Re: [PATCH 1/2 v3] Make diffcore_std only can run once before a diff_flush.

From: Bo Yang <struggleyb.nku@gmail.com>
Date: Apr 23, 2010, 03:55 UTC
Message-ID: <n2m41f08ee11004222055i498174dfi47a5991a08114cd2@mail.gmail.com>
In-Reply-To: <7v1ve743vb.fsf@alter.siamese.dyndns.org>

On Fri, Apr 23, 2010 at 4:41 AM, Junio C Hamano <gitster@pobox.com> wrote:

> It actually is stronger than that; we should never run it more than once,
> and it would be a bug if we did so.  Which codepath tries to call *_std()
> twice?

In command 'git log --follow ...' log_tree_diff call diff_tree_sha1 and then diff_tree_diff_flush, when '--follow' is given, the former function will call try_to_follow_renames, which will call diffcore_std to detect rename. And then, diff_tree_diff_flush call 'diffcore_std' again unconditional. (and I have try to find a condition to make the call, but I fail, so I figure out this patch.)

Breakpoint 1, diffcore_std (options=0xbf9cc044) at diff.c:3748
3748		if (diff_queued_diff.run)
(gdb) bt
#0  diffcore_std (options=0xbf9cc044) at diff.c:3748
#1  0x08124206 in try_to_follow_renames (t1=0xbf9cc130, t2=0xbf9cc11c,
base=0x81571c9 "", opt=0xbf9cc468) at tree-diff.c:358
#2  0x08124480 in diff_tree_sha1 (old=0x9c51d8c
"$\033\222T���\a\035\200T����\210;8\235i", new=0x9c51d2c
"\201�\017<�\v��n]\226{�+�\001\003\232\232\230",
    base=0x81571c9 "", opt=0xbf9cc468) at tree-diff.c:418
#3  0x080e660e in log_tree_diff (opt=0xbf9cc220, commit=0x9c51d28,
log=0xbf9cc1ac) at log-tree.c:536
#4  0x080e668f in log_tree_commit (opt=0xbf9cc220, commit=0x9c51d28)
at log-tree.c:560
#5  0x0807faa1 in cmd_log_walk (rev=0xbf9cc220) at builtin/log.c:237
#6  0x080806e2 in cmd_log (argc=5, argv=0xbf9cc788, prefix=0x0) at
builtin/log.c:481
#7  0x0804b8eb in run_builtin (p=0x8161524, argc=5, argv=0xbf9cc788)
at git.c:260
#8  0x0804ba51 in handle_internal_command (argc=5, argv=0xbf9cc788) at git.c:416
#9  0x0804bb2c in run_argv (argcp=0xbf9cc700, argv=0xbf9cc704) at git.c:458
#10 0x0804bcbe in main (argc=5, argv=0xbf9cc788) at git.c:529
(gdb) c
Continuing.

Breakpoint 1, diffcore_std (options=0xbf9cc468) at diff.c:3748 3748 if (diff_queued_diff.run) (gdb) bt #0 diffcore_std (options=0xbf9cc468) at diff.c:3748 #1 0x080e6356 in log_tree_diff_flush (opt=0xbf9cc220) at log-tree.c:449 #2 0x080e6619 in log_tree_diff (opt=0xbf9cc220, commit=0x9c51d28, log=0xbf9cc1ac) at log-tree.c:537 #3 0x080e668f in log_tree_commit (opt=0xbf9cc220, commit=0x9c51d28) at log-tree.c:560 #4 0x0807faa1 in cmd_log_walk (rev=0xbf9cc220) at builtin/log.c:237 #5 0x080806e2 in cmd_log (argc=5, argv=0xbf9cc788, prefix=0x0) at builtin/log.c:481 #6 0x0804b8eb in run_builtin (p=0x8161524, argc=5, argv=0xbf9cc788) at git.c:260 #7 0x0804ba51 in handle_internal_command (argc=5, argv=0xbf9cc788) at git.c:416 #8 0x0804bb2c in run_argv (argcp=0xbf9cc700, argv=0xbf9cc704) at git.c:458 #9 0x0804bcbe in main (argc=5, argv=0xbf9cc788) at git.c:529 (gdb)

Show 24 quoted lines

> The standard calling sequence is:
>
>  - start from an empty queue.
>
>  - use diff_change() and diff_addremove() to populate the queue.
>
>  - call diffcore_std(). if you need to use a non-standard chain of
>   diffcore transformations, you _could_ call the diffcore_* routines that
>   diffcore_std() calls, if you choose to, but as you found out, some of
>   them are not idempotent operations, and shouldn't be called twice.
>
>  - and finally call diffcore_flush().
>
>> @@ -3745,6 +3742,12 @@ void diffcore_fix_diff_index(struct diff_options *options)
>>
>>  void diffcore_std(struct diff_options *options)
>>  {
>> +     /* We never run this function more than one time, because the
>> +      * rename/copy detection logic can only run once.
>> +      */
>> +     if (diff_queued_diff.run)
>> +             return;
>
> Shouldn't this be a BUG() instead?

Anyone may call diff_tree_sha1 and then diffcore_std, and diff_tree_sha1 may call another diffcore_std if '--follow' given. If this is a BUG, the calling pattern, diff_tree_sha1 -> diffcore_std should all disappear from our code. And this involved much code refactor. And I suggest my way that we avoid the duplicate call actively in diffcore_std.

Show 9 quoted lines

> The trivial rewrite to use this macro is a good idea, but it probably
> should be a separate patch.
>
>> +#define DIFF_QUEUE_CLEAR(q) \
>> +     do { \
>> +             (q)->queue = NULL; \
>> +             (q)->nr = (q)->alloc = (q)->run = 0; \
>> +     } while(0);
>

You mean split this commit into two?

Regards! Bo

-- 
My blog: http://blog.morebits.org