|
If you make a question about some super-fine peephole optimization, an answer that says "Trying to do anything like that is a waste of your time" is an appropriate answer.
Years ago, I could spend days timing and fine-tuning code, testing out various inline assembly variants. Gradually, I came to realize that the compiler would beat me almost every time. Instructions sequences that "looked like" being inefficient, actually run faster when I timed it.
Since those days, CPUs have gotten even bigger caches, more lookahead, hyperthreading and whathaveyou, all confusing tight timing loops to the degree of making them useless. Writing (or generating) assembler code to suppress single instructions was meaningful in the days of true RISCs (including pre-1975 architectures when all machines were RISCs...) running at 1 instruction/cycle with (almost) no exception. Today, we are in a different world.
I really should have spent the time to assembler code the example you bring up, with and without the repeated register load, and time them for you. But I have a very strong gut feeling of what it would show. I am so certain that I do not spend the time to do that for you.
Religious freedom is the freedom to say that two plus two make five.
|
|
|
|
|
I guess I just don't see looking at a new (to me) tech for code generation to see if it's doing what I expect in terms of performance as a waste of time.
To be fair, I also look at the native output of my C++ code. I'm glad I have. Even if not especially the times when it ruined my day, like when I realized how craptastic the ESP32 floating point coprocessor was.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
If you are working on a jitter for one specific CPU, or a gcc code generator for one specific CPU, and your task is to improve the code generating, then you would study common methods for code generating and peephole optimization.
If you are not developing or improving a code generator (whether gcc or jitter), the only reason for studying the architecture of one specific of them is for curiosity. Not for modifying your source code, not even with "minor adjustments".
It can be both educating and interesting to study what resides a couple of layers below the layer you are working on. But you should remember that it is a couple of layers down. You are not working at that layer, and should not try to interfere with it.
(I should mention that I grew up in an OSI protocol world. Not the one where all you know is that some people have something they call 'layers', but one where layers were separated by solid hulls, and service/protocol were as separated as oil and water. An application entity should never fiddle with TCP protocol elements or IP routing, shouldn't even know that they are there! 30+ years of OO programming, interface definitions, private and protected elements -- and still, developers have not learned to keep their fingers out of lower layers, neither in protocol stacks nor in general programming!)
Religious freedom is the freedom to say that two plus two make five.
|
|
|
|
|
Why not download ILSpy[^] and nosey at the produced IL code? Just compile your application in release mode and take a look at the produced IL to see whether it's been optimised. I would hazard a guess that it probably doesn't optimise something like that, but I could be wrong!
|
|
|
|
|
Because I'm not interested in the IL code, but in the post jitted native code.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
Maybe you ought to be more interested in the IL code.
The optimizations that is really significant to you application are not concerned with register loads, but with techniques such as moving invariant code out of loops, doing arithmetic simplifications etc. Peephole optimization (done at code generation) is a combination of "already done, always" and "no real effect on execution time".
I have had a few surprises with C# performance, but they were typically related to data structures, and they were discovered by timing at application level. To pick one example: I suspected that moving a variable out of a single-instance class, making it a static, would make address calculation simpler and faster, compared to addressing a variable within an object instance. I was seriously wrong; that slowed down the application significantly. I could have (maybe should have) dug into the binary code to see what made addressing a static location significantly slower, but as I knew the effect already, I didn't spend the time when I was working on that application.
Religious freedom is the freedom to say that two plus two make five.
|
|
|
|
|
I'm intimately familiar with the IL code already. I both generate code that then gets compiled to it, and I Reflection Emit it directly.
I get that you don't want me to be concerned about the things that I am concerned about. Get that I am anyway.
I already optimized at the application level.
I should add, I inlined one method and got a 20% performance increase. That's strictly jit manipulation. You don't think it's worth it. My tests say otherwise.
And one more thing - not paying attention to this? That along with some broken benchmarks (which shielded me from seeing the performance issues) led me into a huge mess.
Sure if you're writing an e-commerce site you don't have to be concerned with inner loop performance and "performance critical codepaths" because to the degree that you have them, they are measured in seconds to complete or longer.
