Re: [PATCH] remove the unsigned_type language hook

["Richard Guenther" <richard dot guenther at gmail dot com>]

On 5/8/07, Eric Botcazou <ebotcazou@adacore.com> wrote:
> > Eric: I know that somebody recently did a gcov run of gnat1 over the entire
> > AdaCore test suite and acats.  Can you check if the copy_nodes in the
> > code in question are ever executed?
>
> I don't trust this stuff too much. :-)  Instead I put 4 gcc_unreachable () in
> the hot spots of gnat_unsigned_type/gnat_signed_type.  It didn't go very far:
>
> ../../xgcc -B../../ -B/usr/gnat/i686-pc-linux-gnu/bin/
> -isystem /usr/gnat/i686-pc-linux-gnu/include
> -isystem /usr/gnat/i686-pc-linux-gnu/sys-include
> -L/cardiff.a/gnatmail/gcc-head/build-cardiff/obj/gcc/../ld -c -O2 -fPIC
> -gnatpgn  g-alleve.adb -o g-alleve.o
> +===========================GNAT BUG DETECTED==============================+
> | Pro 6.1.0w (20070507-head) (i686-pc-linux-gnu) GCC error:                |
> | in gnat_unsigned_type, at ada/utils.c:2321                               |
> | Error detected at g-alleve.adb:1884:59
>
>    function To_Pixel (Source : unsigned_short) return Pixel_16 is
>
>       --  This conversion should not depend on the host endianess;
>       --  therefore, we cannot use an unchecked conversion.
>
>       Target : Pixel_16;
>
>    begin
>       Target.T := Unsigned_1 (Bits (Source, 0, 0)   mod 2 ** 1);
>       Target.R := Unsigned_5 (Bits (Source, 1, 5)   mod 2 ** 5);
>       Target.G := Unsigned_5 (Bits (Source, 6, 10)  mod 2 ** 5);
>       Target.B := Unsigned_5 (Bits (Source, 11, 15) mod 2 ** 5);
>       return Target;
>    end To_Pixel;
>
>
> Without rebuilding the runtime, ACATS is hurt too:
>
> FAIL:   c354002
> FAIL:   c360002
> FAIL:   c420001
> FAIL:   c450001
> FAIL:   c460008
> FAIL:   c460011
> FAIL:   cxf2a01
> FAIL:   cxf2a02
>
>
> However, like Richard G, I don't fully understand the purpose of the code.  Do
> you have any recollections?

Something that would make more "sense" would be that the generic code
substituted
for this langhook has to handle the case that we ask for an unsigned
variant of a
subtype by making sure to make the unsigned subtype a subtype of the unsigned
variant of the original basetype.  So (pseudocode)

 if (TREE_TYPE (type))
   TREE_TYPE (new_type) = unsigned_type_for (TREE_TYPE(type));

Richard.