Race condition bug in Formatter when requesting formatted changes for a sub-span in a file

[mavasani]

As part of testing the changes in #67818, I have identified a race condition bug in the Formatter wherein it intermittently reports (incorrect) non-zero formatting changes when asked for formatting changes for a sub-span but reports zero formatting changes when asked for full file span.

Test code where I get wrong IDE0055 violations from lightbulb (span-based) invocation:

Below is the local (temporary) product change in FormattingAnalyzerHelper.cs to handle optional FilterSpan from analysis context. For the above case, I am hitting the breakpoint at line formattingBug = true, indicating the formatter race condition bug I explained above:

            var node = context.FilterSpan.HasValue ? root.FindNode(context.FilterSpan.GetValueOrDefault()) : root;
            var formattingChanges = Formatter.GetFormattedTextChanges(node, formattingProvider, options, cancellationToken);

            var formattingBug = false;
            if (formattingChanges.Count > 0 && context.FilterSpan.HasValue)
            {
                var rootChanges = Formatter.GetFormattedTextChanges(root, formattingProvider, options, cancellationToken);
                if (rootChanges.Count == 0)
                    formattingBug = true;
            }

[mavasani]

Turns out this is not a Formatter bug, just that I was using the wrong Formatter API for get formatted text changes.

// Wrong API
var node = context.FilterSpan.HasValue ? root.FindNode(context.FilterSpan.GetValueOrDefault()) : root;
var formattingChanges = Formatter.GetFormattedTextChanges(node, formattingProvider, options, cancellationToken);

// Correct API
var span = context.FilterSpan.HasValue ? context.FilterSpan.GetValueOrDefault() : root.FullSpan;
var spans = SpecializedCollections.SingletonEnumerable(span);
var formattingChanges = Formatter.GetFormattedTextChanges(root, spans, formattingProvider, options, cancellationToken);

This was showing up as a race condition because sometimes the background analysis would compute and cache full document diagnostics before the lightbulb span based diagnostics are requested. After making the above changes, I no longer see this issue.