It seems I cannot commit this fix (You have insufficient Karma!)

Notice:
No regression detection on x86_64 (no failed test)
On ppc64 we now have only 16 failed tests (instead of 25)

You should be able to commit it now

The commit message says:
> On x86_64:
> (long)(double)9223372036854775807+1 = -9223372036854775808
> On ppc64
> (long)(double)9223372036854775807-1 = 9223372036854775807

How much is (long)(double)0x7fffffffffffffffL in that arch? Is it 0x7ffffffffffffffeL ?

The closest doubles near 0x7fffffffffffffff are 0x8000000000000000 and 0x7ffffffffffffffe. So basically ppp64 rounds down while x64 rounds up? If that's the case, there's nothing to "fix" in the code; you'll presumably get this behavior in a lot of other places, not just when casting LONG_MAX to double.

But in any case, I don't see how that is relevant to the patch. The patch changes a (long)dval cast to (long)(unsigned long)dval cast, which is a double-to-long cast, not a long-to-double cast.

> (long)(double)9223372036854775807-1 = 9223372036854775807

Ah sorry, I read that wrong. It can't obviously be rounding down, because the result seems larger; you're subtracting one here when you're adding 1 on top... The part about the closes doubles is wrong too, they're already spaced by 1024 by then.

In any case, your explanation in the commit seems both irrelevant and wrong. AS to the wrong part, in x86_64:

    printf("%ld\n",((long)(double)9223372036854775807)+1);
    double dd = 9223372036854775807;
    printf("%ld\n", ((long)dd) + 1);

gives

-9223372036854775808
-9223372036854775807

So the -9223372036854775808 result is just an artifact of the compiler doing away with the casts completely.

BTW, that's with optimization turned off. If I pass -O3 to gcc , it's still able to figure out it doesn't need to cast anything and the output is -9223372036854775808.

In any case, we should get to the bottom of this. I very much doubt the patch is wrong, but if we figure out the underling cause we may find out the problem is larger.

@cataphract see my latest message on php-cvs ML.

Really strange. C Test (with -O2) give the same result on x86_64 and ppc64 and seems to prove the fix is not needed.

No I need to understand why it improves the PHP result.

I can keep the test builder for some time, so will investigate on this...