The dialog box pops up, and I enter the filename for the app. Still, the gprof view is empty! Sometimes I get an "invalid binary format" messagebox, sometimes not. But I never get any profiling data to display in the gprof Eclipse view. I use MingW on Windows. Any suggestions are most welcome!
Additional date! After building with MingW on Windows using -pg, and running my app. However, when I use the gpof Eclipse view and double click on the same gmon. I have seen semihosting implementations being able to send only a few bytes per second. Ir recommend that you profile only a very small portion of your application first to see if things are workign properly.
Actually I try to profile only main. I have been waiting for about half an hour may be more and nothing has changed. Still the same. To be sure, simply disable it. The point is to reduce the problem as much as possible. Without actually debugging your code, is is very hard to find the problem. A few tips: make sure you have enough stack space allocated, just to be sure there is no stack overflow.
Another problem could be your version of gcc and ARM libraries. If you can, try my code with an NXP K64F board and using the the Kinetis Design Studio: that way you have a working reference project you can compare with. In this case a HardFault may occur or the program will continue working useing changed variable values. As a consequence, there may occur an infinite loop or any other unpredictible behavior. Could you please help me with several questions about your library? While initializing monstartup function uses lowpc and highpc argumets.
As I have understood — these arguments should be PC of the beginning of a piece of code we gonna debug and PC of end of this code. From this perspective, it is a magic number, and I only started from there to save some RAM. Be free to change the memory range for profiling. It turned out to be easier to dump the memory location of the profiling structs using gdb, and generate the gmon file using host-side code adjusting for bit host vs bit target.
But even after that, I see that gprof correctly constructs a symbol table and knows the addresses of everything, yet there are just way too many tags in the output. I thought it might have something to do with the way the PC always points some bytes past the present execution point seems the compiler pretty much picks a number here , but there appears to be no proper consistency.
Seriously thinking of just abandoning this…. I recommend to drop OpenOCD for this. Semihosting worked for me only in an very old 0. So what I recommend is to use a decent debugging interface like the Segger J-Link which a is fast and b it works. And which library newlib? It sounds like you are missing some low level library routines. Not sure, looks like a reset occurs. Are you using my files or something differnt? You have to step with the debugger through that call to see what is going on.
I think you are causing somehow a hard fault. Due to memory problem, i am unable to use proceed, is there a way to send these data in serial without storing them in memory.
To my knowledge this is not possible with gprof, because you have to have some tables in memory. Thanks for the clear info. How to over come this Problem? Hi Hossam, never seen that message. Can you try with instrumenting a very small file, and not the application. Just do it for a single function to see if that works.
Hi Federico, there is no need to store every possible call, it will record only the ones which are executed in the table. The from[] basically maps the from source locations, divided by a factor e. The to[] array is a hash table with overflow linkage too. That tos[] array grows while executing it, and has an upper limit with the number of possible destination addresses. I tried to run profiling in a STM32F All worked fiend but when I generate a gmou.
If not, can you extend the range passed by monstartup? Arglist at 0xf68, args: Locals at 0xf68, Previous frame's sp is 0xf78 Saved registers: r4 at 0xf68, r5 at 0xf6c, r7 at 0xf70, lr at 0xf I would check from where this 0x3afe74f0 is coming from the stack, or does it make any sense?
Or could it be some artefact from a stack overflow? Is there a standard implementation for the profiling library? I have started with such a library but never finished it due lack of time.
The other point to mention is: gprof is really independent of the target: all it does is consuming the data produced, and is not depending at all of the system producing the data.
Please let me know if this makes sense to you or if there is something I am missing. Thanks so much for your hard work. I am looking forward to trying this out.
So here is what I uncovered. Is my thinking correct? Doesn't ip contain the callee address? Thanks so much for your help and your hard work in this space. Like Liked by 1 person. Hi Juan, yes, this is correct. And yes, it returns to the callee the one called by the original function. Thanks for catching this! Will your code work for this processor? Talking about gmon. S, etc. Or do I need to rewrite some parts?
Any help most appreciated. Another question… I was able to profile code for some Tensilica processors some time ago, but I did not need any hardware board to be connected. I just generated the code for the target, and ran it in Eclipse or in command line. This produced gmon. Why is this not possible with ARM, why is a board required?
Then I would say you used a simulator or you were not actually running it on the hardware itself, but on the host PC. So not running Tensilica code at all. First of all thanks for the research on this, it is such a nice tool for embedded developers! About the question from the OP. As my requirements are more of doing unit tests for the application side, I have mocked some of the peripherals.
Hi Erich! Thank you so much for this article! I have a question about retrieving the previous value of PC in profile. Is this due to another function call? Thanks for the reply!
What made you choose 14 as offset? How can I make it 10kHz? You can go as high as you want. But keep in mind that this adds overhead, so depending on your target speed this will or will not work for you. These are not the same. If b you can use this trick , using the Pause button in Eclipse. Added: Maybe it would help to convey some experience of what performance problems are actually like, and where you can expect to find them.
Here are some simple examples:. Where is the hot-spot? Where is the problem? In the sort where string-compare is called. The point where string-compare is called is "cold", but that's where the problem is.
A small number of samples of the call stack pinpoint it with certainty. An app that loads plugins takes a long time to start up. A profiler says basically everything in it is "cold". But, stack samples show a large percentage of time is spent with the stack about 20 layers deep in the process of reading the resource part of plugin dlls for the purpose of translating string constants into the local language.
Investigating further, you find that most of the strings being translated are not the the kind that actually need translation. They were just put in "in case" they might need translation, and were never thought to be something that could cause a performance problem.
Fixing that issue brings a hefty time savings. So it is common to think in terms of "hotspots" and "bottlenecks", but most programs, especially the larger ones, tend to have performance problems in the form of function calls requesting work that doesn't really need to be done.
Fortunately they display themselves on the call stack during the time that they are spending. Stack Overflow for Teams — Collaborate and share knowledge with a private group. Create a free Team What is Teams? Collectives on Stack Overflow. Learn more. Asked 11 years, 11 months ago. Active 9 years, 9 months ago. Viewed 8k times. Improve this question. Pieter Pieter Which compiler are you using on Windows?
0コメント