Re: profiling on sh

[mike stump <mrs at windriver dot com>]

> To: "Kris Warkentin" <kewarken@qnx.com>
> Cc: <gcc@gcc.gnu.org>
> From: Alexandre Oliva <aoliva@redhat.com>
> Date: 16 Mar 2002 16:00:27 -0300

>   /* In C++, we never create builtin functions whose name does not
>      begin with `__'.  Users should be using headers to get prototypes
>      in C++.  It would be nice if we could warn when `-fbuiltin' is
>      used explicitly, but we do not have that information.  */
>   flag_no_builtin = 1;

I disagree fundamentally with this change.  The main thing I don't
like about it, is that it kills performance:

#include <memory.h>

int a, b;

int main() {
    memcpy(&a, &b, sizeof(a));
}

main:
	sethi	%hi(a), %g1
	or	%g1, %lo(a), %o0
	sethi	%hi(b), %g1
	or	%g1, %lo(b), %o1
	call	memcpy, 0
	mov	4, %o2

v:

	sethi	%hi(b), %g1
	ld	[%g1+%lo(b)], %i2
	sethi	%hi(a), %g1
	st	%i2, [%g1+%lo(a)]

This is unacceptable.  Remove non-standard builtins, if you must with
-ansi -pendantic.  Heck, even remove specific unwanted builtins as you
want.  Removing memcpy, in a normal compile, isn't what we want.  That
should only be removed with -fno-builtin, or with -freestanding, or
whatever they call it this week.

What issue were you trying to resolve with this change Mark?  Did you
mean to kill performance of some apps?  :-(

Also, I don't think we should remove alloca from being a builtin with
a normal compile.