Faster C.Run by using less reflect

[dolmen]

In C.Run bypass check of the signature of the Run method of c.TB by using a prefilled table for the common cases of *testing.T, *testing.B and *quicktest.C.

This serie of patches shows the developement process:

• refactor C.Run to extract getRunFuncSignature
• add unit test for getRunFuncSignature
• add a dummy implementation of a caching version (which does no caching at this point) of getRunFuncSignature and add a BenchmarkCRunGetFuncSig to compare them
• implement the real version of getRunFuncSignatureCache (see commit for results)
• enable the use of getRunFuncSignatureCache instead of getRunFuncSignature in C.Run

$ go test -bench BenchmarkCRunGetFuncSig -benchmem
    BenchmarkCRunGetFuncSig/*testing.T-no-cache-4          562579   2142     ns/op   168 B/op             4 allocs/op
    BenchmarkCRunGetFuncSig/*testing.T-with-cache-4     183638598      7.514 ns/op     0 B/op             0 allocs/op
    BenchmarkCRunGetFuncSig/*testing.B-no-cache-4          608715   2626     ns/op   168 B/op             4 allocs/op
    BenchmarkCRunGetFuncSig/*testing.B-with-cache-4     100000000     11.00  ns/op     0 B/op             0 allocs/op
    BenchmarkCRunGetFuncSig/*quicktest.C-no-cache-4        636226   2523     ns/op   168 B/op             4 allocs/op
    BenchmarkCRunGetFuncSig/*quicktest.C-with-cache-4    68991796     48.91  ns/op     0 B/op             0 allocs/op

[dolmen]

@frankban Please merge #157 first. I'll do the rebase of #160 on top.

[dolmen]

@frankban Rebased.

This important optimization (with my other fixes) are worth a release.

[dolmen]

ping @frankban

[frankban]

This looks very interesting. We are discussing with @rogpeppe about other possible ways for introducing this optimization. Thanks for this work!

[dolmen]

I could go further and completely remove the getRunFuncSignature. At this point the cache is statically filled and getRunFuncSignature is only used at runtime for error reporting in the case of a bad usage (and also for the benchmark that would go too): this could be replaced by a simple message displaying the signature of the received function and the list of acceptable ones.

An alternate implementation would use generics to enforce the signature at compile time. But are you ready to push the minimum Go version required from 1.13 to 1.18?

[dolmen]

Well, Go doesn't have generic methods yet, so I don't see how a compile-time check of the subtest signature could be enforced.

[dolmen]

Rebased on top of v1.14.6.

[dolmen]

Replaced by #165 which completely bypass reflect in the common cases for which we used a cache here.