Slashdot Mirror
Optimizations - Programmer vs. Compiler?
Saravana Kannan asks: "I have been coding in C for a while (10 yrs or so) and tend to use short code snippets. As a simple example, take 'if (!ptr)' instead of 'if (ptr==NULL)'. The reason someone might use the former code snippet is because they believe it would result in smaller machine code if the compiler does not do optimizations or is not smart enough to optimize the particular code snippet. IMHO the latter code snippet is clearer than the former, and I would use it in my code if I know for sure that the compiler will optimize it and produce machine code equivalent to the former code snippet. The previous example was easy. What about code that is more complex? Now that compilers have matured over years and have had many improvements, I ask the Slashdot crowd, what they believe the compiler can be trusted to optimize and what must be hand optimized?"
"How would your answer differ (in terms of the level of trust on the compiler) if I'm talking about compilers for Desktops vs. Embedded systems? Compilers for which of the following platforms do you think is more optimized at present - Desktops (because is more commonly used) or Embedded systems (because of need for maximum optimization)? Would be better if you could stick to free (as in beer) and Open Source compilers. Give examples of code optimizations that you think the compiler can/can't be trusted to do."
I think writing clear and easy to understand code is more important in the long run, especially if other people will have to look at it.
The sad truth is that, as far as optimization goes, this isn't where attention is most needed.
Before we start worrying about things like saving two cycles here and there, we need to start teaching people how to select the proper algorithm for the task at hand.
There are too many programmers who spend hours turning their code into unreadable mush for the sake of squeezing a few milliseconds out of a loop that runs on the order of O(n!) or O(2^n).
For 99% of the coders out there, all that needs to be known about code optimization is: pick the right algorithms! Couple this with readable code, and you'll have a program that runs several thousand times faster than it'll ever need to and is easy to maintain--and that's probably all you'll ever need.
Obliteracy: Words with explosions
Hard to measure, but what is the tradeoff between increased speed and increased readability (which is a prerequisite for correctness and maintainability)? And if you can estimate that tradeoff, which is more important to the goals of your application?
As a side note, it is far more important to make sure you are using efficient algorithms and data structures than to make minor local optimizations. I've seen programmers use bizarre local optimization tricks in a module that ran in exponential time rather than log time.
Sheesh, evil *and* a jerk. -- Jade
What about code that is more complex? Now that compilers have matured over years and have had many improvements, I ask the Slashdot crowd, what they believe the compiler can be trusted to optimize and what must be hand optimized?
Programmers cost lots more per hour than computer time. Let the compiler optimize and let the programmers concentrated on developing solid maintainable code.
If you make code too clever in an effort to try to pre-optimize, you end up with code that other people have difficulty understanding. This is leads to lower quality code as it evolves if the people that follow you are not as savvy.
Not only that, but the vast majority of code written today is UI-centric or I/O bound. If you want real optimization, design a harddrive/controller combo that gets you 1 GBps off the physical platter (and at a price that consumers can afford).
1) Code for maintainability
2) Profile your code
3) Optimize the bottlenecks
That said, (!ptr) should be just as maintanable as (ptr == NULL) simply because it is a frequently used 'dialect'. As long as these 'shortcuts' are used throughout the entire codebase they should be familiar enough that they don't get in the way of maintainability.
What you're talking about it micro-optimization.
Compilers are pretty good at that, and you should let them do their job.
Programmers should optimize at a higher level: by their choice of algorithms, organizing the program so that memory access is cache-friendly, making sure various objects don't get destroyed and re-created unnecessarily, that sort of thing.
...are doomed to repeat the biggest trap in computer programming over and over again:
"Premature optimization is the root of all evil"
If there's only one rule in computer programming a person ever learns, "Hoare's dictum" is the one I would choose.
Almost all modern languages have extensive libraries available to handle common programming tasks and can handle the vast majority of optimizations you speak of automatically. This means that 99.99% of the time you shouldn't be thinking about optimizations at all. Unless you're John Carmack or you're writing a new compiler from scratch (and perhaps you are) or involved in a handful of other activities you're making a big big mistake if your spending any time worrying about these things. There are far more important things to worry about, such as writing code that can be understood by others, can easily be units tested, etc.
