initial cut at guarded devirt perf tests

[AndyAyersMS]

I can see these are getting compiled but they don't show up in listing or filtering.

So not sure what I'm doing wrong.

@adamsitnik can you take a look?

[AndyAyersMS]

This more or less reproduces the first couple of graphs in the guarded devirtualization writeup, which I measured in a more ad-hoc fashion.

Perf for the "really bad guess" case (pB = 0) shows a somewhat higher relative loss, but the crossover point is for GDV still somewhere between 0.0 and 0.1....

Method	testInput	no GDV	GDV
Call2	pB = 0.00	1.894 ns	2.321 ns
Call2	pB = 0.10	3.191 ns	2.797 ns
Call2	pB = 0.20	4.036 ns	2.836 ns
Call2	pB = 0.30	4.950 ns	2.731 ns
Call2	pB = 0.40	6.032 ns	2.490 ns
Call2	pB = 0.50	6.381 ns	2.305 ns
Call2	pB = 0.60	5.382 ns	1.964 ns
Call2	pB = 0.70	5.146 ns	1.886 ns
Call2	pB = 0.80	3.735 ns	2.442 ns
Call2	pB = 0.90	2.854 ns	1.900 ns
Call2	pB = 1.00	1.888 ns	0.893 ns

I have similar tests for interface calls that I'll work on converting over. Measuring interface call impact is a bit trickier because VSD has some residual state and so we can't use the same call site for different test inputs.

[AndyAyersMS]

Added the interface versions. Simpler than I thought as BDN seems to run each case in its own process so there is no VSD carryover (one can tell by looking at the no GDV two class results, they will not be symmetric about p=0.5 if VSD is carrying state over from the previous input array).

Interface results:

Method	testInput	noGDV	GDV
Call2	pB = 0.00	2.480 ns	2.8002 ns
Call2	pB = 0.10	4.602 ns	3.2773 ns
Call2	pB = 0.20	6.162 ns	2.9830 ns
Call2	pB = 0.30	7.262 ns	2.8617 ns
Call2	pB = 0.40	8.396 ns	2.7793 ns
Call2	pB = 0.50	8.705 ns	2.3494 ns
Call2	pB = 0.60	8.726 ns	2.1261 ns
Call2	pB = 0.70	7.557 ns	1.9230 ns
Call2	pB = 0.80	5.930 ns	1.7155 ns
Call2	pB = 0.90	4.482 ns	1.5456 ns
Call2	pB = 1.00	2.466 ns	0.7914 ns

[AndyAyersMS]

@adamsitnik think these are ready

[adamsitnik]

LGTM! Thanks for contributing new benchmarks to the perf repo!