fix double cast to int on 32-bit

[remicollet]

2 tests are failing with GCC 8.2 on 32-bit build

zend_dval_to_lval preserves low bits  (32 bit long) [Zend/tests/dval_to_lval_32.phpt]
testing integer underflow (32bit) [Zend/tests/int_underflow_32bit.phpt]

The comment was added in 716da71 by @dstogov

[cmb69]

This fix looks good to me. Without it we would cast negative doubles to unsigned ints, which is undefined behavior.

However, this should be carefully tested on Windows which never defines ZEND_DVAL_TO_LVAL_CAST_OK. A quick check on VC15/x86 succeeds, but probably more checking should be done. cc @weltling

BTW: the language specification states:

If the remainder is less than zero, it is rounded towards infinity and X is added.

However, the current implementation rounds towards zero in this case. Is the implementation or the spec in error?

[dstogov]

The fix is right.
I guess, GCC-8.2 started to use SSE floating point math instruction in 32-bit mode by default.
Older versions used x87 FPU instruction, with more precise internal representation, that didn't need this addition.

Please, merge this.

[remicollet]

@dstogov thanks for the check.

Merged