Re: Additional patch to further improve delete_computation.

[Jeffrey A Law <law at cygnus dot com>]

  In message <199908082226.SAA26637@jwlab.FEITH.COM>you write:
  > This patch allows delete_computation to remove calls to constant
  > functions.  For example using -O2 -fno-schedule-insns2 on a x86
  > platform to compile:
  > 
  >   long
  >   func(long long a, long long b)
  >     {
  >     long long c;
  > 
  >     c = a / b;
  > 
  >     if (c)
  >       return 0;
  > 
  >     c -= c;
  > 
  >     return c;
  >     }
  > 
  > without the patch produces:
  > 
  > 	pushl %ebp
  > 	movl %esp,%ebp
  > 	subl $8,%esp
  > 	movl 16(%ebp),%eax
  > 	movl 20(%ebp),%edx
  > 	pushl %edx
  > 	pushl %eax
  > 	movl 8(%ebp),%eax
  > 	movl 12(%ebp),%edx
  > 	pushl %edx
  > 	pushl %eax
  > 	call __divdi3
  > 	xorl %eax,%eax
  > 	leave
  > 	ret
  > 
  > with the patch:
  > 
  > 	pushl %ebp
  > 	movl %esp,%ebp
  > 	subl $8,%esp
  > 	movl 16(%ebp),%eax
  > 	movl 20(%ebp),%edx
  > 	pushl %edx
  > 	pushl %eax
  > 	movl 8(%ebp),%eax
  > 	movl 12(%ebp),%edx
  > 	pushl %edx
  > 	pushl %eax
  > 	xorl %eax,%eax
  > 	leave
  > 	ret
  > 
  > This patch passes make bootstrap and make check on FreeBSD-3.2 x86.
  > 
  > Notes:
  > 
  >   1) Unfortunately at the moment I don't see a straight forward way
  >      to avoid pushing the arguements on the stack (helpful little
  >      things such as REG_LIBCALL and REG_RETVAL notes are not around),
  >      however with enough analysis it's probably doable.
  > 
  > ChangeLog:
  > 
  > Sun Aug  8 00:03:55 EDT 1999  John Wehle  (john@feith.com)
  > 
  > 	* jump.c (delete_prior_computation): Also check calls
  > 	to constant functions.  Don't bother checking for a
  > 	REG_UNUSED note before adding it.
  > 	(delete_computation): Handle multi-word hard registers
  > 	when synthesizing missing REG_DEAD notes for a register
  > 	which is both set and used by an insn.
I think we've already got the multi-word hard register stuff installed and
note synthesis installed.  Right?

Can you resubmit with just the new stuff.  I suspect it'll be fine, but it's
just easier for me to read and evaluate.

On the subject of killing argument stores -- for machines which do not push
args, but instead use reg+d stores, local dead code elimination has a 
reasonable
chance at eliminating the useless stores.  Global dead store elimination is in
the works, which should allow us to do a pretty good job on those targets.
[ Global dead store elimination is just gcse/lcm run on the reverse flow
  graph :-) ]

For machines which do push arguments, one could write a relatively simple
optimizer that looked for arg pushes followed by a stack pointer adjustment
without a call separating them and eliminate the pushes & stack pointer
adjustment.  That (of course) would be a separate patch :-)

jeff