[LLVMdev] Proposal: add intrinsics for safe division

[Filip Pizlo]

On April 30, 2014 at 9:06:20 PM, Philip Reames (listmail at philipreames.com) wrote:

On 04/29/2014 12:39 PM, Filip Pizlo wrote:
On April 29, 2014 at 11:27:06 AM, Philip Reames (listmail at philipreames.com) wrote:
On 04/29/2014 10:44 AM, Filip Pizlo wrote:
LD;DR: Your desire to use trapping on x86 only further convinces me that Michael's proposed intrinsics are the best way to go.
I'm still not convinced, but am not going to actively oppose it either.  I'm leery of designing a solution with major assumptions we don't have data to backup. 

I worry your assumptions about deoptimization are potentially unsound.  But I don't have data to actually show this (yet).
I *think* I may have been unclear about my assumptions; in particular, my claims with respect to deoptimization are probably more subtle than they appeared.  WebKit can use LLVM and it has divisions and we do all possible deoptimization/profiling/etc tricks for it, so this is grounded in experience.  Forgive me if the rest of this e-mail contains a lecture on things that are obvious - I'll try to err on the side of clarity and completeness since this discussion is sufficiently dense that we run the risk of talking cross-purposes unless some baseline assumptions are established.

I think we're using the same terminology, but with slightly different sets of assumptions.  I'll point this out below where relevant. 

Also, thanks for taking the time to expand.  It help clarify the discussion quite a bit. 
I think we may be converging to an understanding of what you want versus what I want, and I think that there are some points - possibly unrelated to division - that are worth clarifying.  I think that the main difference is that when I say "patchpoint", I am referring to a concrete intrinsic with specific semantics that cannot change without breaking WebKit, while you are using the term to refer to a broad concept, or rather, a class of as-yet-unimplemented intrinsics that share some of the same features with patchpoints but otherwise have incompatible semantics.

Also, when I say that you wouldn't want to use the existing patchpoint to do your trapping deopt, what I mean is that the performance of doing this would suck for reasons that are not related to deoptimization or trapping.  I'm not claiming that deoptimization performs poorly (trust me, I know better) or that trapping to deoptimize is bad (I've done this many, many times and I know better).  I'm saying that with the current patchpoint intrinsics in LLVM, as they are currently specified and implemented, doing it would be a bad idea because you'd have to compromise a bunch of other optimizations to achieve it.

You have essentially described new intrinsics that would make this less of a bad idea and I am interested in your intrinsics, so I'll try to both respond with why patchpoints don't currently give you what you want (and why simply changing patchpoint semantics would be evil) and I'll also try to comment on what I think of the intrinsic that you're effectively proposing.  Long story short, I think you should formally propose your intrinsic even if it's not completely fleshed out.  I think that it's an interesting capability and in its most basic form, it is a simple variation of the current patchpoint/stackmap intrinsics.





On April 29, 2014 at 10:09:49 AM, Philip Reames (listmail at philipreames.com) wrote:

As the discussion has progressed and I've spent more time thinking about the topic, I find myself less and less enthused about the current proposal.  I'm in full support of having idiomatic ways to express safe division, but I'm starting to doubt that using an intrinsic is the right way at the moment.

One case I find myself thinking about is how one would combine profiling information and implicit div-by-zero/overflow checks with this proposal.  I don't really see a clean way.  Ideally, for a "safe div" which never has the exceptional paths taken, you'd like to completely do away with the control flow entirely.  (And rely on hardware traps w/exceptions instead.)  I don't really see a way to represent that type of construct given the current proposal. 
This is a deeper problem and to solve it you'd need a solution to trapping in general.  Let's consider the case of Java.  A Java program may want to catch the arithmetic exception due to divide by zero.  How would you do this with a trap in LLVM IR?  Spill all state that is live at the catch?  Use a patchpoint for the entire division instruction?

We'd likely use something similar to a patchpoint.  You'd need the "abstract vm state" (which is not the compiled frame necessarily) available at the div instruction.  You could then re-enter the interpreter at the specified index (part of the vm state).  We have all most of these mechanisms in place.  Ideally, you'd trigger a recompile and otherwise ensure re-entry into compiled code at the soonest possible moment. 

This requires a lot of runtime support, but we already have most of it implemented for another compiler.  From our perspective, the runtime requirements are not a major blocker. 
Right, you'll use a patchpoint.  That's way more expensive than using a safe division intrinsic with branches, because it's opaque to the optimizer.

