ClojureScript

Incorrect behaviour of (str obj) when obj has valueOf method

Details

  • Type: Defect Defect
  • Status: Reopened Reopened
  • Priority: Minor Minor
  • Resolution: Unresolved
  • Affects Version/s: None
  • Fix Version/s: None
  • Component/s: None
  • Labels:
    None
  • Patch:
    Code

Description

Example

(str #js {"toString" (fn [] "hello") "valueOf" (fn [] 42)}) ; => "42"

The problem in the fact that ClojureScript uses concatenation to convert values to strings and that doesn't work well with objects which have valueOf() method overriden.

Example in js:

var obj = {
    toString: function() { return 'hello'; },
    valueOf: function() { return 42; }
};
console.log(String(obj)); => 'hello'
console.log(obj.toString()); => 'hello'
console.log('' + obj); => '42'

Potential solution might be to use String() function. Using toString() won't work as described in this issue: http://dev.clojure.org/jira/browse/CLJS-847

  1. cljs-890.patch
    19/Jan/15 11:34 PM
    3 kB
    Francis Avila
  2. cljs-core-str-perf.diff
    01/Jan/15 7:28 AM
    1 kB
    Thomas Heller

Activity

Hide
Kevin Neaton added a comment -

Is there a valid use case where toString and valueOf are not in sync? E.g.

(not= (.toString x) (js/String (.valueOf x))

If not, is it "incorrect" for the two methods to be out of sync?

Show
Kevin Neaton added a comment - Is there a valid use case where toString and valueOf are not in sync? E.g.
(not= (.toString x) (js/String (.valueOf x))
If not, is it "incorrect" for the two methods to be out of sync?
Hide
Nikita Beloglazov added a comment -

Here is an example of such use case: https://github.com/processing-js/processing-js/blob/master/src/Objects/Char.js
That's how I found this bug.

Show
Nikita Beloglazov added a comment - Here is an example of such use case: https://github.com/processing-js/processing-js/blob/master/src/Objects/Char.js That's how I found this bug.
Hide
Kevin Neaton added a comment -

Thanks for the link. I see what you mean.

Show
Kevin Neaton added a comment - Thanks for the link. I see what you mean.
Hide
David Nolen added a comment -

The problem with going with String appears to be a massive performance hit to printing http://jsperf.com/string-vs-tostring2/6.

Unless a brilliant idea is proposed this seems best solved / worked around in user code.

Show
David Nolen added a comment - The problem with going with String appears to be a massive performance hit to printing http://jsperf.com/string-vs-tostring2/6. Unless a brilliant idea is proposed this seems best solved / worked around in user code.
Hide
Nikita Beloglazov added a comment -

Append performs better on strings and numbers, but it performs worse on objects so it is not a clear performance hit. If I heavily work with objects and use (str) to convert them into strings then I actually lose on performance with current implementation.
Anyway current implementation of str is incorrect as it doesn't honor toString method. And this is what str function supposed to do. I believe a compiler should be correct first and then worry about performance.

Show
Nikita Beloglazov added a comment - Append performs better on strings and numbers, but it performs worse on objects so it is not a clear performance hit. If I heavily work with objects and use (str) to convert them into strings then I actually lose on performance with current implementation. Anyway current implementation of str is incorrect as it doesn't honor toString method. And this is what str function supposed to do. I believe a compiler should be correct first and then worry about performance.
Hide
David Nolen added a comment -

Sorry going back over this I believe the issue is really that we simply need to backout CLJS-801.

Show
David Nolen added a comment - Sorry going back over this I believe the issue is really that we simply need to backout CLJS-801.
Hide
David Nolen added a comment -

reverted CLJS-801 in master

Show
David Nolen added a comment - reverted CLJS-801 in master
Hide
Francis Avila added a comment -

CLJS-801 only deals with the str macro. Aren't we still going to have str function problem because of CLJS-847? https://github.com/clojure/clojurescript/commit/08b4b1585cf0ef739e903985ee4c6b7fc6c47642 (also Yet if we use toString there, Safari 6.0.5 blows up. Maybe we need {{[o].join('')}}? Depending on where the bug is this may be wrong in Safari 6.0.5 too.