A few years ago I used to write C/C++/asm code extensively and used to be obsessed with performance and optimization. Then, one day, I had an epiphany and started writing code that is about 10 times slower than my old code (different in computer language and style) and infinitely easier to understand and expand. The only time I optimize now is at the very very end of development when I have solid profiler results from the final product that show noticable delays for the end user and this only happens rarely.
Of course, this is just my own personal experience and others may see things differently.
With regard to your example, I can't imagine any modern compiler wouldn't treat the two as equivalent.
However, in your example, I actually prefer "if (!ptr)" to "if (ptr == NULL)", for two reasons. First the latter is more error-prone, because you can accidentally end up with "if (ptr = NULL)". One common solution to avoid that problem is to write "if (NULL == ptr)", but that just doesn't read well to me. Another is to turn on warnings, and let your compiler point out code like that -- but that assumes a decent compiler.
The second, and more important, reason is that to anyone who's been writing C for a while, the compact representation is actually clearer because it's an instantly-recognizable idiom. To me, parsing the "ptr == NULL" format requires a few microseconds of thought to figure out what you're doing. "!ptr" requires none. There are a number of common idioms in C that are strange-looking at first, but soon become just another part of your programming vocabulary. IMO, if you're writing code in a given language, you should write it in the style that is most comfortable to other programmers in that language. I think proper use of idiomatic expressions *enhances* maintainability. Don't try to write Pascal in C, or Java in C++, or COBOL in, well, anything, but that's a separate issue :-)
Oh, and my answer to your more general question about whether or not you should try to write code that is easy for the compiler... no. Don't do that. Write code that is clear and readable to programmers and let the compiler do what it does. If profiling shows that a particular piece of code is too slow, then figure out how to optimize it, whether by tailoring the code, dropping down to assembler, or whatever. But not before.
Note to ACs: I usually delete AC replies without reading them. If you want to talk to me, log in.
"Programs should be written for people to read, and only incidentally for machines to execute."
- Structure and Interpretation of Computer Programs
Not true. Many CPUs have a unary jump-if-zero, or a jump-if-nonzero operation. Thus the comparasin step can be bypassed since you know you're comparing to zero.
However, any compiler worth anything should find that and optimize it very easily in the case where you're comparing to a constant that evaluates to zero.
Don't label something "offtopic" unless you know the topic well enough to tell what's on topic.
...then the code isn't important enough to optimize. Plain and simple.
Never try to optimize anything unless you have measured the speed of the code before optimizing and have measured it again after optimizing.
Optimized code is almost always harder to understand, contains more possible code paths, and more likely to contain bugs than the most straightforward code. It's only worth it if it's really faster...
And you simply cannot tell whether it's faster unless you actually time it. It's absolutely mindboggling how often a change you are certain will speed up the code has no effect, or a truly negligible effect, or slows it down.
This has always been true. In these days of heavily optimized compilers and complex CPUs that are doing branch prediction and God knows what all, it is truer than ever. You cannot tell whether code is fast just by glancing at it. Well, maybe there are processor gurus who can accurately visualize the exact flow of all the bits through the pipeline, but I'm certainly not one of them.
A corollary is that since the optimized code is almost always trickier, harder to understand, and often contains more logic paths than the most straightforward code, you shouldn't optimize unless you are committed to spending the time to write a careful unit-test fixture that exercises everything tricky you've done, and write good comments in the code.
"How to Do Nothing," kids activities, back in print!
The compiler will perform strength reduction in all reasonable instances.
The compiler will raise invariant computations from inner loops in almost all cases that do not involve pointers.
The compiler knows how to optimize integer division in ways I wouldn't have even thought of.
The compiler sometimes "forgets" about a register and produces sub-optimal code for inner loops.
The compiler can't always tell what variable is most important to keep in a register in an inner loop.
