On 02/02/2014 01:20 AM, Pádraig Brady wrote:
> On 01/31/2014 09:44 AM, Niels Möller wrote:
>> address@hidden (Niels Möller) writes:
>>
>>> Pádraig Brady <address@hidden> writes:
>>>> I agree this would be useful and easy enough to add.
>>>> I suppose the interface would be --endian=little|big
>>>
>>> Maybe I can have a look at what it takes.
>>
>> Below is a crude patch (missing: usage message, tests cases, docs,
>> translation). I think it should work fine for floats too. I see no
>> obvious and more beautiful way to do it.
>>
>> (And I think I have copyright assignment papers for coreutils in place,
>> since work on factor some year ago).
>>
>> Regards,
>> /Niels
>>
>> diff --git a/src/od.c b/src/od.c
>> index 514fe50..a71e302 100644
>> --- a/src/od.c
>> +++ b/src/od.c
>> @@ -259,13 +259,16 @@ static enum size_spec
>> integral_type_size[MAX_INTEGRAL_TYPE_SIZE + 1];
>> #define MAX_FP_TYPE_SIZE sizeof (long double)
>> static enum size_spec fp_type_size[MAX_FP_TYPE_SIZE + 1];
>>
>> +bool input_swap;
>> +
>> static char const short_options[] = "A:aBbcDdeFfHhIij:LlN:OoS:st:vw::Xx";
>>
>> /* For long options that have no equivalent short option, use a
>> non-character as a pseudo short option, starting with CHAR_MAX + 1. */
>> enum
>> {
>> - TRADITIONAL_OPTION = CHAR_MAX + 1
>> + TRADITIONAL_OPTION = CHAR_MAX + 1,
>> + ENDIAN_OPTION,
>> };
>>
>> static struct option const long_options[] =
>> @@ -278,6 +281,7 @@ static struct option const long_options[] =
>> {"strings", optional_argument, NULL, 'S'},
>> {"traditional", no_argument, NULL, TRADITIONAL_OPTION},
>> {"width", optional_argument, NULL, 'w'},
>> + {"endian", required_argument, NULL, ENDIAN_OPTION },
>>
>> {GETOPT_HELP_OPTION_DECL},
>> {GETOPT_VERSION_OPTION_DECL},
>> @@ -406,7 +410,21 @@ N (size_t fields, size_t blank, void const *block,
>> \
>> { \
>> int next_pad = pad * (i - 1) / fields; \
>> int adjusted_width = pad_remaining - next_pad + width; \
>> - T x = *p++; \
>> + T x; \
>> + if (input_swap && sizeof(T) > 1) \
>> + { \
>> + int j; \
>> + union { \
>> + T x; \
>> + char b[sizeof(T)]; \
>> + } u; \
>> + for (j = 0; j < sizeof(T); j++) \
>> + u.b[j] = ((const char *) p)[sizeof(T) - 1 - j]; \
>> + x = u.x; \
>> + } \
>> + else \
>> + x = *p; \
>> + p++; \
>> ACTION; \
>> pad_remaining = next_pad; \
>> } \
>> @@ -1664,6 +1682,24 @@ main (int argc, char **argv)
>> traditional = true;
>> break;
>>
>> + case ENDIAN_OPTION:
>> + if (!strcmp (optarg, "big"))
>> + {
>> +#if !WORDS_BIGENDIAN
>> + input_swap = true;
>> +#endif
>> + }
>> + else if (!strcmp (optarg, "little"))
>> + {
>> +#if WORDS_BIGENDIAN
>> + input_swap = true;
>> +#endif
>> + }
>> + else
>> + error (EXIT_FAILURE, 0,
>> + _("bad argument '%s' for --endian option"), optarg);
>> + break;
>> +
>> /* The next several cases map the traditional format
>> specification options to the corresponding modern format
>> specs. GNU od accepts any combination of old- and
>
> That looks good.
> I'll adjust slightly to use XARGMATCH and add some docs/tests.
> I'm travelling at the moment but merge this soon.
Attached in the patch I intend to push in your name.
I changed the option handling to reuse the XARGMATCH functionality.
Also I changed things slightly so as the last --endian option
specified wins. Previously we only set the input_swap variable
to true, never to false. On a related point I set the input_swap
global to be static.
I also added docs to usage() and the texinfo file, and added a test.
BTW I checked if there was any speed difference with the new code.
I wasn't expecting this to be a bottleneck, and true enough
there is only a marginal change. The new code is consistently
a little _faster_ though on my i3-2310M which is a bit surprising.
$ truncate -s1G od.in
$ time od.old -tx8 od.in
5.05 elapsed
$ time od.new -tx8 --endian=bug od.in
4.97 elapsed
My hunch is there is more pretching happening in the new version,
but can't check on this system due to:
$ perf stat -e L1-dcache-prefetches:u true
<not supported> L1-dcache-prefetches:u
For kicks I put in bswap_{16,32,64}() calls which are guaranteed
available by gnulib, but replaced with architecture specific asm
on this system, and the speed regressed back to that of od.old.
thanks,
Pádraig.
od--endian.patch
Description: Text Data
--- End Message ---