, 2013 at 3:59 PM, Dan Burton wrote:

...

I'm wary of "let's not mark it as INLINE because we want the compiler to automagically inline it for us." This seems like saying we should not have type signatures, because we want the type inferencer to figure it out for us. (If you want to test the auto-inliner's wisdom, then just add a setting that ignores INLINE pragmas and see if it inlines the same thing that humans do?)

I don't really care how it's accomplished, but I do think that we should make sure that maybe, either, and bool are inlined, and the most obvious way to accomplish this is to directly mark them INLINE, is it not?

-- Dan Burton

On Mon, Sep 16, 2013 at 1:33 PM, Edward Kmett wrote:

...

Contrary to appearances, I fully agree. =)

On Mon, Sep 16, 2013 at 4:12 PM, Austin Seipp wrote:

...

I'm strongly opposed to this.

Being INLINE happy is not a good thing, it is a bad thing. More often than not, I see people stuffing INLINE all over the place for things that would trivially be unfolded and put in the interface file anyway. This is bad, and it teaches people to just use the INLINE hammer everywhere instead of understanding the actual implications of what the inliner does. It also makes it impossible to actually observe how the inliner behaves and see where it needs tuning: if we just mark everything INLINE, we might as well not have it and make it unconditional.

There are some particular cases where GHC is hesitant to inline small things if it would lead to work duplication, or where the inliner behavior is tweaked and you may want to force it across multiple versions to be sure (lens is a good example of this.) But this is far more rare, and this case is not that. In particular, Joachim checked the 'bool' commit. As expected, the unfolding for bool was put into the interface file for Data.Bool, meaning if you use -O (or just -O0 -fno-ignore-interface-pragmas,) it should be inlined at call sites appropriately when it is used.

If we're going to INLINE things, we need to make sure it actually has an empirical benefit, by looking at the core, and seeing where the inliner is failing. Not just attach it to things because it seems like a good idea. This also helps drive feedback into the inliner so we can see where it fails.

On Mon, Sep 16, 2013 at 2:59 PM, Carter Schonwald wrote:

...

Its come to my attention that maybe, either, and its new sibling bool, all lack the INLINEABLE attribute, or its more aggressive sibling INLINE

this seems like one of those operations where inlining in client use sites is a good option to have, and currently not possible!

theres probably other stuff that would benefit from an INLINEABLE pragma in base, but this is an obvious, simple, "easy win" that I noticed when Oliver's patch got merged into base.

Thoughts? Time scale: sometime this week? (ghc 7.8 merge window is landing!)

cheers

_______________________________________________ Libraries mailing list Libraries@haskell.org http://www.haskell.org/mailman/listinfo/libraries

-- Austin Seipp, Haskell Consultant Well-Typed LLP, http://www.well-typed.com/ _______________________________________________ Libraries mailing list Libraries@haskell.org http://www.haskell.org/mailman/listinfo/libraries

_______________________________________________ Libraries mailing list Libraries@haskell.org http://www.haskell.org/mailman/listinfo/libraries

-- Austin Seipp, Haskell Consultant Well-Typed LLP, http://www.well-typed.com/