Regression in memory allocs when forming NTuple

[jw3126]

The following code allocates on master and 1.7beta2, but does not allocate on 1.6:

Tuple3(itr) = NTuple{3, Float64}(itr)::NTuple{3,Float64}

function doit(itr)
    for x in itr
        Tuple3(x)
    end
end

itr = [randn(3) for _ in 1:10000]
doit(itr)
@time doit(itr)

master:  0.021418 seconds (120.00 k allocations: 2.899 MiB)
julia 1.6:  0.000032 seconds

This lead to regressions in two packages for me:
https://github.com/jw3126/VoxelRayTracers.jl/blob/master/src/core.jl#L71
https://github.com/jw3126/LinearInterpolations.jl/blob/master/src/LinearInterpolations.jl#L407
https://github.com/jw3126/LinearInterpolations.jl/blob/master/test/test_LinearInterpolations.jl#L259

[simeonschaub]

Fixed in #41074, but looks like we forgot to mark that for backporting.

[jw3126]

@simeonschaub I could reproduce with today's master. Note that type inference works, but still it allocates for me. Could you run my example just to double check?

[jw3126]

@simeonschaub this is how it looks on today's master for me:

   _       _ _(_)_     |  Documentation: https://docs.julialang.org
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  Type "?" for help, "]?" for Pkg help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 1.8.0-DEV.155 (2021-07-07)
 _/ |\__'_|_|_|\__'_|  |  Commit 97f817a379* (0 days old master)
|__/                   |

julia> Tuple3(itr) = NTuple{3, Float64}(itr)::NTuple{3,Float64}
Tuple3 (generic function with 1 method)

julia> function doit(itr);
           for x in itr
               Tuple3(x)
           end
       end
doit (generic function with 1 method)

julia> itr = [randn(3) for _ in 1:10000];doit(itr);

julia> @time doit(itr)
  0.023238 seconds (120.00 k allocations: 2.899 MiB)

[simeonschaub]

Ah sorry, I misunderstood. Would be good to fix this.

[JeffBezanson]

Base._totuple was inlined in 1.6, but isn't anymore. The cost is fine, and the method is even marked inline, so I guess it might be due to recursion detection?