Re: bundle boundaries in ia64 assembly output?

[Geert Bosch <bosch at gnat dot com>]

On Jan 24, 2005, at 08:29, Jan Beulich wrote:
> There was a discussion started at
> http://gcc.gnu.org/ml/gcc/2003-12/msg00821.html and leading to the
> idea of adding bundle boundaries (not just bundle types as is still
> getting done) to the assembly output. Searching gcc-patches, however,
> doesn't even seem to locate a rejected or never approved patch, and
> hence I wonder whether implementation of this was ever attempted.

We did try this approach, and inserted proper markers for the begin
and end of bundles. However, for our goal of being able to use the
Intel assembler this was not enough. GCC puts labels, mainly for EH,
in arbitrary locations. The published Intel assembly syntax requires
use of instruction tags instead. These tags have a different syntax.

Fixing this issue would require changes outside of the IA-64 specific
files, as EH symbols and tags get added too late. Because of these
issues and the impression no one else was interested in having
GCC produce assembly in the format documented by Intel, we are now
considering porting GNU as to IVMS after all.

> The reason I'm asking is not because of using ias as the consumer of
> the assembly, but instead to have an additional level of verification 
> of
> the generated output; with a recently submitted patch (don't recall by
> whom) I noticed that bundling in .s and .o differed, and it took me a
> while to realize that this was because an insn had a wrong attribute,
> making the compiler schedule it on the wrong unit. Such problems go by
> unnoticed because without bundle boundaries the assembler silently
> splits in two bundles what was intended to go into one.

I think it would be good to have the explicit bundling, as I distrust
"smart" assemblers. Ideally, GCC would directly output object files,
and optionally object file listings for debugging purposes. Failing 
that,
generating unambiguous assembly code is a reasonable alternative.

I'll see if I can dig up the patch for the explicit bundle markers, if
you're interested.

  -Geert