----- Original Message -----
From: "Jim Paris" <jim@jtan.com>
To: "Justin Carlson" <justinca@ri.cmu.edu>
Cc: <linux-mips@oss.sgi.com>
Sent: Friday, December 07, 2001 2:43 PM
Subject: Re: PATCH: io.h remove detrimental do {...} whiles, add sequence
points, add const modifiers
> > Maybe I missed this, but is there any reason for the patch, other then
> > a personal preference of how to do macros that look like functions?
> > I've seen gcc do strange non-optimal things with functions declared
> > inlines, but I've never seen it generate bad code WRT to do{}while(0)
> > constructs.
> >
> > Unless I'm missing something, this patch looks like a solution in search
> > of a problem...
No Justin. See below.
> In the case of set_io_port_base, I see no real reason. But for the
> out[b,w,l] functions, having the do/while can prevent constructs that
> might otherwise make sense, like
>
> for(i=0;i<10;i++,outb(i,port)) {
> ...
> }
>
> Okay, so it's a bad example, but.. :) Maybe Brad has a better one.
>From drivers/net/wireless/heremes.h:
<snip>
/* Register access convenience macros */
#define hermes_read_reg(hw, off) (inw((hw)->iobase + (off)))
#define hermes_write_reg(hw, off, val) (outw_p((val), (hw)->iobase + (off)))
#define hermes_read_regn(hw, name) (hermes_read_reg((hw), HERMES_##name))
#define hermes_write_regn(hw, name, val) (hermes_write_reg((hw),
HERMES_##name, (val)))
/* Note that for the next two, the count is in 16-bit words, not bytes */
#define hermes_read_data(hw, off, buf, count) (insw((hw)->iobase + (off),
(buf), (count)))
#define hermes_write_data(hw, off, buf, count) (outsw((hw)->iobase + (off),
(buf), (count)))
</snip>
That won't compile with the do {...} while(0)s left in io.h. My patch lets
hermes.h (and all other io code that I've tested) compile.
heremes.h compiles as-is on other platforms. Why should mips snub it for
some dubious value of do {...} while(0)?
Regards,
Brad
|