Re: PING^5: [PATCH] rtl-optimization/110939 Really fix narrow comparison of memory and constant

[Eric Botcazou <botcazou@adacore.com>]

> This is why I got a bit uncertain and hoped to get some feedback whether
> my intuition is correct or not.  Meanwhile I also found a comment in
> the internals book at "14.7 Constant Expression Types" where we have:
> 
>    "Constants generated for modes with fewer bits than in HOST_WIDE_INT
>     must be sign extended to full width (e.g., with gen_int_mode).
>     [...]
>     Note however that values are neither inherently signed nor
>     inherently unsigned; where necessary, signedness is determined by
>     the rtl operation instead."
> 
> At least this and the assert statement document that the normal form of
> a CONST_INT is kind of special w.r.t. unsigned integers.  Is there
> anyone who can shed some light on _why_ such a normal form was chosen?

In RTL integral values have no sign, they just represent a given pattern of 
bits so, in order to have a 1-to-1 mapping, you need to choose a canonical 
form.  The signed form is probably more natural and, since CONST_INTs have no 
mode, the same objects are used for e.g. QImode and HImode, which means that 
you need to sign-extend the bit pattern.

-- 
Eric Botcazou