This replaces the code that scanned a script block for suspicious strings.
The previous implementation:
* Tokenized input (generating many strings for garbage collection)
* Used multiple threads
This approach is based on Rubin-Karp and does not allocate any memory
other than a small array to hold the running hash values.
I tested the new and old approaches on 2200 files in the PowerShell repo.
The old code ran in about 1.8-2.1s (ignoring time spent reading files)
The new code runs in about 0.6s and is more stable due to no garbage.
When dotting a script cmdlet, the locals of automatic variables from the `PSScriptCmdlet` is not set up in the current scope before parameter binding. The fix is to push the locals in `CommandProcessor.OnSetCurrentScope` and pop them in `CommandProcessor.OnRestorePreviousScope`, which will be called from `SetCurrentScopeToExecutionScope` and `RestorePreviousScope` respectively.
Summary of changes:
1. When a new local scope is used, currently we set the locals for `CommandProcessor` right before parameter binding (in `BindCommandLineParametersNoValidation`); we set the locals for `ScriptCommandProcessor` in `Prepare`. I moved both to the constructor, right after the new scope is created so that the code is more consistent.
2. In `CmdletParameterBinderController.cs`, we set up the `PSBoundParameters` and `MyInvocation` variables in `HandleCommandLineDynamicParameters` again, which I think is unnecessary because this method is only called from `BindCommandLineParametersNoValidation`, where the setup is done for the first time.
3. Currently, the locals will be set for dotted script cmdlet in `EnterScope()` and `ExitScope()`. Now, that logic is moved to `OnSetCurrentScope` and `OnRestorePreviousScope` of `CommandProcessor`. This not only makes sure that locals are set before parameter binding, but also is consistent with the `ScriptCommandProcessor`.
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.
Our private implementation of FailFast was likely introduced
when each individual application needed it's own code to
get Watson reports. The CLR takes care of this for us now,
so we don't need our implementation.
Our method CheckForSevereException was also introduced in
the early days of the CLR - the exceptions it was checking
for aren't actually raised the CLR anymore, they just FailFast.
I removed them as there is a tiny bit of overhead (in code size)
and also in the generated code, e.g. dynamic sites called the method.
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.
Jason Shirk