[LLVMdev] optimization assumes malloc return is non-null

Jonathan S. Shapiro shap at eros-os.com
Wed Apr 30 14:20:24 PDT 2008

On Wed, 2008-04-30 at 15:25 -0400, David Vandevoorde wrote:
> On Apr 30, 2008, at 2:47 PM, Jonathan S. Shapiro wrote:
> > Daveed:
> >
> > Perhaps I am looking at the wrong version of the specification.  
> > Section
> > appears to refer to objects having volatile-qualified type.  
> > The
> > type of malloc() is not volatile qualified in the standard library
> > definition.
> ...malloc() is not specified to access a volatile  
> object, modify an object, or modifying a file (directly or  
> indirectly); i.e., it has no side effect from the language point of  
> view.


Good to know that I was looking at the correct section. I do not agree
that your interpretation follows the as-if rule, because I do not agree
with your interpretation of the C library specification of malloc().

The standard library specification of malloc() clearly requires that it
allocates storage, and that such allocation is contingent on storage
availability. Storage availability is, in turn, a function (in part) of
previous calls to malloc() and free(). Even if free() is not called, the
possibility of realloc() implies a need to retain per-malloc() state. In
either case, it follows immediately that malloc() is stateful, and
therefore that any conforming implementation of malloc() must modify at
least one object in the sense of the standard.

If I understand your position correctly, your justification for the
optimization is that the C library standard does not say in so many
words that malloc() modifies an object. I do not believe that any such
overt statement is required in order for it to be clear that malloc() is
stateful. The functional description of malloc() and free() clearly
cannot be satisfied under the C abstract machine without mutation of at
least one object.

Also, I do not read in the way that you do. Paragraph 2 defines
"side effect", but it does not imply any requirement that side effects
be explicitly annotated. What Paragraph 3 gives you is leeway to
optimize standard functions when you proactively know their behavior. A
standard library procedure is not side-effect free for optimization
purposes by virtue of the absence of annotation. It can only be treated
as side-effect free by virtue of proactive knowledge of the
implementation of the procedure. In this case, we clearly have knowledge
of the implementation of malloc, and that knowledge clearly precludes
any possibility that malloc is simultaneously side-effect free and

So it seems clear that this optimization is wrong. By my reading, not
only does the standard fail to justify it under paragraph 3, it
*prohibits* this optimization under under Paragraph 1 because
there is no conforming implementation that is side-effect free.

Exception: there are rare cases where, under whole-program optimization,
it is possible to observe that free() is not called, that there is an
upper bound on the number of possible calls to malloc() and also an
upper bound on the total amount of storage allocated. In this very
unusual case, the compiler can perform a hypothetical inlining of the
known implementation of malloc and then do partial evaluation to
determine that no heap size tracking is required. If so, it can then
legally perform the optimization that is currently being done.

But I don't think that the current compiler is actually doing that
analysis in this case...

> > In general, calls to procedures that are outside the current unit of
> > compilation are presumed to involve side effects performed in the body
> > of the external procedure (at least in the absence of annotation).
> That may often be done in practice, but it's not a language  
> requirement.  In particular, for standard library functions (like  
> malloc) an optimizer can exploit the known behavior of the function.

I disagree. In the malloc() case, the known behavior is side effecting.
In the general case, the compiler cannot assume side-effect freedom
unless it can prove it, and in the absence of implementation knowledge
the standard requires conservatism.

> The same concept exists in C++, and we often refer to it as the "as  
> if" rule; i.e., implementations can do all kinds of things, as long as  
> the effect is "as if" specified by the abstract machine.

Yes. But the C++ version of this is quite different, because in any
situation where new would return NULL it would instead be raising an out
of memory exception. In consequence, the optimization is correct for
operator new whether or not operator new is side effecting.

Setting the matter of the standard entirely aside, the currently
implemented behavior deviates so blatantly from common sense and
universal convention that it really must be viewed as a bug.

Finally, I strongly suspect that LLVM will fail the standard conformance
suites so long as this optimization is retained.


More information about the llvm-dev mailing list