We want to avoid separate `mscorlib.dll` assemblies for 32/64 bits so
the architecture specific code for `n[u]int` and `nfloat` must be
preserved in both cases.
Because a single, slightly larger, assembly is much smaller than two
(slightly smaller) ones.
This is not a common situation since the extraneous preserved API are
often used in the application (or 3rd party code) so, in most cases,
a single `mscorlib.dll` was already used.
This merely close the gap for some cases, like our `link all` application
where this happened.
* [mtouch] Rename temporary directories so the order assemblies move is clearer.
I implemented this myself, but I can never remember in which order assemblies
go from one directory to another during the build.
So number these temporary directories, so that even the most forgetful minds
can understand without having to remember anything.
* [tests] Update test according to temporary directory name change.
* [tests] Improve debug spew for the RebuildTest_WithExtensions test.
* [mtouch/mmp] Store/load if the dynamic registrar is removed or not into the cached link results.
Store/load if the dynamic registrar is removed or not into the cached link
results, so that we generate the correct main.m even if cached linker results
are used.
* [mtouch/mmp] The static registrar must not execute if we're loading cached results from the linker.
The static registrar must not execute if we're loading cached results from the
linker, because the static registrar needs information from the linker that's
not restored from the cache.
* [mtouch/mmp] Share Touch code.
* [mtouch/mmp] Make it possible to touch inexistent files (to create them).
* [mtouch/mmp] Fix tracking of whether the static registrar should run again or not.
The recent changes to support optimizing away the dynamic registrar caused the
Xamarin.MTouch.RebuildTest_WithExtensions test to regress.
The problem
-----------
* The linker now collects and stores information the static registrar needs.
* This information is not restored from disk when the linker realizes that it
can reload previously linked assemblies instead of executing again.
* The static registrar runs again (for another reason).
* The information the static registrar needs isn't available, and incorrect
output follows.
So fix 1: show an error if the static registrar runs when the linker loaded
cached results.
The exact scenario the test ran into is this:
* 1st build: everything is new and everything is built.
* 2nd build: contents of .exe changes, the linker runs again, the static
registrar runs again, but sees that the generated output didn't change, so
it doesn't write the new content to disk (this is an optimization to avoid
compiling the registrar.m file again unless needed).
* 3rd build: only the .exe timestamp changes, the linker sees nothing changes
in the contents of the .exe and loads the previously linked assemblies from
disk, the static registrar sees that the .exe's timestamp is newer than
registrar.m's timestamp and run again, but doesn't produce the right result
because it doesn't have the information it needs.
Considered solutions
--------------------
1. Only track timestamps, not file contents. This is not ideal, since it will
result in more work done: in particular for the case above, it would add a
registrar.m compilation in build #2, and linker rerun + static registrar
rerun + registrar.m compilation + final native link in build #3.
2. Always write the output of the static registrar, even if it hasn't changed.
This is not ideal either, since it will also result in more work done: for
the case above, it would add a registrar.m compilation + final native link
in build #3.
3. Always write the output of the static registrar, but track if it changed or
not, and if it didn't, just touch registrar.o instead of recompiling it.
This only means the final native link in build #3 is added (see #5 for why
this is worse than it sounds).
4. Always write the output of the static registrar, but track it it changed or
not, and if it didn't, just touch registrar.o instead of recompiling it,
and track that too, so that the final native link in build #3 isn't needed
anymore. Unfortunately this may result in incorrect behavior, because now
the msbuild tasks will detect that the executable has changed, and may run
dsymutil + strip again. The executable didn't actually change, which means
it would be the previously stripped executable, and thus we'd end up with
an empty .dSYM because we ran dsymtil on an already stripped executable.
5. Idea #4, but write the output of the final link into a temporary directory
instead of the .app, so that we could track whether we should update the
executable in the .app or not. This is not optimal either, because
executables can be *big* (I've seen multi-GB tvOS bitcode executables), and
extra copies of such files should not be taken lightly.
6. Idea #4, but tell the MSBuild tasks that dsymutil/strip doesn't need to be
rerun even if the timestamp of the executable changed. This might actually
work, but now the solution's become quite complex.
Implemented solution
--------------------
Use stamp files to detect whether a file is up-to-date or not.
In particular:
* When we don't write to a file because the new contents are identical to the
old contents, we now touch a .stamp file. This stamp file means "the
accompanying file was determined to be up-to-date when the stamp was
touched."
* When checking whether a file is up-to-date, also check for the presence of a
.stamp file, and if it exists, use the highest timestamp between the stamp
file and the actual file.
Now the test scenario becomes:
* 1st build: everything is new and everything is built.
* 2nd build: contents of .exe changes, the linker runs again, the static
registrar runs again, but sees that the generated output didn't change, so
it doesn't write the new content to disk, but it creates a registrar.m.stamp
file to indicate the point in time when registrar.m was considered up-to-
date.
* 3rd build: only the .exe timestamp changes, the linker sees nothing changes
in the contents of the .exe and loads the previously linked assemblies from
disk, the static registrar sees that the .exe's timestamp is *older* than
registrar.m.stamp's timestamp and doesn't run again.
We only use the stamp file for source code (registrar.[m|h], main.[m|h],
pinvokes.[m|h]), since using it every time has too much potential for running
into other problems (for instance we should never create .stamp files inside
the .app).
