Benchmarkdotnet: Add annotations for diassembler output

[adamsitnik]

From private conversation with @JosephTremoulet

trying to replicate (in some fashion) the various annotations we report in the checked-jit-produced disasm (symbol resolution would be a good start, resolving helper tokens to their symbolic names also important, structured reporting of GC and EH tables, compilation flags, local variable table…) – open question how you’d reconstruct this information after-the-fact (debug info? event [ETW] streams? maybe new ones that we’d add?), where the code should live (e.g. a library that something like PerfView could also tap into would be cool, maybe a lot of this functionality belongs in CLRMD), etc. Everything we’ve put in the COMPlus_JitDisasm output is something we’ve put there because it’s useful to be able to understand and diff the machine code

[GeorgePlotnikov]

hi @adamsitnik , is this issue still actual? may I grab it? If yes, I would be really appreciated you hinted me where the best point to start.

[adamsitnik]

hi @GeorgePlotnikov !

yes, it's still up for grabs. But the feature itself is rather very BIG.

Personally I would start with some research of what COMPlus_JitDisasm does in Core CLR. Once you know what they are doing and make sure that the data is available outside of CLR itself I would start implementing the features one by one.

@JosephTremoulet do you have any hints? Is there somebody that could point us to the right place?

[JosephTremoulet]

If you want to look at the relevant source code, search in the jit directory for opts.disAsm.

It may be easier to start with examples. If you build a "checked" or "debug" local build of the CoreCLR runtime, and run some code against that with the COMPlus_JitDisasm environment variable set to *, it will emit assembly listings for every method that gets jitted.

[GeorgePlotnikov]

@adamsitnik, @JosephTremoulet great! thanks a lot!

[adamsitnik]

@AndreyAkinshin could you assign @GeorgePlotnikov to this issue?

[AndreyAkinshin]

@GeorgePlotnikov, I added you to the collaborators list. Please, accept the invitation, then I will be able to assign you.

[GeorgePlotnikov]

@AndreyAkinshin done. Thanks.

[AndreyAkinshin]

@GeorgePlotnikov, assigned.

[GeorgePlotnikov]

small update to track this issue: I'm investigating the coreclr disasm output to compare vs BDN and to figure out how it should look like.

[adamsitnik]

@GeorgePlotnikov great, thanks for the update!

[GeorgePlotnikov]

I want to clarify if I'm understanding clearly. Correct me if I'm wrong.

Small example:

c# code

public string Calc()
{
        return "result: " + WithoutStarg(0x11) + WithStarg(0x12);
}

produces via JitDisasm annotation below

; Total bytes of code 23, prolog size 4 for method jitdumptest.Program:Main(ref)
; ============================================================
; Assembly listing for method jitdumptest.Program:Calc():ref
; Emitting BLENDED_CODE for X64 CPU with AVX
; optimized code
; rsp based frame
; partially interruptible
; Final local variable assignments
;
;  V00 tmp0         [V00,T00] (  4,  8   )     ref  ->  rsi         class-hnd exact
;  V01 tmp1         [V01,T01] (  4,  8   )     ref  ->  rdi         class-hnd exact
;  V02 OutArgs      [V02    ] (  1,  1   )  lclBlk (32) [rsp+0x00]
;
; Lcl frame size = 40

that's the goal?

[adamsitnik]

@JosephTremoulet could you give some feedback to @GeorgePlotnikov ?

[JosephTremoulet]

Yes, as much of that header as can be recreated would be helpful, though I don't know what would still be available on the runtime side in the MethodDesc or GC info or debug info.

Aside from the header, there should have been a dump of the actual disassembly that followed, which likewise may have some useful comments, and may be using symbolic names more than BDN currently is.

[GeorgePlotnikov]

@adamsitnik, @JosephTremoulet thank you for replies, will discover the solution

[adamsitnik]

hello @JosephTremoulet

Could you tell us what are the most important things that we should reproduce in BDN?

@GeorgePlotnikov has started working on this issue (draft of the PR #635 ) and we need to know if we can get the most important data

[GeorgePlotnikov]

Hi @adamsitnik, @AndreyAkinshin did you see the results, what do you think about it?

[adamsitnik]

No progress since 2017, we have also switched to Iced for instructions decoding (which might add similar feature in the future) so I am closing the issue