Lexing, or regex searching is not that. If you don't think manipulating the jitter is worth it then why don't you ask microsoft why they mark up their generated regex code with attributes specifically designed to manipulate the jitted code?
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
"I should add, I inlined one method and got a 20% performance increase."
You are not telling, 20% of what? The entire application, or that specific function call?
And: Inlining is not peephole optimization. The jitter didn't do that. The compiler generating the IL did.
Inlining isn't quite at the same level as changing the algorithm, but much closer to that than to register allocation. In another post, I mentioned my experience with trying to make a variable static, rather than local to the single instance. Inlining is more at that level.
I am saying that modifying your source code to affect peephole optimization is a waste of energy. Inlining is at a different level, and might be worth it, especially if the method is small and called only in a few places.
Religious freedom is the freedom to say that two plus two make five.
|
|
|
|
|
20% of my total execution time, lexing a document end to end.
> The jitter didn't do that. The compiler generating the IL did.
Sorry, but that's just categorically false. The method is created and called in the IL code. It's only inline when jitted.
[System.Runtime.CompilerServices.MethodImpl(System.Runtime.CompilerServices.MethodImplOptions.AggressiveInlining)]
Feel free to play around and mark your code up with that attribute. Watch the compiled results, and the jitted results. You'll see the compiler still drops your method in the assembly, and still drops the callvirt opcode to call it. THE JITTER is what inlines it.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
Yes, I guess I was wrong about that. In C#, inlining is not absolute, just a proposal. The compiler/jitter is free to ignore the proposal. That depends on the cost of a call, which varies a lot with CPU architecture. I would guess that on an ARM, a lot more functions are not inlined, even if proposed by the developer, as calls are significantly cheaper on the ARM than on, say, x8x/x64.
Nevertheless, even if the code generator makes the final decision based on that CPUs specific instruction set and instruction timing, inlining is something you relate to at source code level. Compare it to unrolling a tight loop with a small, fixed number of iterations. Or use of #define expressions in C/C++. It is not at the level of which instructions are generated. (Well, of course all source code has an effect on code generated, but not at the level of selecting specific coding techniques.) If a method is inlined on both architecture X and architecture Y, that is the same structural code change, regardless of X and Y instruction set.
I saw the inlining option a generation ago, when it was a new concept. Then it was a directive to be honored, not a meek proposal. That was at a time when you could also direct a variable to reside in a specific register its entire lifetime. Experience showed that the compiler might know better ... (So we started trusting the compilers!).
Note that leaving the decision whether to inline or not might restrict the freedom of the higher level optimizer: If it takes care of the inlining above code generator level, it can e.g. combine common expressions in the inlined code with other code before or after the (inlined) call. While a code generator in principle could do a similar analysis of surrounding code, don't expect, it to be prepared to! The code to possibly be inlined will be inlined in extenso, even if identical expressions were calculated before or after the call. The task of a (code-independent) compiler is to discover such common expressions even when it takes the responsibility for inlining functions, while the code generator does not have a similar responsibily for restructuring the parse tree before generating code.
Religious freedom is the freedom to say that two plus two make five.
|
|
|
|
|
Maybe I didn't choose the best example in the OP, but it was the one most readily in front of me.
I'll say that about this - not knowing if the JITter would "know" that a repeated access off argument zero could be registerized is a fair question. I already know the answer of a traditional compiler. Here, (given my other most recent response to you, with an eye toward #1) the difference in performance would be significant, if my fear were realized about the actual generated code. I predict substantially more than a 20% difference in execution speed given how often I hit that field in my code. I can't easily test that, because I can't make the jitter do the wrong thing. So admittedly, it's a bit post hoc ergo propter hoc, but I wouldn't say it's a wild guess either.
But finding that out was significant. It wasn't about the CPU making adjustments to the microcode. It was higher than that level. The CPU can't figure that out. It requires at the very least peephole optimization, or better. I know a traditional compiler will do it, but I don't know the cost benefit calculus microsoft engage in in order to even decide if they thought it was worth it to do that optimization in the JITter for most purposes - my purposes being somewhat different than most purposes here.
