This is RecordSize plugin for GNU C++ compiler

Its aim is to detect cases where we can save some space by reorganizing fields
within record.

Usage:

g++ -fplugin=recordsize.so some_cpp_source

By default plugin will emit a warning for records for which its minimal size
estimation is less then actual record size. Template instantiations are ignored.

You can tweak plugin behavior using command line switches:

-fplugin-arg-recordsize-process-templates - enables template instantiations
  processing.

-fplugin-arg-recordsize-print-layout - print record layout. That allows to see
  which sizes/offsets/alignments are used by each field of reported records.

-fplugin-arg-recordsize-print-all - print record layout for all (even not
  missized) records we process.

-fplugin-arg-recordsize-dumpfile=filename - dump record information to given
  file. Dump file is appended with new record information on each next GCC
  invocation. This especially useful if you modify CXXFLAGS for some large
  project. After build is finished you can use separate tool rs-report and build
  report for whole project.
  WARNING: There is no source file updates tracking, so you have to remove dump
  after you've modified sources. Otherwise record data won't be updated and
  you'll see outdated output!

Report tool usage:

rs-report dumpfile [skip=skipspec] [sort=sortspec]

By default rs-report will dump all records with layout in some random order.
You can omit printing some type of records by adding 'skip' argument.
Following letters are accepted in 'skipspec':
  e - skip empty records (with bases/fields)
  g - skip 'good' records. Record is 'good' if its actual size is less or equal
      to estimated minimal size. Not-estimated records are also 'good'.
  h - skip handled records (see caveats to know what records aren't handled)
  t - skip template instantiations.

You can force specific sorting by adding 'sort' argument.
Following letters are accepted in 'sortspec':
  d - sort by difference between actual size and estimated minimal size (most
      oversized records first). 'Good' records are left as is.
  s - sort by record actual size.
  n - sort by record name.

Examples:

rs-report dumpfile skip=g sort=d
  Print only oversized records with most oversized first.

rs-report dumpfile skip=et sort=ns
  Print only non-template and non-empty records ordered by name and size.


Caveats:

Not all records are estimated. Currently we ignore records with:
 - bit-fields. Optimal bit-field layout is NP-hard problem (bin packing problem)
 - virtual bases. Vptr and virtual base are placed at the end of derived class,
   so this case is tricky.

Minimal size estimation doesn't try to reorder bases in multiple inheritance case.
It doesn't try to fit any member into padding between bases and first field.