Re: optimization based on CPU type or flags?

[Ryan Hill <dirtyepic at gentoo dot org>]

On Wed, 3 Apr 2013 20:49:40 -0700 (PDT)
Hei Chan <structurechart@yahoo.com> wrote:

> Hi,
> 
> I noticed that I can specify -march=native and -mtune=native.
> 
> I have few questions:
> - On some boxes, I noticed that if I specify -march=native and -mtune=native,
> g++ might "convert" -march=native into -march=corei7-avx along with a list of
> flags such as -mcx16 -msahf -mavx -msse4.2 --param l1-cache-size=32 --param
> l1-cache-line-size=64 --param l2-cache-size=10240 (there are a lot more flags
> but I have to take them out as GCC mailing server marked my mail as spam if I
> listed all), and -mtune=generic.  how come gcc recognize the CPU type as
> corei7-avx but it can't recognize the CPU type for -mtune?

It turns out that for some cpus -mtune=generic can produce better code than
-mtune=<cpu>. Mostly for cpus that don't have a specifically tuned cost model
IIUC.

> - Let's say g++ figures that my CPU support sse4.1, and how does it choose
> between sse4.1 instructions and other set of instructions (assuming both
> achieve the same task)?  Does g++ have a database that it computes the
> numbers of CPU cycles for all the possible combinations before it decides?

Yes there are cost models that are used to decide what instructions to use.  See
gcc/config/i386/i386.c for example.

> - Also, how does the l1-cache-size and l2-cache-size parameters affect the
> optimization?  Only affecting the decision of inline?  Or there are more?

Not sure on that one.

> - Is there any tool that will tell me what CPU specific instructions are used
> (other than going through the object files manually with objdump one by one)?

https://github.com/dirtyepic/scripts/blob/master/analyze-x86 is a clumsy
attempt.  It probably misses a lot and I haven't updated it in a couple
releases though.  If there's another tool like this out there I'd definitely be
interested.


-- 
gcc-porting
toolchain, wxwidgets                                     by design, by neglect
@ gentoo.org                                     for a fact or just for effect

Attachment: signature.asc
Description: PGP signature