Re: [PATCH 2/2] built-in rebase: reinstate `checkout -q` behavior where appropriate
Junio C Hamano writes: > Johannes Schindelin writes: > >>> On the other hand, for a file-scope static that is likely to stay as a >>> non-API function but a local helper, it may not be worth it. >> >> Oh, do you think that `reset_head()` is likely to stay as non-API >> function? I found myself in the need of repeating this tedious >> unpack_trees() dance quite a few times over the past two years, and it is >> *always* daunting because the API is *that* unintuitive. >> >> So I *do* hope that `reset_head()` will actually be moved to reset.[ch] >> eventually, and callsites e.g. in `sequencer.c` will be converted from >> calling `unpack_trees()` to calling `reset_head()` instead. > > As long as the public API function is well thought out, of course it > is OK. Ideally, builtin/reset.c can detect when it is making a hard > reset to HEAD and call that same function. Which may or may not > mean you would also want to tell it to record ORIG_HEAD and remove > MERGE_HEAD and others), perhaps with another bit in that flag word. Nah, forget about that one. It sometimes does branch switching and sometimes does hard reset, which looks like a strange chimera. We shouldn't burden it further by adding "while at it remove cruft, as 'reset --hard' needs to do that" to it.
Re: [PATCH 2/2] built-in rebase: reinstate `checkout -q` behavior where appropriate
Johannes Schindelin writes: >> On the other hand, for a file-scope static that is likely to stay as a >> non-API function but a local helper, it may not be worth it. > > Oh, do you think that `reset_head()` is likely to stay as non-API > function? I found myself in the need of repeating this tedious > unpack_trees() dance quite a few times over the past two years, and it is > *always* daunting because the API is *that* unintuitive. > > So I *do* hope that `reset_head()` will actually be moved to reset.[ch] > eventually, and callsites e.g. in `sequencer.c` will be converted from > calling `unpack_trees()` to calling `reset_head()` instead. As long as the public API function is well thought out, of course it is OK. Ideally, builtin/reset.c can detect when it is making a hard reset to HEAD and call that same function. Which may or may not mean you would also want to tell it to record ORIG_HEAD and remove MERGE_HEAD and others), perhaps with another bit in that flag word.
Re: [PATCH 2/2] built-in rebase: reinstate `checkout -q` behavior where appropriate
Hi Junio, On Mon, 12 Nov 2018, Junio C Hamano wrote: > Johannes Schindelin writes: > > >> > static int reset_head(struct object_id *oid, const char *action, > >> > - const char *switch_to_branch, int detach_head, > >> > + const char *switch_to_branch, > >> > + int detach_head, int reset_hard, > >> > >> It might be worth switching to a single flag variable here. It would > >> make calls like this: > >> > >> > -if (reset_head(>object.oid, "checkout", NULL, 1, > >> > +if (reset_head(>object.oid, "checkout", NULL, 1, > >> > 0, > >> > NULL, msg.buf)) > >> > >> a little more self-documenting (if a little more verbose). > > > > I thought about that, but for just two flags? Well, let me try it and see > > whether I like the result ;-) > > My rule of thumb is that repeating three times is definitely when we > should consolidate separate ones into a single flag word, and twice > is a borderline. > > For two-time repetition, it is often worth fixing when somebody > notices it during the review, as that is a sign that repetition, > even a minor one, disturbed a reader enough to point out. That's my thought exactly, hence I looked into it. The end result is actually easier to read, in particular the commit that re-introduces the `reset --hard` behavior: it no longer touches *all* callsites of `reset_head()` but only the relevant ones. > On the other hand, for a file-scope static that is likely to stay as a > non-API function but a local helper, it may not be worth it. Oh, do you think that `reset_head()` is likely to stay as non-API function? I found myself in the need of repeating this tedious unpack_trees() dance quite a few times over the past two years, and it is *always* daunting because the API is *that* unintuitive. So I *do* hope that `reset_head()` will actually be moved to reset.[ch] eventually, and callsites e.g. in `sequencer.c` will be converted from calling `unpack_trees()` to calling `reset_head()` instead. v2 on the way, Dscho > So I am OK either way, slightly in favor of fixing it while we > remember. > > > >> This one could actually turn into: > >> > >> ret = error(...); > >> goto leave_reset_head; > >> > >> now. We don't have to worry about an uninitialized desc.buffer anymore > >> (as I mentioned in the previous email), because "nr" would be 0. > >> > >> It doesn't save any lines, though (but maybe having a single > >> cleanup/exit point would make things easier to read; I dunno). > > > > But you're right, of course. Consistency makes for easier-to-read code. > > Yup, that does sound good. > > Thanks. >
