From: "George C. Moschovitis" <gmoscho AT alexander DOT cc DOT ece DOT ntua DOT gr>
Subject: optimisation...
To: djgpp AT sun DOT soe DOT clarkson DOT edu (djgpp)
Date: Wed, 28 Jun 1995 11:36:51 +0300 (EET DST)


Hi there...


  i was experimenting with the asm output of gcc and was quite dissapointed
  with the quality of the code. I compiled with the -O3 option... does this 
  include ABSOLUTELY ALL optimisations ? or should i use more switches.
  And what are the switches for the absolutely best speed optimized code ?
  
  btw here is an example:
  the c++ code:
  
inline __dpmi_memblock	__dpmi_allocate_low_memory(int size)
{
__dpmi_memblock	mb;

  mb.l.segment  = __dpmi_allocate_dos_memory((size>>4)+1,&mb.l.selector);
  mb.l.size     = size;
  return mb;
}

void	VBE_InitBuffer()
{ 
  // Allocate a global buffer for communicating with the VBE.
  TransferBuffer = __dpmi_allocate_low_memory(vbeBUFFERSIZE);

  // remember to free the buffer before exiting !
  atexit(VBE_KillBuffer);
}

  produced the asm output:
	.align 2				; why not align 4 ??
.globl _VBE_InitBuffer__Fv
_VBE_InitBuffer__Fv:
	pushl %ebp
	movl %esp,%ebp
	subl $32,%esp
	leal -24(%ebp),%eax
	pushl %eax
	pushl $65
	call ___dpmi_allocate_dos_memory
	movw %ax,-20(%ebp)			; why is 
	movl $1024,-28(%ebp)			; all this crap 
	movl -32(%ebp),%ecx			; generated ?
	movl %ecx,-16(%ebp)			; an optimizing 
	movl $1024,-12(%ebp)			; compiler should
	movl -24(%ebp),%edx			; leave out this 
	movl %edx,-8(%ebp)			; code
	movl -20(%ebp),%eax			; (like Watcom C/C++ does
	movl %eax,-4(%ebp)			; for example)
	addl $8,%esp				
	movl %ecx,_TransferBuffer
	movl $1024,_TransferBuffer+4
	movl %edx,_TransferBuffer+8
	movl %eax,_TransferBuffer+12
	pushl $_VBE_KillBuffer__Fv
	call _atexit
	leave
	ret
  
  is there anyway i can rearange the c++ code for the compiler to produce
  better output ? or any switch i should use ?	
  if I understand correctly gcc fills the temporary object mb even if it is
  not needed. Maybe I am wrong and I havent run many tests but i preferred
  to mail to this list since there are MANY extremely helpfull guys here...
  And something else. From the (admitedly) limited asm outputs i have seen
  gcc doesnt seem to use register passing that much :( On this topic how can
  i tell the compiler to which registers to pass the parameters ?
  Btw i dont think i have seen any size otpimisation switches, aren't here
  any ?  
  
  tmL-
  
ps: btw sorry for my bad english :(

ps2: i allready found that -fomit-frame-pointer gets rid of the ebp/leave crap..
no need telling me abou this...