Fixes these test failures:
1) Failed : Xamarin.MTouch.RebuildTest_WithExtensions("single","",False,System.String[])
single
Expected: <empty>
But was: < "/Users/builder/data/lanes/5746/4123bf7e/source/xamarin-macios/tests/mtouch/bin/Debug/tmp-test-dir/Xamarin.Tests.BundlerTool.CreateTemporaryDirectory371/testApp.app/testApp is modified, timestamp: 2/15/2018 3:04:11 PM > 2/15/2018 3:04:09 PM" >
2) Failed : Xamarin.MTouch.RebuildTest_WithExtensions("dual","armv7,arm64",False,System.String[])
dual
Expected: <empty>
But was: < "/Users/builder/data/lanes/5746/4123bf7e/source/xamarin-macios/tests/mtouch/bin/Debug/tmp-test-dir/Xamarin.Tests.BundlerTool.CreateTemporaryDirectory375/testApp.app/testApp is modified, timestamp: 2/15/2018 3:06:03 PM > 2/15/2018 3:06:00 PM" >
3) Failed : Xamarin.MTouch.RebuildTest_WithExtensions("llvm","armv7+llvm",False,System.String[])
llvm
Expected: <empty>
But was: < "/Users/builder/data/lanes/5746/4123bf7e/source/xamarin-macios/tests/mtouch/bin/Debug/tmp-test-dir/Xamarin.Tests.BundlerTool.CreateTemporaryDirectory379/testApp.app/testApp is modified, timestamp: 2/15/2018 3:07:14 PM > 2/15/2018 3:07:12 PM" >
4) Failed : Xamarin.MTouch.RebuildTest_WithExtensions("debug","",True,System.String[])
debug
Expected: <empty>
But was: < "/Users/builder/data/lanes/5746/4123bf7e/source/xamarin-macios/tests/mtouch/bin/Debug/tmp-test-dir/Xamarin.Tests.BundlerTool.CreateTemporaryDirectory383/testApp.app/testApp is modified, timestamp: 2/15/2018 3:08:16 PM > 2/15/2018 3:08:13 PM" >
5) Failed : Xamarin.MTouch.RebuildTest_WithExtensions("single-framework","",False,System.String[])
single-framework
Expected: <empty>
But was: < "/Users/builder/data/lanes/5746/4123bf7e/source/xamarin-macios/tests/mtouch/bin/Debug/tmp-test-dir/Xamarin.Tests.BundlerTool.CreateTemporaryDirectory387/testApp.app/testApp is modified, timestamp: 2/15/2018 3:09:18 PM > 2/15/2018 3:09:16 PM" >
Fixes https://github.com/xamarin/maccore/issues/641
The linker might remove interfaces that have already been linked away. Make
sure to look for the TypeDefinition for such interfaces among the types that
have already been linked away.
Fixes https://github.com/xamarin/xamarin-macios/issues/3513.
It seems the d15-5 version of VSfM has a non-optimal startup path, where
launching vstool takes almost 4 minutes on the bots, making it difficult to
build & run tests within the allotted time for those tests:
$ time /Applications/Visual\ Studio.app/Contents/MacOS/vstool help
[...]
real 3m30.172s
So bump to the latest d15-6 preview (which takes _only_ 12 seconds to launch
on my machine), hoping it will be enough to make tests build on the bots.
Implement support for optimizing away the dynamic registrar.
* Add a new property (Runtime.DynamicRegistrationSupported) that indicates at
runtime whether dynamic registration is available.
* Check this new property whenever we need dynamic registration.
* Add an optimization to mtouch/mmp that detects whether dynamic registration
is required, and if not, changes Runtime.DynamicRegistrationSupported to
return false.
* Add an optimization to mtouch/mmp that inlines
Runtime.DynamicRegistrationSupported as a constant value.
The end result is that the linker will link away the dynamic registrar if
mtouch/mmp detects that it's not needed.
Benchmark
---------
I've compared the size of entire apps built for device:
|test | Before | After | Diff | % |
|:-----------------------------|-------:|-------:|-------:|------:|
|[monotouch-test/Release][1] | 95.7mb | 95.0mb | -680kb | -0.7% |
|[link sdk/Release][2] | 21.2mb | 20.9mb | -245kb | -1.2% |
|[minimalistic app/Release][3] | 4.58mb | 4.32mb | -259kb | -5.7% |
[1]: https://gist.github.com/rolfbjarne/3871e36d2de8db1a8eee1d9f9276d3d2#monotouch-test-release-wall-optimizations-enabled
[2]: https://gist.github.com/rolfbjarne/3871e36d2de8db1a8eee1d9f9276d3d2#link-sdk
[3]: https://gist.github.com/rolfbjarne/3871e36d2de8db1a8eee1d9f9276d3d2#minimalistic-app
Fixes this test:
Xamarin.ApiTest.ApiTest.AlwaysOptimizable(iOS) : All methods calling optimizable API must be optimizable
The method System.IntPtr ObjCRuntime.BlockLiteral::GetBlockForDelegate(System.Reflection.MethodInfo,System.Object,System.String) calls System.Boolean ObjCRuntime.Runtime::get_DynamicRegistrationSupported(), but it does not have a [BindingImpl (BindingImplOptions.Optimizable)] attribute.
Also skip the BlockLiteral.SetupBlock optimization for this method, because
it's not optimizable (and thus produces a warning users can't ignore).
It seems LLVM freaks out if it finds unreachable exception handlers:
* Assertion at ../../../../external/mono/mono/mini/mini-llvm.c:7406, condition `err == 0' not met
Basic Block in function 'Xamarin_iOS_ObjCRuntime_Class_GetClassHandle_System_Type' does not have terminator!
label %BB7_CALL_HANDLER_TARGET
Mark managed callback functions if they're only used in the dynamic registrar,
and only initialize the corresponding delegate if the dynamic registrar is
available.
This will make it possible for the linker to remove those managed functions if
the dynamic registrar is being removed.
Rolf Bjarne Kvinge
Manuel de la Pena
Sebastien Pouliot
Matt Sylvia
Alex Soto
Zoltan Varga
Vincent Dondain