Re: problem with subregs.

["Tal Agmon" <Tal dot Agmon at nsc dot com>]

Tal Agmon wrote:
>> The compiler can use a pair of HI regs to get a long one but if it uses
a
>> long reg then it cannot refer to it's
>> upper word.

Jim Wilson wrote:
>Try CANNOT_CHANGE_MODE_CLASS.  This has various names depending on what
>gcc version you are looking at.

Thank you, I used the macro CANNOT_CHANGE_MODE_CLASS and it worked on most
cases
(I'm using gcc-3.4-20031029). Here is an example where it fails:

In our architecture pointer size is 32 bits and int size is 16 bits. I
tried to compile a program that
contains the line a = &b - &c;    (a is int) and the compiler issued the
error:

internal compiler error: in ?, at gcc-3.3.2\gcc\function.c:3776
The insn that cause this is:

(insn 350 1020 351 61 007A63F0 (set (reg:HI 79)
        (minus:HI (subreg:HI (reg/v/f:SI 24) 0)
            (subreg:HI (reg/f:SI 17 virtual-stack-vars) 0))) -1 (nil)
    (nil))

The problem here is in the rtl (subreg:HI (reg/f:SI 17 virtual-stack-vars)
0).
The frame pointer in our architecture is a 32 bits reg and changing the
mode here is not allowed
according to the macro.

This bug can be solved by solving 1 or 2:

1. Actually, what I wish to prevent is subregs like (subreg:HI (reg:SI long
register) 2)
                                                     and not subregs like
(subreg:HI (reg:SI long register) 0).
    And by using the macro I prevent both and it cause this bug (and also
some decrease in optimizations).

2. Allowing changing SImode to HImode but with some additional move and
scratch regs.
(reload_in and reload_out did not helped in that because they effects only
moves so they will
not work in the case described here).

Regards,

    Tal.