Closed Bug 900125 Opened 7 years ago Closed 7 years ago

Float32: add Math.fround to the interpreter

Categories

(Core :: JavaScript Engine, defect)

defect
Not set

Tracking

()

RESOLVED FIXED
mozilla26
Tracking Status
relnote-firefox --- 26+

People

(Reporter: bbouvier, Assigned: bbouvier)

References

(Blocks 1 open bug)

Details

(Keywords: dev-doc-complete)

Attachments

(2 files, 6 obsolete files)

Math.ToFloat32 takes a JS value and converts it to a Float32, whenever possible.
This function has been successfully proposed to TC39 recently.
Attached patch WIP (obsolete) — Splinter Review
This is actually trivial.

I see that some math functions belong to the js:: namespace and some other don't.
Waldo or luke, is there any criteria of choice to prefer where to put the function?

Luke, should I add a browser pref for this feature, as it hasn't been accepted by the committee yet?

Waldo, I would like to add some tests for this function. I obviously understand that I should put them in /js/src/tests/, but is there any specific subdirectory where to put them?
Flags: needinfo?(luke)
Flags: needinfo?(jwalden+bmo)
It sounds like there was general support at the most recent TC39 meeting; I forget Dave, was it added to todo list of ES6?  Given that, it seems like, as with Math.imul, we have good reason to add it w/o a pref.  Agreed Dave?  Also is there any strawman/wiki URL we can link to here?
Flags: needinfo?(luke)
Regarding the js:: question: we are incrementally moving all the things to namespace js and un-prefixing, so let's do that (you can put it righ tnext to js::math_imul).  (If you'd like to write a patch to move the remaining extern js_* functions to js::, go for it :)
The JS engine can't depend on browser prefs anyway, because it's browser that depends on js/src, not the other way around.  (We could compile-time pref it based on update channel, I think, but that's slightly messy, and best avoided if we can, as it seems like is probably possible here.)
Flags: needinfo?(jwalden+bmo)
Blocks: 900257
Attached patch Proposed implementation (obsolete) — Splinter Review
Same implementation, just the order of functions has been changed.
Attachment #783894 - Attachment is obsolete: true
Attachment #784118 - Flags: review?(sstangl)
Attached patch Tests (obsolete) — Splinter Review
A few tests. As there is no formal spec right now, I just considered some obvious cases, so I would love to get feedback and / or ideas regarding some other nice tests to add.
Attachment #784119 - Flags: review?(jwalden+bmo)
Flags: needinfo?(dherman)
Comment on attachment 784119 [details] [diff] [review]
Tests

Review of attachment 784119 [details] [diff] [review]:
-----------------------------------------------------------------

Et viola!  ;-)

Probably you could go even further than I suggested, I'm just covering the obvious edge-case bases that occurred to me with a few minutes' thinking.  (And a few hours' drawing up, grrrr.)

::: js/src/tests/ecma_6/Math/toFloat32.js
@@ +1,1 @@
> +// Some tests regarding conversion to Float32

Put a public-domain license block at the top of this -- see js/src/tests/ecma_6/Map/NaN-as-key.js for an example.

@@ +4,5 @@
> +assertEq(Math.toFloat32(NaN), NaN);
> +assertEq(Math.toFloat32(-Infinity), -Infinity);
> +assertEq(Math.toFloat32(Infinity), Infinity);
> +assertEq(Math.toFloat32(-0), -0);
> +assertEq(Math.toFloat32(+0), +0);

We should test the maximum number (float64/double) value and the maximum float32 value:

  function maxValue(exponentWidth, significandWidth)
  {
    var n = 0;
    var maxExp = Math.pow(2, exponentWidth - 1) - 1;
    for (var i = significandWidth; i >= 0; i--)
      n += Math.pow(2, maxExp - i);
    return n;
  }

  var DBL_MAX = maxValue(11, 52);
  assertEq(DBL_MAX, Number.MAX_VALUE); // sanity check

  // Finite as a double, too big for a float
  assertEq(Math.toFloat32(DBL_MAX), Infinity);

  var FLT_MAX = maxValue(8, 23);
  assertEq(Math.toFloat32(FLT_MAX), FLT_MAX);

  assertEq(Math.toFloat32(FLT_MAX + Math.pow(2, Math.pow(2, 8 - 1) - 1 - 23 - 2)), FLT_MAX); // round-nearest rounds down to FLT_MAX
  assertEq(Math.toFloat32(FLT_MAX + Math.pow(2, Math.pow(2, 8 - 1) - 1 - 23 - 1)), Infinity); // no longer rounds down to FLT_MAX

We should also test some denormals:

  function minValue(exponentWidth, significandWidth)
  {
    return Math.pow(2, -(Math.pow(2, exponentWidth - 1) - 2) - significandWidth);
  }

  var DBL_MIN = Math.pow(2, -1074);
  assertEq(DBL_MIN, Number.MIN_VALUE); // sanity check