Re: [PATCH 2/2] built-in rebase: reinstate `checkout -q` behavior where appropriate
Hi Peff, On Fri, 9 Nov 2018, Jeff King wrote: > On Fri, Nov 09, 2018 at 06:21:41PM +0100, Johannes Schindelin wrote: > > > Actually, you got me thinking about the desc.buffer. And I think there is > > one corner case where it could cause a problem: `struct tree_desc desc[2]` > > does not initialize the buffers to NULL. And what if > > fill_tree_descriptor() function returns NULL? Then the buffer is still > > uninitialized. > > > > In practice, our *current* version of fill_tree_descriptor() never returns > > NULL if the oid parameter is non-NULL. It would die() in the error case > > instead (bad!). So to prepare for a less bad version, I'd rather > > initialize the `desc` array and be on the safe (and easier to reason > > about) side. > > Yeah, I agree with all of that. > > One solution would just be to increment only after success: > > if (fill_tree_descriptor([nr], ..) < 0) > goto error; > nr++; /* now we know it's valid! */ > > But there are lots of alternatives. :) True. I simply prefer to initialize it and be done with it. ;-) Ciao, Dscho
Re: [PATCH 2/2] built-in rebase: reinstate `checkout -q` behavior where appropriate
Johannes Schindelin writes: >> > static int reset_head(struct object_id *oid, const char *action, >> > -const char *switch_to_branch, int detach_head, >> > +const char *switch_to_branch, >> > +int detach_head, int reset_hard, >> >> It might be worth switching to a single flag variable here. It would >> make calls like this: >> >> > - if (reset_head(>object.oid, "checkout", NULL, 1, >> > + if (reset_head(>object.oid, "checkout", NULL, 1, 0, >> >NULL, msg.buf)) >> >> a little more self-documenting (if a little more verbose). > > I thought about that, but for just two flags? Well, let me try it and see > whether I like the result ;-) My rule of thumb is that repeating three times is definitely when we should consolidate separate ones into a single flag word, and twice is a borderline. For two-time repetition, it is often worth fixing when somebody notices it during the review, as that is a sign that repetition, even a minor one, disturbed a reader enough to point out. On the other hand, for a file-scope static that is likely to stay as a non-API function but a local helper, it may not be worth it. So I am OK either way, slightly in favor of fixing it while we remember. >> This one could actually turn into: >> >> ret = error(...); >> goto leave_reset_head; >> >> now. We don't have to worry about an uninitialized desc.buffer anymore >> (as I mentioned in the previous email), because "nr" would be 0. >> >> It doesn't save any lines, though (but maybe having a single >> cleanup/exit point would make things easier to read; I dunno). > > But you're right, of course. Consistency makes for easier-to-read code. Yup, that does sound good. Thanks.
Re: [PATCH 2/2] built-in rebase: reinstate `checkout -q` behavior where appropriate
On Fri, Nov 09, 2018 at 06:21:41PM +0100, Johannes Schindelin wrote: > Actually, you got me thinking about the desc.buffer. And I think there is > one corner case where it could cause a problem: `struct tree_desc desc[2]` > does not initialize the buffers to NULL. And what if > fill_tree_descriptor() function returns NULL? Then the buffer is still > uninitialized. > > In practice, our *current* version of fill_tree_descriptor() never returns > NULL if the oid parameter is non-NULL. It would die() in the error case > instead (bad!). So to prepare for a less bad version, I'd rather > initialize the `desc` array and be on the safe (and easier to reason > about) side. Yeah, I agree with all of that. One solution would just be to increment only after success: if (fill_tree_descriptor([nr], ..) < 0) goto error; nr++; /* now we know it's valid! */ But there are lots of alternatives. :) -Peff
Re: [PATCH 2/2] built-in rebase: reinstate `checkout -q` behavior where appropriate
Hi Peff,
On Fri, 9 Nov 2018, Jeff King wrote:
> On Fri, Nov 09, 2018 at 01:34:19AM -0800, Johannes Schindelin via
> GitGitGadget wrote:
>
> > From: Johannes Schindelin
> >
> > When we converted a `git checkout -q $onto^0` call to use
> > `reset_head()`, we inadvertently incurred a change from a twoway_merge
> > to a oneway_merge, as if we wanted a `git reset --hard` instead.
> >
> > This has performance ramifications under certain, though, as the
> > oneway_merge needs to lstat() every single index entry whereas
> > twoway_merge does not.
> >
> > So let's go back to the old behavior.
>
> Makes sense. I didn't think too hard about any possible gotchas with the
> twoway/oneway switch, but if that's what git-checkout was doing before,
> it seems obviously safe.
Right.
