asan: Instrument <retval> stores in callees rather than callers [PR112709]

Hi!

asan currently instruments since PR69276 r6-6758 fix calls which store
the return value into memory on the caller side, before the call it
verifies the memory is writable.
Now PR112709 where we ICE on trying to instrument such calls made me
think about whether that is what we want to do.

There are 3 different cases.

One is when a function returns an aggregate which is passed e.g. in
registers, say like struct S { int a[4]; }; returning on x86_64.
That would be ideally instrumented in between the actual call and
storing of the aggregate into memory, but asan currently mostly
works as a GIMPLE pass and arranging for the instrumentation to happen
at that spot would be really hard.  We could diagnose after the call
but generally asan attempts to diagnose stuff before something is
overwritten rather than after, or keep the current behavior (that is
what this patch does, which has the disadvantage that it can complain
about UB even for functions which never return and so never actually store,
and doesn't check whether the memory wasn't e.g. poisoned during the call)
or could e.g. instrument both before and after the call (that would have
the disadvantage the current state has but at least would check post-factum
the store again afterwards).

Another case is when a function returns an aggregate through a hidden
reference, struct T { int a[128]; }; on x86_64 or even the above struct S
on ia32 as example.  In the actual program such stores happen when storing
something to <retval> or its parts in the callee, because <retval> there
expands to *hidden_retval.  So, IMHO we should instrument those in the
callee rather than caller, that is where the writes are and we can do that
easily.  This is what the patch below does.

And the last case is for builtins/internal functions.  Usually those don't
return aggregates, but in case they'd do and can be expanded inline, it is
better to instrument them in the caller (as before) rather than not
instrumenting the return stores at all.

I had to tweak the expected output on the PR69276 testcase, because
with the patch it keeps previous behavior on x86_64 (structure returned
in registers, stored in the caller, so reported as UB in A::A()),
while on i686 it changed the behavior and is reported as UB in the
vnull::operator vec which stores the structure, A::A() is then a frame
above it in the backtrace.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2024-03-12  Jakub Jelinek  <jakub@redhat.com>

	PR sanitizer/112709
	* asan.cc (has_stmt_been_instrumented_p): Don't instrument call
	stores on the caller side unless it is a call to a builtin or
	internal function or function doesn't return by hidden reference.
	(maybe_instrument_call): Likewise.
	(instrument_derefs): Instrument stores to RESULT_DECL if
	returning by hidden reference.

	* gcc.dg/asan/pr112709-1.c: New test.
	* g++.dg/asan/pr69276.C: Adjust expected output for some targets.


	Jakub

On Tue, 12 Mar 2024, Jakub Jelinek wrote:

> Hi!
> 
> asan currently instruments since PR69276 r6-6758 fix calls which store
> the return value into memory on the caller side, before the call it
> verifies the memory is writable.
> Now PR112709 where we ICE on trying to instrument such calls made me
> think about whether that is what we want to do.
> 
> There are 3 different cases.
> 
> One is when a function returns an aggregate which is passed e.g. in
> registers, say like struct S { int a[4]; }; returning on x86_64.
> That would be ideally instrumented in between the actual call and
> storing of the aggregate into memory, but asan currently mostly
> works as a GIMPLE pass and arranging for the instrumentation to happen
> at that spot would be really hard.  We could diagnose after the call
> but generally asan attempts to diagnose stuff before something is
> overwritten rather than after, or keep the current behavior (that is
> what this patch does, which has the disadvantage that it can complain
> about UB even for functions which never return and so never actually store,
> and doesn't check whether the memory wasn't e.g. poisoned during the call)
> or could e.g. instrument both before and after the call (that would have
> the disadvantage the current state has but at least would check post-factum
> the store again afterwards).
> 
> Another case is when a function returns an aggregate through a hidden
> reference, struct T { int a[128]; }; on x86_64 or even the above struct S
> on ia32 as example.  In the actual program such stores happen when storing
> something to <retval> or its parts in the callee, because <retval> there
> expands to *hidden_retval.  So, IMHO we should instrument those in the
> callee rather than caller, that is where the writes are and we can do that
> easily.  This is what the patch below does.
> 
> And the last case is for builtins/internal functions.  Usually those don't
> return aggregates, but in case they'd do and can be expanded inline, it is
> better to instrument them in the caller (as before) rather than not
> instrumenting the return stores at all.
> 
> I had to tweak the expected output on the PR69276 testcase, because
> with the patch it keeps previous behavior on x86_64 (structure returned
> in registers, stored in the caller, so reported as UB in A::A()),
> while on i686 it changed the behavior and is reported as UB in the
> vnull::operator vec which stores the structure, A::A() is then a frame
> above it in the backtrace.
> 
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

OK.

Thanks,
Richard.

