525379 - nanojit: make the _nIns/_nExitIns swap hack more consistent

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Nicholas Nethercote [inactive]]

Attachment: WIP patch, only X64 backend done so far

We generate native code into two places:  the normal code area, and the
exit area.  We have a hack whereby we swap the _nIns and _nExitIns
pointers when we want to write to the exit area;  this means that the code
responsible for codegen need not be parameterised, it can just use _nIns.
But there are some other related values (codeStart/exitStart,
codeEnd/exitEnd, codeBytes/exitBytes) that we don't swap.  This leads to
odd-looking code like this:

  if (_inExit)
      codeAlloc(exitStart, exitEnd, _nIns verbose_only(, exitBytes)); 
  else
      codeAlloc(codeStart, codeEnd, _nIns verbose_only(, codeBytes));

You might expect to see _nExitIns in the first case rather than _nIns, but
_nIns is right because it's been swapped and so is really _nExitIns.

This patch changes things so that all the relevant values get swapped,
which avoids the need for _inExit tests.  The above code then becomes:

  codeAlloc(codeStart, codeEnd, _nIns verbose_only(, codeBytes));

It's mostly a correctness/simplicity patch, but it should also reduce the
number of instructions because there's one less test in the hot path in
underrunProtect().  I don't see any SS changes, but SS is being particularly
noisy for me today and yesterday.

Comment 1

[Nicholas Nethercote [inactive]]

Attachment: patch

All backends done, tested on x86 and x64.  I'll do the others once it's passed review.

Comment 2

[Nicholas Nethercote [inactive]]

Passed TR's try server:

http://hg.mozilla.org/projects/nanojit-central/rev/bad939491825

Comment 3

[Edwin Smith]

pushed to tamarin http://hg.mozilla.org/tamarin-redux/rev/76ccc148ba1d

Comment 4

[Nicholas Nethercote [inactive]]

http://hg.mozilla.org/tracemonkey/rev/74504312c860

Comment 5

[Nicholas Nethercote [inactive]]

http://hg.mozilla.org/mozilla-central/rev/74504312c860