Am 09.02.21 um 21:46 schrieb Richard Henderson:
On 2/8/21 2:55 PM, Richard Henderson wrote:
Ok, I've reproduced something on a T3 (gcc102.fsffrance.org).
Running the same code side-by-side vs the T5, I get different results.
Brown paper bag time: the T5 build dir lost the --enable-tcg-interpreter flag,
so was testing tcg native.
Big-endian bug wrt an odd api wart in libffi. Fixed thus.
Thanks for solving this. The patch works for me.
BIOS boot time with qemu-system-i386 is about 41 s (with my code which
lacks thread support and ffi it was 40 s).
With qemu-system-x86_64 it is twice as fast, so it looks like in my last
report where I said that the new code had doubled the speed I compared
different system emulations.
Apropos "slow" TCI: on Apple's M1 it is faster than native TCG on most
of my Intel / AMD machines. And it works with current git master while
native TCG still waits for pending patches which fix the memory access.
Stefan
