ctor_copy_dtor test allocates ridiculous amounts of memory

Zack Weinberg zackw@Stanford.EDU
Sat Oct 7 13:00:00 GMT 2000

On Sat, Oct 07, 2000 at 12:34:31PM -0700, Benjamin Kosnik wrote:
> > where csz01 has been initialized to str01.max_size().  Flipping
> > through basic_string.h, it looks like that's a constant equal to
> > one-quarter the max virtual address space - 1GB in this case.
> right
> this is the full test block, in context you can see what the test is 
> doing:
> this test did end up passing for you, correct?

With resource limits applied, it catches std::bad_alloc and proceeds.
Without, it exhausts swap and gets killed.  (Or something else does.)

> > My question to you, then, is - is that a sensible upper limit?  And is
> > it sensible to actually create a string of that size during testing?
> as a pathological corner case, this is legitimate. 
> as for your second question: is it sensible to allocate this kind of 
> memory during testing? I've always thought it ok to test stuff like this 
> (ie pathological cases) but perhaps I'm off base.

I can see why you'd want to test this sort of thing, but I don't think
it's acceptable to allocate that amount of memory during testing.  I'd
have to have more than a gigabyte of swap for the test to succeed as
intended.  I don't think most people have that much.

You'll only catch bad_alloc if there are resource limits; if there
aren't, the OS will pretend it has infinite memory and then kill
random processes when it runs out of memory.  And I mean _random_.
It's depressingly common for it to shoot all the system daemons before
it hits the process causing the problem.  Or it might just deadlock
inside the kernel.  I don't run with resource limits on by default,
because I can't reliably predict how much memory the compiler will
want for any given run.

Also, I tend to run builds in the background while working.  When we
hit that test, everything else on the machine is paged out and the
computer takes ~20sec to accept a single keystroke.  Which is mostly
lame Linux paging heuristics, but still.

As a rule of thumb, I'd say that all execution tests should consume no
more than two megabytes of RAM.  Yes, two.  There's still people out
there with 8 and 16 megabyte systems.


More information about the Gcc-bugs mailing list