PGO: don't instrument edges for known cold places

[EgorBo]

static void Foo(object o)
{
    if (o == null)
        throw new ArgumentNullException();

    // other code
}

; Method Prog:Foo(System.Object)
G_M6323_IG01:
       push     rbp
       sub      rsp, 48
       lea      rbp, [rsp+30H]
       xor      eax, eax
       mov      qword ptr [rbp-08H], rax
       mov      gword ptr [rbp+10H], rcx
       cmp      gword ptr [rbp+10H], 0
       jne      SHORT G_M6323_IG04
G_M6323_IG03:
       inc      dword ptr [(reloc)]    ;;; <------------
       mov      rcx, 0xD1FFAB1E
       call     CORINFO_HELP_NEWSFAST
       mov      gword ptr [rbp-08H], rax
       mov      rcx, gword ptr [rbp-08H]
       call     [System.ArgumentNullException:.ctor():this]
       mov      rcx, gword ptr [rbp-08H]
       call     CORINFO_HELP_THROW
G_M6323_IG04:
       inc      dword ptr [(reloc)]    ;;; <------------
       ...

Currently, we emit 2 edge counters for such patterns. If we assume that throw new is always cold we can skip the other edge too and then re-construct missing weights from other blocks

cc @AndyAyersMS the idea we discussed last week

If the throw new is actually hot then 🤷‍♂️ presumably savings/bad decisions from PGO will be not noticeable compared to EH

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

Issue Details

---

static void Foo(object o)
{
    if (o == null)
        throw new ArgumentNullException();

    // other code
}

; Method Prog:Foo(System.Object)
G_M6323_IG01:
       push     rbp
       sub      rsp, 48
       lea      rbp, [rsp+30H]
       xor      eax, eax
       mov      qword ptr [rbp-08H], rax
       mov      gword ptr [rbp+10H], rcx
       cmp      gword ptr [rbp+10H], 0
       jne      SHORT G_M6323_IG04
G_M6323_IG03:
       inc      dword ptr [(reloc)]    ;;; <------------
       mov      rcx, 0xD1FFAB1E
       call     CORINFO_HELP_NEWSFAST
       mov      gword ptr [rbp-08H], rax
       mov      rcx, gword ptr [rbp-08H]
       call     [System.ArgumentNullException:.ctor():this]
       mov      rcx, gword ptr [rbp-08H]
       call     CORINFO_HELP_THROW
G_M6323_IG04:
       inc      dword ptr [(reloc)]    ;;; <------------
       ...

Currently, we emit 2 edge counters for such patterns. If we assume that throw new is always cold we can skip the other edge too and then re-construct missing weights from other blocks

cc @AndyAyersMS the idea we discussed last week

Author:	EgorBo
Assignees:	-
Labels:	area-CodeGen-coreclr, untriaged
Milestone:	-

[AndyAyersMS]

Not sure what impact this would have. Those extra probes should be cold too, and unless exceptions are the only branching cases we'll still have a similar number of probes in the hot parts of the method.

[EgorBo]

Not sure what impact this would have. Those extra probes should be cold too, and unless exceptions are the only branching cases we'll still have a similar number of probes in the hot parts of the method.

We have quite a few ThrowHelpers with conditions inside - those get instrumented (since aren't inlined) and we hit counters in them in Tier0

[AndyAyersMS]

Not sure what impact this would have. Those extra probes should be cold too, and unless exceptions are the only branching cases we'll still have a similar number of probes in the hot parts of the method.

We have quite a few ThrowHelpers with conditions inside - those get instrumented (since aren't inlined) and we hit counters in them in Tier0

That could be. It should be easy enough to experiment with leaving those probes out (and hence possibly leaving out other probes if the method ends up being minimal) -- we take pains to make sure we are instrumenting the throws. The tricky part is to ensure reconstruction doesn't get messed up.

I'll set something up.

[AndyAyersMS]

#81972 does this for explicit throws.

Calls to throw helpers will be harder to spot, because at the time we make the instrumentation plan we haven't resolved tokens.