What we need to do very specifically is somehow get the return value of the (in ECMASCRIPT-ese) ToString abstract operation on the object (or the underlying ToPrimitive abstract operation with the String hint). String concat with the add operator

Options as I see it are:

  • x.toString() : Bad because of CLJS-847
  • {{[x].join('')}} : Should work (and does right thing for null/undefined), but I think we should test in Safari 6.0.5. Also very slow.
  • String
  • String.prototype.concat
  • String.prototype.slice(x,0) String.prototype.substring(x,0) String.prototype.substr(x, 0)
  • x.toString() normally, but String if we detect that we'll trigger CLJS-847. (Can specialize on startup.)
Show
Francis Avila added a comment - CLJS-801 only deals with the str macro. Aren't we still going to have str function problem because of CLJS-847? https://github.com/clojure/clojurescript/commit/08b4b1585cf0ef739e903985ee4c6b7fc6c47642 (also Yet if we use toString there, Safari 6.0.5 blows up. Maybe we need {{[o].join('')}}? Depending on where the bug is this may be wrong in Safari 6.0.5 too. What we need to do very specifically is somehow get the return value of the (in ECMASCRIPT-ese) ToString abstract operation on the object (or the underlying ToPrimitive abstract operation with the String hint). String concat with the add operator Options as I see it are:
  • x.toString() : Bad because of CLJS-847
  • {{[x].join('')}} : Should work (and does right thing for null/undefined), but I think we should test in Safari 6.0.5. Also very slow.
  • String
  • String.prototype.concat
  • String.prototype.slice(x,0) String.prototype.substring(x,0) String.prototype.substr(x, 0)
  • x.toString() normally, but String if we detect that we'll trigger CLJS-847. (Can specialize on startup.)
Hide
David Nolen added a comment -

Is there any evidence that higher usage of str is actually problematic?

Show
David Nolen added a comment - Is there any evidence that higher usage of str is actually problematic?
Hide
Francis Avila added a comment -

String concat using the addition operator uses an un-hinted ToPrimitive abstract call (which will try x.valueOf() first then x.toString(), usually) and then {{ToString}}s the result of that, so it's not an option unless we are concating primitive values.

Details:

Show
Francis Avila added a comment - String concat using the addition operator uses an un-hinted ToPrimitive abstract call (which will try x.valueOf() first then x.toString(), usually) and then {{ToString}}s the result of that, so it's not an option unless we are concating primitive values. Details:
Hide
David Nolen added a comment -

I'm not really all that concerned about the specification, only if it matters in the wild. If this doesn't affect Safari 6.05 we don't care.

Show
David Nolen added a comment - I'm not really all that concerned about the specification, only if it matters in the wild. If this doesn't affect Safari 6.05 we don't care.
Hide
Francis Avila added a comment -

Is there any evidence that higher usage of str is actually problematic?

Kevin Neaton, who opened CLJS-847, was using a patch in production which only addressed the higher order case and he said the patch fixed the issue for them. He was unaffected by the str macro case because it either used ''+x already (with CLJS-801 applied) or it used {{[x].join('')}} (which hasn't been tested with Safari 6.0.5 yet, but probably works).

So if we had a problem using ''+x with the str macro, we will certainly have a problem with ''+x with a string function as long as CLJS-847 is applied.

I haven't pulled down master yet, but here is a test case which I bet will fail with the CLJS-847 patch:

(def tricky-obj #js {"toString" (fn [] "hello") "valueOf" (fn [] 42)})
(assert (= (apply str tricky-obj) "hello")) ;; will get "42"
Show
Francis Avila added a comment -
Is there any evidence that higher usage of str is actually problematic?
Kevin Neaton, who opened CLJS-847, was using a patch in production which only addressed the higher order case and he said the patch fixed the issue for them. He was unaffected by the str macro case because it either used ''+x already (with CLJS-801 applied) or it used {{[x].join('')}} (which hasn't been tested with Safari 6.0.5 yet, but probably works). So if we had a problem using ''+x with the str macro, we will certainly have a problem with ''+x with a string function as long as CLJS-847 is applied. I haven't pulled down master yet, but here is a test case which I bet will fail with the CLJS-847 patch:
(def tricky-obj #js {"toString" (fn [] "hello") "valueOf" (fn [] 42)})
(assert (= (apply str tricky-obj) "hello")) ;; will get "42"
Hide
Francis Avila added a comment -

I'm not really all that concerned about the specification, only if it matters in the wild. If this doesn't affect Safari 6.05 we don't care.

To be clear, there are two issues here:

CLJS-847: x.toString() fails on Safari 6.0.5. Workaround is ''+x (only done in str macro case).
CLJS-890: ''+x doesn't give expected results for objects which define valueOf. Expectation is that x.toString() is called, instead x.valueOf().toString(). Fix is to use array join instead of string concat in str macro, but it doesn't address the ''+x workaround from CLJS-847.

To make matters worse, it looks like the toString() error on Safari may only be triggered at certain JIT levels!