Other stuff:
x^=y; y^=x; x^=y; optimizes to an XCHG instruction with gcc on x86. I was amazed that it could do that. (Yes, that piece of code exchanges x and y). On the other hand, tmp=x; x=y; y=tmp; doesn't get optimized to an XCHG. Obviously, the compiler is using a Boolean simplifier or identity-prover.
The compiler always assumes a branch will be taken (unless you use certain compiler switches to change this behavior). Thus you should always arrange your conditional tests so that the less-often executed code is within the braces.
Don't be afraid to write complex expressions. Subexpression elimination is almost foolproof in all instances where pointers are NOT involved. It's better to leave your code clear, and let the compiler optimize it.
And ABOVE ALL:
No matter how much the compiler optimizes your code, you can throw it all down the toilet with bad design by screwing the cache utilization. This is EXTREMELY important especially in graphical applications which process huge raster buffers. Row-wise processing is always more efficient than column-wise. Random access will kill your performance. Do not trust the memory allocator to keep your allocations together. Write your own allocator if you are dealing with thousands or millions of small, related chunks of information.
I could go on... But I must also second what others have said, which is to perform algorithmic optimizations FIRST and do not bother with constant-factor optimizations until you are CERTAIN that you are using the best algorithm. If you ignore this advice you might waste a week optimizing a three-line inner loop and then come up with a better algorithm the next week which makes all your hard work redundant.
I make my code easy to read for my own sanity. I've lived out this bash.org quote way too many times.
"Upon attaching the waterblock to my penis, I began to notice that I know nothing about computers." -- JRockway
Premature Optimization is the DEVIL! I repeat, it is the gosh darn DEVIL! Don't do it. Write clear code so that I don't have to spend days trying to figure out what you are trying to do.
The biggest mistake I see in my professional (and unprofessional) life is programmers who try to optimize their code is all sorts of "733+" ways, trying to "trick" the compiler into removing 1 or 2 lines of assembly, yet completely disregard that they are using a map instead of a hash_map, or doing a linear search when they could do a binary search, or doing the same lookup multiple times, when they could do it just once. It's just silly, and goes to show that lots of programmers don't know how to optimize effectively.
Compilers are good. They optimize code well. Don't try to help them out unless you know your code has a definite bottleneck in a tight loop that needs hand tuning. Focus on using correct algorithms and designing your code from a high level to process data efficiently. Write your code in a clear and easy to read manner, so that you or some other programmer can easily figure out what's going on a few months down the line when you need to add fixes or new functionality. These are the ways to build efficient and maintainable systems, not by writing stuff that you could enter in an obfuscated code contest.
+1 Insightful, -1 Troll. What can I say, I'm an Insightful Troll.
With the greatest respect to Linus, but writing a kernel does not make you the authority on programming. It does make you the authority on what particular style you allow in your CVS tree, but that's it.
I certainly agree that loop_counter is a bad name, though. But rather than use i, I prefer to at least make a note of what sort of objects I'm looping through.
For instance:
Code can never be 100% self documenting, but that's no reason not to settle for 0%. Whether you use CamelCase or words_broken_with_underscores is a matter of style, and you should stick with the style of the code base you're working on.
Anyone who can't or won't work with multiple languages or adopt the necessary style for an existing project is a poor programmer. When you create project, you create the rules. When you work on someone else's project, you follow the rules.
For a cheap, fast batch lookups I once wrote a hashed matrix using STL. Loaded all the cells, dynamically typed, added indexes on the data for that run, and then passed around this collection of in-memory tables to our routines. Ran fast and was simple to debug, since all the traversing was O(ln(n)) based (or a variant thereof). Adding serialization, we could distribute to machines overnight dynamically and cut the run to a few minutes - from almost 8 hours.
Until it came time to dipose the memory. The STL slowly crawled tons of our objects, and the C++ dispose pattern was just too inefficient for all the stack hits. So we pointed the library at a custom heap and never disposed the dictionary - we just disposed the heap in bulk.
All written without hesitation for "longhand" syntax. (and btw, its "if ( NULL == var ) " to those that care). The code optimized fine, with just a few choice inlines we got to stick. No reg vars, no assembly piles littering the code.
