Don't use exchange in the hot path of the GC

[gbaraldi]

xchg is a very expensive instruction when working with memory, on x86 (intel 11th gen) it has an implied lock and a latency of 19. A false positive here just means a slight amount of extra work so don't do that.

[oscardssmith]

does this make a noticeable difference in benchmarks?

[vchuravy]

@d-netto I think this is correct and I think @gbaraldi had a benchmark example where this saved a few couple of percent.

[gbaraldi]

On a very simple benchmark just to illustrate. But this reproduces

PR
❯ ~/julia4/julia test.jl -t4 (base)
8.333745 seconds (1.36 G allocations: 40.596 GiB, 46.29% gc time, 10.57% compilation time)

Master
❯ ~/julia7/julia test.jl -t4 (base)
8.707281 seconds (1.36 G allocations: 40.596 GiB, 48.79% gc time, 10.14% compilation time)

[d-netto]

Seems mergeable to me.