Show
Francis Avila added a comment -
I'm not really all that concerned about the specification, only if it matters in the wild. If this doesn't affect Safari 6.05 we don't care.
To be clear, there are two issues here: CLJS-847: x.toString() fails on Safari 6.0.5. Workaround is ''+x (only done in str macro case). CLJS-890: ''+x doesn't give expected results for objects which define valueOf. Expectation is that x.toString() is called, instead x.valueOf().toString(). Fix is to use array join instead of string concat in str macro, but it doesn't address the ''+x workaround from CLJS-847. To make matters worse, it looks like the toString() error on Safari may only be triggered at certain JIT levels!
Hide
Francis Avila added a comment -

Workaround is ''+x (only done in str macro case).

I mean "Workaround is ''+x (only done in str function case)."

Show
Francis Avila added a comment -
Workaround is ''+x (only done in str macro case).
I mean "Workaround is ''+x (only done in str function case)."
Hide
Nikita Beloglazov added a comment -

Can this bug be reopened meanwhile? If I understand correctly the fix should affect https://github.com/clojure/clojurescript/commit/08b4b1585cf0ef739e903985ee4c6b7fc6c47642 but this code still present in HEAD.

Show
Nikita Beloglazov added a comment - Can this bug be reopened meanwhile? If I understand correctly the fix should affect https://github.com/clojure/clojurescript/commit/08b4b1585cf0ef739e903985ee4c6b7fc6c47642 but this code still present in HEAD.
Hide
David Nolen added a comment -
Show
David Nolen added a comment - We've switched to goog.string.buildString in master https://github.com/clojure/clojurescript/commit/94eb8a960fef6aaca4ba44b251cefbfa04d0f6ac
Hide
Nikita Beloglazov added a comment -

Yes, that works. Cool, thanks!

Show
Nikita Beloglazov added a comment - Yes, that works. Cool, thanks!
Hide
Thomas Heller added a comment -

Sorry for re-opening.

I was doing some profiling of my code and noticed a warning in the profiling output about cljs.core/str.

Chromes complains with: "Not optimized. Bad value context for arguments value", looking further at the implementation of goog.string.buildString

goog.string.buildString = function(var_args) {
  return Array.prototype.join.call(arguments, '');
};

Given that we don't ever call it with more than one argument it is probably not best implementation choice.

Maybe skip the call and inline it ala

(defn str
  "With no args, returns the empty string. With one arg x, returns
  x.toString().  (str nil) returns the empty string. With more than
  one arg, returns the concatenation of the str values of the args."
  ([] "")
  ([x] (if (nil? x)
         ""
         (.join #js [x] "")))
  ([x & ys]
    (loop [sb (StringBuffer. (str x)) more ys]
      (if more
        (recur (. sb  (append (str (first more)))) (next more))
        (.toString sb)))))

I didn't follow this issue but why are we not using .toString? The buildString/array approach seems kind of hackish?

I'm not too sure about the overall impact but since cljs.core/str showed up pretty high in my profile I think this should be investigated further.

Show
Thomas Heller added a comment - Sorry for re-opening. I was doing some profiling of my code and noticed a warning in the profiling output about cljs.core/str. Chromes complains with: "Not optimized. Bad value context for arguments value", looking further at the implementation of goog.string.buildString
goog.string.buildString = function(var_args) {
  return Array.prototype.join.call(arguments, '');
};
Given that we don't ever call it with more than one argument it is probably not best implementation choice. Maybe skip the call and inline it ala
(defn str
  "With no args, returns the empty string. With one arg x, returns
  x.toString().  (str nil) returns the empty string. With more than
  one arg, returns the concatenation of the str values of the args."
  ([] "")
  ([x] (if (nil? x)
         ""
         (.join #js [x] "")))
  ([x & ys]
    (loop [sb (StringBuffer. (str x)) more ys]
      (if more
        (recur (. sb  (append (str (first more)))) (next more))
        (.toString sb)))))
I didn't follow this issue but why are we not using .toString? The buildString/array approach seems kind of hackish? I'm not too sure about the overall impact but since cljs.core/str showed up pretty high in my profile I think this should be investigated further.
Hide
Thomas Heller added a comment -

Before:

;;; str
[], (str "1"), 1000000 runs, 254 msecs
[], (str 1), 1000000 runs, 266 msecs
[], (str nil), 1000000 runs, 80 msecs
[], (str "1" "2" "3"), 1000000 runs, 753 msecs

After:

;;; str
[], (str "1"), 1000000 runs, 82 msecs
[], (str 1), 1000000 runs, 86 msecs
[], (str nil), 1000000 runs, 79 msecs
[], (str "1" "2" "3"), 1000000 runs, 242 msecs

But I only tested V8, probably needs some verification.