But this was an in-house business app, and the lifecycles / requirements are different than other products. However, because of the nice algorithms, optimization wasn't difficult, and didn't rely on code tricks. If you're squabbling over code tricks for optimization, you're choosing the wrong algorithm, to me.
I've known some very good first rate programmers who religiously put the constants on the left. I've never known a second rate programmer who did.
Second: Do it later. There are thousands of situations where you can postpone the actual computations. Imagine writing a Matrix class with the invert() method. You can actually postpone calculating the inverse of the matrix until there is a call to access on of the fields in the matrix. Also you can calculate only the field being accessed. Or at some sensible threshold you may assume that the user code will read the entire inverted matrix and you can just calculate the remaining inverted fields... the options are endless.
Most string class implementations already make good use of this rule by only copying their buffers only when the "copied" buffer changes.
Third: Apply minimum algorithmic complexity. If you can use a hashmap instead of a treemap use the hash version it's O(1) vs Olog(n). Use quicksort for just about any kind of sorting you need to do.
Fourth: Cache your data. Download or buy a good caching class or use some facilities your language provides (eg. Java SoftReference class) for basic caching. There are some enormous performance gains that can be realized with smart caching strategies.
Fifth: Optimize using your language constructs. User the register keyword, use language idioms that you know compile into faster code etc... Scratch this rule! If you're applying rules one to four you can forget about this one and still have fast AND readable code.
Your pizza just the way you ought to have it.
"My rule is never comment what the program does, comment why it does it."
Bah. Comments lie. Code never lies.
Need Mercedes parts ?
one product
one customer
420,000 lines
260 staff
no competition
no trade shows
no salespeople selling new features that have never been discussed
It's interesting to talk about their attention to detail, but to hold it up as a model for all software development neglects to consider that they are working under an entirely different set of constraints from most everyone else.
I think the example is fine; you just displayed an assumption that highlights one of the quirks of C.
! means "not" or "inverse of"; it is a boolean function. The variable ptr is a pointer; it is a reference to data, which means it isn't really data itself. !ptr shouldn't compute; a boolean operator should only work on boolean data. But C logical comparators are designed to work on everything. You are just supposed to know that 0 == NULL == false. This supposition is totally arbitrary and doesn't hold up in any language with strong typing.
This is what makes C difficult for beginners. Bad code compiles even though it has logical flaws, and ends up failing in mysterious ways.
The second case makes more sense. Equality is an operator that should work on all types of data. NULL is necessary if you are going to abstract data through the use of pointers or objects. Doing away with NULL would be equivalent to eliminating true and false and using 1 and 0 instead. Or eliminating strings and using sequences of ASCII codes. These substitutions are technically correct but in reality they make code unreadable.
Ten years of programming in the language and you:
1) Don't know when two things are obviously equivalent to any non-brain dead compiler.
2) Think something other than readability matters.
3) Think the non-idiomatic way of doing something is more readable.
But I'm sure I'm just repeating the comments I can't be bothered reading.
Most compilers today will get all the simple stuff like if (!ptr) vs if (NULL == ptr) optimization. Its the more complex things that the compiler cannot "prove" where it has trouble. For example:
void h(int x, int y) {
for (i=0; i < N; i++) {
if (0 != (x & (1 << y))) {
f(i);
} else {
g(i);
}
}
}
Very few compilers will dare simplify this to:
void h(int x, int y) {
if (0 != (x & (1 << y))) {
for (i=0; i < N; i++) f(i);
} else {
for (i=0; i < N; i++) g(i);
}
}
Because the compilers have a hard time realizing that the conditional is constant and should be hoisted to the outside of the for loop. The compiler has the opportunity to perform loop unrolling in the second form that its may not try in the first instance.
You can learn these things from experience, or you can simply figure it out for yourself with the afore mentioned decompilation tools.
Here's my "guide to optimizing":
1) Are you disk I/O bound? You might need to switch to memory mapped files, or you might need to tweak the settings on the ones you have. You might need to use a lower level library to do your I/O. Many C++ iostreams implementations are slow, and many similar libraries involve lots of copying.