I stand by that the question was worth knowing, and that the code would have been worth modifying in that worst case.
Because of the kind of code that it is. I'm not arguing general purpose development here. You know as well as I do that generalized rules aren't meant to cover every specific scenario- that's where the experience to know when to step outside those general rules is worth the cost and the potential risks.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
Another remark to micro-optimization:
In the 1990s, I was teaching basic computer architecture at a community college. To get a hands-on feeling of real registers and memory and such, the students did a few hand-ins of assembler coding. I was strongly stressing that we must strive to make code - especially assembler - as readable as possible. To zero the AX register, you move 0 into it: MOV AX, 0.
One of the students insisted that the proper way to zero AX is XOR AX, AX. No, I did not accept that; that is incomprehensible tribal language. The student insisted: But XOR is faster! So I brought him timing tables for both the 286 and 386 (our machines had 386s) to show him that the instructions took exactly the same number of clock cycles. He returned with timing diagrams for the 8086, showing that XOR would save 1 clock cycle on that processor. He wanted to write software running at maximum speed at all machines, even old 8086s! He was not willing to sacrifice a single clock cycle for the sake of more readable code!
So for the next homework problem, he handed in a listing headed by "This is how we are forced by the lecturer to program it:", and a very readable, neat solution. This was followed by a commented-out section: "And this is how real programmers do it:", with some really messy code.
I didn't care to uncomment the second alternative for the purpose of timing it.
Why did I come to think of this old memory right now?
Religious freedom is the freedom to say that two plus two make five.
|
|
|
|
|
We're not talking about xor ax,ax vs mov ax, 0 though.
We're talking about changing a field reference to a local variable.
You couldn't even see the xor ax,ax vs mov ax,0 in C. You'd have to drop to asm.
In this case, you certainly see it, and the code does something quite different in each case.
It's up to the jitter in this case, to make that leap of intent in the code, because compiler doesn't do it.
I suspected it did make that leap. But I won't apologize for checking, or deciding it's worth finding out.
Also
A) We're talking about generated code
and
B) This change does next to nothing in terms of affecting its readability.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
Caches do wonders. Even what looks as reloading the same from RAM will be cache hits. The performance penalty will be minimal.
Religious freedom is the freedom to say that two plus two make five.
|
|
|
|
|
Yeah, I mean, but cache is inherently somewhat unreliable, in that even at best its availability varies from processor to processor even on the same architecture in terms of amount and sometimes speed. It also is hard to figure cache misses, and sometimes hard to track down why your functions aren't being held in the cache lines - maybe your locality of reference stinks or something.
Look, I get your point about not bothering with this stuff in general. In general I wouldn't.
But this is
1. Critical code in terms of performance. I expect to be able to find tens of thousands of matches through some text in milliseconds. For code like that it is worth doing the legwork to find out how to make it faster. Even 20% is worth it.
2. Generated code, so the priorities aren't strictly about readability or maintainability, but even if they were, the change I proposed doesn't really impact readability. Furthermore, being that it's generated by a programmatic process, it pays even more to make that process produce the fastest code reasonable, given #1 (note I didn't say possible, I said reasonable. A 20% gain as a general rule of thumb is easily reasonable for generated code, even if readability is impacted somewhat). If .NET 8 changes the landscape significantly I'll update my code to reflect that.
3. I'm not doing anything Microsoft hasn't done in terms of optimizing their generated code to produce optimal output. I guarantee you they smoke this thing with benchmarks before every ship, and bits get twiddled. My goal was to beat them at this task, for reasons. So I have to play at the same level they are.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
I see your statement about a 20% gain.
Would you be willing to publish the complete source code of an application so that we can verify that the application runs 20% faster at the user level when we compile the code with a single "inline" either enabled or disabled, but otherwise identical code? Then we could compile it with varying optimization flags to see the effects on the user level timings. We could see the variations between x86 and x64. Maybe there are people here with access to machines running ARM processors so we can see the effect of inlining on those.
