threads / patch / 55897

patchadd_pending_object_with_path(): work around "gcc -O3" complaint

Subject: [PATCH] add_pending_object_with_path(): work around "gcc -O3" complaint

## tl;dr

3 messages between Jun 10, 2021 and Jun 11, 2021. Diffs are folded; open one to read it.

replies: 2people: 2as markdown or json

Jeff King· Jun 10, 2021, 13:06 UTC · lore

When compiling with -O3, some gcc versions (10.2.1 here) complain about an out-of-bounds subscript:

  revision.c: In function ‘do_add_index_objects_to_pending’:
  revision.c:321:22: error: array subscript [1, 2147483647] is outside array bounds of ‘char[1]’ [-Werror=array-bounds]
    321 |   if (0 < len && name[len] && buf.len)
        |                  ~~~~^~~~~

The "len" parameter here comes from calling interpret_branch_name(), which intends to return the number of characters of "name" it parsed.

But the compiler doesn't realize this. It knows the size of the empty string "name" passed in from do_add_index_objects_to_pending(), but it has no clue that the "len" we get back will be constrained to "0" in that case.

And I don't think the warning is telling us about some subtle or clever bug. The implementation of interpret_branch_name() is in another file entirely, and the compiler can't see it (you can even verify there is no clever LTO going on by replacing it with "return 0" and still getting the warning).

We can work around this by replacing our "did we hit the trailing NUL" subscript dereference with a length check. We do not even have to pay the cost for an extra strlen(), as we can pass our new length into interpret_branch_name(), which was converting our "0" into a call to strlen() anyway.

Signed-off-by: Jeff King <peff@peff.net>
---
Together with ab/trace2-squelch-gcc-warning and the ll-merge fix I
posted in [0], this makes the whole project compile cleanly with "-O3"
for me.

I suspect after this patch we could drop the "if len is 0, call strlen() as a courtesy" code from interpret_branch_name() entirely. But the call stack gets pretty hairy (some of these len values come directly from the get_oid functions directly!) so I wasn't brave enough to follow through on that.

It's even possible the current code in interpret_branch_name() could be buggy here, if some caller is using a non-NUL-terminated buffer or a portion of a string, and we mistake a true "0" size for a request to call strlen(). That might be worth poking at, but I didn't want to spend time on it just now.

[0] https://lore.kernel.org/git/YMIKwsEFnkqz6PWa@coredump.intra.peff.net/
 revision.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
Show changes to revision.c +3 −2
diff --git a/revision.c b/revision.c
index 8140561b6c..cddd0542a6 100644
--- a/revision.c
+++ b/revision.c
@@ -316,9 +316,10 @@ static void add_pending_object_with_path(struct rev_info *revs,
 		revs->no_walk = 0;
 	if (revs->reflog_info && obj->type == OBJ_COMMIT) {
 		struct strbuf buf = STRBUF_INIT;
-		int len = interpret_branch_name(name, 0, &buf, &options);
+		size_t namelen = strlen(name);
+		int len = interpret_branch_name(name, namelen, &buf, &options);
 
-		if (0 < len && name[len] && buf.len)
+		if (0 < len && len < namelen && buf.len)
 			strbuf_addstr(&buf, name + len);
 		add_reflog_for_walk(revs->reflog_info,
 				    (struct commit *)obj,

-- 
2.32.0.529.g079a794268
Jeff King· Jun 10, 2021, 13:11 UTC · re: Jeff King · lore

Re: [PATCH] add_pending_object_with_path(): work around "gcc -O3" complaint

On Thu, Jun 10, 2021 at 09:06:44AM -0400, Jeff King wrote:
Show 9 quoted lines
> We can work around this by replacing our "did we hit the trailing NUL"
> subscript dereference with a length check. We do not even have to pay
> the cost for an extra strlen(), as we can pass our new length into
> interpret_branch_name(), which was converting our "0" into a call to
> strlen() anyway.
> [...]
> -		if (0 < len && name[len] && buf.len)
> +		if (0 < len && len < namelen && buf.len)
>  			strbuf_addstr(&buf, name + len);

I guess another option would be to drop the check entirely. It is only protecting us from calling strbuf_addstr() with an empty string, which is a noop anyway (it would not even cause a useless allocation, since we know that buf is non-empty, and that it won't need to grow).

I think I still prefer my original solution, though.
-Peff
Junio C Hamano· Jun 11, 2021, 03:52 UTC · re: Jeff King · lore

Re: [PATCH] add_pending_object_with_path(): work around "gcc -O3" complaint

Jeff King <peff@peff.net> writes:
Show 18 quoted lines
> On Thu, Jun 10, 2021 at 09:06:44AM -0400, Jeff King wrote:
>
>> We can work around this by replacing our "did we hit the trailing NUL"
>> subscript dereference with a length check. We do not even have to pay
>> the cost for an extra strlen(), as we can pass our new length into
>> interpret_branch_name(), which was converting our "0" into a call to
>> strlen() anyway.
>> [...]
>> -		if (0 < len && name[len] && buf.len)
>> +		if (0 < len && len < namelen && buf.len)
>>  			strbuf_addstr(&buf, name + len);
>
> I guess another option would be to drop the check entirely. It is only
> protecting us from calling strbuf_addstr() with an empty string, which
> is a noop anyway (it would not even cause a useless allocation, since we
> know that buf is non-empty, and that it won't need to grow).
>
> I think I still prefer my original solution, though.

It may still work without the guard but it is not apparent to the readers if it just happens to work by accident or by design. At least the guard makes it clear what is going on, I would think.

← back to recent threads