Il 17/04/2014 15:44, Michael S. Tsirkin ha scritto:
I'm fine with this.
And I'm not against a runtime switch to get rid of per-target build of virtio.
I am merely asking for separatable patchsets, e.g. structured like this:
1. implement bi-endian support with no data path overhead for fixed endian
targets
2. cleanup getting rid of per-target build adding minor data path overhead
This way down the line if we see performance regressions it's easy to
revert and verify what causes it.
An alternative is accompanying patchset with performance benchmarking
results.
There is no such thing as complete benchmarks, and splitting the series
in two as you suggest is cleaner (in addition to more clearly
bisectable), so I'd really go that way.
Paolo