This is the mail archive of the gdb-patches@sourceware.org mailing list for the GDB project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: wrong assumptions about pthread_t being numeric


On 09/17/2011 02:30 AM, Pedro Alves wrote:
On Saturday 17 September 2011 00:13:10, John Spencer wrote:
On 09/17/2011 01:00 AM, Pedro Alves wrote:
These are only built natively on solaris and aix respectively, so
let's just leave them alone.

I expected it to be desirable for a product in industrial use to be
standard-compliant and not invoking undefined behavior.
Those files are tied to those platforms' thread_db/libc implementations.
There's absolutely no need to handle some other hipotetical libc that
defines pthread_t diferently there.  If it appears, we'll handle it.

Chances are, some other changes would be necessary to make it really
work, not just build.


exactly. for example in musl's case it is wrong to compare the underlying type (which is a struct pointer) with 0.
in the implementation specific part it would practically be valid to compare it with
NULL, but that's still invalid regarding the definition in POSIX.
so the few places that use a thread id had to be changed everywhere, to not have dozens of special case
sections for any libc/OS combination...


also there are a few spots where the thread_t is numerically compared to some arbitrary int value
(i have not yet evaluated where those originate from). this also can't work when thread_t is no number.
this happens for example in the if statement leading to the posted printf format string errors.


i expect that the number of actual fixes would be small, but it might require another way to get to a thread id.

if the id is only used for textual representation, it would be sufficient if gdb would assign a new unique number
on each successful pthread_create.
if it is also used to find a thread internally, it had to have a list of tid numbers -> pthread_t.



thread-db.c: In function 'find_one_thread':
thread-db.c:295:7: error: format '%ld' expects type 'long int', but
argument 3 has type 'thread_t'
thread-db.c: In function 'attach_thread':
thread-db.c:335:7: error: format '%ld' expects type 'long int', but
argument 3 has type 'thread_t'
thread-db.c:341:9: error: format '%ld' expects type 'long int', but
argument 2 has type 'thread_t'
So just cast it to long, and you're done.

pthread_t could legally be a struct, which you can't just cast to a long.
No need to complicate things for an hipotetical scenario.  The set of
libc's in existence is finite.  If we were to handle a struct pthread_t,
we'd need to be able to print it, and so we'd need some libc specific
way to do it, something autoconf'ed.  There's no need to invent work.

i disagree. adding a proper solution once is superior to creating dozens of special case hacks.
also it saves a lot of time in the long term.


-- JS


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]