Re: [RFC] Don't completely scalarize a record if it contains bit-field (PR tree-optimization/45144)

[Jie Zhang <jie at codesourcery dot com>]

On 08/03/2010 05:00 PM, Richard Guenther wrote:
> On Mon, Aug 2, 2010 at 6:52 PM, Mark Mitchell<mark@codesourcery.com>  wrote:
> > Jie Zhang wrote:
> >
> > > My patch prevents several full scalarizations of records with bit-field
> > > when compiling Linux kernel for x86_64, but none of these causes
> > > differences in final assemblies. I use 2.6.34.1 and the default config
> > > for x86_64. I checked -O2 and -Os.
> >
> > That seems at odds with the statement made previously in this thread
> > that this optimization was essential for Linux kernel performance.
> >
> > If Jie's statement is accurate, then, whether or not this is a "hack",
> > it seems like a win.  I don't see anything wrong with accepting a small,
> > local improvement that has no user-observable negative impact; we can
> > always rip it out and replace it with something better when something
> > better exists.
>
> OTOH no changes in code generation are also not in favor of this
> patch.  Why didn't it improve anything?  Or was that expected?
Only two total scalarizations are prevented by my patch when building 
the linux kernel for the default x86_64 config, which compiles 1646 C files.

One is in io_apic.c:ioapic_write_entry (), "struct IO_APIC_route_entry e".

The other is in fs-writeback.c:bdi_sync_writeback (), "struct 
wb_writeback_args args".

In both cases, those structs are used simply as a whole. GCC can already 
optimize away the parts generated by total scalarization. So there is no 
difference when it's disabled.

> Can you adjust the patch according to Martins suggestion?
OK.


Thanks,
--
Jie Zhang
CodeSourcery