Show
Thomas Heller added a comment - Before:
;;; str
[], (str "1"), 1000000 runs, 254 msecs
[], (str 1), 1000000 runs, 266 msecs
[], (str nil), 1000000 runs, 80 msecs
[], (str "1" "2" "3"), 1000000 runs, 753 msecs
After:
;;; str
[], (str "1"), 1000000 runs, 82 msecs
[], (str 1), 1000000 runs, 86 msecs
[], (str nil), 1000000 runs, 79 msecs
[], (str "1" "2" "3"), 1000000 runs, 242 msecs
But I only tested V8, probably needs some verification.
Hide
Thomas Heller added a comment -
(defn str
  "With no args, returns the empty string. With one arg x, returns
  x.toString().  (str nil) returns the empty string. With more than
  one arg, returns the concatenation of the str values of the args."
  ([] "")
  ([x1]
     (.join #js [x1] ""))
  ([x1 x2]
     (.join #js [x1 x2] ""))
  ([x1 x2 x3]
     (.join #js [x1 x2 x3] ""))
  ([x1 x2 x3 x4]
     (.join #js [x1 x2 x3 x4] ""))
  ...)

Does perform even better.

;;; str
[], (str "1"), 1000000 runs, 40 msecs
[], (str 1), 1000000 runs, 43 msecs
[], (str nil), 1000000 runs, 96 msecs
[], (str "1" "2" "3"), 1000000 runs, 117 msecs

How many args should it inline?

Show
Thomas Heller added a comment -
(defn str
  "With no args, returns the empty string. With one arg x, returns
  x.toString().  (str nil) returns the empty string. With more than
  one arg, returns the concatenation of the str values of the args."
  ([] "")
  ([x1]
     (.join #js [x1] ""))
  ([x1 x2]
     (.join #js [x1 x2] ""))
  ([x1 x2 x3]
     (.join #js [x1 x2 x3] ""))
  ([x1 x2 x3 x4]
     (.join #js [x1 x2 x3 x4] ""))
  ...)
Does perform even better.
;;; str
[], (str "1"), 1000000 runs, 40 msecs
[], (str 1), 1000000 runs, 43 msecs
[], (str nil), 1000000 runs, 96 msecs
[], (str "1" "2" "3"), 1000000 runs, 117 msecs
How many args should it inline?
Hide
David Nolen added a comment -

I'd be OK with up to 4 then variadic.

Show
David Nolen added a comment - I'd be OK with up to 4 then variadic.
Hide
Thomas Heller added a comment -

There is some weird interaction between the code generated by the cljs.core/str macro and function.

The macro generates

(str "hello" 1 "world" :yo nil)

yields

[cljs.core.str("hello"),cljs.core.str((1)),cljs.core.str("world"),cljs.core.str(new cljs.core.Keyword(null,"yo","yo",1207083126)),cljs.core.str(null)].join('');

Given that str with 1 arg will basically unroll to

[["hello"].join(""), ...]

I think it might be safe to completely remove the macro since cljs.core/str would then do the same and the JIT is probably smart enough to figure this out (or even Closure when compiling).

Show
Thomas Heller added a comment - There is some weird interaction between the code generated by the cljs.core/str macro and function. The macro generates
(str "hello" 1 "world" :yo nil)

yields

[cljs.core.str("hello"),cljs.core.str((1)),cljs.core.str("world"),cljs.core.str(new cljs.core.Keyword(null,"yo","yo",1207083126)),cljs.core.str(null)].join('');
Given that str with 1 arg will basically unroll to
[["hello"].join(""), ...]
I think it might be safe to completely remove the macro since cljs.core/str would then do the same and the JIT is probably smart enough to figure this out (or even Closure when compiling).
Hide
Francis Avila added a comment - - edited

Chromes complains with: "Not optimized. Bad value context for arguments value", looking further at the implementation of goog.string.buildString

Chrome complains about the variadic function dispatch code in the same way, see CLJS-916 plus patch.

I think it might be safe to completely remove the macro since cljs.core/str would then do the same and the JIT is probably smart enough to figure this out (or even Closure when compiling).

The Closure compiler is not smart enough to remove the intermediate array, which is why I filed CLJS-801 (which this ticket rolled back). I don't think JITs can do it either.

