[2/3] Allocate data in cached_reg_t
Commit Message
Aarch64 SVE requires a max register size of 256. The current max size in gdb
is 64. This is part of a series demonstrating the replacement of
MAX_REGISTER_SIZE.
In cached_reg_t the data is changed to a pointer, which is allocated using the
size of the register being cached. This pointer must be manually freed when
deleting a DEF_VEC of cached_reg_t's.
Tested on x86.
Ok to commit?
Thanks,
Alan.
2017-01-09 Alan Hayward <alan.hayward@arm.com>
* remote.c (struct cached_reg): Change data into a pointer.
* (stop_reply_dtr): Free data pointers before deleting vector.
(process_stop_reply): Likewise.
(remote_parse_stop_reply): Allocate space for data
Comments
On 01/09/2017 04:56 AM, Alan Hayward wrote:
> Aarch64 SVE requires a max register size of 256. The current max size in gdb
> is 64. This is part of a series demonstrating the replacement of
> MAX_REGISTER_SIZE.
>
> In cached_reg_t the data is changed to a pointer, which is allocated using the
> size of the register being cached. This pointer must be manually freed when
> deleting a DEF_VEC of cached_reg_t's.
>
> Tested on x86.
> Ok to commit?
>
> Thanks,
> Alan.
>
> 2017-01-09 Alan Hayward <alan.hayward@arm.com>
>
> * remote.c (struct cached_reg): Change data into a pointer.
> * (stop_reply_dtr): Free data pointers before deleting vector.
> (process_stop_reply): Likewise.
> (remote_parse_stop_reply): Allocate space for data
>
>
> diff --git a/gdb/remote.c b/gdb/remote.c
> index 6da6eb366ae442354fd6a37741335af9a4a5a056..9247d43b094925ff397eb36b450eaba521adfc99 100644
> --- a/gdb/remote.c
> +++ b/gdb/remote.c
> @@ -6306,7 +6306,7 @@ remote_console_output (char *msg)
> typedef struct cached_reg
> {
> int num;
> - gdb_byte data[MAX_REGISTER_SIZE];
> + gdb_byte *data;
Would it make sense to go C++ and use a data structure that can take
care of variable sizes? Just thinking if that would be easier than
handling allocation/deallocation of the data.
> } cached_reg_t;
>
> DEF_VEC_O(cached_reg_t);
> @@ -6402,6 +6402,13 @@ static void
> stop_reply_dtr (struct notif_event *event)
> {
> struct stop_reply *r = (struct stop_reply *) event;
> + cached_reg_t *reg;
> + int ix;
> +
> + for (ix = 0;
> + VEC_iterate(cached_reg_t, r->regcache, ix, reg);
> + ix++)
> + xfree (reg->data);
>
> VEC_free (cached_reg_t, r->regcache);
> }
> @@ -6974,6 +6981,7 @@ Packet: '%s'\n"),
> {
> struct packet_reg *reg = packet_reg_from_pnum (rsa, pnum);
> cached_reg_t cached_reg;
> + struct gdbarch *gdbarch = target_gdbarch ();
>
> if (reg == NULL)
> error (_("Remote sent bad register number %s: %s\n\
> @@ -6981,14 +6989,14 @@ Packet: '%s'\n"),
> hex_string (pnum), p, buf);
>
> cached_reg.num = reg->regnum;
> + cached_reg.data = (gdb_byte *)
> + xmalloc (register_size (gdbarch, reg->regnum));
>
> p = p1 + 1;
> fieldsize = hex2bin (p, cached_reg.data,
> - register_size (target_gdbarch (),
> - reg->regnum));
> + register_size (gdbarch, reg->regnum));
> p += 2 * fieldsize;
> - if (fieldsize < register_size (target_gdbarch (),
> - reg->regnum))
> + if (fieldsize < register_size (gdbarch, reg->regnum))
> warning (_("Remote reply is too short: %s"), buf);
>
> VEC_safe_push (cached_reg_t, event->regcache, &cached_reg);
> @@ -7211,7 +7219,11 @@ process_stop_reply (struct stop_reply *stop_reply,
> for (ix = 0;
> VEC_iterate(cached_reg_t, stop_reply->regcache, ix, reg);
> ix++)
> + {
> regcache_raw_supply (regcache, reg->num, reg->data);
> + xfree (reg->data);
> + }
> +
> VEC_free (cached_reg_t, stop_reply->regcache);
> }
>
>
>
>
On 17-01-09 14:11:13, Luis Machado wrote:
> >@@ -6306,7 +6306,7 @@ remote_console_output (char *msg)
> > typedef struct cached_reg
> > {
> > int num;
> >- gdb_byte data[MAX_REGISTER_SIZE];
> >+ gdb_byte *data;
>
> Would it make sense to go C++ and use a data structure that can take
> care of variable sizes? Just thinking if that would be easier than
> handling allocation/deallocation of the data.
>
Do you suggest std::vector? We still need to allocate/deallocate it
if we change "gdb_byte *data" to "std::vector<gdb_byte> *data", or we
have to convert cached_reg to class, and do RAII.
class cached_reg
{
public:
cached_reg (struct gdbarch *gdbarch, int num_)
: num (num_)
{
this->data = (gdb_byte *) xmalloc (register_size (gdbarch, num_));
}
~cached_reg ()
{
xfree (this->data);
}
private:
int num;
gdb_byte *data;
};
On 01/10/2017 06:59 AM, Yao Qi wrote:
> On 17-01-09 14:11:13, Luis Machado wrote:
>>> @@ -6306,7 +6306,7 @@ remote_console_output (char *msg)
>>> typedef struct cached_reg
>>> {
>>> int num;
>>> - gdb_byte data[MAX_REGISTER_SIZE];
>>> + gdb_byte *data;
>>
>> Would it make sense to go C++ and use a data structure that can take
>> care of variable sizes? Just thinking if that would be easier than
>> handling allocation/deallocation of the data.
