Re: behavior-preserving optimization in C, was compiler bugs

George Neuner schrieb:


> With two weak definitions of 'x' and a weak definition of 'garbage',
> the compiler did not allocate any storage in the data segment. The
> linker did not object to having no global data (why should it?) and
> produced an executable that crashed with an illegal address
> reference because the system loader didn't assign any pages to the
> empty data segment.


An allocation is not required, because your code doesn't contain
memory reads from the segment. Nonetheless even an empty segment
should have a valid address at runtime.


> I checked the linker's map file to verify that no storage had been
> allocated for variables.


Did the data segment occur in that list at all?
A zero size were perfectly valid, but no segment definition at all would
be illegal.


IMO there must exist a reason, why the loader cannot patch the stored
'&x' offset with the address of the data segment. Most probably it
could not find the according segment symbol in it's load map, because
the linker failed to emit the *definition* of that segment (symbol),
while *referencing* it in the fixup table.


Otherwise, when the segment (symbol) could be found in the load map,
it would be harmless if it's address were unassigned or
illegal. Patching the '&x' offset with an invalid base address would
end up in garbage in the printf() output, but never in an read access
to that address.


> This happened with 4 different compilers under high optimization.


Again an indication of an buggy linker.


DoDi

Return to the comp.compilers page.
Search the comp.compilers archives again.