I am beginning to wonder if we should ignore the Safari 6.0.5 problem in CLJS-847 that started all this string mess. To recap:

  1. CLJS-801 is accepted, which removes [123, x].join('') in the str macro case in favor of ''+123+(cljs.core/str$arity$1 x) style code, which the closure compiler can precompute. At this time, the one-arg cljs.core/str function (not macro) calls toString on its argument.
  2. CLJS-847 is filed. On Safari 6.0.5 at higher JIT levels, calling toString on some things (possibly only unboxed numbers? definitely not strings) throws a TypeError. This is unquestionably a bug in Safari. David fixes by making one-arg cljs.core/str function call js-str instead of toString. js-str uses string-concat ''+x.
  3. However, this breaks for objects that define valueOf (issue in current ticket), because in js ''+x is the same as ''+x.valueOf().toString() not ''+x.toString().
  4. David considers using String() and variations but rejects because of performance hit.
  5. David rolls back CLJS-801 from the string-concat back to the array-join style to fix.
  6. Nikita and I point out that rolling back CLJS-801 only fixes the str macro, not the string function, which still uses js-str and hence string-concat.
  7. David fixes the str function to use goog.string.buildString, which has the behavior of array.join. Behavior is now correct even on Safari 6.0.5.
  8. Thomas points out that buildString uses arguments in a way unoptimizable by v8, and now the str function (not macro) has a performance regression. He suggests using [].join() directly.

So, there's a lot of back and forth on this issue, and it's all because of a bug in Safari 6.0.5 which no one has been able to reproduce first-hand because Safari 6.0.5 is old and rare. For some perspective, Safari 6.0.x was only available on Lion and Mountain Lion between July 25,2012 and June 11,2013. Before July 25,2012 Lion used Safari 5.1.x and there was no Mountain Lion. On June 11, 2013, both Lion and Mountain Lion switched to Safari 6.1.x which does not suffer from the toString TypeError bug (I checked--I have an iMac with Lion on it). The only machines on Safari 6.0.5 are (Mountain) Lion machines which used software updates until late 2012-early 2013 and then stopped. I can't imagine this is a large number of people.

It is theoretically possible for me to run Safari 6.0.x on my Lion machine to actually test this, but I can't find a way to downgrade from 6.1.x.

I think the options are:

  1. Use array.join() for all stringification and take the performance hit (which we should quantify). Include a comment that this is only for Safari 6.0.x (only confirmed second-hand on 6.0.4 and 6.0.5) for future generations, who are going to think this is weird.
  2. Use CLJS-801 and toString (status quo before CLJS-847), and ignore this problem for Safari 6.0.x.
  3. Use CLJS-801, but add a number? check (with comment) to cljs.core/str$arity$1 for Safari 6.0.5. The number case should use js-str, and the rest toString. I think this will work, but again we have no way to test--we really need to get our hands on a Safari 6.0.x browser.

Of course we should benchmark these approaches but my hunch is that 2 is faster than 3 is faster than 1.

Show
Francis Avila added a comment - - edited
Chromes complains with: "Not optimized. Bad value context for arguments value", looking further at the implementation of goog.string.buildString
Chrome complains about the variadic function dispatch code in the same way, see CLJS-916 plus patch.
I think it might be safe to completely remove the macro since cljs.core/str would then do the same and the JIT is probably smart enough to figure this out (or even Closure when compiling).
The Closure compiler is not smart enough to remove the intermediate array, which is why I filed CLJS-801 (which this ticket rolled back). I don't think JITs can do it either. I am beginning to wonder if we should ignore the Safari 6.0.5 problem in CLJS-847 that started all this string mess. To recap:
  1. CLJS-801 is accepted, which removes [123, x].join('') in the str macro case in favor of ''+123+(cljs.core/str$arity$1 x) style code, which the closure compiler can precompute. At this time, the one-arg cljs.core/str function (not macro) calls toString on its argument.
  2. CLJS-847 is filed. On Safari 6.0.5 at higher JIT levels, calling toString on some things (possibly only unboxed numbers? definitely not strings) throws a TypeError. This is unquestionably a bug in Safari. David fixes by making one-arg cljs.core/str function call js-str instead of toString. js-str uses string-concat ''+x.
  3. However, this breaks for objects that define valueOf (issue in current ticket), because in js ''+x is the same as ''+x.valueOf().toString() not ''+x.toString().
  4. David considers using String() and variations but rejects because of performance hit.
  5. David rolls back CLJS-801 from the string-concat back to the array-join style to fix.
  6. Nikita and I point out that rolling back CLJS-801 only fixes the str macro, not the string function, which still uses js-str and hence string-concat.
  7. David fixes the str function to use goog.string.buildString, which has the behavior of array.join. Behavior is now correct even on Safari 6.0.5.
  8. Thomas points out that buildString uses arguments in a way unoptimizable by v8, and now the str function (not macro) has a performance regression. He suggests using [].join() directly.
So, there's a lot of back and forth on this issue, and it's all because of a bug in Safari 6.0.5 which no one has been able to reproduce first-hand because Safari 6.0.5 is old and rare. For some perspective, Safari 6.0.x was only available on Lion and Mountain Lion between July 25,2012 and June 11,2013. Before July 25,2012 Lion used Safari 5.1.x and there was no Mountain Lion. On June 11, 2013, both Lion and Mountain Lion switched to Safari 6.1.x which does not suffer from the toString TypeError bug (I checked--I have an iMac with Lion on it). The only machines on Safari 6.0.5 are (Mountain) Lion machines which used software updates until late 2012-early 2013 and then stopped. I can't imagine this is a large number of people. It is theoretically possible for me to run Safari 6.0.x on my Lion machine to actually test this, but I can't find a way to downgrade from 6.1.x. I think the options are:
  1. Use array.join() for all stringification and take the performance hit (which we should quantify). Include a comment that this is only for Safari 6.0.x (only confirmed second-hand on 6.0.4 and 6.0.5) for future generations, who are going to think this is weird.
  2. Use CLJS-801 and toString (status quo before CLJS-847), and ignore this problem for Safari 6.0.x.
  3. Use CLJS-801, but add a number? check (with comment) to cljs.core/str$arity$1 for Safari 6.0.5. The number case should use js-str, and the rest toString. I think this will work, but again we have no way to test--we really need to get our hands on a Safari 6.0.x browser.
Of course we should benchmark these approaches but my hunch is that 2 is faster than 3 is faster than 1.
Hide
David Nolen added a comment -

