Disable arm-gepopt by default

Wed Apr 22 17:28:15 PDT 2015

Hi James,

Thank you for understanding! I think I won’t be able to describe the situation better than Gerolf has. I agree that it’d be much better if know this earlier, but we’ve just discovered that recently.

Thanks,
Michael

> On Apr 22, 2015, at 12:45 AM, James Molloy <james at jamesmolloy.co.uk> wrote:
> 
> Hi Michael and Gerolf,
> 
> Gerolf, I agree in general terms. However in gobmk these madds being removed are not being strength reduced, they are redundant. Not only that but they hold two registers live as arguments, so we're reducing both number of instructions and register pressure and not increasing any other metric. However the gain will depend on the number of multiply functional units your microarchitecture has.
> 
> Michael, thanks for that explanation. It shows that there really is a problem, where GEP matching is superior to arithmetic matching. It may just be that the pointer arithmetic emitted by GEP lowering differs slightly from canonical form and that is what our instruction selector is tuned for. But it's certainly nontrivial so that, combined with the other issues, makes me inclined to turn it off by default.
> 
> However, I robustly defend the position I took. I don't think pointing the finger at a pass due to one testcase over 9 months later and asking to disable that pass, without looking for an underlying cause, is the best way of improving LLVM. If the pass were newly added, I would feel differently of course.
> 
> Cheers,
> 
> James
> 
> 
> 
> On Wed, 22 Apr 2015 at 03:00 Michael Zolotukhin <mzolotukhin at apple.com <mailto:mzolotukhin at apple.com>> wrote:
> Hi James,
> 
> I’ve spent some time getting familiar with the source and looking into the generated assembly, and here is what I found.
> 
> First, let me start with some examples where the pass did some not-optimzal transformation. The code snippets are from random functions (not necessarily hot ones), but they still should give the idea.
> 
> Case 1:
> gepopt=true+D9148:
> mul	 x8, x8, x9
> stp	x11, x8, [sp, #32]
> add	x8, x25, #8
> str	x8, [sp, #48]
> add	x8, x25, #56
> str	x8, [sp, #16]
> movz	w8, #72
> mul	 x22, x26, x8
> add	x8, x25, x22
> ldrb	w10, [x8, #23]
> sxtb	w9, w10
> ldr	x11, [sp, #48]
> ldr	x11, [x11, x22]
> 
> gepopt=false:
> mul	 x28, x8, x9
> movz	w8, #72
> madd	x8, x26, x8, x25
> ldrb	w10, [x8, #23]
> sxtb	w9, w10
> ldr	x11, [x8, #8]
> 
> In this case we have a load from [x25+8+x26*72] address, and the pass split the expression into two parts: (x25+8) and (x26*72). That increased register pressure, and we had to generate a spill-fill pair for (x25+8) expression. Without the pass we generated madd and a load with a simple addressing mode.
> 
> 
> Case 2:
> Gepopt=true+D9148:
> ldr	x22, [x0, #136]
> orr	w9, wzr, #0x18
> mul	 x8, x8, x9
> add	x23, x22, x8
> str	wzr, [x23]
> orr	x26, x8, #0x4
> str	wzr, [x26, x22]
> add	x24, x8, #8
> str	wzr, [x24, x22]
> add	x25, x8, #16
> ldr	x27, [x25, x22]
> cbz	x27, 0xABCDABCD
> 
> Gepopt=false:
> ldr	x9, [x0, #136]
> orr	w10, wzr, #0x18
> madd	x22, x8, x10, x9
> str	xzr, [x22]
> str	wzr, [x22, #8]
> ldr	x23, [x22, #16]
> cbz	x23, 0xABCDABCD
> 
> In this case the hoisted common part seems to be just x22, and the we generate stores to [x22+x8*0x18], [x22+x8*0x18+4], and [x22+x8*0x18+8]. New addresses also seem to be more confusing for other optimization passes, so we weren’t able to merge adjacent stores, as we did with gepopt=false. We also didn’t fuse ‘add’ and ‘mul’ to ‘madd’ in this case.
> 
> 
> Looking into these examples, I thought that the problem might be that we do this transformation always and unconditionally - we don’t check if it’s beneficial at all. This would work fine if the transformation is free, i.e. if codegen always can roll everything back. But it’s not free, and in cases when we don’t actually hit the target (i.e. hoist a common part of several GEP-expressions), we often lose.
> 
