| 1 |
Eduardo Schoedler wrote: |
| 2 |
|
| 3 |
<SNIP> |
| 4 |
> |
| 5 |
> Hi Beso. |
| 6 |
> |
| 7 |
> http://gcc.gnu.org/onlinedocs/gcc-4.3.2/gcc/i386-and-x86_002d64-Options.html#i386-and-x86_002d64-Options |
| 8 |
> |
| 9 |
> |
| 10 |
> "native |
| 11 |
> This selects the CPU to tune for at compilation time by determining |
| 12 |
> the processor type of the compiling machine. Using -mtune=native will |
| 13 |
> produce code optimized for the local machine under the constraints of |
| 14 |
> the selected instruction set. Using -march=native will enable all |
| 15 |
> instruction subsets supported by the local machine (hence the result |
| 16 |
> might not run on different machines). " |
| 17 |
> |
| 18 |
> Best Regards, |
| 19 |
> Eduardo. |
| 20 |
> |
| 21 |
|
| 22 |
I have Phenom 9950 and use gcc-4.3.2 with CFLAGS="-march=barcelona |
| 23 |
-combine -O2 -pipe" |
| 24 |
|
| 25 |
I tried to compile some small test code with -v -march=native and -v |
| 26 |
-march=barcelona, just to see if there is any difference. |
| 27 |
Compiled code seems to be the same, but that means nothing, since the |
| 28 |
example used was trivia |
| 29 |
|
| 30 |
But there is one difference in gcc's coments. |
| 31 |
|
| 32 |
With -march=native it says: |
| 33 |
|
| 34 |
.. |
| 35 |
COLLECT_GCC_OPTIONS='-v' '-O3' '-pipe' '-o' 'tt2' |
| 36 |
/usr/libexec/gcc/x86_64-pc-linux-gnu/4.3.2/cc1 -quiet -v test.c |
| 37 |
-march=amdfam10 -mcx16 -msahf --param l1-cache-size=64 --param |
| 38 |
l1-cache-line-size=64 -mtune=amdfam10 -quiet -dumpbase test.c -auxbase |
| 39 |
test -O3 -version -o - | |
| 40 |
... |
| 41 |
|
| 42 |
...and with -march=barcelona: |
| 43 |
|
| 44 |
... |
| 45 |
COLLECT_GCC_OPTIONS='-v' '-march=barcelona' '-O3' '-pipe' '-o' 'tt1' |
| 46 |
/usr/libexec/gcc/x86_64-pc-linux-gnu/4.3.2/cc1 -quiet -v test.c -quiet |
| 47 |
-dumpbase test.c -march=barcelona -auxbase test -O3 -version -o - | |
| 48 |
.. |
| 49 |
|
| 50 |
whereas with -march=amdfam10: |
| 51 |
|
| 52 |
... |
| 53 |
COLLECT_GCC_OPTIONS='-v' '-march=amdfam10' '-O3' '-pipe' '-o' 'tt1' |
| 54 |
/usr/libexec/gcc/x86_64-pc-linux-gnu/4.3.2/cc1 -quiet -v test.c -quiet |
| 55 |
-dumpbase test.c -march=amdfam10 -auxbase test -O3 -version -o - | |
| 56 |
... |
| 57 |
|
| 58 |
|
| 59 |
It seems that "barcelona","amdfam10" and "native" are not exactly |
| 60 |
synonymous on my machine. "amdfam10" seems to offer broadest description |
| 61 |
to which "barcelona" adds some extra details, but "native" seems to be |
| 62 |
most detailed,a s it adds details about cache size and cache line size, |
| 63 |
CMPXCHNG16B and LAHF/SAHF instructions. |
| 64 |
|
| 65 |
I also tried to compile "unrar" with those variations, but was unable to |
| 66 |
make difference in final binary- which admittedly doesn't prove anything... |
Archives