Roslyn crashes when reporting an error with an "unusual" compilation unit name

[jskeet]

Version Used: .NET 9.0.100

Steps to Reproduce:

1. Create a new console application (dotnet new console)
2. Edit Program.cs to:

#line 100 ":invalid:"
int x = 5.0;

3. Run dotnet build

Expected Behavior:

Normal error messages without a crash.

Actual Behavior:

C:\Program Files\dotnet\sdk\9.0.100\Roslyn\Microsoft.CSharp.Core.targets(89,5): error MSB6006: "csc.dll" exited with code 1.

Notes:

• Visual Studio indicates csc.dll exiting with code 1 as well
• This is not reproducible in SharpLab - I suspect it's injecting its own #line directives
• Changing ":invalid:" to "valid" produces normal error messages.
• Names of ".", ".." and "..." produce some interesting results
• Changing the code to only produce a warning (e.g. just int x = 5; which results in CS0219) when using ":invalid:" is weird - the build succeeds without issuing a warning at all

[jaredpar]

Normal error messages without a crash.

The compiler is not crashing here, it is exiting normally and displaying the correct diagnostics. Can see this by looking at the binary log of this build

@rainersigwald, @baronfel this appears to be a behavior of msbuild. When this code is executed by the compiler it will produce the following diagnostics:

C:\Users\jaredpar\temp\console\:invalid:(100,9): error CS0266: Cannot implicitly convert type 'double' to 'int'. An explicit conversion exists (are you missing a cast?)
C:\Users\jaredpar\temp\console\:invalid:(100,5): warning CS0219: The variable 'x' is assigned but its value is never used

The compiler produces these and exits normally but the result of build is as displayed in the bug report

error MSB6006: "csc.dll" exited with code 1.

I suspect this is because the problem matcher code in MSBuild does not recognize the output as a compiler diagnostic. Suspect that the inclusion of : in the file path name is breaking the matching code. As a result msbuild doesn't detect any diagnostics and falls back to the generic error message.

This is overall a bit of a gray area. What is legal / illegal differs based on the file system. Even the set returned from Path.GetInvalidFileNameChars is not guaranteed to be complete so the compiler / msbuild couldn't fully account for bad characters. In similar cases we've pushed the burden onto customers to only use names / patterns that are valid for the file system they're working on.

[jskeet]

Apologies for misdiagnosing this as a crash - I assume the "exited with code 1" output is basically suppressed (by both VS and dotnet) when that "host" has already reported errors?

Thanks for looking into it - and to be clear, I don't personally think this is worth spending much time on. It's a corner case of "people who have too much time on their hands writing deliberately weird code". I just wanted to report it while it looked like a crash.

[jaredpar]

Apologies for misdiagnosing this as a crash

It fooled me too for the same reason. I threw it up under complog replay to see what the crash was and got the diagnostic output instead.

I just wanted to report it while it looked like a crash.

Definitely appreciate crash reports, even on obscure cases. 😄

[rainersigwald]

I suspect this is because the problem matcher code in MSBuild does not recognize the output as a compiler diagnostic.

Exactly! Contrast http://www.regexstorm.net/tester?p=%5e%5cs*%28%28%28%3f%3cORIGIN%3e%28%28%28%5cd%2b%3e%29%3f%5ba-zA-Z%5d%3f%3a%5b%5e%3a%5d*%29%7c%28%5b%5e%3a%5d*%29%29%29%3a%29%7c%28%29%29%28%3f%3cSUBCATEGORY%3e%28%28%29%7c%28%5b%5e%3a%5d*%3f+%29%29%29%28%3f%3cCATEGORY%3e%28error%7cwarning%29%29%28+%5cs*%28%3f%3cCODE%3e%5b%5e%3a+%5d*%29%29%3f%5cs*%3a%28%3f%3cTEXT%3e.*%29%24&i=C%3a%5cUsers%5cjaredpar%5ctemp%5cconsole%5c%3ainvalid%3a%28100%2c9%29%3a+error+CS0266%3a+Cannot+implicitly+convert+type+%27double%27+to+%27int%27.+An+explicit+conversion+exists+%28are+you+missing+a+cast%3f%29%0d%0a

with

http://www.regexstorm.net/tester?p=%5e%5cs*%28%28%28%3f%3cORIGIN%3e%28%28%28%5cd%2b%3e%29%3f%5ba-zA-Z%5d%3f%3a%5b%5e%3a%5d*%29%7c%28%5b%5e%3a%5d*%29%29%29%3a%29%7c%28%29%29%28%3f%3cSUBCATEGORY%3e%28%28%29%7c%28%5b%5e%3a%5d*%3f+%29%29%29%28%3f%3cCATEGORY%3e%28error%7cwarning%29%29%28+%5cs*%28%3f%3cCODE%3e%5b%5e%3a+%5d*%29%29%3f%5cs*%3a%28%3f%3cTEXT%3e.*%29%24&i=C%3a%5cUsers%5cjaredpar%5ctemp%5cconsole%5casdf.cs%28100%2c9%29%3a+error+CS0266%3a+Cannot+implicitly+convert+type+%27double%27+to+%27int%27.+An+explicit+conversion+exists+%28are+you+missing+a+cast%3f%29%0d%0a

I'm willing to consider replacing some of those [^:] around here with a more robust system but we'd have to consider backtracking cost.

I think that's probably more reasonable than changing Roslyn to report these differently as of now. @jaredpar want to move this to MSBuild?