JIT: clean up unescaped only stored-to objects

[AndyAyersMS]

We see these in particular after stack allocation. We might be able to clean them up early in the SSA phases.

See comment stream starting here: #84872 (comment)

We also might be able to avoid some of this by recognizing dead allocations in escape analysis, though that might be trickier.

[dotnet-policy-service]

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

[AndyAyersMS]

opportunistic for 10.0

[AndyAyersMS]

There seem to be two categories here to consider:

• Stack allocated objects that are only written to. For normal objects we seemingly do a good job of this now that partial stores are not counted as reads. But for arrays we're not there yet; there is no indexing analog of LCL_FLD/STORE_LCL_FLD so the analysis is more involved and needs to trigger off of IND/STORE_IND understanding what is being read from or written to.

For example

    public static void Array()
    {
        int[] a = new int[10];
        a[3] = 33;
    }

produces

; Method DeadStoreExamples:Array() (FullOpts)
G_M280_IG01:  ;; offset=0x0000
       sub      rsp, 56
						;; size=4 bbWeight=1 PerfScore 0.25

G_M280_IG02:  ;; offset=0x0004
       vxorps   ymm0, ymm0, ymm0
       vmovdqu  ymmword ptr [rsp], ymm0
       vmovdqu  ymmword ptr [rsp+0x18], ymm0
       mov      rax, 0x7FFE555DB870      ; int[]
       mov      qword ptr [rsp], rax
       lea      rax, [rsp]
       mov      dword ptr [rax+0x08], 10
       lea      rax, [rsp]
       mov      dword ptr [rax+0x1C], 33
						;; size=51 bbWeight=1 PerfScore 6.58

G_M280_IG03:  ;; offset=0x0037
       add      rsp, 56
       ret      
						;; size=5 bbWeight=1 PerfScore 1.25
; Total bytes of code: 60

Here we only write to the array storage.

• Heap allocated objects that end up not escaping, as in the example from Fold GDV checks for inlined delegates #84872

G_M43202_IG01:  ;; offset=0x0000
       push     rbx
       sub      rsp, 32
       mov      ebx, ecx
						;; size=7 bbWeight=1 PerfScore 1.50
G_M43202_IG02:  ;; offset=0x0007
       mov      rcx, 0x7FFE61E5D800      ; Program+<>c__DisplayClass0_0
       call     CORINFO_HELP_NEWSFAST
       mov      dword ptr [rax+0x08], ebx
       mov      eax, dword ptr [rax+0x08]
       add      eax, 101
						;; size=24 bbWeight=1 PerfScore 4.50
G_M43202_IG03:  ;; offset=0x001F
       add      rsp, 32
       pop      rbx
       ret  

These may appear to be escaped during escape analysis, but subsequent optimization may lead to them not escaping.

[AndyAyersMS]

Doesn't look like this will happen in .NET 10, so will move to future.