On 09/07, Michael Haggerty wrote:
> On Wed, Sep 6, 2017 at 3:26 AM, Junio C Hamano <gitster@xxxxxxxxx> wrote:
> > Thomas Gummerer <t.gummerer@xxxxxxxxx> writes:
> >
> >> gcc on arch linux (version 7.1.1) warns that a NULL argument is passed
> >> as the second parameter of memcpy.
> >> [...]
> >
> > It is hugely annoying to see a halfway-intelligent compiler forces
> > you to add such pointless asserts.
> >
> > The only way the compiler could error on this is by inferring the
> > fact that new_sha1/old_sha1 could be NULL by looking at the callsite
> > in ref_transaction_update() where these are used as conditionals to
> > set HAVE_NEW/HAVE_OLD that are passed. Even if the compiler were
> > doing the whole-program analysis, the other two callsites of the
> > function pass the address of oid.hash[] in an oid structure so it
> > should know these won't be NULL.
> >
> > [...]
> >
> > I wonder if REF_HAVE_NEW/REF_HAVE_OLD are really needed in these
> > codepaths, though. Perhaps we can instead declare !!new_sha1 means
> > we have the new side and rewrite the above part to
> >
> > if (new_sha1)
> > hashcpy(update->new_oid.hash, new_sha1);
> >
> > without an extra and totally pointless assert()?
>
> The ultimate reason for those flags is that `struct ref_update` embeds
> `new_oid` and `old_oid` directly in the struct, so there is no way to
> set it to "NULL". (The `is_null_sha1` value is used for a different
> purpose.) So those flags keep track of whether the corresponding value
> is specified or absent.
>
> Four of the five callers of `ref_transaction_add_update()` are
> constructing a new `ref_update` from an old one. They currently don't
> have to look into `flags`; they just pass it on (possibly changing a
> bit or two). Implementing your proposal would oblige those callers to
> change from something like
Thanks for the explanation!
> > new_update = ref_transaction_add_update(
> > transaction, "HEAD",
> > update->flags | REF_LOG_ONLY | REF_NODEREF,
> > update->new_oid.hash, update->old_oid.hash,
> > update->msg);
>
> to
>
> > new_update = ref_transaction_add_update(
> > transaction, "HEAD",
> > update->flags | REF_LOG_ONLY | REF_NODEREF,
> > (update->flags & REF_HAVE_NEW) ? update->new_oid.hash : NULL,
> > (update->flags & REF_HAVE_OLD) ? update->old_oid.hash : NULL,
> > update->msg);
>
> It's not the end of the world, but it's annoying.
> `ref_transaction_add_update()` was meant to be a low-level,
> low-overhead way of allocating a `struct ref_update` and add it to a
> transaction.
>
> Another solution (also annoying, but maybe a tad less so) would be to
> change the one iffy caller, `ref_transaction_update()`, to pass in a
> pointer to the null OID for `new_sha1` and `old_sha1` when the
> corresponding flags are turned off. That value would never be looked
> at, but it would hopefully reassure gcc.
>
> I did just realize one thing: `ref_transaction_update()` takes `flags`
> as an argument and alters it using
>
> > flags |= (new_sha1 ? REF_HAVE_NEW : 0) | (old_sha1 ? REF_HAVE_OLD : 0);
>
> Perhaps gcc is *more* intelligent than we give it credit for, and is
> actually worried that the `flags` argument passed in by the caller
> might *already* have one of these bits set. In that case
> `ref_transaction_add_update()` would indeed be called incorrectly.
> Does the warning go away if you change that line to
>
> > if (new_sha1)
> > flags |=REF_HAVE_NEW;
> > else
> > flags &= ~REF_HAVE_NEW;
> > if (old_sha1)
> > flags |=REF_HAVE_OLD;
> > else
> > flags &= ~REF_HAVE_OLD;
>
> ?
Indeed that fixes it, great catch! gcc is indeed smarter than we gave
it credit for, this makes a lot of sense.
Interestingly stripping away the flags fixes the compiler warning:
diff --git a/refs.c b/refs.c
index ba22f4acef..2e6871beac 100644
--- a/refs.c
+++ b/refs.c
@@ -921,6 +921,9 @@ int ref_transaction_update(struct ref_transaction *transaction,
return -1;
}
+ flags &= ~REF_HAVE_NEW;
+ flags &= ~REF_HAVE_OLD;
+
flags |= (new_sha1 ? REF_HAVE_NEW : 0) | (old_sha1 ? REF_HAVE_OLD : 0);
ref_transaction_add_update(transaction, refname, flags,
while checking that the flags are not there in the first place still
leaves the compiler warning (whether I use "die()" or just "return -1"
doesn't matter in this case):
diff --git a/refs.c b/refs.c
index ba22f4acef..62ff283755 100644
--- a/refs.c
+++ b/refs.c
@@ -921,6 +921,9 @@ int ref_transaction_update(struct ref_transaction *transaction,
return -1;
}
+ if ((flags & REF_HAVE_NEW) != 0 || (flags & REF_HAVE_OLD) != 0)
+ die("BUG: passed invalid flag to ref_transaction_update");
+
flags |= (new_sha1 ? REF_HAVE_NEW : 0) | (old_sha1 ? REF_HAVE_OLD : 0);
ref_transaction_add_update(transaction, refname, flags,
> This might be a nice change to have anyway, to isolate
> `ref_transaction_update()` from mistakes by its callers. For that
> matter, one might want to be even more selective about what bits are
> allowed in the `flags` argument to `ref_transaction_update()`'s
> callers:
>
> > flags &= REF_ALLOWED_FLAGS; /* value would need to be determined */
Interesting idea, would we just strip out the flags in this case, or
should we check that the flags are not there in the first place and
error out if they are?
I guess with the second option we would make sure that every caller
cleans up the flags before calling the function in the first place,
but also have a small risk of missing something if we don't test a
particular codepath. I stripping out the flags might actually fix the
compiler warning as well, so that might be the better option here.
I'm happy to try and patch this based on the suggestion above, but
most likely it won't happen before next week, as I'm traveling over
the weekend :)
> Michael
