i#5411 bbdup mem: Shrink per-thread heap usage

[derekbruening]

Moves all drbbdup per-thread memory to unreachable heap.

Moves the drbbdup dynamic case hit table to be dynamically allocated
based on a new drbbdup option "never_enable_dynamic_handling" not
being set. This eliminates 128K per thread for drmemtrace, which
makes a big difference on large applications.

Sanity test:
Before:

  $ bin64/drrun -rstats_to_stderr -t drcachesim -offline -max_global_trace_refs 10K -- suite/tests/bin/client.annotation-concurrency libclient.annotation-concurrency.appdll.so A 4 64 3
              Peak threads under DynamoRIO control :                 5
              Peak vmm blocks for unreachable heap :               767
                Peak vmm blocks for reachable heap :               726
            Peak vmm virtual memory in use (bytes) :          10293248

After:

              Peak threads under DynamoRIO control :                 5
              Peak vmm blocks for unreachable heap :               736
                Peak vmm blocks for reachable heap :               250
            Peak vmm virtual memory in use (bytes) :           8183808

Issue: #5411

[johnfxgalea]

Can you sprinkle some asserts in the code-base where opts.never_enable_dynamic_handling should be false. Maybe somewhere close to the beginning of the following functions:

• init_fp_cache()
• drbbdup_handle_new_case()
• drbbdup_insert_dynamic_handling()

[derekbruening]

Can you sprinkle some asserts in the code-base where opts.never_enable_dynamic_handling should be false. Maybe somewhere close to the beginning of the following functions:

• init_fp_cache()
• drbbdup_handle_new_case()
• drbbdup_insert_dynamic_handling()

Done.

[derekbruening]

win64 failure is burst_traceopts/replaceall hitting #4486