| 1 |
Tomorrow my new Core 2 workstation should be arriving, and reading |
| 2 |
through the 2006.1 amd64 install notes and so forth, the recomended |
| 3 |
CFLAGS for EM64T CPUs is -march=nocona. However as that is tuned for |
| 4 |
64-bit enabled Netburst rather than Conroe, I'm wondering if anyone |
| 5 |
had thoughts on good flags. In particular as (to the extent that I |
| 6 |
know anything about microarchitecture), Conroe seems closer to a K8 |
| 7 |
than a P4, especially with regards to pipeline length and execution |
| 8 |
resources, I'm not sure that P4/Netburst tuning is the right thing to |
| 9 |
do. But the GCC docs say that -march=k8 enables 3dnow, which Intel |
| 10 |
chips don't have. So, right now the "seems likely" options would |
| 11 |
be: |
| 12 |
|
| 13 |
-march=k8 -mno-3dnow -msse3 |
| 14 |
-march=nocona |
| 15 |
-march=pentium-m -m64 [saw this suggested in the forums, seems like a bad idea] |
| 16 |
|
| 17 |
Or can I get away with just something like |
| 18 |
|
| 19 |
-mtune=k8 -msse3 (will this get me 64-bit code?) |
| 20 |
|
| 21 |
To be clear, my main goals here are a) enable generation of all |
| 22 |
instructions the CPU has, b) keep GCC from generating 3dnow,etc |
| 23 |
instructions, and c) get instruction scheduling that is at least |
| 24 |
moderately decent for the uarch I'm using. |
| 25 |
|
| 26 |
Since SSE/SSE2, -mfpmath=sse, and argument passing via registers are |
| 27 |
default in 64-bit mode, I suspect the only other options I might try |
| 28 |
are -momit-leaf-frame-pointer, -frename-registers, and |
| 29 |
-finline-functions. |
| 30 |
|
| 31 |
Ideas? |
| 32 |
|
| 33 |
-Jack |
| 34 |
-- |
| 35 |
gentoo-amd64@g.o mailing list |
Archives