Re: x86/x64 JIT assembler for C++

Phil Carmody asked:

beside [...]

> On that note - does anyone know at what point an unconditional
> short jmp forward becomes faster than a sequence of NOPs for
> various popular arch's?

It depends on the CPU's code-prefetch size and decoder capability.

AMD K7/K8 skips up to three NOPS in zero time and run up to 15 Nops
within one cycle if they are within one code-fetch (after any jmp/call/..)
and if this code-fetch wont cross cache bounds.

An 'EB/E9' jump, may take just one cycle or less, but at the jump
target we have to add a full fetch cycle including pre-decode then,
means 2..5 clock-cycles depending on type, size and count of
instructions found there.

During code evaluation, I always have NOPs in the trailing end
of cache-lines and it doesn't affect timing if I jump over half
a line (32 nops), but it speeds things up if I jump over a full
cache-line (usually to the next 64-byte bound).

Don't know how Intel-CPUs behave on this.

__
wolfgang