584603 - JM: function-statement can disappear if it shadows a subsequent var declaration

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Jesse Ruderman]

var spots = 0;
(function() {
  for (var j = 0; j < 3; ++j) {
    if (j == 0) {
      function x(){}
    } else if (x) {
      ++spots;
    }
  }
})(); var x;
print(spots);

I put "var x" on the same line so the testcase shows the bug even when pasted into an interactive shell.

Result:
  ./js     2
  ./js -j  2
  ./js -m  1   <-- probably wrong
  ./jsc    2

Comment 1

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

http://hg.mozilla.org/users/danderson_mozilla.com/moo/rev/787e35063545

Comment 2

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

wrong bug, sorry

Comment 3

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

The problem is that we've predicted "x" in the "else if" to be a global variable, but it's really not. It's a potentially aliased in the Call object because of the named function definition snuggled in a block.

This is broken in our test suite already. I added an assert to jsinterp.cpp after js_FindIdentifierBase in JSOP_NAME:

    JS_ASSERT_IF(op == JSOP_GETGNAME || op == JSOP_CALLGNAME, obj == obj->getGlobal());

Then doMath.js fails (and another, filed as bug 585524). Turning the DEFFUN into DEFLOCALFUN would break our current (though out-of-spec) semantics. Instead maybe this and bug 585524 should share a common solution of marking which names cannot be optimized.

Comment 5

[Luke Wagner [:luke]]

A simple test case is (which throws 'too much recursion') is:

function f(x) {
    if (x) {
        function f() { return 2; }
        assertEq(f(true), 2);
    }
}
f(true);

Comment 6

[David Mandelin [:dmandelin]]

Attachment: Patch

This patch avoids the JSOP_GETGLOBAL optimization if the JSOP_DEFFUN op occurs. Summarizing earlier discussion, this is necessary because a JSOP_DEFFUN can introduce a shadowing definition, but it will not be seen in the parser because that shadowing is not recorded in the parser; it is only introduced at runtime when the JSOP_DEFFUN is executed.

Comment 7

[David Mandelin [:dmandelin]]

I forgot to say in comment 6 that I verified that the deoptimization does not occur in Sunspider, V8, or Kraken.

Comment 8

[David Mandelin [:dmandelin]]

http://hg.mozilla.org/tracemonkey/rev/96a84cd98d84

Comment 11

[Brian Hackett [Laid off!]]

This still seems to be not working quite right.

function g(v) {
  with (v) {
    var f = function() { print("HI"); }
    f();
  }
  f();
}
g({});

The 'f' function is called twice here, but on disassembly the first call is a CALLNAME and the second is a CALLGNAME.

Comment 12

[Robert Sayre]

http://hg.mozilla.org/mozilla-central/rev/96a84cd98d84

Comment 13

[Brendan Eich [:brendan]]

(In reply to comment #11)
> This still seems to be not working quite right.
> 
> function g(v) {
>   with (v) {
>     var f = function() { print("HI"); }
>     f();
>   }
>   f();
> }
> g({});
> 
> The 'f' function is called twice here, but on disassembly the first call is a
> CALLNAME and the second is a CALLGNAME.

David, doesn't this need a followup bug?

/be

Comment 14

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

Was about to file, but just realized: that test case is different, and is bug 585524.