On 01/29/2012 02:14 PM, Mike Frysinger wrote:
On Saturday 28 January 2012 07:26:59 Anthony G. Basile wrote:
I've run nbench on two amd64 systems both running the same kernel
vanilla-3.2.2.
i don't think nbench is a good benchmark for this as it isn't really testing
what you think it's testing. it's very good at validating math support in the
ISA/ABI, optimized compiler output, and supplementary math implementations in
libgcc. PIE vs non-PIE will still be able to multiply/divide in pretty much
the same amount of time.
I know, but the problem is, what benchmark best approximates common
every day use? So I wrote the following which really hits the problem
hard on x86:
int modfac(int n)
{
if(n==0) return 1;
return n * modfac(n-1);
}
int main()
{
int i;
for( i = 0 ; i < 4096*4096 ; i++ ) modfac(4096);
return 0;
}
Using vanilla kernel 3.2.2, userland built with vanilla toolchain,
gcc-4.5.3-r1, glibc-2.13-r4, binutils-2.21.1-r1, compiling my code
simply as gcc -o test modfac.c, CFLAGS="-O2 -march=i686 -pipe" I get:
time -p ./test
real 327.89
user 327.72
sys 0.00
Keep everything else the same, even the same hardware, but switch to
userland built with hardened gcc-4.5.3-r2 (not -r1 because of the bus
error), I get:
time -p ./test
real 629.68
user 629.37
sys 0.00
The hardware is 8 x "Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz" with 12 GB
ram. That's nearly a factor of 2x but how often does one set up 4k
stack frames in everyday use?
So at least on amd64, I don't think that performance is ever an issue.
yes, most likely on systems where the PIC has hardware support in the ISA, the
performance hit on PIE is typically low.
I have yet to look at x86.
pretty sure this is going to be much more palpable.
-mike
--
Anthony G. Basile, Ph.D.
Gentoo Linux Developer [Hardened]
E-Mail : bluen...@gentoo.org
GnuPG FP : 8040 5A4D 8709 21B1 1A88 33CE 979C AF40 D045 5535
GnuPG ID : D0455535
