[gentoo-user] GCC 4.7 and LTO: it works

[gentoo-user] GCC 4.7 and LTO: it works

[Nikos Chantziaras]

I've emerged system and world with gcc-4.7.0 and LTO. I'm posting from 
it right now :-)  It's a KDE system with 1043 packages installed.

I've posted details on how to do this (including info on how to disable 
LTO for specific packages that don't work with it) here:

http://realnc.blogspot.com/2012/06/building-gentoo-linux-with-gcc-47-and.html

Re: [gentoo-user] GCC 4.7 and LTO: it works

[Paul Hartman]

On Mon, Jun 4, 2012 at 3:19 PM, Nikos Chantziaras <realnc@gmail.com> wrote:
> I've emerged system and world with gcc-4.7.0 and LTO. I'm posting from it
> right now :-)  It's a KDE system with 1043 packages installed.
>
> I've posted details on how to do this (including info on how to disable LTO
> for specific packages that don't work with it) here:
>
> http://realnc.blogspot.com/2012/06/building-gentoo-linux-with-gcc-47-and.html

Thanks. I'm not sure if I'm ready to recompile world yet, but we'll
see where curiosity and boredom lead me. :) I have an older, slower
machine that might benefit more from small optimizations. A few
questions:

Do you have any measure of compile times using lto compared to not using it?

Was there any effect on quality of debugging info in the resulting
binaries? I thought I read at some point there was no (or bad) debug
info with LTO. Maybe I'm thinking about clang, though.

Did you use gold or the standard linker?

[gentoo-user] Re: GCC 4.7 and LTO: it works

[Nikos Chantziaras]

On 05/06/12 00:21, Paul Hartman wrote:
> On Mon, Jun 4, 2012 at 3:19 PM, Nikos Chantziaras<realnc@gmail.com>  wrote:
>> I've emerged system and world with gcc-4.7.0 and LTO. I'm posting from it
>> right now :-)  It's a KDE system with 1043 packages installed.
>>
>> I've posted details on how to do this (including info on how to disable LTO
>> for specific packages that don't work with it) here:
>>
>> http://realnc.blogspot.com/2012/06/building-gentoo-linux-with-gcc-47-and.html
> [...]
> Do you have any measure of compile times using lto compared to not using it?

It was pretty obvious without doing any actual measurement: linking is 
slower with LTO.  Large programs even take several minutes for the link 
step.


> Was there any effect on quality of debugging info in the resulting
> binaries? I thought I read at some point there was no (or bad) debug
> info with LTO. Maybe I'm thinking about clang, though.

Didn't notice anything strange yet.  But I suspect that this isn't 
important to begin with; all we need are backtraces.  Thorough debugging 
symbols are not important for emerged packages.


> Did you use gold or the standard linker?

The standard one.  I didn't actually think about the importance of this. 
  Does gold work better with LTO?

Re: [gentoo-user] Re: GCC 4.7 and LTO: it works

[Paul Hartman]

On Mon, Jun 4, 2012 at 4:39 PM, Nikos Chantziaras <realnc@gmail.com> wrote:
>
>> Did you use gold or the standard linker?
>
>
> The standard one.  I didn't actually think about the importance of this.
>  Does gold work better with LTO?

I don't know much about it, but AFAIK gold is supposed to be several
times faster at linking in general, and when using it in combination
with gcc+LTO the compiler actually offloads some of the LTO processing
to the linker which is supposed to be more efficient.

I've never tried it personally, but I just googled and found this
mentioned on the GCC site:

"As an added feature, LTO will take advantage of the plugin feature in
gold. This allows the compiler to pick up object files that may have
been stored in library archives. To use this feature, you must be
using gold as the linker and enable the use of the plugin by compiling
with gcc -fuse-linker-plugin. This will shift the responsibility of
driving the final stages of compilation from collect2 to gold via the
linker plugin."

