Re: [RFT/RFA] Fix AIX fallout from PR/19653 patch

[Geoff Keating <geoffk at geoffk dot org>]

On 04/04/2006, at 10:52 AM, David Edelsohn wrote:

> 	I will test your patch.  This machinery originally was designed
> and implemented by Geoff, so I would appreciate his comments on your
> analysis of a latent bug.

Actually, the constant_pool_expr_p routine was implemented by Clinton  
Popetz.

His comment, in <http://gcc.gnu.org/ml/gcc-patches/2000-02/ 
msg00223.html>, about these routines I think, was:

> (3) Adds some code to differentiate expressions that reference the
> constant_pool as opposed to those that reference the constant_pool via
> the TOC, which is needed because CSE and reload can try to turn  
> strange
> things into addresses, and input_operand and the force_const_mem  
> code in
> movsi were accepting too much.

I don't see how the proposed patch fixes the problem described.  My  
guess is that it simply causes constant_pool_expr_p to return false  
in the particular cases that are being encountered, suppressing but  
not fixing the problem.

I suspect that each use of constant_pool_expr_p immediately followed  
by ASM_OUTPUT_SPECIAL_POOL_ENTRY_P (get_pool_constant (op), ...),  
that is all but one, should actually be

  GET_CODE (op) == SYMBOL_REF
  && CONSTANT_POOL_ADDRESS_P (op)
  && ASM_OUTPUT_SPECIAL_POOL_ENTRY_P (get_pool_constant (op),  
get_pool_mode (op))

There are three places that should be changed, I think.

Attachment: smime.p7s
Description: S/MIME cryptographic signature