Broadcasted in-place assignment of matrix to subarray 3x slower in 1.11-alpha1 over 1.10 and nightly

[BioTurboNick]

Discovered a slowdown in 1.11-alpha when copying an array to a subarray.

Since it seems nightly doesn't have it it's possible this is fixed and already marked for backport but I haven't found it - apologies for the noise if so.

using BenchmarkTools

A = rand(1000,1000);
B = zeros(1001,1001);

@btime $B[1:end - 1, 1:end - 1] .= $A;
# Julia 1.10.2: 526.900 μs
# Julia 1.11-alpha1: 1.560 ms
# Julia 1.12-DEV.136: 532.800 μs

[giordano]

Duplicate of #53158?

[BioTurboNick]

Ah, I think it is. Thanks. EDIT: I'll take my hands off for the moment, didn't want to close prematurely.

[jishnub]

There's also #53430

[abraemer]

Bisection shows that the regression was introduced by 9aa7980

I'll bisect again to find the fix as well and report back in a couple of hours.

Bisect script

#!/bin/bash
#=
make -s clean >> ~/buildlogs_julia.txt 2>&1
make -s -C deps uninstall >> ~/buildlogs_julia.txt 2>&1
if make -s -j 8 >> ~/buildlogs_julia.txt 2>&1 ; then
	echo "build success"
else
	echo "build failure"
	exit 125
fi
exec ~/julia/bisect/julia/julia --startup=no $0
#exec julia --startup=no $0 # test case v1.10.0
=#
#import Pkg; Pkg.add("BenchmarkTools")
using BenchmarkTools

A = rand(1000,1000);
B = zeros(1001,1001);

b = @benchmark $B[1:end - 1, 1:end - 1] .= $A;
time = minimum(b.times)
@info "" time
if time > 1.5e6 # local runtime is 1e6 before regression
	@info "BAD COMMIT (performance difference)"
	exit(1) # bad commit
else
	@info "GOOD COMMIT (no performance difference)"
	exit(0) # good commit
end

[giordano]

Bisection shows that the regression was introduced by 9aa7980

...which is the same as #53158 (comment), which suggests again this is a duplicate of that ticket.

[abraemer]

Performance for this specific example was fixed by 1a90409 by @jishnub

So this is likely a duplicate of either #40962 or #53430

Bisect script

#!/bin/bash
#=
make -s clean >> ~/buildlogs_julia.txt 2>&1
make -s -C deps uninstall >> ~/buildlogs_julia.txt 2>&1
if make -s -j 8 >> ~/buildlogs_julia.txt 2>&1 ; then
	echo "build success"
else
	echo "build failure"
	exit 125
fi
exec ~/julia/bisect/julia/julia --startup=no $0
=#
#import Pkg; Pkg.add("BenchmarkTools")
using BenchmarkTools

A = rand(1000,1000);
B = zeros(1001,1001);

b = @benchmark $B[1:end - 1, 1:end - 1] .= $A;
time = minimum(b.times)
@info "" time
# we search for the commit that fixes the performance
# so in git bisect lingo:
# performance difference -> normal -> good
# no performance difference -> "bug" -> bad
if time > 1.5e6
	@info "GOOD COMMIT (performance difference)"
	exit(0) # good commit
else
	@info "BAD COMMIT (no performance difference)"
	exit(1) # bad commit
end