This statement is true at the moment, but it shouldn't be.  I think this is our fundamental difference in approach. 

You should be able to write something like:
i32 %res = invoke patchpoint (... x86_trapping_divide, a, b) normal_dest invoke_dest

normal_dest:
  ;; use %res
invoke_dest:
  landingpad
  ;; dispatch edge cases
  ;; this could be unreachable code if you deopt this frame in the trap handler and jump directly to an interpreter or other bit of code
I see.  It sounds like you want a generalization of the "div.with.stackmap" that I thought you wanted - you want to be able to wrap anything in a stackmap.

The current patchpoint intrinsic does not do this, and you run the risk of breaking existing semantics if you changed this.  You'd probably break WebKit, which treats the call target of the patchpoint as nothing more than a quirk - we always pass null.  Also, the current patchpoint treats the callee as an i8* if I remember right and it would be super weird if all LLVM phases had to interpret this i8* by unwrapping a possible bitcast to get to a declared function that may be an intrinsic.  Yuck!  Basically, the call target of existing patchpoints is meant to be a kind of convenience feature rather than the core of the mechanism.

I agree in principle that the intrinsic that you want would be a useful intrinsic.  But let's not call it a patchpoint for the purposes of this discussion, and let's not confuse the discussion by claiming (incorrectly) that the existing patchpoint facility gives you what you want.  It doesn't: patchpoints are designed to make the call target opaque (hence the use of i8*) and there shouldn't be a correlation between what the patchpoint does at run-time and what the called function would have done.  You could make the call target be null (like WebKit does) and the patchpoint should still mean "this code can do anything" because the expectation is that the client JIT will patch over it anyway.

Also, "patchpoint" would probably not be the right term for the intrinsic that you want.  I think that what you want is "call.with.stackmap".  Or maybe "stackmap.wrapper".  Or just "stackmap" - I'd be OK, in principle, with changing the name of the current "stackmap" intrinsic to something that reflects the fact that it's less of a stackmap than what you want.

To summarize.  A patchpoint's main purpose is that you can patch it with arbitrary code.  The current "stackmap" means that you can patch it with arbitrary code and that patching may be destructive to a shadow of machine code bytes, so it's really just like patchpoints - we could change its name to "patchpoint.shadow" for example.

If you were to propose such a stackmap intrinsic, then I think there could be some ways of doing it that wouldn't be too terrible.  Basically you want something that is like a patchpoint in that it reports a stackmap via a side channel, but unlike patchpoints, it doesn't allow arbitrary patching - instead the optimizer should be allowed to assume that the code within the patchpoint will always do the same thing that the call target would have done.  There are downsides to truly doing this.  For example, to make division efficient with such an intrinsic, you'd have to do something that is somewhat worse than just recognizing intrinsics in the optimizer - you'd have to first recognize a call to your "stackmap wrapper" intrinsic and then observe that its call target argument is in turn another intrinsic.  To me personally, that's kind of yucky, but I won't deny that it could be useful.

As to the use of invoke: I don't believe that the use of invoke versus my suggested "branch on a trap predicate" idea are different in any truly meaningful way.  I buy that either would work.



A patchpoint should not require any excess spilling.  If values are live in registers, that should be reflected in the stack map.  (I do not know if this is the case for patchpoint at the moment or not.)
Patchpoints do not require spilling.

My point was that with existing patchpoints, you *either* use a patchpoint for the entire division which makes the division opaque to the optimizer - because a patchpoint means "this code can do anything" - *or* you could spill stuff to the stack prior to your trapping division intrinsic, since spilling is something that you could do as a workaround if you didn't have a patchpoint.

The reason why I brought up spilling at all is that I suspect that spilling all state to the stack might be cheaper - for some systems - than turning the division into a patchpoint.  Turning the division into a patchpoint is horrendously brutal - the patchpoint looks like it clobbers the heap (which a division doesn't do), has to execute (a division is an obvious DCE candidate), cannot be hoisted (hoisting divisions is awesome), etc.  Perhaps most importantly, though, a patchpoint doesn't tell LLVM that you're *doing a division* - so all constant folding, strenght reduction, and algebraic reasoning flies out the window.  On the other hand, spilling all state to the stack is an arguably sound and performant solution to a lot of VM problems.  I've seen JVM implementations that ensure that there is always a copy of state on the stack at some critical points, just because it makes loads of stuff simpler (debugging, profiling, GC, and of course deopt).  I personally prefer to stay away from such a strategy because it's not free.

