Re: [Valgrind-users] POW() function behaving different when running in valgrind?

[John R. <jr...@bi...>]

> Duplicating the discussion from the V8 user group:
>
>     https://groups.google.com/forum/#!topic/v8-users/qRorKvYNq24
>
>     On Fri, Dec 4, 2015 at 3:26 PM, Wilfried Gösgens <do...@gm...> wrote:
>      > Hi everyone
>      > Using the regular d8 shell:
>      >
>      >     # ./d8
>      >     V8 version 4.8.271.5
>      >     d8> Math.pow(0, 0.1);
>      >     0
>      >     d8>
>      >
>      > gives proper results.
>      > But using d8 in valgrind gives:
>      >
>      >     /usr/bin/valgrind.bin --log-file=/tmp/valgrindlog.%p
>      > --show-reachable=yes --leak-check=full ./d8
>      >     Math.pow(0, 0.1);
>      >     V8 version 4.8.271.5
>      >     d8> NaN
>      >     d8>

> Does valgrind replace mathematical functions?

No, but valgrind forces round-to-nearest 64-bit result
for every floating-point operation: no extra precision
nor exponent range for 80-bit extended opcodes,
and no directed rounding (truncate, ceil, floor).  This is
well documented.  Thus if the user code relies on 80-bit mode
or rounding other than round-to-nearest, then "bad results"
may ensue.  The bug is in V8 not diagnosing the actual environment
before proceeding.  Never run a laboratory experiment
for quantitative results without first calibrating the instruments.