Re: [PATCH] Fix ICE on user view-conversion (v2)

[Richard Henderson <rth at redhat dot com>]

Eric Botcazou wrote:
>   /* As an optimization, reject constants that mips_legitimize_move
>      can expand inline.
>
>      Suppose we have a multi-instruction sequence that loads constant C
>      into register R.  If R does not get allocated a hard register, and
>      R is used in an operand that allows both registers and memory
>      references, reload will consider forcing C into memory and using
>      one of the instruction's memory alternatives.  Returning false
>      here will force it to use an input reload instead.  */
>   if (GET_CODE (x) == CONST_INT && LEGITIMATE_CONSTANT_P (x))
>     return true;

I think this is a bug in the mips port.  And looking closer, in
the bfin and frv ports as well.  Everyone else appears to actually
identify those constants that *can't* go to memory.  In particular,
I don't see CONST_INT, CONST_DOUBLE, or CONST_VECTOR being rejected.

These 3 ports are all assuming that force_const_mem is only called
by reload to push non-legitimate constants to memory.  There are
clearly other applications.  This is one; there are several others.

Another option for these ports is if we add another parameter that
specifies that this is being called from reload.  Though, knowing
reload there's probably a cleaner way to handle that.


r~