Extraneous temporary and copy in implicit byref parameter

[asik]

Using the tuple form of a call that has a byref parameter, if the surrounding code is sufficiently complex, the compiler emits an extraneous temporary and copy.

open System
open System.Collections.Generic

[<EntryPoint>]
let main argv = 
    let dict = Dictionary<int, DateTime>()
    let result, elem = dict.TryGetValue(0)
    let b = 
        if result then
            if elem.Day > 0 then
                Console.WriteLine ""
                true
            else false
        else false
    0

Expected behavior

Only one local is generated and no extra copy happens, i.e. generated code is equivalent to:

    let mutable V_1 = DateTime()
    let result = dict.TryGetValue(0, &V_1)
    let b = 
        if result then
            if V_1.Day > 0 then ...

Actual behavior

Looking at the IL, this generates 2 DateTime locals:

       [1] valuetype [mscorlib]System.DateTime V_1,
       [2] valuetype [mscorlib]System.DateTime V_2,

V_1 is passed into the call, then V_1 is copied into V_2, then V_2 is used. In effect this is equivalent to the following code:

    let mutable V_1 = DateTime()
    let result = dict.TryGetValue(0, &V_1)
    let V_2 = V_1
    let b = 
        if result then
            if V_2.Day > 0 then ...

I tried simplifying this example as much as possible, but at this point removing anything results in good generated code, without the copy, so it seems to depend a lot on the surrounding logic.

This makes a big difference in code that heavily relies on passing large structs by reference, e.g. System.Numerics.Vectors.Matrix4x4 etc; it wastes stack space and time copying. The JIT doesn't optimize this away.

Known workarounds

Don't use the tupled form, use a mutable local and pass it explicitly.

Related information

Visual Studio 2017 15.2 (26430.6).

[dsyme]

@asik If you get a chance to try out #5110 on some performance benchmarks for System.Numerics.Vectors.Matrix4x4 etc. that would be great

[cartermp]

Closing this, as #5110 is merged.