[lto] implement awful types_compatible_p hack

Nathan Froyd froydnj@codesourcery.com
Sat Dec 22 21:57:00 GMT 2007


On Sat, Dec 22, 2007 at 11:23:17AM -0800, Chris Lattner wrote:
> On Dec 22, 2007, at 11:08 AM, Nathan Froyd wrote:
> >253.perlbmk now compiles (with -fwrapv) and runs properly.  This  
> >leaves
> >a miscompile of 176.gcc as the only remaining C SPECint benchmark that
> >LTO doesn't work with.
> 
> Cool, what sort of speedups are you seeing?

I haven't done any measurement yet.  I think the binaries are a little
smaller (my evidence is slim--I was debugging and noticed that the LTO
dump files inlined in some places where the standoalone compiler
couldn't), but nothing definite.  I will attempt to get some numbers
next week.

I will also note that I have several hacks in my SPEC tree from when I
was in LTO Shall Only Compile Perfect (e.g. always-prototyped) Code
mode.  Some of the necessary changes to decl merging abanded that goal.
I shall have to remove those hacks and see whether we need more code to
handle what my hacks were supposed to have fixed.

-Nathan



More information about the Gcc-patches mailing list