Hello ! I have a problem which I can't resolve. I wrote program for 89C2051 for one of my friend in Basom Basic compiler (http://www.mcselec.com). Compiled code was about 1945 bytes. Since I prefer C, I transalte Bascom basic to C. But when I compile, generated code was 2313. I try to change optimization parameters, but generated code never go below 2313. Even bigger. Where am I wrong ? Is it possible that stupid Basic make smaller code tne best C compiler ? Then I find out that this code: while (1) { printf ("Hello World\n"); } generate fully 1093 bytes (!) for Atmel 89C2051. 1093 bytes???? Three lines? Same program in Bascom generate only 154 (!) bytes. When I remove all "printf" from my C code, generated code was arond 2200 byte. But, with same action in Basic code (remove all "print") 1340 bytes long code. Exactly same program in Keil C and Bascom basic generate 860 byte smaller code ! I don't belive this. Do I have to program in (stupid) Basic if I want smaller code ? I still belive I did mistake somewhere, so I need help. I will send both C and basic code for interested people. With best regards !
The method how to handle non-constant format strings is obvious, isn't it? Apparently, not obvious enough. The mechanism that you outline is suitable for an interpreted language, but a compiled application has to be self-sufficient. The app can not, at run time, upon encountering a foreign format specifier, ask the compiler for anything. - Dave
Dave, I can't see anything wrong with Hans's description. What he describes happens at compile time. His third point was that if the compiler sees a non-constant string anywhere in the code, then it gives up trying to optimise, and uses the full-featured version of printf().
You can get into situations where, for example, if there are multiple source files, each source module calls for a different "edition" of printf. If this is the case, it is actually up to the linker to resolve which printf gets included. It's fairly easy for the compiler to say, "Oh, this module calls printf and only has an unsigned int in the format string. OK, include printf#1." In another module, however, the compiler might say, "Oh. this module calls printf and uses floating-point format specifiers in the format string (but no unsigned ints are used). OK, include printf#4." Now, the linker has to make sense of all of these different printf functions. And, at the end of the day, the linker includes most of what's in the standard printf. The only applications this would work well for would be the Hello World example. I mean, if I include sprintf in my application, I'm going to use it to my best advantage. I'm not going to try to use a printf SUBSET to save memory. If memory is that precious, I'm going to write my own formatting routines. printf has been an acknowledged memory hog since C was invented. There have been very few successful attempts at reducing its size. For its size, I think it does quiet well. Jon
I just downloaded the BASCOM demo and played around for a while. It's pretty neat. After playing around for a while, I decided to run a real program and compare it to C51. Not wanting to spend the rest of my life on this, I chose the SIEVE. Here are the results: BASCOM: Code Size: 866 Bytes Execution Speed: 23.553726 Seconds Keil (using printf): Code Size: 1,248 Bytes Execution Speed: 6.331546 Seconds Keil (using putchar and puts): Code Size: 524 Bytes Execution Speed: 6.330123 Seconds I did all speed testing with the uVision2 Debugger/Simulator under identical circumstances. From this little comparison, I think you can tell that the printf is what kills the C code--but not the performance. And, the difference between a CUSTOM formatter (using puts and putchar) and printf saves 724 bytes. I'll be more than happy to post the source for these things (of course, the SIEVE source comes with the Keil tools) if anyone wants to check my work. Jon
And...with linker code packing... Keil (using putchar and puts and linker code packing): Code Size: 486 Bytes Execution Speed: 6.330105 Seconds Jon
Jon, Here's a suggestion for the manual. How about adding a note to the printf page, pointing out that puts/putchar use much less code memory for simple output.
Dave, I can't see anything wrong with Hans's description. What he describes happens at compile time. His third point was that if the compiler sees a non-constant string anywhere in the code, then it gives up trying to optimise, and uses the full-featured version of printf(). Hmmm.. okay, I went back and re-read the post that I (somewhat snippily) replied to, and I agree that it's technically possible. However, this feels like something that should be taken into account and compensated for (if necessary) on the application side rather than the tools side. If you start building this level of "special case" consideration into the toolset, you'll wind up with Win51, the first 32-bit (!) GUI OS for embedded 8051 applications, and I, for one, would rather skip that particular nightmare. Not to mention the extra development effort and cost and the price point impact it would no doubt have. IMHO, of course. You could also argue that this is the whole point of having an optimizer. It's late. I give up. =) - Dave
Nice job comparing SIEVE in both compilers. You now know whay I did the first post. But I didn't measere execution time. Where is the swith for linker code packing ? I make another threed post, where I looking Libraries for smart cards (eg. SLE4428). Anyoone help since this is most 'live' thread ? Mladen Bruck
"Where is the swith for linker code packing?" It's in the "Options for target" -> "C51" tab. Stefan
"And...with linker code packing... Keil (using putchar and puts and linker code packing): Code Size: 486 Bytes Execution Speed: 6.330105 Seconds" Jon, That's very interesting. What optimisation level did you use? Stefan
If tools are available that are beneficial to the project then use them. Don't rule things out merely because you've decided at some time that they are not appropriate. Stefan, You are thinking development, I am thinking end product. Using a bit of assembler I managed to change a product so that it could be manufactured for $1.83 less. That does not sound like much till I tell you that 500.000 were made. A modern C compiler can make code very close to (sometimes better than) a seasoned assembler programmer and for code that does not utilize special functions of the hardware C is, indeed, my choice. The REAL problem with C and more so with C++ is that it allow the programmer to obfusciate the code to a level where no single soul can figure out what happens. I have seen 500 character long if stetments. The worst about C is that it is stated to be "self documenting" I would rather buy swampland in Florida than attempt to read C code by someone that believes that. Stefan, I believe very much in "the right tool for the job" and my products now vary from 99% C and 1% assembler to a speed demon that is 25% C and 75% assembler. I have yet to see ANY C++ code that is maintainable and efficient. There is no way anyone is going to convince me that code that inherits and morphs is going to be more maintainable than code that does not. There is no absolute divide between controllers and processors in much the same way that there is no absolute divide between what is and is not suitable for use on each device You are right; however, if you program a microcontroller as a microprocessor and tell me that is the right thing to do, you better have some very good reason. You state about 5 levels into this discussion that you are doing the above for battery life reasons. I can accept that easily enough, but initially you stated it as a global right thing to do, which I can not agree with. Erik
What optimisation level did you use? The very highest one I could - 11. Jon
Where is the swith for linker code packing? The Use LX51 Linker must be enabled as well. Jon
I use C because I can crank things out very quickly. Often in a matter of minutes. This allows me to test and prove ideas and algorithms quickly. And, with the significant reqources available (examples, forum, and so on) C on the 8051 is pretty hard to beat. I mean with google and this forum, I rarely have to post questions asking for things -- I just have to be really, really good using the search tools that are available. Once everything works (or mostly works) I can start optimizing for speed. This is where I use the profiler to figure out where most of the program's time is spent. There's no need to speed up something that's only invoked a few times. Speed-up the stuff that's invoked a thousand times. I usually start optimizing for size from the very beginning. If I know that I'm size limited, I try to figure out what is the biggest memory hog that I'll have, is it data, long term storage, constants, code, or ???? Once I know what the big memory space killer is, I work on reducing that first. Then I work my way to the second biggest, and on down the line. Jon
"There's no need to speed up something that's only invoked a few times." Unless, of course, those few times happen to be in the critical path on which the whole system depends...! Just goes to show that all generalisations are bad...! ;-)
"What optimisation level did you use? The very highest one I could - 11." I'll bet you could shrink it a bit further using level 9. Stefan