> > diff --git a/builtin/rebase.c b/builtin/rebase.c
> > index 6f6d7de156..c1cc50f3f8 100644
> > --- a/builtin/rebase.c
> > +++ b/builtin/rebase.c
> > @@ -523,11 +523,12 @@ finished_rebase:
> > #define GIT_REFLOG_ACTION_ENVIRONMENT "GIT_REFLOG_ACTION"
> >
> > static int reset_head(struct object_id *oid, const char *action,
> > - const char *switch_to_branch, int detach_head,
> > + const char *switch_to_branch,
> > + int detach_head, int reset_hard,
>
> It might be worth switching to a single flag variable here. It would
> make calls like this:
>
> > - if (reset_head(>object.oid, "checkout", NULL, 1,
> > + if (reset_head(>object.oid, "checkout", NULL, 1, 0,
> > NULL, msg.buf))
>
> a little more self-documenting (if a little more verbose).
I thought about that, but for just two flags? Well, let me try it and see
whether I like the result ;-)
> > - if (!oid) {
> > - if (get_oid("HEAD", _oid)) {
> > - rollback_lock_file();
> > - return error(_("could not determine HEAD revision"));
> > - }
> > - oid = _oid;
> > + if (get_oid("HEAD", _oid)) {
> > + rollback_lock_file();
> > + return error(_("could not determine HEAD revision"));
> > }
>
> This one could actually turn into:
>
> ret = error(...);
> goto leave_reset_head;
>
> now. We don't have to worry about an uninitialized desc.buffer anymore
> (as I mentioned in the previous email), because "nr" would be 0.
>
> It doesn't save any lines, though (but maybe having a single
> cleanup/exit point would make things easier to read; I dunno).
But you're right, of course. Consistency makes for easier-to-read code.
> Take all my comments as observations, not objections. This looks OK to
> me either way.
Actually, you got me thinking about the desc.buffer. And I think there is
one corner case where it could cause a problem: `struct tree_desc desc[2]`
does not initialize the buffers to NULL. And what if
fill_tree_descriptor() function returns NULL? Then the buffer is still
uninitialized.
In practice, our *current* version of fill_tree_descriptor() never returns
NULL if the oid parameter is non-NULL. It would die() in the error case
instead (bad!). So to prepare for a less bad version, I'd rather
initialize the `desc` array and be on the safe (and easier to reason
about) side.
Thank you for your helpful review,
Dscho
Re: [PATCH 2/2] built-in rebase: reinstate `checkout -q` behavior where appropriate
On Fri, Nov 09, 2018 at 01:34:19AM -0800, Johannes Schindelin via GitGitGadget
wrote:
> From: Johannes Schindelin
>
> When we converted a `git checkout -q $onto^0` call to use
> `reset_head()`, we inadvertently incurred a change from a twoway_merge
> to a oneway_merge, as if we wanted a `git reset --hard` instead.
>
> This has performance ramifications under certain, though, as the
> oneway_merge needs to lstat() every single index entry whereas
> twoway_merge does not.
>
> So let's go back to the old behavior.
Makes sense. I didn't think too hard about any possible gotchas with the
twoway/oneway switch, but if that's what git-checkout was doing before,
it seems obviously safe.
> diff --git a/builtin/rebase.c b/builtin/rebase.c
> index 6f6d7de156..c1cc50f3f8 100644
> --- a/builtin/rebase.c
> +++ b/builtin/rebase.c
> @@ -523,11 +523,12 @@ finished_rebase:
> #define GIT_REFLOG_ACTION_ENVIRONMENT "GIT_REFLOG_ACTION"
>
> static int reset_head(struct object_id *oid, const char *action,
> - const char *switch_to_branch, int detach_head,
> + const char *switch_to_branch,
> + int detach_head, int reset_hard,
It might be worth switching to a single flag variable here. It would
make calls like this:
> - if (reset_head(>object.oid, "checkout", NULL, 1,
> + if (reset_head(>object.oid, "checkout", NULL, 1, 0,
> NULL, msg.buf))
a little more self-documenting (if a little more verbose).
> - if (!oid) {
> - if (get_oid("HEAD", _oid)) {
> - rollback_lock_file();
> - return error(_("could not determine HEAD revision"));
> - }
> - oid = _oid;
> + if (get_oid("HEAD", _oid)) {
> + rollback_lock_file();
> + return error(_("could not determine HEAD revision"));
> }
This one could actually turn into:
ret = error(...);
goto leave_reset_head;
now. We don't have to worry about an uninitialized desc.buffer anymore
(as I mentioned in the previous email), because "nr" would be 0.
It doesn't save any lines, though (but maybe having a single
cleanup/exit point would make things easier to read; I dunno).
Take all my comments as observations, not objections. This looks OK to
me either way.
-Peff
[PATCH 2/2] built-in rebase: reinstate `checkout -q` behavior where appropriate
From: Johannes Schindelin
When we converted a `git checkout -q $onto^0` call to use
`reset_head()`, we inadvertently incurred a change from a twoway_merge
to a oneway_merge, as if we wanted a `git reset --hard` instead.
