Andreas Bogk <[EMAIL PROTECTED]> writes:
> It is my fear that the existing behaviour of gcc when used without
> -fwrapv breaks a lot of code out there that was written with the
> implicit assumption that signed ints would overflow the way the
> underlying machine integers do. More importantly, some of the code that
> breaks is in checks against integer overflows, and thus relevant to
> security: breaking the checks means opening vulnerabilities. A number
> of IT security professionals shares my view of things here.
I am not interested in starting another lengthy discussion on this
general issue.
I just want to report that I have a working patch to generate warnings
every time gcc modifies code relying on the fact that signed overflow
is undefined, except for cases where signed loop indexes are assumed
to not wrap around. I plan to start submitting this patch soon.
My current intention, subject of course to the opinions of other
maintainers, is to implement a -fstrict-overflow option, along the
lines of -fstrict-aliasing. This will be enabled by default at -O2,
as is the case for -fstrict-aliasing. -fno-strict-overflow will not
be the same as -fwrapv, but it will inhibit optimizations which are
only valid if signed overflow is undefined. The new -Wstrict-overflow
warning will issue warnings for each case where gcc assumes that
signed overflow is undefined.
To be clear, this -Wstrict-overflow option generates a lot of false
positives. That is because there are a lot of cases where gcc will
optimize code in ways that would fail if the numbers involved were
close to INT_MAX or INT_MIN. For example, when compiling the gcc/*.c
files, I get 62 warnings.
I would appreciate any comments about this plan.
Some comments about those 62 warnings. First, each one of those
warnings is a case where gcc implemented an optimization which could
not have been implemented when using -fwrapv. Now, a couple of
examples.
In bb-reorder.c at line 266 we find this:
find_traces_1_round (REG_BR_PROB_BASE * branch_threshold[i] / 1000,
max_entry_frequency * exec_threshold[i] / 1000,
count_threshold, traces, n_traces, i, &heap,
number_of_rounds);
REG_BR_PROB_BASE is a #define macro with the value 10000. gcc
converts (10000 * branch_threshold[i] / 1000) into (10 *
branch_threshold[i]). This minor optimization is only possible
because signed overflow is undefined. If branch_threshold[i] had the
value (INT_MAX / 10 - 1), say, the optimization will cause the
expression to have a different result (assuming standards
twos-complement arithmetic wrapping). I see 17 warnings of this
form.
In expmed.c we find several times (size - 1 < BITS_PER_WORD). gcc
converts this to (size <= BITS_PER_WORD). This optimization is only
possible because signed overflow is undefined. If size has the value
INT_MIN, the optimization will cause the expression to have a
different result. I see 23 warnings of this form.
In fact, all the warnings on gcc/*.c have the same general form: a
mathematical simplification which is only permissible if the values do
not wrap around.
Ian
