A tools thread is one that can access core runtime structures but not touch anything managed (code or heap).
It's intended to enable the profiler to resolve missing backtraces from its helper thread.
THIS IS A HACK.
The right solution would be to not perform this on the profiler at all and simply log IPs and let the decoder
resolve them. But then there's reality and we can't simply break the current format without consequences.
We're taking this hack for now until we can fix the profiler to do something sensible.
As mentioned in commit 789517f8, I'm trying to compare inter-API level
differences in Mono.Android.dll, and want to reduce as much
"meaningless noise" as possible -- changes which don't meaningfully
break API or ABI.
Among those "meaningless" changes are the addition of property setters
to properties. Adding a setter doesn't break API or ABI.
Add a `mono-api-html --ignore-changes-property-setters` option.
--ignore-changes-property-setters:
Ignore the addition of property setters.
If a source property contains a getter and the target property
contains both a getter and a setter, do not flag this as a
change.
As mentioned in commit 789517f8, the intentino of
`mono-api-html --ignore-changes-virtual` is so that when a method is
changed from non-virtual to virtual or override, or from virtual to
override, a "Removed member" message isn't triggered.
However, the previous implementation was too lenient. While *adding*
virtual to the method is fine, *removing* is NOT fine, as that would
be an API breaking change (e.g. all classes which attempt to override
a previously virtual method could no longer do so).
Make the virtual check stricter so that virtual/override to
non-virtual changes are not ignored.
I've been trying to review API diffs in Xamarin.Android between API
levels, but the diffs are "noisy" because Google will change parameter
names at-will, or will change a non-virtual method into a virtual
method, or a virtual method into an override, all of which are
acceptable changes but individually generate a Removed line in
mono-api-html output, and manually removing the *hundreds* of changes
is error-prone.
Add two new option flags to mono-api-html:
--ignore-changes-parameter-names:
Ignore changes in constructor and method parameter names.
If parameter names change between <reference.xml> and
<assembly.xml>, don't report a change for the constructor or
method.
--ignore-changes-virtual:
Ignore changes in "virtual-ness".
If a method or property changes from non-virtual to virtual,
non-virtual to override, or virtual to override, don't report a
change for the member.
This is a jumbo commit that addresses multiple issues with statistical backtrace collection:
1) It's not possible to reliably collect MonoMethod* so we check that they exist before
serializing to the .mlpd file and retrieve in sync context.
2) It was reporting the generic instance MonoMethod* under gsharing. Those methods are
not reported to the profiler since they are never compiled. They would turn out as unknowns.
3) Collecting IL offsets from async context is currently impossible to be done safely, so it's
now disabled. The way we'll do it in the future is by serializing the sequence point table when
JIT'ing.
4) The way we used to collect frames in statistical ignored all frames until the first LMF, which
discards a lot of them under statistical mode. We now properly pass sigcontext and have it been consumed.
In the future we should simply use the callchain mode.
The current signal profiling code depended on invalid linux behavior of setitimer
allowing multiple threads to register for the same timer. Linux kernel 2.6.12 fixed
this and only the main thread gets the signal. Always.
The correct POSIX behavior happens on OSX/iOS.
On Linux we could use timer_create and SIGEV_THREAD_ID but that would not work on iOS/OSX
since it's specific to it. Given that little detail, I'll leave implementing it as an
exercise to the reader.
The solution, which aims to be as portable as possible, is to use a pair of signals and
have the first one fire the second one against all other threads.
We use a pair of signals since it's not well defined which thread will receive the signal
so detecting this from the signal is non-trivial.