> 2024-03-12  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR sanitizer/112709
> 	* asan.cc (has_stmt_been_instrumented_p): Don't instrument call
> 	stores on the caller side unless it is a call to a builtin or
> 	internal function or function doesn't return by hidden reference.
> 	(maybe_instrument_call): Likewise.
> 	(instrument_derefs): Instrument stores to RESULT_DECL if
> 	returning by hidden reference.
> 
> 	* gcc.dg/asan/pr112709-1.c: New test.
> 	* g++.dg/asan/pr69276.C: Adjust expected output for some targets.
> 
> --- gcc/asan.cc.jj	2024-03-06 09:35:04.132894608 +0100
> +++ gcc/asan.cc	2024-03-11 13:49:58.931045179 +0100
> @@ -1372,7 +1372,12 @@ has_stmt_been_instrumented_p (gimple *st
>  	  return true;
>  	}
>      }
> -  else if (is_gimple_call (stmt) && gimple_store_p (stmt))
> +  else if (is_gimple_call (stmt)
> +	   && gimple_store_p (stmt)
> +	   && (gimple_call_builtin_p (stmt)
> +	       || gimple_call_internal_p (stmt)
> +	       || !aggregate_value_p (TREE_TYPE (gimple_call_lhs (stmt)),
> +				      gimple_call_fntype (stmt))))
>      {
>        asan_mem_ref r;
>        asan_mem_ref_init (&r, NULL, 1);
> @@ -2751,7 +2756,9 @@ instrument_derefs (gimple_stmt_iterator
>      return;
>  
>    poly_int64 decl_size;
> -  if ((VAR_P (inner) || TREE_CODE (inner) == RESULT_DECL)
> +  if ((VAR_P (inner)
> +       || (TREE_CODE (inner) == RESULT_DECL
> +	   && !aggregate_value_p (inner, current_function_decl)))
>        && offset == NULL_TREE
>        && DECL_SIZE (inner)
>        && poly_int_tree_p (DECL_SIZE (inner), &decl_size)
> @@ -3023,7 +3030,11 @@ maybe_instrument_call (gimple_stmt_itera
>      }
>  
>    bool instrumented = false;
> -  if (gimple_store_p (stmt))
> +  if (gimple_store_p (stmt)
> +      && (gimple_call_builtin_p (stmt)
> +	  || gimple_call_internal_p (stmt)
> +	  || !aggregate_value_p (TREE_TYPE (gimple_call_lhs (stmt)),
> +				 gimple_call_fntype (stmt))))
>      {
>        tree ref_expr = gimple_call_lhs (stmt);
>        instrument_derefs (iter, ref_expr,
> --- gcc/testsuite/gcc.dg/asan/pr112709-1.c.jj	2024-03-11 13:59:15.300408140 +0100
> +++ gcc/testsuite/gcc.dg/asan/pr112709-1.c	2024-03-11 13:59:58.626813417 +0100
> @@ -0,0 +1,52 @@
> +/* PR sanitizer/112709 */
> +/* { dg-do compile } */
> +/* { dg-options "-fsanitize=address -O2" } */
> +
> +struct S { char c[1024]; };
> +int foo (int);
> +
> +__attribute__((returns_twice, noipa)) struct S
> +bar (int x)
> +{
> +  (void) x;
> +  struct S s = {};
> +  s.c[42] = 42;
> +  return s;
> +}
> +
> +void
> +baz (struct S *p)
> +{
> +  foo (1);
> +  *p = bar (0);
> +}
> +
> +void
> +qux (int x, struct S *p)
> +{
> +  if (x == 25)
> +    x = foo (2);
> +  else if (x == 42)
> +    x = foo (foo (3));
> +  *p = bar (x);
> +}
> +
> +void
> +corge (int x, struct S *p)
> +{
> +  void *q[] = { &&l1, &&l2, &&l3, &&l3 };
> +  if (x == 25)
> +    {
> +    l1:
> +      x = foo (2);
> +    }
> +  else if (x == 42)
> +    {
> +    l2:
> +      x = foo (foo (3));
> +    }
> +l3:
> +  *p = bar (x);
> +  if (x < 4)
> +    goto *q[x & 3];
> +}
> --- gcc/testsuite/g++.dg/asan/pr69276.C.jj	2020-01-14 20:02:46.691611212 +0100
> +++ gcc/testsuite/g++.dg/asan/pr69276.C	2024-03-12 09:09:05.901446463 +0100
> @@ -35,4 +35,5 @@ int main()
>  }
>  
>  /* { dg-output "ERROR: AddressSanitizer: heap-buffer-overflow.*(\n|\r\n|\r)" } */
> -/* { dg-output "    #0 0x\[0-9a-f\]+ +in A::A()" } */
> +/* { dg-output "    #0 0x\[0-9a-f\]+ +in (A::A\\\(\\\)|vnull::operator vec\\\(\\\).*(\n|\r\n|\r)" } */
> +/* { dg-output "    #1 0x\[0-9a-f\]+ +in A::A\\\(\\\))" } */
> 
> 	Jakub
> 
>