Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Tue, 20 May 2014 02:41:36 +0200
From: Szabolcs Nagy <nsz@...t70.net>
To: musl@...ts.openwall.com
Subject: Re: Re: thoughts on reallocarray, explicit_bzero?

* Andy Lutomirski <luto@...capital.net> [2014-05-19 15:08:11 -0700]:
> On 05/19/2014 09:25 AM, Szabolcs Nagy wrote:
> > i'd use a saturated multiplication, because malloc/realloc
> > are not the only places where overflowing size calculations
> > may cause problems and in such cases (size_t)-1 is just as
> > good as a failure and it can be added to your code without
> > portability issues
> > 
> > static size_t sizemul(size_t a, size_t b)
> > {
> > 	return b>1 && a>1 && a>-1/b ? -1 : a*b;
> > }
> 
> Before going nuts trying to optimize this, it may pay to write some
> good-enough helper and to use native compiler support for this, which is
> already available in Clang [1] and should be coming reasonably soon in
> gcc [2].

it's a shame that clang came up with this nonsese

they managed to add 18 new compiler specific builtins,
without actually addressing the practical issue:
easy to use overflow check of size_t multiplication..
(or checking arithmetics of various other non-builtin types)
(the several new multiprecision arithmetics builtins
are bad too but less problematic in practice)

they didn't make it easy to write backward compatible code
either: historically ifdef hackery was used to "detect"
builtins support using the __GNUC__ version macros, but clang
has incompatible versioning and builtins now making the use of
new builtins more painful

(meanwhile a lot of code has idiomatic overflow checks in iso c
which is not recognized by gcc or clang..
and many c parsing tools don't understand the fancy new builtins)

> I suspect that, on all reasonably platforms, if doublesize_t is the
> unsigned type that's twice as wide as size_t, then this isn't too bad
> either:
> 
> doublesize_t total = (doublesize_t)a * (doublesize_t)b;
> if (total > SIZE_MAX)
>   fail;
> 
> For quite a while, gcc has had a 128-bit integer type that works on
> 64-bit platforms, and gcc should always support a 64-bit type on 32-bit
> platforms.  On systems with widening multiply (e.g. x86), even if the
> optimizer doesn't detect the idiom, this is only a few cycles slower
> than the optimal code.

umm __int128 is only supported in gcc since 4.6 i think
(and even after that there were some related brokenness
in hacked toolchains so >=gcc-4.6 is not enough to check)

otherwise yes with doublesize_t it is easy to do
but the point was to do it in c

for doublesize_t you would need configure time checks..
or nasty ifdef hackery.. and in the end you still need the
fallback for implementations without such a type

(the code i showed can be included in any source file where
size_t is defined)

> [1]
> http://clang.llvm.org/docs/LanguageExtensions.html#checked-arithmetic-builtins
> [2] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61129

Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.