On the other hand, branches can be cheap.  A branch on a divide is cheaper than not being able to optimize the divide.



The Value called by a patchpoint should participate in optimization normally.  
I agree that you could have a different intrinsic that behaves like this.

We really want the patchpoint part of the call to be supplemental.  It should still be a call.  It should be constant propagated, transformed, etc..  This is not the case currently.  I've got a couple of off the wall ideas for improving the current status, but I'll agree this is a hardish problem. 

It should be legal to use a patchpoint in an invoke.  It's an ABI issue of how the invoke path gets invoked.  (i.e. side tables for the runtime to lookup, etc..)  This is not possible today, and probably requires a fair amount of work.  Some of it, I've already done and will be sharing shortly.  Other parts, I haven't even thought about. 
Right, it's significantly more complex than either the existing patchpoints or Michael's proposed safe.div.



If you didn't want to use the trapping semantics, you'd insert dedicated control flow _before_ the divide.  This would allow normal optimization of the control flow. 

Notes:
1) This might require a new PATCHPOINT pseudo op in the backend.  Haven't thought much about that yet.
2) I *think* your current intrinsic could be translated into something like this.  (Leaving aside the question of where the deopt state comes from.)  In fact, the more I look at this, the less difference I actually see between the approaches. 



In a lot of languages, a divide produces some result even in the exceptional case and this result requires effectively deoptimizing since the resut won't be the one you would have predicted (double instead of int, or BigInt instead of small int), which sort of means that if the CPU exception occurs you have to be able to reconstruct all state.  A patchpoint could do this, and so could spilling all state to the stack before the divide - but both are very heavy hammers that are sure to be more expensive than just doing a branch.

This isn't necessarily as expensive as you might believe.  I'd recommend reading the Graal project papers on this topic.

Whether deopt or branching is more profitable *in this case*, I can't easily say.  I'm not yet to the point of being able to run that experiment.  We can argue about what "should" be better all we want, but real performance data is the only way to truly know. 
My point may have been confusing.  I know that deoptimization is cheap and WebKit uses it everywhere, including division corner cases, if profiling tells us that it's profitable to do so (which it does, in the common case).  WebKit is a heavy user of deoptimization in general, so you don't need to convince me that it's worth it.

Acknowledged. 
Note that I want *both* deopt *and* branching, because in this case, a branch is the fastest overall way of detecting when to deopt.  In the future, I will want to implement the deopt in terms of branching, and when we do this, I believe that the most sound and performat approach would be using Michael's intrinsics.  This is subtle and I'll try to explain why it's the case.

The point is that you wouldn't want to do deoptimization by spilling state on the main path or by using a patchpoint for the main path of the division.

This is the main point I disagree with.  I don't believe that having a patchpoint on the main path should be any more expensive then the original call.  (see above)
The reason why the patchpoint is expensive is that if you use a patchpoint to implement a division then the optimizer won't be allowed to assume that it's a division, because the whole point of "patchpoint" is to tell the optimizer to piss off.



Worth noting explicitly: I'm assuming that all of your deopt state would already be available for other purposes in nearby code.  It's on the stack or in registers.  I'm assuming that by adding the deopt point, you are not radically changing the set of computations which need to be done.  If that's not the case, you should avoid deopt and instead just inline the slow paths with explicit checks. 
Yes, of course it is.  That's not the issue.



I'll note that given your assumptions about the cost of a patchpoint, the rest of your position makes a lot more sense.  :)  As I spelled out above, I believe this cost is not fundamental. 
You don't want the common path of executing the division to involve a patchpoint instruction, although using a patchpoint or stackmap to implement deoptimization on the failing path is great:

Good: if (division would fail) { call @patchpoint(all of my state) } else { result = a / b }

Given your cost assumptions, I'd agree. 
Not my cost assumptions.  The reason why this is better is that the division is expressed in LLVM IR so that LLVM can do useful things to it - like eliminate it, for example.


Bad: call @patchpoint(all of my state) // patch with a divide instruction - bad because the optimizer has no clue what you're doing and assumes the very worst

Yuck.  Agreed. 
To be clear, this is what you're proposing - except that you're assuming that LLVM will know that you've patched a division because you're expecting the call target to have semantic meaning.  Or, rather, you're expecting that you can make the contents of the patchpoint be a division by having the call target be a division intrinsic.  In the current implementation and as it is currently specified, the call target has no meaning and so you get the yuck that I'm showing.


