FileMapInfo::patch_archived_heap_embedded_pointers() may be called during VM bootstrap. This happens, for example, if the heap size has changed significantly between CDS dump time and run time.

http://hg.openjdk.java.net/jdk/jdk/file/5ac19bd3a1e2/src/hotspot/share/memory/filemap.cpp#l1885

The default CDS archive is created with -Xmx128m to optimize for apps with small heaps (e.g., those used in cloud). However, if we run with a bigger heap (somewhat larger than 2GB), we're likely to use 0-based 3-bit shift oop compression, and thus the archived heap must be patched.

Because the patching is fairly independent, and the patched contents are not needed until the VM loads the first class, it should be safe to do the patching in a GC worker thread while the main VM thread continues with other initialization (such as interpreter generation).

============== about 2.64% degradation or 1.09ms

# No patching (dumptime heap = runtime heap)
$ java -Xshare:dump -Xmx4g
$ perf stat -r 100 java -Xmx4g -version

 Performance counter stats for '/jdk/bld/bench/eva/loader_con0/bin/java -Xmx4g -version' (100 runs):

             56.11 msec task-clock                #    1.360 CPUs utilized            ( +-  0.60% )
               199      context-switches          #    0.004 M/sec                    ( +-  1.04% )
                12      cpu-migrations            #    0.206 K/sec                    ( +-  2.08% )
             2,937      page-faults               #    0.052 M/sec                    ( +-  0.04% )
       137,492,599      cycles                    #    2.451 GHz                      ( +-  0.15% )
        85,290,084      stalled-cycles-frontend   #   62.03% frontend cycles idle     ( +-  0.18% )
        66,524,836      stalled-cycles-backend    #   48.38% backend cycles idle      ( +-  0.21% )
       105,195,769      instructions              #    0.77  insn per cycle         
                                                  #    0.81  stalled cycles per insn  ( +-  0.14% )
        20,009,439      branches                  #  356.625 M/sec                    ( +-  0.15% )
         1,005,784      branch-misses             #    5.03% of all branches          ( +-  0.19% )

          0.041249 +- 0.000282 seconds time elapsed  ( +-  0.68% )

# With patching (dumptime heap != runtime heap)
$ java -Xshare:dump -Xmx128m
$ perf stat -r 100 java -Xmx4g -version

 Performance counter stats for '/jdk/bld/bench/eva/loader_con0/bin/java -Xmx4g -version' (100 runs):

             56.96 msec task-clock                #    1.345 CPUs utilized            ( +-  0.55% )
               202      context-switches          #    0.004 M/sec                    ( +-  0.97% )
                12      cpu-migrations            #    0.212 K/sec                    ( +-  1.90% )
             3,118      page-faults               #    0.055 M/sec                    ( +-  0.04% )
       139,782,303      cycles                    #    2.454 GHz                      ( +-  0.14% )
        86,193,560      stalled-cycles-frontend   #   61.66% frontend cycles idle     ( +-  0.15% )
        67,261,311      stalled-cycles-backend    #   48.12% backend cycles idle      ( +-  0.17% )
       107,975,073      instructions              #    0.77  insn per cycle         
                                                  #    0.80  stalled cycles per insn  ( +-  0.14% )
        20,706,343      branches                  #  363.509 M/sec                    ( +-  0.15% )
         1,023,199      branch-misses             #    4.94% of all branches          ( +-  0.19% )

          0.042338 +- 0.000274 seconds time elapsed  ( +-  0.65% )