Fix for "Too restrictive sanity check"

[Jeffrey A Law]

 In message < 20000419130601.A11198@atrey.karlin.mff.cuni.cz >you write:
  > Hi
  > Interestingly this testcase exhibits two unrelated bugs.  First one is in
  > my sanity checking.  It is done for sibbling calls too, where it is incorre
  > ct
  > since my last calls.c patch (it did before).
  > 
  > Second is stack missalignment, that don't show actually in the code (since
  > another sequence is choosed by sibcall), but the purpose is obvious:
  > first compute_argument_block_size is used to caluculate size of argument
  > block to get proper alignment once they are pused and then precompute_argum
  > ents
  > is called that expands another call resulting in missaligning stack (and
  > adjusting stack_pointer_delta).
  > 
  > The fix is to call compute_argument_block_size later.  There are two
  > uses of args_size in between.  First one is in precompute_arguments.
  > I have patch approved to avoid this, since it is useless. Second is in
  > finalize_must_preallocate.  I've did same change for this code in my home
  > tree.  The rationale is, that finalize_must_preallocate use args_size to
  > decide whether it is better to precompute arguments (and direct structures
  > to it) or not precompute (and require pushing).  Ratio of structs over
  > args size is calculated and when it exceeds 50%, preallocation is done.
  > 
  > Author had unadjusted args size in mind here, so in fact this is third
  > bug fixed by this patch.
  > 
  > Alexandre: this obsolettes you patch, but please try to get approval for
  > the testcase, it is definitly interesting from the calls.c point of
  > view. Thank you for pointing this out.
  > 
  > Honza
  > 
  > Wed Apr 19 12:54:31 MET DST 2000  Jan Hubicka  <jh@suse.cz>
  > 	* calls.c (expand_call): Call compute_argument_block_size right
  > 	before allocating the block; update comment; don't do alignment
  > 	sanity checking for sibbling call; use args_size instead of
  > 	unadjusted_args_size before args_size is adjusted.
This patch is clearly wrong.

compute_argument_block_size modifies the data in args_size which needs to
happen before the call to finalize_must_preallocate.

Please do _not_ install this patch.

jeff