We are not going to ignore Safari 6.0.X. Any decisions made about this ticket will include supporting it.

Show
David Nolen added a comment - We are not going to ignore Safari 6.0.X. Any decisions made about this ticket will include supporting it.
Hide
Francis Avila added a comment - - edited

Update on some research I am doing into this.

I created a jsperf of alternative str implementations that I am trying out. Right now I've only looked at one-arg str. I discovered a few things:

  • {{''+[x]}} is a faster alternative to [x].join('').
  • Advanced compilation can compute {{''+[x]}} at compile time if x is a bool, str, undefined, null, or number, even through function calls! I.e. str_test.str_arr(123) compiles to "123" without macro magic.
  • However, using an intermediate array (even if a preallocated singleton) is still slower than the old (if (nil? x) "" (.toString x))
  • Using a switch statement is as least as fast as the str-tostr baseline, usually faster.
  • I am 99% sure all these implementations (except str-tostr, the baseline, which definitely fails) work on the dreaded Safari 6.0.x. If anyone has this version, point it at the jsperf link above and run the tests. I think Browserstack has this version of Safari.

I'm still investigating the variadic case (str x y z a b c). It might be better to use reduce instead of Stringbuffer+seq. (Stringbuffer just does ''+x now instead of an array-join.)

Show
Francis Avila added a comment - - edited Update on some research I am doing into this. I created a jsperf of alternative str implementations that I am trying out. Right now I've only looked at one-arg str. I discovered a few things:
  • {{''+[x]}} is a faster alternative to [x].join('').
  • Advanced compilation can compute {{''+[x]}} at compile time if x is a bool, str, undefined, null, or number, even through function calls! I.e. str_test.str_arr(123) compiles to "123" without macro magic.
  • However, using an intermediate array (even if a preallocated singleton) is still slower than the old (if (nil? x) "" (.toString x))
  • Using a switch statement is as least as fast as the str-tostr baseline, usually faster.
  • I am 99% sure all these implementations (except str-tostr, the baseline, which definitely fails) work on the dreaded Safari 6.0.x. If anyone has this version, point it at the jsperf link above and run the tests. I think Browserstack has this version of Safari.
I'm still investigating the variadic case (str x y z a b c). It might be better to use reduce instead of Stringbuffer+seq. (Stringbuffer just does ''+x now instead of an array-join.)
Hide
Thomas Heller added a comment -

Sorry, got side-tracked for a bit.

@Francis: Thanks for the recap.

Don't have Safari 6 available either, but seems wrong that we all have to suffer because is minor percentage still has this (667 users of 190k+ on my site). Don't have a solution cause I can't test whether it works, we might try String.concat.

"".concat(obj); // "42"
"".concat(obj, ""); // "hello"
String.prototype.concat(obj, "") // "hello"
String.prototype.concat("", obj) // "hello"

But no idea if String.concat works, also it behaves odd with regards to valueOf.

http://jsperf.com/js-string-concat-variants

Perf is also inconclusive since Firefox appears to be cheating.