  // Too small for a float
  assertEq(Math.toFloat32(DBL_MIN), 0);

  var FLT_MIN = minValue(8, 23);
  assertEq(Math.toFloat32(FLT_MIN), FLT_MIN);

  assertEq(Math.toFloat32(FLT_MIN / 2), 0); // halfway, round-nearest rounds down to 0 (even)
  assertEq(Math.toFloat32(FLT_MIN / 2 + Math.pow(2, -202)), FLT_MIN); // first double > FLT_MIN / 2, rounds up to FLT_MIN

Also some tests that round (even-directed and non-even-directed) to negative zero would be good:

  assertEq(Math.toFloat32(-FLT_MIN), -FLT_MIN);

  assertEq(Math.toFloat32(-FLT_MIN / 2), -0); // halfway, round-nearest rounds up to -0 (even)
  assertEq(Math.toFloat32(-FLT_MIN / 2 - Math.pow(2, -202)), -FLT_MIN); // first double < -FLT_MIN / 2, rounds down to -FLT_MIN

I think all that code will just work, as I did a bunch of manual testing of it (not using your patch, just toFloat32-like operations), but you should test it and be sure.
Attachment #784119 - Flags: review?(jwalden+bmo) → review+
Attached patch More tests (obsolete) — Splinter Review
Awesome tests, Waldo!
They all pass with Math.toFloat32 in the interpreter and also with --ion-eager and the patch in bug 900257.

Yay \o/

Carrying over r+.
Attachment #784119 - Attachment is obsolete: true
Attachment #785240 - Flags: review+
Blocks: 904918
Comment on attachment 784118 [details] [diff] [review]
Proposed implementation

Review of attachment 784118 [details] [diff] [review]:
-----------------------------------------------------------------

::: js/src/jsmath.cpp
@@ +429,5 @@
> +        return false;
> +
> +    float f = x;
> +    x = f;
> +    args.rval().setNumber(x);

setDouble((double)f)
Attachment #784118 - Flags: review?(sstangl) → review+
double(f) or static_cast<double>(f) -- I'd go with the former -- would be better than a C-style cast.
This is my first draft of a spec for Math.toFloat32:

    https://etherpad.mozilla.org/Math-toFloat32

I've got emails out to various stakeholders to check my thinking. Note that there's an implicit ToNumber coercion applied to the argument on entry to the function.

Dave
Flags: needinfo?(dherman)
Note: naming change in the spec to Math.roundFloat32 since it was pointed out that all other toX() funtions are methods on a prototype acting on 'this', so toFloat32 didn't fit in.
Summary: Float32: add Math.ToFloat32 to the interpreter → Float32: add Math.roundFloat32 to the interpreter
(In reply to Dave Herman [:dherman] from comment #11)

After the initial ToNumber on the argument, why wouldn't you delegate to IEEE-754?  Something like

Math.toFloat32(x):
1. Let x be ToNumber(x).
2. Let y be x, converted to single format according to the IEEE-754 specification, using the round to nearest mode.  NOTE: The number y may not be the same as the number x, and it may be infinite when x is finite.  If x is a NaN value, the bit pattern of y is unspecified.
3. Return y.

or so.  Trying to write out an algorithm for this, rather than delegating to IEEE-754, is just asking to get something wrong.
Not that it fits well with asm.js (but I guess that's our problem to solve, not the spec's), but wouldn't it be more idiomatically JS for the method to be Number.prototype.toFloat32(), if the toX concern is valid (which it is)?
(In reply to Jeff Walden [:Waldo] (remove +bmo to email) from comment #14)
Allen made that point.  Yeah, Number.prototype.toFloat32 is just inherently harder for asm.js to lock down so that, after an initial check, you don't need to check anymore (you can, but it's gnarly and it would be slow on normal engines).
carrying over r+ from sstangl (fixed conversion).

This currently implements what Waldo said (modulo the fact that if there's no argument, NaN is returned).
Attachment #784118 - Attachment is obsolete: true
Attachment #793162 - Flags: review+
Attached patch 2 - Tests (obsolete) — Splinter Review
Carrying over r+ from Waldo.

Added a test that Math.roundFloat32() == NaN.
Attachment #785240 - Attachment is obsolete: true
Attachment #793163 - Flags: review+
After discussions, it looks like the name of this function should be Math.fround, for Float rounding. This is consistent with libc's math library. The name can still be changed in the future in a follow-up bug.

Carrying over r+ from sstangl
Attachment #793162 - Attachment is obsolete: true
Attachment #796998 - Flags: review+
Attached patch 2 - TestsSplinter Review
Carrying over r+ from Waldo.
Attachment #793163 - Attachment is obsolete: true
Attachment #796999 - Flags: review+
https://hg.mozilla.org/mozilla-central/rev/fefe46f36a5b
https://hg.mozilla.org/mozilla-central/rev/1882f24d48d5
Status: NEW → RESOLVED
Closed: 7 years ago
Resolution: --- → FIXED
Target Milestone: --- → mozilla26
Summary: Float32: add Math.roundFloat32 to the interpreter → Float32: add Math.fround to the interpreter
Thank you very much, Benjamin. A small info: once documentation is done, we don't remove the dev-doc-needed flag, we switch it to dev-doc-complete, that we can continue to keep track of it ;-)
Nom'ing for relnote 26.
relnote-firefox: --- → ?
You need to log in before you can comment on or make changes to this bug.