Are tail calls slower than non-tail-calls?

[littledan]

We have heard different things from different browsers:

• JSC has reported that their PTC implementation makes existing code run faster.
• Microsoft has a bug explaining why their current calling convention makes tail calls slower than ordinary calls Implement Proper Tail Call [Discussion] chakra-core/ChakraCore#796
• V8 has not found a regression on large benchmarks with our PTC implementation, but it might be possible to find a microbenchmark with a regression when flipping the tail call flag on, even with extensive compiler work that we have done to support PTCs in various optimizing modes. We do not see PTCs as a way to improve performance of existing code.
• SpiderMonkey @shu @efaust ?

I know this is being discussed on various different threads, but I hope we can compile results here.

[msaboff]

I think I may have been misquoted. For JSC, we don't find any performance changes good or bad with our PTC implementation. Certainly there are micro benchmarks that one can construct that would show performance deltas either way.

Given the instrumenting we did in JSC to gather PTC usage in the wild, it is easy to understand why we don't see any performance deltas. Our statistics show that PTC only happen ~5% of the time.

Assuming that the ChakraCore team doesn't regress standard call performance with their PTC implementation, the impact of any slowness of making a PTC will only impact overall performance by 1/20 the PTC slowdown given the incidence of actual PTCs we found.

[rossberg]

FWIW, V8 also has found slight performance regressions with TCE activated. They are most visible (5+%) on artificial micro benchmarks, though. For "real-world" code, the regression was usually barely above noise level.