[GR-43934] Only build full FrameTree when OmitInlinedMethodDebugLineInfo is false.

[graalvmbot]

As a default using -g will create DebugLocationInfo only for the body of compiled method but not for any method inlined into the compiled method. Despite that we always produced a full FrameTree even though only the first two levels are only every visited in this mode. This is a huge waste of time in the debug info generation part of the image built time.

This PR ensures that only when -H:-OmitInlinedMethodDebugLineInfo is used a full FrameTree gets created.

[olpaw]

If DebugCodeInfoMaxDepth or DebugCodeInfoUseSourceMappings are set explicitly by the user (-H:...) they override any adjustment of the settings performed for OmitInlinedMethodDebugLineInfo.

[olpaw]

@adinn ideally I would want to switch to OmitInlinedMethodDebugLineInfo to be disabled per default. But to get there I need more testing and maybe a bit of optimizations in com.oracle.svm.core.code.CompilationResultFrameTree.Builder#build. This will likely require a follow up PR (that will not be backported to 23.0).

[olpaw]

I tested with the example that caused most debuginfo generation time increase. And it turns out now with SourceMappings not being used anymore we can have -H:-OmitInlinedMethodDebugLineInfo as default and we are still much faster than before:

dacapo-9-12-mr1-git+2baec49-fop-pgo-ee -H:-UseOldDebugInfo

• With -H:-OmitInlinedMethodDebugLineInfo

60.49MB (38.51%) for debug info generated in 7.4s

vs. master

213.61MB (68.86%) for debug info generated in 41.0s

• Without -H:-OmitInlinedMethodDebugLineInfo

  41.42MB (30.01%) for debug info generated in 6.2s

vs. master

  54.83MB (36.21%) for debug info generated in 28.1s

[adinn]

Thanks, Paul. This looks good modulo one opportunity to remove a small bit of technical debt.

[adinn]

Local var fileIndex is declared a few lines up as an Integer, a legacy of an earlier version where the index for a given source dir was stored in a per-CU (i.e. per ClassEntry) map. This no longer applies with the latest code where the file index is defined globally across all classes, hence stored as an int in the FileEntry. It would be good to take the opportunity to clean this up by declaring fileIndex as an int.