> I see how this pass can be beneficial for targets like NVPTX which (according to comments in SeparateConstOffsetFromGEP.cpp) don't support complex addressing modes with several registers. However, it’s not that clear whether it’s beneficial or not on targets like ARM64. I’m inclining to believe that turning this pass in the current state on by default was the ‘large hammer’. I still think it could be beneficial, but it needs some more work for it.
> 
> Thanks,
> Michael
> 
>> On Apr 21, 2015, at 11:38 AM, James Molloy <james at jamesmolloy.co.uk <mailto:james at jamesmolloy.co.uk>> wrote:
>> 
>> Hi Michael,
>> 
>> Thanks. I'll take a look tomorrow. At least this is nothing to do with the pass itself- if re association regresses instruction selection, something is fragile!
>> 
>> Will you have time to look at this one today? It's now 19:30 here and I'm away for the evening.
>> 
>> Cheers,
>> 
>> James
>> On Tue, 21 Apr 2015 at 19:36, Michael Zolotukhin <mzolotukhin at apple.com <mailto:mzolotukhin at apple.com>> wrote:
>> Hi James,
>> 
>> It looks like the first patch gives nearly the same results as both patches together (at least, in terms of gains and regressions on geekbench). E.g. the code generated for mandelbrot is the same.
>> 
>> Thanks,
>> Michael
>> 
>>> On Apr 21, 2015, at 11:29 AM, James Molloy <james at jamesmolloy.co.uk <mailto:james at jamesmolloy.co.uk>> wrote:
>>> 
>>> Hi Michael,
>>> 
>>> Jingue pointed out that my second patch did pessimize some code, so do you still see the regressions with my first patch (the reassoc one)?
>>> 
>>> Cheers,
>>> 
>>> James
>>> On Tue, 21 Apr 2015 at 19:26, Michael Zolotukhin <mzolotukhin at apple.com <mailto:mzolotukhin at apple.com>> wrote:
>>> Hi James,
>>> 
>>> Thank you for taking your time to work on it! I tried the patches you posted and they did fix the problem on twofish. However, with them there are other regressions - and while you are at it, could you please take a look at the generated code? Probably, you can come up with another one-line fix.
>>> 
>>> So, one regression is on Geekbench/Mandlerbrot, here is the ‘fast’ and ‘slow’ disassembly files (fast - from reference compiler, slow - reference compiler+D9148+D9149).
>>> 
>>> 
>>> I can see that we generate less ‘madd’ instructions, but in some cases we end up with less-optimal sequences, including ‘mul', ‘add' and ‘or' instead. Also, in some cases we now generate 2 ‘str' instead of 1 ‘stp'.
>>> 
>>> Do you have any ideas what might go wrong this time?
>>> 
>>> Thanks,
>>> Michael
>>> 
>>>> On Apr 21, 2015, at 6:55 AM, James Molloy <james at jamesmolloy.co.uk <mailto:james at jamesmolloy.co.uk>> wrote:
>>>> 
>>>> See:
>>>> 
>>>> http://reviews.llvm.org/D9148 <http://reviews.llvm.org/D9148>
>>>> http://reviews.llvm.org/D9149 <http://reviews.llvm.org/D9149>
>>>> 
>>>> Cheers,
>>>> 
>>>> James
>>>> 
>>>> On Tue, 21 Apr 2015 at 12:06 James Molloy <james at jamesmolloy.co.uk <mailto:james at jamesmolloy.co.uk>> wrote:
>>>> Hi Michael, Gerolf,
>>>> 
>>>> The main goal of this optimization was to reduce the number of multiply-adds emitted when complex GEPs are used. With regards to SPEC, this is particularly prevalent in sjeng and gobmk (although we don't just care about SPEC and there were plenty of other places we caught this idiom).
>>>> 
>>>> The pass is actively reducing the number of multiply-adds in gobmk:
>>>> 
>>>> $ llc engine.ll -O3 -mllvm -aarch64-gep-opt=false -o - |& grep madd | wc -l
>>>>  297
>>>> 
>>>> $ llc engine.ll -O3 -mllvm -aarch64-gep-opt=true -o - |& grep madd | wc -l
>>>>  158
>>>> 
>>>> So it demonstrably is making at least some code better.
>>>> 
>>>> I was able to reproduce your twofish example. Twofish, like most crypto algorithms, has a pattern repeated many times so the problem could just be down to one pattern producing one more instruction, which it turns out is exactly the case.
>>>> 
>>>> The problem is that with gep-opt enabled, we're associating shifts and constants the other way around.
>>>> 
