Make include scanning faster when a target has many include scannables.

Instead of repeatedly scanning all transitively required include scannables, do
one pass from all include scannables; this prevents us revisiting the common
transitive closure of the include scannables.

Additionally, only scan command line includes relatively to a main source file,
not relatively to all source files. This is better than what we had before, but
it's still not exactly right - we should actually scan the command line
includes relatively to the module map file ([]

Brings include scanning times for large TUs with modules down from ~60 seconds
to < 2 seconds.

--
MOS_MIGRATED_REVID=88963159

1 files changed, 7 insertions, 0 deletions

diff --git a/src/main/java/com/google/devtools/build/lib/rules/cpp/IncludeScannable.java b/src/main/java/com/google/devtools/build/lib/rules/cpp/IncludeScannable.java
index 9c70090ac0..d1f89f9657 100644
--- a/src/main/java/com/google/devtools/build/lib/rules/cpp/IncludeScannable.java
+++ b/src/main/java/com/google/devtools/build/lib/rules/cpp/IncludeScannable.java
@@ -68,8 +68,15 @@ public interface IncludeScannable {
   List<String> getCmdlineIncludes();
 
   /**
+   * Returns the artifact relative to which the {@code getCmdlineIncludes()} should be interpreted. 
+   */
+  Artifact getMainIncludeScannerSource();
+  
+  /**
    * Returns an immutable list of sources that the IncludeScanner should scan
    * for this action.
+   * 
+   * <p>Must contain {@code getMainIncludeScannerSource()}.
    */
   Collection<Artifact> getIncludeScannerSources();