>>
>
> Do you suggest std::vector? We still need to allocate/deallocate it
> if we change "gdb_byte *data" to "std::vector<gdb_byte> *data", or we
> have to convert cached_reg to class, and do RAII.
>
Something like std::vector, yes. But it is true we will still need to
allocate the vector itself.
I was pondering about the benefits of not being limited to a specific
register size. Then we wouldn't need to worry about adjusting things as
we have to do now.
As i see it, data[register_size(register_size (gdbarch, num)] wouldn't
be much different than data[MAX_REGISTER_SIZE]. The only thing is that
we're setting the max register size dynamically.
We may or may not need this flexibility right now, but who knows what
weird architectures we may have in the future.
> On 10 Jan 2017, at 17:42, Luis Machado <lgustavo@codesourcery.com> wrote:
>
> On 01/10/2017 06:59 AM, Yao Qi wrote:
>> On 17-01-09 14:11:13, Luis Machado wrote:
>>>> @@ -6306,7 +6306,7 @@ remote_console_output (char *msg)
>>>> typedef struct cached_reg
>>>> {
>>>> int num;
>>>> - gdb_byte data[MAX_REGISTER_SIZE];
>>>> + gdb_byte *data;
>>>
>>> Would it make sense to go C++ and use a data structure that can take
>>> care of variable sizes? Just thinking if that would be easier than
>>> handling allocation/deallocation of the data.
>>>
>>
>> Do you suggest std::vector? We still need to allocate/deallocate it
>> if we change "gdb_byte *data" to "std::vector<gdb_byte> *data", or we
>> have to convert cached_reg to class, and do RAII.
>>
>
> Something like std::vector, yes. But it is true we will still need to allocate the vector itself.
>
> I was pondering about the benefits of not being limited to a specific register size. Then we wouldn't need to worry about adjusting things as we have to do now.
>
> As i see it, data[register_size(register_size (gdbarch, num)] wouldn't be much different than data[MAX_REGISTER_SIZE]. The only thing is that we're setting the max register size dynamically.
>
> We may or may not need this flexibility right now, but who knows what weird architectures we may have in the future.
I avoided using data[register_size(gdbarch, num)] as dynamically sized arrays are a gcc extension - it’s not something we can expect to work on other compilers.
Plus there is a danger of it using up all the stack. Aarch64 SVE will have a max register size of 256, potentially this could grow further in the future.
This is why I used alloca in patches 1/3 and 2/3.
I know that gcc avoid using std::vector because it’s fairly slow. I think gcc also have newer c++ like replacements for VEC, but am not sure what they have that’s suitable. However, I’d suggest that’s something that should be done as part of the work to move gdb to c++, and not this patch.
Alan.
On 17-01-09 10:56:56, Alan Hayward wrote:
> @@ -6402,6 +6402,13 @@ static void
> stop_reply_dtr (struct notif_event *event)
> {
> struct stop_reply *r = (struct stop_reply *) event;
> + cached_reg_t *reg;
> + int ix;
> +
> + for (ix = 0;
> + VEC_iterate(cached_reg_t, r->regcache, ix, reg);
A space is needed before "(".
Patch is good to me.
@@ -6306,7 +6306,7 @@ remote_console_output (char *msg)
typedef struct cached_reg
{
int num;
- gdb_byte data[MAX_REGISTER_SIZE];
+ gdb_byte *data;
} cached_reg_t;
DEF_VEC_O(cached_reg_t);
@@ -6402,6 +6402,13 @@ static void
stop_reply_dtr (struct notif_event *event)
{
struct stop_reply *r = (struct stop_reply *) event;
+ cached_reg_t *reg;
+ int ix;
+
+ for (ix = 0;
+ VEC_iterate(cached_reg_t, r->regcache, ix, reg);
+ ix++)
+ xfree (reg->data);
VEC_free (cached_reg_t, r->regcache);
}
@@ -6974,6 +6981,7 @@ Packet: '%s'\n"),
{
struct packet_reg *reg = packet_reg_from_pnum (rsa, pnum);
cached_reg_t cached_reg;
+ struct gdbarch *gdbarch = target_gdbarch ();
if (reg == NULL)
error (_("Remote sent bad register number %s: %s\n\
@@ -6981,14 +6989,14 @@ Packet: '%s'\n"),
hex_string (pnum), p, buf);
cached_reg.num = reg->regnum;
+ cached_reg.data = (gdb_byte *)
+ xmalloc (register_size (gdbarch, reg->regnum));
p = p1 + 1;
fieldsize = hex2bin (p, cached_reg.data,
- register_size (target_gdbarch (),
- reg->regnum));
+ register_size (gdbarch, reg->regnum));
p += 2 * fieldsize;
- if (fieldsize < register_size (target_gdbarch (),
- reg->regnum))
+ if (fieldsize < register_size (gdbarch, reg->regnum))
warning (_("Remote reply is too short: %s"), buf);
VEC_safe_push (cached_reg_t, event->regcache, &cached_reg);
@@ -7211,7 +7219,11 @@ process_stop_reply (struct stop_reply *stop_reply,
for (ix = 0;
VEC_iterate(cached_reg_t, stop_reply->regcache, ix, reg);
ix++)
+ {
regcache_raw_supply (regcache, reg->num, reg->data);
+ xfree (reg->data);
+ }
+
VEC_free (cached_reg_t, stop_reply->regcache);
}