[PATCH] ix86: wrap constants

[Jan Beulich]

On 04.06.2021 14:26, H.J. Lu wrote:
> On Fri, Jun 4, 2021 at 5:14 AM Jan Beulich <jbeulich@suse.com> wrote:
>>
>> Non-64-bit code should get handled the same with or without BFD64. This
>> wasn't the case though in a number of situations (and quite likely there
>> are more that I haven't spotted yet).
>>
>> It's not very nice to tie the check in md_apply_fix() to object_64bit,
>> but afaict at that time we have no record anymore of the mode an insn
>> was assembled in (it might also have been data). This doesn't look to be
>> the first inconsistency of this kind, though. In x86_cons() it's even
>> less clear what the right approach would be: flag_code shouldn't matter
>> for data emission, but instead we'd need to know from which mode(s) the
>> data actually gets accessed. On this basis, signed_cons() also gets
>> adjusted.
>>
>> gas/
>> 2021-06-XX  Jan Beulich  <jbeulich@suse.com>
>>
>>         * config/tc-i386.c (extend_to_address): New.
>>         (x86_cons, i386_finalize_immediate, md_apply_fix): Use it.
>>         (signed_cons): Use object_64bit.
>>         * testsuite/gas/i386/wrap32.s, testsuite/gas/i386/wrap32-data.d,
>>         testsuite/gas/i386/wrap32-text.d: New.
>>         * testsuite/gas/i386/i386.exp: Run new tests.
>>
>> --- a/gas/config/tc-i386.c
>> +++ b/gas/config/tc-i386.c
>> @@ -2457,6 +2457,19 @@ fits_in_unsigned_long (addressT num ATTR
>>  #endif
>>  }                              /* fits_in_unsigned_long() */
>>
>> +static INLINE valueT extend_to_address (addressT num)
>> +{
>> +#ifdef BFD64
>> +  if (fits_in_unsigned_long(num))
>> +    return (num ^ ((addressT) 1 << 31)) - ((addressT) 1 << 31);
>> +
>> +  if (!fits_in_signed_long (num))
>> +    return num & 0xffffffff;
>> +#endif
> 
> Shouldn't this check apply only to 32-bit objects?

It already does (as far as it should; the check should really
apply to non-64-bit code only) - callers use either !object_64bit
up front, or "flag_code != CODE_64BIT". As per what the description
says, the latter is really how we'd like it to be (because for
32-bit code it shouldn't matter what container it lives in), but
- in x86_cons() applicability of flag_code is at best questionable,
- in md_apply_fix() we don't have that information available.

Jan