>>>> With gep-opt disabled (synthetic code example):
>>>>    add x0, x0, #72
>>>>    ldr x0, [x1, x0, lsl #2]
>>>> 
>>>> With gep-opt enabled:
>>>>    add x0, x1, x0, lsl #2
>>>>    ldr x0, [x0, #72]
>>>> 
>>>> This is actually fine, until another expression reuses a different shift with the same offset:
>>>> 
>>>> With gep-opt disabled:
>>>>    add x0, x0, #72
>>>>    ldr x0, [x1, x0, lsl #2]
>>>>    ldr x2, [x1, x0, lsl #3]
>>>> 
>>>> With gep-opt enabled:
>>>>    add x0, x1, x0, lsl #2
>>>>    add x2, x1, x0, lsl #3
>>>>    ldr x0, [x0, #72]
>>>>    ldr x2, [x0, #72]
>>>> 
>>>> Amusingly, I saw the opposite behaviour on AArch32 on a different testcase and came to the conclusion that expressions with correlated immediate offsets and identical shift amounts were more likely than expressions with correlated shift amounts and identical immediate offsets, but that's a different story.
>>>> 
>>>> Anyway, the reason this is happening is because we're not running reassociate after GEP expansion. If we run reassociate it all goes away, it's a one-line fix.
>>>> 
>>>> The two testcases that were disabled - well, they weren't really disabled, but the behaviour was changed for them. I put that down to a bug in the SeparateConstOffsetFromGEP pass, where it tries to separate a GEP with only one variadic offset:
>>>> 
>>>>   GEP i8* %base, i32 %offset
>>>> 
>>>> Nothing beneficial can come of that, so I've got, again, a single line patch to fix this. I'll push both patches upstream just as soon as I verify them.
>>>> 
>>>> So in summary, I don't see any need to disable this pass. It's doing good work, it just has one or two bugs.
>>>> 
>>>> Cheers,
>>>> 
>>>> James
>>>> 
>>>> On Tue, 21 Apr 2015 at 05:27 Michael Zolotukhin <mzolotukhin at apple.com <mailto:mzolotukhin at apple.com>> wrote:
>>>> Hi Tim and others,
>>>> 
>>>> So, the entire discussion is moving around the assumption that this pass is beneficial in some cases, but we don’t have such data. Please note that I don’t want to say that the assumption is wrong, but as far as I understand, only Hao observed some gains. Thus, until we hear from him, we have real slow downs versus presumable gains.
>>>> 
>>>> As for the looking into the code later - I’m not an expert in this area, and I don’t even have a motivational example for this optimization, so I’d barely be a right person for it. I’d be glad to assist with collecting some additional data/running tests if that might help though.
>>>> 
>>>> Thanks,
>>>> Michael
>>>> 
>>>> 
>>>> > On Apr 20, 2015, at 8:40 PM, Tim Northover <t.p.northover at gmail.com <mailto:t.p.northover at gmail.com>> wrote:
>>>> >
>>>> > On 20 April 2015 at 20:34, Gerolf Hoflehner <ghoflehner at apple.com <mailto:ghoflehner at apple.com>> wrote:
>>>> >> It shouldn’t come across as ‘extreme’.
>>>> >
>>>> > The "extreme" part wasn't so much the tentative disabling as "Could
>>>> > you and/or Hao take this on after the code is disabled/backed out?". I
>>>> > realise it wasn't an ultimatum, but it could very easily turn into
>>>> > "we're washing our hands of this", which I'd object to.
>>>> >
>>>> > I assume the pass was committed with some kind of metric that it
>>>> > improved, and now we have a counter-example. It's down to everyone to
>>>> > decide what the best way to proceed is.
>>>> >
>>>> > Cheers.
>>>> >
>>>> > Tim.
>>>> >
>>>> > _______________________________________________
>>>> > llvm-commits mailing list
>>>> > llvm-commits at cs.uiuc.edu <mailto:llvm-commits at cs.uiuc.edu>
>>>> > http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits <http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits>
>>>> 
>>>> 
>>>> _______________________________________________
>>>> llvm-commits mailing list
>>>> llvm-commits at cs.uiuc.edu <mailto:llvm-commits at cs.uiuc.edu>
>>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits <http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits>
>>> 
>> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150422/791817dd/attachment.html>