Instruction instrumentation

[Karlos]

Hi,

Is it possible, perhaps using the debugger, to profile instruction counts, by type, for some duration? I don't mean cycle exact or anything, just a set of opcode/count pairs over some time interval.

[Toni Wilen]

Doing only that would most likely return totally useless results..

How do you define an "instruction"? Counts per unique opcode words would be easy enough but that can't be used to separate all addressing modes. Or do you only need to separate "base" instructions? (MOVE, NOP, SUB etc)

And what about OS routines? Filesystem handlers? Device drivers? You can't detect when you are running OS code and when "your" code easily. One possibility is to use "uaedbg" mode where whole binary gets loaded in normally non-existing RAM space.

Not taking these in account would return totally useless information. Unless program takes over the system.

[Karlos]

Originally Posted by Toni Wilen

Doing only that would most likely return totally useless results..

How do you define an "instruction"? Counts per unique opcode words would be easy enough but that can't be used to separate all addressing modes. Or do you only need to separate "base" instructions? (MOVE, NOP, SUB etc)

And what about OS routines? Filesystem handlers? Device drivers? You can't detect when you are running OS code and when "your" code easily. One possibility is to use "uaedbg" mode where whole binary gets loaded in normally non-existing RAM space.

Not taking these in account would return totally useless information. Unless program takes over the system.

I get that, I was looking for something broad strokes on an a CPU bound application that, until recently did take over the system. As for how it's collated, by opcode is fine but I'd be looking to collate that further. I'm looking for the counts of particular time consuming operations, like division.

More ideally, I would like to be able to call a trap of some kind that would initiate the instrumentation and a later trap that would disable it and log it somewhere. The scope of this code is small and doesn't call any library routines between the points I want to instrument. Now any kind of task switch or other interrupt in the duration of the interval will skew the results but it's not particularly likely in this case and I could just as easily invoke Disable() and Forbid() outside the region I'm instrumenting.

[Toni Wilen]

Originally Posted by Karlos

I get that, I was looking for something broad strokes on an a CPU bound application that, until recently did take over the system. As for how it's collated, by opcode is fine but I'd be looking to collate that further. I'm looking for the counts of particular time consuming operations, like division.

More ideally, I would like to be able to call a trap of some kind that would initiate the instrumentation and a later trap that would disable it and log it somewhere. The scope of this code is small and doesn't call any library routines between the points I want to instrument. Now any kind of task switch or other interrupt in the duration of the interval will skew the results but it's not particularly likely in this case and I could just as easily invoke Disable() and Forbid() outside the region I'm instrumenting.

This sounds doable. Remind me after 4.10.1 is out.

[Karlos]

Originally Posted by Toni Wilen

This sounds doable. Remind me after 4.10.1 is out.

Sounds good. I'll try to remember!