JIT: preliminary version of profile-based inline policy

[AndyAyersMS]

Add a new inline policy that can be used when a method has profile data.

It uses the call site frequency to boost profitability. Size and per-call
benefit are currently using the estimates done by the model policy.

Not on by default. Set COMPlus_JitInlinePolicyProfile=1 to enable.
Add testing to weekly experimental runs.

[AndyAyersMS]

cc @dotnet/jit-contrib

[EgorBo]

So it fixes #41923 (I've just checked) 🎉

Also, is it expected that it's now able to inline methods with more than MAX_BASIC_BLOCKS bbs?
because I've just tested this:

string Foo() => "hello";

[MethodImpl(MethodImplOptions.NoInlining)]
bool DoWork()
{
    return Foo() == "hello";
}

Currently, inliner gives up here because string.Equals(string,string) has 7 basic blocks and ends up with this for DoWork:

G_M53056_IG01:
       sub      rsp, 40
						;; bbWeight=1    PerfScore 0.25
G_M53056_IG02:
       mov      rcx, 0xD1FFAB1E
       mov      rdx, gword ptr [rcx]
       mov      rcx, rdx
       call     System.String:Equals(System.String,System.String):bool
       nop      
						;; bbWeight=1    PerfScore 3.75
G_M53056_IG03:
       add      rsp, 40
       ret

Your branch (with profile):

G_M53056_IG01:              ;; offset=0000H
                                                ;; bbWeight=1    PerfScore 0.00
G_M53056_IG02:              ;; offset=0000H
       B801000000           mov      eax, 1
                                                ;; bbWeight=1    PerfScore 0.25
G_M53056_IG03:              ;; offset=0005H
       C3                   ret

(basically, fixes #33338)

[BruceForstall]

One fix needed in config value string, otherwise LGTM

[BruceForstall]

nit: copy-paste comment

[BruceForstall]

String is wrong here: JitInlinePolicyProfile => JitInlinePolicyProfileThreshold

[AndyAyersMS]

Thanks for spotting this.

[AndyAyersMS]

is it expected that it's now able to inline methods with more than MAX_BASIC_BLOCKS bbs

Yes. I'm aiming to have very few arbitrary early limits on what might be inlined, to let the profile data and heuristics determine what's best.

We will need to have some limit on the max IL size we'll consider so that we don't spend time analyzing large methods that can't possibly be good inlines. Not sure what the right value for that limit is just yet; I have it set to 1000 bytes here.

[AndyAyersMS]

Failure seems unrelated; none of this code should be enabled by default.