in-place broadcast (e.g. `.+=`) significanctly slower for reinterpreted array

[Moelf]

on 1.9-beta4 and master branch (0b8e8fc):

julia> using BenchmarkTools

julia> @btime ary.+=Int32(1) setup = (ary = rand(Int32, 10^5));
  3.389 μs (0 allocations: 0 bytes)

julia> @btime ary.+=Int32(1) setup = (ary = reinterpret(Int32, rand(UInt8, 10^5*4)));
  31.680 μs (0 allocations: 0 bytes)

not sure if it's just another facet of:

• ntoh / bswap are 10x slower when operating in-place on reinterpret array #42227

[Moelf]

a possible lead (#43153 (comment)):

julia> ary = reinterpret(Int32, rand(UInt8, 10^5*4))

julia> f(ary) = ary .+= Int32(1);

julia> @code_llvm f(ary)

....
vector.memcheck92:                                ; preds = %pass19.lr.ph.split.us
  %scevgep93 = getelementptr i8, i8* %12, i64 4
  %bound097 = icmp ult i8* %12, %scevgep93
  br i1 %bound097, label %scalar.ph101, label %vector.ph104
...

although I don't see the conflict flag variable

[Moelf]

#44186 doesn't help

[N5N3]

IIRC, LLVM's vectorizer would insert an overly strict memcheck unless it can prove dest === src.
Unfortunately, our broadcast_unalias mechanism fools the complier.
Since #44186 doesn't help, we still need ivdep to force LLVM skipping the unneeded memcheck.

[Moelf]

Since #44186 doesn't help, we still need ivdep to force LLVM skipping the unneeded memcheck.

what does "need" mean (where should this happen?) I would love to but I'm not sure I know enough about LLVM to help

[N5N3]

I mean adding @simd ivdep into broadcast kernal. It's the simplest but unsafe fix. Such change might be safer once we have compile-time infer_effect, we can add ivdep once we confirm that the broadcast function is effect free and the inputs contains no memory overlapping.