Worse: spill all state to the stack; call @trapping.div(a, b) // the spills will hurt you far more than a branch, so this should be avoided

I'm assuming this is an explicit spill rather than simply recording a stack map *at the div*.  If so, agreed. 
I suppose we could imagine a fourth option that involves a patchpoint to pick up the state and a trapping divide instrinsic.  But a trapping divide intrinsic alone is not enough.  Consider this:

result = call @trapping.div(a, b); call @stackmap(all of my state)

As soon as these are separate instructions, you have no guarantees that the state that the stackmap reports is sound for the point at which the div would trap. 

This is the closest to what I'd propose, except that the two calls would be merged into a single patchpoint.  Isn't the entire point of a patchpoint to record the stack map for a call?  
No.  It would be bad if that's what the documentation says.  That's not at all how WebKit uses it or probably any IC client would use it.

Patchpoints are designed to be inline assembly on steroids.  They're there to allow the client JIT to tell LLVM to piss off.

(Well, ignoring the actual patching part..)  Why not write this as:
patchpoint(..., trapping.div, a, b);

Is there something I'm missing here?

Just to note: I fully agree that the two call proposal is unsound and should be avoided. 

So, the division itself shouldn't be a trapping instruction and instead you want to detect the bad case with a branch.

To be clear:

- Whether you use deoptimization for division or anything else - like WebKit has done since before any of the Graal papers were written - is mostly unrelated to how you represent the division, unless you wanted to add a new intrinsic that is like a trapping-division-with-stackmap:

result = call @trapping.div.with.stackmap(a, b, ... all of my state ...)

Now, maybe you do want such an intrinsic, in which case you should propose it! 

Given what we already have with patchpoints, I don't think a merged intrinsic is necessary.  (See above).  I believe we have all the parts to build this solution, and that - in theory - they should compose neatly.

p.s. The bits I was referring to was not deopt per se.  It was particularly which set of deopt state you used for each deopt point.  That's a bit of tangent for the rest of the discussion now though. 
The reason why I haven't proposed it is that I think that long-term, the currently proposed intrinsics are a better path to getting the trapping optimizations.  See my previous mail, where I show how we could tell LLVM what the failing path is (which may have deoptimization code that uses a stackmap or whatever), what the trapping predicate is (it comes from the safe.div intrinsic), and the fact that trapping is wise (branch weights).

- If you want to do the deoptimization with a trap, then your only choice currently is to use a patchpoint for the main path of the division.  This will be slower than using a branch to an OSR exit basic block, because you're making the division itself opaque to the optimizer (bad) just to get rid of a branch (which was probably cheap to begin with).

Hence, what you want to do - one way or another, regardless of whether this proposed intrinsic is added - is to branch on the corner case condition, and have the slow case of the branch go to a basic block that deoptimizes.  Unless of course you have profiling that says that the case does happen often, in which case you can have that basic block handle the corner case inline without leaving optimized code (FWIW, we do have such paths in WebKit and they are useful).

So the question for me is whether the branching involves explicit control flow or is hidden inside an intrinsic.  I prefer for it to be within an intrinsic because it:

- allows the optimizer to do more interesting things in the common cases, like hoisting the entire division.

- will give us a clearer path for implementing trapping optimizations in the future.

- is an immediate win on ARM.

I'd be curious to hear what specific idea you have about how to implement trap-based deoptimization with your trapping division intrinsic for x86 - maybe it's different from the two "bad" idioms I showed above.

I hope my explanation above helps.  If not, ask, and I'll try to explain more clearly. 
I think I understand it.  I think that the only issue is that:

- Patchpoints currently don't do what you want.

- If you made patchpoints do what you want then you'd break WebKit - not to mention anyone who wants to use them for inline caches.

So it seems like you want a new intrinsic.  You should officially propose this new intrinsic, particularly since the core semantic differences are not so great from what we have now.  OTOH, if you truly believe that patchpoints should just be changed to your semantics in a way that does break WebKit, then that's probably also something you should get off your chest. ;-)



One point just for clarity; I don't believe this effects the conclusions of our discussion so far.  I'm also fairly sure that you (Filip) are aware of this, but want to spell it out for other readers. 