Generally speaking: 20% gain is essential in benchmark tests. If you set up a double-blind test (so that not even those managing the test knows), presenting users for a test that may either be a 100% speed one or a 120% speed one, and ask them what they see, you cannot expect a very high hit rate.
Religious freedom is the freedom to say that two plus two make five.
|
|
|
|
|
Thanks for making me check my work, even if I feel a little foolish. My initial claim was based on code that has long since evolved. The short answer is I cannot reproduce those initial findings in my current efforts.
I'm not yet saying that claim doesn't hold up. I can't. I could still be missing something pretty big, and my classes have some forks in their derivation chains for compiled vs generated vs runtime runners - i need to comb over that stuff before I can say the claim *doesn't* hold up either, because right now I'm not seeing any difference, and that's almost certainly not right. I have some conditional compiles and other things I need to check but I'm not in the headspace for running all that down at the moment. I probably will a little later tonight and repost unless I get distracted. Feel free to ping me if you don't hear from me about it. I feel a little silly not having some repro code available at this moment, but I didn't think I'd be backing this claim up in the lounge when i tested it either.  Either way, I need to actually focus on proving or disproving that claim at this point as a task in and of itself, because it's not just revealing itself to me at this point.
The Benchmarks project here is what i'm using, and FALIB_SMALLER *should* remove the inlining attribute from the runtime runners. The real deal though is the generated runners and I need to modify both the benchmarks and the generator code to allow for that option.
GitHub - codewitch-honey-crisis/VisualFA: A fast C# DFA regular expression engine[^]
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
"real world implications for some C# code that my library generates"
I don't think this is a good line of research for that reason.
They will change the JIT in the future. I wouldn't be surprised if there are minor version updates that change how it works.
So how are you going to validate that optimizations that you put into place for one single version will continue to be valid for every version in the future and in the past?
|
|
|
|
|
If it's such a significant difference in the generated code then yes.
Especially because in the case I outlined (turns out it does register access after all though) it would require relatively minor adjustments to my generated code to avoid that potential performance pitfall, and do so without significantly impacting readability.
I don't like to wait around and hope that Microsoft will one day do the right thing. I've worked there.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
I don't know of any developer using the gcc compiler suite who studies one or more of the code generators (quite a bunch is available) to learn how it works, in order to modify their source code to make one specific code generator produce some specific binary code. Not even "minor code adjustments".
The code generator part(s) of gcc is a close parallel to the dotNet jitter. The IL is analogous to the API between the gcc source code parsers (and overall optimizer) and the gcc code generator. When you switch to a newer version of a gcc compiler, you do not adapt your C, C++, Fortran, or whatever, code for making one specific code generator create the very best code. Well, maybe you would do it, but I never met or heard of anyone else who would even consider adapting HLL source code to one specific gcc code generator.
...With one possible exception: Way back in time, when you would go to NetNews (aka. Usenet) for discussions, there was one developer who very intensely claimed that the C compiler for DEC VAX was completely useless! There was this one machine instruction that he wanted the compiler to generate for his C code, but he had found no way to force the compiler to do that. So the compiler was complete garbage!
The discussion involved some very experienced VAX programmers, who could certify that this machine instruction would not at all speed up execution, or reduce the code size. It would have no advantages whatsoever to use that instruction. Yet the insistent developer continued insisting that when he wants that instruction, it is the compiler's d**n responsibility to provide a way to generate it.
I guess that this fellow would go along with you in modifying the source code to fit one specific code generator.
This happened in an age when offline digital storage was limited to (expensive) floppies, and URLs were not yet invented. I found the arguing from this fellow to be so funny that I did preserve it in a printout, where I can also find the specific instruction in question (I have forgotten which one), but that printout is buried deep down in one of my historical IT scrapbooks in the basement. I am not digging up that tonight.
Religious freedom is the freedom to say that two plus two make five.
|
|
|
|
|
You're comparing something that involves a total rewrite with a change that makes Advance() take an additional parameter, which it uses instead of current.
So really, you're blowing this out of proportion.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
What I am saying is: Leave peephole optimizing to the compiler/code generator, and trust it at that.
