Profiling WinUAE with Visual Studio 2013

[mark_k]

I've been playing/experimenting with the profiling facility in Visual Studio 2013.

In my Windows 7 VirtualBox VM I couldn't get the sampling method to work; no samples are captured. Am I doing something wrong or is VirtualBox incompatible with Visual Studio sampling? [With sampling, Visual Studio would periodically check which function is being run. Over time it builds up a statistical picture of the functions which use most CPU time.]

The instrumentation profiling method did work. With that method, a special version of the executable is built, which allows all function calls (and the time taken for each) to be recorded and analysed. However that logging adds a significant amount of overhead to the program.

On starting the instrumented winuae.exe, it takes ages (like about 10 minutes) until the settings window appears. While that isn't a big issue for the normal winuae.exe, maybe this has shown up something which could improve startup time slightly.

The profiling results showed a huge number of calls to free() in this code path:
default_prefs() ⇒ inputdevice_default_prefs() ⇒ inputdevice_default_kb_all() ⇒ set_kbr_default() ⇒ clear_id() ⇒ xfree()

Take a look at clear_id():

static void clear_id (struct uae_input_device *id)
{
#ifndef	_DEBUG
	int i, j;
	for (i = 0; i < MAX_INPUT_DEVICE_EVENTS; i++) {
		for (j = 0; j < MAX_INPUT_SUB_EVENT_ALL; j++)
			xfree (id->custom[i][j]);
	}
#endif
	TCHAR *cn = id->configname;
	TCHAR *n = id->name;
	memset (id, 0, sizeof (struct uae_input_device));
	id->configname = cn;
	id->name = n;
}

To reduce the number of function calls you could change it to only call xfree() when the argument is non-zero. But look at the loop in set_kbr_default() which calls clear_id():

	for (i = 0; i < MAX_INPUT_DEVICE_EVENTS; i++) {
		clear_id (kbr);
		kbr->extra[i] = -1;
	}

and compare with the for loop in clear_id() itself. Are you doing a lot of unnecessary free()ing with the two nested for (i = 0; i < MAX_INPUT_DEVICE_EVENTS; i++)???

[Toni Wilen]

AFAIK there is option to disable part of c-library allocation function tracing which is the horribly slow part.

I am not going to add extra "if (address) free(address)" because first function of free() is NULL check!

[mark_k]

clear_id() basically does this:
- free all memory entries in the custom[][] array
- zero out the entire uae_input_device structure, except for configname and name fields.

In set_kbr_default() there is this though:

for (i = 0; i < MAX_INPUT_DEVICE_EVENTS; i++) {
	clear_id (kbr);
	kbr->extra[i] = -1;
}

So for each of the 256 loop iterations (MAX_INPUT_DEVICE_EVENTS is 256) you call clear_id() then set the extra[i] field to -1. But on the next loop iteration, because clear_id() zeros out the structure, the previous extra[i] field is also zeroed. And all the custom[][] memory regions will have been freed on the first loop iteration.

So, would it make sense to move clear_id(kbr) out of the loop, doing this instead:

clear_id (kbr);
for (i = 0; i < MAX_INPUT_DEVICE_EVENTS; i++) {
	kbr->extra[i] = -1;
}

Update: I built WinUAE with that change, and now with instrumented profiling the settings window opens in about 6-7 seconds (on my 6-year-old laptop).

[mark_k]

I found a blog post from the Visual Studio Profile Team which mentions using profiling in virtual environments. Looks like Visual Studio specifically supports VMWare, Hyper-V and Virtual PC. So it is possible to use the sampling profiling method, just not in VirtualBox.

So was the issue I noted in set_kbr_default() a bug?