This has performance ramifications under certain, though, as the
oneway_merge needs to lstat() every single index entry whereas
twoway_merge does not.
So let's go back to the old behavior.
Signed-off-by: Johannes Schindelin
---
builtin/rebase.c | 45 ++---
1 file changed, 26 insertions(+), 19 deletions(-)
diff --git a/builtin/rebase.c b/builtin/rebase.c
index 6f6d7de156..c1cc50f3f8 100644
--- a/builtin/rebase.c
+++ b/builtin/rebase.c
@@ -523,11 +523,12 @@ finished_rebase:
#define GIT_REFLOG_ACTION_ENVIRONMENT "GIT_REFLOG_ACTION"
static int reset_head(struct object_id *oid, const char *action,
- const char *switch_to_branch, int detach_head,
+ const char *switch_to_branch,
+ int detach_head, int reset_hard,
const char *reflog_orig_head, const char *reflog_head)
{
struct object_id head_oid;
- struct tree_desc desc;
+ struct tree_desc desc[2];
struct lock_file lock = LOCK_INIT;
struct unpack_trees_options unpack_tree_opts;
struct tree *tree;
@@ -536,7 +537,7 @@ static int reset_head(struct object_id *oid, const char
*action,
size_t prefix_len;
struct object_id *orig = NULL, oid_orig,
*old_orig = NULL, oid_old_orig;
- int ret = 0;
+ int ret = 0, nr = 0;
if (switch_to_branch && !starts_with(switch_to_branch, "refs/"))
BUG("Not a fully qualified branch: '%s'", switch_to_branch);
@@ -544,20 +545,20 @@ static int reset_head(struct object_id *oid, const char
*action,
if (hold_locked_index(, LOCK_REPORT_ON_ERROR) < 0)
return -1;
- if (!oid) {
- if (get_oid("HEAD", _oid)) {
- rollback_lock_file();
- return error(_("could not determine HEAD revision"));
- }
- oid = _oid;
+ if (get_oid("HEAD", _oid)) {
+ rollback_lock_file();
+ return error(_("could not determine HEAD revision"));
}
+ if (!oid)
+ oid = _oid;
+
memset(_tree_opts, 0, sizeof(unpack_tree_opts));
setup_unpack_trees_porcelain(_tree_opts, action);
unpack_tree_opts.head_idx = 1;
unpack_tree_opts.src_index = the_repository->index;
unpack_tree_opts.dst_index = the_repository->index;
- unpack_tree_opts.fn = oneway_merge;
+ unpack_tree_opts.fn = reset_hard ? oneway_merge : twoway_merge;
unpack_tree_opts.update = 1;
unpack_tree_opts.merge = 1;
if (!detach_head)
@@ -568,12 +569,17 @@ static int reset_head(struct object_id *oid, const char
*action,
return error(_("could not read index"));
}
- if (!fill_tree_descriptor(, oid)) {
+ if (!reset_hard && !fill_tree_descriptor([nr++], _oid)) {
+ ret = error(_("failed to find tree of %s"), oid_to_hex(oid));
+ goto leave_reset_head;
+ }
+
+ if (!fill_tree_descriptor([nr++], oid)) {
ret = error(_("failed to find tree of %s"), oid_to_hex(oid));
goto leave_reset_head;
}
- if (unpack_trees(1, , _tree_opts)) {
+ if (unpack_trees(nr, desc, _tree_opts)) {
ret = -1;
goto leave_reset_head;
}
@@ -621,7 +627,8 @@ static int reset_head(struct object_id *oid, const char
*action,
leave_reset_head:
strbuf_release();
rollback_lock_file();
- free((void *)desc.buffer);
+ while (nr)
+ free((void *)desc[--nr].buffer);
return ret;
}
@@ -999,7 +1006,7 @@ int cmd_rebase(int argc, const char **argv, const char
*prefix)
rerere_clear(_rr);
string_list_clear(_rr, 1);
- if (reset_head(NULL, "reset", NULL, 0, NULL, NULL) < 0)
+ if (reset_head(NULL, "reset", NULL, 0, 1, NULL, NULL) < 0)
die(_("could not discard worktree changes"));
if (read_basic_state())
exit(1);
@@ -1015,7 +1022,7 @@ int cmd_rebase(int argc, const char **argv, const char
*prefix)
if (read_basic_state())
exit(1);
if (reset_head(_head, "reset",
- options.head_name, 0, NULL, NULL) < 0)
+ options.head_name, 0, 1, NULL, NULL) < 0)
die(_("could not move back to %s"),
oid_to_hex(_head));
ret = finish_rebase();
@@ -1379,7 +1386,7 @@ int cmd_rebase(int argc, const char **argv, const char
*prefix)
write_file(autostash, "%s",
