Side-effectful foreach callback mysteriously eliminated on nightly (miscompilation?) #48807
Description
I discovered this while trying to determine why Flux's nightly CI was failing. Have only managed to reduce it to the following:
using ChainRulesCore
function mwe(x, v::Val{S}) where {S}
function inner(ys)
l = size(x, 1)
dx = fill!(similar(x, float(eltype(x)), l * S, size(x, 2)), false)
gs = ntuple(v) do n
@view dx[begin+(l*(n-1)):(l*n), :]
end
foreach(gs, ys) do g, y
y isa AbstractZero && return
copy!(g, y)
end
# Works
# for (g, y) in zip(gs, ys)
# y isa AbstractZero && continue
# copy!(g, y)
# end
return dx
end
inner
end
x = zeros(2, 3)
f = mwe(x, Val(3))
ys = (a=ones(2, 3), b=ZeroTangent(), c=ZeroTangent())
f(ys)Expected output:
6×3 Matrix{Float64}:
1.0 1.0 1.0
1.0 1.0 1.0
0.0 0.0 0.0
0.0 0.0 0.0
0.0 0.0 0.0
0.0 0.0 0.0
Actual output:
6×3 Matrix{Float64}:
0.0 0.0 0.0
0.0 0.0 0.0
0.0 0.0 0.0
0.0 0.0 0.0
0.0 0.0 0.0
0.0 0.0 0.0
What makes this mysterious is what seemingly minor changes can restore the expected behaviour:
- Using the loop instead of
foreach - Defining a custom abstract type + 2 zero-sized struct subtypes mirroring ChainRulesCore's
AbstractZero+NoTangent/ZeroTangent. This only seems to work with the CRC types. - Using a Vector instead of a NamedTuple. Note that the original example used
ChainRulesCore.Tangent, but I'm not sure if that indicates anything. - Adding
Core.donotdelete(false)in theforeachcallback. This suggests to me that the call to said callback is being completely eliminated somehow. Unfortunately Cthulhu won't let me descend into theCore._apply_iterate(Base.iterate, f, z)call that would actually call said callback, so the best I have to go on here is the following remarks:
Call result type was widened because the return value is unused [constprop]
Disabled by argument and rettype heuristics [constprop] Disabled by entry heuristic (unimprovable result)
Versioninfo:
Julia Version 1.10.0-DEV.661
Commit 0b8e8fc216a (2023-02-25 23:01 UTC)
Platform Info:
OS: Linux (x86_64-linux-gnu)
CPU: 8 × Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz
WORD_SIZE: 64
LIBM: libopenlibm
LLVM: libLLVM-14.0.6 (ORCJIT, haswell)
Threads: 1 on 8 virtual cores