Re: Interpreters and indirect-branch prediction

On Wed, 13 Nov 2024 8:20:27 +0000, Anton Ertl wrote:

mitchalsup@aol.com (MitchAlsup1) writes:

For something like a bytecode interpreter, the prediction accuracy
of the jump predictor is going to be epically bad.

>
And with "jump predictor", you probably mean the indirect-branch
predictor.
>
On hardware of up to around 2005 where the indirect branches were at
best predicted by BTBs with 2-bit counters, a "bytecode interpreter"
with one shared indirect branch (as generated by the usual
for...{switch...{...}} idiom in C) results in misprediction rates of

80% [ertl&gregg03jilp].

That certainly qualifies as bad.

                        With one a non-shared indirect branch per VM
instruction implementation (as in typical threaded-code
implementations), we measured 50%-61% mispredictions for BTB with
2-bit counters in simulation [ertl&gregg03jilp] and similar results on
actual CPUs.

50% misprediction rate qualifies as bad.

>
We simulated 2-level history-based indirect branch predictors
[ertl&gregg03jilp], and they typically have less than 10%
mispredictions.

10% misprediction rate qualifies as "not so bad".