2) Are you socket I/O (or similar) bound? If so, you may need to rewrite with asynchronous I/O. This can be a PITA. Suck it up.
3) Are your threads spending all their time sitting in locks waiting for other threads? One, make sure you're using an appropriate number of worker threads optimized by the number of CPUs the host has. If you've already got the right number of threads, this can be a really tough decision. Presumably, the threads are helping your program readability, and trying to rework things into fewer threads is often a *bad idea*.
4) Are you spending all your time in malloc/new/constructor free/delete/deconstructor? Maybe you need to keep things on the stack, use a garbage collector, use reference counted objects, use pooled memory techniques, etc. In the right places, switching from some "string" library to const char* and stack buffers can give a huge benefit. Make sure, of course, that you use the "n" version of all standard string functions (the ones that take the size of the buffer as an argument) to avoid buffer overruns.
5) Are you spending all of your time in some system call? Like maybe some kind of WriteTextToScreen or FillRectangleWithPattern type of thing? For drawing code in general, try buffering things that are algorithmically generated in bitmaps, and only regenerate the parts that change. Then just blit together the pieces for your final output. Perhaps you need to rely on hardware transparency support for fast layer compositing. You might need fewer system level windows so you draw more in one function. Maybe you need to reduce your frame rate.
6) Are you using memcpy as appropriate?
If any of the previous items are true, you have no business worrying about the compiler. However, once you've gotten this far, you can start worrying about optimizing your code line by line.
7) Since you've gotten this far, the line(s) of code you're worried about are all inside some loop that gets run. A lot. They may be inside a function that's called from a loop too, of course. So, a few things to consider. A) You may need to use templates to get code that is optimized for the appropriate data type. B) You may need to split off a more focused version of the function from the general purpose function if it's also used in non-critical areas. This has negative maintainance ramifications. C) Do the bonehead obvious stuff like moving everything out of the loop that you can. D) Look at the assembly actually generated by your compiler. If you're not confortable with this, you have no business doing further optimization.
After looking at the assembler, then you'll know if the following are important. In my experience, they are.
1) Change array indexing logic to pointer logic:
can change to:
This eliminates lots of redundant addition. All of those stuff[i] = val type of statements tend to generate:
mov
No, you're adopting a black or white approach. You are, in essence, saying that you don't need to comment at all. The original poster was saying that comments needed to be everywhere, on everything. I believe in a middle ground approach.
I comment things that are non intuitive. I comment things that I *think* may be non intuitive. I comment things that I think someone else might have some difficulty understanding, because I happened to be deep into a code burn and consequently wrote something pretty tight, pretty sweet, but also pretty obfuscated. Finally, I comment things that I think *I* may not understand when I go back and look at the code again 3 months from now.
I don't comment every single line... I don't comment simple data structures, loops "/* this is a for loop using the integer variable I */" etc which would be stupid. I do however disassemble the complex portions of my code, describe how I'm dispatching events and best of all *why* I decided to do things a certain way instead of a different way.
I have, however, been handed 30k lines of code with zero documentation and not a single comment anywhere in it, with absolutely no clue at all how it worked and no access to the original programmer and been told "We need such and such fixed|updated|added by friday" and had to spend the entire week basically tracing every single line of code to figure out that the original programmer must have been smoking crack with NO indication of why he wrote things how he did and NO help when he decided to be exceedingly "clever"
in his code. That time was wasted.
Would it have killed him to simply put a comment block explaining his event dispatch model? Or to tell me what his functions and methods did and best of all why they did it?
There *is* a middle ground, believe it or not.
-- Gary F.
This issue is in like this,
You need to understand the language, both syntax AND semantics you are using
this ranges from the simple to mind-bending e.g. C++ (I am convinced that not even Bjarne Stroustrup understands this evil language);
at that point you have two bi-furcations (a) interpreted languages eg Java, Perl, PHP and Python -v- (b) cpmpiled languages, and (c) finally DIY (do it your self) Assembler
So: what does it amount to in practice? A) Rock Bottom, understand the architecture, including virtual memory, architecture and instruction set issues, read and understand the chip data sheet. Hard! See bottom line, architecture dependand code in Linux, bsd ...
