JIT: Switch to rangecheck when profitable

[BoyBaykiller]

https://godbolt.org/z/ovnb41bfv

The current switch normalization logic (inserting GT_SUB) is purely correctness driven to fit the jump table. Consider:

static bool Test1(int a)
{
    return a == 300 || a == 301 || a == 302;
}
static bool Test2(int a)
{
    return a == 3 || a == 4 || a == 5;
}

Test1 would be normalized by switch-recognition because it has numbers larger than SWITCH_MAX_DISTANCE while Test2 wouldn't. This difference would lead to one being collapsed into range-check while the other wouldn't:

G_M12381_IG02:  ;; offset=0x0000
       sub      ecx, 300
       cmp      ecx, 3
       setb     al
       movzx    rax, al

G_M1246_IG02:  ;; offset=0x0000
       cmp      ecx, 5
       ja       SHORT G_M1246_IG06
						;; size=5 bbWeight=1 PerfScore 1.25
G_M1246_IG03:  ;; offset=0x0005
       mov      eax, 7
       bt       eax, ecx
       jb       SHORT G_M1246_IG06
						;; size=10 bbWeight=0.97 PerfScore 1.70
G_M1246_IG04:  ;; offset=0x000F
       mov      eax, 1
						;; size=5 bbWeight=0.25 PerfScore 0.06
G_M1246_IG05:  ;; offset=0x0014
       ret      
						;; size=1 bbWeight=0.25 PerfScore 0.25
G_M1246_IG06:  ;; offset=0x0015
       xor      eax, eax
						;; size=2 bbWeight=0.25 PerfScore 0.06
G_M1246_IG07:  ;; offset=0x0017
       ret      

Here I am just adding an additional check to normalize the switch whenever the numbers form a continous range which fixes that.

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[BoyBaykiller]

Unfortunately, this identity check is needed as nothing after "recognize switch" phase optimizes out the SUB(x, 0)

[BoyBaykiller]

Our canonical form would be Add(switchValue, -minValue), should I use that instead?
Then the if-chain shown in the PR description would match codegen for manual switch

[BoyBaykiller]

@AndyAyersMS PTAL. Simple change. diffs