This change moves powershell to .NET Core 2.0. Major changes are:
1. PowerShell assemblies are now targeting `netcoreapp2.0`. We are using `microsoft.netcore.app-2.0.0-preview1-001913-00`, which is from dotnet-core build 4/4/17. We cannot target `netstandard2.0` because the packages `System.Reflection.Emit` and `System.Reflection.Emit.Lightweight`, which are needed for powershell class, cannot be referenced when targeting `netstandard2.0`.
2. Refactor code to remove most CLR stub types and extension types.
3. Update build scripts to enable CI builds. The `-cache` section is specified to depend on `appveyor.yml`, so the cache will be invalidated if `appveyor.yml` is changed.
4. Ship `netcoreapp` reference assemblies with powershell to fix the issues in `Add-Type` (#2764). By default `Add-Type` will reference all those reference assemblies when compiling C# code. If `-ReferenceAssembly` is specified, then we search reference assemblies first, then the framework runtime assemblies, and lastly the loaded assemblies (possibly a third-party one that was already loaded).
5. `dotnet publish` generates executable on Unix platforms, but doesn't set "x" permission and thus it cannot execute. Currently, the "x" permission is set in the build script, `dotnet/cli` issue [#6286](https://github.com/dotnet/cli/issues/6286) is tracking this.
6. Replace the use of some APIs with the ones that take `SecureString`.
7. osx.10.12 is required to update to `netcoreapp2.0` because `dotnet-cli` 2.0.0-preview only works on osx.10.12.
8. Add dependency to `System.ValueTuple` to work around a ambiguous type identity issue in coreclr. The issue is tracked by `dotnet/corefx` [#17797](https://github.com/dotnet/corefx/issues/17797). When moving to newer version of `netcoreapp2.0`, we need to verify if this dependency is still needed.
In rare cases, it was possible to cause high contention on a lock used
while compiling code to run in the interpreter.
This fixes the problem in a couple different ways:
* Use ConditionalWeakTable which has finer granularity in locking
* Avoid dictionary lookups in the most common cases
- There really aren't too many real cases, I could have covered them
all, but keeping the code generic is useful for the future.
Not quite related, but I noticed we didn't cache certain interpreter
instructions and it made sense to add a cache because we generate a lot
of array-init instructions.
Mostly using R# to automatically refactor, with some manual
fixups where it wouldn't work automatically (e.g. xml comments
on fields) or some it seemed to miss.
Some minor code reformatting was also done on properties near
other stuff I was manually inspecting.
I (Jason Shirk) ran https://github.com/dotnet/codeformatter with the default rules, basically:
codeformatter /nocopyright "/c:DEBUG,UNIX,CORECLR" @files.rsp
This caused a few problems building, which were fixed up manually.
Notable changes:
`this.` is removed unless needed to disambiguate.
private instance fields are renamed to have a `_` prefix.
private static fields are renamed to have a `s_` prefix.
I left some projects alone (like PackageManagement) and also left some generated code alone.
Ilya