And in gentoo you can switch to gold as explained on the wiki:
https://wiki.gentoo.org/wiki/Gold

[gentoo-user] Re: GCC 4.7 and LTO: it works

[Nikos Chantziaras]

On 05/06/12 01:37, Paul Hartman wrote:
> On Mon, Jun 4, 2012 at 4:39 PM, Nikos Chantziaras<realnc@gmail.com>  wrote:
>>
>>> Did you use gold or the standard linker?
>>
>>
>> The standard one.  I didn't actually think about the importance of this.
>>   Does gold work better with LTO?
>
> I don't know much about it, but AFAIK gold is supposed to be several
> times faster at linking in general, and when using it in combination
> with gcc+LTO the compiler actually offloads some of the LTO processing
> to the linker which is supposed to be more efficient.

Sounds like it's worth trying, but one thing doesn't look good; Diego's 
blog is full of articles about Gold breakage:

http://blog.flameeyes.eu/tag/gold

Re: [gentoo-user] GCC 4.7 and LTO: it works

[Stefan G. Weichinger]

Am 04.06.2012 22:19, schrieb Nikos Chantziaras:
> I've emerged system and world with gcc-4.7.0 and LTO. I'm posting from
> it right now :-)  It's a KDE system with 1043 packages installed.
> 
> I've posted details on how to do this (including info on how to disable
> LTO for specific packages that don't work with it) here:
> 
> http://realnc.blogspot.com/2012/06/building-gentoo-linux-with-gcc-47-and.html

Gave it a try out of curiosity ... got @system emerged OK now ...

Did you also apply it to compiling the kernel?
Worth it?

Stefan

[gentoo-user] Re: GCC 4.7 and LTO: it works

[Nikos Chantziaras]

On 05/06/12 17:28, Stefan G. Weichinger wrote:
> Am 04.06.2012 22:19, schrieb Nikos Chantziaras:
>> I've emerged system and world with gcc-4.7.0 and LTO. I'm posting from
>> it right now :-)  It's a KDE system with 1043 packages installed.
>>
>> I've posted details on how to do this (including info on how to disable
>> LTO for specific packages that don't work with it) here:
>>
>> http://realnc.blogspot.com/2012/06/building-gentoo-linux-with-gcc-47-and.html
>
> Gave it a try out of curiosity ... got @system emerged OK now ...
>
> Did you also apply it to compiling the kernel?
> Worth it?

Kernel builds and runs fine with 4.7.  But I didn't try to use LTO on 
it.  I actually never try to force any custom CFLAGS on the kernel 
build.  I trust upstream to pick safe and fast ones.

Re: [gentoo-user] Re: GCC 4.7 and LTO: it works

[Stefan G. Weichinger]

Am 05.06.2012 17:54, schrieb Nikos Chantziaras:

> Kernel builds and runs fine with 4.7.  But I didn't try to use LTO on
> it.  I actually never try to force any custom CFLAGS on the kernel
> build.  I trust upstream to pick safe and fast ones.

Good choice.

Yes, I also run gentoo-sources-3.4.0 (waiting for 3.4.1 already)
compiled with gcc-4.7.0 successfully for days now.

S

Re: [gentoo-user] Re: GCC 4.7 and LTO: it works

[Stefan G. Weichinger]

Am 2012-06-05 18:56, schrieb Stefan G. Weichinger:
> Am 05.06.2012 17:54, schrieb Nikos Chantziaras:
> 
>> Kernel builds and runs fine with 4.7.  But I didn't try to use LTO on
>> it.  I actually never try to force any custom CFLAGS on the kernel
>> build.  I trust upstream to pick safe and fast ones.
> 
> Good choice.
> 
> Yes, I also run gentoo-sources-3.4.0 (waiting for 3.4.1 already)
> compiled with gcc-4.7.0 successfully for days now.

had to add

sys-process/procps   (3.3.3)

to my

/etc/portage/package.env

fyi