B) use 'gcc -S' and write the code in C, hand improve the assembler output, this is what I normally do, but you need to keep an open mind otherwise you miss things, I once took a compute intensive algorithm for the M68020 and made it run 10'000 times faster using this approach
C)consider hardware optimisation; strictly price/performance.
In terms of optimizing, generally compilers do a pretty good job, however there are several areas that no compiler I know of can help.
1. Choose the right algorithm. For example, in an embedded project I worked on an engineer used a linked list to store thousands of fields that must be added and deleted. While adding is fast, it didn't scale for deleting. Changed it to a hash table and it sped it up significantly.
2. Know your data and how it is used. Knowing how to organize your data and access it can make a huge difference. As a previous poster pointed out, sequential memory accesses are much faster than random accesses. I had to do some 90 degree image rotation code. The simple solution just used a couple for loops when copying the pixels from one buffer to another. In another, I took into account the processor cache and how memory is accessed and broke it down into tiles. The first algorithm, while simple and elegant ran at 30 frames per second. The other ran at over 200 frames per second. Looking at the code the first algorithm should be faster since the code is simpler. Both algorithms operate in O(N) time, where N=width * height.
Further optimization attempts to hint to the CPU cache about memory made no difference (Athlon XP 1700+). The only possible way I see to speed it up further would be to write it in hand-coded assembler.
3. Reduce the number of system calls if possible. Some operating systems can be very painful when calling the kernel. Group reads and writes together so fewer calls are made.
4. Profile your code to find bottlenecks.
5. Try and keep a tradeoff between memory usage and performance. A smaller tightly packed data set will execute faster with CPU caches and will reduce page faults when loading and starting up.
6. Try debugging your code at the assembler level, stepping through it. It will help you better understand your compiler.
7. Don't bother trying to optimize things like getting every ounce of performance when the next function you call will be very slow. I.e. in one section of MS DOS's source code which was hand-coded assembly language it was calculating the cluster or sector of the disk to access. First the code checked if it was running on a 16-bit or 32-bit CPU. Next it took the 16-bit or 32-bit path for multiplication, then it read from the disk. Why the hell write all this code to check the CPU if it's 16 or 32 bit for the multiply when the frigging disk is going to be slow. They should have just stuck with the 16-bit multiply rather than be clever.
In general applications with GCC, I rarely see much difference between -O2 or -O3. For that matter, I often don't see a noticable difference between -O0 and -O3 for a lot of code.
I only see improvements in some very CPU intensive multimedia code. I also saw a significant improvement in some multimedia code when I told the compiler to generate code for an Ultrasparc rather than the default, but that's because the pre-ultrasparc code didn't use a multiply instruction.
-Aaron
This post is encrypted twice with ROT-13. Documenting or attempting to crack this encryption is illegal.
Do not perform minor optimizations without first: a) Determining there is a performance problem b) Profiling your code to determine what areas should be optimized.
This does not mean that you should choose naive algortithms for the problem at hand. Choosing the proper algorithm for the problem at hand is always important.
Hand-optimized code should be reserved for those times when you have profiled your code with reasonable inputs and have shown that the lack of clarity is compensated for by the increased performance.
The example you gave is a perfect example of a hand optimization that is completely worthless with today's compilers.
"I ask the Slashdot crowd, what they believe the compiler can be trusted to optimize and what must be hand optimized? Give examples of code optimizations that you think the compiler can/can't be trusted to do."
/.
Somehow 99% of the readers took this to mean "What is the difference between NULL and the zero bit pattern and do you think it is a good idea to write clear code and do the profile/algorithm change cycle until there is nothing left to optimize or should I write low level optimized code from the start?"
sigh.. I've only found two comments with code so far after going through hundreds of posts. This is possibly the worst signal to noise ratio I've witnessed on