You seem to be assuming that compiled code needs to explicitly branch to a point where deopt state is known to exit a compiled frame.  
This is a slightly unclear characterization of my assumptions.  Our JIT does deoptimization without explicit branches for many, many things.  You should look at it some time, it's pretty fancy. ;-)

Worth noting is that you can also exit a compiled frame on a trap (without an explicitly condition/branch!) if the deopt state is known at the point you take the trap.  This "exit frame on trap" behavior shows up with null pointer exceptions as well.  I'll note that both compilers in OpenJDK support some combination of "exit-on-trap" conditions for division and null dereferences.  The two differ on exactly what strategies they use in each case though.  :)
Yeah, and I've also implemented VMs that do this - and I endorse the basic idea.  I know what you want, and my only point is that the existing patchpoints only give you this if you're willing to make a huge compromise: namely, that you're willing to make the division (or heap load for the null case) completely opaque to the compiler to the point that GVN, LICM, SCCP, and all algebraic reasoning have to give up on optimizing it.  The point of using LLVM is that it can optimize code.  It can optimize branches and divisions pretty well.  So, eliminating an explicit branch by replacing it with a construct that appears opaque to the optimizer is not a smart trade-off.

You could add a new intrinsic that, like patchpoints, records the layout of state in a side-table, but that is used as a kind of wrapper for operations that LLVM understands.  This may or may not be hairy - you seem to have sort of acknowledged that it's got some complexity and I've also pointed out some possible issues.  If this is something that you want, you should propose it so that others know what you're talking about.  One danger of how we're discussing this right now is that you're overloading patchpoints to mean the thing you want them to mean rather than what they actually mean, which makes it seem like we don't need Michael's intrinsics on the grounds that patchpoints already offer a solution.  They don't already offer a solution precisely because patchpoints don't do what your intrinsics would do.



I'm not really arguing that either scheme is "better" in all cases.  I'm simply arguing that we should support both and allow optimization and tuning between them.  As far as I can tell, you seem to be assuming that an explicit branch to known exit point is always superior.


Ok, back to the topic at hand...

With regards to the current proposal, I'm going to take a step back.  You guys seem to have already looked in this in a fair amount of depth.  I'm not necessarily convinced you've come to the best solution, but at some point, we need to make forward progress.  What you have is clearly better than nothing. 

Please go ahead and submit your current approach.  We can come back and revise later if we really need to. 

I do request the following changes:
- Mark it clearly as experimental.

- Either don't specify the value computed in the edge cases, or allow those values to be specified as constant arguments to the call.  This allows efficient lowering to x86's div instruction if you want to make use of the trapping semantics. 
Once again: how would you use this to get trapping semantics without throwing all of LLVM's optimizations out the window, in the absence of the kind of patchpoint-like intrinsic that you want?  I ask just to make sure that we're on the same page.



Finally, as for performance data, which part of this do you want performance data for?  I concede that I don't have performance data for using Michael's new intrinsic.  Part of what the intrinsic accomplishes is it gives a less ugly way of doing something that is already possible with target intrinsics on ARM.  I think it would be great if you could get those semantics - along with a known-good implementation - on other architectures as well.

I would be very interested in seeing data comparing two schemes:
- Explicit control flow emited by the frontend
- The safe.div intrinsic emitted by the frontend, desugared in CodeGenPrep

My strong suspicion is that each would preform well in some cases and not in others.  At least on x86.  Since the edge-checks are essentially free on ARM, the second scheme would probably be strictly superior there. 

I am NOT asking that we block submission on this data however. 

But this discussion has also involved suggestions that we should use trapping to implement deoptimization, and the main point of my message is to strongly argue against anything like this given the current state of trapping semantics and how patchpoints work.  My point is that using traps for division corner cases would either be unsound (see the stackmap after the trap, above), or would require you to do things that are obviously inefficient.  If you truly believe that the branch to detect division slow paths is more expensive than spilling all bytecode state to the stack or using a patchpoint for the division, then I could probably hack something up in WebKit to show you the performance implications.  (Or you could do it yourself, the code is open source...)

In a couple of months, I'll probably have the performance data to discuss this for real.  When that happens, let's pick this up and continue the debate.  Alternatively, if you want to chat this over more with a beer in hand at the social next week, let me know.  In the meantime, let's not stall the current proposal any more. 

Philip

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20140430/8cddb746/attachment.html>