Regression: mul! of views allocates in v1.9+ for arrays of 4 or higher dimension, not in v1.8.5

[pablosanjose]

This seems like a regression: in Julia v1.9-rc2 and master

julia> using LinearAlgebra, BenchmarkTools

julia> A = rand(ComplexF64,4,4,1000,1000);

julia> B = similar(A);

julia> a,b = (view(B,:,:,1,1),view(A,:,:,1,1));

julia> @btime mul!($b,$a,$a);
  321.792 ns (10 allocations: 608 bytes)

This is on an M1 Max

julia> versioninfo()
Julia Version 1.9.0-rc2
Commit 72aec423c2a (2023-04-01 10:41 UTC)
Platform Info:
  OS: macOS (arm64-apple-darwin21.3.0)
  CPU: 10 × Apple M1 Max
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-14.0.6 (ORCJIT, apple-m1)
  Threads: 1 on 8 virtual cores
Environment:
  JULIA_EDITOR = code

But in v1.8.5 we don't allocate

julia> using LinearAlgebra, BenchmarkTools

julia> A = rand(ComplexF64,4,4,1000,1000);

julia> B = similar(A);

julia> a,b = (view(B,:,:,1,1),view(A,:,:,1,1));

julia> @btime mul!($b,$a,$a);
  97.076 ns (0 allocations: 0 bytes)

Note that if A has dimensions 3 or less, this issue does not arise.

[KristofferC]

I haven't benchmarked this but one difference is the following:

julia> A = rand(ComplexF64,4,4,1000,1000);

julia> B = similar(A);

julia> a,b = (view(B,:,:,1,1),view(A,:,:,1,1));

julia> f(A,B,C) = Base.require_one_based_indexing(A,B,C)
f (generic function with 1 method)

julia> @code_warntype optimize=true f(a, a, b)

This used to constant fold in 1.8 to true but on 1.9 it calls:

1 ─ %1  = invoke Base._any_tuple(Base.has_offset_axes::typeof(Base.has_offset_axes), false::Bool, A::SubArray{ComplexF64, 2, Array{ComplexF64, 4}, ...

[KristofferC]

Seems to be f718238 cc @vtjnash

(but again, I have not confirmed that this is the actual cause of the speed or allocation regression here, just the change in inlining)

Seems to indeed be where the regression comes from:

Before:

  185.500 ns (0 allocations: 0 bytes)

After:

 509.057 ns (10 allocations: 608 bytes)

[pablosanjose]

Is it clear to anyone what would be the good strategy to fix this?

[pablosanjose]

Friendly bump, in case anyone with some available time knows what could be done here.

This is a fairly severe regression that has our simulation code stuck on 1.8.5 because we need views of 5-dimensional arrays, and the runtime penalty is too large. With current nightly it's become even worse:

julia> @btime mul!($b,$a,$a);
  493.124 ns (10 allocations: 608 bytes)

[pablosanjose]

It seems the root cause of this is the following

julia> A = view(rand(ComplexF64, 4, 4, 4, 4), :, :, 1, 1);

julia> @allocations Base.require_one_based_indexing(A,A,A)  # second run
11

This should be 1. So require_one_based_indexing is sometimes not elided as expected. The PR #45260 turns out to be insufficient for this particular combination of views of high-dimensional arrays.