On Thu, May 1, 2008 at 5:10 PM, Daniel Berlin <[EMAIL PROTECTED]> wrote:
> > Sure, a malloc call has side-effects, so a DCE pass cannot just remove it.
> > Only struct-reorg knows that it has replaced all side-effects with others.
>
> Malloc only has side effects if the result is used.
>
>
> For example, LLVM will transform
>
>
> int main(int argc, char** argv){
> if(malloc(sizeof(int)) == NULL){ return 0; }
> else{ return 1; }
> }
>
> into return 1
>
> There is a thread going on in the llvm mailing lists about this right
> now, and so far all the text people can find in standards says this is
> okay (though I think susv2/POSIX says differently).
>
> Chris says:
> "
> LLVM should not (and does not, afaik) assume the malloc succeeds in
> general.
>
> If LLVM is able to eliminate all users of the malloc assuming the
> malloc succeeded (as in this case), then it is safe to assume the malloc
> returned success."
BTW, the argument goes something like this:
Transforming malloc into:
our_malloc(int size)
{
if (callsite == the one being compared against null)
{
static char buf[size] __attribute__(maximally_aligned)
return (void *)buf;
}
else
return malloc(size)
}
is legal, and would produce a result that is always non-null.
Thus, if you can eliminate all users of your malloc under the
assumption that malloc returns non-null, you can do so and delete the
malloc.
This also means that things like if (malloc(x) == 0x12345678) <do
something> would *not* be transformed into "if (1)", only those cases
where the behavior of malloc is checked for nullness.
