Re: [PATCH] Complex MEM adjust_address instead of subreg (PR middle-end/21742)

[David Edelsohn <dje at watson dot ibm dot com>]

>>>>> Richard Henderson writes:

Richard> You shouldn't have had to do this; simplify_gen_subreg is supposed
Richard> to have simplified the subreg of memory into a plain MEM, via 
Richard> adjust_address.

Richard> Someone needs to look to see why this isn't working.

	simplify_gen_subreg() calls simplify_subreg() for most of the
work, including applying adjust_address() to MEM.  That code is protected
by

  if (MEM_P (op)
      && ! mode_dependent_address_p (XEXP (op, 0))
      && ...)

The address is

(plus:SI (reg/f:SI 114 virtual-stack-vars)
    (const_int 33112 [0x8158]))

The PowerPC port considers all addresses with offsets to be
mode-dependent.  The current comment explains this because of Altivec:

"On the RS/6000 this is true of all integral offsets (since AltiVec modes
don't allow them)"

and previously it was because of offsets to multi-word addresses:

"On the RS/6000 this is true if the address is valid with a zero offset
but not with an offset of four (this means it cannot be used as an address
for DImode or DFmode)"

	How can the mode-dependent restriction be addressed?

Thanks, David