Show
Thomas Heller added a comment - Sorry, got side-tracked for a bit. @Francis: Thanks for the recap. Don't have Safari 6 available either, but seems wrong that we all have to suffer because is minor percentage still has this (667 users of 190k+ on my site). Don't have a solution cause I can't test whether it works, we might try String.concat.
"".concat(obj); // "42"
"".concat(obj, ""); // "hello"
String.prototype.concat(obj, "") // "hello"
String.prototype.concat("", obj) // "hello"
But no idea if String.concat works, also it behaves odd with regards to valueOf. http://jsperf.com/js-string-concat-variants Perf is also inconclusive since Firefox appears to be cheating.
Hide
Francis Avila added a comment -

Tested that jsperf with Safari 6.0.5 using Browserstack, results are there.

Note I could not reproduce CLJS-847 because str-tostr does not fail as expected. I will try harder now that I have a browser to test.

Show
Francis Avila added a comment - Tested that jsperf with Safari 6.0.5 using Browserstack, results are there. Note I could not reproduce CLJS-847 because str-tostr does not fail as expected. I will try harder now that I have a browser to test.
Hide
Francis Avila added a comment -

Still cannot reproduce CLJS-847.

This script includes my attempt at a minimum reproducible case. My theory was that certain types at higher jit levels would fail. I could not get any case to fail. I also tried flapping back and forth between types and using one type at a time, but still no failures.

In this thread I found this "minimal" script which the OP said he could get to fail reliably. I could not get it to fail. However the original post was from feb 15, 2013, which means the Safari he was using would have to be 6.0.2 or lower.

Hypotheses:

  1. This error does not affect 6.0.5 but maybe 6.0.4 or lower.
  2. BrowserStack's system somehow mitigates the bug, meaning we need a "real" Lion Safari 6.0.x to test.
  3. These tests only fail under the correct phase of the moon.

So I can code up a patch for str using the str-switch implementation (which is at least a bit faster on some browsers), but I have no idea if it may fail on Safari 6.0.5. I only know that it works so far. CLJS-801 should also be safe to reapply because the root cause of all issues is the implementation 1-arity of the cljs.core/str function.

I have also asked for Kevin's help back in CLJS-847. (Kevin was the original reporter of the Safari 6.0.x issue.)

Show
Francis Avila added a comment - Still cannot reproduce CLJS-847. This script includes my attempt at a minimum reproducible case. My theory was that certain types at higher jit levels would fail. I could not get any case to fail. I also tried flapping back and forth between types and using one type at a time, but still no failures. In this thread I found this "minimal" script which the OP said he could get to fail reliably. I could not get it to fail. However the original post was from feb 15, 2013, which means the Safari he was using would have to be 6.0.2 or lower. Hypotheses:
  1. This error does not affect 6.0.5 but maybe 6.0.4 or lower.
  2. BrowserStack's system somehow mitigates the bug, meaning we need a "real" Lion Safari 6.0.x to test.
  3. These tests only fail under the correct phase of the moon.
So I can code up a patch for str using the str-switch implementation (which is at least a bit faster on some browsers), but I have no idea if it may fail on Safari 6.0.5. I only know that it works so far. CLJS-801 should also be safe to reapply because the root cause of all issues is the implementation 1-arity of the cljs.core/str function. I have also asked for Kevin's help back in CLJS-847. (Kevin was the original reporter of the Safari 6.0.x issue.)
Hide
Francis Avila added a comment -

Made a jsperf of variadic cases. Chrome seems to really prefer IReduce to seq+stringbuilder for vectors (other collections not tested), but there is no difference or a small slowdown on other browsers. Not sure if it's worth it.

Also updated arity-one cases with a str using switch and never using toString. Nearly 50% slower than using switch or toString on Chrome, smaller on Safari.

In terms of safety str-switch-notostr does not use toString at all so is probably safer. I think str-switch will likely work too, though, and is significantly faster. However I haven't been able to get any TypeErrors in Safari 6.0.5 so it's anyone's guess.

I suggest something like this as a new str (which doesn't use reduce, but could):

