556181 - Make math-partial sums not waste time with GETXPROP

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[David Mandelin [:dmandelin]]

This is a placeholder bug so I can externalize the need to make sure this gets fixed eventually:

SunSpider's math-partial-sums inadvertently mostly uses global variables in its hot loop. There are lots of statements like |a2 += e| where |a2| is a free variable (and is a global at run time). SM implements this more-or-less like this:

  bindname a2
  dup
  getxprop a2
  [[code for expr e]]
  setname a2
  pop

This gives pretty bad perf and makes life hard on jits generally. It can probably be fixed as part of a general overhaul of free variables and globals. We should be able to prove that a2 must be a global and emit code more like this:

  getglobal a2
  [[code for expr e]]
  setglobal a2
  pop

But this recommendation might not be good. The key point is we have to make sure we get this to run fast one way or another.

Comment 1

[Brendan Eich [:brendan]]

See bug 321757 for the genesis of GETXPROP -- it's required that this throw:

foo += 42;

where foo is not yet defined (by var or by an assignment).

/be

Comment 2

[Brendan Eich [:brendan]]

Of course the old interpreter's bytecode sequence is suboptimal -- don't let me dissuade you from going much faster in JM.

One might hope for a SS fix to declare 9 out of 10 local variables correctly :-P.

/be

Comment 3

[Andreas Gal :gal]

The tracer turns these into pretty efficient code. JM should be able to use a PIC to a similar effect.

Comment 4

[David Mandelin [:dmandelin]]

Yes, you are both right that we might be able to fix this up just fine without modifying the bytecode at all. I just needed a reminder to make sure we do fix it up in JM eventually. :-)

Comment 5

[David Anderson [:dvander] - inactive, e-mail if emergency]

This went away with all of JM's changes to globals (bug 561923 and follow-ups).