TODO.txt

 * flesh out dist bundle readme
 * add high-level "interact" method for fun nodejs demo

* fixes from after 2.5.0 merge:
   * clean up thread-local-ref handling in jit backend
   * AbstractFailDescr _attrs_
      * ISTM that the x86 backend was using _x86_adr_jump_offset,
        and is now using a declared adr_jump_offset
      * we should use rd_locs for faillocs, adr_jump_offset for block ref,
        then store the extra descr details in a hash on the block.
   * leaked allocations in jit backend tests
      * we may want to avoid mallocing in the jit backend?
   * failing tests:
      * test_runner -> many
      * translation tests require new errno-saving logic
   * lots of new unaligned stores, perhaps some new memory-zeroing thing?
       * OP_RAW_MEMCLEAR ?
       * RPyField(...) = 0L; ?
   * make release-debug wants to call gcc for some strange reason...

developer niceties:
  * high-level intro to the various modules, where they live
    and what they do


SOON:

 * get all the testcases passing again

 * name the memory file after a hash of its contents, to avoid
   weirdness from caching an incorrect version of the file

 * async script loading, so it compiles off main thread


improve the jitted code:
 * generate switch or if or something else depending on the conditions
   that we're actually processing, rather than always doing a switch.
 * avoid lots of pointless swapping of variables
 * tighten up code around func calls
    * don't spill things to frame if they're already there
 * remove unneeded conversions to unsigned e.g. in write barrier?
 * reduce number of comparisons in write barrier
 * somehow reduce overhead of dynCall_XXX()?
   * maybe send relevant funcs into the jitted code via plt?

* almost all JIT overhead is spent in support.jitRecompile() - up to several seconds in many cases!
   * this is dominated by a handful of large functions that generate
     "compiled slowly" warnings under odinmonkey.
   * we need the equivalent of the outliner:
      * separately-compiled helper functions, e.g. guard failure helpers
      * split inner loops into their own functions


EVENTUALLY:

* pypy thinks we have st_atim, when we don't

* sizeof off_t is not longlong

* use emscripten_get_now() for performance timing?

* use emscripten_sleep() for sleeping

* have to figure out the "stack check" thing and what I need to do to
  make it work in the jit, and in js in general (I think this is the thing
  that raises a nice python-level error when you recurse too much)

* using closure on the compiled shell