(defn str
 ([x]
  (case (js* "typeof ~{}" x)
   "string" x
   "object" (if (identical? nil x) "" (.toString x))
   ("boolean" "number") (js-str x)
   "undefined" ""
   (js-str #js [x])))                                       ;; insurance against Safari 6.0.x TypeError bug.
 ([a b] (js* "~{}+~{}" (str a) (str b)))
 ([a b c] (js* "~{}+~{}+~{}" (str a) (str b) (str c)))
 ([a b c d] (js* "~{}+~{}+~{}+~{}" (str a) (str b) (str c) (str d)))
 ([a b c d & more]
  (loop [s (str a b c d) [e f g h & r] more]
   (let [s' (js* "~{}+~{}+~{}+~{}+~{}" s e f g h)]
    (if (nil? r)
     s'
     (recur s' r))))))
Show
Francis Avila added a comment - Made a jsperf of variadic cases. Chrome seems to really prefer IReduce to seq+stringbuilder for vectors (other collections not tested), but there is no difference or a small slowdown on other browsers. Not sure if it's worth it. Also updated arity-one cases with a str using switch and never using toString. Nearly 50% slower than using switch or toString on Chrome, smaller on Safari. In terms of safety str-switch-notostr does not use toString at all so is probably safer. I think str-switch will likely work too, though, and is significantly faster. However I haven't been able to get any TypeErrors in Safari 6.0.5 so it's anyone's guess. I suggest something like this as a new str (which doesn't use reduce, but could):
(defn str
 ([x]
  (case (js* "typeof ~{}" x)
   "string" x
   "object" (if (identical? nil x) "" (.toString x))
   ("boolean" "number") (js-str x)
   "undefined" ""
   (js-str #js [x])))                                       ;; insurance against Safari 6.0.x TypeError bug.
 ([a b] (js* "~{}+~{}" (str a) (str b)))
 ([a b c] (js* "~{}+~{}+~{}" (str a) (str b) (str c)))
 ([a b c d] (js* "~{}+~{}+~{}+~{}" (str a) (str b) (str c) (str d)))
 ([a b c d & more]
  (loop [s (str a b c d) [e f g h & r] more]
   (let [s' (js* "~{}+~{}+~{}+~{}+~{}" s e f g h)]
    (if (nil? r)
     s'
     (recur s' r))))))
Hide
Francis Avila added a comment - - edited

First cut of a possible patch that resolves this while not breaking CLJS-847. Should wait for confirmation that this does not break Safari 6.0.x.

Show
Francis Avila added a comment - - edited First cut of a possible patch that resolves this while not breaking CLJS-847. Should wait for confirmation that this does not break Safari 6.0.x.
Hide
Francis Avila added a comment -

Oops forgot tests.

Show
Francis Avila added a comment - Oops forgot tests.
Hide
Francis Avila added a comment -

Update in CLJS-847: original reporter was not able to reproduce his original bug report in Safari 6.0.x running in BrowserStack. This may be because of BrowserStack, but it's the best we have.

Given how hard this bug is to reproduce, how few people it affects, and how significant the performance regression is, I still think we should go back to the simple (if (nil? x) "" (.toString x)) implementation. However, you could also try the patch on this ticket (using a typeof switch), which at least (handwaving) might fix this bug in Safari 6.0.x and is a little faster than a simple .toString in Chrome and not much slower elsewhere. (The reason I think it might avoid this bug in Safari is that it avoids calling .toString on non-Objects.)

Show
Francis Avila added a comment - Update in CLJS-847: original reporter was not able to reproduce his original bug report in Safari 6.0.x running in BrowserStack. This may be because of BrowserStack, but it's the best we have. Given how hard this bug is to reproduce, how few people it affects, and how significant the performance regression is, I still think we should go back to the simple (if (nil? x) "" (.toString x)) implementation. However, you could also try the patch on this ticket (using a typeof switch), which at least (handwaving) might fix this bug in Safari 6.0.x and is a little faster than a simple .toString in Chrome and not much slower elsewhere. (The reason I think it might avoid this bug in Safari is that it avoids calling .toString on non-Objects.)
Hide
Antonin Hildebrand added a comment -

I wonder if you considered swapping str function at runtime during CLJS init phase.

Implement str function using plain .toString() call (original solution). And at startup check for Safari 6.0.x presence and optionally swap str for implementation wrapping .toString() call in a try-catch block silencing TypeError exceptions by falling back to Safari 6.0.x friendly .toString() alternative.

We would get correct semantics in all cases. And price would be just slower printing execution on Safari 6.0.x not on all systems.

Show
Antonin Hildebrand added a comment - I wonder if you considered swapping str function at runtime during CLJS init phase. Implement str function using plain .toString() call (original solution). And at startup check for Safari 6.0.x presence and optionally swap str for implementation wrapping .toString() call in a try-catch block silencing TypeError exceptions by falling back to Safari 6.0.x friendly .toString() alternative. We would get correct semantics in all cases. And price would be just slower printing execution on Safari 6.0.x not on all systems.
Hide
Mike Fikes added a comment -

cljs-890.patch no longer applies

Show
Mike Fikes added a comment - cljs-890.patch no longer applies

People

Vote (2)
Watch (6)

Dates

  • Created:
    Updated: