Re: [patch] canonicalize unsigned [1,MAX] ranges into ~[0,0]

[Iain Sandoe <idsandoe at googlemail dot com>]

Hi Aldy,

Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE> wrote:
>> 
>> On 10/15/19 7:58 AM, Rainer Orth wrote:
>>> Hi Aldy,
>>> 
>>>>>> ~[0,0] has been the accepted way for a long time, I'd really prefer to
>>>>>> keep that (for now).
>>>>> It has.  Very true.  But I don't necessarily think that means we should
>>>>> be introducing even more of 'em.
>>> [...]
>>>> Happily, normalizing into ~0 for signed and [1,MAX] for unsigned,
>>>> simplified the patch because it on longer needs tweaks to
>>>> ranges_from_anti_range.
>>>> 
>>>> OK for trunk?
>>> 
>>> the new testcase FAILs on several (all?) 32-bit targets:
>>> 
>>> +FAIL: gcc.dg/tree-ssa/evrp4.c scan-tree-dump evrp "\\\\[1B, -1B\\\\]"
>> 
>> That's unfortunate.
>> 
>> Is this the only test that is failing?
> 
> it's the only on on Solaris/SPARC and Solaris/x86.  Haven't checked
> other affected targets, though.
> 
>>> I'm seeing this on 32-bit i386-pc-solaris2.11 and sparc-sun-solaris2.11,
>>> with more reports for armv8l, pru, and s390x.
>>> 
>>> Comparing the dumps between 64 and 32-bit, I see
>>> 
>>> -_1: int * [1B, -1B]
>>> +_1: int * [1B, 4294967295B]
>> 
>> I wonder why 32-bit targets at displaying 4294967295 instead of -1.  Or are
>> pointers 64-bits here?
> 
> No, it's a pure 32-bit target.  The compiler is 32-bit, too, but
> bi-arch (32 and 64-bit).

identical result on i686-darwin9, also a pure32b target (with a 64b multilb).
Iain