361312 - difference output on tamarin avmplus executable between Mac and Windows when running Math.pow(Infinity,-1)

Status: VERIFIED FIXED

(Tamarin Graveyard :: Virtual Machine, defect)

Description

[Dan Schaffer]

There are 5 failing testcases on mac on the compliance tests.  The reason is the following code differs between mac and windows.  

 print(Math.pow(Infinity,-1));
 print(Math.pow(-Infinity,-1));
 print(Math.pow(-Infinity,-2));

Windows Output:
 NaN
 NaN
 NaN

Mac Output:
 0
 0
 0

the pow.abc code is attached to this bug.

Comment 1

[Dan Schaffer]

Attachment: pow.as

Comment 2

[Dan Schaffer]

Attachment: pow.abc

Comment 3

[Brendan Eich [:brendan]]

See http://lxr.mozilla.org/mozilla/source/js/src/jsmath.c#301 or so:

#if !JS_USE_FDLIBM_MATH
    /*
     * Because C99 and ECMA specify different behavior for pow(),
     * we need to wrap the libm call to make it ECMA compliant.
     */
    if (!JSDOUBLE_IS_FINITE(y) && (x == 1.0 || x == -1.0)) {
        *rval = DOUBLE_TO_JSVAL(cx->runtime->jsNaN);
        return JS_TRUE;
    }
    /* pow(x, +-0) is always 1, even for x = NaN. */
    if (y == 0) {
        *rval = JSVAL_ONE;
        return JS_TRUE;
    }
#endif

/be

Comment 4

[Dan Schaffer]

yeah, the expected behavior should be:
pow(Infinity,-1)=0 

so tamarin windows is wrong and tamarin mac is correct. I will update the 5 failing testcases to expect = 0

tested on JavaScript 1.7 (mozilla/js) and it is correct.

Comment 5

[Werner Sharp]

Attachment: Fix for Math.pow problem plus 3 other changes

1.  MathUtils.cpp change to fix Math.pow edge cases
2.  AvmCore.cpp change to fix interpreted mode comparison of negative zero and zero (or other theoretical cases where an integer value ends up in a double atom type)
3. 64-bit fix for GCHiddenPointer
4. Some CodegenMIR.cpp optimizations for int/uint operations with mult/add/sub operators.  If result is int/uint and inputs are int/uint, we can safely do integer math instead of promoting to double all the time.

Comment 6

[Dan Schaffer]

Attachment: e15_8_2_13.as test fixes

patch for mozilla/js/tamarin/test/ecma3/Math/e15_8_2_13.as fixes 5 failing Windows testcases Math.pow(Infinity,-1)=0

Comment 7

[Chris Peyer]

Transfer Bug from Core->Tamarin to Tamarin Product.