This is the mail archive of the mailing list for the GCC 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: locating unsigned type for non-standard precision

Richard Guenther wrote:

>> I've run into another issue supporting a 20-bit integer for which I'd
>> appreciate a hint.  With this code:
>>   typedef long int __attribute__((__a20__)) int20_t;
>>   int20_t xi;
>>   int20_t addit () { xi += 0x54321L; }
>> xi ends up in mode PSImode, which is a MODE_PARTIAL_INT with 20 bits of
>> precision and 32-bit width.
>> convert() notices that, because the constant in the add expression is
>> SImode, there's an SImode add being truncated to a PSImode result, and
>> pushes the truncation down into the operands.
>> The problem is this ends up in convert_to_integer, which detects that the
>> signed operation might overflow so calls unsigned_type_for() to get the
>> unsigned variant.
>> Unfortunately, this ends up in c_common_type_for_size(), which knows nothing
>> about PSImode, and returns an unsigned type with 32 bits of precision when
>> asked for one with 20 bits of precision.
> That's expected - this function returns a type that is suitable for holding all
> values, not a type that has necessarily matching precision.  If the caller
> wants such type it needs to verify what the function returned.  Which seems
> to me to be the correct fix for this (premature) optimization in
> convert_to_integer.
> Richard.
>>   The expression is rewritten with
>> the 32-bit constant integer recast to the 32-bit unsigned type (instead
>> of the 20-bit one it might have used), and infinite recursion results.

This is already filed as

It works with 4.8 trunk but crashes with 4.7.
Did not yet track what changes made it work with 4.8, though.
Unfortunately, noone remembers :-(


>> Is it proper for c_common_type_for_size() to know about partial int modes
>> and return the best one available?  Is there a hook that would allow me to
>> do this customized in my back-end?  Or is there another way to get
>> unsigned_type_for() to return the "right" type for an unusual integer
>> precision?
>> Thanks.
>> Peter

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