[IL Emit] Improve analysis to reduce emit of tail.

[En3Tho]

https://github.com/En3Tho/RandomBenchmarks/blob/master/src/RandomBenchmarks/FSharpBenchmarks/UnexpectedTailCalls.fs - here is the benchmark with repro

There are 2 GSeq modules: first is in file with benchmark and second is in another project, but they basically have same code inside them

For some reason F# compiler emits no tailcalls (as expected) when calling GSeq functions from the same file
But it emits .tailcall prefix when calling same GSeq functions from another project

Examples:

.method public static int32
    noTailCalls(
      class [System.Collections]System.Collections.Generic.List`1<int32> list
    ) cil managed
  {
    .maxstack 8

    // [29 8 - 29 17]
    IL_0000: ldc.i4.0
    IL_0001: ldsfld       class Tailcalls/noTailCalls@29 Tailcalls/noTailCalls@29::@_instance
    IL_0006: ldarg.0      // list
    IL_0007: callvirt     instance valuetype [System.Collections]System.Collections.Generic.List`1/Enumerator<!0/*int32*/> class [System.Collections]System.Collections.Generic.List`1<int32>::GetEnumerator()
    IL_000c: call         !!0/*int32*/ Tailcalls/GSeq::fold<int32, int32, valuetype [System.Collections]System.Collections.Generic.List`1/Enumerator<int32>>(!!0/*int32*/, class [FSharp.Core]Microsoft.FSharp.Core.FSharpFunc`2<!!0/*int32*/, class [FSharp.Core]Microsoft.FSharp.Core.FSharpFunc`2<!!1/*int32*/, !!0/*int32*/>>, !!2/*valuetype [System.Collections]System.Collections.Generic.List`1/Enumerator<int32>*/)
    IL_0011: ret

  } // end of method Tailcalls::noTailCalls

.method public static int32
    whyTailCalls(
      class [System.Collections]System.Collections.Generic.List`1<int32> list
    ) cil managed
  {
    .maxstack 8

    // [34 8 - 34 47]
    IL_0000: ldc.i4.0
    IL_0001: ldsfld       class Tailcalls/whyTailCalls@34 Tailcalls/whyTailCalls@34::@_instance
    IL_0006: ldarg.0      // list
    IL_0007: callvirt     instance valuetype [System.Collections]System.Collections.Generic.List`1/Enumerator<!0/*int32*/> class [System.Collections]System.Collections.Generic.List`1<int32>::GetEnumerator()
    IL_000c: tail.
    IL_000e: call         !!0/*int32*/ [Lib.FSharp]Lib.FSharp.GenericEnumerators.GSeq::fold<int32, int32, valuetype [System.Collections]System.Collections.Generic.List`1/Enumerator<int32>>(!!0/*int32*/, class [FSharp.Core]Microsoft.FSharp.Core.FSharpFunc`2<!!0/*int32*/, class [FSharp.Core]Microsoft.FSharp.Core.FSharpFunc`2<!!1/*int32*/, !!0/*int32*/>>, !!2/*valuetype [System.Collections]System.Collections.Generic.List`1/Enumerator<int32>*/)
    IL_0013: ret

  } // end of method Tailcalls::whyTailCalls

Unexpected tailcall makes function run 2-3 times slower and blows asm size 2x because of additional tailcall dispatch helpers being generated by jit

[dsyme]

For some reason F# compiler emits no tailcalls (as expected) when calling GSeq functions from the same file

But it emits .tailcall prefix when calling same GSeq functions from another project

I believe this is by design. There are cases where F# decides to omit .tail because it does an analysis that this is not needed. This analysis may not always be available across assembly boundaries.

[En3Tho]

Method	Job	EnvironmentVariables	Count	Mean	Error	StdDev	Code Size	Allocated
ListGSeqFoldNoTailCall	DynamicPGO	DOTNET_TC_QuickJitForLoop=1	100	201.1 ns	1.78 ns	1.49 ns	264 B	-
ListGSeqTailCall	DynamicPGO	DOTNET_TC_QuickJitForLoop=1	100	281.3 ns	5.61 ns	5.51 ns	823 B	-
ListCSharpFold	DynamicPGO	DOTNET_TC_QuickJitForLoop=1	100	181.4 ns	3.64 ns	3.23 ns	264 B	-

[En3Tho]

Here is a simple benchmark showcasing that when .tail analysis fail, we potentially take a huge hit in asm size and perf. I don't know which functions take this hit but the amount might be huge. The only difference is this:

// GSeq from this assembly
[<Benchmark>]
member this.ListGSeqFoldNoTailCall() =
        this.List
        |> GSeq.getEnumerator
        |> GSeq.fold 0 ^ fun x y -> x + y

// GSeq from another assembly
[<Benchmark>]
member this.ListGSeqTailCall() =
        this.List
        |> GSeq.getEnumerator
        |> Lib.FSharp.GenericEnumerators.GSeq.fold 0 ^ fun x y -> x + y

I believe that if it's possible, we should improve .tail analysis.