We have been doing that kind of optimizations since the 1960s (starting in the late 1950s!). It is routine work. Any reasonably well trained code generator developer will handle it well using his left hand. If you think you can improve on it, you are most likely wrong. And even if you manage to dig up some special case, for the reduction in time in the execution time of some user level operation, "percent" is likely to be a much too large unit.
Spend your optimizing efforts on considering algorithms, and not the least: data structures. These are way more essential to user perceived execution speed that register allocation. Do timing at user level, not at instruction level.
Religious freedom is the freedom to say that two plus two make five.
|
|
|
|
|
The other comment reminded me of what I did long ago.
The C compiler could be configured to emit assembler. I did that, optimized it, then I used that instead of the original C code in the build.
You could certainly do that here.
|
|
|
|
|
I could, and indeed I do one small optimization with my Reflect Emit based compiler that isn't possible - at least readily in C#.
if((codepoint>='A' && codepoint<='Z') || codepoint=='_' || (codepoint>='a' && codepoint<='a')
The comparison ranges are in sorted order left to right. So rather than run through all of the || conditions, I short circuit if the minimum of the next range in the series is greater than the codepoint. It's easy to do in IL since all of this is already resolved to a series of jumps. Not so easy in C#.
But I did it there because it was a minor change, and didn't really impact anything. I'd be far more hesitant to create a total fork in my compiler vs. source generator. The performance benefits would have to be compelling.
Fortunately, I didn't need to do that here, because my fears were not realized in the end. The JITter was smart enough to optimize that code. But if it wasn't, I could have reorganized my generated source code to produce more efficient IL, in that it would translate to more efficient native code on most platforms.
I'd have preferred that approach as it keeps things from being black boxed.
Check out my IoT graphics library here:
https://honeythecodewitch.com/gfx
And my IoT UI/User Experience library here:
https://honeythecodewitch.com/uix
|
|
|
|
|
jschell wrote: The C compiler could be configured to emit assembler. I did that, optimized it, then I used that instead of the original C code in the build.
Did you do much timing to verify the speed gain at the user level? (E.g. from the moment the user supplies some input up to the next request for input from the user. For some applications, you must use other criteria, such as the time from startup to the first screenful has been drawn.)
There are certainly compilers out there that are build by people not very experienced or knowledgeable about compiler optimization. On the other hand, I guess that the true experts keep some tricks up their sleeves. But I'd guess that 98% of all optimizing tricks that have a measurable effect has been documented in publicly available sources. If the compiler writer reads the proper sources and do what they tell him to, there is so little left for "even better" hand optimizing that it certainly isn't worth the effort.
Usually, the compiler knows a lot more than you do about the instruction set, cache, virtual memory etc. It has a lot more background for making the ideal code than you. If you change its proposed code, you will often just be doing it in a different way, but not measurably faster. In the worst case, your 'optimization' may even slow down the code. In the very best case, you may be able to prove a nanosecond or two faster execution - but you need a bag of tricks (not all of them very clean) to 'prove' a speedup at user level.
I have experience with all three of these alternatives. Back in history, you might get a real speedup. First, because on old CPUs the total execution time could be calculated by adding together the cycle counts for every instruction executed. Second, The Book of Well Known Standard Optimizing Tricks was a lot thinner then. The closer we get to the CPUs of today, with a long list of hardware speedup techniques (all sorts of prefetch and lookahead, caches, execution pipelines, speculative execution, ...), the more blurry the 'Add together the timing of all instructions executed' becomes. And The Book is quite thick.
Implement the tricks in the book, and there is not much left to do. But that implementation is the responsibility of the compiler writer.
If you must dig up a really bad compiler written by someone who has never hear of optimization, in order to prove that you can do hand optimizing that has some effect, then you should rather find a better compiler.
Religious freedom is the freedom to say that two plus two make five.
|
|
|
|
|
General 
News 
Suggestion 
Question 
Bug 
Answer 
Joke 
Praise 
Rant 
Admin 
Submit an article or tip
