working around "QImode can alias everything" missed optimizations

working around "QImode can alias everything" missed optimizations

[Doug Evans]

version: current cvs sources as of around noon today
target: i386-linux

If you make `buf' in this program a short*, gcc -O2 -fomit-frame-pointer
optimizes `foo' just fine.  If `buf' is a char*, gcc does poorly
(I'm guessing because it can't know that buf may point to itself
and QImode values are allowed to alias anything).

So how do I work around this?
How do I rewrite this code so that I get something like the following?

foo:
	movl	4(%esp), %edx
	movl	(%edx), %eax
	movw	$1, (%eax)
	movw	$2, 1(%eax)
	movw	$3, 2(%eax)
	addl	$3, %eax
	movl	%eax, (%edx)
	ret

typedef struct {
    char* buf;
} tc_t;

static inline void
emit_1 (tc_t* tc, int c)
{
    *tc->buf = c;
    ++tc->buf;
}

void
foo (tc_t* p)
{
    emit_1 (p, 1);
    emit_1 (p, 2);
    emit_1 (p, 3);
}

Re: working around "QImode can alias everything" missed optimizations

[Richard Henderson]

On Fri, May 26, 2000 at 03:43:55PM -0700, Doug Evans wrote:
> (I'm guessing because it can't know that buf may point to itself
> and QImode values are allowed to alias anything).

Yep.

> So how do I work around this?

You can't.  :-(



r~

Re: working around "QImode can alias everything" missed optimizations

[Doug Evans]

Richard Henderson writes:
 > On Fri, May 26, 2000 at 03:43:55PM -0700, Doug Evans wrote:
 > > (I'm guessing because it can't know that buf may point to itself
 > > and QImode values are allowed to alias anything).
 > 
 > Yep.
 > 
 > > So how do I work around this?
 > 
 > You can't.  :-(

I was afraid you were going to say that.

Any interest in fixing this [dunno how yet]?
GCC produces great code for non-QImode values (well done!).
It has to be agreed that QImode values shouldn't be penalized (right?).
[not that I would change behaviour of course, but rather that some
new way of specifying a "ya, I'm a QImode (ptr) value but, gosh darn it,
I don't alias everything" value needs to be found.  Alternatively, one
might come up with another way, but this way seems at first glance
like the right way to go.  I tried playing with `restrict' but I
don't know enough about it yet to say whether that has a chance
of working - I reread jfc's original email regarding his `restrict'
patch and he said there'd be problems with inline fns].

Re: working around "QImode can alias everything" missedoptimizations

[Mark Mitchell]

>>>>> "Doug" == Doug Evans <dje@transmeta.com> writes:

    Doug> right way to go.  I tried playing with `restrict' but I
    Doug> don't know enough about it yet to say whether that has a
    Doug> chance of working - I reread jfc's original email regarding
    Doug> his `restrict' patch and he said there'd be problems with
    Doug> inline fns].

I think `restrict' will do what you want -- in theory.  Whether the
GCC implementation (which is, I believe, based more on my work that
John Carr's) will do it or not is unclear.  The current implementation
of restrict is, shall we way, limited.  Information is not propogated
around the compiler as widely as it should be.

--
Mark Mitchell                   mark@codesourcery.com
CodeSourcery, LLC               http://www.codesourcery.com

Re: working around "QImode can alias everything" missed optimizations

[Toon Moene]

Mark Mitchell wrote:

> I think `restrict' will do what you want -- in theory.  Whether the
> GCC implementation (which is, I believe, based more on my work that
> John Carr's) will do it or not is unclear.

The original work on alias analysis by John Carr contained an
implementation of `restrict'.  I know, because when we used his work as
an additional patch to the gcc backend (delivered "with" g77 back in
March 1997), we had to be careful to remove the C frontend stuff that
implemented `restrict' from our patch ...

-- 
Toon Moene - mailto:toon@moene.indiv.nluug.nl - phoneto: +31 346 214290
Saturnushof 14, 3738 XG  Maartensdijk, The Netherlands
GNU Fortran 77: http://gcc.gnu.org/onlinedocs/g77_news.html
GNU Fortran 95: http://g95.sourceforge.net/ (under construction)